program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3402.3.2"}, {"coremlc-version", "3402.4.1"}})]
{
    func main<ios18>(tensor<fp16, [1, 1, ?, ?]> causal_mask, tensor<int32, [1, ?]> input_ids, state<tensor<fp16, [28, 1, 2, 2048, 128]>> key_cache, state<tensor<fp16, [28, 1, 2, 2048, 128]>> value_cache) [FlexibleShapeInformation = tuple<tuple<string, dict<string, tensor<int32, [?]>>>, tuple<string, dict<string, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"causal_mask", [1, 1, 1, 1]}, {"input_ids", [1, 1]}}), ("RangeDims", {{"causal_mask", [[1, 1], [1, 1], [1, 2048], [1, 2048]]}, {"input_ids", [[1, 1], [1, 2048]]}})))] {
            tensor<int32, [4]> var_7_shape_cast_fp16 = shape(x = causal_mask)[name = string("op_7_shape_cast_fp16")];
            int32 gather_0_axis_0 = const()[name = string("gather_0_axis_0"), val = int32(0)];
            int32 gather_0_batch_dims_0 = const()[name = string("gather_0_batch_dims_0"), val = int32(0)];
            bool gather_0_validate_indices_0 = const()[name = string("gather_0_validate_indices_0"), val = bool(false)];
            string var_7_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_7_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")];
            uint16 select_0_to_uint16 = const()[name = string("select_0_to_uint16"), val = uint16(3)];
            tensor<int16, [4]> var_7_shape_cast_fp16_to_int16 = cast(dtype = var_7_shape_cast_fp16_to_int16_dtype_0, x = var_7_shape_cast_fp16)[name = string("cast_176")];
            int16 gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_7_shape_cast_fp16_to_int16)[name = string("gather_0_cast_uint16")];
            string gather_0_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_0_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_10_shape = shape(x = input_ids)[name = string("op_10_shape")];
            int32 gather_1_axis_0 = const()[name = string("gather_1_axis_0"), val = int32(0)];
            int32 gather_1_batch_dims_0 = const()[name = string("gather_1_batch_dims_0"), val = int32(0)];
            bool gather_1_validate_indices_0 = const()[name = string("gather_1_validate_indices_0"), val = bool(false)];
            string var_10_shape_to_uint16_dtype_0 = const()[name = string("op_10_shape_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_1_to_uint16 = const()[name = string("select_1_to_uint16"), val = uint16(1)];
            tensor<uint16, [2]> var_10_shape_to_uint16 = cast(dtype = var_10_shape_to_uint16_dtype_0, x = var_10_shape)[name = string("cast_175")];
            uint16 gather_1_cast_uint16 = gather(axis = gather_1_axis_0, batch_dims = gather_1_batch_dims_0, indices = select_1_to_uint16, validate_indices = gather_1_validate_indices_0, x = var_10_shape_to_uint16)[name = string("gather_1_cast_uint16")];
            string gather_1_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_1_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = string("cast_173")];
            int32 gather_1_cast_uint16_to_int32 = cast(dtype = gather_1_cast_uint16_to_int32_dtype_0, x = gather_1_cast_uint16)[name = string("cast_174")];
            int32 past_seen_tokens = sub(x = gather_0_cast_uint16_to_int32, y = gather_1_cast_uint16_to_int32)[name = string("past_seen_tokens")];
            int32 var_72 = const()[name = string("op_72"), val = int32(-1)];
            int32 var_77 = const()[name = string("op_77"), val = int32(6)];
            int32 inputs_embeds_axis_0 = const()[name = string("inputs_embeds_axis_0"), val = int32(0)];
            int32 inputs_embeds_batch_dims_0 = const()[name = string("inputs_embeds_batch_dims_0"), val = int32(0)];
            bool inputs_embeds_validate_indices_0 = const()[name = string("inputs_embeds_validate_indices_0"), val = bool(false)];
            tensor<fp16, [151936, 1536]> model_model_embed_tokens_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [151936, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [151936, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116686976))))[name = string("model_model_embed_tokens_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = input_ids, validate_indices = inputs_embeds_validate_indices_0, x = model_model_embed_tokens_weight_to_fp16_quantized)[name = string("inputs_embeds_cast_fp16")];
            tensor<int32, [3]> var_174_shape_cast_fp16 = shape(x = inputs_embeds_cast_fp16)[name = string("op_174_shape_cast_fp16")];
            int32 gather_2_axis_0 = const()[name = string("gather_2_axis_0"), val = int32(0)];
            int32 gather_2_batch_dims_0 = const()[name = string("gather_2_batch_dims_0"), val = int32(0)];
            bool gather_2_validate_indices_0 = const()[name = string("gather_2_validate_indices_0"), val = bool(false)];
            string var_174_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_174_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_2_to_uint16 = const()[name = string("select_2_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_174_shape_cast_fp16_to_uint16 = cast(dtype = var_174_shape_cast_fp16_to_uint16_dtype_0, x = var_174_shape_cast_fp16)[name = string("cast_172")];
            uint16 gather_2_cast_uint16 = gather(axis = gather_2_axis_0, batch_dims = gather_2_batch_dims_0, indices = select_2_to_uint16, validate_indices = gather_2_validate_indices_0, x = var_174_shape_cast_fp16_to_uint16)[name = string("gather_2_cast_uint16")];
            string gather_2_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_2_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_2_cast_uint16_to_int32 = cast(dtype = gather_2_cast_uint16_to_int32_dtype_0, x = gather_2_cast_uint16)[name = string("cast_171")];
            int32 var_176 = add(x = past_seen_tokens, y = gather_2_cast_uint16_to_int32)[name = string("op_176")];
            int32 const_0 = const()[name = string("const_0"), val = int32(1)];
            tensor<int32, [?]> cache_position = range_1d(end = var_176, start = past_seen_tokens, step = const_0)[name = string("cache_position")];
            tensor<int32, [1]> position_ids_axes_0 = const()[name = string("position_ids_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1, ?]> position_ids = expand_dims(axes = position_ids_axes_0, x = cache_position)[name = string("position_ids")];
            tensor<int32, [1]> var_189_axes_0 = const()[name = string("op_189_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1, ?]> var_189 = expand_dims(axes = var_189_axes_0, x = position_ids)[name = string("op_189")];
            bool var_194_transpose_x_0 = const()[name = string("op_194_transpose_x_0"), val = bool(false)];
            bool var_194_transpose_y_0 = const()[name = string("op_194_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 64, 1]> const_2_to_fp16 = const()[name = string("const_2_to_fp16"), val = tensor<fp16, [1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131272896)))];
            string cast_2_to_fp16_dtype_0 = const()[name = string("cast_2_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 1, ?]> var_189_to_fp16 = cast(dtype = cast_2_to_fp16_dtype_0, x = var_189)[name = string("cast_170")];
            tensor<fp16, [1, 64, ?]> var_194_cast_fp16 = matmul(transpose_x = var_194_transpose_x_0, transpose_y = var_194_transpose_y_0, x = const_2_to_fp16, y = var_189_to_fp16)[name = string("op_194_cast_fp16")];
            tensor<int32, [3]> freqs_perm_0 = const()[name = string("freqs_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            bool emb_interleave_0 = const()[name = string("emb_interleave_0"), val = bool(false)];
            tensor<fp16, [1, ?, 64]> freqs_cast_fp16 = transpose(perm = freqs_perm_0, x = var_194_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, ?, 128]> emb_cast_fp16 = concat(axis = var_72, interleave = emb_interleave_0, values = (freqs_cast_fp16, freqs_cast_fp16))[name = string("emb_cast_fp16")];
            tensor<fp16, [1, ?, 128]> cos_1_cast_fp16 = cos(x = emb_cast_fp16)[name = string("cos_1_cast_fp16")];
            tensor<fp16, [1, ?, 128]> sin_1_cast_fp16 = sin(x = emb_cast_fp16)[name = string("sin_1_cast_fp16")];
            fp16 var_78_promoted_to_fp16 = const()[name = string("op_78_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_215_cast_fp16 = pow(x = inputs_embeds_cast_fp16, y = var_78_promoted_to_fp16)[name = string("op_215_cast_fp16")];
            tensor<int32, [1]> variance_1_axes_0 = const()[name = string("variance_1_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_1_keep_dims_0 = const()[name = string("variance_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_1_cast_fp16 = reduce_mean(axes = variance_1_axes_0, keep_dims = variance_1_keep_dims_0, x = var_215_cast_fp16)[name = string("variance_1_cast_fp16")];
            fp16 var_218_to_fp16 = const()[name = string("op_218_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_219_cast_fp16 = add(x = variance_1_cast_fp16, y = var_218_to_fp16)[name = string("op_219_cast_fp16")];
            fp32 var_220_epsilon_0 = const()[name = string("op_220_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_220_cast_fp16 = rsqrt(epsilon = var_220_epsilon_0, x = var_219_cast_fp16)[name = string("op_220_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_3_cast_fp16 = mul(x = inputs_embeds_cast_fp16, y = var_220_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131273088)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_7_cast_fp16 = mul(x = model_model_layers_0_input_layernorm_weight_to_fp16, y = hidden_states_3_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131276224))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132455936))))[name = string("model_model_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132603456)))];
            tensor<fp16, [1, ?, 1536]> linear_0_cast_fp16 = linear(bias = model_model_layers_0_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_7_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132606592))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132803264))))[name = string("model_model_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_0_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_0_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132827904)))];
            tensor<fp16, [1, ?, 256]> linear_1_cast_fp16 = linear(bias = model_model_layers_0_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_7_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132828480))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133025152))))[name = string("model_model_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133049792)))];
            tensor<fp16, [1, ?, 256]> linear_2_cast_fp16 = linear(bias = model_model_layers_0_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_7_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_243_cast_fp16 = reshape(shape = concat_0x, x = linear_0_cast_fp16)[name = string("op_243_cast_fp16")];
            tensor<int32, [4]> q_1_perm_0 = const()[name = string("q_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_246_cast_fp16 = reshape(shape = concat_1x, x = linear_1_cast_fp16)[name = string("op_246_cast_fp16")];
            tensor<int32, [4]> k_1_perm_0 = const()[name = string("k_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_249_cast_fp16 = reshape(shape = concat_2x, x = linear_2_cast_fp16)[name = string("op_249_cast_fp16")];
            tensor<int32, [4]> v_state_1_perm_0 = const()[name = string("v_state_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> cos_7_axes_0 = const()[name = string("cos_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, ?, 128]> cos_7_cast_fp16 = expand_dims(axes = cos_7_axes_0, x = cos_1_cast_fp16)[name = string("cos_7_cast_fp16")];
