diff --git "a/Llama-2-7b-hf_chunk12.mlmodelc/model.mil" "b/Llama-2-7b-hf_chunk12.mlmodelc/model.mil" --- "a/Llama-2-7b-hf_chunk12.mlmodelc/model.mil" +++ "b/Llama-2-7b-hf_chunk12.mlmodelc/model.mil" @@ -1,7 +1,7 @@ program(1.0) -[buildInfo = dict, tensor>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "2.1.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "7.2"}})] +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3304.5.2"}, {"coremlc-version", "3304.6.2"}, {"coremltools-component-torch", "2.1.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.0b1"}})] { - func main(tensor cos, tensor k_cache_0, tensor k_cache_1, tensor mask, tensor sin, tensor v_cache_0, tensor v_cache_1, tensor x) [CoreML_InputDefaultValues = dict, tensor>({{"k_cache_0", 0}, {"k_cache_1", 0}, {"v_cache_0", 0}, {"v_cache_1", 0}})] { + func main(tensor cos, tensor k_cache_0, tensor k_cache_1, tensor mask, tensor sin, tensor v_cache_0, tensor v_cache_1, tensor x) [CoreML_InputDefaultValues = dict, tensor>({{"k_cache_0", 0}, {"k_cache_1", 0}, {"v_cache_0", 0}, {"v_cache_1", 0}})] { tensor blocks_0_attn_q_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8388736))), name = tensor("blocks_0_attn_q_proj_weight_palettized_cast_fp16"), shape = tensor([4096, 4096, 1, 1])]; tensor blocks_0_attn_k_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8388864))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16777536))), name = tensor("blocks_0_attn_k_proj_weight_palettized_cast_fp16"), shape = tensor([4096, 4096, 1, 1])]; tensor blocks_0_attn_v_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16777664))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25166336))), name = tensor("blocks_0_attn_v_proj_weight_palettized_cast_fp16"), shape = tensor([4096, 4096, 1, 1])]; @@ -16,273 +16,1545 @@ program(1.0) tensor blocks_1_mlp_fc_1_weight_palettized_cast_fp16 = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134744192))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157288640))), name = tensor("blocks_1_mlp_fc_1_weight_palettized_cast_fp16"), shape = tensor([11008, 4096, 1, 1])]; tensor blocks_1_mlp_fc_2_weight_palettized_cast_fp16 = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157288768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179833216))), name = tensor("blocks_1_mlp_fc_2_weight_palettized_cast_fp16"), shape = tensor([11008, 4096, 1, 1])]; tensor blocks_1_mlp_proj_weight_palettized_cast_fp16 = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179833344))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202377792))), name = tensor("blocks_1_mlp_proj_weight_palettized_cast_fp16"), shape = tensor([4096, 11008, 1, 1])]; - tensor var_14 = const()[name = tensor("op_14"), val = tensor(3)]; - tensor var_19 = const()[name = tensor("op_19"), val = tensor(-2)]; - tensor var_21 = const()[name = tensor("op_21"), val = tensor(-1)]; - tensor var_28 = const()[name = tensor("op_28"), val = tensor(1)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; - tensor var_37_cast_fp16 = mul(x = x, y = x)[name = tensor("op_37_cast_fp16")]; - tensor var_38 = const()[name = tensor("op_38"), val = tensor([1])]; - tensor norm_x_1_cast_fp16 = reduce_mean(axes = var_38, keep_dims = var_29, x = var_37_cast_fp16)[name = tensor("norm_x_1_cast_fp16")]; - tensor var_40_to_fp16 = const()[name = tensor("op_40_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_41_cast_fp16 = add(x = norm_x_1_cast_fp16, y = var_40_to_fp16)[name = tensor("op_41_cast_fp16")]; - tensor var_42_epsilon_0_to_fp16 = const()[name = tensor("op_42_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; - tensor var_42_cast_fp16 = rsqrt(epsilon = var_42_epsilon_0_to_fp16, x = var_41_cast_fp16)[name = tensor("op_42_cast_fp16")]; - tensor x_normed_1_cast_fp16 = mul(x = x, y = var_42_cast_fp16)[name = tensor("x_normed_1_cast_fp16")]; - tensor blocks_0_norm_1_weight_to_fp16 = const()[name = tensor("blocks_0_norm_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202377920)))]; - tensor x_5_cast_fp16 = mul(x = x_normed_1_cast_fp16, y = blocks_0_norm_1_weight_to_fp16)[name = tensor("x_5_cast_fp16")]; - tensor var_54 = const()[name = tensor("op_54"), val = tensor([1, 1])]; - tensor var_56 = const()[name = tensor("op_56"), val = tensor([1, 1])]; - tensor var_58_pad_type_0 = const()[name = tensor("op_58_pad_type_0"), val = tensor("custom")]; - tensor var_58_pad_0 = const()[name = tensor("op_58_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor var_58_cast_fp16 = conv(dilations = var_56, groups = var_28, pad = var_58_pad_0, pad_type = var_58_pad_type_0, strides = var_54, weight = blocks_0_attn_q_proj_weight_palettized_cast_fp16, x = x_5_cast_fp16)[name = tensor("op_58_cast_fp16")]; - tensor blocks_0_attn_q_proj_output_scales_to_fp16 = const()[name = tensor("blocks_0_attn_q_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202386176)))]; - tensor q_1_cast_fp16 = mul(x = var_58_cast_fp16, y = blocks_0_attn_q_proj_output_scales_to_fp16)[name = tensor("q_1_cast_fp16")]; - tensor var_62 = const()[name = tensor("op_62"), val = tensor([1, 1])]; - tensor var_64 = const()[name = tensor("op_64"), val = tensor([1, 1])]; - tensor var_66_pad_type_0 = const()[name = tensor("op_66_pad_type_0"), val = tensor("custom")]; - tensor var_66_pad_0 = const()[name = tensor("op_66_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor var_66_cast_fp16 = conv(dilations = var_64, groups = var_28, pad = var_66_pad_0, pad_type = var_66_pad_type_0, strides = var_62, weight = blocks_0_attn_k_proj_weight_palettized_cast_fp16, x = x_5_cast_fp16)[name = tensor("op_66_cast_fp16")]; - tensor blocks_0_attn_k_proj_output_scales_to_fp16 = const()[name = tensor("blocks_0_attn_k_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202394432)))]; - tensor k_1_cast_fp16 = mul(x = var_66_cast_fp16, y = blocks_0_attn_k_proj_output_scales_to_fp16)[name = tensor("k_1_cast_fp16")]; - tensor var_70 = const()[name = tensor("op_70"), val = tensor([1, 1])]; - tensor var_72 = const()[name = tensor("op_72"), val = tensor([1, 1])]; - tensor var_74_pad_type_0 = const()[name = tensor("op_74_pad_type_0"), val = tensor("custom")]; - tensor var_74_pad_0 = const()[name = tensor("op_74_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor var_74_cast_fp16 = conv(dilations = var_72, groups = var_28, pad = var_74_pad_0, pad_type = var_74_pad_type_0, strides = var_70, weight = blocks_0_attn_v_proj_weight_palettized_cast_fp16, x = x_5_cast_fp16)[name = tensor("op_74_cast_fp16")]; - tensor blocks_0_attn_v_proj_output_scales_to_fp16 = const()[name = tensor("blocks_0_attn_v_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202402688)))]; - tensor v_1_cast_fp16 = mul(x = var_74_cast_fp16, y = blocks_0_attn_v_proj_output_scales_to_fp16)[name = tensor("v_1_cast_fp16")]; - tensor var_76 = const()[name = tensor("op_76"), val = tensor([1, 32, 128, 64])]; - tensor q_3_cast_fp16 = reshape(shape = var_76, x = q_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; - tensor var_78 = const()[name = tensor("op_78"), val = tensor([1, 32, 128, 64])]; - tensor k_3_cast_fp16 = reshape(shape = var_78, x = k_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; - tensor var_80 = const()[name = tensor("op_80"), val = tensor([1, 32, 128, 64])]; - tensor new_v_cache_0 = reshape(shape = var_80, x = v_1_cast_fp16)[name = tensor("v_3_cast_fp16")]; - tensor var_92_begin_0 = const()[name = tensor("op_92_begin_0"), val = tensor([0, 0, 0, 0])]; - tensor var_92_end_0 = const()[name = tensor("op_92_end_0"), val = tensor([1, 32, 64, 64])]; - tensor var_92_end_mask_0 = const()[name = tensor("op_92_end_mask_0"), val = tensor([true, true, false, true])]; - tensor var_92_cast_fp16 = slice_by_index(begin = var_92_begin_0, end = var_92_end_0, end_mask = var_92_end_mask_0, x = q_3_cast_fp16)[name = tensor("op_92_cast_fp16")]; - tensor var_98_begin_0 = const()[name = tensor("op_98_begin_0"), val = tensor([0, 0, 64, 0])]; - tensor var_98_end_0 = const()[name = tensor("op_98_end_0"), val = tensor([1, 32, 128, 64])]; - tensor var_98_end_mask_0 = const()[name = tensor("op_98_end_mask_0"), val = tensor([true, true, true, true])]; - tensor var_98_cast_fp16 = slice_by_index(begin = var_98_begin_0, end = var_98_end_0, end_mask = var_98_end_mask_0, x = q_3_cast_fp16)[name = tensor("op_98_cast_fp16")]; - tensor const_3_promoted_to_fp16 = const()[name = tensor("const_3_promoted_to_fp16"), val = tensor(-0x1p+0)]; - tensor var_100_cast_fp16 = mul(x = var_98_cast_fp16, y = const_3_promoted_to_fp16)[name = tensor("op_100_cast_fp16")]; + tensor var_13 = const()[name = tensor("op_13"), val = tensor(-1)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(-2)]; + tensor var_19 = const()[name = tensor("op_19"), val = tensor(-3)]; + tensor var_60 = const()[name = tensor("op_60"), val = tensor(1)]; + tensor var_63 = const()[name = tensor("op_63"), val = tensor(true)]; + tensor x_eps_1_interleave_0 = const()[name = tensor("x_eps_1_interleave_0"), val = tensor(false)]; + tensor eps_chan_1_to_fp16 = const()[name = tensor("eps_chan_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202377920)))]; + tensor x_eps_1_cast_fp16 = concat(axis = var_60, interleave = x_eps_1_interleave_0, values = (x, eps_chan_1_to_fp16))[name = tensor("x_eps_1_cast_fp16")]; + tensor norm_x_1_axes_0 = const()[name = tensor("norm_x_1_axes_0"), val = tensor([1])]; + tensor norm_x_1_cast_fp16 = reduce_l2_norm(axes = norm_x_1_axes_0, keep_dims = var_63, x = x_eps_1_cast_fp16)[name = tensor("norm_x_1_cast_fp16")]; + tensor x_normed_1_cast_fp16 = real_div(x = x, y = norm_x_1_cast_fp16)[name = tensor("x_normed_1_cast_fp16")]; + tensor var_87_to_fp16 = const()[name = tensor("op_87_to_fp16"), val = tensor(0x1p+6)]; + tensor x_normed_3_cast_fp16 = mul(x = x_normed_1_cast_fp16, y = var_87_to_fp16)[name = tensor("x_normed_3_cast_fp16")]; + tensor blocks_0_norm_1_weight_to_fp16 = const()[name = tensor("blocks_0_norm_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202378112)))]; + tensor x_5_cast_fp16 = mul(x = x_normed_3_cast_fp16, y = blocks_0_norm_1_weight_to_fp16)[name = tensor("x_5_cast_fp16")]; + tensor var_109 = const()[name = tensor("op_109"), val = tensor([1, 4096, 1, -1])]; + tensor input_1_cast_fp16 = reshape(shape = var_109, x = x_5_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_113 = const()[name = tensor("op_113"), val = tensor([1, 1])]; + tensor var_115 = const()[name = tensor("op_115"), val = tensor([1, 1])]; + tensor var_117_pad_type_0 = const()[name = tensor("op_117_pad_type_0"), val = tensor("custom")]; + tensor var_117_pad_0 = const()[name = tensor("op_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_117_cast_fp16 = conv(dilations = var_115, groups = var_60, pad = var_117_pad_0, pad_type = var_117_pad_type_0, strides = var_113, weight = blocks_0_attn_q_proj_weight_palettized_cast_fp16, x = input_1_cast_fp16)[name = tensor("op_117_cast_fp16")]; + tensor blocks_0_attn_q_proj_output_scales_to_fp16 = const()[name = tensor("blocks_0_attn_q_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202386368)))]; + tensor q_1_cast_fp16 = mul(x = var_117_cast_fp16, y = blocks_0_attn_q_proj_output_scales_to_fp16)[name = tensor("q_1_cast_fp16")]; + tensor var_121 = const()[name = tensor("op_121"), val = tensor([1, 1])]; + tensor var_123 = const()[name = tensor("op_123"), val = tensor([1, 1])]; + tensor var_125_pad_type_0 = const()[name = tensor("op_125_pad_type_0"), val = tensor("custom")]; + tensor var_125_pad_0 = const()[name = tensor("op_125_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_125_cast_fp16 = conv(dilations = var_123, groups = var_60, pad = var_125_pad_0, pad_type = var_125_pad_type_0, strides = var_121, weight = blocks_0_attn_k_proj_weight_palettized_cast_fp16, x = input_1_cast_fp16)[name = tensor("op_125_cast_fp16")]; + tensor blocks_0_attn_k_proj_output_scales_to_fp16 = const()[name = tensor("blocks_0_attn_k_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202394624)))]; + tensor k_1_cast_fp16 = mul(x = var_125_cast_fp16, y = blocks_0_attn_k_proj_output_scales_to_fp16)[name = tensor("k_1_cast_fp16")]; + tensor var_129 = const()[name = tensor("op_129"), val = tensor([1, 1])]; + tensor var_131 = const()[name = tensor("op_131"), val = tensor([1, 1])]; + tensor var_133_pad_type_0 = const()[name = tensor("op_133_pad_type_0"), val = tensor("custom")]; + tensor var_133_pad_0 = const()[name = tensor("op_133_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_133_cast_fp16 = conv(dilations = var_131, groups = var_60, pad = var_133_pad_0, pad_type = var_133_pad_type_0, strides = var_129, weight = blocks_0_attn_v_proj_weight_palettized_cast_fp16, x = input_1_cast_fp16)[name = tensor("op_133_cast_fp16")]; + tensor blocks_0_attn_v_proj_output_scales_to_fp16 = const()[name = tensor("blocks_0_attn_v_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202402880)))]; + tensor v_1_cast_fp16 = mul(x = var_133_cast_fp16, y = blocks_0_attn_v_proj_output_scales_to_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_135 = const()[name = tensor("op_135"), val = tensor([1, 32, 128, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_135, x = q_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_137 = const()[name = tensor("op_137"), val = tensor([1, 32, 128, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_137, x = k_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_151_begin_0 = const()[name = tensor("op_151_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_151_end_0 = const()[name = tensor("op_151_end_0"), val = tensor([1, 32, 64, 64])]; + tensor var_151_end_mask_0 = const()[name = tensor("op_151_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_151_cast_fp16 = slice_by_index(begin = var_151_begin_0, end = var_151_end_0, end_mask = var_151_end_mask_0, x = q_3_cast_fp16)[name = tensor("op_151_cast_fp16")]; + tensor var_157_begin_0 = const()[name = tensor("op_157_begin_0"), val = tensor([0, 0, 64, 0])]; + tensor var_157_end_0 = const()[name = tensor("op_157_end_0"), val = tensor([1, 32, 128, 64])]; + tensor var_157_end_mask_0 = const()[name = tensor("op_157_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_157_cast_fp16 = slice_by_index(begin = var_157_begin_0, end = var_157_end_0, end_mask = var_157_end_mask_0, x = q_3_cast_fp16)[name = tensor("op_157_cast_fp16")]; + tensor const_11_promoted_to_fp16 = const()[name = tensor("const_11_promoted_to_fp16"), val = tensor(-0x1p+0)]; + tensor var_159_cast_fp16 = mul(x = var_157_cast_fp16, y = const_11_promoted_to_fp16)[name = tensor("op_159_cast_fp16")]; tensor rotated_1_interleave_0 = const()[name = tensor("rotated_1_interleave_0"), val = tensor(false)]; - tensor rotated_1_cast_fp16 = concat(axis = var_19, interleave = rotated_1_interleave_0, values = (var_100_cast_fp16, var_92_cast_fp16))[name = tensor("rotated_1_cast_fp16")]; - tensor var_103_cast_fp16 = mul(x = q_3_cast_fp16, y = cos)[name = tensor("op_103_cast_fp16")]; - tensor var_104_cast_fp16 = mul(x = rotated_1_cast_fp16, y = sin)[name = tensor("op_104_cast_fp16")]; - tensor roped_1_cast_fp16 = add(x = var_103_cast_fp16, y = var_104_cast_fp16)[name = tensor("roped_1_cast_fp16")]; - tensor var_117_begin_0 = const()[name = tensor("op_117_begin_0"), val = tensor([0, 0, 0, 0])]; - tensor var_117_end_0 = const()[name = tensor("op_117_end_0"), val = tensor([1, 32, 64, 64])]; - tensor var_117_end_mask_0 = const()[name = tensor("op_117_end_mask_0"), val = tensor([true, true, false, true])]; - tensor var_117_cast_fp16 = slice_by_index(begin = var_117_begin_0, end = var_117_end_0, end_mask = var_117_end_mask_0, x = k_3_cast_fp16)[name = tensor("op_117_cast_fp16")]; - tensor var_123_begin_0 = const()[name = tensor("op_123_begin_0"), val = tensor([0, 0, 64, 0])]; - tensor var_123_end_0 = const()[name = tensor("op_123_end_0"), val = tensor([1, 32, 128, 64])]; - tensor var_123_end_mask_0 = const()[name = tensor("op_123_end_mask_0"), val = tensor([true, true, true, true])]; - tensor var_123_cast_fp16 = slice_by_index(begin = var_123_begin_0, end = var_123_end_0, end_mask = var_123_end_mask_0, x = k_3_cast_fp16)[name = tensor("op_123_cast_fp16")]; - tensor const_5_promoted_to_fp16 = const()[name = tensor("const_5_promoted_to_fp16"), val = tensor(-0x1p+0)]; - tensor var_125_cast_fp16 = mul(x = var_123_cast_fp16, y = const_5_promoted_to_fp16)[name = tensor("op_125_cast_fp16")]; + tensor rotated_1_cast_fp16 = concat(axis = var_17, interleave = rotated_1_interleave_0, values = (var_159_cast_fp16, var_151_cast_fp16))[name = tensor("rotated_1_cast_fp16")]; + tensor var_162_cast_fp16 = mul(x = q_3_cast_fp16, y = cos)[name = tensor("op_162_cast_fp16")]; + tensor var_163_cast_fp16 = mul(x = rotated_1_cast_fp16, y = sin)[name = tensor("op_163_cast_fp16")]; + tensor roped_1_cast_fp16 = add(x = var_162_cast_fp16, y = var_163_cast_fp16)[name = tensor("roped_1_cast_fp16")]; + tensor var_176_begin_0 = const()[name = tensor("op_176_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_176_end_0 = const()[name = tensor("op_176_end_0"), val = tensor([1, 32, 64, 64])]; + tensor var_176_end_mask_0 = const()[name = tensor("op_176_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_176_cast_fp16 = slice_by_index(begin = var_176_begin_0, end = var_176_end_0, end_mask = var_176_end_mask_0, x = k_3_cast_fp16)[name = tensor("op_176_cast_fp16")]; + tensor var_182_begin_0 = const()[name = tensor("op_182_begin_0"), val = tensor([0, 0, 64, 0])]; + tensor var_182_end_0 = const()[name = tensor("op_182_end_0"), val = tensor([1, 32, 128, 64])]; + tensor var_182_end_mask_0 = const()[name = tensor("op_182_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_182_cast_fp16 = slice_by_index(begin = var_182_begin_0, end = var_182_end_0, end_mask = var_182_end_mask_0, x = k_3_cast_fp16)[name = tensor("op_182_cast_fp16")]; + tensor const_13_promoted_to_fp16 = const()[name = tensor("const_13_promoted_to_fp16"), val = tensor(-0x1p+0)]; + tensor var_184_cast_fp16 = mul(x = var_182_cast_fp16, y = const_13_promoted_to_fp16)[name = tensor("op_184_cast_fp16")]; tensor rotated_3_interleave_0 = const()[name = tensor("rotated_3_interleave_0"), val = tensor(false)]; - tensor rotated_3_cast_fp16 = concat(axis = var_19, interleave = rotated_3_interleave_0, values = (var_125_cast_fp16, var_117_cast_fp16))[name = tensor("rotated_3_cast_fp16")]; - tensor var_128_cast_fp16 = mul(x = k_3_cast_fp16, y = cos)[name = tensor("op_128_cast_fp16")]; - tensor var_129_cast_fp16 = mul(x = rotated_3_cast_fp16, y = sin)[name = tensor("op_129_cast_fp16")]; - tensor roped_3_cast_fp16 = add(x = var_128_cast_fp16, y = var_129_cast_fp16)[name = tensor("roped_3_cast_fp16")]; - tensor q_5_interleave_0 = const()[name = tensor("q_5_interleave_0"), val = tensor(false)]; - tensor q_5_cast_fp16 = concat(axis = var_19, interleave = q_5_interleave_0, values = roped_1_cast_fp16)[name = tensor("q_5_cast_fp16")]; - tensor k_5_interleave_0 = const()[name = tensor("k_5_interleave_0"), val = tensor(false)]; - tensor new_k_cache_0 = concat(axis = var_19, interleave = k_5_interleave_0, values = roped_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; - tensor k_7_interleave_0 = const()[name = tensor("k_7_interleave_0"), val = tensor(false)]; - tensor k_7_cast_fp16 = concat(axis = var_21, interleave = k_7_interleave_0, values = (k_cache_0, new_k_cache_0))[name = tensor("k_7_cast_fp16")]; - tensor v_5_interleave_0 = const()[name = tensor("v_5_interleave_0"), val = tensor(false)]; - tensor v_5_cast_fp16 = concat(axis = var_21, interleave = v_5_interleave_0, values = (v_cache_0, new_v_cache_0))[name = tensor("v_5_cast_fp16")]; - tensor var_151_to_fp16 = const()[name = tensor("op_151_to_fp16"), val = tensor(0x1.6ap-4)]; - tensor var_152_cast_fp16 = mul(x = q_5_cast_fp16, y = var_151_to_fp16)[name = tensor("op_152_cast_fp16")]; - tensor attn_weights_1_transpose_x_0 = const()[name = tensor("attn_weights_1_transpose_x_0"), val = tensor(true)]; - tensor attn_weights_1_transpose_y_0 = const()[name = tensor("attn_weights_1_transpose_y_0"), val = tensor(false)]; - tensor attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_0, transpose_y = attn_weights_1_transpose_y_0, x = var_152_cast_fp16, y = k_7_cast_fp16)[name = tensor("attn_weights_1_cast_fp16")]; - tensor attn_weights_3_cast_fp16 = add(x = attn_weights_1_cast_fp16, y = mask)[name = tensor("attn_weights_3_cast_fp16")]; - tensor var_160_cast_fp16 = softmax(axis = var_14, x = attn_weights_3_cast_fp16)[name = tensor("op_160_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 = v_5_cast_fp16, y = var_160_cast_fp16)[name = tensor("attn_1_cast_fp16")]; - tensor var_164 = const()[name = tensor("op_164"), val = tensor([1, 4096, 1, -1])]; - tensor input_1_cast_fp16 = reshape(shape = var_164, x = attn_1_cast_fp16)[name = tensor("input_1_cast_fp16")]; - tensor var_168 = const()[name = tensor("op_168"), val = tensor([1, 1])]; - tensor var_170 = const()[name = tensor("op_170"), val = tensor([1, 1])]; - tensor var_172_pad_type_0 = const()[name = tensor("op_172_pad_type_0"), val = tensor("custom")]; - tensor var_172_pad_0 = const()[name = tensor("op_172_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor var_172_cast_fp16 = conv(dilations = var_170, groups = var_28, pad = var_172_pad_0, pad_type = var_172_pad_type_0, strides = var_168, weight = blocks_0_attn_proj_weight_palettized_cast_fp16, x = input_1_cast_fp16)[name = tensor("op_172_cast_fp16")]; - tensor blocks_0_attn_proj_output_scales_to_fp16 = const()[name = tensor("blocks_0_attn_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202410944)))]; - tensor attention_output_1_cast_fp16 = mul(x = var_172_cast_fp16, y = blocks_0_attn_proj_output_scales_to_fp16)[name = tensor("attention_output_1_cast_fp16")]; - tensor x_11_cast_fp16 = add(x = attention_output_1_cast_fp16, y = x)[name = tensor("x_11_cast_fp16")]; - tensor var_181_cast_fp16 = mul(x = x_11_cast_fp16, y = x_11_cast_fp16)[name = tensor("op_181_cast_fp16")]; - tensor var_182 = const()[name = tensor("op_182"), val = tensor([1])]; - tensor norm_x_3_cast_fp16 = reduce_mean(axes = var_182, keep_dims = var_29, x = var_181_cast_fp16)[name = tensor("norm_x_3_cast_fp16")]; - tensor var_184_to_fp16 = const()[name = tensor("op_184_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_185_cast_fp16 = add(x = norm_x_3_cast_fp16, y = var_184_to_fp16)[name = tensor("op_185_cast_fp16")]; - tensor var_186_epsilon_0_to_fp16 = const()[name = tensor("op_186_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; - tensor var_186_cast_fp16 = rsqrt(epsilon = var_186_epsilon_0_to_fp16, x = var_185_cast_fp16)[name = tensor("op_186_cast_fp16")]; - tensor x_normed_5_cast_fp16 = mul(x = x_11_cast_fp16, y = var_186_cast_fp16)[name = tensor("x_normed_5_cast_fp16")]; - tensor blocks_0_norm_2_weight_to_fp16 = const()[name = tensor("blocks_0_norm_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202419200)))]; - tensor input_3_cast_fp16 = mul(x = x_normed_5_cast_fp16, y = blocks_0_norm_2_weight_to_fp16)[name = tensor("input_3_cast_fp16")]; - tensor var_198 = const()[name = tensor("op_198"), val = tensor([1, 1])]; - tensor var_200 = const()[name = tensor("op_200"), val = tensor([1, 1])]; - tensor var_202_pad_type_0 = const()[name = tensor("op_202_pad_type_0"), val = tensor("custom")]; - tensor var_202_pad_0 = const()[name = tensor("op_202_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor var_202_cast_fp16 = conv(dilations = var_200, groups = var_28, pad = var_202_pad_0, pad_type = var_202_pad_type_0, strides = var_198, weight = blocks_0_mlp_fc_1_weight_palettized_cast_fp16, x = input_3_cast_fp16)[name = tensor("op_202_cast_fp16")]; - tensor blocks_0_mlp_fc_1_output_scales_to_fp16 = const()[name = tensor("blocks_0_mlp_fc_1_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202427456)))]; - tensor input_5_cast_fp16 = mul(x = var_202_cast_fp16, y = blocks_0_mlp_fc_1_output_scales_to_fp16)[name = tensor("input_5_cast_fp16")]; - tensor var_206 = const()[name = tensor("op_206"), val = tensor([1, 1])]; - tensor var_208 = const()[name = tensor("op_208"), val = tensor([1, 1])]; - tensor var_210_pad_type_0 = const()[name = tensor("op_210_pad_type_0"), val = tensor("custom")]; - tensor var_210_pad_0 = const()[name = tensor("op_210_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor var_210_cast_fp16 = conv(dilations = var_208, groups = var_28, pad = var_210_pad_0, pad_type = var_210_pad_type_0, strides = var_206, weight = blocks_0_mlp_fc_2_weight_palettized_cast_fp16, x = input_3_cast_fp16)[name = tensor("op_210_cast_fp16")]; - tensor blocks_0_mlp_fc_2_output_scales_to_fp16 = const()[name = tensor("blocks_0_mlp_fc_2_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202449536)))]; - tensor x_fc_2_1_cast_fp16 = mul(x = var_210_cast_fp16, y = blocks_0_mlp_fc_2_output_scales_to_fp16)[name = tensor("x_fc_2_1_cast_fp16")]; - tensor var_212_cast_fp16 = silu(x = input_5_cast_fp16)[name = tensor("op_212_cast_fp16")]; - tensor input_7_cast_fp16 = mul(x = var_212_cast_fp16, y = x_fc_2_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; - tensor var_216 = const()[name = tensor("op_216"), val = tensor([1, 1])]; - tensor var_218 = const()[name = tensor("op_218"), val = tensor([1, 1])]; - tensor var_220_pad_type_0 = const()[name = tensor("op_220_pad_type_0"), val = tensor("custom")]; - tensor var_220_pad_0 = const()[name = tensor("op_220_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor var_220_cast_fp16 = conv(dilations = var_218, groups = var_28, pad = var_220_pad_0, pad_type = var_220_pad_type_0, strides = var_216, weight = blocks_0_mlp_proj_weight_palettized_cast_fp16, x = input_7_cast_fp16)[name = tensor("op_220_cast_fp16")]; - tensor blocks_0_mlp_proj_output_scales_to_fp16 = const()[name = tensor("blocks_0_mlp_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202471616)))]; - tensor var_221_cast_fp16 = mul(x = var_220_cast_fp16, y = blocks_0_mlp_proj_output_scales_to_fp16)[name = tensor("op_221_cast_fp16")]; - tensor x_15_cast_fp16 = add(x = var_221_cast_fp16, y = x_11_cast_fp16)[name = tensor("x_15_cast_fp16")]; - tensor var_228 = const()[name = tensor("op_228"), val = tensor(3)]; - tensor var_233 = const()[name = tensor("op_233"), val = tensor(-2)]; - tensor var_235 = const()[name = tensor("op_235"), val = tensor(-1)]; - tensor var_242 = const()[name = tensor("op_242"), val = tensor(1)]; - tensor var_243 = const()[name = tensor("op_243"), val = tensor(true)]; - tensor var_250_cast_fp16 = mul(x = x_15_cast_fp16, y = x_15_cast_fp16)[name = tensor("op_250_cast_fp16")]; - tensor var_251 = const()[name = tensor("op_251"), val = tensor([1])]; - tensor norm_x_5_cast_fp16 = reduce_mean(axes = var_251, keep_dims = var_243, x = var_250_cast_fp16)[name = tensor("norm_x_5_cast_fp16")]; - tensor var_253_to_fp16 = const()[name = tensor("op_253_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_254_cast_fp16 = add(x = norm_x_5_cast_fp16, y = var_253_to_fp16)[name = tensor("op_254_cast_fp16")]; - tensor var_255_epsilon_0_to_fp16 = const()[name = tensor("op_255_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; - tensor var_255_cast_fp16 = rsqrt(epsilon = var_255_epsilon_0_to_fp16, x = var_254_cast_fp16)[name = tensor("op_255_cast_fp16")]; - tensor x_normed_9_cast_fp16 = mul(x = x_15_cast_fp16, y = var_255_cast_fp16)[name = tensor("x_normed_9_cast_fp16")]; - tensor blocks_1_norm_1_weight_to_fp16 = const()[name = tensor("blocks_1_norm_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202479872)))]; - tensor x_19_cast_fp16 = mul(x = x_normed_9_cast_fp16, y = blocks_1_norm_1_weight_to_fp16)[name = tensor("x_19_cast_fp16")]; - tensor var_270 = const()[name = tensor("op_270"), val = tensor([1, 1])]; - tensor var_272 = const()[name = tensor("op_272"), val = tensor([1, 1])]; - tensor var_274_pad_type_0 = const()[name = tensor("op_274_pad_type_0"), val = tensor("custom")]; - tensor var_274_pad_0 = const()[name = tensor("op_274_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor var_274_cast_fp16 = conv(dilations = var_272, groups = var_242, pad = var_274_pad_0, pad_type = var_274_pad_type_0, strides = var_270, weight = blocks_1_attn_q_proj_weight_palettized_cast_fp16, x = x_19_cast_fp16)[name = tensor("op_274_cast_fp16")]; - tensor blocks_1_attn_q_proj_output_scales_to_fp16 = const()[name = tensor("blocks_1_attn_q_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202488128)))]; - tensor q_7_cast_fp16 = mul(x = var_274_cast_fp16, y = blocks_1_attn_q_proj_output_scales_to_fp16)[name = tensor("q_7_cast_fp16")]; - tensor var_278 = const()[name = tensor("op_278"), val = tensor([1, 1])]; - tensor var_280 = const()[name = tensor("op_280"), val = tensor([1, 1])]; - tensor var_282_pad_type_0 = const()[name = tensor("op_282_pad_type_0"), val = tensor("custom")]; - tensor var_282_pad_0 = const()[name = tensor("op_282_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor var_282_cast_fp16 = conv(dilations = var_280, groups = var_242, pad = var_282_pad_0, pad_type = var_282_pad_type_0, strides = var_278, weight = blocks_1_attn_k_proj_weight_palettized_cast_fp16, x = x_19_cast_fp16)[name = tensor("op_282_cast_fp16")]; - tensor blocks_1_attn_k_proj_output_scales_to_fp16 = const()[name = tensor("blocks_1_attn_k_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202496384)))]; - tensor k_9_cast_fp16 = mul(x = var_282_cast_fp16, y = blocks_1_attn_k_proj_output_scales_to_fp16)[name = tensor("k_9_cast_fp16")]; - tensor var_286 = const()[name = tensor("op_286"), val = tensor([1, 1])]; - tensor var_288 = const()[name = tensor("op_288"), val = tensor([1, 1])]; - tensor var_290_pad_type_0 = const()[name = tensor("op_290_pad_type_0"), val = tensor("custom")]; - tensor var_290_pad_0 = const()[name = tensor("op_290_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor var_290_cast_fp16 = conv(dilations = var_288, groups = var_242, pad = var_290_pad_0, pad_type = var_290_pad_type_0, strides = var_286, weight = blocks_1_attn_v_proj_weight_palettized_cast_fp16, x = x_19_cast_fp16)[name = tensor("op_290_cast_fp16")]; - tensor blocks_1_attn_v_proj_output_scales_to_fp16 = const()[name = tensor("blocks_1_attn_v_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202504640)))]; - tensor v_7_cast_fp16 = mul(x = var_290_cast_fp16, y = blocks_1_attn_v_proj_output_scales_to_fp16)[name = tensor("v_7_cast_fp16")]; - tensor var_292 = const()[name = tensor("op_292"), val = tensor([1, 32, 128, 64])]; - tensor q_9_cast_fp16 = reshape(shape = var_292, x = q_7_cast_fp16)[name = tensor("q_9_cast_fp16")]; - tensor var_294 = const()[name = tensor("op_294"), val = tensor([1, 32, 128, 64])]; - tensor k_11_cast_fp16 = reshape(shape = var_294, x = k_9_cast_fp16)[name = tensor("k_11_cast_fp16")]; - tensor var_296 = const()[name = tensor("op_296"), val = tensor([1, 32, 128, 64])]; - tensor new_v_cache_1 = reshape(shape = var_296, x = v_7_cast_fp16)[name = tensor("v_9_cast_fp16")]; - tensor var_308_begin_0 = const()[name = tensor("op_308_begin_0"), val = tensor([0, 0, 0, 0])]; - tensor var_308_end_0 = const()[name = tensor("op_308_end_0"), val = tensor([1, 32, 64, 64])]; - tensor var_308_end_mask_0 = const()[name = tensor("op_308_end_mask_0"), val = tensor([true, true, false, true])]; - tensor var_308_cast_fp16 = slice_by_index(begin = var_308_begin_0, end = var_308_end_0, end_mask = var_308_end_mask_0, x = q_9_cast_fp16)[name = tensor("op_308_cast_fp16")]; - tensor var_314_begin_0 = const()[name = tensor("op_314_begin_0"), val = tensor([0, 0, 64, 0])]; - tensor var_314_end_0 = const()[name = tensor("op_314_end_0"), val = tensor([1, 32, 128, 64])]; - tensor var_314_end_mask_0 = const()[name = tensor("op_314_end_mask_0"), val = tensor([true, true, true, true])]; - tensor var_314_cast_fp16 = slice_by_index(begin = var_314_begin_0, end = var_314_end_0, end_mask = var_314_end_mask_0, x = q_9_cast_fp16)[name = tensor("op_314_cast_fp16")]; - tensor const_10_promoted_to_fp16 = const()[name = tensor("const_10_promoted_to_fp16"), val = tensor(-0x1p+0)]; - tensor var_316_cast_fp16 = mul(x = var_314_cast_fp16, y = const_10_promoted_to_fp16)[name = tensor("op_316_cast_fp16")]; + tensor rotated_3_cast_fp16 = concat(axis = var_17, interleave = rotated_3_interleave_0, values = (var_184_cast_fp16, var_176_cast_fp16))[name = tensor("rotated_3_cast_fp16")]; + tensor var_187_cast_fp16 = mul(x = k_3_cast_fp16, y = cos)[name = tensor("op_187_cast_fp16")]; + tensor var_188_cast_fp16 = mul(x = rotated_3_cast_fp16, y = sin)[name = tensor("op_188_cast_fp16")]; + tensor roped_3_cast_fp16 = add(x = var_187_cast_fp16, y = var_188_cast_fp16)[name = tensor("roped_3_cast_fp16")]; + tensor var_191 = const()[name = tensor("op_191"), val = tensor([1, 4096, 1, 64])]; + tensor var_192_cast_fp16 = reshape(shape = var_191, x = roped_3_cast_fp16)[name = tensor("op_192_cast_fp16")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, -1, 2, -3])]; + tensor var_194 = const()[name = tensor("op_194"), val = tensor([1, 4096, 1, 64])]; + tensor new_v_cache_0 = reshape(shape = var_194, x = v_1_cast_fp16)[name = tensor("new_v_cache_0_type_fp32_cast_fp16")]; + tensor k_9_interleave_0 = const()[name = tensor("k_9_interleave_0"), val = tensor(false)]; + tensor new_k_cache_0 = transpose(perm = k_7_perm_0, x = var_192_cast_fp16)[name = tensor("transpose_1")]; + tensor k_9_cast_fp16 = concat(axis = var_19, interleave = k_9_interleave_0, values = (k_cache_0, new_k_cache_0))[name = tensor("k_9_cast_fp16")]; + tensor v_7_interleave_0 = const()[name = tensor("v_7_interleave_0"), val = tensor(false)]; + tensor v_7_cast_fp16 = concat(axis = var_13, interleave = v_7_interleave_0, values = (v_cache_0, new_v_cache_0))[name = tensor("v_7_cast_fp16")]; + tensor var_201 = const()[name = tensor("op_201"), val = tensor([1, 4096, 1, -1])]; + tensor q_7_cast_fp16 = reshape(shape = var_201, x = roped_1_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor var_206_begin_0 = const()[name = tensor("op_206_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_206_end_0 = const()[name = tensor("op_206_end_0"), val = tensor([1, 128, 1, 64])]; + tensor var_206_end_mask_0 = const()[name = tensor("op_206_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_206_cast_fp16 = slice_by_index(begin = var_206_begin_0, end = var_206_end_0, end_mask = var_206_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_206_cast_fp16")]; + tensor var_210_begin_0 = const()[name = tensor("op_210_begin_0"), val = tensor([0, 128, 0, 0])]; + tensor var_210_end_0 = const()[name = tensor("op_210_end_0"), val = tensor([1, 256, 1, 64])]; + tensor var_210_end_mask_0 = const()[name = tensor("op_210_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_210_cast_fp16 = slice_by_index(begin = var_210_begin_0, end = var_210_end_0, end_mask = var_210_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_210_cast_fp16")]; + tensor var_214_begin_0 = const()[name = tensor("op_214_begin_0"), val = tensor([0, 256, 0, 0])]; + tensor var_214_end_0 = const()[name = tensor("op_214_end_0"), val = tensor([1, 384, 1, 64])]; + tensor var_214_end_mask_0 = const()[name = tensor("op_214_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_214_cast_fp16 = slice_by_index(begin = var_214_begin_0, end = var_214_end_0, end_mask = var_214_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_214_cast_fp16")]; + tensor var_218_begin_0 = const()[name = tensor("op_218_begin_0"), val = tensor([0, 384, 0, 0])]; + tensor var_218_end_0 = const()[name = tensor("op_218_end_0"), val = tensor([1, 512, 1, 64])]; + tensor var_218_end_mask_0 = const()[name = tensor("op_218_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_218_cast_fp16 = slice_by_index(begin = var_218_begin_0, end = var_218_end_0, end_mask = var_218_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_218_cast_fp16")]; + tensor var_222_begin_0 = const()[name = tensor("op_222_begin_0"), val = tensor([0, 512, 0, 0])]; + tensor var_222_end_0 = const()[name = tensor("op_222_end_0"), val = tensor([1, 640, 1, 64])]; + tensor var_222_end_mask_0 = const()[name = tensor("op_222_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_222_cast_fp16 = slice_by_index(begin = var_222_begin_0, end = var_222_end_0, end_mask = var_222_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_222_cast_fp16")]; + tensor var_226_begin_0 = const()[name = tensor("op_226_begin_0"), val = tensor([0, 640, 0, 0])]; + tensor var_226_end_0 = const()[name = tensor("op_226_end_0"), val = tensor([1, 768, 1, 64])]; + tensor var_226_end_mask_0 = const()[name = tensor("op_226_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_226_cast_fp16 = slice_by_index(begin = var_226_begin_0, end = var_226_end_0, end_mask = var_226_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_226_cast_fp16")]; + tensor var_230_begin_0 = const()[name = tensor("op_230_begin_0"), val = tensor([0, 768, 0, 0])]; + tensor var_230_end_0 = const()[name = tensor("op_230_end_0"), val = tensor([1, 896, 1, 64])]; + tensor var_230_end_mask_0 = const()[name = tensor("op_230_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_230_cast_fp16 = slice_by_index(begin = var_230_begin_0, end = var_230_end_0, end_mask = var_230_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_230_cast_fp16")]; + tensor var_234_begin_0 = const()[name = tensor("op_234_begin_0"), val = tensor([0, 896, 0, 0])]; + tensor var_234_end_0 = const()[name = tensor("op_234_end_0"), val = tensor([1, 1024, 1, 64])]; + tensor var_234_end_mask_0 = const()[name = tensor("op_234_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_234_cast_fp16 = slice_by_index(begin = var_234_begin_0, end = var_234_end_0, end_mask = var_234_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_234_cast_fp16")]; + tensor var_238_begin_0 = const()[name = tensor("op_238_begin_0"), val = tensor([0, 1024, 0, 0])]; + tensor var_238_end_0 = const()[name = tensor("op_238_end_0"), val = tensor([1, 1152, 1, 64])]; + tensor var_238_end_mask_0 = const()[name = tensor("op_238_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_238_cast_fp16 = slice_by_index(begin = var_238_begin_0, end = var_238_end_0, end_mask = var_238_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor var_242_begin_0 = const()[name = tensor("op_242_begin_0"), val = tensor([0, 1152, 0, 0])]; + tensor var_242_end_0 = const()[name = tensor("op_242_end_0"), val = tensor([1, 1280, 1, 64])]; + tensor var_242_end_mask_0 = const()[name = tensor("op_242_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_242_cast_fp16 = slice_by_index(begin = var_242_begin_0, end = var_242_end_0, end_mask = var_242_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_242_cast_fp16")]; + tensor var_246_begin_0 = const()[name = tensor("op_246_begin_0"), val = tensor([0, 1280, 0, 0])]; + tensor var_246_end_0 = const()[name = tensor("op_246_end_0"), val = tensor([1, 1408, 1, 64])]; + tensor var_246_end_mask_0 = const()[name = tensor("op_246_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_246_cast_fp16 = slice_by_index(begin = var_246_begin_0, end = var_246_end_0, end_mask = var_246_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_246_cast_fp16")]; + tensor var_250_begin_0 = const()[name = tensor("op_250_begin_0"), val = tensor([0, 1408, 0, 0])]; + tensor var_250_end_0 = const()[name = tensor("op_250_end_0"), val = tensor([1, 1536, 1, 64])]; + tensor var_250_end_mask_0 = const()[name = tensor("op_250_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_250_cast_fp16 = slice_by_index(begin = var_250_begin_0, end = var_250_end_0, end_mask = var_250_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_250_cast_fp16")]; + tensor var_254_begin_0 = const()[name = tensor("op_254_begin_0"), val = tensor([0, 1536, 0, 0])]; + tensor var_254_end_0 = const()[name = tensor("op_254_end_0"), val = tensor([1, 1664, 1, 64])]; + tensor var_254_end_mask_0 = const()[name = tensor("op_254_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_254_cast_fp16 = slice_by_index(begin = var_254_begin_0, end = var_254_end_0, end_mask = var_254_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor var_258_begin_0 = const()[name = tensor("op_258_begin_0"), val = tensor([0, 1664, 0, 0])]; + tensor var_258_end_0 = const()[name = tensor("op_258_end_0"), val = tensor([1, 1792, 1, 64])]; + tensor var_258_end_mask_0 = const()[name = tensor("op_258_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_258_cast_fp16 = slice_by_index(begin = var_258_begin_0, end = var_258_end_0, end_mask = var_258_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_258_cast_fp16")]; + tensor var_262_begin_0 = const()[name = tensor("op_262_begin_0"), val = tensor([0, 1792, 0, 0])]; + tensor var_262_end_0 = const()[name = tensor("op_262_end_0"), val = tensor([1, 1920, 1, 64])]; + tensor var_262_end_mask_0 = const()[name = tensor("op_262_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_262_cast_fp16 = slice_by_index(begin = var_262_begin_0, end = var_262_end_0, end_mask = var_262_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_262_cast_fp16")]; + tensor var_266_begin_0 = const()[name = tensor("op_266_begin_0"), val = tensor([0, 1920, 0, 0])]; + tensor var_266_end_0 = const()[name = tensor("op_266_end_0"), val = tensor([1, 2048, 1, 64])]; + tensor var_266_end_mask_0 = const()[name = tensor("op_266_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_266_cast_fp16 = slice_by_index(begin = var_266_begin_0, end = var_266_end_0, end_mask = var_266_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_266_cast_fp16")]; + tensor var_270_begin_0 = const()[name = tensor("op_270_begin_0"), val = tensor([0, 2048, 0, 0])]; + tensor var_270_end_0 = const()[name = tensor("op_270_end_0"), val = tensor([1, 2176, 1, 64])]; + tensor var_270_end_mask_0 = const()[name = tensor("op_270_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_270_cast_fp16 = slice_by_index(begin = var_270_begin_0, end = var_270_end_0, end_mask = var_270_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_270_cast_fp16")]; + tensor var_274_begin_0 = const()[name = tensor("op_274_begin_0"), val = tensor([0, 2176, 0, 0])]; + tensor var_274_end_0 = const()[name = tensor("op_274_end_0"), val = tensor([1, 2304, 1, 64])]; + tensor var_274_end_mask_0 = const()[name = tensor("op_274_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_274_cast_fp16 = slice_by_index(begin = var_274_begin_0, end = var_274_end_0, end_mask = var_274_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_274_cast_fp16")]; + tensor var_278_begin_0 = const()[name = tensor("op_278_begin_0"), val = tensor([0, 2304, 0, 0])]; + tensor var_278_end_0 = const()[name = tensor("op_278_end_0"), val = tensor([1, 2432, 1, 64])]; + tensor var_278_end_mask_0 = const()[name = tensor("op_278_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_278_cast_fp16 = slice_by_index(begin = var_278_begin_0, end = var_278_end_0, end_mask = var_278_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_278_cast_fp16")]; + tensor var_282_begin_0 = const()[name = tensor("op_282_begin_0"), val = tensor([0, 2432, 0, 0])]; + tensor var_282_end_0 = const()[name = tensor("op_282_end_0"), val = tensor([1, 2560, 1, 64])]; + tensor var_282_end_mask_0 = const()[name = tensor("op_282_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_282_cast_fp16 = slice_by_index(begin = var_282_begin_0, end = var_282_end_0, end_mask = var_282_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_282_cast_fp16")]; + tensor var_286_begin_0 = const()[name = tensor("op_286_begin_0"), val = tensor([0, 2560, 0, 0])]; + tensor var_286_end_0 = const()[name = tensor("op_286_end_0"), val = tensor([1, 2688, 1, 64])]; + tensor var_286_end_mask_0 = const()[name = tensor("op_286_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_286_cast_fp16 = slice_by_index(begin = var_286_begin_0, end = var_286_end_0, end_mask = var_286_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_286_cast_fp16")]; + tensor var_290_begin_0 = const()[name = tensor("op_290_begin_0"), val = tensor([0, 2688, 0, 0])]; + tensor var_290_end_0 = const()[name = tensor("op_290_end_0"), val = tensor([1, 2816, 1, 64])]; + tensor var_290_end_mask_0 = const()[name = tensor("op_290_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_290_cast_fp16 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_290_cast_fp16")]; + tensor var_294_begin_0 = const()[name = tensor("op_294_begin_0"), val = tensor([0, 2816, 0, 0])]; + tensor var_294_end_0 = const()[name = tensor("op_294_end_0"), val = tensor([1, 2944, 1, 64])]; + tensor var_294_end_mask_0 = const()[name = tensor("op_294_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_294_cast_fp16 = slice_by_index(begin = var_294_begin_0, end = var_294_end_0, end_mask = var_294_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_294_cast_fp16")]; + tensor var_298_begin_0 = const()[name = tensor("op_298_begin_0"), val = tensor([0, 2944, 0, 0])]; + tensor var_298_end_0 = const()[name = tensor("op_298_end_0"), val = tensor([1, 3072, 1, 64])]; + tensor var_298_end_mask_0 = const()[name = tensor("op_298_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_298_cast_fp16 = slice_by_index(begin = var_298_begin_0, end = var_298_end_0, end_mask = var_298_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor var_302_begin_0 = const()[name = tensor("op_302_begin_0"), val = tensor([0, 3072, 0, 0])]; + tensor var_302_end_0 = const()[name = tensor("op_302_end_0"), val = tensor([1, 3200, 1, 64])]; + tensor var_302_end_mask_0 = const()[name = tensor("op_302_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_302_cast_fp16 = slice_by_index(begin = var_302_begin_0, end = var_302_end_0, end_mask = var_302_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_302_cast_fp16")]; + tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([0, 3200, 0, 0])]; + tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([1, 3328, 1, 64])]; + tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_306_cast_fp16 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_306_cast_fp16")]; + tensor var_310_begin_0 = const()[name = tensor("op_310_begin_0"), val = tensor([0, 3328, 0, 0])]; + tensor var_310_end_0 = const()[name = tensor("op_310_end_0"), val = tensor([1, 3456, 1, 64])]; + tensor var_310_end_mask_0 = const()[name = tensor("op_310_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_310_cast_fp16 = slice_by_index(begin = var_310_begin_0, end = var_310_end_0, end_mask = var_310_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_310_cast_fp16")]; + tensor var_314_begin_0 = const()[name = tensor("op_314_begin_0"), val = tensor([0, 3456, 0, 0])]; + tensor var_314_end_0 = const()[name = tensor("op_314_end_0"), val = tensor([1, 3584, 1, 64])]; + tensor var_314_end_mask_0 = const()[name = tensor("op_314_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_314_cast_fp16 = slice_by_index(begin = var_314_begin_0, end = var_314_end_0, end_mask = var_314_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_314_cast_fp16")]; + tensor var_318_begin_0 = const()[name = tensor("op_318_begin_0"), val = tensor([0, 3584, 0, 0])]; + tensor var_318_end_0 = const()[name = tensor("op_318_end_0"), val = tensor([1, 3712, 1, 64])]; + tensor var_318_end_mask_0 = const()[name = tensor("op_318_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_318_cast_fp16 = slice_by_index(begin = var_318_begin_0, end = var_318_end_0, end_mask = var_318_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_318_cast_fp16")]; + tensor var_322_begin_0 = const()[name = tensor("op_322_begin_0"), val = tensor([0, 3712, 0, 0])]; + tensor var_322_end_0 = const()[name = tensor("op_322_end_0"), val = tensor([1, 3840, 1, 64])]; + tensor var_322_end_mask_0 = const()[name = tensor("op_322_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_322_cast_fp16 = slice_by_index(begin = var_322_begin_0, end = var_322_end_0, end_mask = var_322_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_322_cast_fp16")]; + tensor var_326_begin_0 = const()[name = tensor("op_326_begin_0"), val = tensor([0, 3840, 0, 0])]; + tensor var_326_end_0 = const()[name = tensor("op_326_end_0"), val = tensor([1, 3968, 1, 64])]; + tensor var_326_end_mask_0 = const()[name = tensor("op_326_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_326_cast_fp16 = slice_by_index(begin = var_326_begin_0, end = var_326_end_0, end_mask = var_326_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_326_cast_fp16")]; + tensor var_330_begin_0 = const()[name = tensor("op_330_begin_0"), val = tensor([0, 3968, 0, 0])]; + tensor var_330_end_0 = const()[name = tensor("op_330_end_0"), val = tensor([1, 4096, 1, 64])]; + tensor var_330_end_mask_0 = const()[name = tensor("op_330_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_330_cast_fp16 = slice_by_index(begin = var_330_begin_0, end = var_330_end_0, end_mask = var_330_end_mask_0, x = q_7_cast_fp16)[name = tensor("op_330_cast_fp16")]; + tensor var_336_begin_0 = const()[name = tensor("op_336_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_336_end_0 = const()[name = tensor("op_336_end_0"), val = tensor([1, 512, 1, 128])]; + tensor var_336_end_mask_0 = const()[name = tensor("op_336_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_336_cast_fp16 = slice_by_index(begin = var_336_begin_0, end = var_336_end_0, end_mask = var_336_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_336_cast_fp16")]; + tensor var_340_begin_0 = const()[name = tensor("op_340_begin_0"), val = tensor([0, 0, 0, 128])]; + tensor var_340_end_0 = const()[name = tensor("op_340_end_0"), val = tensor([1, 512, 1, 256])]; + tensor var_340_end_mask_0 = const()[name = tensor("op_340_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_340_cast_fp16 = slice_by_index(begin = var_340_begin_0, end = var_340_end_0, end_mask = var_340_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor var_344_begin_0 = const()[name = tensor("op_344_begin_0"), val = tensor([0, 0, 0, 256])]; + tensor var_344_end_0 = const()[name = tensor("op_344_end_0"), val = tensor([1, 512, 1, 384])]; + tensor var_344_end_mask_0 = const()[name = tensor("op_344_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_344_cast_fp16 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_344_cast_fp16")]; + tensor var_348_begin_0 = const()[name = tensor("op_348_begin_0"), val = tensor([0, 0, 0, 384])]; + tensor var_348_end_0 = const()[name = tensor("op_348_end_0"), val = tensor([1, 512, 1, 512])]; + tensor var_348_end_mask_0 = const()[name = tensor("op_348_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_348_cast_fp16 = slice_by_index(begin = var_348_begin_0, end = var_348_end_0, end_mask = var_348_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_352_begin_0 = const()[name = tensor("op_352_begin_0"), val = tensor([0, 0, 0, 512])]; + tensor var_352_end_0 = const()[name = tensor("op_352_end_0"), val = tensor([1, 512, 1, 640])]; + tensor var_352_end_mask_0 = const()[name = tensor("op_352_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_352_cast_fp16 = slice_by_index(begin = var_352_begin_0, end = var_352_end_0, end_mask = var_352_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_352_cast_fp16")]; + tensor var_356_begin_0 = const()[name = tensor("op_356_begin_0"), val = tensor([0, 0, 0, 640])]; + tensor var_356_end_0 = const()[name = tensor("op_356_end_0"), val = tensor([1, 512, 1, 768])]; + tensor var_356_end_mask_0 = const()[name = tensor("op_356_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_356_cast_fp16 = slice_by_index(begin = var_356_begin_0, end = var_356_end_0, end_mask = var_356_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_356_cast_fp16")]; + tensor var_360_begin_0 = const()[name = tensor("op_360_begin_0"), val = tensor([0, 0, 0, 768])]; + tensor var_360_end_0 = const()[name = tensor("op_360_end_0"), val = tensor([1, 512, 1, 896])]; + tensor var_360_end_mask_0 = const()[name = tensor("op_360_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_360_cast_fp16 = slice_by_index(begin = var_360_begin_0, end = var_360_end_0, end_mask = var_360_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_360_cast_fp16")]; + tensor var_364_begin_0 = const()[name = tensor("op_364_begin_0"), val = tensor([0, 0, 0, 896])]; + tensor var_364_end_0 = const()[name = tensor("op_364_end_0"), val = tensor([1, 512, 1, 1024])]; + tensor var_364_end_mask_0 = const()[name = tensor("op_364_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_364_cast_fp16 = slice_by_index(begin = var_364_begin_0, end = var_364_end_0, end_mask = var_364_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_364_cast_fp16")]; + tensor var_368_begin_0 = const()[name = tensor("op_368_begin_0"), val = tensor([0, 0, 0, 1024])]; + tensor var_368_end_0 = const()[name = tensor("op_368_end_0"), val = tensor([1, 512, 1, 1152])]; + tensor var_368_end_mask_0 = const()[name = tensor("op_368_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_368_cast_fp16 = slice_by_index(begin = var_368_begin_0, end = var_368_end_0, end_mask = var_368_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_368_cast_fp16")]; + tensor var_372_begin_0 = const()[name = tensor("op_372_begin_0"), val = tensor([0, 0, 0, 1152])]; + tensor var_372_end_0 = const()[name = tensor("op_372_end_0"), val = tensor([1, 512, 1, 1280])]; + tensor var_372_end_mask_0 = const()[name = tensor("op_372_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_372_cast_fp16 = slice_by_index(begin = var_372_begin_0, end = var_372_end_0, end_mask = var_372_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_372_cast_fp16")]; + tensor var_376_begin_0 = const()[name = tensor("op_376_begin_0"), val = tensor([0, 0, 0, 1280])]; + tensor var_376_end_0 = const()[name = tensor("op_376_end_0"), val = tensor([1, 512, 1, 1408])]; + tensor var_376_end_mask_0 = const()[name = tensor("op_376_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_376_cast_fp16 = slice_by_index(begin = var_376_begin_0, end = var_376_end_0, end_mask = var_376_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_376_cast_fp16")]; + tensor var_380_begin_0 = const()[name = tensor("op_380_begin_0"), val = tensor([0, 0, 0, 1408])]; + tensor var_380_end_0 = const()[name = tensor("op_380_end_0"), val = tensor([1, 512, 1, 1536])]; + tensor var_380_end_mask_0 = const()[name = tensor("op_380_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_380_cast_fp16 = slice_by_index(begin = var_380_begin_0, end = var_380_end_0, end_mask = var_380_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_380_cast_fp16")]; + tensor var_384_begin_0 = const()[name = tensor("op_384_begin_0"), val = tensor([0, 0, 0, 1536])]; + tensor var_384_end_0 = const()[name = tensor("op_384_end_0"), val = tensor([1, 512, 1, 1664])]; + tensor var_384_end_mask_0 = const()[name = tensor("op_384_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_384_cast_fp16 = slice_by_index(begin = var_384_begin_0, end = var_384_end_0, end_mask = var_384_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor var_388_begin_0 = const()[name = tensor("op_388_begin_0"), val = tensor([0, 0, 0, 1664])]; + tensor var_388_end_0 = const()[name = tensor("op_388_end_0"), val = tensor([1, 512, 1, 1792])]; + tensor var_388_end_mask_0 = const()[name = tensor("op_388_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_388_cast_fp16 = slice_by_index(begin = var_388_begin_0, end = var_388_end_0, end_mask = var_388_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_388_cast_fp16")]; + tensor var_392_begin_0 = const()[name = tensor("op_392_begin_0"), val = tensor([0, 0, 0, 1792])]; + tensor var_392_end_0 = const()[name = tensor("op_392_end_0"), val = tensor([1, 512, 1, 1920])]; + tensor var_392_end_mask_0 = const()[name = tensor("op_392_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_392_cast_fp16 = slice_by_index(begin = var_392_begin_0, end = var_392_end_0, end_mask = var_392_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_392_cast_fp16")]; + tensor var_396_begin_0 = const()[name = tensor("op_396_begin_0"), val = tensor([0, 0, 0, 1920])]; + tensor var_396_end_0 = const()[name = tensor("op_396_end_0"), val = tensor([1, 512, 1, 2048])]; + tensor var_396_end_mask_0 = const()[name = tensor("op_396_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_396_cast_fp16 = slice_by_index(begin = var_396_begin_0, end = var_396_end_0, end_mask = var_396_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_396_cast_fp16")]; + tensor var_400_begin_0 = const()[name = tensor("op_400_begin_0"), val = tensor([0, 0, 0, 2048])]; + tensor var_400_end_0 = const()[name = tensor("op_400_end_0"), val = tensor([1, 512, 1, 2176])]; + tensor var_400_end_mask_0 = const()[name = tensor("op_400_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_400_cast_fp16 = slice_by_index(begin = var_400_begin_0, end = var_400_end_0, end_mask = var_400_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_400_cast_fp16")]; + tensor var_404_begin_0 = const()[name = tensor("op_404_begin_0"), val = tensor([0, 0, 0, 2176])]; + tensor var_404_end_0 = const()[name = tensor("op_404_end_0"), val = tensor([1, 512, 1, 2304])]; + tensor var_404_end_mask_0 = const()[name = tensor("op_404_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_404_cast_fp16 = slice_by_index(begin = var_404_begin_0, end = var_404_end_0, end_mask = var_404_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_404_cast_fp16")]; + tensor var_408_begin_0 = const()[name = tensor("op_408_begin_0"), val = tensor([0, 0, 0, 2304])]; + tensor var_408_end_0 = const()[name = tensor("op_408_end_0"), val = tensor([1, 512, 1, 2432])]; + tensor var_408_end_mask_0 = const()[name = tensor("op_408_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_408_cast_fp16 = slice_by_index(begin = var_408_begin_0, end = var_408_end_0, end_mask = var_408_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_408_cast_fp16")]; + tensor var_412_begin_0 = const()[name = tensor("op_412_begin_0"), val = tensor([0, 0, 0, 2432])]; + tensor var_412_end_0 = const()[name = tensor("op_412_end_0"), val = tensor([1, 512, 1, 2560])]; + tensor var_412_end_mask_0 = const()[name = tensor("op_412_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_412_cast_fp16 = slice_by_index(begin = var_412_begin_0, end = var_412_end_0, end_mask = var_412_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_412_cast_fp16")]; + tensor var_416_begin_0 = const()[name = tensor("op_416_begin_0"), val = tensor([0, 0, 0, 2560])]; + tensor var_416_end_0 = const()[name = tensor("op_416_end_0"), val = tensor([1, 512, 1, 2688])]; + tensor var_416_end_mask_0 = const()[name = tensor("op_416_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_416_cast_fp16 = slice_by_index(begin = var_416_begin_0, end = var_416_end_0, end_mask = var_416_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_416_cast_fp16")]; + tensor var_420_begin_0 = const()[name = tensor("op_420_begin_0"), val = tensor([0, 0, 0, 2688])]; + tensor var_420_end_0 = const()[name = tensor("op_420_end_0"), val = tensor([1, 512, 1, 2816])]; + tensor var_420_end_mask_0 = const()[name = tensor("op_420_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_420_cast_fp16 = slice_by_index(begin = var_420_begin_0, end = var_420_end_0, end_mask = var_420_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_420_cast_fp16")]; + tensor var_424_begin_0 = const()[name = tensor("op_424_begin_0"), val = tensor([0, 0, 0, 2816])]; + tensor var_424_end_0 = const()[name = tensor("op_424_end_0"), val = tensor([1, 512, 1, 2944])]; + tensor var_424_end_mask_0 = const()[name = tensor("op_424_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_424_cast_fp16 = slice_by_index(begin = var_424_begin_0, end = var_424_end_0, end_mask = var_424_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_424_cast_fp16")]; + tensor var_428_begin_0 = const()[name = tensor("op_428_begin_0"), val = tensor([0, 0, 0, 2944])]; + tensor var_428_end_0 = const()[name = tensor("op_428_end_0"), val = tensor([1, 512, 1, 3072])]; + tensor var_428_end_mask_0 = const()[name = tensor("op_428_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_428_cast_fp16 = slice_by_index(begin = var_428_begin_0, end = var_428_end_0, end_mask = var_428_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_428_cast_fp16")]; + tensor var_432_begin_0 = const()[name = tensor("op_432_begin_0"), val = tensor([0, 0, 0, 3072])]; + tensor var_432_end_0 = const()[name = tensor("op_432_end_0"), val = tensor([1, 512, 1, 3200])]; + tensor var_432_end_mask_0 = const()[name = tensor("op_432_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_432_cast_fp16 = slice_by_index(begin = var_432_begin_0, end = var_432_end_0, end_mask = var_432_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_432_cast_fp16")]; + tensor var_436_begin_0 = const()[name = tensor("op_436_begin_0"), val = tensor([0, 0, 0, 3200])]; + tensor var_436_end_0 = const()[name = tensor("op_436_end_0"), val = tensor([1, 512, 1, 3328])]; + tensor var_436_end_mask_0 = const()[name = tensor("op_436_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_436_cast_fp16 = slice_by_index(begin = var_436_begin_0, end = var_436_end_0, end_mask = var_436_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_436_cast_fp16")]; + tensor var_440_begin_0 = const()[name = tensor("op_440_begin_0"), val = tensor([0, 0, 0, 3328])]; + tensor var_440_end_0 = const()[name = tensor("op_440_end_0"), val = tensor([1, 512, 1, 3456])]; + tensor var_440_end_mask_0 = const()[name = tensor("op_440_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_440_cast_fp16 = slice_by_index(begin = var_440_begin_0, end = var_440_end_0, end_mask = var_440_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_440_cast_fp16")]; + tensor var_444_begin_0 = const()[name = tensor("op_444_begin_0"), val = tensor([0, 0, 0, 3456])]; + tensor var_444_end_0 = const()[name = tensor("op_444_end_0"), val = tensor([1, 512, 1, 3584])]; + tensor var_444_end_mask_0 = const()[name = tensor("op_444_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_444_cast_fp16 = slice_by_index(begin = var_444_begin_0, end = var_444_end_0, end_mask = var_444_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_444_cast_fp16")]; + tensor var_448_begin_0 = const()[name = tensor("op_448_begin_0"), val = tensor([0, 0, 0, 3584])]; + tensor var_448_end_0 = const()[name = tensor("op_448_end_0"), val = tensor([1, 512, 1, 3712])]; + tensor var_448_end_mask_0 = const()[name = tensor("op_448_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_448_cast_fp16 = slice_by_index(begin = var_448_begin_0, end = var_448_end_0, end_mask = var_448_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_448_cast_fp16")]; + tensor var_452_begin_0 = const()[name = tensor("op_452_begin_0"), val = tensor([0, 0, 0, 3712])]; + tensor var_452_end_0 = const()[name = tensor("op_452_end_0"), val = tensor([1, 512, 1, 3840])]; + tensor var_452_end_mask_0 = const()[name = tensor("op_452_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_452_cast_fp16 = slice_by_index(begin = var_452_begin_0, end = var_452_end_0, end_mask = var_452_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_452_cast_fp16")]; + tensor var_456_begin_0 = const()[name = tensor("op_456_begin_0"), val = tensor([0, 0, 0, 3840])]; + tensor var_456_end_0 = const()[name = tensor("op_456_end_0"), val = tensor([1, 512, 1, 3968])]; + tensor var_456_end_mask_0 = const()[name = tensor("op_456_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_456_cast_fp16 = slice_by_index(begin = var_456_begin_0, end = var_456_end_0, end_mask = var_456_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_456_cast_fp16")]; + tensor var_460_begin_0 = const()[name = tensor("op_460_begin_0"), val = tensor([0, 0, 0, 3968])]; + tensor var_460_end_0 = const()[name = tensor("op_460_end_0"), val = tensor([1, 512, 1, 4096])]; + tensor var_460_end_mask_0 = const()[name = tensor("op_460_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_460_cast_fp16 = slice_by_index(begin = var_460_begin_0, end = var_460_end_0, end_mask = var_460_end_mask_0, x = k_9_cast_fp16)[name = tensor("op_460_cast_fp16")]; + tensor var_462_begin_0 = const()[name = tensor("op_462_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_462_end_0 = const()[name = tensor("op_462_end_0"), val = tensor([1, 128, 1, 512])]; + tensor var_462_end_mask_0 = const()[name = tensor("op_462_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_462_cast_fp16 = slice_by_index(begin = var_462_begin_0, end = var_462_end_0, end_mask = var_462_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_462_cast_fp16")]; + tensor var_466_begin_0 = const()[name = tensor("op_466_begin_0"), val = tensor([0, 128, 0, 0])]; + tensor var_466_end_0 = const()[name = tensor("op_466_end_0"), val = tensor([1, 256, 1, 512])]; + tensor var_466_end_mask_0 = const()[name = tensor("op_466_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_466_cast_fp16 = slice_by_index(begin = var_466_begin_0, end = var_466_end_0, end_mask = var_466_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_466_cast_fp16")]; + tensor var_470_begin_0 = const()[name = tensor("op_470_begin_0"), val = tensor([0, 256, 0, 0])]; + tensor var_470_end_0 = const()[name = tensor("op_470_end_0"), val = tensor([1, 384, 1, 512])]; + tensor var_470_end_mask_0 = const()[name = tensor("op_470_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_470_cast_fp16 = slice_by_index(begin = var_470_begin_0, end = var_470_end_0, end_mask = var_470_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_470_cast_fp16")]; + tensor var_474_begin_0 = const()[name = tensor("op_474_begin_0"), val = tensor([0, 384, 0, 0])]; + tensor var_474_end_0 = const()[name = tensor("op_474_end_0"), val = tensor([1, 512, 1, 512])]; + tensor var_474_end_mask_0 = const()[name = tensor("op_474_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_474_cast_fp16 = slice_by_index(begin = var_474_begin_0, end = var_474_end_0, end_mask = var_474_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_474_cast_fp16")]; + tensor var_478_begin_0 = const()[name = tensor("op_478_begin_0"), val = tensor([0, 512, 0, 0])]; + tensor var_478_end_0 = const()[name = tensor("op_478_end_0"), val = tensor([1, 640, 1, 512])]; + tensor var_478_end_mask_0 = const()[name = tensor("op_478_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_478_cast_fp16 = slice_by_index(begin = var_478_begin_0, end = var_478_end_0, end_mask = var_478_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_478_cast_fp16")]; + tensor var_482_begin_0 = const()[name = tensor("op_482_begin_0"), val = tensor([0, 640, 0, 0])]; + tensor var_482_end_0 = const()[name = tensor("op_482_end_0"), val = tensor([1, 768, 1, 512])]; + tensor var_482_end_mask_0 = const()[name = tensor("op_482_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_482_cast_fp16 = slice_by_index(begin = var_482_begin_0, end = var_482_end_0, end_mask = var_482_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_482_cast_fp16")]; + tensor var_486_begin_0 = const()[name = tensor("op_486_begin_0"), val = tensor([0, 768, 0, 0])]; + tensor var_486_end_0 = const()[name = tensor("op_486_end_0"), val = tensor([1, 896, 1, 512])]; + tensor var_486_end_mask_0 = const()[name = tensor("op_486_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_486_cast_fp16 = slice_by_index(begin = var_486_begin_0, end = var_486_end_0, end_mask = var_486_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_486_cast_fp16")]; + tensor var_490_begin_0 = const()[name = tensor("op_490_begin_0"), val = tensor([0, 896, 0, 0])]; + tensor var_490_end_0 = const()[name = tensor("op_490_end_0"), val = tensor([1, 1024, 1, 512])]; + tensor var_490_end_mask_0 = const()[name = tensor("op_490_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_490_cast_fp16 = slice_by_index(begin = var_490_begin_0, end = var_490_end_0, end_mask = var_490_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_490_cast_fp16")]; + tensor var_494_begin_0 = const()[name = tensor("op_494_begin_0"), val = tensor([0, 1024, 0, 0])]; + tensor var_494_end_0 = const()[name = tensor("op_494_end_0"), val = tensor([1, 1152, 1, 512])]; + tensor var_494_end_mask_0 = const()[name = tensor("op_494_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_494_cast_fp16 = slice_by_index(begin = var_494_begin_0, end = var_494_end_0, end_mask = var_494_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_494_cast_fp16")]; + tensor var_498_begin_0 = const()[name = tensor("op_498_begin_0"), val = tensor([0, 1152, 0, 0])]; + tensor var_498_end_0 = const()[name = tensor("op_498_end_0"), val = tensor([1, 1280, 1, 512])]; + tensor var_498_end_mask_0 = const()[name = tensor("op_498_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_498_cast_fp16 = slice_by_index(begin = var_498_begin_0, end = var_498_end_0, end_mask = var_498_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_498_cast_fp16")]; + tensor var_502_begin_0 = const()[name = tensor("op_502_begin_0"), val = tensor([0, 1280, 0, 0])]; + tensor var_502_end_0 = const()[name = tensor("op_502_end_0"), val = tensor([1, 1408, 1, 512])]; + tensor var_502_end_mask_0 = const()[name = tensor("op_502_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_502_cast_fp16 = slice_by_index(begin = var_502_begin_0, end = var_502_end_0, end_mask = var_502_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_502_cast_fp16")]; + tensor var_506_begin_0 = const()[name = tensor("op_506_begin_0"), val = tensor([0, 1408, 0, 0])]; + tensor var_506_end_0 = const()[name = tensor("op_506_end_0"), val = tensor([1, 1536, 1, 512])]; + tensor var_506_end_mask_0 = const()[name = tensor("op_506_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_506_cast_fp16 = slice_by_index(begin = var_506_begin_0, end = var_506_end_0, end_mask = var_506_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor var_510_begin_0 = const()[name = tensor("op_510_begin_0"), val = tensor([0, 1536, 0, 0])]; + tensor var_510_end_0 = const()[name = tensor("op_510_end_0"), val = tensor([1, 1664, 1, 512])]; + tensor var_510_end_mask_0 = const()[name = tensor("op_510_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_510_cast_fp16 = slice_by_index(begin = var_510_begin_0, end = var_510_end_0, end_mask = var_510_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_510_cast_fp16")]; + tensor var_514_begin_0 = const()[name = tensor("op_514_begin_0"), val = tensor([0, 1664, 0, 0])]; + tensor var_514_end_0 = const()[name = tensor("op_514_end_0"), val = tensor([1, 1792, 1, 512])]; + tensor var_514_end_mask_0 = const()[name = tensor("op_514_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_514_cast_fp16 = slice_by_index(begin = var_514_begin_0, end = var_514_end_0, end_mask = var_514_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_514_cast_fp16")]; + tensor var_518_begin_0 = const()[name = tensor("op_518_begin_0"), val = tensor([0, 1792, 0, 0])]; + tensor var_518_end_0 = const()[name = tensor("op_518_end_0"), val = tensor([1, 1920, 1, 512])]; + tensor var_518_end_mask_0 = const()[name = tensor("op_518_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_518_cast_fp16 = slice_by_index(begin = var_518_begin_0, end = var_518_end_0, end_mask = var_518_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_518_cast_fp16")]; + tensor var_522_begin_0 = const()[name = tensor("op_522_begin_0"), val = tensor([0, 1920, 0, 0])]; + tensor var_522_end_0 = const()[name = tensor("op_522_end_0"), val = tensor([1, 2048, 1, 512])]; + tensor var_522_end_mask_0 = const()[name = tensor("op_522_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_522_cast_fp16 = slice_by_index(begin = var_522_begin_0, end = var_522_end_0, end_mask = var_522_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_522_cast_fp16")]; + tensor var_526_begin_0 = const()[name = tensor("op_526_begin_0"), val = tensor([0, 2048, 0, 0])]; + tensor var_526_end_0 = const()[name = tensor("op_526_end_0"), val = tensor([1, 2176, 1, 512])]; + tensor var_526_end_mask_0 = const()[name = tensor("op_526_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_526_cast_fp16 = slice_by_index(begin = var_526_begin_0, end = var_526_end_0, end_mask = var_526_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_526_cast_fp16")]; + tensor var_530_begin_0 = const()[name = tensor("op_530_begin_0"), val = tensor([0, 2176, 0, 0])]; + tensor var_530_end_0 = const()[name = tensor("op_530_end_0"), val = tensor([1, 2304, 1, 512])]; + tensor var_530_end_mask_0 = const()[name = tensor("op_530_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_530_cast_fp16 = slice_by_index(begin = var_530_begin_0, end = var_530_end_0, end_mask = var_530_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_530_cast_fp16")]; + tensor var_534_begin_0 = const()[name = tensor("op_534_begin_0"), val = tensor([0, 2304, 0, 0])]; + tensor var_534_end_0 = const()[name = tensor("op_534_end_0"), val = tensor([1, 2432, 1, 512])]; + tensor var_534_end_mask_0 = const()[name = tensor("op_534_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_534_cast_fp16 = slice_by_index(begin = var_534_begin_0, end = var_534_end_0, end_mask = var_534_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_534_cast_fp16")]; + tensor var_538_begin_0 = const()[name = tensor("op_538_begin_0"), val = tensor([0, 2432, 0, 0])]; + tensor var_538_end_0 = const()[name = tensor("op_538_end_0"), val = tensor([1, 2560, 1, 512])]; + tensor var_538_end_mask_0 = const()[name = tensor("op_538_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_538_cast_fp16 = slice_by_index(begin = var_538_begin_0, end = var_538_end_0, end_mask = var_538_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_538_cast_fp16")]; + tensor var_542_begin_0 = const()[name = tensor("op_542_begin_0"), val = tensor([0, 2560, 0, 0])]; + tensor var_542_end_0 = const()[name = tensor("op_542_end_0"), val = tensor([1, 2688, 1, 512])]; + tensor var_542_end_mask_0 = const()[name = tensor("op_542_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_542_cast_fp16 = slice_by_index(begin = var_542_begin_0, end = var_542_end_0, end_mask = var_542_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_542_cast_fp16")]; + tensor var_546_begin_0 = const()[name = tensor("op_546_begin_0"), val = tensor([0, 2688, 0, 0])]; + tensor var_546_end_0 = const()[name = tensor("op_546_end_0"), val = tensor([1, 2816, 1, 512])]; + tensor var_546_end_mask_0 = const()[name = tensor("op_546_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_546_cast_fp16 = slice_by_index(begin = var_546_begin_0, end = var_546_end_0, end_mask = var_546_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_546_cast_fp16")]; + tensor var_550_begin_0 = const()[name = tensor("op_550_begin_0"), val = tensor([0, 2816, 0, 0])]; + tensor var_550_end_0 = const()[name = tensor("op_550_end_0"), val = tensor([1, 2944, 1, 512])]; + tensor var_550_end_mask_0 = const()[name = tensor("op_550_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_550_cast_fp16 = slice_by_index(begin = var_550_begin_0, end = var_550_end_0, end_mask = var_550_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_550_cast_fp16")]; + tensor var_554_begin_0 = const()[name = tensor("op_554_begin_0"), val = tensor([0, 2944, 0, 0])]; + tensor var_554_end_0 = const()[name = tensor("op_554_end_0"), val = tensor([1, 3072, 1, 512])]; + tensor var_554_end_mask_0 = const()[name = tensor("op_554_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_554_cast_fp16 = slice_by_index(begin = var_554_begin_0, end = var_554_end_0, end_mask = var_554_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_554_cast_fp16")]; + tensor var_558_begin_0 = const()[name = tensor("op_558_begin_0"), val = tensor([0, 3072, 0, 0])]; + tensor var_558_end_0 = const()[name = tensor("op_558_end_0"), val = tensor([1, 3200, 1, 512])]; + tensor var_558_end_mask_0 = const()[name = tensor("op_558_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_558_cast_fp16 = slice_by_index(begin = var_558_begin_0, end = var_558_end_0, end_mask = var_558_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_558_cast_fp16")]; + tensor var_562_begin_0 = const()[name = tensor("op_562_begin_0"), val = tensor([0, 3200, 0, 0])]; + tensor var_562_end_0 = const()[name = tensor("op_562_end_0"), val = tensor([1, 3328, 1, 512])]; + tensor var_562_end_mask_0 = const()[name = tensor("op_562_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_562_cast_fp16 = slice_by_index(begin = var_562_begin_0, end = var_562_end_0, end_mask = var_562_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_562_cast_fp16")]; + tensor var_566_begin_0 = const()[name = tensor("op_566_begin_0"), val = tensor([0, 3328, 0, 0])]; + tensor var_566_end_0 = const()[name = tensor("op_566_end_0"), val = tensor([1, 3456, 1, 512])]; + tensor var_566_end_mask_0 = const()[name = tensor("op_566_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_566_cast_fp16 = slice_by_index(begin = var_566_begin_0, end = var_566_end_0, end_mask = var_566_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_566_cast_fp16")]; + tensor var_570_begin_0 = const()[name = tensor("op_570_begin_0"), val = tensor([0, 3456, 0, 0])]; + tensor var_570_end_0 = const()[name = tensor("op_570_end_0"), val = tensor([1, 3584, 1, 512])]; + tensor var_570_end_mask_0 = const()[name = tensor("op_570_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_570_cast_fp16 = slice_by_index(begin = var_570_begin_0, end = var_570_end_0, end_mask = var_570_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_570_cast_fp16")]; + tensor var_574_begin_0 = const()[name = tensor("op_574_begin_0"), val = tensor([0, 3584, 0, 0])]; + tensor var_574_end_0 = const()[name = tensor("op_574_end_0"), val = tensor([1, 3712, 1, 512])]; + tensor var_574_end_mask_0 = const()[name = tensor("op_574_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_574_cast_fp16 = slice_by_index(begin = var_574_begin_0, end = var_574_end_0, end_mask = var_574_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor var_578_begin_0 = const()[name = tensor("op_578_begin_0"), val = tensor([0, 3712, 0, 0])]; + tensor var_578_end_0 = const()[name = tensor("op_578_end_0"), val = tensor([1, 3840, 1, 512])]; + tensor var_578_end_mask_0 = const()[name = tensor("op_578_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_578_cast_fp16 = slice_by_index(begin = var_578_begin_0, end = var_578_end_0, end_mask = var_578_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_578_cast_fp16")]; + tensor var_582_begin_0 = const()[name = tensor("op_582_begin_0"), val = tensor([0, 3840, 0, 0])]; + tensor var_582_end_0 = const()[name = tensor("op_582_end_0"), val = tensor([1, 3968, 1, 512])]; + tensor var_582_end_mask_0 = const()[name = tensor("op_582_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_582_cast_fp16 = slice_by_index(begin = var_582_begin_0, end = var_582_end_0, end_mask = var_582_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_582_cast_fp16")]; + tensor var_586_begin_0 = const()[name = tensor("op_586_begin_0"), val = tensor([0, 3968, 0, 0])]; + tensor var_586_end_0 = const()[name = tensor("op_586_end_0"), val = tensor([1, 4096, 1, 512])]; + tensor var_586_end_mask_0 = const()[name = tensor("op_586_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_586_cast_fp16 = slice_by_index(begin = var_586_begin_0, end = var_586_end_0, end_mask = var_586_end_mask_0, x = v_7_cast_fp16)[name = tensor("op_586_cast_fp16")]; + tensor var_590_equation_0 = const()[name = tensor("op_590_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_590_cast_fp16 = einsum(equation = var_590_equation_0, values = (var_336_cast_fp16, var_206_cast_fp16))[name = tensor("op_590_cast_fp16")]; + tensor var_591_to_fp16 = const()[name = tensor("op_591_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_592_cast_fp16 = mul(x = var_590_cast_fp16, y = var_591_to_fp16)[name = tensor("op_592_cast_fp16")]; + tensor var_594_equation_0 = const()[name = tensor("op_594_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_594_cast_fp16 = einsum(equation = var_594_equation_0, values = (var_340_cast_fp16, var_210_cast_fp16))[name = tensor("op_594_cast_fp16")]; + tensor var_595_to_fp16 = const()[name = tensor("op_595_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_596_cast_fp16 = mul(x = var_594_cast_fp16, y = var_595_to_fp16)[name = tensor("op_596_cast_fp16")]; + tensor var_598_equation_0 = const()[name = tensor("op_598_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_598_cast_fp16 = einsum(equation = var_598_equation_0, values = (var_344_cast_fp16, var_214_cast_fp16))[name = tensor("op_598_cast_fp16")]; + tensor var_599_to_fp16 = const()[name = tensor("op_599_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_600_cast_fp16 = mul(x = var_598_cast_fp16, y = var_599_to_fp16)[name = tensor("op_600_cast_fp16")]; + tensor var_602_equation_0 = const()[name = tensor("op_602_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_602_cast_fp16 = einsum(equation = var_602_equation_0, values = (var_348_cast_fp16, var_218_cast_fp16))[name = tensor("op_602_cast_fp16")]; + tensor var_603_to_fp16 = const()[name = tensor("op_603_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_604_cast_fp16 = mul(x = var_602_cast_fp16, y = var_603_to_fp16)[name = tensor("op_604_cast_fp16")]; + tensor var_606_equation_0 = const()[name = tensor("op_606_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_606_cast_fp16 = einsum(equation = var_606_equation_0, values = (var_352_cast_fp16, var_222_cast_fp16))[name = tensor("op_606_cast_fp16")]; + tensor var_607_to_fp16 = const()[name = tensor("op_607_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_608_cast_fp16 = mul(x = var_606_cast_fp16, y = var_607_to_fp16)[name = tensor("op_608_cast_fp16")]; + tensor var_610_equation_0 = const()[name = tensor("op_610_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_610_cast_fp16 = einsum(equation = var_610_equation_0, values = (var_356_cast_fp16, var_226_cast_fp16))[name = tensor("op_610_cast_fp16")]; + tensor var_611_to_fp16 = const()[name = tensor("op_611_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_612_cast_fp16 = mul(x = var_610_cast_fp16, y = var_611_to_fp16)[name = tensor("op_612_cast_fp16")]; + tensor var_614_equation_0 = const()[name = tensor("op_614_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_614_cast_fp16 = einsum(equation = var_614_equation_0, values = (var_360_cast_fp16, var_230_cast_fp16))[name = tensor("op_614_cast_fp16")]; + tensor var_615_to_fp16 = const()[name = tensor("op_615_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_616_cast_fp16 = mul(x = var_614_cast_fp16, y = var_615_to_fp16)[name = tensor("op_616_cast_fp16")]; + tensor var_618_equation_0 = const()[name = tensor("op_618_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_618_cast_fp16 = einsum(equation = var_618_equation_0, values = (var_364_cast_fp16, var_234_cast_fp16))[name = tensor("op_618_cast_fp16")]; + tensor var_619_to_fp16 = const()[name = tensor("op_619_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_620_cast_fp16 = mul(x = var_618_cast_fp16, y = var_619_to_fp16)[name = tensor("op_620_cast_fp16")]; + tensor var_622_equation_0 = const()[name = tensor("op_622_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_622_cast_fp16 = einsum(equation = var_622_equation_0, values = (var_368_cast_fp16, var_238_cast_fp16))[name = tensor("op_622_cast_fp16")]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_624_cast_fp16 = mul(x = var_622_cast_fp16, y = var_623_to_fp16)[name = tensor("op_624_cast_fp16")]; + tensor var_626_equation_0 = const()[name = tensor("op_626_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_626_cast_fp16 = einsum(equation = var_626_equation_0, values = (var_372_cast_fp16, var_242_cast_fp16))[name = tensor("op_626_cast_fp16")]; + tensor var_627_to_fp16 = const()[name = tensor("op_627_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_628_cast_fp16 = mul(x = var_626_cast_fp16, y = var_627_to_fp16)[name = tensor("op_628_cast_fp16")]; + tensor var_630_equation_0 = const()[name = tensor("op_630_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_630_cast_fp16 = einsum(equation = var_630_equation_0, values = (var_376_cast_fp16, var_246_cast_fp16))[name = tensor("op_630_cast_fp16")]; + tensor var_631_to_fp16 = const()[name = tensor("op_631_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_632_cast_fp16 = mul(x = var_630_cast_fp16, y = var_631_to_fp16)[name = tensor("op_632_cast_fp16")]; + tensor var_634_equation_0 = const()[name = tensor("op_634_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_634_cast_fp16 = einsum(equation = var_634_equation_0, values = (var_380_cast_fp16, var_250_cast_fp16))[name = tensor("op_634_cast_fp16")]; + tensor var_635_to_fp16 = const()[name = tensor("op_635_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_636_cast_fp16 = mul(x = var_634_cast_fp16, y = var_635_to_fp16)[name = tensor("op_636_cast_fp16")]; + tensor var_638_equation_0 = const()[name = tensor("op_638_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_638_cast_fp16 = einsum(equation = var_638_equation_0, values = (var_384_cast_fp16, var_254_cast_fp16))[name = tensor("op_638_cast_fp16")]; + tensor var_639_to_fp16 = const()[name = tensor("op_639_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_640_cast_fp16 = mul(x = var_638_cast_fp16, y = var_639_to_fp16)[name = tensor("op_640_cast_fp16")]; + tensor var_642_equation_0 = const()[name = tensor("op_642_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_642_cast_fp16 = einsum(equation = var_642_equation_0, values = (var_388_cast_fp16, var_258_cast_fp16))[name = tensor("op_642_cast_fp16")]; + tensor var_643_to_fp16 = const()[name = tensor("op_643_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_644_cast_fp16 = mul(x = var_642_cast_fp16, y = var_643_to_fp16)[name = tensor("op_644_cast_fp16")]; + tensor var_646_equation_0 = const()[name = tensor("op_646_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_646_cast_fp16 = einsum(equation = var_646_equation_0, values = (var_392_cast_fp16, var_262_cast_fp16))[name = tensor("op_646_cast_fp16")]; + tensor var_647_to_fp16 = const()[name = tensor("op_647_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_648_cast_fp16 = mul(x = var_646_cast_fp16, y = var_647_to_fp16)[name = tensor("op_648_cast_fp16")]; + tensor var_650_equation_0 = const()[name = tensor("op_650_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_650_cast_fp16 = einsum(equation = var_650_equation_0, values = (var_396_cast_fp16, var_266_cast_fp16))[name = tensor("op_650_cast_fp16")]; + tensor var_651_to_fp16 = const()[name = tensor("op_651_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_652_cast_fp16 = mul(x = var_650_cast_fp16, y = var_651_to_fp16)[name = tensor("op_652_cast_fp16")]; + tensor var_654_equation_0 = const()[name = tensor("op_654_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_654_cast_fp16 = einsum(equation = var_654_equation_0, values = (var_400_cast_fp16, var_270_cast_fp16))[name = tensor("op_654_cast_fp16")]; + tensor var_655_to_fp16 = const()[name = tensor("op_655_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_656_cast_fp16 = mul(x = var_654_cast_fp16, y = var_655_to_fp16)[name = tensor("op_656_cast_fp16")]; + tensor var_658_equation_0 = const()[name = tensor("op_658_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_658_cast_fp16 = einsum(equation = var_658_equation_0, values = (var_404_cast_fp16, var_274_cast_fp16))[name = tensor("op_658_cast_fp16")]; + tensor var_659_to_fp16 = const()[name = tensor("op_659_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_660_cast_fp16 = mul(x = var_658_cast_fp16, y = var_659_to_fp16)[name = tensor("op_660_cast_fp16")]; + tensor var_662_equation_0 = const()[name = tensor("op_662_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_662_cast_fp16 = einsum(equation = var_662_equation_0, values = (var_408_cast_fp16, var_278_cast_fp16))[name = tensor("op_662_cast_fp16")]; + tensor var_663_to_fp16 = const()[name = tensor("op_663_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_664_cast_fp16 = mul(x = var_662_cast_fp16, y = var_663_to_fp16)[name = tensor("op_664_cast_fp16")]; + tensor var_666_equation_0 = const()[name = tensor("op_666_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_666_cast_fp16 = einsum(equation = var_666_equation_0, values = (var_412_cast_fp16, var_282_cast_fp16))[name = tensor("op_666_cast_fp16")]; + tensor var_667_to_fp16 = const()[name = tensor("op_667_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_668_cast_fp16 = mul(x = var_666_cast_fp16, y = var_667_to_fp16)[name = tensor("op_668_cast_fp16")]; + tensor var_670_equation_0 = const()[name = tensor("op_670_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_670_cast_fp16 = einsum(equation = var_670_equation_0, values = (var_416_cast_fp16, var_286_cast_fp16))[name = tensor("op_670_cast_fp16")]; + tensor var_671_to_fp16 = const()[name = tensor("op_671_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_672_cast_fp16 = mul(x = var_670_cast_fp16, y = var_671_to_fp16)[name = tensor("op_672_cast_fp16")]; + tensor var_674_equation_0 = const()[name = tensor("op_674_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_674_cast_fp16 = einsum(equation = var_674_equation_0, values = (var_420_cast_fp16, var_290_cast_fp16))[name = tensor("op_674_cast_fp16")]; + tensor var_675_to_fp16 = const()[name = tensor("op_675_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_676_cast_fp16 = mul(x = var_674_cast_fp16, y = var_675_to_fp16)[name = tensor("op_676_cast_fp16")]; + tensor var_678_equation_0 = const()[name = tensor("op_678_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_678_cast_fp16 = einsum(equation = var_678_equation_0, values = (var_424_cast_fp16, var_294_cast_fp16))[name = tensor("op_678_cast_fp16")]; + tensor var_679_to_fp16 = const()[name = tensor("op_679_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_680_cast_fp16 = mul(x = var_678_cast_fp16, y = var_679_to_fp16)[name = tensor("op_680_cast_fp16")]; + tensor var_682_equation_0 = const()[name = tensor("op_682_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_682_cast_fp16 = einsum(equation = var_682_equation_0, values = (var_428_cast_fp16, var_298_cast_fp16))[name = tensor("op_682_cast_fp16")]; + tensor var_683_to_fp16 = const()[name = tensor("op_683_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_684_cast_fp16 = mul(x = var_682_cast_fp16, y = var_683_to_fp16)[name = tensor("op_684_cast_fp16")]; + tensor var_686_equation_0 = const()[name = tensor("op_686_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_686_cast_fp16 = einsum(equation = var_686_equation_0, values = (var_432_cast_fp16, var_302_cast_fp16))[name = tensor("op_686_cast_fp16")]; + tensor var_687_to_fp16 = const()[name = tensor("op_687_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_688_cast_fp16 = mul(x = var_686_cast_fp16, y = var_687_to_fp16)[name = tensor("op_688_cast_fp16")]; + tensor var_690_equation_0 = const()[name = tensor("op_690_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_690_cast_fp16 = einsum(equation = var_690_equation_0, values = (var_436_cast_fp16, var_306_cast_fp16))[name = tensor("op_690_cast_fp16")]; + tensor var_691_to_fp16 = const()[name = tensor("op_691_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_692_cast_fp16 = mul(x = var_690_cast_fp16, y = var_691_to_fp16)[name = tensor("op_692_cast_fp16")]; + tensor var_694_equation_0 = const()[name = tensor("op_694_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_694_cast_fp16 = einsum(equation = var_694_equation_0, values = (var_440_cast_fp16, var_310_cast_fp16))[name = tensor("op_694_cast_fp16")]; + tensor var_695_to_fp16 = const()[name = tensor("op_695_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_696_cast_fp16 = mul(x = var_694_cast_fp16, y = var_695_to_fp16)[name = tensor("op_696_cast_fp16")]; + tensor var_698_equation_0 = const()[name = tensor("op_698_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_698_cast_fp16 = einsum(equation = var_698_equation_0, values = (var_444_cast_fp16, var_314_cast_fp16))[name = tensor("op_698_cast_fp16")]; + tensor var_699_to_fp16 = const()[name = tensor("op_699_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_700_cast_fp16 = mul(x = var_698_cast_fp16, y = var_699_to_fp16)[name = tensor("op_700_cast_fp16")]; + tensor var_702_equation_0 = const()[name = tensor("op_702_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_702_cast_fp16 = einsum(equation = var_702_equation_0, values = (var_448_cast_fp16, var_318_cast_fp16))[name = tensor("op_702_cast_fp16")]; + tensor var_703_to_fp16 = const()[name = tensor("op_703_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_704_cast_fp16 = mul(x = var_702_cast_fp16, y = var_703_to_fp16)[name = tensor("op_704_cast_fp16")]; + tensor var_706_equation_0 = const()[name = tensor("op_706_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_706_cast_fp16 = einsum(equation = var_706_equation_0, values = (var_452_cast_fp16, var_322_cast_fp16))[name = tensor("op_706_cast_fp16")]; + tensor var_707_to_fp16 = const()[name = tensor("op_707_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_708_cast_fp16 = mul(x = var_706_cast_fp16, y = var_707_to_fp16)[name = tensor("op_708_cast_fp16")]; + tensor var_710_equation_0 = const()[name = tensor("op_710_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_710_cast_fp16 = einsum(equation = var_710_equation_0, values = (var_456_cast_fp16, var_326_cast_fp16))[name = tensor("op_710_cast_fp16")]; + tensor var_711_to_fp16 = const()[name = tensor("op_711_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_712_cast_fp16 = mul(x = var_710_cast_fp16, y = var_711_to_fp16)[name = tensor("op_712_cast_fp16")]; + tensor var_714_equation_0 = const()[name = tensor("op_714_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_714_cast_fp16 = einsum(equation = var_714_equation_0, values = (var_460_cast_fp16, var_330_cast_fp16))[name = tensor("op_714_cast_fp16")]; + tensor var_715_to_fp16 = const()[name = tensor("op_715_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_716_cast_fp16 = mul(x = var_714_cast_fp16, y = var_715_to_fp16)[name = tensor("op_716_cast_fp16")]; + tensor aw_1_cast_fp16 = add(x = var_592_cast_fp16, y = mask)[name = tensor("aw_1_cast_fp16")]; + tensor aw_3_cast_fp16 = add(x = var_596_cast_fp16, y = mask)[name = tensor("aw_3_cast_fp16")]; + tensor aw_5_cast_fp16 = add(x = var_600_cast_fp16, y = mask)[name = tensor("aw_5_cast_fp16")]; + tensor aw_7_cast_fp16 = add(x = var_604_cast_fp16, y = mask)[name = tensor("aw_7_cast_fp16")]; + tensor aw_9_cast_fp16 = add(x = var_608_cast_fp16, y = mask)[name = tensor("aw_9_cast_fp16")]; + tensor aw_11_cast_fp16 = add(x = var_612_cast_fp16, y = mask)[name = tensor("aw_11_cast_fp16")]; + tensor aw_13_cast_fp16 = add(x = var_616_cast_fp16, y = mask)[name = tensor("aw_13_cast_fp16")]; + tensor aw_15_cast_fp16 = add(x = var_620_cast_fp16, y = mask)[name = tensor("aw_15_cast_fp16")]; + tensor aw_17_cast_fp16 = add(x = var_624_cast_fp16, y = mask)[name = tensor("aw_17_cast_fp16")]; + tensor aw_19_cast_fp16 = add(x = var_628_cast_fp16, y = mask)[name = tensor("aw_19_cast_fp16")]; + tensor aw_21_cast_fp16 = add(x = var_632_cast_fp16, y = mask)[name = tensor("aw_21_cast_fp16")]; + tensor aw_23_cast_fp16 = add(x = var_636_cast_fp16, y = mask)[name = tensor("aw_23_cast_fp16")]; + tensor aw_25_cast_fp16 = add(x = var_640_cast_fp16, y = mask)[name = tensor("aw_25_cast_fp16")]; + tensor aw_27_cast_fp16 = add(x = var_644_cast_fp16, y = mask)[name = tensor("aw_27_cast_fp16")]; + tensor aw_29_cast_fp16 = add(x = var_648_cast_fp16, y = mask)[name = tensor("aw_29_cast_fp16")]; + tensor aw_31_cast_fp16 = add(x = var_652_cast_fp16, y = mask)[name = tensor("aw_31_cast_fp16")]; + tensor aw_33_cast_fp16 = add(x = var_656_cast_fp16, y = mask)[name = tensor("aw_33_cast_fp16")]; + tensor aw_35_cast_fp16 = add(x = var_660_cast_fp16, y = mask)[name = tensor("aw_35_cast_fp16")]; + tensor aw_37_cast_fp16 = add(x = var_664_cast_fp16, y = mask)[name = tensor("aw_37_cast_fp16")]; + tensor aw_39_cast_fp16 = add(x = var_668_cast_fp16, y = mask)[name = tensor("aw_39_cast_fp16")]; + tensor aw_41_cast_fp16 = add(x = var_672_cast_fp16, y = mask)[name = tensor("aw_41_cast_fp16")]; + tensor aw_43_cast_fp16 = add(x = var_676_cast_fp16, y = mask)[name = tensor("aw_43_cast_fp16")]; + tensor aw_45_cast_fp16 = add(x = var_680_cast_fp16, y = mask)[name = tensor("aw_45_cast_fp16")]; + tensor aw_47_cast_fp16 = add(x = var_684_cast_fp16, y = mask)[name = tensor("aw_47_cast_fp16")]; + tensor aw_49_cast_fp16 = add(x = var_688_cast_fp16, y = mask)[name = tensor("aw_49_cast_fp16")]; + tensor aw_51_cast_fp16 = add(x = var_692_cast_fp16, y = mask)[name = tensor("aw_51_cast_fp16")]; + tensor aw_53_cast_fp16 = add(x = var_696_cast_fp16, y = mask)[name = tensor("aw_53_cast_fp16")]; + tensor aw_55_cast_fp16 = add(x = var_700_cast_fp16, y = mask)[name = tensor("aw_55_cast_fp16")]; + tensor aw_57_cast_fp16 = add(x = var_704_cast_fp16, y = mask)[name = tensor("aw_57_cast_fp16")]; + tensor aw_59_cast_fp16 = add(x = var_708_cast_fp16, y = mask)[name = tensor("aw_59_cast_fp16")]; + tensor aw_61_cast_fp16 = add(x = var_712_cast_fp16, y = mask)[name = tensor("aw_61_cast_fp16")]; + tensor aw_63_cast_fp16 = add(x = var_716_cast_fp16, y = mask)[name = tensor("aw_63_cast_fp16")]; + tensor var_749_cast_fp16 = softmax(axis = var_60, x = aw_1_cast_fp16)[name = tensor("op_749_cast_fp16")]; + tensor var_750_cast_fp16 = softmax(axis = var_60, x = aw_3_cast_fp16)[name = tensor("op_750_cast_fp16")]; + tensor var_751_cast_fp16 = softmax(axis = var_60, x = aw_5_cast_fp16)[name = tensor("op_751_cast_fp16")]; + tensor var_752_cast_fp16 = softmax(axis = var_60, x = aw_7_cast_fp16)[name = tensor("op_752_cast_fp16")]; + tensor var_753_cast_fp16 = softmax(axis = var_60, x = aw_9_cast_fp16)[name = tensor("op_753_cast_fp16")]; + tensor var_754_cast_fp16 = softmax(axis = var_60, x = aw_11_cast_fp16)[name = tensor("op_754_cast_fp16")]; + tensor var_755_cast_fp16 = softmax(axis = var_60, x = aw_13_cast_fp16)[name = tensor("op_755_cast_fp16")]; + tensor var_756_cast_fp16 = softmax(axis = var_60, x = aw_15_cast_fp16)[name = tensor("op_756_cast_fp16")]; + tensor var_757_cast_fp16 = softmax(axis = var_60, x = aw_17_cast_fp16)[name = tensor("op_757_cast_fp16")]; + tensor var_758_cast_fp16 = softmax(axis = var_60, x = aw_19_cast_fp16)[name = tensor("op_758_cast_fp16")]; + tensor var_759_cast_fp16 = softmax(axis = var_60, x = aw_21_cast_fp16)[name = tensor("op_759_cast_fp16")]; + tensor var_760_cast_fp16 = softmax(axis = var_60, x = aw_23_cast_fp16)[name = tensor("op_760_cast_fp16")]; + tensor var_761_cast_fp16 = softmax(axis = var_60, x = aw_25_cast_fp16)[name = tensor("op_761_cast_fp16")]; + tensor var_762_cast_fp16 = softmax(axis = var_60, x = aw_27_cast_fp16)[name = tensor("op_762_cast_fp16")]; + tensor var_763_cast_fp16 = softmax(axis = var_60, x = aw_29_cast_fp16)[name = tensor("op_763_cast_fp16")]; + tensor var_764_cast_fp16 = softmax(axis = var_60, x = aw_31_cast_fp16)[name = tensor("op_764_cast_fp16")]; + tensor var_765_cast_fp16 = softmax(axis = var_60, x = aw_33_cast_fp16)[name = tensor("op_765_cast_fp16")]; + tensor var_766_cast_fp16 = softmax(axis = var_60, x = aw_35_cast_fp16)[name = tensor("op_766_cast_fp16")]; + tensor var_767_cast_fp16 = softmax(axis = var_60, x = aw_37_cast_fp16)[name = tensor("op_767_cast_fp16")]; + tensor var_768_cast_fp16 = softmax(axis = var_60, x = aw_39_cast_fp16)[name = tensor("op_768_cast_fp16")]; + tensor var_769_cast_fp16 = softmax(axis = var_60, x = aw_41_cast_fp16)[name = tensor("op_769_cast_fp16")]; + tensor var_770_cast_fp16 = softmax(axis = var_60, x = aw_43_cast_fp16)[name = tensor("op_770_cast_fp16")]; + tensor var_771_cast_fp16 = softmax(axis = var_60, x = aw_45_cast_fp16)[name = tensor("op_771_cast_fp16")]; + tensor var_772_cast_fp16 = softmax(axis = var_60, x = aw_47_cast_fp16)[name = tensor("op_772_cast_fp16")]; + tensor var_773_cast_fp16 = softmax(axis = var_60, x = aw_49_cast_fp16)[name = tensor("op_773_cast_fp16")]; + tensor var_774_cast_fp16 = softmax(axis = var_60, x = aw_51_cast_fp16)[name = tensor("op_774_cast_fp16")]; + tensor var_775_cast_fp16 = softmax(axis = var_60, x = aw_53_cast_fp16)[name = tensor("op_775_cast_fp16")]; + tensor var_776_cast_fp16 = softmax(axis = var_60, x = aw_55_cast_fp16)[name = tensor("op_776_cast_fp16")]; + tensor var_777_cast_fp16 = softmax(axis = var_60, x = aw_57_cast_fp16)[name = tensor("op_777_cast_fp16")]; + tensor var_778_cast_fp16 = softmax(axis = var_60, x = aw_59_cast_fp16)[name = tensor("op_778_cast_fp16")]; + tensor var_779_cast_fp16 = softmax(axis = var_60, x = aw_61_cast_fp16)[name = tensor("op_779_cast_fp16")]; + tensor var_780_cast_fp16 = softmax(axis = var_60, x = aw_63_cast_fp16)[name = tensor("op_780_cast_fp16")]; + tensor var_782_equation_0 = const()[name = tensor("op_782_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_782_cast_fp16 = einsum(equation = var_782_equation_0, values = (var_462_cast_fp16, var_749_cast_fp16))[name = tensor("op_782_cast_fp16")]; + tensor var_784_equation_0 = const()[name = tensor("op_784_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_784_cast_fp16 = einsum(equation = var_784_equation_0, values = (var_466_cast_fp16, var_750_cast_fp16))[name = tensor("op_784_cast_fp16")]; + tensor var_786_equation_0 = const()[name = tensor("op_786_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_786_cast_fp16 = einsum(equation = var_786_equation_0, values = (var_470_cast_fp16, var_751_cast_fp16))[name = tensor("op_786_cast_fp16")]; + tensor var_788_equation_0 = const()[name = tensor("op_788_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_788_cast_fp16 = einsum(equation = var_788_equation_0, values = (var_474_cast_fp16, var_752_cast_fp16))[name = tensor("op_788_cast_fp16")]; + tensor var_790_equation_0 = const()[name = tensor("op_790_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_790_cast_fp16 = einsum(equation = var_790_equation_0, values = (var_478_cast_fp16, var_753_cast_fp16))[name = tensor("op_790_cast_fp16")]; + tensor var_792_equation_0 = const()[name = tensor("op_792_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_792_cast_fp16 = einsum(equation = var_792_equation_0, values = (var_482_cast_fp16, var_754_cast_fp16))[name = tensor("op_792_cast_fp16")]; + tensor var_794_equation_0 = const()[name = tensor("op_794_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_794_cast_fp16 = einsum(equation = var_794_equation_0, values = (var_486_cast_fp16, var_755_cast_fp16))[name = tensor("op_794_cast_fp16")]; + tensor var_796_equation_0 = const()[name = tensor("op_796_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_796_cast_fp16 = einsum(equation = var_796_equation_0, values = (var_490_cast_fp16, var_756_cast_fp16))[name = tensor("op_796_cast_fp16")]; + tensor var_798_equation_0 = const()[name = tensor("op_798_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_798_cast_fp16 = einsum(equation = var_798_equation_0, values = (var_494_cast_fp16, var_757_cast_fp16))[name = tensor("op_798_cast_fp16")]; + tensor var_800_equation_0 = const()[name = tensor("op_800_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_800_cast_fp16 = einsum(equation = var_800_equation_0, values = (var_498_cast_fp16, var_758_cast_fp16))[name = tensor("op_800_cast_fp16")]; + tensor var_802_equation_0 = const()[name = tensor("op_802_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_802_cast_fp16 = einsum(equation = var_802_equation_0, values = (var_502_cast_fp16, var_759_cast_fp16))[name = tensor("op_802_cast_fp16")]; + tensor var_804_equation_0 = const()[name = tensor("op_804_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_804_cast_fp16 = einsum(equation = var_804_equation_0, values = (var_506_cast_fp16, var_760_cast_fp16))[name = tensor("op_804_cast_fp16")]; + tensor var_806_equation_0 = const()[name = tensor("op_806_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_806_cast_fp16 = einsum(equation = var_806_equation_0, values = (var_510_cast_fp16, var_761_cast_fp16))[name = tensor("op_806_cast_fp16")]; + tensor var_808_equation_0 = const()[name = tensor("op_808_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_808_cast_fp16 = einsum(equation = var_808_equation_0, values = (var_514_cast_fp16, var_762_cast_fp16))[name = tensor("op_808_cast_fp16")]; + tensor var_810_equation_0 = const()[name = tensor("op_810_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_810_cast_fp16 = einsum(equation = var_810_equation_0, values = (var_518_cast_fp16, var_763_cast_fp16))[name = tensor("op_810_cast_fp16")]; + tensor var_812_equation_0 = const()[name = tensor("op_812_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_812_cast_fp16 = einsum(equation = var_812_equation_0, values = (var_522_cast_fp16, var_764_cast_fp16))[name = tensor("op_812_cast_fp16")]; + tensor var_814_equation_0 = const()[name = tensor("op_814_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_814_cast_fp16 = einsum(equation = var_814_equation_0, values = (var_526_cast_fp16, var_765_cast_fp16))[name = tensor("op_814_cast_fp16")]; + tensor var_816_equation_0 = const()[name = tensor("op_816_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_816_cast_fp16 = einsum(equation = var_816_equation_0, values = (var_530_cast_fp16, var_766_cast_fp16))[name = tensor("op_816_cast_fp16")]; + tensor var_818_equation_0 = const()[name = tensor("op_818_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_818_cast_fp16 = einsum(equation = var_818_equation_0, values = (var_534_cast_fp16, var_767_cast_fp16))[name = tensor("op_818_cast_fp16")]; + tensor var_820_equation_0 = const()[name = tensor("op_820_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_820_cast_fp16 = einsum(equation = var_820_equation_0, values = (var_538_cast_fp16, var_768_cast_fp16))[name = tensor("op_820_cast_fp16")]; + tensor var_822_equation_0 = const()[name = tensor("op_822_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_822_cast_fp16 = einsum(equation = var_822_equation_0, values = (var_542_cast_fp16, var_769_cast_fp16))[name = tensor("op_822_cast_fp16")]; + tensor var_824_equation_0 = const()[name = tensor("op_824_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_824_cast_fp16 = einsum(equation = var_824_equation_0, values = (var_546_cast_fp16, var_770_cast_fp16))[name = tensor("op_824_cast_fp16")]; + tensor var_826_equation_0 = const()[name = tensor("op_826_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_826_cast_fp16 = einsum(equation = var_826_equation_0, values = (var_550_cast_fp16, var_771_cast_fp16))[name = tensor("op_826_cast_fp16")]; + tensor var_828_equation_0 = const()[name = tensor("op_828_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_828_cast_fp16 = einsum(equation = var_828_equation_0, values = (var_554_cast_fp16, var_772_cast_fp16))[name = tensor("op_828_cast_fp16")]; + tensor var_830_equation_0 = const()[name = tensor("op_830_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_830_cast_fp16 = einsum(equation = var_830_equation_0, values = (var_558_cast_fp16, var_773_cast_fp16))[name = tensor("op_830_cast_fp16")]; + tensor var_832_equation_0 = const()[name = tensor("op_832_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_832_cast_fp16 = einsum(equation = var_832_equation_0, values = (var_562_cast_fp16, var_774_cast_fp16))[name = tensor("op_832_cast_fp16")]; + tensor var_834_equation_0 = const()[name = tensor("op_834_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_834_cast_fp16 = einsum(equation = var_834_equation_0, values = (var_566_cast_fp16, var_775_cast_fp16))[name = tensor("op_834_cast_fp16")]; + tensor var_836_equation_0 = const()[name = tensor("op_836_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_836_cast_fp16 = einsum(equation = var_836_equation_0, values = (var_570_cast_fp16, var_776_cast_fp16))[name = tensor("op_836_cast_fp16")]; + tensor var_838_equation_0 = const()[name = tensor("op_838_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_838_cast_fp16 = einsum(equation = var_838_equation_0, values = (var_574_cast_fp16, var_777_cast_fp16))[name = tensor("op_838_cast_fp16")]; + tensor var_840_equation_0 = const()[name = tensor("op_840_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_840_cast_fp16 = einsum(equation = var_840_equation_0, values = (var_578_cast_fp16, var_778_cast_fp16))[name = tensor("op_840_cast_fp16")]; + tensor var_842_equation_0 = const()[name = tensor("op_842_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_842_cast_fp16 = einsum(equation = var_842_equation_0, values = (var_582_cast_fp16, var_779_cast_fp16))[name = tensor("op_842_cast_fp16")]; + tensor var_844_equation_0 = const()[name = tensor("op_844_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_844_cast_fp16 = einsum(equation = var_844_equation_0, values = (var_586_cast_fp16, var_780_cast_fp16))[name = tensor("op_844_cast_fp16")]; + tensor x_11_interleave_0 = const()[name = tensor("x_11_interleave_0"), val = tensor(false)]; + tensor x_11_cast_fp16 = concat(axis = var_60, interleave = x_11_interleave_0, values = (var_782_cast_fp16, var_784_cast_fp16, var_786_cast_fp16, var_788_cast_fp16, var_790_cast_fp16, var_792_cast_fp16, var_794_cast_fp16, var_796_cast_fp16, var_798_cast_fp16, var_800_cast_fp16, var_802_cast_fp16, var_804_cast_fp16, var_806_cast_fp16, var_808_cast_fp16, var_810_cast_fp16, var_812_cast_fp16, var_814_cast_fp16, var_816_cast_fp16, var_818_cast_fp16, var_820_cast_fp16, var_822_cast_fp16, var_824_cast_fp16, var_826_cast_fp16, var_828_cast_fp16, var_830_cast_fp16, var_832_cast_fp16, var_834_cast_fp16, var_836_cast_fp16, var_838_cast_fp16, var_840_cast_fp16, var_842_cast_fp16, var_844_cast_fp16))[name = tensor("x_11_cast_fp16")]; + tensor var_849 = const()[name = tensor("op_849"), val = tensor([1, 4096, -1, 8])]; + tensor input_3_cast_fp16 = reshape(shape = var_849, x = x_11_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 1])]; + tensor var_855 = const()[name = tensor("op_855"), val = tensor([1, 1])]; + tensor var_857_pad_type_0 = const()[name = tensor("op_857_pad_type_0"), val = tensor("custom")]; + tensor var_857_pad_0 = const()[name = tensor("op_857_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_857_cast_fp16 = conv(dilations = var_855, groups = var_60, pad = var_857_pad_0, pad_type = var_857_pad_type_0, strides = var_853, weight = blocks_0_attn_proj_weight_palettized_cast_fp16, x = input_3_cast_fp16)[name = tensor("op_857_cast_fp16")]; + tensor blocks_0_attn_proj_output_scales_to_fp16 = const()[name = tensor("blocks_0_attn_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202411136)))]; + tensor attention_output_1_cast_fp16 = mul(x = var_857_cast_fp16, y = blocks_0_attn_proj_output_scales_to_fp16)[name = tensor("attention_output_1_cast_fp16")]; + tensor x_13_cast_fp16 = add(x = attention_output_1_cast_fp16, y = x)[name = tensor("x_13_cast_fp16")]; + tensor x_eps_3_interleave_0 = const()[name = tensor("x_eps_3_interleave_0"), val = tensor(false)]; + tensor eps_chan_3_to_fp16 = const()[name = tensor("eps_chan_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202419392)))]; + tensor x_eps_3_cast_fp16 = concat(axis = var_60, interleave = x_eps_3_interleave_0, values = (x_13_cast_fp16, eps_chan_3_to_fp16))[name = tensor("x_eps_3_cast_fp16")]; + tensor norm_x_3_axes_0 = const()[name = tensor("norm_x_3_axes_0"), val = tensor([1])]; + tensor norm_x_3_cast_fp16 = reduce_l2_norm(axes = norm_x_3_axes_0, keep_dims = var_63, x = x_eps_3_cast_fp16)[name = tensor("norm_x_3_cast_fp16")]; + tensor x_normed_7_cast_fp16 = real_div(x = x_13_cast_fp16, y = norm_x_3_cast_fp16)[name = tensor("x_normed_7_cast_fp16")]; + tensor var_882_to_fp16 = const()[name = tensor("op_882_to_fp16"), val = tensor(0x1p+6)]; + tensor x_normed_9_cast_fp16 = mul(x = x_normed_7_cast_fp16, y = var_882_to_fp16)[name = tensor("x_normed_9_cast_fp16")]; + tensor blocks_0_norm_2_weight_to_fp16 = const()[name = tensor("blocks_0_norm_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202419584)))]; + tensor input_5_cast_fp16 = mul(x = x_normed_9_cast_fp16, y = blocks_0_norm_2_weight_to_fp16)[name = tensor("input_5_cast_fp16")]; + tensor var_894 = const()[name = tensor("op_894"), val = tensor([1, 1])]; + tensor var_896 = const()[name = tensor("op_896"), val = tensor([1, 1])]; + tensor var_898_pad_type_0 = const()[name = tensor("op_898_pad_type_0"), val = tensor("custom")]; + tensor var_898_pad_0 = const()[name = tensor("op_898_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_898_cast_fp16 = conv(dilations = var_896, groups = var_60, pad = var_898_pad_0, pad_type = var_898_pad_type_0, strides = var_894, weight = blocks_0_mlp_fc_1_weight_palettized_cast_fp16, x = input_5_cast_fp16)[name = tensor("op_898_cast_fp16")]; + tensor blocks_0_mlp_fc_1_output_scales_to_fp16 = const()[name = tensor("blocks_0_mlp_fc_1_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202427840)))]; + tensor input_7_cast_fp16 = mul(x = var_898_cast_fp16, y = blocks_0_mlp_fc_1_output_scales_to_fp16)[name = tensor("input_7_cast_fp16")]; + tensor var_902 = const()[name = tensor("op_902"), val = tensor([1, 1])]; + tensor var_904 = const()[name = tensor("op_904"), val = tensor([1, 1])]; + tensor var_906_pad_type_0 = const()[name = tensor("op_906_pad_type_0"), val = tensor("custom")]; + tensor var_906_pad_0 = const()[name = tensor("op_906_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_906_cast_fp16 = conv(dilations = var_904, groups = var_60, pad = var_906_pad_0, pad_type = var_906_pad_type_0, strides = var_902, weight = blocks_0_mlp_fc_2_weight_palettized_cast_fp16, x = input_5_cast_fp16)[name = tensor("op_906_cast_fp16")]; + tensor blocks_0_mlp_fc_2_output_scales_to_fp16 = const()[name = tensor("blocks_0_mlp_fc_2_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202449920)))]; + tensor x_fc_2_1_cast_fp16 = mul(x = var_906_cast_fp16, y = blocks_0_mlp_fc_2_output_scales_to_fp16)[name = tensor("x_fc_2_1_cast_fp16")]; + tensor var_908_cast_fp16 = silu(x = input_7_cast_fp16)[name = tensor("op_908_cast_fp16")]; + tensor input_9_cast_fp16 = mul(x = var_908_cast_fp16, y = x_fc_2_1_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor var_912 = const()[name = tensor("op_912"), val = tensor([1, 1])]; + tensor var_914 = const()[name = tensor("op_914"), val = tensor([1, 1])]; + tensor var_916_pad_type_0 = const()[name = tensor("op_916_pad_type_0"), val = tensor("custom")]; + tensor var_916_pad_0 = const()[name = tensor("op_916_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_916_cast_fp16 = conv(dilations = var_914, groups = var_60, pad = var_916_pad_0, pad_type = var_916_pad_type_0, strides = var_912, weight = blocks_0_mlp_proj_weight_palettized_cast_fp16, x = input_9_cast_fp16)[name = tensor("op_916_cast_fp16")]; + tensor blocks_0_mlp_proj_output_scales_to_fp16 = const()[name = tensor("blocks_0_mlp_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202472000)))]; + tensor var_917_cast_fp16 = mul(x = var_916_cast_fp16, y = blocks_0_mlp_proj_output_scales_to_fp16)[name = tensor("op_917_cast_fp16")]; + tensor x_17_cast_fp16 = add(x = var_917_cast_fp16, y = x_13_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor var_923 = const()[name = tensor("op_923"), val = tensor(-1)]; + tensor var_927 = const()[name = tensor("op_927"), val = tensor(-2)]; + tensor var_929 = const()[name = tensor("op_929"), val = tensor(-3)]; + tensor var_970 = const()[name = tensor("op_970"), val = tensor(1)]; + tensor var_973 = const()[name = tensor("op_973"), val = tensor(true)]; + tensor x_eps_5_interleave_0 = const()[name = tensor("x_eps_5_interleave_0"), val = tensor(false)]; + tensor eps_chan_5_to_fp16 = const()[name = tensor("eps_chan_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202480256)))]; + tensor x_eps_5_cast_fp16 = concat(axis = var_970, interleave = x_eps_5_interleave_0, values = (x_17_cast_fp16, eps_chan_5_to_fp16))[name = tensor("x_eps_5_cast_fp16")]; + tensor norm_x_5_axes_0 = const()[name = tensor("norm_x_5_axes_0"), val = tensor([1])]; + tensor norm_x_5_cast_fp16 = reduce_l2_norm(axes = norm_x_5_axes_0, keep_dims = var_973, x = x_eps_5_cast_fp16)[name = tensor("norm_x_5_cast_fp16")]; + tensor x_normed_13_cast_fp16 = real_div(x = x_17_cast_fp16, y = norm_x_5_cast_fp16)[name = tensor("x_normed_13_cast_fp16")]; + tensor var_996_to_fp16 = const()[name = tensor("op_996_to_fp16"), val = tensor(0x1p+6)]; + tensor x_normed_15_cast_fp16 = mul(x = x_normed_13_cast_fp16, y = var_996_to_fp16)[name = tensor("x_normed_15_cast_fp16")]; + tensor blocks_1_norm_1_weight_to_fp16 = const()[name = tensor("blocks_1_norm_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202480448)))]; + tensor x_21_cast_fp16 = mul(x = x_normed_15_cast_fp16, y = blocks_1_norm_1_weight_to_fp16)[name = tensor("x_21_cast_fp16")]; + tensor var_1021 = const()[name = tensor("op_1021"), val = tensor([1, 4096, 1, -1])]; + tensor input_11_cast_fp16 = reshape(shape = var_1021, x = x_21_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_1025 = const()[name = tensor("op_1025"), val = tensor([1, 1])]; + tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([1, 1])]; + tensor var_1029_pad_type_0 = const()[name = tensor("op_1029_pad_type_0"), val = tensor("custom")]; + tensor var_1029_pad_0 = const()[name = tensor("op_1029_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1029_cast_fp16 = conv(dilations = var_1027, groups = var_970, pad = var_1029_pad_0, pad_type = var_1029_pad_type_0, strides = var_1025, weight = blocks_1_attn_q_proj_weight_palettized_cast_fp16, x = input_11_cast_fp16)[name = tensor("op_1029_cast_fp16")]; + tensor blocks_1_attn_q_proj_output_scales_to_fp16 = const()[name = tensor("blocks_1_attn_q_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202488704)))]; + tensor q_9_cast_fp16 = mul(x = var_1029_cast_fp16, y = blocks_1_attn_q_proj_output_scales_to_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_1033 = const()[name = tensor("op_1033"), val = tensor([1, 1])]; + tensor var_1035 = const()[name = tensor("op_1035"), val = tensor([1, 1])]; + tensor var_1037_pad_type_0 = const()[name = tensor("op_1037_pad_type_0"), val = tensor("custom")]; + tensor var_1037_pad_0 = const()[name = tensor("op_1037_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1037_cast_fp16 = conv(dilations = var_1035, groups = var_970, pad = var_1037_pad_0, pad_type = var_1037_pad_type_0, strides = var_1033, weight = blocks_1_attn_k_proj_weight_palettized_cast_fp16, x = input_11_cast_fp16)[name = tensor("op_1037_cast_fp16")]; + tensor blocks_1_attn_k_proj_output_scales_to_fp16 = const()[name = tensor("blocks_1_attn_k_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202496960)))]; + tensor k_11_cast_fp16 = mul(x = var_1037_cast_fp16, y = blocks_1_attn_k_proj_output_scales_to_fp16)[name = tensor("k_11_cast_fp16")]; + tensor var_1041 = const()[name = tensor("op_1041"), val = tensor([1, 1])]; + tensor var_1043 = const()[name = tensor("op_1043"), val = tensor([1, 1])]; + tensor var_1045_pad_type_0 = const()[name = tensor("op_1045_pad_type_0"), val = tensor("custom")]; + tensor var_1045_pad_0 = const()[name = tensor("op_1045_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1045_cast_fp16 = conv(dilations = var_1043, groups = var_970, pad = var_1045_pad_0, pad_type = var_1045_pad_type_0, strides = var_1041, weight = blocks_1_attn_v_proj_weight_palettized_cast_fp16, x = input_11_cast_fp16)[name = tensor("op_1045_cast_fp16")]; + tensor blocks_1_attn_v_proj_output_scales_to_fp16 = const()[name = tensor("blocks_1_attn_v_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202505216)))]; + tensor v_11_cast_fp16 = mul(x = var_1045_cast_fp16, y = blocks_1_attn_v_proj_output_scales_to_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_1047 = const()[name = tensor("op_1047"), val = tensor([1, 32, 128, 64])]; + tensor q_11_cast_fp16 = reshape(shape = var_1047, x = q_9_cast_fp16)[name = tensor("q_11_cast_fp16")]; + tensor var_1049 = const()[name = tensor("op_1049"), val = tensor([1, 32, 128, 64])]; + tensor k_13_cast_fp16 = reshape(shape = var_1049, x = k_11_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor var_1063_begin_0 = const()[name = tensor("op_1063_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1063_end_0 = const()[name = tensor("op_1063_end_0"), val = tensor([1, 32, 64, 64])]; + tensor var_1063_end_mask_0 = const()[name = tensor("op_1063_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1063_cast_fp16 = slice_by_index(begin = var_1063_begin_0, end = var_1063_end_0, end_mask = var_1063_end_mask_0, x = q_11_cast_fp16)[name = tensor("op_1063_cast_fp16")]; + tensor var_1069_begin_0 = const()[name = tensor("op_1069_begin_0"), val = tensor([0, 0, 64, 0])]; + tensor var_1069_end_0 = const()[name = tensor("op_1069_end_0"), val = tensor([1, 32, 128, 64])]; + tensor var_1069_end_mask_0 = const()[name = tensor("op_1069_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1069_cast_fp16 = slice_by_index(begin = var_1069_begin_0, end = var_1069_end_0, end_mask = var_1069_end_mask_0, x = q_11_cast_fp16)[name = tensor("op_1069_cast_fp16")]; + tensor const_32_promoted_to_fp16 = const()[name = tensor("const_32_promoted_to_fp16"), val = tensor(-0x1p+0)]; + tensor var_1071_cast_fp16 = mul(x = var_1069_cast_fp16, y = const_32_promoted_to_fp16)[name = tensor("op_1071_cast_fp16")]; tensor rotated_5_interleave_0 = const()[name = tensor("rotated_5_interleave_0"), val = tensor(false)]; - tensor rotated_5_cast_fp16 = concat(axis = var_233, interleave = rotated_5_interleave_0, values = (var_316_cast_fp16, var_308_cast_fp16))[name = tensor("rotated_5_cast_fp16")]; - tensor var_319_cast_fp16 = mul(x = q_9_cast_fp16, y = cos)[name = tensor("op_319_cast_fp16")]; - tensor var_320_cast_fp16 = mul(x = rotated_5_cast_fp16, y = sin)[name = tensor("op_320_cast_fp16")]; - tensor roped_5_cast_fp16 = add(x = var_319_cast_fp16, y = var_320_cast_fp16)[name = tensor("roped_5_cast_fp16")]; - tensor var_333_begin_0 = const()[name = tensor("op_333_begin_0"), val = tensor([0, 0, 0, 0])]; - tensor var_333_end_0 = const()[name = tensor("op_333_end_0"), val = tensor([1, 32, 64, 64])]; - tensor var_333_end_mask_0 = const()[name = tensor("op_333_end_mask_0"), val = tensor([true, true, false, true])]; - tensor var_333_cast_fp16 = slice_by_index(begin = var_333_begin_0, end = var_333_end_0, end_mask = var_333_end_mask_0, x = k_11_cast_fp16)[name = tensor("op_333_cast_fp16")]; - tensor var_339_begin_0 = const()[name = tensor("op_339_begin_0"), val = tensor([0, 0, 64, 0])]; - tensor var_339_end_0 = const()[name = tensor("op_339_end_0"), val = tensor([1, 32, 128, 64])]; - tensor var_339_end_mask_0 = const()[name = tensor("op_339_end_mask_0"), val = tensor([true, true, true, true])]; - tensor var_339_cast_fp16 = slice_by_index(begin = var_339_begin_0, end = var_339_end_0, end_mask = var_339_end_mask_0, x = k_11_cast_fp16)[name = tensor("op_339_cast_fp16")]; - tensor const_12_promoted_to_fp16 = const()[name = tensor("const_12_promoted_to_fp16"), val = tensor(-0x1p+0)]; - tensor var_341_cast_fp16 = mul(x = var_339_cast_fp16, y = const_12_promoted_to_fp16)[name = tensor("op_341_cast_fp16")]; + tensor rotated_5_cast_fp16 = concat(axis = var_927, interleave = rotated_5_interleave_0, values = (var_1071_cast_fp16, var_1063_cast_fp16))[name = tensor("rotated_5_cast_fp16")]; + tensor var_1074_cast_fp16 = mul(x = q_11_cast_fp16, y = cos)[name = tensor("op_1074_cast_fp16")]; + tensor var_1075_cast_fp16 = mul(x = rotated_5_cast_fp16, y = sin)[name = tensor("op_1075_cast_fp16")]; + tensor roped_5_cast_fp16 = add(x = var_1074_cast_fp16, y = var_1075_cast_fp16)[name = tensor("roped_5_cast_fp16")]; + tensor var_1088_begin_0 = const()[name = tensor("op_1088_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1088_end_0 = const()[name = tensor("op_1088_end_0"), val = tensor([1, 32, 64, 64])]; + tensor var_1088_end_mask_0 = const()[name = tensor("op_1088_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1088_cast_fp16 = slice_by_index(begin = var_1088_begin_0, end = var_1088_end_0, end_mask = var_1088_end_mask_0, x = k_13_cast_fp16)[name = tensor("op_1088_cast_fp16")]; + tensor var_1094_begin_0 = const()[name = tensor("op_1094_begin_0"), val = tensor([0, 0, 64, 0])]; + tensor var_1094_end_0 = const()[name = tensor("op_1094_end_0"), val = tensor([1, 32, 128, 64])]; + tensor var_1094_end_mask_0 = const()[name = tensor("op_1094_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1094_cast_fp16 = slice_by_index(begin = var_1094_begin_0, end = var_1094_end_0, end_mask = var_1094_end_mask_0, x = k_13_cast_fp16)[name = tensor("op_1094_cast_fp16")]; + tensor const_34_promoted_to_fp16 = const()[name = tensor("const_34_promoted_to_fp16"), val = tensor(-0x1p+0)]; + tensor var_1096_cast_fp16 = mul(x = var_1094_cast_fp16, y = const_34_promoted_to_fp16)[name = tensor("op_1096_cast_fp16")]; tensor rotated_interleave_0 = const()[name = tensor("rotated_interleave_0"), val = tensor(false)]; - tensor rotated_cast_fp16 = concat(axis = var_233, interleave = rotated_interleave_0, values = (var_341_cast_fp16, var_333_cast_fp16))[name = tensor("rotated_cast_fp16")]; - tensor var_344_cast_fp16 = mul(x = k_11_cast_fp16, y = cos)[name = tensor("op_344_cast_fp16")]; - tensor var_345_cast_fp16 = mul(x = rotated_cast_fp16, y = sin)[name = tensor("op_345_cast_fp16")]; - tensor roped_cast_fp16 = add(x = var_344_cast_fp16, y = var_345_cast_fp16)[name = tensor("roped_cast_fp16")]; - tensor q_interleave_0 = const()[name = tensor("q_interleave_0"), val = tensor(false)]; - tensor q_cast_fp16 = concat(axis = var_233, interleave = q_interleave_0, values = roped_5_cast_fp16)[name = tensor("q_cast_fp16")]; - tensor k_13_interleave_0 = const()[name = tensor("k_13_interleave_0"), val = tensor(false)]; - tensor new_k_cache_1 = concat(axis = var_233, interleave = k_13_interleave_0, values = roped_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor rotated_cast_fp16 = concat(axis = var_927, interleave = rotated_interleave_0, values = (var_1096_cast_fp16, var_1088_cast_fp16))[name = tensor("rotated_cast_fp16")]; + tensor var_1099_cast_fp16 = mul(x = k_13_cast_fp16, y = cos)[name = tensor("op_1099_cast_fp16")]; + tensor var_1100_cast_fp16 = mul(x = rotated_cast_fp16, y = sin)[name = tensor("op_1100_cast_fp16")]; + tensor roped_cast_fp16 = add(x = var_1099_cast_fp16, y = var_1100_cast_fp16)[name = tensor("roped_cast_fp16")]; + tensor var_1103 = const()[name = tensor("op_1103"), val = tensor([1, 4096, 1, 64])]; + tensor var_1104_cast_fp16 = reshape(shape = var_1103, x = roped_cast_fp16)[name = tensor("op_1104_cast_fp16")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, -1, 2, -3])]; + tensor var_1106 = const()[name = tensor("op_1106"), val = tensor([1, 4096, 1, 64])]; + tensor new_v_cache_1 = reshape(shape = var_1106, x = v_11_cast_fp16)[name = tensor("new_v_cache_1_type_fp32_cast_fp16")]; tensor k_interleave_0 = const()[name = tensor("k_interleave_0"), val = tensor(false)]; - tensor k_cast_fp16 = concat(axis = var_235, interleave = k_interleave_0, values = (k_cache_1, new_k_cache_1))[name = tensor("k_cast_fp16")]; - tensor v_interleave_0 = const()[name = tensor("v_interleave_0"), val = tensor(false)]; - tensor v_cast_fp16 = concat(axis = var_235, interleave = v_interleave_0, values = (v_cache_1, new_v_cache_1))[name = tensor("v_cast_fp16")]; - tensor var_367_to_fp16 = const()[name = tensor("op_367_to_fp16"), val = tensor(0x1.6ap-4)]; - tensor var_368_cast_fp16 = mul(x = q_cast_fp16, y = var_367_to_fp16)[name = tensor("op_368_cast_fp16")]; - tensor attn_weights_5_transpose_x_0 = const()[name = tensor("attn_weights_5_transpose_x_0"), val = tensor(true)]; - tensor attn_weights_5_transpose_y_0 = const()[name = tensor("attn_weights_5_transpose_y_0"), val = tensor(false)]; - tensor attn_weights_5_cast_fp16 = matmul(transpose_x = attn_weights_5_transpose_x_0, transpose_y = attn_weights_5_transpose_y_0, x = var_368_cast_fp16, y = k_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; - tensor attn_weights_cast_fp16 = add(x = attn_weights_5_cast_fp16, y = mask)[name = tensor("attn_weights_cast_fp16")]; - tensor var_376_cast_fp16 = softmax(axis = var_228, x = attn_weights_cast_fp16)[name = tensor("op_376_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 = v_cast_fp16, y = var_376_cast_fp16)[name = tensor("attn_3_cast_fp16")]; - tensor var_380 = const()[name = tensor("op_380"), val = tensor([1, 4096, 1, -1])]; - tensor input_9_cast_fp16 = reshape(shape = var_380, x = attn_3_cast_fp16)[name = tensor("input_9_cast_fp16")]; - tensor var_384 = const()[name = tensor("op_384"), val = tensor([1, 1])]; - tensor var_386 = const()[name = tensor("op_386"), val = tensor([1, 1])]; - tensor var_388_pad_type_0 = const()[name = tensor("op_388_pad_type_0"), val = tensor("custom")]; - tensor var_388_pad_0 = const()[name = tensor("op_388_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor var_388_cast_fp16 = conv(dilations = var_386, groups = var_242, pad = var_388_pad_0, pad_type = var_388_pad_type_0, strides = var_384, weight = blocks_1_attn_proj_weight_palettized_cast_fp16, x = input_9_cast_fp16)[name = tensor("op_388_cast_fp16")]; - tensor blocks_1_attn_proj_output_scales_to_fp16 = const()[name = tensor("blocks_1_attn_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202512896)))]; - tensor attention_output_cast_fp16 = mul(x = var_388_cast_fp16, y = blocks_1_attn_proj_output_scales_to_fp16)[name = tensor("attention_output_cast_fp16")]; - tensor x_25_cast_fp16 = add(x = attention_output_cast_fp16, y = x_15_cast_fp16)[name = tensor("x_25_cast_fp16")]; - tensor var_397_cast_fp16 = mul(x = x_25_cast_fp16, y = x_25_cast_fp16)[name = tensor("op_397_cast_fp16")]; - tensor var_398 = const()[name = tensor("op_398"), val = tensor([1])]; - tensor norm_x_cast_fp16 = reduce_mean(axes = var_398, keep_dims = var_243, x = var_397_cast_fp16)[name = tensor("norm_x_cast_fp16")]; - tensor var_400_to_fp16 = const()[name = tensor("op_400_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_401_cast_fp16 = add(x = norm_x_cast_fp16, y = var_400_to_fp16)[name = tensor("op_401_cast_fp16")]; - tensor var_402_epsilon_0_to_fp16 = const()[name = tensor("op_402_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; - tensor var_402_cast_fp16 = rsqrt(epsilon = var_402_epsilon_0_to_fp16, x = var_401_cast_fp16)[name = tensor("op_402_cast_fp16")]; - tensor x_normed_13_cast_fp16 = mul(x = x_25_cast_fp16, y = var_402_cast_fp16)[name = tensor("x_normed_13_cast_fp16")]; - tensor blocks_1_norm_2_weight_to_fp16 = const()[name = tensor("blocks_1_norm_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202521152)))]; - tensor input_11_cast_fp16 = mul(x = x_normed_13_cast_fp16, y = blocks_1_norm_2_weight_to_fp16)[name = tensor("input_11_cast_fp16")]; - tensor var_414 = const()[name = tensor("op_414"), val = tensor([1, 1])]; - tensor var_416 = const()[name = tensor("op_416"), val = tensor([1, 1])]; - tensor var_418_pad_type_0 = const()[name = tensor("op_418_pad_type_0"), val = tensor("custom")]; - tensor var_418_pad_0 = const()[name = tensor("op_418_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor var_418_cast_fp16 = conv(dilations = var_416, groups = var_242, pad = var_418_pad_0, pad_type = var_418_pad_type_0, strides = var_414, weight = blocks_1_mlp_fc_1_weight_palettized_cast_fp16, x = input_11_cast_fp16)[name = tensor("op_418_cast_fp16")]; - tensor blocks_1_mlp_fc_1_output_scales_to_fp16 = const()[name = tensor("blocks_1_mlp_fc_1_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202529408)))]; - tensor input_13_cast_fp16 = mul(x = var_418_cast_fp16, y = blocks_1_mlp_fc_1_output_scales_to_fp16)[name = tensor("input_13_cast_fp16")]; - tensor var_422 = const()[name = tensor("op_422"), val = tensor([1, 1])]; - tensor var_424 = const()[name = tensor("op_424"), val = tensor([1, 1])]; - tensor var_426_pad_type_0 = const()[name = tensor("op_426_pad_type_0"), val = tensor("custom")]; - tensor var_426_pad_0 = const()[name = tensor("op_426_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor var_426_cast_fp16 = conv(dilations = var_424, groups = var_242, pad = var_426_pad_0, pad_type = var_426_pad_type_0, strides = var_422, weight = blocks_1_mlp_fc_2_weight_palettized_cast_fp16, x = input_11_cast_fp16)[name = tensor("op_426_cast_fp16")]; - tensor blocks_1_mlp_fc_2_output_scales_to_fp16 = const()[name = tensor("blocks_1_mlp_fc_2_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202551488)))]; - tensor x_fc_2_cast_fp16 = mul(x = var_426_cast_fp16, y = blocks_1_mlp_fc_2_output_scales_to_fp16)[name = tensor("x_fc_2_cast_fp16")]; - tensor var_428_cast_fp16 = silu(x = input_13_cast_fp16)[name = tensor("op_428_cast_fp16")]; - tensor input_cast_fp16 = mul(x = var_428_cast_fp16, y = x_fc_2_cast_fp16)[name = tensor("input_cast_fp16")]; - tensor var_432 = const()[name = tensor("op_432"), val = tensor([1, 1])]; - tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 1])]; - tensor var_436_pad_type_0 = const()[name = tensor("op_436_pad_type_0"), val = tensor("custom")]; - tensor var_436_pad_0 = const()[name = tensor("op_436_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor var_436_cast_fp16 = conv(dilations = var_434, groups = var_242, pad = var_436_pad_0, pad_type = var_436_pad_type_0, strides = var_432, weight = blocks_1_mlp_proj_weight_palettized_cast_fp16, x = input_cast_fp16)[name = tensor("op_436_cast_fp16")]; - tensor blocks_1_mlp_proj_output_scales_to_fp16 = const()[name = tensor("blocks_1_mlp_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202573568)))]; - tensor var_437_cast_fp16 = mul(x = var_436_cast_fp16, y = blocks_1_mlp_proj_output_scales_to_fp16)[name = tensor("op_437_cast_fp16")]; - tensor new_x = add(x = var_437_cast_fp16, y = x_25_cast_fp16)[name = tensor("op_438_cast_fp16")]; + tensor new_k_cache_1 = transpose(perm = k_17_perm_0, x = var_1104_cast_fp16)[name = tensor("transpose_0")]; + tensor k_cast_fp16 = concat(axis = var_929, interleave = k_interleave_0, values = (k_cache_1, new_k_cache_1))[name = tensor("k_cast_fp16")]; + tensor v_17_interleave_0 = const()[name = tensor("v_17_interleave_0"), val = tensor(false)]; + tensor v_17_cast_fp16 = concat(axis = var_923, interleave = v_17_interleave_0, values = (v_cache_1, new_v_cache_1))[name = tensor("v_17_cast_fp16")]; + tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([1, 4096, 1, -1])]; + tensor q_cast_fp16 = reshape(shape = var_1113, x = roped_5_cast_fp16)[name = tensor("q_cast_fp16")]; + tensor var_1118_begin_0 = const()[name = tensor("op_1118_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1118_end_0 = const()[name = tensor("op_1118_end_0"), val = tensor([1, 128, 1, 64])]; + tensor var_1118_end_mask_0 = const()[name = tensor("op_1118_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1118_cast_fp16 = slice_by_index(begin = var_1118_begin_0, end = var_1118_end_0, end_mask = var_1118_end_mask_0, x = q_cast_fp16)[name = tensor("op_1118_cast_fp16")]; + tensor var_1122_begin_0 = const()[name = tensor("op_1122_begin_0"), val = tensor([0, 128, 0, 0])]; + tensor var_1122_end_0 = const()[name = tensor("op_1122_end_0"), val = tensor([1, 256, 1, 64])]; + tensor var_1122_end_mask_0 = const()[name = tensor("op_1122_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1122_cast_fp16 = slice_by_index(begin = var_1122_begin_0, end = var_1122_end_0, end_mask = var_1122_end_mask_0, x = q_cast_fp16)[name = tensor("op_1122_cast_fp16")]; + tensor var_1126_begin_0 = const()[name = tensor("op_1126_begin_0"), val = tensor([0, 256, 0, 0])]; + tensor var_1126_end_0 = const()[name = tensor("op_1126_end_0"), val = tensor([1, 384, 1, 64])]; + tensor var_1126_end_mask_0 = const()[name = tensor("op_1126_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1126_cast_fp16 = slice_by_index(begin = var_1126_begin_0, end = var_1126_end_0, end_mask = var_1126_end_mask_0, x = q_cast_fp16)[name = tensor("op_1126_cast_fp16")]; + tensor var_1130_begin_0 = const()[name = tensor("op_1130_begin_0"), val = tensor([0, 384, 0, 0])]; + tensor var_1130_end_0 = const()[name = tensor("op_1130_end_0"), val = tensor([1, 512, 1, 64])]; + tensor var_1130_end_mask_0 = const()[name = tensor("op_1130_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1130_cast_fp16 = slice_by_index(begin = var_1130_begin_0, end = var_1130_end_0, end_mask = var_1130_end_mask_0, x = q_cast_fp16)[name = tensor("op_1130_cast_fp16")]; + tensor var_1134_begin_0 = const()[name = tensor("op_1134_begin_0"), val = tensor([0, 512, 0, 0])]; + tensor var_1134_end_0 = const()[name = tensor("op_1134_end_0"), val = tensor([1, 640, 1, 64])]; + tensor var_1134_end_mask_0 = const()[name = tensor("op_1134_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1134_cast_fp16 = slice_by_index(begin = var_1134_begin_0, end = var_1134_end_0, end_mask = var_1134_end_mask_0, x = q_cast_fp16)[name = tensor("op_1134_cast_fp16")]; + tensor var_1138_begin_0 = const()[name = tensor("op_1138_begin_0"), val = tensor([0, 640, 0, 0])]; + tensor var_1138_end_0 = const()[name = tensor("op_1138_end_0"), val = tensor([1, 768, 1, 64])]; + tensor var_1138_end_mask_0 = const()[name = tensor("op_1138_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1138_cast_fp16 = slice_by_index(begin = var_1138_begin_0, end = var_1138_end_0, end_mask = var_1138_end_mask_0, x = q_cast_fp16)[name = tensor("op_1138_cast_fp16")]; + tensor var_1142_begin_0 = const()[name = tensor("op_1142_begin_0"), val = tensor([0, 768, 0, 0])]; + tensor var_1142_end_0 = const()[name = tensor("op_1142_end_0"), val = tensor([1, 896, 1, 64])]; + tensor var_1142_end_mask_0 = const()[name = tensor("op_1142_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1142_cast_fp16 = slice_by_index(begin = var_1142_begin_0, end = var_1142_end_0, end_mask = var_1142_end_mask_0, x = q_cast_fp16)[name = tensor("op_1142_cast_fp16")]; + tensor var_1146_begin_0 = const()[name = tensor("op_1146_begin_0"), val = tensor([0, 896, 0, 0])]; + tensor var_1146_end_0 = const()[name = tensor("op_1146_end_0"), val = tensor([1, 1024, 1, 64])]; + tensor var_1146_end_mask_0 = const()[name = tensor("op_1146_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1146_cast_fp16 = slice_by_index(begin = var_1146_begin_0, end = var_1146_end_0, end_mask = var_1146_end_mask_0, x = q_cast_fp16)[name = tensor("op_1146_cast_fp16")]; + tensor var_1150_begin_0 = const()[name = tensor("op_1150_begin_0"), val = tensor([0, 1024, 0, 0])]; + tensor var_1150_end_0 = const()[name = tensor("op_1150_end_0"), val = tensor([1, 1152, 1, 64])]; + tensor var_1150_end_mask_0 = const()[name = tensor("op_1150_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1150_cast_fp16 = slice_by_index(begin = var_1150_begin_0, end = var_1150_end_0, end_mask = var_1150_end_mask_0, x = q_cast_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor var_1154_begin_0 = const()[name = tensor("op_1154_begin_0"), val = tensor([0, 1152, 0, 0])]; + tensor var_1154_end_0 = const()[name = tensor("op_1154_end_0"), val = tensor([1, 1280, 1, 64])]; + tensor var_1154_end_mask_0 = const()[name = tensor("op_1154_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1154_cast_fp16 = slice_by_index(begin = var_1154_begin_0, end = var_1154_end_0, end_mask = var_1154_end_mask_0, x = q_cast_fp16)[name = tensor("op_1154_cast_fp16")]; + tensor var_1158_begin_0 = const()[name = tensor("op_1158_begin_0"), val = tensor([0, 1280, 0, 0])]; + tensor var_1158_end_0 = const()[name = tensor("op_1158_end_0"), val = tensor([1, 1408, 1, 64])]; + tensor var_1158_end_mask_0 = const()[name = tensor("op_1158_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1158_cast_fp16 = slice_by_index(begin = var_1158_begin_0, end = var_1158_end_0, end_mask = var_1158_end_mask_0, x = q_cast_fp16)[name = tensor("op_1158_cast_fp16")]; + tensor var_1162_begin_0 = const()[name = tensor("op_1162_begin_0"), val = tensor([0, 1408, 0, 0])]; + tensor var_1162_end_0 = const()[name = tensor("op_1162_end_0"), val = tensor([1, 1536, 1, 64])]; + tensor var_1162_end_mask_0 = const()[name = tensor("op_1162_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1162_cast_fp16 = slice_by_index(begin = var_1162_begin_0, end = var_1162_end_0, end_mask = var_1162_end_mask_0, x = q_cast_fp16)[name = tensor("op_1162_cast_fp16")]; + tensor var_1166_begin_0 = const()[name = tensor("op_1166_begin_0"), val = tensor([0, 1536, 0, 0])]; + tensor var_1166_end_0 = const()[name = tensor("op_1166_end_0"), val = tensor([1, 1664, 1, 64])]; + tensor var_1166_end_mask_0 = const()[name = tensor("op_1166_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1166_cast_fp16 = slice_by_index(begin = var_1166_begin_0, end = var_1166_end_0, end_mask = var_1166_end_mask_0, x = q_cast_fp16)[name = tensor("op_1166_cast_fp16")]; + tensor var_1170_begin_0 = const()[name = tensor("op_1170_begin_0"), val = tensor([0, 1664, 0, 0])]; + tensor var_1170_end_0 = const()[name = tensor("op_1170_end_0"), val = tensor([1, 1792, 1, 64])]; + tensor var_1170_end_mask_0 = const()[name = tensor("op_1170_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1170_cast_fp16 = slice_by_index(begin = var_1170_begin_0, end = var_1170_end_0, end_mask = var_1170_end_mask_0, x = q_cast_fp16)[name = tensor("op_1170_cast_fp16")]; + tensor var_1174_begin_0 = const()[name = tensor("op_1174_begin_0"), val = tensor([0, 1792, 0, 0])]; + tensor var_1174_end_0 = const()[name = tensor("op_1174_end_0"), val = tensor([1, 1920, 1, 64])]; + tensor var_1174_end_mask_0 = const()[name = tensor("op_1174_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1174_cast_fp16 = slice_by_index(begin = var_1174_begin_0, end = var_1174_end_0, end_mask = var_1174_end_mask_0, x = q_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor var_1178_begin_0 = const()[name = tensor("op_1178_begin_0"), val = tensor([0, 1920, 0, 0])]; + tensor var_1178_end_0 = const()[name = tensor("op_1178_end_0"), val = tensor([1, 2048, 1, 64])]; + tensor var_1178_end_mask_0 = const()[name = tensor("op_1178_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1178_cast_fp16 = slice_by_index(begin = var_1178_begin_0, end = var_1178_end_0, end_mask = var_1178_end_mask_0, x = q_cast_fp16)[name = tensor("op_1178_cast_fp16")]; + tensor var_1182_begin_0 = const()[name = tensor("op_1182_begin_0"), val = tensor([0, 2048, 0, 0])]; + tensor var_1182_end_0 = const()[name = tensor("op_1182_end_0"), val = tensor([1, 2176, 1, 64])]; + tensor var_1182_end_mask_0 = const()[name = tensor("op_1182_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1182_cast_fp16 = slice_by_index(begin = var_1182_begin_0, end = var_1182_end_0, end_mask = var_1182_end_mask_0, x = q_cast_fp16)[name = tensor("op_1182_cast_fp16")]; + tensor var_1186_begin_0 = const()[name = tensor("op_1186_begin_0"), val = tensor([0, 2176, 0, 0])]; + tensor var_1186_end_0 = const()[name = tensor("op_1186_end_0"), val = tensor([1, 2304, 1, 64])]; + tensor var_1186_end_mask_0 = const()[name = tensor("op_1186_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1186_cast_fp16 = slice_by_index(begin = var_1186_begin_0, end = var_1186_end_0, end_mask = var_1186_end_mask_0, x = q_cast_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor var_1190_begin_0 = const()[name = tensor("op_1190_begin_0"), val = tensor([0, 2304, 0, 0])]; + tensor var_1190_end_0 = const()[name = tensor("op_1190_end_0"), val = tensor([1, 2432, 1, 64])]; + tensor var_1190_end_mask_0 = const()[name = tensor("op_1190_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1190_cast_fp16 = slice_by_index(begin = var_1190_begin_0, end = var_1190_end_0, end_mask = var_1190_end_mask_0, x = q_cast_fp16)[name = tensor("op_1190_cast_fp16")]; + tensor var_1194_begin_0 = const()[name = tensor("op_1194_begin_0"), val = tensor([0, 2432, 0, 0])]; + tensor var_1194_end_0 = const()[name = tensor("op_1194_end_0"), val = tensor([1, 2560, 1, 64])]; + tensor var_1194_end_mask_0 = const()[name = tensor("op_1194_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1194_cast_fp16 = slice_by_index(begin = var_1194_begin_0, end = var_1194_end_0, end_mask = var_1194_end_mask_0, x = q_cast_fp16)[name = tensor("op_1194_cast_fp16")]; + tensor var_1198_begin_0 = const()[name = tensor("op_1198_begin_0"), val = tensor([0, 2560, 0, 0])]; + tensor var_1198_end_0 = const()[name = tensor("op_1198_end_0"), val = tensor([1, 2688, 1, 64])]; + tensor var_1198_end_mask_0 = const()[name = tensor("op_1198_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1198_cast_fp16 = slice_by_index(begin = var_1198_begin_0, end = var_1198_end_0, end_mask = var_1198_end_mask_0, x = q_cast_fp16)[name = tensor("op_1198_cast_fp16")]; + tensor var_1202_begin_0 = const()[name = tensor("op_1202_begin_0"), val = tensor([0, 2688, 0, 0])]; + tensor var_1202_end_0 = const()[name = tensor("op_1202_end_0"), val = tensor([1, 2816, 1, 64])]; + tensor var_1202_end_mask_0 = const()[name = tensor("op_1202_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1202_cast_fp16 = slice_by_index(begin = var_1202_begin_0, end = var_1202_end_0, end_mask = var_1202_end_mask_0, x = q_cast_fp16)[name = tensor("op_1202_cast_fp16")]; + tensor var_1206_begin_0 = const()[name = tensor("op_1206_begin_0"), val = tensor([0, 2816, 0, 0])]; + tensor var_1206_end_0 = const()[name = tensor("op_1206_end_0"), val = tensor([1, 2944, 1, 64])]; + tensor var_1206_end_mask_0 = const()[name = tensor("op_1206_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1206_cast_fp16 = slice_by_index(begin = var_1206_begin_0, end = var_1206_end_0, end_mask = var_1206_end_mask_0, x = q_cast_fp16)[name = tensor("op_1206_cast_fp16")]; + tensor var_1210_begin_0 = const()[name = tensor("op_1210_begin_0"), val = tensor([0, 2944, 0, 0])]; + tensor var_1210_end_0 = const()[name = tensor("op_1210_end_0"), val = tensor([1, 3072, 1, 64])]; + tensor var_1210_end_mask_0 = const()[name = tensor("op_1210_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1210_cast_fp16 = slice_by_index(begin = var_1210_begin_0, end = var_1210_end_0, end_mask = var_1210_end_mask_0, x = q_cast_fp16)[name = tensor("op_1210_cast_fp16")]; + tensor var_1214_begin_0 = const()[name = tensor("op_1214_begin_0"), val = tensor([0, 3072, 0, 0])]; + tensor var_1214_end_0 = const()[name = tensor("op_1214_end_0"), val = tensor([1, 3200, 1, 64])]; + tensor var_1214_end_mask_0 = const()[name = tensor("op_1214_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1214_cast_fp16 = slice_by_index(begin = var_1214_begin_0, end = var_1214_end_0, end_mask = var_1214_end_mask_0, x = q_cast_fp16)[name = tensor("op_1214_cast_fp16")]; + tensor var_1218_begin_0 = const()[name = tensor("op_1218_begin_0"), val = tensor([0, 3200, 0, 0])]; + tensor var_1218_end_0 = const()[name = tensor("op_1218_end_0"), val = tensor([1, 3328, 1, 64])]; + tensor var_1218_end_mask_0 = const()[name = tensor("op_1218_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1218_cast_fp16 = slice_by_index(begin = var_1218_begin_0, end = var_1218_end_0, end_mask = var_1218_end_mask_0, x = q_cast_fp16)[name = tensor("op_1218_cast_fp16")]; + tensor var_1222_begin_0 = const()[name = tensor("op_1222_begin_0"), val = tensor([0, 3328, 0, 0])]; + tensor var_1222_end_0 = const()[name = tensor("op_1222_end_0"), val = tensor([1, 3456, 1, 64])]; + tensor var_1222_end_mask_0 = const()[name = tensor("op_1222_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1222_cast_fp16 = slice_by_index(begin = var_1222_begin_0, end = var_1222_end_0, end_mask = var_1222_end_mask_0, x = q_cast_fp16)[name = tensor("op_1222_cast_fp16")]; + tensor var_1226_begin_0 = const()[name = tensor("op_1226_begin_0"), val = tensor([0, 3456, 0, 0])]; + tensor var_1226_end_0 = const()[name = tensor("op_1226_end_0"), val = tensor([1, 3584, 1, 64])]; + tensor var_1226_end_mask_0 = const()[name = tensor("op_1226_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1226_cast_fp16 = slice_by_index(begin = var_1226_begin_0, end = var_1226_end_0, end_mask = var_1226_end_mask_0, x = q_cast_fp16)[name = tensor("op_1226_cast_fp16")]; + tensor var_1230_begin_0 = const()[name = tensor("op_1230_begin_0"), val = tensor([0, 3584, 0, 0])]; + tensor var_1230_end_0 = const()[name = tensor("op_1230_end_0"), val = tensor([1, 3712, 1, 64])]; + tensor var_1230_end_mask_0 = const()[name = tensor("op_1230_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1230_cast_fp16 = slice_by_index(begin = var_1230_begin_0, end = var_1230_end_0, end_mask = var_1230_end_mask_0, x = q_cast_fp16)[name = tensor("op_1230_cast_fp16")]; + tensor var_1234_begin_0 = const()[name = tensor("op_1234_begin_0"), val = tensor([0, 3712, 0, 0])]; + tensor var_1234_end_0 = const()[name = tensor("op_1234_end_0"), val = tensor([1, 3840, 1, 64])]; + tensor var_1234_end_mask_0 = const()[name = tensor("op_1234_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1234_cast_fp16 = slice_by_index(begin = var_1234_begin_0, end = var_1234_end_0, end_mask = var_1234_end_mask_0, x = q_cast_fp16)[name = tensor("op_1234_cast_fp16")]; + tensor var_1238_begin_0 = const()[name = tensor("op_1238_begin_0"), val = tensor([0, 3840, 0, 0])]; + tensor var_1238_end_0 = const()[name = tensor("op_1238_end_0"), val = tensor([1, 3968, 1, 64])]; + tensor var_1238_end_mask_0 = const()[name = tensor("op_1238_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1238_cast_fp16 = slice_by_index(begin = var_1238_begin_0, end = var_1238_end_0, end_mask = var_1238_end_mask_0, x = q_cast_fp16)[name = tensor("op_1238_cast_fp16")]; + tensor var_1242_begin_0 = const()[name = tensor("op_1242_begin_0"), val = tensor([0, 3968, 0, 0])]; + tensor var_1242_end_0 = const()[name = tensor("op_1242_end_0"), val = tensor([1, 4096, 1, 64])]; + tensor var_1242_end_mask_0 = const()[name = tensor("op_1242_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1242_cast_fp16 = slice_by_index(begin = var_1242_begin_0, end = var_1242_end_0, end_mask = var_1242_end_mask_0, x = q_cast_fp16)[name = tensor("op_1242_cast_fp16")]; + tensor var_1248_begin_0 = const()[name = tensor("op_1248_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1248_end_0 = const()[name = tensor("op_1248_end_0"), val = tensor([1, 512, 1, 128])]; + tensor var_1248_end_mask_0 = const()[name = tensor("op_1248_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1248_cast_fp16 = slice_by_index(begin = var_1248_begin_0, end = var_1248_end_0, end_mask = var_1248_end_mask_0, x = k_cast_fp16)[name = tensor("op_1248_cast_fp16")]; + tensor var_1252_begin_0 = const()[name = tensor("op_1252_begin_0"), val = tensor([0, 0, 0, 128])]; + tensor var_1252_end_0 = const()[name = tensor("op_1252_end_0"), val = tensor([1, 512, 1, 256])]; + tensor var_1252_end_mask_0 = const()[name = tensor("op_1252_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1252_cast_fp16 = slice_by_index(begin = var_1252_begin_0, end = var_1252_end_0, end_mask = var_1252_end_mask_0, x = k_cast_fp16)[name = tensor("op_1252_cast_fp16")]; + tensor var_1256_begin_0 = const()[name = tensor("op_1256_begin_0"), val = tensor([0, 0, 0, 256])]; + tensor var_1256_end_0 = const()[name = tensor("op_1256_end_0"), val = tensor([1, 512, 1, 384])]; + tensor var_1256_end_mask_0 = const()[name = tensor("op_1256_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1256_cast_fp16 = slice_by_index(begin = var_1256_begin_0, end = var_1256_end_0, end_mask = var_1256_end_mask_0, x = k_cast_fp16)[name = tensor("op_1256_cast_fp16")]; + tensor var_1260_begin_0 = const()[name = tensor("op_1260_begin_0"), val = tensor([0, 0, 0, 384])]; + tensor var_1260_end_0 = const()[name = tensor("op_1260_end_0"), val = tensor([1, 512, 1, 512])]; + tensor var_1260_end_mask_0 = const()[name = tensor("op_1260_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1260_cast_fp16 = slice_by_index(begin = var_1260_begin_0, end = var_1260_end_0, end_mask = var_1260_end_mask_0, x = k_cast_fp16)[name = tensor("op_1260_cast_fp16")]; + tensor var_1264_begin_0 = const()[name = tensor("op_1264_begin_0"), val = tensor([0, 0, 0, 512])]; + tensor var_1264_end_0 = const()[name = tensor("op_1264_end_0"), val = tensor([1, 512, 1, 640])]; + tensor var_1264_end_mask_0 = const()[name = tensor("op_1264_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1264_cast_fp16 = slice_by_index(begin = var_1264_begin_0, end = var_1264_end_0, end_mask = var_1264_end_mask_0, x = k_cast_fp16)[name = tensor("op_1264_cast_fp16")]; + tensor var_1268_begin_0 = const()[name = tensor("op_1268_begin_0"), val = tensor([0, 0, 0, 640])]; + tensor var_1268_end_0 = const()[name = tensor("op_1268_end_0"), val = tensor([1, 512, 1, 768])]; + tensor var_1268_end_mask_0 = const()[name = tensor("op_1268_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1268_cast_fp16 = slice_by_index(begin = var_1268_begin_0, end = var_1268_end_0, end_mask = var_1268_end_mask_0, x = k_cast_fp16)[name = tensor("op_1268_cast_fp16")]; + tensor var_1272_begin_0 = const()[name = tensor("op_1272_begin_0"), val = tensor([0, 0, 0, 768])]; + tensor var_1272_end_0 = const()[name = tensor("op_1272_end_0"), val = tensor([1, 512, 1, 896])]; + tensor var_1272_end_mask_0 = const()[name = tensor("op_1272_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1272_cast_fp16 = slice_by_index(begin = var_1272_begin_0, end = var_1272_end_0, end_mask = var_1272_end_mask_0, x = k_cast_fp16)[name = tensor("op_1272_cast_fp16")]; + tensor var_1276_begin_0 = const()[name = tensor("op_1276_begin_0"), val = tensor([0, 0, 0, 896])]; + tensor var_1276_end_0 = const()[name = tensor("op_1276_end_0"), val = tensor([1, 512, 1, 1024])]; + tensor var_1276_end_mask_0 = const()[name = tensor("op_1276_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1276_cast_fp16 = slice_by_index(begin = var_1276_begin_0, end = var_1276_end_0, end_mask = var_1276_end_mask_0, x = k_cast_fp16)[name = tensor("op_1276_cast_fp16")]; + tensor var_1280_begin_0 = const()[name = tensor("op_1280_begin_0"), val = tensor([0, 0, 0, 1024])]; + tensor var_1280_end_0 = const()[name = tensor("op_1280_end_0"), val = tensor([1, 512, 1, 1152])]; + tensor var_1280_end_mask_0 = const()[name = tensor("op_1280_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1280_cast_fp16 = slice_by_index(begin = var_1280_begin_0, end = var_1280_end_0, end_mask = var_1280_end_mask_0, x = k_cast_fp16)[name = tensor("op_1280_cast_fp16")]; + tensor var_1284_begin_0 = const()[name = tensor("op_1284_begin_0"), val = tensor([0, 0, 0, 1152])]; + tensor var_1284_end_0 = const()[name = tensor("op_1284_end_0"), val = tensor([1, 512, 1, 1280])]; + tensor var_1284_end_mask_0 = const()[name = tensor("op_1284_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1284_cast_fp16 = slice_by_index(begin = var_1284_begin_0, end = var_1284_end_0, end_mask = var_1284_end_mask_0, x = k_cast_fp16)[name = tensor("op_1284_cast_fp16")]; + tensor var_1288_begin_0 = const()[name = tensor("op_1288_begin_0"), val = tensor([0, 0, 0, 1280])]; + tensor var_1288_end_0 = const()[name = tensor("op_1288_end_0"), val = tensor([1, 512, 1, 1408])]; + tensor var_1288_end_mask_0 = const()[name = tensor("op_1288_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1288_cast_fp16 = slice_by_index(begin = var_1288_begin_0, end = var_1288_end_0, end_mask = var_1288_end_mask_0, x = k_cast_fp16)[name = tensor("op_1288_cast_fp16")]; + tensor var_1292_begin_0 = const()[name = tensor("op_1292_begin_0"), val = tensor([0, 0, 0, 1408])]; + tensor var_1292_end_0 = const()[name = tensor("op_1292_end_0"), val = tensor([1, 512, 1, 1536])]; + tensor var_1292_end_mask_0 = const()[name = tensor("op_1292_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1292_cast_fp16 = slice_by_index(begin = var_1292_begin_0, end = var_1292_end_0, end_mask = var_1292_end_mask_0, x = k_cast_fp16)[name = tensor("op_1292_cast_fp16")]; + tensor var_1296_begin_0 = const()[name = tensor("op_1296_begin_0"), val = tensor([0, 0, 0, 1536])]; + tensor var_1296_end_0 = const()[name = tensor("op_1296_end_0"), val = tensor([1, 512, 1, 1664])]; + tensor var_1296_end_mask_0 = const()[name = tensor("op_1296_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1296_cast_fp16 = slice_by_index(begin = var_1296_begin_0, end = var_1296_end_0, end_mask = var_1296_end_mask_0, x = k_cast_fp16)[name = tensor("op_1296_cast_fp16")]; + tensor var_1300_begin_0 = const()[name = tensor("op_1300_begin_0"), val = tensor([0, 0, 0, 1664])]; + tensor var_1300_end_0 = const()[name = tensor("op_1300_end_0"), val = tensor([1, 512, 1, 1792])]; + tensor var_1300_end_mask_0 = const()[name = tensor("op_1300_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1300_cast_fp16 = slice_by_index(begin = var_1300_begin_0, end = var_1300_end_0, end_mask = var_1300_end_mask_0, x = k_cast_fp16)[name = tensor("op_1300_cast_fp16")]; + tensor var_1304_begin_0 = const()[name = tensor("op_1304_begin_0"), val = tensor([0, 0, 0, 1792])]; + tensor var_1304_end_0 = const()[name = tensor("op_1304_end_0"), val = tensor([1, 512, 1, 1920])]; + tensor var_1304_end_mask_0 = const()[name = tensor("op_1304_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1304_cast_fp16 = slice_by_index(begin = var_1304_begin_0, end = var_1304_end_0, end_mask = var_1304_end_mask_0, x = k_cast_fp16)[name = tensor("op_1304_cast_fp16")]; + tensor var_1308_begin_0 = const()[name = tensor("op_1308_begin_0"), val = tensor([0, 0, 0, 1920])]; + tensor var_1308_end_0 = const()[name = tensor("op_1308_end_0"), val = tensor([1, 512, 1, 2048])]; + tensor var_1308_end_mask_0 = const()[name = tensor("op_1308_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1308_cast_fp16 = slice_by_index(begin = var_1308_begin_0, end = var_1308_end_0, end_mask = var_1308_end_mask_0, x = k_cast_fp16)[name = tensor("op_1308_cast_fp16")]; + tensor var_1312_begin_0 = const()[name = tensor("op_1312_begin_0"), val = tensor([0, 0, 0, 2048])]; + tensor var_1312_end_0 = const()[name = tensor("op_1312_end_0"), val = tensor([1, 512, 1, 2176])]; + tensor var_1312_end_mask_0 = const()[name = tensor("op_1312_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1312_cast_fp16 = slice_by_index(begin = var_1312_begin_0, end = var_1312_end_0, end_mask = var_1312_end_mask_0, x = k_cast_fp16)[name = tensor("op_1312_cast_fp16")]; + tensor var_1316_begin_0 = const()[name = tensor("op_1316_begin_0"), val = tensor([0, 0, 0, 2176])]; + tensor var_1316_end_0 = const()[name = tensor("op_1316_end_0"), val = tensor([1, 512, 1, 2304])]; + tensor var_1316_end_mask_0 = const()[name = tensor("op_1316_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1316_cast_fp16 = slice_by_index(begin = var_1316_begin_0, end = var_1316_end_0, end_mask = var_1316_end_mask_0, x = k_cast_fp16)[name = tensor("op_1316_cast_fp16")]; + tensor var_1320_begin_0 = const()[name = tensor("op_1320_begin_0"), val = tensor([0, 0, 0, 2304])]; + tensor var_1320_end_0 = const()[name = tensor("op_1320_end_0"), val = tensor([1, 512, 1, 2432])]; + tensor var_1320_end_mask_0 = const()[name = tensor("op_1320_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1320_cast_fp16 = slice_by_index(begin = var_1320_begin_0, end = var_1320_end_0, end_mask = var_1320_end_mask_0, x = k_cast_fp16)[name = tensor("op_1320_cast_fp16")]; + tensor var_1324_begin_0 = const()[name = tensor("op_1324_begin_0"), val = tensor([0, 0, 0, 2432])]; + tensor var_1324_end_0 = const()[name = tensor("op_1324_end_0"), val = tensor([1, 512, 1, 2560])]; + tensor var_1324_end_mask_0 = const()[name = tensor("op_1324_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1324_cast_fp16 = slice_by_index(begin = var_1324_begin_0, end = var_1324_end_0, end_mask = var_1324_end_mask_0, x = k_cast_fp16)[name = tensor("op_1324_cast_fp16")]; + tensor var_1328_begin_0 = const()[name = tensor("op_1328_begin_0"), val = tensor([0, 0, 0, 2560])]; + tensor var_1328_end_0 = const()[name = tensor("op_1328_end_0"), val = tensor([1, 512, 1, 2688])]; + tensor var_1328_end_mask_0 = const()[name = tensor("op_1328_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1328_cast_fp16 = slice_by_index(begin = var_1328_begin_0, end = var_1328_end_0, end_mask = var_1328_end_mask_0, x = k_cast_fp16)[name = tensor("op_1328_cast_fp16")]; + tensor var_1332_begin_0 = const()[name = tensor("op_1332_begin_0"), val = tensor([0, 0, 0, 2688])]; + tensor var_1332_end_0 = const()[name = tensor("op_1332_end_0"), val = tensor([1, 512, 1, 2816])]; + tensor var_1332_end_mask_0 = const()[name = tensor("op_1332_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1332_cast_fp16 = slice_by_index(begin = var_1332_begin_0, end = var_1332_end_0, end_mask = var_1332_end_mask_0, x = k_cast_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1336_begin_0 = const()[name = tensor("op_1336_begin_0"), val = tensor([0, 0, 0, 2816])]; + tensor var_1336_end_0 = const()[name = tensor("op_1336_end_0"), val = tensor([1, 512, 1, 2944])]; + tensor var_1336_end_mask_0 = const()[name = tensor("op_1336_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1336_cast_fp16 = slice_by_index(begin = var_1336_begin_0, end = var_1336_end_0, end_mask = var_1336_end_mask_0, x = k_cast_fp16)[name = tensor("op_1336_cast_fp16")]; + tensor var_1340_begin_0 = const()[name = tensor("op_1340_begin_0"), val = tensor([0, 0, 0, 2944])]; + tensor var_1340_end_0 = const()[name = tensor("op_1340_end_0"), val = tensor([1, 512, 1, 3072])]; + tensor var_1340_end_mask_0 = const()[name = tensor("op_1340_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1340_cast_fp16 = slice_by_index(begin = var_1340_begin_0, end = var_1340_end_0, end_mask = var_1340_end_mask_0, x = k_cast_fp16)[name = tensor("op_1340_cast_fp16")]; + tensor var_1344_begin_0 = const()[name = tensor("op_1344_begin_0"), val = tensor([0, 0, 0, 3072])]; + tensor var_1344_end_0 = const()[name = tensor("op_1344_end_0"), val = tensor([1, 512, 1, 3200])]; + tensor var_1344_end_mask_0 = const()[name = tensor("op_1344_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1344_cast_fp16 = slice_by_index(begin = var_1344_begin_0, end = var_1344_end_0, end_mask = var_1344_end_mask_0, x = k_cast_fp16)[name = tensor("op_1344_cast_fp16")]; + tensor var_1348_begin_0 = const()[name = tensor("op_1348_begin_0"), val = tensor([0, 0, 0, 3200])]; + tensor var_1348_end_0 = const()[name = tensor("op_1348_end_0"), val = tensor([1, 512, 1, 3328])]; + tensor var_1348_end_mask_0 = const()[name = tensor("op_1348_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1348_cast_fp16 = slice_by_index(begin = var_1348_begin_0, end = var_1348_end_0, end_mask = var_1348_end_mask_0, x = k_cast_fp16)[name = tensor("op_1348_cast_fp16")]; + tensor var_1352_begin_0 = const()[name = tensor("op_1352_begin_0"), val = tensor([0, 0, 0, 3328])]; + tensor var_1352_end_0 = const()[name = tensor("op_1352_end_0"), val = tensor([1, 512, 1, 3456])]; + tensor var_1352_end_mask_0 = const()[name = tensor("op_1352_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1352_cast_fp16 = slice_by_index(begin = var_1352_begin_0, end = var_1352_end_0, end_mask = var_1352_end_mask_0, x = k_cast_fp16)[name = tensor("op_1352_cast_fp16")]; + tensor var_1356_begin_0 = const()[name = tensor("op_1356_begin_0"), val = tensor([0, 0, 0, 3456])]; + tensor var_1356_end_0 = const()[name = tensor("op_1356_end_0"), val = tensor([1, 512, 1, 3584])]; + tensor var_1356_end_mask_0 = const()[name = tensor("op_1356_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1356_cast_fp16 = slice_by_index(begin = var_1356_begin_0, end = var_1356_end_0, end_mask = var_1356_end_mask_0, x = k_cast_fp16)[name = tensor("op_1356_cast_fp16")]; + tensor