program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}})] { func main(tensor cache_length, tensor decoder_key_padding_mask, tensor encoder_output_embeds, tensor input_ids, tensor key_cache, tensor kv_cache_update_mask, tensor value_cache) { tensor var_80_axis_0 = const()[name = tensor("op_80_axis_0"), val = tensor(0)]; tensor var_80_batch_dims_0 = const()[name = tensor("op_80_batch_dims_0"), val = tensor(0)]; tensor embed_tokens_weight_to_fp16 = const()[name = tensor("embed_tokens_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; tensor var_80_cast_fp16 = gather(axis = var_80_axis_0, batch_dims = var_80_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor("op_80_cast_fp16")]; tensor var_84_axis_0 = const()[name = tensor("op_84_axis_0"), val = tensor(0)]; tensor var_84_batch_dims_0 = const()[name = tensor("op_84_batch_dims_0"), val = tensor(0)]; tensor embed_positions_weight_to_fp16 = const()[name = tensor("embed_positions_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132777088)))]; tensor var_84_cast_fp16 = gather(axis = var_84_axis_0, batch_dims = var_84_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor("op_84_cast_fp16")]; tensor hidden_states_1_cast_fp16 = add(x = var_80_cast_fp16, y = var_84_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; tensor var_98_axes_0 = const()[name = tensor("op_98_axes_0"), val = tensor([2])]; tensor var_98_cast_fp16 = expand_dims(axes = var_98_axes_0, x = hidden_states_1_cast_fp16)[name = tensor("op_98_cast_fp16")]; tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([3])]; tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_98_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280])]; tensor var_103_axis_0 = const()[name = tensor("op_103_axis_0"), val = tensor(1)]; tensor var_103_cast_fp16_0, tensor var_103_cast_fp16_1, tensor var_103_cast_fp16_2, tensor var_103_cast_fp16_3, tensor var_103_cast_fp16_4, tensor var_103_cast_fp16_5, tensor var_103_cast_fp16_6, tensor var_103_cast_fp16_7, tensor var_103_cast_fp16_8, tensor var_103_cast_fp16_9, tensor var_103_cast_fp16_10, tensor var_103_cast_fp16_11, tensor var_103_cast_fp16_12, tensor var_103_cast_fp16_13, tensor var_103_cast_fp16_14, tensor var_103_cast_fp16_15, tensor var_103_cast_fp16_16, tensor var_103_cast_fp16_17, tensor var_103_cast_fp16_18, tensor var_103_cast_fp16_19, tensor var_103_cast_fp16_20, tensor var_103_cast_fp16_21, tensor var_103_cast_fp16_22, tensor var_103_cast_fp16_23, tensor var_103_cast_fp16_24, tensor var_103_cast_fp16_25, tensor var_103_cast_fp16_26, tensor var_103_cast_fp16_27, tensor var_103_cast_fp16_28, tensor var_103_cast_fp16_29, tensor var_103_cast_fp16_30, tensor var_103_cast_fp16_31 = split(axis = var_103_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor("op_103_cast_fp16")]; tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280])]; tensor var_138_axis_0 = const()[name = tensor("op_138_axis_0"), val = tensor(1)]; tensor var_138_cast_fp16_0, tensor var_138_cast_fp16_1, tensor var_138_cast_fp16_2, tensor var_138_cast_fp16_3, tensor var_138_cast_fp16_4, tensor var_138_cast_fp16_5, tensor var_138_cast_fp16_6, tensor var_138_cast_fp16_7, tensor var_138_cast_fp16_8, tensor var_138_cast_fp16_9, tensor var_138_cast_fp16_10, tensor var_138_cast_fp16_11, tensor var_138_cast_fp16_12, tensor var_138_cast_fp16_13, tensor var_138_cast_fp16_14, tensor var_138_cast_fp16_15, tensor var_138_cast_fp16_16, tensor var_138_cast_fp16_17, tensor var_138_cast_fp16_18, tensor var_138_cast_fp16_19, tensor var_138_cast_fp16_20, tensor var_138_cast_fp16_21, tensor var_138_cast_fp16_22, tensor var_138_cast_fp16_23, tensor var_138_cast_fp16_24, tensor var_138_cast_fp16_25, tensor var_138_cast_fp16_26, tensor var_138_cast_fp16_27, tensor var_138_cast_fp16_28, tensor var_138_cast_fp16_29, tensor var_138_cast_fp16_30, tensor var_138_cast_fp16_31 = split(axis = var_138_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor("op_138_cast_fp16")]; tensor var_176 = const()[name = tensor("op_176"), val = tensor(3)]; tensor var_183 = const()[name = tensor("op_183"), val = tensor(1)]; tensor var_184 = const()[name = tensor("op_184"), val = tensor(true)]; tensor var_196 = const()[name = tensor("op_196"), val = tensor([1])]; tensor channels_mean_1_cast_fp16 = reduce_mean(axes = var_196, keep_dims = var_184, x = inputs_1_cast_fp16)[name = tensor("channels_mean_1_cast_fp16")]; tensor zero_mean_1_cast_fp16 = sub(x = inputs_1_cast_fp16, y = channels_mean_1_cast_fp16)[name = tensor("zero_mean_1_cast_fp16")]; tensor zero_mean_sq_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = zero_mean_1_cast_fp16)[name = tensor("zero_mean_sq_1_cast_fp16")]; tensor var_200 = const()[name = tensor("op_200"), val = tensor([1])]; tensor var_201_cast_fp16 = reduce_mean(axes = var_200, keep_dims = var_184, x = zero_mean_sq_1_cast_fp16)[name = tensor("op_201_cast_fp16")]; tensor var_202_to_fp16 = const()[name = tensor("op_202_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_203_cast_fp16 = add(x = var_201_cast_fp16, y = var_202_to_fp16)[name = tensor("op_203_cast_fp16")]; tensor denom_1_epsilon_0_to_fp16 = const()[name = tensor("denom_1_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_1_cast_fp16 = rsqrt(epsilon = denom_1_epsilon_0_to_fp16, x = var_203_cast_fp16)[name = tensor("denom_1_cast_fp16")]; tensor out_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = denom_1_cast_fp16)[name = tensor("out_1_cast_fp16")]; tensor obj_1_mean_0_to_fp16 = const()[name = tensor("obj_1_mean_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133924032)))]; tensor obj_1_variance_0_to_fp16 = const()[name = tensor("obj_1_variance_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133926656)))]; tensor obj_1_gamma_0_to_fp16 = const()[name = tensor("obj_1_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133929280)))]; tensor obj_1_beta_0_to_fp16 = const()[name = tensor("obj_1_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133931904)))]; tensor obj_1_epsilon_0_to_fp16 = const()[name = tensor("obj_1_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor("obj_1_cast_fp16")]; tensor layers_0_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133934528)))]; tensor input_1_cast_fp16 = sub(x = obj_1_cast_fp16, y = layers_0_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_1_cast_fp16")]; tensor var_222 = const()[name = tensor("op_222"), val = tensor([1, 1])]; tensor var_224 = const()[name = tensor("op_224"), val = tensor([1, 1])]; tensor x_1_pad_type_0 = const()[name = tensor("x_1_pad_type_0"), val = tensor("custom")]; tensor x_1_pad_0 = const()[name = tensor("x_1_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133937152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134756416))), name = tensor("layers_0_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_0_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134756544)))]; tensor x_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_module_bias_to_fp16, dilations = var_224, groups = var_183, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = var_222, weight = layers_0_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; tensor layers_0_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134759168)))]; tensor query_1_cast_fp16 = mul(x = x_1_cast_fp16, y = layers_0_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_1_cast_fp16")]; tensor var_234 = const()[name = tensor("op_234"), val = tensor([1, 1])]; tensor var_236 = const()[name = tensor("op_236"), val = tensor([1, 1])]; tensor x_3_pad_type_0 = const()[name = tensor("x_3_pad_type_0"), val = tensor("custom")]; tensor x_3_pad_0 = const()[name = tensor("x_3_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134761792))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135581056))), name = tensor("layers_0_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_0_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135581184)))]; tensor x_3_cast_fp16 = conv(bias = layers_0_self_attn_k_proj_module_bias_to_fp16, dilations = var_236, groups = var_183, pad = x_3_pad_0, pad_type = x_3_pad_type_0, strides = var_234, weight = layers_0_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; tensor layers_0_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_0_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135583808)))]; tensor current_key_1_cast_fp16 = mul(x = x_3_cast_fp16, y = layers_0_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_1_cast_fp16")]; tensor var_246 = const()[name = tensor("op_246"), val = tensor([1, 1])]; tensor var_248 = const()[name = tensor("op_248"), val = tensor([1, 1])]; tensor x_5_pad_type_0 = const()[name = tensor("x_5_pad_type_0"), val = tensor("custom")]; tensor x_5_pad_0 = const()[name = tensor("x_5_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135586432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136405696))), name = tensor("layers_0_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_0_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136405824)))]; tensor x_5_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_module_bias_to_fp16, dilations = var_248, groups = var_183, pad = x_5_pad_0, pad_type = x_5_pad_type_0, strides = var_246, weight = layers_0_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_1_cast_fp16)[name = tensor("x_5_cast_fp16")]; tensor layers_0_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136408448)))]; tensor current_value_1_cast_fp16 = mul(x = x_5_cast_fp16, y = layers_0_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_1_cast_fp16")]; tensor var_253_axes_0 = const()[name = tensor("op_253_axes_0"), val = tensor([1])]; tensor var_253_cast_fp16 = expand_dims(axes = var_253_axes_0, x = kv_cache_update_mask)[name = tensor("op_253_cast_fp16")]; tensor var_254_axes_0 = const()[name = tensor("op_254_axes_0"), val = tensor([2])]; tensor var_254_cast_fp16 = expand_dims(axes = var_254_axes_0, x = var_253_cast_fp16)[name = tensor("op_254_cast_fp16")]; tensor var_256_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_256_cast_fp16")]; tensor var_177_to_fp16 = const()[name = tensor("op_177_to_fp16"), val = tensor(0x1p+0)]; tensor var_257_cast_fp16 = sub(x = var_177_to_fp16, y = var_254_cast_fp16)[name = tensor("op_257_cast_fp16")]; tensor var_258_cast_fp16 = mul(x = var_103_cast_fp16_0, y = var_257_cast_fp16)[name = tensor("op_258_cast_fp16")]; tensor key_1_cast_fp16 = add(x = var_256_cast_fp16, y = var_258_cast_fp16)[name = tensor("key_1_cast_fp16")]; tensor var_260_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_260_cast_fp16")]; tensor var_262_cast_fp16 = mul(x = var_138_cast_fp16_0, y = var_257_cast_fp16)[name = tensor("op_262_cast_fp16")]; tensor value_1_cast_fp16 = add(x = var_260_cast_fp16, y = var_262_cast_fp16)[name = tensor("value_1_cast_fp16")]; tensor var_265 = const()[name = tensor("op_265"), val = tensor([1, 20, 64, -1])]; tensor var_266_cast_fp16 = reshape(shape = var_265, x = query_1_cast_fp16)[name = tensor("op_266_cast_fp16")]; tensor var_267_to_fp16 = const()[name = tensor("op_267_to_fp16"), val = tensor(0x1p-3)]; tensor var_268_cast_fp16 = mul(x = var_266_cast_fp16, y = var_267_to_fp16)[name = tensor("op_268_cast_fp16")]; tensor var_269 = const()[name = tensor("op_269"), val = tensor([1, 20, 64, -1])]; tensor var_270_cast_fp16 = reshape(shape = var_269, x = key_1_cast_fp16)[name = tensor("op_270_cast_fp16")]; tensor mh_w_1_transpose_x_0 = const()[name = tensor("mh_w_1_transpose_x_0"), val = tensor(true)]; tensor mh_w_1_transpose_y_0 = const()[name = tensor("mh_w_1_transpose_y_0"), val = tensor(false)]; tensor mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_268_cast_fp16, y = var_270_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; tensor var_274_axes_0 = const()[name = tensor("op_274_axes_0"), val = tensor([1])]; tensor var_274_cast_fp16 = expand_dims(axes = var_274_axes_0, x = decoder_key_padding_mask)[name = tensor("op_274_cast_fp16")]; tensor var_275_axes_0 = const()[name = tensor("op_275_axes_0"), val = tensor([2])]; tensor var_275_cast_fp16 = expand_dims(axes = var_275_axes_0, x = var_274_cast_fp16)[name = tensor("op_275_cast_fp16")]; tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; tensor var_278_cast_fp16 = softmax(axis = var_176, x = mh_w_3_cast_fp16)[name = tensor("op_278_cast_fp16")]; tensor var_279 = const()[name = tensor("op_279"), val = tensor([1, 20, 64, -1])]; tensor var_280_cast_fp16 = reshape(shape = var_279, x = value_1_cast_fp16)[name = tensor("op_280_cast_fp16")]; tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(true)]; tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_280_cast_fp16, y = var_278_cast_fp16)[name = tensor("attn_1_cast_fp16")]; tensor var_283 = const()[name = tensor("op_283"), val = tensor([1, 1280, 1, -1])]; tensor x_7_cast_fp16 = reshape(shape = var_283, x = attn_1_cast_fp16)[name = tensor("x_7_cast_fp16")]; tensor layers_0_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136411072)))]; tensor input_7_cast_fp16 = sub(x = x_7_cast_fp16, y = layers_0_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_7_cast_fp16")]; tensor var_291 = const()[name = tensor("op_291"), val = tensor([1, 1])]; tensor var_293 = const()[name = tensor("op_293"), val = tensor([1, 1])]; tensor x_9_pad_type_0 = const()[name = tensor("x_9_pad_type_0"), val = tensor("custom")]; tensor x_9_pad_0 = const()[name = tensor("x_9_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136413696))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137232960))), name = tensor("layers_0_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_0_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137233088)))]; tensor x_9_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_module_bias_to_fp16, dilations = var_293, groups = var_183, pad = x_9_pad_0, pad_type = x_9_pad_type_0, strides = var_291, weight = layers_0_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_7_cast_fp16)[name = tensor("x_9_cast_fp16")]; tensor layers_0_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137235712)))]; tensor obj_7_cast_fp16 = mul(x = x_9_cast_fp16, y = layers_0_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_7_cast_fp16")]; tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; tensor var_304 = const()[name = tensor("op_304"), val = tensor([1])]; tensor channels_mean_3_cast_fp16 = reduce_mean(axes = var_304, keep_dims = var_184, x = inputs_3_cast_fp16)[name = tensor("channels_mean_3_cast_fp16")]; tensor zero_mean_3_cast_fp16 = sub(x = inputs_3_cast_fp16, y = channels_mean_3_cast_fp16)[name = tensor("zero_mean_3_cast_fp16")]; tensor zero_mean_sq_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = zero_mean_3_cast_fp16)[name = tensor("zero_mean_sq_3_cast_fp16")]; tensor var_308 = const()[name = tensor("op_308"), val = tensor([1])]; tensor var_309_cast_fp16 = reduce_mean(axes = var_308, keep_dims = var_184, x = zero_mean_sq_3_cast_fp16)[name = tensor("op_309_cast_fp16")]; tensor var_310_to_fp16 = const()[name = tensor("op_310_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_311_cast_fp16 = add(x = var_309_cast_fp16, y = var_310_to_fp16)[name = tensor("op_311_cast_fp16")]; tensor denom_3_epsilon_0_to_fp16 = const()[name = tensor("denom_3_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_3_cast_fp16 = rsqrt(epsilon = denom_3_epsilon_0_to_fp16, x = var_311_cast_fp16)[name = tensor("denom_3_cast_fp16")]; tensor out_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = denom_3_cast_fp16)[name = tensor("out_3_cast_fp16")]; tensor obj_9_gamma_0_to_fp16 = const()[name = tensor("obj_9_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137238336)))]; tensor obj_9_beta_0_to_fp16 = const()[name = tensor("obj_9_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137240960)))]; tensor obj_9_epsilon_0_to_fp16 = const()[name = tensor("obj_9_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor("obj_9_cast_fp16")]; tensor layers_0_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137243584)))]; tensor input_9_cast_fp16 = sub(x = obj_9_cast_fp16, y = layers_0_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_9_cast_fp16")]; tensor var_330 = const()[name = tensor("op_330"), val = tensor([1, 1])]; tensor var_332 = const()[name = tensor("op_332"), val = tensor([1, 1])]; tensor x_11_pad_type_0 = const()[name = tensor("x_11_pad_type_0"), val = tensor("custom")]; tensor x_11_pad_0 = const()[name = tensor("x_11_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137246208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138065472))), name = tensor("layers_0_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_0_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138065600)))]; tensor x_11_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_332, groups = var_183, pad = x_11_pad_0, pad_type = x_11_pad_type_0, strides = var_330, weight = layers_0_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = tensor("x_11_cast_fp16")]; tensor layers_0_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138068224)))]; tensor query_3_cast_fp16 = mul(x = x_11_cast_fp16, y = layers_0_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_3_cast_fp16")]; tensor layers_0_encoder_attn_k_proj_input_shift_to_fp16 = const()[name = tensor("layers_0_encoder_attn_k_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138070848)))]; tensor input_11_cast_fp16 = sub(x = encoder_output_embeds, y = layers_0_encoder_attn_k_proj_input_shift_to_fp16)[name = tensor("input_11_cast_fp16")]; tensor var_342 = const()[name = tensor("op_342"), val = tensor([1, 1])]; tensor var_344 = const()[name = tensor("op_344"), val = tensor([1, 1])]; tensor x_13_pad_type_0 = const()[name = tensor("x_13_pad_type_0"), val = tensor("custom")]; tensor x_13_pad_0 = const()[name = tensor("x_13_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138073472))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138892736))), name = tensor("layers_0_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_0_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138892864)))]; tensor x_13_cast_fp16 = conv(bias = layers_0_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_344, groups = var_183, pad = x_13_pad_0, pad_type = x_13_pad_type_0, strides = var_342, weight = layers_0_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_13_cast_fp16")]; tensor layers_0_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_0_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138895488)))]; tensor key_3_cast_fp16 = mul(x = x_13_cast_fp16, y = layers_0_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_3_cast_fp16")]; tensor var_354 = const()[name = tensor("op_354"), val = tensor([1, 1])]; tensor var_356 = const()[name = tensor("op_356"), val = tensor([1, 1])]; tensor x_15_pad_type_0 = const()[name = tensor("x_15_pad_type_0"), val = tensor("custom")]; tensor x_15_pad_0 = const()[name = tensor("x_15_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138898112))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139717376))), name = tensor("layers_0_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_0_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139717504)))]; tensor x_15_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_356, groups = var_183, pad = x_15_pad_0, pad_type = x_15_pad_type_0, strides = var_354, weight = layers_0_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_15_cast_fp16")]; tensor layers_0_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139720128)))]; tensor value_3_cast_fp16 = mul(x = x_15_cast_fp16, y = layers_0_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_3_cast_fp16")]; tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 20, 64, -1])]; tensor var_362_cast_fp16 = reshape(shape = var_361, x = query_3_cast_fp16)[name = tensor("op_362_cast_fp16")]; tensor var_363_to_fp16 = const()[name = tensor("op_363_to_fp16"), val = tensor(0x1p-3)]; tensor var_364_cast_fp16 = mul(x = var_362_cast_fp16, y = var_363_to_fp16)[name = tensor("op_364_cast_fp16")]; tensor var_365 = const()[name = tensor("op_365"), val = tensor([1, 20, 64, -1])]; tensor var_366_cast_fp16 = reshape(shape = var_365, x = key_3_cast_fp16)[name = tensor("op_366_cast_fp16")]; tensor mh_w_5_transpose_x_0 = const()[name = tensor("mh_w_5_transpose_x_0"), val = tensor(true)]; tensor mh_w_5_transpose_y_0 = const()[name = tensor("mh_w_5_transpose_y_0"), val = tensor(false)]; tensor mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_364_cast_fp16, y = var_366_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; tensor obj_13_cast_fp16 = softmax(axis = var_176, x = mh_w_5_cast_fp16)[name = tensor("obj_13_cast_fp16")]; tensor var_370 = const()[name = tensor("op_370"), val = tensor([1, 20, 64, -1])]; tensor var_371_cast_fp16 = reshape(shape = var_370, x = value_3_cast_fp16)[name = tensor("op_371_cast_fp16")]; tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(true)]; tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_371_cast_fp16, y = obj_13_cast_fp16)[name = tensor("attn_3_cast_fp16")]; tensor var_374 = const()[name = tensor("op_374"), val = tensor([1, 1280, 1, -1])]; tensor x_17_cast_fp16 = reshape(shape = var_374, x = attn_3_cast_fp16)[name = tensor("x_17_cast_fp16")]; tensor layers_0_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139722752)))]; tensor input_15_cast_fp16 = sub(x = x_17_cast_fp16, y = layers_0_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_15_cast_fp16")]; tensor var_382 = const()[name = tensor("op_382"), val = tensor([1, 1])]; tensor var_384 = const()[name = tensor("op_384"), val = tensor([1, 1])]; tensor x_19_pad_type_0 = const()[name = tensor("x_19_pad_type_0"), val = tensor("custom")]; tensor x_19_pad_0 = const()[name = tensor("x_19_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139725376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140544640))), name = tensor("layers_0_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_0_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140544768)))]; tensor x_19_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_384, groups = var_183, pad = x_19_pad_0, pad_type = x_19_pad_type_0, strides = var_382, weight = layers_0_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = tensor("x_19_cast_fp16")]; tensor layers_0_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140547392)))]; tensor obj_11_cast_fp16 = mul(x = x_19_cast_fp16, y = layers_0_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_11_cast_fp16")]; tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; tensor var_391 = const()[name = tensor("op_391"), val = tensor([1])]; tensor channels_mean_5_cast_fp16 = reduce_mean(axes = var_391, keep_dims = var_184, x = inputs_5_cast_fp16)[name = tensor("channels_mean_5_cast_fp16")]; tensor zero_mean_5_cast_fp16 = sub(x = inputs_5_cast_fp16, y = channels_mean_5_cast_fp16)[name = tensor("zero_mean_5_cast_fp16")]; tensor zero_mean_sq_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = zero_mean_5_cast_fp16)[name = tensor("zero_mean_sq_5_cast_fp16")]; tensor var_395 = const()[name = tensor("op_395"), val = tensor([1])]; tensor var_396_cast_fp16 = reduce_mean(axes = var_395, keep_dims = var_184, x = zero_mean_sq_5_cast_fp16)[name = tensor("op_396_cast_fp16")]; tensor var_397_to_fp16 = const()[name = tensor("op_397_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_398_cast_fp16 = add(x = var_396_cast_fp16, y = var_397_to_fp16)[name = tensor("op_398_cast_fp16")]; tensor denom_5_epsilon_0_to_fp16 = const()[name = tensor("denom_5_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_5_cast_fp16 = rsqrt(epsilon = denom_5_epsilon_0_to_fp16, x = var_398_cast_fp16)[name = tensor("denom_5_cast_fp16")]; tensor out_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = denom_5_cast_fp16)[name = tensor("out_5_cast_fp16")]; tensor x_21_gamma_0_to_fp16 = const()[name = tensor("x_21_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140550016)))]; tensor x_21_beta_0_to_fp16 = const()[name = tensor("x_21_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140552640)))]; tensor x_21_epsilon_0_to_fp16 = const()[name = tensor("x_21_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_21_cast_fp16 = batch_norm(beta = x_21_beta_0_to_fp16, epsilon = x_21_epsilon_0_to_fp16, gamma = x_21_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor("x_21_cast_fp16")]; tensor layers_0_fc1_input_shift_to_fp16 = const()[name = tensor("layers_0_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140555264)))]; tensor input_17_cast_fp16 = sub(x = x_21_cast_fp16, y = layers_0_fc1_input_shift_to_fp16)[name = tensor("input_17_cast_fp16")]; tensor var_413 = const()[name = tensor("op_413"), val = tensor([1, 1])]; tensor var_415 = const()[name = tensor("op_415"), val = tensor([1, 1])]; tensor x_23_pad_type_0 = const()[name = tensor("x_23_pad_type_0"), val = tensor("custom")]; tensor x_23_pad_0 = const()[name = tensor("x_23_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140557888))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143834752))), name = tensor("layers_0_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_0_fc1_module_bias_to_fp16 = const()[name = tensor("layers_0_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143834880)))]; tensor x_23_cast_fp16 = conv(bias = layers_0_fc1_module_bias_to_fp16, dilations = var_415, groups = var_183, pad = x_23_pad_0, pad_type = x_23_pad_type_0, strides = var_413, weight = layers_0_fc1_module_weight_to_fp16_palettized, x = input_17_cast_fp16)[name = tensor("x_23_cast_fp16")]; tensor layers_0_fc1_output_scale_to_fp16 = const()[name = tensor("layers_0_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143845184)))]; tensor input_19_cast_fp16 = mul(x = x_23_cast_fp16, y = layers_0_fc1_output_scale_to_fp16)[name = tensor("input_19_cast_fp16")]; tensor x_25_mode_0 = const()[name = tensor("x_25_mode_0"), val = tensor("EXACT")]; tensor x_25_cast_fp16 = gelu(mode = x_25_mode_0, x = input_19_cast_fp16)[name = tensor("x_25_cast_fp16")]; tensor layers_0_fc2_input_shift_to_fp16 = const()[name = tensor("layers_0_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143855488)))]; tensor input_21_cast_fp16 = sub(x = x_25_cast_fp16, y = layers_0_fc2_input_shift_to_fp16)[name = tensor("input_21_cast_fp16")]; tensor var_426 = const()[name = tensor("op_426"), val = tensor([1, 1])]; tensor var_428 = const()[name = tensor("op_428"), val = tensor([1, 1])]; tensor x_27_pad_type_0 = const()[name = tensor("x_27_pad_type_0"), val = tensor("custom")]; tensor x_27_pad_0 = const()[name = tensor("x_27_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_0_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143865792))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147142656))), name = tensor("layers_0_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_0_fc2_module_bias_to_fp16 = const()[name = tensor("layers_0_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147142784)))]; tensor x_27_cast_fp16 = conv(bias = layers_0_fc2_module_bias_to_fp16, dilations = var_428, groups = var_183, pad = x_27_pad_0, pad_type = x_27_pad_type_0, strides = var_426, weight = layers_0_fc2_module_weight_to_fp16_palettized, x = input_21_cast_fp16)[name = tensor("x_27_cast_fp16")]; tensor layers_0_fc2_output_scale_to_fp16 = const()[name = tensor("layers_0_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147145408)))]; tensor hidden_states_3_cast_fp16 = mul(x = x_27_cast_fp16, y = layers_0_fc2_output_scale_to_fp16)[name = tensor("hidden_states_3_cast_fp16")]; tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; tensor var_442 = const()[name = tensor("op_442"), val = tensor(3)]; tensor var_449 = const()[name = tensor("op_449"), val = tensor(1)]; tensor var_450 = const()[name = tensor("op_450"), val = tensor(true)]; tensor var_462 = const()[name = tensor("op_462"), val = tensor([1])]; tensor channels_mean_7_cast_fp16 = reduce_mean(axes = var_462, keep_dims = var_450, x = inputs_7_cast_fp16)[name = tensor("channels_mean_7_cast_fp16")]; tensor zero_mean_7_cast_fp16 = sub(x = inputs_7_cast_fp16, y = channels_mean_7_cast_fp16)[name = tensor("zero_mean_7_cast_fp16")]; tensor zero_mean_sq_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = zero_mean_7_cast_fp16)[name = tensor("zero_mean_sq_7_cast_fp16")]; tensor var_466 = const()[name = tensor("op_466"), val = tensor([1])]; tensor var_467_cast_fp16 = reduce_mean(axes = var_466, keep_dims = var_450, x = zero_mean_sq_7_cast_fp16)[name = tensor("op_467_cast_fp16")]; tensor var_468_to_fp16 = const()[name = tensor("op_468_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_469_cast_fp16 = add(x = var_467_cast_fp16, y = var_468_to_fp16)[name = tensor("op_469_cast_fp16")]; tensor denom_7_epsilon_0_to_fp16 = const()[name = tensor("denom_7_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_7_cast_fp16 = rsqrt(epsilon = denom_7_epsilon_0_to_fp16, x = var_469_cast_fp16)[name = tensor("denom_7_cast_fp16")]; tensor out_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = denom_7_cast_fp16)[name = tensor("out_7_cast_fp16")]; tensor obj_15_gamma_0_to_fp16 = const()[name = tensor("obj_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147148032)))]; tensor obj_15_beta_0_to_fp16 = const()[name = tensor("obj_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147150656)))]; tensor obj_15_epsilon_0_to_fp16 = const()[name = tensor("obj_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor("obj_15_cast_fp16")]; tensor layers_1_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147153280)))]; tensor input_23_cast_fp16 = sub(x = obj_15_cast_fp16, y = layers_1_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_23_cast_fp16")]; tensor var_488 = const()[name = tensor("op_488"), val = tensor([1, 1])]; tensor var_490 = const()[name = tensor("op_490"), val = tensor([1, 1])]; tensor x_29_pad_type_0 = const()[name = tensor("x_29_pad_type_0"), val = tensor("custom")]; tensor x_29_pad_0 = const()[name = tensor("x_29_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147155904))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147975168))), name = tensor("layers_1_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_1_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147975296)))]; tensor x_29_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_module_bias_to_fp16, dilations = var_490, groups = var_449, pad = x_29_pad_0, pad_type = x_29_pad_type_0, strides = var_488, weight = layers_1_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_23_cast_fp16)[name = tensor("x_29_cast_fp16")]; tensor layers_1_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147977920)))]; tensor query_5_cast_fp16 = mul(x = x_29_cast_fp16, y = layers_1_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_5_cast_fp16")]; tensor var_500 = const()[name = tensor("op_500"), val = tensor([1, 1])]; tensor var_502 = const()[name = tensor("op_502"), val = tensor([1, 1])]; tensor x_31_pad_type_0 = const()[name = tensor("x_31_pad_type_0"), val = tensor("custom")]; tensor x_31_pad_0 = const()[name = tensor("x_31_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147980544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148799808))), name = tensor("layers_1_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_1_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148799936)))]; tensor x_31_cast_fp16 = conv(bias = layers_1_self_attn_k_proj_module_bias_to_fp16, dilations = var_502, groups = var_449, pad = x_31_pad_0, pad_type = x_31_pad_type_0, strides = var_500, weight = layers_1_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_23_cast_fp16)[name = tensor("x_31_cast_fp16")]; tensor layers_1_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_1_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148802560)))]; tensor current_key_3_cast_fp16 = mul(x = x_31_cast_fp16, y = layers_1_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_3_cast_fp16")]; tensor var_512 = const()[name = tensor("op_512"), val = tensor([1, 1])]; tensor var_514 = const()[name = tensor("op_514"), val = tensor([1, 1])]; tensor x_33_pad_type_0 = const()[name = tensor("x_33_pad_type_0"), val = tensor("custom")]; tensor x_33_pad_0 = const()[name = tensor("x_33_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148805184))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149624448))), name = tensor("layers_1_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_1_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149624576)))]; tensor x_33_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_module_bias_to_fp16, dilations = var_514, groups = var_449, pad = x_33_pad_0, pad_type = x_33_pad_type_0, strides = var_512, weight = layers_1_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_23_cast_fp16)[name = tensor("x_33_cast_fp16")]; tensor layers_1_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149627200)))]; tensor current_value_3_cast_fp16 = mul(x = x_33_cast_fp16, y = layers_1_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_3_cast_fp16")]; tensor var_522_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_522_cast_fp16")]; tensor var_524_cast_fp16 = mul(x = var_103_cast_fp16_1, y = var_257_cast_fp16)[name = tensor("op_524_cast_fp16")]; tensor key_5_cast_fp16 = add(x = var_522_cast_fp16, y = var_524_cast_fp16)[name = tensor("key_5_cast_fp16")]; tensor var_526_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_526_cast_fp16")]; tensor var_528_cast_fp16 = mul(x = var_138_cast_fp16_1, y = var_257_cast_fp16)[name = tensor("op_528_cast_fp16")]; tensor value_5_cast_fp16 = add(x = var_526_cast_fp16, y = var_528_cast_fp16)[name = tensor("value_5_cast_fp16")]; tensor var_531 = const()[name = tensor("op_531"), val = tensor([1, 20, 64, -1])]; tensor var_532_cast_fp16 = reshape(shape = var_531, x = query_5_cast_fp16)[name = tensor("op_532_cast_fp16")]; tensor var_533_to_fp16 = const()[name = tensor("op_533_to_fp16"), val = tensor(0x1p-3)]; tensor var_534_cast_fp16 = mul(x = var_532_cast_fp16, y = var_533_to_fp16)[name = tensor("op_534_cast_fp16")]; tensor var_535 = const()[name = tensor("op_535"), val = tensor([1, 20, 64, -1])]; tensor var_536_cast_fp16 = reshape(shape = var_535, x = key_5_cast_fp16)[name = tensor("op_536_cast_fp16")]; tensor mh_w_7_transpose_x_0 = const()[name = tensor("mh_w_7_transpose_x_0"), val = tensor(true)]; tensor mh_w_7_transpose_y_0 = const()[name = tensor("mh_w_7_transpose_y_0"), val = tensor(false)]; tensor mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_534_cast_fp16, y = var_536_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; tensor var_544_cast_fp16 = softmax(axis = var_442, x = mh_w_9_cast_fp16)[name = tensor("op_544_cast_fp16")]; tensor var_545 = const()[name = tensor("op_545"), val = tensor([1, 20, 64, -1])]; tensor var_546_cast_fp16 = reshape(shape = var_545, x = value_5_cast_fp16)[name = tensor("op_546_cast_fp16")]; tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(true)]; tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_546_cast_fp16, y = var_544_cast_fp16)[name = tensor("attn_5_cast_fp16")]; tensor var_549 = const()[name = tensor("op_549"), val = tensor([1, 1280, 1, -1])]; tensor x_35_cast_fp16 = reshape(shape = var_549, x = attn_5_cast_fp16)[name = tensor("x_35_cast_fp16")]; tensor layers_1_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149629824)))]; tensor input_29_cast_fp16 = sub(x = x_35_cast_fp16, y = layers_1_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_29_cast_fp16")]; tensor var_557 = const()[name = tensor("op_557"), val = tensor([1, 1])]; tensor var_559 = const()[name = tensor("op_559"), val = tensor([1, 1])]; tensor x_37_pad_type_0 = const()[name = tensor("x_37_pad_type_0"), val = tensor("custom")]; tensor x_37_pad_0 = const()[name = tensor("x_37_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149632448))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150451712))), name = tensor("layers_1_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_1_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150451840)))]; tensor x_37_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_module_bias_to_fp16, dilations = var_559, groups = var_449, pad = x_37_pad_0, pad_type = x_37_pad_type_0, strides = var_557, weight = layers_1_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = tensor("x_37_cast_fp16")]; tensor layers_1_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150454464)))]; tensor obj_21_cast_fp16 = mul(x = x_37_cast_fp16, y = layers_1_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_21_cast_fp16")]; tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; tensor var_570 = const()[name = tensor("op_570"), val = tensor([1])]; tensor channels_mean_9_cast_fp16 = reduce_mean(axes = var_570, keep_dims = var_450, x = inputs_9_cast_fp16)[name = tensor("channels_mean_9_cast_fp16")]; tensor zero_mean_9_cast_fp16 = sub(x = inputs_9_cast_fp16, y = channels_mean_9_cast_fp16)[name = tensor("zero_mean_9_cast_fp16")]; tensor zero_mean_sq_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = zero_mean_9_cast_fp16)[name = tensor("zero_mean_sq_9_cast_fp16")]; tensor var_574 = const()[name = tensor("op_574"), val = tensor([1])]; tensor var_575_cast_fp16 = reduce_mean(axes = var_574, keep_dims = var_450, x = zero_mean_sq_9_cast_fp16)[name = tensor("op_575_cast_fp16")]; tensor var_576_to_fp16 = const()[name = tensor("op_576_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_577_cast_fp16 = add(x = var_575_cast_fp16, y = var_576_to_fp16)[name = tensor("op_577_cast_fp16")]; tensor denom_9_epsilon_0_to_fp16 = const()[name = tensor("denom_9_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_9_cast_fp16 = rsqrt(epsilon = denom_9_epsilon_0_to_fp16, x = var_577_cast_fp16)[name = tensor("denom_9_cast_fp16")]; tensor out_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = denom_9_cast_fp16)[name = tensor("out_9_cast_fp16")]; tensor obj_23_gamma_0_to_fp16 = const()[name = tensor("obj_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150457088)))]; tensor obj_23_beta_0_to_fp16 = const()[name = tensor("obj_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150459712)))]; tensor obj_23_epsilon_0_to_fp16 = const()[name = tensor("obj_23_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor("obj_23_cast_fp16")]; tensor layers_1_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150462336)))]; tensor input_31_cast_fp16 = sub(x = obj_23_cast_fp16, y = layers_1_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_31_cast_fp16")]; tensor var_596 = const()[name = tensor("op_596"), val = tensor([1, 1])]; tensor var_598 = const()[name = tensor("op_598"), val = tensor([1, 1])]; tensor x_39_pad_type_0 = const()[name = tensor("x_39_pad_type_0"), val = tensor("custom")]; tensor x_39_pad_0 = const()[name = tensor("x_39_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150464960))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151284224))), name = tensor("layers_1_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_1_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151284352)))]; tensor x_39_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_598, groups = var_449, pad = x_39_pad_0, pad_type = x_39_pad_type_0, strides = var_596, weight = layers_1_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_31_cast_fp16)[name = tensor("x_39_cast_fp16")]; tensor layers_1_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151286976)))]; tensor query_7_cast_fp16 = mul(x = x_39_cast_fp16, y = layers_1_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_7_cast_fp16")]; tensor var_608 = const()[name = tensor("op_608"), val = tensor([1, 1])]; tensor var_610 = const()[name = tensor("op_610"), val = tensor([1, 1])]; tensor x_41_pad_type_0 = const()[name = tensor("x_41_pad_type_0"), val = tensor("custom")]; tensor x_41_pad_0 = const()[name = tensor("x_41_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151289600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152108864))), name = tensor("layers_1_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_1_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152108992)))]; tensor x_41_cast_fp16 = conv(bias = layers_1_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_610, groups = var_449, pad = x_41_pad_0, pad_type = x_41_pad_type_0, strides = var_608, weight = layers_1_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_41_cast_fp16")]; tensor layers_1_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_1_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152111616)))]; tensor key_7_cast_fp16 = mul(x = x_41_cast_fp16, y = layers_1_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_7_cast_fp16")]; tensor var_620 = const()[name = tensor("op_620"), val = tensor([1, 1])]; tensor var_622 = const()[name = tensor("op_622"), val = tensor([1, 1])]; tensor x_43_pad_type_0 = const()[name = tensor("x_43_pad_type_0"), val = tensor("custom")]; tensor x_43_pad_0 = const()[name = tensor("x_43_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152114240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152933504))), name = tensor("layers_1_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_1_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152933632)))]; tensor x_43_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_622, groups = var_449, pad = x_43_pad_0, pad_type = x_43_pad_type_0, strides = var_620, weight = layers_1_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_43_cast_fp16")]; tensor layers_1_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152936256)))]; tensor value_7_cast_fp16 = mul(x = x_43_cast_fp16, y = layers_1_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_7_cast_fp16")]; tensor var_627 = const()[name = tensor("op_627"), val = tensor([1, 20, 64, -1])]; tensor var_628_cast_fp16 = reshape(shape = var_627, x = query_7_cast_fp16)[name = tensor("op_628_cast_fp16")]; tensor var_629_to_fp16 = const()[name = tensor("op_629_to_fp16"), val = tensor(0x1p-3)]; tensor var_630_cast_fp16 = mul(x = var_628_cast_fp16, y = var_629_to_fp16)[name = tensor("op_630_cast_fp16")]; tensor var_631 = const()[name = tensor("op_631"), val = tensor([1, 20, 64, -1])]; tensor var_632_cast_fp16 = reshape(shape = var_631, x = key_7_cast_fp16)[name = tensor("op_632_cast_fp16")]; tensor mh_w_11_transpose_x_0 = const()[name = tensor("mh_w_11_transpose_x_0"), val = tensor(true)]; tensor mh_w_11_transpose_y_0 = const()[name = tensor("mh_w_11_transpose_y_0"), val = tensor(false)]; tensor mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_630_cast_fp16, y = var_632_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; tensor obj_27_cast_fp16 = softmax(axis = var_442, x = mh_w_11_cast_fp16)[name = tensor("obj_27_cast_fp16")]; tensor var_636 = const()[name = tensor("op_636"), val = tensor([1, 20, 64, -1])]; tensor var_637_cast_fp16 = reshape(shape = var_636, x = value_7_cast_fp16)[name = tensor("op_637_cast_fp16")]; tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(true)]; tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_637_cast_fp16, y = obj_27_cast_fp16)[name = tensor("attn_7_cast_fp16")]; tensor var_640 = const()[name = tensor("op_640"), val = tensor([1, 1280, 1, -1])]; tensor x_45_cast_fp16 = reshape(shape = var_640, x = attn_7_cast_fp16)[name = tensor("x_45_cast_fp16")]; tensor layers_1_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152938880)))]; tensor input_37_cast_fp16 = sub(x = x_45_cast_fp16, y = layers_1_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_37_cast_fp16")]; tensor var_648 = const()[name = tensor("op_648"), val = tensor([1, 1])]; tensor var_650 = const()[name = tensor("op_650"), val = tensor([1, 1])]; tensor x_47_pad_type_0 = const()[name = tensor("x_47_pad_type_0"), val = tensor("custom")]; tensor x_47_pad_0 = const()[name = tensor("x_47_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152941504))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153760768))), name = tensor("layers_1_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_1_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153760896)))]; tensor x_47_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_650, groups = var_449, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = var_648, weight = layers_1_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_37_cast_fp16)[name = tensor("x_47_cast_fp16")]; tensor layers_1_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153763520)))]; tensor obj_25_cast_fp16 = mul(x = x_47_cast_fp16, y = layers_1_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_25_cast_fp16")]; tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; tensor var_657 = const()[name = tensor("op_657"), val = tensor([1])]; tensor channels_mean_11_cast_fp16 = reduce_mean(axes = var_657, keep_dims = var_450, x = inputs_11_cast_fp16)[name = tensor("channels_mean_11_cast_fp16")]; tensor zero_mean_11_cast_fp16 = sub(x = inputs_11_cast_fp16, y = channels_mean_11_cast_fp16)[name = tensor("zero_mean_11_cast_fp16")]; tensor zero_mean_sq_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = zero_mean_11_cast_fp16)[name = tensor("zero_mean_sq_11_cast_fp16")]; tensor var_661 = const()[name = tensor("op_661"), val = tensor([1])]; tensor var_662_cast_fp16 = reduce_mean(axes = var_661, keep_dims = var_450, x = zero_mean_sq_11_cast_fp16)[name = tensor("op_662_cast_fp16")]; tensor var_663_to_fp16 = const()[name = tensor("op_663_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_664_cast_fp16 = add(x = var_662_cast_fp16, y = var_663_to_fp16)[name = tensor("op_664_cast_fp16")]; tensor denom_11_epsilon_0_to_fp16 = const()[name = tensor("denom_11_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_11_cast_fp16 = rsqrt(epsilon = denom_11_epsilon_0_to_fp16, x = var_664_cast_fp16)[name = tensor("denom_11_cast_fp16")]; tensor out_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = denom_11_cast_fp16)[name = tensor("out_11_cast_fp16")]; tensor x_49_gamma_0_to_fp16 = const()[name = tensor("x_49_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153766144)))]; tensor x_49_beta_0_to_fp16 = const()[name = tensor("x_49_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153768768)))]; tensor x_49_epsilon_0_to_fp16 = const()[name = tensor("x_49_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_49_cast_fp16 = batch_norm(beta = x_49_beta_0_to_fp16, epsilon = x_49_epsilon_0_to_fp16, gamma = x_49_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor("x_49_cast_fp16")]; tensor layers_1_fc1_input_shift_to_fp16 = const()[name = tensor("layers_1_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153771392)))]; tensor input_39_cast_fp16 = sub(x = x_49_cast_fp16, y = layers_1_fc1_input_shift_to_fp16)[name = tensor("input_39_cast_fp16")]; tensor var_679 = const()[name = tensor("op_679"), val = tensor([1, 1])]; tensor var_681 = const()[name = tensor("op_681"), val = tensor([1, 1])]; tensor x_51_pad_type_0 = const()[name = tensor("x_51_pad_type_0"), val = tensor("custom")]; tensor x_51_pad_0 = const()[name = tensor("x_51_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153774016))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157050880))), name = tensor("layers_1_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_1_fc1_module_bias_to_fp16 = const()[name = tensor("layers_1_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157051008)))]; tensor x_51_cast_fp16 = conv(bias = layers_1_fc1_module_bias_to_fp16, dilations = var_681, groups = var_449, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = var_679, weight = layers_1_fc1_module_weight_to_fp16_palettized, x = input_39_cast_fp16)[name = tensor("x_51_cast_fp16")]; tensor layers_1_fc1_output_scale_to_fp16 = const()[name = tensor("layers_1_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157061312)))]; tensor input_41_cast_fp16 = mul(x = x_51_cast_fp16, y = layers_1_fc1_output_scale_to_fp16)[name = tensor("input_41_cast_fp16")]; tensor x_53_mode_0 = const()[name = tensor("x_53_mode_0"), val = tensor("EXACT")]; tensor x_53_cast_fp16 = gelu(mode = x_53_mode_0, x = input_41_cast_fp16)[name = tensor("x_53_cast_fp16")]; tensor layers_1_fc2_input_shift_to_fp16 = const()[name = tensor("layers_1_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157071616)))]; tensor input_43_cast_fp16 = sub(x = x_53_cast_fp16, y = layers_1_fc2_input_shift_to_fp16)[name = tensor("input_43_cast_fp16")]; tensor var_692 = const()[name = tensor("op_692"), val = tensor([1, 1])]; tensor var_694 = const()[name = tensor("op_694"), val = tensor([1, 1])]; tensor x_55_pad_type_0 = const()[name = tensor("x_55_pad_type_0"), val = tensor("custom")]; tensor x_55_pad_0 = const()[name = tensor("x_55_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_1_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157081920))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160358784))), name = tensor("layers_1_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_1_fc2_module_bias_to_fp16 = const()[name = tensor("layers_1_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160358912)))]; tensor x_55_cast_fp16 = conv(bias = layers_1_fc2_module_bias_to_fp16, dilations = var_694, groups = var_449, pad = x_55_pad_0, pad_type = x_55_pad_type_0, strides = var_692, weight = layers_1_fc2_module_weight_to_fp16_palettized, x = input_43_cast_fp16)[name = tensor("x_55_cast_fp16")]; tensor layers_1_fc2_output_scale_to_fp16 = const()[name = tensor("layers_1_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160361536)))]; tensor hidden_states_5_cast_fp16 = mul(x = x_55_cast_fp16, y = layers_1_fc2_output_scale_to_fp16)[name = tensor("hidden_states_5_cast_fp16")]; tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; tensor var_708 = const()[name = tensor("op_708"), val = tensor(3)]; tensor var_715 = const()[name = tensor("op_715"), val = tensor(1)]; tensor var_716 = const()[name = tensor("op_716"), val = tensor(true)]; tensor var_728 = const()[name = tensor("op_728"), val = tensor([1])]; tensor channels_mean_13_cast_fp16 = reduce_mean(axes = var_728, keep_dims = var_716, x = inputs_13_cast_fp16)[name = tensor("channels_mean_13_cast_fp16")]; tensor zero_mean_13_cast_fp16 = sub(x = inputs_13_cast_fp16, y = channels_mean_13_cast_fp16)[name = tensor("zero_mean_13_cast_fp16")]; tensor zero_mean_sq_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = zero_mean_13_cast_fp16)[name = tensor("zero_mean_sq_13_cast_fp16")]; tensor var_732 = const()[name = tensor("op_732"), val = tensor([1])]; tensor var_733_cast_fp16 = reduce_mean(axes = var_732, keep_dims = var_716, x = zero_mean_sq_13_cast_fp16)[name = tensor("op_733_cast_fp16")]; tensor var_734_to_fp16 = const()[name = tensor("op_734_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_735_cast_fp16 = add(x = var_733_cast_fp16, y = var_734_to_fp16)[name = tensor("op_735_cast_fp16")]; tensor denom_13_epsilon_0_to_fp16 = const()[name = tensor("denom_13_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_13_cast_fp16 = rsqrt(epsilon = denom_13_epsilon_0_to_fp16, x = var_735_cast_fp16)[name = tensor("denom_13_cast_fp16")]; tensor out_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = denom_13_cast_fp16)[name = tensor("out_13_cast_fp16")]; tensor obj_29_gamma_0_to_fp16 = const()[name = tensor("obj_29_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160364160)))]; tensor obj_29_beta_0_to_fp16 = const()[name = tensor("obj_29_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160366784)))]; tensor obj_29_epsilon_0_to_fp16 = const()[name = tensor("obj_29_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor("obj_29_cast_fp16")]; tensor layers_2_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160369408)))]; tensor input_45_cast_fp16 = sub(x = obj_29_cast_fp16, y = layers_2_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_45_cast_fp16")]; tensor var_754 = const()[name = tensor("op_754"), val = tensor([1, 1])]; tensor var_756 = const()[name = tensor("op_756"), val = tensor([1, 1])]; tensor x_57_pad_type_0 = const()[name = tensor("x_57_pad_type_0"), val = tensor("custom")]; tensor x_57_pad_0 = const()[name = tensor("x_57_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160372032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161191296))), name = tensor("layers_2_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_2_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161191424)))]; tensor x_57_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_module_bias_to_fp16, dilations = var_756, groups = var_715, pad = x_57_pad_0, pad_type = x_57_pad_type_0, strides = var_754, weight = layers_2_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_45_cast_fp16)[name = tensor("x_57_cast_fp16")]; tensor layers_2_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161194048)))]; tensor query_9_cast_fp16 = mul(x = x_57_cast_fp16, y = layers_2_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_9_cast_fp16")]; tensor var_766 = const()[name = tensor("op_766"), val = tensor([1, 1])]; tensor var_768 = const()[name = tensor("op_768"), val = tensor([1, 1])]; tensor x_59_pad_type_0 = const()[name = tensor("x_59_pad_type_0"), val = tensor("custom")]; tensor x_59_pad_0 = const()[name = tensor("x_59_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161196672))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162015936))), name = tensor("layers_2_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_2_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162016064)))]; tensor x_59_cast_fp16 = conv(bias = layers_2_self_attn_k_proj_module_bias_to_fp16, dilations = var_768, groups = var_715, pad = x_59_pad_0, pad_type = x_59_pad_type_0, strides = var_766, weight = layers_2_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_45_cast_fp16)[name = tensor("x_59_cast_fp16")]; tensor layers_2_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_2_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162018688)))]; tensor current_key_5_cast_fp16 = mul(x = x_59_cast_fp16, y = layers_2_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_5_cast_fp16")]; tensor var_778 = const()[name = tensor("op_778"), val = tensor([1, 1])]; tensor var_780 = const()[name = tensor("op_780"), val = tensor([1, 1])]; tensor x_61_pad_type_0 = const()[name = tensor("x_61_pad_type_0"), val = tensor("custom")]; tensor x_61_pad_0 = const()[name = tensor("x_61_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162021312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162840576))), name = tensor("layers_2_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_2_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162840704)))]; tensor x_61_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_module_bias_to_fp16, dilations = var_780, groups = var_715, pad = x_61_pad_0, pad_type = x_61_pad_type_0, strides = var_778, weight = layers_2_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_45_cast_fp16)[name = tensor("x_61_cast_fp16")]; tensor layers_2_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162843328)))]; tensor current_value_5_cast_fp16 = mul(x = x_61_cast_fp16, y = layers_2_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_5_cast_fp16")]; tensor var_788_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_788_cast_fp16")]; tensor var_790_cast_fp16 = mul(x = var_103_cast_fp16_2, y = var_257_cast_fp16)[name = tensor("op_790_cast_fp16")]; tensor key_9_cast_fp16 = add(x = var_788_cast_fp16, y = var_790_cast_fp16)[name = tensor("key_9_cast_fp16")]; tensor var_792_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_792_cast_fp16")]; tensor var_794_cast_fp16 = mul(x = var_138_cast_fp16_2, y = var_257_cast_fp16)[name = tensor("op_794_cast_fp16")]; tensor value_9_cast_fp16 = add(x = var_792_cast_fp16, y = var_794_cast_fp16)[name = tensor("value_9_cast_fp16")]; tensor var_797 = const()[name = tensor("op_797"), val = tensor([1, 20, 64, -1])]; tensor var_798_cast_fp16 = reshape(shape = var_797, x = query_9_cast_fp16)[name = tensor("op_798_cast_fp16")]; tensor var_799_to_fp16 = const()[name = tensor("op_799_to_fp16"), val = tensor(0x1p-3)]; tensor var_800_cast_fp16 = mul(x = var_798_cast_fp16, y = var_799_to_fp16)[name = tensor("op_800_cast_fp16")]; tensor var_801 = const()[name = tensor("op_801"), val = tensor([1, 20, 64, -1])]; tensor var_802_cast_fp16 = reshape(shape = var_801, x = key_9_cast_fp16)[name = tensor("op_802_cast_fp16")]; tensor mh_w_13_transpose_x_0 = const()[name = tensor("mh_w_13_transpose_x_0"), val = tensor(true)]; tensor mh_w_13_transpose_y_0 = const()[name = tensor("mh_w_13_transpose_y_0"), val = tensor(false)]; tensor mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_800_cast_fp16, y = var_802_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; tensor var_810_cast_fp16 = softmax(axis = var_708, x = mh_w_15_cast_fp16)[name = tensor("op_810_cast_fp16")]; tensor var_811 = const()[name = tensor("op_811"), val = tensor([1, 20, 64, -1])]; tensor var_812_cast_fp16 = reshape(shape = var_811, x = value_9_cast_fp16)[name = tensor("op_812_cast_fp16")]; tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(true)]; tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_812_cast_fp16, y = var_810_cast_fp16)[name = tensor("attn_9_cast_fp16")]; tensor var_815 = const()[name = tensor("op_815"), val = tensor([1, 1280, 1, -1])]; tensor x_63_cast_fp16 = reshape(shape = var_815, x = attn_9_cast_fp16)[name = tensor("x_63_cast_fp16")]; tensor layers_2_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162845952)))]; tensor input_51_cast_fp16 = sub(x = x_63_cast_fp16, y = layers_2_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_51_cast_fp16")]; tensor var_823 = const()[name = tensor("op_823"), val = tensor([1, 1])]; tensor var_825 = const()[name = tensor("op_825"), val = tensor([1, 1])]; tensor x_65_pad_type_0 = const()[name = tensor("x_65_pad_type_0"), val = tensor("custom")]; tensor x_65_pad_0 = const()[name = tensor("x_65_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162848576))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163667840))), name = tensor("layers_2_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_2_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163667968)))]; tensor x_65_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_module_bias_to_fp16, dilations = var_825, groups = var_715, pad = x_65_pad_0, pad_type = x_65_pad_type_0, strides = var_823, weight = layers_2_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = tensor("x_65_cast_fp16")]; tensor layers_2_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163670592)))]; tensor obj_35_cast_fp16 = mul(x = x_65_cast_fp16, y = layers_2_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_35_cast_fp16")]; tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; tensor var_836 = const()[name = tensor("op_836"), val = tensor([1])]; tensor channels_mean_15_cast_fp16 = reduce_mean(axes = var_836, keep_dims = var_716, x = inputs_15_cast_fp16)[name = tensor("channels_mean_15_cast_fp16")]; tensor zero_mean_15_cast_fp16 = sub(x = inputs_15_cast_fp16, y = channels_mean_15_cast_fp16)[name = tensor("zero_mean_15_cast_fp16")]; tensor zero_mean_sq_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = zero_mean_15_cast_fp16)[name = tensor("zero_mean_sq_15_cast_fp16")]; tensor var_840 = const()[name = tensor("op_840"), val = tensor([1])]; tensor var_841_cast_fp16 = reduce_mean(axes = var_840, keep_dims = var_716, x = zero_mean_sq_15_cast_fp16)[name = tensor("op_841_cast_fp16")]; tensor var_842_to_fp16 = const()[name = tensor("op_842_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_843_cast_fp16 = add(x = var_841_cast_fp16, y = var_842_to_fp16)[name = tensor("op_843_cast_fp16")]; tensor denom_15_epsilon_0_to_fp16 = const()[name = tensor("denom_15_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_15_cast_fp16 = rsqrt(epsilon = denom_15_epsilon_0_to_fp16, x = var_843_cast_fp16)[name = tensor("denom_15_cast_fp16")]; tensor out_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = denom_15_cast_fp16)[name = tensor("out_15_cast_fp16")]; tensor obj_37_gamma_0_to_fp16 = const()[name = tensor("obj_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163673216)))]; tensor obj_37_beta_0_to_fp16 = const()[name = tensor("obj_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163675840)))]; tensor obj_37_epsilon_0_to_fp16 = const()[name = tensor("obj_37_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor("obj_37_cast_fp16")]; tensor layers_2_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163678464)))]; tensor input_53_cast_fp16 = sub(x = obj_37_cast_fp16, y = layers_2_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_53_cast_fp16")]; tensor var_862 = const()[name = tensor("op_862"), val = tensor([1, 1])]; tensor var_864 = const()[name = tensor("op_864"), val = tensor([1, 1])]; tensor x_67_pad_type_0 = const()[name = tensor("x_67_pad_type_0"), val = tensor("custom")]; tensor x_67_pad_0 = const()[name = tensor("x_67_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163681088))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164500352))), name = tensor("layers_2_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_2_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164500480)))]; tensor x_67_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_864, groups = var_715, pad = x_67_pad_0, pad_type = x_67_pad_type_0, strides = var_862, weight = layers_2_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_53_cast_fp16)[name = tensor("x_67_cast_fp16")]; tensor layers_2_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164503104)))]; tensor query_11_cast_fp16 = mul(x = x_67_cast_fp16, y = layers_2_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_11_cast_fp16")]; tensor var_874 = const()[name = tensor("op_874"), val = tensor([1, 1])]; tensor var_876 = const()[name = tensor("op_876"), val = tensor([1, 1])]; tensor x_69_pad_type_0 = const()[name = tensor("x_69_pad_type_0"), val = tensor("custom")]; tensor x_69_pad_0 = const()[name = tensor("x_69_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164505728))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165324992))), name = tensor("layers_2_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_2_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165325120)))]; tensor x_69_cast_fp16 = conv(bias = layers_2_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_876, groups = var_715, pad = x_69_pad_0, pad_type = x_69_pad_type_0, strides = var_874, weight = layers_2_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_69_cast_fp16")]; tensor layers_2_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_2_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165327744)))]; tensor key_11_cast_fp16 = mul(x = x_69_cast_fp16, y = layers_2_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_11_cast_fp16")]; tensor var_886 = const()[name = tensor("op_886"), val = tensor([1, 1])]; tensor var_888 = const()[name = tensor("op_888"), val = tensor([1, 1])]; tensor x_71_pad_type_0 = const()[name = tensor("x_71_pad_type_0"), val = tensor("custom")]; tensor x_71_pad_0 = const()[name = tensor("x_71_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165330368))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166149632))), name = tensor("layers_2_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_2_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166149760)))]; tensor x_71_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_888, groups = var_715, pad = x_71_pad_0, pad_type = x_71_pad_type_0, strides = var_886, weight = layers_2_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_71_cast_fp16")]; tensor layers_2_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166152384)))]; tensor value_11_cast_fp16 = mul(x = x_71_cast_fp16, y = layers_2_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_11_cast_fp16")]; tensor var_893 = const()[name = tensor("op_893"), val = tensor([1, 20, 64, -1])]; tensor var_894_cast_fp16 = reshape(shape = var_893, x = query_11_cast_fp16)[name = tensor("op_894_cast_fp16")]; tensor var_895_to_fp16 = const()[name = tensor("op_895_to_fp16"), val = tensor(0x1p-3)]; tensor var_896_cast_fp16 = mul(x = var_894_cast_fp16, y = var_895_to_fp16)[name = tensor("op_896_cast_fp16")]; tensor var_897 = const()[name = tensor("op_897"), val = tensor([1, 20, 64, -1])]; tensor var_898_cast_fp16 = reshape(shape = var_897, x = key_11_cast_fp16)[name = tensor("op_898_cast_fp16")]; tensor mh_w_17_transpose_x_0 = const()[name = tensor("mh_w_17_transpose_x_0"), val = tensor(true)]; tensor mh_w_17_transpose_y_0 = const()[name = tensor("mh_w_17_transpose_y_0"), val = tensor(false)]; tensor mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_896_cast_fp16, y = var_898_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; tensor obj_41_cast_fp16 = softmax(axis = var_708, x = mh_w_17_cast_fp16)[name = tensor("obj_41_cast_fp16")]; tensor var_902 = const()[name = tensor("op_902"), val = tensor([1, 20, 64, -1])]; tensor var_903_cast_fp16 = reshape(shape = var_902, x = value_11_cast_fp16)[name = tensor("op_903_cast_fp16")]; tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(true)]; tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_903_cast_fp16, y = obj_41_cast_fp16)[name = tensor("attn_11_cast_fp16")]; tensor var_906 = const()[name = tensor("op_906"), val = tensor([1, 1280, 1, -1])]; tensor x_73_cast_fp16 = reshape(shape = var_906, x = attn_11_cast_fp16)[name = tensor("x_73_cast_fp16")]; tensor layers_2_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166155008)))]; tensor input_59_cast_fp16 = sub(x = x_73_cast_fp16, y = layers_2_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_59_cast_fp16")]; tensor var_914 = const()[name = tensor("op_914"), val = tensor([1, 1])]; tensor var_916 = const()[name = tensor("op_916"), val = tensor([1, 1])]; tensor x_75_pad_type_0 = const()[name = tensor("x_75_pad_type_0"), val = tensor("custom")]; tensor x_75_pad_0 = const()[name = tensor("x_75_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166157632))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166976896))), name = tensor("layers_2_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_2_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166977024)))]; tensor x_75_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_916, groups = var_715, pad = x_75_pad_0, pad_type = x_75_pad_type_0, strides = var_914, weight = layers_2_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_59_cast_fp16)[name = tensor("x_75_cast_fp16")]; tensor layers_2_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166979648)))]; tensor obj_39_cast_fp16 = mul(x = x_75_cast_fp16, y = layers_2_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_39_cast_fp16")]; tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; tensor var_923 = const()[name = tensor("op_923"), val = tensor([1])]; tensor channels_mean_17_cast_fp16 = reduce_mean(axes = var_923, keep_dims = var_716, x = inputs_17_cast_fp16)[name = tensor("channels_mean_17_cast_fp16")]; tensor zero_mean_17_cast_fp16 = sub(x = inputs_17_cast_fp16, y = channels_mean_17_cast_fp16)[name = tensor("zero_mean_17_cast_fp16")]; tensor zero_mean_sq_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = zero_mean_17_cast_fp16)[name = tensor("zero_mean_sq_17_cast_fp16")]; tensor var_927 = const()[name = tensor("op_927"), val = tensor([1])]; tensor var_928_cast_fp16 = reduce_mean(axes = var_927, keep_dims = var_716, x = zero_mean_sq_17_cast_fp16)[name = tensor("op_928_cast_fp16")]; tensor var_929_to_fp16 = const()[name = tensor("op_929_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_930_cast_fp16 = add(x = var_928_cast_fp16, y = var_929_to_fp16)[name = tensor("op_930_cast_fp16")]; tensor denom_17_epsilon_0_to_fp16 = const()[name = tensor("denom_17_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_17_cast_fp16 = rsqrt(epsilon = denom_17_epsilon_0_to_fp16, x = var_930_cast_fp16)[name = tensor("denom_17_cast_fp16")]; tensor out_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = denom_17_cast_fp16)[name = tensor("out_17_cast_fp16")]; tensor x_77_gamma_0_to_fp16 = const()[name = tensor("x_77_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166982272)))]; tensor x_77_beta_0_to_fp16 = const()[name = tensor("x_77_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166984896)))]; tensor x_77_epsilon_0_to_fp16 = const()[name = tensor("x_77_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_77_cast_fp16 = batch_norm(beta = x_77_beta_0_to_fp16, epsilon = x_77_epsilon_0_to_fp16, gamma = x_77_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor("x_77_cast_fp16")]; tensor layers_2_fc1_input_shift_to_fp16 = const()[name = tensor("layers_2_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166987520)))]; tensor input_61_cast_fp16 = sub(x = x_77_cast_fp16, y = layers_2_fc1_input_shift_to_fp16)[name = tensor("input_61_cast_fp16")]; tensor var_945 = const()[name = tensor("op_945"), val = tensor([1, 1])]; tensor var_947 = const()[name = tensor("op_947"), val = tensor([1, 1])]; tensor x_79_pad_type_0 = const()[name = tensor("x_79_pad_type_0"), val = tensor("custom")]; tensor x_79_pad_0 = const()[name = tensor("x_79_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166990144))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170267008))), name = tensor("layers_2_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_2_fc1_module_bias_to_fp16 = const()[name = tensor("layers_2_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170267136)))]; tensor x_79_cast_fp16 = conv(bias = layers_2_fc1_module_bias_to_fp16, dilations = var_947, groups = var_715, pad = x_79_pad_0, pad_type = x_79_pad_type_0, strides = var_945, weight = layers_2_fc1_module_weight_to_fp16_palettized, x = input_61_cast_fp16)[name = tensor("x_79_cast_fp16")]; tensor layers_2_fc1_output_scale_to_fp16 = const()[name = tensor("layers_2_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170277440)))]; tensor input_63_cast_fp16 = mul(x = x_79_cast_fp16, y = layers_2_fc1_output_scale_to_fp16)[name = tensor("input_63_cast_fp16")]; tensor x_81_mode_0 = const()[name = tensor("x_81_mode_0"), val = tensor("EXACT")]; tensor x_81_cast_fp16 = gelu(mode = x_81_mode_0, x = input_63_cast_fp16)[name = tensor("x_81_cast_fp16")]; tensor layers_2_fc2_input_shift_to_fp16 = const()[name = tensor("layers_2_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170287744)))]; tensor input_65_cast_fp16 = sub(x = x_81_cast_fp16, y = layers_2_fc2_input_shift_to_fp16)[name = tensor("input_65_cast_fp16")]; tensor var_958 = const()[name = tensor("op_958"), val = tensor([1, 1])]; tensor var_960 = const()[name = tensor("op_960"), val = tensor([1, 1])]; tensor x_83_pad_type_0 = const()[name = tensor("x_83_pad_type_0"), val = tensor("custom")]; tensor x_83_pad_0 = const()[name = tensor("x_83_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_2_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170298048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173574912))), name = tensor("layers_2_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_2_fc2_module_bias_to_fp16 = const()[name = tensor("layers_2_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173575040)))]; tensor x_83_cast_fp16 = conv(bias = layers_2_fc2_module_bias_to_fp16, dilations = var_960, groups = var_715, pad = x_83_pad_0, pad_type = x_83_pad_type_0, strides = var_958, weight = layers_2_fc2_module_weight_to_fp16_palettized, x = input_65_cast_fp16)[name = tensor("x_83_cast_fp16")]; tensor layers_2_fc2_output_scale_to_fp16 = const()[name = tensor("layers_2_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173577664)))]; tensor hidden_states_7_cast_fp16 = mul(x = x_83_cast_fp16, y = layers_2_fc2_output_scale_to_fp16)[name = tensor("hidden_states_7_cast_fp16")]; tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; tensor var_974 = const()[name = tensor("op_974"), val = tensor(3)]; tensor var_981 = const()[name = tensor("op_981"), val = tensor(1)]; tensor var_982 = const()[name = tensor("op_982"), val = tensor(true)]; tensor var_994 = const()[name = tensor("op_994"), val = tensor([1])]; tensor channels_mean_19_cast_fp16 = reduce_mean(axes = var_994, keep_dims = var_982, x = inputs_19_cast_fp16)[name = tensor("channels_mean_19_cast_fp16")]; tensor zero_mean_19_cast_fp16 = sub(x = inputs_19_cast_fp16, y = channels_mean_19_cast_fp16)[name = tensor("zero_mean_19_cast_fp16")]; tensor zero_mean_sq_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = zero_mean_19_cast_fp16)[name = tensor("zero_mean_sq_19_cast_fp16")]; tensor var_998 = const()[name = tensor("op_998"), val = tensor([1])]; tensor var_999_cast_fp16 = reduce_mean(axes = var_998, keep_dims = var_982, x = zero_mean_sq_19_cast_fp16)[name = tensor("op_999_cast_fp16")]; tensor var_1000_to_fp16 = const()[name = tensor("op_1000_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1001_cast_fp16 = add(x = var_999_cast_fp16, y = var_1000_to_fp16)[name = tensor("op_1001_cast_fp16")]; tensor denom_19_epsilon_0_to_fp16 = const()[name = tensor("denom_19_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_19_cast_fp16 = rsqrt(epsilon = denom_19_epsilon_0_to_fp16, x = var_1001_cast_fp16)[name = tensor("denom_19_cast_fp16")]; tensor out_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = denom_19_cast_fp16)[name = tensor("out_19_cast_fp16")]; tensor obj_43_gamma_0_to_fp16 = const()[name = tensor("obj_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173580288)))]; tensor obj_43_beta_0_to_fp16 = const()[name = tensor("obj_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173582912)))]; tensor obj_43_epsilon_0_to_fp16 = const()[name = tensor("obj_43_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor("obj_43_cast_fp16")]; tensor layers_3_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173585536)))]; tensor input_67_cast_fp16 = sub(x = obj_43_cast_fp16, y = layers_3_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_67_cast_fp16")]; tensor var_1020 = const()[name = tensor("op_1020"), val = tensor([1, 1])]; tensor var_1022 = const()[name = tensor("op_1022"), val = tensor([1, 1])]; tensor x_85_pad_type_0 = const()[name = tensor("x_85_pad_type_0"), val = tensor("custom")]; tensor x_85_pad_0 = const()[name = tensor("x_85_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173588160))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174407424))), name = tensor("layers_3_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_3_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174407552)))]; tensor x_85_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_module_bias_to_fp16, dilations = var_1022, groups = var_981, pad = x_85_pad_0, pad_type = x_85_pad_type_0, strides = var_1020, weight = layers_3_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_67_cast_fp16)[name = tensor("x_85_cast_fp16")]; tensor layers_3_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174410176)))]; tensor query_13_cast_fp16 = mul(x = x_85_cast_fp16, y = layers_3_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_13_cast_fp16")]; tensor var_1032 = const()[name = tensor("op_1032"), val = tensor([1, 1])]; tensor var_1034 = const()[name = tensor("op_1034"), val = tensor([1, 1])]; tensor x_87_pad_type_0 = const()[name = tensor("x_87_pad_type_0"), val = tensor("custom")]; tensor x_87_pad_0 = const()[name = tensor("x_87_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174412800))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175232064))), name = tensor("layers_3_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_3_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175232192)))]; tensor x_87_cast_fp16 = conv(bias = layers_3_self_attn_k_proj_module_bias_to_fp16, dilations = var_1034, groups = var_981, pad = x_87_pad_0, pad_type = x_87_pad_type_0, strides = var_1032, weight = layers_3_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_67_cast_fp16)[name = tensor("x_87_cast_fp16")]; tensor layers_3_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_3_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175234816)))]; tensor current_key_7_cast_fp16 = mul(x = x_87_cast_fp16, y = layers_3_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_7_cast_fp16")]; tensor var_1044 = const()[name = tensor("op_1044"), val = tensor([1, 1])]; tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 1])]; tensor x_89_pad_type_0 = const()[name = tensor("x_89_pad_type_0"), val = tensor("custom")]; tensor x_89_pad_0 = const()[name = tensor("x_89_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175237440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176056704))), name = tensor("layers_3_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_3_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176056832)))]; tensor x_89_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_module_bias_to_fp16, dilations = var_1046, groups = var_981, pad = x_89_pad_0, pad_type = x_89_pad_type_0, strides = var_1044, weight = layers_3_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_67_cast_fp16)[name = tensor("x_89_cast_fp16")]; tensor layers_3_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176059456)))]; tensor current_value_7_cast_fp16 = mul(x = x_89_cast_fp16, y = layers_3_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_7_cast_fp16")]; tensor var_1054_cast_fp16 = mul(x = current_key_7_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_1054_cast_fp16")]; tensor var_1056_cast_fp16 = mul(x = var_103_cast_fp16_3, y = var_257_cast_fp16)[name = tensor("op_1056_cast_fp16")]; tensor key_13_cast_fp16 = add(x = var_1054_cast_fp16, y = var_1056_cast_fp16)[name = tensor("key_13_cast_fp16")]; tensor var_1058_cast_fp16 = mul(x = current_value_7_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_1058_cast_fp16")]; tensor var_1060_cast_fp16 = mul(x = var_138_cast_fp16_3, y = var_257_cast_fp16)[name = tensor("op_1060_cast_fp16")]; tensor value_13_cast_fp16 = add(x = var_1058_cast_fp16, y = var_1060_cast_fp16)[name = tensor("value_13_cast_fp16")]; tensor var_1063 = const()[name = tensor("op_1063"), val = tensor([1, 20, 64, -1])]; tensor var_1064_cast_fp16 = reshape(shape = var_1063, x = query_13_cast_fp16)[name = tensor("op_1064_cast_fp16")]; tensor var_1065_to_fp16 = const()[name = tensor("op_1065_to_fp16"), val = tensor(0x1p-3)]; tensor var_1066_cast_fp16 = mul(x = var_1064_cast_fp16, y = var_1065_to_fp16)[name = tensor("op_1066_cast_fp16")]; tensor var_1067 = const()[name = tensor("op_1067"), val = tensor([1, 20, 64, -1])]; tensor var_1068_cast_fp16 = reshape(shape = var_1067, x = key_13_cast_fp16)[name = tensor("op_1068_cast_fp16")]; tensor mh_w_19_transpose_x_0 = const()[name = tensor("mh_w_19_transpose_x_0"), val = tensor(true)]; tensor mh_w_19_transpose_y_0 = const()[name = tensor("mh_w_19_transpose_y_0"), val = tensor(false)]; tensor mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_1066_cast_fp16, y = var_1068_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; tensor mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; tensor var_1076_cast_fp16 = softmax(axis = var_974, x = mh_w_21_cast_fp16)[name = tensor("op_1076_cast_fp16")]; tensor var_1077 = const()[name = tensor("op_1077"), val = tensor([1, 20, 64, -1])]; tensor var_1078_cast_fp16 = reshape(shape = var_1077, x = value_13_cast_fp16)[name = tensor("op_1078_cast_fp16")]; tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(true)]; tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_1078_cast_fp16, y = var_1076_cast_fp16)[name = tensor("attn_13_cast_fp16")]; tensor var_1081 = const()[name = tensor("op_1081"), val = tensor([1, 1280, 1, -1])]; tensor x_91_cast_fp16 = reshape(shape = var_1081, x = attn_13_cast_fp16)[name = tensor("x_91_cast_fp16")]; tensor layers_3_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176062080)))]; tensor input_73_cast_fp16 = sub(x = x_91_cast_fp16, y = layers_3_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_73_cast_fp16")]; tensor var_1089 = const()[name = tensor("op_1089"), val = tensor([1, 1])]; tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([1, 1])]; tensor x_93_pad_type_0 = const()[name = tensor("x_93_pad_type_0"), val = tensor("custom")]; tensor x_93_pad_0 = const()[name = tensor("x_93_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176064704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176883968))), name = tensor("layers_3_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_3_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176884096)))]; tensor x_93_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_module_bias_to_fp16, dilations = var_1091, groups = var_981, pad = x_93_pad_0, pad_type = x_93_pad_type_0, strides = var_1089, weight = layers_3_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_73_cast_fp16)[name = tensor("x_93_cast_fp16")]; tensor layers_3_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176886720)))]; tensor obj_49_cast_fp16 = mul(x = x_93_cast_fp16, y = layers_3_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_49_cast_fp16")]; tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; tensor var_1102 = const()[name = tensor("op_1102"), val = tensor([1])]; tensor channels_mean_21_cast_fp16 = reduce_mean(axes = var_1102, keep_dims = var_982, x = inputs_21_cast_fp16)[name = tensor("channels_mean_21_cast_fp16")]; tensor zero_mean_21_cast_fp16 = sub(x = inputs_21_cast_fp16, y = channels_mean_21_cast_fp16)[name = tensor("zero_mean_21_cast_fp16")]; tensor zero_mean_sq_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = zero_mean_21_cast_fp16)[name = tensor("zero_mean_sq_21_cast_fp16")]; tensor var_1106 = const()[name = tensor("op_1106"), val = tensor([1])]; tensor var_1107_cast_fp16 = reduce_mean(axes = var_1106, keep_dims = var_982, x = zero_mean_sq_21_cast_fp16)[name = tensor("op_1107_cast_fp16")]; tensor var_1108_to_fp16 = const()[name = tensor("op_1108_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1109_cast_fp16 = add(x = var_1107_cast_fp16, y = var_1108_to_fp16)[name = tensor("op_1109_cast_fp16")]; tensor denom_21_epsilon_0_to_fp16 = const()[name = tensor("denom_21_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_21_cast_fp16 = rsqrt(epsilon = denom_21_epsilon_0_to_fp16, x = var_1109_cast_fp16)[name = tensor("denom_21_cast_fp16")]; tensor out_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = denom_21_cast_fp16)[name = tensor("out_21_cast_fp16")]; tensor obj_51_gamma_0_to_fp16 = const()[name = tensor("obj_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176889344)))]; tensor obj_51_beta_0_to_fp16 = const()[name = tensor("obj_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176891968)))]; tensor obj_51_epsilon_0_to_fp16 = const()[name = tensor("obj_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor("obj_51_cast_fp16")]; tensor layers_3_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176894592)))]; tensor input_75_cast_fp16 = sub(x = obj_51_cast_fp16, y = layers_3_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_75_cast_fp16")]; tensor var_1128 = const()[name = tensor("op_1128"), val = tensor([1, 1])]; tensor var_1130 = const()[name = tensor("op_1130"), val = tensor([1, 1])]; tensor x_95_pad_type_0 = const()[name = tensor("x_95_pad_type_0"), val = tensor("custom")]; tensor x_95_pad_0 = const()[name = tensor("x_95_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176897216))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177716480))), name = tensor("layers_3_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_3_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177716608)))]; tensor x_95_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_1130, groups = var_981, pad = x_95_pad_0, pad_type = x_95_pad_type_0, strides = var_1128, weight = layers_3_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = tensor("x_95_cast_fp16")]; tensor layers_3_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177719232)))]; tensor query_15_cast_fp16 = mul(x = x_95_cast_fp16, y = layers_3_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_15_cast_fp16")]; tensor var_1140 = const()[name = tensor("op_1140"), val = tensor([1, 1])]; tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([1, 1])]; tensor x_97_pad_type_0 = const()[name = tensor("x_97_pad_type_0"), val = tensor("custom")]; tensor x_97_pad_0 = const()[name = tensor("x_97_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177721856))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178541120))), name = tensor("layers_3_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_3_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178541248)))]; tensor x_97_cast_fp16 = conv(bias = layers_3_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_1142, groups = var_981, pad = x_97_pad_0, pad_type = x_97_pad_type_0, strides = var_1140, weight = layers_3_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_97_cast_fp16")]; tensor layers_3_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_3_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178543872)))]; tensor key_15_cast_fp16 = mul(x = x_97_cast_fp16, y = layers_3_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_15_cast_fp16")]; tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 1])]; tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 1])]; tensor x_99_pad_type_0 = const()[name = tensor("x_99_pad_type_0"), val = tensor("custom")]; tensor x_99_pad_0 = const()[name = tensor("x_99_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178546496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179365760))), name = tensor("layers_3_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_3_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179365888)))]; tensor x_99_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_1154, groups = var_981, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = var_1152, weight = layers_3_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_99_cast_fp16")]; tensor layers_3_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179368512)))]; tensor value_15_cast_fp16 = mul(x = x_99_cast_fp16, y = layers_3_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_15_cast_fp16")]; tensor var_1159 = const()[name = tensor("op_1159"), val = tensor([1, 20, 64, -1])]; tensor var_1160_cast_fp16 = reshape(shape = var_1159, x = query_15_cast_fp16)[name = tensor("op_1160_cast_fp16")]; tensor var_1161_to_fp16 = const()[name = tensor("op_1161_to_fp16"), val = tensor(0x1p-3)]; tensor var_1162_cast_fp16 = mul(x = var_1160_cast_fp16, y = var_1161_to_fp16)[name = tensor("op_1162_cast_fp16")]; tensor var_1163 = const()[name = tensor("op_1163"), val = tensor([1, 20, 64, -1])]; tensor var_1164_cast_fp16 = reshape(shape = var_1163, x = key_15_cast_fp16)[name = tensor("op_1164_cast_fp16")]; tensor mh_w_23_transpose_x_0 = const()[name = tensor("mh_w_23_transpose_x_0"), val = tensor(true)]; tensor mh_w_23_transpose_y_0 = const()[name = tensor("mh_w_23_transpose_y_0"), val = tensor(false)]; tensor mh_w_23_cast_fp16 = matmul(transpose_x = mh_w_23_transpose_x_0, transpose_y = mh_w_23_transpose_y_0, x = var_1162_cast_fp16, y = var_1164_cast_fp16)[name = tensor("mh_w_23_cast_fp16")]; tensor obj_55_cast_fp16 = softmax(axis = var_974, x = mh_w_23_cast_fp16)[name = tensor("obj_55_cast_fp16")]; tensor var_1168 = const()[name = tensor("op_1168"), val = tensor([1, 20, 64, -1])]; tensor var_1169_cast_fp16 = reshape(shape = var_1168, x = value_15_cast_fp16)[name = tensor("op_1169_cast_fp16")]; tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(true)]; tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_1169_cast_fp16, y = obj_55_cast_fp16)[name = tensor("attn_15_cast_fp16")]; tensor var_1172 = const()[name = tensor("op_1172"), val = tensor([1, 1280, 1, -1])]; tensor x_101_cast_fp16 = reshape(shape = var_1172, x = attn_15_cast_fp16)[name = tensor("x_101_cast_fp16")]; tensor layers_3_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179371136)))]; tensor input_81_cast_fp16 = sub(x = x_101_cast_fp16, y = layers_3_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_81_cast_fp16")]; tensor var_1180 = const()[name = tensor("op_1180"), val = tensor([1, 1])]; tensor var_1182 = const()[name = tensor("op_1182"), val = tensor([1, 1])]; tensor x_103_pad_type_0 = const()[name = tensor("x_103_pad_type_0"), val = tensor("custom")]; tensor x_103_pad_0 = const()[name = tensor("x_103_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179373760))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180193024))), name = tensor("layers_3_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_3_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180193152)))]; tensor x_103_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_1182, groups = var_981, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = var_1180, weight = layers_3_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_81_cast_fp16)[name = tensor("x_103_cast_fp16")]; tensor layers_3_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180195776)))]; tensor obj_53_cast_fp16 = mul(x = x_103_cast_fp16, y = layers_3_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_53_cast_fp16")]; tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; tensor var_1189 = const()[name = tensor("op_1189"), val = tensor([1])]; tensor channels_mean_23_cast_fp16 = reduce_mean(axes = var_1189, keep_dims = var_982, x = inputs_23_cast_fp16)[name = tensor("channels_mean_23_cast_fp16")]; tensor zero_mean_23_cast_fp16 = sub(x = inputs_23_cast_fp16, y = channels_mean_23_cast_fp16)[name = tensor("zero_mean_23_cast_fp16")]; tensor zero_mean_sq_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = zero_mean_23_cast_fp16)[name = tensor("zero_mean_sq_23_cast_fp16")]; tensor var_1193 = const()[name = tensor("op_1193"), val = tensor([1])]; tensor var_1194_cast_fp16 = reduce_mean(axes = var_1193, keep_dims = var_982, x = zero_mean_sq_23_cast_fp16)[name = tensor("op_1194_cast_fp16")]; tensor var_1195_to_fp16 = const()[name = tensor("op_1195_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1196_cast_fp16 = add(x = var_1194_cast_fp16, y = var_1195_to_fp16)[name = tensor("op_1196_cast_fp16")]; tensor denom_23_epsilon_0_to_fp16 = const()[name = tensor("denom_23_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_23_cast_fp16 = rsqrt(epsilon = denom_23_epsilon_0_to_fp16, x = var_1196_cast_fp16)[name = tensor("denom_23_cast_fp16")]; tensor out_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = denom_23_cast_fp16)[name = tensor("out_23_cast_fp16")]; tensor x_105_gamma_0_to_fp16 = const()[name = tensor("x_105_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180198400)))]; tensor x_105_beta_0_to_fp16 = const()[name = tensor("x_105_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180201024)))]; tensor x_105_epsilon_0_to_fp16 = const()[name = tensor("x_105_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_105_cast_fp16 = batch_norm(beta = x_105_beta_0_to_fp16, epsilon = x_105_epsilon_0_to_fp16, gamma = x_105_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor("x_105_cast_fp16")]; tensor layers_3_fc1_input_shift_to_fp16 = const()[name = tensor("layers_3_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180203648)))]; tensor input_83_cast_fp16 = sub(x = x_105_cast_fp16, y = layers_3_fc1_input_shift_to_fp16)[name = tensor("input_83_cast_fp16")]; tensor var_1211 = const()[name = tensor("op_1211"), val = tensor([1, 1])]; tensor var_1213 = const()[name = tensor("op_1213"), val = tensor([1, 1])]; tensor x_107_pad_type_0 = const()[name = tensor("x_107_pad_type_0"), val = tensor("custom")]; tensor x_107_pad_0 = const()[name = tensor("x_107_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180206272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183483136))), name = tensor("layers_3_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_3_fc1_module_bias_to_fp16 = const()[name = tensor("layers_3_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183483264)))]; tensor x_107_cast_fp16 = conv(bias = layers_3_fc1_module_bias_to_fp16, dilations = var_1213, groups = var_981, pad = x_107_pad_0, pad_type = x_107_pad_type_0, strides = var_1211, weight = layers_3_fc1_module_weight_to_fp16_palettized, x = input_83_cast_fp16)[name = tensor("x_107_cast_fp16")]; tensor layers_3_fc1_output_scale_to_fp16 = const()[name = tensor("layers_3_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183493568)))]; tensor input_85_cast_fp16 = mul(x = x_107_cast_fp16, y = layers_3_fc1_output_scale_to_fp16)[name = tensor("input_85_cast_fp16")]; tensor x_109_mode_0 = const()[name = tensor("x_109_mode_0"), val = tensor("EXACT")]; tensor x_109_cast_fp16 = gelu(mode = x_109_mode_0, x = input_85_cast_fp16)[name = tensor("x_109_cast_fp16")]; tensor layers_3_fc2_input_shift_to_fp16 = const()[name = tensor("layers_3_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183503872)))]; tensor input_87_cast_fp16 = sub(x = x_109_cast_fp16, y = layers_3_fc2_input_shift_to_fp16)[name = tensor("input_87_cast_fp16")]; tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([1, 1])]; tensor var_1226 = const()[name = tensor("op_1226"), val = tensor([1, 1])]; tensor x_111_pad_type_0 = const()[name = tensor("x_111_pad_type_0"), val = tensor("custom")]; tensor x_111_pad_0 = const()[name = tensor("x_111_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_3_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183514176))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186791040))), name = tensor("layers_3_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_3_fc2_module_bias_to_fp16 = const()[name = tensor("layers_3_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186791168)))]; tensor x_111_cast_fp16 = conv(bias = layers_3_fc2_module_bias_to_fp16, dilations = var_1226, groups = var_981, pad = x_111_pad_0, pad_type = x_111_pad_type_0, strides = var_1224, weight = layers_3_fc2_module_weight_to_fp16_palettized, x = input_87_cast_fp16)[name = tensor("x_111_cast_fp16")]; tensor layers_3_fc2_output_scale_to_fp16 = const()[name = tensor("layers_3_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186793792)))]; tensor hidden_states_9_cast_fp16 = mul(x = x_111_cast_fp16, y = layers_3_fc2_output_scale_to_fp16)[name = tensor("hidden_states_9_cast_fp16")]; tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; tensor var_1240 = const()[name = tensor("op_1240"), val = tensor(3)]; tensor var_1247 = const()[name = tensor("op_1247"), val = tensor(1)]; tensor var_1248 = const()[name = tensor("op_1248"), val = tensor(true)]; tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([1])]; tensor channels_mean_25_cast_fp16 = reduce_mean(axes = var_1260, keep_dims = var_1248, x = inputs_25_cast_fp16)[name = tensor("channels_mean_25_cast_fp16")]; tensor zero_mean_25_cast_fp16 = sub(x = inputs_25_cast_fp16, y = channels_mean_25_cast_fp16)[name = tensor("zero_mean_25_cast_fp16")]; tensor zero_mean_sq_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = zero_mean_25_cast_fp16)[name = tensor("zero_mean_sq_25_cast_fp16")]; tensor var_1264 = const()[name = tensor("op_1264"), val = tensor([1])]; tensor var_1265_cast_fp16 = reduce_mean(axes = var_1264, keep_dims = var_1248, x = zero_mean_sq_25_cast_fp16)[name = tensor("op_1265_cast_fp16")]; tensor var_1266_to_fp16 = const()[name = tensor("op_1266_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1267_cast_fp16 = add(x = var_1265_cast_fp16, y = var_1266_to_fp16)[name = tensor("op_1267_cast_fp16")]; tensor denom_25_epsilon_0_to_fp16 = const()[name = tensor("denom_25_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_25_cast_fp16 = rsqrt(epsilon = denom_25_epsilon_0_to_fp16, x = var_1267_cast_fp16)[name = tensor("denom_25_cast_fp16")]; tensor out_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = denom_25_cast_fp16)[name = tensor("out_25_cast_fp16")]; tensor obj_57_gamma_0_to_fp16 = const()[name = tensor("obj_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186796416)))]; tensor obj_57_beta_0_to_fp16 = const()[name = tensor("obj_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186799040)))]; tensor obj_57_epsilon_0_to_fp16 = const()[name = tensor("obj_57_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_57_cast_fp16 = batch_norm(beta = obj_57_beta_0_to_fp16, epsilon = obj_57_epsilon_0_to_fp16, gamma = obj_57_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor("obj_57_cast_fp16")]; tensor layers_4_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186801664)))]; tensor input_89_cast_fp16 = sub(x = obj_57_cast_fp16, y = layers_4_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_89_cast_fp16")]; tensor var_1286 = const()[name = tensor("op_1286"), val = tensor([1, 1])]; tensor var_1288 = const()[name = tensor("op_1288"), val = tensor([1, 1])]; tensor x_113_pad_type_0 = const()[name = tensor("x_113_pad_type_0"), val = tensor("custom")]; tensor x_113_pad_0 = const()[name = tensor("x_113_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186804288))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187623552))), name = tensor("layers_4_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_4_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187623680)))]; tensor x_113_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_module_bias_to_fp16, dilations = var_1288, groups = var_1247, pad = x_113_pad_0, pad_type = x_113_pad_type_0, strides = var_1286, weight = layers_4_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_89_cast_fp16)[name = tensor("x_113_cast_fp16")]; tensor layers_4_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187626304)))]; tensor query_17_cast_fp16 = mul(x = x_113_cast_fp16, y = layers_4_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_17_cast_fp16")]; tensor var_1298 = const()[name = tensor("op_1298"), val = tensor([1, 1])]; tensor var_1300 = const()[name = tensor("op_1300"), val = tensor([1, 1])]; tensor x_115_pad_type_0 = const()[name = tensor("x_115_pad_type_0"), val = tensor("custom")]; tensor x_115_pad_0 = const()[name = tensor("x_115_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187628928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188448192))), name = tensor("layers_4_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_4_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188448320)))]; tensor x_115_cast_fp16 = conv(bias = layers_4_self_attn_k_proj_module_bias_to_fp16, dilations = var_1300, groups = var_1247, pad = x_115_pad_0, pad_type = x_115_pad_type_0, strides = var_1298, weight = layers_4_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_89_cast_fp16)[name = tensor("x_115_cast_fp16")]; tensor layers_4_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_4_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188450944)))]; tensor current_key_9_cast_fp16 = mul(x = x_115_cast_fp16, y = layers_4_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_9_cast_fp16")]; tensor var_1310 = const()[name = tensor("op_1310"), val = tensor([1, 1])]; tensor var_1312 = const()[name = tensor("op_1312"), val = tensor([1, 1])]; tensor x_117_pad_type_0 = const()[name = tensor("x_117_pad_type_0"), val = tensor("custom")]; tensor x_117_pad_0 = const()[name = tensor("x_117_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188453568))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189272832))), name = tensor("layers_4_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_4_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189272960)))]; tensor x_117_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_module_bias_to_fp16, dilations = var_1312, groups = var_1247, pad = x_117_pad_0, pad_type = x_117_pad_type_0, strides = var_1310, weight = layers_4_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_89_cast_fp16)[name = tensor("x_117_cast_fp16")]; tensor layers_4_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189275584)))]; tensor current_value_9_cast_fp16 = mul(x = x_117_cast_fp16, y = layers_4_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_9_cast_fp16")]; tensor var_1320_cast_fp16 = mul(x = current_key_9_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_1320_cast_fp16")]; tensor var_1322_cast_fp16 = mul(x = var_103_cast_fp16_4, y = var_257_cast_fp16)[name = tensor("op_1322_cast_fp16")]; tensor