            tensor<int32, [1]> sin_7_axes_0 = const()[name = string("sin_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, ?, 128]> sin_7_cast_fp16 = expand_dims(axes = sin_7_axes_0, x = sin_1_cast_fp16)[name = string("sin_7_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = var_243_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 12, ?, 128]> var_253_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_7_cast_fp16)[name = string("op_253_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_1_cast_fp16)[name = string("x2_1_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_264_cast_fp16 = mul(x = x2_1_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_264_cast_fp16")];
            bool var_266_interleave_0 = const()[name = string("op_266_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_266_cast_fp16 = concat(axis = var_72, interleave = var_266_interleave_0, values = (var_264_cast_fp16, x1_1_cast_fp16))[name = string("op_266_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_267_cast_fp16 = mul(x = var_266_cast_fp16, y = sin_7_cast_fp16)[name = string("op_267_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_3_cast_fp16 = add(x = var_253_cast_fp16, y = var_267_cast_fp16)[name = string("query_states_3_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = var_246_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 2, ?, 128]> var_269_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_7_cast_fp16)[name = string("op_269_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_1_cast_fp16)[name = string("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_1_cast_fp16)[name = string("x2_3_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_280_cast_fp16 = mul(x = x2_3_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_280_cast_fp16")];
            bool var_282_interleave_0 = const()[name = string("op_282_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_282_cast_fp16 = concat(axis = var_72, interleave = var_282_interleave_0, values = (var_280_cast_fp16, x1_3_cast_fp16))[name = string("op_282_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_283_cast_fp16 = mul(x = var_282_cast_fp16, y = sin_7_cast_fp16)[name = string("op_283_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_1_cast_fp16 = add(x = var_269_cast_fp16, y = var_283_cast_fp16)[name = string("k_state_1_cast_fp16")];
            tensor<int32, [1]> var_285_shape = shape(x = cache_position)[name = string("op_285_shape")];
            int32 gather_10_axis_0 = const()[name = string("gather_10_axis_0"), val = int32(0)];
            int32 gather_10_batch_dims_0 = const()[name = string("gather_10_batch_dims_0"), val = int32(0)];
            bool gather_10_validate_indices_0 = const()[name = string("gather_10_validate_indices_0"), val = bool(false)];
            string var_285_shape_to_uint16_dtype_0 = const()[name = string("op_285_shape_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_10_to_uint16 = const()[name = string("select_10_to_uint16"), val = uint16(0)];
            tensor<uint16, [1]> var_285_shape_to_uint16 = cast(dtype = var_285_shape_to_uint16_dtype_0, x = var_285_shape)[name = string("cast_169")];
            uint16 gather_10_cast_uint16 = gather(axis = gather_10_axis_0, batch_dims = gather_10_batch_dims_0, indices = select_10_to_uint16, validate_indices = gather_10_validate_indices_0, x = var_285_shape_to_uint16)[name = string("gather_10_cast_uint16")];
            string gather_10_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_10_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_10_cast_uint16_to_int32 = cast(dtype = gather_10_cast_uint16_to_int32_dtype_0, x = gather_10_cast_uint16)[name = string("cast_168")];
            int32 end_1 = add(x = past_seen_tokens, y = gather_10_cast_uint16_to_int32)[name = string("end_1")];
            tensor<fp16, [28, 1, 2, 2048, 128]> read_state_0 = read_state(input = key_cache)[name = string("read_state_0")];
            tensor<int32, [1]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_2_axes_0 = const()[name = string("expand_dims_2_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_2 = expand_dims(axes = expand_dims_2_axes_0, x = past_seen_tokens)[name = string("expand_dims_2")];
            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1]> expand_dims_5_axes_0 = const()[name = string("expand_dims_5_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_5 = expand_dims(axes = expand_dims_5_axes_0, x = end_1)[name = string("expand_dims_5")];
            tensor<int32, [1]> concat_5_values0_0 = const()[name = string("concat_5_values0_0"), val = tensor<int32, [1]>([0])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (concat_5_values0_0, expand_dims_0, expand_dims_1, expand_dims_2, expand_dims_3))[name = string("concat_5")];
            tensor<int32, [1]> concat_6_values0_0 = const()[name = string("concat_6_values0_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values1_0 = const()[name = string("concat_6_values1_0"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_6_values4_0 = const()[name = string("concat_6_values4_0"), val = tensor<int32, [1]>([0])];
            int32 concat_6_axis_0 = const()[name = string("concat_6_axis_0"), val = int32(0)];
            bool concat_6_interleave_0 = const()[name = string("concat_6_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (concat_6_values0_0, concat_6_values1_0, expand_dims_4, expand_dims_5, concat_6_values4_0))[name = string("concat_6")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_1_stride_0 = const()[name = string("key_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = key_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_1_squeeze_mask_0, stride = key_cache_internal_tensor_assign_1_stride_0, update = k_state_1_cast_fp16, x = read_state_0)[name = string("key_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_1_cast_fp16, input = key_cache)[name = string("coreml_update_state_56_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_56 = read_state(input = key_cache)[name = string("coreml_update_state_56")];
            tensor<fp16, [28, 1, 2, 2048, 128]> read_state_1 = read_state(input = value_cache)[name = string("read_state_1")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_1_stride_0 = const()[name = string("value_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_1_cast_fp16 = transpose(perm = v_state_1_perm_0, x = var_249_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_5, begin_mask = value_cache_internal_tensor_assign_1_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_1_squeeze_mask_0, stride = value_cache_internal_tensor_assign_1_stride_0, update = v_state_1_cast_fp16, x = read_state_1)[name = string("value_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_1_cast_fp16, input = value_cache)[name = string("coreml_update_state_57_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_57 = read_state(input = value_cache)[name = string("coreml_update_state_57")];
            tensor<int32, [5]> var_306_begin_0 = const()[name = string("op_306_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_306_end_0 = const()[name = string("op_306_end_0"), val = tensor<int32, [5]>([1, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_306_end_mask_0 = const()[name = string("op_306_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_306_squeeze_mask_0 = const()[name = string("op_306_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_306_cast_fp16 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, squeeze_mask = var_306_squeeze_mask_0, x = coreml_update_state_56)[name = string("op_306_cast_fp16")];
            int32 concat_11_values0_0 = const()[name = string("concat_11_values0_0"), val = int32(1)];
            int32 concat_11_values1_0 = const()[name = string("concat_11_values1_0"), val = int32(2)];
            int32 concat_11_values3_0 = const()[name = string("concat_11_values3_0"), val = int32(128)];
            int32 concat_11_axis_0 = const()[name = string("concat_11_axis_0"), val = int32(0)];
            bool concat_11_interleave_0 = const()[name = string("concat_11_interleave_0"), val = bool(false)];
            tensor<int32, [4]> concat_11 = concat(axis = concat_11_axis_0, interleave = concat_11_interleave_0, values = (concat_11_values0_0, concat_11_values1_0, end_1, concat_11_values3_0))[name = string("concat_11")];
            tensor<int32, [4]> var_309_begin_0 = const()[name = string("op_309_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_309_end_mask_0 = const()[name = string("op_309_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_309_cast_fp16 = slice_by_index(begin = var_309_begin_0, end = concat_11, end_mask = var_309_end_mask_0, x = var_306_cast_fp16)[name = string("op_309_cast_fp16")];
            tensor<int32, [5]> var_311_begin_0 = const()[name = string("op_311_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_311_end_0 = const()[name = string("op_311_end_0"), val = tensor<int32, [5]>([1, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_311_end_mask_0 = const()[name = string("op_311_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_311_squeeze_mask_0 = const()[name = string("op_311_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_311_cast_fp16 = slice_by_index(begin = var_311_begin_0, end = var_311_end_0, end_mask = var_311_end_mask_0, squeeze_mask = var_311_squeeze_mask_0, x = coreml_update_state_57)[name = string("op_311_cast_fp16")];
            tensor<int32, [4]> var_314_begin_0 = const()[name = string("op_314_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_314_end_mask_0 = const()[name = string("op_314_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_314_cast_fp16 = slice_by_index(begin = var_314_begin_0, end = concat_11, end_mask = var_314_end_mask_0, x = var_311_cast_fp16)[name = string("op_314_cast_fp16")];
            tensor<int32, [4]> var_316_shape_cast_fp16 = shape(x = var_309_cast_fp16)[name = string("op_316_shape_cast_fp16")];
            int32 gather_13 = const()[name = string("gather_13"), val = int32(1)];
            int32 gather_14 = const()[name = string("gather_14"), val = int32(2)];
            int32 gather_15_axis_0 = const()[name = string("gather_15_axis_0"), val = int32(0)];
            int32 gather_15_batch_dims_0 = const()[name = string("gather_15_batch_dims_0"), val = int32(0)];
            bool gather_15_validate_indices_0 = const()[name = string("gather_15_validate_indices_0"), val = bool(false)];
            string var_316_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_316_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_15_to_uint16 = const()[name = string("select_15_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_316_shape_cast_fp16_to_uint16 = cast(dtype = var_316_shape_cast_fp16_to_uint16_dtype_0, x = var_316_shape_cast_fp16)[name = string("cast_167")];
            uint16 gather_15_cast_uint16 = gather(axis = gather_15_axis_0, batch_dims = gather_15_batch_dims_0, indices = select_15_to_uint16, validate_indices = gather_15_validate_indices_0, x = var_316_shape_cast_fp16_to_uint16)[name = string("gather_15_cast_uint16")];
            string gather_15_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_15_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_16 = const()[name = string("gather_16"), val = int32(128)];
            tensor<int32, [1]> var_323_axes_0 = const()[name = string("op_323_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_323_cast_fp16 = expand_dims(axes = var_323_axes_0, x = var_309_cast_fp16)[name = string("op_323_cast_fp16")];