var_1360_begin_0 = const()[name = tensor("op_1360_begin_0"), val = tensor([0, 0, 0, 3584])]; + tensor var_1360_end_0 = const()[name = tensor("op_1360_end_0"), val = tensor([1, 512, 1, 3712])]; + tensor var_1360_end_mask_0 = const()[name = tensor("op_1360_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1360_cast_fp16 = slice_by_index(begin = var_1360_begin_0, end = var_1360_end_0, end_mask = var_1360_end_mask_0, x = k_cast_fp16)[name = tensor("op_1360_cast_fp16")]; + tensor var_1364_begin_0 = const()[name = tensor("op_1364_begin_0"), val = tensor([0, 0, 0, 3712])]; + tensor var_1364_end_0 = const()[name = tensor("op_1364_end_0"), val = tensor([1, 512, 1, 3840])]; + tensor var_1364_end_mask_0 = const()[name = tensor("op_1364_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1364_cast_fp16 = slice_by_index(begin = var_1364_begin_0, end = var_1364_end_0, end_mask = var_1364_end_mask_0, x = k_cast_fp16)[name = tensor("op_1364_cast_fp16")]; + tensor var_1368_begin_0 = const()[name = tensor("op_1368_begin_0"), val = tensor([0, 0, 0, 3840])]; + tensor var_1368_end_0 = const()[name = tensor("op_1368_end_0"), val = tensor([1, 512, 1, 3968])]; + tensor var_1368_end_mask_0 = const()[name = tensor("op_1368_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1368_cast_fp16 = slice_by_index(begin = var_1368_begin_0, end = var_1368_end_0, end_mask = var_1368_end_mask_0, x = k_cast_fp16)[name = tensor("op_1368_cast_fp16")]; + tensor var_1372_begin_0 = const()[name = tensor("op_1372_begin_0"), val = tensor([0, 0, 0, 3968])]; + tensor var_1372_end_0 = const()[name = tensor("op_1372_end_0"), val = tensor([1, 512, 1, 4096])]; + tensor var_1372_end_mask_0 = const()[name = tensor("op_1372_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1372_cast_fp16 = slice_by_index(begin = var_1372_begin_0, end = var_1372_end_0, end_mask = var_1372_end_mask_0, x = k_cast_fp16)[name = tensor("op_1372_cast_fp16")]; + tensor var_1374_begin_0 = const()[name = tensor("op_1374_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1374_end_0 = const()[name = tensor("op_1374_end_0"), val = tensor([1, 128, 1, 512])]; + tensor var_1374_end_mask_0 = const()[name = tensor("op_1374_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1374_cast_fp16 = slice_by_index(begin = var_1374_begin_0, end = var_1374_end_0, end_mask = var_1374_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1374_cast_fp16")]; + tensor var_1378_begin_0 = const()[name = tensor("op_1378_begin_0"), val = tensor([0, 128, 0, 0])]; + tensor var_1378_end_0 = const()[name = tensor("op_1378_end_0"), val = tensor([1, 256, 1, 512])]; + tensor var_1378_end_mask_0 = const()[name = tensor("op_1378_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1378_cast_fp16 = slice_by_index(begin = var_1378_begin_0, end = var_1378_end_0, end_mask = var_1378_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1378_cast_fp16")]; + tensor var_1382_begin_0 = const()[name = tensor("op_1382_begin_0"), val = tensor([0, 256, 0, 0])]; + tensor var_1382_end_0 = const()[name = tensor("op_1382_end_0"), val = tensor([1, 384, 1, 512])]; + tensor var_1382_end_mask_0 = const()[name = tensor("op_1382_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1382_cast_fp16 = slice_by_index(begin = var_1382_begin_0, end = var_1382_end_0, end_mask = var_1382_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1382_cast_fp16")]; + tensor var_1386_begin_0 = const()[name = tensor("op_1386_begin_0"), val = tensor([0, 384, 0, 0])]; + tensor var_1386_end_0 = const()[name = tensor("op_1386_end_0"), val = tensor([1, 512, 1, 512])]; + tensor var_1386_end_mask_0 = const()[name = tensor("op_1386_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1386_cast_fp16 = slice_by_index(begin = var_1386_begin_0, end = var_1386_end_0, end_mask = var_1386_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1386_cast_fp16")]; + tensor var_1390_begin_0 = const()[name = tensor("op_1390_begin_0"), val = tensor([0, 512, 0, 0])]; + tensor var_1390_end_0 = const()[name = tensor("op_1390_end_0"), val = tensor([1, 640, 1, 512])]; + tensor var_1390_end_mask_0 = const()[name = tensor("op_1390_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1390_cast_fp16 = slice_by_index(begin = var_1390_begin_0, end = var_1390_end_0, end_mask = var_1390_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1390_cast_fp16")]; + tensor var_1394_begin_0 = const()[name = tensor("op_1394_begin_0"), val = tensor([0, 640, 0, 0])]; + tensor var_1394_end_0 = const()[name = tensor("op_1394_end_0"), val = tensor([1, 768, 1, 512])]; + tensor var_1394_end_mask_0 = const()[name = tensor("op_1394_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1394_cast_fp16 = slice_by_index(begin = var_1394_begin_0, end = var_1394_end_0, end_mask = var_1394_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1394_cast_fp16")]; + tensor var_1398_begin_0 = const()[name = tensor("op_1398_begin_0"), val = tensor([0, 768, 0, 0])]; + tensor var_1398_end_0 = const()[name = tensor("op_1398_end_0"), val = tensor([1, 896, 1, 512])]; + tensor var_1398_end_mask_0 = const()[name = tensor("op_1398_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1398_cast_fp16 = slice_by_index(begin = var_1398_begin_0, end = var_1398_end_0, end_mask = var_1398_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1398_cast_fp16")]; + tensor var_1402_begin_0 = const()[name = tensor("op_1402_begin_0"), val = tensor([0, 896, 0, 0])]; + tensor var_1402_end_0 = const()[name = tensor("op_1402_end_0"), val = tensor([1, 1024, 1, 512])]; + tensor var_1402_end_mask_0 = const()[name = tensor("op_1402_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1402_cast_fp16 = slice_by_index(begin = var_1402_begin_0, end = var_1402_end_0, end_mask = var_1402_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1402_cast_fp16")]; + tensor var_1406_begin_0 = const()[name = tensor("op_1406_begin_0"), val = tensor([0, 1024, 0, 0])]; + tensor var_1406_end_0 = const()[name = tensor("op_1406_end_0"), val = tensor([1, 1152, 1, 512])]; + tensor var_1406_end_mask_0 = const()[name = tensor("op_1406_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1406_cast_fp16 = slice_by_index(begin = var_1406_begin_0, end = var_1406_end_0, end_mask = var_1406_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1406_cast_fp16")]; + tensor var_1410_begin_0 = const()[name = tensor("op_1410_begin_0"), val = tensor([0, 1152, 0, 0])]; + tensor var_1410_end_0 = const()[name = tensor("op_1410_end_0"), val = tensor([1, 1280, 1, 512])]; + tensor var_1410_end_mask_0 = const()[name = tensor("op_1410_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1410_cast_fp16 = slice_by_index(begin = var_1410_begin_0, end = var_1410_end_0, end_mask = var_1410_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1410_cast_fp16")]; + tensor var_1414_begin_0 = const()[name = tensor("op_1414_begin_0"), val = tensor([0, 1280, 0, 0])]; + tensor var_1414_end_0 = const()[name = tensor("op_1414_end_0"), val = tensor([1, 1408, 1, 512])]; + tensor var_1414_end_mask_0 = const()[name = tensor("op_1414_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1414_cast_fp16 = slice_by_index(begin = var_1414_begin_0, end = var_1414_end_0, end_mask = var_1414_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1414_cast_fp16")]; + tensor var_1418_begin_0 = const()[name = tensor("op_1418_begin_0"), val = tensor([0, 1408, 0, 0])]; + tensor var_1418_end_0 = const()[name = tensor("op_1418_end_0"), val = tensor([1, 1536, 1, 512])]; + tensor var_1418_end_mask_0 = const()[name = tensor("op_1418_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1418_cast_fp16 = slice_by_index(begin = var_1418_begin_0, end = var_1418_end_0, end_mask = var_1418_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1418_cast_fp16")]; + tensor var_1422_begin_0 = const()[name = tensor("op_1422_begin_0"), val = tensor([0, 1536, 0, 0])]; + tensor var_1422_end_0 = const()[name = tensor("op_1422_end_0"), val = tensor([1, 1664, 1, 512])]; + tensor var_1422_end_mask_0 = const()[name = tensor("op_1422_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1422_cast_fp16 = slice_by_index(begin = var_1422_begin_0, end = var_1422_end_0, end_mask = var_1422_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1422_cast_fp16")]; + tensor var_1426_begin_0 = const()[name = tensor("op_1426_begin_0"), val = tensor([0, 1664, 0, 0])]; + tensor var_1426_end_0 = const()[name = tensor("op_1426_end_0"), val = tensor([1, 1792, 1, 512])]; + tensor var_1426_end_mask_0 = const()[name = tensor("op_1426_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1426_cast_fp16 = slice_by_index(begin = var_1426_begin_0, end = var_1426_end_0, end_mask = var_1426_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1426_cast_fp16")]; + tensor var_1430_begin_0 = const()[name = tensor("op_1430_begin_0"), val = tensor([0, 1792, 0, 0])]; + tensor var_1430_end_0 = const()[name = tensor("op_1430_end_0"), val = tensor([1, 1920, 1, 512])]; + tensor var_1430_end_mask_0 = const()[name = tensor("op_1430_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1430_cast_fp16 = slice_by_index(begin = var_1430_begin_0, end = var_1430_end_0, end_mask = var_1430_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1430_cast_fp16")]; + tensor var_1434_begin_0 = const()[name = tensor("op_1434_begin_0"), val = tensor([0, 1920, 0, 0])]; + tensor var_1434_end_0 = const()[name = tensor("op_1434_end_0"), val = tensor([1, 2048, 1, 512])]; + tensor var_1434_end_mask_0 = const()[name = tensor("op_1434_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1434_cast_fp16 = slice_by_index(begin = var_1434_begin_0, end = var_1434_end_0, end_mask = var_1434_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1434_cast_fp16")]; + tensor var_1438_begin_0 = const()[name = tensor("op_1438_begin_0"), val = tensor([0, 2048, 0, 0])]; + tensor var_1438_end_0 = const()[name = tensor("op_1438_end_0"), val = tensor([1, 2176, 1, 512])]; + tensor var_1438_end_mask_0 = const()[name = tensor("op_1438_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1438_cast_fp16 = slice_by_index(begin = var_1438_begin_0, end = var_1438_end_0, end_mask = var_1438_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor var_1442_begin_0 = const()[name = tensor("op_1442_begin_0"), val = tensor([0, 2176, 0, 0])]; + tensor var_1442_end_0 = const()[name = tensor("op_1442_end_0"), val = tensor([1, 2304, 1, 512])]; + tensor var_1442_end_mask_0 = const()[name = tensor("op_1442_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1442_cast_fp16 = slice_by_index(begin = var_1442_begin_0, end = var_1442_end_0, end_mask = var_1442_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1442_cast_fp16")]; + tensor var_1446_begin_0 = const()[name = tensor("op_1446_begin_0"), val = tensor([0, 2304, 0, 0])]; + tensor var_1446_end_0 = const()[name = tensor("op_1446_end_0"), val = tensor([1, 2432, 1, 512])]; + tensor var_1446_end_mask_0 = const()[name = tensor("op_1446_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1446_cast_fp16 = slice_by_index(begin = var_1446_begin_0, end = var_1446_end_0, end_mask = var_1446_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1446_cast_fp16")]; + tensor var_1450_begin_0 = const()[name = tensor("op_1450_begin_0"), val = tensor([0, 2432, 0, 0])]; + tensor var_1450_end_0 = const()[name = tensor("op_1450_end_0"), val = tensor([1, 2560, 1, 512])]; + tensor var_1450_end_mask_0 = const()[name = tensor("op_1450_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1450_cast_fp16 = slice_by_index(begin = var_1450_begin_0, end = var_1450_end_0, end_mask = var_1450_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1450_cast_fp16")]; + tensor var_1454_begin_0 = const()[name = tensor("op_1454_begin_0"), val = tensor([0, 2560, 0, 0])]; + tensor var_1454_end_0 = const()[name = tensor("op_1454_end_0"), val = tensor([1, 2688, 1, 512])]; + tensor var_1454_end_mask_0 = const()[name = tensor("op_1454_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1454_cast_fp16 = slice_by_index(begin = var_1454_begin_0, end = var_1454_end_0, end_mask = var_1454_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1454_cast_fp16")]; + tensor var_1458_begin_0 = const()[name = tensor("op_1458_begin_0"), val = tensor([0, 2688, 0, 0])]; + tensor var_1458_end_0 = const()[name = tensor("op_1458_end_0"), val = tensor([1, 2816, 1, 512])]; + tensor var_1458_end_mask_0 = const()[name = tensor("op_1458_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1458_cast_fp16 = slice_by_index(begin = var_1458_begin_0, end = var_1458_end_0, end_mask = var_1458_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1458_cast_fp16")]; + tensor var_1462_begin_0 = const()[name = tensor("op_1462_begin_0"), val = tensor([0, 2816, 0, 0])]; + tensor var_1462_end_0 = const()[name = tensor("op_1462_end_0"), val = tensor([1, 2944, 1, 512])]; + tensor var_1462_end_mask_0 = const()[name = tensor("op_1462_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1462_cast_fp16 = slice_by_index(begin = var_1462_begin_0, end = var_1462_end_0, end_mask = var_1462_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1462_cast_fp16")]; + tensor var_1466_begin_0 = const()[name = tensor("op_1466_begin_0"), val = tensor([0, 2944, 0, 0])]; + tensor var_1466_end_0 = const()[name = tensor("op_1466_end_0"), val = tensor([1, 3072, 1, 512])]; + tensor var_1466_end_mask_0 = const()[name = tensor("op_1466_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1466_cast_fp16 = slice_by_index(begin = var_1466_begin_0, end = var_1466_end_0, end_mask = var_1466_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1466_cast_fp16")]; + tensor var_1470_begin_0 = const()[name = tensor("op_1470_begin_0"), val = tensor([0, 3072, 0, 0])]; + tensor var_1470_end_0 = const()[name = tensor("op_1470_end_0"), val = tensor([1, 3200, 1, 512])]; + tensor var_1470_end_mask_0 = const()[name = tensor("op_1470_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1470_cast_fp16 = slice_by_index(begin = var_1470_begin_0, end = var_1470_end_0, end_mask = var_1470_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1470_cast_fp16")]; + tensor var_1474_begin_0 = const()[name = tensor("op_1474_begin_0"), val = tensor([0, 3200, 0, 0])]; + tensor var_1474_end_0 = const()[name = tensor("op_1474_end_0"), val = tensor([1, 3328, 1, 512])]; + tensor var_1474_end_mask_0 = const()[name = tensor("op_1474_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1474_cast_fp16 = slice_by_index(begin = var_1474_begin_0, end = var_1474_end_0, end_mask = var_1474_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1474_cast_fp16")]; + tensor var_1478_begin_0 = const()[name = tensor("op_1478_begin_0"), val = tensor([0, 3328, 0, 0])]; + tensor var_1478_end_0 = const()[name = tensor("op_1478_end_0"), val = tensor([1, 3456, 1, 512])]; + tensor var_1478_end_mask_0 = const()[name = tensor("op_1478_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1478_cast_fp16 = slice_by_index(begin = var_1478_begin_0, end = var_1478_end_0, end_mask = var_1478_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1478_cast_fp16")]; + tensor var_1482_begin_0 = const()[name = tensor("op_1482_begin_0"), val = tensor([0, 3456, 0, 0])]; + tensor var_1482_end_0 = const()[name = tensor("op_1482_end_0"), val = tensor([1, 3584, 1, 512])]; + tensor var_1482_end_mask_0 = const()[name = tensor("op_1482_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1482_cast_fp16 = slice_by_index(begin = var_1482_begin_0, end = var_1482_end_0, end_mask = var_1482_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1482_cast_fp16")]; + tensor var_1486_begin_0 = const()[name = tensor("op_1486_begin_0"), val = tensor([0, 3584, 0, 0])]; + tensor var_1486_end_0 = const()[name = tensor("op_1486_end_0"), val = tensor([1, 3712, 1, 512])]; + tensor var_1486_end_mask_0 = const()[name = tensor("op_1486_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1486_cast_fp16 = slice_by_index(begin = var_1486_begin_0, end = var_1486_end_0, end_mask = var_1486_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1486_cast_fp16")]; + tensor var_1490_begin_0 = const()[name = tensor("op_1490_begin_0"), val = tensor([0, 3712, 0, 0])]; + tensor var_1490_end_0 = const()[name = tensor("op_1490_end_0"), val = tensor([1, 3840, 1, 512])]; + tensor var_1490_end_mask_0 = const()[name = tensor("op_1490_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1490_cast_fp16 = slice_by_index(begin = var_1490_begin_0, end = var_1490_end_0, end_mask = var_1490_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1490_cast_fp16")]; + tensor var_1494_begin_0 = const()[name = tensor("op_1494_begin_0"), val = tensor([0, 3840, 0, 0])]; + tensor var_1494_end_0 = const()[name = tensor("op_1494_end_0"), val = tensor([1, 3968, 1, 512])]; + tensor var_1494_end_mask_0 = const()[name = tensor("op_1494_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1494_cast_fp16 = slice_by_index(begin = var_1494_begin_0, end = var_1494_end_0, end_mask = var_1494_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1494_cast_fp16")]; + tensor var_1498_begin_0 = const()[name = tensor("op_1498_begin_0"), val = tensor([0, 3968, 0, 0])]; + tensor var_1498_end_0 = const()[name = tensor("op_1498_end_0"), val = tensor([1, 4096, 1, 512])]; + tensor var_1498_end_mask_0 = const()[name = tensor("op_1498_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1498_cast_fp16 = slice_by_index(begin = var_1498_begin_0, end = var_1498_end_0, end_mask = var_1498_end_mask_0, x = v_17_cast_fp16)[name = tensor("op_1498_cast_fp16")]; + tensor var_1502_equation_0 = const()[name = tensor("op_1502_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1502_cast_fp16 = einsum(equation = var_1502_equation_0, values = (var_1248_cast_fp16, var_1118_cast_fp16))[name = tensor("op_1502_cast_fp16")]; + tensor var_1503_to_fp16 = const()[name = tensor("op_1503_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1504_cast_fp16 = mul(x = var_1502_cast_fp16, y = var_1503_to_fp16)[name = tensor("op_1504_cast_fp16")]; + tensor var_1506_equation_0 = const()[name = tensor("op_1506_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1506_cast_fp16 = einsum(equation = var_1506_equation_0, values = (var_1252_cast_fp16, var_1122_cast_fp16))[name = tensor("op_1506_cast_fp16")]; + tensor var_1507_to_fp16 = const()[name = tensor("op_1507_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1508_cast_fp16 = mul(x = var_1506_cast_fp16, y = var_1507_to_fp16)[name = tensor("op_1508_cast_fp16")]; + tensor var_1510_equation_0 = const()[name = tensor("op_1510_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1510_cast_fp16 = einsum(equation = var_1510_equation_0, values = (var_1256_cast_fp16, var_1126_cast_fp16))[name = tensor("op_1510_cast_fp16")]; + tensor var_1511_to_fp16 = const()[name = tensor("op_1511_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1512_cast_fp16 = mul(x = var_1510_cast_fp16, y = var_1511_to_fp16)[name = tensor("op_1512_cast_fp16")]; + tensor var_1514_equation_0 = const()[name = tensor("op_1514_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1514_cast_fp16 = einsum(equation = var_1514_equation_0, values = (var_1260_cast_fp16, var_1130_cast_fp16))[name = tensor("op_1514_cast_fp16")]; + tensor var_1515_to_fp16 = const()[name = tensor("op_1515_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1516_cast_fp16 = mul(x = var_1514_cast_fp16, y = var_1515_to_fp16)[name = tensor("op_1516_cast_fp16")]; + tensor var_1518_equation_0 = const()[name = tensor("op_1518_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1518_cast_fp16 = einsum(equation = var_1518_equation_0, values = (var_1264_cast_fp16, var_1134_cast_fp16))[name = tensor("op_1518_cast_fp16")]; + tensor var_1519_to_fp16 = const()[name = tensor("op_1519_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1520_cast_fp16 = mul(x = var_1518_cast_fp16, y = var_1519_to_fp16)[name = tensor("op_1520_cast_fp16")]; + tensor var_1522_equation_0 = const()[name = tensor("op_1522_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1522_cast_fp16 = einsum(equation = var_1522_equation_0, values = (var_1268_cast_fp16, var_1138_cast_fp16))[name = tensor("op_1522_cast_fp16")]; + tensor var_1523_to_fp16 = const()[name = tensor("op_1523_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1524_cast_fp16 = mul(x = var_1522_cast_fp16, y = var_1523_to_fp16)[name = tensor("op_1524_cast_fp16")]; + tensor var_1526_equation_0 = const()[name = tensor("op_1526_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1526_cast_fp16 = einsum(equation = var_1526_equation_0, values = (var_1272_cast_fp16, var_1142_cast_fp16))[name = tensor("op_1526_cast_fp16")]; + tensor var_1527_to_fp16 = const()[name = tensor("op_1527_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1528_cast_fp16 = mul(x = var_1526_cast_fp16, y = var_1527_to_fp16)[name = tensor("op_1528_cast_fp16")]; + tensor var_1530_equation_0 = const()[name = tensor("op_1530_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1530_cast_fp16 = einsum(equation = var_1530_equation_0, values = (var_1276_cast_fp16, var_1146_cast_fp16))[name = tensor("op_1530_cast_fp16")]; + tensor var_1531_to_fp16 = const()[name = tensor("op_1531_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1532_cast_fp16 = mul(x = var_1530_cast_fp16, y = var_1531_to_fp16)[name = tensor("op_1532_cast_fp16")]; + tensor var_1534_equation_0 = const()[name = tensor("op_1534_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1534_cast_fp16 = einsum(equation = var_1534_equation_0, values = (var_1280_cast_fp16, var_1150_cast_fp16))[name = tensor("op_1534_cast_fp16")]; + tensor var_1535_to_fp16 = const()[name = tensor("op_1535_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1536_cast_fp16 = mul(x = var_1534_cast_fp16, y = var_1535_to_fp16)[name = tensor("op_1536_cast_fp16")]; + tensor var_1538_equation_0 = const()[name = tensor("op_1538_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1538_cast_fp16 = einsum(equation = var_1538_equation_0, values = (var_1284_cast_fp16, var_1154_cast_fp16))[name = tensor("op_1538_cast_fp16")]; + tensor var_1539_to_fp16 = const()[name = tensor("op_1539_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1540_cast_fp16 = mul(x = var_1538_cast_fp16, y = var_1539_to_fp16)[name = tensor("op_1540_cast_fp16")]; + tensor var_1542_equation_0 = const()[name = tensor("op_1542_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1542_cast_fp16 = einsum(equation = var_1542_equation_0, values = (var_1288_cast_fp16, var_1158_cast_fp16))[name = tensor("op_1542_cast_fp16")]; + tensor var_1543_to_fp16 = const()[name = tensor("op_1543_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1544_cast_fp16 = mul(x = var_1542_cast_fp16, y = var_1543_to_fp16)[name = tensor("op_1544_cast_fp16")]; + tensor var_1546_equation_0 = const()[name = tensor("op_1546_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1546_cast_fp16 = einsum(equation = var_1546_equation_0, values = (var_1292_cast_fp16, var_1162_cast_fp16))[name = tensor("op_1546_cast_fp16")]; + tensor var_1547_to_fp16 = const()[name = tensor("op_1547_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1548_cast_fp16 = mul(x = var_1546_cast_fp16, y = var_1547_to_fp16)[name = tensor("op_1548_cast_fp16")]; + tensor var_1550_equation_0 = const()[name = tensor("op_1550_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1550_cast_fp16 = einsum(equation = var_1550_equation_0, values = (var_1296_cast_fp16, var_1166_cast_fp16))[name = tensor("op_1550_cast_fp16")]; + tensor var_1551_to_fp16 = const()[name = tensor("op_1551_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1552_cast_fp16 = mul(x = var_1550_cast_fp16, y = var_1551_to_fp16)[name = tensor("op_1552_cast_fp16")]; + tensor var_1554_equation_0 = const()[name = tensor("op_1554_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1554_cast_fp16 = einsum(equation = var_1554_equation_0, values = (var_1300_cast_fp16, var_1170_cast_fp16))[name = tensor("op_1554_cast_fp16")]; + tensor var_1555_to_fp16 = const()[name = tensor("op_1555_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1556_cast_fp16 = mul(x = var_1554_cast_fp16, y = var_1555_to_fp16)[name = tensor("op_1556_cast_fp16")]; + tensor var_1558_equation_0 = const()[name = tensor("op_1558_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1558_cast_fp16 = einsum(equation = var_1558_equation_0, values = (var_1304_cast_fp16, var_1174_cast_fp16))[name = tensor("op_1558_cast_fp16")]; + tensor var_1559_to_fp16 = const()[name = tensor("op_1559_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1560_cast_fp16 = mul(x = var_1558_cast_fp16, y = var_1559_to_fp16)[name = tensor("op_1560_cast_fp16")]; + tensor var_1562_equation_0 = const()[name = tensor("op_1562_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1562_cast_fp16 = einsum(equation = var_1562_equation_0, values = (var_1308_cast_fp16, var_1178_cast_fp16))[name = tensor("op_1562_cast_fp16")]; + tensor var_1563_to_fp16 = const()[name = tensor("op_1563_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1564_cast_fp16 = mul(x = var_1562_cast_fp16, y = var_1563_to_fp16)[name = tensor("op_1564_cast_fp16")]; + tensor var_1566_equation_0 = const()[name = tensor("op_1566_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1566_cast_fp16 = einsum(equation = var_1566_equation_0, values = (var_1312_cast_fp16, var_1182_cast_fp16))[name = tensor("op_1566_cast_fp16")]; + tensor var_1567_to_fp16 = const()[name = tensor("op_1567_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1568_cast_fp16 = mul(x = var_1566_cast_fp16, y = var_1567_to_fp16)[name = tensor("op_1568_cast_fp16")]; + tensor var_1570_equation_0 = const()[name = tensor("op_1570_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1570_cast_fp16 = einsum(equation = var_1570_equation_0, values = (var_1316_cast_fp16, var_1186_cast_fp16))[name = tensor("op_1570_cast_fp16")]; + tensor var_1571_to_fp16 = const()[name = tensor("op_1571_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1572_cast_fp16 = mul(x = var_1570_cast_fp16, y = var_1571_to_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor var_1574_equation_0 = const()[name = tensor("op_1574_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1574_cast_fp16 = einsum(equation = var_1574_equation_0, values = (var_1320_cast_fp16, var_1190_cast_fp16))[name = tensor("op_1574_cast_fp16")]; + tensor var_1575_to_fp16 = const()[name = tensor("op_1575_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1576_cast_fp16 = mul(x = var_1574_cast_fp16, y = var_1575_to_fp16)[name = tensor("op_1576_cast_fp16")]; + tensor var_1578_equation_0 = const()[name = tensor("op_1578_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1578_cast_fp16 = einsum(equation = var_1578_equation_0, values = (var_1324_cast_fp16, var_1194_cast_fp16))[name = tensor("op_1578_cast_fp16")]; + tensor var_1579_to_fp16 = const()[name = tensor("op_1579_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1580_cast_fp16 = mul(x = var_1578_cast_fp16, y = var_1579_to_fp16)[name = tensor("op_1580_cast_fp16")]; + tensor var_1582_equation_0 = const()[name = tensor("op_1582_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1582_cast_fp16 = einsum(equation = var_1582_equation_0, values = (var_1328_cast_fp16, var_1198_cast_fp16))[name = tensor("op_1582_cast_fp16")]; + tensor var_1583_to_fp16 = const()[name = tensor("op_1583_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1584_cast_fp16 = mul(x = var_1582_cast_fp16, y = var_1583_to_fp16)[name = tensor("op_1584_cast_fp16")]; + tensor var_1586_equation_0 = const()[name = tensor("op_1586_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1586_cast_fp16 = einsum(equation = var_1586_equation_0, values = (var_1332_cast_fp16, var_1202_cast_fp16))[name = tensor("op_1586_cast_fp16")]; + tensor var_1587_to_fp16 = const()[name = tensor("op_1587_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1588_cast_fp16 = mul(x = var_1586_cast_fp16, y = var_1587_to_fp16)[name = tensor("op_1588_cast_fp16")]; + tensor var_1590_equation_0 = const()[name = tensor("op_1590_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1590_cast_fp16 = einsum(equation = var_1590_equation_0, values = (var_1336_cast_fp16, var_1206_cast_fp16))[name = tensor("op_1590_cast_fp16")]; + tensor var_1591_to_fp16 = const()[name = tensor("op_1591_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1592_cast_fp16 = mul(x = var_1590_cast_fp16, y = var_1591_to_fp16)[name = tensor("op_1592_cast_fp16")]; + tensor var_1594_equation_0 = const()[name = tensor("op_1594_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1594_cast_fp16 = einsum(equation = var_1594_equation_0, values = (var_1340_cast_fp16, var_1210_cast_fp16))[name = tensor("op_1594_cast_fp16")]; + tensor var_1595_to_fp16 = const()[name = tensor("op_1595_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1596_cast_fp16 = mul(x = var_1594_cast_fp16, y = var_1595_to_fp16)[name = tensor("op_1596_cast_fp16")]; + tensor var_1598_equation_0 = const()[name = tensor("op_1598_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1598_cast_fp16 = einsum(equation = var_1598_equation_0, values = (var_1344_cast_fp16, var_1214_cast_fp16))[name = tensor("op_1598_cast_fp16")]; + tensor var_1599_to_fp16 = const()[name = tensor("op_1599_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1600_cast_fp16 = mul(x = var_1598_cast_fp16, y = var_1599_to_fp16)[name = tensor("op_1600_cast_fp16")]; + tensor var_1602_equation_0 = const()[name = tensor("op_1602_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1602_cast_fp16 = einsum(equation = var_1602_equation_0, values = (var_1348_cast_fp16, var_1218_cast_fp16))[name = tensor("op_1602_cast_fp16")]; + tensor var_1603_to_fp16 = const()[name = tensor("op_1603_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1604_cast_fp16 = mul(x = var_1602_cast_fp16, y = var_1603_to_fp16)[name = tensor("op_1604_cast_fp16")]; + tensor var_1606_equation_0 = const()[name = tensor("op_1606_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1606_cast_fp16 = einsum(equation = var_1606_equation_0, values = (var_1352_cast_fp16, var_1222_cast_fp16))[name = tensor("op_1606_cast_fp16")]; + tensor var_1607_to_fp16 = const()[name = tensor("op_1607_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1608_cast_fp16 = mul(x = var_1606_cast_fp16, y = var_1607_to_fp16)[name = tensor("op_1608_cast_fp16")]; + tensor var_1610_equation_0 = const()[name = tensor("op_1610_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1610_cast_fp16 = einsum(equation = var_1610_equation_0, values = (var_1356_cast_fp16, var_1226_cast_fp16))[name = tensor("op_1610_cast_fp16")]; + tensor var_1611_to_fp16 = const()[name = tensor("op_1611_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1612_cast_fp16 = mul(x = var_1610_cast_fp16, y = var_1611_to_fp16)[name = tensor("op_1612_cast_fp16")]; + tensor var_1614_equation_0 = const()[name = tensor("op_1614_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1614_cast_fp16 = einsum(equation = var_1614_equation_0, values = (var_1360_cast_fp16, var_1230_cast_fp16))[name = tensor("op_1614_cast_fp16")]; + tensor var_1615_to_fp16 = const()[name = tensor("op_1615_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1616_cast_fp16 = mul(x = var_1614_cast_fp16, y = var_1615_to_fp16)[name = tensor("op_1616_cast_fp16")]; + tensor var_1618_equation_0 = const()[name = tensor("op_1618_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1618_cast_fp16 = einsum(equation = var_1618_equation_0, values = (var_1364_cast_fp16, var_1234_cast_fp16))[name = tensor("op_1618_cast_fp16")]; + tensor var_1619_to_fp16 = const()[name = tensor("op_1619_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1620_cast_fp16 = mul(x = var_1618_cast_fp16, y = var_1619_to_fp16)[name = tensor("op_1620_cast_fp16")]; + tensor var_1622_equation_0 = const()[name = tensor("op_1622_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1622_cast_fp16 = einsum(equation = var_1622_equation_0, values = (var_1368_cast_fp16, var_1238_cast_fp16))[name = tensor("op_1622_cast_fp16")]; + tensor var_1623_to_fp16 = const()[name = tensor("op_1623_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1624_cast_fp16 = mul(x = var_1622_cast_fp16, y = var_1623_to_fp16)[name = tensor("op_1624_cast_fp16")]; + tensor var_1626_equation_0 = const()[name = tensor("op_1626_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor var_1626_cast_fp16 = einsum(equation = var_1626_equation_0, values = (var_1372_cast_fp16, var_1242_cast_fp16))[name = tensor("op_1626_cast_fp16")]; + tensor var_1627_to_fp16 = const()[name = tensor("op_1627_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1628_cast_fp16 = mul(x = var_1626_cast_fp16, y = var_1627_to_fp16)[name = tensor("op_1628_cast_fp16")]; + tensor aw_65_cast_fp16 = add(x = var_1504_cast_fp16, y = mask)[name = tensor("aw_65_cast_fp16")]; + tensor aw_67_cast_fp16 = add(x = var_1508_cast_fp16, y = mask)[name = tensor("aw_67_cast_fp16")]; + tensor aw_69_cast_fp16 = add(x = var_1512_cast_fp16, y = mask)[name = tensor("aw_69_cast_fp16")]; + tensor aw_71_cast_fp16 = add(x = var_1516_cast_fp16, y = mask)[name = tensor("aw_71_cast_fp16")]; + tensor aw_73_cast_fp16 = add(x = var_1520_cast_fp16, y = mask)[name = tensor("aw_73_cast_fp16")]; + tensor aw_75_cast_fp16 = add(x = var_1524_cast_fp16, y = mask)[name = tensor("aw_75_cast_fp16")]; + tensor aw_77_cast_fp16 = add(x = var_1528_cast_fp16, y = mask)[name = tensor("aw_77_cast_fp16")]; + tensor aw_79_cast_fp16 = add(x = var_1532_cast_fp16, y = mask)[name = tensor("aw_79_cast_fp16")]; + tensor aw_81_cast_fp16 = add(x = var_1536_cast_fp16, y = mask)[name = tensor("aw_81_cast_fp16")]; + tensor aw_83_cast_fp16 = add(x = var_1540_cast_fp16, y = mask)[name = tensor("aw_83_cast_fp16")]; + tensor aw_85_cast_fp16 = add(x = var_1544_cast_fp16, y = mask)[name = tensor("aw_85_cast_fp16")]; + tensor aw_87_cast_fp16 = add(x = var_1548_cast_fp16, y = mask)[name = tensor("aw_87_cast_fp16")]; + tensor aw_89_cast_fp16 = add(x = var_1552_cast_fp16, y = mask)[name = tensor("aw_89_cast_fp16")]; + tensor aw_91_cast_fp16 = add(x = var_1556_cast_fp16, y = mask)[name = tensor("aw_91_cast_fp16")]; + tensor aw_93_cast_fp16 = add(x = var_1560_cast_fp16, y = mask)[name = tensor("aw_93_cast_fp16")]; + tensor aw_95_cast_fp16 = add(x = var_1564_cast_fp16, y = mask)[name = tensor("aw_95_cast_fp16")]; + tensor aw_97_cast_fp16 = add(x = var_1568_cast_fp16, y = mask)[name = tensor("aw_97_cast_fp16")]; + tensor aw_99_cast_fp16 = add(x = var_1572_cast_fp16, y = mask)[name = tensor("aw_99_cast_fp16")]; + tensor aw_101_cast_fp16 = add(x = var_1576_cast_fp16, y = mask)[name = tensor("aw_101_cast_fp16")]; + tensor aw_103_cast_fp16 = add(x = var_1580_cast_fp16, y = mask)[name = tensor("aw_103_cast_fp16")]; + tensor aw_105_cast_fp16 = add(x = var_1584_cast_fp16, y = mask)[name = tensor("aw_105_cast_fp16")]; + tensor aw_107_cast_fp16 = add(x = var_1588_cast_fp16, y = mask)[name = tensor("aw_107_cast_fp16")]; + tensor aw_109_cast_fp16 = add(x = var_1592_cast_fp16, y = mask)[name = tensor("aw_109_cast_fp16")]; + tensor aw_111_cast_fp16 = add(x = var_1596_cast_fp16, y = mask)[name = tensor("aw_111_cast_fp16")]; + tensor aw_113_cast_fp16 = add(x = var_1600_cast_fp16, y = mask)[name = tensor("aw_113_cast_fp16")]; + tensor aw_115_cast_fp16 = add(x = var_1604_cast_fp16, y = mask)[name = tensor("aw_115_cast_fp16")]; + tensor aw_117_cast_fp16 = add(x = var_1608_cast_fp16, y = mask)[name = tensor("aw_117_cast_fp16")]; + tensor aw_119_cast_fp16 = add(x = var_1612_cast_fp16, y = mask)[name = tensor("aw_119_cast_fp16")]; + tensor aw_121_cast_fp16 = add(x = var_1616_cast_fp16, y = mask)[name = tensor("aw_121_cast_fp16")]; + tensor aw_123_cast_fp16 = add(x = var_1620_cast_fp16, y = mask)[name = tensor("aw_123_cast_fp16")]; + tensor aw_125_cast_fp16 = add(x = var_1624_cast_fp16, y = mask)[name = tensor("aw_125_cast_fp16")]; + tensor aw_cast_fp16 = add(x = var_1628_cast_fp16, y = mask)[name = tensor("aw_cast_fp16")]; + tensor var_1661_cast_fp16 = softmax(axis = var_970, x = aw_65_cast_fp16)[name = tensor("op_1661_cast_fp16")]; + tensor var_1662_cast_fp16 = softmax(axis = var_970, x = aw_67_cast_fp16)[name = tensor("op_1662_cast_fp16")]; + tensor var_1663_cast_fp16 = softmax(axis = var_970, x = aw_69_cast_fp16)[name = tensor("op_1663_cast_fp16")]; + tensor var_1664_cast_fp16 = softmax(axis = var_970, x = aw_71_cast_fp16)[name = tensor("op_1664_cast_fp16")]; + tensor var_1665_cast_fp16 = softmax(axis = var_970, x = aw_73_cast_fp16)[name = tensor("op_1665_cast_fp16")]; + tensor var_1666_cast_fp16 = softmax(axis = var_970, x = aw_75_cast_fp16)[name = tensor("op_1666_cast_fp16")]; + tensor var_1667_cast_fp16 = softmax(axis = var_970, x = aw_77_cast_fp16)[name = tensor("op_1667_cast_fp16")]; + tensor var_1668_cast_fp16 = softmax(axis = var_970, x = aw_79_cast_fp16)[name = tensor("op_1668_cast_fp16")]; + tensor var_1669_cast_fp16 = softmax(axis = var_970, x = aw_81_cast_fp16)[name = tensor("op_1669_cast_fp16")]; + tensor var_1670_cast_fp16 = softmax(axis = var_970, x = aw_83_cast_fp16)[name = tensor("op_1670_cast_fp16")]; + tensor var_1671_cast_fp16 = softmax(axis = var_970, x = aw_85_cast_fp16)[name = tensor("op_1671_cast_fp16")]; + tensor var_1672_cast_fp16 = softmax(axis = var_970, x = aw_87_cast_fp16)[name = tensor("op_1672_cast_fp16")]; + tensor var_1673_cast_fp16 = softmax(axis = var_970, x = aw_89_cast_fp16)[name = tensor("op_1673_cast_fp16")]; + tensor var_1674_cast_fp16 = softmax(axis = var_970, x = aw_91_cast_fp16)[name = tensor("op_1674_cast_fp16")]; + tensor var_1675_cast_fp16 = softmax(axis = var_970, x = aw_93_cast_fp16)[name = tensor("op_1675_cast_fp16")]; + tensor var_1676_cast_fp16 = softmax(axis = var_970, x = aw_95_cast_fp16)[name = tensor("op_1676_cast_fp16")]; + tensor var_1677_cast_fp16 = softmax(axis = var_970, x = aw_97_cast_fp16)[name = tensor("op_1677_cast_fp16")]; + tensor var_1678_cast_fp16 = softmax(axis = var_970, x = aw_99_cast_fp16)[name = tensor("op_1678_cast_fp16")]; + tensor var_1679_cast_fp16 = softmax(axis = var_970, x = aw_101_cast_fp16)[name = tensor("op_1679_cast_fp16")]; + tensor var_1680_cast_fp16 = softmax(axis = var_970, x = aw_103_cast_fp16)[name = tensor("op_1680_cast_fp16")]; + tensor var_1681_cast_fp16 = softmax(axis = var_970, x = aw_105_cast_fp16)[name = tensor("op_1681_cast_fp16")]; + tensor var_1682_cast_fp16 = softmax(axis = var_970, x = aw_107_cast_fp16)[name = tensor("op_1682_cast_fp16")]; + tensor var_1683_cast_fp16 = softmax(axis = var_970, x = aw_109_cast_fp16)[name = tensor("op_1683_cast_fp16")]; + tensor var_1684_cast_fp16 = softmax(axis = var_970, x = aw_111_cast_fp16)[name = tensor("op_1684_cast_fp16")]; + tensor var_1685_cast_fp16 = softmax(axis = var_970, x = aw_113_cast_fp16)[name = tensor("op_1685_cast_fp16")]; + tensor var_1686_cast_fp16 = softmax(axis = var_970, x = aw_115_cast_fp16)[name = tensor("op_1686_cast_fp16")]; + tensor var_1687_cast_fp16 = softmax(axis = var_970, x = aw_117_cast_fp16)[name = tensor("op_1687_cast_fp16")]; + tensor var_1688_cast_fp16 = softmax(axis = var_970, x = aw_119_cast_fp16)[name = tensor("op_1688_cast_fp16")]; + tensor var_1689_cast_fp16 = softmax(axis = var_970, x = aw_121_cast_fp16)[name = tensor("op_1689_cast_fp16")]; + tensor var_1690_cast_fp16 = softmax(axis = var_970, x = aw_123_cast_fp16)[name = tensor("op_1690_cast_fp16")]; + tensor var_1691_cast_fp16 = softmax(axis = var_970, x = aw_125_cast_fp16)[name = tensor("op_1691_cast_fp16")]; + tensor var_1692_cast_fp16 = softmax(axis = var_970, x = aw_cast_fp16)[name = tensor("op_1692_cast_fp16")]; + tensor var_1694_equation_0 = const()[name = tensor("op_1694_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1694_cast_fp16 = einsum(equation = var_1694_equation_0, values = (var_1374_cast_fp16, var_1661_cast_fp16))[name = tensor("op_1694_cast_fp16")]; + tensor var_1696_equation_0 = const()[name = tensor("op_1696_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1696_cast_fp16 = einsum(equation = var_1696_equation_0, values = (var_1378_cast_fp16, var_1662_cast_fp16))[name = tensor("op_1696_cast_fp16")]; + tensor var_1698_equation_0 = const()[name = tensor("op_1698_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1698_cast_fp16 = einsum(equation = var_1698_equation_0, values = (var_1382_cast_fp16, var_1663_cast_fp16))[name = tensor("op_1698_cast_fp16")]; + tensor var_1700_equation_0 = const()[name = tensor("op_1700_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1700_cast_fp16 = einsum(equation = var_1700_equation_0, values = (var_1386_cast_fp16, var_1664_cast_fp16))[name = tensor("op_1700_cast_fp16")]; + tensor var_1702_equation_0 = const()[name = tensor("op_1702_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1702_cast_fp16 = einsum(equation = var_1702_equation_0, values = (var_1390_cast_fp16, var_1665_cast_fp16))[name = tensor("op_1702_cast_fp16")]; + tensor var_1704_equation_0 = const()[name = tensor("op_1704_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1704_cast_fp16 = einsum(equation = var_1704_equation_0, values = (var_1394_cast_fp16, var_1666_cast_fp16))[name = tensor("op_1704_cast_fp16")]; + tensor var_1706_equation_0 = const()[name = tensor("op_1706_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1706_cast_fp16 = einsum(equation = var_1706_equation_0, values = (var_1398_cast_fp16, var_1667_cast_fp16))[name = tensor("op_1706_cast_fp16")]; + tensor var_1708_equation_0 = const()[name = tensor("op_1708_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1708_cast_fp16 = einsum(equation = var_1708_equation_0, values = (var_1402_cast_fp16, var_1668_cast_fp16))[name = tensor("op_1708_cast_fp16")]; + tensor var_1710_equation_0 = const()[name = tensor("op_1710_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1710_cast_fp16 = einsum(equation = var_1710_equation_0, values = (var_1406_cast_fp16, var_1669_cast_fp16))[name = tensor("op_1710_cast_fp16")]; + tensor var_1712_equation_0 = const()[name = tensor("op_1712_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1712_cast_fp16 = einsum(equation = var_1712_equation_0, values = (var_1410_cast_fp16, var_1670_cast_fp16))[name = tensor("op_1712_cast_fp16")]; + tensor var_1714_equation_0 = const()[name = tensor("op_1714_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1714_cast_fp16 = einsum(equation = var_1714_equation_0, values = (var_1414_cast_fp16, var_1671_cast_fp16))[name = tensor("op_1714_cast_fp16")]; + tensor var_1716_equation_0 = const()[name = tensor("op_1716_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1716_cast_fp16 = einsum(equation = var_1716_equation_0, values = (var_1418_cast_fp16, var_1672_cast_fp16))[name = tensor("op_1716_cast_fp16")]; + tensor var_1718_equation_0 = const()[name = tensor("op_1718_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1718_cast_fp16 = einsum(equation = var_1718_equation_0, values = (var_1422_cast_fp16, var_1673_cast_fp16))[name = tensor("op_1718_cast_fp16")]; + tensor var_1720_equation_0 = const()[name = tensor("op_1720_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1720_cast_fp16 = einsum(equation = var_1720_equation_0, values = (var_1426_cast_fp16, var_1674_cast_fp16))[name = tensor("op_1720_cast_fp16")]; + tensor var_1722_equation_0 = const()[name = tensor("op_1722_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1722_cast_fp16 = einsum(equation = var_1722_equation_0, values = (var_1430_cast_fp16, var_1675_cast_fp16))[name = tensor("op_1722_cast_fp16")]; + tensor var_1724_equation_0 = const()[name = tensor("op_1724_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1724_cast_fp16 = einsum(equation = var_1724_equation_0, values = (var_1434_cast_fp16, var_1676_cast_fp16))[name = tensor("op_1724_cast_fp16")]; + tensor var_1726_equation_0 = const()[name = tensor("op_1726_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1726_cast_fp16 = einsum(equation = var_1726_equation_0, values = (var_1438_cast_fp16, var_1677_cast_fp16))[name = tensor("op_1726_cast_fp16")]; + tensor var_1728_equation_0 = const()[name = tensor("op_1728_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1728_cast_fp16 = einsum(equation = var_1728_equation_0, values = (var_1442_cast_fp16, var_1678_cast_fp16))[name = tensor("op_1728_cast_fp16")]; + tensor var_1730_equation_0 = const()[name = tensor("op_1730_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1730_cast_fp16 = einsum(equation = var_1730_equation_0, values = (var_1446_cast_fp16, var_1679_cast_fp16))[name = tensor("op_1730_cast_fp16")]; + tensor var_1732_equation_0 = const()[name = tensor("op_1732_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1732_cast_fp16 = einsum(equation = var_1732_equation_0, values = (var_1450_cast_fp16, var_1680_cast_fp16))[name = tensor("op_1732_cast_fp16")]; + tensor var_1734_equation_0 = const()[name = tensor("op_1734_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1734_cast_fp16 = einsum(equation = var_1734_equation_0, values = (var_1454_cast_fp16, var_1681_cast_fp16))[name = tensor("op_1734_cast_fp16")]; + tensor var_1736_equation_0 = const()[name = tensor("op_1736_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1736_cast_fp16 = einsum(equation = var_1736_equation_0, values = (var_1458_cast_fp16, var_1682_cast_fp16))[name = tensor("op_1736_cast_fp16")]; + tensor var_1738_equation_0 = const()[name = tensor("op_1738_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1738_cast_fp16 = einsum(equation = var_1738_equation_0, values = (var_1462_cast_fp16, var_1683_cast_fp16))[name = tensor("op_1738_cast_fp16")]; + tensor var_1740_equation_0 = const()[name = tensor("op_1740_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1740_cast_fp16 = einsum(equation = var_1740_equation_0, values = (var_1466_cast_fp16, var_1684_cast_fp16))[name = tensor("op_1740_cast_fp16")]; + tensor var_1742_equation_0 = const()[name = tensor("op_1742_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1742_cast_fp16 = einsum(equation = var_1742_equation_0, values = (var_1470_cast_fp16, var_1685_cast_fp16))[name = tensor("op_1742_cast_fp16")]; + tensor var_1744_equation_0 = const()[name = tensor("op_1744_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1744_cast_fp16 = einsum(equation = var_1744_equation_0, values = (var_1474_cast_fp16, var_1686_cast_fp16))[name = tensor("op_1744_cast_fp16")]; + tensor var_1746_equation_0 = const()[name = tensor("op_1746_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1746_cast_fp16 = einsum(equation = var_1746_equation_0, values = (var_1478_cast_fp16, var_1687_cast_fp16))[name = tensor("op_1746_cast_fp16")]; + tensor var_1748_equation_0 = const()[name = tensor("op_1748_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1748_cast_fp16 = einsum(equation = var_1748_equation_0, values = (var_1482_cast_fp16, var_1688_cast_fp16))[name = tensor("op_1748_cast_fp16")]; + tensor var_1750_equation_0 = const()[name = tensor("op_1750_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1750_cast_fp16 = einsum(equation = var_1750_equation_0, values = (var_1486_cast_fp16, var_1689_cast_fp16))[name = tensor("op_1750_cast_fp16")]; + tensor var_1752_equation_0 = const()[name = tensor("op_1752_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1752_cast_fp16 = einsum(equation = var_1752_equation_0, values = (var_1490_cast_fp16, var_1690_cast_fp16))[name = tensor("op_1752_cast_fp16")]; + tensor var_1754_equation_0 = const()[name = tensor("op_1754_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1754_cast_fp16 = einsum(equation = var_1754_equation_0, values = (var_1494_cast_fp16, var_1691_cast_fp16))[name = tensor("op_1754_cast_fp16")]; + tensor var_1756_equation_0 = const()[name = tensor("op_1756_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1756_cast_fp16 = einsum(equation = var_1756_equation_0, values = (var_1498_cast_fp16, var_1692_cast_fp16))[name = tensor("op_1756_cast_fp16")]; + tensor x_27_interleave_0 = const()[name = tensor("x_27_interleave_0"), val = tensor(false)]; + tensor x_27_cast_fp16 = concat(axis = var_970, interleave = x_27_interleave_0, values = (var_1694_cast_fp16, var_1696_cast_fp16, var_1698_cast_fp16, var_1700_cast_fp16, var_1702_cast_fp16, var_1704_cast_fp16, var_1706_cast_fp16, var_1708_cast_fp16, var_1710_cast_fp16, var_1712_cast_fp16, var_1714_cast_fp16, var_1716_cast_fp16, var_1718_cast_fp16, var_1720_cast_fp16, var_1722_cast_fp16, var_1724_cast_fp16, var_1726_cast_fp16, var_1728_cast_fp16, var_1730_cast_fp16, var_1732_cast_fp16, var_1734_cast_fp16, var_1736_cast_fp16, var_1738_cast_fp16, var_1740_cast_fp16, var_1742_cast_fp16, var_1744_cast_fp16, var_1746_cast_fp16, var_1748_cast_fp16, var_1750_cast_fp16, var_1752_cast_fp16, var_1754_cast_fp16, var_1756_cast_fp16))[name = tensor("x_27_cast_fp16")]; + tensor var_1761 = const()[name = tensor("op_1761"), val = tensor([1, 4096, -1, 8])]; + tensor input_13_cast_fp16 = reshape(shape = var_1761, x = x_27_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor var_1765 = const()[name = tensor("op_1765"), val = tensor([1, 1])]; + tensor var_1767 = const()[name = tensor("op_1767"), val = tensor([1, 1])]; + tensor var_1769_pad_type_0 = const()[name = tensor("op_1769_pad_type_0"), val = tensor("custom")]; + tensor var_1769_pad_0 = const()[name = tensor("op_1769_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1769_cast_fp16 = conv(dilations = var_1767, groups = var_970, pad = var_1769_pad_0, pad_type = var_1769_pad_type_0, strides = var_1765, weight = blocks_1_attn_proj_weight_palettized_cast_fp16, x = input_13_cast_fp16)[name = tensor("op_1769_cast_fp16")]; + tensor blocks_1_attn_proj_output_scales_to_fp16 = const()[name = tensor("blocks_1_attn_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202513472)))]; + tensor attention_output_cast_fp16 = mul(x = var_1769_cast_fp16, y = blocks_1_attn_proj_output_scales_to_fp16)[name = tensor("attention_output_cast_fp16")]; + tensor x_29_cast_fp16 = add(x = attention_output_cast_fp16, y = x_17_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor x_eps_interleave_0 = const()[name = tensor("x_eps_interleave_0"), val = tensor(false)]; + tensor eps_chan_to_fp16 = const()[name = tensor("eps_chan_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202521728)))]; + tensor x_eps_cast_fp16 = concat(axis = var_970, interleave = x_eps_interleave_0, values = (x_29_cast_fp16, eps_chan_to_fp16))[name = tensor("x_eps_cast_fp16")]; + tensor norm_x_axes_0 = const()[name = tensor("norm_x_axes_0"), val = tensor([1])]; + tensor norm_x_cast_fp16 = reduce_l2_norm(axes = norm_x_axes_0, keep_dims = var_973, x = x_eps_cast_fp16)[name = tensor("norm_x_cast_fp16")]; + tensor x_normed_19_cast_fp16 = real_div(x = x_29_cast_fp16, y = norm_x_cast_fp16)[name = tensor("x_normed_19_cast_fp16")]; + tensor var_1794_to_fp16 = const()[name = tensor("op_1794_to_fp16"), val = tensor(0x1p+6)]; + tensor x_normed_21_cast_fp16 = mul(x = x_normed_19_cast_fp16, y = var_1794_to_fp16)[name = tensor("x_normed_21_cast_fp16")]; + tensor blocks_1_norm_2_weight_to_fp16 = const()[name = tensor("blocks_1_norm_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202521920)))]; + tensor input_15_cast_fp16 = mul(x = x_normed_21_cast_fp16, y = blocks_1_norm_2_weight_to_fp16)[name = tensor("input_15_cast_fp16")]; + tensor var_1806 = const()[name = tensor("op_1806"), val = tensor([1, 1])]; + tensor var_1808 = const()[name = tensor("op_1808"), val = tensor([1, 1])]; + tensor var_1810_pad_type_0 = const()[name = tensor("op_1810_pad_type_0"), val = tensor("custom")]; + tensor var_1810_pad_0 = const()[name = tensor("op_1810_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1810_cast_fp16 = conv(dilations = var_1808, groups = var_970, pad = var_1810_pad_0, pad_type = var_1810_pad_type_0, strides = var_1806, weight = blocks_1_mlp_fc_1_weight_palettized_cast_fp16, x = input_15_cast_fp16)[name = tensor("op_1810_cast_fp16")]; + tensor blocks_1_mlp_fc_1_output_scales_to_fp16 = const()[name = tensor("blocks_1_mlp_fc_1_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202530176)))]; + tensor input_17_cast_fp16 = mul(x = var_1810_cast_fp16, y = blocks_1_mlp_fc_1_output_scales_to_fp16)[name = tensor("input_17_cast_fp16")]; + tensor var_1814 = const()[name = tensor("op_1814"), val = tensor([1, 1])]; + tensor var_1816 = const()[name = tensor("op_1816"), val = tensor([1, 1])]; + tensor var_1818_pad_type_0 = const()[name = tensor("op_1818_pad_type_0"), val = tensor("custom")]; + tensor var_1818_pad_0 = const()[name = tensor("op_1818_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1818_cast_fp16 = conv(dilations = var_1816, groups = var_970, pad = var_1818_pad_0, pad_type = var_1818_pad_type_0, strides = var_1814, weight = blocks_1_mlp_fc_2_weight_palettized_cast_fp16, x = input_15_cast_fp16)[name = tensor("op_1818_cast_fp16")]; + tensor blocks_1_mlp_fc_2_output_scales_to_fp16 = const()[name = tensor("blocks_1_mlp_fc_2_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202552256)))]; + tensor x_fc_2_cast_fp16 = mul(x = var_1818_cast_fp16, y = blocks_1_mlp_fc_2_output_scales_to_fp16)[name = tensor("x_fc_2_cast_fp16")]; + tensor var_1820_cast_fp16 = silu(x = input_17_cast_fp16)[name = tensor("op_1820_cast_fp16")]; + tensor input_cast_fp16 = mul(x = var_1820_cast_fp16, y = x_fc_2_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_1824 = const()[name = tensor("op_1824"), val = tensor([1, 1])]; + tensor var_1826 = const()[name = tensor("op_1826"), val = tensor([1, 1])]; + tensor var_1828_pad_type_0 = const()[name = tensor("op_1828_pad_type_0"), val = tensor("custom")]; + tensor var_1828_pad_0 = const()[name = tensor("op_1828_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1828_cast_fp16 = conv(dilations = var_1826, groups = var_970, pad = var_1828_pad_0, pad_type = var_1828_pad_type_0, strides = var_1824, weight = blocks_1_mlp_proj_weight_palettized_cast_fp16, x = input_cast_fp16)[name = tensor("op_1828_cast_fp16")]; + tensor blocks_1_mlp_proj_output_scales_to_fp16 = const()[name = tensor("blocks_1_mlp_proj_output_scales_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202574336)))]; + tensor var_1829_cast_fp16 = mul(x = var_1828_cast_fp16, y = blocks_1_mlp_proj_output_scales_to_fp16)[name = tensor("op_1829_cast_fp16")]; + tensor new_x = add(x = var_1829_cast_fp16, y = x_29_cast_fp16)[name = tensor("op_1830_cast_fp16")]; } -> (new_x, new_k_cache_0, new_k_cache_1, new_v_cache_0, new_v_cache_1); } \ No newline at end of file