key_17_cast_fp16 = add(x = var_1320_cast_fp16, y = var_1322_cast_fp16)[name = tensor("key_17_cast_fp16")]; tensor var_1324_cast_fp16 = mul(x = current_value_9_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_1324_cast_fp16")]; tensor var_1326_cast_fp16 = mul(x = var_138_cast_fp16_4, y = var_257_cast_fp16)[name = tensor("op_1326_cast_fp16")]; tensor value_17_cast_fp16 = add(x = var_1324_cast_fp16, y = var_1326_cast_fp16)[name = tensor("value_17_cast_fp16")]; tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([1, 20, 64, -1])]; tensor var_1330_cast_fp16 = reshape(shape = var_1329, x = query_17_cast_fp16)[name = tensor("op_1330_cast_fp16")]; tensor var_1331_to_fp16 = const()[name = tensor("op_1331_to_fp16"), val = tensor(0x1p-3)]; tensor var_1332_cast_fp16 = mul(x = var_1330_cast_fp16, y = var_1331_to_fp16)[name = tensor("op_1332_cast_fp16")]; tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([1, 20, 64, -1])]; tensor var_1334_cast_fp16 = reshape(shape = var_1333, x = key_17_cast_fp16)[name = tensor("op_1334_cast_fp16")]; tensor mh_w_25_transpose_x_0 = const()[name = tensor("mh_w_25_transpose_x_0"), val = tensor(true)]; tensor mh_w_25_transpose_y_0 = const()[name = tensor("mh_w_25_transpose_y_0"), val = tensor(false)]; tensor mh_w_25_cast_fp16 = matmul(transpose_x = mh_w_25_transpose_x_0, transpose_y = mh_w_25_transpose_y_0, x = var_1332_cast_fp16, y = var_1334_cast_fp16)[name = tensor("mh_w_25_cast_fp16")]; tensor mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_27_cast_fp16")]; tensor var_1342_cast_fp16 = softmax(axis = var_1240, x = mh_w_27_cast_fp16)[name = tensor("op_1342_cast_fp16")]; tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([1, 20, 64, -1])]; tensor var_1344_cast_fp16 = reshape(shape = var_1343, x = value_17_cast_fp16)[name = tensor("op_1344_cast_fp16")]; tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(true)]; tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_1344_cast_fp16, y = var_1342_cast_fp16)[name = tensor("attn_17_cast_fp16")]; tensor var_1347 = const()[name = tensor("op_1347"), val = tensor([1, 1280, 1, -1])]; tensor x_119_cast_fp16 = reshape(shape = var_1347, x = attn_17_cast_fp16)[name = tensor("x_119_cast_fp16")]; tensor layers_4_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189278208)))]; tensor input_95_cast_fp16 = sub(x = x_119_cast_fp16, y = layers_4_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_95_cast_fp16")]; tensor var_1355 = const()[name = tensor("op_1355"), val = tensor([1, 1])]; tensor var_1357 = const()[name = tensor("op_1357"), val = tensor([1, 1])]; tensor x_121_pad_type_0 = const()[name = tensor("x_121_pad_type_0"), val = tensor("custom")]; tensor x_121_pad_0 = const()[name = tensor("x_121_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189280832))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190100096))), name = tensor("layers_4_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_4_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190100224)))]; tensor x_121_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_module_bias_to_fp16, dilations = var_1357, groups = var_1247, pad = x_121_pad_0, pad_type = x_121_pad_type_0, strides = var_1355, weight = layers_4_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_95_cast_fp16)[name = tensor("x_121_cast_fp16")]; tensor layers_4_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190102848)))]; tensor obj_63_cast_fp16 = mul(x = x_121_cast_fp16, y = layers_4_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_63_cast_fp16")]; tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_63_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([1])]; tensor channels_mean_27_cast_fp16 = reduce_mean(axes = var_1368, keep_dims = var_1248, x = inputs_27_cast_fp16)[name = tensor("channels_mean_27_cast_fp16")]; tensor zero_mean_27_cast_fp16 = sub(x = inputs_27_cast_fp16, y = channels_mean_27_cast_fp16)[name = tensor("zero_mean_27_cast_fp16")]; tensor zero_mean_sq_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = zero_mean_27_cast_fp16)[name = tensor("zero_mean_sq_27_cast_fp16")]; tensor var_1372 = const()[name = tensor("op_1372"), val = tensor([1])]; tensor var_1373_cast_fp16 = reduce_mean(axes = var_1372, keep_dims = var_1248, x = zero_mean_sq_27_cast_fp16)[name = tensor("op_1373_cast_fp16")]; tensor var_1374_to_fp16 = const()[name = tensor("op_1374_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1375_cast_fp16 = add(x = var_1373_cast_fp16, y = var_1374_to_fp16)[name = tensor("op_1375_cast_fp16")]; tensor denom_27_epsilon_0_to_fp16 = const()[name = tensor("denom_27_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_27_cast_fp16 = rsqrt(epsilon = denom_27_epsilon_0_to_fp16, x = var_1375_cast_fp16)[name = tensor("denom_27_cast_fp16")]; tensor out_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = denom_27_cast_fp16)[name = tensor("out_27_cast_fp16")]; tensor obj_65_gamma_0_to_fp16 = const()[name = tensor("obj_65_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190105472)))]; tensor obj_65_beta_0_to_fp16 = const()[name = tensor("obj_65_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190108096)))]; tensor obj_65_epsilon_0_to_fp16 = const()[name = tensor("obj_65_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_65_cast_fp16 = batch_norm(beta = obj_65_beta_0_to_fp16, epsilon = obj_65_epsilon_0_to_fp16, gamma = obj_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor("obj_65_cast_fp16")]; tensor layers_4_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190110720)))]; tensor input_97_cast_fp16 = sub(x = obj_65_cast_fp16, y = layers_4_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_97_cast_fp16")]; tensor var_1394 = const()[name = tensor("op_1394"), val = tensor([1, 1])]; tensor var_1396 = const()[name = tensor("op_1396"), val = tensor([1, 1])]; tensor x_123_pad_type_0 = const()[name = tensor("x_123_pad_type_0"), val = tensor("custom")]; tensor x_123_pad_0 = const()[name = tensor("x_123_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190113344))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190932608))), name = tensor("layers_4_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_4_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190932736)))]; tensor x_123_cast_fp16 = conv(bias = layers_4_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_1396, groups = var_1247, pad = x_123_pad_0, pad_type = x_123_pad_type_0, strides = var_1394, weight = layers_4_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_97_cast_fp16)[name = tensor("x_123_cast_fp16")]; tensor layers_4_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190935360)))]; tensor query_19_cast_fp16 = mul(x = x_123_cast_fp16, y = layers_4_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_19_cast_fp16")]; tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([1, 1])]; tensor var_1408 = const()[name = tensor("op_1408"), val = tensor([1, 1])]; tensor x_125_pad_type_0 = const()[name = tensor("x_125_pad_type_0"), val = tensor("custom")]; tensor x_125_pad_0 = const()[name = tensor("x_125_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190937984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191757248))), name = tensor("layers_4_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_4_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191757376)))]; tensor x_125_cast_fp16 = conv(bias = layers_4_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_1408, groups = var_1247, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = var_1406, weight = layers_4_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_125_cast_fp16")]; tensor layers_4_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_4_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191760000)))]; tensor key_19_cast_fp16 = mul(x = x_125_cast_fp16, y = layers_4_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_19_cast_fp16")]; tensor var_1418 = const()[name = tensor("op_1418"), val = tensor([1, 1])]; tensor var_1420 = const()[name = tensor("op_1420"), val = tensor([1, 1])]; tensor x_127_pad_type_0 = const()[name = tensor("x_127_pad_type_0"), val = tensor("custom")]; tensor x_127_pad_0 = const()[name = tensor("x_127_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191762624))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192581888))), name = tensor("layers_4_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_4_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192582016)))]; tensor x_127_cast_fp16 = conv(bias = layers_4_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_1420, groups = var_1247, pad = x_127_pad_0, pad_type = x_127_pad_type_0, strides = var_1418, weight = layers_4_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_127_cast_fp16")]; tensor layers_4_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_4_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192584640)))]; tensor value_19_cast_fp16 = mul(x = x_127_cast_fp16, y = layers_4_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_19_cast_fp16")]; tensor var_1425 = const()[name = tensor("op_1425"), val = tensor([1, 20, 64, -1])]; tensor var_1426_cast_fp16 = reshape(shape = var_1425, x = query_19_cast_fp16)[name = tensor("op_1426_cast_fp16")]; tensor var_1427_to_fp16 = const()[name = tensor("op_1427_to_fp16"), val = tensor(0x1p-3)]; tensor var_1428_cast_fp16 = mul(x = var_1426_cast_fp16, y = var_1427_to_fp16)[name = tensor("op_1428_cast_fp16")]; tensor var_1429 = const()[name = tensor("op_1429"), val = tensor([1, 20, 64, -1])]; tensor var_1430_cast_fp16 = reshape(shape = var_1429, x = key_19_cast_fp16)[name = tensor("op_1430_cast_fp16")]; tensor mh_w_29_transpose_x_0 = const()[name = tensor("mh_w_29_transpose_x_0"), val = tensor(true)]; tensor mh_w_29_transpose_y_0 = const()[name = tensor("mh_w_29_transpose_y_0"), val = tensor(false)]; tensor mh_w_29_cast_fp16 = matmul(transpose_x = mh_w_29_transpose_x_0, transpose_y = mh_w_29_transpose_y_0, x = var_1428_cast_fp16, y = var_1430_cast_fp16)[name = tensor("mh_w_29_cast_fp16")]; tensor obj_69_cast_fp16 = softmax(axis = var_1240, x = mh_w_29_cast_fp16)[name = tensor("obj_69_cast_fp16")]; tensor var_1434 = const()[name = tensor("op_1434"), val = tensor([1, 20, 64, -1])]; tensor var_1435_cast_fp16 = reshape(shape = var_1434, x = value_19_cast_fp16)[name = tensor("op_1435_cast_fp16")]; tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(true)]; tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_1435_cast_fp16, y = obj_69_cast_fp16)[name = tensor("attn_19_cast_fp16")]; tensor var_1438 = const()[name = tensor("op_1438"), val = tensor([1, 1280, 1, -1])]; tensor x_129_cast_fp16 = reshape(shape = var_1438, x = attn_19_cast_fp16)[name = tensor("x_129_cast_fp16")]; tensor layers_4_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192587264)))]; tensor input_103_cast_fp16 = sub(x = x_129_cast_fp16, y = layers_4_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_103_cast_fp16")]; tensor var_1446 = const()[name = tensor("op_1446"), val = tensor([1, 1])]; tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([1, 1])]; tensor x_131_pad_type_0 = const()[name = tensor("x_131_pad_type_0"), val = tensor("custom")]; tensor x_131_pad_0 = const()[name = tensor("x_131_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192589888))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193409152))), name = tensor("layers_4_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_4_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193409280)))]; tensor x_131_cast_fp16 = conv(bias = layers_4_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_1448, groups = var_1247, pad = x_131_pad_0, pad_type = x_131_pad_type_0, strides = var_1446, weight = layers_4_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_103_cast_fp16)[name = tensor("x_131_cast_fp16")]; tensor layers_4_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193411904)))]; tensor obj_67_cast_fp16 = mul(x = x_131_cast_fp16, y = layers_4_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_67_cast_fp16")]; tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = obj_67_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([1])]; tensor channels_mean_29_cast_fp16 = reduce_mean(axes = var_1455, keep_dims = var_1248, x = inputs_29_cast_fp16)[name = tensor("channels_mean_29_cast_fp16")]; tensor zero_mean_29_cast_fp16 = sub(x = inputs_29_cast_fp16, y = channels_mean_29_cast_fp16)[name = tensor("zero_mean_29_cast_fp16")]; tensor zero_mean_sq_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = zero_mean_29_cast_fp16)[name = tensor("zero_mean_sq_29_cast_fp16")]; tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([1])]; tensor var_1460_cast_fp16 = reduce_mean(axes = var_1459, keep_dims = var_1248, x = zero_mean_sq_29_cast_fp16)[name = tensor("op_1460_cast_fp16")]; tensor var_1461_to_fp16 = const()[name = tensor("op_1461_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1462_cast_fp16 = add(x = var_1460_cast_fp16, y = var_1461_to_fp16)[name = tensor("op_1462_cast_fp16")]; tensor denom_29_epsilon_0_to_fp16 = const()[name = tensor("denom_29_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_29_cast_fp16 = rsqrt(epsilon = denom_29_epsilon_0_to_fp16, x = var_1462_cast_fp16)[name = tensor("denom_29_cast_fp16")]; tensor out_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = denom_29_cast_fp16)[name = tensor("out_29_cast_fp16")]; tensor x_133_gamma_0_to_fp16 = const()[name = tensor("x_133_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193414528)))]; tensor x_133_beta_0_to_fp16 = const()[name = tensor("x_133_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193417152)))]; tensor x_133_epsilon_0_to_fp16 = const()[name = tensor("x_133_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_133_cast_fp16 = batch_norm(beta = x_133_beta_0_to_fp16, epsilon = x_133_epsilon_0_to_fp16, gamma = x_133_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor("x_133_cast_fp16")]; tensor layers_4_fc1_input_shift_to_fp16 = const()[name = tensor("layers_4_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193419776)))]; tensor input_105_cast_fp16 = sub(x = x_133_cast_fp16, y = layers_4_fc1_input_shift_to_fp16)[name = tensor("input_105_cast_fp16")]; tensor var_1477 = const()[name = tensor("op_1477"), val = tensor([1, 1])]; tensor var_1479 = const()[name = tensor("op_1479"), val = tensor([1, 1])]; tensor x_135_pad_type_0 = const()[name = tensor("x_135_pad_type_0"), val = tensor("custom")]; tensor x_135_pad_0 = const()[name = tensor("x_135_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193422400))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196699264))), name = tensor("layers_4_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_4_fc1_module_bias_to_fp16 = const()[name = tensor("layers_4_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196699392)))]; tensor x_135_cast_fp16 = conv(bias = layers_4_fc1_module_bias_to_fp16, dilations = var_1479, groups = var_1247, pad = x_135_pad_0, pad_type = x_135_pad_type_0, strides = var_1477, weight = layers_4_fc1_module_weight_to_fp16_palettized, x = input_105_cast_fp16)[name = tensor("x_135_cast_fp16")]; tensor layers_4_fc1_output_scale_to_fp16 = const()[name = tensor("layers_4_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196709696)))]; tensor input_107_cast_fp16 = mul(x = x_135_cast_fp16, y = layers_4_fc1_output_scale_to_fp16)[name = tensor("input_107_cast_fp16")]; tensor x_137_mode_0 = const()[name = tensor("x_137_mode_0"), val = tensor("EXACT")]; tensor x_137_cast_fp16 = gelu(mode = x_137_mode_0, x = input_107_cast_fp16)[name = tensor("x_137_cast_fp16")]; tensor layers_4_fc2_input_shift_to_fp16 = const()[name = tensor("layers_4_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196720000)))]; tensor input_109_cast_fp16 = sub(x = x_137_cast_fp16, y = layers_4_fc2_input_shift_to_fp16)[name = tensor("input_109_cast_fp16")]; tensor var_1490 = const()[name = tensor("op_1490"), val = tensor([1, 1])]; tensor var_1492 = const()[name = tensor("op_1492"), val = tensor([1, 1])]; tensor x_139_pad_type_0 = const()[name = tensor("x_139_pad_type_0"), val = tensor("custom")]; tensor x_139_pad_0 = const()[name = tensor("x_139_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_4_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196730304))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200007168))), name = tensor("layers_4_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_4_fc2_module_bias_to_fp16 = const()[name = tensor("layers_4_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200007296)))]; tensor x_139_cast_fp16 = conv(bias = layers_4_fc2_module_bias_to_fp16, dilations = var_1492, groups = var_1247, pad = x_139_pad_0, pad_type = x_139_pad_type_0, strides = var_1490, weight = layers_4_fc2_module_weight_to_fp16_palettized, x = input_109_cast_fp16)[name = tensor("x_139_cast_fp16")]; tensor layers_4_fc2_output_scale_to_fp16 = const()[name = tensor("layers_4_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200009920)))]; tensor hidden_states_11_cast_fp16 = mul(x = x_139_cast_fp16, y = layers_4_fc2_output_scale_to_fp16)[name = tensor("hidden_states_11_cast_fp16")]; tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = hidden_states_11_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; tensor var_1506 = const()[name = tensor("op_1506"), val = tensor(3)]; tensor var_1513 = const()[name = tensor("op_1513"), val = tensor(1)]; tensor var_1514 = const()[name = tensor("op_1514"), val = tensor(true)]; tensor var_1526 = const()[name = tensor("op_1526"), val = tensor([1])]; tensor channels_mean_31_cast_fp16 = reduce_mean(axes = var_1526, keep_dims = var_1514, x = inputs_31_cast_fp16)[name = tensor("channels_mean_31_cast_fp16")]; tensor zero_mean_31_cast_fp16 = sub(x = inputs_31_cast_fp16, y = channels_mean_31_cast_fp16)[name = tensor("zero_mean_31_cast_fp16")]; tensor zero_mean_sq_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = zero_mean_31_cast_fp16)[name = tensor("zero_mean_sq_31_cast_fp16")]; tensor var_1530 = const()[name = tensor("op_1530"), val = tensor([1])]; tensor var_1531_cast_fp16 = reduce_mean(axes = var_1530, keep_dims = var_1514, x = zero_mean_sq_31_cast_fp16)[name = tensor("op_1531_cast_fp16")]; tensor var_1532_to_fp16 = const()[name = tensor("op_1532_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1533_cast_fp16 = add(x = var_1531_cast_fp16, y = var_1532_to_fp16)[name = tensor("op_1533_cast_fp16")]; tensor denom_31_epsilon_0_to_fp16 = const()[name = tensor("denom_31_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_31_cast_fp16 = rsqrt(epsilon = denom_31_epsilon_0_to_fp16, x = var_1533_cast_fp16)[name = tensor("denom_31_cast_fp16")]; tensor out_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = denom_31_cast_fp16)[name = tensor("out_31_cast_fp16")]; tensor obj_71_gamma_0_to_fp16 = const()[name = tensor("obj_71_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200012544)))]; tensor obj_71_beta_0_to_fp16 = const()[name = tensor("obj_71_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200015168)))]; tensor obj_71_epsilon_0_to_fp16 = const()[name = tensor("obj_71_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_71_cast_fp16 = batch_norm(beta = obj_71_beta_0_to_fp16, epsilon = obj_71_epsilon_0_to_fp16, gamma = obj_71_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor("obj_71_cast_fp16")]; tensor layers_5_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200017792)))]; tensor input_111_cast_fp16 = sub(x = obj_71_cast_fp16, y = layers_5_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_111_cast_fp16")]; tensor var_1552 = const()[name = tensor("op_1552"), val = tensor([1, 1])]; tensor var_1554 = const()[name = tensor("op_1554"), val = tensor([1, 1])]; tensor x_141_pad_type_0 = const()[name = tensor("x_141_pad_type_0"), val = tensor("custom")]; tensor x_141_pad_0 = const()[name = tensor("x_141_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200020416))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200839680))), name = tensor("layers_5_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_5_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200839808)))]; tensor x_141_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_module_bias_to_fp16, dilations = var_1554, groups = var_1513, pad = x_141_pad_0, pad_type = x_141_pad_type_0, strides = var_1552, weight = layers_5_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = tensor("x_141_cast_fp16")]; tensor layers_5_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200842432)))]; tensor query_21_cast_fp16 = mul(x = x_141_cast_fp16, y = layers_5_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_21_cast_fp16")]; tensor var_1564 = const()[name = tensor("op_1564"), val = tensor([1, 1])]; tensor var_1566 = const()[name = tensor("op_1566"), val = tensor([1, 1])]; tensor x_143_pad_type_0 = const()[name = tensor("x_143_pad_type_0"), val = tensor("custom")]; tensor x_143_pad_0 = const()[name = tensor("x_143_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200845056))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201664320))), name = tensor("layers_5_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_5_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201664448)))]; tensor x_143_cast_fp16 = conv(bias = layers_5_self_attn_k_proj_module_bias_to_fp16, dilations = var_1566, groups = var_1513, pad = x_143_pad_0, pad_type = x_143_pad_type_0, strides = var_1564, weight = layers_5_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = tensor("x_143_cast_fp16")]; tensor layers_5_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_5_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201667072)))]; tensor current_key_11_cast_fp16 = mul(x = x_143_cast_fp16, y = layers_5_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_11_cast_fp16")]; tensor var_1576 = const()[name = tensor("op_1576"), val = tensor([1, 1])]; tensor var_1578 = const()[name = tensor("op_1578"), val = tensor([1, 1])]; tensor x_145_pad_type_0 = const()[name = tensor("x_145_pad_type_0"), val = tensor("custom")]; tensor x_145_pad_0 = const()[name = tensor("x_145_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201669696))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202488960))), name = tensor("layers_5_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_5_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202489088)))]; tensor x_145_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_module_bias_to_fp16, dilations = var_1578, groups = var_1513, pad = x_145_pad_0, pad_type = x_145_pad_type_0, strides = var_1576, weight = layers_5_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = tensor("x_145_cast_fp16")]; tensor layers_5_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202491712)))]; tensor current_value_11_cast_fp16 = mul(x = x_145_cast_fp16, y = layers_5_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_11_cast_fp16")]; tensor var_1586_cast_fp16 = mul(x = current_key_11_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_1586_cast_fp16")]; tensor var_1588_cast_fp16 = mul(x = var_103_cast_fp16_5, y = var_257_cast_fp16)[name = tensor("op_1588_cast_fp16")]; tensor key_21_cast_fp16 = add(x = var_1586_cast_fp16, y = var_1588_cast_fp16)[name = tensor("key_21_cast_fp16")]; tensor var_1590_cast_fp16 = mul(x = current_value_11_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_1590_cast_fp16")]; tensor var_1592_cast_fp16 = mul(x = var_138_cast_fp16_5, y = var_257_cast_fp16)[name = tensor("op_1592_cast_fp16")]; tensor value_21_cast_fp16 = add(x = var_1590_cast_fp16, y = var_1592_cast_fp16)[name = tensor("value_21_cast_fp16")]; tensor var_1595 = const()[name = tensor("op_1595"), val = tensor([1, 20, 64, -1])]; tensor var_1596_cast_fp16 = reshape(shape = var_1595, x = query_21_cast_fp16)[name = tensor("op_1596_cast_fp16")]; tensor var_1597_to_fp16 = const()[name = tensor("op_1597_to_fp16"), val = tensor(0x1p-3)]; tensor var_1598_cast_fp16 = mul(x = var_1596_cast_fp16, y = var_1597_to_fp16)[name = tensor("op_1598_cast_fp16")]; tensor var_1599 = const()[name = tensor("op_1599"), val = tensor([1, 20, 64, -1])]; tensor var_1600_cast_fp16 = reshape(shape = var_1599, x = key_21_cast_fp16)[name = tensor("op_1600_cast_fp16")]; tensor mh_w_31_transpose_x_0 = const()[name = tensor("mh_w_31_transpose_x_0"), val = tensor(true)]; tensor mh_w_31_transpose_y_0 = const()[name = tensor("mh_w_31_transpose_y_0"), val = tensor(false)]; tensor mh_w_31_cast_fp16 = matmul(transpose_x = mh_w_31_transpose_x_0, transpose_y = mh_w_31_transpose_y_0, x = var_1598_cast_fp16, y = var_1600_cast_fp16)[name = tensor("mh_w_31_cast_fp16")]; tensor mh_w_33_cast_fp16 = add(x = mh_w_31_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_33_cast_fp16")]; tensor var_1608_cast_fp16 = softmax(axis = var_1506, x = mh_w_33_cast_fp16)[name = tensor("op_1608_cast_fp16")]; tensor var_1609 = const()[name = tensor("op_1609"), val = tensor([1, 20, 64, -1])]; tensor var_1610_cast_fp16 = reshape(shape = var_1609, x = value_21_cast_fp16)[name = tensor("op_1610_cast_fp16")]; tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(true)]; tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_1610_cast_fp16, y = var_1608_cast_fp16)[name = tensor("attn_21_cast_fp16")]; tensor var_1613 = const()[name = tensor("op_1613"), val = tensor([1, 1280, 1, -1])]; tensor x_147_cast_fp16 = reshape(shape = var_1613, x = attn_21_cast_fp16)[name = tensor("x_147_cast_fp16")]; tensor layers_5_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202494336)))]; tensor input_117_cast_fp16 = sub(x = x_147_cast_fp16, y = layers_5_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_117_cast_fp16")]; tensor var_1621 = const()[name = tensor("op_1621"), val = tensor([1, 1])]; tensor var_1623 = const()[name = tensor("op_1623"), val = tensor([1, 1])]; tensor x_149_pad_type_0 = const()[name = tensor("x_149_pad_type_0"), val = tensor("custom")]; tensor x_149_pad_0 = const()[name = tensor("x_149_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202496960))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203316224))), name = tensor("layers_5_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_5_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203316352)))]; tensor x_149_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_module_bias_to_fp16, dilations = var_1623, groups = var_1513, pad = x_149_pad_0, pad_type = x_149_pad_type_0, strides = var_1621, weight = layers_5_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_117_cast_fp16)[name = tensor("x_149_cast_fp16")]; tensor layers_5_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203318976)))]; tensor obj_77_cast_fp16 = mul(x = x_149_cast_fp16, y = layers_5_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_77_cast_fp16")]; tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = obj_77_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; tensor var_1634 = const()[name = tensor("op_1634"), val = tensor([1])]; tensor channels_mean_33_cast_fp16 = reduce_mean(axes = var_1634, keep_dims = var_1514, x = inputs_33_cast_fp16)[name = tensor("channels_mean_33_cast_fp16")]; tensor zero_mean_33_cast_fp16 = sub(x = inputs_33_cast_fp16, y = channels_mean_33_cast_fp16)[name = tensor("zero_mean_33_cast_fp16")]; tensor zero_mean_sq_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = zero_mean_33_cast_fp16)[name = tensor("zero_mean_sq_33_cast_fp16")]; tensor var_1638 = const()[name = tensor("op_1638"), val = tensor([1])]; tensor var_1639_cast_fp16 = reduce_mean(axes = var_1638, keep_dims = var_1514, x = zero_mean_sq_33_cast_fp16)[name = tensor("op_1639_cast_fp16")]; tensor var_1640_to_fp16 = const()[name = tensor("op_1640_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1641_cast_fp16 = add(x = var_1639_cast_fp16, y = var_1640_to_fp16)[name = tensor("op_1641_cast_fp16")]; tensor denom_33_epsilon_0_to_fp16 = const()[name = tensor("denom_33_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_33_cast_fp16 = rsqrt(epsilon = denom_33_epsilon_0_to_fp16, x = var_1641_cast_fp16)[name = tensor("denom_33_cast_fp16")]; tensor out_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = denom_33_cast_fp16)[name = tensor("out_33_cast_fp16")]; tensor obj_79_gamma_0_to_fp16 = const()[name = tensor("obj_79_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203321600)))]; tensor obj_79_beta_0_to_fp16 = const()[name = tensor("obj_79_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203324224)))]; tensor obj_79_epsilon_0_to_fp16 = const()[name = tensor("obj_79_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_79_cast_fp16 = batch_norm(beta = obj_79_beta_0_to_fp16, epsilon = obj_79_epsilon_0_to_fp16, gamma = obj_79_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor("obj_79_cast_fp16")]; tensor layers_5_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203326848)))]; tensor input_119_cast_fp16 = sub(x = obj_79_cast_fp16, y = layers_5_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_119_cast_fp16")]; tensor var_1660 = const()[name = tensor("op_1660"), val = tensor([1, 1])]; tensor var_1662 = const()[name = tensor("op_1662"), val = tensor([1, 1])]; tensor x_151_pad_type_0 = const()[name = tensor("x_151_pad_type_0"), val = tensor("custom")]; tensor x_151_pad_0 = const()[name = tensor("x_151_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203329472))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204148736))), name = tensor("layers_5_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_5_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204148864)))]; tensor x_151_cast_fp16 = conv(bias = layers_5_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_1662, groups = var_1513, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = var_1660, weight = layers_5_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_119_cast_fp16)[name = tensor("x_151_cast_fp16")]; tensor layers_5_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204151488)))]; tensor query_23_cast_fp16 = mul(x = x_151_cast_fp16, y = layers_5_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_23_cast_fp16")]; tensor var_1672 = const()[name = tensor("op_1672"), val = tensor([1, 1])]; tensor var_1674 = const()[name = tensor("op_1674"), val = tensor([1, 1])]; tensor x_153_pad_type_0 = const()[name = tensor("x_153_pad_type_0"), val = tensor("custom")]; tensor x_153_pad_0 = const()[name = tensor("x_153_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204154112))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204973376))), name = tensor("layers_5_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_5_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204973504)))]; tensor x_153_cast_fp16 = conv(bias = layers_5_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_1674, groups = var_1513, pad = x_153_pad_0, pad_type = x_153_pad_type_0, strides = var_1672, weight = layers_5_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_153_cast_fp16")]; tensor layers_5_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_5_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204976128)))]; tensor key_23_cast_fp16 = mul(x = x_153_cast_fp16, y = layers_5_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_23_cast_fp16")]; tensor var_1684 = const()[name = tensor("op_1684"), val = tensor([1, 1])]; tensor var_1686 = const()[name = tensor("op_1686"), val = tensor([1, 1])]; tensor x_155_pad_type_0 = const()[name = tensor("x_155_pad_type_0"), val = tensor("custom")]; tensor x_155_pad_0 = const()[name = tensor("x_155_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204978752))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205798016))), name = tensor("layers_5_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_5_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205798144)))]; tensor x_155_cast_fp16 = conv(bias = layers_5_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_1686, groups = var_1513, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = var_1684, weight = layers_5_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_155_cast_fp16")]; tensor layers_5_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205800768)))]; tensor value_23_cast_fp16 = mul(x = x_155_cast_fp16, y = layers_5_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_23_cast_fp16")]; tensor var_1691 = const()[name = tensor("op_1691"), val = tensor([1, 20, 64, -1])]; tensor var_1692_cast_fp16 = reshape(shape = var_1691, x = query_23_cast_fp16)[name = tensor("op_1692_cast_fp16")]; tensor var_1693_to_fp16 = const()[name = tensor("op_1693_to_fp16"), val = tensor(0x1p-3)]; tensor var_1694_cast_fp16 = mul(x = var_1692_cast_fp16, y = var_1693_to_fp16)[name = tensor("op_1694_cast_fp16")]; tensor var_1695 = const()[name = tensor("op_1695"), val = tensor([1, 20, 64, -1])]; tensor var_1696_cast_fp16 = reshape(shape = var_1695, x = key_23_cast_fp16)[name = tensor("op_1696_cast_fp16")]; tensor mh_w_35_transpose_x_0 = const()[name = tensor("mh_w_35_transpose_x_0"), val = tensor(true)]; tensor mh_w_35_transpose_y_0 = const()[name = tensor("mh_w_35_transpose_y_0"), val = tensor(false)]; tensor mh_w_35_cast_fp16 = matmul(transpose_x = mh_w_35_transpose_x_0, transpose_y = mh_w_35_transpose_y_0, x = var_1694_cast_fp16, y = var_1696_cast_fp16)[name = tensor("mh_w_35_cast_fp16")]; tensor obj_83_cast_fp16 = softmax(axis = var_1506, x = mh_w_35_cast_fp16)[name = tensor("obj_83_cast_fp16")]; tensor var_1700 = const()[name = tensor("op_1700"), val = tensor([1, 20, 64, -1])]; tensor var_1701_cast_fp16 = reshape(shape = var_1700, x = value_23_cast_fp16)[name = tensor("op_1701_cast_fp16")]; tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(true)]; tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = var_1701_cast_fp16, y = obj_83_cast_fp16)[name = tensor("attn_23_cast_fp16")]; tensor var_1704 = const()[name = tensor("op_1704"), val = tensor([1, 1280, 1, -1])]; tensor x_157_cast_fp16 = reshape(shape = var_1704, x = attn_23_cast_fp16)[name = tensor("x_157_cast_fp16")]; tensor layers_5_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205803392)))]; tensor input_125_cast_fp16 = sub(x = x_157_cast_fp16, y = layers_5_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_125_cast_fp16")]; tensor var_1712 = const()[name = tensor("op_1712"), val = tensor([1, 1])]; tensor var_1714 = const()[name = tensor("op_1714"), val = tensor([1, 1])]; tensor x_159_pad_type_0 = const()[name = tensor("x_159_pad_type_0"), val = tensor("custom")]; tensor x_159_pad_0 = const()[name = tensor("x_159_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205806016))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206625280))), name = tensor("layers_5_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_5_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206625408)))]; tensor x_159_cast_fp16 = conv(bias = layers_5_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_1714, groups = var_1513, pad = x_159_pad_0, pad_type = x_159_pad_type_0, strides = var_1712, weight = layers_5_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_125_cast_fp16)[name = tensor("x_159_cast_fp16")]; tensor layers_5_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206628032)))]; tensor obj_81_cast_fp16 = mul(x = x_159_cast_fp16, y = layers_5_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_81_cast_fp16")]; tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_81_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; tensor var_1721 = const()[name = tensor("op_1721"), val = tensor([1])]; tensor channels_mean_35_cast_fp16 = reduce_mean(axes = var_1721, keep_dims = var_1514, x = inputs_35_cast_fp16)[name = tensor("channels_mean_35_cast_fp16")]; tensor zero_mean_35_cast_fp16 = sub(x = inputs_35_cast_fp16, y = channels_mean_35_cast_fp16)[name = tensor("zero_mean_35_cast_fp16")]; tensor zero_mean_sq_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = zero_mean_35_cast_fp16)[name = tensor("zero_mean_sq_35_cast_fp16")]; tensor var_1725 = const()[name = tensor("op_1725"), val = tensor([1])]; tensor var_1726_cast_fp16 = reduce_mean(axes = var_1725, keep_dims = var_1514, x = zero_mean_sq_35_cast_fp16)[name = tensor("op_1726_cast_fp16")]; tensor var_1727_to_fp16 = const()[name = tensor("op_1727_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1728_cast_fp16 = add(x = var_1726_cast_fp16, y = var_1727_to_fp16)[name = tensor("op_1728_cast_fp16")]; tensor denom_35_epsilon_0_to_fp16 = const()[name = tensor("denom_35_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_35_cast_fp16 = rsqrt(epsilon = denom_35_epsilon_0_to_fp16, x = var_1728_cast_fp16)[name = tensor("denom_35_cast_fp16")]; tensor out_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = denom_35_cast_fp16)[name = tensor("out_35_cast_fp16")]; tensor x_161_gamma_0_to_fp16 = const()[name = tensor("x_161_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206630656)))]; tensor x_161_beta_0_to_fp16 = const()[name = tensor("x_161_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206633280)))]; tensor x_161_epsilon_0_to_fp16 = const()[name = tensor("x_161_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_161_cast_fp16 = batch_norm(beta = x_161_beta_0_to_fp16, epsilon = x_161_epsilon_0_to_fp16, gamma = x_161_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor("x_161_cast_fp16")]; tensor layers_5_fc1_input_shift_to_fp16 = const()[name = tensor("layers_5_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206635904)))]; tensor input_127_cast_fp16 = sub(x = x_161_cast_fp16, y = layers_5_fc1_input_shift_to_fp16)[name = tensor("input_127_cast_fp16")]; tensor var_1743 = const()[name = tensor("op_1743"), val = tensor([1, 1])]; tensor var_1745 = const()[name = tensor("op_1745"), val = tensor([1, 1])]; tensor x_163_pad_type_0 = const()[name = tensor("x_163_pad_type_0"), val = tensor("custom")]; tensor x_163_pad_0 = const()[name = tensor("x_163_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206638528))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209915392))), name = tensor("layers_5_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_5_fc1_module_bias_to_fp16 = const()[name = tensor("layers_5_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209915520)))]; tensor x_163_cast_fp16 = conv(bias = layers_5_fc1_module_bias_to_fp16, dilations = var_1745, groups = var_1513, pad = x_163_pad_0, pad_type = x_163_pad_type_0, strides = var_1743, weight = layers_5_fc1_module_weight_to_fp16_palettized, x = input_127_cast_fp16)[name = tensor("x_163_cast_fp16")]; tensor layers_5_fc1_output_scale_to_fp16 = const()[name = tensor("layers_5_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209925824)))]; tensor input_129_cast_fp16 = mul(x = x_163_cast_fp16, y = layers_5_fc1_output_scale_to_fp16)[name = tensor("input_129_cast_fp16")]; tensor x_165_mode_0 = const()[name = tensor("x_165_mode_0"), val = tensor("EXACT")]; tensor x_165_cast_fp16 = gelu(mode = x_165_mode_0, x = input_129_cast_fp16)[name = tensor("x_165_cast_fp16")]; tensor layers_5_fc2_input_shift_to_fp16 = const()[name = tensor("layers_5_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209936128)))]; tensor input_131_cast_fp16 = sub(x = x_165_cast_fp16, y = layers_5_fc2_input_shift_to_fp16)[name = tensor("input_131_cast_fp16")]; tensor var_1756 = const()[name = tensor("op_1756"), val = tensor([1, 1])]; tensor var_1758 = const()[name = tensor("op_1758"), val = tensor([1, 1])]; tensor x_167_pad_type_0 = const()[name = tensor("x_167_pad_type_0"), val = tensor("custom")]; tensor x_167_pad_0 = const()[name = tensor("x_167_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_5_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209946432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213223296))), name = tensor("layers_5_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_5_fc2_module_bias_to_fp16 = const()[name = tensor("layers_5_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213223424)))]; tensor x_167_cast_fp16 = conv(bias = layers_5_fc2_module_bias_to_fp16, dilations = var_1758, groups = var_1513, pad = x_167_pad_0, pad_type = x_167_pad_type_0, strides = var_1756, weight = layers_5_fc2_module_weight_to_fp16_palettized, x = input_131_cast_fp16)[name = tensor("x_167_cast_fp16")]; tensor layers_5_fc2_output_scale_to_fp16 = const()[name = tensor("layers_5_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213226048)))]; tensor hidden_states_13_cast_fp16 = mul(x = x_167_cast_fp16, y = layers_5_fc2_output_scale_to_fp16)[name = tensor("hidden_states_13_cast_fp16")]; tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; tensor var_1772 = const()[name = tensor("op_1772"), val = tensor(3)]; tensor var_1779 = const()[name = tensor("op_1779"), val = tensor(1)]; tensor var_1780 = const()[name = tensor("op_1780"), val = tensor(true)]; tensor var_1792 = const()[name = tensor("op_1792"), val = tensor([1])]; tensor channels_mean_37_cast_fp16 = reduce_mean(axes = var_1792, keep_dims = var_1780, x = inputs_37_cast_fp16)[name = tensor("channels_mean_37_cast_fp16")]; tensor zero_mean_37_cast_fp16 = sub(x = inputs_37_cast_fp16, y = channels_mean_37_cast_fp16)[name = tensor("zero_mean_37_cast_fp16")]; tensor zero_mean_sq_37_cast_fp16 = mul(x = zero_mean_37_cast_fp16, y = zero_mean_37_cast_fp16)[name = tensor("zero_mean_sq_37_cast_fp16")]; tensor var_1796 = const()[name = tensor("op_1796"), val = tensor([1])]; tensor var_1797_cast_fp16 = reduce_mean(axes = var_1796, keep_dims = var_1780, x = zero_mean_sq_37_cast_fp16)[name = tensor("op_1797_cast_fp16")]; tensor var_1798_to_fp16 = const()[name = tensor("op_1798_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1799_cast_fp16 = add(x = var_1797_cast_fp16, y = var_1798_to_fp16)[name = tensor("op_1799_cast_fp16")]; tensor denom_37_epsilon_0_to_fp16 = const()[name = tensor("denom_37_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_37_cast_fp16 = rsqrt(epsilon = denom_37_epsilon_0_to_fp16, x = var_1799_cast_fp16)[name = tensor("denom_37_cast_fp16")]; tensor out_37_cast_fp16 = mul(x = zero_mean_37_cast_fp16, y = denom_37_cast_fp16)[name = tensor("out_37_cast_fp16")]; tensor obj_85_gamma_0_to_fp16 = const()[name = tensor("obj_85_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213228672)))]; tensor obj_85_beta_0_to_fp16 = const()[name = tensor("obj_85_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213231296)))]; tensor obj_85_epsilon_0_to_fp16 = const()[name = tensor("obj_85_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_85_cast_fp16 = batch_norm(beta = obj_85_beta_0_to_fp16, epsilon = obj_85_epsilon_0_to_fp16, gamma = obj_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor("obj_85_cast_fp16")]; tensor layers_6_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213233920)))]; tensor input_133_cast_fp16 = sub(x = obj_85_cast_fp16, y = layers_6_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_133_cast_fp16")]; tensor var_1818 = const()[name = tensor("op_1818"), val = tensor([1, 1])]; tensor var_1820 = const()[name = tensor("op_1820"), val = tensor([1, 1])]; tensor x_169_pad_type_0 = const()[name = tensor("x_169_pad_type_0"), val = tensor("custom")]; tensor x_169_pad_0 = const()[name = tensor("x_169_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213236544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214055808))), name = tensor("layers_6_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_6_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214055936)))]; tensor x_169_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_module_bias_to_fp16, dilations = var_1820, groups = var_1779, pad = x_169_pad_0, pad_type = x_169_pad_type_0, strides = var_1818, weight = layers_6_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_133_cast_fp16)[name = tensor("x_169_cast_fp16")]; tensor layers_6_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214058560)))]; tensor query_25_cast_fp16 = mul(x = x_169_cast_fp16, y = layers_6_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_25_cast_fp16")]; tensor var_1830 = const()[name = tensor("op_1830"), val = tensor([1, 1])]; tensor var_1832 = const()[name = tensor("op_1832"), val = tensor([1, 1])]; tensor x_171_pad_type_0 = const()[name = tensor("x_171_pad_type_0"), val = tensor("custom")]; tensor x_171_pad_0 = const()[name = tensor("x_171_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214061184))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214880448))), name = tensor("layers_6_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_6_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214880576)))]; tensor x_171_cast_fp16 = conv(bias = layers_6_self_attn_k_proj_module_bias_to_fp16, dilations = var_1832, groups = var_1779, pad = x_171_pad_0, pad_type = x_171_pad_type_0, strides = var_1830, weight = layers_6_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_133_cast_fp16)[name = tensor("x_171_cast_fp16")]; tensor layers_6_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_6_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214883200)))]; tensor current_key_13_cast_fp16 = mul(x = x_171_cast_fp16, y = layers_6_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_13_cast_fp16")]; tensor var_1842 = const()[name = tensor("op_1842"), val = tensor([1, 1])]; tensor var_1844 = const()[name = tensor("op_1844"), val = tensor([1, 1])]; tensor x_173_pad_type_0 = const()[name = tensor("x_173_pad_type_0"), val = tensor("custom")]; tensor x_173_pad_0 = const()[name = tensor("x_173_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214885824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215705088))), name = tensor("layers_6_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_6_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215705216)))]; tensor x_173_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_module_bias_to_fp16, dilations = var_1844, groups = var_1779, pad = x_173_pad_0, pad_type = x_173_pad_type_0, strides = var_1842, weight = layers_6_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_133_cast_fp16)[name = tensor("x_173_cast_fp16")]; tensor layers_6_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215707840)))]; tensor current_value_13_cast_fp16 = mul(x = x_173_cast_fp16, y = layers_6_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_13_cast_fp16")]; tensor var_1852_cast_fp16 = mul(x = current_key_13_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_1852_cast_fp16")]; tensor var_1854_cast_fp16 = mul(x = var_103_cast_fp16_6, y = var_257_cast_fp16)[name = tensor("op_1854_cast_fp16")]; tensor key_25_cast_fp16 = add(x = var_1852_cast_fp16, y = var_1854_cast_fp16)[name = tensor("key_25_cast_fp16")]; tensor var_1856_cast_fp16 = mul(x = current_value_13_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_1856_cast_fp16")]; tensor var_1858_cast_fp16 = mul(x = var_138_cast_fp16_6, y = var_257_cast_fp16)[name = tensor("op_1858_cast_fp16")]; tensor value_25_cast_fp16 = add(x = var_1856_cast_fp16, y = var_1858_cast_fp16)[name = tensor("value_25_cast_fp16")]; tensor var_1861 = const()[name = tensor("op_1861"), val = tensor([1, 20, 64, -1])]; tensor var_1862_cast_fp16 = reshape(shape = var_1861, x = query_25_cast_fp16)[name = tensor("op_1862_cast_fp16")]; tensor var_1863_to_fp16 = const()[name = tensor("op_1863_to_fp16"), val = tensor(0x1p-3)]; tensor var_1864_cast_fp16 = mul(x = var_1862_cast_fp16, y = var_1863_to_fp16)[name = tensor("op_1864_cast_fp16")]; tensor var_1865 = const()[name = tensor("op_1865"), val = tensor([1, 20, 64, -1])]; tensor var_1866_cast_fp16 = reshape(shape = var_1865, x = key_25_cast_fp16)[name = tensor("op_1866_cast_fp16")]; tensor mh_w_37_transpose_x_0 = const()[name = tensor("mh_w_37_transpose_x_0"), val = tensor(true)]; tensor mh_w_37_transpose_y_0 = const()[name = tensor("mh_w_37_transpose_y_0"), val = tensor(false)]; tensor mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_1864_cast_fp16, y = var_1866_cast_fp16)[name = tensor("mh_w_37_cast_fp16")]; tensor mh_w_39_cast_fp16 = add(x = mh_w_37_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_39_cast_fp16")]; tensor var_1874_cast_fp16 = softmax(axis = var_1772, x = mh_w_39_cast_fp16)[name = tensor("op_1874_cast_fp16")]; tensor var_1875 = const()[name = tensor("op_1875"), val = tensor([1, 20, 64, -1])]; tensor var_1876_cast_fp16 = reshape(shape = var_1875, x = value_25_cast_fp16)[name = tensor("op_1876_cast_fp16")]; tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(true)]; tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_1876_cast_fp16, y = var_1874_cast_fp16)[name = tensor("attn_25_cast_fp16")]; tensor var_1879 = const()[name = tensor("op_1879"), val = tensor([1, 1280, 1, -1])]; tensor x_175_cast_fp16 = reshape(shape = var_1879, x = attn_25_cast_fp16)[name = tensor("x_175_cast_fp16")]; tensor layers_6_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215710464)))]; tensor input_139_cast_fp16 = sub(x = x_175_cast_fp16, y = layers_6_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_139_cast_fp16")]; tensor var_1887 = const()[name = tensor("op_1887"), val = tensor([1, 1])]; tensor var_1889 = const()[name = tensor("op_1889"), val = tensor([1, 1])]; tensor x_177_pad_type_0 = const()[name = tensor("x_177_pad_type_0"), val = tensor("custom")]; tensor x_177_pad_0 = const()[name = tensor("x_177_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215713088))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216532352))), name = tensor("layers_6_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_6_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216532480)))]; tensor x_177_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_module_bias_to_fp16, dilations = var_1889, groups = var_1779, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = var_1887, weight = layers_6_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_139_cast_fp16)[name = tensor("x_177_cast_fp16")]; tensor layers_6_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216535104)))]; tensor obj_91_cast_fp16 = mul(x = x_177_cast_fp16, y = layers_6_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_91_cast_fp16")]; tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_91_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; tensor var_1900 = const()[name = tensor("op_1900"), val = tensor([1])]; tensor channels_mean_39_cast_fp16 = reduce_mean(axes = var_1900, keep_dims = var_1780, x = inputs_39_cast_fp16)[name = tensor("channels_mean_39_cast_fp16")]; tensor zero_mean_39_cast_fp16 = sub(x = inputs_39_cast_fp16, y = channels_mean_39_cast_fp16)[name = tensor("zero_mean_39_cast_fp16")]; tensor zero_mean_sq_39_cast_fp16 = mul(x = zero_mean_39_cast_fp16, y = zero_mean_39_cast_fp16)[name = tensor("zero_mean_sq_39_cast_fp16")]; tensor var_1904 = const()[name = tensor("op_1904"), val = tensor([1])]; tensor var_1905_cast_fp16 = reduce_mean(axes = var_1904, keep_dims = var_1780, x = zero_mean_sq_39_cast_fp16)[name = tensor("op_1905_cast_fp16")]; tensor var_1906_to_fp16 = const()[name = tensor("op_1906_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1907_cast_fp16 = add(x = var_1905_cast_fp16, y = var_1906_to_fp16)[name = tensor("op_1907_cast_fp16")]; tensor denom_39_epsilon_0_to_fp16 = const()[name = tensor("denom_39_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_39_cast_fp16 = rsqrt(epsilon = denom_39_epsilon_0_to_fp16, x = var_1907_cast_fp16)[name = tensor("denom_39_cast_fp16")]; tensor out_39_cast_fp16 = mul(x = zero_mean_39_cast_fp16, y = denom_39_cast_fp16)[name = tensor("out_39_cast_fp16")]; tensor obj_93_gamma_0_to_fp16 = const()[name = tensor("obj_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216537728)))]; tensor obj_93_beta_0_to_fp16 = const()[name = tensor("obj_93_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216540352)))]; tensor obj_93_epsilon_0_to_fp16 = const()[name = tensor("obj_93_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_93_cast_fp16 = batch_norm(beta = obj_93_beta_0_to_fp16, epsilon = obj_93_epsilon_0_to_fp16, gamma = obj_93_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor("obj_93_cast_fp16")]; tensor layers_6_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_6_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216542976)))]; tensor input_141_cast_fp16 = sub(x = obj_93_cast_fp16, y = layers_6_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_141_cast_fp16")]; tensor var_1926 = const()[name = tensor("op_1926"), val = tensor([1, 1])]; tensor var_1928 = const()[name = tensor("op_1928"), val = tensor([1, 1])]; tensor x_179_pad_type_0 = const()[name = tensor("x_179_pad_type_0"), val = tensor("custom")]; tensor x_179_pad_0 = const()[name = tensor("x_179_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216545600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217364864))), name = tensor("layers_6_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_6_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217364992)))]; tensor x_179_cast_fp16 = conv(bias = layers_6_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_1928, groups = var_1779, pad = x_179_pad_0, pad_type = x_179_pad_type_0, strides = var_1926, weight = layers_6_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_141_cast_fp16)[name = tensor("x_179_cast_fp16")]; tensor layers_6_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_6_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217367616)))]; tensor query_27_cast_fp16 = mul(x = x_179_cast_fp16, y = layers_6_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_27_cast_fp16")]; tensor var_1938 = const()[name = tensor("op_1938"), val = tensor([1, 1])]; tensor var_1940 = const()[name = tensor("op_1940"), val = tensor([1, 1])]; tensor x_181_pad_type_0 = const()[name = tensor("x_181_pad_type_0"), val = tensor("custom")]; tensor x_181_pad_0 = const()[name = tensor("x_181_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217370240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218189504))), name = tensor("layers_6_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_6_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218189632)))]; tensor x_181_cast_fp16 = conv(bias = layers_6_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_1940, groups = var_1779, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = var_1938, weight = layers_6_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_181_cast_fp16")]; tensor layers_6_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_6_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218192256)))]; tensor key_27_cast_fp16 = mul(x = x_181_cast_fp16, y = layers_6_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_27_cast_fp16")]; tensor var_1950 = const()[name = tensor("op_1950"), val = tensor([1, 1])]; tensor var_1952 = const()[name = tensor("op_1952"), val = tensor([1, 1])]; tensor x_183_pad_type_0 = const()[name = tensor("x_183_pad_type_0"), val = tensor("custom")]; tensor x_183_pad_0 = const()[name = tensor("x_183_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218194880))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219014144))), name = tensor("layers_6_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_6_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219014272)))]; tensor x_183_cast_fp16 = conv(bias = layers_6_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_1952, groups = var_1779, pad = x_183_pad_0, pad_type = x_183_pad_type_0, strides = var_1950, weight = layers_6_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_183_cast_fp16")]; tensor layers_6_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_6_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219016896)))]; tensor value_27_cast_fp16 = mul(x = x_183_cast_fp16, y = layers_6_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_27_cast_fp16")]; tensor var_1957 = const()[name = tensor("op_1957"), val = tensor([1, 20, 64, -1])]; tensor var_1958_cast_fp16 = reshape(shape = var_1957, x = query_27_cast_fp16)[name = tensor("op_1958_cast_fp16")]; tensor var_1959_to_fp16 = const()[name = tensor("op_1959_to_fp16"), val = tensor(0x1p-3)]; tensor var_1960_cast_fp16 = mul(x = var_1958_cast_fp16, y = var_1959_to_fp16)[name = tensor("op_1960_cast_fp16")]; tensor var_1961 = const()[name = tensor("op_1961"), val = tensor([1, 20, 64, -1])]; tensor var_1962_cast_fp16 = reshape(shape = var_1961, x = key_27_cast_fp16)[name = tensor("op_1962_cast_fp16")]; tensor mh_w_41_transpose_x_0 = const()[name = tensor("mh_w_41_transpose_x_0"), val = tensor(true)]; tensor mh_w_41_transpose_y_0 = const()[name = tensor("mh_w_41_transpose_y_0"), val = tensor(false)]; tensor mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_1960_cast_fp16, y = var_1962_cast_fp16)[name = tensor("mh_w_41_cast_fp16")]; tensor obj_97_cast_fp16 = softmax(axis = var_1772, x = mh_w_41_cast_fp16)[name = tensor("obj_97_cast_fp16")]; tensor var_1966 = const()[name = tensor("op_1966"), val = tensor([1, 20, 64, -1])]; tensor var_1967_cast_fp16 = reshape(shape = var_1966, x = value_27_cast_fp16)[name = tensor("op_1967_cast_fp16")]; tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(true)]; tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_1967_cast_fp16, y = obj_97_cast_fp16)[name = tensor("attn_27_cast_fp16")]; tensor var_1970 = const()[name = tensor("op_1970"), val = tensor([1, 1280, 1, -1])]; tensor x_185_cast_fp16 = reshape(shape = var_1970, x = attn_27_cast_fp16)[name = tensor("x_185_cast_fp16")]; tensor layers_6_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_6_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219019520)))]; tensor input_147_cast_fp16 = sub(x = x_185_cast_fp16, y = layers_6_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_147_cast_fp16")]; tensor var_1978 = const()[name = tensor("op_1978"), val = tensor([1, 1])]; tensor var_1980 = const()[name = tensor("op_1980"), val = tensor([1, 1])]; tensor x_187_pad_type_0 = const()[name = tensor("x_187_pad_type_0"), val = tensor("custom")]; tensor x_187_pad_0 = const()[name = tensor("x_187_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219022144))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219841408))), name = tensor("layers_6_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_6_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219841536)))]; tensor x_187_cast_fp16 = conv(bias = layers_6_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_1980, groups = var_1779, pad = x_187_pad_0, pad_type = x_187_pad_type_0, strides = var_1978, weight = layers_6_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = tensor("x_187_cast_fp16")]; tensor layers_6_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_6_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219844160)))]; tensor obj_95_cast_fp16 = mul(x = x_187_cast_fp16, y = layers_6_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_95_cast_fp16")]; tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = obj_95_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; tensor var_1987 = const()[name = tensor("op_1987"), val = tensor([1])]; tensor channels_mean_41_cast_fp16 = reduce_mean(axes = var_1987, keep_dims = var_1780, x = inputs_41_cast_fp16)[name = tensor("channels_mean_41_cast_fp16")]; tensor zero_mean_41_cast_fp16 = sub(x = inputs_41_cast_fp16, y = channels_mean_41_cast_fp16)[name = tensor("zero_mean_41_cast_fp16")]; tensor zero_mean_sq_41_cast_fp16 = mul(x = zero_mean_41_cast_fp16, y = zero_mean_41_cast_fp16)[name = tensor("zero_mean_sq_41_cast_fp16")]; tensor var_1991 = const()[name = tensor("op_1991"), val = tensor([1])]; tensor var_1992_cast_fp16 = reduce_mean(axes = var_1991, keep_dims = var_1780, x = zero_mean_sq_41_cast_fp16)[name = tensor("op_1992_cast_fp16")]; tensor var_1993_to_fp16 = const()[name = tensor("op_1993_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_1994_cast_fp16 = add(x = var_1992_cast_fp16, y = var_1993_to_fp16)[name = tensor("op_1994_cast_fp16")]; tensor denom_41_epsilon_0_to_fp16 = const()[name = tensor("denom_41_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_41_cast_fp16 = rsqrt(epsilon = denom_41_epsilon_0_to_fp16, x = var_1994_cast_fp16)[name = tensor("denom_41_cast_fp16")]; tensor out_41_cast_fp16 = mul(x = zero_mean_41_cast_fp16, y = denom_41_cast_fp16)[name = tensor("out_41_cast_fp16")]; tensor x_189_gamma_0_to_fp16 = const()[name = tensor("x_189_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219846784)))]; tensor x_189_beta_0_to_fp16 = const()[name = tensor("x_189_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219849408)))]; tensor x_189_epsilon_0_to_fp16 = const()[name = tensor("x_189_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_189_cast_fp16 = batch_norm(beta = x_189_beta_0_to_fp16, epsilon = x_189_epsilon_0_to_fp16, gamma = x_189_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor("x_189_cast_fp16")]; tensor layers_6_fc1_input_shift_to_fp16 = const()[name = tensor("layers_6_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219852032)))]; tensor input_149_cast_fp16 = sub(x = x_189_cast_fp16, y = layers_6_fc1_input_shift_to_fp16)[name = tensor("input_149_cast_fp16")]; tensor var_2009 = const()[name = tensor("op_2009"), val = tensor([1, 1])]; tensor var_2011 = const()[name = tensor("op_2011"), val = tensor([1, 1])]; tensor x_191_pad_type_0 = const()[name = tensor("x_191_pad_type_0"), val = tensor("custom")]; tensor x_191_pad_0 = const()[name = tensor("x_191_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219854656))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223131520))), name = tensor("layers_6_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_6_fc1_module_bias_to_fp16 = const()[name = tensor("layers_6_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223131648)))]; tensor x_191_cast_fp16 = conv(bias = layers_6_fc1_module_bias_to_fp16, dilations = var_2011, groups = var_1779, pad = x_191_pad_0, pad_type = x_191_pad_type_0, strides = var_2009, weight = layers_6_fc1_module_weight_to_fp16_palettized, x = input_149_cast_fp16)[name = tensor("x_191_cast_fp16")]; tensor layers_6_fc1_output_scale_to_fp16 = const()[name = tensor("layers_6_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223141952)))]; tensor input_151_cast_fp16 = mul(x = x_191_cast_fp16, y = layers_6_fc1_output_scale_to_fp16)[name = tensor("input_151_cast_fp16")]; tensor x_193_mode_0 = const()[name = tensor("x_193_mode_0"), val = tensor("EXACT")]; tensor x_193_cast_fp16 = gelu(mode = x_193_mode_0, x = input_151_cast_fp16)[name = tensor("x_193_cast_fp16")]; tensor layers_6_fc2_input_shift_to_fp16 = const()[name = tensor("layers_6_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223152256)))]; tensor input_153_cast_fp16 = sub(x = x_193_cast_fp16, y = layers_6_fc2_input_shift_to_fp16)[name = tensor("input_153_cast_fp16")]; tensor var_2022 = const()[name = tensor("op_2022"), val = tensor([1, 1])]; tensor var_2024 = const()[name = tensor("op_2024"), val = tensor([1, 1])]; tensor x_195_pad_type_0 = const()[name = tensor("x_195_pad_type_0"), val = tensor("custom")]; tensor x_195_pad_0 = const()[name = tensor("x_195_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_6_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223162560))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226439424))), name = tensor("layers_6_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_6_fc2_module_bias_to_fp16 = const()[name = tensor("layers_6_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226439552)))]; tensor x_195_cast_fp16 = conv(bias = layers_6_fc2_module_bias_to_fp16, dilations = var_2024, groups = var_1779, pad = x_195_pad_0, pad_type = x_195_pad_type_0, strides = var_2022, weight = layers_6_fc2_module_weight_to_fp16_palettized, x = input_153_cast_fp16)[name = tensor("x_195_cast_fp16")]; tensor layers_6_fc2_output_scale_to_fp16 = const()[name = tensor("layers_6_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226442176)))]; tensor hidden_states_15_cast_fp16 = mul(x = x_195_cast_fp16, y = layers_6_fc2_output_scale_to_fp16)[name = tensor("hidden_states_15_cast_fp16")]; tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = hidden_states_15_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; tensor var_2038 = const()[name = tensor("op_2038"), val = tensor(3)]; tensor var_2045 = const()[name = tensor("op_2045"), val = tensor(1)]; tensor var_2046 = const()[name = tensor("op_2046"), val = tensor(true)]; tensor var_2058 = const()[name = tensor("op_2058"), val = tensor([1])]; tensor channels_mean_43_cast_fp16 = reduce_mean(axes = var_2058, keep_dims = var_2046, x = inputs_43_cast_fp16)[name = tensor("channels_mean_43_cast_fp16")]; tensor zero_mean_43_cast_fp16 = sub(x = inputs_43_cast_fp16, y = channels_mean_43_cast_fp16)[name = tensor("zero_mean_43_cast_fp16")]; tensor zero_mean_sq_43_cast_fp16 = mul(x = zero_mean_43_cast_fp16, y = zero_mean_43_cast_fp16)[name = tensor("zero_mean_sq_43_cast_fp16")]; tensor var_2062 = const()[name = tensor("op_2062"), val = tensor([1])]; tensor var_2063_cast_fp16 = reduce_mean(axes = var_2062, keep_dims = var_2046, x = zero_mean_sq_43_cast_fp16)[name = tensor("op_2063_cast_fp16")]; tensor var_2064_to_fp16 = const()[name = tensor("op_2064_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2065_cast_fp16 = add(x = var_2063_cast_fp16, y = var_2064_to_fp16)[name = tensor("op_2065_cast_fp16")]; tensor denom_43_epsilon_0_to_fp16 = const()[name = tensor("denom_43_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_43_cast_fp16 = rsqrt(epsilon = denom_43_epsilon_0_to_fp16, x = var_2065_cast_fp16)[name = tensor("denom_43_cast_fp16")]; tensor out_43_cast_fp16 = mul(x = zero_mean_43_cast_fp16, y = denom_43_cast_fp16)[name = tensor("out_43_cast_fp16")]; tensor obj_99_gamma_0_to_fp16 = const()[name = tensor("obj_99_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226444800)))]; tensor obj_99_beta_0_to_fp16 = const()[name = tensor("obj_99_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226447424)))]; tensor obj_99_epsilon_0_to_fp16 = const()[name = tensor("obj_99_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_99_cast_fp16 = batch_norm(beta = obj_99_beta_0_to_fp16, epsilon = obj_99_epsilon_0_to_fp16, gamma = obj_99_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor("obj_99_cast_fp16")]; tensor layers_7_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226450048)))]; tensor input_155_cast_fp16 = sub(x = obj_99_cast_fp16, y = layers_7_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_155_cast_fp16")]; tensor var_2084 = const()[name = tensor("op_2084"), val = tensor([1, 1])]; tensor var_2086 = const()[name = tensor("op_2086"), val = tensor([1, 1])]; tensor x_197_pad_type_0 = const()[name = tensor("x_197_pad_type_0"), val = tensor("custom")]; tensor x_197_pad_0 = const()[name = tensor("x_197_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226452672))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227271936))), name = tensor("layers_7_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_7_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227272064)))]; tensor x_197_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_module_bias_to_fp16, dilations = var_2086, groups = var_2045, pad = x_197_pad_0, pad_type = x_197_pad_type_0, strides = var_2084, weight = layers_7_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_155_cast_fp16)[name = tensor("x_197_cast_fp16")]; tensor layers_7_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227274688)))]; tensor query_29_cast_fp16 = mul(x = x_197_cast_fp16, y = layers_7_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_29_cast_fp16")]; tensor var_2096 = const()[name = tensor("op_2096"), val = tensor([1, 1])]; tensor var_2098 = const()[name = tensor("op_2098"), val = tensor([1, 1])]; tensor x_199_pad_type_0 = const()[name = tensor("x_199_pad_type_0"), val = tensor("custom")]; tensor x_199_pad_0 = const()[name = tensor("x_199_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227277312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228096576))), name = tensor("layers_7_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_7_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228096704)))]; tensor x_199_cast_fp16 = conv(bias = layers_7_self_attn_k_proj_module_bias_to_fp16, dilations = var_2098, groups = var_2045, pad = x_199_pad_0, pad_type = x_199_pad_type_0, strides = var_2096, weight = layers_7_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_155_cast_fp16)[name = tensor("x_199_cast_fp16")]; tensor layers_7_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_7_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228099328)))]; tensor current_key_15_cast_fp16 = mul(x = x_199_cast_fp16, y = layers_7_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_15_cast_fp16")]; tensor var_2108 = const()[name = tensor("op_2108"), val = tensor([1, 1])]; tensor var_2110 = const()[name = tensor("op_2110"), val = tensor([1, 1])]; tensor x_201_pad_type_0 = const()[name = tensor("x_201_pad_type_0"), val = tensor("custom")]; tensor x_201_pad_0 = const()[name = tensor("x_201_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228101952))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228921216))), name = tensor("layers_7_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_7_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228921344)))]; tensor x_201_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_module_bias_to_fp16, dilations = var_2110, groups = var_2045, pad = x_201_pad_0, pad_type = x_201_pad_type_0, strides = var_2108, weight = layers_7_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_155_cast_fp16)[name = tensor("x_201_cast_fp16")]; tensor layers_7_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228923968)))]; tensor current_value_15_cast_fp16 = mul(x = x_201_cast_fp16, y = layers_7_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_15_cast_fp16")]; tensor var_2118_cast_fp16 = mul(x = current_key_15_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_2118_cast_fp16")]; tensor var_2120_cast_fp16 = mul(x = var_103_cast_fp16_7, y = var_257_cast_fp16)[name = tensor("op_2120_cast_fp16")]; tensor key_29_cast_fp16 = add(x = var_2118_cast_fp16, y = var_2120_cast_fp16)[name = tensor("key_29_cast_fp16")]; tensor var_2122_cast_fp16 = mul(x = current_value_15_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_2122_cast_fp16")]; tensor var_2124_cast_fp16 = mul(x = var_138_cast_fp16_7, y = var_257_cast_fp16)[name = tensor("op_2124_cast_fp16")]; tensor value_29_cast_fp16 = add(x = var_2122_cast_fp16, y = var_2124_cast_fp16)[name = tensor("value_29_cast_fp16")]; tensor var_2127 = const()[name = tensor("op_2127"), val = tensor([1, 20, 64, -1])]; tensor var_2128_cast_fp16 = reshape(shape = var_2127, x = query_29_cast_fp16)[name = tensor("op_2128_cast_fp16")]; tensor var_2129_to_fp16 = const()[name = tensor("op_2129_to_fp16"), val = tensor(0x1p-3)]; tensor var_2130_cast_fp16 = mul(x = var_2128_cast_fp16, y = var_2129_to_fp16)[name = tensor("op_2130_cast_fp16")]; tensor var_2131 = const()[name = tensor("op_2131"), val = tensor([1, 20, 64, -1])]; tensor var_2132_cast_fp16 = reshape(shape = var_2131, x = key_29_cast_fp16)[name = tensor("op_2132_cast_fp16")]; tensor mh_w_43_transpose_x_0 = const()[name = tensor("mh_w_43_transpose_x_0"), val = tensor(true)]; tensor mh_w_43_transpose_y_0 = const()[name = tensor("mh_w_43_transpose_y_0"), val = tensor(false)]; tensor mh_w_43_cast_fp16 = matmul(transpose_x = mh_w_43_transpose_x_0, transpose_y = mh_w_43_transpose_y_0, x = var_2130_cast_fp16, y = var_2132_cast_fp16)[name = tensor("mh_w_43_cast_fp16")]; tensor mh_w_45_cast_fp16 = add(x = mh_w_43_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_45_cast_fp16")]; tensor var_2140_cast_fp16 = softmax(axis = var_2038, x = mh_w_45_cast_fp16)[name = tensor("op_2140_cast_fp16")]; tensor var_2141 = const()[name = tensor("op_2141"), val = tensor([1, 20, 64, -1])]; tensor var_2142_cast_fp16 = reshape(shape = var_2141, x = value_29_cast_fp16)[name = tensor("op_2142_cast_fp16")]; tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(true)]; tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_2142_cast_fp16, y = var_2140_cast_fp16)[name = tensor("attn_29_cast_fp16")]; tensor var_2145 = const()[name = tensor("op_2145"), val = tensor([1, 1280, 1, -1])]; tensor x_203_cast_fp16 = reshape(shape = var_2145, x = attn_29_cast_fp16)[name = tensor("x_203_cast_fp16")]; tensor layers_7_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228926592)))]; tensor input_161_cast_fp16 = sub(x = x_203_cast_fp16, y = layers_7_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_161_cast_fp16")]; tensor var_2153 = const()[name = tensor("op_2153"), val = tensor([1, 1])]; tensor var_2155 = const()[name = tensor("op_2155"), val = tensor([1, 1])]; tensor x_205_pad_type_0 = const()[name = tensor("x_205_pad_type_0"), val = tensor("custom")]; tensor x_205_pad_0 = const()[name = tensor("x_205_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228929216))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229748480))), name = tensor("layers_7_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_7_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229748608)))]; tensor x_205_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_module_bias_to_fp16, dilations = var_2155, groups = var_2045, pad = x_205_pad_0, pad_type = x_205_pad_type_0, strides = var_2153, weight = layers_7_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_161_cast_fp16)[name = tensor("x_205_cast_fp16")]; tensor layers_7_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229751232)))]; tensor obj_105_cast_fp16 = mul(x = x_205_cast_fp16, y = layers_7_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_105_cast_fp16")]; tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = obj_105_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; tensor var_2166 = const()[name = tensor("op_2166"), val = tensor([1])]; tensor channels_mean_45_cast_fp16 = reduce_mean(axes = var_2166, keep_dims = var_2046, x = inputs_45_cast_fp16)[name = tensor("channels_mean_45_cast_fp16")]; tensor zero_mean_45_cast_fp16 = sub(x = inputs_45_cast_fp16, y = channels_mean_45_cast_fp16)[name = tensor("zero_mean_45_cast_fp16")]; tensor zero_mean_sq_45_cast_fp16 = mul(x = zero_mean_45_cast_fp16, y = zero_mean_45_cast_fp16)[name = tensor("zero_mean_sq_45_cast_fp16")]; tensor var_2170 = const()[name = tensor("op_2170"), val = tensor([1])]; tensor var_2171_cast_fp16 = reduce_mean(axes = var_2170, keep_dims = var_2046, x = zero_mean_sq_45_cast_fp16)[name = tensor("op_2171_cast_fp16")]; tensor var_2172_to_fp16 = const()[name = tensor("op_2172_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2173_cast_fp16 = add(x = var_2171_cast_fp16, y = var_2172_to_fp16)[name = tensor("op_2173_cast_fp16")]; tensor denom_45_epsilon_0_to_fp16 = const()[name = tensor("denom_45_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_45_cast_fp16 = rsqrt(epsilon = denom_45_epsilon_0_to_fp16, x = var_2173_cast_fp16)[name = tensor("denom_45_cast_fp16")]; tensor out_45_cast_fp16 = mul(x = zero_mean_45_cast_fp16, y = denom_45_cast_fp16)[name = tensor("out_45_cast_fp16")]; tensor obj_107_gamma_0_to_fp16 = const()[name = tensor("obj_107_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229753856)))]; tensor obj_107_beta_0_to_fp16 = const()[name = tensor("obj_107_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229756480)))]; tensor obj_107_epsilon_0_to_fp16 = const()[name = tensor("obj_107_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_107_cast_fp16 = batch_norm(beta = obj_107_beta_0_to_fp16, epsilon = obj_107_epsilon_0_to_fp16, gamma = obj_107_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor("obj_107_cast_fp16")]; tensor layers_7_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_7_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229759104)))]; tensor input_163_cast_fp16 = sub(x = obj_107_cast_fp16, y = layers_7_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_163_cast_fp16")]; tensor var_2192 = const()[name = tensor("op_2192"), val = tensor([1, 1])]; tensor var_2194 = const()[name = tensor("op_2194"), val = tensor([1, 1])]; tensor x_207_pad_type_0 = const()[name = tensor("x_207_pad_type_0"), val = tensor("custom")]; tensor x_207_pad_0 = const()[name = tensor("x_207_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229761728))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230580992))), name = tensor("layers_7_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_7_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230581120)))]; tensor x_207_cast_fp16 = conv(bias = layers_7_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_2194, groups = var_2045, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = var_2192, weight = layers_7_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_163_cast_fp16)[name = tensor("x_207_cast_fp16")]; tensor layers_7_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_7_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230583744)))]; tensor query_31_cast_fp16 = mul(x = x_207_cast_fp16, y = layers_7_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_31_cast_fp16")]; tensor var_2204 = const()[name = tensor("op_2204"), val = tensor([1, 1])]; tensor var_2206 = const()[name = tensor("op_2206"), val = tensor([1, 1])]; tensor x_209_pad_type_0 = const()[name = tensor("x_209_pad_type_0"), val = tensor("custom")]; tensor x_209_pad_0 = const()[name = tensor("x_209_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230586368))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231405632))), name = tensor("layers_7_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_7_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231405760)))]; tensor x_209_cast_fp16 = conv(bias = layers_7_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_2206, groups = var_2045, pad = x_209_pad_0, pad_type = x_209_pad_type_0, strides = var_2204, weight = layers_7_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_209_cast_fp16")]; tensor layers_7_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_7_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231408384)))]; tensor key_31_cast_fp16 = mul(x = x_209_cast_fp16, y = layers_7_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_31_cast_fp16")]; tensor var_2216 = const()[name = tensor("op_2216"), val = tensor([1, 1])]; tensor var_2218 = const()[name = tensor("op_2218"), val = tensor([1, 1])]; tensor x_211_pad_type_0 = const()[name = tensor("x_211_pad_type_0"), val = tensor("custom")]; tensor x_211_pad_0 = const()[name = tensor("x_211_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231411008))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232230272))), name = tensor("layers_7_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_7_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232230400)))]; tensor x_211_cast_fp16 = conv(bias = layers_7_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_2218, groups = var_2045, pad = x_211_pad_0, pad_type = x_211_pad_type_0, strides = var_2216, weight = layers_7_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_211_cast_fp16")]; tensor layers_7_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_7_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232233024)))]; tensor value_31_cast_fp16 = mul(x = x_211_cast_fp16, y = layers_7_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_31_cast_fp16")]; tensor var_2223 = const()[name = tensor("op_2223"), val = tensor([1, 20, 64, -1])]; tensor var_2224_cast_fp16 = reshape(shape = var_2223, x = query_31_cast_fp16)[name = tensor("op_2224_cast_fp16")]; tensor var_2225_to_fp16 = const()[name = tensor("op_2225_to_fp16"), val = tensor(0x1p-3)]; tensor var_2226_cast_fp16 = mul(x = var_2224_cast_fp16, y = var_2225_to_fp16)[name = tensor("op_2226_cast_fp16")]; tensor var_2227 = const()[name = tensor("op_2227"), val = tensor([1, 20, 64, -1])]; tensor var_2228_cast_fp16 = reshape(shape = var_2227, x = key_31_cast_fp16)[name = tensor("op_2228_cast_fp16")]; tensor mh_w_47_transpose_x_0 = const()[name = tensor("mh_w_47_transpose_x_0"), val = tensor(true)]; tensor mh_w_47_transpose_y_0 = const()[name = tensor("mh_w_47_transpose_y_0"), val = tensor(false)]; tensor mh_w_47_cast_fp16 = matmul(transpose_x = mh_w_47_transpose_x_0, transpose_y = mh_w_47_transpose_y_0, x = var_2226_cast_fp16, y = var_2228_cast_fp16)[name = tensor("mh_w_47_cast_fp16")]; tensor obj_111_cast_fp16 = softmax(axis = var_2038, x = mh_w_47_cast_fp16)[name = tensor("obj_111_cast_fp16")]; tensor var_2232 = const()[name = tensor("op_2232"), val = tensor([1, 20, 64, -1])]; tensor var_2233_cast_fp16 = reshape(shape = var_2232, x = value_31_cast_fp16)[name = tensor("op_2233_cast_fp16")]; tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(true)]; tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_2233_cast_fp16, y = obj_111_cast_fp16)[name = tensor("attn_31_cast_fp16")]; tensor var_2236 = const()[name = tensor("op_2236"), val = tensor([1, 1280, 1, -1])]; tensor x_213_cast_fp16 = reshape(shape = var_2236, x = attn_31_cast_fp16)[name = tensor("x_213_cast_fp16")]; tensor layers_7_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_7_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232235648)))]; tensor input_169_cast_fp16 = sub(x = x_213_cast_fp16, y = layers_7_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_169_cast_fp16")]; tensor var_2244 = const()[name = tensor("op_2244"), val = tensor([1, 1])]; tensor var_2246 = const()[name = tensor("op_2246"), val = tensor([1, 1])]; tensor x_215_pad_type_0 = const()[name = tensor("x_215_pad_type_0"), val = tensor("custom")]; tensor x_215_pad_0 = const()[name = tensor("x_215_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232238272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233057536))), name = tensor("layers_7_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_7_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233057664)))]; tensor x_215_cast_fp16 = conv(bias = layers_7_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_2246, groups = var_2045, pad = x_215_pad_0, pad_type = x_215_pad_type_0, strides = var_2244, weight = layers_7_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = tensor("x_215_cast_fp16")]; tensor layers_7_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_7_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233060288)))]; tensor obj_109_cast_fp16 = mul(x = x_215_cast_fp16, y = layers_7_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_109_cast_fp16")]; tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_109_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; tensor var_2256 = const()[name = tensor("op_2256"), val = tensor([1])]; tensor channels_mean_47_cast_fp16 = reduce_mean(axes = var_2256, keep_dims = var_2046, x = inputs_47_cast_fp16)[name = tensor("channels_mean_47_cast_fp16")]; tensor zero_mean_47_cast_fp16 = sub(x = inputs_47_cast_fp16, y = channels_mean_47_cast_fp16)[name = tensor("zero_mean_47_cast_fp16")]; tensor zero_mean_sq_47_cast_fp16 = mul(x = zero_mean_47_cast_fp16, y = zero_mean_47_cast_fp16)[name = tensor("zero_mean_sq_47_cast_fp16")]; tensor var_2260 = const()[name = tensor("op_2260"), val = tensor([1])]; tensor var_2261_cast_fp16 = reduce_mean(axes = var_2260, keep_dims = var_2046, x = zero_mean_sq_47_cast_fp16)[name = tensor("op_2261_cast_fp16")]; tensor var_2262_to_fp16 = const()[name = tensor("op_2262_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2263_cast_fp16 = add(x = var_2261_cast_fp16, y = var_2262_to_fp16)[name = tensor("op_2263_cast_fp16")]; tensor denom_47_epsilon_0_to_fp16 = const()[name = tensor("denom_47_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_47_cast_fp16 = rsqrt(epsilon = denom_47_epsilon_0_to_fp16, x = var_2263_cast_fp16)[name = tensor("denom_47_cast_fp16")]; tensor out_47_cast_fp16 = mul(x = zero_mean_47_cast_fp16, y = denom_47_cast_fp16)[name = tensor("out_47_cast_fp16")]; tensor x_217_gamma_0_to_fp16 = const()[name = tensor("x_217_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233062912)))]; tensor x_217_beta_0_to_fp16 = const()[name = tensor("x_217_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233065536)))]; tensor x_217_epsilon_0_to_fp16 = const()[name = tensor("x_217_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_217_cast_fp16 = batch_norm(beta = x_217_beta_0_to_fp16, epsilon = x_217_epsilon_0_to_fp16, gamma = x_217_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor("x_217_cast_fp16")]; tensor layers_7_fc1_input_shift_to_fp16 = const()[name = tensor("layers_7_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233068160)))]; tensor input_171_cast_fp16 = sub(x = x_217_cast_fp16, y = layers_7_fc1_input_shift_to_fp16)[name = tensor("input_171_cast_fp16")]; tensor var_2278 = const()[name = tensor("op_2278"), val = tensor([1, 1])]; tensor var_2280 = const()[name = tensor("op_2280"), val = tensor([1, 1])]; tensor x_219_pad_type_0 = const()[name = tensor("x_219_pad_type_0"), val = tensor("custom")]; tensor x_219_pad_0 = const()[name = tensor("x_219_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233070784))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236347648))), name = tensor("layers_7_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_7_fc1_module_bias_to_fp16 = const()[name = tensor("layers_7_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236347776)))]; tensor x_219_cast_fp16 = conv(bias = layers_7_fc1_module_bias_to_fp16, dilations = var_2280, groups = var_2045, pad = x_219_pad_0, pad_type = x_219_pad_type_0, strides = var_2278, weight = layers_7_fc1_module_weight_to_fp16_palettized, x = input_171_cast_fp16)[name = tensor("x_219_cast_fp16")]; tensor layers_7_fc1_output_scale_to_fp16 = const()[name = tensor("layers_7_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236358080)))]; tensor input_173_cast_fp16 = mul(x = x_219_cast_fp16, y = layers_7_fc1_output_scale_to_fp16)[name = tensor("input_173_cast_fp16")]; tensor x_221_mode_0 = const()[name = tensor("x_221_mode_0"), val = tensor("EXACT")]; tensor x_221_cast_fp16 = gelu(mode = x_221_mode_0, x = input_173_cast_fp16)[name = tensor("x_221_cast_fp16")]; tensor layers_7_fc2_input_shift_to_fp16 = const()[name = tensor("layers_7_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236368384)))]; tensor input_175_cast_fp16 = sub(x = x_221_cast_fp16, y = layers_7_fc2_input_shift_to_fp16)[name = tensor("input_175_cast_fp16")]; tensor var_2291 = const()[name = tensor("op_2291"), val = tensor([1, 1])]; tensor var_2293 = const()[name = tensor("op_2293"), val = tensor([1, 1])]; tensor x_223_pad_type_0 = const()[name = tensor("x_223_pad_type_0"), val = tensor("custom")]; tensor x_223_pad_0 = const()[name = tensor("x_223_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_7_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236378688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239655552))), name = tensor("layers_7_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_7_fc2_module_bias_to_fp16 = const()[name = tensor("layers_7_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239655680)))]; tensor x_223_cast_fp16 = conv(bias = layers_7_fc2_module_bias_to_fp16, dilations = var_2293, groups = var_2045, pad = x_223_pad_0, pad_type = x_223_pad_type_0, strides = var_2291, weight = layers_7_fc2_module_weight_to_fp16_palettized, x = input_175_cast_fp16)[name = tensor("x_223_cast_fp16")]; tensor layers_7_fc2_output_scale_to_fp16 = const()[name = tensor("layers_7_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239658304)))]; tensor hidden_states_17_cast_fp16 = mul(x = x_223_cast_fp16, y = layers_7_fc2_output_scale_to_fp16)[name = tensor("hidden_states_17_cast_fp16")]; tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_17_cast_fp16)[name = tensor("inputs_49_cast_fp16")]; tensor var_2308 = const()[name = tensor("op_2308"), val = tensor(3)]; tensor var_2315 = const()[name = tensor("op_2315"), val = tensor(1)]; tensor var_2316 = const()[name = tensor("op_2316"), val = tensor(true)]; tensor var_2328 = const()[name = tensor("op_2328"), val = tensor([1])]; tensor channels_mean_49_cast_fp16 = reduce_mean(axes = var_2328, keep_dims = var_2316, x = inputs_49_cast_fp16)[name = tensor("channels_mean_49_cast_fp16")]; tensor zero_mean_49_cast_fp16 = sub(x = inputs_49_cast_fp16, y = channels_mean_49_cast_fp16)[name = tensor("zero_mean_49_cast_fp16")]; tensor zero_mean_sq_49_cast_fp16 = mul(x = zero_mean_49_cast_fp16, y = zero_mean_49_cast_fp16)[name = tensor("zero_mean_sq_49_cast_fp16")]; tensor var_2332 = const()[name = tensor("op_2332"), val = tensor([1])]; tensor var_2333_cast_fp16 = reduce_mean(axes = var_2332, keep_dims = var_2316, x = zero_mean_sq_49_cast_fp16)[name = tensor("op_2333_cast_fp16")]; tensor var_2334_to_fp16 = const()[name = tensor("op_2334_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2335_cast_fp16 = add(x = var_2333_cast_fp16, y = var_2334_to_fp16)[name = tensor("op_2335_cast_fp16")]; tensor denom_49_epsilon_0_to_fp16 = const()[name = tensor("denom_49_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_49_cast_fp16 = rsqrt(epsilon = denom_49_epsilon_0_to_fp16, x = var_2335_cast_fp16)[name = tensor("denom_49_cast_fp16")]; tensor out_49_cast_fp16 = mul(x = zero_mean_49_cast_fp16, y = denom_49_cast_fp16)[name = tensor("out_49_cast_fp16")]; tensor obj_113_gamma_0_to_fp16 = const()[name = tensor("obj_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239660928)))]; tensor obj_113_beta_0_to_fp16 = const()[name = tensor("obj_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239663552)))]; tensor obj_113_epsilon_0_to_fp16 = const()[name = tensor("obj_113_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_113_cast_fp16 = batch_norm(beta = obj_113_beta_0_to_fp16, epsilon = obj_113_epsilon_0_to_fp16, gamma = obj_113_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor("obj_113_cast_fp16")]; tensor layers_8_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239666176)))]; tensor input_177_cast_fp16 = sub(x = obj_113_cast_fp16, y = layers_8_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_177_cast_fp16")]; tensor var_2354 = const()[name = tensor("op_2354"), val = tensor([1, 1])]; tensor var_2356 = const()[name = tensor("op_2356"), val = tensor([1, 1])]; tensor x_225_pad_type_0 = const()[name = tensor("x_225_pad_type_0"), val = tensor("custom")]; tensor x_225_pad_0 = const()[name = tensor("x_225_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239668800))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240488064))), name = tensor("layers_8_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_8_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240488192)))]; tensor x_225_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_module_bias_to_fp16, dilations = var_2356, groups = var_2315, pad = x_225_pad_0, pad_type = x_225_pad_type_0, strides = var_2354, weight = layers_8_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_177_cast_fp16)[name = tensor("x_225_cast_fp16")]; tensor layers_8_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240490816)))]; tensor query_33_cast_fp16 = mul(x = x_225_cast_fp16, y = layers_8_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_33_cast_fp16")]; tensor var_2366 = const()[name = tensor("op_2366"), val = tensor([1, 1])]; tensor var_2368 = const()[name = tensor("op_2368"), val = tensor([1, 1])]; tensor x_227_pad_type_0 = const()[name = tensor("x_227_pad_type_0"), val = tensor("custom")]; tensor x_227_pad_0 = const()[name = tensor("x_227_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240493440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(241312704))), name = tensor("layers_8_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_8_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(241312832)))]; tensor x_227_cast_fp16 = conv(bias = layers_8_self_attn_k_proj_module_bias_to_fp16, dilations = var_2368, groups = var_2315, pad = x_227_pad_0, pad_type = x_227_pad_type_0, strides = var_2366, weight = layers_8_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_177_cast_fp16)[name = tensor("x_227_cast_fp16")]; tensor layers_8_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_8_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(241315456)))]; tensor current_key_17_cast_fp16 = mul(x = x_227_cast_fp16, y = layers_8_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_17_cast_fp16")]; tensor var_2378 = const()[name = tensor("op_2378"), val = tensor([1, 1])]; tensor var_2380 = const()[name = tensor("op_2380"), val = tensor([1, 1])]; tensor x_229_pad_type_0 = const()[name = tensor("x_229_pad_type_0"), val = tensor("custom")]; tensor x_229_pad_0 = const()[name = tensor("x_229_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(241318080))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242137344))), name = tensor("layers_8_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_8_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242137472)))]; tensor x_229_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_module_bias_to_fp16, dilations = var_2380, groups = var_2315, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = var_2378, weight = layers_8_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_177_cast_fp16)[name = tensor("x_229_cast_fp16")]; tensor layers_8_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242140096)))]; tensor current_value_17_cast_fp16 = mul(x = x_229_cast_fp16, y = layers_8_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_17_cast_fp16")]; tensor var_2388_cast_fp16 = mul(x = current_key_17_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_2388_cast_fp16")]; tensor var_2390_cast_fp16 = mul(x = var_103_cast_fp16_8, y = var_257_cast_fp16)[name = tensor("op_2390_cast_fp16")]; tensor key_33_cast_fp16 = add(x = var_2388_cast_fp16, y = var_2390_cast_fp16)[name = tensor("key_33_cast_fp16")]; tensor var_2392_cast_fp16 = mul(x = current_value_17_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_2392_cast_fp16")]; tensor var_2394_cast_fp16 = mul(x = var_138_cast_fp16_8, y = var_257_cast_fp16)[name = tensor("op_2394_cast_fp16")]; tensor value_33_cast_fp16 = add(x = var_2392_cast_fp16, y = var_2394_cast_fp16)[name = tensor("value_33_cast_fp16")]; tensor var_2397 = const()[name = tensor("op_2397"), val = tensor([1, 20, 64, -1])]; tensor var_2398_cast_fp16 = reshape(shape = var_2397, x = query_33_cast_fp16)[name = tensor("op_2398_cast_fp16")]; tensor var_2399_to_fp16 = const()[name = tensor("op_2399_to_fp16"), val = tensor(0x1p-3)]; tensor var_2400_cast_fp16 = mul(x = var_2398_cast_fp16, y = var_2399_to_fp16)[name = tensor("op_2400_cast_fp16")]; tensor var_2401 = const()[name = tensor("op_2401"), val = tensor([1, 20, 64, -1])]; tensor var_2402_cast_fp16 = reshape(shape = var_2401, x = key_33_cast_fp16)[name = tensor("op_2402_cast_fp16")]; tensor mh_w_49_transpose_x_0 = const()[name = tensor("mh_w_49_transpose_x_0"), val = tensor(true)]; tensor mh_w_49_transpose_y_0 = const()[name = tensor("mh_w_49_transpose_y_0"), val = tensor(false)]; tensor mh_w_49_cast_fp16 = matmul(transpose_x = mh_w_49_transpose_x_0, transpose_y = mh_w_49_transpose_y_0, x = var_2400_cast_fp16, y = var_2402_cast_fp16)[name = tensor("mh_w_49_cast_fp16")]; tensor mh_w_51_cast_fp16 = add(x = mh_w_49_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_51_cast_fp16")]; tensor var_2410_cast_fp16 = softmax(axis = var_2308, x = mh_w_51_cast_fp16)[name = tensor("op_2410_cast_fp16")]; tensor var_2411 = const()[name = tensor("op_2411"), val = tensor([1, 20, 64, -1])]; tensor var_2412_cast_fp16 = reshape(shape = var_2411, x = value_33_cast_fp16)[name = tensor("op_2412_cast_fp16")]; tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(true)]; tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_2412_cast_fp16, y = var_2410_cast_fp16)[name = tensor("attn_33_cast_fp16")]; tensor var_2415 = const()[name = tensor("op_2415"), val = tensor([1, 1280, 1, -1])]; tensor x_231_cast_fp16 = reshape(shape = var_2415, x = attn_33_cast_fp16)[name = tensor("x_231_cast_fp16")]; tensor layers_8_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242142720)))]; tensor input_183_cast_fp16 = sub(x = x_231_cast_fp16, y = layers_8_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_183_cast_fp16")]; tensor var_2423 = const()[name = tensor("op_2423"), val = tensor([1, 1])]; tensor var_2425 = const()[name = tensor("op_2425"), val = tensor([1, 1])]; tensor x_233_pad_type_0 = const()[name = tensor("x_233_pad_type_0"), val = tensor("custom")]; tensor x_233_pad_0 = const()[name = tensor("x_233_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242145344))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242964608))), name = tensor("layers_8_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_8_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242964736)))]; tensor x_233_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_module_bias_to_fp16, dilations = var_2425, groups = var_2315, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = var_2423, weight = layers_8_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = tensor("x_233_cast_fp16")]; tensor layers_8_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242967360)))]; tensor obj_119_cast_fp16 = mul(x = x_233_cast_fp16, y = layers_8_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_119_cast_fp16")]; tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_119_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; tensor var_2436 = const()[name = tensor("op_2436"), val = tensor([1])]; tensor channels_mean_51_cast_fp16 = reduce_mean(axes = var_2436, keep_dims = var_2316, x = inputs_51_cast_fp16)[name = tensor("channels_mean_51_cast_fp16")]; tensor zero_mean_51_cast_fp16 = sub(x = inputs_51_cast_fp16, y = channels_mean_51_cast_fp16)[name = tensor("zero_mean_51_cast_fp16")]; tensor zero_mean_sq_51_cast_fp16 = mul(x = zero_mean_51_cast_fp16, y = zero_mean_51_cast_fp16)[name = tensor("zero_mean_sq_51_cast_fp16")]; tensor var_2440 = const()[name = tensor("op_2440"), val = tensor([1])]; tensor var_2441_cast_fp16 = reduce_mean(axes = var_2440, keep_dims = var_2316, x = zero_mean_sq_51_cast_fp16)[name = tensor("op_2441_cast_fp16")]; tensor var_2442_to_fp16 = const()[name = tensor("op_2442_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2443_cast_fp16 = add(x = var_2441_cast_fp16, y = var_2442_to_fp16)[name = tensor("op_2443_cast_fp16")]; tensor denom_51_epsilon_0_to_fp16 = const()[name = tensor("denom_51_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_51_cast_fp16 = rsqrt(epsilon = denom_51_epsilon_0_to_fp16, x = var_2443_cast_fp16)[name = tensor("denom_51_cast_fp16")]; tensor out_51_cast_fp16 = mul(x = zero_mean_51_cast_fp16, y = denom_51_cast_fp16)[name = tensor("out_51_cast_fp16")]; tensor obj_121_gamma_0_to_fp16 = const()[name = tensor("obj_121_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242969984)))]; tensor obj_121_beta_0_to_fp16 = const()[name = tensor("obj_121_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242972608)))]; tensor obj_121_epsilon_0_to_fp16 = const()[name = tensor("obj_121_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_121_cast_fp16 = batch_norm(beta = obj_121_beta_0_to_fp16, epsilon = obj_121_epsilon_0_to_fp16, gamma = obj_121_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_51_cast_fp16)[name = tensor("obj_121_cast_fp16")]; tensor layers_8_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_8_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242975232)))]; tensor input_185_cast_fp16 = sub(x = obj_121_cast_fp16, y = layers_8_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_185_cast_fp16")]; tensor var_2462 = const()[name = tensor("op_2462"), val = tensor([1, 1])]; tensor var_2464 = const()[name = tensor("op_2464"), val = tensor([1, 1])]; tensor x_235_pad_type_0 = const()[name = tensor("x_235_pad_type_0"), val = tensor("custom")]; tensor x_235_pad_0 = const()[name = tensor("x_235_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242977856))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243797120))), name = tensor("layers_8_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_8_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243797248)))]; tensor