            tensor<int32, [5]> shape_17_cast_fp16 = shape(x = var_323_cast_fp16)[name = string("shape_17_cast_fp16")];
            int32 concat_13_axis_0 = const()[name = string("concat_13_axis_0"), val = int32(0)];
            bool concat_13_interleave_0 = const()[name = string("concat_13_interleave_0"), val = bool(false)];
            int32 gather_15_cast_uint16_to_int32 = cast(dtype = gather_15_cast_uint16_to_int32_dtype_0, x = gather_15_cast_uint16)[name = string("cast_166")];
            tensor<int32, [5]> concat_13 = concat(axis = concat_13_axis_0, interleave = concat_13_interleave_0, values = (gather_13, gather_14, var_77, gather_15_cast_uint16_to_int32, gather_16))[name = string("concat_13")];
            tensor<int32, [5]> real_div_0 = real_div(x = concat_13, y = shape_17_cast_fp16)[name = string("real_div_0")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_11_cast_fp16 = tile(reps = real_div_0, x = var_323_cast_fp16)[name = string("hidden_states_11_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_3_cast_fp16 = reshape(shape = concat_14x, x = hidden_states_11_cast_fp16)[name = string("key_states_3_cast_fp16")];
            tensor<int32, [4]> var_333_shape_cast_fp16 = shape(x = var_314_cast_fp16)[name = string("op_333_shape_cast_fp16")];
            int32 gather_17 = const()[name = string("gather_17"), val = int32(1)];
            int32 gather_18 = const()[name = string("gather_18"), val = int32(2)];
            int32 gather_19_axis_0 = const()[name = string("gather_19_axis_0"), val = int32(0)];
            int32 gather_19_batch_dims_0 = const()[name = string("gather_19_batch_dims_0"), val = int32(0)];
            bool gather_19_validate_indices_0 = const()[name = string("gather_19_validate_indices_0"), val = bool(false)];
            string var_333_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_333_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_19_to_uint16 = const()[name = string("select_19_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_333_shape_cast_fp16_to_uint16 = cast(dtype = var_333_shape_cast_fp16_to_uint16_dtype_0, x = var_333_shape_cast_fp16)[name = string("cast_165")];
            uint16 gather_19_cast_uint16 = gather(axis = gather_19_axis_0, batch_dims = gather_19_batch_dims_0, indices = select_19_to_uint16, validate_indices = gather_19_validate_indices_0, x = var_333_shape_cast_fp16_to_uint16)[name = string("gather_19_cast_uint16")];
            string gather_19_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_19_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_20 = const()[name = string("gather_20"), val = int32(128)];
            tensor<int32, [1]> var_340_axes_0 = const()[name = string("op_340_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_340_cast_fp16 = expand_dims(axes = var_340_axes_0, x = var_314_cast_fp16)[name = string("op_340_cast_fp16")];
            tensor<int32, [5]> shape_22_cast_fp16 = shape(x = var_340_cast_fp16)[name = string("shape_22_cast_fp16")];
            int32 concat_15_axis_0 = const()[name = string("concat_15_axis_0"), val = int32(0)];
            bool concat_15_interleave_0 = const()[name = string("concat_15_interleave_0"), val = bool(false)];
            int32 gather_19_cast_uint16_to_int32 = cast(dtype = gather_19_cast_uint16_to_int32_dtype_0, x = gather_19_cast_uint16)[name = string("cast_164")];
            tensor<int32, [5]> concat_15 = concat(axis = concat_15_axis_0, interleave = concat_15_interleave_0, values = (gather_17, gather_18, var_77, gather_19_cast_uint16_to_int32, gather_20))[name = string("concat_15")];
            tensor<int32, [5]> real_div_1 = real_div(x = concat_15, y = shape_22_cast_fp16)[name = string("real_div_1")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_15_cast_fp16 = tile(reps = real_div_1, x = var_340_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            tensor<int32, [4]> concat_16x = const()[name = string("concat_16x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_3_cast_fp16 = reshape(shape = concat_16x, x = hidden_states_15_cast_fp16)[name = string("value_states_3_cast_fp16")];
            tensor<int32, [4]> var_350_shape_cast_fp16 = shape(x = key_states_3_cast_fp16)[name = string("op_350_shape_cast_fp16")];
            int32 gather_21_axis_0 = const()[name = string("gather_21_axis_0"), val = int32(0)];
            int32 gather_21_batch_dims_0 = const()[name = string("gather_21_batch_dims_0"), val = int32(0)];
            bool gather_21_validate_indices_0 = const()[name = string("gather_21_validate_indices_0"), val = bool(false)];
            string var_350_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_350_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_21_to_uint16 = const()[name = string("select_21_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_350_shape_cast_fp16_to_uint16 = cast(dtype = var_350_shape_cast_fp16_to_uint16_dtype_0, x = var_350_shape_cast_fp16)[name = string("cast_163")];
            uint16 gather_21_cast_uint16 = gather(axis = gather_21_axis_0, batch_dims = gather_21_batch_dims_0, indices = select_21_to_uint16, validate_indices = gather_21_validate_indices_0, x = var_350_shape_cast_fp16_to_uint16)[name = string("gather_21_cast_uint16")];
            string gather_21_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_21_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_17_values0_0 = const()[name = string("concat_17_values0_0"), val = int32(1)];
            int32 concat_17_values1_0 = const()[name = string("concat_17_values1_0"), val = int32(1)];
            int32 concat_17_values2_0 = const()[name = string("concat_17_values2_0"), val = int32(0)];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            int32 gather_21_cast_uint16_to_int32 = cast(dtype = gather_21_cast_uint16_to_int32_dtype_0, x = gather_21_cast_uint16)[name = string("cast_162")];
            tensor<int32, [4]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (concat_17_values0_0, concat_17_values1_0, concat_17_values2_0, gather_21_cast_uint16_to_int32))[name = string("concat_17")];
            tensor<int32, [4]> causal_mask_3_begin_0 = const()[name = string("causal_mask_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_3_end_mask_0 = const()[name = string("causal_mask_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_3_cast_fp16 = slice_by_index(begin = causal_mask_3_begin_0, end = concat_17, end_mask = causal_mask_3_end_mask_0, x = causal_mask)[name = string("causal_mask_3_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_1_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_3_cast_fp16, key = key_states_3_cast_fp16, query = query_states_3_cast_fp16, value = value_states_3_cast_fp16)[name = string("attn_output_1_cast_fp16")];
            tensor<int32, [4]> var_356_perm_0 = const()[name = string("op_356_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_18x = const()[name = string("concat_18x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_356_cast_fp16 = transpose(perm = var_356_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, ?, 1536]> input_1_cast_fp16 = reshape(shape = concat_18x, x = var_356_cast_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133050368))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134230080))))[name = string("model_model_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> linear_3_bias_0_to_fp16 = const()[name = string("linear_3_bias_0_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134377600)))];
            tensor<fp16, [1, ?, 1536]> linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_19_cast_fp16 = add(x = inputs_embeds_cast_fp16, y = linear_3_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            fp16 var_78_promoted_1_to_fp16 = const()[name = string("op_78_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_365_cast_fp16 = pow(x = hidden_states_19_cast_fp16, y = var_78_promoted_1_to_fp16)[name = string("op_365_cast_fp16")];
            tensor<int32, [1]> variance_3_axes_0 = const()[name = string("variance_3_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_3_keep_dims_0 = const()[name = string("variance_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_3_cast_fp16 = reduce_mean(axes = variance_3_axes_0, keep_dims = variance_3_keep_dims_0, x = var_365_cast_fp16)[name = string("variance_3_cast_fp16")];
            fp16 var_368_to_fp16 = const()[name = string("op_368_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_369_cast_fp16 = add(x = variance_3_cast_fp16, y = var_368_to_fp16)[name = string("op_369_cast_fp16")];
            fp32 var_370_epsilon_0 = const()[name = string("op_370_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_370_cast_fp16 = rsqrt(epsilon = var_370_epsilon_0, x = var_369_cast_fp16)[name = string("op_370_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_23_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = var_370_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134380736)))];
            tensor<fp16, [1, ?, 1536]> input_3_cast_fp16 = mul(x = model_model_layers_0_post_attention_layernorm_weight_to_fp16, y = hidden_states_23_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134383872))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141265216))))[name = string("model_model_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [8960]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142125440)))];
            tensor<fp16, [1, ?, 8960]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_382_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_382_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142143424))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149024768))))[name = string("model_model_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_7_cast_fp16 = mul(x = var_382_cast_fp16, y = linear_5_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149884992))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156766336))))[name = string("model_model_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_6_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_29_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = linear_6_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            fp16 var_78_promoted_2_to_fp16 = const()[name = string("op_78_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_395_cast_fp16 = pow(x = hidden_states_29_cast_fp16, y = var_78_promoted_2_to_fp16)[name = string("op_395_cast_fp16")];
            tensor<int32, [1]> variance_5_axes_0 = const()[name = string("variance_5_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_5_keep_dims_0 = const()[name = string("variance_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_5_cast_fp16 = reduce_mean(axes = variance_5_axes_0, keep_dims = variance_5_keep_dims_0, x = var_395_cast_fp16)[name = string("variance_5_cast_fp16")];
            fp16 var_398_to_fp16 = const()[name = string("op_398_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_399_cast_fp16 = add(x = variance_5_cast_fp16, y = var_398_to_fp16)[name = string("op_399_cast_fp16")];
            fp32 var_400_epsilon_0 = const()[name = string("op_400_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_400_cast_fp16 = rsqrt(epsilon = var_400_epsilon_0, x = var_399_cast_fp16)[name = string("op_400_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_33_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = var_400_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157626560)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_37_cast_fp16 = mul(x = model_model_layers_1_input_layernorm_weight_to_fp16, y = hidden_states_33_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157629696))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158809408))))[name = string("model_model_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158956928)))];
            tensor<fp16, [1, ?, 1536]> linear_7_cast_fp16 = linear(bias = model_model_layers_1_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_37_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158960064))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159156736))))[name = string("model_model_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_1_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_1_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159181376)))];
            tensor<fp16, [1, ?, 256]> linear_8_cast_fp16 = linear(bias = model_model_layers_1_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_37_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159181952))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159378624))))[name = string("model_model_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159403264)))];
            tensor<fp16, [1, ?, 256]> linear_9_cast_fp16 = linear(bias = model_model_layers_1_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_37_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> concat_19x = const()[name = string("concat_19x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_423_cast_fp16 = reshape(shape = concat_19x, x = linear_7_cast_fp16)[name = string("op_423_cast_fp16")];
            tensor<int32, [4]> q_3_perm_0 = const()[name = string("q_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_20x = const()[name = string("concat_20x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_426_cast_fp16 = reshape(shape = concat_20x, x = linear_8_cast_fp16)[name = string("op_426_cast_fp16")];
            tensor<int32, [4]> k_3_perm_0 = const()[name = string("k_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_21x = const()[name = string("concat_21x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_429_cast_fp16 = reshape(shape = concat_21x, x = linear_9_cast_fp16)[name = string("op_429_cast_fp16")];
            tensor<int32, [4]> v_state_3_perm_0 = const()[name = string("v_state_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = var_423_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 12, ?, 128]> var_433_cast_fp16 = mul(x = q_3_cast_fp16, y = cos_7_cast_fp16)[name = string("op_433_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_3_cast_fp16)[name = string("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_3_cast_fp16)[name = string("x2_5_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_444_cast_fp16 = mul(x = x2_5_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_444_cast_fp16")];
            bool var_446_interleave_0 = const()[name = string("op_446_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_446_cast_fp16 = concat(axis = var_72, interleave = var_446_interleave_0, values = (var_444_cast_fp16, x1_5_cast_fp16))[name = string("op_446_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_447_cast_fp16 = mul(x = var_446_cast_fp16, y = sin_7_cast_fp16)[name = string("op_447_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_7_cast_fp16 = add(x = var_433_cast_fp16, y = var_447_cast_fp16)[name = string("query_states_7_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = var_426_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [1, 2, ?, 128]> var_449_cast_fp16 = mul(x = k_3_cast_fp16, y = cos_7_cast_fp16)[name = string("op_449_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_3_cast_fp16)[name = string("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_3_cast_fp16)[name = string("x2_7_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_460_cast_fp16 = mul(x = x2_7_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_460_cast_fp16")];
            bool var_462_interleave_0 = const()[name = string("op_462_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_462_cast_fp16 = concat(axis = var_72, interleave = var_462_interleave_0, values = (var_460_cast_fp16, x1_7_cast_fp16))[name = string("op_462_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_463_cast_fp16 = mul(x = var_462_cast_fp16, y = sin_7_cast_fp16)[name = string("op_463_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_3_cast_fp16 = add(x = var_449_cast_fp16, y = var_463_cast_fp16)[name = string("k_state_3_cast_fp16")];
            tensor<int32, [1]> expand_dims_12 = const()[name = string("expand_dims_12"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_13 = const()[name = string("expand_dims_13"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_15 = const()[name = string("expand_dims_15"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_24_values0_0 = const()[name = string("concat_24_values0_0"), val = tensor<int32, [1]>([1])];
            int32 concat_24_axis_0 = const()[name = string("concat_24_axis_0"), val = int32(0)];
            bool concat_24_interleave_0 = const()[name = string("concat_24_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_24 = concat(axis = concat_24_axis_0, interleave = concat_24_interleave_0, values = (concat_24_values0_0, expand_dims_12, expand_dims_13, expand_dims_2, expand_dims_15))[name = string("concat_24")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_2_stride_0 = const()[name = string("key_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_24, begin_mask = key_cache_internal_tensor_assign_2_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_2_squeeze_mask_0, stride = key_cache_internal_tensor_assign_2_stride_0, update = k_state_3_cast_fp16, x = coreml_update_state_56)[name = string("key_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_2_cast_fp16, input = key_cache)[name = string("coreml_update_state_58_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_58 = read_state(input = key_cache)[name = string("coreml_update_state_58")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_2_stride_0 = const()[name = string("value_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_3_cast_fp16 = transpose(perm = v_state_3_perm_0, x = var_429_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_24, begin_mask = value_cache_internal_tensor_assign_2_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_2_squeeze_mask_0, stride = value_cache_internal_tensor_assign_2_stride_0, update = v_state_3_cast_fp16, x = coreml_update_state_57)[name = string("value_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_2_cast_fp16, input = value_cache)[name = string("coreml_update_state_59_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_59 = read_state(input = value_cache)[name = string("coreml_update_state_59")];
            tensor<int32, [5]> var_486_begin_0 = const()[name = string("op_486_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> var_486_end_0 = const()[name = string("op_486_end_0"), val = tensor<int32, [5]>([2, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_486_end_mask_0 = const()[name = string("op_486_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_486_squeeze_mask_0 = const()[name = string("op_486_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_486_cast_fp16 = slice_by_index(begin = var_486_begin_0, end = var_486_end_0, end_mask = var_486_end_mask_0, squeeze_mask = var_486_squeeze_mask_0, x = coreml_update_state_58)[name = string("op_486_cast_fp16")];
            tensor<int32, [4]> var_489_begin_0 = const()[name = string("op_489_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_489_end_mask_0 = const()[name = string("op_489_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_489_cast_fp16 = slice_by_index(begin = var_489_begin_0, end = concat_11, end_mask = var_489_end_mask_0, x = var_486_cast_fp16)[name = string("op_489_cast_fp16")];
            tensor<int32, [5]> var_491_begin_0 = const()[name = string("op_491_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> var_491_end_0 = const()[name = string("op_491_end_0"), val = tensor<int32, [5]>([2, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_491_end_mask_0 = const()[name = string("op_491_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_491_squeeze_mask_0 = const()[name = string("op_491_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_491_cast_fp16 = slice_by_index(begin = var_491_begin_0, end = var_491_end_0, end_mask = var_491_end_mask_0, squeeze_mask = var_491_squeeze_mask_0, x = coreml_update_state_59)[name = string("op_491_cast_fp16")];
            tensor<int32, [4]> var_494_begin_0 = const()[name = string("op_494_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_494_end_mask_0 = const()[name = string("op_494_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_494_cast_fp16 = slice_by_index(begin = var_494_begin_0, end = concat_11, end_mask = var_494_end_mask_0, x = var_491_cast_fp16)[name = string("op_494_cast_fp16")];
            tensor<int32, [4]> var_496_shape_cast_fp16 = shape(x = var_489_cast_fp16)[name = string("op_496_shape_cast_fp16")];
            int32 gather_31 = const()[name = string("gather_31"), val = int32(1)];
            int32 gather_32 = const()[name = string("gather_32"), val = int32(2)];
            int32 gather_33_axis_0 = const()[name = string("gather_33_axis_0"), val = int32(0)];
            int32 gather_33_batch_dims_0 = const()[name = string("gather_33_batch_dims_0"), val = int32(0)];
            bool gather_33_validate_indices_0 = const()[name = string("gather_33_validate_indices_0"), val = bool(false)];
            string var_496_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_496_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_33_to_uint16 = const()[name = string("select_33_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_496_shape_cast_fp16_to_uint16 = cast(dtype = var_496_shape_cast_fp16_to_uint16_dtype_0, x = var_496_shape_cast_fp16)[name = string("cast_161")];
            uint16 gather_33_cast_uint16 = gather(axis = gather_33_axis_0, batch_dims = gather_33_batch_dims_0, indices = select_33_to_uint16, validate_indices = gather_33_validate_indices_0, x = var_496_shape_cast_fp16_to_uint16)[name = string("gather_33_cast_uint16")];
            string gather_33_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_33_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_34 = const()[name = string("gather_34"), val = int32(128)];
            tensor<int32, [1]> var_503_axes_0 = const()[name = string("op_503_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_503_cast_fp16 = expand_dims(axes = var_503_axes_0, x = var_489_cast_fp16)[name = string("op_503_cast_fp16")];
            tensor<int32, [5]> shape_37_cast_fp16 = shape(x = var_503_cast_fp16)[name = string("shape_37_cast_fp16")];
            int32 concat_32_axis_0 = const()[name = string("concat_32_axis_0"), val = int32(0)];
            bool concat_32_interleave_0 = const()[name = string("concat_32_interleave_0"), val = bool(false)];
            int32 gather_33_cast_uint16_to_int32 = cast(dtype = gather_33_cast_uint16_to_int32_dtype_0, x = gather_33_cast_uint16)[name = string("cast_160")];
            tensor<int32, [5]> concat_32 = concat(axis = concat_32_axis_0, interleave = concat_32_interleave_0, values = (gather_31, gather_32, var_77, gather_33_cast_uint16_to_int32, gather_34))[name = string("concat_32")];
            tensor<int32, [5]> real_div_2 = real_div(x = concat_32, y = shape_37_cast_fp16)[name = string("real_div_2")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_41_cast_fp16 = tile(reps = real_div_2, x = var_503_cast_fp16)[name = string("hidden_states_41_cast_fp16")];