x_235_cast_fp16 = conv(bias = layers_8_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_2464, groups = var_2315, pad = x_235_pad_0, pad_type = x_235_pad_type_0, strides = var_2462, weight = layers_8_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_185_cast_fp16)[name = tensor("x_235_cast_fp16")]; tensor layers_8_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_8_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243799872)))]; tensor query_35_cast_fp16 = mul(x = x_235_cast_fp16, y = layers_8_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_35_cast_fp16")]; tensor var_2474 = const()[name = tensor("op_2474"), val = tensor([1, 1])]; tensor var_2476 = const()[name = tensor("op_2476"), val = tensor([1, 1])]; tensor x_237_pad_type_0 = const()[name = tensor("x_237_pad_type_0"), val = tensor("custom")]; tensor x_237_pad_0 = const()[name = tensor("x_237_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243802496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244621760))), name = tensor("layers_8_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_8_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244621888)))]; tensor x_237_cast_fp16 = conv(bias = layers_8_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_2476, groups = var_2315, pad = x_237_pad_0, pad_type = x_237_pad_type_0, strides = var_2474, weight = layers_8_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_237_cast_fp16")]; tensor layers_8_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_8_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244624512)))]; tensor key_35_cast_fp16 = mul(x = x_237_cast_fp16, y = layers_8_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_35_cast_fp16")]; tensor var_2486 = const()[name = tensor("op_2486"), val = tensor([1, 1])]; tensor var_2488 = const()[name = tensor("op_2488"), val = tensor([1, 1])]; tensor x_239_pad_type_0 = const()[name = tensor("x_239_pad_type_0"), val = tensor("custom")]; tensor x_239_pad_0 = const()[name = tensor("x_239_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244627136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245446400))), name = tensor("layers_8_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_8_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245446528)))]; tensor x_239_cast_fp16 = conv(bias = layers_8_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_2488, groups = var_2315, pad = x_239_pad_0, pad_type = x_239_pad_type_0, strides = var_2486, weight = layers_8_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_239_cast_fp16")]; tensor layers_8_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_8_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245449152)))]; tensor value_35_cast_fp16 = mul(x = x_239_cast_fp16, y = layers_8_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_35_cast_fp16")]; tensor var_2493 = const()[name = tensor("op_2493"), val = tensor([1, 20, 64, -1])]; tensor var_2494_cast_fp16 = reshape(shape = var_2493, x = query_35_cast_fp16)[name = tensor("op_2494_cast_fp16")]; tensor var_2495_to_fp16 = const()[name = tensor("op_2495_to_fp16"), val = tensor(0x1p-3)]; tensor var_2496_cast_fp16 = mul(x = var_2494_cast_fp16, y = var_2495_to_fp16)[name = tensor("op_2496_cast_fp16")]; tensor var_2497 = const()[name = tensor("op_2497"), val = tensor([1, 20, 64, -1])]; tensor var_2498_cast_fp16 = reshape(shape = var_2497, x = key_35_cast_fp16)[name = tensor("op_2498_cast_fp16")]; tensor mh_w_53_transpose_x_0 = const()[name = tensor("mh_w_53_transpose_x_0"), val = tensor(true)]; tensor mh_w_53_transpose_y_0 = const()[name = tensor("mh_w_53_transpose_y_0"), val = tensor(false)]; tensor mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_2496_cast_fp16, y = var_2498_cast_fp16)[name = tensor("mh_w_53_cast_fp16")]; tensor obj_125_cast_fp16 = softmax(axis = var_2308, x = mh_w_53_cast_fp16)[name = tensor("obj_125_cast_fp16")]; tensor var_2502 = const()[name = tensor("op_2502"), val = tensor([1, 20, 64, -1])]; tensor var_2503_cast_fp16 = reshape(shape = var_2502, x = value_35_cast_fp16)[name = tensor("op_2503_cast_fp16")]; tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(true)]; tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_2503_cast_fp16, y = obj_125_cast_fp16)[name = tensor("attn_35_cast_fp16")]; tensor var_2506 = const()[name = tensor("op_2506"), val = tensor([1, 1280, 1, -1])]; tensor x_241_cast_fp16 = reshape(shape = var_2506, x = attn_35_cast_fp16)[name = tensor("x_241_cast_fp16")]; tensor layers_8_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_8_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245451776)))]; tensor input_191_cast_fp16 = sub(x = x_241_cast_fp16, y = layers_8_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_191_cast_fp16")]; tensor var_2514 = const()[name = tensor("op_2514"), val = tensor([1, 1])]; tensor var_2516 = const()[name = tensor("op_2516"), val = tensor([1, 1])]; tensor x_243_pad_type_0 = const()[name = tensor("x_243_pad_type_0"), val = tensor("custom")]; tensor x_243_pad_0 = const()[name = tensor("x_243_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245454400))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246273664))), name = tensor("layers_8_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_8_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246273792)))]; tensor x_243_cast_fp16 = conv(bias = layers_8_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_2516, groups = var_2315, pad = x_243_pad_0, pad_type = x_243_pad_type_0, strides = var_2514, weight = layers_8_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_191_cast_fp16)[name = tensor("x_243_cast_fp16")]; tensor layers_8_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_8_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246276416)))]; tensor obj_123_cast_fp16 = mul(x = x_243_cast_fp16, y = layers_8_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_123_cast_fp16")]; tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = obj_123_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; tensor var_2523 = const()[name = tensor("op_2523"), val = tensor([1])]; tensor channels_mean_53_cast_fp16 = reduce_mean(axes = var_2523, keep_dims = var_2316, x = inputs_53_cast_fp16)[name = tensor("channels_mean_53_cast_fp16")]; tensor zero_mean_53_cast_fp16 = sub(x = inputs_53_cast_fp16, y = channels_mean_53_cast_fp16)[name = tensor("zero_mean_53_cast_fp16")]; tensor zero_mean_sq_53_cast_fp16 = mul(x = zero_mean_53_cast_fp16, y = zero_mean_53_cast_fp16)[name = tensor("zero_mean_sq_53_cast_fp16")]; tensor var_2527 = const()[name = tensor("op_2527"), val = tensor([1])]; tensor var_2528_cast_fp16 = reduce_mean(axes = var_2527, keep_dims = var_2316, x = zero_mean_sq_53_cast_fp16)[name = tensor("op_2528_cast_fp16")]; tensor var_2529_to_fp16 = const()[name = tensor("op_2529_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2530_cast_fp16 = add(x = var_2528_cast_fp16, y = var_2529_to_fp16)[name = tensor("op_2530_cast_fp16")]; tensor denom_53_epsilon_0_to_fp16 = const()[name = tensor("denom_53_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_53_cast_fp16 = rsqrt(epsilon = denom_53_epsilon_0_to_fp16, x = var_2530_cast_fp16)[name = tensor("denom_53_cast_fp16")]; tensor out_53_cast_fp16 = mul(x = zero_mean_53_cast_fp16, y = denom_53_cast_fp16)[name = tensor("out_53_cast_fp16")]; tensor x_245_gamma_0_to_fp16 = const()[name = tensor("x_245_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246279040)))]; tensor x_245_beta_0_to_fp16 = const()[name = tensor("x_245_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246281664)))]; tensor x_245_epsilon_0_to_fp16 = const()[name = tensor("x_245_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_245_cast_fp16 = batch_norm(beta = x_245_beta_0_to_fp16, epsilon = x_245_epsilon_0_to_fp16, gamma = x_245_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_53_cast_fp16)[name = tensor("x_245_cast_fp16")]; tensor layers_8_fc1_input_shift_to_fp16 = const()[name = tensor("layers_8_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246284288)))]; tensor input_193_cast_fp16 = sub(x = x_245_cast_fp16, y = layers_8_fc1_input_shift_to_fp16)[name = tensor("input_193_cast_fp16")]; tensor var_2545 = const()[name = tensor("op_2545"), val = tensor([1, 1])]; tensor var_2547 = const()[name = tensor("op_2547"), val = tensor([1, 1])]; tensor x_247_pad_type_0 = const()[name = tensor("x_247_pad_type_0"), val = tensor("custom")]; tensor x_247_pad_0 = const()[name = tensor("x_247_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246286912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249563776))), name = tensor("layers_8_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_8_fc1_module_bias_to_fp16 = const()[name = tensor("layers_8_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249563904)))]; tensor x_247_cast_fp16 = conv(bias = layers_8_fc1_module_bias_to_fp16, dilations = var_2547, groups = var_2315, pad = x_247_pad_0, pad_type = x_247_pad_type_0, strides = var_2545, weight = layers_8_fc1_module_weight_to_fp16_palettized, x = input_193_cast_fp16)[name = tensor("x_247_cast_fp16")]; tensor layers_8_fc1_output_scale_to_fp16 = const()[name = tensor("layers_8_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249574208)))]; tensor input_195_cast_fp16 = mul(x = x_247_cast_fp16, y = layers_8_fc1_output_scale_to_fp16)[name = tensor("input_195_cast_fp16")]; tensor x_249_mode_0 = const()[name = tensor("x_249_mode_0"), val = tensor("EXACT")]; tensor x_249_cast_fp16 = gelu(mode = x_249_mode_0, x = input_195_cast_fp16)[name = tensor("x_249_cast_fp16")]; tensor layers_8_fc2_input_shift_to_fp16 = const()[name = tensor("layers_8_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249584512)))]; tensor input_197_cast_fp16 = sub(x = x_249_cast_fp16, y = layers_8_fc2_input_shift_to_fp16)[name = tensor("input_197_cast_fp16")]; tensor var_2558 = const()[name = tensor("op_2558"), val = tensor([1, 1])]; tensor var_2560 = const()[name = tensor("op_2560"), val = tensor([1, 1])]; tensor x_251_pad_type_0 = const()[name = tensor("x_251_pad_type_0"), val = tensor("custom")]; tensor x_251_pad_0 = const()[name = tensor("x_251_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_8_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249594816))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252871680))), name = tensor("layers_8_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_8_fc2_module_bias_to_fp16 = const()[name = tensor("layers_8_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252871808)))]; tensor x_251_cast_fp16 = conv(bias = layers_8_fc2_module_bias_to_fp16, dilations = var_2560, groups = var_2315, pad = x_251_pad_0, pad_type = x_251_pad_type_0, strides = var_2558, weight = layers_8_fc2_module_weight_to_fp16_palettized, x = input_197_cast_fp16)[name = tensor("x_251_cast_fp16")]; tensor layers_8_fc2_output_scale_to_fp16 = const()[name = tensor("layers_8_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252874432)))]; tensor hidden_states_19_cast_fp16 = mul(x = x_251_cast_fp16, y = layers_8_fc2_output_scale_to_fp16)[name = tensor("hidden_states_19_cast_fp16")]; tensor inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = hidden_states_19_cast_fp16)[name = tensor("inputs_55_cast_fp16")]; tensor var_2574 = const()[name = tensor("op_2574"), val = tensor(3)]; tensor var_2581 = const()[name = tensor("op_2581"), val = tensor(1)]; tensor var_2582 = const()[name = tensor("op_2582"), val = tensor(true)]; tensor var_2594 = const()[name = tensor("op_2594"), val = tensor([1])]; tensor channels_mean_55_cast_fp16 = reduce_mean(axes = var_2594, keep_dims = var_2582, x = inputs_55_cast_fp16)[name = tensor("channels_mean_55_cast_fp16")]; tensor zero_mean_55_cast_fp16 = sub(x = inputs_55_cast_fp16, y = channels_mean_55_cast_fp16)[name = tensor("zero_mean_55_cast_fp16")]; tensor zero_mean_sq_55_cast_fp16 = mul(x = zero_mean_55_cast_fp16, y = zero_mean_55_cast_fp16)[name = tensor("zero_mean_sq_55_cast_fp16")]; tensor var_2598 = const()[name = tensor("op_2598"), val = tensor([1])]; tensor var_2599_cast_fp16 = reduce_mean(axes = var_2598, keep_dims = var_2582, x = zero_mean_sq_55_cast_fp16)[name = tensor("op_2599_cast_fp16")]; tensor var_2600_to_fp16 = const()[name = tensor("op_2600_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2601_cast_fp16 = add(x = var_2599_cast_fp16, y = var_2600_to_fp16)[name = tensor("op_2601_cast_fp16")]; tensor denom_55_epsilon_0_to_fp16 = const()[name = tensor("denom_55_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_55_cast_fp16 = rsqrt(epsilon = denom_55_epsilon_0_to_fp16, x = var_2601_cast_fp16)[name = tensor("denom_55_cast_fp16")]; tensor out_55_cast_fp16 = mul(x = zero_mean_55_cast_fp16, y = denom_55_cast_fp16)[name = tensor("out_55_cast_fp16")]; tensor obj_127_gamma_0_to_fp16 = const()[name = tensor("obj_127_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252877056)))]; tensor obj_127_beta_0_to_fp16 = const()[name = tensor("obj_127_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252879680)))]; tensor obj_127_epsilon_0_to_fp16 = const()[name = tensor("obj_127_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_127_cast_fp16 = batch_norm(beta = obj_127_beta_0_to_fp16, epsilon = obj_127_epsilon_0_to_fp16, gamma = obj_127_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor("obj_127_cast_fp16")]; tensor layers_9_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252882304)))]; tensor input_199_cast_fp16 = sub(x = obj_127_cast_fp16, y = layers_9_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_199_cast_fp16")]; tensor var_2620 = const()[name = tensor("op_2620"), val = tensor([1, 1])]; tensor var_2622 = const()[name = tensor("op_2622"), val = tensor([1, 1])]; tensor x_253_pad_type_0 = const()[name = tensor("x_253_pad_type_0"), val = tensor("custom")]; tensor x_253_pad_0 = const()[name = tensor("x_253_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252884928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253704192))), name = tensor("layers_9_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_9_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253704320)))]; tensor x_253_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_module_bias_to_fp16, dilations = var_2622, groups = var_2581, pad = x_253_pad_0, pad_type = x_253_pad_type_0, strides = var_2620, weight = layers_9_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = tensor("x_253_cast_fp16")]; tensor layers_9_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253706944)))]; tensor query_37_cast_fp16 = mul(x = x_253_cast_fp16, y = layers_9_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_37_cast_fp16")]; tensor var_2632 = const()[name = tensor("op_2632"), val = tensor([1, 1])]; tensor var_2634 = const()[name = tensor("op_2634"), val = tensor([1, 1])]; tensor x_255_pad_type_0 = const()[name = tensor("x_255_pad_type_0"), val = tensor("custom")]; tensor x_255_pad_0 = const()[name = tensor("x_255_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253709568))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254528832))), name = tensor("layers_9_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_9_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254528960)))]; tensor x_255_cast_fp16 = conv(bias = layers_9_self_attn_k_proj_module_bias_to_fp16, dilations = var_2634, groups = var_2581, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = var_2632, weight = layers_9_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = tensor("x_255_cast_fp16")]; tensor layers_9_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_9_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254531584)))]; tensor current_key_19_cast_fp16 = mul(x = x_255_cast_fp16, y = layers_9_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_19_cast_fp16")]; tensor var_2644 = const()[name = tensor("op_2644"), val = tensor([1, 1])]; tensor var_2646 = const()[name = tensor("op_2646"), val = tensor([1, 1])]; tensor x_257_pad_type_0 = const()[name = tensor("x_257_pad_type_0"), val = tensor("custom")]; tensor x_257_pad_0 = const()[name = tensor("x_257_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254534208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255353472))), name = tensor("layers_9_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_9_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255353600)))]; tensor x_257_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_module_bias_to_fp16, dilations = var_2646, groups = var_2581, pad = x_257_pad_0, pad_type = x_257_pad_type_0, strides = var_2644, weight = layers_9_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = tensor("x_257_cast_fp16")]; tensor layers_9_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255356224)))]; tensor current_value_19_cast_fp16 = mul(x = x_257_cast_fp16, y = layers_9_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_19_cast_fp16")]; tensor var_2654_cast_fp16 = mul(x = current_key_19_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_2654_cast_fp16")]; tensor var_2656_cast_fp16 = mul(x = var_103_cast_fp16_9, y = var_257_cast_fp16)[name = tensor("op_2656_cast_fp16")]; tensor key_37_cast_fp16 = add(x = var_2654_cast_fp16, y = var_2656_cast_fp16)[name = tensor("key_37_cast_fp16")]; tensor var_2658_cast_fp16 = mul(x = current_value_19_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_2658_cast_fp16")]; tensor var_2660_cast_fp16 = mul(x = var_138_cast_fp16_9, y = var_257_cast_fp16)[name = tensor("op_2660_cast_fp16")]; tensor value_37_cast_fp16 = add(x = var_2658_cast_fp16, y = var_2660_cast_fp16)[name = tensor("value_37_cast_fp16")]; tensor var_2663 = const()[name = tensor("op_2663"), val = tensor([1, 20, 64, -1])]; tensor var_2664_cast_fp16 = reshape(shape = var_2663, x = query_37_cast_fp16)[name = tensor("op_2664_cast_fp16")]; tensor var_2665_to_fp16 = const()[name = tensor("op_2665_to_fp16"), val = tensor(0x1p-3)]; tensor var_2666_cast_fp16 = mul(x = var_2664_cast_fp16, y = var_2665_to_fp16)[name = tensor("op_2666_cast_fp16")]; tensor var_2667 = const()[name = tensor("op_2667"), val = tensor([1, 20, 64, -1])]; tensor var_2668_cast_fp16 = reshape(shape = var_2667, x = key_37_cast_fp16)[name = tensor("op_2668_cast_fp16")]; tensor mh_w_55_transpose_x_0 = const()[name = tensor("mh_w_55_transpose_x_0"), val = tensor(true)]; tensor mh_w_55_transpose_y_0 = const()[name = tensor("mh_w_55_transpose_y_0"), val = tensor(false)]; tensor mh_w_55_cast_fp16 = matmul(transpose_x = mh_w_55_transpose_x_0, transpose_y = mh_w_55_transpose_y_0, x = var_2666_cast_fp16, y = var_2668_cast_fp16)[name = tensor("mh_w_55_cast_fp16")]; tensor mh_w_57_cast_fp16 = add(x = mh_w_55_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_57_cast_fp16")]; tensor var_2676_cast_fp16 = softmax(axis = var_2574, x = mh_w_57_cast_fp16)[name = tensor("op_2676_cast_fp16")]; tensor var_2677 = const()[name = tensor("op_2677"), val = tensor([1, 20, 64, -1])]; tensor var_2678_cast_fp16 = reshape(shape = var_2677, x = value_37_cast_fp16)[name = tensor("op_2678_cast_fp16")]; tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(true)]; tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_2678_cast_fp16, y = var_2676_cast_fp16)[name = tensor("attn_37_cast_fp16")]; tensor var_2681 = const()[name = tensor("op_2681"), val = tensor([1, 1280, 1, -1])]; tensor x_259_cast_fp16 = reshape(shape = var_2681, x = attn_37_cast_fp16)[name = tensor("x_259_cast_fp16")]; tensor layers_9_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255358848)))]; tensor input_205_cast_fp16 = sub(x = x_259_cast_fp16, y = layers_9_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_205_cast_fp16")]; tensor var_2689 = const()[name = tensor("op_2689"), val = tensor([1, 1])]; tensor var_2691 = const()[name = tensor("op_2691"), val = tensor([1, 1])]; tensor x_261_pad_type_0 = const()[name = tensor("x_261_pad_type_0"), val = tensor("custom")]; tensor x_261_pad_0 = const()[name = tensor("x_261_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255361472))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256180736))), name = tensor("layers_9_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_9_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256180864)))]; tensor x_261_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_module_bias_to_fp16, dilations = var_2691, groups = var_2581, pad = x_261_pad_0, pad_type = x_261_pad_type_0, strides = var_2689, weight = layers_9_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_205_cast_fp16)[name = tensor("x_261_cast_fp16")]; tensor layers_9_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256183488)))]; tensor obj_133_cast_fp16 = mul(x = x_261_cast_fp16, y = layers_9_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_133_cast_fp16")]; tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = obj_133_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; tensor var_2702 = const()[name = tensor("op_2702"), val = tensor([1])]; tensor channels_mean_57_cast_fp16 = reduce_mean(axes = var_2702, keep_dims = var_2582, x = inputs_57_cast_fp16)[name = tensor("channels_mean_57_cast_fp16")]; tensor zero_mean_57_cast_fp16 = sub(x = inputs_57_cast_fp16, y = channels_mean_57_cast_fp16)[name = tensor("zero_mean_57_cast_fp16")]; tensor zero_mean_sq_57_cast_fp16 = mul(x = zero_mean_57_cast_fp16, y = zero_mean_57_cast_fp16)[name = tensor("zero_mean_sq_57_cast_fp16")]; tensor var_2706 = const()[name = tensor("op_2706"), val = tensor([1])]; tensor var_2707_cast_fp16 = reduce_mean(axes = var_2706, keep_dims = var_2582, x = zero_mean_sq_57_cast_fp16)[name = tensor("op_2707_cast_fp16")]; tensor var_2708_to_fp16 = const()[name = tensor("op_2708_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2709_cast_fp16 = add(x = var_2707_cast_fp16, y = var_2708_to_fp16)[name = tensor("op_2709_cast_fp16")]; tensor denom_57_epsilon_0_to_fp16 = const()[name = tensor("denom_57_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_57_cast_fp16 = rsqrt(epsilon = denom_57_epsilon_0_to_fp16, x = var_2709_cast_fp16)[name = tensor("denom_57_cast_fp16")]; tensor out_57_cast_fp16 = mul(x = zero_mean_57_cast_fp16, y = denom_57_cast_fp16)[name = tensor("out_57_cast_fp16")]; tensor obj_135_gamma_0_to_fp16 = const()[name = tensor("obj_135_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256186112)))]; tensor obj_135_beta_0_to_fp16 = const()[name = tensor("obj_135_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256188736)))]; tensor obj_135_epsilon_0_to_fp16 = const()[name = tensor("obj_135_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_135_cast_fp16 = batch_norm(beta = obj_135_beta_0_to_fp16, epsilon = obj_135_epsilon_0_to_fp16, gamma = obj_135_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor("obj_135_cast_fp16")]; tensor layers_9_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_9_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256191360)))]; tensor input_207_cast_fp16 = sub(x = obj_135_cast_fp16, y = layers_9_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_207_cast_fp16")]; tensor var_2728 = const()[name = tensor("op_2728"), val = tensor([1, 1])]; tensor var_2730 = const()[name = tensor("op_2730"), val = tensor([1, 1])]; tensor x_263_pad_type_0 = const()[name = tensor("x_263_pad_type_0"), val = tensor("custom")]; tensor x_263_pad_0 = const()[name = tensor("x_263_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256193984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257013248))), name = tensor("layers_9_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_9_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257013376)))]; tensor x_263_cast_fp16 = conv(bias = layers_9_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_2730, groups = var_2581, pad = x_263_pad_0, pad_type = x_263_pad_type_0, strides = var_2728, weight = layers_9_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_207_cast_fp16)[name = tensor("x_263_cast_fp16")]; tensor layers_9_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_9_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257016000)))]; tensor query_39_cast_fp16 = mul(x = x_263_cast_fp16, y = layers_9_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_39_cast_fp16")]; tensor var_2740 = const()[name = tensor("op_2740"), val = tensor([1, 1])]; tensor var_2742 = const()[name = tensor("op_2742"), val = tensor([1, 1])]; tensor x_265_pad_type_0 = const()[name = tensor("x_265_pad_type_0"), val = tensor("custom")]; tensor x_265_pad_0 = const()[name = tensor("x_265_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257018624))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257837888))), name = tensor("layers_9_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_9_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257838016)))]; tensor x_265_cast_fp16 = conv(bias = layers_9_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_2742, groups = var_2581, pad = x_265_pad_0, pad_type = x_265_pad_type_0, strides = var_2740, weight = layers_9_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_265_cast_fp16")]; tensor layers_9_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_9_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257840640)))]; tensor key_39_cast_fp16 = mul(x = x_265_cast_fp16, y = layers_9_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_39_cast_fp16")]; tensor var_2752 = const()[name = tensor("op_2752"), val = tensor([1, 1])]; tensor var_2754 = const()[name = tensor("op_2754"), val = tensor([1, 1])]; tensor x_267_pad_type_0 = const()[name = tensor("x_267_pad_type_0"), val = tensor("custom")]; tensor x_267_pad_0 = const()[name = tensor("x_267_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257843264))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258662528))), name = tensor("layers_9_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_9_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258662656)))]; tensor x_267_cast_fp16 = conv(bias = layers_9_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_2754, groups = var_2581, pad = x_267_pad_0, pad_type = x_267_pad_type_0, strides = var_2752, weight = layers_9_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_267_cast_fp16")]; tensor layers_9_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_9_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258665280)))]; tensor value_39_cast_fp16 = mul(x = x_267_cast_fp16, y = layers_9_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_39_cast_fp16")]; tensor var_2759 = const()[name = tensor("op_2759"), val = tensor([1, 20, 64, -1])]; tensor var_2760_cast_fp16 = reshape(shape = var_2759, x = query_39_cast_fp16)[name = tensor("op_2760_cast_fp16")]; tensor var_2761_to_fp16 = const()[name = tensor("op_2761_to_fp16"), val = tensor(0x1p-3)]; tensor var_2762_cast_fp16 = mul(x = var_2760_cast_fp16, y = var_2761_to_fp16)[name = tensor("op_2762_cast_fp16")]; tensor var_2763 = const()[name = tensor("op_2763"), val = tensor([1, 20, 64, -1])]; tensor var_2764_cast_fp16 = reshape(shape = var_2763, x = key_39_cast_fp16)[name = tensor("op_2764_cast_fp16")]; tensor mh_w_59_transpose_x_0 = const()[name = tensor("mh_w_59_transpose_x_0"), val = tensor(true)]; tensor mh_w_59_transpose_y_0 = const()[name = tensor("mh_w_59_transpose_y_0"), val = tensor(false)]; tensor mh_w_59_cast_fp16 = matmul(transpose_x = mh_w_59_transpose_x_0, transpose_y = mh_w_59_transpose_y_0, x = var_2762_cast_fp16, y = var_2764_cast_fp16)[name = tensor("mh_w_59_cast_fp16")]; tensor obj_139_cast_fp16 = softmax(axis = var_2574, x = mh_w_59_cast_fp16)[name = tensor("obj_139_cast_fp16")]; tensor var_2768 = const()[name = tensor("op_2768"), val = tensor([1, 20, 64, -1])]; tensor var_2769_cast_fp16 = reshape(shape = var_2768, x = value_39_cast_fp16)[name = tensor("op_2769_cast_fp16")]; tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(true)]; tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_2769_cast_fp16, y = obj_139_cast_fp16)[name = tensor("attn_39_cast_fp16")]; tensor var_2772 = const()[name = tensor("op_2772"), val = tensor([1, 1280, 1, -1])]; tensor x_269_cast_fp16 = reshape(shape = var_2772, x = attn_39_cast_fp16)[name = tensor("x_269_cast_fp16")]; tensor layers_9_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_9_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258667904)))]; tensor input_213_cast_fp16 = sub(x = x_269_cast_fp16, y = layers_9_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_213_cast_fp16")]; tensor var_2780 = const()[name = tensor("op_2780"), val = tensor([1, 1])]; tensor var_2782 = const()[name = tensor("op_2782"), val = tensor([1, 1])]; tensor x_271_pad_type_0 = const()[name = tensor("x_271_pad_type_0"), val = tensor("custom")]; tensor x_271_pad_0 = const()[name = tensor("x_271_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258670528))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259489792))), name = tensor("layers_9_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_9_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259489920)))]; tensor x_271_cast_fp16 = conv(bias = layers_9_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_2782, groups = var_2581, pad = x_271_pad_0, pad_type = x_271_pad_type_0, strides = var_2780, weight = layers_9_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = tensor("x_271_cast_fp16")]; tensor layers_9_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_9_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259492544)))]; tensor obj_137_cast_fp16 = mul(x = x_271_cast_fp16, y = layers_9_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_137_cast_fp16")]; tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_137_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; tensor var_2789 = const()[name = tensor("op_2789"), val = tensor([1])]; tensor channels_mean_59_cast_fp16 = reduce_mean(axes = var_2789, keep_dims = var_2582, x = inputs_59_cast_fp16)[name = tensor("channels_mean_59_cast_fp16")]; tensor zero_mean_59_cast_fp16 = sub(x = inputs_59_cast_fp16, y = channels_mean_59_cast_fp16)[name = tensor("zero_mean_59_cast_fp16")]; tensor zero_mean_sq_59_cast_fp16 = mul(x = zero_mean_59_cast_fp16, y = zero_mean_59_cast_fp16)[name = tensor("zero_mean_sq_59_cast_fp16")]; tensor var_2793 = const()[name = tensor("op_2793"), val = tensor([1])]; tensor var_2794_cast_fp16 = reduce_mean(axes = var_2793, keep_dims = var_2582, x = zero_mean_sq_59_cast_fp16)[name = tensor("op_2794_cast_fp16")]; tensor var_2795_to_fp16 = const()[name = tensor("op_2795_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2796_cast_fp16 = add(x = var_2794_cast_fp16, y = var_2795_to_fp16)[name = tensor("op_2796_cast_fp16")]; tensor denom_59_epsilon_0_to_fp16 = const()[name = tensor("denom_59_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_59_cast_fp16 = rsqrt(epsilon = denom_59_epsilon_0_to_fp16, x = var_2796_cast_fp16)[name = tensor("denom_59_cast_fp16")]; tensor out_59_cast_fp16 = mul(x = zero_mean_59_cast_fp16, y = denom_59_cast_fp16)[name = tensor("out_59_cast_fp16")]; tensor x_273_gamma_0_to_fp16 = const()[name = tensor("x_273_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259495168)))]; tensor x_273_beta_0_to_fp16 = const()[name = tensor("x_273_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259497792)))]; tensor x_273_epsilon_0_to_fp16 = const()[name = tensor("x_273_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_273_cast_fp16 = batch_norm(beta = x_273_beta_0_to_fp16, epsilon = x_273_epsilon_0_to_fp16, gamma = x_273_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_59_cast_fp16)[name = tensor("x_273_cast_fp16")]; tensor layers_9_fc1_input_shift_to_fp16 = const()[name = tensor("layers_9_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259500416)))]; tensor input_215_cast_fp16 = sub(x = x_273_cast_fp16, y = layers_9_fc1_input_shift_to_fp16)[name = tensor("input_215_cast_fp16")]; tensor var_2811 = const()[name = tensor("op_2811"), val = tensor([1, 1])]; tensor var_2813 = const()[name = tensor("op_2813"), val = tensor([1, 1])]; tensor x_275_pad_type_0 = const()[name = tensor("x_275_pad_type_0"), val = tensor("custom")]; tensor x_275_pad_0 = const()[name = tensor("x_275_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259503040))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262779904))), name = tensor("layers_9_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_9_fc1_module_bias_to_fp16 = const()[name = tensor("layers_9_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262780032)))]; tensor x_275_cast_fp16 = conv(bias = layers_9_fc1_module_bias_to_fp16, dilations = var_2813, groups = var_2581, pad = x_275_pad_0, pad_type = x_275_pad_type_0, strides = var_2811, weight = layers_9_fc1_module_weight_to_fp16_palettized, x = input_215_cast_fp16)[name = tensor("x_275_cast_fp16")]; tensor layers_9_fc1_output_scale_to_fp16 = const()[name = tensor("layers_9_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262790336)))]; tensor input_217_cast_fp16 = mul(x = x_275_cast_fp16, y = layers_9_fc1_output_scale_to_fp16)[name = tensor("input_217_cast_fp16")]; tensor x_277_mode_0 = const()[name = tensor("x_277_mode_0"), val = tensor("EXACT")]; tensor x_277_cast_fp16 = gelu(mode = x_277_mode_0, x = input_217_cast_fp16)[name = tensor("x_277_cast_fp16")]; tensor layers_9_fc2_input_shift_to_fp16 = const()[name = tensor("layers_9_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262800640)))]; tensor input_219_cast_fp16 = sub(x = x_277_cast_fp16, y = layers_9_fc2_input_shift_to_fp16)[name = tensor("input_219_cast_fp16")]; tensor var_2824 = const()[name = tensor("op_2824"), val = tensor([1, 1])]; tensor var_2826 = const()[name = tensor("op_2826"), val = tensor([1, 1])]; tensor x_279_pad_type_0 = const()[name = tensor("x_279_pad_type_0"), val = tensor("custom")]; tensor x_279_pad_0 = const()[name = tensor("x_279_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_9_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262810944))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266087808))), name = tensor("layers_9_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_9_fc2_module_bias_to_fp16 = const()[name = tensor("layers_9_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266087936)))]; tensor x_279_cast_fp16 = conv(bias = layers_9_fc2_module_bias_to_fp16, dilations = var_2826, groups = var_2581, pad = x_279_pad_0, pad_type = x_279_pad_type_0, strides = var_2824, weight = layers_9_fc2_module_weight_to_fp16_palettized, x = input_219_cast_fp16)[name = tensor("x_279_cast_fp16")]; tensor layers_9_fc2_output_scale_to_fp16 = const()[name = tensor("layers_9_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266090560)))]; tensor hidden_states_21_cast_fp16 = mul(x = x_279_cast_fp16, y = layers_9_fc2_output_scale_to_fp16)[name = tensor("hidden_states_21_cast_fp16")]; tensor inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = hidden_states_21_cast_fp16)[name = tensor("inputs_61_cast_fp16")]; tensor var_2840 = const()[name = tensor("op_2840"), val = tensor(3)]; tensor var_2847 = const()[name = tensor("op_2847"), val = tensor(1)]; tensor var_2848 = const()[name = tensor("op_2848"), val = tensor(true)]; tensor var_2860 = const()[name = tensor("op_2860"), val = tensor([1])]; tensor channels_mean_61_cast_fp16 = reduce_mean(axes = var_2860, keep_dims = var_2848, x = inputs_61_cast_fp16)[name = tensor("channels_mean_61_cast_fp16")]; tensor zero_mean_61_cast_fp16 = sub(x = inputs_61_cast_fp16, y = channels_mean_61_cast_fp16)[name = tensor("zero_mean_61_cast_fp16")]; tensor zero_mean_sq_61_cast_fp16 = mul(x = zero_mean_61_cast_fp16, y = zero_mean_61_cast_fp16)[name = tensor("zero_mean_sq_61_cast_fp16")]; tensor var_2864 = const()[name = tensor("op_2864"), val = tensor([1])]; tensor var_2865_cast_fp16 = reduce_mean(axes = var_2864, keep_dims = var_2848, x = zero_mean_sq_61_cast_fp16)[name = tensor("op_2865_cast_fp16")]; tensor var_2866_to_fp16 = const()[name = tensor("op_2866_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2867_cast_fp16 = add(x = var_2865_cast_fp16, y = var_2866_to_fp16)[name = tensor("op_2867_cast_fp16")]; tensor denom_61_epsilon_0_to_fp16 = const()[name = tensor("denom_61_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_61_cast_fp16 = rsqrt(epsilon = denom_61_epsilon_0_to_fp16, x = var_2867_cast_fp16)[name = tensor("denom_61_cast_fp16")]; tensor out_61_cast_fp16 = mul(x = zero_mean_61_cast_fp16, y = denom_61_cast_fp16)[name = tensor("out_61_cast_fp16")]; tensor obj_141_gamma_0_to_fp16 = const()[name = tensor("obj_141_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266093184)))]; tensor obj_141_beta_0_to_fp16 = const()[name = tensor("obj_141_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266095808)))]; tensor obj_141_epsilon_0_to_fp16 = const()[name = tensor("obj_141_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_141_cast_fp16 = batch_norm(beta = obj_141_beta_0_to_fp16, epsilon = obj_141_epsilon_0_to_fp16, gamma = obj_141_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_61_cast_fp16)[name = tensor("obj_141_cast_fp16")]; tensor layers_10_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266098432)))]; tensor input_221_cast_fp16 = sub(x = obj_141_cast_fp16, y = layers_10_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_221_cast_fp16")]; tensor var_2886 = const()[name = tensor("op_2886"), val = tensor([1, 1])]; tensor var_2888 = const()[name = tensor("op_2888"), val = tensor([1, 1])]; tensor x_281_pad_type_0 = const()[name = tensor("x_281_pad_type_0"), val = tensor("custom")]; tensor x_281_pad_0 = const()[name = tensor("x_281_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266101056))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266920320))), name = tensor("layers_10_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_10_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266920448)))]; tensor x_281_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_module_bias_to_fp16, dilations = var_2888, groups = var_2847, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = var_2886, weight = layers_10_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = tensor("x_281_cast_fp16")]; tensor layers_10_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266923072)))]; tensor query_41_cast_fp16 = mul(x = x_281_cast_fp16, y = layers_10_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_41_cast_fp16")]; tensor var_2898 = const()[name = tensor("op_2898"), val = tensor([1, 1])]; tensor var_2900 = const()[name = tensor("op_2900"), val = tensor([1, 1])]; tensor x_283_pad_type_0 = const()[name = tensor("x_283_pad_type_0"), val = tensor("custom")]; tensor x_283_pad_0 = const()[name = tensor("x_283_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266925696))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267744960))), name = tensor("layers_10_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_10_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267745088)))]; tensor x_283_cast_fp16 = conv(bias = layers_10_self_attn_k_proj_module_bias_to_fp16, dilations = var_2900, groups = var_2847, pad = x_283_pad_0, pad_type = x_283_pad_type_0, strides = var_2898, weight = layers_10_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = tensor("x_283_cast_fp16")]; tensor layers_10_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_10_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267747712)))]; tensor current_key_21_cast_fp16 = mul(x = x_283_cast_fp16, y = layers_10_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_21_cast_fp16")]; tensor var_2910 = const()[name = tensor("op_2910"), val = tensor([1, 1])]; tensor var_2912 = const()[name = tensor("op_2912"), val = tensor([1, 1])]; tensor x_285_pad_type_0 = const()[name = tensor("x_285_pad_type_0"), val = tensor("custom")]; tensor x_285_pad_0 = const()[name = tensor("x_285_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267750336))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268569600))), name = tensor("layers_10_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_10_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268569728)))]; tensor x_285_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_module_bias_to_fp16, dilations = var_2912, groups = var_2847, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = var_2910, weight = layers_10_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = tensor("x_285_cast_fp16")]; tensor layers_10_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268572352)))]; tensor current_value_21_cast_fp16 = mul(x = x_285_cast_fp16, y = layers_10_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_21_cast_fp16")]; tensor var_2920_cast_fp16 = mul(x = current_key_21_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_2920_cast_fp16")]; tensor var_2922_cast_fp16 = mul(x = var_103_cast_fp16_10, y = var_257_cast_fp16)[name = tensor("op_2922_cast_fp16")]; tensor key_41_cast_fp16 = add(x = var_2920_cast_fp16, y = var_2922_cast_fp16)[name = tensor("key_41_cast_fp16")]; tensor var_2924_cast_fp16 = mul(x = current_value_21_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_2924_cast_fp16")]; tensor var_2926_cast_fp16 = mul(x = var_138_cast_fp16_10, y = var_257_cast_fp16)[name = tensor("op_2926_cast_fp16")]; tensor value_41_cast_fp16 = add(x = var_2924_cast_fp16, y = var_2926_cast_fp16)[name = tensor("value_41_cast_fp16")]; tensor var_2929 = const()[name = tensor("op_2929"), val = tensor([1, 20, 64, -1])]; tensor var_2930_cast_fp16 = reshape(shape = var_2929, x = query_41_cast_fp16)[name = tensor("op_2930_cast_fp16")]; tensor var_2931_to_fp16 = const()[name = tensor("op_2931_to_fp16"), val = tensor(0x1p-3)]; tensor var_2932_cast_fp16 = mul(x = var_2930_cast_fp16, y = var_2931_to_fp16)[name = tensor("op_2932_cast_fp16")]; tensor var_2933 = const()[name = tensor("op_2933"), val = tensor([1, 20, 64, -1])]; tensor var_2934_cast_fp16 = reshape(shape = var_2933, x = key_41_cast_fp16)[name = tensor("op_2934_cast_fp16")]; tensor mh_w_61_transpose_x_0 = const()[name = tensor("mh_w_61_transpose_x_0"), val = tensor(true)]; tensor mh_w_61_transpose_y_0 = const()[name = tensor("mh_w_61_transpose_y_0"), val = tensor(false)]; tensor mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_2932_cast_fp16, y = var_2934_cast_fp16)[name = tensor("mh_w_61_cast_fp16")]; tensor mh_w_63_cast_fp16 = add(x = mh_w_61_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_63_cast_fp16")]; tensor var_2942_cast_fp16 = softmax(axis = var_2840, x = mh_w_63_cast_fp16)[name = tensor("op_2942_cast_fp16")]; tensor var_2943 = const()[name = tensor("op_2943"), val = tensor([1, 20, 64, -1])]; tensor var_2944_cast_fp16 = reshape(shape = var_2943, x = value_41_cast_fp16)[name = tensor("op_2944_cast_fp16")]; tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(true)]; tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_2944_cast_fp16, y = var_2942_cast_fp16)[name = tensor("attn_41_cast_fp16")]; tensor var_2947 = const()[name = tensor("op_2947"), val = tensor([1, 1280, 1, -1])]; tensor x_287_cast_fp16 = reshape(shape = var_2947, x = attn_41_cast_fp16)[name = tensor("x_287_cast_fp16")]; tensor layers_10_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268574976)))]; tensor input_227_cast_fp16 = sub(x = x_287_cast_fp16, y = layers_10_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_227_cast_fp16")]; tensor var_2955 = const()[name = tensor("op_2955"), val = tensor([1, 1])]; tensor var_2957 = const()[name = tensor("op_2957"), val = tensor([1, 1])]; tensor x_289_pad_type_0 = const()[name = tensor("x_289_pad_type_0"), val = tensor("custom")]; tensor x_289_pad_0 = const()[name = tensor("x_289_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268577600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269396864))), name = tensor("layers_10_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_10_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269396992)))]; tensor x_289_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_module_bias_to_fp16, dilations = var_2957, groups = var_2847, pad = x_289_pad_0, pad_type = x_289_pad_type_0, strides = var_2955, weight = layers_10_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_227_cast_fp16)[name = tensor("x_289_cast_fp16")]; tensor layers_10_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269399616)))]; tensor obj_147_cast_fp16 = mul(x = x_289_cast_fp16, y = layers_10_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_147_cast_fp16")]; tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_147_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; tensor var_2968 = const()[name = tensor("op_2968"), val = tensor([1])]; tensor channels_mean_63_cast_fp16 = reduce_mean(axes = var_2968, keep_dims = var_2848, x = inputs_63_cast_fp16)[name = tensor("channels_mean_63_cast_fp16")]; tensor zero_mean_63_cast_fp16 = sub(x = inputs_63_cast_fp16, y = channels_mean_63_cast_fp16)[name = tensor("zero_mean_63_cast_fp16")]; tensor zero_mean_sq_63_cast_fp16 = mul(x = zero_mean_63_cast_fp16, y = zero_mean_63_cast_fp16)[name = tensor("zero_mean_sq_63_cast_fp16")]; tensor var_2972 = const()[name = tensor("op_2972"), val = tensor([1])]; tensor var_2973_cast_fp16 = reduce_mean(axes = var_2972, keep_dims = var_2848, x = zero_mean_sq_63_cast_fp16)[name = tensor("op_2973_cast_fp16")]; tensor var_2974_to_fp16 = const()[name = tensor("op_2974_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_2975_cast_fp16 = add(x = var_2973_cast_fp16, y = var_2974_to_fp16)[name = tensor("op_2975_cast_fp16")]; tensor denom_63_epsilon_0_to_fp16 = const()[name = tensor("denom_63_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_63_cast_fp16 = rsqrt(epsilon = denom_63_epsilon_0_to_fp16, x = var_2975_cast_fp16)[name = tensor("denom_63_cast_fp16")]; tensor out_63_cast_fp16 = mul(x = zero_mean_63_cast_fp16, y = denom_63_cast_fp16)[name = tensor("out_63_cast_fp16")]; tensor obj_149_gamma_0_to_fp16 = const()[name = tensor("obj_149_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269402240)))]; tensor obj_149_beta_0_to_fp16 = const()[name = tensor("obj_149_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269404864)))]; tensor obj_149_epsilon_0_to_fp16 = const()[name = tensor("obj_149_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_149_cast_fp16 = batch_norm(beta = obj_149_beta_0_to_fp16, epsilon = obj_149_epsilon_0_to_fp16, gamma = obj_149_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_63_cast_fp16)[name = tensor("obj_149_cast_fp16")]; tensor layers_10_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_10_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269407488)))]; tensor input_229_cast_fp16 = sub(x = obj_149_cast_fp16, y = layers_10_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_229_cast_fp16")]; tensor var_2994 = const()[name = tensor("op_2994"), val = tensor([1, 1])]; tensor var_2996 = const()[name = tensor("op_2996"), val = tensor([1, 1])]; tensor x_291_pad_type_0 = const()[name = tensor("x_291_pad_type_0"), val = tensor("custom")]; tensor x_291_pad_0 = const()[name = tensor("x_291_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269410112))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270229376))), name = tensor("layers_10_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_10_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270229504)))]; tensor x_291_cast_fp16 = conv(bias = layers_10_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_2996, groups = var_2847, pad = x_291_pad_0, pad_type = x_291_pad_type_0, strides = var_2994, weight = layers_10_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_229_cast_fp16)[name = tensor("x_291_cast_fp16")]; tensor layers_10_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_10_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270232128)))]; tensor query_43_cast_fp16 = mul(x = x_291_cast_fp16, y = layers_10_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_43_cast_fp16")]; tensor var_3006 = const()[name = tensor("op_3006"), val = tensor([1, 1])]; tensor var_3008 = const()[name = tensor("op_3008"), val = tensor([1, 1])]; tensor x_293_pad_type_0 = const()[name = tensor("x_293_pad_type_0"), val = tensor("custom")]; tensor x_293_pad_0 = const()[name = tensor("x_293_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270234752))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(271054016))), name = tensor("layers_10_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_10_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(271054144)))]; tensor x_293_cast_fp16 = conv(bias = layers_10_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_3008, groups = var_2847, pad = x_293_pad_0, pad_type = x_293_pad_type_0, strides = var_3006, weight = layers_10_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_293_cast_fp16")]; tensor layers_10_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_10_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(271056768)))]; tensor key_43_cast_fp16 = mul(x = x_293_cast_fp16, y = layers_10_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_43_cast_fp16")]; tensor var_3018 = const()[name = tensor("op_3018"), val = tensor([1, 1])]; tensor var_3020 = const()[name = tensor("op_3020"), val = tensor([1, 1])]; tensor x_295_pad_type_0 = const()[name = tensor("x_295_pad_type_0"), val = tensor("custom")]; tensor x_295_pad_0 = const()[name = tensor("x_295_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(271059392))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(271878656))), name = tensor("layers_10_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_10_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(271878784)))]; tensor x_295_cast_fp16 = conv(bias = layers_10_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_3020, groups = var_2847, pad = x_295_pad_0, pad_type = x_295_pad_type_0, strides = var_3018, weight = layers_10_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_295_cast_fp16")]; tensor layers_10_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_10_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(271881408)))]; tensor value_43_cast_fp16 = mul(x = x_295_cast_fp16, y = layers_10_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_43_cast_fp16")]; tensor var_3025 = const()[name = tensor("op_3025"), val = tensor([1, 20, 64, -1])]; tensor var_3026_cast_fp16 = reshape(shape = var_3025, x = query_43_cast_fp16)[name = tensor("op_3026_cast_fp16")]; tensor var_3027_to_fp16 = const()[name = tensor("op_3027_to_fp16"), val = tensor(0x1p-3)]; tensor var_3028_cast_fp16 = mul(x = var_3026_cast_fp16, y = var_3027_to_fp16)[name = tensor("op_3028_cast_fp16")]; tensor var_3029 = const()[name = tensor("op_3029"), val = tensor([1, 20, 64, -1])]; tensor var_3030_cast_fp16 = reshape(shape = var_3029, x = key_43_cast_fp16)[name = tensor("op_3030_cast_fp16")]; tensor mh_w_65_transpose_x_0 = const()[name = tensor("mh_w_65_transpose_x_0"), val = tensor(true)]; tensor mh_w_65_transpose_y_0 = const()[name = tensor("mh_w_65_transpose_y_0"), val = tensor(false)]; tensor mh_w_65_cast_fp16 = matmul(transpose_x = mh_w_65_transpose_x_0, transpose_y = mh_w_65_transpose_y_0, x = var_3028_cast_fp16, y = var_3030_cast_fp16)[name = tensor("mh_w_65_cast_fp16")]; tensor obj_153_cast_fp16 = softmax(axis = var_2840, x = mh_w_65_cast_fp16)[name = tensor("obj_153_cast_fp16")]; tensor var_3034 = const()[name = tensor("op_3034"), val = tensor([1, 20, 64, -1])]; tensor var_3035_cast_fp16 = reshape(shape = var_3034, x = value_43_cast_fp16)[name = tensor("op_3035_cast_fp16")]; tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(true)]; tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_3035_cast_fp16, y = obj_153_cast_fp16)[name = tensor("attn_43_cast_fp16")]; tensor var_3038 = const()[name = tensor("op_3038"), val = tensor([1, 1280, 1, -1])]; tensor x_297_cast_fp16 = reshape(shape = var_3038, x = attn_43_cast_fp16)[name = tensor("x_297_cast_fp16")]; tensor layers_10_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_10_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(271884032)))]; tensor input_235_cast_fp16 = sub(x = x_297_cast_fp16, y = layers_10_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_235_cast_fp16")]; tensor var_3046 = const()[name = tensor("op_3046"), val = tensor([1, 1])]; tensor var_3048 = const()[name = tensor("op_3048"), val = tensor([1, 1])]; tensor x_299_pad_type_0 = const()[name = tensor("x_299_pad_type_0"), val = tensor("custom")]; tensor x_299_pad_0 = const()[name = tensor("x_299_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(271886656))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272705920))), name = tensor("layers_10_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_10_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272706048)))]; tensor x_299_cast_fp16 = conv(bias = layers_10_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_3048, groups = var_2847, pad = x_299_pad_0, pad_type = x_299_pad_type_0, strides = var_3046, weight = layers_10_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_235_cast_fp16)[name = tensor("x_299_cast_fp16")]; tensor layers_10_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_10_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272708672)))]; tensor obj_151_cast_fp16 = mul(x = x_299_cast_fp16, y = layers_10_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_151_cast_fp16")]; tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = obj_151_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; tensor var_3058 = const()[name = tensor("op_3058"), val = tensor([1])]; tensor channels_mean_65_cast_fp16 = reduce_mean(axes = var_3058, keep_dims = var_2848, x = inputs_65_cast_fp16)[name = tensor("channels_mean_65_cast_fp16")]; tensor zero_mean_65_cast_fp16 = sub(x = inputs_65_cast_fp16, y = channels_mean_65_cast_fp16)[name = tensor("zero_mean_65_cast_fp16")]; tensor zero_mean_sq_65_cast_fp16 = mul(x = zero_mean_65_cast_fp16, y = zero_mean_65_cast_fp16)[name = tensor("zero_mean_sq_65_cast_fp16")]; tensor var_3062 = const()[name = tensor("op_3062"), val = tensor([1])]; tensor var_3063_cast_fp16 = reduce_mean(axes = var_3062, keep_dims = var_2848, x = zero_mean_sq_65_cast_fp16)[name = tensor("op_3063_cast_fp16")]; tensor var_3064_to_fp16 = const()[name = tensor("op_3064_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_3065_cast_fp16 = add(x = var_3063_cast_fp16, y = var_3064_to_fp16)[name = tensor("op_3065_cast_fp16")]; tensor denom_65_epsilon_0_to_fp16 = const()[name = tensor("denom_65_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_65_cast_fp16 = rsqrt(epsilon = denom_65_epsilon_0_to_fp16, x = var_3065_cast_fp16)[name = tensor("denom_65_cast_fp16")]; tensor out_65_cast_fp16 = mul(x = zero_mean_65_cast_fp16, y = denom_65_cast_fp16)[name = tensor("out_65_cast_fp16")]; tensor x_301_gamma_0_to_fp16 = const()[name = tensor("x_301_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272711296)))]; tensor x_301_beta_0_to_fp16 = const()[name = tensor("x_301_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272713920)))]; tensor x_301_epsilon_0_to_fp16 = const()[name = tensor("x_301_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_301_cast_fp16 = batch_norm(beta = x_301_beta_0_to_fp16, epsilon = x_301_epsilon_0_to_fp16, gamma = x_301_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_65_cast_fp16)[name = tensor("x_301_cast_fp16")]; tensor layers_10_fc1_input_shift_to_fp16 = const()[name = tensor("layers_10_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272716544)))]; tensor input_237_cast_fp16 = sub(x = x_301_cast_fp16, y = layers_10_fc1_input_shift_to_fp16)[name = tensor("input_237_cast_fp16")]; tensor var_3080 = const()[name = tensor("op_3080"), val = tensor([1, 1])]; tensor var_3082 = const()[name = tensor("op_3082"), val = tensor([1, 1])]; tensor x_303_pad_type_0 = const()[name = tensor("x_303_pad_type_0"), val = tensor("custom")]; tensor x_303_pad_0 = const()[name = tensor("x_303_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272719168))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(275996032))), name = tensor("layers_10_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_10_fc1_module_bias_to_fp16 = const()[name = tensor("layers_10_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(275996160)))]; tensor x_303_cast_fp16 = conv(bias = layers_10_fc1_module_bias_to_fp16, dilations = var_3082, groups = var_2847, pad = x_303_pad_0, pad_type = x_303_pad_type_0, strides = var_3080, weight = layers_10_fc1_module_weight_to_fp16_palettized, x = input_237_cast_fp16)[name = tensor("x_303_cast_fp16")]; tensor layers_10_fc1_output_scale_to_fp16 = const()[name = tensor("layers_10_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276006464)))]; tensor input_239_cast_fp16 = mul(x = x_303_cast_fp16, y = layers_10_fc1_output_scale_to_fp16)[name = tensor("input_239_cast_fp16")]; tensor x_305_mode_0 = const()[name = tensor("x_305_mode_0"), val = tensor("EXACT")]; tensor x_305_cast_fp16 = gelu(mode = x_305_mode_0, x = input_239_cast_fp16)[name = tensor("x_305_cast_fp16")]; tensor layers_10_fc2_input_shift_to_fp16 = const()[name = tensor("layers_10_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276016768)))]; tensor input_241_cast_fp16 = sub(x = x_305_cast_fp16, y = layers_10_fc2_input_shift_to_fp16)[name = tensor("input_241_cast_fp16")]; tensor var_3093 = const()[name = tensor("op_3093"), val = tensor([1, 1])]; tensor var_3095 = const()[name = tensor("op_3095"), val = tensor([1, 1])]; tensor x_307_pad_type_0 = const()[name = tensor("x_307_pad_type_0"), val = tensor("custom")]; tensor x_307_pad_0 = const()[name = tensor("x_307_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_10_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276027072))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279303936))), name = tensor("layers_10_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_10_fc2_module_bias_to_fp16 = const()[name = tensor("layers_10_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279304064)))]; tensor x_307_cast_fp16 = conv(bias = layers_10_fc2_module_bias_to_fp16, dilations = var_3095, groups = var_2847, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = var_3093, weight = layers_10_fc2_module_weight_to_fp16_palettized, x = input_241_cast_fp16)[name = tensor("x_307_cast_fp16")]; tensor layers_10_fc2_output_scale_to_fp16 = const()[name = tensor("layers_10_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279306688)))]; tensor hidden_states_23_cast_fp16 = mul(x = x_307_cast_fp16, y = layers_10_fc2_output_scale_to_fp16)[name = tensor("hidden_states_23_cast_fp16")]; tensor inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = hidden_states_23_cast_fp16)[name = tensor("inputs_67_cast_fp16")]; tensor var_3110 = const()[name = tensor("op_3110"), val = tensor(3)]; tensor var_3117 = const()[name = tensor("op_3117"), val = tensor(1)]; tensor var_3118 = const()[name = tensor("op_3118"), val = tensor(true)]; tensor var_3130 = const()[name = tensor("op_3130"), val = tensor([1])]; tensor channels_mean_67_cast_fp16 = reduce_mean(axes = var_3130, keep_dims = var_3118, x = inputs_67_cast_fp16)[name = tensor("channels_mean_67_cast_fp16")]; tensor zero_mean_67_cast_fp16 = sub(x = inputs_67_cast_fp16, y = channels_mean_67_cast_fp16)[name = tensor("zero_mean_67_cast_fp16")]; tensor zero_mean_sq_67_cast_fp16 = mul(x = zero_mean_67_cast_fp16, y = zero_mean_67_cast_fp16)[name = tensor("zero_mean_sq_67_cast_fp16")]; tensor var_3134 = const()[name = tensor("op_3134"), val = tensor([1])]; tensor var_3135_cast_fp16 = reduce_mean(axes = var_3134, keep_dims = var_3118, x = zero_mean_sq_67_cast_fp16)[name = tensor("op_3135_cast_fp16")]; tensor var_3136_to_fp16 = const()[name = tensor("op_3136_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_3137_cast_fp16 = add(x = var_3135_cast_fp16, y = var_3136_to_fp16)[name = tensor("op_3137_cast_fp16")]; tensor denom_67_epsilon_0_to_fp16 = const()[name = tensor("denom_67_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_67_cast_fp16 = rsqrt(epsilon = denom_67_epsilon_0_to_fp16, x = var_3137_cast_fp16)[name = tensor("denom_67_cast_fp16")]; tensor out_67_cast_fp16 = mul(x = zero_mean_67_cast_fp16, y = denom_67_cast_fp16)[name = tensor("out_67_cast_fp16")]; tensor obj_155_gamma_0_to_fp16 = const()[name = tensor("obj_155_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279309312)))]; tensor obj_155_beta_0_to_fp16 = const()[name = tensor("obj_155_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279311936)))]; tensor obj_155_epsilon_0_to_fp16 = const()[name = tensor("obj_155_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_155_cast_fp16 = batch_norm(beta = obj_155_beta_0_to_fp16, epsilon = obj_155_epsilon_0_to_fp16, gamma = obj_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor("obj_155_cast_fp16")]; tensor layers_11_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279314560)))]; tensor input_243_cast_fp16 = sub(x = obj_155_cast_fp16, y = layers_11_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_243_cast_fp16")]; tensor var_3156 = const()[name = tensor("op_3156"), val = tensor([1, 1])]; tensor var_3158 = const()[name = tensor("op_3158"), val = tensor([1, 1])]; tensor x_309_pad_type_0 = const()[name = tensor("x_309_pad_type_0"), val = tensor("custom")]; tensor x_309_pad_0 = const()[name = tensor("x_309_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279317184))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280136448))), name = tensor("layers_11_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_11_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280136576)))]; tensor x_309_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_module_bias_to_fp16, dilations = var_3158, groups = var_3117, pad = x_309_pad_0, pad_type = x_309_pad_type_0, strides = var_3156, weight = layers_11_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_243_cast_fp16)[name = tensor("x_309_cast_fp16")]; tensor layers_11_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280139200)))]; tensor query_45_cast_fp16 = mul(x = x_309_cast_fp16, y = layers_11_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_45_cast_fp16")]; tensor var_3168 = const()[name = tensor("op_3168"), val = tensor([1, 1])]; tensor var_3170 = const()[name = tensor("op_3170"), val = tensor([1, 1])]; tensor x_311_pad_type_0 = const()[name = tensor("x_311_pad_type_0"), val = tensor("custom")]; tensor x_311_pad_0 = const()[name = tensor("x_311_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280141824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280961088))), name = tensor("layers_11_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_11_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280961216)))]; tensor x_311_cast_fp16 = conv(bias = layers_11_self_attn_k_proj_module_bias_to_fp16, dilations = var_3170, groups = var_3117, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = var_3168, weight = layers_11_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_243_cast_fp16)[name = tensor("x_311_cast_fp16")]; tensor layers_11_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_11_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280963840)))]; tensor current_key_23_cast_fp16 = mul(x = x_311_cast_fp16, y = layers_11_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_23_cast_fp16")]; tensor var_3180 = const()[name = tensor("op_3180"), val = tensor([1, 1])]; tensor var_3182 = const()[name = tensor("op_3182"), val = tensor([1, 1])]; tensor x_313_pad_type_0 = const()[name = tensor("x_313_pad_type_0"), val = tensor("custom")]; tensor x_313_pad_0 = const()[name = tensor("x_313_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280966464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281785728))), name = tensor("layers_11_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_11_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281785856)))]; tensor x_313_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_module_bias_to_fp16, dilations = var_3182, groups = var_3117, pad = x_313_pad_0, pad_type = x_313_pad_type_0, strides = var_3180, weight = layers_11_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_243_cast_fp16)[name = tensor("x_313_cast_fp16")]; tensor layers_11_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281788480)))]; tensor current_value_23_cast_fp16 = mul(x = x_313_cast_fp16, y = layers_11_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_23_cast_fp16")]; tensor var_3190_cast_fp16 = mul(x = current_key_23_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_3190_cast_fp16")]; tensor var_3192_cast_fp16 = mul(x = var_103_cast_fp16_11, y = var_257_cast_fp16)[name = tensor("op_3192_cast_fp16")]; tensor key_45_cast_fp16 = add(x = var_3190_cast_fp16, y = var_3192_cast_fp16)[name = tensor("key_45_cast_fp16")]; tensor var_3194_cast_fp16 = mul(x = current_value_23_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_3194_cast_fp16")]; tensor var_3196_cast_fp16 = mul(x = var_138_cast_fp16_11, y = var_257_cast_fp16)[name = tensor("op_3196_cast_fp16")]; tensor value_45_cast_fp16 = add(x = var_3194_cast_fp16, y = var_3196_cast_fp16)[name = tensor("value_45_cast_fp16")]; tensor var_3199 = const()[name = tensor("op_3199"), val = tensor([1, 20, 64, -1])]; tensor var_3200_cast_fp16 = reshape(shape = var_3199, x = query_45_cast_fp16)[name = tensor("op_3200_cast_fp16")]; tensor var_3201_to_fp16 = const()[name = tensor("op_3201_to_fp16"), val = tensor(0x1p-3)]; tensor var_3202_cast_fp16 = mul(x = var_3200_cast_fp16, y = var_3201_to_fp16)[name = tensor("op_3202_cast_fp16")]; tensor var_3203 = const()[name = tensor("op_3203"), val = tensor([1, 20, 64, -1])]; tensor var_3204_cast_fp16 = reshape(shape = var_3203, x = key_45_cast_fp16)[name = tensor("op_3204_cast_fp16")]; tensor mh_w_67_transpose_x_0 = const()[name = tensor("mh_w_67_transpose_x_0"), val = tensor(true)]; tensor mh_w_67_transpose_y_0 = const()[name = tensor("mh_w_67_transpose_y_0"), val = tensor(false)]; tensor mh_w_67_cast_fp16 = matmul(transpose_x = mh_w_67_transpose_x_0, transpose_y = mh_w_67_transpose_y_0, x = var_3202_cast_fp16, y = var_3204_cast_fp16)[name = tensor("mh_w_67_cast_fp16")]; tensor mh_w_69_cast_fp16 = add(x = mh_w_67_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_69_cast_fp16")]; tensor var_3212_cast_fp16 = softmax(axis = var_3110, x = mh_w_69_cast_fp16)[name = tensor("op_3212_cast_fp16")]; tensor var_3213 = const()[name = tensor("op_3213"), val = tensor([1, 20, 64, -1])]; tensor var_3214_cast_fp16 = reshape(shape = var_3213, x = value_45_cast_fp16)[name = tensor("op_3214_cast_fp16")]; tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(true)]; tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_3214_cast_fp16, y = var_3212_cast_fp16)[name = tensor("attn_45_cast_fp16")]; tensor var_3217 = const()[name = tensor("op_3217"), val = tensor([1, 1280, 1, -1])]; tensor x_315_cast_fp16 = reshape(shape = var_3217, x = attn_45_cast_fp16)[name = tensor("x_315_cast_fp16")]; tensor layers_11_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281791104)))]; tensor input_249_cast_fp16 = sub(x = x_315_cast_fp16, y = layers_11_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_249_cast_fp16")]; tensor var_3225 = const()[name = tensor("op_3225"), val = tensor([1, 1])]; tensor var_3227 = const()[name = tensor("op_3227"), val = tensor([1, 1])]; tensor x_317_pad_type_0 = const()[name = tensor("x_317_pad_type_0"), val = tensor("custom")]; tensor x_317_pad_0 = const()[name = tensor("x_317_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281793728))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282612992))), name = tensor("layers_11_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_11_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282613120)))]; tensor x_317_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_module_bias_to_fp16, dilations = var_3227, groups = var_3117, pad = x_317_pad_0, pad_type = x_317_pad_type_0, strides = var_3225, weight = layers_11_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_249_cast_fp16)[name = tensor("x_317_cast_fp16")]; tensor layers_11_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282615744)))]; tensor obj_161_cast_fp16 = mul(x = x_317_cast_fp16, y = layers_11_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_161_cast_fp16")]; tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = obj_161_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; tensor var_3238 = const()[name = tensor("op_3238"), val = tensor([1])]; tensor channels_mean_69_cast_fp16 = reduce_mean(axes = var_3238, keep_dims = var_3118, x = inputs_69_cast_fp16)[name = tensor("channels_mean_69_cast_fp16")]; tensor zero_mean_69_cast_fp16 = sub(x = inputs_69_cast_fp16, y = channels_mean_69_cast_fp16)[name = tensor("zero_mean_69_cast_fp16")]; tensor zero_mean_sq_69_cast_fp16 = mul(x = zero_mean_69_cast_fp16, y = zero_mean_69_cast_fp16)[name = tensor("zero_mean_sq_69_cast_fp16")]; tensor var_3242 = const()[name = tensor("op_3242"), val = tensor([1])]; tensor var_3243_cast_fp16 = reduce_mean(axes = var_3242, keep_dims = var_3118, x = zero_mean_sq_69_cast_fp16)[name = tensor("op_3243_cast_fp16")]; tensor var_3244_to_fp16 = const()[name = tensor("op_3244_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_3245_cast_fp16 = add(x = var_3243_cast_fp16, y = var_3244_to_fp16)[name = tensor("op_3245_cast_fp16")]; tensor denom_69_epsilon_0_to_fp16 = const()[name = tensor("denom_69_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_69_cast_fp16 = rsqrt(epsilon = denom_69_epsilon_0_to_fp16, x = var_3245_cast_fp16)[name = tensor("denom_69_cast_fp16")]; tensor out_69_cast_fp16 = mul(x = zero_mean_69_cast_fp16, y = denom_69_cast_fp16)[name = tensor("out_69_cast_fp16")]; tensor obj_163_gamma_0_to_fp16 = const()[name = tensor("obj_163_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282618368)))]; tensor obj_163_beta_0_to_fp16 = const()[name = tensor("obj_163_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282620992)))]; tensor obj_163_epsilon_0_to_fp16 = const()[name = tensor("obj_163_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_163_cast_fp16 = batch_norm(beta = obj_163_beta_0_to_fp16, epsilon = obj_163_epsilon_0_to_fp16, gamma = obj_163_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor("obj_163_cast_fp16")]; tensor layers_11_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_11_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282623616)))]; tensor input_251_cast_fp16 = sub(x = obj_163_cast_fp16, y = layers_11_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_251_cast_fp16")]; tensor var_3264 = const()[name = tensor("op_3264"), val = tensor([1, 1])]; tensor var_3266 = const()[name = tensor("op_3266"), val = tensor([1, 1])]; tensor x_319_pad_type_0 = const()[name = tensor("x_319_pad_type_0"), val = tensor("custom")]; tensor x_319_pad_0 = const()[name = tensor("x_319_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282626240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283445504))), name = tensor("layers_11_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_11_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283445632)))]; tensor x_319_cast_fp16 = conv(bias = layers_11_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_3266, groups = var_3117, pad = x_319_pad_0, pad_type = x_319_pad_type_0, strides = var_3264, weight = layers_11_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = tensor("x_319_cast_fp16")]; tensor layers_11_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_11_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283448256)))]; tensor query_47_cast_fp16 = mul(x = x_319_cast_fp16, y = layers_11_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_47_cast_fp16")]; tensor var_3276 = const()[name = tensor("op_3276"), val = tensor([1, 1])]; tensor var_3278 = const()[name = tensor("op_3278"), val = tensor([1, 1])]; tensor x_321_pad_type_0 = const()[name = tensor("x_321_pad_type_0"), val = tensor("custom")]; tensor x_321_pad_0 = const()[name = tensor("x_321_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283450880))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284270144))), name = tensor("layers_11_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_11_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284270272)))]; tensor x_321_cast_fp16 = conv(bias = layers_11_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_3278, groups = var_3117, pad = x_321_pad_0, pad_type = x_321_pad_type_0, strides = var_3276, weight = layers_11_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_321_cast_fp16")]; tensor layers_11_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_11_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284272896)))]; tensor key_47_cast_fp16 = mul(x = x_321_cast_fp16, y = layers_11_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_47_cast_fp16")]; tensor var_3288 = const()[name = tensor("op_3288"), val = tensor([1, 1])]; tensor var_3290 = const()[name = tensor("op_3290"), val = tensor([1, 1])]; tensor x_323_pad_type_0 = const()[name = tensor("x_323_pad_type_0"), val = tensor("custom")]; tensor x_323_pad_0 = const()[name = tensor("x_323_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284275520))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285094784))), name = tensor("layers_11_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_11_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285094912)))]; tensor x_323_cast_fp16 = conv(bias = layers_11_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_3290, groups = var_3117, pad = x_323_pad_0, pad_type = x_323_pad_type_0, strides = var_3288, weight = layers_11_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_323_cast_fp16")]; tensor layers_11_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_11_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285097536)))]; tensor value_47_cast_fp16 = mul(x = x_323_cast_fp16, y = layers_11_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_47_cast_fp16")]; tensor var_3295 = const()[name = tensor("op_3295"), val = tensor([1, 20, 64, -1])]; tensor var_3296_cast_fp16 = reshape(shape = var_3295, x = query_47_cast_fp16)[name = tensor("op_3296_cast_fp16")]; tensor var_3297_to_fp16 = const()[name = tensor("op_3297_to_fp16"), val = tensor(0x1p-3)]; tensor var_3298_cast_fp16 = mul(x = var_3296_cast_fp16, y = var_3297_to_fp16)[name = tensor("op_3298_cast_fp16")]; tensor var_3299 = const()[name = tensor("op_3299"), val = tensor([1, 20, 64, -1])]; tensor var_3300_cast_fp16 = reshape(shape = var_3299, x = key_47_cast_fp16)[name = tensor("op_3300_cast_fp16")]; tensor mh_w_71_transpose_x_0 = const()[name = tensor("mh_w_71_transpose_x_0"), val = tensor(true)]; tensor mh_w_71_transpose_y_0 = const()[name = tensor("mh_w_71_transpose_y_0"), val = tensor(false)]; tensor mh_w_71_cast_fp16 = matmul(transpose_x = mh_w_71_transpose_x_0, transpose_y = mh_w_71_transpose_y_0, x = var_3298_cast_fp16, y = var_3300_cast_fp16)[name = tensor("mh_w_71_cast_fp16")]; tensor obj_167_cast_fp16 = softmax(axis = var_3110, x = mh_w_71_cast_fp16)[name = tensor("obj_167_cast_fp16")]; tensor var_3304 = const()[name = tensor("op_3304"), val = tensor([1, 20, 64, -1])]; tensor var_3305_cast_fp16 = reshape(shape = var_3304, x = value_47_cast_fp16)[name = tensor("op_3305_cast_fp16")]; tensor attn_47_transpose_x_0 = const()[name = tensor("attn_47_transpose_x_0"), val = tensor(false)]; tensor attn_47_transpose_y_0 = const()[name = tensor("attn_47_transpose_y_0"), val = tensor(true)]; tensor attn_47_cast_fp16 = matmul(transpose_x = attn_47_transpose_x_0, transpose_y = attn_47_transpose_y_0, x = var_3305_cast_fp16, y = obj_167_cast_fp16)[name = tensor("attn_47_cast_fp16")]; tensor var_3308 = const()[name = tensor("op_3308"), val = tensor([1, 1280, 1, -1])]; tensor x_325_cast_fp16 = reshape(shape = var_3308, x = attn_47_cast_fp16)[name = tensor("x_325_cast_fp16")]; tensor layers_11_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_11_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285100160)))]; tensor input_257_cast_fp16 = sub(x = x_325_cast_fp16, y = layers_11_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_257_cast_fp16")]; tensor var_3316 = const()[name = tensor("op_3316"), val = tensor([1, 1])]; tensor var_3318 = const()[name = tensor("op_3318"), val = tensor([1, 1])]; tensor x_327_pad_type_0 = const()[name = tensor("x_327_pad_type_0"), val = tensor("custom")]; tensor x_327_pad_0 = const()[name = tensor("x_327_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285102784))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285922048))), name = tensor("layers_11_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_11_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285922176)))]; tensor x_327_cast_fp16 = conv(bias = layers_11_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_3318, groups = var_3117, pad = x_327_pad_0, pad_type = x_327_pad_type_0, strides = var_3316, weight = layers_11_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_257_cast_fp16)[name = tensor("x_327_cast_fp16")]; tensor layers_11_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_11_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285924800)))]; tensor obj_165_cast_fp16 = mul(x = x_327_cast_fp16, y = layers_11_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_165_cast_fp16")]; tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_165_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; tensor var_3325 = const()[name = tensor("op_3325"), val = tensor([1])]; tensor channels_mean_71_cast_fp16 = reduce_mean(axes = var_3325, keep_dims = var_3118, x = inputs_71_cast_fp16)[name = tensor("channels_mean_71_cast_fp16")]; tensor zero_mean_71_cast_fp16 = sub(x = inputs_71_cast_fp16, y = channels_mean_71_cast_fp16)[name = tensor("zero_mean_71_cast_fp16")]; tensor zero_mean_sq_71_cast_fp16 = mul(x = zero_mean_71_cast_fp16, y = zero_mean_71_cast_fp16)[name = tensor("zero_mean_sq_71_cast_fp16")]; tensor var_3329 = const()[name = tensor("op_3329"), val = tensor([1])]; tensor var_3330_cast_fp16 = reduce_mean(axes = var_3329, keep_dims = var_3118, x = zero_mean_sq_71_cast_fp16)[name = tensor("op_3330_cast_fp16")]; tensor var_3331_to_fp16 = const()[name = tensor("op_3331_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_3332_cast_fp16 = add(x = var_3330_cast_fp16, y = var_3331_to_fp16)[name = tensor("op_3332_cast_fp16")]; tensor denom_71_epsilon_0_to_fp16 = const()[name = tensor("denom_71_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_71_cast_fp16 = rsqrt(epsilon = denom_71_epsilon_0_to_fp16, x = var_3332_cast_fp16)[name = tensor("denom_71_cast_fp16")]; tensor out_71_cast_fp16 = mul(x = zero_mean_71_cast_fp16, y = denom_71_cast_fp16)[name = tensor("out_71_cast_fp16")]; tensor x_329_gamma_0_to_fp16 = const()[name = tensor("x_329_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285927424)))]; tensor x_329_beta_0_to_fp16 = const()[name = tensor("x_329_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285930048)))]; tensor x_329_epsilon_0_to_fp16 = const()[name = tensor("x_329_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_329_cast_fp16 = batch_norm(beta = x_329_beta_0_to_fp16, epsilon = x_329_epsilon_0_to_fp16, gamma = x_329_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_71_cast_fp16)[name = tensor("x_329_cast_fp16")]; tensor layers_11_fc1_input_shift_to_fp16 = const()[name = tensor("layers_11_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285932672)))]; tensor input_259_cast_fp16 = sub(x = x_329_cast_fp16, y = layers_11_fc1_input_shift_to_fp16)[name = tensor("input_259_cast_fp16")]; tensor var_3347 = const()[name = tensor("op_3347"), val = tensor([1, 1])]; tensor var_3349 = const()[name = tensor("op_3349"), val = tensor([1, 1])]; tensor x_331_pad_type_0 = const()[name = tensor("x_331_pad_type_0"), val = tensor("custom")]; tensor x_331_pad_0 = const()[name = tensor("x_331_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285935296))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289212160))), name = tensor("layers_11_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_11_fc1_module_bias_to_fp16 = const()[name = tensor("layers_11_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289212288)))]; tensor x_331_cast_fp16 = conv(bias = layers_11_fc1_module_bias_to_fp16, dilations = var_3349, groups = var_3117, pad = x_331_pad_0, pad_type = x_331_pad_type_0, strides = var_3347, weight = layers_11_fc1_module_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = tensor("x_331_cast_fp16")]; tensor layers_11_fc1_output_scale_to_fp16 = const()[name = tensor("layers_11_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289222592)))]; tensor input_261_cast_fp16 = mul(x = x_331_cast_fp16, y = layers_11_fc1_output_scale_to_fp16)[name = tensor("input_261_cast_fp16")]; tensor x_333_mode_0 = const()[name = tensor("x_333_mode_0"), val = tensor("EXACT")]; tensor x_333_cast_fp16 = gelu(mode = x_333_mode_0, x = input_261_cast_fp16)[name = tensor("x_333_cast_fp16")]; tensor layers_11_fc2_input_shift_to_fp16 = const()[name = tensor("layers_11_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289232896)))]; tensor input_263_cast_fp16 = sub(x = x_333_cast_fp16, y = layers_11_fc2_input_shift_to_fp16)[name = tensor("input_263_cast_fp16")]; tensor var_3360 = const()[name = tensor("op_3360"), val = tensor([1, 1])]; tensor var_3362 = const()[name = tensor("op_3362"), val = tensor([1, 1])]; tensor x_335_pad_type_0 = const()[name = tensor("x_335_pad_type_0"), val = tensor("custom")]; tensor x_335_pad_0 = const()[name = tensor("x_335_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_11_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289243200))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292520064))), name = tensor("layers_11_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_11_fc2_module_bias_to_fp16 = const()[name = tensor("layers_11_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292520192)))]; tensor x_335_cast_fp16 = conv(bias = layers_11_fc2_module_bias_to_fp16, dilations = var_3362, groups = var_3117, pad = x_335_pad_0, pad_type = x_335_pad_type_0, strides = var_3360, weight = layers_11_fc2_module_weight_to_fp16_palettized, x = input_263_cast_fp16)[name = tensor("x_335_cast_fp16")]; tensor layers_11_fc2_output_scale_to_fp16 = const()[name = tensor("layers_11_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292522816)))]; tensor hidden_states_25_cast_fp16 = mul(x = x_335_cast_fp16, y = layers_11_fc2_output_scale_to_fp16)[name = tensor("hidden_states_25_cast_fp16")]; tensor inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor("inputs_73_cast_fp16")]; tensor var_3376 = const()[name = tensor("op_3376"), val = tensor(3)]; tensor var_3383 = const()[name = tensor("op_3383"), val = tensor(1)]; tensor var_3384 = const()[name = tensor("op_3384"), val = tensor(true)]; tensor var_3396 = const()[name = tensor("op_3396"), val = tensor([1])]; tensor channels_mean_73_cast_fp16 = reduce_mean(axes = var_3396, keep_dims = var_3384, x = inputs_73_cast_fp16)[name = tensor("channels_mean_73_cast_fp16")]; tensor zero_mean_73_cast_fp16 = sub(x = inputs_73_cast_fp16, y = channels_mean_73_cast_fp16)[name = tensor("zero_mean_73_cast_fp16")]; tensor zero_mean_sq_73_cast_fp16 = mul(x = zero_mean_73_cast_fp16, y = zero_mean_73_cast_fp16)[name = tensor("zero_mean_sq_73_cast_fp16")]; tensor var_3400 = const()[name = tensor("op_3400"), val = tensor([1])]; tensor var_3401_cast_fp16 = reduce_mean(axes = var_3400, keep_dims = var_3384, x = zero_mean_sq_73_cast_fp16)[name = tensor("op_3401_cast_fp16")]; tensor var_3402_to_fp16 = const()[name = tensor("op_3402_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_3403_cast_fp16 = add(x = var_3401_cast_fp16, y = var_3402_to_fp16)[name = tensor("op_3403_cast_fp16")]; tensor denom_73_epsilon_0_to_fp16 = const()[name = tensor("denom_73_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_73_cast_fp16 = rsqrt(epsilon = denom_73_epsilon_0_to_fp16, x = var_3403_cast_fp16)[name = tensor("denom_73_cast_fp16")]; tensor out_73_cast_fp16 = mul(x = zero_mean_73_cast_fp16, y = denom_73_cast_fp16)[name = tensor("out_73_cast_fp16")]; tensor obj_169_gamma_0_to_fp16 = const()[name = tensor("obj_169_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292525440)))]; tensor obj_169_beta_0_to_fp16 = const()[name = tensor("obj_169_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292528064)))]; tensor obj_169_epsilon_0_to_fp16 = const()[name = tensor("obj_169_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_169_cast_fp16 = batch_norm(beta = obj_169_beta_0_to_fp16, epsilon = obj_169_epsilon_0_to_fp16, gamma = obj_169_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_73_cast_fp16)[name = tensor("obj_169_cast_fp16")]; tensor layers_12_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292530688)))]; tensor input_265_cast_fp16 = sub(x = obj_169_cast_fp16, y = layers_12_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_265_cast_fp16")]; tensor var_3422 = const()[name = tensor("op_3422"), val = tensor([1, 1])]; tensor var_3424 = const()[name = tensor("op_3424"), val = tensor([1, 1])]; tensor x_337_pad_type_0 = const()[name = tensor("x_337_pad_type_0"), val = tensor("custom")]; tensor x_337_pad_0 = const()[name = tensor("x_337_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_12_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292533312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293352576))), name = tensor("layers_12_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_12_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293352704)))]; tensor x_337_cast_fp16 = conv(bias = layers_12_self_attn_q_proj_module_bias_to_fp16, dilations = var_3424, groups = var_3383, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = var_3422, weight = layers_12_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_265_cast_fp16)[name = tensor("x_337_cast_fp16")]; tensor layers_12_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293355328)))]; tensor query_49_cast_fp16 = mul(x = x_337_cast_fp16, y = layers_12_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_49_cast_fp16")]; tensor var_3434 = const()[name = tensor("op_3434"), val = tensor([1, 1])]; tensor var_3436 = const()[name = tensor("op_3436"), val = tensor([1, 1])]; tensor x_339_pad_type_0 = const()[name = tensor("x_339_pad_type_0"), val = tensor("custom")]; tensor x_339_pad_0 = const()[name = tensor("x_339_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_12_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293357952))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294177216))), name = tensor("layers_12_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_12_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294177344)))]; tensor x_339_cast_fp16 = conv(bias = layers_12_self_attn_k_proj_module_bias_to_fp16, dilations = var_3436, groups = var_3383, pad = x_339_pad_0, pad_type = x_339_pad_type_0, strides = var_3434, weight = layers_12_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_265_cast_fp16)[name = tensor("x_339_cast_fp16")]; tensor layers_12_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_12_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294179968)))]; tensor current_key_25_cast_fp16 = mul(x = x_339_cast_fp16, y = layers_12_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_25_cast_fp16")]; tensor var_3446 = const()[name = tensor("op_3446"), val = tensor([1, 1])]; tensor var_3448 = const()[name = tensor("op_3448"), val = tensor([1, 1])]; tensor x_341_pad_type_0 = const()[name = tensor("x_341_pad_type_0"), val = tensor("custom")]; tensor x_341_pad_0 = const()[name = tensor("x_341_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_12_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294182592))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295001856))), name = tensor("layers_12_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_12_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295001984)))]; tensor x_341_cast_fp16 = conv(bias = layers_12_self_attn_v_proj_module_bias_to_fp16, dilations = var_3448, groups = var_3383, pad = x_341_pad_0, pad_type = x_341_pad_type_0, strides = var_3446, weight = layers_12_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_265_cast_fp16)[name = tensor("x_341_cast_fp16")]; tensor layers_12_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295004608)))]; tensor current_value_25_cast_fp16 = mul(x = x_341_cast_fp16, y = layers_12_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_25_cast_fp16")]; tensor var_3456_cast_fp16 = mul(x = current_key_25_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_3456_cast_fp16")]; tensor var_3458_cast_fp16 = mul(x = var_103_cast_fp16_12, y = var_257_cast_fp16)[name = tensor("op_3458_cast_fp16")]; tensor key_49_cast_fp16 = add(x = var_3456_cast_fp16, y = var_3458_cast_fp16)[name = tensor("key_49_cast_fp16")]; tensor var_3460_cast_fp16 = mul(x = current_value_25_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_3460_cast_fp16")]; tensor var_3462_cast_fp16 = mul(x = var_138_cast_fp16_12, y = var_257_cast_fp16)[name = tensor("op_3462_cast_fp16")]; tensor value_49_cast_fp16 = add(x = var_3460_cast_fp16, y = var_3462_cast_fp16)[name = tensor("value_49_cast_fp16")]; tensor var_3465 = const()[name = tensor("op_3465"), val = tensor([1, 20, 64, -1])]; tensor var_3466_cast_fp16 = reshape(shape = var_3465, x = query_49_cast_fp16)[name = tensor("op_3466_cast_fp16")]; tensor var_3467_to_fp16 = const()[name = tensor("op_3467_to_fp16"), val = tensor(0x1p-3)]; tensor var_3468_cast_fp16 = mul(x = var_3466_cast_fp16, y = var_3467_to_fp16)[name = tensor("op_3468_cast_fp16")]; tensor var_3469 = const()[name = tensor("op_3469"), val = tensor([1, 20, 64, -1])]; tensor var_3470_cast_fp16 = reshape(shape = var_3469, x = key_49_cast_fp16)[name = tensor("op_3470_cast_fp16")]; tensor mh_w_73_transpose_x_0 = const()[name = tensor("mh_w_73_transpose_x_0"), val = tensor(true)]; tensor mh_w_73_transpose_y_0 = const()[name = tensor("mh_w_73_transpose_y_0"), val = tensor(false)]; tensor mh_w_73_cast_fp16 = matmul(transpose_x = mh_w_73_transpose_x_0, transpose_y = mh_w_73_transpose_y_0, x = var_3468_cast_fp16, y = var_3470_cast_fp16)[name = tensor("mh_w_73_cast_fp16")]; tensor mh_w_75_cast_fp16 = add(x = mh_w_73_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_75_cast_fp16")]; tensor var_3478_cast_fp16 = softmax(axis = var_3376, x = mh_w_75_cast_fp16)[name = tensor("op_3478_cast_fp16")]; tensor var_3479 = const()[name = tensor("op_3479"), val = tensor([1, 20, 64, -1])]; tensor var_3480_cast_fp16 = reshape(shape = var_3479, x = value_49_cast_fp16)[name = tensor("op_3480_cast_fp16")]; tensor attn_49_transpose_x_0 = const()[name = tensor("attn_49_transpose_x_0"), val = tensor(false)]; tensor attn_49_transpose_y_0 = const()[name = tensor("attn_49_transpose_y_0"), val = tensor(true)]; tensor attn_49_cast_fp16 = matmul(transpose_x = attn_49_transpose_x_0, transpose_y = attn_49_transpose_y_0, x = var_3480_cast_fp16, y = var_3478_cast_fp16)[name = tensor("attn_49_cast_fp16")]; tensor var_3483 = const()[name = tensor("op_3483"), val = tensor([1, 1280, 1, -1])]; tensor x_343_cast_fp16 = reshape(shape = var_3483, x = attn_49_cast_fp16)[name = tensor("x_343_cast_fp16")]; tensor layers_12_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295007232)))]; tensor input_271_cast_fp16 = sub(x = x_343_cast_fp16, y = layers_12_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_271_cast_fp16")]; tensor var_3491 = const()[name = tensor("op_3491"), val = tensor([1, 1])]; tensor var_3493 = const()[name = tensor("op_3493"), val = tensor([1, 1])]; tensor x_345_pad_type_0 = const()[name = tensor("x_345_pad_type_0"), val = tensor("custom")]; tensor x_345_pad_0 = const()[name = tensor("x_345_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_12_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295009856))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295829120))), name = tensor("layers_12_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_12_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295829248)))]; tensor x_345_cast_fp16 = conv(bias = layers_12_self_attn_o_proj_module_bias_to_fp16, dilations = var_3493, groups = var_3383, pad = x_345_pad_0, pad_type = x_345_pad_type_0, strides = var_3491, weight = layers_12_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_271_cast_fp16)[name = tensor("x_345_cast_fp16")]; tensor layers_12_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295831872)))]; tensor obj_175_cast_fp16 = mul(x = x_345_cast_fp16, y = layers_12_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_175_cast_fp16")]; tensor inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = obj_175_cast_fp16)[name = tensor("inputs_75_cast_fp16")]; tensor var_3504 = const()[name = tensor("op_3504"), val = tensor([1])]; tensor channels_mean_75_cast_fp16 = reduce_mean(axes = var_3504, keep_dims = var_3384, x = inputs_75_cast_fp16)[name = tensor("channels_mean_75_cast_fp16")]; tensor zero_mean_75_cast_fp16 = sub(x = inputs_75_cast_fp16, y = channels_mean_75_cast_fp16)[name = tensor("zero_mean_75_cast_fp16")]; tensor zero_mean_sq_75_cast_fp16 = mul(x = zero_mean_75_cast_fp16, y = zero_mean_75_cast_fp16)[name = tensor("zero_mean_sq_75_cast_fp16")]; tensor var_3508 = const()[name = tensor("op_3508"), val = tensor([1])]; tensor var_3509_cast_fp16 = reduce_mean(axes = var_3508, keep_dims = var_3384, x = zero_mean_sq_75_cast_fp16)[name = tensor("op_3509_cast_fp16")]; tensor var_3510_to_fp16 = const()[name = tensor("op_3510_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_3511_cast_fp16 = add(x = var_3509_cast_fp16, y = var_3510_to_fp16)[name = tensor("op_3511_cast_fp16")]; tensor denom_75_epsilon_0_to_fp16 = const()[name = tensor("denom_75_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_75_cast_fp16 = rsqrt(epsilon = denom_75_epsilon_0_to_fp16, x = var_3511_cast_fp16)[name = tensor("denom_75_cast_fp16")]; tensor out_75_cast_fp16 = mul(x = zero_mean_75_cast_fp16, y = denom_75_cast_fp16)[name = tensor("out_75_cast_fp16")]; tensor obj_177_gamma_0_to_fp16 = const()[name = tensor("obj_177_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295834496)))]; tensor obj_177_beta_0_to_fp16 = const()[name = tensor("obj_177_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295837120)))]; tensor obj_177_epsilon_0_to_fp16 = const()[name = tensor("obj_177_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_177_cast_fp16 = batch_norm(beta = obj_177_beta_0_to_fp16, epsilon = obj_177_epsilon_0_to_fp16, gamma = obj_177_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_75_cast_fp16)[name = tensor("obj_177_cast_fp16")]; tensor layers_12_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_12_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295839744)))]; tensor input_273_cast_fp16 = sub(x = obj_177_cast_fp16, y = layers_12_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_273_cast_fp16")]; tensor var_3530 = const()[name = tensor("op_3530"), val = tensor([1, 1])]; tensor var_3532 = const()[name = tensor("op_3532"), val = tensor([1, 1])]; tensor x_347_pad_type_0 = const()[name = tensor("x_347_pad_type_0"), val = tensor("custom")]; tensor x_347_pad_0 = const()[name = tensor("x_347_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_12_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295842368))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296661632))), name = tensor("layers_12_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_12_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296661760)))]; tensor x_347_cast_fp16 = conv(bias = layers_12_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_3532, groups = var_3383, pad = x_347_pad_0, pad_type = x_347_pad_type_0, strides = var_3530, weight = layers_12_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_273_cast_fp16)[name = tensor("x_347_cast_fp16")]; tensor layers_12_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_12_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296664384)))]; tensor query_51_cast_fp16 = mul(x = x_347_cast_fp16, y = layers_12_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_51_cast_fp16")]; tensor var_3542 = const()[name = tensor("op_3542"), val = tensor([1, 1])]; tensor var_3544 = const()[name = tensor("op_3544"), val = tensor([1, 1])]; tensor x_349_pad_type_0 = const()[name = tensor("x_349_pad_type_0"), val = tensor("custom")]; tensor x_349_pad_0 = const()[name = tensor("x_349_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_12_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296667008))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297486272))), name = tensor("layers_12_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_12_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297486400)))]; tensor x_349_cast_fp16 = conv(bias = layers_12_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_3544, groups = var_3383, pad = x_349_pad_0, pad_type = x_349_pad_type_0, strides = var_3542, weight = layers_12_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_349_cast_fp16")]; tensor layers_12_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_12_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297489024)))]; tensor key_51_cast_fp16 = mul(x = x_349_cast_fp16, y = layers_12_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_51_cast_fp16")]; tensor var_3554 = const()[name = tensor("op_3554"), val = tensor([1, 1])]; tensor var_3556 = const()[name = tensor("op_3556"), val = tensor([1, 1])]; tensor x_351_pad_type_0 = const()[name = tensor("x_351_pad_type_0"), val = tensor("custom")]; tensor x_351_pad_0 = const()[name = tensor("x_351_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_12_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297491648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298310912))), name = tensor("layers_12_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_12_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298311040)))]; tensor x_351_cast_fp16 = conv(bias = layers_12_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_3556, groups = var_3383, pad = x_351_pad_0, pad_type = x_351_pad_type_0, strides = var_3554, weight = layers_12_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_351_cast_fp16")]; tensor layers_12_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_12_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298313664)))]; tensor value_51_cast_fp16 = mul(x = x_351_cast_fp16, y = layers_12_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_51_cast_fp16")]; tensor var_3561 = const()[name = tensor("op_3561"), val = tensor([1, 20, 64, -1])]; tensor var_3562_cast_fp16 = reshape(shape = var_3561, x = query_51_cast_fp16)[name = tensor("op_3562_cast_fp16")]; tensor var_3563_to_fp16 = const()[name = tensor("op_3563_to_fp16"), val = tensor(0x1p-3)]; tensor var_3564_cast_fp16 = mul(x = var_3562_cast_fp16, y = var_3563_to_fp16)[name = tensor("op_3564_cast_fp16")]; tensor var_3565 = const()[name = tensor("op_3565"), val = tensor([1, 20, 64, -1])]; tensor var_3566_cast_fp16 = reshape(shape = var_3565, x = key_51_cast_fp16)[name = tensor("op_3566_cast_fp16")]; tensor mh_w_77_transpose_x_0 = const()[name = tensor("mh_w_77_transpose_x_0"), val = tensor(true)]; tensor mh_w_77_transpose_y_0 = const()[name = tensor("mh_w_77_transpose_y_0"), val = tensor(false)]; tensor mh_w_77_cast_fp16 = matmul(transpose_x = mh_w_77_transpose_x_0, transpose_y = mh_w_77_transpose_y_0, x = var_3564_cast_fp16, y = var_3566_cast_fp16)[name = tensor("mh_w_77_cast_fp16")]; tensor obj_181_cast_fp16 = softmax(axis = var_3376, x = mh_w_77_cast_fp16)[name = tensor("obj_181_cast_fp16")]; tensor var_3570 = const()[name = tensor("op_3570"), val = tensor([1, 20, 64, -1])]; tensor var_3571_cast_fp16 = reshape(shape = var_3570, x = value_51_cast_fp16)[name = tensor("op_3571_cast_fp16")]; tensor attn_51_transpose_x_0 = const()[name = tensor("attn_51_transpose_x_0"), val = tensor(false)]; tensor attn_51_transpose_y_0 = const()[name = tensor("attn_51_transpose_y_0"), val = tensor(true)]; tensor attn_51_cast_fp16 = matmul(transpose_x = attn_51_transpose_x_0, transpose_y = attn_51_transpose_y_0, x = var_3571_cast_fp16, y = obj_181_cast_fp16)[name = tensor("attn_51_cast_fp16")]; tensor var_3574 = const()[name = tensor("op_3574"), val = tensor([1, 1280, 1, -1])]; tensor x_353_cast_fp16 = reshape(shape = var_3574, x = attn_51_cast_fp16)[name = tensor("x_353_cast_fp16")]; tensor layers_12_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_12_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298316288)))]; tensor input_279_cast_fp16 = sub(x = x_353_cast_fp16, y = layers_12_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_279_cast_fp16")]; tensor var_3582 = const()[name = tensor("op_3582"), val = tensor([1, 1])]; tensor var_3584 = const()[name = tensor("op_3584"), val = tensor([1, 1])]; tensor x_355_pad_type_0 = const()[name = tensor("x_355_pad_type_0"), val = tensor("custom")]; tensor x_355_pad_0 = const()[name = tensor("x_355_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_12_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298318912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299138176))), name = tensor("layers_12_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_12_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299138304)))]; tensor x_355_cast_fp16 = conv(bias = layers_12_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_3584, groups = var_3383, pad = x_355_pad_0, pad_type = x_355_pad_type_0, strides = var_3582, weight = layers_12_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_279_cast_fp16)[name = tensor("x_355_cast_fp16")]; tensor layers_12_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_12_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299140928)))]; tensor obj_179_cast_fp16 = mul(x = x_355_cast_fp16, y = layers_12_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_179_cast_fp16")]; tensor inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = obj_179_cast_fp16)[name = tensor("inputs_77_cast_fp16")]; tensor var_3594 = const()[name = tensor("op_3594"), val = tensor([1])]; tensor channels_mean_77_cast_fp16 = reduce_mean(axes = var_3594, keep_dims = var_3384, x = inputs_77_cast_fp16)[name = tensor("channels_mean_77_cast_fp16")]; tensor zero_mean_77_cast_fp16 = sub(x = inputs_77_cast_fp16, y = channels_mean_77_cast_fp16)[name = tensor("zero_mean_77_cast_fp16")]; tensor zero_mean_sq_77_cast_fp16 = mul(x = zero_mean_77_cast_fp16, y = zero_mean_77_cast_fp16)[name = tensor("zero_mean_sq_77_cast_fp16")]; tensor var_3598 = const()[name = tensor("op_3598"), val = tensor([1])]; tensor var_3599_cast_fp16 = reduce_mean(axes = var_3598, keep_dims = var_3384, x = zero_mean_sq_77_cast_fp16)[name = tensor("op_3599_cast_fp16")]; tensor var_3600_to_fp16 = const()[name = tensor("op_3600_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_3601_cast_fp16 = add(x = var_3599_cast_fp16, y = var_3600_to_fp16)[name = tensor("op_3601_cast_fp16")]; tensor denom_77_epsilon_0_to_fp16 = const()[name = tensor("denom_77_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_77_cast_fp16 = rsqrt(epsilon = denom_77_epsilon_0_to_fp16, x = var_3601_cast_fp16)[name = tensor("denom_77_cast_fp16")]; tensor out_77_cast_fp16 = mul(x = zero_mean_77_cast_fp16, y = denom_77_cast_fp16)[name = tensor("out_77_cast_fp16")]; tensor x_357_gamma_0_to_fp16 = const()[name = tensor("x_357_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299143552)))]; tensor x_357_beta_0_to_fp16 = const()[name = tensor("x_357_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299146176)))]; tensor x_357_epsilon_0_to_fp16 = const()[name = tensor("x_357_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_357_cast_fp16 = batch_norm(beta = x_357_beta_0_to_fp16, epsilon = x_357_epsilon_0_to_fp16, gamma = x_357_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_77_cast_fp16)[name = tensor("x_357_cast_fp16")]; tensor layers_12_fc1_input_shift_to_fp16 = const()[name = tensor("layers_12_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299148800)))]; tensor input_281_cast_fp16 = sub(x = x_357_cast_fp16, y = layers_12_fc1_input_shift_to_fp16)[name = tensor("input_281_cast_fp16")]; tensor var_3616 = const()[name = tensor("op_3616"), val = tensor([1, 1])]; tensor var_3618 = const()[name = tensor("op_3618"), val = tensor([1, 1])]; tensor x_359_pad_type_0 = const()[name = tensor("x_359_pad_type_0"), val = tensor("custom")]; tensor x_359_pad_0 = const()[name = tensor("x_359_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_12_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299151424))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302428288))), name = tensor("layers_12_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_12_fc1_module_bias_to_fp16 = const()[name = tensor("layers_12_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302428416)))]; tensor x_359_cast_fp16 = conv(bias = layers_12_fc1_module_bias_to_fp16, dilations = var_3618, groups = var_3383, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = var_3616, weight = layers_12_fc1_module_weight_to_fp16_palettized, x = input_281_cast_fp16)[name = tensor("x_359_cast_fp16")]; tensor layers_12_fc1_output_scale_to_fp16 = const()[name = tensor("layers_12_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302438720)))]; tensor input_283_cast_fp16 = mul(x = x_359_cast_fp16, y = layers_12_fc1_output_scale_to_fp16)[name = tensor("input_283_cast_fp16")]; tensor x_361_mode_0 = const()[name = tensor("x_361_mode_0"), val = tensor("EXACT")]; tensor x_361_cast_fp16 = gelu(mode = x_361_mode_0, x = input_283_cast_fp16)[name = tensor("x_361_cast_fp16")]; tensor layers_12_fc2_input_shift_to_fp16 = const()[name = tensor("layers_12_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302449024)))]; tensor input_285_cast_fp16 = sub(x = x_361_cast_fp16, y = layers_12_fc2_input_shift_to_fp16)[name = tensor("input_285_cast_fp16")]; tensor var_3629 = const()[name = tensor("op_3629"), val = tensor([1, 1])]; tensor var_3631 = const()[name = tensor("op_3631"), val = tensor([1, 1])]; tensor x_363_pad_type_0 = const()[name = tensor("x_363_pad_type_0"), val = tensor("custom")]; tensor x_363_pad_0 = const()[name = tensor("x_363_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_12_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302459328))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(305736192))), name = tensor("layers_12_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_12_fc2_module_bias_to_fp16 = const()[name = tensor("layers_12_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(305736320)))]; tensor x_363_cast_fp16 = conv(bias = layers_12_fc2_module_bias_to_fp16, dilations = var_3631, groups = var_3383, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = var_3629, weight = layers_12_fc2_module_weight_to_fp16_palettized, x = input_285_cast_fp16)[name = tensor("x_363_cast_fp16")]; tensor layers_12_fc2_output_scale_to_fp16 = const()[name = tensor("layers_12_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(305738944)))]; tensor hidden_states_27_cast_fp16 = mul(x = x_363_cast_fp16, y = layers_12_fc2_output_scale_to_fp16)[name = tensor("hidden_states_27_cast_fp16")]; tensor inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = hidden_states_27_cast_fp16)[name = tensor("inputs_79_cast_fp16")]; tensor var_3646 = const()[name = tensor("op_3646"), val = tensor(3)]; tensor var_3653 = const()[name = tensor("op_3653"), val = tensor(1)]; tensor var_3654 = const()[name = tensor("op_3654"), val = tensor(true)]; tensor var_3666 = const()[name = tensor("op_3666"), val = tensor([1])]; tensor channels_mean_79_cast_fp16 = reduce_mean(axes = var_3666, keep_dims = var_3654, x = inputs_79_cast_fp16)[name = tensor("channels_mean_79_cast_fp16")]; tensor zero_mean_79_cast_fp16 = sub(x = inputs_79_cast_fp16, y = channels_mean_79_cast_fp16)[name = tensor("zero_mean_79_cast_fp16")]; tensor zero_mean_sq_79_cast_fp16 = mul(x = zero_mean_79_cast_fp16, y = zero_mean_79_cast_fp16)[name = tensor("zero_mean_sq_79_cast_fp16")]; tensor var_3670 = const()[name = tensor("op_3670"), val = tensor([1])]; tensor var_3671_cast_fp16 = reduce_mean(axes = var_3670, keep_dims = var_3654, x = zero_mean_sq_79_cast_fp16)[name = tensor("op_3671_cast_fp16")]; tensor var_3672_to_fp16 = const()[name = tensor("op_3672_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_3673_cast_fp16 = add(x = var_3671_cast_fp16, y = var_3672_to_fp16)[name = tensor("op_3673_cast_fp16")]; tensor denom_79_epsilon_0_to_fp16 = const()[name = tensor("denom_79_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_79_cast_fp16 = rsqrt(epsilon = denom_79_epsilon_0_to_fp16, x = var_3673_cast_fp16)[name = tensor("denom_79_cast_fp16")]; tensor out_79_cast_fp16 = mul(x = zero_mean_79_cast_fp16, y = denom_79_cast_fp16)[name = tensor("out_79_cast_fp16")]; tensor obj_183_gamma_0_to_fp16 = const()[name = tensor("obj_183_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(305741568)))]; tensor obj_183_beta_0_to_fp16 = const()[name = tensor("obj_183_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(305744192)))]; tensor obj_183_epsilon_0_to_fp16 = const()[name = tensor("obj_183_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_183_cast_fp16 = batch_norm(beta = obj_183_beta_0_to_fp16, epsilon = obj_183_epsilon_0_to_fp16, gamma = obj_183_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_79_cast_fp16)[name = tensor("obj_183_cast_fp16")]; tensor layers_13_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(305746816)))]; tensor input_287_cast_fp16 = sub(x = obj_183_cast_fp16, y = layers_13_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_287_cast_fp16")]; tensor var_3692 = const()[name = tensor("op_3692"), val = tensor([1, 1])]; tensor var_3694 = const()[name = tensor("op_3694"), val = tensor([1, 1])]; tensor x_365_pad_type_0 = const()[name = tensor("x_365_pad_type_0"), val = tensor("custom")]; tensor x_365_pad_0 = const()[name = tensor("x_365_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_13_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(305749440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306568704))), name = tensor("layers_13_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_13_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306568832)))]; tensor x_365_cast_fp16 = conv(bias = layers_13_self_attn_q_proj_module_bias_to_fp16, dilations = var_3694, groups = var_3653, pad = x_365_pad_0, pad_type = x_365_pad_type_0, strides = var_3692, weight = layers_13_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = tensor("x_365_cast_fp16")]; tensor layers_13_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306571456)))]; tensor query_53_cast_fp16 = mul(x = x_365_cast_fp16, y = layers_13_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_53_cast_fp16")]; tensor var_3704 = const()[name = tensor("op_3704"), val = tensor([1, 1])]; tensor var_3706 = const()[name = tensor("op_3706"), val = tensor([1, 1])]; tensor x_367_pad_type_0 = const()[name = tensor("x_367_pad_type_0"), val = tensor("custom")]; tensor x_367_pad_0 = const()[name = tensor("x_367_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_13_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306574080))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307393344))), name = tensor("layers_13_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_13_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307393472)))]; tensor x_367_cast_fp16 = conv(bias = layers_13_self_attn_k_proj_module_bias_to_fp16, dilations = var_3706, groups = var_3653, pad = x_367_pad_0, pad_type = x_367_pad_type_0, strides = var_3704, weight = layers_13_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = tensor("x_367_cast_fp16")]; tensor layers_13_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_13_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307396096)))]; tensor current_key_27_cast_fp16 = mul(x = x_367_cast_fp16, y = layers_13_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_27_cast_fp16")]; tensor var_3716 = const()[name = tensor("op_3716"), val = tensor([1, 1])]; tensor var_3718 = const()[name = tensor("op_3718"), val = tensor([1, 1])]; tensor x_369_pad_type_0 = const()[name = tensor("x_369_pad_type_0"), val = tensor("custom")]; tensor x_369_pad_0 = const()[name = tensor("x_369_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_13_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307398720))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308217984))), name = tensor("layers_13_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_13_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308218112)))]; tensor x_369_cast_fp16 = conv(bias = layers_13_self_attn_v_proj_module_bias_to_fp16, dilations = var_3718, groups = var_3653, pad = x_369_pad_0, pad_type = x_369_pad_type_0, strides = var_3716, weight = layers_13_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = tensor("x_369_cast_fp16")]; tensor layers_13_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308220736)))]; tensor current_value_27_cast_fp16 = mul(x = x_369_cast_fp16, y = layers_13_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_27_cast_fp16")]; tensor var_3726_cast_fp16 = mul(x = current_key_27_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_3726_cast_fp16")]; tensor var_3728_cast_fp16 = mul(x = var_103_cast_fp16_13, y = var_257_cast_fp16)[name = tensor("op_3728_cast_fp16")]; tensor key_53_cast_fp16 = add(x = var_3726_cast_fp16, y = var_3728_cast_fp16)[name = tensor("key_53_cast_fp16")]; tensor var_3730_cast_fp16 = mul(x = current_value_27_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_3730_cast_fp16")]; tensor var_3732_cast_fp16 = mul(x = var_138_cast_fp16_13, y = var_257_cast_fp16)[name = tensor("op_3732_cast_fp16")]; tensor value_53_cast_fp16 = add(x = var_3730_cast_fp16, y = var_3732_cast_fp16)[name = tensor("value_53_cast_fp16")]; tensor var_3735 = const()[name = tensor("op_3735"), val = tensor([1, 20, 64, -1])]; tensor var_3736_cast_fp16 = reshape(shape = var_3735, x = query_53_cast_fp16)[name = tensor("op_3736_cast_fp16")]; tensor var_3737_to_fp16 = const()[name = tensor("op_3737_to_fp16"), val = tensor(0x1p-3)]; tensor var_3738_cast_fp16 = mul(x = var_3736_cast_fp16, y = var_3737_to_fp16)[name = tensor("op_3738_cast_fp16")]; tensor var_3739 = const()[name = tensor("op_3739"), val = tensor([1, 20, 64, -1])]; tensor var_3740_cast_fp16 = reshape(shape = var_3739, x = key_53_cast_fp16)[name = tensor("op_3740_cast_fp16")]; tensor mh_w_79_transpose_x_0 = const()[name = tensor("mh_w_79_transpose_x_0"), val = tensor(true)]; tensor mh_w_79_transpose_y_0 = const()[name = tensor("mh_w_79_transpose_y_0"), val = tensor(false)]; tensor mh_w_79_cast_fp16 = matmul(transpose_x = mh_w_79_transpose_x_0, transpose_y = mh_w_79_transpose_y_0, x = var_3738_cast_fp16, y = var_3740_cast_fp16)[name = tensor("mh_w_79_cast_fp16")]; tensor mh_w_81_cast_fp16 = add(x = mh_w_79_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_81_cast_fp16")]; tensor var_3748_cast_fp16 = softmax(axis = var_3646, x = mh_w_81_cast_fp16)[name = tensor("op_3748_cast_fp16")]; tensor var_3749 = const()[name = tensor("op_3749"), val = tensor([1, 20, 64, -1])]; tensor var_3750_cast_fp16 = reshape(shape = var_3749, x = value_53_cast_fp16)[name = tensor("op_3750_cast_fp16")]; tensor attn_53_transpose_x_0 = const()[name = tensor("attn_53_transpose_x_0"), val = tensor(false)]; tensor attn_53_transpose_y_0 = const()[name = tensor("attn_53_transpose_y_0"), val = tensor(true)]; tensor attn_53_cast_fp16 = matmul(transpose_x = attn_53_transpose_x_0, transpose_y = attn_53_transpose_y_0, x = var_3750_cast_fp16, y = var_3748_cast_fp16)[name = tensor("attn_53_cast_fp16")]; tensor var_3753 = const()[name = tensor("op_3753"), val = tensor([1, 1280, 1, -1])]; tensor x_371_cast_fp16 = reshape(shape = var_3753, x = attn_53_cast_fp16)[name = tensor("x_371_cast_fp16")]; tensor layers_13_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308223360)))]; tensor input_293_cast_fp16 = sub(x = x_371_cast_fp16, y = layers_13_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_293_cast_fp16")]; tensor var_3761 = const()[name = tensor("op_3761"), val = tensor([1, 1])]; tensor var_3763 = const()[name = tensor("op_3763"), val = tensor([1, 1])]; tensor x_373_pad_type_0 = const()[name = tensor("x_373_pad_type_0"), val = tensor("custom")]; tensor x_373_pad_0 = const()[name = tensor("x_373_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_13_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308225984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309045248))), name = tensor("layers_13_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_13_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309045376)))]; tensor x_373_cast_fp16 = conv(bias = layers_13_self_attn_o_proj_module_bias_to_fp16, dilations = var_3763, groups = var_3653, pad = x_373_pad_0, pad_type = x_373_pad_type_0, strides = var_3761, weight = layers_13_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_293_cast_fp16)[name = tensor("x_373_cast_fp16")]; tensor layers_13_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309048000)))]; tensor obj_189_cast_fp16 = mul(x = x_373_cast_fp16, y = layers_13_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_189_cast_fp16")]; tensor inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = obj_189_cast_fp16)[name = tensor("inputs_81_cast_fp16")]; tensor var_3774 = const()[name = tensor("op_3774"), val = tensor([1])]; tensor channels_mean_81_cast_fp16 = reduce_mean(axes = var_3774, keep_dims = var_3654, x = inputs_81_cast_fp16)[name = tensor("channels_mean_81_cast_fp16")]; tensor zero_mean_81_cast_fp16 = sub(x = inputs_81_cast_fp16, y = channels_mean_81_cast_fp16)[name = tensor("zero_mean_81_cast_fp16")]; tensor zero_mean_sq_81_cast_fp16 = mul(x = zero_mean_81_cast_fp16, y = zero_mean_81_cast_fp16)[name = tensor("zero_mean_sq_81_cast_fp16")]; tensor var_3778 = const()[name = tensor("op_3778"), val = tensor([1])]; tensor var_3779_cast_fp16 = reduce_mean(axes = var_3778, keep_dims = var_3654, x = zero_mean_sq_81_cast_fp16)[name = tensor("op_3779_cast_fp16")]; tensor var_3780_to_fp16 = const()[name = tensor("op_3780_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_3781_cast_fp16 = add(x = var_3779_cast_fp16, y = var_3780_to_fp16)[name = tensor("op_3781_cast_fp16")]; tensor denom_81_epsilon_0_to_fp16 = const()[name = tensor("denom_81_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_81_cast_fp16 = rsqrt(epsilon = denom_81_epsilon_0_to_fp16, x = var_3781_cast_fp16)[name = tensor("denom_81_cast_fp16")]; tensor out_81_cast_fp16 = mul(x = zero_mean_81_cast_fp16, y = denom_81_cast_fp16)[name = tensor("out_81_cast_fp16")]; tensor obj_191_gamma_0_to_fp16 = const()[name = tensor("obj_191_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309050624)))]; tensor obj_191_beta_0_to_fp16 = const()[name = tensor("obj_191_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309053248)))]; tensor obj_191_epsilon_0_to_fp16 = const()[name = tensor("obj_191_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_191_cast_fp16 = batch_norm(beta = obj_191_beta_0_to_fp16, epsilon = obj_191_epsilon_0_to_fp16, gamma = obj_191_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_81_cast_fp16)[name = tensor("obj_191_cast_fp16")]; tensor layers_13_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_13_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309055872)))]; tensor input_295_cast_fp16 = sub(x = obj_191_cast_fp16, y = layers_13_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_295_cast_fp16")]; tensor var_3800 = const()[name = tensor("op_3800"), val = tensor([1, 1])]; tensor var_3802 = const()[name = tensor("op_3802"), val = tensor([1, 1])]; tensor x_375_pad_type_0 = const()[name = tensor("x_375_pad_type_0"), val = tensor("custom")]; tensor x_375_pad_0 = const()[name = tensor("x_375_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_13_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309058496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309877760))), name = tensor("layers_13_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_13_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309877888)))]; tensor x_375_cast_fp16 = conv(bias = layers_13_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_3802, groups = var_3653, pad = x_375_pad_0, pad_type = x_375_pad_type_0, strides = var_3800, weight = layers_13_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_295_cast_fp16)[name = tensor("x_375_cast_fp16")]; tensor layers_13_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_13_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309880512)))]; tensor query_55_cast_fp16 = mul(x = x_375_cast_fp16, y = layers_13_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_55_cast_fp16")]; tensor var_3812 = const()[name = tensor("op_3812"), val = tensor([1, 1])]; tensor var_3814 = const()[name = tensor("op_3814"), val = tensor([1, 1])]; tensor x_377_pad_type_0 = const()[name = tensor("x_377_pad_type_0"), val = tensor("custom")]; tensor x_377_pad_0 = const()[name = tensor("x_377_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_13_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309883136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310702400))), name = tensor("layers_13_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_13_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310702528)))]; tensor x_377_cast_fp16 = conv(bias = layers_13_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_3814, groups = var_3653, pad = x_377_pad_0, pad_type = x_377_pad_type_0, strides = var_3812, weight = layers_13_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_377_cast_fp16")]; tensor layers_13_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_13_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310705152)))]; tensor key_55_cast_fp16 = mul(x = x_377_cast_fp16, y = layers_13_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_55_cast_fp16")]; tensor var_3824 = const()[name = tensor("op_3824"), val = tensor([1, 1])]; tensor var_3826 = const()[name = tensor("op_3826"), val = tensor([1, 1])]; tensor x_379_pad_type_0 = const()[name = tensor("x_379_pad_type_0"), val = tensor("custom")]; tensor x_379_pad_0 = const()[name = tensor("x_379_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_13_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310707776))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311527040))), name = tensor("layers_13_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_13_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311527168)))]; tensor x_379_cast_fp16 = conv(bias = layers_13_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_3826, groups = var_3653, pad = x_379_pad_0, pad_type = x_379_pad_type_0, strides = var_3824, weight = layers_13_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_379_cast_fp16")]; tensor layers_13_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_13_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311529792)))]; tensor value_55_cast_fp16 = mul(x = x_379_cast_fp16, y = layers_13_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_55_cast_fp16")]; tensor var_3831 = const()[name = tensor("op_3831"), val = tensor([1, 20, 64, -1])]; tensor var_3832_cast_fp16 = reshape(shape = var_3831, x = query_55_cast_fp16)[name = tensor("op_3832_cast_fp16")]; tensor var_3833_to_fp16 = const()[name = tensor("op_3833_to_fp16"), val = tensor(0x1p-3)]; tensor var_3834_cast_fp16 = mul(x = var_3832_cast_fp16, y = var_3833_to_fp16)[name = tensor("op_3834_cast_fp16")]; tensor var_3835 = const()[name = tensor("op_3835"), val = tensor([1, 20, 64, -1])]; tensor var_3836_cast_fp16 = reshape(shape = var_3835, x = key_55_cast_fp16)[name = tensor("op_3836_cast_fp16")]; tensor mh_w_83_transpose_x_0 = const()[name = tensor("mh_w_83_transpose_x_0"), val = tensor(true)]; tensor mh_w_83_transpose_y_0 = const()[name = tensor("mh_w_83_transpose_y_0"), val = tensor(false)]; tensor mh_w_83_cast_fp16 = matmul(transpose_x = mh_w_83_transpose_x_0, transpose_y = mh_w_83_transpose_y_0, x = var_3834_cast_fp16, y = var_3836_cast_fp16)[name = tensor("mh_w_83_cast_fp16")]; tensor obj_195_cast_fp16 = softmax(axis = var_3646, x = mh_w_83_cast_fp16)[name = tensor("obj_195_cast_fp16")]; tensor var_3840 = const()[name = tensor("op_3840"), val = tensor([1, 20, 64, -1])]; tensor var_3841_cast_fp16 = reshape(shape = var_3840, x = value_55_cast_fp16)[name = tensor("op_3841_cast_fp16")]; tensor attn_55_transpose_x_0 = const()[name = tensor("attn_55_transpose_x_0"), val = tensor(false)]; tensor attn_55_transpose_y_0 = const()[name = tensor("attn_55_transpose_y_0"), val = tensor(true)]; tensor attn_55_cast_fp16 = matmul(transpose_x = attn_55_transpose_x_0, transpose_y = attn_55_transpose_y_0, x = var_3841_cast_fp16, y = obj_195_cast_fp16)[name = tensor("attn_55_cast_fp16")]; tensor var_3844 = const()[name = tensor("op_3844"), val = tensor([1, 1280, 1, -1])]; tensor x_381_cast_fp16 = reshape(shape = var_3844, x = attn_55_cast_fp16)[name = tensor("x_381_cast_fp16")]; tensor layers_13_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_13_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311532416)))]; tensor input_301_cast_fp16 = sub(x = x_381_cast_fp16, y = layers_13_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_301_cast_fp16")]; tensor var_3852 = const()[name = tensor("op_3852"), val = tensor([1, 1])]; tensor var_3854 = const()[name = tensor("op_3854"), val = tensor([1, 1])]; tensor x_383_pad_type_0 = const()[name = tensor("x_383_pad_type_0"), val = tensor("custom")]; tensor x_383_pad_0 = const()[name = tensor("x_383_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_13_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311535040))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312354304))), name = tensor("layers_13_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_13_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312354432)))]; tensor x_383_cast_fp16 = conv(bias = layers_13_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_3854, groups = var_3653, pad = x_383_pad_0, pad_type = x_383_pad_type_0, strides = var_3852, weight = layers_13_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_301_cast_fp16)[name = tensor("x_383_cast_fp16")]; tensor layers_13_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_13_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312357056)))]; tensor obj_193_cast_fp16 = mul(x = x_383_cast_fp16, y = layers_13_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_193_cast_fp16")]; tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = obj_193_cast_fp16)[name = tensor("inputs_83_cast_fp16")]; tensor var_3864 = const()[name = tensor("op_3864"), val = tensor([1])]; tensor channels_mean_83_cast_fp16 = reduce_mean(axes = var_3864, keep_dims = var_3654, x = inputs_83_cast_fp16)[name = tensor("channels_mean_83_cast_fp16")]; tensor zero_mean_83_cast_fp16 = sub(x = inputs_83_cast_fp16, y = channels_mean_83_cast_fp16)[name = tensor("zero_mean_83_cast_fp16")]; tensor zero_mean_sq_83_cast_fp16 = mul(x = zero_mean_83_cast_fp16, y = zero_mean_83_cast_fp16)[name = tensor("zero_mean_sq_83_cast_fp16")]; tensor var_3868 = const()[name = tensor("op_3868"), val = tensor([1])]; tensor var_3869_cast_fp16 = reduce_mean(axes = var_3868, keep_dims = var_3654, x = zero_mean_sq_83_cast_fp16)[name = tensor("op_3869_cast_fp16")]; tensor var_3870_to_fp16 = const()[name = tensor("op_3870_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_3871_cast_fp16 = add(x = var_3869_cast_fp16, y = var_3870_to_fp16)[name = tensor("op_3871_cast_fp16")]; tensor denom_83_epsilon_0_to_fp16 = const()[name = tensor("denom_83_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_83_cast_fp16 = rsqrt(epsilon = denom_83_epsilon_0_to_fp16, x = var_3871_cast_fp16)[name = tensor("denom_83_cast_fp16")]; tensor out_83_cast_fp16 = mul(x = zero_mean_83_cast_fp16, y = denom_83_cast_fp16)[name = tensor("out_83_cast_fp16")]; tensor x_385_gamma_0_to_fp16 = const()[name = tensor("x_385_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312359680)))]; tensor x_385_beta_0_to_fp16 = const()[name = tensor("x_385_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312362304)))]; tensor x_385_epsilon_0_to_fp16 = const()[name = tensor("x_385_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_385_cast_fp16 = batch_norm(beta = x_385_beta_0_to_fp16, epsilon = x_385_epsilon_0_to_fp16, gamma = x_385_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_83_cast_fp16)[name = tensor("x_385_cast_fp16")]; tensor layers_13_fc1_input_shift_to_fp16 = const()[name = tensor("layers_13_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312364928)))]; tensor input_303_cast_fp16 = sub(x = x_385_cast_fp16, y = layers_13_fc1_input_shift_to_fp16)[name = tensor("input_303_cast_fp16")]; tensor var_3886 = const()[name = tensor("op_3886"), val = tensor([1, 1])]; tensor var_3888 = const()[name = tensor("op_3888"), val = tensor([1, 1])]; tensor x_387_pad_type_0 = const()[name = tensor("x_387_pad_type_0"), val = tensor("custom")]; tensor x_387_pad_0 = const()[name = tensor("x_387_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_13_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312367552))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(315644416))), name = tensor("layers_13_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_13_fc1_module_bias_to_fp16 = const()[name = tensor("layers_13_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(315644544)))]; tensor x_387_cast_fp16 = conv(bias = layers_13_fc1_module_bias_to_fp16, dilations = var_3888, groups = var_3653, pad = x_387_pad_0, pad_type = x_387_pad_type_0, strides = var_3886, weight = layers_13_fc1_module_weight_to_fp16_palettized, x = input_303_cast_fp16)[name = tensor("x_387_cast_fp16")]; tensor layers_13_fc1_output_scale_to_fp16 = const()[name = tensor("layers_13_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(315654848)))]; tensor input_305_cast_fp16 = mul(x = x_387_cast_fp16, y = layers_13_fc1_output_scale_to_fp16)[name = tensor("input_305_cast_fp16")]; tensor x_389_mode_0 = const()[name = tensor("x_389_mode_0"), val = tensor("EXACT")]; tensor x_389_cast_fp16 = gelu(mode = x_389_mode_0, x = input_305_cast_fp16)[name = tensor("x_389_cast_fp16")]; tensor layers_13_fc2_input_shift_to_fp16 = const()[name = tensor("layers_13_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(315665152)))]; tensor input_307_cast_fp16 = sub(x = x_389_cast_fp16, y = layers_13_fc2_input_shift_to_fp16)[name = tensor("input_307_cast_fp16")]; tensor var_3899 = const()[name = tensor("op_3899"), val = tensor([1, 1])]; tensor var_3901 = const()[name = tensor("op_3901"), val = tensor([1, 1])]; tensor x_391_pad_type_0 = const()[name = tensor("x_391_pad_type_0"), val = tensor("custom")]; tensor x_391_pad_0 = const()[name = tensor("x_391_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_13_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(315675456))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318952320))), name = tensor("layers_13_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_13_fc2_module_bias_to_fp16 = const()[name = tensor("layers_13_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318952448)))]; tensor x_391_cast_fp16 = conv(bias = layers_13_fc2_module_bias_to_fp16, dilations = var_3901, groups = var_3653, pad = x_391_pad_0, pad_type = x_391_pad_type_0, strides = var_3899, weight = layers_13_fc2_module_weight_to_fp16_palettized, x = input_307_cast_fp16)[name = tensor("x_391_cast_fp16")]; tensor layers_13_fc2_output_scale_to_fp16 = const()[name = tensor("layers_13_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318955072)))]; tensor hidden_states_29_cast_fp16 = mul(x = x_391_cast_fp16, y = layers_13_fc2_output_scale_to_fp16)[name = tensor("hidden_states_29_cast_fp16")]; tensor inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = hidden_states_29_cast_fp16)[name = tensor("inputs_85_cast_fp16")]; tensor var_3916 = const()[name = tensor("op_3916"), val = tensor(3)]; tensor var_3923 = const()[name = tensor("op_3923"), val = tensor(1)]; tensor var_3924 = const()[name = tensor("op_3924"), val = tensor(true)]; tensor var_3936 = const()[name = tensor("op_3936"), val = tensor([1])]; tensor channels_mean_85_cast_fp16 = reduce_mean(axes = var_3936, keep_dims = var_3924, x = inputs_85_cast_fp16)[name = tensor("channels_mean_85_cast_fp16")]; tensor zero_mean_85_cast_fp16 = sub(x = inputs_85_cast_fp16, y = channels_mean_85_cast_fp16)[name = tensor("zero_mean_85_cast_fp16")]; tensor zero_mean_sq_85_cast_fp16 = mul(x = zero_mean_85_cast_fp16, y = zero_mean_85_cast_fp16)[name = tensor("zero_mean_sq_85_cast_fp16")]; tensor var_3940 = const()[name = tensor("op_3940"), val = tensor([1])]; tensor var_3941_cast_fp16 = reduce_mean(axes = var_3940, keep_dims = var_3924, x = zero_mean_sq_85_cast_fp16)[name = tensor("op_3941_cast_fp16")]; tensor var_3942_to_fp16 = const()[name = tensor("op_3942_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_3943_cast_fp16 = add(x = var_3941_cast_fp16, y = var_3942_to_fp16)[name = tensor("op_3943_cast_fp16")]; tensor denom_85_epsilon_0_to_fp16 = const()[name = tensor("denom_85_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_85_cast_fp16 = rsqrt(epsilon = denom_85_epsilon_0_to_fp16, x = var_3943_cast_fp16)[name = tensor("denom_85_cast_fp16")]; tensor out_85_cast_fp16 = mul(x = zero_mean_85_cast_fp16, y = denom_85_cast_fp16)[name = tensor("out_85_cast_fp16")]; tensor obj_197_gamma_0_to_fp16 = const()[name = tensor("obj_197_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318957696)))]; tensor obj_197_beta_0_to_fp16 = const()[name = tensor("obj_197_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318960320)))]; tensor obj_197_epsilon_0_to_fp16 = const()[name = tensor("obj_197_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_197_cast_fp16 = batch_norm(beta = obj_197_beta_0_to_fp16, epsilon = obj_197_epsilon_0_to_fp16, gamma = obj_197_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_85_cast_fp16)[name = tensor("obj_197_cast_fp16")]; tensor layers_14_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318962944)))]; tensor input_309_cast_fp16 = sub(x = obj_197_cast_fp16, y = layers_14_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_309_cast_fp16")]; tensor var_3962 = const()[name = tensor("op_3962"), val = tensor([1, 1])]; tensor var_3964 = const()[name = tensor("op_3964"), val = tensor([1, 1])]; tensor x_393_pad_type_0 = const()[name = tensor("x_393_pad_type_0"), val = tensor("custom")]; tensor x_393_pad_0 = const()[name = tensor("x_393_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_14_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318965568))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319784832))), name = tensor("layers_14_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_14_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319784960)))]; tensor x_393_cast_fp16 = conv(bias = layers_14_self_attn_q_proj_module_bias_to_fp16, dilations = var_3964, groups = var_3923, pad = x_393_pad_0, pad_type = x_393_pad_type_0, strides = var_3962, weight = layers_14_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_309_cast_fp16)[name = tensor("x_393_cast_fp16")]; tensor layers_14_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319787584)))]; tensor query_57_cast_fp16 = mul(x = x_393_cast_fp16, y = layers_14_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_57_cast_fp16")]; tensor var_3974 = const()[name = tensor("op_3974"), val = tensor([1, 1])]; tensor var_3976 = const()[name = tensor("op_3976"), val = tensor([1, 1])]; tensor x_395_pad_type_0 = const()[name = tensor("x_395_pad_type_0"), val = tensor("custom")]; tensor x_395_pad_0 = const()[name = tensor("x_395_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_14_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319790208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320609472))), name = tensor("layers_14_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_14_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320609600)))]; tensor x_395_cast_fp16 = conv(bias = layers_14_self_attn_k_proj_module_bias_to_fp16, dilations = var_3976, groups = var_3923, pad = x_395_pad_0, pad_type = x_395_pad_type_0, strides = var_3974, weight = layers_14_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_309_cast_fp16)[name = tensor("x_395_cast_fp16")]; tensor layers_14_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_14_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320612224)))]; tensor current_key_29_cast_fp16 = mul(x = x_395_cast_fp16, y = layers_14_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_29_cast_fp16")]; tensor var_3986 = const()[name = tensor("op_3986"), val = tensor([1, 1])]; tensor var_3988 = const()[name = tensor("op_3988"), val = tensor([1, 1])]; tensor x_397_pad_type_0 = const()[name = tensor("x_397_pad_type_0"), val = tensor("custom")]; tensor x_397_pad_0 = const()[name = tensor("x_397_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_14_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320614848))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321434112))), name = tensor("layers_14_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_14_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321434240)))]; tensor x_397_cast_fp16 = conv(bias = layers_14_self_attn_v_proj_module_bias_to_fp16, dilations = var_3988, groups = var_3923, pad = x_397_pad_0, pad_type = x_397_pad_type_0, strides = var_3986, weight = layers_14_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_309_cast_fp16)[name = tensor("x_397_cast_fp16")]; tensor layers_14_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321436864)))]; tensor current_value_29_cast_fp16 = mul(x = x_397_cast_fp16, y = layers_14_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_29_cast_fp16")]; tensor var_3996_cast_fp16 = mul(x = current_key_29_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_3996_cast_fp16")]; tensor var_3998_cast_fp16 = mul(x = var_103_cast_fp16_14, y = var_257_cast_fp16)[name = tensor("op_3998_cast_fp16")]; tensor key_57_cast_fp16 = add(x = var_3996_cast_fp16, y = var_3998_cast_fp16)[name = tensor("key_57_cast_fp16")]; tensor var_4000_cast_fp16 = mul(x = current_value_29_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_4000_cast_fp16")]; tensor var_4002_cast_fp16 = mul(x = var_138_cast_fp16_14, y = var_257_cast_fp16)[name = tensor("op_4002_cast_fp16")]; tensor value_57_cast_fp16 = add(x = var_4000_cast_fp16, y = var_4002_cast_fp16)[name = tensor("value_57_cast_fp16")]; tensor var_4005 = const()[name = tensor("op_4005"), val = tensor([1, 20, 64, -1])]; tensor var_4006_cast_fp16 = reshape(shape = var_4005, x = query_57_cast_fp16)[name = tensor("op_4006_cast_fp16")]; tensor var_4007_to_fp16 = const()[name = tensor("op_4007_to_fp16"), val = tensor(0x1p-3)]; tensor var_4008_cast_fp16 = mul(x = var_4006_cast_fp16, y = var_4007_to_fp16)[name = tensor("op_4008_cast_fp16")]; tensor var_4009 = const()[name = tensor("op_4009"), val = tensor([1, 20, 64, -1])]; tensor var_4010_cast_fp16 = reshape(shape = var_4009, x = key_57_cast_fp16)[name = tensor("op_4010_cast_fp16")]; tensor mh_w_85_transpose_x_0 = const()[name = tensor("mh_w_85_transpose_x_0"), val = tensor(true)]; tensor mh_w_85_transpose_y_0 = const()[name = tensor("mh_w_85_transpose_y_0"), val = tensor(false)]; tensor mh_w_85_cast_fp16 = matmul(transpose_x = mh_w_85_transpose_x_0, transpose_y = mh_w_85_transpose_y_0, x = var_4008_cast_fp16, y = var_4010_cast_fp16)[name = tensor("mh_w_85_cast_fp16")]; tensor mh_w_87_cast_fp16 = add(x = mh_w_85_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_87_cast_fp16")]; tensor var_4018_cast_fp16 = softmax(axis = var_3916, x = mh_w_87_cast_fp16)[name = tensor("op_4018_cast_fp16")]; tensor var_4019 = const()[name = tensor("op_4019"), val = tensor([1, 20, 64, -1])]; tensor var_4020_cast_fp16 = reshape(shape = var_4019, x = value_57_cast_fp16)[name = tensor("op_4020_cast_fp16")]; tensor attn_57_transpose_x_0 = const()[name = tensor("attn_57_transpose_x_0"), val = tensor(false)]; tensor attn_57_transpose_y_0 = const()[name = tensor("attn_57_transpose_y_0"), val = tensor(true)]; tensor attn_57_cast_fp16 = matmul(transpose_x = attn_57_transpose_x_0, transpose_y = attn_57_transpose_y_0, x = var_4020_cast_fp16, y = var_4018_cast_fp16)[name = tensor("attn_57_cast_fp16")]; tensor var_4023 = const()[name = tensor("op_4023"), val = tensor([1, 1280, 1, -1])]; tensor x_399_cast_fp16 = reshape(shape = var_4023, x = attn_57_cast_fp16)[name = tensor("x_399_cast_fp16")]; tensor layers_14_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321439488)))]; tensor input_315_cast_fp16 = sub(x = x_399_cast_fp16, y = layers_14_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_315_cast_fp16")]; tensor var_4031 = const()[name = tensor("op_4031"), val = tensor([1, 1])]; tensor var_4033 = const()[name = tensor("op_4033"), val = tensor([1, 1])]; tensor x_401_pad_type_0 = const()[name = tensor("x_401_pad_type_0"), val = tensor("custom")]; tensor x_401_pad_0 = const()[name = tensor("x_401_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_14_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321442112))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322261376))), name = tensor("layers_14_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_14_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322261504)))]; tensor x_401_cast_fp16 = conv(bias = layers_14_self_attn_o_proj_module_bias_to_fp16, dilations = var_4033, groups = var_3923, pad = x_401_pad_0, pad_type = x_401_pad_type_0, strides = var_4031, weight = layers_14_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_315_cast_fp16)[name = tensor("x_401_cast_fp16")]; tensor layers_14_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322264128)))]; tensor obj_203_cast_fp16 = mul(x = x_401_cast_fp16, y = layers_14_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_203_cast_fp16")]; tensor inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = obj_203_cast_fp16)[name = tensor("inputs_87_cast_fp16")]; tensor var_4044 = const()[name = tensor("op_4044"), val = tensor([1])]; tensor channels_mean_87_cast_fp16 = reduce_mean(axes = var_4044, keep_dims = var_3924, x = inputs_87_cast_fp16)[name = tensor("channels_mean_87_cast_fp16")]; tensor zero_mean_87_cast_fp16 = sub(x = inputs_87_cast_fp16, y = channels_mean_87_cast_fp16)[name = tensor("zero_mean_87_cast_fp16")]; tensor zero_mean_sq_87_cast_fp16 = mul(x = zero_mean_87_cast_fp16, y = zero_mean_87_cast_fp16)[name = tensor("zero_mean_sq_87_cast_fp16")]; tensor var_4048 = const()[name = tensor("op_4048"), val = tensor([1])]; tensor var_4049_cast_fp16 = reduce_mean(axes = var_4048, keep_dims = var_3924, x = zero_mean_sq_87_cast_fp16)[name = tensor("op_4049_cast_fp16")]; tensor var_4050_to_fp16 = const()[name = tensor("op_4050_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_4051_cast_fp16 = add(x = var_4049_cast_fp16, y = var_4050_to_fp16)[name = tensor("op_4051_cast_fp16")]; tensor denom_87_epsilon_0_to_fp16 = const()[name = tensor("denom_87_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_87_cast_fp16 = rsqrt(epsilon = denom_87_epsilon_0_to_fp16, x = var_4051_cast_fp16)[name = tensor("denom_87_cast_fp16")]; tensor out_87_cast_fp16 = mul(x = zero_mean_87_cast_fp16, y = denom_87_cast_fp16)[name = tensor("out_87_cast_fp16")]; tensor obj_205_gamma_0_to_fp16 = const()[name = tensor("obj_205_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322266752)))]; tensor obj_205_beta_0_to_fp16 = const()[name = tensor("obj_205_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322269376)))]; tensor obj_205_epsilon_0_to_fp16 = const()[name = tensor("obj_205_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_205_cast_fp16 = batch_norm(beta = obj_205_beta_0_to_fp16, epsilon = obj_205_epsilon_0_to_fp16, gamma = obj_205_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_87_cast_fp16)[name = tensor("obj_205_cast_fp16")]; tensor layers_14_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_14_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322272000)))]; tensor input_317_cast_fp16 = sub(x = obj_205_cast_fp16, y = layers_14_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_317_cast_fp16")]; tensor var_4070 = const()[name = tensor("op_4070"), val = tensor([1, 1])]; tensor var_4072 = const()[name = tensor("op_4072"), val = tensor([1, 1])]; tensor x_403_pad_type_0 = const()[name = tensor("x_403_pad_type_0"), val = tensor("custom")]; tensor x_403_pad_0 = const()[name = tensor("x_403_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_14_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322274624))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323093888))), name = tensor("layers_14_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_14_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323094016)))]; tensor x_403_cast_fp16 = conv(bias = layers_14_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_4072, groups = var_3923, pad = x_403_pad_0, pad_type = x_403_pad_type_0, strides = var_4070, weight = layers_14_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_317_cast_fp16)[name = tensor("x_403_cast_fp16")]; tensor layers_14_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_14_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323096640)))]; tensor query_59_cast_fp16 = mul(x = x_403_cast_fp16, y = layers_14_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_59_cast_fp16")]; tensor var_4082 = const()[name = tensor("op_4082"), val = tensor([1, 1])]; tensor var_4084 = const()[name = tensor("op_4084"), val = tensor([1, 1])]; tensor x_405_pad_type_0 = const()[name = tensor("x_405_pad_type_0"), val = tensor("custom")]; tensor x_405_pad_0 = const()[name = tensor("x_405_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_14_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323099264))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323918528))), name = tensor("layers_14_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_14_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323918656)))]; tensor x_405_cast_fp16 = conv(bias = layers_14_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_4084, groups = var_3923, pad = x_405_pad_0, pad_type = x_405_pad_type_0, strides = var_4082, weight = layers_14_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_405_cast_fp16")]; tensor layers_14_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_14_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323921280)))]; tensor key_59_cast_fp16 = mul(x = x_405_cast_fp16, y = layers_14_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_59_cast_fp16")]; tensor var_4094 = const()[name = tensor("op_4094"), val = tensor([1, 1])]; tensor var_4096 = const()[name = tensor("op_4096"), val = tensor([1, 1])]; tensor x_407_pad_type_0 = const()[name = tensor("x_407_pad_type_0"), val = tensor("custom")]; tensor x_407_pad_0 = const()[name = tensor("x_407_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_14_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323923904))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324743168))), name = tensor("layers_14_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_14_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324743296)))]; tensor x_407_cast_fp16 = conv(bias = layers_14_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_4096, groups = var_3923, pad = x_407_pad_0, pad_type = x_407_pad_type_0, strides = var_4094, weight = layers_14_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_407_cast_fp16")]; tensor layers_14_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_14_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324745920)))]; tensor value_59_cast_fp16 = mul(x = x_407_cast_fp16, y = layers_14_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_59_cast_fp16")]; tensor var_4101 = const()[name = tensor("op_4101"), val = tensor([1, 20, 64, -1])]; tensor var_4102_cast_fp16 = reshape(shape = var_4101, x = query_59_cast_fp16)[name = tensor("op_4102_cast_fp16")]; tensor var_4103_to_fp16 = const()[name = tensor("op_4103_to_fp16"), val = tensor(0x1p-3)]; tensor var_4104_cast_fp16 = mul(x = var_4102_cast_fp16, y = var_4103_to_fp16)[name = tensor("op_4104_cast_fp16")]; tensor var_4105 = const()[name = tensor("op_4105"), val = tensor([1, 20, 64, -1])]; tensor var_4106_cast_fp16 = reshape(shape = var_4105, x = key_59_cast_fp16)[name = tensor("op_4106_cast_fp16")]; tensor mh_w_89_transpose_x_0 = const()[name = tensor("mh_w_89_transpose_x_0"), val = tensor(true)]; tensor mh_w_89_transpose_y_0 = const()[name = tensor("mh_w_89_transpose_y_0"), val = tensor(false)]; tensor mh_w_89_cast_fp16 = matmul(transpose_x = mh_w_89_transpose_x_0, transpose_y = mh_w_89_transpose_y_0, x = var_4104_cast_fp16, y = var_4106_cast_fp16)[name = tensor("mh_w_89_cast_fp16")]; tensor obj_209_cast_fp16 = softmax(axis = var_3916, x = mh_w_89_cast_fp16)[name = tensor("obj_209_cast_fp16")]; tensor var_4110 = const()[name = tensor("op_4110"), val = tensor([1, 20, 64, -1])]; tensor var_4111_cast_fp16 = reshape(shape = var_4110, x = value_59_cast_fp16)[name = tensor("op_4111_cast_fp16")]; tensor attn_59_transpose_x_0 = const()[name = tensor("attn_59_transpose_x_0"), val = tensor(false)]; tensor attn_59_transpose_y_0 = const()[name = tensor("attn_59_transpose_y_0"), val = tensor(true)]; tensor attn_59_cast_fp16 = matmul(transpose_x = attn_59_transpose_x_0, transpose_y = attn_59_transpose_y_0, x = var_4111_cast_fp16, y = obj_209_cast_fp16)[name = tensor("attn_59_cast_fp16")]; tensor var_4114 = const()[name = tensor("op_4114"), val = tensor([1, 1280, 1, -1])]; tensor x_409_cast_fp16 = reshape(shape = var_4114, x = attn_59_cast_fp16)[name = tensor("x_409_cast_fp16")]; tensor layers_14_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_14_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324748544)))]; tensor input_323_cast_fp16 = sub(x = x_409_cast_fp16, y = layers_14_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_323_cast_fp16")]; tensor var_4122 = const()[name = tensor("op_4122"), val = tensor([1, 1])]; tensor var_4124 = const()[name = tensor("op_4124"), val = tensor([1, 1])]; tensor x_411_pad_type_0 = const()[name = tensor("x_411_pad_type_0"), val = tensor("custom")]; tensor x_411_pad_0 = const()[name = tensor("x_411_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_14_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324751168))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325570432))), name = tensor("layers_14_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_14_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325570560)))]; tensor x_411_cast_fp16 = conv(bias = layers_14_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_4124, groups = var_3923, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = var_4122, weight = layers_14_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_323_cast_fp16)[name = tensor("x_411_cast_fp16")]; tensor layers_14_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_14_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325573184)))]; tensor obj_207_cast_fp16 = mul(x = x_411_cast_fp16, y = layers_14_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_207_cast_fp16")]; tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = obj_207_cast_fp16)[name = tensor("inputs_89_cast_fp16")]; tensor var_4131 = const()[name = tensor("op_4131"), val = tensor([1])]; tensor channels_mean_89_cast_fp16 = reduce_mean(axes = var_4131, keep_dims = var_3924, x = inputs_89_cast_fp16)[name = tensor("channels_mean_89_cast_fp16")]; tensor zero_mean_89_cast_fp16 = sub(x = inputs_89_cast_fp16, y = channels_mean_89_cast_fp16)[name = tensor("zero_mean_89_cast_fp16")]; tensor zero_mean_sq_89_cast_fp16 = mul(x = zero_mean_89_cast_fp16, y = zero_mean_89_cast_fp16)[name = tensor("zero_mean_sq_89_cast_fp16")]; tensor var_4135 = const()[name = tensor("op_4135"), val = tensor([1])]; tensor var_4136_cast_fp16 = reduce_mean(axes = var_4135, keep_dims = var_3924, x = zero_mean_sq_89_cast_fp16)[name = tensor("op_4136_cast_fp16")]; tensor var_4137_to_fp16 = const()[name = tensor("op_4137_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_4138_cast_fp16 = add(x = var_4136_cast_fp16, y = var_4137_to_fp16)[name = tensor("op_4138_cast_fp16")]; tensor denom_89_epsilon_0_to_fp16 = const()[name = tensor("denom_89_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_89_cast_fp16 = rsqrt(epsilon = denom_89_epsilon_0_to_fp16, x = var_4138_cast_fp16)[name = tensor("denom_89_cast_fp16")]; tensor out_89_cast_fp16 = mul(x = zero_mean_89_cast_fp16, y = denom_89_cast_fp16)[name = tensor("out_89_cast_fp16")]; tensor x_413_gamma_0_to_fp16 = const()[name = tensor("x_413_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325575808)))]; tensor x_413_beta_0_to_fp16 = const()[name = tensor("x_413_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325578432)))]; tensor x_413_epsilon_0_to_fp16 = const()[name = tensor("x_413_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_413_cast_fp16 = batch_norm(beta = x_413_beta_0_to_fp16, epsilon = x_413_epsilon_0_to_fp16, gamma = x_413_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_89_cast_fp16)[name = tensor("x_413_cast_fp16")]; tensor layers_14_fc1_input_shift_to_fp16 = const()[name = tensor("layers_14_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325581056)))]; tensor input_325_cast_fp16 = sub(x = x_413_cast_fp16, y = layers_14_fc1_input_shift_to_fp16)[name = tensor("input_325_cast_fp16")]; tensor var_4153 = const()[name = tensor("op_4153"), val = tensor([1, 1])]; tensor var_4155 = const()[name = tensor("op_4155"), val = tensor([1, 1])]; tensor x_415_pad_type_0 = const()[name = tensor("x_415_pad_type_0"), val = tensor("custom")]; tensor x_415_pad_0 = const()[name = tensor("x_415_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_14_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325583680))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328860544))), name = tensor("layers_14_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_14_fc1_module_bias_to_fp16 = const()[name = tensor("layers_14_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328860672)))]; tensor x_415_cast_fp16 = conv(bias = layers_14_fc1_module_bias_to_fp16, dilations = var_4155, groups = var_3923, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = var_4153, weight = layers_14_fc1_module_weight_to_fp16_palettized, x = input_325_cast_fp16)[name = tensor("x_415_cast_fp16")]; tensor layers_14_fc1_output_scale_to_fp16 = const()[name = tensor("layers_14_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328870976)))]; tensor input_327_cast_fp16 = mul(x = x_415_cast_fp16, y = layers_14_fc1_output_scale_to_fp16)[name = tensor("input_327_cast_fp16")]; tensor x_417_mode_0 = const()[name = tensor("x_417_mode_0"), val = tensor("EXACT")]; tensor x_417_cast_fp16 = gelu(mode = x_417_mode_0, x = input_327_cast_fp16)[name = tensor("x_417_cast_fp16")]; tensor layers_14_fc2_input_shift_to_fp16 = const()[name = tensor("layers_14_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328881280)))]; tensor input_329_cast_fp16 = sub(x = x_417_cast_fp16, y = layers_14_fc2_input_shift_to_fp16)[name = tensor("input_329_cast_fp16")]; tensor var_4166 = const()[name = tensor("op_4166"), val = tensor([1, 1])]; tensor var_4168 = const()[name = tensor("op_4168"), val = tensor([1, 1])]; tensor x_419_pad_type_0 = const()[name = tensor("x_419_pad_type_0"), val = tensor("custom")]; tensor x_419_pad_0 = const()[name = tensor("x_419_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_14_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328891584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332168448))), name = tensor("layers_14_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_14_fc2_module_bias_to_fp16 = const()[name = tensor("layers_14_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332168576)))]; tensor x_419_cast_fp16 = conv(bias = layers_14_fc2_module_bias_to_fp16, dilations = var_4168, groups = var_3923, pad = x_419_pad_0, pad_type = x_419_pad_type_0, strides = var_4166, weight = layers_14_fc2_module_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = tensor("x_419_cast_fp16")]; tensor layers_14_fc2_output_scale_to_fp16 = const()[name = tensor("layers_14_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332171200)))]; tensor hidden_states_31_cast_fp16 = mul(x = x_419_cast_fp16, y = layers_14_fc2_output_scale_to_fp16)[name = tensor("hidden_states_31_cast_fp16")]; tensor inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = hidden_states_31_cast_fp16)[name = tensor("inputs_91_cast_fp16")]; tensor var_4182 = const()[name = tensor("op_4182"), val = tensor(3)]; tensor var_4189 = const()[name = tensor("op_4189"), val = tensor(1)]; tensor var_4190 = const()[name = tensor("op_4190"), val = tensor(true)]; tensor var_4202 = const()[name = tensor("op_4202"), val = tensor([1])]; tensor channels_mean_91_cast_fp16 = reduce_mean(axes = var_4202, keep_dims = var_4190, x = inputs_91_cast_fp16)[name = tensor("channels_mean_91_cast_fp16")]; tensor zero_mean_91_cast_fp16 = sub(x = inputs_91_cast_fp16, y = channels_mean_91_cast_fp16)[name = tensor("zero_mean_91_cast_fp16")]; tensor zero_mean_sq_91_cast_fp16 = mul(x = zero_mean_91_cast_fp16, y = zero_mean_91_cast_fp16)[name = tensor("zero_mean_sq_91_cast_fp16")]; tensor var_4206 = const()[name = tensor("op_4206"), val = tensor([1])]; tensor var_4207_cast_fp16 = reduce_mean(axes = var_4206, keep_dims = var_4190, x = zero_mean_sq_91_cast_fp16)[name = tensor("op_4207_cast_fp16")]; tensor var_4208_to_fp16 = const()[name = tensor("op_4208_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_4209_cast_fp16 = add(x = var_4207_cast_fp16, y = var_4208_to_fp16)[name = tensor("op_4209_cast_fp16")]; tensor denom_91_epsilon_0_to_fp16 = const()[name = tensor("denom_91_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_91_cast_fp16 = rsqrt(epsilon = denom_91_epsilon_0_to_fp16, x = var_4209_cast_fp16)[name = tensor("denom_91_cast_fp16")]; tensor out_91_cast_fp16 = mul(x = zero_mean_91_cast_fp16, y = denom_91_cast_fp16)[name = tensor("out_91_cast_fp16")]; tensor obj_211_gamma_0_to_fp16 = const()[name = tensor("obj_211_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332173824)))]; tensor obj_211_beta_0_to_fp16 = const()[name = tensor("obj_211_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332176448)))]; tensor obj_211_epsilon_0_to_fp16 = const()[name = tensor("obj_211_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_211_cast_fp16 = batch_norm(beta = obj_211_beta_0_to_fp16, epsilon = obj_211_epsilon_0_to_fp16, gamma = obj_211_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_91_cast_fp16)[name = tensor("obj_211_cast_fp16")]; tensor layers_15_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332179072)))]; tensor input_331_cast_fp16 = sub(x = obj_211_cast_fp16, y = layers_15_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_331_cast_fp16")]; tensor var_4228 = const()[name = tensor("op_4228"), val = tensor([1, 1])]; tensor var_4230 = const()[name = tensor("op_4230"), val = tensor([1, 1])]; tensor x_421_pad_type_0 = const()[name = tensor("x_421_pad_type_0"), val = tensor("custom")]; tensor x_421_pad_0 = const()[name = tensor("x_421_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_15_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332181696))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333000960))), name = tensor("layers_15_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_15_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333001088)))]; tensor x_421_cast_fp16 = conv(bias = layers_15_self_attn_q_proj_module_bias_to_fp16, dilations = var_4230, groups = var_4189, pad = x_421_pad_0, pad_type = x_421_pad_type_0, strides = var_4228, weight = layers_15_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_331_cast_fp16)[name = tensor("x_421_cast_fp16")]; tensor layers_15_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333003712)))]; tensor query_61_cast_fp16 = mul(x = x_421_cast_fp16, y = layers_15_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_61_cast_fp16")]; tensor var_4240 = const()[name = tensor("op_4240"), val = tensor([1, 1])]; tensor var_4242 = const()[name = tensor("op_4242"), val = tensor([1, 1])]; tensor x_423_pad_type_0 = const()[name = tensor("x_423_pad_type_0"), val = tensor("custom")]; tensor x_423_pad_0 = const()[name = tensor("x_423_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_15_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333006336))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333825600))), name = tensor("layers_15_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_15_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333825728)))]; tensor x_423_cast_fp16 = conv(bias = layers_15_self_attn_k_proj_module_bias_to_fp16, dilations = var_4242, groups = var_4189, pad = x_423_pad_0, pad_type = x_423_pad_type_0, strides = var_4240, weight = layers_15_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_331_cast_fp16)[name = tensor("x_423_cast_fp16")]; tensor layers_15_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_15_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333828352)))]; tensor current_key_31_cast_fp16 = mul(x = x_423_cast_fp16, y = layers_15_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_31_cast_fp16")]; tensor var_4252 = const()[name = tensor("op_4252"), val = tensor([1, 1])]; tensor var_4254 = const()[name = tensor("op_4254"), val = tensor([1, 1])]; tensor x_425_pad_type_0 = const()[name = tensor("x_425_pad_type_0"), val = tensor("custom")]; tensor x_425_pad_0 = const()[name = tensor("x_425_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_15_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333830976))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334650240))), name = tensor("layers_15_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_15_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334650368)))]; tensor x_425_cast_fp16 = conv(bias = layers_15_self_attn_v_proj_module_bias_to_fp16, dilations = var_4254, groups = var_4189, pad = x_425_pad_0, pad_type = x_425_pad_type_0, strides = var_4252, weight = layers_15_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_331_cast_fp16)[name = tensor("x_425_cast_fp16")]; tensor layers_15_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334652992)))]; tensor current_value_31_cast_fp16 = mul(x = x_425_cast_fp16, y = layers_15_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_31_cast_fp16")]; tensor var_4262_cast_fp16 = mul(x = current_key_31_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_4262_cast_fp16")]; tensor var_4264_cast_fp16 = mul(x = var_103_cast_fp16_15, y = var_257_cast_fp16)[name = tensor("op_4264_cast_fp16")]; tensor key_61_cast_fp16 = add(x = var_4262_cast_fp16, y = var_4264_cast_fp16)[name = tensor("key_61_cast_fp16")]; tensor var_4266_cast_fp16 = mul(x = current_value_31_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_4266_cast_fp16")]; tensor var_4268_cast_fp16 = mul(x = var_138_cast_fp16_15, y = var_257_cast_fp16)[name = tensor("op_4268_cast_fp16")]; tensor value_61_cast_fp16 = add(x = var_4266_cast_fp16, y = var_4268_cast_fp16)[name = tensor("value_61_cast_fp16")]; tensor var_4271 = const()[name = tensor("op_4271"), val = tensor([1, 20, 64, -1])]; tensor var_4272_cast_fp16 = reshape(shape = var_4271, x = query_61_cast_fp16)[name = tensor("op_4272_cast_fp16")]; tensor var_4273_to_fp16 = const()[name = tensor("op_4273_to_fp16"), val = tensor(0x1p-3)]; tensor var_4274_cast_fp16 = mul(x = var_4272_cast_fp16, y = var_4273_to_fp16)[name = tensor("op_4274_cast_fp16")]; tensor var_4275 = const()[name = tensor("op_4275"), val = tensor([1, 20, 64, -1])]; tensor var_4276_cast_fp16 = reshape(shape = var_4275, x = key_61_cast_fp16)[name = tensor("op_4276_cast_fp16")]; tensor mh_w_91_transpose_x_0 = const()[name = tensor("mh_w_91_transpose_x_0"), val = tensor(true)]; tensor mh_w_91_transpose_y_0 = const()[name = tensor("mh_w_91_transpose_y_0"), val = tensor(false)]; tensor mh_w_91_cast_fp16 = matmul(transpose_x = mh_w_91_transpose_x_0, transpose_y = mh_w_91_transpose_y_0, x = var_4274_cast_fp16, y = var_4276_cast_fp16)[name = tensor("mh_w_91_cast_fp16")]; tensor mh_w_93_cast_fp16 = add(x = mh_w_91_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_93_cast_fp16")]; tensor var_4284_cast_fp16 = softmax(axis = var_4182, x = mh_w_93_cast_fp16)[name = tensor("op_4284_cast_fp16")]; tensor var_4285 = const()[name = tensor("op_4285"), val = tensor([1, 20, 64, -1])]; tensor var_4286_cast_fp16 = reshape(shape = var_4285, x = value_61_cast_fp16)[name = tensor("op_4286_cast_fp16")]; tensor attn_61_transpose_x_0 = const()[name = tensor("attn_61_transpose_x_0"), val = tensor(false)]; tensor attn_61_transpose_y_0 = const()[name = tensor("attn_61_transpose_y_0"), val = tensor(true)]; tensor attn_61_cast_fp16 = matmul(transpose_x = attn_61_transpose_x_0, transpose_y = attn_61_transpose_y_0, x = var_4286_cast_fp16, y = var_4284_cast_fp16)[name = tensor("attn_61_cast_fp16")]; tensor var_4289 = const()[name = tensor("op_4289"), val = tensor([1, 1280, 1, -1])]; tensor x_427_cast_fp16 = reshape(shape = var_4289, x = attn_61_cast_fp16)[name = tensor("x_427_cast_fp16")]; tensor layers_15_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334655616)))]; tensor input_337_cast_fp16 = sub(x = x_427_cast_fp16, y = layers_15_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_337_cast_fp16")]; tensor var_4297 = const()[name = tensor("op_4297"), val = tensor([1, 1])]; tensor var_4299 = const()[name = tensor("op_4299"), val = tensor([1, 1])]; tensor x_429_pad_type_0 = const()[name = tensor("x_429_pad_type_0"), val = tensor("custom")]; tensor x_429_pad_0 = const()[name = tensor("x_429_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_15_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334658240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335477504))), name = tensor("layers_15_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_15_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335477632)))]; tensor x_429_cast_fp16 = conv(bias = layers_15_self_attn_o_proj_module_bias_to_fp16, dilations = var_4299, groups = var_4189, pad = x_429_pad_0, pad_type = x_429_pad_type_0, strides = var_4297, weight = layers_15_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_337_cast_fp16)[name = tensor("x_429_cast_fp16")]; tensor layers_15_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335480256)))]; tensor obj_217_cast_fp16 = mul(x = x_429_cast_fp16, y = layers_15_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_217_cast_fp16")]; tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = obj_217_cast_fp16)[name = tensor("inputs_93_cast_fp16")]; tensor var_4310 = const()[name = tensor("op_4310"), val = tensor([1])]; tensor channels_mean_93_cast_fp16 = reduce_mean(axes = var_4310, keep_dims = var_4190, x = inputs_93_cast_fp16)[name = tensor("channels_mean_93_cast_fp16")]; tensor zero_mean_93_cast_fp16 = sub(x = inputs_93_cast_fp16, y = channels_mean_93_cast_fp16)[name = tensor("zero_mean_93_cast_fp16")]; tensor zero_mean_sq_93_cast_fp16 = mul(x = zero_mean_93_cast_fp16, y = zero_mean_93_cast_fp16)[name = tensor("zero_mean_sq_93_cast_fp16")]; tensor var_4314 = const()[name = tensor("op_4314"), val = tensor([1])]; tensor var_4315_cast_fp16 = reduce_mean(axes = var_4314, keep_dims = var_4190, x = zero_mean_sq_93_cast_fp16)[name = tensor("op_4315_cast_fp16")]; tensor var_4316_to_fp16 = const()[name = tensor("op_4316_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_4317_cast_fp16 = add(x = var_4315_cast_fp16, y = var_4316_to_fp16)[name = tensor("op_4317_cast_fp16")]; tensor denom_93_epsilon_0_to_fp16 = const()[name = tensor("denom_93_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_93_cast_fp16 = rsqrt(epsilon = denom_93_epsilon_0_to_fp16, x = var_4317_cast_fp16)[name = tensor("denom_93_cast_fp16")]; tensor out_93_cast_fp16 = mul(x = zero_mean_93_cast_fp16, y = denom_93_cast_fp16)[name = tensor("out_93_cast_fp16")]; tensor obj_219_gamma_0_to_fp16 = const()[name = tensor("obj_219_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335482880)))]; tensor obj_219_beta_0_to_fp16 = const()[name = tensor("obj_219_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335485504)))]; tensor obj_219_epsilon_0_to_fp16 = const()[name = tensor("obj_219_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_219_cast_fp16 = batch_norm(beta = obj_219_beta_0_to_fp16, epsilon = obj_219_epsilon_0_to_fp16, gamma = obj_219_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_93_cast_fp16)[name = tensor("obj_219_cast_fp16")]; tensor layers_15_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_15_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335488128)))]; tensor input_339_cast_fp16 = sub(x = obj_219_cast_fp16, y = layers_15_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_339_cast_fp16")]; tensor var_4336 = const()[name = tensor("op_4336"), val = tensor([1, 1])]; tensor var_4338 = const()[name = tensor("op_4338"), val = tensor([1, 1])]; tensor x_431_pad_type_0 = const()[name = tensor("x_431_pad_type_0"), val = tensor("custom")]; tensor x_431_pad_0 = const()[name = tensor("x_431_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_15_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335490752))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336310016))), name = tensor("layers_15_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_15_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336310144)))]; tensor x_431_cast_fp16 = conv(bias = layers_15_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_4338, groups = var_4189, pad = x_431_pad_0, pad_type = x_431_pad_type_0, strides = var_4336, weight = layers_15_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_339_cast_fp16)[name = tensor("x_431_cast_fp16")]; tensor layers_15_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_15_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336312768)))]; tensor query_63_cast_fp16 = mul(x = x_431_cast_fp16, y = layers_15_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_63_cast_fp16")]; tensor var_4348 = const()[name = tensor("op_4348"), val = tensor([1, 1])]; tensor var_4350 = const()[name = tensor("op_4350"), val = tensor([1, 1])]; tensor x_433_pad_type_0 = const()[name = tensor("x_433_pad_type_0"), val = tensor("custom")]; tensor x_433_pad_0 = const()[name = tensor("x_433_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_15_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336315392))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337134656))), name = tensor("layers_15_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_15_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337134784)))]; tensor x_433_cast_fp16 = conv(bias = layers_15_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_4350, groups = var_4189, pad = x_433_pad_0, pad_type = x_433_pad_type_0, strides = var_4348, weight = layers_15_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_433_cast_fp16")]; tensor layers_15_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_15_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337137408)))]; tensor key_63_cast_fp16 = mul(x = x_433_cast_fp16, y = layers_15_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_63_cast_fp16")]; tensor var_4360 = const()[name = tensor("op_4360"), val = tensor([1, 1])]; tensor var_4362 = const()[name = tensor("op_4362"), val = tensor([1, 1])]; tensor x_435_pad_type_0 = const()[name = tensor("x_435_pad_type_0"), val = tensor("custom")]; tensor x_435_pad_0 = const()[name = tensor("x_435_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_15_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337140032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337959296))), name = tensor("layers_15_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_15_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337959424)))]; tensor x_435_cast_fp16 = conv(bias = layers_15_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_4362, groups = var_4189, pad = x_435_pad_0, pad_type = x_435_pad_type_0, strides = var_4360, weight = layers_15_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_435_cast_fp16")]; tensor layers_15_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_15_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337962048)))]; tensor value_63_cast_fp16 = mul(x = x_435_cast_fp16, y = layers_15_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_63_cast_fp16")]; tensor var_4367 = const()[name = tensor("op_4367"), val = tensor([1, 20, 64, -1])]; tensor var_4368_cast_fp16 = reshape(shape = var_4367, x = query_63_cast_fp16)[name = tensor("op_4368_cast_fp16")]; tensor var_4369_to_fp16 = const()[name = tensor("op_4369_to_fp16"), val = tensor(0x1p-3)]; tensor var_4370_cast_fp16 = mul(x = var_4368_cast_fp16, y = var_4369_to_fp16)[name = tensor("op_4370_cast_fp16")]; tensor var_4371 = const()[name = tensor("op_4371"), val = tensor([1, 20, 64, -1])]; tensor var_4372_cast_fp16 = reshape(shape = var_4371, x = key_63_cast_fp16)[name = tensor("op_4372_cast_fp16")]; tensor mh_w_95_transpose_x_0 = const()[name = tensor("mh_w_95_transpose_x_0"), val = tensor(true)]; tensor mh_w_95_transpose_y_0 = const()[name = tensor("mh_w_95_transpose_y_0"), val = tensor(false)]; tensor mh_w_95_cast_fp16 = matmul(transpose_x = mh_w_95_transpose_x_0, transpose_y = mh_w_95_transpose_y_0, x = var_4370_cast_fp16, y = var_4372_cast_fp16)[name = tensor("mh_w_95_cast_fp16")]; tensor obj_223_cast_fp16 = softmax(axis = var_4182, x = mh_w_95_cast_fp16)[name = tensor("obj_223_cast_fp16")]; tensor var_4376 = const()[name = tensor("op_4376"), val = tensor([1, 20, 64, -1])]; tensor var_4377_cast_fp16 = reshape(shape = var_4376, x = value_63_cast_fp16)[name = tensor("op_4377_cast_fp16")]; tensor attn_63_transpose_x_0 = const()[name = tensor("attn_63_transpose_x_0"), val = tensor(false)]; tensor attn_63_transpose_y_0 = const()[name = tensor("attn_63_transpose_y_0"), val = tensor(true)]; tensor attn_63_cast_fp16 = matmul(transpose_x = attn_63_transpose_x_0, transpose_y = attn_63_transpose_y_0, x = var_4377_cast_fp16, y = obj_223_cast_fp16)[name = tensor("attn_63_cast_fp16")]; tensor var_4380 = const()[name = tensor("op_4380"), val = tensor([1, 1280, 1, -1])]; tensor x_437_cast_fp16 = reshape(shape = var_4380, x = attn_63_cast_fp16)[name = tensor("x_437_cast_fp16")]; tensor layers_15_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_15_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337964672)))]; tensor input_345_cast_fp16 = sub(x = x_437_cast_fp16, y = layers_15_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_345_cast_fp16")]; tensor var_4388 = const()[name = tensor("op_4388"), val = tensor([1, 1])]; tensor var_4390 = const()[name = tensor("op_4390"), val = tensor([1, 1])]; tensor x_439_pad_type_0 = const()[name = tensor("x_439_pad_type_0"), val = tensor("custom")]; tensor x_439_pad_0 = const()[name = tensor("x_439_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_15_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337967296))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338786560))), name = tensor("layers_15_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_15_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338786688)))]; tensor x_439_cast_fp16 = conv(bias = layers_15_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_4390, groups = var_4189, pad = x_439_pad_0, pad_type = x_439_pad_type_0, strides = var_4388, weight = layers_15_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_345_cast_fp16)[name = tensor("x_439_cast_fp16")]; tensor layers_15_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_15_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338789312)))]; tensor obj_221_cast_fp16 = mul(x = x_439_cast_fp16, y = layers_15_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_221_cast_fp16")]; tensor inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = obj_221_cast_fp16)[name = tensor("inputs_95_cast_fp16")]; tensor var_4397 = const()[name = tensor("op_4397"), val = tensor([1])]; tensor channels_mean_95_cast_fp16 = reduce_mean(axes = var_4397, keep_dims = var_4190, x = inputs_95_cast_fp16)[name = tensor("channels_mean_95_cast_fp16")]; tensor zero_mean_95_cast_fp16 = sub(x = inputs_95_cast_fp16, y = channels_mean_95_cast_fp16)[name = tensor("zero_mean_95_cast_fp16")]; tensor zero_mean_sq_95_cast_fp16 = mul(x = zero_mean_95_cast_fp16, y = zero_mean_95_cast_fp16)[name = tensor("zero_mean_sq_95_cast_fp16")]; tensor var_4401 = const()[name = tensor("op_4401"), val = tensor([1])]; tensor var_4402_cast_fp16 = reduce_mean(axes = var_4401, keep_dims = var_4190, x = zero_mean_sq_95_cast_fp16)[name = tensor("op_4402_cast_fp16")]; tensor var_4403_to_fp16 = const()[name = tensor("op_4403_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_4404_cast_fp16 = add(x = var_4402_cast_fp16, y = var_4403_to_fp16)[name = tensor("op_4404_cast_fp16")]; tensor denom_95_epsilon_0_to_fp16 = const()[name = tensor("denom_95_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_95_cast_fp16 = rsqrt(epsilon = denom_95_epsilon_0_to_fp16, x = var_4404_cast_fp16)[name = tensor("denom_95_cast_fp16")]; tensor out_95_cast_fp16 = mul(x = zero_mean_95_cast_fp16, y = denom_95_cast_fp16)[name = tensor("out_95_cast_fp16")]; tensor x_441_gamma_0_to_fp16 = const()[name = tensor("x_441_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338791936)))]; tensor x_441_beta_0_to_fp16 = const()[name = tensor("x_441_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338794560)))]; tensor x_441_epsilon_0_to_fp16 = const()[name = tensor("x_441_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_441_cast_fp16 = batch_norm(beta = x_441_beta_0_to_fp16, epsilon = x_441_epsilon_0_to_fp16, gamma = x_441_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_95_cast_fp16)[name = tensor("x_441_cast_fp16")]; tensor layers_15_fc1_input_shift_to_fp16 = const()[name = tensor("layers_15_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338797184)))]; tensor input_347_cast_fp16 = sub(x = x_441_cast_fp16, y = layers_15_fc1_input_shift_to_fp16)[name = tensor("input_347_cast_fp16")]; tensor var_4419 = const()[name = tensor("op_4419"), val = tensor([1, 1])]; tensor var_4421 = const()[name = tensor("op_4421"), val = tensor([1, 1])]; tensor x_443_pad_type_0 = const()[name = tensor("x_443_pad_type_0"), val = tensor("custom")]; tensor x_443_pad_0 = const()[name = tensor("x_443_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_15_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338799808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342076672))), name = tensor("layers_15_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_15_fc1_module_bias_to_fp16 = const()[name = tensor("layers_15_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342076800)))]; tensor x_443_cast_fp16 = conv(bias = layers_15_fc1_module_bias_to_fp16, dilations = var_4421, groups = var_4189, pad = x_443_pad_0, pad_type = x_443_pad_type_0, strides = var_4419, weight = layers_15_fc1_module_weight_to_fp16_palettized, x = input_347_cast_fp16)[name = tensor("x_443_cast_fp16")]; tensor layers_15_fc1_output_scale_to_fp16 = const()[name = tensor("layers_15_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342087104)))]; tensor input_349_cast_fp16 = mul(x = x_443_cast_fp16, y = layers_15_fc1_output_scale_to_fp16)[name = tensor("input_349_cast_fp16")]; tensor x_445_mode_0 = const()[name = tensor("x_445_mode_0"), val = tensor("EXACT")]; tensor x_445_cast_fp16 = gelu(mode = x_445_mode_0, x = input_349_cast_fp16)[name = tensor("x_445_cast_fp16")]; tensor layers_15_fc2_input_shift_to_fp16 = const()[name = tensor("layers_15_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342097408)))]; tensor input_351_cast_fp16 = sub(x = x_445_cast_fp16, y = layers_15_fc2_input_shift_to_fp16)[name = tensor("input_351_cast_fp16")]; tensor var_4432 = const()[name = tensor("op_4432"), val = tensor([1, 1])]; tensor var_4434 = const()[name = tensor("op_4434"), val = tensor([1, 1])]; tensor x_447_pad_type_0 = const()[name = tensor("x_447_pad_type_0"), val = tensor("custom")]; tensor x_447_pad_0 = const()[name = tensor("x_447_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_15_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342107712))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345384576))), name = tensor("layers_15_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_15_fc2_module_bias_to_fp16 = const()[name = tensor("layers_15_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345384704)))]; tensor x_447_cast_fp16 = conv(bias = layers_15_fc2_module_bias_to_fp16, dilations = var_4434, groups = var_4189, pad = x_447_pad_0, pad_type = x_447_pad_type_0, strides = var_4432, weight = layers_15_fc2_module_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = tensor("x_447_cast_fp16")]; tensor layers_15_fc2_output_scale_to_fp16 = const()[name = tensor("layers_15_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345387328)))]; tensor hidden_states_33_cast_fp16 = mul(x = x_447_cast_fp16, y = layers_15_fc2_output_scale_to_fp16)[name = tensor("hidden_states_33_cast_fp16")]; tensor inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = hidden_states_33_cast_fp16)[name = tensor("inputs_97_cast_fp16")]; tensor var_4448 = const()[name = tensor("op_4448"), val = tensor(3)]; tensor var_4455 = const()[name = tensor("op_4455"), val = tensor(1)]; tensor var_4456 = const()[name = tensor("op_4456"), val = tensor(true)]; tensor var_4468 = const()[name = tensor("op_4468"), val = tensor([1])]; tensor channels_mean_97_cast_fp16 = reduce_mean(axes = var_4468, keep_dims = var_4456, x = inputs_97_cast_fp16)[name = tensor("channels_mean_97_cast_fp16")]; tensor zero_mean_97_cast_fp16 = sub(x = inputs_97_cast_fp16, y = channels_mean_97_cast_fp16)[name = tensor("zero_mean_97_cast_fp16")]; tensor zero_mean_sq_97_cast_fp16 = mul(x = zero_mean_97_cast_fp16, y = zero_mean_97_cast_fp16)[name = tensor("zero_mean_sq_97_cast_fp16")]; tensor var_4472 = const()[name = tensor("op_4472"), val = tensor([1])]; tensor var_4473_cast_fp16 = reduce_mean(axes = var_4472, keep_dims = var_4456, x = zero_mean_sq_97_cast_fp16)[name = tensor("op_4473_cast_fp16")]; tensor var_4474_to_fp16 = const()[name = tensor("op_4474_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_4475_cast_fp16 = add(x = var_4473_cast_fp16, y = var_4474_to_fp16)[name = tensor("op_4475_cast_fp16")]; tensor denom_97_epsilon_0_to_fp16 = const()[name = tensor("denom_97_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_97_cast_fp16 = rsqrt(epsilon = denom_97_epsilon_0_to_fp16, x = var_4475_cast_fp16)[name = tensor("denom_97_cast_fp16")]; tensor out_97_cast_fp16 = mul(x = zero_mean_97_cast_fp16, y = denom_97_cast_fp16)[name = tensor("out_97_cast_fp16")]; tensor obj_225_gamma_0_to_fp16 = const()[name = tensor("obj_225_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345389952)))]; tensor obj_225_beta_0_to_fp16 = const()[name = tensor("obj_225_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345392576)))]; tensor obj_225_epsilon_0_to_fp16 = const()[name = tensor("obj_225_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_225_cast_fp16 = batch_norm(beta = obj_225_beta_0_to_fp16, epsilon = obj_225_epsilon_0_to_fp16, gamma = obj_225_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_97_cast_fp16)[name = tensor("obj_225_cast_fp16")]; tensor layers_16_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345395200)))]; tensor input_353_cast_fp16 = sub(x = obj_225_cast_fp16, y = layers_16_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_353_cast_fp16")]; tensor var_4494 = const()[name = tensor("op_4494"), val = tensor([1, 1])]; tensor var_4496 = const()[name = tensor("op_4496"), val = tensor([1, 1])]; tensor x_449_pad_type_0 = const()[name = tensor("x_449_pad_type_0"), val = tensor("custom")]; tensor x_449_pad_0 = const()[name = tensor("x_449_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_16_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345397824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(346217088))), name = tensor("layers_16_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_16_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(346217216)))]; tensor x_449_cast_fp16 = conv(bias = layers_16_self_attn_q_proj_module_bias_to_fp16, dilations = var_4496, groups = var_4455, pad = x_449_pad_0, pad_type = x_449_pad_type_0, strides = var_4494, weight = layers_16_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_353_cast_fp16)[name = tensor("x_449_cast_fp16")]; tensor layers_16_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(346219840)))]; tensor query_65_cast_fp16 = mul(x = x_449_cast_fp16, y = layers_16_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_65_cast_fp16")]; tensor var_4506 = const()[name = tensor("op_4506"), val = tensor([1, 1])]; tensor var_4508 = const()[name = tensor("op_4508"), val = tensor([1, 1])]; tensor x_451_pad_type_0 = const()[name = tensor("x_451_pad_type_0"), val = tensor("custom")]; tensor x_451_pad_0 = const()[name = tensor("x_451_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_16_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(346222464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347041728))), name = tensor("layers_16_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_16_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347041856)))]; tensor x_451_cast_fp16 = conv(bias = layers_16_self_attn_k_proj_module_bias_to_fp16, dilations = var_4508, groups = var_4455, pad = x_451_pad_0, pad_type = x_451_pad_type_0, strides = var_4506, weight = layers_16_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_353_cast_fp16)[name = tensor("x_451_cast_fp16")]; tensor layers_16_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_16_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347044480)))]; tensor current_key_33_cast_fp16 = mul(x = x_451_cast_fp16, y = layers_16_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_33_cast_fp16")]; tensor var_4518 = const()[name = tensor("op_4518"), val = tensor([1, 1])]; tensor var_4520 = const()[name = tensor("op_4520"), val = tensor([1, 1])]; tensor x_453_pad_type_0 = const()[name = tensor("x_453_pad_type_0"), val = tensor("custom")]; tensor x_453_pad_0 = const()[name = tensor("x_453_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_16_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347047104))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347866368))), name = tensor("layers_16_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_16_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347866496)))]; tensor x_453_cast_fp16 = conv(bias = layers_16_self_attn_v_proj_module_bias_to_fp16, dilations = var_4520, groups = var_4455, pad = x_453_pad_0, pad_type = x_453_pad_type_0, strides = var_4518, weight = layers_16_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_353_cast_fp16)[name = tensor("x_453_cast_fp16")]; tensor layers_16_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347869120)))]; tensor current_value_33_cast_fp16 = mul(x = x_453_cast_fp16, y = layers_16_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_33_cast_fp16")]; tensor var_4528_cast_fp16 = mul(x = current_key_33_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_4528_cast_fp16")]; tensor var_4530_cast_fp16 = mul(x = var_103_cast_fp16_16, y = var_257_cast_fp16)[name = tensor("op_4530_cast_fp16")]; tensor key_65_cast_fp16 = add(x = var_4528_cast_fp16, y = var_4530_cast_fp16)[name = tensor("key_65_cast_fp16")]; tensor var_4532_cast_fp16 = mul(x = current_value_33_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_4532_cast_fp16")]; tensor var_4534_cast_fp16 = mul(x = var_138_cast_fp16_16, y = var_257_cast_fp16)[name = tensor("op_4534_cast_fp16")]; tensor value_65_cast_fp16 = add(x = var_4532_cast_fp16, y = var_4534_cast_fp16)[name = tensor("value_65_cast_fp16")]; tensor var_4537 = const()[name = tensor("op_4537"), val = tensor([1, 20, 64, -1])]; tensor var_4538_cast_fp16 = reshape(shape = var_4537, x = query_65_cast_fp16)[name = tensor("op_4538_cast_fp16")]; tensor var_4539_to_fp16 = const()[name = tensor("op_4539_to_fp16"), val = tensor(0x1p-3)]; tensor var_4540_cast_fp16 = mul(x = var_4538_cast_fp16, y = var_4539_to_fp16)[name = tensor("op_4540_cast_fp16")]; tensor var_4541 = const()[name = tensor("op_4541"), val = tensor([1, 20, 64, -1])]; tensor var_4542_cast_fp16 = reshape(shape = var_4541, x = key_65_cast_fp16)[name = tensor("op_4542_cast_fp16")]; tensor mh_w_97_transpose_x_0 = const()[name = tensor("mh_w_97_transpose_x_0"), val = tensor(true)]; tensor mh_w_97_transpose_y_0 = const()[name = tensor("mh_w_97_transpose_y_0"), val = tensor(false)]; tensor mh_w_97_cast_fp16 = matmul(transpose_x = mh_w_97_transpose_x_0, transpose_y = mh_w_97_transpose_y_0, x = var_4540_cast_fp16, y = var_4542_cast_fp16)[name = tensor("mh_w_97_cast_fp16")]; tensor mh_w_99_cast_fp16 = add(x = mh_w_97_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_99_cast_fp16")]; tensor var_4550_cast_fp16 = softmax(axis = var_4448, x = mh_w_99_cast_fp16)[name = tensor("op_4550_cast_fp16")]; tensor var_4551 = const()[name = tensor("op_4551"), val = tensor([1, 20, 64, -1])]; tensor var_4552_cast_fp16 = reshape(shape = var_4551, x = value_65_cast_fp16)[name = tensor("op_4552_cast_fp16")]; tensor attn_65_transpose_x_0 = const()[name = tensor("attn_65_transpose_x_0"), val = tensor(false)]; tensor attn_65_transpose_y_0 = const()[name = tensor("attn_65_transpose_y_0"), val = tensor(true)]; tensor attn_65_cast_fp16 = matmul(transpose_x = attn_65_transpose_x_0, transpose_y = attn_65_transpose_y_0, x = var_4552_cast_fp16, y = var_4550_cast_fp16)[name = tensor("attn_65_cast_fp16")]; tensor var_4555 = const()[name = tensor("op_4555"), val = tensor([1, 1280, 1, -1])]; tensor x_455_cast_fp16 = reshape(shape = var_4555, x = attn_65_cast_fp16)[name = tensor("x_455_cast_fp16")]; tensor layers_16_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347871744)))]; tensor input_359_cast_fp16 = sub(x = x_455_cast_fp16, y = layers_16_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_359_cast_fp16")]; tensor var_4563 = const()[name = tensor("op_4563"), val = tensor([1, 1])]; tensor var_4565 = const()[name = tensor("op_4565"), val = tensor([1, 1])]; tensor x_457_pad_type_0 = const()[name = tensor("x_457_pad_type_0"), val = tensor("custom")]; tensor x_457_pad_0 = const()[name = tensor("x_457_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_16_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347874368))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348693632))), name = tensor("layers_16_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_16_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348693760)))]; tensor x_457_cast_fp16 = conv(bias = layers_16_self_attn_o_proj_module_bias_to_fp16, dilations = var_4565, groups = var_4455, pad = x_457_pad_0, pad_type = x_457_pad_type_0, strides = var_4563, weight = layers_16_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_359_cast_fp16)[name = tensor("x_457_cast_fp16")]; tensor layers_16_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348696384)))]; tensor obj_231_cast_fp16 = mul(x = x_457_cast_fp16, y = layers_16_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_231_cast_fp16")]; tensor inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = obj_231_cast_fp16)[name = tensor("inputs_99_cast_fp16")]; tensor var_4576 = const()[name = tensor("op_4576"), val = tensor([1])]; tensor channels_mean_99_cast_fp16 = reduce_mean(axes = var_4576, keep_dims = var_4456, x = inputs_99_cast_fp16)[name = tensor("channels_mean_99_cast_fp16")]; tensor zero_mean_99_cast_fp16 = sub(x = inputs_99_cast_fp16, y = channels_mean_99_cast_fp16)[name = tensor("zero_mean_99_cast_fp16")]; tensor zero_mean_sq_99_cast_fp16 = mul(x = zero_mean_99_cast_fp16, y = zero_mean_99_cast_fp16)[name = tensor("zero_mean_sq_99_cast_fp16")]; tensor var_4580 = const()[name = tensor("op_4580"), val = tensor([1])]; tensor var_4581_cast_fp16 = reduce_mean(axes = var_4580, keep_dims = var_4456, x = zero_mean_sq_99_cast_fp16)[name = tensor("op_4581_cast_fp16")]; tensor var_4582_to_fp16 = const()[name = tensor("op_4582_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_4583_cast_fp16 = add(x = var_4581_cast_fp16, y = var_4582_to_fp16)[name = tensor("op_4583_cast_fp16")]; tensor denom_99_epsilon_0_to_fp16 = const()[name = tensor("denom_99_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_99_cast_fp16 = rsqrt(epsilon = denom_99_epsilon_0_to_fp16, x = var_4583_cast_fp16)[name = tensor("denom_99_cast_fp16")]; tensor out_99_cast_fp16 = mul(x = zero_mean_99_cast_fp16, y = denom_99_cast_fp16)[name = tensor("out_99_cast_fp16")]; tensor obj_233_gamma_0_to_fp16 = const()[name = tensor("obj_233_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348699008)))]; tensor obj_233_beta_0_to_fp16 = const()[name = tensor("obj_233_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348701632)))]; tensor obj_233_epsilon_0_to_fp16 = const()[name = tensor("obj_233_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_233_cast_fp16 = batch_norm(beta = obj_233_beta_0_to_fp16, epsilon = obj_233_epsilon_0_to_fp16, gamma = obj_233_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_99_cast_fp16)[name = tensor("obj_233_cast_fp16")]; tensor layers_16_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_16_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348704256)))]; tensor input_361_cast_fp16 = sub(x = obj_233_cast_fp16, y = layers_16_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_361_cast_fp16")]; tensor var_4602 = const()[name = tensor("op_4602"), val = tensor([1, 1])]; tensor var_4604 = const()[name = tensor("op_4604"), val = tensor([1, 1])]; tensor x_459_pad_type_0 = const()[name = tensor("x_459_pad_type_0"), val = tensor("custom")]; tensor x_459_pad_0 = const()[name = tensor("x_459_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_16_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348706880))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(349526144))), name = tensor("layers_16_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_16_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(349526272)))]; tensor x_459_cast_fp16 = conv(bias = layers_16_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_4604, groups = var_4455, pad = x_459_pad_0, pad_type = x_459_pad_type_0, strides = var_4602, weight = layers_16_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_361_cast_fp16)[name = tensor("x_459_cast_fp16")]; tensor layers_16_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_16_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(349528896)))]; tensor query_67_cast_fp16 = mul(x = x_459_cast_fp16, y = layers_16_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_67_cast_fp16")]; tensor var_4614 = const()[name = tensor("op_4614"), val = tensor([1, 1])]; tensor var_4616 = const()[name = tensor("op_4616"), val = tensor([1, 1])]; tensor x_461_pad_type_0 = const()[name = tensor("x_461_pad_type_0"), val = tensor("custom")]; tensor x_461_pad_0 = const()[name = tensor("x_461_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_16_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(349531520))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350350784))), name = tensor("layers_16_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_16_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350350912)))]; tensor x_461_cast_fp16 = conv(bias = layers_16_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_4616, groups = var_4455, pad = x_461_pad_0, pad_type = x_461_pad_type_0, strides = var_4614, weight = layers_16_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_461_cast_fp16")]; tensor layers_16_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_16_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350353536)))]; tensor key_67_cast_fp16 = mul(x = x_461_cast_fp16, y = layers_16_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_67_cast_fp16")]; tensor var_4626 = const()[name = tensor("op_4626"), val = tensor([1, 1])]; tensor var_4628 = const()[name = tensor("op_4628"), val = tensor([1, 1])]; tensor x_463_pad_type_0 = const()[name = tensor("x_463_pad_type_0"), val = tensor("custom")]; tensor x_463_pad_0 = const()[name = tensor("x_463_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_16_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350356160))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351175424))), name = tensor("layers_16_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_16_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351175552)))]; tensor x_463_cast_fp16 = conv(bias = layers_16_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_4628, groups = var_4455, pad = x_463_pad_0, pad_type = x_463_pad_type_0, strides = var_4626, weight = layers_16_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_463_cast_fp16")]; tensor layers_16_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_16_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351178176)))]; tensor value_67_cast_fp16 = mul(x = x_463_cast_fp16, y = layers_16_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_67_cast_fp16")]; tensor var_4633 = const()[name = tensor("op_4633"), val = tensor([1, 20, 64, -1])]; tensor var_4634_cast_fp16 = reshape(shape = var_4633, x = query_67_cast_fp16)[name = tensor("op_4634_cast_fp16")]; tensor var_4635_to_fp16 = const()[name = tensor("op_4635_to_fp16"), val = tensor(0x1p-3)]; tensor var_4636_cast_fp16 = mul(x = var_4634_cast_fp16, y = var_4635_to_fp16)[name = tensor("op_4636_cast_fp16")]; tensor var_4637 = const()[name = tensor("op_4637"), val = tensor([1, 20, 64, -1])]; tensor var_4638_cast_fp16 = reshape(shape = var_4637, x = key_67_cast_fp16)[name = tensor("op_4638_cast_fp16")]; tensor mh_w_101_transpose_x_0 = const()[name = tensor("mh_w_101_transpose_x_0"), val = tensor(true)]; tensor mh_w_101_transpose_y_0 = const()[name = tensor("mh_w_101_transpose_y_0"), val = tensor(false)]; tensor mh_w_101_cast_fp16 = matmul(transpose_x = mh_w_101_transpose_x_0, transpose_y = mh_w_101_transpose_y_0, x = var_4636_cast_fp16, y = var_4638_cast_fp16)[name = tensor("mh_w_101_cast_fp16")]; tensor obj_237_cast_fp16 = softmax(axis = var_4448, x = mh_w_101_cast_fp16)[name = tensor("obj_237_cast_fp16")]; tensor var_4642 = const()[name = tensor("op_4642"), val = tensor([1, 20, 64, -1])]; tensor var_4643_cast_fp16 = reshape(shape = var_4642, x = value_67_cast_fp16)[name = tensor("op_4643_cast_fp16")]; tensor attn_67_transpose_x_0 = const()[name = tensor("attn_67_transpose_x_0"), val = tensor(false)]; tensor attn_67_transpose_y_0 = const()[name = tensor("attn_67_transpose_y_0"), val = tensor(true)]; tensor attn_67_cast_fp16 = matmul(transpose_x = attn_67_transpose_x_0, transpose_y = attn_67_transpose_y_0, x = var_4643_cast_fp16, y = obj_237_cast_fp16)[name = tensor("attn_67_cast_fp16")]; tensor var_4646 = const()[name = tensor("op_4646"), val = tensor([1, 1280, 1, -1])]; tensor x_465_cast_fp16 = reshape(shape = var_4646, x = attn_67_cast_fp16)[name = tensor("x_465_cast_fp16")]; tensor layers_16_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_16_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351180800)))]; tensor input_367_cast_fp16 = sub(x = x_465_cast_fp16, y = layers_16_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_367_cast_fp16")]; tensor var_4654 = const()[name = tensor("op_4654"), val = tensor([1, 1])]; tensor var_4656 = const()[name = tensor("op_4656"), val = tensor([1, 1])]; tensor x_467_pad_type_0 = const()[name = tensor("x_467_pad_type_0"), val = tensor("custom")]; tensor x_467_pad_0 = const()[name = tensor("x_467_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_16_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351183424))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352002688))), name = tensor("layers_16_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_16_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352002816)))]; tensor x_467_cast_fp16 = conv(bias = layers_16_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_4656, groups = var_4455, pad = x_467_pad_0, pad_type = x_467_pad_type_0, strides = var_4654, weight = layers_16_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_367_cast_fp16)[name = tensor("x_467_cast_fp16")]; tensor layers_16_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_16_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352005440)))]; tensor obj_235_cast_fp16 = mul(x = x_467_cast_fp16, y = layers_16_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_235_cast_fp16")]; tensor inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = obj_235_cast_fp16)[name = tensor("inputs_101_cast_fp16")]; tensor var_4666 = const()[name = tensor("op_4666"), val = tensor([1])]; tensor channels_mean_101_cast_fp16 = reduce_mean(axes = var_4666, keep_dims = var_4456, x = inputs_101_cast_fp16)[name = tensor("channels_mean_101_cast_fp16")]; tensor zero_mean_101_cast_fp16 = sub(x = inputs_101_cast_fp16, y = channels_mean_101_cast_fp16)[name = tensor("zero_mean_101_cast_fp16")]; tensor zero_mean_sq_101_cast_fp16 = mul(x = zero_mean_101_cast_fp16, y = zero_mean_101_cast_fp16)[name = tensor("zero_mean_sq_101_cast_fp16")]; tensor var_4670 = const()[name = tensor("op_4670"), val = tensor([1])]; tensor var_4671_cast_fp16 = reduce_mean(axes = var_4670, keep_dims = var_4456, x = zero_mean_sq_101_cast_fp16)[name = tensor("op_4671_cast_fp16")]; tensor var_4672_to_fp16 = const()[name = tensor("op_4672_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_4673_cast_fp16 = add(x = var_4671_cast_fp16, y = var_4672_to_fp16)[name = tensor("op_4673_cast_fp16")]; tensor denom_101_epsilon_0_to_fp16 = const()[name = tensor("denom_101_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_101_cast_fp16 = rsqrt(epsilon = denom_101_epsilon_0_to_fp16, x = var_4673_cast_fp16)[name = tensor("denom_101_cast_fp16")]; tensor out_101_cast_fp16 = mul(x = zero_mean_101_cast_fp16, y = denom_101_cast_fp16)[name = tensor("out_101_cast_fp16")]; tensor x_469_gamma_0_to_fp16 = const()[name = tensor("x_469_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352008064)))]; tensor x_469_beta_0_to_fp16 = const()[name = tensor("x_469_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352010688)))]; tensor x_469_epsilon_0_to_fp16 = const()[name = tensor("x_469_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_469_cast_fp16 = batch_norm(beta = x_469_beta_0_to_fp16, epsilon = x_469_epsilon_0_to_fp16, gamma = x_469_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_101_cast_fp16)[name = tensor("x_469_cast_fp16")]; tensor layers_16_fc1_input_shift_to_fp16 = const()[name = tensor("layers_16_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352013312)))]; tensor input_369_cast_fp16 = sub(x = x_469_cast_fp16, y = layers_16_fc1_input_shift_to_fp16)[name = tensor("input_369_cast_fp16")]; tensor var_4688 = const()[name = tensor("op_4688"), val = tensor([1, 1])]; tensor var_4690 = const()[name = tensor("op_4690"), val = tensor([1, 1])]; tensor x_471_pad_type_0 = const()[name = tensor("x_471_pad_type_0"), val = tensor("custom")]; tensor x_471_pad_0 = const()[name = tensor("x_471_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_16_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352015936))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355292800))), name = tensor("layers_16_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_16_fc1_module_bias_to_fp16 = const()[name = tensor("layers_16_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355292928)))]; tensor x_471_cast_fp16 = conv(bias = layers_16_fc1_module_bias_to_fp16, dilations = var_4690, groups = var_4455, pad = x_471_pad_0, pad_type = x_471_pad_type_0, strides = var_4688, weight = layers_16_fc1_module_weight_to_fp16_palettized, x = input_369_cast_fp16)[name = tensor("x_471_cast_fp16")]; tensor layers_16_fc1_output_scale_to_fp16 = const()[name = tensor("layers_16_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355303232)))]; tensor input_371_cast_fp16 = mul(x = x_471_cast_fp16, y = layers_16_fc1_output_scale_to_fp16)[name = tensor("input_371_cast_fp16")]; tensor x_473_mode_0 = const()[name = tensor("x_473_mode_0"), val = tensor("EXACT")]; tensor x_473_cast_fp16 = gelu(mode = x_473_mode_0, x = input_371_cast_fp16)[name = tensor("x_473_cast_fp16")]; tensor layers_16_fc2_input_shift_to_fp16 = const()[name = tensor("layers_16_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355313536)))]; tensor input_373_cast_fp16 = sub(x = x_473_cast_fp16, y = layers_16_fc2_input_shift_to_fp16)[name = tensor("input_373_cast_fp16")]; tensor var_4701 = const()[name = tensor("op_4701"), val = tensor([1, 1])]; tensor var_4703 = const()[name = tensor("op_4703"), val = tensor([1, 1])]; tensor x_475_pad_type_0 = const()[name = tensor("x_475_pad_type_0"), val = tensor("custom")]; tensor x_475_pad_0 = const()[name = tensor("x_475_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_16_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355323840))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358600704))), name = tensor("layers_16_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_16_fc2_module_bias_to_fp16 = const()[name = tensor("layers_16_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358600832)))]; tensor x_475_cast_fp16 = conv(bias = layers_16_fc2_module_bias_to_fp16, dilations = var_4703, groups = var_4455, pad = x_475_pad_0, pad_type = x_475_pad_type_0, strides = var_4701, weight = layers_16_fc2_module_weight_to_fp16_palettized, x = input_373_cast_fp16)[name = tensor("x_475_cast_fp16")]; tensor layers_16_fc2_output_scale_to_fp16 = const()[name = tensor("layers_16_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358603456)))]; tensor hidden_states_35_cast_fp16 = mul(x = x_475_cast_fp16, y = layers_16_fc2_output_scale_to_fp16)[name = tensor("hidden_states_35_cast_fp16")]; tensor inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = hidden_states_35_cast_fp16)[name = tensor("inputs_103_cast_fp16")]; tensor var_4718 = const()[name = tensor("op_4718"), val = tensor(3)]; tensor var_4725 = const()[name = tensor("op_4725"), val = tensor(1)]; tensor var_4726 = const()[name = tensor("op_4726"), val = tensor(true)]; tensor var_4738 = const()[name = tensor("op_4738"), val = tensor([1])]; tensor channels_mean_103_cast_fp16 = reduce_mean(axes = var_4738, keep_dims = var_4726, x = inputs_103_cast_fp16)[name = tensor("channels_mean_103_cast_fp16")]; tensor zero_mean_103_cast_fp16 = sub(x = inputs_103_cast_fp16, y = channels_mean_103_cast_fp16)[name = tensor("zero_mean_103_cast_fp16")]; tensor zero_mean_sq_103_cast_fp16 = mul(x = zero_mean_103_cast_fp16, y = zero_mean_103_cast_fp16)[name = tensor("zero_mean_sq_103_cast_fp16")]; tensor var_4742 = const()[name = tensor("op_4742"), val = tensor([1])]; tensor var_4743_cast_fp16 = reduce_mean(axes = var_4742, keep_dims = var_4726, x = zero_mean_sq_103_cast_fp16)[name = tensor("op_4743_cast_fp16")]; tensor var_4744_to_fp16 = const()[name = tensor("op_4744_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_4745_cast_fp16 = add(x = var_4743_cast_fp16, y = var_4744_to_fp16)[name = tensor("op_4745_cast_fp16")]; tensor denom_103_epsilon_0_to_fp16 = const()[name = tensor("denom_103_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_103_cast_fp16 = rsqrt(epsilon = denom_103_epsilon_0_to_fp16, x = var_4745_cast_fp16)[name = tensor("denom_103_cast_fp16")]; tensor out_103_cast_fp16 = mul(x = zero_mean_103_cast_fp16, y = denom_103_cast_fp16)[name = tensor("out_103_cast_fp16")]; tensor obj_239_gamma_0_to_fp16 = const()[name = tensor("obj_239_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358606080)))]; tensor obj_239_beta_0_to_fp16 = const()[name = tensor("obj_239_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358608704)))]; tensor obj_239_epsilon_0_to_fp16 = const()[name = tensor("obj_239_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_239_cast_fp16 = batch_norm(beta = obj_239_beta_0_to_fp16, epsilon = obj_239_epsilon_0_to_fp16, gamma = obj_239_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_103_cast_fp16)[name = tensor("obj_239_cast_fp16")]; tensor layers_17_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358611328)))]; tensor input_375_cast_fp16 = sub(x = obj_239_cast_fp16, y = layers_17_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_375_cast_fp16")]; tensor var_4764 = const()[name = tensor("op_4764"), val = tensor([1, 1])]; tensor var_4766 = const()[name = tensor("op_4766"), val = tensor([1, 1])]; tensor x_477_pad_type_0 = const()[name = tensor("x_477_pad_type_0"), val = tensor("custom")]; tensor x_477_pad_0 = const()[name = tensor("x_477_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_17_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358613952))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359433216))), name = tensor("layers_17_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_17_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359433344)))]; tensor x_477_cast_fp16 = conv(bias = layers_17_self_attn_q_proj_module_bias_to_fp16, dilations = var_4766, groups = var_4725, pad = x_477_pad_0, pad_type = x_477_pad_type_0, strides = var_4764, weight = layers_17_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_375_cast_fp16)[name = tensor("x_477_cast_fp16")]; tensor layers_17_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359435968)))]; tensor query_69_cast_fp16 = mul(x = x_477_cast_fp16, y = layers_17_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_69_cast_fp16")]; tensor var_4776 = const()[name = tensor("op_4776"), val = tensor([1, 1])]; tensor var_4778 = const()[name = tensor("op_4778"), val = tensor([1, 1])]; tensor x_479_pad_type_0 = const()[name = tensor("x_479_pad_type_0"), val = tensor("custom")]; tensor x_479_pad_0 = const()[name = tensor("x_479_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_17_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359438592))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360257856))), name = tensor("layers_17_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_17_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360257984)))]; tensor x_479_cast_fp16 = conv(bias = layers_17_self_attn_k_proj_module_bias_to_fp16, dilations = var_4778, groups = var_4725, pad = x_479_pad_0, pad_type = x_479_pad_type_0, strides = var_4776, weight = layers_17_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_375_cast_fp16)[name = tensor("x_479_cast_fp16")]; tensor layers_17_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_17_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360260608)))]; tensor current_key_35_cast_fp16 = mul(x = x_479_cast_fp16, y = layers_17_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_35_cast_fp16")]; tensor var_4788 = const()[name = tensor("op_4788"), val = tensor([1, 1])]; tensor var_4790 = const()[name = tensor("op_4790"), val = tensor([1, 1])]; tensor x_481_pad_type_0 = const()[name = tensor("x_481_pad_type_0"), val = tensor("custom")]; tensor x_481_pad_0 = const()[name = tensor("x_481_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_17_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360263232))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361082496))), name = tensor("layers_17_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_17_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361082624)))]; tensor x_481_cast_fp16 = conv(bias = layers_17_self_attn_v_proj_module_bias_to_fp16, dilations = var_4790, groups = var_4725, pad = x_481_pad_0, pad_type = x_481_pad_type_0, strides = var_4788, weight = layers_17_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_375_cast_fp16)[name = tensor("x_481_cast_fp16")]; tensor layers_17_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361085248)))]; tensor current_value_35_cast_fp16 = mul(x = x_481_cast_fp16, y = layers_17_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_35_cast_fp16")]; tensor var_4798_cast_fp16 = mul(x = current_key_35_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_4798_cast_fp16")]; tensor var_4800_cast_fp16 = mul(x = var_103_cast_fp16_17, y = var_257_cast_fp16)[name = tensor("op_4800_cast_fp16")]; tensor key_69_cast_fp16 = add(x = var_4798_cast_fp16, y = var_4800_cast_fp16)[name = tensor("key_69_cast_fp16")]; tensor var_4802_cast_fp16 = mul(x = current_value_35_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_4802_cast_fp16")]; tensor var_4804_cast_fp16 = mul(x = var_138_cast_fp16_17, y = var_257_cast_fp16)[name = tensor("op_4804_cast_fp16")]; tensor value_69_cast_fp16 = add(x = var_4802_cast_fp16, y = var_4804_cast_fp16)[name = tensor("value_69_cast_fp16")]; tensor var_4807 = const()[name = tensor("op_4807"), val = tensor([1, 20, 64, -1])]; tensor var_4808_cast_fp16 = reshape(shape = var_4807, x = query_69_cast_fp16)[name = tensor("op_4808_cast_fp16")]; tensor var_4809_to_fp16 = const()[name = tensor("op_4809_to_fp16"), val = tensor(0x1p-3)]; tensor var_4810_cast_fp16 = mul(x = var_4808_cast_fp16, y = var_4809_to_fp16)[name = tensor("op_4810_cast_fp16")]; tensor var_4811 = const()[name = tensor("op_4811"), val = tensor([1, 20, 64, -1])]; tensor var_4812_cast_fp16 = reshape(shape = var_4811, x = key_69_cast_fp16)[name = tensor("op_4812_cast_fp16")]; tensor mh_w_103_transpose_x_0 = const()[name = tensor("mh_w_103_transpose_x_0"), val = tensor(true)]; tensor mh_w_103_transpose_y_0 = const()[name = tensor("mh_w_103_transpose_y_0"), val = tensor(false)]; tensor mh_w_103_cast_fp16 = matmul(transpose_x = mh_w_103_transpose_x_0, transpose_y = mh_w_103_transpose_y_0, x = var_4810_cast_fp16, y = var_4812_cast_fp16)[name = tensor("mh_w_103_cast_fp16")]; tensor mh_w_105_cast_fp16 = add(x = mh_w_103_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_105_cast_fp16")]; tensor var_4820_cast_fp16 = softmax(axis = var_4718, x = mh_w_105_cast_fp16)[name = tensor("op_4820_cast_fp16")]; tensor var_4821 = const()[name = tensor("op_4821"), val = tensor([1, 20, 64, -1])]; tensor var_4822_cast_fp16 = reshape(shape = var_4821, x = value_69_cast_fp16)[name = tensor("op_4822_cast_fp16")]; tensor attn_69_transpose_x_0 = const()[name = tensor("attn_69_transpose_x_0"), val = tensor(false)]; tensor attn_69_transpose_y_0 = const()[name = tensor("attn_69_transpose_y_0"), val = tensor(true)]; tensor attn_69_cast_fp16 = matmul(transpose_x = attn_69_transpose_x_0, transpose_y = attn_69_transpose_y_0, x = var_4822_cast_fp16, y = var_4820_cast_fp16)[name = tensor("attn_69_cast_fp16")]; tensor var_4825 = const()[name = tensor("op_4825"), val = tensor([1, 1280, 1, -1])]; tensor x_483_cast_fp16 = reshape(shape = var_4825, x = attn_69_cast_fp16)[name = tensor("x_483_cast_fp16")]; tensor layers_17_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361087872)))]; tensor input_381_cast_fp16 = sub(x = x_483_cast_fp16, y = layers_17_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_381_cast_fp16")]; tensor var_4833 = const()[name = tensor("op_4833"), val = tensor([1, 1])]; tensor var_4835 = const()[name = tensor("op_4835"), val = tensor([1, 1])]; tensor x_485_pad_type_0 = const()[name = tensor("x_485_pad_type_0"), val = tensor("custom")]; tensor x_485_pad_0 = const()[name = tensor("x_485_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_17_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361090496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361909760))), name = tensor("layers_17_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_17_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361909888)))]; tensor x_485_cast_fp16 = conv(bias = layers_17_self_attn_o_proj_module_bias_to_fp16, dilations = var_4835, groups = var_4725, pad = x_485_pad_0, pad_type = x_485_pad_type_0, strides = var_4833, weight = layers_17_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_381_cast_fp16)[name = tensor("x_485_cast_fp16")]; tensor layers_17_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361912512)))]; tensor obj_245_cast_fp16 = mul(x = x_485_cast_fp16, y = layers_17_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_245_cast_fp16")]; tensor inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = obj_245_cast_fp16)[name = tensor("inputs_105_cast_fp16")]; tensor var_4846 = const()[name = tensor("op_4846"), val = tensor([1])]; tensor channels_mean_105_cast_fp16 = reduce_mean(axes = var_4846, keep_dims = var_4726, x = inputs_105_cast_fp16)[name = tensor("channels_mean_105_cast_fp16")]; tensor zero_mean_105_cast_fp16 = sub(x = inputs_105_cast_fp16, y = channels_mean_105_cast_fp16)[name = tensor("zero_mean_105_cast_fp16")]; tensor zero_mean_sq_105_cast_fp16 = mul(x = zero_mean_105_cast_fp16, y = zero_mean_105_cast_fp16)[name = tensor("zero_mean_sq_105_cast_fp16")]; tensor var_4850 = const()[name = tensor("op_4850"), val = tensor([1])]; tensor var_4851_cast_fp16 = reduce_mean(axes = var_4850, keep_dims = var_4726, x = zero_mean_sq_105_cast_fp16)[name = tensor("op_4851_cast_fp16")]; tensor var_4852_to_fp16 = const()[name = tensor("op_4852_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_4853_cast_fp16 = add(x = var_4851_cast_fp16, y = var_4852_to_fp16)[name = tensor("op_4853_cast_fp16")]; tensor denom_105_epsilon_0_to_fp16 = const()[name = tensor("denom_105_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_105_cast_fp16 = rsqrt(epsilon = denom_105_epsilon_0_to_fp16, x = var_4853_cast_fp16)[name = tensor("denom_105_cast_fp16")]; tensor out_105_cast_fp16 = mul(x = zero_mean_105_cast_fp16, y = denom_105_cast_fp16)[name = tensor("out_105_cast_fp16")]; tensor obj_247_gamma_0_to_fp16 = const()[name = tensor("obj_247_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361915136)))]; tensor obj_247_beta_0_to_fp16 = const()[name = tensor("obj_247_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361917760)))]; tensor obj_247_epsilon_0_to_fp16 = const()[name = tensor("obj_247_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_247_cast_fp16 = batch_norm(beta = obj_247_beta_0_to_fp16, epsilon = obj_247_epsilon_0_to_fp16, gamma = obj_247_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_105_cast_fp16)[name = tensor("obj_247_cast_fp16")]; tensor layers_17_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_17_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361920384)))]; tensor input_383_cast_fp16 = sub(x = obj_247_cast_fp16, y = layers_17_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_383_cast_fp16")]; tensor var_4872 = const()[name = tensor("op_4872"), val = tensor([1, 1])]; tensor var_4874 = const()[name = tensor("op_4874"), val = tensor([1, 1])]; tensor x_487_pad_type_0 = const()[name = tensor("x_487_pad_type_0"), val = tensor("custom")]; tensor x_487_pad_0 = const()[name = tensor("x_487_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_17_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361923008))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362742272))), name = tensor("layers_17_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_17_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362742400)))]; tensor x_487_cast_fp16 = conv(bias = layers_17_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_4874, groups = var_4725, pad = x_487_pad_0, pad_type = x_487_pad_type_0, strides = var_4872, weight = layers_17_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_383_cast_fp16)[name = tensor("x_487_cast_fp16")]; tensor layers_17_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_17_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362745024)))]; tensor query_71_cast_fp16 = mul(x = x_487_cast_fp16, y = layers_17_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_71_cast_fp16")]; tensor var_4884 = const()[name = tensor("op_4884"), val = tensor([1, 1])]; tensor var_4886 = const()[name = tensor("op_4886"), val = tensor([1, 1])]; tensor x_489_pad_type_0 = const()[name = tensor("x_489_pad_type_0"), val = tensor("custom")]; tensor x_489_pad_0 = const()[name = tensor("x_489_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_17_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362747648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363566912))), name = tensor("layers_17_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_17_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363567040)))]; tensor x_489_cast_fp16 = conv(bias = layers_17_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_4886, groups = var_4725, pad = x_489_pad_0, pad_type = x_489_pad_type_0, strides = var_4884, weight = layers_17_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_489_cast_fp16")]; tensor layers_17_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_17_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363569664)))]; tensor key_71_cast_fp16 = mul(x = x_489_cast_fp16, y = layers_17_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_71_cast_fp16")]; tensor var_4896 = const()[name = tensor("op_4896"), val = tensor([1, 1])]; tensor var_4898 = const()[name = tensor("op_4898"), val = tensor([1, 1])]; tensor x_491_pad_type_0 = const()[name = tensor("x_491_pad_type_0"), val = tensor("custom")]; tensor x_491_pad_0 = const()[name = tensor("x_491_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_17_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363572288))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364391552))), name = tensor("layers_17_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_17_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364391680)))]; tensor x_491_cast_fp16 = conv(bias = layers_17_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_4898, groups = var_4725, pad = x_491_pad_0, pad_type = x_491_pad_type_0, strides = var_4896, weight = layers_17_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_491_cast_fp16")]; tensor layers_17_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_17_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364394304)))]; tensor value_71_cast_fp16 = mul(x = x_491_cast_fp16, y = layers_17_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_71_cast_fp16")]; tensor var_4903 = const()[name = tensor("op_4903"), val = tensor([1, 20, 64, -1])]; tensor var_4904_cast_fp16 = reshape(shape = var_4903, x = query_71_cast_fp16)[name = tensor("op_4904_cast_fp16")]; tensor var_4905_to_fp16 = const()[name = tensor("op_4905_to_fp16"), val = tensor(0x1p-3)]; tensor var_4906_cast_fp16 = mul(x = var_4904_cast_fp16, y = var_4905_to_fp16)[name = tensor("op_4906_cast_fp16")]; tensor var_4907 = const()[name = tensor("op_4907"), val = tensor([1, 20, 64, -1])]; tensor var_4908_cast_fp16 = reshape(shape = var_4907, x = key_71_cast_fp16)[name = tensor("op_4908_cast_fp16")]; tensor mh_w_107_transpose_x_0 = const()[name = tensor("mh_w_107_transpose_x_0"), val = tensor(true)]; tensor mh_w_107_transpose_y_0 = const()[name = tensor("mh_w_107_transpose_y_0"), val = tensor(false)]; tensor mh_w_107_cast_fp16 = matmul(transpose_x = mh_w_107_transpose_x_0, transpose_y = mh_w_107_transpose_y_0, x = var_4906_cast_fp16, y = var_4908_cast_fp16)[name = tensor("mh_w_107_cast_fp16")]; tensor obj_251_cast_fp16 = softmax(axis = var_4718, x = mh_w_107_cast_fp16)[name = tensor("obj_251_cast_fp16")]; tensor var_4912 = const()[name = tensor("op_4912"), val = tensor([1, 20, 64, -1])]; tensor var_4913_cast_fp16 = reshape(shape = var_4912, x = value_71_cast_fp16)[name = tensor("op_4913_cast_fp16")]; tensor attn_71_transpose_x_0 = const()[name = tensor("attn_71_transpose_x_0"), val = tensor(false)]; tensor attn_71_transpose_y_0 = const()[name = tensor("attn_71_transpose_y_0"), val = tensor(true)]; tensor attn_71_cast_fp16 = matmul(transpose_x = attn_71_transpose_x_0, transpose_y = attn_71_transpose_y_0, x = var_4913_cast_fp16, y = obj_251_cast_fp16)[name = tensor("attn_71_cast_fp16")]; tensor var_4916 = const()[name = tensor("op_4916"), val = tensor([1, 1280, 1, -1])]; tensor x_493_cast_fp16 = reshape(shape = var_4916, x = attn_71_cast_fp16)[name = tensor("x_493_cast_fp16")]; tensor layers_17_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_17_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364396928)))]; tensor input_389_cast_fp16 = sub(x = x_493_cast_fp16, y = layers_17_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_389_cast_fp16")]; tensor var_4924 = const()[name = tensor("op_4924"), val = tensor([1, 1])]; tensor var_4926 = const()[name = tensor("op_4926"), val = tensor([1, 1])]; tensor x_495_pad_type_0 = const()[name = tensor("x_495_pad_type_0"), val = tensor("custom")]; tensor x_495_pad_0 = const()[name = tensor("x_495_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_17_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364399552))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(365218816))), name = tensor("layers_17_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_17_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(365218944)))]; tensor x_495_cast_fp16 = conv(bias = layers_17_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_4926, groups = var_4725, pad = x_495_pad_0, pad_type = x_495_pad_type_0, strides = var_4924, weight = layers_17_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_389_cast_fp16)[name = tensor("x_495_cast_fp16")]; tensor layers_17_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_17_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(365221568)))]; tensor obj_249_cast_fp16 = mul(x = x_495_cast_fp16, y = layers_17_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_249_cast_fp16")]; tensor inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = obj_249_cast_fp16)[name = tensor("inputs_107_cast_fp16")]; tensor var_4936 = const()[name = tensor("op_4936"), val = tensor([1])]; tensor channels_mean_107_cast_fp16 = reduce_mean(axes = var_4936, keep_dims = var_4726, x = inputs_107_cast_fp16)[name = tensor("channels_mean_107_cast_fp16")]; tensor zero_mean_107_cast_fp16 = sub(x = inputs_107_cast_fp16, y = channels_mean_107_cast_fp16)[name = tensor("zero_mean_107_cast_fp16")]; tensor zero_mean_sq_107_cast_fp16 = mul(x = zero_mean_107_cast_fp16, y = zero_mean_107_cast_fp16)[name = tensor("zero_mean_sq_107_cast_fp16")]; tensor var_4940 = const()[name = tensor("op_4940"), val = tensor([1])]; tensor var_4941_cast_fp16 = reduce_mean(axes = var_4940, keep_dims = var_4726, x = zero_mean_sq_107_cast_fp16)[name = tensor("op_4941_cast_fp16")]; tensor var_4942_to_fp16 = const()[name = tensor("op_4942_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_4943_cast_fp16 = add(x = var_4941_cast_fp16, y = var_4942_to_fp16)[name = tensor("op_4943_cast_fp16")]; tensor denom_107_epsilon_0_to_fp16 = const()[name = tensor("denom_107_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_107_cast_fp16 = rsqrt(epsilon = denom_107_epsilon_0_to_fp16, x = var_4943_cast_fp16)[name = tensor("denom_107_cast_fp16")]; tensor out_107_cast_fp16 = mul(x = zero_mean_107_cast_fp16, y = denom_107_cast_fp16)[name = tensor("out_107_cast_fp16")]; tensor x_497_gamma_0_to_fp16 = const()[name = tensor("x_497_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(365224192)))]; tensor x_497_beta_0_to_fp16 = const()[name = tensor("x_497_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(365226816)))]; tensor x_497_epsilon_0_to_fp16 = const()[name = tensor("x_497_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_497_cast_fp16 = batch_norm(beta = x_497_beta_0_to_fp16, epsilon = x_497_epsilon_0_to_fp16, gamma = x_497_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_107_cast_fp16)[name = tensor("x_497_cast_fp16")]; tensor layers_17_fc1_input_shift_to_fp16 = const()[name = tensor("layers_17_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(365229440)))]; tensor input_391_cast_fp16 = sub(x = x_497_cast_fp16, y = layers_17_fc1_input_shift_to_fp16)[name = tensor("input_391_cast_fp16")]; tensor var_4958 = const()[name = tensor("op_4958"), val = tensor([1, 1])]; tensor var_4960 = const()[name = tensor("op_4960"), val = tensor([1, 1])]; tensor x_499_pad_type_0 = const()[name = tensor("x_499_pad_type_0"), val = tensor("custom")]; tensor x_499_pad_0 = const()[name = tensor("x_499_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_17_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(365232064))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368508928))), name = tensor("layers_17_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_17_fc1_module_bias_to_fp16 = const()[name = tensor("layers_17_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368509056)))]; tensor x_499_cast_fp16 = conv(bias = layers_17_fc1_module_bias_to_fp16, dilations = var_4960, groups = var_4725, pad = x_499_pad_0, pad_type = x_499_pad_type_0, strides = var_4958, weight = layers_17_fc1_module_weight_to_fp16_palettized, x = input_391_cast_fp16)[name = tensor("x_499_cast_fp16")]; tensor layers_17_fc1_output_scale_to_fp16 = const()[name = tensor("layers_17_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368519360)))]; tensor input_393_cast_fp16 = mul(x = x_499_cast_fp16, y = layers_17_fc1_output_scale_to_fp16)[name = tensor("input_393_cast_fp16")]; tensor x_501_mode_0 = const()[name = tensor("x_501_mode_0"), val = tensor("EXACT")]; tensor x_501_cast_fp16 = gelu(mode = x_501_mode_0, x = input_393_cast_fp16)[name = tensor("x_501_cast_fp16")]; tensor layers_17_fc2_input_shift_to_fp16 = const()[name = tensor("layers_17_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368529664)))]; tensor input_395_cast_fp16 = sub(x = x_501_cast_fp16, y = layers_17_fc2_input_shift_to_fp16)[name = tensor("input_395_cast_fp16")]; tensor var_4971 = const()[name = tensor("op_4971"), val = tensor([1, 1])]; tensor var_4973 = const()[name = tensor("op_4973"), val = tensor([1, 1])]; tensor x_503_pad_type_0 = const()[name = tensor("x_503_pad_type_0"), val = tensor("custom")]; tensor x_503_pad_0 = const()[name = tensor("x_503_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_17_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368539968))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371816832))), name = tensor("layers_17_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_17_fc2_module_bias_to_fp16 = const()[name = tensor("layers_17_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371816960)))]; tensor x_503_cast_fp16 = conv(bias = layers_17_fc2_module_bias_to_fp16, dilations = var_4973, groups = var_4725, pad = x_503_pad_0, pad_type = x_503_pad_type_0, strides = var_4971, weight = layers_17_fc2_module_weight_to_fp16_palettized, x = input_395_cast_fp16)[name = tensor("x_503_cast_fp16")]; tensor layers_17_fc2_output_scale_to_fp16 = const()[name = tensor("layers_17_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371819584)))]; tensor hidden_states_37_cast_fp16 = mul(x = x_503_cast_fp16, y = layers_17_fc2_output_scale_to_fp16)[name = tensor("hidden_states_37_cast_fp16")]; tensor inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = hidden_states_37_cast_fp16)[name = tensor("inputs_109_cast_fp16")]; tensor var_4988 = const()[name = tensor("op_4988"), val = tensor(3)]; tensor var_4995 = const()[name = tensor("op_4995"), val = tensor(1)]; tensor var_4996 = const()[name = tensor("op_4996"), val = tensor(true)]; tensor var_5008 = const()[name = tensor("op_5008"), val = tensor([1])]; tensor channels_mean_109_cast_fp16 = reduce_mean(axes = var_5008, keep_dims = var_4996, x = inputs_109_cast_fp16)[name = tensor("channels_mean_109_cast_fp16")]; tensor zero_mean_109_cast_fp16 = sub(x = inputs_109_cast_fp16, y = channels_mean_109_cast_fp16)[name = tensor("zero_mean_109_cast_fp16")]; tensor zero_mean_sq_109_cast_fp16 = mul(x = zero_mean_109_cast_fp16, y = zero_mean_109_cast_fp16)[name = tensor("zero_mean_sq_109_cast_fp16")]; tensor var_5012 = const()[name = tensor("op_5012"), val = tensor([1])]; tensor var_5013_cast_fp16 = reduce_mean(axes = var_5012, keep_dims = var_4996, x = zero_mean_sq_109_cast_fp16)[name = tensor("op_5013_cast_fp16")]; tensor var_5014_to_fp16 = const()[name = tensor("op_5014_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_5015_cast_fp16 = add(x = var_5013_cast_fp16, y = var_5014_to_fp16)[name = tensor("op_5015_cast_fp16")]; tensor denom_109_epsilon_0_to_fp16 = const()[name = tensor("denom_109_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_109_cast_fp16 = rsqrt(epsilon = denom_109_epsilon_0_to_fp16, x = var_5015_cast_fp16)[name = tensor("denom_109_cast_fp16")]; tensor out_109_cast_fp16 = mul(x = zero_mean_109_cast_fp16, y = denom_109_cast_fp16)[name = tensor("out_109_cast_fp16")]; tensor obj_253_gamma_0_to_fp16 = const()[name = tensor("obj_253_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371822208)))]; tensor obj_253_beta_0_to_fp16 = const()[name = tensor("obj_253_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371824832)))]; tensor obj_253_epsilon_0_to_fp16 = const()[name = tensor("obj_253_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_253_cast_fp16 = batch_norm(beta = obj_253_beta_0_to_fp16, epsilon = obj_253_epsilon_0_to_fp16, gamma = obj_253_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_109_cast_fp16)[name = tensor("obj_253_cast_fp16")]; tensor layers_18_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371827456)))]; tensor input_397_cast_fp16 = sub(x = obj_253_cast_fp16, y = layers_18_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_397_cast_fp16")]; tensor var_5034 = const()[name = tensor("op_5034"), val = tensor([1, 1])]; tensor var_5036 = const()[name = tensor("op_5036"), val = tensor([1, 1])]; tensor x_505_pad_type_0 = const()[name = tensor("x_505_pad_type_0"), val = tensor("custom")]; tensor x_505_pad_0 = const()[name = tensor("x_505_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_18_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371830080))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(372649344))), name = tensor("layers_18_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_18_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(372649472)))]; tensor x_505_cast_fp16 = conv(bias = layers_18_self_attn_q_proj_module_bias_to_fp16, dilations = var_5036, groups = var_4995, pad = x_505_pad_0, pad_type = x_505_pad_type_0, strides = var_5034, weight = layers_18_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = tensor("x_505_cast_fp16")]; tensor layers_18_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(372652096)))]; tensor query_73_cast_fp16 = mul(x = x_505_cast_fp16, y = layers_18_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_73_cast_fp16")]; tensor var_5046 = const()[name = tensor("op_5046"), val = tensor([1, 1])]; tensor var_5048 = const()[name = tensor("op_5048"), val = tensor([1, 1])]; tensor x_507_pad_type_0 = const()[name = tensor("x_507_pad_type_0"), val = tensor("custom")]; tensor x_507_pad_0 = const()[name = tensor("x_507_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_18_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(372654720))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373473984))), name = tensor("layers_18_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_18_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373474112)))]; tensor x_507_cast_fp16 = conv(bias = layers_18_self_attn_k_proj_module_bias_to_fp16, dilations = var_5048, groups = var_4995, pad = x_507_pad_0, pad_type = x_507_pad_type_0, strides = var_5046, weight = layers_18_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = tensor("x_507_cast_fp16")]; tensor layers_18_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_18_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373476736)))]; tensor current_key_37_cast_fp16 = mul(x = x_507_cast_fp16, y = layers_18_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_37_cast_fp16")]; tensor var_5058 = const()[name = tensor("op_5058"), val = tensor([1, 1])]; tensor var_5060 = const()[name = tensor("op_5060"), val = tensor([1, 1])]; tensor x_509_pad_type_0 = const()[name = tensor("x_509_pad_type_0"), val = tensor("custom")]; tensor x_509_pad_0 = const()[name = tensor("x_509_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_18_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373479360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(374298624))), name = tensor("layers_18_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_18_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(374298752)))]; tensor x_509_cast_fp16 = conv(bias = layers_18_self_attn_v_proj_module_bias_to_fp16, dilations = var_5060, groups = var_4995, pad = x_509_pad_0, pad_type = x_509_pad_type_0, strides = var_5058, weight = layers_18_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = tensor("x_509_cast_fp16")]; tensor layers_18_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(374301376)))]; tensor current_value_37_cast_fp16 = mul(x = x_509_cast_fp16, y = layers_18_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_37_cast_fp16")]; tensor var_5068_cast_fp16 = mul(x = current_key_37_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_5068_cast_fp16")]; tensor var_5070_cast_fp16 = mul(x = var_103_cast_fp16_18, y = var_257_cast_fp16)[name = tensor("op_5070_cast_fp16")]; tensor key_73_cast_fp16 = add(x = var_5068_cast_fp16, y = var_5070_cast_fp16)[name = tensor("key_73_cast_fp16")]; tensor var_5072_cast_fp16 = mul(x = current_value_37_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_5072_cast_fp16")]; tensor var_5074_cast_fp16 = mul(x = var_138_cast_fp16_18, y = var_257_cast_fp16)[name = tensor("op_5074_cast_fp16")]; tensor value_73_cast_fp16 = add(x = var_5072_cast_fp16, y = var_5074_cast_fp16)[name = tensor("value_73_cast_fp16")]; tensor var_5077 = const()[name = tensor("op_5077"), val = tensor([1, 20, 64, -1])]; tensor var_5078_cast_fp16 = reshape(shape = var_5077, x = query_73_cast_fp16)[name = tensor("op_5078_cast_fp16")]; tensor var_5079_to_fp16 = const()[name = tensor("op_5079_to_fp16"), val = tensor(0x1p-3)]; tensor var_5080_cast_fp16 = mul(x = var_5078_cast_fp16, y = var_5079_to_fp16)[name = tensor("op_5080_cast_fp16")]; tensor var_5081 = const()[name = tensor("op_5081"), val = tensor([1, 20, 64, -1])]; tensor var_5082_cast_fp16 = reshape(shape = var_5081, x = key_73_cast_fp16)[name = tensor("op_5082_cast_fp16")]; tensor mh_w_109_transpose_x_0 = const()[name = tensor("mh_w_109_transpose_x_0"), val = tensor(true)]; tensor mh_w_109_transpose_y_0 = const()[name = tensor("mh_w_109_transpose_y_0"), val = tensor(false)]; tensor mh_w_109_cast_fp16 = matmul(transpose_x = mh_w_109_transpose_x_0, transpose_y = mh_w_109_transpose_y_0, x = var_5080_cast_fp16, y = var_5082_cast_fp16)[name = tensor("mh_w_109_cast_fp16")]; tensor mh_w_111_cast_fp16 = add(x = mh_w_109_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_111_cast_fp16")]; tensor var_5090_cast_fp16 = softmax(axis = var_4988, x = mh_w_111_cast_fp16)[name = tensor("op_5090_cast_fp16")]; tensor var_5091 = const()[name = tensor("op_5091"), val = tensor([1, 20, 64, -1])]; tensor var_5092_cast_fp16 = reshape(shape = var_5091, x = value_73_cast_fp16)[name = tensor("op_5092_cast_fp16")]; tensor attn_73_transpose_x_0 = const()[name = tensor("attn_73_transpose_x_0"), val = tensor(false)]; tensor attn_73_transpose_y_0 = const()[name = tensor("attn_73_transpose_y_0"), val = tensor(true)]; tensor attn_73_cast_fp16 = matmul(transpose_x = attn_73_transpose_x_0, transpose_y = attn_73_transpose_y_0, x = var_5092_cast_fp16, y = var_5090_cast_fp16)[name = tensor("attn_73_cast_fp16")]; tensor var_5095 = const()[name = tensor("op_5095"), val = tensor([1, 1280, 1, -1])]; tensor x_511_cast_fp16 = reshape(shape = var_5095, x = attn_73_cast_fp16)[name = tensor("x_511_cast_fp16")]; tensor layers_18_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(374304000)))]; tensor input_403_cast_fp16 = sub(x = x_511_cast_fp16, y = layers_18_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_403_cast_fp16")]; tensor var_5103 = const()[name = tensor("op_5103"), val = tensor([1, 1])]; tensor var_5105 = const()[name = tensor("op_5105"), val = tensor([1, 1])]; tensor x_513_pad_type_0 = const()[name = tensor("x_513_pad_type_0"), val = tensor("custom")]; tensor x_513_pad_0 = const()[name = tensor("x_513_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_18_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(374306624))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375125888))), name = tensor("layers_18_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_18_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375126016)))]; tensor x_513_cast_fp16 = conv(bias = layers_18_self_attn_o_proj_module_bias_to_fp16, dilations = var_5105, groups = var_4995, pad = x_513_pad_0, pad_type = x_513_pad_type_0, strides = var_5103, weight = layers_18_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = tensor("x_513_cast_fp16")]; tensor layers_18_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375128640)))]; tensor obj_259_cast_fp16 = mul(x = x_513_cast_fp16, y = layers_18_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_259_cast_fp16")]; tensor inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = obj_259_cast_fp16)[name = tensor("inputs_111_cast_fp16")]; tensor var_5116 = const()[name = tensor("op_5116"), val = tensor([1])]; tensor channels_mean_111_cast_fp16 = reduce_mean(axes = var_5116, keep_dims = var_4996, x = inputs_111_cast_fp16)[name = tensor("channels_mean_111_cast_fp16")]; tensor zero_mean_111_cast_fp16 = sub(x = inputs_111_cast_fp16, y = channels_mean_111_cast_fp16)[name = tensor("zero_mean_111_cast_fp16")]; tensor zero_mean_sq_111_cast_fp16 = mul(x = zero_mean_111_cast_fp16, y = zero_mean_111_cast_fp16)[name = tensor("zero_mean_sq_111_cast_fp16")]; tensor var_5120 = const()[name = tensor("op_5120"), val = tensor([1])]; tensor var_5121_cast_fp16 = reduce_mean(axes = var_5120, keep_dims = var_4996, x = zero_mean_sq_111_cast_fp16)[name = tensor("op_5121_cast_fp16")]; tensor var_5122_to_fp16 = const()[name = tensor("op_5122_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_5123_cast_fp16 = add(x = var_5121_cast_fp16, y = var_5122_to_fp16)[name = tensor("op_5123_cast_fp16")]; tensor denom_111_epsilon_0_to_fp16 = const()[name = tensor("denom_111_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_111_cast_fp16 = rsqrt(epsilon = denom_111_epsilon_0_to_fp16, x = var_5123_cast_fp16)[name = tensor("denom_111_cast_fp16")]; tensor out_111_cast_fp16 = mul(x = zero_mean_111_cast_fp16, y = denom_111_cast_fp16)[name = tensor("out_111_cast_fp16")]; tensor obj_261_gamma_0_to_fp16 = const()[name = tensor("obj_261_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375131264)))]; tensor obj_261_beta_0_to_fp16 = const()[name = tensor("obj_261_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375133888)))]; tensor obj_261_epsilon_0_to_fp16 = const()[name = tensor("obj_261_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_261_cast_fp16 = batch_norm(beta = obj_261_beta_0_to_fp16, epsilon = obj_261_epsilon_0_to_fp16, gamma = obj_261_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_111_cast_fp16)[name = tensor("obj_261_cast_fp16")]; tensor layers_18_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_18_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375136512)))]; tensor input_405_cast_fp16 = sub(x = obj_261_cast_fp16, y = layers_18_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_405_cast_fp16")]; tensor var_5142 = const()[name = tensor("op_5142"), val = tensor([1, 1])]; tensor var_5144 = const()[name = tensor("op_5144"), val = tensor([1, 1])]; tensor x_515_pad_type_0 = const()[name = tensor("x_515_pad_type_0"), val = tensor("custom")]; tensor x_515_pad_0 = const()[name = tensor("x_515_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_18_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375139136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375958400))), name = tensor("layers_18_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_18_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375958528)))]; tensor x_515_cast_fp16 = conv(bias = layers_18_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_5144, groups = var_4995, pad = x_515_pad_0, pad_type = x_515_pad_type_0, strides = var_5142, weight = layers_18_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_405_cast_fp16)[name = tensor("x_515_cast_fp16")]; tensor layers_18_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_18_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375961152)))]; tensor query_75_cast_fp16 = mul(x = x_515_cast_fp16, y = layers_18_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_75_cast_fp16")]; tensor var_5154 = const()[name = tensor("op_5154"), val = tensor([1, 1])]; tensor var_5156 = const()[name = tensor("op_5156"), val = tensor([1, 1])]; tensor x_517_pad_type_0 = const()[name = tensor("x_517_pad_type_0"), val = tensor("custom")]; tensor x_517_pad_0 = const()[name = tensor("x_517_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_18_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375963776))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(376783040))), name = tensor("layers_18_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_18_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(376783168)))]; tensor x_517_cast_fp16 = conv(bias = layers_18_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_5156, groups = var_4995, pad = x_517_pad_0, pad_type = x_517_pad_type_0, strides = var_5154, weight = layers_18_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_517_cast_fp16")]; tensor layers_18_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_18_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(376785792)))]; tensor key_75_cast_fp16 = mul(x = x_517_cast_fp16, y = layers_18_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_75_cast_fp16")]; tensor var_5166 = const()[name = tensor("op_5166"), val = tensor([1, 1])]; tensor var_5168 = const()[name = tensor("op_5168"), val = tensor([1, 1])]; tensor x_519_pad_type_0 = const()[name = tensor("x_519_pad_type_0"), val = tensor("custom")]; tensor x_519_pad_0 = const()[name = tensor("x_519_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_18_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(376788416))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377607680))), name = tensor("layers_18_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_18_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377607808)))]; tensor x_519_cast_fp16 = conv(bias = layers_18_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_5168, groups = var_4995, pad = x_519_pad_0, pad_type = x_519_pad_type_0, strides = var_5166, weight = layers_18_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_519_cast_fp16")]; tensor layers_18_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_18_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377610432)))]; tensor value_75_cast_fp16 = mul(x = x_519_cast_fp16, y = layers_18_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_75_cast_fp16")]; tensor var_5173 = const()[name = tensor("op_5173"), val = tensor([1, 20, 64, -1])]; tensor var_5174_cast_fp16 = reshape(shape = var_5173, x = query_75_cast_fp16)[name = tensor("op_5174_cast_fp16")]; tensor var_5175_to_fp16 = const()[name = tensor("op_5175_to_fp16"), val = tensor(0x1p-3)]; tensor var_5176_cast_fp16 = mul(x = var_5174_cast_fp16, y = var_5175_to_fp16)[name = tensor("op_5176_cast_fp16")]; tensor var_5177 = const()[name = tensor("op_5177"), val = tensor([1, 20, 64, -1])]; tensor var_5178_cast_fp16 = reshape(shape = var_5177, x = key_75_cast_fp16)[name = tensor("op_5178_cast_fp16")]; tensor mh_w_113_transpose_x_0 = const()[name = tensor("mh_w_113_transpose_x_0"), val = tensor(true)]; tensor mh_w_113_transpose_y_0 = const()[name = tensor("mh_w_113_transpose_y_0"), val = tensor(false)]; tensor mh_w_113_cast_fp16 = matmul(transpose_x = mh_w_113_transpose_x_0, transpose_y = mh_w_113_transpose_y_0, x = var_5176_cast_fp16, y = var_5178_cast_fp16)[name = tensor("mh_w_113_cast_fp16")]; tensor obj_265_cast_fp16 = softmax(axis = var_4988, x = mh_w_113_cast_fp16)[name = tensor("obj_265_cast_fp16")]; tensor var_5182 = const()[name = tensor("op_5182"), val = tensor([1, 20, 64, -1])]; tensor var_5183_cast_fp16 = reshape(shape = var_5182, x = value_75_cast_fp16)[name = tensor("op_5183_cast_fp16")]; tensor attn_75_transpose_x_0 = const()[name = tensor("attn_75_transpose_x_0"), val = tensor(false)]; tensor attn_75_transpose_y_0 = const()[name = tensor("attn_75_transpose_y_0"), val = tensor(true)]; tensor attn_75_cast_fp16 = matmul(transpose_x = attn_75_transpose_x_0, transpose_y = attn_75_transpose_y_0, x = var_5183_cast_fp16, y = obj_265_cast_fp16)[name = tensor("attn_75_cast_fp16")]; tensor var_5186 = const()[name = tensor("op_5186"), val = tensor([1, 1280, 1, -1])]; tensor x_521_cast_fp16 = reshape(shape = var_5186, x = attn_75_cast_fp16)[name = tensor("x_521_cast_fp16")]; tensor layers_18_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_18_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377613056)))]; tensor input_411_cast_fp16 = sub(x = x_521_cast_fp16, y = layers_18_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_411_cast_fp16")]; tensor var_5194 = const()[name = tensor("op_5194"), val = tensor([1, 1])]; tensor var_5196 = const()[name = tensor("op_5196"), val = tensor([1, 1])]; tensor x_523_pad_type_0 = const()[name = tensor("x_523_pad_type_0"), val = tensor("custom")]; tensor x_523_pad_0 = const()[name = tensor("x_523_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_18_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377615680))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378434944))), name = tensor("layers_18_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_18_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378435072)))]; tensor x_523_cast_fp16 = conv(bias = layers_18_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_5196, groups = var_4995, pad = x_523_pad_0, pad_type = x_523_pad_type_0, strides = var_5194, weight = layers_18_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_411_cast_fp16)[name = tensor("x_523_cast_fp16")]; tensor layers_18_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_18_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378437696)))]; tensor obj_263_cast_fp16 = mul(x = x_523_cast_fp16, y = layers_18_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_263_cast_fp16")]; tensor inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = obj_263_cast_fp16)[name = tensor("inputs_113_cast_fp16")]; tensor var_5203 = const()[name = tensor("op_5203"), val = tensor([1])]; tensor channels_mean_113_cast_fp16 = reduce_mean(axes = var_5203, keep_dims = var_4996, x = inputs_113_cast_fp16)[name = tensor("channels_mean_113_cast_fp16")]; tensor zero_mean_113_cast_fp16 = sub(x = inputs_113_cast_fp16, y = channels_mean_113_cast_fp16)[name = tensor("zero_mean_113_cast_fp16")]; tensor zero_mean_sq_113_cast_fp16 = mul(x = zero_mean_113_cast_fp16, y = zero_mean_113_cast_fp16)[name = tensor("zero_mean_sq_113_cast_fp16")]; tensor var_5207 = const()[name = tensor("op_5207"), val = tensor([1])]; tensor var_5208_cast_fp16 = reduce_mean(axes = var_5207, keep_dims = var_4996, x = zero_mean_sq_113_cast_fp16)[name = tensor("op_5208_cast_fp16")]; tensor var_5209_to_fp16 = const()[name = tensor("op_5209_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_5210_cast_fp16 = add(x = var_5208_cast_fp16, y = var_5209_to_fp16)[name = tensor("op_5210_cast_fp16")]; tensor denom_113_epsilon_0_to_fp16 = const()[name = tensor("denom_113_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_113_cast_fp16 = rsqrt(epsilon = denom_113_epsilon_0_to_fp16, x = var_5210_cast_fp16)[name = tensor("denom_113_cast_fp16")]; tensor out_113_cast_fp16 = mul(x = zero_mean_113_cast_fp16, y = denom_113_cast_fp16)[name = tensor("out_113_cast_fp16")]; tensor x_525_gamma_0_to_fp16 = const()[name = tensor("x_525_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378440320)))]; tensor x_525_beta_0_to_fp16 = const()[name = tensor("x_525_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378442944)))]; tensor x_525_epsilon_0_to_fp16 = const()[name = tensor("x_525_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_525_cast_fp16 = batch_norm(beta = x_525_beta_0_to_fp16, epsilon = x_525_epsilon_0_to_fp16, gamma = x_525_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_113_cast_fp16)[name = tensor("x_525_cast_fp16")]; tensor layers_18_fc1_input_shift_to_fp16 = const()[name = tensor("layers_18_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378445568)))]; tensor input_413_cast_fp16 = sub(x = x_525_cast_fp16, y = layers_18_fc1_input_shift_to_fp16)[name = tensor("input_413_cast_fp16")]; tensor var_5225 = const()[name = tensor("op_5225"), val = tensor([1, 1])]; tensor var_5227 = const()[name = tensor("op_5227"), val = tensor([1, 1])]; tensor x_527_pad_type_0 = const()[name = tensor("x_527_pad_type_0"), val = tensor("custom")]; tensor x_527_pad_0 = const()[name = tensor("x_527_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_18_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378448192))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381725056))), name = tensor("layers_18_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_18_fc1_module_bias_to_fp16 = const()[name = tensor("layers_18_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381725184)))]; tensor x_527_cast_fp16 = conv(bias = layers_18_fc1_module_bias_to_fp16, dilations = var_5227, groups = var_4995, pad = x_527_pad_0, pad_type = x_527_pad_type_0, strides = var_5225, weight = layers_18_fc1_module_weight_to_fp16_palettized, x = input_413_cast_fp16)[name = tensor("x_527_cast_fp16")]; tensor layers_18_fc1_output_scale_to_fp16 = const()[name = tensor("layers_18_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381735488)))]; tensor input_415_cast_fp16 = mul(x = x_527_cast_fp16, y = layers_18_fc1_output_scale_to_fp16)[name = tensor("input_415_cast_fp16")]; tensor x_529_mode_0 = const()[name = tensor("x_529_mode_0"), val = tensor("EXACT")]; tensor x_529_cast_fp16 = gelu(mode = x_529_mode_0, x = input_415_cast_fp16)[name = tensor("x_529_cast_fp16")]; tensor layers_18_fc2_input_shift_to_fp16 = const()[name = tensor("layers_18_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381745792)))]; tensor input_417_cast_fp16 = sub(x = x_529_cast_fp16, y = layers_18_fc2_input_shift_to_fp16)[name = tensor("input_417_cast_fp16")]; tensor var_5238 = const()[name = tensor("op_5238"), val = tensor([1, 1])]; tensor var_5240 = const()[name = tensor("op_5240"), val = tensor([1, 1])]; tensor x_531_pad_type_0 = const()[name = tensor("x_531_pad_type_0"), val = tensor("custom")]; tensor x_531_pad_0 = const()[name = tensor("x_531_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_18_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381756096))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385032960))), name = tensor("layers_18_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_18_fc2_module_bias_to_fp16 = const()[name = tensor("layers_18_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385033088)))]; tensor x_531_cast_fp16 = conv(bias = layers_18_fc2_module_bias_to_fp16, dilations = var_5240, groups = var_4995, pad = x_531_pad_0, pad_type = x_531_pad_type_0, strides = var_5238, weight = layers_18_fc2_module_weight_to_fp16_palettized, x = input_417_cast_fp16)[name = tensor("x_531_cast_fp16")]; tensor layers_18_fc2_output_scale_to_fp16 = const()[name = tensor("layers_18_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385035712)))]; tensor hidden_states_39_cast_fp16 = mul(x = x_531_cast_fp16, y = layers_18_fc2_output_scale_to_fp16)[name = tensor("hidden_states_39_cast_fp16")]; tensor inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = hidden_states_39_cast_fp16)[name = tensor("inputs_115_cast_fp16")]; tensor var_5254 = const()[name = tensor("op_5254"), val = tensor(3)]; tensor var_5261 = const()[name = tensor("op_5261"), val = tensor(1)]; tensor var_5262 = const()[name = tensor("op_5262"), val = tensor(true)]; tensor var_5274 = const()[name = tensor("op_5274"), val = tensor([1])]; tensor channels_mean_115_cast_fp16 = reduce_mean(axes = var_5274, keep_dims = var_5262, x = inputs_115_cast_fp16)[name = tensor("channels_mean_115_cast_fp16")]; tensor zero_mean_115_cast_fp16 = sub(x = inputs_115_cast_fp16, y = channels_mean_115_cast_fp16)[name = tensor("zero_mean_115_cast_fp16")]; tensor zero_mean_sq_115_cast_fp16 = mul(x = zero_mean_115_cast_fp16, y = zero_mean_115_cast_fp16)[name = tensor("zero_mean_sq_115_cast_fp16")]; tensor var_5278 = const()[name = tensor("op_5278"), val = tensor([1])]; tensor var_5279_cast_fp16 = reduce_mean(axes = var_5278, keep_dims = var_5262, x = zero_mean_sq_115_cast_fp16)[name = tensor("op_5279_cast_fp16")]; tensor var_5280_to_fp16 = const()[name = tensor("op_5280_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_5281_cast_fp16 = add(x = var_5279_cast_fp16, y = var_5280_to_fp16)[name = tensor("op_5281_cast_fp16")]; tensor denom_115_epsilon_0_to_fp16 = const()[name = tensor("denom_115_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_115_cast_fp16 = rsqrt(epsilon = denom_115_epsilon_0_to_fp16, x = var_5281_cast_fp16)[name = tensor("denom_115_cast_fp16")]; tensor out_115_cast_fp16 = mul(x = zero_mean_115_cast_fp16, y = denom_115_cast_fp16)[name = tensor("out_115_cast_fp16")]; tensor obj_267_gamma_0_to_fp16 = const()[name = tensor("obj_267_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385038336)))]; tensor obj_267_beta_0_to_fp16 = const()[name = tensor("obj_267_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385040960)))]; tensor obj_267_epsilon_0_to_fp16 = const()[name = tensor("obj_267_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_267_cast_fp16 = batch_norm(beta = obj_267_beta_0_to_fp16, epsilon = obj_267_epsilon_0_to_fp16, gamma = obj_267_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_115_cast_fp16)[name = tensor("obj_267_cast_fp16")]; tensor layers_19_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385043584)))]; tensor input_419_cast_fp16 = sub(x = obj_267_cast_fp16, y = layers_19_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_419_cast_fp16")]; tensor var_5300 = const()[name = tensor("op_5300"), val = tensor([1, 1])]; tensor var_5302 = const()[name = tensor("op_5302"), val = tensor([1, 1])]; tensor x_533_pad_type_0 = const()[name = tensor("x_533_pad_type_0"), val = tensor("custom")]; tensor x_533_pad_0 = const()[name = tensor("x_533_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_19_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385046208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385865472))), name = tensor("layers_19_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_19_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385865600)))]; tensor x_533_cast_fp16 = conv(bias = layers_19_self_attn_q_proj_module_bias_to_fp16, dilations = var_5302, groups = var_5261, pad = x_533_pad_0, pad_type = x_533_pad_type_0, strides = var_5300, weight = layers_19_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_419_cast_fp16)[name = tensor("x_533_cast_fp16")]; tensor layers_19_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385868224)))]; tensor query_77_cast_fp16 = mul(x = x_533_cast_fp16, y = layers_19_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_77_cast_fp16")]; tensor var_5312 = const()[name = tensor("op_5312"), val = tensor([1, 1])]; tensor var_5314 = const()[name = tensor("op_5314"), val = tensor([1, 1])]; tensor x_535_pad_type_0 = const()[name = tensor("x_535_pad_type_0"), val = tensor("custom")]; tensor x_535_pad_0 = const()[name = tensor("x_535_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_19_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385870848))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386690112))), name = tensor("layers_19_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_19_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386690240)))]; tensor x_535_cast_fp16 = conv(bias = layers_19_self_attn_k_proj_module_bias_to_fp16, dilations = var_5314, groups = var_5261, pad = x_535_pad_0, pad_type = x_535_pad_type_0, strides = var_5312, weight = layers_19_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_419_cast_fp16)[name = tensor("x_535_cast_fp16")]; tensor layers_19_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_19_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386692864)))]; tensor current_key_39_cast_fp16 = mul(x = x_535_cast_fp16, y = layers_19_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_39_cast_fp16")]; tensor var_5324 = const()[name = tensor("op_5324"), val = tensor([1, 1])]; tensor var_5326 = const()[name = tensor("op_5326"), val = tensor([1, 1])]; tensor x_537_pad_type_0 = const()[name = tensor("x_537_pad_type_0"), val = tensor("custom")]; tensor x_537_pad_0 = const()[name = tensor("x_537_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_19_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386695488))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387514752))), name = tensor("layers_19_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_19_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387514880)))]; tensor x_537_cast_fp16 = conv(bias = layers_19_self_attn_v_proj_module_bias_to_fp16, dilations = var_5326, groups = var_5261, pad = x_537_pad_0, pad_type = x_537_pad_type_0, strides = var_5324, weight = layers_19_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_419_cast_fp16)[name = tensor("x_537_cast_fp16")]; tensor layers_19_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387517504)))]; tensor current_value_39_cast_fp16 = mul(x = x_537_cast_fp16, y = layers_19_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_39_cast_fp16")]; tensor var_5334_cast_fp16 = mul(x = current_key_39_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_5334_cast_fp16")]; tensor var_5336_cast_fp16 = mul(x = var_103_cast_fp16_19, y = var_257_cast_fp16)[name = tensor("op_5336_cast_fp16")]; tensor key_77_cast_fp16 = add(x = var_5334_cast_fp16, y = var_5336_cast_fp16)[name = tensor("key_77_cast_fp16")]; tensor var_5338_cast_fp16 = mul(x = current_value_39_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_5338_cast_fp16")]; tensor var_5340_cast_fp16 = mul(x = var_138_cast_fp16_19, y = var_257_cast_fp16)[name = tensor("op_5340_cast_fp16")]; tensor value_77_cast_fp16 = add(x = var_5338_cast_fp16, y = var_5340_cast_fp16)[name = tensor("value_77_cast_fp16")]; tensor var_5343 = const()[name = tensor("op_5343"), val = tensor([1, 20, 64, -1])]; tensor var_5344_cast_fp16 = reshape(shape = var_5343, x = query_77_cast_fp16)[name = tensor("op_5344_cast_fp16")]; tensor var_5345_to_fp16 = const()[name = tensor("op_5345_to_fp16"), val = tensor(0x1p-3)]; tensor var_5346_cast_fp16 = mul(x = var_5344_cast_fp16, y = var_5345_to_fp16)[name = tensor("op_5346_cast_fp16")]; tensor var_5347 = const()[name = tensor("op_5347"), val = tensor([1, 20, 64, -1])]; tensor var_5348_cast_fp16 = reshape(shape = var_5347, x = key_77_cast_fp16)[name = tensor("op_5348_cast_fp16")]; tensor mh_w_115_transpose_x_0 = const()[name = tensor("mh_w_115_transpose_x_0"), val = tensor(true)]; tensor mh_w_115_transpose_y_0 = const()[name = tensor("mh_w_115_transpose_y_0"), val = tensor(false)]; tensor mh_w_115_cast_fp16 = matmul(transpose_x = mh_w_115_transpose_x_0, transpose_y = mh_w_115_transpose_y_0, x = var_5346_cast_fp16, y = var_5348_cast_fp16)[name = tensor("mh_w_115_cast_fp16")]; tensor mh_w_117_cast_fp16 = add(x = mh_w_115_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_117_cast_fp16")]; tensor var_5356_cast_fp16 = softmax(axis = var_5254, x = mh_w_117_cast_fp16)[name = tensor("op_5356_cast_fp16")]; tensor var_5357 = const()[name = tensor("op_5357"), val = tensor([1, 20, 64, -1])]; tensor var_5358_cast_fp16 = reshape(shape = var_5357, x = value_77_cast_fp16)[name = tensor("op_5358_cast_fp16")]; tensor attn_77_transpose_x_0 = const()[name = tensor("attn_77_transpose_x_0"), val = tensor(false)]; tensor attn_77_transpose_y_0 = const()[name = tensor("attn_77_transpose_y_0"), val = tensor(true)]; tensor attn_77_cast_fp16 = matmul(transpose_x = attn_77_transpose_x_0, transpose_y = attn_77_transpose_y_0, x = var_5358_cast_fp16, y = var_5356_cast_fp16)[name = tensor("attn_77_cast_fp16")]; tensor var_5361 = const()[name = tensor("op_5361"), val = tensor([1, 1280, 1, -1])]; tensor x_539_cast_fp16 = reshape(shape = var_5361, x = attn_77_cast_fp16)[name = tensor("x_539_cast_fp16")]; tensor layers_19_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387520128)))]; tensor input_425_cast_fp16 = sub(x = x_539_cast_fp16, y = layers_19_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_425_cast_fp16")]; tensor var_5369 = const()[name = tensor("op_5369"), val = tensor([1, 1])]; tensor var_5371 = const()[name = tensor("op_5371"), val = tensor([1, 1])]; tensor x_541_pad_type_0 = const()[name = tensor("x_541_pad_type_0"), val = tensor("custom")]; tensor x_541_pad_0 = const()[name = tensor("x_541_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_19_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387522752))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(388342016))), name = tensor("layers_19_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_19_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(388342144)))]; tensor x_541_cast_fp16 = conv(bias = layers_19_self_attn_o_proj_module_bias_to_fp16, dilations = var_5371, groups = var_5261, pad = x_541_pad_0, pad_type = x_541_pad_type_0, strides = var_5369, weight = layers_19_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_425_cast_fp16)[name = tensor("x_541_cast_fp16")]; tensor layers_19_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(388344768)))]; tensor obj_273_cast_fp16 = mul(x = x_541_cast_fp16, y = layers_19_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_273_cast_fp16")]; tensor inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = obj_273_cast_fp16)[name = tensor("inputs_117_cast_fp16")]; tensor var_5382 = const()[name = tensor("op_5382"), val = tensor([1])]; tensor channels_mean_117_cast_fp16 = reduce_mean(axes = var_5382, keep_dims = var_5262, x = inputs_117_cast_fp16)[name = tensor("channels_mean_117_cast_fp16")]; tensor zero_mean_117_cast_fp16 = sub(x = inputs_117_cast_fp16, y = channels_mean_117_cast_fp16)[name = tensor("zero_mean_117_cast_fp16")]; tensor zero_mean_sq_117_cast_fp16 = mul(x = zero_mean_117_cast_fp16, y = zero_mean_117_cast_fp16)[name = tensor("zero_mean_sq_117_cast_fp16")]; tensor var_5386 = const()[name = tensor("op_5386"), val = tensor([1])]; tensor var_5387_cast_fp16 = reduce_mean(axes = var_5386, keep_dims = var_5262, x = zero_mean_sq_117_cast_fp16)[name = tensor("op_5387_cast_fp16")]; tensor var_5388_to_fp16 = const()[name = tensor("op_5388_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_5389_cast_fp16 = add(x = var_5387_cast_fp16, y = var_5388_to_fp16)[name = tensor("op_5389_cast_fp16")]; tensor denom_117_epsilon_0_to_fp16 = const()[name = tensor("denom_117_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_117_cast_fp16 = rsqrt(epsilon = denom_117_epsilon_0_to_fp16, x = var_5389_cast_fp16)[name = tensor("denom_117_cast_fp16")]; tensor out_117_cast_fp16 = mul(x = zero_mean_117_cast_fp16, y = denom_117_cast_fp16)[name = tensor("out_117_cast_fp16")]; tensor obj_275_gamma_0_to_fp16 = const()[name = tensor("obj_275_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(388347392)))]; tensor obj_275_beta_0_to_fp16 = const()[name = tensor("obj_275_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(388350016)))]; tensor obj_275_epsilon_0_to_fp16 = const()[name = tensor("obj_275_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_275_cast_fp16 = batch_norm(beta = obj_275_beta_0_to_fp16, epsilon = obj_275_epsilon_0_to_fp16, gamma = obj_275_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_117_cast_fp16)[name = tensor("obj_275_cast_fp16")]; tensor layers_19_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_19_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(388352640)))]; tensor input_427_cast_fp16 = sub(x = obj_275_cast_fp16, y = layers_19_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_427_cast_fp16")]; tensor var_5408 = const()[name = tensor("op_5408"), val = tensor([1, 1])]; tensor var_5410 = const()[name = tensor("op_5410"), val = tensor([1, 1])]; tensor x_543_pad_type_0 = const()[name = tensor("x_543_pad_type_0"), val = tensor("custom")]; tensor x_543_pad_0 = const()[name = tensor("x_543_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_19_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(388355264))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389174528))), name = tensor("layers_19_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_19_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389174656)))]; tensor x_543_cast_fp16 = conv(bias = layers_19_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_5410, groups = var_5261, pad = x_543_pad_0, pad_type = x_543_pad_type_0, strides = var_5408, weight = layers_19_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_427_cast_fp16)[name = tensor("x_543_cast_fp16")]; tensor layers_19_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_19_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389177280)))]; tensor query_79_cast_fp16 = mul(x = x_543_cast_fp16, y = layers_19_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_79_cast_fp16")]; tensor var_5420 = const()[name = tensor("op_5420"), val = tensor([1, 1])]; tensor var_5422 = const()[name = tensor("op_5422"), val = tensor([1, 1])]; tensor x_545_pad_type_0 = const()[name = tensor("x_545_pad_type_0"), val = tensor("custom")]; tensor x_545_pad_0 = const()[name = tensor("x_545_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_19_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389179904))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389999168))), name = tensor("layers_19_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_19_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389999296)))]; tensor x_545_cast_fp16 = conv(bias = layers_19_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_5422, groups = var_5261, pad = x_545_pad_0, pad_type = x_545_pad_type_0, strides = var_5420, weight = layers_19_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_545_cast_fp16")]; tensor layers_19_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_19_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(390001920)))]; tensor key_79_cast_fp16 = mul(x = x_545_cast_fp16, y = layers_19_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_79_cast_fp16")]; tensor var_5432 = const()[name = tensor("op_5432"), val = tensor([1, 1])]; tensor var_5434 = const()[name = tensor("op_5434"), val = tensor([1, 1])]; tensor x_547_pad_type_0 = const()[name = tensor("x_547_pad_type_0"), val = tensor("custom")]; tensor x_547_pad_0 = const()[name = tensor("x_547_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_19_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(390004544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(390823808))), name = tensor("layers_19_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_19_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(390823936)))]; tensor x_547_cast_fp16 = conv(bias = layers_19_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_5434, groups = var_5261, pad = x_547_pad_0, pad_type = x_547_pad_type_0, strides = var_5432, weight = layers_19_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_547_cast_fp16")]; tensor layers_19_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_19_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(390826560)))]; tensor value_79_cast_fp16 = mul(x = x_547_cast_fp16, y = layers_19_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_79_cast_fp16")]; tensor var_5439 = const()[name = tensor("op_5439"), val = tensor([1, 20, 64, -1])]; tensor var_5440_cast_fp16 = reshape(shape = var_5439, x = query_79_cast_fp16)[name = tensor("op_5440_cast_fp16")]; tensor var_5441_to_fp16 = const()[name = tensor("op_5441_to_fp16"), val = tensor(0x1p-3)]; tensor var_5442_cast_fp16 = mul(x = var_5440_cast_fp16, y = var_5441_to_fp16)[name = tensor("op_5442_cast_fp16")]; tensor var_5443 = const()[name = tensor("op_5443"), val = tensor([1, 20, 64, -1])]; tensor var_5444_cast_fp16 = reshape(shape = var_5443, x = key_79_cast_fp16)[name = tensor("op_5444_cast_fp16")]; tensor mh_w_119_transpose_x_0 = const()[name = tensor("mh_w_119_transpose_x_0"), val = tensor(true)]; tensor mh_w_119_transpose_y_0 = const()[name = tensor("mh_w_119_transpose_y_0"), val = tensor(false)]; tensor mh_w_119_cast_fp16 = matmul(transpose_x = mh_w_119_transpose_x_0, transpose_y = mh_w_119_transpose_y_0, x = var_5442_cast_fp16, y = var_5444_cast_fp16)[name = tensor("mh_w_119_cast_fp16")]; tensor obj_279_cast_fp16 = softmax(axis = var_5254, x = mh_w_119_cast_fp16)[name = tensor("obj_279_cast_fp16")]; tensor var_5448 = const()[name = tensor("op_5448"), val = tensor([1, 20, 64, -1])]; tensor var_5449_cast_fp16 = reshape(shape = var_5448, x = value_79_cast_fp16)[name = tensor("op_5449_cast_fp16")]; tensor attn_79_transpose_x_0 = const()[name = tensor("attn_79_transpose_x_0"), val = tensor(false)]; tensor attn_79_transpose_y_0 = const()[name = tensor("attn_79_transpose_y_0"), val = tensor(true)]; tensor attn_79_cast_fp16 = matmul(transpose_x = attn_79_transpose_x_0, transpose_y = attn_79_transpose_y_0, x = var_5449_cast_fp16, y = obj_279_cast_fp16)[name = tensor("attn_79_cast_fp16")]; tensor var_5452 = const()[name = tensor("op_5452"), val = tensor([1, 1280, 1, -1])]; tensor x_549_cast_fp16 = reshape(shape = var_5452, x = attn_79_cast_fp16)[name = tensor("x_549_cast_fp16")]; tensor layers_19_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_19_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(390829184)))]; tensor input_433_cast_fp16 = sub(x = x_549_cast_fp16, y = layers_19_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_433_cast_fp16")]; tensor var_5460 = const()[name = tensor("op_5460"), val = tensor([1, 1])]; tensor var_5462 = const()[name = tensor("op_5462"), val = tensor([1, 1])]; tensor x_551_pad_type_0 = const()[name = tensor("x_551_pad_type_0"), val = tensor("custom")]; tensor x_551_pad_0 = const()[name = tensor("x_551_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_19_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(390831808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391651072))), name = tensor("layers_19_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_19_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391651200)))]; tensor x_551_cast_fp16 = conv(bias = layers_19_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_5462, groups = var_5261, pad = x_551_pad_0, pad_type = x_551_pad_type_0, strides = var_5460, weight = layers_19_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_433_cast_fp16)[name = tensor("x_551_cast_fp16")]; tensor layers_19_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_19_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391653824)))]; tensor obj_277_cast_fp16 = mul(x = x_551_cast_fp16, y = layers_19_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_277_cast_fp16")]; tensor inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = obj_277_cast_fp16)[name = tensor("inputs_119_cast_fp16")]; tensor var_5472 = const()[name = tensor("op_5472"), val = tensor([1])]; tensor channels_mean_119_cast_fp16 = reduce_mean(axes = var_5472, keep_dims = var_5262, x = inputs_119_cast_fp16)[name = tensor("channels_mean_119_cast_fp16")]; tensor zero_mean_119_cast_fp16 = sub(x = inputs_119_cast_fp16, y = channels_mean_119_cast_fp16)[name = tensor("zero_mean_119_cast_fp16")]; tensor zero_mean_sq_119_cast_fp16 = mul(x = zero_mean_119_cast_fp16, y = zero_mean_119_cast_fp16)[name = tensor("zero_mean_sq_119_cast_fp16")]; tensor var_5476 = const()[name = tensor("op_5476"), val = tensor([1])]; tensor var_5477_cast_fp16 = reduce_mean(axes = var_5476, keep_dims = var_5262, x = zero_mean_sq_119_cast_fp16)[name = tensor("op_5477_cast_fp16")]; tensor var_5478_to_fp16 = const()[name = tensor("op_5478_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_5479_cast_fp16 = add(x = var_5477_cast_fp16, y = var_5478_to_fp16)[name = tensor("op_5479_cast_fp16")]; tensor denom_119_epsilon_0_to_fp16 = const()[name = tensor("denom_119_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_119_cast_fp16 = rsqrt(epsilon = denom_119_epsilon_0_to_fp16, x = var_5479_cast_fp16)[name = tensor("denom_119_cast_fp16")]; tensor out_119_cast_fp16 = mul(x = zero_mean_119_cast_fp16, y = denom_119_cast_fp16)[name = tensor("out_119_cast_fp16")]; tensor x_553_gamma_0_to_fp16 = const()[name = tensor("x_553_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391656448)))]; tensor x_553_beta_0_to_fp16 = const()[name = tensor("x_553_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391659072)))]; tensor x_553_epsilon_0_to_fp16 = const()[name = tensor("x_553_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_553_cast_fp16 = batch_norm(beta = x_553_beta_0_to_fp16, epsilon = x_553_epsilon_0_to_fp16, gamma = x_553_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_119_cast_fp16)[name = tensor("x_553_cast_fp16")]; tensor layers_19_fc1_input_shift_to_fp16 = const()[name = tensor("layers_19_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391661696)))]; tensor input_435_cast_fp16 = sub(x = x_553_cast_fp16, y = layers_19_fc1_input_shift_to_fp16)[name = tensor("input_435_cast_fp16")]; tensor var_5494 = const()[name = tensor("op_5494"), val = tensor([1, 1])]; tensor var_5496 = const()[name = tensor("op_5496"), val = tensor([1, 1])]; tensor x_555_pad_type_0 = const()[name = tensor("x_555_pad_type_0"), val = tensor("custom")]; tensor x_555_pad_0 = const()[name = tensor("x_555_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_19_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391664320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394941184))), name = tensor("layers_19_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_19_fc1_module_bias_to_fp16 = const()[name = tensor("layers_19_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394941312)))]; tensor x_555_cast_fp16 = conv(bias = layers_19_fc1_module_bias_to_fp16, dilations = var_5496, groups = var_5261, pad = x_555_pad_0, pad_type = x_555_pad_type_0, strides = var_5494, weight = layers_19_fc1_module_weight_to_fp16_palettized, x = input_435_cast_fp16)[name = tensor("x_555_cast_fp16")]; tensor layers_19_fc1_output_scale_to_fp16 = const()[name = tensor("layers_19_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394951616)))]; tensor input_437_cast_fp16 = mul(x = x_555_cast_fp16, y = layers_19_fc1_output_scale_to_fp16)[name = tensor("input_437_cast_fp16")]; tensor x_557_mode_0 = const()[name = tensor("x_557_mode_0"), val = tensor("EXACT")]; tensor x_557_cast_fp16 = gelu(mode = x_557_mode_0, x = input_437_cast_fp16)[name = tensor("x_557_cast_fp16")]; tensor layers_19_fc2_input_shift_to_fp16 = const()[name = tensor("layers_19_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394961920)))]; tensor input_439_cast_fp16 = sub(x = x_557_cast_fp16, y = layers_19_fc2_input_shift_to_fp16)[name = tensor("input_439_cast_fp16")]; tensor var_5507 = const()[name = tensor("op_5507"), val = tensor([1, 1])]; tensor var_5509 = const()[name = tensor("op_5509"), val = tensor([1, 1])]; tensor x_559_pad_type_0 = const()[name = tensor("x_559_pad_type_0"), val = tensor("custom")]; tensor x_559_pad_0 = const()[name = tensor("x_559_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_19_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394972224))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398249088))), name = tensor("layers_19_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_19_fc2_module_bias_to_fp16 = const()[name = tensor("layers_19_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398249216)))]; tensor x_559_cast_fp16 = conv(bias = layers_19_fc2_module_bias_to_fp16, dilations = var_5509, groups = var_5261, pad = x_559_pad_0, pad_type = x_559_pad_type_0, strides = var_5507, weight = layers_19_fc2_module_weight_to_fp16_palettized, x = input_439_cast_fp16)[name = tensor("x_559_cast_fp16")]; tensor layers_19_fc2_output_scale_to_fp16 = const()[name = tensor("layers_19_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398251840)))]; tensor hidden_states_41_cast_fp16 = mul(x = x_559_cast_fp16, y = layers_19_fc2_output_scale_to_fp16)[name = tensor("hidden_states_41_cast_fp16")]; tensor inputs_121_cast_fp16 = add(x = inputs_119_cast_fp16, y = hidden_states_41_cast_fp16)[name = tensor("inputs_121_cast_fp16")]; tensor var_5524 = const()[name = tensor("op_5524"), val = tensor(3)]; tensor var_5531 = const()[name = tensor("op_5531"), val = tensor(1)]; tensor var_5532 = const()[name = tensor("op_5532"), val = tensor(true)]; tensor var_5544 = const()[name = tensor("op_5544"), val = tensor([1])]; tensor channels_mean_121_cast_fp16 = reduce_mean(axes = var_5544, keep_dims = var_5532, x = inputs_121_cast_fp16)[name = tensor("channels_mean_121_cast_fp16")]; tensor zero_mean_121_cast_fp16 = sub(x = inputs_121_cast_fp16, y = channels_mean_121_cast_fp16)[name = tensor("zero_mean_121_cast_fp16")]; tensor zero_mean_sq_121_cast_fp16 = mul(x = zero_mean_121_cast_fp16, y = zero_mean_121_cast_fp16)[name = tensor("zero_mean_sq_121_cast_fp16")]; tensor var_5548 = const()[name = tensor("op_5548"), val = tensor([1])]; tensor var_5549_cast_fp16 = reduce_mean(axes = var_5548, keep_dims = var_5532, x = zero_mean_sq_121_cast_fp16)[name = tensor("op_5549_cast_fp16")]; tensor var_5550_to_fp16 = const()[name = tensor("op_5550_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_5551_cast_fp16 = add(x = var_5549_cast_fp16, y = var_5550_to_fp16)[name = tensor("op_5551_cast_fp16")]; tensor denom_121_epsilon_0_to_fp16 = const()[name = tensor("denom_121_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_121_cast_fp16 = rsqrt(epsilon = denom_121_epsilon_0_to_fp16, x = var_5551_cast_fp16)[name = tensor("denom_121_cast_fp16")]; tensor out_121_cast_fp16 = mul(x = zero_mean_121_cast_fp16, y = denom_121_cast_fp16)[name = tensor("out_121_cast_fp16")]; tensor obj_281_gamma_0_to_fp16 = const()[name = tensor("obj_281_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398254464)))]; tensor obj_281_beta_0_to_fp16 = const()[name = tensor("obj_281_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398257088)))]; tensor obj_281_epsilon_0_to_fp16 = const()[name = tensor("obj_281_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_281_cast_fp16 = batch_norm(beta = obj_281_beta_0_to_fp16, epsilon = obj_281_epsilon_0_to_fp16, gamma = obj_281_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_121_cast_fp16)[name = tensor("obj_281_cast_fp16")]; tensor layers_20_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398259712)))]; tensor input_441_cast_fp16 = sub(x = obj_281_cast_fp16, y = layers_20_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_441_cast_fp16")]; tensor var_5570 = const()[name = tensor("op_5570"), val = tensor([1, 1])]; tensor var_5572 = const()[name = tensor("op_5572"), val = tensor([1, 1])]; tensor x_561_pad_type_0 = const()[name = tensor("x_561_pad_type_0"), val = tensor("custom")]; tensor x_561_pad_0 = const()[name = tensor("x_561_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_20_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398262336))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(399081600))), name = tensor("layers_20_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_20_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(399081728)))]; tensor x_561_cast_fp16 = conv(bias = layers_20_self_attn_q_proj_module_bias_to_fp16, dilations = var_5572, groups = var_5531, pad = x_561_pad_0, pad_type = x_561_pad_type_0, strides = var_5570, weight = layers_20_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_441_cast_fp16)[name = tensor("x_561_cast_fp16")]; tensor layers_20_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(399084352)))]; tensor query_81_cast_fp16 = mul(x = x_561_cast_fp16, y = layers_20_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_81_cast_fp16")]; tensor var_5582 = const()[name = tensor("op_5582"), val = tensor([1, 1])]; tensor var_5584 = const()[name = tensor("op_5584"), val = tensor([1, 1])]; tensor x_563_pad_type_0 = const()[name = tensor("x_563_pad_type_0"), val = tensor("custom")]; tensor x_563_pad_0 = const()[name = tensor("x_563_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_20_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(399086976))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(399906240))), name = tensor("layers_20_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_20_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(399906368)))]; tensor x_563_cast_fp16 = conv(bias = layers_20_self_attn_k_proj_module_bias_to_fp16, dilations = var_5584, groups = var_5531, pad = x_563_pad_0, pad_type = x_563_pad_type_0, strides = var_5582, weight = layers_20_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_441_cast_fp16)[name = tensor("x_563_cast_fp16")]; tensor layers_20_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_20_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(399908992)))]; tensor current_key_41_cast_fp16 = mul(x = x_563_cast_fp16, y = layers_20_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_41_cast_fp16")]; tensor var_5594 = const()[name = tensor("op_5594"), val = tensor([1, 1])]; tensor var_5596 = const()[name = tensor("op_5596"), val = tensor([1, 1])]; tensor x_565_pad_type_0 = const()[name = tensor("x_565_pad_type_0"), val = tensor("custom")]; tensor x_565_pad_0 = const()[name = tensor("x_565_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_20_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(399911616))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(400730880))), name = tensor("layers_20_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_20_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(400731008)))]; tensor x_565_cast_fp16 = conv(bias = layers_20_self_attn_v_proj_module_bias_to_fp16, dilations = var_5596, groups = var_5531, pad = x_565_pad_0, pad_type = x_565_pad_type_0, strides = var_5594, weight = layers_20_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_441_cast_fp16)[name = tensor("x_565_cast_fp16")]; tensor layers_20_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(400733632)))]; tensor current_value_41_cast_fp16 = mul(x = x_565_cast_fp16, y = layers_20_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_41_cast_fp16")]; tensor var_5604_cast_fp16 = mul(x = current_key_41_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_5604_cast_fp16")]; tensor var_5606_cast_fp16 = mul(x = var_103_cast_fp16_20, y = var_257_cast_fp16)[name = tensor("op_5606_cast_fp16")]; tensor key_81_cast_fp16 = add(x = var_5604_cast_fp16, y = var_5606_cast_fp16)[name = tensor("key_81_cast_fp16")]; tensor var_5608_cast_fp16 = mul(x = current_value_41_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_5608_cast_fp16")]; tensor var_5610_cast_fp16 = mul(x = var_138_cast_fp16_20, y = var_257_cast_fp16)[name = tensor("op_5610_cast_fp16")]; tensor value_81_cast_fp16 = add(x = var_5608_cast_fp16, y = var_5610_cast_fp16)[name = tensor("value_81_cast_fp16")]; tensor var_5613 = const()[name = tensor("op_5613"), val = tensor([1, 20, 64, -1])]; tensor var_5614_cast_fp16 = reshape(shape = var_5613, x = query_81_cast_fp16)[name = tensor("op_5614_cast_fp16")]; tensor var_5615_to_fp16 = const()[name = tensor("op_5615_to_fp16"), val = tensor(0x1p-3)]; tensor var_5616_cast_fp16 = mul(x = var_5614_cast_fp16, y = var_5615_to_fp16)[name = tensor("op_5616_cast_fp16")]; tensor var_5617 = const()[name = tensor("op_5617"), val = tensor([1, 20, 64, -1])]; tensor var_5618_cast_fp16 = reshape(shape = var_5617, x = key_81_cast_fp16)[name = tensor("op_5618_cast_fp16")]; tensor mh_w_121_transpose_x_0 = const()[name = tensor("mh_w_121_transpose_x_0"), val = tensor(true)]; tensor mh_w_121_transpose_y_0 = const()[name = tensor("mh_w_121_transpose_y_0"), val = tensor(false)]; tensor mh_w_121_cast_fp16 = matmul(transpose_x = mh_w_121_transpose_x_0, transpose_y = mh_w_121_transpose_y_0, x = var_5616_cast_fp16, y = var_5618_cast_fp16)[name = tensor("mh_w_121_cast_fp16")]; tensor mh_w_123_cast_fp16 = add(x = mh_w_121_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_123_cast_fp16")]; tensor var_5626_cast_fp16 = softmax(axis = var_5524, x = mh_w_123_cast_fp16)[name = tensor("op_5626_cast_fp16")]; tensor var_5627 = const()[name = tensor("op_5627"), val = tensor([1, 20, 64, -1])]; tensor var_5628_cast_fp16 = reshape(shape = var_5627, x = value_81_cast_fp16)[name = tensor("op_5628_cast_fp16")]; tensor attn_81_transpose_x_0 = const()[name = tensor("attn_81_transpose_x_0"), val = tensor(false)]; tensor attn_81_transpose_y_0 = const()[name = tensor("attn_81_transpose_y_0"), val = tensor(true)]; tensor attn_81_cast_fp16 = matmul(transpose_x = attn_81_transpose_x_0, transpose_y = attn_81_transpose_y_0, x = var_5628_cast_fp16, y = var_5626_cast_fp16)[name = tensor("attn_81_cast_fp16")]; tensor var_5631 = const()[name = tensor("op_5631"), val = tensor([1, 1280, 1, -1])]; tensor x_567_cast_fp16 = reshape(shape = var_5631, x = attn_81_cast_fp16)[name = tensor("x_567_cast_fp16")]; tensor layers_20_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(400736256)))]; tensor input_447_cast_fp16 = sub(x = x_567_cast_fp16, y = layers_20_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_447_cast_fp16")]; tensor var_5639 = const()[name = tensor("op_5639"), val = tensor([1, 1])]; tensor var_5641 = const()[name = tensor("op_5641"), val = tensor([1, 1])]; tensor x_569_pad_type_0 = const()[name = tensor("x_569_pad_type_0"), val = tensor("custom")]; tensor x_569_pad_0 = const()[name = tensor("x_569_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_20_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(400738880))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(401558144))), name = tensor("layers_20_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_20_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(401558272)))]; tensor x_569_cast_fp16 = conv(bias = layers_20_self_attn_o_proj_module_bias_to_fp16, dilations = var_5641, groups = var_5531, pad = x_569_pad_0, pad_type = x_569_pad_type_0, strides = var_5639, weight = layers_20_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = tensor("x_569_cast_fp16")]; tensor layers_20_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(401560896)))]; tensor obj_287_cast_fp16 = mul(x = x_569_cast_fp16, y = layers_20_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_287_cast_fp16")]; tensor inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = obj_287_cast_fp16)[name = tensor("inputs_123_cast_fp16")]; tensor var_5652 = const()[name = tensor("op_5652"), val = tensor([1])]; tensor channels_mean_123_cast_fp16 = reduce_mean(axes = var_5652, keep_dims = var_5532, x = inputs_123_cast_fp16)[name = tensor("channels_mean_123_cast_fp16")]; tensor zero_mean_123_cast_fp16 = sub(x = inputs_123_cast_fp16, y = channels_mean_123_cast_fp16)[name = tensor("zero_mean_123_cast_fp16")]; tensor zero_mean_sq_123_cast_fp16 = mul(x = zero_mean_123_cast_fp16, y = zero_mean_123_cast_fp16)[name = tensor("zero_mean_sq_123_cast_fp16")]; tensor var_5656 = const()[name = tensor("op_5656"), val = tensor([1])]; tensor var_5657_cast_fp16 = reduce_mean(axes = var_5656, keep_dims = var_5532, x = zero_mean_sq_123_cast_fp16)[name = tensor("op_5657_cast_fp16")]; tensor var_5658_to_fp16 = const()[name = tensor("op_5658_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_5659_cast_fp16 = add(x = var_5657_cast_fp16, y = var_5658_to_fp16)[name = tensor("op_5659_cast_fp16")]; tensor denom_123_epsilon_0_to_fp16 = const()[name = tensor("denom_123_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_123_cast_fp16 = rsqrt(epsilon = denom_123_epsilon_0_to_fp16, x = var_5659_cast_fp16)[name = tensor("denom_123_cast_fp16")]; tensor out_123_cast_fp16 = mul(x = zero_mean_123_cast_fp16, y = denom_123_cast_fp16)[name = tensor("out_123_cast_fp16")]; tensor obj_289_gamma_0_to_fp16 = const()[name = tensor("obj_289_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(401563520)))]; tensor obj_289_beta_0_to_fp16 = const()[name = tensor("obj_289_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(401566144)))]; tensor obj_289_epsilon_0_to_fp16 = const()[name = tensor("obj_289_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_289_cast_fp16 = batch_norm(beta = obj_289_beta_0_to_fp16, epsilon = obj_289_epsilon_0_to_fp16, gamma = obj_289_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_123_cast_fp16)[name = tensor("obj_289_cast_fp16")]; tensor layers_20_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_20_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(401568768)))]; tensor input_449_cast_fp16 = sub(x = obj_289_cast_fp16, y = layers_20_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_449_cast_fp16")]; tensor var_5678 = const()[name = tensor("op_5678"), val = tensor([1, 1])]; tensor var_5680 = const()[name = tensor("op_5680"), val = tensor([1, 1])]; tensor x_571_pad_type_0 = const()[name = tensor("x_571_pad_type_0"), val = tensor("custom")]; tensor x_571_pad_0 = const()[name = tensor("x_571_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_20_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(401571392))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402390656))), name = tensor("layers_20_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_20_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402390784)))]; tensor x_571_cast_fp16 = conv(bias = layers_20_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_5680, groups = var_5531, pad = x_571_pad_0, pad_type = x_571_pad_type_0, strides = var_5678, weight = layers_20_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_449_cast_fp16)[name = tensor("x_571_cast_fp16")]; tensor layers_20_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_20_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402393408)))]; tensor query_83_cast_fp16 = mul(x = x_571_cast_fp16, y = layers_20_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_83_cast_fp16")]; tensor var_5690 = const()[name = tensor("op_5690"), val = tensor([1, 1])]; tensor var_5692 = const()[name = tensor("op_5692"), val = tensor([1, 1])]; tensor x_573_pad_type_0 = const()[name = tensor("x_573_pad_type_0"), val = tensor("custom")]; tensor x_573_pad_0 = const()[name = tensor("x_573_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_20_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402396032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403215296))), name = tensor("layers_20_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_20_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403215424)))]; tensor x_573_cast_fp16 = conv(bias = layers_20_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_5692, groups = var_5531, pad = x_573_pad_0, pad_type = x_573_pad_type_0, strides = var_5690, weight = layers_20_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_573_cast_fp16")]; tensor layers_20_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_20_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403218048)))]; tensor key_83_cast_fp16 = mul(x = x_573_cast_fp16, y = layers_20_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_83_cast_fp16")]; tensor var_5702 = const()[name = tensor("op_5702"), val = tensor([1, 1])]; tensor var_5704 = const()[name = tensor("op_5704"), val = tensor([1, 1])]; tensor x_575_pad_type_0 = const()[name = tensor("x_575_pad_type_0"), val = tensor("custom")]; tensor x_575_pad_0 = const()[name = tensor("x_575_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_20_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403220672))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404039936))), name = tensor("layers_20_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_20_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404040064)))]; tensor x_575_cast_fp16 = conv(bias = layers_20_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_5704, groups = var_5531, pad = x_575_pad_0, pad_type = x_575_pad_type_0, strides = var_5702, weight = layers_20_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_575_cast_fp16")]; tensor layers_20_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_20_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404042688)))]; tensor value_83_cast_fp16 = mul(x = x_575_cast_fp16, y = layers_20_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_83_cast_fp16")]; tensor var_5709 = const()[name = tensor("op_5709"), val = tensor([1, 20, 64, -1])]; tensor var_5710_cast_fp16 = reshape(shape = var_5709, x = query_83_cast_fp16)[name = tensor("op_5710_cast_fp16")]; tensor var_5711_to_fp16 = const()[name = tensor("op_5711_to_fp16"), val = tensor(0x1p-3)]; tensor var_5712_cast_fp16 = mul(x = var_5710_cast_fp16, y = var_5711_to_fp16)[name = tensor("op_5712_cast_fp16")]; tensor var_5713 = const()[name = tensor("op_5713"), val = tensor([1, 20, 64, -1])]; tensor var_5714_cast_fp16 = reshape(shape = var_5713, x = key_83_cast_fp16)[name = tensor("op_5714_cast_fp16")]; tensor mh_w_125_transpose_x_0 = const()[name = tensor("mh_w_125_transpose_x_0"), val = tensor(true)]; tensor mh_w_125_transpose_y_0 = const()[name = tensor("mh_w_125_transpose_y_0"), val = tensor(false)]; tensor mh_w_125_cast_fp16 = matmul(transpose_x = mh_w_125_transpose_x_0, transpose_y = mh_w_125_transpose_y_0, x = var_5712_cast_fp16, y = var_5714_cast_fp16)[name = tensor("mh_w_125_cast_fp16")]; tensor obj_293_cast_fp16 = softmax(axis = var_5524, x = mh_w_125_cast_fp16)[name = tensor("obj_293_cast_fp16")]; tensor var_5718 = const()[name = tensor("op_5718"), val = tensor([1, 20, 64, -1])]; tensor var_5719_cast_fp16 = reshape(shape = var_5718, x = value_83_cast_fp16)[name = tensor("op_5719_cast_fp16")]; tensor attn_83_transpose_x_0 = const()[name = tensor("attn_83_transpose_x_0"), val = tensor(false)]; tensor attn_83_transpose_y_0 = const()[name = tensor("attn_83_transpose_y_0"), val = tensor(true)]; tensor attn_83_cast_fp16 = matmul(transpose_x = attn_83_transpose_x_0, transpose_y = attn_83_transpose_y_0, x = var_5719_cast_fp16, y = obj_293_cast_fp16)[name = tensor("attn_83_cast_fp16")]; tensor var_5722 = const()[name = tensor("op_5722"), val = tensor([1, 1280, 1, -1])]; tensor x_577_cast_fp16 = reshape(shape = var_5722, x = attn_83_cast_fp16)[name = tensor("x_577_cast_fp16")]; tensor layers_20_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_20_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404045312)))]; tensor input_455_cast_fp16 = sub(x = x_577_cast_fp16, y = layers_20_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_455_cast_fp16")]; tensor var_5730 = const()[name = tensor("op_5730"), val = tensor([1, 1])]; tensor var_5732 = const()[name = tensor("op_5732"), val = tensor([1, 1])]; tensor x_579_pad_type_0 = const()[name = tensor("x_579_pad_type_0"), val = tensor("custom")]; tensor x_579_pad_0 = const()[name = tensor("x_579_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_20_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404047936))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404867200))), name = tensor("layers_20_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_20_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404867328)))]; tensor x_579_cast_fp16 = conv(bias = layers_20_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_5732, groups = var_5531, pad = x_579_pad_0, pad_type = x_579_pad_type_0, strides = var_5730, weight = layers_20_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = tensor("x_579_cast_fp16")]; tensor layers_20_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_20_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404869952)))]; tensor obj_291_cast_fp16 = mul(x = x_579_cast_fp16, y = layers_20_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_291_cast_fp16")]; tensor inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = obj_291_cast_fp16)[name = tensor("inputs_125_cast_fp16")]; tensor var_5739 = const()[name = tensor("op_5739"), val = tensor([1])]; tensor channels_mean_125_cast_fp16 = reduce_mean(axes = var_5739, keep_dims = var_5532, x = inputs_125_cast_fp16)[name = tensor("channels_mean_125_cast_fp16")]; tensor zero_mean_125_cast_fp16 = sub(x = inputs_125_cast_fp16, y = channels_mean_125_cast_fp16)[name = tensor("zero_mean_125_cast_fp16")]; tensor zero_mean_sq_125_cast_fp16 = mul(x = zero_mean_125_cast_fp16, y = zero_mean_125_cast_fp16)[name = tensor("zero_mean_sq_125_cast_fp16")]; tensor var_5743 = const()[name = tensor("op_5743"), val = tensor([1])]; tensor var_5744_cast_fp16 = reduce_mean(axes = var_5743, keep_dims = var_5532, x = zero_mean_sq_125_cast_fp16)[name = tensor("op_5744_cast_fp16")]; tensor var_5745_to_fp16 = const()[name = tensor("op_5745_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_5746_cast_fp16 = add(x = var_5744_cast_fp16, y = var_5745_to_fp16)[name = tensor("op_5746_cast_fp16")]; tensor denom_125_epsilon_0_to_fp16 = const()[name = tensor("denom_125_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_125_cast_fp16 = rsqrt(epsilon = denom_125_epsilon_0_to_fp16, x = var_5746_cast_fp16)[name = tensor("denom_125_cast_fp16")]; tensor out_125_cast_fp16 = mul(x = zero_mean_125_cast_fp16, y = denom_125_cast_fp16)[name = tensor("out_125_cast_fp16")]; tensor x_581_gamma_0_to_fp16 = const()[name = tensor("x_581_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404872576)))]; tensor x_581_beta_0_to_fp16 = const()[name = tensor("x_581_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404875200)))]; tensor x_581_epsilon_0_to_fp16 = const()[name = tensor("x_581_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_581_cast_fp16 = batch_norm(beta = x_581_beta_0_to_fp16, epsilon = x_581_epsilon_0_to_fp16, gamma = x_581_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_125_cast_fp16)[name = tensor("x_581_cast_fp16")]; tensor layers_20_fc1_input_shift_to_fp16 = const()[name = tensor("layers_20_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404877824)))]; tensor input_457_cast_fp16 = sub(x = x_581_cast_fp16, y = layers_20_fc1_input_shift_to_fp16)[name = tensor("input_457_cast_fp16")]; tensor var_5761 = const()[name = tensor("op_5761"), val = tensor([1, 1])]; tensor var_5763 = const()[name = tensor("op_5763"), val = tensor([1, 1])]; tensor x_583_pad_type_0 = const()[name = tensor("x_583_pad_type_0"), val = tensor("custom")]; tensor x_583_pad_0 = const()[name = tensor("x_583_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_20_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404880448))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408157312))), name = tensor("layers_20_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_20_fc1_module_bias_to_fp16 = const()[name = tensor("layers_20_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408157440)))]; tensor x_583_cast_fp16 = conv(bias = layers_20_fc1_module_bias_to_fp16, dilations = var_5763, groups = var_5531, pad = x_583_pad_0, pad_type = x_583_pad_type_0, strides = var_5761, weight = layers_20_fc1_module_weight_to_fp16_palettized, x = input_457_cast_fp16)[name = tensor("x_583_cast_fp16")]; tensor layers_20_fc1_output_scale_to_fp16 = const()[name = tensor("layers_20_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408167744)))]; tensor input_459_cast_fp16 = mul(x = x_583_cast_fp16, y = layers_20_fc1_output_scale_to_fp16)[name = tensor("input_459_cast_fp16")]; tensor x_585_mode_0 = const()[name = tensor("x_585_mode_0"), val = tensor("EXACT")]; tensor x_585_cast_fp16 = gelu(mode = x_585_mode_0, x = input_459_cast_fp16)[name = tensor("x_585_cast_fp16")]; tensor layers_20_fc2_input_shift_to_fp16 = const()[name = tensor("layers_20_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408178048)))]; tensor input_461_cast_fp16 = sub(x = x_585_cast_fp16, y = layers_20_fc2_input_shift_to_fp16)[name = tensor("input_461_cast_fp16")]; tensor var_5774 = const()[name = tensor("op_5774"), val = tensor([1, 1])]; tensor var_5776 = const()[name = tensor("op_5776"), val = tensor([1, 1])]; tensor x_587_pad_type_0 = const()[name = tensor("x_587_pad_type_0"), val = tensor("custom")]; tensor x_587_pad_0 = const()[name = tensor("x_587_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_20_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408188352))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411465216))), name = tensor("layers_20_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_20_fc2_module_bias_to_fp16 = const()[name = tensor("layers_20_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411465344)))]; tensor x_587_cast_fp16 = conv(bias = layers_20_fc2_module_bias_to_fp16, dilations = var_5776, groups = var_5531, pad = x_587_pad_0, pad_type = x_587_pad_type_0, strides = var_5774, weight = layers_20_fc2_module_weight_to_fp16_palettized, x = input_461_cast_fp16)[name = tensor("x_587_cast_fp16")]; tensor layers_20_fc2_output_scale_to_fp16 = const()[name = tensor("layers_20_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411467968)))]; tensor hidden_states_43_cast_fp16 = mul(x = x_587_cast_fp16, y = layers_20_fc2_output_scale_to_fp16)[name = tensor("hidden_states_43_cast_fp16")]; tensor inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = hidden_states_43_cast_fp16)[name = tensor("inputs_127_cast_fp16")]; tensor var_5790 = const()[name = tensor("op_5790"), val = tensor(3)]; tensor var_5797 = const()[name = tensor("op_5797"), val = tensor(1)]; tensor var_5798 = const()[name = tensor("op_5798"), val = tensor(true)]; tensor var_5810 = const()[name = tensor("op_5810"), val = tensor([1])]; tensor channels_mean_127_cast_fp16 = reduce_mean(axes = var_5810, keep_dims = var_5798, x = inputs_127_cast_fp16)[name = tensor("channels_mean_127_cast_fp16")]; tensor zero_mean_127_cast_fp16 = sub(x = inputs_127_cast_fp16, y = channels_mean_127_cast_fp16)[name = tensor("zero_mean_127_cast_fp16")]; tensor zero_mean_sq_127_cast_fp16 = mul(x = zero_mean_127_cast_fp16, y = zero_mean_127_cast_fp16)[name = tensor("zero_mean_sq_127_cast_fp16")]; tensor var_5814 = const()[name = tensor("op_5814"), val = tensor([1])]; tensor var_5815_cast_fp16 = reduce_mean(axes = var_5814, keep_dims = var_5798, x = zero_mean_sq_127_cast_fp16)[name = tensor("op_5815_cast_fp16")]; tensor var_5816_to_fp16 = const()[name = tensor("op_5816_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_5817_cast_fp16 = add(x = var_5815_cast_fp16, y = var_5816_to_fp16)[name = tensor("op_5817_cast_fp16")]; tensor denom_127_epsilon_0_to_fp16 = const()[name = tensor("denom_127_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_127_cast_fp16 = rsqrt(epsilon = denom_127_epsilon_0_to_fp16, x = var_5817_cast_fp16)[name = tensor("denom_127_cast_fp16")]; tensor out_127_cast_fp16 = mul(x = zero_mean_127_cast_fp16, y = denom_127_cast_fp16)[name = tensor("out_127_cast_fp16")]; tensor obj_295_gamma_0_to_fp16 = const()[name = tensor("obj_295_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411470592)))]; tensor obj_295_beta_0_to_fp16 = const()[name = tensor("obj_295_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411473216)))]; tensor obj_295_epsilon_0_to_fp16 = const()[name = tensor("obj_295_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_295_cast_fp16 = batch_norm(beta = obj_295_beta_0_to_fp16, epsilon = obj_295_epsilon_0_to_fp16, gamma = obj_295_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_127_cast_fp16)[name = tensor("obj_295_cast_fp16")]; tensor layers_21_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411475840)))]; tensor input_463_cast_fp16 = sub(x = obj_295_cast_fp16, y = layers_21_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_463_cast_fp16")]; tensor var_5836 = const()[name = tensor("op_5836"), val = tensor([1, 1])]; tensor var_5838 = const()[name = tensor("op_5838"), val = tensor([1, 1])]; tensor x_589_pad_type_0 = const()[name = tensor("x_589_pad_type_0"), val = tensor("custom")]; tensor x_589_pad_0 = const()[name = tensor("x_589_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_21_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411478464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412297728))), name = tensor("layers_21_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_21_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412297856)))]; tensor x_589_cast_fp16 = conv(bias = layers_21_self_attn_q_proj_module_bias_to_fp16, dilations = var_5838, groups = var_5797, pad = x_589_pad_0, pad_type = x_589_pad_type_0, strides = var_5836, weight = layers_21_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_463_cast_fp16)[name = tensor("x_589_cast_fp16")]; tensor layers_21_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412300480)))]; tensor query_85_cast_fp16 = mul(x = x_589_cast_fp16, y = layers_21_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_85_cast_fp16")]; tensor var_5848 = const()[name = tensor("op_5848"), val = tensor([1, 1])]; tensor var_5850 = const()[name = tensor("op_5850"), val = tensor([1, 1])]; tensor x_591_pad_type_0 = const()[name = tensor("x_591_pad_type_0"), val = tensor("custom")]; tensor x_591_pad_0 = const()[name = tensor("x_591_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_21_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412303104))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(413122368))), name = tensor("layers_21_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_21_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(413122496)))]; tensor x_591_cast_fp16 = conv(bias = layers_21_self_attn_k_proj_module_bias_to_fp16, dilations = var_5850, groups = var_5797, pad = x_591_pad_0, pad_type = x_591_pad_type_0, strides = var_5848, weight = layers_21_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_463_cast_fp16)[name = tensor("x_591_cast_fp16")]; tensor layers_21_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_21_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(413125120)))]; tensor current_key_43_cast_fp16 = mul(x = x_591_cast_fp16, y = layers_21_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_43_cast_fp16")]; tensor var_5860 = const()[name = tensor("op_5860"), val = tensor([1, 1])]; tensor var_5862 = const()[name = tensor("op_5862"), val = tensor([1, 1])]; tensor x_593_pad_type_0 = const()[name = tensor("x_593_pad_type_0"), val = tensor("custom")]; tensor x_593_pad_0 = const()[name = tensor("x_593_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_21_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(413127744))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(413947008))), name = tensor("layers_21_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_21_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(413947136)))]; tensor x_593_cast_fp16 = conv(bias = layers_21_self_attn_v_proj_module_bias_to_fp16, dilations = var_5862, groups = var_5797, pad = x_593_pad_0, pad_type = x_593_pad_type_0, strides = var_5860, weight = layers_21_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_463_cast_fp16)[name = tensor("x_593_cast_fp16")]; tensor layers_21_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(413949760)))]; tensor current_value_43_cast_fp16 = mul(x = x_593_cast_fp16, y = layers_21_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_43_cast_fp16")]; tensor var_5870_cast_fp16 = mul(x = current_key_43_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_5870_cast_fp16")]; tensor var_5872_cast_fp16 = mul(x = var_103_cast_fp16_21, y = var_257_cast_fp16)[name = tensor("op_5872_cast_fp16")]; tensor key_85_cast_fp16 = add(x = var_5870_cast_fp16, y = var_5872_cast_fp16)[name = tensor("key_85_cast_fp16")]; tensor var_5874_cast_fp16 = mul(x = current_value_43_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_5874_cast_fp16")]; tensor var_5876_cast_fp16 = mul(x = var_138_cast_fp16_21, y = var_257_cast_fp16)[name = tensor("op_5876_cast_fp16")]; tensor value_85_cast_fp16 = add(x = var_5874_cast_fp16, y = var_5876_cast_fp16)[name = tensor("value_85_cast_fp16")]; tensor var_5879 = const()[name = tensor("op_5879"), val = tensor([1, 20, 64, -1])]; tensor var_5880_cast_fp16 = reshape(shape = var_5879, x = query_85_cast_fp16)[name = tensor("op_5880_cast_fp16")]; tensor var_5881_to_fp16 = const()[name = tensor("op_5881_to_fp16"), val = tensor(0x1p-3)]; tensor var_5882_cast_fp16 = mul(x = var_5880_cast_fp16, y = var_5881_to_fp16)[name = tensor("op_5882_cast_fp16")]; tensor var_5883 = const()[name = tensor("op_5883"), val = tensor([1, 20, 64, -1])]; tensor var_5884_cast_fp16 = reshape(shape = var_5883, x = key_85_cast_fp16)[name = tensor("op_5884_cast_fp16")]; tensor mh_w_127_transpose_x_0 = const()[name = tensor("mh_w_127_transpose_x_0"), val = tensor(true)]; tensor mh_w_127_transpose_y_0 = const()[name = tensor("mh_w_127_transpose_y_0"), val = tensor(false)]; tensor mh_w_127_cast_fp16 = matmul(transpose_x = mh_w_127_transpose_x_0, transpose_y = mh_w_127_transpose_y_0, x = var_5882_cast_fp16, y = var_5884_cast_fp16)[name = tensor("mh_w_127_cast_fp16")]; tensor mh_w_129_cast_fp16 = add(x = mh_w_127_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_129_cast_fp16")]; tensor var_5892_cast_fp16 = softmax(axis = var_5790, x = mh_w_129_cast_fp16)[name = tensor("op_5892_cast_fp16")]; tensor var_5893 = const()[name = tensor("op_5893"), val = tensor([1, 20, 64, -1])]; tensor var_5894_cast_fp16 = reshape(shape = var_5893, x = value_85_cast_fp16)[name = tensor("op_5894_cast_fp16")]; tensor attn_85_transpose_x_0 = const()[name = tensor("attn_85_transpose_x_0"), val = tensor(false)]; tensor attn_85_transpose_y_0 = const()[name = tensor("attn_85_transpose_y_0"), val = tensor(true)]; tensor attn_85_cast_fp16 = matmul(transpose_x = attn_85_transpose_x_0, transpose_y = attn_85_transpose_y_0, x = var_5894_cast_fp16, y = var_5892_cast_fp16)[name = tensor("attn_85_cast_fp16")]; tensor var_5897 = const()[name = tensor("op_5897"), val = tensor([1, 1280, 1, -1])]; tensor x_595_cast_fp16 = reshape(shape = var_5897, x = attn_85_cast_fp16)[name = tensor("x_595_cast_fp16")]; tensor layers_21_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(413952384)))]; tensor input_469_cast_fp16 = sub(x = x_595_cast_fp16, y = layers_21_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_469_cast_fp16")]; tensor var_5905 = const()[name = tensor("op_5905"), val = tensor([1, 1])]; tensor var_5907 = const()[name = tensor("op_5907"), val = tensor([1, 1])]; tensor x_597_pad_type_0 = const()[name = tensor("x_597_pad_type_0"), val = tensor("custom")]; tensor x_597_pad_0 = const()[name = tensor("x_597_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_21_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(413955008))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414774272))), name = tensor("layers_21_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_21_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414774400)))]; tensor x_597_cast_fp16 = conv(bias = layers_21_self_attn_o_proj_module_bias_to_fp16, dilations = var_5907, groups = var_5797, pad = x_597_pad_0, pad_type = x_597_pad_type_0, strides = var_5905, weight = layers_21_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_469_cast_fp16)[name = tensor("x_597_cast_fp16")]; tensor layers_21_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414777024)))]; tensor obj_301_cast_fp16 = mul(x = x_597_cast_fp16, y = layers_21_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_301_cast_fp16")]; tensor inputs_129_cast_fp16 = add(x = inputs_127_cast_fp16, y = obj_301_cast_fp16)[name = tensor("inputs_129_cast_fp16")]; tensor var_5918 = const()[name = tensor("op_5918"), val = tensor([1])]; tensor channels_mean_129_cast_fp16 = reduce_mean(axes = var_5918, keep_dims = var_5798, x = inputs_129_cast_fp16)[name = tensor("channels_mean_129_cast_fp16")]; tensor zero_mean_129_cast_fp16 = sub(x = inputs_129_cast_fp16, y = channels_mean_129_cast_fp16)[name = tensor("zero_mean_129_cast_fp16")]; tensor zero_mean_sq_129_cast_fp16 = mul(x = zero_mean_129_cast_fp16, y = zero_mean_129_cast_fp16)[name = tensor("zero_mean_sq_129_cast_fp16")]; tensor var_5922 = const()[name = tensor("op_5922"), val = tensor([1])]; tensor var_5923_cast_fp16 = reduce_mean(axes = var_5922, keep_dims = var_5798, x = zero_mean_sq_129_cast_fp16)[name = tensor("op_5923_cast_fp16")]; tensor var_5924_to_fp16 = const()[name = tensor("op_5924_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_5925_cast_fp16 = add(x = var_5923_cast_fp16, y = var_5924_to_fp16)[name = tensor("op_5925_cast_fp16")]; tensor denom_129_epsilon_0_to_fp16 = const()[name = tensor("denom_129_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_129_cast_fp16 = rsqrt(epsilon = denom_129_epsilon_0_to_fp16, x = var_5925_cast_fp16)[name = tensor("denom_129_cast_fp16")]; tensor out_129_cast_fp16 = mul(x = zero_mean_129_cast_fp16, y = denom_129_cast_fp16)[name = tensor("out_129_cast_fp16")]; tensor obj_303_gamma_0_to_fp16 = const()[name = tensor("obj_303_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414779648)))]; tensor obj_303_beta_0_to_fp16 = const()[name = tensor("obj_303_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414782272)))]; tensor obj_303_epsilon_0_to_fp16 = const()[name = tensor("obj_303_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_303_cast_fp16 = batch_norm(beta = obj_303_beta_0_to_fp16, epsilon = obj_303_epsilon_0_to_fp16, gamma = obj_303_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_129_cast_fp16)[name = tensor("obj_303_cast_fp16")]; tensor layers_21_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_21_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414784896)))]; tensor input_471_cast_fp16 = sub(x = obj_303_cast_fp16, y = layers_21_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_471_cast_fp16")]; tensor var_5944 = const()[name = tensor("op_5944"), val = tensor([1, 1])]; tensor var_5946 = const()[name = tensor("op_5946"), val = tensor([1, 1])]; tensor x_599_pad_type_0 = const()[name = tensor("x_599_pad_type_0"), val = tensor("custom")]; tensor x_599_pad_0 = const()[name = tensor("x_599_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_21_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414787520))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(415606784))), name = tensor("layers_21_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_21_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(415606912)))]; tensor x_599_cast_fp16 = conv(bias = layers_21_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_5946, groups = var_5797, pad = x_599_pad_0, pad_type = x_599_pad_type_0, strides = var_5944, weight = layers_21_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_471_cast_fp16)[name = tensor("x_599_cast_fp16")]; tensor layers_21_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_21_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(415609536)))]; tensor query_87_cast_fp16 = mul(x = x_599_cast_fp16, y = layers_21_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_87_cast_fp16")]; tensor var_5956 = const()[name = tensor("op_5956"), val = tensor([1, 1])]; tensor var_5958 = const()[name = tensor("op_5958"), val = tensor([1, 1])]; tensor x_601_pad_type_0 = const()[name = tensor("x_601_pad_type_0"), val = tensor("custom")]; tensor x_601_pad_0 = const()[name = tensor("x_601_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_21_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(415612160))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416431424))), name = tensor("layers_21_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_21_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416431552)))]; tensor x_601_cast_fp16 = conv(bias = layers_21_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_5958, groups = var_5797, pad = x_601_pad_0, pad_type = x_601_pad_type_0, strides = var_5956, weight = layers_21_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_601_cast_fp16")]; tensor layers_21_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_21_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416434176)))]; tensor key_87_cast_fp16 = mul(x = x_601_cast_fp16, y = layers_21_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_87_cast_fp16")]; tensor var_5968 = const()[name = tensor("op_5968"), val = tensor([1, 1])]; tensor var_5970 = const()[name = tensor("op_5970"), val = tensor([1, 1])]; tensor x_603_pad_type_0 = const()[name = tensor("x_603_pad_type_0"), val = tensor("custom")]; tensor x_603_pad_0 = const()[name = tensor("x_603_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_21_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416436800))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417256064))), name = tensor("layers_21_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_21_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417256192)))]; tensor x_603_cast_fp16 = conv(bias = layers_21_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_5970, groups = var_5797, pad = x_603_pad_0, pad_type = x_603_pad_type_0, strides = var_5968, weight = layers_21_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_603_cast_fp16")]; tensor layers_21_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_21_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417258816)))]; tensor value_87_cast_fp16 = mul(x = x_603_cast_fp16, y = layers_21_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_87_cast_fp16")]; tensor var_5975 = const()[name = tensor("op_5975"), val = tensor([1, 20, 64, -1])]; tensor var_5976_cast_fp16 = reshape(shape = var_5975, x = query_87_cast_fp16)[name = tensor("op_5976_cast_fp16")]; tensor var_5977_to_fp16 = const()[name = tensor("op_5977_to_fp16"), val = tensor(0x1p-3)]; tensor var_5978_cast_fp16 = mul(x = var_5976_cast_fp16, y = var_5977_to_fp16)[name = tensor("op_5978_cast_fp16")]; tensor var_5979 = const()[name = tensor("op_5979"), val = tensor([1, 20, 64, -1])]; tensor var_5980_cast_fp16 = reshape(shape = var_5979, x = key_87_cast_fp16)[name = tensor("op_5980_cast_fp16")]; tensor mh_w_131_transpose_x_0 = const()[name = tensor("mh_w_131_transpose_x_0"), val = tensor(true)]; tensor mh_w_131_transpose_y_0 = const()[name = tensor("mh_w_131_transpose_y_0"), val = tensor(false)]; tensor mh_w_131_cast_fp16 = matmul(transpose_x = mh_w_131_transpose_x_0, transpose_y = mh_w_131_transpose_y_0, x = var_5978_cast_fp16, y = var_5980_cast_fp16)[name = tensor("mh_w_131_cast_fp16")]; tensor obj_307_cast_fp16 = softmax(axis = var_5790, x = mh_w_131_cast_fp16)[name = tensor("obj_307_cast_fp16")]; tensor var_5984 = const()[name = tensor("op_5984"), val = tensor([1, 20, 64, -1])]; tensor var_5985_cast_fp16 = reshape(shape = var_5984, x = value_87_cast_fp16)[name = tensor("op_5985_cast_fp16")]; tensor attn_87_transpose_x_0 = const()[name = tensor("attn_87_transpose_x_0"), val = tensor(false)]; tensor attn_87_transpose_y_0 = const()[name = tensor("attn_87_transpose_y_0"), val = tensor(true)]; tensor attn_87_cast_fp16 = matmul(transpose_x = attn_87_transpose_x_0, transpose_y = attn_87_transpose_y_0, x = var_5985_cast_fp16, y = obj_307_cast_fp16)[name = tensor("attn_87_cast_fp16")]; tensor var_5988 = const()[name = tensor("op_5988"), val = tensor([1, 1280, 1, -1])]; tensor x_605_cast_fp16 = reshape(shape = var_5988, x = attn_87_cast_fp16)[name = tensor("x_605_cast_fp16")]; tensor layers_21_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_21_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417261440)))]; tensor input_477_cast_fp16 = sub(x = x_605_cast_fp16, y = layers_21_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_477_cast_fp16")]; tensor var_5996 = const()[name = tensor("op_5996"), val = tensor([1, 1])]; tensor var_5998 = const()[name = tensor("op_5998"), val = tensor([1, 1])]; tensor x_607_pad_type_0 = const()[name = tensor("x_607_pad_type_0"), val = tensor("custom")]; tensor x_607_pad_0 = const()[name = tensor("x_607_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_21_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417264064))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(418083328))), name = tensor("layers_21_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_21_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(418083456)))]; tensor x_607_cast_fp16 = conv(bias = layers_21_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_5998, groups = var_5797, pad = x_607_pad_0, pad_type = x_607_pad_type_0, strides = var_5996, weight = layers_21_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_477_cast_fp16)[name = tensor("x_607_cast_fp16")]; tensor layers_21_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_21_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(418086080)))]; tensor obj_305_cast_fp16 = mul(x = x_607_cast_fp16, y = layers_21_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_305_cast_fp16")]; tensor inputs_131_cast_fp16 = add(x = inputs_129_cast_fp16, y = obj_305_cast_fp16)[name = tensor("inputs_131_cast_fp16")]; tensor var_6008 = const()[name = tensor("op_6008"), val = tensor([1])]; tensor channels_mean_131_cast_fp16 = reduce_mean(axes = var_6008, keep_dims = var_5798, x = inputs_131_cast_fp16)[name = tensor("channels_mean_131_cast_fp16")]; tensor zero_mean_131_cast_fp16 = sub(x = inputs_131_cast_fp16, y = channels_mean_131_cast_fp16)[name = tensor("zero_mean_131_cast_fp16")]; tensor zero_mean_sq_131_cast_fp16 = mul(x = zero_mean_131_cast_fp16, y = zero_mean_131_cast_fp16)[name = tensor("zero_mean_sq_131_cast_fp16")]; tensor var_6012 = const()[name = tensor("op_6012"), val = tensor([1])]; tensor var_6013_cast_fp16 = reduce_mean(axes = var_6012, keep_dims = var_5798, x = zero_mean_sq_131_cast_fp16)[name = tensor("op_6013_cast_fp16")]; tensor var_6014_to_fp16 = const()[name = tensor("op_6014_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_6015_cast_fp16 = add(x = var_6013_cast_fp16, y = var_6014_to_fp16)[name = tensor("op_6015_cast_fp16")]; tensor denom_131_epsilon_0_to_fp16 = const()[name = tensor("denom_131_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_131_cast_fp16 = rsqrt(epsilon = denom_131_epsilon_0_to_fp16, x = var_6015_cast_fp16)[name = tensor("denom_131_cast_fp16")]; tensor out_131_cast_fp16 = mul(x = zero_mean_131_cast_fp16, y = denom_131_cast_fp16)[name = tensor("out_131_cast_fp16")]; tensor x_609_gamma_0_to_fp16 = const()[name = tensor("x_609_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(418088704)))]; tensor x_609_beta_0_to_fp16 = const()[name = tensor("x_609_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(418091328)))]; tensor x_609_epsilon_0_to_fp16 = const()[name = tensor("x_609_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_609_cast_fp16 = batch_norm(beta = x_609_beta_0_to_fp16, epsilon = x_609_epsilon_0_to_fp16, gamma = x_609_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_131_cast_fp16)[name = tensor("x_609_cast_fp16")]; tensor layers_21_fc1_input_shift_to_fp16 = const()[name = tensor("layers_21_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(418093952)))]; tensor input_479_cast_fp16 = sub(x = x_609_cast_fp16, y = layers_21_fc1_input_shift_to_fp16)[name = tensor("input_479_cast_fp16")]; tensor var_6030 = const()[name = tensor("op_6030"), val = tensor([1, 1])]; tensor var_6032 = const()[name = tensor("op_6032"), val = tensor([1, 1])]; tensor x_611_pad_type_0 = const()[name = tensor("x_611_pad_type_0"), val = tensor("custom")]; tensor x_611_pad_0 = const()[name = tensor("x_611_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_21_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(418096576))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421373440))), name = tensor("layers_21_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_21_fc1_module_bias_to_fp16 = const()[name = tensor("layers_21_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421373568)))]; tensor x_611_cast_fp16 = conv(bias = layers_21_fc1_module_bias_to_fp16, dilations = var_6032, groups = var_5797, pad = x_611_pad_0, pad_type = x_611_pad_type_0, strides = var_6030, weight = layers_21_fc1_module_weight_to_fp16_palettized, x = input_479_cast_fp16)[name = tensor("x_611_cast_fp16")]; tensor layers_21_fc1_output_scale_to_fp16 = const()[name = tensor("layers_21_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421383872)))]; tensor input_481_cast_fp16 = mul(x = x_611_cast_fp16, y = layers_21_fc1_output_scale_to_fp16)[name = tensor("input_481_cast_fp16")]; tensor x_613_mode_0 = const()[name = tensor("x_613_mode_0"), val = tensor("EXACT")]; tensor x_613_cast_fp16 = gelu(mode = x_613_mode_0, x = input_481_cast_fp16)[name = tensor("x_613_cast_fp16")]; tensor layers_21_fc2_input_shift_to_fp16 = const()[name = tensor("layers_21_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421394176)))]; tensor input_483_cast_fp16 = sub(x = x_613_cast_fp16, y = layers_21_fc2_input_shift_to_fp16)[name = tensor("input_483_cast_fp16")]; tensor var_6043 = const()[name = tensor("op_6043"), val = tensor([1, 1])]; tensor var_6045 = const()[name = tensor("op_6045"), val = tensor([1, 1])]; tensor x_615_pad_type_0 = const()[name = tensor("x_615_pad_type_0"), val = tensor("custom")]; tensor x_615_pad_0 = const()[name = tensor("x_615_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_21_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421404480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424681344))), name = tensor("layers_21_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_21_fc2_module_bias_to_fp16 = const()[name = tensor("layers_21_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424681472)))]; tensor x_615_cast_fp16 = conv(bias = layers_21_fc2_module_bias_to_fp16, dilations = var_6045, groups = var_5797, pad = x_615_pad_0, pad_type = x_615_pad_type_0, strides = var_6043, weight = layers_21_fc2_module_weight_to_fp16_palettized, x = input_483_cast_fp16)[name = tensor("x_615_cast_fp16")]; tensor layers_21_fc2_output_scale_to_fp16 = const()[name = tensor("layers_21_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424684096)))]; tensor hidden_states_45_cast_fp16 = mul(x = x_615_cast_fp16, y = layers_21_fc2_output_scale_to_fp16)[name = tensor("hidden_states_45_cast_fp16")]; tensor inputs_133_cast_fp16 = add(x = inputs_131_cast_fp16, y = hidden_states_45_cast_fp16)[name = tensor("inputs_133_cast_fp16")]; tensor var_6060 = const()[name = tensor("op_6060"), val = tensor(3)]; tensor var_6067 = const()[name = tensor("op_6067"), val = tensor(1)]; tensor var_6068 = const()[name = tensor("op_6068"), val = tensor(true)]; tensor var_6080 = const()[name = tensor("op_6080"), val = tensor([1])]; tensor channels_mean_133_cast_fp16 = reduce_mean(axes = var_6080, keep_dims = var_6068, x = inputs_133_cast_fp16)[name = tensor("channels_mean_133_cast_fp16")]; tensor zero_mean_133_cast_fp16 = sub(x = inputs_133_cast_fp16, y = channels_mean_133_cast_fp16)[name = tensor("zero_mean_133_cast_fp16")]; tensor zero_mean_sq_133_cast_fp16 = mul(x = zero_mean_133_cast_fp16, y = zero_mean_133_cast_fp16)[name = tensor("zero_mean_sq_133_cast_fp16")]; tensor var_6084 = const()[name = tensor("op_6084"), val = tensor([1])]; tensor var_6085_cast_fp16 = reduce_mean(axes = var_6084, keep_dims = var_6068, x = zero_mean_sq_133_cast_fp16)[name = tensor("op_6085_cast_fp16")]; tensor var_6086_to_fp16 = const()[name = tensor("op_6086_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_6087_cast_fp16 = add(x = var_6085_cast_fp16, y = var_6086_to_fp16)[name = tensor("op_6087_cast_fp16")]; tensor denom_133_epsilon_0_to_fp16 = const()[name = tensor("denom_133_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_133_cast_fp16 = rsqrt(epsilon = denom_133_epsilon_0_to_fp16, x = var_6087_cast_fp16)[name = tensor("denom_133_cast_fp16")]; tensor out_133_cast_fp16 = mul(x = zero_mean_133_cast_fp16, y = denom_133_cast_fp16)[name = tensor("out_133_cast_fp16")]; tensor obj_309_gamma_0_to_fp16 = const()[name = tensor("obj_309_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424686720)))]; tensor obj_309_beta_0_to_fp16 = const()[name = tensor("obj_309_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424689344)))]; tensor obj_309_epsilon_0_to_fp16 = const()[name = tensor("obj_309_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_309_cast_fp16 = batch_norm(beta = obj_309_beta_0_to_fp16, epsilon = obj_309_epsilon_0_to_fp16, gamma = obj_309_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_133_cast_fp16)[name = tensor("obj_309_cast_fp16")]; tensor layers_22_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424691968)))]; tensor input_485_cast_fp16 = sub(x = obj_309_cast_fp16, y = layers_22_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_485_cast_fp16")]; tensor var_6106 = const()[name = tensor("op_6106"), val = tensor([1, 1])]; tensor var_6108 = const()[name = tensor("op_6108"), val = tensor([1, 1])]; tensor x_617_pad_type_0 = const()[name = tensor("x_617_pad_type_0"), val = tensor("custom")]; tensor x_617_pad_0 = const()[name = tensor("x_617_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_22_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424694592))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(425513856))), name = tensor("layers_22_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_22_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(425513984)))]; tensor x_617_cast_fp16 = conv(bias = layers_22_self_attn_q_proj_module_bias_to_fp16, dilations = var_6108, groups = var_6067, pad = x_617_pad_0, pad_type = x_617_pad_type_0, strides = var_6106, weight = layers_22_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = tensor("x_617_cast_fp16")]; tensor layers_22_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(425516608)))]; tensor query_89_cast_fp16 = mul(x = x_617_cast_fp16, y = layers_22_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_89_cast_fp16")]; tensor var_6118 = const()[name = tensor("op_6118"), val = tensor([1, 1])]; tensor var_6120 = const()[name = tensor("op_6120"), val = tensor([1, 1])]; tensor x_619_pad_type_0 = const()[name = tensor("x_619_pad_type_0"), val = tensor("custom")]; tensor x_619_pad_0 = const()[name = tensor("x_619_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_22_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(425519232))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426338496))), name = tensor("layers_22_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_22_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426338624)))]; tensor x_619_cast_fp16 = conv(bias = layers_22_self_attn_k_proj_module_bias_to_fp16, dilations = var_6120, groups = var_6067, pad = x_619_pad_0, pad_type = x_619_pad_type_0, strides = var_6118, weight = layers_22_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = tensor("x_619_cast_fp16")]; tensor layers_22_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_22_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426341248)))]; tensor current_key_45_cast_fp16 = mul(x = x_619_cast_fp16, y = layers_22_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_45_cast_fp16")]; tensor var_6130 = const()[name = tensor("op_6130"), val = tensor([1, 1])]; tensor var_6132 = const()[name = tensor("op_6132"), val = tensor([1, 1])]; tensor x_621_pad_type_0 = const()[name = tensor("x_621_pad_type_0"), val = tensor("custom")]; tensor x_621_pad_0 = const()[name = tensor("x_621_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_22_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426343872))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427163136))), name = tensor("layers_22_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_22_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427163264)))]; tensor x_621_cast_fp16 = conv(bias = layers_22_self_attn_v_proj_module_bias_to_fp16, dilations = var_6132, groups = var_6067, pad = x_621_pad_0, pad_type = x_621_pad_type_0, strides = var_6130, weight = layers_22_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = tensor("x_621_cast_fp16")]; tensor layers_22_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427165888)))]; tensor current_value_45_cast_fp16 = mul(x = x_621_cast_fp16, y = layers_22_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_45_cast_fp16")]; tensor var_6140_cast_fp16 = mul(x = current_key_45_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_6140_cast_fp16")]; tensor var_6142_cast_fp16 = mul(x = var_103_cast_fp16_22, y = var_257_cast_fp16)[name = tensor("op_6142_cast_fp16")]; tensor key_89_cast_fp16 = add(x = var_6140_cast_fp16, y = var_6142_cast_fp16)[name = tensor("key_89_cast_fp16")]; tensor var_6144_cast_fp16 = mul(x = current_value_45_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_6144_cast_fp16")]; tensor var_6146_cast_fp16 = mul(x = var_138_cast_fp16_22, y = var_257_cast_fp16)[name = tensor("op_6146_cast_fp16")]; tensor value_89_cast_fp16 = add(x = var_6144_cast_fp16, y = var_6146_cast_fp16)[name = tensor("value_89_cast_fp16")]; tensor var_6149 = const()[name = tensor("op_6149"), val = tensor([1, 20, 64, -1])]; tensor var_6150_cast_fp16 = reshape(shape = var_6149, x = query_89_cast_fp16)[name = tensor("op_6150_cast_fp16")]; tensor var_6151_to_fp16 = const()[name = tensor("op_6151_to_fp16"), val = tensor(0x1p-3)]; tensor var_6152_cast_fp16 = mul(x = var_6150_cast_fp16, y = var_6151_to_fp16)[name = tensor("op_6152_cast_fp16")]; tensor var_6153 = const()[name = tensor("op_6153"), val = tensor([1, 20, 64, -1])]; tensor var_6154_cast_fp16 = reshape(shape = var_6153, x = key_89_cast_fp16)[name = tensor("op_6154_cast_fp16")]; tensor mh_w_133_transpose_x_0 = const()[name = tensor("mh_w_133_transpose_x_0"), val = tensor(true)]; tensor mh_w_133_transpose_y_0 = const()[name = tensor("mh_w_133_transpose_y_0"), val = tensor(false)]; tensor mh_w_133_cast_fp16 = matmul(transpose_x = mh_w_133_transpose_x_0, transpose_y = mh_w_133_transpose_y_0, x = var_6152_cast_fp16, y = var_6154_cast_fp16)[name = tensor("mh_w_133_cast_fp16")]; tensor mh_w_135_cast_fp16 = add(x = mh_w_133_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_135_cast_fp16")]; tensor var_6162_cast_fp16 = softmax(axis = var_6060, x = mh_w_135_cast_fp16)[name = tensor("op_6162_cast_fp16")]; tensor var_6163 = const()[name = tensor("op_6163"), val = tensor([1, 20, 64, -1])]; tensor var_6164_cast_fp16 = reshape(shape = var_6163, x = value_89_cast_fp16)[name = tensor("op_6164_cast_fp16")]; tensor attn_89_transpose_x_0 = const()[name = tensor("attn_89_transpose_x_0"), val = tensor(false)]; tensor attn_89_transpose_y_0 = const()[name = tensor("attn_89_transpose_y_0"), val = tensor(true)]; tensor attn_89_cast_fp16 = matmul(transpose_x = attn_89_transpose_x_0, transpose_y = attn_89_transpose_y_0, x = var_6164_cast_fp16, y = var_6162_cast_fp16)[name = tensor("attn_89_cast_fp16")]; tensor var_6167 = const()[name = tensor("op_6167"), val = tensor([1, 1280, 1, -1])]; tensor x_623_cast_fp16 = reshape(shape = var_6167, x = attn_89_cast_fp16)[name = tensor("x_623_cast_fp16")]; tensor layers_22_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427168512)))]; tensor input_491_cast_fp16 = sub(x = x_623_cast_fp16, y = layers_22_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_491_cast_fp16")]; tensor var_6175 = const()[name = tensor("op_6175"), val = tensor([1, 1])]; tensor var_6177 = const()[name = tensor("op_6177"), val = tensor([1, 1])]; tensor x_625_pad_type_0 = const()[name = tensor("x_625_pad_type_0"), val = tensor("custom")]; tensor x_625_pad_0 = const()[name = tensor("x_625_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_22_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427171136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427990400))), name = tensor("layers_22_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_22_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427990528)))]; tensor x_625_cast_fp16 = conv(bias = layers_22_self_attn_o_proj_module_bias_to_fp16, dilations = var_6177, groups = var_6067, pad = x_625_pad_0, pad_type = x_625_pad_type_0, strides = var_6175, weight = layers_22_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_491_cast_fp16)[name = tensor("x_625_cast_fp16")]; tensor layers_22_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427993152)))]; tensor obj_315_cast_fp16 = mul(x = x_625_cast_fp16, y = layers_22_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_315_cast_fp16")]; tensor inputs_135_cast_fp16 = add(x = inputs_133_cast_fp16, y = obj_315_cast_fp16)[name = tensor("inputs_135_cast_fp16")]; tensor var_6188 = const()[name = tensor("op_6188"), val = tensor([1])]; tensor channels_mean_135_cast_fp16 = reduce_mean(axes = var_6188, keep_dims = var_6068, x = inputs_135_cast_fp16)[name = tensor("channels_mean_135_cast_fp16")]; tensor zero_mean_135_cast_fp16 = sub(x = inputs_135_cast_fp16, y = channels_mean_135_cast_fp16)[name = tensor("zero_mean_135_cast_fp16")]; tensor zero_mean_sq_135_cast_fp16 = mul(x = zero_mean_135_cast_fp16, y = zero_mean_135_cast_fp16)[name = tensor("zero_mean_sq_135_cast_fp16")]; tensor var_6192 = const()[name = tensor("op_6192"), val = tensor([1])]; tensor var_6193_cast_fp16 = reduce_mean(axes = var_6192, keep_dims = var_6068, x = zero_mean_sq_135_cast_fp16)[name = tensor("op_6193_cast_fp16")]; tensor var_6194_to_fp16 = const()[name = tensor("op_6194_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_6195_cast_fp16 = add(x = var_6193_cast_fp16, y = var_6194_to_fp16)[name = tensor("op_6195_cast_fp16")]; tensor denom_135_epsilon_0_to_fp16 = const()[name = tensor("denom_135_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_135_cast_fp16 = rsqrt(epsilon = denom_135_epsilon_0_to_fp16, x = var_6195_cast_fp16)[name = tensor("denom_135_cast_fp16")]; tensor out_135_cast_fp16 = mul(x = zero_mean_135_cast_fp16, y = denom_135_cast_fp16)[name = tensor("out_135_cast_fp16")]; tensor obj_317_gamma_0_to_fp16 = const()[name = tensor("obj_317_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427995776)))]; tensor obj_317_beta_0_to_fp16 = const()[name = tensor("obj_317_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427998400)))]; tensor obj_317_epsilon_0_to_fp16 = const()[name = tensor("obj_317_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_317_cast_fp16 = batch_norm(beta = obj_317_beta_0_to_fp16, epsilon = obj_317_epsilon_0_to_fp16, gamma = obj_317_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_135_cast_fp16)[name = tensor("obj_317_cast_fp16")]; tensor layers_22_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_22_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428001024)))]; tensor input_493_cast_fp16 = sub(x = obj_317_cast_fp16, y = layers_22_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_493_cast_fp16")]; tensor var_6214 = const()[name = tensor("op_6214"), val = tensor([1, 1])]; tensor var_6216 = const()[name = tensor("op_6216"), val = tensor([1, 1])]; tensor x_627_pad_type_0 = const()[name = tensor("x_627_pad_type_0"), val = tensor("custom")]; tensor x_627_pad_0 = const()[name = tensor("x_627_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_22_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428003648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428822912))), name = tensor("layers_22_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_22_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428823040)))]; tensor x_627_cast_fp16 = conv(bias = layers_22_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_6216, groups = var_6067, pad = x_627_pad_0, pad_type = x_627_pad_type_0, strides = var_6214, weight = layers_22_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_493_cast_fp16)[name = tensor("x_627_cast_fp16")]; tensor layers_22_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_22_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428825664)))]; tensor query_91_cast_fp16 = mul(x = x_627_cast_fp16, y = layers_22_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_91_cast_fp16")]; tensor var_6226 = const()[name = tensor("op_6226"), val = tensor([1, 1])]; tensor var_6228 = const()[name = tensor("op_6228"), val = tensor([1, 1])]; tensor x_629_pad_type_0 = const()[name = tensor("x_629_pad_type_0"), val = tensor("custom")]; tensor x_629_pad_0 = const()[name = tensor("x_629_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_22_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428828288))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429647552))), name = tensor("layers_22_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_22_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429647680)))]; tensor x_629_cast_fp16 = conv(bias = layers_22_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_6228, groups = var_6067, pad = x_629_pad_0, pad_type = x_629_pad_type_0, strides = var_6226, weight = layers_22_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_629_cast_fp16")]; tensor layers_22_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_22_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429650304)))]; tensor key_91_cast_fp16 = mul(x = x_629_cast_fp16, y = layers_22_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_91_cast_fp16")]; tensor var_6238 = const()[name = tensor("op_6238"), val = tensor([1, 1])]; tensor var_6240 = const()[name = tensor("op_6240"), val = tensor([1, 1])]; tensor x_631_pad_type_0 = const()[name = tensor("x_631_pad_type_0"), val = tensor("custom")]; tensor x_631_pad_0 = const()[name = tensor("x_631_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_22_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429652928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(430472192))), name = tensor("layers_22_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_22_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(430472320)))]; tensor x_631_cast_fp16 = conv(bias = layers_22_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_6240, groups = var_6067, pad = x_631_pad_0, pad_type = x_631_pad_type_0, strides = var_6238, weight = layers_22_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_631_cast_fp16")]; tensor layers_22_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_22_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(430474944)))]; tensor value_91_cast_fp16 = mul(x = x_631_cast_fp16, y = layers_22_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_91_cast_fp16")]; tensor var_6245 = const()[name = tensor("op_6245"), val = tensor([1, 20, 64, -1])]; tensor var_6246_cast_fp16 = reshape(shape = var_6245, x = query_91_cast_fp16)[name = tensor("op_6246_cast_fp16")]; tensor var_6247_to_fp16 = const()[name = tensor("op_6247_to_fp16"), val = tensor(0x1p-3)]; tensor var_6248_cast_fp16 = mul(x = var_6246_cast_fp16, y = var_6247_to_fp16)[name = tensor("op_6248_cast_fp16")]; tensor var_6249 = const()[name = tensor("op_6249"), val = tensor([1, 20, 64, -1])]; tensor var_6250_cast_fp16 = reshape(shape = var_6249, x = key_91_cast_fp16)[name = tensor("op_6250_cast_fp16")]; tensor mh_w_137_transpose_x_0 = const()[name = tensor("mh_w_137_transpose_x_0"), val = tensor(true)]; tensor mh_w_137_transpose_y_0 = const()[name = tensor("mh_w_137_transpose_y_0"), val = tensor(false)]; tensor mh_w_137_cast_fp16 = matmul(transpose_x = mh_w_137_transpose_x_0, transpose_y = mh_w_137_transpose_y_0, x = var_6248_cast_fp16, y = var_6250_cast_fp16)[name = tensor("mh_w_137_cast_fp16")]; tensor obj_321_cast_fp16 = softmax(axis = var_6060, x = mh_w_137_cast_fp16)[name = tensor("obj_321_cast_fp16")]; tensor var_6254 = const()[name = tensor("op_6254"), val = tensor([1, 20, 64, -1])]; tensor var_6255_cast_fp16 = reshape(shape = var_6254, x = value_91_cast_fp16)[name = tensor("op_6255_cast_fp16")]; tensor attn_91_transpose_x_0 = const()[name = tensor("attn_91_transpose_x_0"), val = tensor(false)]; tensor attn_91_transpose_y_0 = const()[name = tensor("attn_91_transpose_y_0"), val = tensor(true)]; tensor attn_91_cast_fp16 = matmul(transpose_x = attn_91_transpose_x_0, transpose_y = attn_91_transpose_y_0, x = var_6255_cast_fp16, y = obj_321_cast_fp16)[name = tensor("attn_91_cast_fp16")]; tensor var_6258 = const()[name = tensor("op_6258"), val = tensor([1, 1280, 1, -1])]; tensor x_633_cast_fp16 = reshape(shape = var_6258, x = attn_91_cast_fp16)[name = tensor("x_633_cast_fp16")]; tensor layers_22_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_22_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(430477568)))]; tensor input_499_cast_fp16 = sub(x = x_633_cast_fp16, y = layers_22_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_499_cast_fp16")]; tensor var_6266 = const()[name = tensor("op_6266"), val = tensor([1, 1])]; tensor var_6268 = const()[name = tensor("op_6268"), val = tensor([1, 1])]; tensor x_635_pad_type_0 = const()[name = tensor("x_635_pad_type_0"), val = tensor("custom")]; tensor x_635_pad_0 = const()[name = tensor("x_635_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_22_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(430480192))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(431299456))), name = tensor("layers_22_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_22_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(431299584)))]; tensor x_635_cast_fp16 = conv(bias = layers_22_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_6268, groups = var_6067, pad = x_635_pad_0, pad_type = x_635_pad_type_0, strides = var_6266, weight = layers_22_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_499_cast_fp16)[name = tensor("x_635_cast_fp16")]; tensor layers_22_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_22_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(431302208)))]; tensor obj_319_cast_fp16 = mul(x = x_635_cast_fp16, y = layers_22_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_319_cast_fp16")]; tensor inputs_137_cast_fp16 = add(x = inputs_135_cast_fp16, y = obj_319_cast_fp16)[name = tensor("inputs_137_cast_fp16")]; tensor var_6275 = const()[name = tensor("op_6275"), val = tensor([1])]; tensor channels_mean_137_cast_fp16 = reduce_mean(axes = var_6275, keep_dims = var_6068, x = inputs_137_cast_fp16)[name = tensor("channels_mean_137_cast_fp16")]; tensor zero_mean_137_cast_fp16 = sub(x = inputs_137_cast_fp16, y = channels_mean_137_cast_fp16)[name = tensor("zero_mean_137_cast_fp16")]; tensor zero_mean_sq_137_cast_fp16 = mul(x = zero_mean_137_cast_fp16, y = zero_mean_137_cast_fp16)[name = tensor("zero_mean_sq_137_cast_fp16")]; tensor var_6279 = const()[name = tensor("op_6279"), val = tensor([1])]; tensor var_6280_cast_fp16 = reduce_mean(axes = var_6279, keep_dims = var_6068, x = zero_mean_sq_137_cast_fp16)[name = tensor("op_6280_cast_fp16")]; tensor var_6281_to_fp16 = const()[name = tensor("op_6281_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_6282_cast_fp16 = add(x = var_6280_cast_fp16, y = var_6281_to_fp16)[name = tensor("op_6282_cast_fp16")]; tensor denom_137_epsilon_0_to_fp16 = const()[name = tensor("denom_137_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_137_cast_fp16 = rsqrt(epsilon = denom_137_epsilon_0_to_fp16, x = var_6282_cast_fp16)[name = tensor("denom_137_cast_fp16")]; tensor out_137_cast_fp16 = mul(x = zero_mean_137_cast_fp16, y = denom_137_cast_fp16)[name = tensor("out_137_cast_fp16")]; tensor x_637_gamma_0_to_fp16 = const()[name = tensor("x_637_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(431304832)))]; tensor x_637_beta_0_to_fp16 = const()[name = tensor("x_637_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(431307456)))]; tensor x_637_epsilon_0_to_fp16 = const()[name = tensor("x_637_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_637_cast_fp16 = batch_norm(beta = x_637_beta_0_to_fp16, epsilon = x_637_epsilon_0_to_fp16, gamma = x_637_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_137_cast_fp16)[name = tensor("x_637_cast_fp16")]; tensor layers_22_fc1_input_shift_to_fp16 = const()[name = tensor("layers_22_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(431310080)))]; tensor input_501_cast_fp16 = sub(x = x_637_cast_fp16, y = layers_22_fc1_input_shift_to_fp16)[name = tensor("input_501_cast_fp16")]; tensor var_6297 = const()[name = tensor("op_6297"), val = tensor([1, 1])]; tensor var_6299 = const()[name = tensor("op_6299"), val = tensor([1, 1])]; tensor x_639_pad_type_0 = const()[name = tensor("x_639_pad_type_0"), val = tensor("custom")]; tensor x_639_pad_0 = const()[name = tensor("x_639_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_22_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(431312704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434589568))), name = tensor("layers_22_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_22_fc1_module_bias_to_fp16 = const()[name = tensor("layers_22_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434589696)))]; tensor x_639_cast_fp16 = conv(bias = layers_22_fc1_module_bias_to_fp16, dilations = var_6299, groups = var_6067, pad = x_639_pad_0, pad_type = x_639_pad_type_0, strides = var_6297, weight = layers_22_fc1_module_weight_to_fp16_palettized, x = input_501_cast_fp16)[name = tensor("x_639_cast_fp16")]; tensor layers_22_fc1_output_scale_to_fp16 = const()[name = tensor("layers_22_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434600000)))]; tensor input_503_cast_fp16 = mul(x = x_639_cast_fp16, y = layers_22_fc1_output_scale_to_fp16)[name = tensor("input_503_cast_fp16")]; tensor x_641_mode_0 = const()[name = tensor("x_641_mode_0"), val = tensor("EXACT")]; tensor x_641_cast_fp16 = gelu(mode = x_641_mode_0, x = input_503_cast_fp16)[name = tensor("x_641_cast_fp16")]; tensor layers_22_fc2_input_shift_to_fp16 = const()[name = tensor("layers_22_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434610304)))]; tensor input_505_cast_fp16 = sub(x = x_641_cast_fp16, y = layers_22_fc2_input_shift_to_fp16)[name = tensor("input_505_cast_fp16")]; tensor var_6310 = const()[name = tensor("op_6310"), val = tensor([1, 1])]; tensor var_6312 = const()[name = tensor("op_6312"), val = tensor([1, 1])]; tensor x_643_pad_type_0 = const()[name = tensor("x_643_pad_type_0"), val = tensor("custom")]; tensor x_643_pad_0 = const()[name = tensor("x_643_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_22_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434620608))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437897472))), name = tensor("layers_22_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_22_fc2_module_bias_to_fp16 = const()[name = tensor("layers_22_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437897600)))]; tensor x_643_cast_fp16 = conv(bias = layers_22_fc2_module_bias_to_fp16, dilations = var_6312, groups = var_6067, pad = x_643_pad_0, pad_type = x_643_pad_type_0, strides = var_6310, weight = layers_22_fc2_module_weight_to_fp16_palettized, x = input_505_cast_fp16)[name = tensor("x_643_cast_fp16")]; tensor layers_22_fc2_output_scale_to_fp16 = const()[name = tensor("layers_22_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437900224)))]; tensor hidden_states_47_cast_fp16 = mul(x = x_643_cast_fp16, y = layers_22_fc2_output_scale_to_fp16)[name = tensor("hidden_states_47_cast_fp16")]; tensor inputs_139_cast_fp16 = add(x = inputs_137_cast_fp16, y = hidden_states_47_cast_fp16)[name = tensor("inputs_139_cast_fp16")]; tensor var_6326 = const()[name = tensor("op_6326"), val = tensor(3)]; tensor var_6333 = const()[name = tensor("op_6333"), val = tensor(1)]; tensor var_6334 = const()[name = tensor("op_6334"), val = tensor(true)]; tensor var_6346 = const()[name = tensor("op_6346"), val = tensor([1])]; tensor channels_mean_139_cast_fp16 = reduce_mean(axes = var_6346, keep_dims = var_6334, x = inputs_139_cast_fp16)[name = tensor("channels_mean_139_cast_fp16")]; tensor zero_mean_139_cast_fp16 = sub(x = inputs_139_cast_fp16, y = channels_mean_139_cast_fp16)[name = tensor("zero_mean_139_cast_fp16")]; tensor zero_mean_sq_139_cast_fp16 = mul(x = zero_mean_139_cast_fp16, y = zero_mean_139_cast_fp16)[name = tensor("zero_mean_sq_139_cast_fp16")]; tensor var_6350 = const()[name = tensor("op_6350"), val = tensor([1])]; tensor var_6351_cast_fp16 = reduce_mean(axes = var_6350, keep_dims = var_6334, x = zero_mean_sq_139_cast_fp16)[name = tensor("op_6351_cast_fp16")]; tensor var_6352_to_fp16 = const()[name = tensor("op_6352_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_6353_cast_fp16 = add(x = var_6351_cast_fp16, y = var_6352_to_fp16)[name = tensor("op_6353_cast_fp16")]; tensor denom_139_epsilon_0_to_fp16 = const()[name = tensor("denom_139_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_139_cast_fp16 = rsqrt(epsilon = denom_139_epsilon_0_to_fp16, x = var_6353_cast_fp16)[name = tensor("denom_139_cast_fp16")]; tensor out_139_cast_fp16 = mul(x = zero_mean_139_cast_fp16, y = denom_139_cast_fp16)[name = tensor("out_139_cast_fp16")]; tensor obj_323_gamma_0_to_fp16 = const()[name = tensor("obj_323_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437902848)))]; tensor obj_323_beta_0_to_fp16 = const()[name = tensor("obj_323_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437905472)))]; tensor obj_323_epsilon_0_to_fp16 = const()[name = tensor("obj_323_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_323_cast_fp16 = batch_norm(beta = obj_323_beta_0_to_fp16, epsilon = obj_323_epsilon_0_to_fp16, gamma = obj_323_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_139_cast_fp16)[name = tensor("obj_323_cast_fp16")]; tensor layers_23_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437908096)))]; tensor input_507_cast_fp16 = sub(x = obj_323_cast_fp16, y = layers_23_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_507_cast_fp16")]; tensor var_6372 = const()[name = tensor("op_6372"), val = tensor([1, 1])]; tensor var_6374 = const()[name = tensor("op_6374"), val = tensor([1, 1])]; tensor x_645_pad_type_0 = const()[name = tensor("x_645_pad_type_0"), val = tensor("custom")]; tensor x_645_pad_0 = const()[name = tensor("x_645_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_23_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437910720))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438729984))), name = tensor("layers_23_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_23_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438730112)))]; tensor x_645_cast_fp16 = conv(bias = layers_23_self_attn_q_proj_module_bias_to_fp16, dilations = var_6374, groups = var_6333, pad = x_645_pad_0, pad_type = x_645_pad_type_0, strides = var_6372, weight = layers_23_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = tensor("x_645_cast_fp16")]; tensor layers_23_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438732736)))]; tensor query_93_cast_fp16 = mul(x = x_645_cast_fp16, y = layers_23_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_93_cast_fp16")]; tensor var_6384 = const()[name = tensor("op_6384"), val = tensor([1, 1])]; tensor var_6386 = const()[name = tensor("op_6386"), val = tensor([1, 1])]; tensor x_647_pad_type_0 = const()[name = tensor("x_647_pad_type_0"), val = tensor("custom")]; tensor x_647_pad_0 = const()[name = tensor("x_647_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_23_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438735360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439554624))), name = tensor("layers_23_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_23_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439554752)))]; tensor x_647_cast_fp16 = conv(bias = layers_23_self_attn_k_proj_module_bias_to_fp16, dilations = var_6386, groups = var_6333, pad = x_647_pad_0, pad_type = x_647_pad_type_0, strides = var_6384, weight = layers_23_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = tensor("x_647_cast_fp16")]; tensor layers_23_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_23_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439557376)))]; tensor current_key_47_cast_fp16 = mul(x = x_647_cast_fp16, y = layers_23_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_47_cast_fp16")]; tensor var_6396 = const()[name = tensor("op_6396"), val = tensor([1, 1])]; tensor var_6398 = const()[name = tensor("op_6398"), val = tensor([1, 1])]; tensor x_649_pad_type_0 = const()[name = tensor("x_649_pad_type_0"), val = tensor("custom")]; tensor x_649_pad_0 = const()[name = tensor("x_649_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_23_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439560000))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440379264))), name = tensor("layers_23_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_23_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440379392)))]; tensor x_649_cast_fp16 = conv(bias = layers_23_self_attn_v_proj_module_bias_to_fp16, dilations = var_6398, groups = var_6333, pad = x_649_pad_0, pad_type = x_649_pad_type_0, strides = var_6396, weight = layers_23_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = tensor("x_649_cast_fp16")]; tensor layers_23_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440382016)))]; tensor current_value_47_cast_fp16 = mul(x = x_649_cast_fp16, y = layers_23_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_47_cast_fp16")]; tensor var_6406_cast_fp16 = mul(x = current_key_47_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_6406_cast_fp16")]; tensor var_6408_cast_fp16 = mul(x = var_103_cast_fp16_23, y = var_257_cast_fp16)[name = tensor("op_6408_cast_fp16")]; tensor key_93_cast_fp16 = add(x = var_6406_cast_fp16, y = var_6408_cast_fp16)[name = tensor("key_93_cast_fp16")]; tensor var_6410_cast_fp16 = mul(x = current_value_47_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_6410_cast_fp16")]; tensor var_6412_cast_fp16 = mul(x = var_138_cast_fp16_23, y = var_257_cast_fp16)[name = tensor("op_6412_cast_fp16")]; tensor value_93_cast_fp16 = add(x = var_6410_cast_fp16, y = var_6412_cast_fp16)[name = tensor("value_93_cast_fp16")]; tensor var_6415 = const()[name = tensor("op_6415"), val = tensor([1, 20, 64, -1])]; tensor var_6416_cast_fp16 = reshape(shape = var_6415, x = query_93_cast_fp16)[name = tensor("op_6416_cast_fp16")]; tensor var_6417_to_fp16 = const()[name = tensor("op_6417_to_fp16"), val = tensor(0x1p-3)]; tensor var_6418_cast_fp16 = mul(x = var_6416_cast_fp16, y = var_6417_to_fp16)[name = tensor("op_6418_cast_fp16")]; tensor var_6419 = const()[name = tensor("op_6419"), val = tensor([1, 20, 64, -1])]; tensor var_6420_cast_fp16 = reshape(shape = var_6419, x = key_93_cast_fp16)[name = tensor("op_6420_cast_fp16")]; tensor mh_w_139_transpose_x_0 = const()[name = tensor("mh_w_139_transpose_x_0"), val = tensor(true)]; tensor mh_w_139_transpose_y_0 = const()[name = tensor("mh_w_139_transpose_y_0"), val = tensor(false)]; tensor mh_w_139_cast_fp16 = matmul(transpose_x = mh_w_139_transpose_x_0, transpose_y = mh_w_139_transpose_y_0, x = var_6418_cast_fp16, y = var_6420_cast_fp16)[name = tensor("mh_w_139_cast_fp16")]; tensor mh_w_141_cast_fp16 = add(x = mh_w_139_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_141_cast_fp16")]; tensor var_6428_cast_fp16 = softmax(axis = var_6326, x = mh_w_141_cast_fp16)[name = tensor("op_6428_cast_fp16")]; tensor var_6429 = const()[name = tensor("op_6429"), val = tensor([1, 20, 64, -1])]; tensor var_6430_cast_fp16 = reshape(shape = var_6429, x = value_93_cast_fp16)[name = tensor("op_6430_cast_fp16")]; tensor attn_93_transpose_x_0 = const()[name = tensor("attn_93_transpose_x_0"), val = tensor(false)]; tensor attn_93_transpose_y_0 = const()[name = tensor("attn_93_transpose_y_0"), val = tensor(true)]; tensor attn_93_cast_fp16 = matmul(transpose_x = attn_93_transpose_x_0, transpose_y = attn_93_transpose_y_0, x = var_6430_cast_fp16, y = var_6428_cast_fp16)[name = tensor("attn_93_cast_fp16")]; tensor var_6433 = const()[name = tensor("op_6433"), val = tensor([1, 1280, 1, -1])]; tensor x_651_cast_fp16 = reshape(shape = var_6433, x = attn_93_cast_fp16)[name = tensor("x_651_cast_fp16")]; tensor layers_23_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440384640)))]; tensor input_513_cast_fp16 = sub(x = x_651_cast_fp16, y = layers_23_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_513_cast_fp16")]; tensor var_6441 = const()[name = tensor("op_6441"), val = tensor([1, 1])]; tensor var_6443 = const()[name = tensor("op_6443"), val = tensor([1, 1])]; tensor x_653_pad_type_0 = const()[name = tensor("x_653_pad_type_0"), val = tensor("custom")]; tensor x_653_pad_0 = const()[name = tensor("x_653_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_23_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440387264))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441206528))), name = tensor("layers_23_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_23_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441206656)))]; tensor x_653_cast_fp16 = conv(bias = layers_23_self_attn_o_proj_module_bias_to_fp16, dilations = var_6443, groups = var_6333, pad = x_653_pad_0, pad_type = x_653_pad_type_0, strides = var_6441, weight = layers_23_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_513_cast_fp16)[name = tensor("x_653_cast_fp16")]; tensor layers_23_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441209280)))]; tensor obj_329_cast_fp16 = mul(x = x_653_cast_fp16, y = layers_23_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_329_cast_fp16")]; tensor inputs_141_cast_fp16 = add(x = inputs_139_cast_fp16, y = obj_329_cast_fp16)[name = tensor("inputs_141_cast_fp16")]; tensor var_6454 = const()[name = tensor("op_6454"), val = tensor([1])]; tensor channels_mean_141_cast_fp16 = reduce_mean(axes = var_6454, keep_dims = var_6334, x = inputs_141_cast_fp16)[name = tensor("channels_mean_141_cast_fp16")]; tensor zero_mean_141_cast_fp16 = sub(x = inputs_141_cast_fp16, y = channels_mean_141_cast_fp16)[name = tensor("zero_mean_141_cast_fp16")]; tensor zero_mean_sq_141_cast_fp16 = mul(x = zero_mean_141_cast_fp16, y = zero_mean_141_cast_fp16)[name = tensor("zero_mean_sq_141_cast_fp16")]; tensor var_6458 = const()[name = tensor("op_6458"), val = tensor([1])]; tensor var_6459_cast_fp16 = reduce_mean(axes = var_6458, keep_dims = var_6334, x = zero_mean_sq_141_cast_fp16)[name = tensor("op_6459_cast_fp16")]; tensor var_6460_to_fp16 = const()[name = tensor("op_6460_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_6461_cast_fp16 = add(x = var_6459_cast_fp16, y = var_6460_to_fp16)[name = tensor("op_6461_cast_fp16")]; tensor denom_141_epsilon_0_to_fp16 = const()[name = tensor("denom_141_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_141_cast_fp16 = rsqrt(epsilon = denom_141_epsilon_0_to_fp16, x = var_6461_cast_fp16)[name = tensor("denom_141_cast_fp16")]; tensor out_141_cast_fp16 = mul(x = zero_mean_141_cast_fp16, y = denom_141_cast_fp16)[name = tensor("out_141_cast_fp16")]; tensor obj_331_gamma_0_to_fp16 = const()[name = tensor("obj_331_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441211904)))]; tensor obj_331_beta_0_to_fp16 = const()[name = tensor("obj_331_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441214528)))]; tensor obj_331_epsilon_0_to_fp16 = const()[name = tensor("obj_331_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_331_cast_fp16 = batch_norm(beta = obj_331_beta_0_to_fp16, epsilon = obj_331_epsilon_0_to_fp16, gamma = obj_331_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_141_cast_fp16)[name = tensor("obj_331_cast_fp16")]; tensor layers_23_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_23_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441217152)))]; tensor input_515_cast_fp16 = sub(x = obj_331_cast_fp16, y = layers_23_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_515_cast_fp16")]; tensor var_6480 = const()[name = tensor("op_6480"), val = tensor([1, 1])]; tensor var_6482 = const()[name = tensor("op_6482"), val = tensor([1, 1])]; tensor x_655_pad_type_0 = const()[name = tensor("x_655_pad_type_0"), val = tensor("custom")]; tensor x_655_pad_0 = const()[name = tensor("x_655_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_23_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441219776))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(442039040))), name = tensor("layers_23_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_23_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(442039168)))]; tensor x_655_cast_fp16 = conv(bias = layers_23_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_6482, groups = var_6333, pad = x_655_pad_0, pad_type = x_655_pad_type_0, strides = var_6480, weight = layers_23_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_515_cast_fp16)[name = tensor("x_655_cast_fp16")]; tensor layers_23_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_23_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(442041792)))]; tensor query_95_cast_fp16 = mul(x = x_655_cast_fp16, y = layers_23_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_95_cast_fp16")]; tensor var_6492 = const()[name = tensor("op_6492"), val = tensor([1, 1])]; tensor var_6494 = const()[name = tensor("op_6494"), val = tensor([1, 1])]; tensor x_657_pad_type_0 = const()[name = tensor("x_657_pad_type_0"), val = tensor("custom")]; tensor x_657_pad_0 = const()[name = tensor("x_657_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_23_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(442044416))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(442863680))), name = tensor("layers_23_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_23_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(442863808)))]; tensor x_657_cast_fp16 = conv(bias = layers_23_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_6494, groups = var_6333, pad = x_657_pad_0, pad_type = x_657_pad_type_0, strides = var_6492, weight = layers_23_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_657_cast_fp16")]; tensor layers_23_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_23_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(442866432)))]; tensor key_95_cast_fp16 = mul(x = x_657_cast_fp16, y = layers_23_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_95_cast_fp16")]; tensor var_6504 = const()[name = tensor("op_6504"), val = tensor([1, 1])]; tensor var_6506 = const()[name = tensor("op_6506"), val = tensor([1, 1])]; tensor x_659_pad_type_0 = const()[name = tensor("x_659_pad_type_0"), val = tensor("custom")]; tensor x_659_pad_0 = const()[name = tensor("x_659_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_23_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(442869056))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(443688320))), name = tensor("layers_23_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_23_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(443688448)))]; tensor x_659_cast_fp16 = conv(bias = layers_23_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_6506, groups = var_6333, pad = x_659_pad_0, pad_type = x_659_pad_type_0, strides = var_6504, weight = layers_23_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_659_cast_fp16")]; tensor layers_23_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_23_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(443691072)))]; tensor value_95_cast_fp16 = mul(x = x_659_cast_fp16, y = layers_23_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_95_cast_fp16")]; tensor var_6511 = const()[name = tensor("op_6511"), val = tensor([1, 20, 64, -1])]; tensor var_6512_cast_fp16 = reshape(shape = var_6511, x = query_95_cast_fp16)[name = tensor("op_6512_cast_fp16")]; tensor var_6513_to_fp16 = const()[name = tensor("op_6513_to_fp16"), val = tensor(0x1p-3)]; tensor var_6514_cast_fp16 = mul(x = var_6512_cast_fp16, y = var_6513_to_fp16)[name = tensor("op_6514_cast_fp16")]; tensor var_6515 = const()[name = tensor("op_6515"), val = tensor([1, 20, 64, -1])]; tensor var_6516_cast_fp16 = reshape(shape = var_6515, x = key_95_cast_fp16)[name = tensor("op_6516_cast_fp16")]; tensor mh_w_143_transpose_x_0 = const()[name = tensor("mh_w_143_transpose_x_0"), val = tensor(true)]; tensor mh_w_143_transpose_y_0 = const()[name = tensor("mh_w_143_transpose_y_0"), val = tensor(false)]; tensor mh_w_143_cast_fp16 = matmul(transpose_x = mh_w_143_transpose_x_0, transpose_y = mh_w_143_transpose_y_0, x = var_6514_cast_fp16, y = var_6516_cast_fp16)[name = tensor("mh_w_143_cast_fp16")]; tensor obj_335_cast_fp16 = softmax(axis = var_6326, x = mh_w_143_cast_fp16)[name = tensor("obj_335_cast_fp16")]; tensor var_6520 = const()[name = tensor("op_6520"), val = tensor([1, 20, 64, -1])]; tensor var_6521_cast_fp16 = reshape(shape = var_6520, x = value_95_cast_fp16)[name = tensor("op_6521_cast_fp16")]; tensor attn_95_transpose_x_0 = const()[name = tensor("attn_95_transpose_x_0"), val = tensor(false)]; tensor attn_95_transpose_y_0 = const()[name = tensor("attn_95_transpose_y_0"), val = tensor(true)]; tensor attn_95_cast_fp16 = matmul(transpose_x = attn_95_transpose_x_0, transpose_y = attn_95_transpose_y_0, x = var_6521_cast_fp16, y = obj_335_cast_fp16)[name = tensor("attn_95_cast_fp16")]; tensor var_6524 = const()[name = tensor("op_6524"), val = tensor([1, 1280, 1, -1])]; tensor x_661_cast_fp16 = reshape(shape = var_6524, x = attn_95_cast_fp16)[name = tensor("x_661_cast_fp16")]; tensor layers_23_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_23_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(443693696)))]; tensor input_521_cast_fp16 = sub(x = x_661_cast_fp16, y = layers_23_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_521_cast_fp16")]; tensor var_6532 = const()[name = tensor("op_6532"), val = tensor([1, 1])]; tensor var_6534 = const()[name = tensor("op_6534"), val = tensor([1, 1])]; tensor x_663_pad_type_0 = const()[name = tensor("x_663_pad_type_0"), val = tensor("custom")]; tensor x_663_pad_0 = const()[name = tensor("x_663_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_23_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(443696320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444515584))), name = tensor("layers_23_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_23_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444515712)))]; tensor x_663_cast_fp16 = conv(bias = layers_23_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_6534, groups = var_6333, pad = x_663_pad_0, pad_type = x_663_pad_type_0, strides = var_6532, weight = layers_23_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_521_cast_fp16)[name = tensor("x_663_cast_fp16")]; tensor layers_23_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_23_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444518336)))]; tensor obj_333_cast_fp16 = mul(x = x_663_cast_fp16, y = layers_23_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_333_cast_fp16")]; tensor inputs_143_cast_fp16 = add(x = inputs_141_cast_fp16, y = obj_333_cast_fp16)[name = tensor("inputs_143_cast_fp16")]; tensor var_6541 = const()[name = tensor("op_6541"), val = tensor([1])]; tensor channels_mean_143_cast_fp16 = reduce_mean(axes = var_6541, keep_dims = var_6334, x = inputs_143_cast_fp16)[name = tensor("channels_mean_143_cast_fp16")]; tensor zero_mean_143_cast_fp16 = sub(x = inputs_143_cast_fp16, y = channels_mean_143_cast_fp16)[name = tensor("zero_mean_143_cast_fp16")]; tensor zero_mean_sq_143_cast_fp16 = mul(x = zero_mean_143_cast_fp16, y = zero_mean_143_cast_fp16)[name = tensor("zero_mean_sq_143_cast_fp16")]; tensor var_6545 = const()[name = tensor("op_6545"), val = tensor([1])]; tensor var_6546_cast_fp16 = reduce_mean(axes = var_6545, keep_dims = var_6334, x = zero_mean_sq_143_cast_fp16)[name = tensor("op_6546_cast_fp16")]; tensor var_6547_to_fp16 = const()[name = tensor("op_6547_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_6548_cast_fp16 = add(x = var_6546_cast_fp16, y = var_6547_to_fp16)[name = tensor("op_6548_cast_fp16")]; tensor denom_143_epsilon_0_to_fp16 = const()[name = tensor("denom_143_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_143_cast_fp16 = rsqrt(epsilon = denom_143_epsilon_0_to_fp16, x = var_6548_cast_fp16)[name = tensor("denom_143_cast_fp16")]; tensor out_143_cast_fp16 = mul(x = zero_mean_143_cast_fp16, y = denom_143_cast_fp16)[name = tensor("out_143_cast_fp16")]; tensor x_665_gamma_0_to_fp16 = const()[name = tensor("x_665_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444520960)))]; tensor x_665_beta_0_to_fp16 = const()[name = tensor("x_665_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444523584)))]; tensor x_665_epsilon_0_to_fp16 = const()[name = tensor("x_665_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_665_cast_fp16 = batch_norm(beta = x_665_beta_0_to_fp16, epsilon = x_665_epsilon_0_to_fp16, gamma = x_665_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_143_cast_fp16)[name = tensor("x_665_cast_fp16")]; tensor layers_23_fc1_input_shift_to_fp16 = const()[name = tensor("layers_23_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444526208)))]; tensor input_523_cast_fp16 = sub(x = x_665_cast_fp16, y = layers_23_fc1_input_shift_to_fp16)[name = tensor("input_523_cast_fp16")]; tensor var_6563 = const()[name = tensor("op_6563"), val = tensor([1, 1])]; tensor var_6565 = const()[name = tensor("op_6565"), val = tensor([1, 1])]; tensor x_667_pad_type_0 = const()[name = tensor("x_667_pad_type_0"), val = tensor("custom")]; tensor x_667_pad_0 = const()[name = tensor("x_667_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_23_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444528832))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447805696))), name = tensor("layers_23_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_23_fc1_module_bias_to_fp16 = const()[name = tensor("layers_23_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447805824)))]; tensor x_667_cast_fp16 = conv(bias = layers_23_fc1_module_bias_to_fp16, dilations = var_6565, groups = var_6333, pad = x_667_pad_0, pad_type = x_667_pad_type_0, strides = var_6563, weight = layers_23_fc1_module_weight_to_fp16_palettized, x = input_523_cast_fp16)[name = tensor("x_667_cast_fp16")]; tensor layers_23_fc1_output_scale_to_fp16 = const()[name = tensor("layers_23_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447816128)))]; tensor input_525_cast_fp16 = mul(x = x_667_cast_fp16, y = layers_23_fc1_output_scale_to_fp16)[name = tensor("input_525_cast_fp16")]; tensor x_669_mode_0 = const()[name = tensor("x_669_mode_0"), val = tensor("EXACT")]; tensor x_669_cast_fp16 = gelu(mode = x_669_mode_0, x = input_525_cast_fp16)[name = tensor("x_669_cast_fp16")]; tensor layers_23_fc2_input_shift_to_fp16 = const()[name = tensor("layers_23_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447826432)))]; tensor input_527_cast_fp16 = sub(x = x_669_cast_fp16, y = layers_23_fc2_input_shift_to_fp16)[name = tensor("input_527_cast_fp16")]; tensor var_6576 = const()[name = tensor("op_6576"), val = tensor([1, 1])]; tensor var_6578 = const()[name = tensor("op_6578"), val = tensor([1, 1])]; tensor x_671_pad_type_0 = const()[name = tensor("x_671_pad_type_0"), val = tensor("custom")]; tensor x_671_pad_0 = const()[name = tensor("x_671_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_23_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447836736))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451113600))), name = tensor("layers_23_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_23_fc2_module_bias_to_fp16 = const()[name = tensor("layers_23_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451113728)))]; tensor x_671_cast_fp16 = conv(bias = layers_23_fc2_module_bias_to_fp16, dilations = var_6578, groups = var_6333, pad = x_671_pad_0, pad_type = x_671_pad_type_0, strides = var_6576, weight = layers_23_fc2_module_weight_to_fp16_palettized, x = input_527_cast_fp16)[name = tensor("x_671_cast_fp16")]; tensor layers_23_fc2_output_scale_to_fp16 = const()[name = tensor("layers_23_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451116352)))]; tensor hidden_states_49_cast_fp16 = mul(x = x_671_cast_fp16, y = layers_23_fc2_output_scale_to_fp16)[name = tensor("hidden_states_49_cast_fp16")]; tensor inputs_145_cast_fp16 = add(x = inputs_143_cast_fp16, y = hidden_states_49_cast_fp16)[name = tensor("inputs_145_cast_fp16")]; tensor var_6592 = const()[name = tensor("op_6592"), val = tensor(3)]; tensor var_6599 = const()[name = tensor("op_6599"), val = tensor(1)]; tensor var_6600 = const()[name = tensor("op_6600"), val = tensor(true)]; tensor var_6612 = const()[name = tensor("op_6612"), val = tensor([1])]; tensor channels_mean_145_cast_fp16 = reduce_mean(axes = var_6612, keep_dims = var_6600, x = inputs_145_cast_fp16)[name = tensor("channels_mean_145_cast_fp16")]; tensor zero_mean_145_cast_fp16 = sub(x = inputs_145_cast_fp16, y = channels_mean_145_cast_fp16)[name = tensor("zero_mean_145_cast_fp16")]; tensor zero_mean_sq_145_cast_fp16 = mul(x = zero_mean_145_cast_fp16, y = zero_mean_145_cast_fp16)[name = tensor("zero_mean_sq_145_cast_fp16")]; tensor var_6616 = const()[name = tensor("op_6616"), val = tensor([1])]; tensor var_6617_cast_fp16 = reduce_mean(axes = var_6616, keep_dims = var_6600, x = zero_mean_sq_145_cast_fp16)[name = tensor("op_6617_cast_fp16")]; tensor var_6618_to_fp16 = const()[name = tensor("op_6618_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_6619_cast_fp16 = add(x = var_6617_cast_fp16, y = var_6618_to_fp16)[name = tensor("op_6619_cast_fp16")]; tensor denom_145_epsilon_0_to_fp16 = const()[name = tensor("denom_145_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_145_cast_fp16 = rsqrt(epsilon = denom_145_epsilon_0_to_fp16, x = var_6619_cast_fp16)[name = tensor("denom_145_cast_fp16")]; tensor out_145_cast_fp16 = mul(x = zero_mean_145_cast_fp16, y = denom_145_cast_fp16)[name = tensor("out_145_cast_fp16")]; tensor obj_337_gamma_0_to_fp16 = const()[name = tensor("obj_337_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451118976)))]; tensor obj_337_beta_0_to_fp16 = const()[name = tensor("obj_337_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451121600)))]; tensor obj_337_epsilon_0_to_fp16 = const()[name = tensor("obj_337_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_337_cast_fp16 = batch_norm(beta = obj_337_beta_0_to_fp16, epsilon = obj_337_epsilon_0_to_fp16, gamma = obj_337_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_145_cast_fp16)[name = tensor("obj_337_cast_fp16")]; tensor layers_24_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_24_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451124224)))]; tensor input_529_cast_fp16 = sub(x = obj_337_cast_fp16, y = layers_24_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_529_cast_fp16")]; tensor var_6638 = const()[name = tensor("op_6638"), val = tensor([1, 1])]; tensor var_6640 = const()[name = tensor("op_6640"), val = tensor([1, 1])]; tensor x_673_pad_type_0 = const()[name = tensor("x_673_pad_type_0"), val = tensor("custom")]; tensor x_673_pad_0 = const()[name = tensor("x_673_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_24_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451126848))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451946112))), name = tensor("layers_24_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_24_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451946240)))]; tensor x_673_cast_fp16 = conv(bias = layers_24_self_attn_q_proj_module_bias_to_fp16, dilations = var_6640, groups = var_6599, pad = x_673_pad_0, pad_type = x_673_pad_type_0, strides = var_6638, weight = layers_24_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_529_cast_fp16)[name = tensor("x_673_cast_fp16")]; tensor layers_24_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_24_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451948864)))]; tensor query_97_cast_fp16 = mul(x = x_673_cast_fp16, y = layers_24_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_97_cast_fp16")]; tensor var_6650 = const()[name = tensor("op_6650"), val = tensor([1, 1])]; tensor var_6652 = const()[name = tensor("op_6652"), val = tensor([1, 1])]; tensor x_675_pad_type_0 = const()[name = tensor("x_675_pad_type_0"), val = tensor("custom")]; tensor x_675_pad_0 = const()[name = tensor("x_675_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_24_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451951488))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452770752))), name = tensor("layers_24_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_24_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452770880)))]; tensor x_675_cast_fp16 = conv(bias = layers_24_self_attn_k_proj_module_bias_to_fp16, dilations = var_6652, groups = var_6599, pad = x_675_pad_0, pad_type = x_675_pad_type_0, strides = var_6650, weight = layers_24_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_529_cast_fp16)[name = tensor("x_675_cast_fp16")]; tensor layers_24_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_24_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452773504)))]; tensor current_key_49_cast_fp16 = mul(x = x_675_cast_fp16, y = layers_24_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_49_cast_fp16")]; tensor var_6662 = const()[name = tensor("op_6662"), val = tensor([1, 1])]; tensor var_6664 = const()[name = tensor("op_6664"), val = tensor([1, 1])]; tensor x_677_pad_type_0 = const()[name = tensor("x_677_pad_type_0"), val = tensor("custom")]; tensor x_677_pad_0 = const()[name = tensor("x_677_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_24_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452776128))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453595392))), name = tensor("layers_24_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_24_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453595520)))]; tensor x_677_cast_fp16 = conv(bias = layers_24_self_attn_v_proj_module_bias_to_fp16, dilations = var_6664, groups = var_6599, pad = x_677_pad_0, pad_type = x_677_pad_type_0, strides = var_6662, weight = layers_24_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_529_cast_fp16)[name = tensor("x_677_cast_fp16")]; tensor layers_24_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_24_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453598144)))]; tensor current_value_49_cast_fp16 = mul(x = x_677_cast_fp16, y = layers_24_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_49_cast_fp16")]; tensor var_6672_cast_fp16 = mul(x = current_key_49_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_6672_cast_fp16")]; tensor var_6674_cast_fp16 = mul(x = var_103_cast_fp16_24, y = var_257_cast_fp16)[name = tensor("op_6674_cast_fp16")]; tensor key_97_cast_fp16 = add(x = var_6672_cast_fp16, y = var_6674_cast_fp16)[name = tensor("key_97_cast_fp16")]; tensor var_6676_cast_fp16 = mul(x = current_value_49_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_6676_cast_fp16")]; tensor var_6678_cast_fp16 = mul(x = var_138_cast_fp16_24, y = var_257_cast_fp16)[name = tensor("op_6678_cast_fp16")]; tensor value_97_cast_fp16 = add(x = var_6676_cast_fp16, y = var_6678_cast_fp16)[name = tensor("value_97_cast_fp16")]; tensor var_6681 = const()[name = tensor("op_6681"), val = tensor([1, 20, 64, -1])]; tensor var_6682_cast_fp16 = reshape(shape = var_6681, x = query_97_cast_fp16)[name = tensor("op_6682_cast_fp16")]; tensor var_6683_to_fp16 = const()[name = tensor("op_6683_to_fp16"), val = tensor(0x1p-3)]; tensor var_6684_cast_fp16 = mul(x = var_6682_cast_fp16, y = var_6683_to_fp16)[name = tensor("op_6684_cast_fp16")]; tensor var_6685 = const()[name = tensor("op_6685"), val = tensor([1, 20, 64, -1])]; tensor var_6686_cast_fp16 = reshape(shape = var_6685, x = key_97_cast_fp16)[name = tensor("op_6686_cast_fp16")]; tensor mh_w_145_transpose_x_0 = const()[name = tensor("mh_w_145_transpose_x_0"), val = tensor(true)]; tensor mh_w_145_transpose_y_0 = const()[name = tensor("mh_w_145_transpose_y_0"), val = tensor(false)]; tensor mh_w_145_cast_fp16 = matmul(transpose_x = mh_w_145_transpose_x_0, transpose_y = mh_w_145_transpose_y_0, x = var_6684_cast_fp16, y = var_6686_cast_fp16)[name = tensor("mh_w_145_cast_fp16")]; tensor mh_w_147_cast_fp16 = add(x = mh_w_145_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_147_cast_fp16")]; tensor var_6694_cast_fp16 = softmax(axis = var_6592, x = mh_w_147_cast_fp16)[name = tensor("op_6694_cast_fp16")]; tensor var_6695 = const()[name = tensor("op_6695"), val = tensor([1, 20, 64, -1])]; tensor var_6696_cast_fp16 = reshape(shape = var_6695, x = value_97_cast_fp16)[name = tensor("op_6696_cast_fp16")]; tensor attn_97_transpose_x_0 = const()[name = tensor("attn_97_transpose_x_0"), val = tensor(false)]; tensor attn_97_transpose_y_0 = const()[name = tensor("attn_97_transpose_y_0"), val = tensor(true)]; tensor attn_97_cast_fp16 = matmul(transpose_x = attn_97_transpose_x_0, transpose_y = attn_97_transpose_y_0, x = var_6696_cast_fp16, y = var_6694_cast_fp16)[name = tensor("attn_97_cast_fp16")]; tensor var_6699 = const()[name = tensor("op_6699"), val = tensor([1, 1280, 1, -1])]; tensor x_679_cast_fp16 = reshape(shape = var_6699, x = attn_97_cast_fp16)[name = tensor("x_679_cast_fp16")]; tensor layers_24_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_24_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453600768)))]; tensor input_535_cast_fp16 = sub(x = x_679_cast_fp16, y = layers_24_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_535_cast_fp16")]; tensor var_6707 = const()[name = tensor("op_6707"), val = tensor([1, 1])]; tensor var_6709 = const()[name = tensor("op_6709"), val = tensor([1, 1])]; tensor x_681_pad_type_0 = const()[name = tensor("x_681_pad_type_0"), val = tensor("custom")]; tensor x_681_pad_0 = const()[name = tensor("x_681_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_24_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453603392))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454422656))), name = tensor("layers_24_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_24_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454422784)))]; tensor x_681_cast_fp16 = conv(bias = layers_24_self_attn_o_proj_module_bias_to_fp16, dilations = var_6709, groups = var_6599, pad = x_681_pad_0, pad_type = x_681_pad_type_0, strides = var_6707, weight = layers_24_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = tensor("x_681_cast_fp16")]; tensor layers_24_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_24_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454425408)))]; tensor obj_343_cast_fp16 = mul(x = x_681_cast_fp16, y = layers_24_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_343_cast_fp16")]; tensor inputs_147_cast_fp16 = add(x = inputs_145_cast_fp16, y = obj_343_cast_fp16)[name = tensor("inputs_147_cast_fp16")]; tensor var_6720 = const()[name = tensor("op_6720"), val = tensor([1])]; tensor channels_mean_147_cast_fp16 = reduce_mean(axes = var_6720, keep_dims = var_6600, x = inputs_147_cast_fp16)[name = tensor("channels_mean_147_cast_fp16")]; tensor zero_mean_147_cast_fp16 = sub(x = inputs_147_cast_fp16, y = channels_mean_147_cast_fp16)[name = tensor("zero_mean_147_cast_fp16")]; tensor zero_mean_sq_147_cast_fp16 = mul(x = zero_mean_147_cast_fp16, y = zero_mean_147_cast_fp16)[name = tensor("zero_mean_sq_147_cast_fp16")]; tensor var_6724 = const()[name = tensor("op_6724"), val = tensor([1])]; tensor var_6725_cast_fp16 = reduce_mean(axes = var_6724, keep_dims = var_6600, x = zero_mean_sq_147_cast_fp16)[name = tensor("op_6725_cast_fp16")]; tensor var_6726_to_fp16 = const()[name = tensor("op_6726_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_6727_cast_fp16 = add(x = var_6725_cast_fp16, y = var_6726_to_fp16)[name = tensor("op_6727_cast_fp16")]; tensor denom_147_epsilon_0_to_fp16 = const()[name = tensor("denom_147_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_147_cast_fp16 = rsqrt(epsilon = denom_147_epsilon_0_to_fp16, x = var_6727_cast_fp16)[name = tensor("denom_147_cast_fp16")]; tensor out_147_cast_fp16 = mul(x = zero_mean_147_cast_fp16, y = denom_147_cast_fp16)[name = tensor("out_147_cast_fp16")]; tensor obj_345_gamma_0_to_fp16 = const()[name = tensor("obj_345_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454428032)))]; tensor obj_345_beta_0_to_fp16 = const()[name = tensor("obj_345_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454430656)))]; tensor obj_345_epsilon_0_to_fp16 = const()[name = tensor("obj_345_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_345_cast_fp16 = batch_norm(beta = obj_345_beta_0_to_fp16, epsilon = obj_345_epsilon_0_to_fp16, gamma = obj_345_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_147_cast_fp16)[name = tensor("obj_345_cast_fp16")]; tensor layers_24_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_24_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454433280)))]; tensor input_537_cast_fp16 = sub(x = obj_345_cast_fp16, y = layers_24_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_537_cast_fp16")]; tensor var_6746 = const()[name = tensor("op_6746"), val = tensor([1, 1])]; tensor var_6748 = const()[name = tensor("op_6748"), val = tensor([1, 1])]; tensor x_683_pad_type_0 = const()[name = tensor("x_683_pad_type_0"), val = tensor("custom")]; tensor x_683_pad_0 = const()[name = tensor("x_683_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_24_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454435904))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455255168))), name = tensor("layers_24_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_24_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_24_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455255296)))]; tensor x_683_cast_fp16 = conv(bias = layers_24_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_6748, groups = var_6599, pad = x_683_pad_0, pad_type = x_683_pad_type_0, strides = var_6746, weight = layers_24_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = tensor("x_683_cast_fp16")]; tensor layers_24_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_24_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455257920)))]; tensor query_99_cast_fp16 = mul(x = x_683_cast_fp16, y = layers_24_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_99_cast_fp16")]; tensor var_6758 = const()[name = tensor("op_6758"), val = tensor([1, 1])]; tensor var_6760 = const()[name = tensor("op_6760"), val = tensor([1, 1])]; tensor x_685_pad_type_0 = const()[name = tensor("x_685_pad_type_0"), val = tensor("custom")]; tensor x_685_pad_0 = const()[name = tensor("x_685_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_24_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455260544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(456079808))), name = tensor("layers_24_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_24_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_24_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(456079936)))]; tensor x_685_cast_fp16 = conv(bias = layers_24_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_6760, groups = var_6599, pad = x_685_pad_0, pad_type = x_685_pad_type_0, strides = var_6758, weight = layers_24_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_685_cast_fp16")]; tensor layers_24_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_24_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(456082560)))]; tensor key_99_cast_fp16 = mul(x = x_685_cast_fp16, y = layers_24_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_99_cast_fp16")]; tensor var_6770 = const()[name = tensor("op_6770"), val = tensor([1, 1])]; tensor var_6772 = const()[name = tensor("op_6772"), val = tensor([1, 1])]; tensor x_687_pad_type_0 = const()[name = tensor("x_687_pad_type_0"), val = tensor("custom")]; tensor x_687_pad_0 = const()[name = tensor("x_687_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_24_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(456085184))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(456904448))), name = tensor("layers_24_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_24_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_24_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(456904576)))]; tensor x_687_cast_fp16 = conv(bias = layers_24_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_6772, groups = var_6599, pad = x_687_pad_0, pad_type = x_687_pad_type_0, strides = var_6770, weight = layers_24_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_687_cast_fp16")]; tensor layers_24_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_24_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(456907200)))]; tensor value_99_cast_fp16 = mul(x = x_687_cast_fp16, y = layers_24_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_99_cast_fp16")]; tensor var_6777 = const()[name = tensor("op_6777"), val = tensor([1, 20, 64, -1])]; tensor var_6778_cast_fp16 = reshape(shape = var_6777, x = query_99_cast_fp16)[name = tensor("op_6778_cast_fp16")]; tensor var_6779_to_fp16 = const()[name = tensor("op_6779_to_fp16"), val = tensor(0x1p-3)]; tensor var_6780_cast_fp16 = mul(x = var_6778_cast_fp16, y = var_6779_to_fp16)[name = tensor("op_6780_cast_fp16")]; tensor var_6781 = const()[name = tensor("op_6781"), val = tensor([1, 20, 64, -1])]; tensor var_6782_cast_fp16 = reshape(shape = var_6781, x = key_99_cast_fp16)[name = tensor("op_6782_cast_fp16")]; tensor mh_w_149_transpose_x_0 = const()[name = tensor("mh_w_149_transpose_x_0"), val = tensor(true)]; tensor mh_w_149_transpose_y_0 = const()[name = tensor("mh_w_149_transpose_y_0"), val = tensor(false)]; tensor mh_w_149_cast_fp16 = matmul(transpose_x = mh_w_149_transpose_x_0, transpose_y = mh_w_149_transpose_y_0, x = var_6780_cast_fp16, y = var_6782_cast_fp16)[name = tensor("mh_w_149_cast_fp16")]; tensor obj_349_cast_fp16 = softmax(axis = var_6592, x = mh_w_149_cast_fp16)[name = tensor("obj_349_cast_fp16")]; tensor var_6786 = const()[name = tensor("op_6786"), val = tensor([1, 20, 64, -1])]; tensor var_6787_cast_fp16 = reshape(shape = var_6786, x = value_99_cast_fp16)[name = tensor("op_6787_cast_fp16")]; tensor attn_99_transpose_x_0 = const()[name = tensor("attn_99_transpose_x_0"), val = tensor(false)]; tensor attn_99_transpose_y_0 = const()[name = tensor("attn_99_transpose_y_0"), val = tensor(true)]; tensor attn_99_cast_fp16 = matmul(transpose_x = attn_99_transpose_x_0, transpose_y = attn_99_transpose_y_0, x = var_6787_cast_fp16, y = obj_349_cast_fp16)[name = tensor("attn_99_cast_fp16")]; tensor var_6790 = const()[name = tensor("op_6790"), val = tensor([1, 1280, 1, -1])]; tensor x_689_cast_fp16 = reshape(shape = var_6790, x = attn_99_cast_fp16)[name = tensor("x_689_cast_fp16")]; tensor layers_24_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_24_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(456909824)))]; tensor input_543_cast_fp16 = sub(x = x_689_cast_fp16, y = layers_24_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_543_cast_fp16")]; tensor var_6798 = const()[name = tensor("op_6798"), val = tensor([1, 1])]; tensor var_6800 = const()[name = tensor("op_6800"), val = tensor([1, 1])]; tensor x_691_pad_type_0 = const()[name = tensor("x_691_pad_type_0"), val = tensor("custom")]; tensor x_691_pad_0 = const()[name = tensor("x_691_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_24_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(456912448))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457731712))), name = tensor("layers_24_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_24_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_24_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457731840)))]; tensor x_691_cast_fp16 = conv(bias = layers_24_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_6800, groups = var_6599, pad = x_691_pad_0, pad_type = x_691_pad_type_0, strides = var_6798, weight = layers_24_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_543_cast_fp16)[name = tensor("x_691_cast_fp16")]; tensor layers_24_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_24_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457734464)))]; tensor obj_347_cast_fp16 = mul(x = x_691_cast_fp16, y = layers_24_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_347_cast_fp16")]; tensor inputs_149_cast_fp16 = add(x = inputs_147_cast_fp16, y = obj_347_cast_fp16)[name = tensor("inputs_149_cast_fp16")]; tensor var_6810 = const()[name = tensor("op_6810"), val = tensor([1])]; tensor channels_mean_149_cast_fp16 = reduce_mean(axes = var_6810, keep_dims = var_6600, x = inputs_149_cast_fp16)[name = tensor("channels_mean_149_cast_fp16")]; tensor zero_mean_149_cast_fp16 = sub(x = inputs_149_cast_fp16, y = channels_mean_149_cast_fp16)[name = tensor("zero_mean_149_cast_fp16")]; tensor zero_mean_sq_149_cast_fp16 = mul(x = zero_mean_149_cast_fp16, y = zero_mean_149_cast_fp16)[name = tensor("zero_mean_sq_149_cast_fp16")]; tensor var_6814 = const()[name = tensor("op_6814"), val = tensor([1])]; tensor var_6815_cast_fp16 = reduce_mean(axes = var_6814, keep_dims = var_6600, x = zero_mean_sq_149_cast_fp16)[name = tensor("op_6815_cast_fp16")]; tensor var_6816_to_fp16 = const()[name = tensor("op_6816_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_6817_cast_fp16 = add(x = var_6815_cast_fp16, y = var_6816_to_fp16)[name = tensor("op_6817_cast_fp16")]; tensor denom_149_epsilon_0_to_fp16 = const()[name = tensor("denom_149_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_149_cast_fp16 = rsqrt(epsilon = denom_149_epsilon_0_to_fp16, x = var_6817_cast_fp16)[name = tensor("denom_149_cast_fp16")]; tensor out_149_cast_fp16 = mul(x = zero_mean_149_cast_fp16, y = denom_149_cast_fp16)[name = tensor("out_149_cast_fp16")]; tensor x_693_gamma_0_to_fp16 = const()[name = tensor("x_693_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457737088)))]; tensor x_693_beta_0_to_fp16 = const()[name = tensor("x_693_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457739712)))]; tensor x_693_epsilon_0_to_fp16 = const()[name = tensor("x_693_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_693_cast_fp16 = batch_norm(beta = x_693_beta_0_to_fp16, epsilon = x_693_epsilon_0_to_fp16, gamma = x_693_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_149_cast_fp16)[name = tensor("x_693_cast_fp16")]; tensor layers_24_fc1_input_shift_to_fp16 = const()[name = tensor("layers_24_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457742336)))]; tensor input_545_cast_fp16 = sub(x = x_693_cast_fp16, y = layers_24_fc1_input_shift_to_fp16)[name = tensor("input_545_cast_fp16")]; tensor var_6832 = const()[name = tensor("op_6832"), val = tensor([1, 1])]; tensor var_6834 = const()[name = tensor("op_6834"), val = tensor([1, 1])]; tensor x_695_pad_type_0 = const()[name = tensor("x_695_pad_type_0"), val = tensor("custom")]; tensor x_695_pad_0 = const()[name = tensor("x_695_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_24_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457744960))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461021824))), name = tensor("layers_24_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_24_fc1_module_bias_to_fp16 = const()[name = tensor("layers_24_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461021952)))]; tensor x_695_cast_fp16 = conv(bias = layers_24_fc1_module_bias_to_fp16, dilations = var_6834, groups = var_6599, pad = x_695_pad_0, pad_type = x_695_pad_type_0, strides = var_6832, weight = layers_24_fc1_module_weight_to_fp16_palettized, x = input_545_cast_fp16)[name = tensor("x_695_cast_fp16")]; tensor layers_24_fc1_output_scale_to_fp16 = const()[name = tensor("layers_24_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461032256)))]; tensor input_547_cast_fp16 = mul(x = x_695_cast_fp16, y = layers_24_fc1_output_scale_to_fp16)[name = tensor("input_547_cast_fp16")]; tensor x_697_mode_0 = const()[name = tensor("x_697_mode_0"), val = tensor("EXACT")]; tensor x_697_cast_fp16 = gelu(mode = x_697_mode_0, x = input_547_cast_fp16)[name = tensor("x_697_cast_fp16")]; tensor layers_24_fc2_input_shift_to_fp16 = const()[name = tensor("layers_24_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461042560)))]; tensor input_549_cast_fp16 = sub(x = x_697_cast_fp16, y = layers_24_fc2_input_shift_to_fp16)[name = tensor("input_549_cast_fp16")]; tensor var_6845 = const()[name = tensor("op_6845"), val = tensor([1, 1])]; tensor var_6847 = const()[name = tensor("op_6847"), val = tensor([1, 1])]; tensor x_699_pad_type_0 = const()[name = tensor("x_699_pad_type_0"), val = tensor("custom")]; tensor x_699_pad_0 = const()[name = tensor("x_699_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_24_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461052864))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(464329728))), name = tensor("layers_24_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_24_fc2_module_bias_to_fp16 = const()[name = tensor("layers_24_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(464329856)))]; tensor x_699_cast_fp16 = conv(bias = layers_24_fc2_module_bias_to_fp16, dilations = var_6847, groups = var_6599, pad = x_699_pad_0, pad_type = x_699_pad_type_0, strides = var_6845, weight = layers_24_fc2_module_weight_to_fp16_palettized, x = input_549_cast_fp16)[name = tensor("x_699_cast_fp16")]; tensor layers_24_fc2_output_scale_to_fp16 = const()[name = tensor("layers_24_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(464332480)))]; tensor hidden_states_51_cast_fp16 = mul(x = x_699_cast_fp16, y = layers_24_fc2_output_scale_to_fp16)[name = tensor("hidden_states_51_cast_fp16")]; tensor inputs_151_cast_fp16 = add(x = inputs_149_cast_fp16, y = hidden_states_51_cast_fp16)[name = tensor("inputs_151_cast_fp16")]; tensor var_6862 = const()[name = tensor("op_6862"), val = tensor(3)]; tensor var_6869 = const()[name = tensor("op_6869"), val = tensor(1)]; tensor var_6870 = const()[name = tensor("op_6870"), val = tensor(true)]; tensor var_6882 = const()[name = tensor("op_6882"), val = tensor([1])]; tensor channels_mean_151_cast_fp16 = reduce_mean(axes = var_6882, keep_dims = var_6870, x = inputs_151_cast_fp16)[name = tensor("channels_mean_151_cast_fp16")]; tensor zero_mean_151_cast_fp16 = sub(x = inputs_151_cast_fp16, y = channels_mean_151_cast_fp16)[name = tensor("zero_mean_151_cast_fp16")]; tensor zero_mean_sq_151_cast_fp16 = mul(x = zero_mean_151_cast_fp16, y = zero_mean_151_cast_fp16)[name = tensor("zero_mean_sq_151_cast_fp16")]; tensor var_6886 = const()[name = tensor("op_6886"), val = tensor([1])]; tensor var_6887_cast_fp16 = reduce_mean(axes = var_6886, keep_dims = var_6870, x = zero_mean_sq_151_cast_fp16)[name = tensor("op_6887_cast_fp16")]; tensor var_6888_to_fp16 = const()[name = tensor("op_6888_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_6889_cast_fp16 = add(x = var_6887_cast_fp16, y = var_6888_to_fp16)[name = tensor("op_6889_cast_fp16")]; tensor denom_151_epsilon_0_to_fp16 = const()[name = tensor("denom_151_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_151_cast_fp16 = rsqrt(epsilon = denom_151_epsilon_0_to_fp16, x = var_6889_cast_fp16)[name = tensor("denom_151_cast_fp16")]; tensor out_151_cast_fp16 = mul(x = zero_mean_151_cast_fp16, y = denom_151_cast_fp16)[name = tensor("out_151_cast_fp16")]; tensor obj_351_gamma_0_to_fp16 = const()[name = tensor("obj_351_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(464335104)))]; tensor obj_351_beta_0_to_fp16 = const()[name = tensor("obj_351_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(464337728)))]; tensor obj_351_epsilon_0_to_fp16 = const()[name = tensor("obj_351_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_351_cast_fp16 = batch_norm(beta = obj_351_beta_0_to_fp16, epsilon = obj_351_epsilon_0_to_fp16, gamma = obj_351_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_151_cast_fp16)[name = tensor("obj_351_cast_fp16")]; tensor layers_25_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_25_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(464340352)))]; tensor input_551_cast_fp16 = sub(x = obj_351_cast_fp16, y = layers_25_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_551_cast_fp16")]; tensor var_6908 = const()[name = tensor("op_6908"), val = tensor([1, 1])]; tensor var_6910 = const()[name = tensor("op_6910"), val = tensor([1, 1])]; tensor x_701_pad_type_0 = const()[name = tensor("x_701_pad_type_0"), val = tensor("custom")]; tensor x_701_pad_0 = const()[name = tensor("x_701_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_25_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(464342976))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465162240))), name = tensor("layers_25_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_25_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465162368)))]; tensor x_701_cast_fp16 = conv(bias = layers_25_self_attn_q_proj_module_bias_to_fp16, dilations = var_6910, groups = var_6869, pad = x_701_pad_0, pad_type = x_701_pad_type_0, strides = var_6908, weight = layers_25_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_551_cast_fp16)[name = tensor("x_701_cast_fp16")]; tensor layers_25_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_25_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465164992)))]; tensor query_101_cast_fp16 = mul(x = x_701_cast_fp16, y = layers_25_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_101_cast_fp16")]; tensor var_6920 = const()[name = tensor("op_6920"), val = tensor([1, 1])]; tensor var_6922 = const()[name = tensor("op_6922"), val = tensor([1, 1])]; tensor x_703_pad_type_0 = const()[name = tensor("x_703_pad_type_0"), val = tensor("custom")]; tensor x_703_pad_0 = const()[name = tensor("x_703_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_25_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465167616))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465986880))), name = tensor("layers_25_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_25_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465987008)))]; tensor x_703_cast_fp16 = conv(bias = layers_25_self_attn_k_proj_module_bias_to_fp16, dilations = var_6922, groups = var_6869, pad = x_703_pad_0, pad_type = x_703_pad_type_0, strides = var_6920, weight = layers_25_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_551_cast_fp16)[name = tensor("x_703_cast_fp16")]; tensor layers_25_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_25_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465989632)))]; tensor current_key_51_cast_fp16 = mul(x = x_703_cast_fp16, y = layers_25_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_51_cast_fp16")]; tensor var_6932 = const()[name = tensor("op_6932"), val = tensor([1, 1])]; tensor var_6934 = const()[name = tensor("op_6934"), val = tensor([1, 1])]; tensor x_705_pad_type_0 = const()[name = tensor("x_705_pad_type_0"), val = tensor("custom")]; tensor x_705_pad_0 = const()[name = tensor("x_705_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_25_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465992256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466811520))), name = tensor("layers_25_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_25_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466811648)))]; tensor x_705_cast_fp16 = conv(bias = layers_25_self_attn_v_proj_module_bias_to_fp16, dilations = var_6934, groups = var_6869, pad = x_705_pad_0, pad_type = x_705_pad_type_0, strides = var_6932, weight = layers_25_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_551_cast_fp16)[name = tensor("x_705_cast_fp16")]; tensor layers_25_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_25_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466814272)))]; tensor current_value_51_cast_fp16 = mul(x = x_705_cast_fp16, y = layers_25_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_51_cast_fp16")]; tensor var_6942_cast_fp16 = mul(x = current_key_51_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_6942_cast_fp16")]; tensor var_6944_cast_fp16 = mul(x = var_103_cast_fp16_25, y = var_257_cast_fp16)[name = tensor("op_6944_cast_fp16")]; tensor key_101_cast_fp16 = add(x = var_6942_cast_fp16, y = var_6944_cast_fp16)[name = tensor("key_101_cast_fp16")]; tensor var_6946_cast_fp16 = mul(x = current_value_51_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_6946_cast_fp16")]; tensor var_6948_cast_fp16 = mul(x = var_138_cast_fp16_25, y = var_257_cast_fp16)[name = tensor("op_6948_cast_fp16")]; tensor value_101_cast_fp16 = add(x = var_6946_cast_fp16, y = var_6948_cast_fp16)[name = tensor("value_101_cast_fp16")]; tensor var_6951 = const()[name = tensor("op_6951"), val = tensor([1, 20, 64, -1])]; tensor var_6952_cast_fp16 = reshape(shape = var_6951, x = query_101_cast_fp16)[name = tensor("op_6952_cast_fp16")]; tensor var_6953_to_fp16 = const()[name = tensor("op_6953_to_fp16"), val = tensor(0x1p-3)]; tensor var_6954_cast_fp16 = mul(x = var_6952_cast_fp16, y = var_6953_to_fp16)[name = tensor("op_6954_cast_fp16")]; tensor var_6955 = const()[name = tensor("op_6955"), val = tensor([1, 20, 64, -1])]; tensor var_6956_cast_fp16 = reshape(shape = var_6955, x = key_101_cast_fp16)[name = tensor("op_6956_cast_fp16")]; tensor mh_w_151_transpose_x_0 = const()[name = tensor("mh_w_151_transpose_x_0"), val = tensor(true)]; tensor mh_w_151_transpose_y_0 = const()[name = tensor("mh_w_151_transpose_y_0"), val = tensor(false)]; tensor mh_w_151_cast_fp16 = matmul(transpose_x = mh_w_151_transpose_x_0, transpose_y = mh_w_151_transpose_y_0, x = var_6954_cast_fp16, y = var_6956_cast_fp16)[name = tensor("mh_w_151_cast_fp16")]; tensor mh_w_153_cast_fp16 = add(x = mh_w_151_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_153_cast_fp16")]; tensor var_6964_cast_fp16 = softmax(axis = var_6862, x = mh_w_153_cast_fp16)[name = tensor("op_6964_cast_fp16")]; tensor var_6965 = const()[name = tensor("op_6965"), val = tensor([1, 20, 64, -1])]; tensor var_6966_cast_fp16 = reshape(shape = var_6965, x = value_101_cast_fp16)[name = tensor("op_6966_cast_fp16")]; tensor attn_101_transpose_x_0 = const()[name = tensor("attn_101_transpose_x_0"), val = tensor(false)]; tensor attn_101_transpose_y_0 = const()[name = tensor("attn_101_transpose_y_0"), val = tensor(true)]; tensor attn_101_cast_fp16 = matmul(transpose_x = attn_101_transpose_x_0, transpose_y = attn_101_transpose_y_0, x = var_6966_cast_fp16, y = var_6964_cast_fp16)[name = tensor("attn_101_cast_fp16")]; tensor var_6969 = const()[name = tensor("op_6969"), val = tensor([1, 1280, 1, -1])]; tensor x_707_cast_fp16 = reshape(shape = var_6969, x = attn_101_cast_fp16)[name = tensor("x_707_cast_fp16")]; tensor layers_25_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_25_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466816896)))]; tensor input_557_cast_fp16 = sub(x = x_707_cast_fp16, y = layers_25_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_557_cast_fp16")]; tensor var_6977 = const()[name = tensor("op_6977"), val = tensor([1, 1])]; tensor var_6979 = const()[name = tensor("op_6979"), val = tensor([1, 1])]; tensor x_709_pad_type_0 = const()[name = tensor("x_709_pad_type_0"), val = tensor("custom")]; tensor x_709_pad_0 = const()[name = tensor("x_709_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_25_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466819520))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467638784))), name = tensor("layers_25_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_25_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467638912)))]; tensor x_709_cast_fp16 = conv(bias = layers_25_self_attn_o_proj_module_bias_to_fp16, dilations = var_6979, groups = var_6869, pad = x_709_pad_0, pad_type = x_709_pad_type_0, strides = var_6977, weight = layers_25_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_557_cast_fp16)[name = tensor("x_709_cast_fp16")]; tensor layers_25_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_25_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467641536)))]; tensor obj_357_cast_fp16 = mul(x = x_709_cast_fp16, y = layers_25_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_357_cast_fp16")]; tensor inputs_153_cast_fp16 = add(x = inputs_151_cast_fp16, y = obj_357_cast_fp16)[name = tensor("inputs_153_cast_fp16")]; tensor var_6990 = const()[name = tensor("op_6990"), val = tensor([1])]; tensor channels_mean_153_cast_fp16 = reduce_mean(axes = var_6990, keep_dims = var_6870, x = inputs_153_cast_fp16)[name = tensor("channels_mean_153_cast_fp16")]; tensor zero_mean_153_cast_fp16 = sub(x = inputs_153_cast_fp16, y = channels_mean_153_cast_fp16)[name = tensor("zero_mean_153_cast_fp16")]; tensor zero_mean_sq_153_cast_fp16 = mul(x = zero_mean_153_cast_fp16, y = zero_mean_153_cast_fp16)[name = tensor("zero_mean_sq_153_cast_fp16")]; tensor var_6994 = const()[name = tensor("op_6994"), val = tensor([1])]; tensor var_6995_cast_fp16 = reduce_mean(axes = var_6994, keep_dims = var_6870, x = zero_mean_sq_153_cast_fp16)[name = tensor("op_6995_cast_fp16")]; tensor var_6996_to_fp16 = const()[name = tensor("op_6996_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_6997_cast_fp16 = add(x = var_6995_cast_fp16, y = var_6996_to_fp16)[name = tensor("op_6997_cast_fp16")]; tensor denom_153_epsilon_0_to_fp16 = const()[name = tensor("denom_153_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_153_cast_fp16 = rsqrt(epsilon = denom_153_epsilon_0_to_fp16, x = var_6997_cast_fp16)[name = tensor("denom_153_cast_fp16")]; tensor out_153_cast_fp16 = mul(x = zero_mean_153_cast_fp16, y = denom_153_cast_fp16)[name = tensor("out_153_cast_fp16")]; tensor obj_359_gamma_0_to_fp16 = const()[name = tensor("obj_359_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467644160)))]; tensor obj_359_beta_0_to_fp16 = const()[name = tensor("obj_359_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467646784)))]; tensor obj_359_epsilon_0_to_fp16 = const()[name = tensor("obj_359_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_359_cast_fp16 = batch_norm(beta = obj_359_beta_0_to_fp16, epsilon = obj_359_epsilon_0_to_fp16, gamma = obj_359_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_153_cast_fp16)[name = tensor("obj_359_cast_fp16")]; tensor layers_25_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_25_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467649408)))]; tensor input_559_cast_fp16 = sub(x = obj_359_cast_fp16, y = layers_25_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_559_cast_fp16")]; tensor var_7016 = const()[name = tensor("op_7016"), val = tensor([1, 1])]; tensor var_7018 = const()[name = tensor("op_7018"), val = tensor([1, 1])]; tensor x_711_pad_type_0 = const()[name = tensor("x_711_pad_type_0"), val = tensor("custom")]; tensor x_711_pad_0 = const()[name = tensor("x_711_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_25_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467652032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468471296))), name = tensor("layers_25_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_25_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_25_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468471424)))]; tensor x_711_cast_fp16 = conv(bias = layers_25_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_7018, groups = var_6869, pad = x_711_pad_0, pad_type = x_711_pad_type_0, strides = var_7016, weight = layers_25_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = tensor("x_711_cast_fp16")]; tensor layers_25_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_25_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468474048)))]; tensor query_103_cast_fp16 = mul(x = x_711_cast_fp16, y = layers_25_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_103_cast_fp16")]; tensor var_7028 = const()[name = tensor("op_7028"), val = tensor([1, 1])]; tensor var_7030 = const()[name = tensor("op_7030"), val = tensor([1, 1])]; tensor x_713_pad_type_0 = const()[name = tensor("x_713_pad_type_0"), val = tensor("custom")]; tensor x_713_pad_0 = const()[name = tensor("x_713_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_25_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468476672))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469295936))), name = tensor("layers_25_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_25_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_25_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469296064)))]; tensor x_713_cast_fp16 = conv(bias = layers_25_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_7030, groups = var_6869, pad = x_713_pad_0, pad_type = x_713_pad_type_0, strides = var_7028, weight = layers_25_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_713_cast_fp16")]; tensor layers_25_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_25_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469298688)))]; tensor key_103_cast_fp16 = mul(x = x_713_cast_fp16, y = layers_25_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_103_cast_fp16")]; tensor var_7040 = const()[name = tensor("op_7040"), val = tensor([1, 1])]; tensor var_7042 = const()[name = tensor("op_7042"), val = tensor([1, 1])]; tensor x_715_pad_type_0 = const()[name = tensor("x_715_pad_type_0"), val = tensor("custom")]; tensor x_715_pad_0 = const()[name = tensor("x_715_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_25_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469301312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470120576))), name = tensor("layers_25_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_25_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_25_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470120704)))]; tensor x_715_cast_fp16 = conv(bias = layers_25_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_7042, groups = var_6869, pad = x_715_pad_0, pad_type = x_715_pad_type_0, strides = var_7040, weight = layers_25_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_715_cast_fp16")]; tensor layers_25_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_25_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470123328)))]; tensor value_103_cast_fp16 = mul(x = x_715_cast_fp16, y = layers_25_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_103_cast_fp16")]; tensor var_7047 = const()[name = tensor("op_7047"), val = tensor([1, 20, 64, -1])]; tensor var_7048_cast_fp16 = reshape(shape = var_7047, x = query_103_cast_fp16)[name = tensor("op_7048_cast_fp16")]; tensor var_7049_to_fp16 = const()[name = tensor("op_7049_to_fp16"), val = tensor(0x1p-3)]; tensor var_7050_cast_fp16 = mul(x = var_7048_cast_fp16, y = var_7049_to_fp16)[name = tensor("op_7050_cast_fp16")]; tensor var_7051 = const()[name = tensor("op_7051"), val = tensor([1, 20, 64, -1])]; tensor var_7052_cast_fp16 = reshape(shape = var_7051, x = key_103_cast_fp16)[name = tensor("op_7052_cast_fp16")]; tensor mh_w_155_transpose_x_0 = const()[name = tensor("mh_w_155_transpose_x_0"), val = tensor(true)]; tensor mh_w_155_transpose_y_0 = const()[name = tensor("mh_w_155_transpose_y_0"), val = tensor(false)]; tensor mh_w_155_cast_fp16 = matmul(transpose_x = mh_w_155_transpose_x_0, transpose_y = mh_w_155_transpose_y_0, x = var_7050_cast_fp16, y = var_7052_cast_fp16)[name = tensor("mh_w_155_cast_fp16")]; tensor obj_363_cast_fp16 = softmax(axis = var_6862, x = mh_w_155_cast_fp16)[name = tensor("obj_363_cast_fp16")]; tensor var_7056 = const()[name = tensor("op_7056"), val = tensor([1, 20, 64, -1])]; tensor var_7057_cast_fp16 = reshape(shape = var_7056, x = value_103_cast_fp16)[name = tensor("op_7057_cast_fp16")]; tensor attn_103_transpose_x_0 = const()[name = tensor("attn_103_transpose_x_0"), val = tensor(false)]; tensor attn_103_transpose_y_0 = const()[name = tensor("attn_103_transpose_y_0"), val = tensor(true)]; tensor attn_103_cast_fp16 = matmul(transpose_x = attn_103_transpose_x_0, transpose_y = attn_103_transpose_y_0, x = var_7057_cast_fp16, y = obj_363_cast_fp16)[name = tensor("attn_103_cast_fp16")]; tensor var_7060 = const()[name = tensor("op_7060"), val = tensor([1, 1280, 1, -1])]; tensor x_717_cast_fp16 = reshape(shape = var_7060, x = attn_103_cast_fp16)[name = tensor("x_717_cast_fp16")]; tensor layers_25_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_25_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470125952)))]; tensor input_565_cast_fp16 = sub(x = x_717_cast_fp16, y = layers_25_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_565_cast_fp16")]; tensor var_7068 = const()[name = tensor("op_7068"), val = tensor([1, 1])]; tensor var_7070 = const()[name = tensor("op_7070"), val = tensor([1, 1])]; tensor x_719_pad_type_0 = const()[name = tensor("x_719_pad_type_0"), val = tensor("custom")]; tensor x_719_pad_0 = const()[name = tensor("x_719_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_25_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470128576))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470947840))), name = tensor("layers_25_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_25_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_25_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470947968)))]; tensor x_719_cast_fp16 = conv(bias = layers_25_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_7070, groups = var_6869, pad = x_719_pad_0, pad_type = x_719_pad_type_0, strides = var_7068, weight = layers_25_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_565_cast_fp16)[name = tensor("x_719_cast_fp16")]; tensor layers_25_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_25_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470950592)))]; tensor obj_361_cast_fp16 = mul(x = x_719_cast_fp16, y = layers_25_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_361_cast_fp16")]; tensor inputs_155_cast_fp16 = add(x = inputs_153_cast_fp16, y = obj_361_cast_fp16)[name = tensor("inputs_155_cast_fp16")]; tensor var_7080 = const()[name = tensor("op_7080"), val = tensor([1])]; tensor channels_mean_155_cast_fp16 = reduce_mean(axes = var_7080, keep_dims = var_6870, x = inputs_155_cast_fp16)[name = tensor("channels_mean_155_cast_fp16")]; tensor zero_mean_155_cast_fp16 = sub(x = inputs_155_cast_fp16, y = channels_mean_155_cast_fp16)[name = tensor("zero_mean_155_cast_fp16")]; tensor zero_mean_sq_155_cast_fp16 = mul(x = zero_mean_155_cast_fp16, y = zero_mean_155_cast_fp16)[name = tensor("zero_mean_sq_155_cast_fp16")]; tensor var_7084 = const()[name = tensor("op_7084"), val = tensor([1])]; tensor var_7085_cast_fp16 = reduce_mean(axes = var_7084, keep_dims = var_6870, x = zero_mean_sq_155_cast_fp16)[name = tensor("op_7085_cast_fp16")]; tensor var_7086_to_fp16 = const()[name = tensor("op_7086_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_7087_cast_fp16 = add(x = var_7085_cast_fp16, y = var_7086_to_fp16)[name = tensor("op_7087_cast_fp16")]; tensor denom_155_epsilon_0_to_fp16 = const()[name = tensor("denom_155_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_155_cast_fp16 = rsqrt(epsilon = denom_155_epsilon_0_to_fp16, x = var_7087_cast_fp16)[name = tensor("denom_155_cast_fp16")]; tensor out_155_cast_fp16 = mul(x = zero_mean_155_cast_fp16, y = denom_155_cast_fp16)[name = tensor("out_155_cast_fp16")]; tensor x_721_gamma_0_to_fp16 = const()[name = tensor("x_721_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470953216)))]; tensor x_721_beta_0_to_fp16 = const()[name = tensor("x_721_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470955840)))]; tensor x_721_epsilon_0_to_fp16 = const()[name = tensor("x_721_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_721_cast_fp16 = batch_norm(beta = x_721_beta_0_to_fp16, epsilon = x_721_epsilon_0_to_fp16, gamma = x_721_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_155_cast_fp16)[name = tensor("x_721_cast_fp16")]; tensor layers_25_fc1_input_shift_to_fp16 = const()[name = tensor("layers_25_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470958464)))]; tensor input_567_cast_fp16 = sub(x = x_721_cast_fp16, y = layers_25_fc1_input_shift_to_fp16)[name = tensor("input_567_cast_fp16")]; tensor var_7102 = const()[name = tensor("op_7102"), val = tensor([1, 1])]; tensor var_7104 = const()[name = tensor("op_7104"), val = tensor([1, 1])]; tensor x_723_pad_type_0 = const()[name = tensor("x_723_pad_type_0"), val = tensor("custom")]; tensor x_723_pad_0 = const()[name = tensor("x_723_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_25_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470961088))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(474237952))), name = tensor("layers_25_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_25_fc1_module_bias_to_fp16 = const()[name = tensor("layers_25_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(474238080)))]; tensor x_723_cast_fp16 = conv(bias = layers_25_fc1_module_bias_to_fp16, dilations = var_7104, groups = var_6869, pad = x_723_pad_0, pad_type = x_723_pad_type_0, strides = var_7102, weight = layers_25_fc1_module_weight_to_fp16_palettized, x = input_567_cast_fp16)[name = tensor("x_723_cast_fp16")]; tensor layers_25_fc1_output_scale_to_fp16 = const()[name = tensor("layers_25_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(474248384)))]; tensor input_569_cast_fp16 = mul(x = x_723_cast_fp16, y = layers_25_fc1_output_scale_to_fp16)[name = tensor("input_569_cast_fp16")]; tensor x_725_mode_0 = const()[name = tensor("x_725_mode_0"), val = tensor("EXACT")]; tensor x_725_cast_fp16 = gelu(mode = x_725_mode_0, x = input_569_cast_fp16)[name = tensor("x_725_cast_fp16")]; tensor layers_25_fc2_input_shift_to_fp16 = const()[name = tensor("layers_25_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(474258688)))]; tensor input_571_cast_fp16 = sub(x = x_725_cast_fp16, y = layers_25_fc2_input_shift_to_fp16)[name = tensor("input_571_cast_fp16")]; tensor var_7115 = const()[name = tensor("op_7115"), val = tensor([1, 1])]; tensor var_7117 = const()[name = tensor("op_7117"), val = tensor([1, 1])]; tensor x_727_pad_type_0 = const()[name = tensor("x_727_pad_type_0"), val = tensor("custom")]; tensor x_727_pad_0 = const()[name = tensor("x_727_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_25_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(474268992))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477545856))), name = tensor("layers_25_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_25_fc2_module_bias_to_fp16 = const()[name = tensor("layers_25_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477545984)))]; tensor x_727_cast_fp16 = conv(bias = layers_25_fc2_module_bias_to_fp16, dilations = var_7117, groups = var_6869, pad = x_727_pad_0, pad_type = x_727_pad_type_0, strides = var_7115, weight = layers_25_fc2_module_weight_to_fp16_palettized, x = input_571_cast_fp16)[name = tensor("x_727_cast_fp16")]; tensor layers_25_fc2_output_scale_to_fp16 = const()[name = tensor("layers_25_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477548608)))]; tensor hidden_states_53_cast_fp16 = mul(x = x_727_cast_fp16, y = layers_25_fc2_output_scale_to_fp16)[name = tensor("hidden_states_53_cast_fp16")]; tensor inputs_157_cast_fp16 = add(x = inputs_155_cast_fp16, y = hidden_states_53_cast_fp16)[name = tensor("inputs_157_cast_fp16")]; tensor var_7132 = const()[name = tensor("op_7132"), val = tensor(3)]; tensor var_7139 = const()[name = tensor("op_7139"), val = tensor(1)]; tensor var_7140 = const()[name = tensor("op_7140"), val = tensor(true)]; tensor var_7152 = const()[name = tensor("op_7152"), val = tensor([1])]; tensor channels_mean_157_cast_fp16 = reduce_mean(axes = var_7152, keep_dims = var_7140, x = inputs_157_cast_fp16)[name = tensor("channels_mean_157_cast_fp16")]; tensor zero_mean_157_cast_fp16 = sub(x = inputs_157_cast_fp16, y = channels_mean_157_cast_fp16)[name = tensor("zero_mean_157_cast_fp16")]; tensor zero_mean_sq_157_cast_fp16 = mul(x = zero_mean_157_cast_fp16, y = zero_mean_157_cast_fp16)[name = tensor("zero_mean_sq_157_cast_fp16")]; tensor var_7156 = const()[name = tensor("op_7156"), val = tensor([1])]; tensor var_7157_cast_fp16 = reduce_mean(axes = var_7156, keep_dims = var_7140, x = zero_mean_sq_157_cast_fp16)[name = tensor("op_7157_cast_fp16")]; tensor var_7158_to_fp16 = const()[name = tensor("op_7158_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_7159_cast_fp16 = add(x = var_7157_cast_fp16, y = var_7158_to_fp16)[name = tensor("op_7159_cast_fp16")]; tensor denom_157_epsilon_0_to_fp16 = const()[name = tensor("denom_157_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_157_cast_fp16 = rsqrt(epsilon = denom_157_epsilon_0_to_fp16, x = var_7159_cast_fp16)[name = tensor("denom_157_cast_fp16")]; tensor out_157_cast_fp16 = mul(x = zero_mean_157_cast_fp16, y = denom_157_cast_fp16)[name = tensor("out_157_cast_fp16")]; tensor obj_365_gamma_0_to_fp16 = const()[name = tensor("obj_365_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477551232)))]; tensor obj_365_beta_0_to_fp16 = const()[name = tensor("obj_365_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477553856)))]; tensor obj_365_epsilon_0_to_fp16 = const()[name = tensor("obj_365_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_365_cast_fp16 = batch_norm(beta = obj_365_beta_0_to_fp16, epsilon = obj_365_epsilon_0_to_fp16, gamma = obj_365_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_157_cast_fp16)[name = tensor("obj_365_cast_fp16")]; tensor layers_26_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_26_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477556480)))]; tensor input_573_cast_fp16 = sub(x = obj_365_cast_fp16, y = layers_26_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_573_cast_fp16")]; tensor var_7178 = const()[name = tensor("op_7178"), val = tensor([1, 1])]; tensor var_7180 = const()[name = tensor("op_7180"), val = tensor([1, 1])]; tensor x_729_pad_type_0 = const()[name = tensor("x_729_pad_type_0"), val = tensor("custom")]; tensor x_729_pad_0 = const()[name = tensor("x_729_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_26_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477559104))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478378368))), name = tensor("layers_26_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_26_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478378496)))]; tensor x_729_cast_fp16 = conv(bias = layers_26_self_attn_q_proj_module_bias_to_fp16, dilations = var_7180, groups = var_7139, pad = x_729_pad_0, pad_type = x_729_pad_type_0, strides = var_7178, weight = layers_26_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_573_cast_fp16)[name = tensor("x_729_cast_fp16")]; tensor layers_26_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_26_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478381120)))]; tensor query_105_cast_fp16 = mul(x = x_729_cast_fp16, y = layers_26_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_105_cast_fp16")]; tensor var_7190 = const()[name = tensor("op_7190"), val = tensor([1, 1])]; tensor var_7192 = const()[name = tensor("op_7192"), val = tensor([1, 1])]; tensor x_731_pad_type_0 = const()[name = tensor("x_731_pad_type_0"), val = tensor("custom")]; tensor x_731_pad_0 = const()[name = tensor("x_731_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_26_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478383744))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479203008))), name = tensor("layers_26_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_26_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479203136)))]; tensor x_731_cast_fp16 = conv(bias = layers_26_self_attn_k_proj_module_bias_to_fp16, dilations = var_7192, groups = var_7139, pad = x_731_pad_0, pad_type = x_731_pad_type_0, strides = var_7190, weight = layers_26_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_573_cast_fp16)[name = tensor("x_731_cast_fp16")]; tensor layers_26_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_26_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479205760)))]; tensor current_key_53_cast_fp16 = mul(x = x_731_cast_fp16, y = layers_26_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_53_cast_fp16")]; tensor var_7202 = const()[name = tensor("op_7202"), val = tensor([1, 1])]; tensor var_7204 = const()[name = tensor("op_7204"), val = tensor([1, 1])]; tensor x_733_pad_type_0 = const()[name = tensor("x_733_pad_type_0"), val = tensor("custom")]; tensor x_733_pad_0 = const()[name = tensor("x_733_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_26_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479208384))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480027648))), name = tensor("layers_26_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_26_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480027776)))]; tensor x_733_cast_fp16 = conv(bias = layers_26_self_attn_v_proj_module_bias_to_fp16, dilations = var_7204, groups = var_7139, pad = x_733_pad_0, pad_type = x_733_pad_type_0, strides = var_7202, weight = layers_26_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_573_cast_fp16)[name = tensor("x_733_cast_fp16")]; tensor layers_26_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_26_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480030400)))]; tensor current_value_53_cast_fp16 = mul(x = x_733_cast_fp16, y = layers_26_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_53_cast_fp16")]; tensor var_7212_cast_fp16 = mul(x = current_key_53_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_7212_cast_fp16")]; tensor var_7214_cast_fp16 = mul(x = var_103_cast_fp16_26, y = var_257_cast_fp16)[name = tensor("op_7214_cast_fp16")]; tensor key_105_cast_fp16 = add(x = var_7212_cast_fp16, y = var_7214_cast_fp16)[name = tensor("key_105_cast_fp16")]; tensor var_7216_cast_fp16 = mul(x = current_value_53_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_7216_cast_fp16")]; tensor var_7218_cast_fp16 = mul(x = var_138_cast_fp16_26, y = var_257_cast_fp16)[name = tensor("op_7218_cast_fp16")]; tensor value_105_cast_fp16 = add(x = var_7216_cast_fp16, y = var_7218_cast_fp16)[name = tensor("value_105_cast_fp16")]; tensor var_7221 = const()[name = tensor("op_7221"), val = tensor([1, 20, 64, -1])]; tensor var_7222_cast_fp16 = reshape(shape = var_7221, x = query_105_cast_fp16)[name = tensor("op_7222_cast_fp16")]; tensor var_7223_to_fp16 = const()[name = tensor("op_7223_to_fp16"), val = tensor(0x1p-3)]; tensor var_7224_cast_fp16 = mul(x = var_7222_cast_fp16, y = var_7223_to_fp16)[name = tensor("op_7224_cast_fp16")]; tensor var_7225 = const()[name = tensor("op_7225"), val = tensor([1, 20, 64, -1])]; tensor var_7226_cast_fp16 = reshape(shape = var_7225, x = key_105_cast_fp16)[name = tensor("op_7226_cast_fp16")]; tensor mh_w_157_transpose_x_0 = const()[name = tensor("mh_w_157_transpose_x_0"), val = tensor(true)]; tensor mh_w_157_transpose_y_0 = const()[name = tensor("mh_w_157_transpose_y_0"), val = tensor(false)]; tensor mh_w_157_cast_fp16 = matmul(transpose_x = mh_w_157_transpose_x_0, transpose_y = mh_w_157_transpose_y_0, x = var_7224_cast_fp16, y = var_7226_cast_fp16)[name = tensor("mh_w_157_cast_fp16")]; tensor mh_w_159_cast_fp16 = add(x = mh_w_157_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_159_cast_fp16")]; tensor var_7234_cast_fp16 = softmax(axis = var_7132, x = mh_w_159_cast_fp16)[name = tensor("op_7234_cast_fp16")]; tensor var_7235 = const()[name = tensor("op_7235"), val = tensor([1, 20, 64, -1])]; tensor var_7236_cast_fp16 = reshape(shape = var_7235, x = value_105_cast_fp16)[name = tensor("op_7236_cast_fp16")]; tensor attn_105_transpose_x_0 = const()[name = tensor("attn_105_transpose_x_0"), val = tensor(false)]; tensor attn_105_transpose_y_0 = const()[name = tensor("attn_105_transpose_y_0"), val = tensor(true)]; tensor attn_105_cast_fp16 = matmul(transpose_x = attn_105_transpose_x_0, transpose_y = attn_105_transpose_y_0, x = var_7236_cast_fp16, y = var_7234_cast_fp16)[name = tensor("attn_105_cast_fp16")]; tensor var_7239 = const()[name = tensor("op_7239"), val = tensor([1, 1280, 1, -1])]; tensor x_735_cast_fp16 = reshape(shape = var_7239, x = attn_105_cast_fp16)[name = tensor("x_735_cast_fp16")]; tensor layers_26_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_26_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480033024)))]; tensor input_579_cast_fp16 = sub(x = x_735_cast_fp16, y = layers_26_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_579_cast_fp16")]; tensor var_7247 = const()[name = tensor("op_7247"), val = tensor([1, 1])]; tensor var_7249 = const()[name = tensor("op_7249"), val = tensor([1, 1])]; tensor x_737_pad_type_0 = const()[name = tensor("x_737_pad_type_0"), val = tensor("custom")]; tensor x_737_pad_0 = const()[name = tensor("x_737_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_26_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480035648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480854912))), name = tensor("layers_26_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_26_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480855040)))]; tensor x_737_cast_fp16 = conv(bias = layers_26_self_attn_o_proj_module_bias_to_fp16, dilations = var_7249, groups = var_7139, pad = x_737_pad_0, pad_type = x_737_pad_type_0, strides = var_7247, weight = layers_26_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_579_cast_fp16)[name = tensor("x_737_cast_fp16")]; tensor layers_26_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_26_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480857664)))]; tensor obj_371_cast_fp16 = mul(x = x_737_cast_fp16, y = layers_26_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_371_cast_fp16")]; tensor inputs_159_cast_fp16 = add(x = inputs_157_cast_fp16, y = obj_371_cast_fp16)[name = tensor("inputs_159_cast_fp16")]; tensor var_7260 = const()[name = tensor("op_7260"), val = tensor([1])]; tensor channels_mean_159_cast_fp16 = reduce_mean(axes = var_7260, keep_dims = var_7140, x = inputs_159_cast_fp16)[name = tensor("channels_mean_159_cast_fp16")]; tensor zero_mean_159_cast_fp16 = sub(x = inputs_159_cast_fp16, y = channels_mean_159_cast_fp16)[name = tensor("zero_mean_159_cast_fp16")]; tensor zero_mean_sq_159_cast_fp16 = mul(x = zero_mean_159_cast_fp16, y = zero_mean_159_cast_fp16)[name = tensor("zero_mean_sq_159_cast_fp16")]; tensor var_7264 = const()[name = tensor("op_7264"), val = tensor([1])]; tensor var_7265_cast_fp16 = reduce_mean(axes = var_7264, keep_dims = var_7140, x = zero_mean_sq_159_cast_fp16)[name = tensor("op_7265_cast_fp16")]; tensor var_7266_to_fp16 = const()[name = tensor("op_7266_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_7267_cast_fp16 = add(x = var_7265_cast_fp16, y = var_7266_to_fp16)[name = tensor("op_7267_cast_fp16")]; tensor denom_159_epsilon_0_to_fp16 = const()[name = tensor("denom_159_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_159_cast_fp16 = rsqrt(epsilon = denom_159_epsilon_0_to_fp16, x = var_7267_cast_fp16)[name = tensor("denom_159_cast_fp16")]; tensor out_159_cast_fp16 = mul(x = zero_mean_159_cast_fp16, y = denom_159_cast_fp16)[name = tensor("out_159_cast_fp16")]; tensor obj_373_gamma_0_to_fp16 = const()[name = tensor("obj_373_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480860288)))]; tensor obj_373_beta_0_to_fp16 = const()[name = tensor("obj_373_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480862912)))]; tensor obj_373_epsilon_0_to_fp16 = const()[name = tensor("obj_373_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_373_cast_fp16 = batch_norm(beta = obj_373_beta_0_to_fp16, epsilon = obj_373_epsilon_0_to_fp16, gamma = obj_373_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_159_cast_fp16)[name = tensor("obj_373_cast_fp16")]; tensor layers_26_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_26_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480865536)))]; tensor input_581_cast_fp16 = sub(x = obj_373_cast_fp16, y = layers_26_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_581_cast_fp16")]; tensor var_7286 = const()[name = tensor("op_7286"), val = tensor([1, 1])]; tensor var_7288 = const()[name = tensor("op_7288"), val = tensor([1, 1])]; tensor x_739_pad_type_0 = const()[name = tensor("x_739_pad_type_0"), val = tensor("custom")]; tensor x_739_pad_0 = const()[name = tensor("x_739_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_26_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480868160))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(481687424))), name = tensor("layers_26_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_26_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_26_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(481687552)))]; tensor x_739_cast_fp16 = conv(bias = layers_26_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_7288, groups = var_7139, pad = x_739_pad_0, pad_type = x_739_pad_type_0, strides = var_7286, weight = layers_26_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = tensor("x_739_cast_fp16")]; tensor layers_26_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_26_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(481690176)))]; tensor query_107_cast_fp16 = mul(x = x_739_cast_fp16, y = layers_26_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_107_cast_fp16")]; tensor var_7298 = const()[name = tensor("op_7298"), val = tensor([1, 1])]; tensor var_7300 = const()[name = tensor("op_7300"), val = tensor([1, 1])]; tensor x_741_pad_type_0 = const()[name = tensor("x_741_pad_type_0"), val = tensor("custom")]; tensor x_741_pad_0 = const()[name = tensor("x_741_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_26_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(481692800))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(482512064))), name = tensor("layers_26_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_26_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_26_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(482512192)))]; tensor x_741_cast_fp16 = conv(bias = layers_26_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_7300, groups = var_7139, pad = x_741_pad_0, pad_type = x_741_pad_type_0, strides = var_7298, weight = layers_26_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_741_cast_fp16")]; tensor layers_26_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_26_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(482514816)))]; tensor key_107_cast_fp16 = mul(x = x_741_cast_fp16, y = layers_26_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_107_cast_fp16")]; tensor var_7310 = const()[name = tensor("op_7310"), val = tensor([1, 1])]; tensor var_7312 = const()[name = tensor("op_7312"), val = tensor([1, 1])]; tensor x_743_pad_type_0 = const()[name = tensor("x_743_pad_type_0"), val = tensor("custom")]; tensor x_743_pad_0 = const()[name = tensor("x_743_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_26_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(482517440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483336704))), name = tensor("layers_26_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_26_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_26_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483336832)))]; tensor x_743_cast_fp16 = conv(bias = layers_26_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_7312, groups = var_7139, pad = x_743_pad_0, pad_type = x_743_pad_type_0, strides = var_7310, weight = layers_26_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_743_cast_fp16")]; tensor layers_26_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_26_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483339456)))]; tensor value_107_cast_fp16 = mul(x = x_743_cast_fp16, y = layers_26_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_107_cast_fp16")]; tensor var_7317 = const()[name = tensor("op_7317"), val = tensor([1, 20, 64, -1])]; tensor var_7318_cast_fp16 = reshape(shape = var_7317, x = query_107_cast_fp16)[name = tensor("op_7318_cast_fp16")]; tensor var_7319_to_fp16 = const()[name = tensor("op_7319_to_fp16"), val = tensor(0x1p-3)]; tensor var_7320_cast_fp16 = mul(x = var_7318_cast_fp16, y = var_7319_to_fp16)[name = tensor("op_7320_cast_fp16")]; tensor var_7321 = const()[name = tensor("op_7321"), val = tensor([1, 20, 64, -1])]; tensor var_7322_cast_fp16 = reshape(shape = var_7321, x = key_107_cast_fp16)[name = tensor("op_7322_cast_fp16")]; tensor mh_w_161_transpose_x_0 = const()[name = tensor("mh_w_161_transpose_x_0"), val = tensor(true)]; tensor mh_w_161_transpose_y_0 = const()[name = tensor("mh_w_161_transpose_y_0"), val = tensor(false)]; tensor mh_w_161_cast_fp16 = matmul(transpose_x = mh_w_161_transpose_x_0, transpose_y = mh_w_161_transpose_y_0, x = var_7320_cast_fp16, y = var_7322_cast_fp16)[name = tensor("mh_w_161_cast_fp16")]; tensor obj_377_cast_fp16 = softmax(axis = var_7132, x = mh_w_161_cast_fp16)[name = tensor("obj_377_cast_fp16")]; tensor var_7326 = const()[name = tensor("op_7326"), val = tensor([1, 20, 64, -1])]; tensor var_7327_cast_fp16 = reshape(shape = var_7326, x = value_107_cast_fp16)[name = tensor("op_7327_cast_fp16")]; tensor attn_107_transpose_x_0 = const()[name = tensor("attn_107_transpose_x_0"), val = tensor(false)]; tensor attn_107_transpose_y_0 = const()[name = tensor("attn_107_transpose_y_0"), val = tensor(true)]; tensor attn_107_cast_fp16 = matmul(transpose_x = attn_107_transpose_x_0, transpose_y = attn_107_transpose_y_0, x = var_7327_cast_fp16, y = obj_377_cast_fp16)[name = tensor("attn_107_cast_fp16")]; tensor var_7330 = const()[name = tensor("op_7330"), val = tensor([1, 1280, 1, -1])]; tensor x_745_cast_fp16 = reshape(shape = var_7330, x = attn_107_cast_fp16)[name = tensor("x_745_cast_fp16")]; tensor layers_26_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_26_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483342080)))]; tensor input_587_cast_fp16 = sub(x = x_745_cast_fp16, y = layers_26_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_587_cast_fp16")]; tensor var_7338 = const()[name = tensor("op_7338"), val = tensor([1, 1])]; tensor var_7340 = const()[name = tensor("op_7340"), val = tensor([1, 1])]; tensor x_747_pad_type_0 = const()[name = tensor("x_747_pad_type_0"), val = tensor("custom")]; tensor x_747_pad_0 = const()[name = tensor("x_747_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_26_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483344704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484163968))), name = tensor("layers_26_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_26_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_26_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484164096)))]; tensor x_747_cast_fp16 = conv(bias = layers_26_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_7340, groups = var_7139, pad = x_747_pad_0, pad_type = x_747_pad_type_0, strides = var_7338, weight = layers_26_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_587_cast_fp16)[name = tensor("x_747_cast_fp16")]; tensor layers_26_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_26_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484166720)))]; tensor obj_375_cast_fp16 = mul(x = x_747_cast_fp16, y = layers_26_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_375_cast_fp16")]; tensor inputs_161_cast_fp16 = add(x = inputs_159_cast_fp16, y = obj_375_cast_fp16)[name = tensor("inputs_161_cast_fp16")]; tensor var_7347 = const()[name = tensor("op_7347"), val = tensor([1])]; tensor channels_mean_161_cast_fp16 = reduce_mean(axes = var_7347, keep_dims = var_7140, x = inputs_161_cast_fp16)[name = tensor("channels_mean_161_cast_fp16")]; tensor zero_mean_161_cast_fp16 = sub(x = inputs_161_cast_fp16, y = channels_mean_161_cast_fp16)[name = tensor("zero_mean_161_cast_fp16")]; tensor zero_mean_sq_161_cast_fp16 = mul(x = zero_mean_161_cast_fp16, y = zero_mean_161_cast_fp16)[name = tensor("zero_mean_sq_161_cast_fp16")]; tensor var_7351 = const()[name = tensor("op_7351"), val = tensor([1])]; tensor var_7352_cast_fp16 = reduce_mean(axes = var_7351, keep_dims = var_7140, x = zero_mean_sq_161_cast_fp16)[name = tensor("op_7352_cast_fp16")]; tensor var_7353_to_fp16 = const()[name = tensor("op_7353_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_7354_cast_fp16 = add(x = var_7352_cast_fp16, y = var_7353_to_fp16)[name = tensor("op_7354_cast_fp16")]; tensor denom_161_epsilon_0_to_fp16 = const()[name = tensor("denom_161_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_161_cast_fp16 = rsqrt(epsilon = denom_161_epsilon_0_to_fp16, x = var_7354_cast_fp16)[name = tensor("denom_161_cast_fp16")]; tensor out_161_cast_fp16 = mul(x = zero_mean_161_cast_fp16, y = denom_161_cast_fp16)[name = tensor("out_161_cast_fp16")]; tensor x_749_gamma_0_to_fp16 = const()[name = tensor("x_749_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484169344)))]; tensor x_749_beta_0_to_fp16 = const()[name = tensor("x_749_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484171968)))]; tensor x_749_epsilon_0_to_fp16 = const()[name = tensor("x_749_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_749_cast_fp16 = batch_norm(beta = x_749_beta_0_to_fp16, epsilon = x_749_epsilon_0_to_fp16, gamma = x_749_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_161_cast_fp16)[name = tensor("x_749_cast_fp16")]; tensor layers_26_fc1_input_shift_to_fp16 = const()[name = tensor("layers_26_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484174592)))]; tensor input_589_cast_fp16 = sub(x = x_749_cast_fp16, y = layers_26_fc1_input_shift_to_fp16)[name = tensor("input_589_cast_fp16")]; tensor var_7369 = const()[name = tensor("op_7369"), val = tensor([1, 1])]; tensor var_7371 = const()[name = tensor("op_7371"), val = tensor([1, 1])]; tensor x_751_pad_type_0 = const()[name = tensor("x_751_pad_type_0"), val = tensor("custom")]; tensor x_751_pad_0 = const()[name = tensor("x_751_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_26_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484177216))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487454080))), name = tensor("layers_26_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_26_fc1_module_bias_to_fp16 = const()[name = tensor("layers_26_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487454208)))]; tensor x_751_cast_fp16 = conv(bias = layers_26_fc1_module_bias_to_fp16, dilations = var_7371, groups = var_7139, pad = x_751_pad_0, pad_type = x_751_pad_type_0, strides = var_7369, weight = layers_26_fc1_module_weight_to_fp16_palettized, x = input_589_cast_fp16)[name = tensor("x_751_cast_fp16")]; tensor layers_26_fc1_output_scale_to_fp16 = const()[name = tensor("layers_26_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487464512)))]; tensor input_591_cast_fp16 = mul(x = x_751_cast_fp16, y = layers_26_fc1_output_scale_to_fp16)[name = tensor("input_591_cast_fp16")]; tensor x_753_mode_0 = const()[name = tensor("x_753_mode_0"), val = tensor("EXACT")]; tensor x_753_cast_fp16 = gelu(mode = x_753_mode_0, x = input_591_cast_fp16)[name = tensor("x_753_cast_fp16")]; tensor layers_26_fc2_input_shift_to_fp16 = const()[name = tensor("layers_26_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487474816)))]; tensor input_593_cast_fp16 = sub(x = x_753_cast_fp16, y = layers_26_fc2_input_shift_to_fp16)[name = tensor("input_593_cast_fp16")]; tensor var_7382 = const()[name = tensor("op_7382"), val = tensor([1, 1])]; tensor var_7384 = const()[name = tensor("op_7384"), val = tensor([1, 1])]; tensor x_755_pad_type_0 = const()[name = tensor("x_755_pad_type_0"), val = tensor("custom")]; tensor x_755_pad_0 = const()[name = tensor("x_755_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_26_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487485120))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490761984))), name = tensor("layers_26_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_26_fc2_module_bias_to_fp16 = const()[name = tensor("layers_26_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490762112)))]; tensor x_755_cast_fp16 = conv(bias = layers_26_fc2_module_bias_to_fp16, dilations = var_7384, groups = var_7139, pad = x_755_pad_0, pad_type = x_755_pad_type_0, strides = var_7382, weight = layers_26_fc2_module_weight_to_fp16_palettized, x = input_593_cast_fp16)[name = tensor("x_755_cast_fp16")]; tensor layers_26_fc2_output_scale_to_fp16 = const()[name = tensor("layers_26_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490764736)))]; tensor hidden_states_55_cast_fp16 = mul(x = x_755_cast_fp16, y = layers_26_fc2_output_scale_to_fp16)[name = tensor("hidden_states_55_cast_fp16")]; tensor inputs_163_cast_fp16 = add(x = inputs_161_cast_fp16, y = hidden_states_55_cast_fp16)[name = tensor("inputs_163_cast_fp16")]; tensor var_7398 = const()[name = tensor("op_7398"), val = tensor(3)]; tensor var_7405 = const()[name = tensor("op_7405"), val = tensor(1)]; tensor var_7406 = const()[name = tensor("op_7406"), val = tensor(true)]; tensor var_7418 = const()[name = tensor("op_7418"), val = tensor([1])]; tensor channels_mean_163_cast_fp16 = reduce_mean(axes = var_7418, keep_dims = var_7406, x = inputs_163_cast_fp16)[name = tensor("channels_mean_163_cast_fp16")]; tensor zero_mean_163_cast_fp16 = sub(x = inputs_163_cast_fp16, y = channels_mean_163_cast_fp16)[name = tensor("zero_mean_163_cast_fp16")]; tensor zero_mean_sq_163_cast_fp16 = mul(x = zero_mean_163_cast_fp16, y = zero_mean_163_cast_fp16)[name = tensor("zero_mean_sq_163_cast_fp16")]; tensor var_7422 = const()[name = tensor("op_7422"), val = tensor([1])]; tensor var_7423_cast_fp16 = reduce_mean(axes = var_7422, keep_dims = var_7406, x = zero_mean_sq_163_cast_fp16)[name = tensor("op_7423_cast_fp16")]; tensor var_7424_to_fp16 = const()[name = tensor("op_7424_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_7425_cast_fp16 = add(x = var_7423_cast_fp16, y = var_7424_to_fp16)[name = tensor("op_7425_cast_fp16")]; tensor denom_163_epsilon_0_to_fp16 = const()[name = tensor("denom_163_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_163_cast_fp16 = rsqrt(epsilon = denom_163_epsilon_0_to_fp16, x = var_7425_cast_fp16)[name = tensor("denom_163_cast_fp16")]; tensor out_163_cast_fp16 = mul(x = zero_mean_163_cast_fp16, y = denom_163_cast_fp16)[name = tensor("out_163_cast_fp16")]; tensor obj_379_gamma_0_to_fp16 = const()[name = tensor("obj_379_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490767360)))]; tensor obj_379_beta_0_to_fp16 = const()[name = tensor("obj_379_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490769984)))]; tensor obj_379_epsilon_0_to_fp16 = const()[name = tensor("obj_379_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_379_cast_fp16 = batch_norm(beta = obj_379_beta_0_to_fp16, epsilon = obj_379_epsilon_0_to_fp16, gamma = obj_379_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_163_cast_fp16)[name = tensor("obj_379_cast_fp16")]; tensor layers_27_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_27_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490772608)))]; tensor input_595_cast_fp16 = sub(x = obj_379_cast_fp16, y = layers_27_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_595_cast_fp16")]; tensor var_7444 = const()[name = tensor("op_7444"), val = tensor([1, 1])]; tensor var_7446 = const()[name = tensor("op_7446"), val = tensor([1, 1])]; tensor x_757_pad_type_0 = const()[name = tensor("x_757_pad_type_0"), val = tensor("custom")]; tensor x_757_pad_0 = const()[name = tensor("x_757_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_27_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490775232))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(491594496))), name = tensor("layers_27_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_27_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(491594624)))]; tensor x_757_cast_fp16 = conv(bias = layers_27_self_attn_q_proj_module_bias_to_fp16, dilations = var_7446, groups = var_7405, pad = x_757_pad_0, pad_type = x_757_pad_type_0, strides = var_7444, weight = layers_27_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_595_cast_fp16)[name = tensor("x_757_cast_fp16")]; tensor layers_27_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_27_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(491597248)))]; tensor query_109_cast_fp16 = mul(x = x_757_cast_fp16, y = layers_27_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_109_cast_fp16")]; tensor var_7456 = const()[name = tensor("op_7456"), val = tensor([1, 1])]; tensor var_7458 = const()[name = tensor("op_7458"), val = tensor([1, 1])]; tensor x_759_pad_type_0 = const()[name = tensor("x_759_pad_type_0"), val = tensor("custom")]; tensor x_759_pad_0 = const()[name = tensor("x_759_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_27_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(491599872))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492419136))), name = tensor("layers_27_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_27_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492419264)))]; tensor x_759_cast_fp16 = conv(bias = layers_27_self_attn_k_proj_module_bias_to_fp16, dilations = var_7458, groups = var_7405, pad = x_759_pad_0, pad_type = x_759_pad_type_0, strides = var_7456, weight = layers_27_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_595_cast_fp16)[name = tensor("x_759_cast_fp16")]; tensor layers_27_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_27_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492421888)))]; tensor current_key_55_cast_fp16 = mul(x = x_759_cast_fp16, y = layers_27_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_55_cast_fp16")]; tensor var_7468 = const()[name = tensor("op_7468"), val = tensor([1, 1])]; tensor var_7470 = const()[name = tensor("op_7470"), val = tensor([1, 1])]; tensor x_761_pad_type_0 = const()[name = tensor("x_761_pad_type_0"), val = tensor("custom")]; tensor x_761_pad_0 = const()[name = tensor("x_761_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_27_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492424512))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493243776))), name = tensor("layers_27_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_27_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493243904)))]; tensor x_761_cast_fp16 = conv(bias = layers_27_self_attn_v_proj_module_bias_to_fp16, dilations = var_7470, groups = var_7405, pad = x_761_pad_0, pad_type = x_761_pad_type_0, strides = var_7468, weight = layers_27_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_595_cast_fp16)[name = tensor("x_761_cast_fp16")]; tensor layers_27_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_27_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493246528)))]; tensor current_value_55_cast_fp16 = mul(x = x_761_cast_fp16, y = layers_27_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_55_cast_fp16")]; tensor var_7478_cast_fp16 = mul(x = current_key_55_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_7478_cast_fp16")]; tensor var_7480_cast_fp16 = mul(x = var_103_cast_fp16_27, y = var_257_cast_fp16)[name = tensor("op_7480_cast_fp16")]; tensor key_109_cast_fp16 = add(x = var_7478_cast_fp16, y = var_7480_cast_fp16)[name = tensor("key_109_cast_fp16")]; tensor var_7482_cast_fp16 = mul(x = current_value_55_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_7482_cast_fp16")]; tensor var_7484_cast_fp16 = mul(x = var_138_cast_fp16_27, y = var_257_cast_fp16)[name = tensor("op_7484_cast_fp16")]; tensor value_109_cast_fp16 = add(x = var_7482_cast_fp16, y = var_7484_cast_fp16)[name = tensor("value_109_cast_fp16")]; tensor var_7487 = const()[name = tensor("op_7487"), val = tensor([1, 20, 64, -1])]; tensor var_7488_cast_fp16 = reshape(shape = var_7487, x = query_109_cast_fp16)[name = tensor("op_7488_cast_fp16")]; tensor var_7489_to_fp16 = const()[name = tensor("op_7489_to_fp16"), val = tensor(0x1p-3)]; tensor var_7490_cast_fp16 = mul(x = var_7488_cast_fp16, y = var_7489_to_fp16)[name = tensor("op_7490_cast_fp16")]; tensor var_7491 = const()[name = tensor("op_7491"), val = tensor([1, 20, 64, -1])]; tensor var_7492_cast_fp16 = reshape(shape = var_7491, x = key_109_cast_fp16)[name = tensor("op_7492_cast_fp16")]; tensor mh_w_163_transpose_x_0 = const()[name = tensor("mh_w_163_transpose_x_0"), val = tensor(true)]; tensor mh_w_163_transpose_y_0 = const()[name = tensor("mh_w_163_transpose_y_0"), val = tensor(false)]; tensor mh_w_163_cast_fp16 = matmul(transpose_x = mh_w_163_transpose_x_0, transpose_y = mh_w_163_transpose_y_0, x = var_7490_cast_fp16, y = var_7492_cast_fp16)[name = tensor("mh_w_163_cast_fp16")]; tensor mh_w_165_cast_fp16 = add(x = mh_w_163_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_165_cast_fp16")]; tensor var_7500_cast_fp16 = softmax(axis = var_7398, x = mh_w_165_cast_fp16)[name = tensor("op_7500_cast_fp16")]; tensor var_7501 = const()[name = tensor("op_7501"), val = tensor([1, 20, 64, -1])]; tensor var_7502_cast_fp16 = reshape(shape = var_7501, x = value_109_cast_fp16)[name = tensor("op_7502_cast_fp16")]; tensor attn_109_transpose_x_0 = const()[name = tensor("attn_109_transpose_x_0"), val = tensor(false)]; tensor attn_109_transpose_y_0 = const()[name = tensor("attn_109_transpose_y_0"), val = tensor(true)]; tensor attn_109_cast_fp16 = matmul(transpose_x = attn_109_transpose_x_0, transpose_y = attn_109_transpose_y_0, x = var_7502_cast_fp16, y = var_7500_cast_fp16)[name = tensor("attn_109_cast_fp16")]; tensor var_7505 = const()[name = tensor("op_7505"), val = tensor([1, 1280, 1, -1])]; tensor x_763_cast_fp16 = reshape(shape = var_7505, x = attn_109_cast_fp16)[name = tensor("x_763_cast_fp16")]; tensor layers_27_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_27_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493249152)))]; tensor input_601_cast_fp16 = sub(x = x_763_cast_fp16, y = layers_27_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_601_cast_fp16")]; tensor var_7513 = const()[name = tensor("op_7513"), val = tensor([1, 1])]; tensor var_7515 = const()[name = tensor("op_7515"), val = tensor([1, 1])]; tensor x_765_pad_type_0 = const()[name = tensor("x_765_pad_type_0"), val = tensor("custom")]; tensor x_765_pad_0 = const()[name = tensor("x_765_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_27_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493251776))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494071040))), name = tensor("layers_27_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_27_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494071168)))]; tensor x_765_cast_fp16 = conv(bias = layers_27_self_attn_o_proj_module_bias_to_fp16, dilations = var_7515, groups = var_7405, pad = x_765_pad_0, pad_type = x_765_pad_type_0, strides = var_7513, weight = layers_27_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_601_cast_fp16)[name = tensor("x_765_cast_fp16")]; tensor layers_27_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_27_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494073792)))]; tensor obj_385_cast_fp16 = mul(x = x_765_cast_fp16, y = layers_27_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_385_cast_fp16")]; tensor inputs_165_cast_fp16 = add(x = inputs_163_cast_fp16, y = obj_385_cast_fp16)[name = tensor("inputs_165_cast_fp16")]; tensor var_7526 = const()[name = tensor("op_7526"), val = tensor([1])]; tensor channels_mean_165_cast_fp16 = reduce_mean(axes = var_7526, keep_dims = var_7406, x = inputs_165_cast_fp16)[name = tensor("channels_mean_165_cast_fp16")]; tensor zero_mean_165_cast_fp16 = sub(x = inputs_165_cast_fp16, y = channels_mean_165_cast_fp16)[name = tensor("zero_mean_165_cast_fp16")]; tensor zero_mean_sq_165_cast_fp16 = mul(x = zero_mean_165_cast_fp16, y = zero_mean_165_cast_fp16)[name = tensor("zero_mean_sq_165_cast_fp16")]; tensor var_7530 = const()[name = tensor("op_7530"), val = tensor([1])]; tensor var_7531_cast_fp16 = reduce_mean(axes = var_7530, keep_dims = var_7406, x = zero_mean_sq_165_cast_fp16)[name = tensor("op_7531_cast_fp16")]; tensor var_7532_to_fp16 = const()[name = tensor("op_7532_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_7533_cast_fp16 = add(x = var_7531_cast_fp16, y = var_7532_to_fp16)[name = tensor("op_7533_cast_fp16")]; tensor denom_165_epsilon_0_to_fp16 = const()[name = tensor("denom_165_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_165_cast_fp16 = rsqrt(epsilon = denom_165_epsilon_0_to_fp16, x = var_7533_cast_fp16)[name = tensor("denom_165_cast_fp16")]; tensor out_165_cast_fp16 = mul(x = zero_mean_165_cast_fp16, y = denom_165_cast_fp16)[name = tensor("out_165_cast_fp16")]; tensor obj_387_gamma_0_to_fp16 = const()[name = tensor("obj_387_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494076416)))]; tensor obj_387_beta_0_to_fp16 = const()[name = tensor("obj_387_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494079040)))]; tensor obj_387_epsilon_0_to_fp16 = const()[name = tensor("obj_387_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_387_cast_fp16 = batch_norm(beta = obj_387_beta_0_to_fp16, epsilon = obj_387_epsilon_0_to_fp16, gamma = obj_387_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_165_cast_fp16)[name = tensor("obj_387_cast_fp16")]; tensor layers_27_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_27_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494081664)))]; tensor input_603_cast_fp16 = sub(x = obj_387_cast_fp16, y = layers_27_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_603_cast_fp16")]; tensor var_7552 = const()[name = tensor("op_7552"), val = tensor([1, 1])]; tensor var_7554 = const()[name = tensor("op_7554"), val = tensor([1, 1])]; tensor x_767_pad_type_0 = const()[name = tensor("x_767_pad_type_0"), val = tensor("custom")]; tensor x_767_pad_0 = const()[name = tensor("x_767_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_27_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494084288))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494903552))), name = tensor("layers_27_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_27_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_27_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494903680)))]; tensor x_767_cast_fp16 = conv(bias = layers_27_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_7554, groups = var_7405, pad = x_767_pad_0, pad_type = x_767_pad_type_0, strides = var_7552, weight = layers_27_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_603_cast_fp16)[name = tensor("x_767_cast_fp16")]; tensor layers_27_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_27_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494906304)))]; tensor query_111_cast_fp16 = mul(x = x_767_cast_fp16, y = layers_27_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_111_cast_fp16")]; tensor var_7564 = const()[name = tensor("op_7564"), val = tensor([1, 1])]; tensor var_7566 = const()[name = tensor("op_7566"), val = tensor([1, 1])]; tensor x_769_pad_type_0 = const()[name = tensor("x_769_pad_type_0"), val = tensor("custom")]; tensor x_769_pad_0 = const()[name = tensor("x_769_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_27_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494908928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495728192))), name = tensor("layers_27_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_27_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_27_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495728320)))]; tensor x_769_cast_fp16 = conv(bias = layers_27_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_7566, groups = var_7405, pad = x_769_pad_0, pad_type = x_769_pad_type_0, strides = var_7564, weight = layers_27_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_769_cast_fp16")]; tensor layers_27_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_27_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495730944)))]; tensor key_111_cast_fp16 = mul(x = x_769_cast_fp16, y = layers_27_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_111_cast_fp16")]; tensor var_7576 = const()[name = tensor("op_7576"), val = tensor([1, 1])]; tensor var_7578 = const()[name = tensor("op_7578"), val = tensor([1, 1])]; tensor x_771_pad_type_0 = const()[name = tensor("x_771_pad_type_0"), val = tensor("custom")]; tensor x_771_pad_0 = const()[name = tensor("x_771_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_27_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495733568))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496552832))), name = tensor("layers_27_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_27_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_27_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496552960)))]; tensor x_771_cast_fp16 = conv(bias = layers_27_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_7578, groups = var_7405, pad = x_771_pad_0, pad_type = x_771_pad_type_0, strides = var_7576, weight = layers_27_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_771_cast_fp16")]; tensor layers_27_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_27_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496555584)))]; tensor value_111_cast_fp16 = mul(x = x_771_cast_fp16, y = layers_27_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_111_cast_fp16")]; tensor var_7583 = const()[name = tensor("op_7583"), val = tensor([1, 20, 64, -1])]; tensor var_7584_cast_fp16 = reshape(shape = var_7583, x = query_111_cast_fp16)[name = tensor("op_7584_cast_fp16")]; tensor var_7585_to_fp16 = const()[name = tensor("op_7585_to_fp16"), val = tensor(0x1p-3)]; tensor var_7586_cast_fp16 = mul(x = var_7584_cast_fp16, y = var_7585_to_fp16)[name = tensor("op_7586_cast_fp16")]; tensor var_7587 = const()[name = tensor("op_7587"), val = tensor([1, 20, 64, -1])]; tensor var_7588_cast_fp16 = reshape(shape = var_7587, x = key_111_cast_fp16)[name = tensor("op_7588_cast_fp16")]; tensor mh_w_167_transpose_x_0 = const()[name = tensor("mh_w_167_transpose_x_0"), val = tensor(true)]; tensor mh_w_167_transpose_y_0 = const()[name = tensor("mh_w_167_transpose_y_0"), val = tensor(false)]; tensor mh_w_167_cast_fp16 = matmul(transpose_x = mh_w_167_transpose_x_0, transpose_y = mh_w_167_transpose_y_0, x = var_7586_cast_fp16, y = var_7588_cast_fp16)[name = tensor("mh_w_167_cast_fp16")]; tensor obj_391_cast_fp16 = softmax(axis = var_7398, x = mh_w_167_cast_fp16)[name = tensor("obj_391_cast_fp16")]; tensor var_7592 = const()[name = tensor("op_7592"), val = tensor([1, 20, 64, -1])]; tensor var_7593_cast_fp16 = reshape(shape = var_7592, x = value_111_cast_fp16)[name = tensor("op_7593_cast_fp16")]; tensor attn_111_transpose_x_0 = const()[name = tensor("attn_111_transpose_x_0"), val = tensor(false)]; tensor attn_111_transpose_y_0 = const()[name = tensor("attn_111_transpose_y_0"), val = tensor(true)]; tensor attn_111_cast_fp16 = matmul(transpose_x = attn_111_transpose_x_0, transpose_y = attn_111_transpose_y_0, x = var_7593_cast_fp16, y = obj_391_cast_fp16)[name = tensor("attn_111_cast_fp16")]; tensor var_7596 = const()[name = tensor("op_7596"), val = tensor([1, 1280, 1, -1])]; tensor x_773_cast_fp16 = reshape(shape = var_7596, x = attn_111_cast_fp16)[name = tensor("x_773_cast_fp16")]; tensor layers_27_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_27_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496558208)))]; tensor input_609_cast_fp16 = sub(x = x_773_cast_fp16, y = layers_27_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_609_cast_fp16")]; tensor var_7604 = const()[name = tensor("op_7604"), val = tensor([1, 1])]; tensor var_7606 = const()[name = tensor("op_7606"), val = tensor([1, 1])]; tensor x_775_pad_type_0 = const()[name = tensor("x_775_pad_type_0"), val = tensor("custom")]; tensor x_775_pad_0 = const()[name = tensor("x_775_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_27_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496560832))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(497380096))), name = tensor("layers_27_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_27_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_27_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(497380224)))]; tensor x_775_cast_fp16 = conv(bias = layers_27_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_7606, groups = var_7405, pad = x_775_pad_0, pad_type = x_775_pad_type_0, strides = var_7604, weight = layers_27_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_609_cast_fp16)[name = tensor("x_775_cast_fp16")]; tensor layers_27_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_27_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(497382848)))]; tensor obj_389_cast_fp16 = mul(x = x_775_cast_fp16, y = layers_27_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_389_cast_fp16")]; tensor inputs_167_cast_fp16 = add(x = inputs_165_cast_fp16, y = obj_389_cast_fp16)[name = tensor("inputs_167_cast_fp16")]; tensor var_7613 = const()[name = tensor("op_7613"), val = tensor([1])]; tensor channels_mean_167_cast_fp16 = reduce_mean(axes = var_7613, keep_dims = var_7406, x = inputs_167_cast_fp16)[name = tensor("channels_mean_167_cast_fp16")]; tensor zero_mean_167_cast_fp16 = sub(x = inputs_167_cast_fp16, y = channels_mean_167_cast_fp16)[name = tensor("zero_mean_167_cast_fp16")]; tensor zero_mean_sq_167_cast_fp16 = mul(x = zero_mean_167_cast_fp16, y = zero_mean_167_cast_fp16)[name = tensor("zero_mean_sq_167_cast_fp16")]; tensor var_7617 = const()[name = tensor("op_7617"), val = tensor([1])]; tensor var_7618_cast_fp16 = reduce_mean(axes = var_7617, keep_dims = var_7406, x = zero_mean_sq_167_cast_fp16)[name = tensor("op_7618_cast_fp16")]; tensor var_7619_to_fp16 = const()[name = tensor("op_7619_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_7620_cast_fp16 = add(x = var_7618_cast_fp16, y = var_7619_to_fp16)[name = tensor("op_7620_cast_fp16")]; tensor denom_167_epsilon_0_to_fp16 = const()[name = tensor("denom_167_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_167_cast_fp16 = rsqrt(epsilon = denom_167_epsilon_0_to_fp16, x = var_7620_cast_fp16)[name = tensor("denom_167_cast_fp16")]; tensor out_167_cast_fp16 = mul(x = zero_mean_167_cast_fp16, y = denom_167_cast_fp16)[name = tensor("out_167_cast_fp16")]; tensor x_777_gamma_0_to_fp16 = const()[name = tensor("x_777_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(497385472)))]; tensor x_777_beta_0_to_fp16 = const()[name = tensor("x_777_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(497388096)))]; tensor x_777_epsilon_0_to_fp16 = const()[name = tensor("x_777_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_777_cast_fp16 = batch_norm(beta = x_777_beta_0_to_fp16, epsilon = x_777_epsilon_0_to_fp16, gamma = x_777_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_167_cast_fp16)[name = tensor("x_777_cast_fp16")]; tensor layers_27_fc1_input_shift_to_fp16 = const()[name = tensor("layers_27_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(497390720)))]; tensor input_611_cast_fp16 = sub(x = x_777_cast_fp16, y = layers_27_fc1_input_shift_to_fp16)[name = tensor("input_611_cast_fp16")]; tensor var_7635 = const()[name = tensor("op_7635"), val = tensor([1, 1])]; tensor var_7637 = const()[name = tensor("op_7637"), val = tensor([1, 1])]; tensor x_779_pad_type_0 = const()[name = tensor("x_779_pad_type_0"), val = tensor("custom")]; tensor x_779_pad_0 = const()[name = tensor("x_779_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_27_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(497393344))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500670208))), name = tensor("layers_27_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_27_fc1_module_bias_to_fp16 = const()[name = tensor("layers_27_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500670336)))]; tensor x_779_cast_fp16 = conv(bias = layers_27_fc1_module_bias_to_fp16, dilations = var_7637, groups = var_7405, pad = x_779_pad_0, pad_type = x_779_pad_type_0, strides = var_7635, weight = layers_27_fc1_module_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = tensor("x_779_cast_fp16")]; tensor layers_27_fc1_output_scale_to_fp16 = const()[name = tensor("layers_27_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500680640)))]; tensor input_613_cast_fp16 = mul(x = x_779_cast_fp16, y = layers_27_fc1_output_scale_to_fp16)[name = tensor("input_613_cast_fp16")]; tensor x_781_mode_0 = const()[name = tensor("x_781_mode_0"), val = tensor("EXACT")]; tensor x_781_cast_fp16 = gelu(mode = x_781_mode_0, x = input_613_cast_fp16)[name = tensor("x_781_cast_fp16")]; tensor layers_27_fc2_input_shift_to_fp16 = const()[name = tensor("layers_27_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500690944)))]; tensor input_615_cast_fp16 = sub(x = x_781_cast_fp16, y = layers_27_fc2_input_shift_to_fp16)[name = tensor("input_615_cast_fp16")]; tensor var_7648 = const()[name = tensor("op_7648"), val = tensor([1, 1])]; tensor var_7650 = const()[name = tensor("op_7650"), val = tensor([1, 1])]; tensor x_783_pad_type_0 = const()[name = tensor("x_783_pad_type_0"), val = tensor("custom")]; tensor x_783_pad_0 = const()[name = tensor("x_783_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_27_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500701248))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503978112))), name = tensor("layers_27_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_27_fc2_module_bias_to_fp16 = const()[name = tensor("layers_27_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503978240)))]; tensor x_783_cast_fp16 = conv(bias = layers_27_fc2_module_bias_to_fp16, dilations = var_7650, groups = var_7405, pad = x_783_pad_0, pad_type = x_783_pad_type_0, strides = var_7648, weight = layers_27_fc2_module_weight_to_fp16_palettized, x = input_615_cast_fp16)[name = tensor("x_783_cast_fp16")]; tensor layers_27_fc2_output_scale_to_fp16 = const()[name = tensor("layers_27_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503980864)))]; tensor hidden_states_57_cast_fp16 = mul(x = x_783_cast_fp16, y = layers_27_fc2_output_scale_to_fp16)[name = tensor("hidden_states_57_cast_fp16")]; tensor inputs_169_cast_fp16 = add(x = inputs_167_cast_fp16, y = hidden_states_57_cast_fp16)[name = tensor("inputs_169_cast_fp16")]; tensor var_7664 = const()[name = tensor("op_7664"), val = tensor(3)]; tensor var_7671 = const()[name = tensor("op_7671"), val = tensor(1)]; tensor var_7672 = const()[name = tensor("op_7672"), val = tensor(true)]; tensor var_7684 = const()[name = tensor("op_7684"), val = tensor([1])]; tensor channels_mean_169_cast_fp16 = reduce_mean(axes = var_7684, keep_dims = var_7672, x = inputs_169_cast_fp16)[name = tensor("channels_mean_169_cast_fp16")]; tensor zero_mean_169_cast_fp16 = sub(x = inputs_169_cast_fp16, y = channels_mean_169_cast_fp16)[name = tensor("zero_mean_169_cast_fp16")]; tensor zero_mean_sq_169_cast_fp16 = mul(x = zero_mean_169_cast_fp16, y = zero_mean_169_cast_fp16)[name = tensor("zero_mean_sq_169_cast_fp16")]; tensor var_7688 = const()[name = tensor("op_7688"), val = tensor([1])]; tensor var_7689_cast_fp16 = reduce_mean(axes = var_7688, keep_dims = var_7672, x = zero_mean_sq_169_cast_fp16)[name = tensor("op_7689_cast_fp16")]; tensor var_7690_to_fp16 = const()[name = tensor("op_7690_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_7691_cast_fp16 = add(x = var_7689_cast_fp16, y = var_7690_to_fp16)[name = tensor("op_7691_cast_fp16")]; tensor denom_169_epsilon_0_to_fp16 = const()[name = tensor("denom_169_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_169_cast_fp16 = rsqrt(epsilon = denom_169_epsilon_0_to_fp16, x = var_7691_cast_fp16)[name = tensor("denom_169_cast_fp16")]; tensor out_169_cast_fp16 = mul(x = zero_mean_169_cast_fp16, y = denom_169_cast_fp16)[name = tensor("out_169_cast_fp16")]; tensor obj_393_gamma_0_to_fp16 = const()[name = tensor("obj_393_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503983488)))]; tensor obj_393_beta_0_to_fp16 = const()[name = tensor("obj_393_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503986112)))]; tensor obj_393_epsilon_0_to_fp16 = const()[name = tensor("obj_393_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_393_cast_fp16 = batch_norm(beta = obj_393_beta_0_to_fp16, epsilon = obj_393_epsilon_0_to_fp16, gamma = obj_393_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_169_cast_fp16)[name = tensor("obj_393_cast_fp16")]; tensor layers_28_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_28_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503988736)))]; tensor input_617_cast_fp16 = sub(x = obj_393_cast_fp16, y = layers_28_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_617_cast_fp16")]; tensor var_7710 = const()[name = tensor("op_7710"), val = tensor([1, 1])]; tensor var_7712 = const()[name = tensor("op_7712"), val = tensor([1, 1])]; tensor x_785_pad_type_0 = const()[name = tensor("x_785_pad_type_0"), val = tensor("custom")]; tensor x_785_pad_0 = const()[name = tensor("x_785_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_28_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503991360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504810624))), name = tensor("layers_28_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_28_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504810752)))]; tensor x_785_cast_fp16 = conv(bias = layers_28_self_attn_q_proj_module_bias_to_fp16, dilations = var_7712, groups = var_7671, pad = x_785_pad_0, pad_type = x_785_pad_type_0, strides = var_7710, weight = layers_28_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_617_cast_fp16)[name = tensor("x_785_cast_fp16")]; tensor layers_28_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_28_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504813376)))]; tensor query_113_cast_fp16 = mul(x = x_785_cast_fp16, y = layers_28_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_113_cast_fp16")]; tensor var_7722 = const()[name = tensor("op_7722"), val = tensor([1, 1])]; tensor var_7724 = const()[name = tensor("op_7724"), val = tensor([1, 1])]; tensor x_787_pad_type_0 = const()[name = tensor("x_787_pad_type_0"), val = tensor("custom")]; tensor x_787_pad_0 = const()[name = tensor("x_787_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_28_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504816000))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(505635264))), name = tensor("layers_28_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_28_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(505635392)))]; tensor x_787_cast_fp16 = conv(bias = layers_28_self_attn_k_proj_module_bias_to_fp16, dilations = var_7724, groups = var_7671, pad = x_787_pad_0, pad_type = x_787_pad_type_0, strides = var_7722, weight = layers_28_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_617_cast_fp16)[name = tensor("x_787_cast_fp16")]; tensor layers_28_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_28_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(505638016)))]; tensor current_key_57_cast_fp16 = mul(x = x_787_cast_fp16, y = layers_28_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_57_cast_fp16")]; tensor var_7734 = const()[name = tensor("op_7734"), val = tensor([1, 1])]; tensor var_7736 = const()[name = tensor("op_7736"), val = tensor([1, 1])]; tensor x_789_pad_type_0 = const()[name = tensor("x_789_pad_type_0"), val = tensor("custom")]; tensor x_789_pad_0 = const()[name = tensor("x_789_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_28_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(505640640))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506459904))), name = tensor("layers_28_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_28_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506460032)))]; tensor x_789_cast_fp16 = conv(bias = layers_28_self_attn_v_proj_module_bias_to_fp16, dilations = var_7736, groups = var_7671, pad = x_789_pad_0, pad_type = x_789_pad_type_0, strides = var_7734, weight = layers_28_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_617_cast_fp16)[name = tensor("x_789_cast_fp16")]; tensor layers_28_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_28_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506462656)))]; tensor current_value_57_cast_fp16 = mul(x = x_789_cast_fp16, y = layers_28_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_57_cast_fp16")]; tensor var_7744_cast_fp16 = mul(x = current_key_57_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_7744_cast_fp16")]; tensor var_7746_cast_fp16 = mul(x = var_103_cast_fp16_28, y = var_257_cast_fp16)[name = tensor("op_7746_cast_fp16")]; tensor key_113_cast_fp16 = add(x = var_7744_cast_fp16, y = var_7746_cast_fp16)[name = tensor("key_113_cast_fp16")]; tensor var_7748_cast_fp16 = mul(x = current_value_57_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_7748_cast_fp16")]; tensor var_7750_cast_fp16 = mul(x = var_138_cast_fp16_28, y = var_257_cast_fp16)[name = tensor("op_7750_cast_fp16")]; tensor value_113_cast_fp16 = add(x = var_7748_cast_fp16, y = var_7750_cast_fp16)[name = tensor("value_113_cast_fp16")]; tensor var_7753 = const()[name = tensor("op_7753"), val = tensor([1, 20, 64, -1])]; tensor var_7754_cast_fp16 = reshape(shape = var_7753, x = query_113_cast_fp16)[name = tensor("op_7754_cast_fp16")]; tensor var_7755_to_fp16 = const()[name = tensor("op_7755_to_fp16"), val = tensor(0x1p-3)]; tensor var_7756_cast_fp16 = mul(x = var_7754_cast_fp16, y = var_7755_to_fp16)[name = tensor("op_7756_cast_fp16")]; tensor var_7757 = const()[name = tensor("op_7757"), val = tensor([1, 20, 64, -1])]; tensor var_7758_cast_fp16 = reshape(shape = var_7757, x = key_113_cast_fp16)[name = tensor("op_7758_cast_fp16")]; tensor mh_w_169_transpose_x_0 = const()[name = tensor("mh_w_169_transpose_x_0"), val = tensor(true)]; tensor mh_w_169_transpose_y_0 = const()[name = tensor("mh_w_169_transpose_y_0"), val = tensor(false)]; tensor mh_w_169_cast_fp16 = matmul(transpose_x = mh_w_169_transpose_x_0, transpose_y = mh_w_169_transpose_y_0, x = var_7756_cast_fp16, y = var_7758_cast_fp16)[name = tensor("mh_w_169_cast_fp16")]; tensor mh_w_171_cast_fp16 = add(x = mh_w_169_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_171_cast_fp16")]; tensor var_7766_cast_fp16 = softmax(axis = var_7664, x = mh_w_171_cast_fp16)[name = tensor("op_7766_cast_fp16")]; tensor var_7767 = const()[name = tensor("op_7767"), val = tensor([1, 20, 64, -1])]; tensor var_7768_cast_fp16 = reshape(shape = var_7767, x = value_113_cast_fp16)[name = tensor("op_7768_cast_fp16")]; tensor attn_113_transpose_x_0 = const()[name = tensor("attn_113_transpose_x_0"), val = tensor(false)]; tensor attn_113_transpose_y_0 = const()[name = tensor("attn_113_transpose_y_0"), val = tensor(true)]; tensor attn_113_cast_fp16 = matmul(transpose_x = attn_113_transpose_x_0, transpose_y = attn_113_transpose_y_0, x = var_7768_cast_fp16, y = var_7766_cast_fp16)[name = tensor("attn_113_cast_fp16")]; tensor var_7771 = const()[name = tensor("op_7771"), val = tensor([1, 1280, 1, -1])]; tensor x_791_cast_fp16 = reshape(shape = var_7771, x = attn_113_cast_fp16)[name = tensor("x_791_cast_fp16")]; tensor layers_28_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_28_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506465280)))]; tensor input_623_cast_fp16 = sub(x = x_791_cast_fp16, y = layers_28_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_623_cast_fp16")]; tensor var_7779 = const()[name = tensor("op_7779"), val = tensor([1, 1])]; tensor var_7781 = const()[name = tensor("op_7781"), val = tensor([1, 1])]; tensor x_793_pad_type_0 = const()[name = tensor("x_793_pad_type_0"), val = tensor("custom")]; tensor x_793_pad_0 = const()[name = tensor("x_793_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_28_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506467904))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507287168))), name = tensor("layers_28_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_28_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507287296)))]; tensor x_793_cast_fp16 = conv(bias = layers_28_self_attn_o_proj_module_bias_to_fp16, dilations = var_7781, groups = var_7671, pad = x_793_pad_0, pad_type = x_793_pad_type_0, strides = var_7779, weight = layers_28_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_623_cast_fp16)[name = tensor("x_793_cast_fp16")]; tensor layers_28_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_28_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507289920)))]; tensor obj_399_cast_fp16 = mul(x = x_793_cast_fp16, y = layers_28_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_399_cast_fp16")]; tensor inputs_171_cast_fp16 = add(x = inputs_169_cast_fp16, y = obj_399_cast_fp16)[name = tensor("inputs_171_cast_fp16")]; tensor var_7792 = const()[name = tensor("op_7792"), val = tensor([1])]; tensor channels_mean_171_cast_fp16 = reduce_mean(axes = var_7792, keep_dims = var_7672, x = inputs_171_cast_fp16)[name = tensor("channels_mean_171_cast_fp16")]; tensor zero_mean_171_cast_fp16 = sub(x = inputs_171_cast_fp16, y = channels_mean_171_cast_fp16)[name = tensor("zero_mean_171_cast_fp16")]; tensor zero_mean_sq_171_cast_fp16 = mul(x = zero_mean_171_cast_fp16, y = zero_mean_171_cast_fp16)[name = tensor("zero_mean_sq_171_cast_fp16")]; tensor var_7796 = const()[name = tensor("op_7796"), val = tensor([1])]; tensor var_7797_cast_fp16 = reduce_mean(axes = var_7796, keep_dims = var_7672, x = zero_mean_sq_171_cast_fp16)[name = tensor("op_7797_cast_fp16")]; tensor var_7798_to_fp16 = const()[name = tensor("op_7798_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_7799_cast_fp16 = add(x = var_7797_cast_fp16, y = var_7798_to_fp16)[name = tensor("op_7799_cast_fp16")]; tensor denom_171_epsilon_0_to_fp16 = const()[name = tensor("denom_171_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_171_cast_fp16 = rsqrt(epsilon = denom_171_epsilon_0_to_fp16, x = var_7799_cast_fp16)[name = tensor("denom_171_cast_fp16")]; tensor out_171_cast_fp16 = mul(x = zero_mean_171_cast_fp16, y = denom_171_cast_fp16)[name = tensor("out_171_cast_fp16")]; tensor obj_401_gamma_0_to_fp16 = const()[name = tensor("obj_401_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507292544)))]; tensor obj_401_beta_0_to_fp16 = const()[name = tensor("obj_401_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507295168)))]; tensor obj_401_epsilon_0_to_fp16 = const()[name = tensor("obj_401_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_401_cast_fp16 = batch_norm(beta = obj_401_beta_0_to_fp16, epsilon = obj_401_epsilon_0_to_fp16, gamma = obj_401_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_171_cast_fp16)[name = tensor("obj_401_cast_fp16")]; tensor layers_28_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_28_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507297792)))]; tensor input_625_cast_fp16 = sub(x = obj_401_cast_fp16, y = layers_28_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_625_cast_fp16")]; tensor var_7818 = const()[name = tensor("op_7818"), val = tensor([1, 1])]; tensor var_7820 = const()[name = tensor("op_7820"), val = tensor([1, 1])]; tensor x_795_pad_type_0 = const()[name = tensor("x_795_pad_type_0"), val = tensor("custom")]; tensor x_795_pad_0 = const()[name = tensor("x_795_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_28_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507300416))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508119680))), name = tensor("layers_28_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_28_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_28_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508119808)))]; tensor x_795_cast_fp16 = conv(bias = layers_28_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_7820, groups = var_7671, pad = x_795_pad_0, pad_type = x_795_pad_type_0, strides = var_7818, weight = layers_28_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_625_cast_fp16)[name = tensor("x_795_cast_fp16")]; tensor layers_28_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_28_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508122432)))]; tensor query_115_cast_fp16 = mul(x = x_795_cast_fp16, y = layers_28_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_115_cast_fp16")]; tensor var_7830 = const()[name = tensor("op_7830"), val = tensor([1, 1])]; tensor var_7832 = const()[name = tensor("op_7832"), val = tensor([1, 1])]; tensor x_797_pad_type_0 = const()[name = tensor("x_797_pad_type_0"), val = tensor("custom")]; tensor x_797_pad_0 = const()[name = tensor("x_797_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_28_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508125056))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508944320))), name = tensor("layers_28_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_28_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_28_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508944448)))]; tensor x_797_cast_fp16 = conv(bias = layers_28_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_7832, groups = var_7671, pad = x_797_pad_0, pad_type = x_797_pad_type_0, strides = var_7830, weight = layers_28_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_797_cast_fp16")]; tensor layers_28_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_28_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508947072)))]; tensor key_115_cast_fp16 = mul(x = x_797_cast_fp16, y = layers_28_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_115_cast_fp16")]; tensor var_7842 = const()[name = tensor("op_7842"), val = tensor([1, 1])]; tensor var_7844 = const()[name = tensor("op_7844"), val = tensor([1, 1])]; tensor x_799_pad_type_0 = const()[name = tensor("x_799_pad_type_0"), val = tensor("custom")]; tensor x_799_pad_0 = const()[name = tensor("x_799_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_28_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508949696))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509768960))), name = tensor("layers_28_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_28_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_28_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509769088)))]; tensor x_799_cast_fp16 = conv(bias = layers_28_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_7844, groups = var_7671, pad = x_799_pad_0, pad_type = x_799_pad_type_0, strides = var_7842, weight = layers_28_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_799_cast_fp16")]; tensor layers_28_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_28_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509771712)))]; tensor value_115_cast_fp16 = mul(x = x_799_cast_fp16, y = layers_28_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_115_cast_fp16")]; tensor var_7849 = const()[name = tensor("op_7849"), val = tensor([1, 20, 64, -1])]; tensor var_7850_cast_fp16 = reshape(shape = var_7849, x = query_115_cast_fp16)[name = tensor("op_7850_cast_fp16")]; tensor var_7851_to_fp16 = const()[name = tensor("op_7851_to_fp16"), val = tensor(0x1p-3)]; tensor var_7852_cast_fp16 = mul(x = var_7850_cast_fp16, y = var_7851_to_fp16)[name = tensor("op_7852_cast_fp16")]; tensor var_7853 = const()[name = tensor("op_7853"), val = tensor([1, 20, 64, -1])]; tensor var_7854_cast_fp16 = reshape(shape = var_7853, x = key_115_cast_fp16)[name = tensor("op_7854_cast_fp16")]; tensor mh_w_173_transpose_x_0 = const()[name = tensor("mh_w_173_transpose_x_0"), val = tensor(true)]; tensor mh_w_173_transpose_y_0 = const()[name = tensor("mh_w_173_transpose_y_0"), val = tensor(false)]; tensor mh_w_173_cast_fp16 = matmul(transpose_x = mh_w_173_transpose_x_0, transpose_y = mh_w_173_transpose_y_0, x = var_7852_cast_fp16, y = var_7854_cast_fp16)[name = tensor("mh_w_173_cast_fp16")]; tensor obj_405_cast_fp16 = softmax(axis = var_7664, x = mh_w_173_cast_fp16)[name = tensor("obj_405_cast_fp16")]; tensor var_7858 = const()[name = tensor("op_7858"), val = tensor([1, 20, 64, -1])]; tensor var_7859_cast_fp16 = reshape(shape = var_7858, x = value_115_cast_fp16)[name = tensor("op_7859_cast_fp16")]; tensor attn_115_transpose_x_0 = const()[name = tensor("attn_115_transpose_x_0"), val = tensor(false)]; tensor attn_115_transpose_y_0 = const()[name = tensor("attn_115_transpose_y_0"), val = tensor(true)]; tensor attn_115_cast_fp16 = matmul(transpose_x = attn_115_transpose_x_0, transpose_y = attn_115_transpose_y_0, x = var_7859_cast_fp16, y = obj_405_cast_fp16)[name = tensor("attn_115_cast_fp16")]; tensor var_7862 = const()[name = tensor("op_7862"), val = tensor([1, 1280, 1, -1])]; tensor x_801_cast_fp16 = reshape(shape = var_7862, x = attn_115_cast_fp16)[name = tensor("x_801_cast_fp16")]; tensor layers_28_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_28_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509774336)))]; tensor input_631_cast_fp16 = sub(x = x_801_cast_fp16, y = layers_28_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_631_cast_fp16")]; tensor var_7870 = const()[name = tensor("op_7870"), val = tensor([1, 1])]; tensor var_7872 = const()[name = tensor("op_7872"), val = tensor([1, 1])]; tensor x_803_pad_type_0 = const()[name = tensor("x_803_pad_type_0"), val = tensor("custom")]; tensor x_803_pad_0 = const()[name = tensor("x_803_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_28_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509776960))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510596224))), name = tensor("layers_28_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_28_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_28_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510596352)))]; tensor x_803_cast_fp16 = conv(bias = layers_28_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_7872, groups = var_7671, pad = x_803_pad_0, pad_type = x_803_pad_type_0, strides = var_7870, weight = layers_28_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_631_cast_fp16)[name = tensor("x_803_cast_fp16")]; tensor layers_28_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_28_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510598976)))]; tensor obj_403_cast_fp16 = mul(x = x_803_cast_fp16, y = layers_28_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_403_cast_fp16")]; tensor inputs_173_cast_fp16 = add(x = inputs_171_cast_fp16, y = obj_403_cast_fp16)[name = tensor("inputs_173_cast_fp16")]; tensor var_7879 = const()[name = tensor("op_7879"), val = tensor([1])]; tensor channels_mean_173_cast_fp16 = reduce_mean(axes = var_7879, keep_dims = var_7672, x = inputs_173_cast_fp16)[name = tensor("channels_mean_173_cast_fp16")]; tensor zero_mean_173_cast_fp16 = sub(x = inputs_173_cast_fp16, y = channels_mean_173_cast_fp16)[name = tensor("zero_mean_173_cast_fp16")]; tensor zero_mean_sq_173_cast_fp16 = mul(x = zero_mean_173_cast_fp16, y = zero_mean_173_cast_fp16)[name = tensor("zero_mean_sq_173_cast_fp16")]; tensor var_7883 = const()[name = tensor("op_7883"), val = tensor([1])]; tensor var_7884_cast_fp16 = reduce_mean(axes = var_7883, keep_dims = var_7672, x = zero_mean_sq_173_cast_fp16)[name = tensor("op_7884_cast_fp16")]; tensor var_7885_to_fp16 = const()[name = tensor("op_7885_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_7886_cast_fp16 = add(x = var_7884_cast_fp16, y = var_7885_to_fp16)[name = tensor("op_7886_cast_fp16")]; tensor denom_173_epsilon_0_to_fp16 = const()[name = tensor("denom_173_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_173_cast_fp16 = rsqrt(epsilon = denom_173_epsilon_0_to_fp16, x = var_7886_cast_fp16)[name = tensor("denom_173_cast_fp16")]; tensor out_173_cast_fp16 = mul(x = zero_mean_173_cast_fp16, y = denom_173_cast_fp16)[name = tensor("out_173_cast_fp16")]; tensor x_805_gamma_0_to_fp16 = const()[name = tensor("x_805_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510601600)))]; tensor x_805_beta_0_to_fp16 = const()[name = tensor("x_805_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510604224)))]; tensor x_805_epsilon_0_to_fp16 = const()[name = tensor("x_805_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_805_cast_fp16 = batch_norm(beta = x_805_beta_0_to_fp16, epsilon = x_805_epsilon_0_to_fp16, gamma = x_805_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_173_cast_fp16)[name = tensor("x_805_cast_fp16")]; tensor layers_28_fc1_input_shift_to_fp16 = const()[name = tensor("layers_28_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510606848)))]; tensor input_633_cast_fp16 = sub(x = x_805_cast_fp16, y = layers_28_fc1_input_shift_to_fp16)[name = tensor("input_633_cast_fp16")]; tensor var_7901 = const()[name = tensor("op_7901"), val = tensor([1, 1])]; tensor var_7903 = const()[name = tensor("op_7903"), val = tensor([1, 1])]; tensor x_807_pad_type_0 = const()[name = tensor("x_807_pad_type_0"), val = tensor("custom")]; tensor x_807_pad_0 = const()[name = tensor("x_807_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_28_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510609472))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513886336))), name = tensor("layers_28_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_28_fc1_module_bias_to_fp16 = const()[name = tensor("layers_28_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513886464)))]; tensor x_807_cast_fp16 = conv(bias = layers_28_fc1_module_bias_to_fp16, dilations = var_7903, groups = var_7671, pad = x_807_pad_0, pad_type = x_807_pad_type_0, strides = var_7901, weight = layers_28_fc1_module_weight_to_fp16_palettized, x = input_633_cast_fp16)[name = tensor("x_807_cast_fp16")]; tensor layers_28_fc1_output_scale_to_fp16 = const()[name = tensor("layers_28_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513896768)))]; tensor input_635_cast_fp16 = mul(x = x_807_cast_fp16, y = layers_28_fc1_output_scale_to_fp16)[name = tensor("input_635_cast_fp16")]; tensor x_809_mode_0 = const()[name = tensor("x_809_mode_0"), val = tensor("EXACT")]; tensor x_809_cast_fp16 = gelu(mode = x_809_mode_0, x = input_635_cast_fp16)[name = tensor("x_809_cast_fp16")]; tensor layers_28_fc2_input_shift_to_fp16 = const()[name = tensor("layers_28_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513907072)))]; tensor input_637_cast_fp16 = sub(x = x_809_cast_fp16, y = layers_28_fc2_input_shift_to_fp16)[name = tensor("input_637_cast_fp16")]; tensor var_7914 = const()[name = tensor("op_7914"), val = tensor([1, 1])]; tensor var_7916 = const()[name = tensor("op_7916"), val = tensor([1, 1])]; tensor x_811_pad_type_0 = const()[name = tensor("x_811_pad_type_0"), val = tensor("custom")]; tensor x_811_pad_0 = const()[name = tensor("x_811_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_28_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513917376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517194240))), name = tensor("layers_28_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_28_fc2_module_bias_to_fp16 = const()[name = tensor("layers_28_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517194368)))]; tensor x_811_cast_fp16 = conv(bias = layers_28_fc2_module_bias_to_fp16, dilations = var_7916, groups = var_7671, pad = x_811_pad_0, pad_type = x_811_pad_type_0, strides = var_7914, weight = layers_28_fc2_module_weight_to_fp16_palettized, x = input_637_cast_fp16)[name = tensor("x_811_cast_fp16")]; tensor layers_28_fc2_output_scale_to_fp16 = const()[name = tensor("layers_28_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517196992)))]; tensor hidden_states_59_cast_fp16 = mul(x = x_811_cast_fp16, y = layers_28_fc2_output_scale_to_fp16)[name = tensor("hidden_states_59_cast_fp16")]; tensor inputs_175_cast_fp16 = add(x = inputs_173_cast_fp16, y = hidden_states_59_cast_fp16)[name = tensor("inputs_175_cast_fp16")]; tensor var_7930 = const()[name = tensor("op_7930"), val = tensor(3)]; tensor var_7937 = const()[name = tensor("op_7937"), val = tensor(1)]; tensor var_7938 = const()[name = tensor("op_7938"), val = tensor(true)]; tensor var_7950 = const()[name = tensor("op_7950"), val = tensor([1])]; tensor channels_mean_175_cast_fp16 = reduce_mean(axes = var_7950, keep_dims = var_7938, x = inputs_175_cast_fp16)[name = tensor("channels_mean_175_cast_fp16")]; tensor zero_mean_175_cast_fp16 = sub(x = inputs_175_cast_fp16, y = channels_mean_175_cast_fp16)[name = tensor("zero_mean_175_cast_fp16")]; tensor zero_mean_sq_175_cast_fp16 = mul(x = zero_mean_175_cast_fp16, y = zero_mean_175_cast_fp16)[name = tensor("zero_mean_sq_175_cast_fp16")]; tensor var_7954 = const()[name = tensor("op_7954"), val = tensor([1])]; tensor var_7955_cast_fp16 = reduce_mean(axes = var_7954, keep_dims = var_7938, x = zero_mean_sq_175_cast_fp16)[name = tensor("op_7955_cast_fp16")]; tensor var_7956_to_fp16 = const()[name = tensor("op_7956_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_7957_cast_fp16 = add(x = var_7955_cast_fp16, y = var_7956_to_fp16)[name = tensor("op_7957_cast_fp16")]; tensor denom_175_epsilon_0_to_fp16 = const()[name = tensor("denom_175_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_175_cast_fp16 = rsqrt(epsilon = denom_175_epsilon_0_to_fp16, x = var_7957_cast_fp16)[name = tensor("denom_175_cast_fp16")]; tensor out_175_cast_fp16 = mul(x = zero_mean_175_cast_fp16, y = denom_175_cast_fp16)[name = tensor("out_175_cast_fp16")]; tensor obj_407_gamma_0_to_fp16 = const()[name = tensor("obj_407_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517199616)))]; tensor obj_407_beta_0_to_fp16 = const()[name = tensor("obj_407_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517202240)))]; tensor obj_407_epsilon_0_to_fp16 = const()[name = tensor("obj_407_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_407_cast_fp16 = batch_norm(beta = obj_407_beta_0_to_fp16, epsilon = obj_407_epsilon_0_to_fp16, gamma = obj_407_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_175_cast_fp16)[name = tensor("obj_407_cast_fp16")]; tensor layers_29_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_29_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517204864)))]; tensor input_639_cast_fp16 = sub(x = obj_407_cast_fp16, y = layers_29_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_639_cast_fp16")]; tensor var_7976 = const()[name = tensor("op_7976"), val = tensor([1, 1])]; tensor var_7978 = const()[name = tensor("op_7978"), val = tensor([1, 1])]; tensor x_813_pad_type_0 = const()[name = tensor("x_813_pad_type_0"), val = tensor("custom")]; tensor x_813_pad_0 = const()[name = tensor("x_813_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_29_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517207488))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(518026752))), name = tensor("layers_29_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_29_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(518026880)))]; tensor x_813_cast_fp16 = conv(bias = layers_29_self_attn_q_proj_module_bias_to_fp16, dilations = var_7978, groups = var_7937, pad = x_813_pad_0, pad_type = x_813_pad_type_0, strides = var_7976, weight = layers_29_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_639_cast_fp16)[name = tensor("x_813_cast_fp16")]; tensor layers_29_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_29_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(518029504)))]; tensor query_117_cast_fp16 = mul(x = x_813_cast_fp16, y = layers_29_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_117_cast_fp16")]; tensor var_7988 = const()[name = tensor("op_7988"), val = tensor([1, 1])]; tensor var_7990 = const()[name = tensor("op_7990"), val = tensor([1, 1])]; tensor x_815_pad_type_0 = const()[name = tensor("x_815_pad_type_0"), val = tensor("custom")]; tensor x_815_pad_0 = const()[name = tensor("x_815_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_29_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(518032128))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(518851392))), name = tensor("layers_29_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_29_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(518851520)))]; tensor x_815_cast_fp16 = conv(bias = layers_29_self_attn_k_proj_module_bias_to_fp16, dilations = var_7990, groups = var_7937, pad = x_815_pad_0, pad_type = x_815_pad_type_0, strides = var_7988, weight = layers_29_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_639_cast_fp16)[name = tensor("x_815_cast_fp16")]; tensor layers_29_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_29_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(518854144)))]; tensor current_key_59_cast_fp16 = mul(x = x_815_cast_fp16, y = layers_29_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_59_cast_fp16")]; tensor var_8000 = const()[name = tensor("op_8000"), val = tensor([1, 1])]; tensor var_8002 = const()[name = tensor("op_8002"), val = tensor([1, 1])]; tensor x_817_pad_type_0 = const()[name = tensor("x_817_pad_type_0"), val = tensor("custom")]; tensor x_817_pad_0 = const()[name = tensor("x_817_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_29_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(518856768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(519676032))), name = tensor("layers_29_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_29_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(519676160)))]; tensor x_817_cast_fp16 = conv(bias = layers_29_self_attn_v_proj_module_bias_to_fp16, dilations = var_8002, groups = var_7937, pad = x_817_pad_0, pad_type = x_817_pad_type_0, strides = var_8000, weight = layers_29_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_639_cast_fp16)[name = tensor("x_817_cast_fp16")]; tensor layers_29_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_29_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(519678784)))]; tensor current_value_59_cast_fp16 = mul(x = x_817_cast_fp16, y = layers_29_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_59_cast_fp16")]; tensor var_8010_cast_fp16 = mul(x = current_key_59_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_8010_cast_fp16")]; tensor var_8012_cast_fp16 = mul(x = var_103_cast_fp16_29, y = var_257_cast_fp16)[name = tensor("op_8012_cast_fp16")]; tensor key_117_cast_fp16 = add(x = var_8010_cast_fp16, y = var_8012_cast_fp16)[name = tensor("key_117_cast_fp16")]; tensor var_8014_cast_fp16 = mul(x = current_value_59_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_8014_cast_fp16")]; tensor var_8016_cast_fp16 = mul(x = var_138_cast_fp16_29, y = var_257_cast_fp16)[name = tensor("op_8016_cast_fp16")]; tensor value_117_cast_fp16 = add(x = var_8014_cast_fp16, y = var_8016_cast_fp16)[name = tensor("value_117_cast_fp16")]; tensor var_8019 = const()[name = tensor("op_8019"), val = tensor([1, 20, 64, -1])]; tensor var_8020_cast_fp16 = reshape(shape = var_8019, x = query_117_cast_fp16)[name = tensor("op_8020_cast_fp16")]; tensor var_8021_to_fp16 = const()[name = tensor("op_8021_to_fp16"), val = tensor(0x1p-3)]; tensor var_8022_cast_fp16 = mul(x = var_8020_cast_fp16, y = var_8021_to_fp16)[name = tensor("op_8022_cast_fp16")]; tensor var_8023 = const()[name = tensor("op_8023"), val = tensor([1, 20, 64, -1])]; tensor var_8024_cast_fp16 = reshape(shape = var_8023, x = key_117_cast_fp16)[name = tensor("op_8024_cast_fp16")]; tensor mh_w_175_transpose_x_0 = const()[name = tensor("mh_w_175_transpose_x_0"), val = tensor(true)]; tensor mh_w_175_transpose_y_0 = const()[name = tensor("mh_w_175_transpose_y_0"), val = tensor(false)]; tensor mh_w_175_cast_fp16 = matmul(transpose_x = mh_w_175_transpose_x_0, transpose_y = mh_w_175_transpose_y_0, x = var_8022_cast_fp16, y = var_8024_cast_fp16)[name = tensor("mh_w_175_cast_fp16")]; tensor mh_w_177_cast_fp16 = add(x = mh_w_175_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_177_cast_fp16")]; tensor var_8032_cast_fp16 = softmax(axis = var_7930, x = mh_w_177_cast_fp16)[name = tensor("op_8032_cast_fp16")]; tensor var_8033 = const()[name = tensor("op_8033"), val = tensor([1, 20, 64, -1])]; tensor var_8034_cast_fp16 = reshape(shape = var_8033, x = value_117_cast_fp16)[name = tensor("op_8034_cast_fp16")]; tensor attn_117_transpose_x_0 = const()[name = tensor("attn_117_transpose_x_0"), val = tensor(false)]; tensor attn_117_transpose_y_0 = const()[name = tensor("attn_117_transpose_y_0"), val = tensor(true)]; tensor attn_117_cast_fp16 = matmul(transpose_x = attn_117_transpose_x_0, transpose_y = attn_117_transpose_y_0, x = var_8034_cast_fp16, y = var_8032_cast_fp16)[name = tensor("attn_117_cast_fp16")]; tensor var_8037 = const()[name = tensor("op_8037"), val = tensor([1, 1280, 1, -1])]; tensor x_819_cast_fp16 = reshape(shape = var_8037, x = attn_117_cast_fp16)[name = tensor("x_819_cast_fp16")]; tensor layers_29_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_29_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(519681408)))]; tensor input_645_cast_fp16 = sub(x = x_819_cast_fp16, y = layers_29_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_645_cast_fp16")]; tensor var_8045 = const()[name = tensor("op_8045"), val = tensor([1, 1])]; tensor var_8047 = const()[name = tensor("op_8047"), val = tensor([1, 1])]; tensor x_821_pad_type_0 = const()[name = tensor("x_821_pad_type_0"), val = tensor("custom")]; tensor x_821_pad_0 = const()[name = tensor("x_821_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_29_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(519684032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520503296))), name = tensor("layers_29_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_29_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520503424)))]; tensor x_821_cast_fp16 = conv(bias = layers_29_self_attn_o_proj_module_bias_to_fp16, dilations = var_8047, groups = var_7937, pad = x_821_pad_0, pad_type = x_821_pad_type_0, strides = var_8045, weight = layers_29_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_645_cast_fp16)[name = tensor("x_821_cast_fp16")]; tensor layers_29_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_29_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520506048)))]; tensor obj_413_cast_fp16 = mul(x = x_821_cast_fp16, y = layers_29_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_413_cast_fp16")]; tensor inputs_177_cast_fp16 = add(x = inputs_175_cast_fp16, y = obj_413_cast_fp16)[name = tensor("inputs_177_cast_fp16")]; tensor var_8058 = const()[name = tensor("op_8058"), val = tensor([1])]; tensor channels_mean_177_cast_fp16 = reduce_mean(axes = var_8058, keep_dims = var_7938, x = inputs_177_cast_fp16)[name = tensor("channels_mean_177_cast_fp16")]; tensor zero_mean_177_cast_fp16 = sub(x = inputs_177_cast_fp16, y = channels_mean_177_cast_fp16)[name = tensor("zero_mean_177_cast_fp16")]; tensor zero_mean_sq_177_cast_fp16 = mul(x = zero_mean_177_cast_fp16, y = zero_mean_177_cast_fp16)[name = tensor("zero_mean_sq_177_cast_fp16")]; tensor var_8062 = const()[name = tensor("op_8062"), val = tensor([1])]; tensor var_8063_cast_fp16 = reduce_mean(axes = var_8062, keep_dims = var_7938, x = zero_mean_sq_177_cast_fp16)[name = tensor("op_8063_cast_fp16")]; tensor var_8064_to_fp16 = const()[name = tensor("op_8064_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_8065_cast_fp16 = add(x = var_8063_cast_fp16, y = var_8064_to_fp16)[name = tensor("op_8065_cast_fp16")]; tensor denom_177_epsilon_0_to_fp16 = const()[name = tensor("denom_177_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_177_cast_fp16 = rsqrt(epsilon = denom_177_epsilon_0_to_fp16, x = var_8065_cast_fp16)[name = tensor("denom_177_cast_fp16")]; tensor out_177_cast_fp16 = mul(x = zero_mean_177_cast_fp16, y = denom_177_cast_fp16)[name = tensor("out_177_cast_fp16")]; tensor obj_415_gamma_0_to_fp16 = const()[name = tensor("obj_415_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520508672)))]; tensor obj_415_beta_0_to_fp16 = const()[name = tensor("obj_415_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520511296)))]; tensor obj_415_epsilon_0_to_fp16 = const()[name = tensor("obj_415_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_415_cast_fp16 = batch_norm(beta = obj_415_beta_0_to_fp16, epsilon = obj_415_epsilon_0_to_fp16, gamma = obj_415_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_177_cast_fp16)[name = tensor("obj_415_cast_fp16")]; tensor layers_29_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_29_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520513920)))]; tensor input_647_cast_fp16 = sub(x = obj_415_cast_fp16, y = layers_29_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_647_cast_fp16")]; tensor var_8084 = const()[name = tensor("op_8084"), val = tensor([1, 1])]; tensor var_8086 = const()[name = tensor("op_8086"), val = tensor([1, 1])]; tensor x_823_pad_type_0 = const()[name = tensor("x_823_pad_type_0"), val = tensor("custom")]; tensor x_823_pad_0 = const()[name = tensor("x_823_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_29_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520516544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521335808))), name = tensor("layers_29_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_29_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_29_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521335936)))]; tensor x_823_cast_fp16 = conv(bias = layers_29_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_8086, groups = var_7937, pad = x_823_pad_0, pad_type = x_823_pad_type_0, strides = var_8084, weight = layers_29_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_647_cast_fp16)[name = tensor("x_823_cast_fp16")]; tensor layers_29_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_29_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521338560)))]; tensor query_119_cast_fp16 = mul(x = x_823_cast_fp16, y = layers_29_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_119_cast_fp16")]; tensor var_8096 = const()[name = tensor("op_8096"), val = tensor([1, 1])]; tensor var_8098 = const()[name = tensor("op_8098"), val = tensor([1, 1])]; tensor x_825_pad_type_0 = const()[name = tensor("x_825_pad_type_0"), val = tensor("custom")]; tensor x_825_pad_0 = const()[name = tensor("x_825_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_29_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521341184))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522160448))), name = tensor("layers_29_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_29_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_29_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522160576)))]; tensor x_825_cast_fp16 = conv(bias = layers_29_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_8098, groups = var_7937, pad = x_825_pad_0, pad_type = x_825_pad_type_0, strides = var_8096, weight = layers_29_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_825_cast_fp16")]; tensor layers_29_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_29_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522163200)))]; tensor key_119_cast_fp16 = mul(x = x_825_cast_fp16, y = layers_29_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_119_cast_fp16")]; tensor var_8108 = const()[name = tensor("op_8108"), val = tensor([1, 1])]; tensor var_8110 = const()[name = tensor("op_8110"), val = tensor([1, 1])]; tensor x_827_pad_type_0 = const()[name = tensor("x_827_pad_type_0"), val = tensor("custom")]; tensor x_827_pad_0 = const()[name = tensor("x_827_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_29_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522165824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522985088))), name = tensor("layers_29_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_29_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_29_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522985216)))]; tensor x_827_cast_fp16 = conv(bias = layers_29_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_8110, groups = var_7937, pad = x_827_pad_0, pad_type = x_827_pad_type_0, strides = var_8108, weight = layers_29_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_827_cast_fp16")]; tensor layers_29_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_29_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522987840)))]; tensor value_119_cast_fp16 = mul(x = x_827_cast_fp16, y = layers_29_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_119_cast_fp16")]; tensor var_8115 = const()[name = tensor("op_8115"), val = tensor([1, 20, 64, -1])]; tensor var_8116_cast_fp16 = reshape(shape = var_8115, x = query_119_cast_fp16)[name = tensor("op_8116_cast_fp16")]; tensor var_8117_to_fp16 = const()[name = tensor("op_8117_to_fp16"), val = tensor(0x1p-3)]; tensor var_8118_cast_fp16 = mul(x = var_8116_cast_fp16, y = var_8117_to_fp16)[name = tensor("op_8118_cast_fp16")]; tensor var_8119 = const()[name = tensor("op_8119"), val = tensor([1, 20, 64, -1])]; tensor var_8120_cast_fp16 = reshape(shape = var_8119, x = key_119_cast_fp16)[name = tensor("op_8120_cast_fp16")]; tensor mh_w_179_transpose_x_0 = const()[name = tensor("mh_w_179_transpose_x_0"), val = tensor(true)]; tensor mh_w_179_transpose_y_0 = const()[name = tensor("mh_w_179_transpose_y_0"), val = tensor(false)]; tensor mh_w_179_cast_fp16 = matmul(transpose_x = mh_w_179_transpose_x_0, transpose_y = mh_w_179_transpose_y_0, x = var_8118_cast_fp16, y = var_8120_cast_fp16)[name = tensor("mh_w_179_cast_fp16")]; tensor obj_419_cast_fp16 = softmax(axis = var_7930, x = mh_w_179_cast_fp16)[name = tensor("obj_419_cast_fp16")]; tensor var_8124 = const()[name = tensor("op_8124"), val = tensor([1, 20, 64, -1])]; tensor var_8125_cast_fp16 = reshape(shape = var_8124, x = value_119_cast_fp16)[name = tensor("op_8125_cast_fp16")]; tensor attn_119_transpose_x_0 = const()[name = tensor("attn_119_transpose_x_0"), val = tensor(false)]; tensor attn_119_transpose_y_0 = const()[name = tensor("attn_119_transpose_y_0"), val = tensor(true)]; tensor attn_119_cast_fp16 = matmul(transpose_x = attn_119_transpose_x_0, transpose_y = attn_119_transpose_y_0, x = var_8125_cast_fp16, y = obj_419_cast_fp16)[name = tensor("attn_119_cast_fp16")]; tensor var_8128 = const()[name = tensor("op_8128"), val = tensor([1, 1280, 1, -1])]; tensor x_829_cast_fp16 = reshape(shape = var_8128, x = attn_119_cast_fp16)[name = tensor("x_829_cast_fp16")]; tensor layers_29_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_29_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522990464)))]; tensor input_653_cast_fp16 = sub(x = x_829_cast_fp16, y = layers_29_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_653_cast_fp16")]; tensor var_8136 = const()[name = tensor("op_8136"), val = tensor([1, 1])]; tensor var_8138 = const()[name = tensor("op_8138"), val = tensor([1, 1])]; tensor x_831_pad_type_0 = const()[name = tensor("x_831_pad_type_0"), val = tensor("custom")]; tensor x_831_pad_0 = const()[name = tensor("x_831_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_29_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522993088))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(523812352))), name = tensor("layers_29_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_29_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_29_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(523812480)))]; tensor x_831_cast_fp16 = conv(bias = layers_29_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_8138, groups = var_7937, pad = x_831_pad_0, pad_type = x_831_pad_type_0, strides = var_8136, weight = layers_29_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_653_cast_fp16)[name = tensor("x_831_cast_fp16")]; tensor layers_29_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_29_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(523815104)))]; tensor obj_417_cast_fp16 = mul(x = x_831_cast_fp16, y = layers_29_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_417_cast_fp16")]; tensor inputs_179_cast_fp16 = add(x = inputs_177_cast_fp16, y = obj_417_cast_fp16)[name = tensor("inputs_179_cast_fp16")]; tensor var_8145 = const()[name = tensor("op_8145"), val = tensor([1])]; tensor channels_mean_179_cast_fp16 = reduce_mean(axes = var_8145, keep_dims = var_7938, x = inputs_179_cast_fp16)[name = tensor("channels_mean_179_cast_fp16")]; tensor zero_mean_179_cast_fp16 = sub(x = inputs_179_cast_fp16, y = channels_mean_179_cast_fp16)[name = tensor("zero_mean_179_cast_fp16")]; tensor zero_mean_sq_179_cast_fp16 = mul(x = zero_mean_179_cast_fp16, y = zero_mean_179_cast_fp16)[name = tensor("zero_mean_sq_179_cast_fp16")]; tensor var_8149 = const()[name = tensor("op_8149"), val = tensor([1])]; tensor var_8150_cast_fp16 = reduce_mean(axes = var_8149, keep_dims = var_7938, x = zero_mean_sq_179_cast_fp16)[name = tensor("op_8150_cast_fp16")]; tensor var_8151_to_fp16 = const()[name = tensor("op_8151_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_8152_cast_fp16 = add(x = var_8150_cast_fp16, y = var_8151_to_fp16)[name = tensor("op_8152_cast_fp16")]; tensor denom_179_epsilon_0_to_fp16 = const()[name = tensor("denom_179_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_179_cast_fp16 = rsqrt(epsilon = denom_179_epsilon_0_to_fp16, x = var_8152_cast_fp16)[name = tensor("denom_179_cast_fp16")]; tensor out_179_cast_fp16 = mul(x = zero_mean_179_cast_fp16, y = denom_179_cast_fp16)[name = tensor("out_179_cast_fp16")]; tensor x_833_gamma_0_to_fp16 = const()[name = tensor("x_833_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(523817728)))]; tensor x_833_beta_0_to_fp16 = const()[name = tensor("x_833_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(523820352)))]; tensor x_833_epsilon_0_to_fp16 = const()[name = tensor("x_833_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_833_cast_fp16 = batch_norm(beta = x_833_beta_0_to_fp16, epsilon = x_833_epsilon_0_to_fp16, gamma = x_833_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_179_cast_fp16)[name = tensor("x_833_cast_fp16")]; tensor layers_29_fc1_input_shift_to_fp16 = const()[name = tensor("layers_29_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(523822976)))]; tensor input_655_cast_fp16 = sub(x = x_833_cast_fp16, y = layers_29_fc1_input_shift_to_fp16)[name = tensor("input_655_cast_fp16")]; tensor var_8167 = const()[name = tensor("op_8167"), val = tensor([1, 1])]; tensor var_8169 = const()[name = tensor("op_8169"), val = tensor([1, 1])]; tensor x_835_pad_type_0 = const()[name = tensor("x_835_pad_type_0"), val = tensor("custom")]; tensor x_835_pad_0 = const()[name = tensor("x_835_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_29_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(523825600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527102464))), name = tensor("layers_29_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_29_fc1_module_bias_to_fp16 = const()[name = tensor("layers_29_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527102592)))]; tensor x_835_cast_fp16 = conv(bias = layers_29_fc1_module_bias_to_fp16, dilations = var_8169, groups = var_7937, pad = x_835_pad_0, pad_type = x_835_pad_type_0, strides = var_8167, weight = layers_29_fc1_module_weight_to_fp16_palettized, x = input_655_cast_fp16)[name = tensor("x_835_cast_fp16")]; tensor layers_29_fc1_output_scale_to_fp16 = const()[name = tensor("layers_29_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527112896)))]; tensor input_657_cast_fp16 = mul(x = x_835_cast_fp16, y = layers_29_fc1_output_scale_to_fp16)[name = tensor("input_657_cast_fp16")]; tensor x_837_mode_0 = const()[name = tensor("x_837_mode_0"), val = tensor("EXACT")]; tensor x_837_cast_fp16 = gelu(mode = x_837_mode_0, x = input_657_cast_fp16)[name = tensor("x_837_cast_fp16")]; tensor layers_29_fc2_input_shift_to_fp16 = const()[name = tensor("layers_29_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527123200)))]; tensor input_659_cast_fp16 = sub(x = x_837_cast_fp16, y = layers_29_fc2_input_shift_to_fp16)[name = tensor("input_659_cast_fp16")]; tensor var_8180 = const()[name = tensor("op_8180"), val = tensor([1, 1])]; tensor var_8182 = const()[name = tensor("op_8182"), val = tensor([1, 1])]; tensor x_839_pad_type_0 = const()[name = tensor("x_839_pad_type_0"), val = tensor("custom")]; tensor x_839_pad_0 = const()[name = tensor("x_839_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_29_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527133504))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530410368))), name = tensor("layers_29_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_29_fc2_module_bias_to_fp16 = const()[name = tensor("layers_29_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530410496)))]; tensor x_839_cast_fp16 = conv(bias = layers_29_fc2_module_bias_to_fp16, dilations = var_8182, groups = var_7937, pad = x_839_pad_0, pad_type = x_839_pad_type_0, strides = var_8180, weight = layers_29_fc2_module_weight_to_fp16_palettized, x = input_659_cast_fp16)[name = tensor("x_839_cast_fp16")]; tensor layers_29_fc2_output_scale_to_fp16 = const()[name = tensor("layers_29_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530413120)))]; tensor hidden_states_61_cast_fp16 = mul(x = x_839_cast_fp16, y = layers_29_fc2_output_scale_to_fp16)[name = tensor("hidden_states_61_cast_fp16")]; tensor inputs_181_cast_fp16 = add(x = inputs_179_cast_fp16, y = hidden_states_61_cast_fp16)[name = tensor("inputs_181_cast_fp16")]; tensor var_8196 = const()[name = tensor("op_8196"), val = tensor(3)]; tensor var_8203 = const()[name = tensor("op_8203"), val = tensor(1)]; tensor var_8204 = const()[name = tensor("op_8204"), val = tensor(true)]; tensor var_8216 = const()[name = tensor("op_8216"), val = tensor([1])]; tensor channels_mean_181_cast_fp16 = reduce_mean(axes = var_8216, keep_dims = var_8204, x = inputs_181_cast_fp16)[name = tensor("channels_mean_181_cast_fp16")]; tensor zero_mean_181_cast_fp16 = sub(x = inputs_181_cast_fp16, y = channels_mean_181_cast_fp16)[name = tensor("zero_mean_181_cast_fp16")]; tensor zero_mean_sq_181_cast_fp16 = mul(x = zero_mean_181_cast_fp16, y = zero_mean_181_cast_fp16)[name = tensor("zero_mean_sq_181_cast_fp16")]; tensor var_8220 = const()[name = tensor("op_8220"), val = tensor([1])]; tensor var_8221_cast_fp16 = reduce_mean(axes = var_8220, keep_dims = var_8204, x = zero_mean_sq_181_cast_fp16)[name = tensor("op_8221_cast_fp16")]; tensor var_8222_to_fp16 = const()[name = tensor("op_8222_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_8223_cast_fp16 = add(x = var_8221_cast_fp16, y = var_8222_to_fp16)[name = tensor("op_8223_cast_fp16")]; tensor denom_181_epsilon_0_to_fp16 = const()[name = tensor("denom_181_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_181_cast_fp16 = rsqrt(epsilon = denom_181_epsilon_0_to_fp16, x = var_8223_cast_fp16)[name = tensor("denom_181_cast_fp16")]; tensor out_181_cast_fp16 = mul(x = zero_mean_181_cast_fp16, y = denom_181_cast_fp16)[name = tensor("out_181_cast_fp16")]; tensor obj_421_gamma_0_to_fp16 = const()[name = tensor("obj_421_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530415744)))]; tensor obj_421_beta_0_to_fp16 = const()[name = tensor("obj_421_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530418368)))]; tensor obj_421_epsilon_0_to_fp16 = const()[name = tensor("obj_421_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_421_cast_fp16 = batch_norm(beta = obj_421_beta_0_to_fp16, epsilon = obj_421_epsilon_0_to_fp16, gamma = obj_421_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_181_cast_fp16)[name = tensor("obj_421_cast_fp16")]; tensor layers_30_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_30_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530420992)))]; tensor input_661_cast_fp16 = sub(x = obj_421_cast_fp16, y = layers_30_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_661_cast_fp16")]; tensor var_8242 = const()[name = tensor("op_8242"), val = tensor([1, 1])]; tensor var_8244 = const()[name = tensor("op_8244"), val = tensor([1, 1])]; tensor x_841_pad_type_0 = const()[name = tensor("x_841_pad_type_0"), val = tensor("custom")]; tensor x_841_pad_0 = const()[name = tensor("x_841_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_30_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530423616))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(531242880))), name = tensor("layers_30_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_30_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(531243008)))]; tensor x_841_cast_fp16 = conv(bias = layers_30_self_attn_q_proj_module_bias_to_fp16, dilations = var_8244, groups = var_8203, pad = x_841_pad_0, pad_type = x_841_pad_type_0, strides = var_8242, weight = layers_30_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_661_cast_fp16)[name = tensor("x_841_cast_fp16")]; tensor layers_30_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_30_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(531245632)))]; tensor query_121_cast_fp16 = mul(x = x_841_cast_fp16, y = layers_30_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_121_cast_fp16")]; tensor var_8254 = const()[name = tensor("op_8254"), val = tensor([1, 1])]; tensor var_8256 = const()[name = tensor("op_8256"), val = tensor([1, 1])]; tensor x_843_pad_type_0 = const()[name = tensor("x_843_pad_type_0"), val = tensor("custom")]; tensor x_843_pad_0 = const()[name = tensor("x_843_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_30_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(531248256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532067520))), name = tensor("layers_30_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_30_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532067648)))]; tensor x_843_cast_fp16 = conv(bias = layers_30_self_attn_k_proj_module_bias_to_fp16, dilations = var_8256, groups = var_8203, pad = x_843_pad_0, pad_type = x_843_pad_type_0, strides = var_8254, weight = layers_30_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_661_cast_fp16)[name = tensor("x_843_cast_fp16")]; tensor layers_30_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_30_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532070272)))]; tensor current_key_61_cast_fp16 = mul(x = x_843_cast_fp16, y = layers_30_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_61_cast_fp16")]; tensor var_8266 = const()[name = tensor("op_8266"), val = tensor([1, 1])]; tensor var_8268 = const()[name = tensor("op_8268"), val = tensor([1, 1])]; tensor x_845_pad_type_0 = const()[name = tensor("x_845_pad_type_0"), val = tensor("custom")]; tensor x_845_pad_0 = const()[name = tensor("x_845_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_30_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532072896))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532892160))), name = tensor("layers_30_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_30_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532892288)))]; tensor x_845_cast_fp16 = conv(bias = layers_30_self_attn_v_proj_module_bias_to_fp16, dilations = var_8268, groups = var_8203, pad = x_845_pad_0, pad_type = x_845_pad_type_0, strides = var_8266, weight = layers_30_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_661_cast_fp16)[name = tensor("x_845_cast_fp16")]; tensor layers_30_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_30_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532894912)))]; tensor current_value_61_cast_fp16 = mul(x = x_845_cast_fp16, y = layers_30_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_61_cast_fp16")]; tensor var_8276_cast_fp16 = mul(x = current_key_61_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_8276_cast_fp16")]; tensor var_8278_cast_fp16 = mul(x = var_103_cast_fp16_30, y = var_257_cast_fp16)[name = tensor("op_8278_cast_fp16")]; tensor key_121_cast_fp16 = add(x = var_8276_cast_fp16, y = var_8278_cast_fp16)[name = tensor("key_121_cast_fp16")]; tensor var_8280_cast_fp16 = mul(x = current_value_61_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_8280_cast_fp16")]; tensor var_8282_cast_fp16 = mul(x = var_138_cast_fp16_30, y = var_257_cast_fp16)[name = tensor("op_8282_cast_fp16")]; tensor value_121_cast_fp16 = add(x = var_8280_cast_fp16, y = var_8282_cast_fp16)[name = tensor("value_121_cast_fp16")]; tensor var_8285 = const()[name = tensor("op_8285"), val = tensor([1, 20, 64, -1])]; tensor var_8286_cast_fp16 = reshape(shape = var_8285, x = query_121_cast_fp16)[name = tensor("op_8286_cast_fp16")]; tensor var_8287_to_fp16 = const()[name = tensor("op_8287_to_fp16"), val = tensor(0x1p-3)]; tensor var_8288_cast_fp16 = mul(x = var_8286_cast_fp16, y = var_8287_to_fp16)[name = tensor("op_8288_cast_fp16")]; tensor var_8289 = const()[name = tensor("op_8289"), val = tensor([1, 20, 64, -1])]; tensor var_8290_cast_fp16 = reshape(shape = var_8289, x = key_121_cast_fp16)[name = tensor("op_8290_cast_fp16")]; tensor mh_w_181_transpose_x_0 = const()[name = tensor("mh_w_181_transpose_x_0"), val = tensor(true)]; tensor mh_w_181_transpose_y_0 = const()[name = tensor("mh_w_181_transpose_y_0"), val = tensor(false)]; tensor mh_w_181_cast_fp16 = matmul(transpose_x = mh_w_181_transpose_x_0, transpose_y = mh_w_181_transpose_y_0, x = var_8288_cast_fp16, y = var_8290_cast_fp16)[name = tensor("mh_w_181_cast_fp16")]; tensor mh_w_183_cast_fp16 = add(x = mh_w_181_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_183_cast_fp16")]; tensor var_8298_cast_fp16 = softmax(axis = var_8196, x = mh_w_183_cast_fp16)[name = tensor("op_8298_cast_fp16")]; tensor var_8299 = const()[name = tensor("op_8299"), val = tensor([1, 20, 64, -1])]; tensor var_8300_cast_fp16 = reshape(shape = var_8299, x = value_121_cast_fp16)[name = tensor("op_8300_cast_fp16")]; tensor attn_121_transpose_x_0 = const()[name = tensor("attn_121_transpose_x_0"), val = tensor(false)]; tensor attn_121_transpose_y_0 = const()[name = tensor("attn_121_transpose_y_0"), val = tensor(true)]; tensor attn_121_cast_fp16 = matmul(transpose_x = attn_121_transpose_x_0, transpose_y = attn_121_transpose_y_0, x = var_8300_cast_fp16, y = var_8298_cast_fp16)[name = tensor("attn_121_cast_fp16")]; tensor var_8303 = const()[name = tensor("op_8303"), val = tensor([1, 1280, 1, -1])]; tensor x_847_cast_fp16 = reshape(shape = var_8303, x = attn_121_cast_fp16)[name = tensor("x_847_cast_fp16")]; tensor layers_30_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_30_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532897536)))]; tensor input_667_cast_fp16 = sub(x = x_847_cast_fp16, y = layers_30_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_667_cast_fp16")]; tensor var_8311 = const()[name = tensor("op_8311"), val = tensor([1, 1])]; tensor var_8313 = const()[name = tensor("op_8313"), val = tensor([1, 1])]; tensor x_849_pad_type_0 = const()[name = tensor("x_849_pad_type_0"), val = tensor("custom")]; tensor x_849_pad_0 = const()[name = tensor("x_849_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_30_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532900160))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(533719424))), name = tensor("layers_30_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_30_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(533719552)))]; tensor x_849_cast_fp16 = conv(bias = layers_30_self_attn_o_proj_module_bias_to_fp16, dilations = var_8313, groups = var_8203, pad = x_849_pad_0, pad_type = x_849_pad_type_0, strides = var_8311, weight = layers_30_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_667_cast_fp16)[name = tensor("x_849_cast_fp16")]; tensor layers_30_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_30_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(533722176)))]; tensor obj_427_cast_fp16 = mul(x = x_849_cast_fp16, y = layers_30_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_427_cast_fp16")]; tensor inputs_183_cast_fp16 = add(x = inputs_181_cast_fp16, y = obj_427_cast_fp16)[name = tensor("inputs_183_cast_fp16")]; tensor var_8324 = const()[name = tensor("op_8324"), val = tensor([1])]; tensor channels_mean_183_cast_fp16 = reduce_mean(axes = var_8324, keep_dims = var_8204, x = inputs_183_cast_fp16)[name = tensor("channels_mean_183_cast_fp16")]; tensor zero_mean_183_cast_fp16 = sub(x = inputs_183_cast_fp16, y = channels_mean_183_cast_fp16)[name = tensor("zero_mean_183_cast_fp16")]; tensor zero_mean_sq_183_cast_fp16 = mul(x = zero_mean_183_cast_fp16, y = zero_mean_183_cast_fp16)[name = tensor("zero_mean_sq_183_cast_fp16")]; tensor var_8328 = const()[name = tensor("op_8328"), val = tensor([1])]; tensor var_8329_cast_fp16 = reduce_mean(axes = var_8328, keep_dims = var_8204, x = zero_mean_sq_183_cast_fp16)[name = tensor("op_8329_cast_fp16")]; tensor var_8330_to_fp16 = const()[name = tensor("op_8330_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_8331_cast_fp16 = add(x = var_8329_cast_fp16, y = var_8330_to_fp16)[name = tensor("op_8331_cast_fp16")]; tensor denom_183_epsilon_0_to_fp16 = const()[name = tensor("denom_183_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_183_cast_fp16 = rsqrt(epsilon = denom_183_epsilon_0_to_fp16, x = var_8331_cast_fp16)[name = tensor("denom_183_cast_fp16")]; tensor out_183_cast_fp16 = mul(x = zero_mean_183_cast_fp16, y = denom_183_cast_fp16)[name = tensor("out_183_cast_fp16")]; tensor obj_429_gamma_0_to_fp16 = const()[name = tensor("obj_429_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(533724800)))]; tensor obj_429_beta_0_to_fp16 = const()[name = tensor("obj_429_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(533727424)))]; tensor obj_429_epsilon_0_to_fp16 = const()[name = tensor("obj_429_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_429_cast_fp16 = batch_norm(beta = obj_429_beta_0_to_fp16, epsilon = obj_429_epsilon_0_to_fp16, gamma = obj_429_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_183_cast_fp16)[name = tensor("obj_429_cast_fp16")]; tensor layers_30_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_30_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(533730048)))]; tensor input_669_cast_fp16 = sub(x = obj_429_cast_fp16, y = layers_30_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_669_cast_fp16")]; tensor var_8350 = const()[name = tensor("op_8350"), val = tensor([1, 1])]; tensor var_8352 = const()[name = tensor("op_8352"), val = tensor([1, 1])]; tensor x_851_pad_type_0 = const()[name = tensor("x_851_pad_type_0"), val = tensor("custom")]; tensor x_851_pad_0 = const()[name = tensor("x_851_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_30_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(533732672))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(534551936))), name = tensor("layers_30_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_30_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_30_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(534552064)))]; tensor x_851_cast_fp16 = conv(bias = layers_30_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_8352, groups = var_8203, pad = x_851_pad_0, pad_type = x_851_pad_type_0, strides = var_8350, weight = layers_30_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_669_cast_fp16)[name = tensor("x_851_cast_fp16")]; tensor layers_30_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_30_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(534554688)))]; tensor query_123_cast_fp16 = mul(x = x_851_cast_fp16, y = layers_30_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_123_cast_fp16")]; tensor var_8362 = const()[name = tensor("op_8362"), val = tensor([1, 1])]; tensor var_8364 = const()[name = tensor("op_8364"), val = tensor([1, 1])]; tensor x_853_pad_type_0 = const()[name = tensor("x_853_pad_type_0"), val = tensor("custom")]; tensor x_853_pad_0 = const()[name = tensor("x_853_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_30_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(534557312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535376576))), name = tensor("layers_30_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_30_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_30_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535376704)))]; tensor x_853_cast_fp16 = conv(bias = layers_30_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_8364, groups = var_8203, pad = x_853_pad_0, pad_type = x_853_pad_type_0, strides = var_8362, weight = layers_30_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_853_cast_fp16")]; tensor layers_30_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_30_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535379328)))]; tensor key_123_cast_fp16 = mul(x = x_853_cast_fp16, y = layers_30_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_123_cast_fp16")]; tensor var_8374 = const()[name = tensor("op_8374"), val = tensor([1, 1])]; tensor var_8376 = const()[name = tensor("op_8376"), val = tensor([1, 1])]; tensor x_855_pad_type_0 = const()[name = tensor("x_855_pad_type_0"), val = tensor("custom")]; tensor x_855_pad_0 = const()[name = tensor("x_855_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_30_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535381952))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536201216))), name = tensor("layers_30_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_30_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_30_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536201344)))]; tensor x_855_cast_fp16 = conv(bias = layers_30_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_8376, groups = var_8203, pad = x_855_pad_0, pad_type = x_855_pad_type_0, strides = var_8374, weight = layers_30_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_855_cast_fp16")]; tensor layers_30_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_30_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536203968)))]; tensor value_123_cast_fp16 = mul(x = x_855_cast_fp16, y = layers_30_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_123_cast_fp16")]; tensor var_8381 = const()[name = tensor("op_8381"), val = tensor([1, 20, 64, -1])]; tensor var_8382_cast_fp16 = reshape(shape = var_8381, x = query_123_cast_fp16)[name = tensor("op_8382_cast_fp16")]; tensor var_8383_to_fp16 = const()[name = tensor("op_8383_to_fp16"), val = tensor(0x1p-3)]; tensor var_8384_cast_fp16 = mul(x = var_8382_cast_fp16, y = var_8383_to_fp16)[name = tensor("op_8384_cast_fp16")]; tensor var_8385 = const()[name = tensor("op_8385"), val = tensor([1, 20, 64, -1])]; tensor var_8386_cast_fp16 = reshape(shape = var_8385, x = key_123_cast_fp16)[name = tensor("op_8386_cast_fp16")]; tensor mh_w_185_transpose_x_0 = const()[name = tensor("mh_w_185_transpose_x_0"), val = tensor(true)]; tensor mh_w_185_transpose_y_0 = const()[name = tensor("mh_w_185_transpose_y_0"), val = tensor(false)]; tensor mh_w_185_cast_fp16 = matmul(transpose_x = mh_w_185_transpose_x_0, transpose_y = mh_w_185_transpose_y_0, x = var_8384_cast_fp16, y = var_8386_cast_fp16)[name = tensor("mh_w_185_cast_fp16")]; tensor obj_433_cast_fp16 = softmax(axis = var_8196, x = mh_w_185_cast_fp16)[name = tensor("obj_433_cast_fp16")]; tensor var_8390 = const()[name = tensor("op_8390"), val = tensor([1, 20, 64, -1])]; tensor var_8391_cast_fp16 = reshape(shape = var_8390, x = value_123_cast_fp16)[name = tensor("op_8391_cast_fp16")]; tensor attn_123_transpose_x_0 = const()[name = tensor("attn_123_transpose_x_0"), val = tensor(false)]; tensor attn_123_transpose_y_0 = const()[name = tensor("attn_123_transpose_y_0"), val = tensor(true)]; tensor attn_123_cast_fp16 = matmul(transpose_x = attn_123_transpose_x_0, transpose_y = attn_123_transpose_y_0, x = var_8391_cast_fp16, y = obj_433_cast_fp16)[name = tensor("attn_123_cast_fp16")]; tensor var_8394 = const()[name = tensor("op_8394"), val = tensor([1, 1280, 1, -1])]; tensor x_857_cast_fp16 = reshape(shape = var_8394, x = attn_123_cast_fp16)[name = tensor("x_857_cast_fp16")]; tensor layers_30_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_30_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536206592)))]; tensor input_675_cast_fp16 = sub(x = x_857_cast_fp16, y = layers_30_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_675_cast_fp16")]; tensor var_8402 = const()[name = tensor("op_8402"), val = tensor([1, 1])]; tensor var_8404 = const()[name = tensor("op_8404"), val = tensor([1, 1])]; tensor x_859_pad_type_0 = const()[name = tensor("x_859_pad_type_0"), val = tensor("custom")]; tensor x_859_pad_0 = const()[name = tensor("x_859_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_30_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536209216))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537028480))), name = tensor("layers_30_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_30_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_30_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537028608)))]; tensor x_859_cast_fp16 = conv(bias = layers_30_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_8404, groups = var_8203, pad = x_859_pad_0, pad_type = x_859_pad_type_0, strides = var_8402, weight = layers_30_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_675_cast_fp16)[name = tensor("x_859_cast_fp16")]; tensor layers_30_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_30_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537031232)))]; tensor obj_431_cast_fp16 = mul(x = x_859_cast_fp16, y = layers_30_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_431_cast_fp16")]; tensor inputs_185_cast_fp16 = add(x = inputs_183_cast_fp16, y = obj_431_cast_fp16)[name = tensor("inputs_185_cast_fp16")]; tensor var_8411 = const()[name = tensor("op_8411"), val = tensor([1])]; tensor channels_mean_185_cast_fp16 = reduce_mean(axes = var_8411, keep_dims = var_8204, x = inputs_185_cast_fp16)[name = tensor("channels_mean_185_cast_fp16")]; tensor zero_mean_185_cast_fp16 = sub(x = inputs_185_cast_fp16, y = channels_mean_185_cast_fp16)[name = tensor("zero_mean_185_cast_fp16")]; tensor zero_mean_sq_185_cast_fp16 = mul(x = zero_mean_185_cast_fp16, y = zero_mean_185_cast_fp16)[name = tensor("zero_mean_sq_185_cast_fp16")]; tensor var_8415 = const()[name = tensor("op_8415"), val = tensor([1])]; tensor var_8416_cast_fp16 = reduce_mean(axes = var_8415, keep_dims = var_8204, x = zero_mean_sq_185_cast_fp16)[name = tensor("op_8416_cast_fp16")]; tensor var_8417_to_fp16 = const()[name = tensor("op_8417_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_8418_cast_fp16 = add(x = var_8416_cast_fp16, y = var_8417_to_fp16)[name = tensor("op_8418_cast_fp16")]; tensor denom_185_epsilon_0_to_fp16 = const()[name = tensor("denom_185_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_185_cast_fp16 = rsqrt(epsilon = denom_185_epsilon_0_to_fp16, x = var_8418_cast_fp16)[name = tensor("denom_185_cast_fp16")]; tensor out_185_cast_fp16 = mul(x = zero_mean_185_cast_fp16, y = denom_185_cast_fp16)[name = tensor("out_185_cast_fp16")]; tensor x_861_gamma_0_to_fp16 = const()[name = tensor("x_861_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537033856)))]; tensor x_861_beta_0_to_fp16 = const()[name = tensor("x_861_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537036480)))]; tensor x_861_epsilon_0_to_fp16 = const()[name = tensor("x_861_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_861_cast_fp16 = batch_norm(beta = x_861_beta_0_to_fp16, epsilon = x_861_epsilon_0_to_fp16, gamma = x_861_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_185_cast_fp16)[name = tensor("x_861_cast_fp16")]; tensor layers_30_fc1_input_shift_to_fp16 = const()[name = tensor("layers_30_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537039104)))]; tensor input_677_cast_fp16 = sub(x = x_861_cast_fp16, y = layers_30_fc1_input_shift_to_fp16)[name = tensor("input_677_cast_fp16")]; tensor var_8433 = const()[name = tensor("op_8433"), val = tensor([1, 1])]; tensor var_8435 = const()[name = tensor("op_8435"), val = tensor([1, 1])]; tensor x_863_pad_type_0 = const()[name = tensor("x_863_pad_type_0"), val = tensor("custom")]; tensor x_863_pad_0 = const()[name = tensor("x_863_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_30_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537041728))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540318592))), name = tensor("layers_30_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_30_fc1_module_bias_to_fp16 = const()[name = tensor("layers_30_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540318720)))]; tensor x_863_cast_fp16 = conv(bias = layers_30_fc1_module_bias_to_fp16, dilations = var_8435, groups = var_8203, pad = x_863_pad_0, pad_type = x_863_pad_type_0, strides = var_8433, weight = layers_30_fc1_module_weight_to_fp16_palettized, x = input_677_cast_fp16)[name = tensor("x_863_cast_fp16")]; tensor layers_30_fc1_output_scale_to_fp16 = const()[name = tensor("layers_30_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540329024)))]; tensor input_679_cast_fp16 = mul(x = x_863_cast_fp16, y = layers_30_fc1_output_scale_to_fp16)[name = tensor("input_679_cast_fp16")]; tensor x_865_mode_0 = const()[name = tensor("x_865_mode_0"), val = tensor("EXACT")]; tensor x_865_cast_fp16 = gelu(mode = x_865_mode_0, x = input_679_cast_fp16)[name = tensor("x_865_cast_fp16")]; tensor layers_30_fc2_input_shift_to_fp16 = const()[name = tensor("layers_30_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540339328)))]; tensor input_681_cast_fp16 = sub(x = x_865_cast_fp16, y = layers_30_fc2_input_shift_to_fp16)[name = tensor("input_681_cast_fp16")]; tensor var_8446 = const()[name = tensor("op_8446"), val = tensor([1, 1])]; tensor var_8448 = const()[name = tensor("op_8448"), val = tensor([1, 1])]; tensor x_867_pad_type_0 = const()[name = tensor("x_867_pad_type_0"), val = tensor("custom")]; tensor x_867_pad_0 = const()[name = tensor("x_867_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_30_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540349632))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543626496))), name = tensor("layers_30_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_30_fc2_module_bias_to_fp16 = const()[name = tensor("layers_30_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543626624)))]; tensor x_867_cast_fp16 = conv(bias = layers_30_fc2_module_bias_to_fp16, dilations = var_8448, groups = var_8203, pad = x_867_pad_0, pad_type = x_867_pad_type_0, strides = var_8446, weight = layers_30_fc2_module_weight_to_fp16_palettized, x = input_681_cast_fp16)[name = tensor("x_867_cast_fp16")]; tensor layers_30_fc2_output_scale_to_fp16 = const()[name = tensor("layers_30_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543629248)))]; tensor hidden_states_63_cast_fp16 = mul(x = x_867_cast_fp16, y = layers_30_fc2_output_scale_to_fp16)[name = tensor("hidden_states_63_cast_fp16")]; tensor inputs_187_cast_fp16 = add(x = inputs_185_cast_fp16, y = hidden_states_63_cast_fp16)[name = tensor("inputs_187_cast_fp16")]; tensor var_8462 = const()[name = tensor("op_8462"), val = tensor(3)]; tensor var_8469 = const()[name = tensor("op_8469"), val = tensor(1)]; tensor var_8470 = const()[name = tensor("op_8470"), val = tensor(true)]; tensor var_8482 = const()[name = tensor("op_8482"), val = tensor([1])]; tensor channels_mean_187_cast_fp16 = reduce_mean(axes = var_8482, keep_dims = var_8470, x = inputs_187_cast_fp16)[name = tensor("channels_mean_187_cast_fp16")]; tensor zero_mean_187_cast_fp16 = sub(x = inputs_187_cast_fp16, y = channels_mean_187_cast_fp16)[name = tensor("zero_mean_187_cast_fp16")]; tensor zero_mean_sq_187_cast_fp16 = mul(x = zero_mean_187_cast_fp16, y = zero_mean_187_cast_fp16)[name = tensor("zero_mean_sq_187_cast_fp16")]; tensor var_8486 = const()[name = tensor("op_8486"), val = tensor([1])]; tensor var_8487_cast_fp16 = reduce_mean(axes = var_8486, keep_dims = var_8470, x = zero_mean_sq_187_cast_fp16)[name = tensor("op_8487_cast_fp16")]; tensor var_8488_to_fp16 = const()[name = tensor("op_8488_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_8489_cast_fp16 = add(x = var_8487_cast_fp16, y = var_8488_to_fp16)[name = tensor("op_8489_cast_fp16")]; tensor denom_187_epsilon_0_to_fp16 = const()[name = tensor("denom_187_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_187_cast_fp16 = rsqrt(epsilon = denom_187_epsilon_0_to_fp16, x = var_8489_cast_fp16)[name = tensor("denom_187_cast_fp16")]; tensor out_187_cast_fp16 = mul(x = zero_mean_187_cast_fp16, y = denom_187_cast_fp16)[name = tensor("out_187_cast_fp16")]; tensor obj_435_gamma_0_to_fp16 = const()[name = tensor("obj_435_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543631872)))]; tensor obj_435_beta_0_to_fp16 = const()[name = tensor("obj_435_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543634496)))]; tensor obj_435_epsilon_0_to_fp16 = const()[name = tensor("obj_435_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_435_cast_fp16 = batch_norm(beta = obj_435_beta_0_to_fp16, epsilon = obj_435_epsilon_0_to_fp16, gamma = obj_435_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_187_cast_fp16)[name = tensor("obj_435_cast_fp16")]; tensor layers_31_self_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_31_self_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543637120)))]; tensor input_683_cast_fp16 = sub(x = obj_435_cast_fp16, y = layers_31_self_attn_q_proj_input_shift_to_fp16)[name = tensor("input_683_cast_fp16")]; tensor var_8508 = const()[name = tensor("op_8508"), val = tensor([1, 1])]; tensor var_8510 = const()[name = tensor("op_8510"), val = tensor([1, 1])]; tensor x_869_pad_type_0 = const()[name = tensor("x_869_pad_type_0"), val = tensor("custom")]; tensor x_869_pad_0 = const()[name = tensor("x_869_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_31_self_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543639744))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(544459008))), name = tensor("layers_31_self_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_31_self_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(544459136)))]; tensor x_869_cast_fp16 = conv(bias = layers_31_self_attn_q_proj_module_bias_to_fp16, dilations = var_8510, groups = var_8469, pad = x_869_pad_0, pad_type = x_869_pad_type_0, strides = var_8508, weight = layers_31_self_attn_q_proj_module_weight_to_fp16_palettized, x = input_683_cast_fp16)[name = tensor("x_869_cast_fp16")]; tensor layers_31_self_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_31_self_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(544461760)))]; tensor query_125_cast_fp16 = mul(x = x_869_cast_fp16, y = layers_31_self_attn_q_proj_output_scale_to_fp16)[name = tensor("query_125_cast_fp16")]; tensor var_8520 = const()[name = tensor("op_8520"), val = tensor([1, 1])]; tensor var_8522 = const()[name = tensor("op_8522"), val = tensor([1, 1])]; tensor x_871_pad_type_0 = const()[name = tensor("x_871_pad_type_0"), val = tensor("custom")]; tensor x_871_pad_0 = const()[name = tensor("x_871_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_31_self_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(544464384))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545283648))), name = tensor("layers_31_self_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_31_self_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545283776)))]; tensor x_871_cast_fp16 = conv(bias = layers_31_self_attn_k_proj_module_bias_to_fp16, dilations = var_8522, groups = var_8469, pad = x_871_pad_0, pad_type = x_871_pad_type_0, strides = var_8520, weight = layers_31_self_attn_k_proj_module_weight_to_fp16_palettized, x = input_683_cast_fp16)[name = tensor("x_871_cast_fp16")]; tensor layers_31_self_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_31_self_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545286400)))]; tensor current_key_cast_fp16 = mul(x = x_871_cast_fp16, y = layers_31_self_attn_k_proj_output_scale_to_fp16)[name = tensor("current_key_cast_fp16")]; tensor var_8532 = const()[name = tensor("op_8532"), val = tensor([1, 1])]; tensor var_8534 = const()[name = tensor("op_8534"), val = tensor([1, 1])]; tensor x_873_pad_type_0 = const()[name = tensor("x_873_pad_type_0"), val = tensor("custom")]; tensor x_873_pad_0 = const()[name = tensor("x_873_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_31_self_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545289024))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546108288))), name = tensor("layers_31_self_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_31_self_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546108416)))]; tensor x_873_cast_fp16 = conv(bias = layers_31_self_attn_v_proj_module_bias_to_fp16, dilations = var_8534, groups = var_8469, pad = x_873_pad_0, pad_type = x_873_pad_type_0, strides = var_8532, weight = layers_31_self_attn_v_proj_module_weight_to_fp16_palettized, x = input_683_cast_fp16)[name = tensor("x_873_cast_fp16")]; tensor layers_31_self_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_31_self_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546111040)))]; tensor current_value_cast_fp16 = mul(x = x_873_cast_fp16, y = layers_31_self_attn_v_proj_output_scale_to_fp16)[name = tensor("current_value_cast_fp16")]; tensor var_8542_cast_fp16 = mul(x = current_key_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_8542_cast_fp16")]; tensor var_8544_cast_fp16 = mul(x = var_103_cast_fp16_31, y = var_257_cast_fp16)[name = tensor("op_8544_cast_fp16")]; tensor key_125_cast_fp16 = add(x = var_8542_cast_fp16, y = var_8544_cast_fp16)[name = tensor("key_125_cast_fp16")]; tensor var_8546_cast_fp16 = mul(x = current_value_cast_fp16, y = var_254_cast_fp16)[name = tensor("op_8546_cast_fp16")]; tensor var_8548_cast_fp16 = mul(x = var_138_cast_fp16_31, y = var_257_cast_fp16)[name = tensor("op_8548_cast_fp16")]; tensor value_125_cast_fp16 = add(x = var_8546_cast_fp16, y = var_8548_cast_fp16)[name = tensor("value_125_cast_fp16")]; tensor var_8551 = const()[name = tensor("op_8551"), val = tensor([1, 20, 64, -1])]; tensor var_8552_cast_fp16 = reshape(shape = var_8551, x = query_125_cast_fp16)[name = tensor("op_8552_cast_fp16")]; tensor var_8553_to_fp16 = const()[name = tensor("op_8553_to_fp16"), val = tensor(0x1p-3)]; tensor var_8554_cast_fp16 = mul(x = var_8552_cast_fp16, y = var_8553_to_fp16)[name = tensor("op_8554_cast_fp16")]; tensor var_8555 = const()[name = tensor("op_8555"), val = tensor([1, 20, 64, -1])]; tensor var_8556_cast_fp16 = reshape(shape = var_8555, x = key_125_cast_fp16)[name = tensor("op_8556_cast_fp16")]; tensor mh_w_187_transpose_x_0 = const()[name = tensor("mh_w_187_transpose_x_0"), val = tensor(true)]; tensor mh_w_187_transpose_y_0 = const()[name = tensor("mh_w_187_transpose_y_0"), val = tensor(false)]; tensor mh_w_187_cast_fp16 = matmul(transpose_x = mh_w_187_transpose_x_0, transpose_y = mh_w_187_transpose_y_0, x = var_8554_cast_fp16, y = var_8556_cast_fp16)[name = tensor("mh_w_187_cast_fp16")]; tensor mh_w_189_cast_fp16 = add(x = mh_w_187_cast_fp16, y = var_275_cast_fp16)[name = tensor("mh_w_189_cast_fp16")]; tensor var_8564_cast_fp16 = softmax(axis = var_8462, x = mh_w_189_cast_fp16)[name = tensor("op_8564_cast_fp16")]; tensor var_8565 = const()[name = tensor("op_8565"), val = tensor([1, 20, 64, -1])]; tensor var_8566_cast_fp16 = reshape(shape = var_8565, x = value_125_cast_fp16)[name = tensor("op_8566_cast_fp16")]; tensor attn_125_transpose_x_0 = const()[name = tensor("attn_125_transpose_x_0"), val = tensor(false)]; tensor attn_125_transpose_y_0 = const()[name = tensor("attn_125_transpose_y_0"), val = tensor(true)]; tensor attn_125_cast_fp16 = matmul(transpose_x = attn_125_transpose_x_0, transpose_y = attn_125_transpose_y_0, x = var_8566_cast_fp16, y = var_8564_cast_fp16)[name = tensor("attn_125_cast_fp16")]; tensor var_8569 = const()[name = tensor("op_8569"), val = tensor([1, 1280, 1, -1])]; tensor x_875_cast_fp16 = reshape(shape = var_8569, x = attn_125_cast_fp16)[name = tensor("x_875_cast_fp16")]; tensor layers_31_self_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_31_self_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546113664)))]; tensor input_689_cast_fp16 = sub(x = x_875_cast_fp16, y = layers_31_self_attn_o_proj_input_shift_to_fp16)[name = tensor("input_689_cast_fp16")]; tensor var_8577 = const()[name = tensor("op_8577"), val = tensor([1, 1])]; tensor var_8579 = const()[name = tensor("op_8579"), val = tensor([1, 1])]; tensor x_877_pad_type_0 = const()[name = tensor("x_877_pad_type_0"), val = tensor("custom")]; tensor x_877_pad_0 = const()[name = tensor("x_877_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_31_self_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546116288))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546935552))), name = tensor("layers_31_self_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_31_self_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546935680)))]; tensor x_877_cast_fp16 = conv(bias = layers_31_self_attn_o_proj_module_bias_to_fp16, dilations = var_8579, groups = var_8469, pad = x_877_pad_0, pad_type = x_877_pad_type_0, strides = var_8577, weight = layers_31_self_attn_o_proj_module_weight_to_fp16_palettized, x = input_689_cast_fp16)[name = tensor("x_877_cast_fp16")]; tensor layers_31_self_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_31_self_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546938304)))]; tensor obj_441_cast_fp16 = mul(x = x_877_cast_fp16, y = layers_31_self_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_441_cast_fp16")]; tensor inputs_189_cast_fp16 = add(x = inputs_187_cast_fp16, y = obj_441_cast_fp16)[name = tensor("inputs_189_cast_fp16")]; tensor var_8590 = const()[name = tensor("op_8590"), val = tensor([1])]; tensor channels_mean_189_cast_fp16 = reduce_mean(axes = var_8590, keep_dims = var_8470, x = inputs_189_cast_fp16)[name = tensor("channels_mean_189_cast_fp16")]; tensor zero_mean_189_cast_fp16 = sub(x = inputs_189_cast_fp16, y = channels_mean_189_cast_fp16)[name = tensor("zero_mean_189_cast_fp16")]; tensor zero_mean_sq_189_cast_fp16 = mul(x = zero_mean_189_cast_fp16, y = zero_mean_189_cast_fp16)[name = tensor("zero_mean_sq_189_cast_fp16")]; tensor var_8594 = const()[name = tensor("op_8594"), val = tensor([1])]; tensor var_8595_cast_fp16 = reduce_mean(axes = var_8594, keep_dims = var_8470, x = zero_mean_sq_189_cast_fp16)[name = tensor("op_8595_cast_fp16")]; tensor var_8596_to_fp16 = const()[name = tensor("op_8596_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_8597_cast_fp16 = add(x = var_8595_cast_fp16, y = var_8596_to_fp16)[name = tensor("op_8597_cast_fp16")]; tensor denom_189_epsilon_0_to_fp16 = const()[name = tensor("denom_189_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_189_cast_fp16 = rsqrt(epsilon = denom_189_epsilon_0_to_fp16, x = var_8597_cast_fp16)[name = tensor("denom_189_cast_fp16")]; tensor out_189_cast_fp16 = mul(x = zero_mean_189_cast_fp16, y = denom_189_cast_fp16)[name = tensor("out_189_cast_fp16")]; tensor obj_443_gamma_0_to_fp16 = const()[name = tensor("obj_443_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546940928)))]; tensor obj_443_beta_0_to_fp16 = const()[name = tensor("obj_443_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546943552)))]; tensor obj_443_epsilon_0_to_fp16 = const()[name = tensor("obj_443_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_443_cast_fp16 = batch_norm(beta = obj_443_beta_0_to_fp16, epsilon = obj_443_epsilon_0_to_fp16, gamma = obj_443_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_189_cast_fp16)[name = tensor("obj_443_cast_fp16")]; tensor layers_31_encoder_attn_q_proj_input_shift_to_fp16 = const()[name = tensor("layers_31_encoder_attn_q_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546946176)))]; tensor input_691_cast_fp16 = sub(x = obj_443_cast_fp16, y = layers_31_encoder_attn_q_proj_input_shift_to_fp16)[name = tensor("input_691_cast_fp16")]; tensor var_8616 = const()[name = tensor("op_8616"), val = tensor([1, 1])]; tensor var_8618 = const()[name = tensor("op_8618"), val = tensor([1, 1])]; tensor x_879_pad_type_0 = const()[name = tensor("x_879_pad_type_0"), val = tensor("custom")]; tensor x_879_pad_0 = const()[name = tensor("x_879_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_31_encoder_attn_q_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546948800))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547768064))), name = tensor("layers_31_encoder_attn_q_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_31_encoder_attn_q_proj_module_bias_to_fp16 = const()[name = tensor("layers_31_encoder_attn_q_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547768192)))]; tensor x_879_cast_fp16 = conv(bias = layers_31_encoder_attn_q_proj_module_bias_to_fp16, dilations = var_8618, groups = var_8469, pad = x_879_pad_0, pad_type = x_879_pad_type_0, strides = var_8616, weight = layers_31_encoder_attn_q_proj_module_weight_to_fp16_palettized, x = input_691_cast_fp16)[name = tensor("x_879_cast_fp16")]; tensor layers_31_encoder_attn_q_proj_output_scale_to_fp16 = const()[name = tensor("layers_31_encoder_attn_q_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547770816)))]; tensor query_cast_fp16 = mul(x = x_879_cast_fp16, y = layers_31_encoder_attn_q_proj_output_scale_to_fp16)[name = tensor("query_cast_fp16")]; tensor var_8628 = const()[name = tensor("op_8628"), val = tensor([1, 1])]; tensor var_8630 = const()[name = tensor("op_8630"), val = tensor([1, 1])]; tensor x_881_pad_type_0 = const()[name = tensor("x_881_pad_type_0"), val = tensor("custom")]; tensor x_881_pad_0 = const()[name = tensor("x_881_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_31_encoder_attn_k_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547773440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(548592704))), name = tensor("layers_31_encoder_attn_k_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_31_encoder_attn_k_proj_module_bias_to_fp16 = const()[name = tensor("layers_31_encoder_attn_k_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(548592832)))]; tensor x_881_cast_fp16 = conv(bias = layers_31_encoder_attn_k_proj_module_bias_to_fp16, dilations = var_8630, groups = var_8469, pad = x_881_pad_0, pad_type = x_881_pad_type_0, strides = var_8628, weight = layers_31_encoder_attn_k_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_881_cast_fp16")]; tensor layers_31_encoder_attn_k_proj_output_scale_to_fp16 = const()[name = tensor("layers_31_encoder_attn_k_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(548595456)))]; tensor key_cast_fp16 = mul(x = x_881_cast_fp16, y = layers_31_encoder_attn_k_proj_output_scale_to_fp16)[name = tensor("key_cast_fp16")]; tensor var_8640 = const()[name = tensor("op_8640"), val = tensor([1, 1])]; tensor var_8642 = const()[name = tensor("op_8642"), val = tensor([1, 1])]; tensor x_883_pad_type_0 = const()[name = tensor("x_883_pad_type_0"), val = tensor("custom")]; tensor x_883_pad_0 = const()[name = tensor("x_883_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_31_encoder_attn_v_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(548598080))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549417344))), name = tensor("layers_31_encoder_attn_v_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_31_encoder_attn_v_proj_module_bias_to_fp16 = const()[name = tensor("layers_31_encoder_attn_v_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549417472)))]; tensor x_883_cast_fp16 = conv(bias = layers_31_encoder_attn_v_proj_module_bias_to_fp16, dilations = var_8642, groups = var_8469, pad = x_883_pad_0, pad_type = x_883_pad_type_0, strides = var_8640, weight = layers_31_encoder_attn_v_proj_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("x_883_cast_fp16")]; tensor layers_31_encoder_attn_v_proj_output_scale_to_fp16 = const()[name = tensor("layers_31_encoder_attn_v_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549420096)))]; tensor value_cast_fp16 = mul(x = x_883_cast_fp16, y = layers_31_encoder_attn_v_proj_output_scale_to_fp16)[name = tensor("value_cast_fp16")]; tensor var_8647 = const()[name = tensor("op_8647"), val = tensor([1, 20, 64, -1])]; tensor var_8648_cast_fp16 = reshape(shape = var_8647, x = query_cast_fp16)[name = tensor("op_8648_cast_fp16")]; tensor var_8649_to_fp16 = const()[name = tensor("op_8649_to_fp16"), val = tensor(0x1p-3)]; tensor var_8650_cast_fp16 = mul(x = var_8648_cast_fp16, y = var_8649_to_fp16)[name = tensor("op_8650_cast_fp16")]; tensor var_8651 = const()[name = tensor("op_8651"), val = tensor([1, 20, 64, -1])]; tensor var_8652_cast_fp16 = reshape(shape = var_8651, x = key_cast_fp16)[name = tensor("op_8652_cast_fp16")]; tensor mh_w_transpose_x_0 = const()[name = tensor("mh_w_transpose_x_0"), val = tensor(true)]; tensor mh_w_transpose_y_0 = const()[name = tensor("mh_w_transpose_y_0"), val = tensor(false)]; tensor mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_8650_cast_fp16, y = var_8652_cast_fp16)[name = tensor("mh_w_cast_fp16")]; tensor obj_447_cast_fp16 = softmax(axis = var_8462, x = mh_w_cast_fp16)[name = tensor("obj_447_cast_fp16")]; tensor var_8656 = const()[name = tensor("op_8656"), val = tensor([1, 20, 64, -1])]; tensor var_8657_cast_fp16 = reshape(shape = var_8656, x = value_cast_fp16)[name = tensor("op_8657_cast_fp16")]; tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(true)]; tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_8657_cast_fp16, y = obj_447_cast_fp16)[name = tensor("attn_cast_fp16")]; tensor var_8660 = const()[name = tensor("op_8660"), val = tensor([1, 1280, 1, -1])]; tensor x_885_cast_fp16 = reshape(shape = var_8660, x = attn_cast_fp16)[name = tensor("x_885_cast_fp16")]; tensor layers_31_encoder_attn_o_proj_input_shift_to_fp16 = const()[name = tensor("layers_31_encoder_attn_o_proj_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549422720)))]; tensor input_697_cast_fp16 = sub(x = x_885_cast_fp16, y = layers_31_encoder_attn_o_proj_input_shift_to_fp16)[name = tensor("input_697_cast_fp16")]; tensor var_8668 = const()[name = tensor("op_8668"), val = tensor([1, 1])]; tensor var_8670 = const()[name = tensor("op_8670"), val = tensor([1, 1])]; tensor x_887_pad_type_0 = const()[name = tensor("x_887_pad_type_0"), val = tensor("custom")]; tensor x_887_pad_0 = const()[name = tensor("x_887_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_31_encoder_attn_o_proj_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549425344))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550244608))), name = tensor("layers_31_encoder_attn_o_proj_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_31_encoder_attn_o_proj_module_bias_to_fp16 = const()[name = tensor("layers_31_encoder_attn_o_proj_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550244736)))]; tensor x_887_cast_fp16 = conv(bias = layers_31_encoder_attn_o_proj_module_bias_to_fp16, dilations = var_8670, groups = var_8469, pad = x_887_pad_0, pad_type = x_887_pad_type_0, strides = var_8668, weight = layers_31_encoder_attn_o_proj_module_weight_to_fp16_palettized, x = input_697_cast_fp16)[name = tensor("x_887_cast_fp16")]; tensor layers_31_encoder_attn_o_proj_output_scale_to_fp16 = const()[name = tensor("layers_31_encoder_attn_o_proj_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550247360)))]; tensor obj_445_cast_fp16 = mul(x = x_887_cast_fp16, y = layers_31_encoder_attn_o_proj_output_scale_to_fp16)[name = tensor("obj_445_cast_fp16")]; tensor inputs_191_cast_fp16 = add(x = inputs_189_cast_fp16, y = obj_445_cast_fp16)[name = tensor("inputs_191_cast_fp16")]; tensor var_8677 = const()[name = tensor("op_8677"), val = tensor([1])]; tensor channels_mean_191_cast_fp16 = reduce_mean(axes = var_8677, keep_dims = var_8470, x = inputs_191_cast_fp16)[name = tensor("channels_mean_191_cast_fp16")]; tensor zero_mean_191_cast_fp16 = sub(x = inputs_191_cast_fp16, y = channels_mean_191_cast_fp16)[name = tensor("zero_mean_191_cast_fp16")]; tensor zero_mean_sq_191_cast_fp16 = mul(x = zero_mean_191_cast_fp16, y = zero_mean_191_cast_fp16)[name = tensor("zero_mean_sq_191_cast_fp16")]; tensor var_8681 = const()[name = tensor("op_8681"), val = tensor([1])]; tensor var_8682_cast_fp16 = reduce_mean(axes = var_8681, keep_dims = var_8470, x = zero_mean_sq_191_cast_fp16)[name = tensor("op_8682_cast_fp16")]; tensor var_8683_to_fp16 = const()[name = tensor("op_8683_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_8684_cast_fp16 = add(x = var_8682_cast_fp16, y = var_8683_to_fp16)[name = tensor("op_8684_cast_fp16")]; tensor denom_191_epsilon_0_to_fp16 = const()[name = tensor("denom_191_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_191_cast_fp16 = rsqrt(epsilon = denom_191_epsilon_0_to_fp16, x = var_8684_cast_fp16)[name = tensor("denom_191_cast_fp16")]; tensor out_191_cast_fp16 = mul(x = zero_mean_191_cast_fp16, y = denom_191_cast_fp16)[name = tensor("out_191_cast_fp16")]; tensor x_889_gamma_0_to_fp16 = const()[name = tensor("x_889_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550249984)))]; tensor x_889_beta_0_to_fp16 = const()[name = tensor("x_889_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550252608)))]; tensor x_889_epsilon_0_to_fp16 = const()[name = tensor("x_889_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor x_889_cast_fp16 = batch_norm(beta = x_889_beta_0_to_fp16, epsilon = x_889_epsilon_0_to_fp16, gamma = x_889_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_191_cast_fp16)[name = tensor("x_889_cast_fp16")]; tensor layers_31_fc1_input_shift_to_fp16 = const()[name = tensor("layers_31_fc1_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550255232)))]; tensor input_699_cast_fp16 = sub(x = x_889_cast_fp16, y = layers_31_fc1_input_shift_to_fp16)[name = tensor("input_699_cast_fp16")]; tensor var_8699 = const()[name = tensor("op_8699"), val = tensor([1, 1])]; tensor var_8701 = const()[name = tensor("op_8701"), val = tensor([1, 1])]; tensor x_891_pad_type_0 = const()[name = tensor("x_891_pad_type_0"), val = tensor("custom")]; tensor x_891_pad_0 = const()[name = tensor("x_891_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_31_fc1_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550257856))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553534720))), name = tensor("layers_31_fc1_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_31_fc1_module_bias_to_fp16 = const()[name = tensor("layers_31_fc1_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553534848)))]; tensor x_891_cast_fp16 = conv(bias = layers_31_fc1_module_bias_to_fp16, dilations = var_8701, groups = var_8469, pad = x_891_pad_0, pad_type = x_891_pad_type_0, strides = var_8699, weight = layers_31_fc1_module_weight_to_fp16_palettized, x = input_699_cast_fp16)[name = tensor("x_891_cast_fp16")]; tensor layers_31_fc1_output_scale_to_fp16 = const()[name = tensor("layers_31_fc1_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553545152)))]; tensor input_701_cast_fp16 = mul(x = x_891_cast_fp16, y = layers_31_fc1_output_scale_to_fp16)[name = tensor("input_701_cast_fp16")]; tensor x_893_mode_0 = const()[name = tensor("x_893_mode_0"), val = tensor("EXACT")]; tensor x_893_cast_fp16 = gelu(mode = x_893_mode_0, x = input_701_cast_fp16)[name = tensor("x_893_cast_fp16")]; tensor layers_31_fc2_input_shift_to_fp16 = const()[name = tensor("layers_31_fc2_input_shift_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553555456)))]; tensor input_cast_fp16 = sub(x = x_893_cast_fp16, y = layers_31_fc2_input_shift_to_fp16)[name = tensor("input_cast_fp16")]; tensor var_8712 = const()[name = tensor("op_8712"), val = tensor([1, 1])]; tensor var_8714 = const()[name = tensor("op_8714"), val = tensor([1, 1])]; tensor x_pad_type_0 = const()[name = tensor("x_pad_type_0"), val = tensor("custom")]; tensor x_pad_0 = const()[name = tensor("x_pad_0"), val = tensor([0, 0, 0, 0])]; tensor layers_31_fc2_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553565760))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556842624))), name = tensor("layers_31_fc2_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_31_fc2_module_bias_to_fp16 = const()[name = tensor("layers_31_fc2_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556842752)))]; tensor x_cast_fp16 = conv(bias = layers_31_fc2_module_bias_to_fp16, dilations = var_8714, groups = var_8469, pad = x_pad_0, pad_type = x_pad_type_0, strides = var_8712, weight = layers_31_fc2_module_weight_to_fp16_palettized, x = input_cast_fp16)[name = tensor("x_cast_fp16")]; tensor layers_31_fc2_output_scale_to_fp16 = const()[name = tensor("layers_31_fc2_output_scale_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556845376)))]; tensor hidden_states_65_cast_fp16 = mul(x = x_cast_fp16, y = layers_31_fc2_output_scale_to_fp16)[name = tensor("hidden_states_65_cast_fp16")]; tensor inputs_cast_fp16 = add(x = inputs_191_cast_fp16, y = hidden_states_65_cast_fp16)[name = tensor("inputs_cast_fp16")]; tensor var_8725 = const()[name = tensor("op_8725"), val = tensor(true)]; tensor var_8729 = const()[name = tensor("op_8729"), val = tensor([1])]; tensor channels_mean_cast_fp16 = reduce_mean(axes = var_8729, keep_dims = var_8725, x = inputs_cast_fp16)[name = tensor("channels_mean_cast_fp16")]; tensor zero_mean_cast_fp16 = sub(x = inputs_cast_fp16, y = channels_mean_cast_fp16)[name = tensor("zero_mean_cast_fp16")]; tensor zero_mean_sq_cast_fp16 = mul(x = zero_mean_cast_fp16, y = zero_mean_cast_fp16)[name = tensor("zero_mean_sq_cast_fp16")]; tensor var_8733 = const()[name = tensor("op_8733"), val = tensor([1])]; tensor var_8734_cast_fp16 = reduce_mean(axes = var_8733, keep_dims = var_8725, x = zero_mean_sq_cast_fp16)[name = tensor("op_8734_cast_fp16")]; tensor var_8735_to_fp16 = const()[name = tensor("op_8735_to_fp16"), val = tensor(0x1.5p-17)]; tensor var_8736_cast_fp16 = add(x = var_8734_cast_fp16, y = var_8735_to_fp16)[name = tensor("op_8736_cast_fp16")]; tensor denom_epsilon_0_to_fp16 = const()[name = tensor("denom_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; tensor denom_cast_fp16 = rsqrt(epsilon = denom_epsilon_0_to_fp16, x = var_8736_cast_fp16)[name = tensor("denom_cast_fp16")]; tensor out_cast_fp16 = mul(x = zero_mean_cast_fp16, y = denom_cast_fp16)[name = tensor("out_cast_fp16")]; tensor hidden_states_gamma_0_to_fp16 = const()[name = tensor("hidden_states_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556848000)))]; tensor hidden_states_beta_0_to_fp16 = const()[name = tensor("hidden_states_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556850624)))]; tensor hidden_states_epsilon_0_to_fp16 = const()[name = tensor("hidden_states_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor("hidden_states_cast_fp16")]; tensor var_8746_axes_0 = const()[name = tensor("op_8746_axes_0"), val = tensor([2])]; tensor var_8746_cast_fp16 = squeeze(axes = var_8746_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_8746_cast_fp16")]; tensor var_8749_perm_0 = const()[name = tensor("op_8749_perm_0"), val = tensor([0, 2, 1])]; tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556853248)))]; tensor transpose_0 = transpose(perm = var_8749_perm_0, x = var_8746_cast_fp16)[name = tensor("transpose_0")]; tensor logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = transpose_0)[name = tensor("linear_0_cast_fp16")]; tensor var_8753 = const()[name = tensor("op_8753"), val = tensor(1)]; tensor obj_451_interleave_0 = const()[name = tensor("obj_451_interleave_0"), val = tensor(false)]; tensor key_cache_updates = concat(axis = var_8753, interleave = obj_451_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_7_cast_fp16, current_key_9_cast_fp16, current_key_11_cast_fp16, current_key_13_cast_fp16, current_key_15_cast_fp16, current_key_17_cast_fp16, current_key_19_cast_fp16, current_key_21_cast_fp16, current_key_23_cast_fp16, current_key_25_cast_fp16, current_key_27_cast_fp16, current_key_29_cast_fp16, current_key_31_cast_fp16, current_key_33_cast_fp16, current_key_35_cast_fp16, current_key_37_cast_fp16, current_key_39_cast_fp16, current_key_41_cast_fp16, current_key_43_cast_fp16, current_key_45_cast_fp16, current_key_47_cast_fp16, current_key_49_cast_fp16, current_key_51_cast_fp16, current_key_53_cast_fp16, current_key_55_cast_fp16, current_key_57_cast_fp16, current_key_59_cast_fp16, current_key_61_cast_fp16, current_key_cast_fp16))[name = tensor("obj_451_cast_fp16")]; tensor var_8756 = const()[name = tensor("op_8756"), val = tensor(1)]; tensor obj_453_interleave_0 = const()[name = tensor("obj_453_interleave_0"), val = tensor(false)]; tensor value_cache_updates = concat(axis = var_8756, interleave = obj_453_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_7_cast_fp16, current_value_9_cast_fp16, current_value_11_cast_fp16, current_value_13_cast_fp16, current_value_15_cast_fp16, current_value_17_cast_fp16, current_value_19_cast_fp16, current_value_21_cast_fp16, current_value_23_cast_fp16, current_value_25_cast_fp16, current_value_27_cast_fp16, current_value_29_cast_fp16, current_value_31_cast_fp16, current_value_33_cast_fp16, current_value_35_cast_fp16, current_value_37_cast_fp16, current_value_39_cast_fp16, current_value_41_cast_fp16, current_value_43_cast_fp16, current_value_45_cast_fp16, current_value_47_cast_fp16, current_value_49_cast_fp16, current_value_51_cast_fp16, current_value_53_cast_fp16, current_value_55_cast_fp16, current_value_57_cast_fp16, current_value_59_cast_fp16, current_value_61_cast_fp16, current_value_cast_fp16))[name = tensor("obj_453_cast_fp16")]; tensor var_8767_begin_0 = const()[name = tensor("op_8767_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_8767_end_0 = const()[name = tensor("op_8767_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_8767_end_mask_0 = const()[name = tensor("op_8767_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_8767_cast_fp16 = slice_by_index(begin = var_8767_begin_0, end = var_8767_end_0, end_mask = var_8767_end_mask_0, x = obj_111_cast_fp16)[name = tensor("op_8767_cast_fp16")]; tensor var_8770_begin_0 = const()[name = tensor("op_8770_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_8770_end_0 = const()[name = tensor("op_8770_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_8770_end_mask_0 = const()[name = tensor("op_8770_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_8770_squeeze_mask_0 = const()[name = tensor("op_8770_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_8770_cast_fp16 = slice_by_index(begin = var_8770_begin_0, end = var_8770_end_0, end_mask = var_8770_end_mask_0, squeeze_mask = var_8770_squeeze_mask_0, x = var_8767_cast_fp16)[name = tensor("op_8770_cast_fp16")]; tensor var_8785_begin_0 = const()[name = tensor("op_8785_begin_0"), val = tensor([0, 17, 0, 0])]; tensor var_8785_end_0 = const()[name = tensor("op_8785_end_0"), val = tensor([1, 18, 1, 1500])]; tensor var_8785_end_mask_0 = const()[name = tensor("op_8785_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_8785_cast_fp16 = slice_by_index(begin = var_8785_begin_0, end = var_8785_end_0, end_mask = var_8785_end_mask_0, x = obj_153_cast_fp16)[name = tensor("op_8785_cast_fp16")]; tensor var_8788_begin_0 = const()[name = tensor("op_8788_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_8788_end_0 = const()[name = tensor("op_8788_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_8788_end_mask_0 = const()[name = tensor("op_8788_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_8788_squeeze_mask_0 = const()[name = tensor("op_8788_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_8788_cast_fp16 = slice_by_index(begin = var_8788_begin_0, end = var_8788_end_0, end_mask = var_8788_end_mask_0, squeeze_mask = var_8788_squeeze_mask_0, x = var_8785_cast_fp16)[name = tensor("op_8788_cast_fp16")]; tensor var_8803_begin_0 = const()[name = tensor("op_8803_begin_0"), val = tensor([0, 18, 0, 0])]; tensor var_8803_end_0 = const()[name = tensor("op_8803_end_0"), val = tensor([1, 19, 1, 1500])]; tensor var_8803_end_mask_0 = const()[name = tensor("op_8803_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_8803_cast_fp16 = slice_by_index(begin = var_8803_begin_0, end = var_8803_end_0, end_mask = var_8803_end_mask_0, x = obj_181_cast_fp16)[name = tensor("op_8803_cast_fp16")]; tensor var_8806_begin_0 = const()[name = tensor("op_8806_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_8806_end_0 = const()[name = tensor("op_8806_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_8806_end_mask_0 = const()[name = tensor("op_8806_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_8806_squeeze_mask_0 = const()[name = tensor("op_8806_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_8806_cast_fp16 = slice_by_index(begin = var_8806_begin_0, end = var_8806_end_0, end_mask = var_8806_end_mask_0, squeeze_mask = var_8806_squeeze_mask_0, x = var_8803_cast_fp16)[name = tensor("op_8806_cast_fp16")]; tensor var_8821_begin_0 = const()[name = tensor("op_8821_begin_0"), val = tensor([0, 12, 0, 0])]; tensor var_8821_end_0 = const()[name = tensor("op_8821_end_0"), val = tensor([1, 13, 1, 1500])]; tensor var_8821_end_mask_0 = const()[name = tensor("op_8821_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_8821_cast_fp16 = slice_by_index(begin = var_8821_begin_0, end = var_8821_end_0, end_mask = var_8821_end_mask_0, x = obj_195_cast_fp16)[name = tensor("op_8821_cast_fp16")]; tensor var_8824_begin_0 = const()[name = tensor("op_8824_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_8824_end_0 = const()[name = tensor("op_8824_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_8824_end_mask_0 = const()[name = tensor("op_8824_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_8824_squeeze_mask_0 = const()[name = tensor("op_8824_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_8824_cast_fp16 = slice_by_index(begin = var_8824_begin_0, end = var_8824_end_0, end_mask = var_8824_end_mask_0, squeeze_mask = var_8824_squeeze_mask_0, x = var_8821_cast_fp16)[name = tensor("op_8824_cast_fp16")]; tensor var_8839_begin_0 = const()[name = tensor("op_8839_begin_0"), val = tensor([0, 1, 0, 0])]; tensor var_8839_end_0 = const()[name = tensor("op_8839_end_0"), val = tensor([1, 2, 1, 1500])]; tensor var_8839_end_mask_0 = const()[name = tensor("op_8839_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_8839_cast_fp16 = slice_by_index(begin = var_8839_begin_0, end = var_8839_end_0, end_mask = var_8839_end_mask_0, x = obj_237_cast_fp16)[name = tensor("op_8839_cast_fp16")]; tensor var_8842_begin_0 = const()[name = tensor("op_8842_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_8842_end_0 = const()[name = tensor("op_8842_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_8842_end_mask_0 = const()[name = tensor("op_8842_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_8842_squeeze_mask_0 = const()[name = tensor("op_8842_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_8842_cast_fp16 = slice_by_index(begin = var_8842_begin_0, end = var_8842_end_0, end_mask = var_8842_end_mask_0, squeeze_mask = var_8842_squeeze_mask_0, x = var_8839_cast_fp16)[name = tensor("op_8842_cast_fp16")]; tensor var_8857_begin_0 = const()[name = tensor("op_8857_begin_0"), val = tensor([0, 14, 0, 0])]; tensor var_8857_end_0 = const()[name = tensor("op_8857_end_0"), val = tensor([1, 15, 1, 1500])]; tensor var_8857_end_mask_0 = const()[name = tensor("op_8857_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_8857_cast_fp16 = slice_by_index(begin = var_8857_begin_0, end = var_8857_end_0, end_mask = var_8857_end_mask_0, x = obj_251_cast_fp16)[name = tensor("op_8857_cast_fp16")]; tensor var_8860_begin_0 = const()[name = tensor("op_8860_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_8860_end_0 = const()[name = tensor("op_8860_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_8860_end_mask_0 = const()[name = tensor("op_8860_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_8860_squeeze_mask_0 = const()[name = tensor("op_8860_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_8860_cast_fp16 = slice_by_index(begin = var_8860_begin_0, end = var_8860_end_0, end_mask = var_8860_end_mask_0, squeeze_mask = var_8860_squeeze_mask_0, x = var_8857_cast_fp16)[name = tensor("op_8860_cast_fp16")]; tensor var_8875_begin_0 = const()[name = tensor("op_8875_begin_0"), val = tensor([0, 11, 0, 0])]; tensor var_8875_end_0 = const()[name = tensor("op_8875_end_0"), val = tensor([1, 12, 1, 1500])]; tensor var_8875_end_mask_0 = const()[name = tensor("op_8875_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_8875_cast_fp16 = slice_by_index(begin = var_8875_begin_0, end = var_8875_end_0, end_mask = var_8875_end_mask_0, x = obj_279_cast_fp16)[name = tensor("op_8875_cast_fp16")]; tensor var_8878_begin_0 = const()[name = tensor("op_8878_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_8878_end_0 = const()[name = tensor("op_8878_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_8878_end_mask_0 = const()[name = tensor("op_8878_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_8878_squeeze_mask_0 = const()[name = tensor("op_8878_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_8878_cast_fp16 = slice_by_index(begin = var_8878_begin_0, end = var_8878_end_0, end_mask = var_8878_end_mask_0, squeeze_mask = var_8878_squeeze_mask_0, x = var_8875_cast_fp16)[name = tensor("op_8878_cast_fp16")]; tensor var_8893_begin_0 = const()[name = tensor("op_8893_begin_0"), val = tensor([0, 4, 0, 0])]; tensor var_8893_end_0 = const()[name = tensor("op_8893_end_0"), val = tensor([1, 5, 1, 1500])]; tensor var_8893_end_mask_0 = const()[name = tensor("op_8893_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_8893_cast_fp16 = slice_by_index(begin = var_8893_begin_0, end = var_8893_end_0, end_mask = var_8893_end_mask_0, x = obj_307_cast_fp16)[name = tensor("op_8893_cast_fp16")]; tensor var_8896_begin_0 = const()[name = tensor("op_8896_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_8896_end_0 = const()[name = tensor("op_8896_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_8896_end_mask_0 = const()[name = tensor("op_8896_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_8896_squeeze_mask_0 = const()[name = tensor("op_8896_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_8896_cast_fp16 = slice_by_index(begin = var_8896_begin_0, end = var_8896_end_0, end_mask = var_8896_end_mask_0, squeeze_mask = var_8896_squeeze_mask_0, x = var_8893_cast_fp16)[name = tensor("op_8896_cast_fp16")]; tensor var_8911_begin_0 = const()[name = tensor("op_8911_begin_0"), val = tensor([0, 1, 0, 0])]; tensor var_8911_end_0 = const()[name = tensor("op_8911_end_0"), val = tensor([1, 2, 1, 1500])]; tensor var_8911_end_mask_0 = const()[name = tensor("op_8911_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_8911_cast_fp16 = slice_by_index(begin = var_8911_begin_0, end = var_8911_end_0, end_mask = var_8911_end_mask_0, x = obj_349_cast_fp16)[name = tensor("op_8911_cast_fp16")]; tensor var_8914_begin_0 = const()[name = tensor("op_8914_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_8914_end_0 = const()[name = tensor("op_8914_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_8914_end_mask_0 = const()[name = tensor("op_8914_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_8914_squeeze_mask_0 = const()[name = tensor("op_8914_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_8914_cast_fp16 = slice_by_index(begin = var_8914_begin_0, end = var_8914_end_0, end_mask = var_8914_end_mask_0, squeeze_mask = var_8914_squeeze_mask_0, x = var_8911_cast_fp16)[name = tensor("op_8914_cast_fp16")]; tensor var_8929_begin_0 = const()[name = tensor("op_8929_begin_0"), val = tensor([0, 6, 0, 0])]; tensor var_8929_end_0 = const()[name = tensor("op_8929_end_0"), val = tensor([1, 7, 1, 1500])]; tensor var_8929_end_mask_0 = const()[name = tensor("op_8929_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_8929_cast_fp16 = slice_by_index(begin = var_8929_begin_0, end = var_8929_end_0, end_mask = var_8929_end_mask_0, x = obj_363_cast_fp16)[name = tensor("op_8929_cast_fp16")]; tensor var_8932_begin_0 = const()[name = tensor("op_8932_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_8932_end_0 = const()[name = tensor("op_8932_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_8932_end_mask_0 = const()[name = tensor("op_8932_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_8932_squeeze_mask_0 = const()[name = tensor("op_8932_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_8932_cast_fp16 = slice_by_index(begin = var_8932_begin_0, end = var_8932_end_0, end_mask = var_8932_end_mask_0, squeeze_mask = var_8932_squeeze_mask_0, x = var_8929_cast_fp16)[name = tensor("op_8932_cast_fp16")]; tensor var_8939 = const()[name = tensor("op_8939"), val = tensor(1)]; tensor var_8940_interleave_0 = const()[name = tensor("op_8940_interleave_0"), val = tensor(false)]; tensor var_8940_cast_fp16 = concat(axis = var_8939, interleave = var_8940_interleave_0, values = (var_8770_cast_fp16, var_8788_cast_fp16, var_8806_cast_fp16, var_8824_cast_fp16, var_8842_cast_fp16, var_8860_cast_fp16, var_8878_cast_fp16, var_8896_cast_fp16, var_8914_cast_fp16, var_8932_cast_fp16))[name = tensor("op_8940_cast_fp16")]; tensor var_8942 = const()[name = tensor("op_8942"), val = tensor([1])]; tensor var_8943 = const()[name = tensor("op_8943"), val = tensor(false)]; tensor alignment_heads_weights = reduce_mean(axes = var_8942, keep_dims = var_8943, x = var_8940_cast_fp16)[name = tensor("obj_cast_fp16")]; } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); }