            tensor<int32, [4]> concat_33x = const()[name = string("concat_33x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_7_cast_fp16 = reshape(shape = concat_33x, x = hidden_states_41_cast_fp16)[name = string("key_states_7_cast_fp16")];
            tensor<int32, [4]> var_513_shape_cast_fp16 = shape(x = var_494_cast_fp16)[name = string("op_513_shape_cast_fp16")];
            int32 gather_35 = const()[name = string("gather_35"), val = int32(1)];
            int32 gather_36 = const()[name = string("gather_36"), val = int32(2)];
            int32 gather_37_axis_0 = const()[name = string("gather_37_axis_0"), val = int32(0)];
            int32 gather_37_batch_dims_0 = const()[name = string("gather_37_batch_dims_0"), val = int32(0)];
            bool gather_37_validate_indices_0 = const()[name = string("gather_37_validate_indices_0"), val = bool(false)];
            string var_513_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_513_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_37_to_uint16 = const()[name = string("select_37_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_513_shape_cast_fp16_to_uint16 = cast(dtype = var_513_shape_cast_fp16_to_uint16_dtype_0, x = var_513_shape_cast_fp16)[name = string("cast_159")];
            uint16 gather_37_cast_uint16 = gather(axis = gather_37_axis_0, batch_dims = gather_37_batch_dims_0, indices = select_37_to_uint16, validate_indices = gather_37_validate_indices_0, x = var_513_shape_cast_fp16_to_uint16)[name = string("gather_37_cast_uint16")];
            string gather_37_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_37_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_38 = const()[name = string("gather_38"), val = int32(128)];
            tensor<int32, [1]> var_520_axes_0 = const()[name = string("op_520_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_520_cast_fp16 = expand_dims(axes = var_520_axes_0, x = var_494_cast_fp16)[name = string("op_520_cast_fp16")];
            tensor<int32, [5]> shape_42_cast_fp16 = shape(x = var_520_cast_fp16)[name = string("shape_42_cast_fp16")];
            int32 concat_34_axis_0 = const()[name = string("concat_34_axis_0"), val = int32(0)];
            bool concat_34_interleave_0 = const()[name = string("concat_34_interleave_0"), val = bool(false)];
            int32 gather_37_cast_uint16_to_int32 = cast(dtype = gather_37_cast_uint16_to_int32_dtype_0, x = gather_37_cast_uint16)[name = string("cast_158")];
            tensor<int32, [5]> concat_34 = concat(axis = concat_34_axis_0, interleave = concat_34_interleave_0, values = (gather_35, gather_36, var_77, gather_37_cast_uint16_to_int32, gather_38))[name = string("concat_34")];
            tensor<int32, [5]> real_div_3 = real_div(x = concat_34, y = shape_42_cast_fp16)[name = string("real_div_3")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_45_cast_fp16 = tile(reps = real_div_3, x = var_520_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_7_cast_fp16 = reshape(shape = concat_35x, x = hidden_states_45_cast_fp16)[name = string("value_states_7_cast_fp16")];
            tensor<int32, [4]> var_530_shape_cast_fp16 = shape(x = key_states_7_cast_fp16)[name = string("op_530_shape_cast_fp16")];
            int32 gather_39_axis_0 = const()[name = string("gather_39_axis_0"), val = int32(0)];
            int32 gather_39_batch_dims_0 = const()[name = string("gather_39_batch_dims_0"), val = int32(0)];
            bool gather_39_validate_indices_0 = const()[name = string("gather_39_validate_indices_0"), val = bool(false)];
            string var_530_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_530_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_39_to_uint16 = const()[name = string("select_39_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_530_shape_cast_fp16_to_uint16 = cast(dtype = var_530_shape_cast_fp16_to_uint16_dtype_0, x = var_530_shape_cast_fp16)[name = string("cast_157")];
            uint16 gather_39_cast_uint16 = gather(axis = gather_39_axis_0, batch_dims = gather_39_batch_dims_0, indices = select_39_to_uint16, validate_indices = gather_39_validate_indices_0, x = var_530_shape_cast_fp16_to_uint16)[name = string("gather_39_cast_uint16")];
            string gather_39_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_39_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_36_values0_0 = const()[name = string("concat_36_values0_0"), val = int32(1)];
            int32 concat_36_values1_0 = const()[name = string("concat_36_values1_0"), val = int32(1)];
            int32 concat_36_values2_0 = const()[name = string("concat_36_values2_0"), val = int32(0)];
            int32 concat_36_axis_0 = const()[name = string("concat_36_axis_0"), val = int32(0)];
            bool concat_36_interleave_0 = const()[name = string("concat_36_interleave_0"), val = bool(false)];
            int32 gather_39_cast_uint16_to_int32 = cast(dtype = gather_39_cast_uint16_to_int32_dtype_0, x = gather_39_cast_uint16)[name = string("cast_156")];
            tensor<int32, [4]> concat_36 = concat(axis = concat_36_axis_0, interleave = concat_36_interleave_0, values = (concat_36_values0_0, concat_36_values1_0, concat_36_values2_0, gather_39_cast_uint16_to_int32))[name = string("concat_36")];
            tensor<int32, [4]> causal_mask_5_begin_0 = const()[name = string("causal_mask_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_5_end_mask_0 = const()[name = string("causal_mask_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_5_cast_fp16 = slice_by_index(begin = causal_mask_5_begin_0, end = concat_36, end_mask = causal_mask_5_end_mask_0, x = causal_mask)[name = string("causal_mask_5_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_5_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_5_cast_fp16, key = key_states_7_cast_fp16, query = query_states_7_cast_fp16, value = value_states_7_cast_fp16)[name = string("attn_output_5_cast_fp16")];
            tensor<int32, [4]> var_536_perm_0 = const()[name = string("op_536_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_536_cast_fp16 = transpose(perm = var_536_perm_0, x = attn_output_5_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [1, ?, 1536]> input_9_cast_fp16 = reshape(shape = concat_37x, x = var_536_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159403840))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160583552))))[name = string("model_model_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_10_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_49_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = linear_10_cast_fp16)[name = string("hidden_states_49_cast_fp16")];
            fp16 var_78_promoted_3_to_fp16 = const()[name = string("op_78_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_545_cast_fp16 = pow(x = hidden_states_49_cast_fp16, y = var_78_promoted_3_to_fp16)[name = string("op_545_cast_fp16")];
            tensor<int32, [1]> variance_7_axes_0 = const()[name = string("variance_7_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_7_keep_dims_0 = const()[name = string("variance_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_7_cast_fp16 = reduce_mean(axes = variance_7_axes_0, keep_dims = variance_7_keep_dims_0, x = var_545_cast_fp16)[name = string("variance_7_cast_fp16")];
            fp16 var_548_to_fp16 = const()[name = string("op_548_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_549_cast_fp16 = add(x = variance_7_cast_fp16, y = var_548_to_fp16)[name = string("op_549_cast_fp16")];
            fp32 var_550_epsilon_0 = const()[name = string("op_550_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_550_cast_fp16 = rsqrt(epsilon = var_550_epsilon_0, x = var_549_cast_fp16)[name = string("op_550_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_53_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = var_550_cast_fp16)[name = string("hidden_states_53_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160731072)))];
            tensor<fp16, [1, ?, 1536]> input_11_cast_fp16 = mul(x = model_model_layers_1_post_attention_layernorm_weight_to_fp16, y = hidden_states_53_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160734208))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167615552))))[name = string("model_model_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_562_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_562_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168475776))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175357120))))[name = string("model_model_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_15_cast_fp16 = mul(x = var_562_cast_fp16, y = linear_12_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176217344))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183098688))))[name = string("model_model_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_13_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_59_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = linear_13_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            fp16 var_78_promoted_4_to_fp16 = const()[name = string("op_78_promoted_4_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_575_cast_fp16 = pow(x = hidden_states_59_cast_fp16, y = var_78_promoted_4_to_fp16)[name = string("op_575_cast_fp16")];
            tensor<int32, [1]> variance_9_axes_0 = const()[name = string("variance_9_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_9_keep_dims_0 = const()[name = string("variance_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_9_cast_fp16 = reduce_mean(axes = variance_9_axes_0, keep_dims = variance_9_keep_dims_0, x = var_575_cast_fp16)[name = string("variance_9_cast_fp16")];
            fp16 var_578_to_fp16 = const()[name = string("op_578_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_579_cast_fp16 = add(x = variance_9_cast_fp16, y = var_578_to_fp16)[name = string("op_579_cast_fp16")];
            fp32 var_580_epsilon_0 = const()[name = string("op_580_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_580_cast_fp16 = rsqrt(epsilon = var_580_epsilon_0, x = var_579_cast_fp16)[name = string("op_580_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_63_cast_fp16 = mul(x = hidden_states_59_cast_fp16, y = var_580_cast_fp16)[name = string("hidden_states_63_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183958912)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_67_cast_fp16 = mul(x = model_model_layers_2_input_layernorm_weight_to_fp16, y = hidden_states_63_cast_fp16)[name = string("hidden_states_67_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183962048))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185141760))))[name = string("model_model_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185289280)))];
            tensor<fp16, [1, ?, 1536]> linear_14_cast_fp16 = linear(bias = model_model_layers_2_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_67_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185292416))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185489088))))[name = string("model_model_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_2_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_2_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185513728)))];
            tensor<fp16, [1, ?, 256]> linear_15_cast_fp16 = linear(bias = model_model_layers_2_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_67_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185514304))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185710976))))[name = string("model_model_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185735616)))];
            tensor<fp16, [1, ?, 256]> linear_16_cast_fp16 = linear(bias = model_model_layers_2_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_67_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_603_cast_fp16 = reshape(shape = concat_38x, x = linear_14_cast_fp16)[name = string("op_603_cast_fp16")];
            tensor<int32, [4]> q_5_perm_0 = const()[name = string("q_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_39x = const()[name = string("concat_39x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_606_cast_fp16 = reshape(shape = concat_39x, x = linear_15_cast_fp16)[name = string("op_606_cast_fp16")];
            tensor<int32, [4]> k_5_perm_0 = const()[name = string("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_40x = const()[name = string("concat_40x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_609_cast_fp16 = reshape(shape = concat_40x, x = linear_16_cast_fp16)[name = string("op_609_cast_fp16")];
            tensor<int32, [4]> v_state_5_perm_0 = const()[name = string("v_state_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_5_cast_fp16 = transpose(perm = q_5_perm_0, x = var_603_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 12, ?, 128]> var_613_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_7_cast_fp16)[name = string("op_613_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_5_cast_fp16)[name = string("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_5_cast_fp16)[name = string("x2_9_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_624_cast_fp16 = mul(x = x2_9_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_624_cast_fp16")];
            bool var_626_interleave_0 = const()[name = string("op_626_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_626_cast_fp16 = concat(axis = var_72, interleave = var_626_interleave_0, values = (var_624_cast_fp16, x1_9_cast_fp16))[name = string("op_626_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_627_cast_fp16 = mul(x = var_626_cast_fp16, y = sin_7_cast_fp16)[name = string("op_627_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_11_cast_fp16 = add(x = var_613_cast_fp16, y = var_627_cast_fp16)[name = string("query_states_11_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_5_cast_fp16 = transpose(perm = k_5_perm_0, x = var_606_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 2, ?, 128]> var_629_cast_fp16 = mul(x = k_5_cast_fp16, y = cos_7_cast_fp16)[name = string("op_629_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_5_cast_fp16)[name = string("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_5_cast_fp16)[name = string("x2_11_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_640_cast_fp16 = mul(x = x2_11_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_640_cast_fp16")];
            bool var_642_interleave_0 = const()[name = string("op_642_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_642_cast_fp16 = concat(axis = var_72, interleave = var_642_interleave_0, values = (var_640_cast_fp16, x1_11_cast_fp16))[name = string("op_642_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_643_cast_fp16 = mul(x = var_642_cast_fp16, y = sin_7_cast_fp16)[name = string("op_643_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_5_cast_fp16 = add(x = var_629_cast_fp16, y = var_643_cast_fp16)[name = string("k_state_5_cast_fp16")];
            tensor<int32, [1]> expand_dims_24 = const()[name = string("expand_dims_24"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_25 = const()[name = string("expand_dims_25"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_27 = const()[name = string("expand_dims_27"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_43_values0_0 = const()[name = string("concat_43_values0_0"), val = tensor<int32, [1]>([2])];
            int32 concat_43_axis_0 = const()[name = string("concat_43_axis_0"), val = int32(0)];
            bool concat_43_interleave_0 = const()[name = string("concat_43_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_43 = concat(axis = concat_43_axis_0, interleave = concat_43_interleave_0, values = (concat_43_values0_0, expand_dims_24, expand_dims_25, expand_dims_2, expand_dims_27))[name = string("concat_43")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_3_stride_0 = const()[name = string("key_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_43, begin_mask = key_cache_internal_tensor_assign_3_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_3_squeeze_mask_0, stride = key_cache_internal_tensor_assign_3_stride_0, update = k_state_5_cast_fp16, x = coreml_update_state_58)[name = string("key_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_3_cast_fp16, input = key_cache)[name = string("coreml_update_state_60_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_60 = read_state(input = key_cache)[name = string("coreml_update_state_60")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_3_stride_0 = const()[name = string("value_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_5_cast_fp16 = transpose(perm = v_state_5_perm_0, x = var_609_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_43, begin_mask = value_cache_internal_tensor_assign_3_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_3_squeeze_mask_0, stride = value_cache_internal_tensor_assign_3_stride_0, update = v_state_5_cast_fp16, x = coreml_update_state_59)[name = string("value_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_3_cast_fp16, input = value_cache)[name = string("coreml_update_state_61_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_61 = read_state(input = value_cache)[name = string("coreml_update_state_61")];
            tensor<int32, [5]> var_666_begin_0 = const()[name = string("op_666_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> var_666_end_0 = const()[name = string("op_666_end_0"), val = tensor<int32, [5]>([3, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_666_end_mask_0 = const()[name = string("op_666_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_666_squeeze_mask_0 = const()[name = string("op_666_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_666_cast_fp16 = slice_by_index(begin = var_666_begin_0, end = var_666_end_0, end_mask = var_666_end_mask_0, squeeze_mask = var_666_squeeze_mask_0, x = coreml_update_state_60)[name = string("op_666_cast_fp16")];
            tensor<int32, [4]> var_669_begin_0 = const()[name = string("op_669_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_669_end_mask_0 = const()[name = string("op_669_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_669_cast_fp16 = slice_by_index(begin = var_669_begin_0, end = concat_11, end_mask = var_669_end_mask_0, x = var_666_cast_fp16)[name = string("op_669_cast_fp16")];
            tensor<int32, [5]> var_671_begin_0 = const()[name = string("op_671_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> var_671_end_0 = const()[name = string("op_671_end_0"), val = tensor<int32, [5]>([3, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_671_end_mask_0 = const()[name = string("op_671_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_671_squeeze_mask_0 = const()[name = string("op_671_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_671_cast_fp16 = slice_by_index(begin = var_671_begin_0, end = var_671_end_0, end_mask = var_671_end_mask_0, squeeze_mask = var_671_squeeze_mask_0, x = coreml_update_state_61)[name = string("op_671_cast_fp16")];
            tensor<int32, [4]> var_674_begin_0 = const()[name = string("op_674_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_674_end_mask_0 = const()[name = string("op_674_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_674_cast_fp16 = slice_by_index(begin = var_674_begin_0, end = concat_11, end_mask = var_674_end_mask_0, x = var_671_cast_fp16)[name = string("op_674_cast_fp16")];
            tensor<int32, [4]> var_676_shape_cast_fp16 = shape(x = var_669_cast_fp16)[name = string("op_676_shape_cast_fp16")];
            int32 gather_49 = const()[name = string("gather_49"), val = int32(1)];
            int32 gather_50 = const()[name = string("gather_50"), val = int32(2)];
            int32 gather_51_axis_0 = const()[name = string("gather_51_axis_0"), val = int32(0)];
            int32 gather_51_batch_dims_0 = const()[name = string("gather_51_batch_dims_0"), val = int32(0)];
            bool gather_51_validate_indices_0 = const()[name = string("gather_51_validate_indices_0"), val = bool(false)];
            string var_676_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_676_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_51_to_uint16 = const()[name = string("select_51_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_676_shape_cast_fp16_to_uint16 = cast(dtype = var_676_shape_cast_fp16_to_uint16_dtype_0, x = var_676_shape_cast_fp16)[name = string("cast_155")];
            uint16 gather_51_cast_uint16 = gather(axis = gather_51_axis_0, batch_dims = gather_51_batch_dims_0, indices = select_51_to_uint16, validate_indices = gather_51_validate_indices_0, x = var_676_shape_cast_fp16_to_uint16)[name = string("gather_51_cast_uint16")];
            string gather_51_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_51_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_52 = const()[name = string("gather_52"), val = int32(128)];
            tensor<int32, [1]> var_683_axes_0 = const()[name = string("op_683_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_683_cast_fp16 = expand_dims(axes = var_683_axes_0, x = var_669_cast_fp16)[name = string("op_683_cast_fp16")];
            tensor<int32, [5]> shape_57_cast_fp16 = shape(x = var_683_cast_fp16)[name = string("shape_57_cast_fp16")];
            int32 concat_51_axis_0 = const()[name = string("concat_51_axis_0"), val = int32(0)];
            bool concat_51_interleave_0 = const()[name = string("concat_51_interleave_0"), val = bool(false)];
            int32 gather_51_cast_uint16_to_int32 = cast(dtype = gather_51_cast_uint16_to_int32_dtype_0, x = gather_51_cast_uint16)[name = string("cast_154")];
            tensor<int32, [5]> concat_51 = concat(axis = concat_51_axis_0, interleave = concat_51_interleave_0, values = (gather_49, gather_50, var_77, gather_51_cast_uint16_to_int32, gather_52))[name = string("concat_51")];
            tensor<int32, [5]> real_div_4 = real_div(x = concat_51, y = shape_57_cast_fp16)[name = string("real_div_4")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_71_cast_fp16 = tile(reps = real_div_4, x = var_683_cast_fp16)[name = string("hidden_states_71_cast_fp16")];
            tensor<int32, [4]> concat_52x = const()[name = string("concat_52x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_11_cast_fp16 = reshape(shape = concat_52x, x = hidden_states_71_cast_fp16)[name = string("key_states_11_cast_fp16")];
            tensor<int32, [4]> var_693_shape_cast_fp16 = shape(x = var_674_cast_fp16)[name = string("op_693_shape_cast_fp16")];
            int32 gather_53 = const()[name = string("gather_53"), val = int32(1)];
            int32 gather_54 = const()[name = string("gather_54"), val = int32(2)];
            int32 gather_55_axis_0 = const()[name = string("gather_55_axis_0"), val = int32(0)];
            int32 gather_55_batch_dims_0 = const()[name = string("gather_55_batch_dims_0"), val = int32(0)];
            bool gather_55_validate_indices_0 = const()[name = string("gather_55_validate_indices_0"), val = bool(false)];
            string var_693_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_693_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_55_to_uint16 = const()[name = string("select_55_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_693_shape_cast_fp16_to_uint16 = cast(dtype = var_693_shape_cast_fp16_to_uint16_dtype_0, x = var_693_shape_cast_fp16)[name = string("cast_153")];
            uint16 gather_55_cast_uint16 = gather(axis = gather_55_axis_0, batch_dims = gather_55_batch_dims_0, indices = select_55_to_uint16, validate_indices = gather_55_validate_indices_0, x = var_693_shape_cast_fp16_to_uint16)[name = string("gather_55_cast_uint16")];
            string gather_55_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_55_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_56 = const()[name = string("gather_56"), val = int32(128)];
            tensor<int32, [1]> var_700_axes_0 = const()[name = string("op_700_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_700_cast_fp16 = expand_dims(axes = var_700_axes_0, x = var_674_cast_fp16)[name = string("op_700_cast_fp16")];
            tensor<int32, [5]> shape_62_cast_fp16 = shape(x = var_700_cast_fp16)[name = string("shape_62_cast_fp16")];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            int32 gather_55_cast_uint16_to_int32 = cast(dtype = gather_55_cast_uint16_to_int32_dtype_0, x = gather_55_cast_uint16)[name = string("cast_152")];
            tensor<int32, [5]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (gather_53, gather_54, var_77, gather_55_cast_uint16_to_int32, gather_56))[name = string("concat_53")];
            tensor<int32, [5]> real_div_5 = real_div(x = concat_53, y = shape_62_cast_fp16)[name = string("real_div_5")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_75_cast_fp16 = tile(reps = real_div_5, x = var_700_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            tensor<int32, [4]> concat_54x = const()[name = string("concat_54x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_11_cast_fp16 = reshape(shape = concat_54x, x = hidden_states_75_cast_fp16)[name = string("value_states_11_cast_fp16")];
            tensor<int32, [4]> var_710_shape_cast_fp16 = shape(x = key_states_11_cast_fp16)[name = string("op_710_shape_cast_fp16")];
            int32 gather_57_axis_0 = const()[name = string("gather_57_axis_0"), val = int32(0)];
            int32 gather_57_batch_dims_0 = const()[name = string("gather_57_batch_dims_0"), val = int32(0)];
            bool gather_57_validate_indices_0 = const()[name = string("gather_57_validate_indices_0"), val = bool(false)];
            string var_710_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_710_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_57_to_uint16 = const()[name = string("select_57_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_710_shape_cast_fp16_to_uint16 = cast(dtype = var_710_shape_cast_fp16_to_uint16_dtype_0, x = var_710_shape_cast_fp16)[name = string("cast_151")];
            uint16 gather_57_cast_uint16 = gather(axis = gather_57_axis_0, batch_dims = gather_57_batch_dims_0, indices = select_57_to_uint16, validate_indices = gather_57_validate_indices_0, x = var_710_shape_cast_fp16_to_uint16)[name = string("gather_57_cast_uint16")];
            string gather_57_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_57_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_55_values0_0 = const()[name = string("concat_55_values0_0"), val = int32(1)];
            int32 concat_55_values1_0 = const()[name = string("concat_55_values1_0"), val = int32(1)];
            int32 concat_55_values2_0 = const()[name = string("concat_55_values2_0"), val = int32(0)];
            int32 concat_55_axis_0 = const()[name = string("concat_55_axis_0"), val = int32(0)];
            bool concat_55_interleave_0 = const()[name = string("concat_55_interleave_0"), val = bool(false)];
            int32 gather_57_cast_uint16_to_int32 = cast(dtype = gather_57_cast_uint16_to_int32_dtype_0, x = gather_57_cast_uint16)[name = string("cast_150")];
            tensor<int32, [4]> concat_55 = concat(axis = concat_55_axis_0, interleave = concat_55_interleave_0, values = (concat_55_values0_0, concat_55_values1_0, concat_55_values2_0, gather_57_cast_uint16_to_int32))[name = string("concat_55")];
            tensor<int32, [4]> causal_mask_7_begin_0 = const()[name = string("causal_mask_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_7_end_mask_0 = const()[name = string("causal_mask_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_7_cast_fp16 = slice_by_index(begin = causal_mask_7_begin_0, end = concat_55, end_mask = causal_mask_7_end_mask_0, x = causal_mask)[name = string("causal_mask_7_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_9_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_7_cast_fp16, key = key_states_11_cast_fp16, query = query_states_11_cast_fp16, value = value_states_11_cast_fp16)[name = string("attn_output_9_cast_fp16")];
            tensor<int32, [4]> var_716_perm_0 = const()[name = string("op_716_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_56x = const()[name = string("concat_56x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_716_cast_fp16 = transpose(perm = var_716_perm_0, x = attn_output_9_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, ?, 1536]> input_17_cast_fp16 = reshape(shape = concat_56x, x = var_716_cast_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185736192))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186915904))))[name = string("model_model_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_17_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_79_cast_fp16 = add(x = hidden_states_59_cast_fp16, y = linear_17_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            fp16 var_78_promoted_5_to_fp16 = const()[name = string("op_78_promoted_5_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_725_cast_fp16 = pow(x = hidden_states_79_cast_fp16, y = var_78_promoted_5_to_fp16)[name = string("op_725_cast_fp16")];
            tensor<int32, [1]> variance_11_axes_0 = const()[name = string("variance_11_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_11_keep_dims_0 = const()[name = string("variance_11_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_11_cast_fp16 = reduce_mean(axes = variance_11_axes_0, keep_dims = variance_11_keep_dims_0, x = var_725_cast_fp16)[name = string("variance_11_cast_fp16")];
            fp16 var_728_to_fp16 = const()[name = string("op_728_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_729_cast_fp16 = add(x = variance_11_cast_fp16, y = var_728_to_fp16)[name = string("op_729_cast_fp16")];
            fp32 var_730_epsilon_0 = const()[name = string("op_730_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_730_cast_fp16 = rsqrt(epsilon = var_730_epsilon_0, x = var_729_cast_fp16)[name = string("op_730_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_83_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = var_730_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187063424)))];
            tensor<fp16, [1, ?, 1536]> input_19_cast_fp16 = mul(x = model_model_layers_2_post_attention_layernorm_weight_to_fp16, y = hidden_states_83_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187066560))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193947904))))[name = string("model_model_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_742_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_742_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194808128))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201689472))))[name = string("model_model_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_23_cast_fp16 = mul(x = var_742_cast_fp16, y = linear_19_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202549696))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209431040))))[name = string("model_model_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_20_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_89_cast_fp16 = add(x = hidden_states_79_cast_fp16, y = linear_20_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            fp16 var_78_promoted_6_to_fp16 = const()[name = string("op_78_promoted_6_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_755_cast_fp16 = pow(x = hidden_states_89_cast_fp16, y = var_78_promoted_6_to_fp16)[name = string("op_755_cast_fp16")];
            tensor<int32, [1]> variance_13_axes_0 = const()[name = string("variance_13_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_13_keep_dims_0 = const()[name = string("variance_13_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_13_cast_fp16 = reduce_mean(axes = variance_13_axes_0, keep_dims = variance_13_keep_dims_0, x = var_755_cast_fp16)[name = string("variance_13_cast_fp16")];
            fp16 var_758_to_fp16 = const()[name = string("op_758_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_759_cast_fp16 = add(x = variance_13_cast_fp16, y = var_758_to_fp16)[name = string("op_759_cast_fp16")];
            fp32 var_760_epsilon_0 = const()[name = string("op_760_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_760_cast_fp16 = rsqrt(epsilon = var_760_epsilon_0, x = var_759_cast_fp16)[name = string("op_760_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_93_cast_fp16 = mul(x = hidden_states_89_cast_fp16, y = var_760_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210291264)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_97_cast_fp16 = mul(x = model_model_layers_3_input_layernorm_weight_to_fp16, y = hidden_states_93_cast_fp16)[name = string("hidden_states_97_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210294400))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211474112))))[name = string("model_model_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211621632)))];
            tensor<fp16, [1, ?, 1536]> linear_21_cast_fp16 = linear(bias = model_model_layers_3_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_97_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211624768))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211821440))))[name = string("model_model_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_3_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_3_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211846080)))];
            tensor<fp16, [1, ?, 256]> linear_22_cast_fp16 = linear(bias = model_model_layers_3_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_97_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211846656))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212043328))))[name = string("model_model_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212067968)))];
            tensor<fp16, [1, ?, 256]> linear_23_cast_fp16 = linear(bias = model_model_layers_3_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_97_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> concat_57x = const()[name = string("concat_57x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_783_cast_fp16 = reshape(shape = concat_57x, x = linear_21_cast_fp16)[name = string("op_783_cast_fp16")];
            tensor<int32, [4]> q_7_perm_0 = const()[name = string("q_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_58x = const()[name = string("concat_58x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_786_cast_fp16 = reshape(shape = concat_58x, x = linear_22_cast_fp16)[name = string("op_786_cast_fp16")];
            tensor<int32, [4]> k_7_perm_0 = const()[name = string("k_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_59x = const()[name = string("concat_59x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_789_cast_fp16 = reshape(shape = concat_59x, x = linear_23_cast_fp16)[name = string("op_789_cast_fp16")];
            tensor<int32, [4]> v_state_7_perm_0 = const()[name = string("v_state_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_7_cast_fp16 = transpose(perm = q_7_perm_0, x = var_783_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 12, ?, 128]> var_793_cast_fp16 = mul(x = q_7_cast_fp16, y = cos_7_cast_fp16)[name = string("op_793_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_7_cast_fp16)[name = string("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_7_cast_fp16)[name = string("x2_13_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_804_cast_fp16 = mul(x = x2_13_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_804_cast_fp16")];
            bool var_806_interleave_0 = const()[name = string("op_806_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_806_cast_fp16 = concat(axis = var_72, interleave = var_806_interleave_0, values = (var_804_cast_fp16, x1_13_cast_fp16))[name = string("op_806_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_807_cast_fp16 = mul(x = var_806_cast_fp16, y = sin_7_cast_fp16)[name = string("op_807_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_15_cast_fp16 = add(x = var_793_cast_fp16, y = var_807_cast_fp16)[name = string("query_states_15_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_7_cast_fp16 = transpose(perm = k_7_perm_0, x = var_786_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 2, ?, 128]> var_809_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_7_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_7_cast_fp16)[name = string("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_7_cast_fp16)[name = string("x2_15_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_820_cast_fp16 = mul(x = x2_15_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_820_cast_fp16")];
            bool var_822_interleave_0 = const()[name = string("op_822_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_822_cast_fp16 = concat(axis = var_72, interleave = var_822_interleave_0, values = (var_820_cast_fp16, x1_15_cast_fp16))[name = string("op_822_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_823_cast_fp16 = mul(x = var_822_cast_fp16, y = sin_7_cast_fp16)[name = string("op_823_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_7_cast_fp16 = add(x = var_809_cast_fp16, y = var_823_cast_fp16)[name = string("k_state_7_cast_fp16")];
            tensor<int32, [1]> expand_dims_36 = const()[name = string("expand_dims_36"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_37 = const()[name = string("expand_dims_37"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_39 = const()[name = string("expand_dims_39"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_62_values0_0 = const()[name = string("concat_62_values0_0"), val = tensor<int32, [1]>([3])];
            int32 concat_62_axis_0 = const()[name = string("concat_62_axis_0"), val = int32(0)];
            bool concat_62_interleave_0 = const()[name = string("concat_62_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_62 = concat(axis = concat_62_axis_0, interleave = concat_62_interleave_0, values = (concat_62_values0_0, expand_dims_36, expand_dims_37, expand_dims_2, expand_dims_39))[name = string("concat_62")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_4_stride_0 = const()[name = string("key_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_62, begin_mask = key_cache_internal_tensor_assign_4_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_4_squeeze_mask_0, stride = key_cache_internal_tensor_assign_4_stride_0, update = k_state_7_cast_fp16, x = coreml_update_state_60)[name = string("key_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_4_cast_fp16, input = key_cache)[name = string("coreml_update_state_62_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_62 = read_state(input = key_cache)[name = string("coreml_update_state_62")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_4_stride_0 = const()[name = string("value_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_7_cast_fp16 = transpose(perm = v_state_7_perm_0, x = var_789_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_62, begin_mask = value_cache_internal_tensor_assign_4_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_4_squeeze_mask_0, stride = value_cache_internal_tensor_assign_4_stride_0, update = v_state_7_cast_fp16, x = coreml_update_state_61)[name = string("value_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_4_cast_fp16, input = value_cache)[name = string("coreml_update_state_63_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_63 = read_state(input = value_cache)[name = string("coreml_update_state_63")];
            tensor<int32, [5]> var_846_begin_0 = const()[name = string("op_846_begin_0"), val = tensor<int32, [5]>([3, 0, 0, 0, 0])];
            tensor<int32, [5]> var_846_end_0 = const()[name = string("op_846_end_0"), val = tensor<int32, [5]>([4, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_846_end_mask_0 = const()[name = string("op_846_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_846_squeeze_mask_0 = const()[name = string("op_846_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_846_cast_fp16 = slice_by_index(begin = var_846_begin_0, end = var_846_end_0, end_mask = var_846_end_mask_0, squeeze_mask = var_846_squeeze_mask_0, x = coreml_update_state_62)[name = string("op_846_cast_fp16")];
            tensor<int32, [4]> var_849_begin_0 = const()[name = string("op_849_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_849_end_mask_0 = const()[name = string("op_849_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_849_cast_fp16 = slice_by_index(begin = var_849_begin_0, end = concat_11, end_mask = var_849_end_mask_0, x = var_846_cast_fp16)[name = string("op_849_cast_fp16")];
            tensor<int32, [5]> var_851_begin_0 = const()[name = string("op_851_begin_0"), val = tensor<int32, [5]>([3, 0, 0, 0, 0])];
            tensor<int32, [5]> var_851_end_0 = const()[name = string("op_851_end_0"), val = tensor<int32, [5]>([4, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_851_end_mask_0 = const()[name = string("op_851_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_851_squeeze_mask_0 = const()[name = string("op_851_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_851_cast_fp16 = slice_by_index(begin = var_851_begin_0, end = var_851_end_0, end_mask = var_851_end_mask_0, squeeze_mask = var_851_squeeze_mask_0, x = coreml_update_state_63)[name = string("op_851_cast_fp16")];
            tensor<int32, [4]> var_854_begin_0 = const()[name = string("op_854_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_854_end_mask_0 = const()[name = string("op_854_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_854_cast_fp16 = slice_by_index(begin = var_854_begin_0, end = concat_11, end_mask = var_854_end_mask_0, x = var_851_cast_fp16)[name = string("op_854_cast_fp16")];
            tensor<int32, [4]> var_856_shape_cast_fp16 = shape(x = var_849_cast_fp16)[name = string("op_856_shape_cast_fp16")];
            int32 gather_67 = const()[name = string("gather_67"), val = int32(1)];
            int32 gather_68 = const()[name = string("gather_68"), val = int32(2)];
            int32 gather_69_axis_0 = const()[name = string("gather_69_axis_0"), val = int32(0)];
            int32 gather_69_batch_dims_0 = const()[name = string("gather_69_batch_dims_0"), val = int32(0)];
            bool gather_69_validate_indices_0 = const()[name = string("gather_69_validate_indices_0"), val = bool(false)];
            string var_856_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_856_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_69_to_uint16 = const()[name = string("select_69_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_856_shape_cast_fp16_to_uint16 = cast(dtype = var_856_shape_cast_fp16_to_uint16_dtype_0, x = var_856_shape_cast_fp16)[name = string("cast_149")];
            uint16 gather_69_cast_uint16 = gather(axis = gather_69_axis_0, batch_dims = gather_69_batch_dims_0, indices = select_69_to_uint16, validate_indices = gather_69_validate_indices_0, x = var_856_shape_cast_fp16_to_uint16)[name = string("gather_69_cast_uint16")];
            string gather_69_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_69_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_70 = const()[name = string("gather_70"), val = int32(128)];
            tensor<int32, [1]> var_863_axes_0 = const()[name = string("op_863_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_863_cast_fp16 = expand_dims(axes = var_863_axes_0, x = var_849_cast_fp16)[name = string("op_863_cast_fp16")];
            tensor<int32, [5]> shape_77_cast_fp16 = shape(x = var_863_cast_fp16)[name = string("shape_77_cast_fp16")];
            int32 concat_70_axis_0 = const()[name = string("concat_70_axis_0"), val = int32(0)];
            bool concat_70_interleave_0 = const()[name = string("concat_70_interleave_0"), val = bool(false)];
            int32 gather_69_cast_uint16_to_int32 = cast(dtype = gather_69_cast_uint16_to_int32_dtype_0, x = gather_69_cast_uint16)[name = string("cast_148")];
            tensor<int32, [5]> concat_70 = concat(axis = concat_70_axis_0, interleave = concat_70_interleave_0, values = (gather_67, gather_68, var_77, gather_69_cast_uint16_to_int32, gather_70))[name = string("concat_70")];
            tensor<int32, [5]> real_div_6 = real_div(x = concat_70, y = shape_77_cast_fp16)[name = string("real_div_6")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_101_cast_fp16 = tile(reps = real_div_6, x = var_863_cast_fp16)[name = string("hidden_states_101_cast_fp16")];
            tensor<int32, [4]> concat_71x = const()[name = string("concat_71x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_15_cast_fp16 = reshape(shape = concat_71x, x = hidden_states_101_cast_fp16)[name = string("key_states_15_cast_fp16")];
            tensor<int32, [4]> var_873_shape_cast_fp16 = shape(x = var_854_cast_fp16)[name = string("op_873_shape_cast_fp16")];
            int32 gather_71 = const()[name = string("gather_71"), val = int32(1)];
            int32 gather_72 = const()[name = string("gather_72"), val = int32(2)];
            int32 gather_73_axis_0 = const()[name = string("gather_73_axis_0"), val = int32(0)];
            int32 gather_73_batch_dims_0 = const()[name = string("gather_73_batch_dims_0"), val = int32(0)];
            bool gather_73_validate_indices_0 = const()[name = string("gather_73_validate_indices_0"), val = bool(false)];
            string var_873_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_873_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_73_to_uint16 = const()[name = string("select_73_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_873_shape_cast_fp16_to_uint16 = cast(dtype = var_873_shape_cast_fp16_to_uint16_dtype_0, x = var_873_shape_cast_fp16)[name = string("cast_147")];
            uint16 gather_73_cast_uint16 = gather(axis = gather_73_axis_0, batch_dims = gather_73_batch_dims_0, indices = select_73_to_uint16, validate_indices = gather_73_validate_indices_0, x = var_873_shape_cast_fp16_to_uint16)[name = string("gather_73_cast_uint16")];
            string gather_73_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_73_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_74 = const()[name = string("gather_74"), val = int32(128)];
            tensor<int32, [1]> var_880_axes_0 = const()[name = string("op_880_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_880_cast_fp16 = expand_dims(axes = var_880_axes_0, x = var_854_cast_fp16)[name = string("op_880_cast_fp16")];
            tensor<int32, [5]> shape_82_cast_fp16 = shape(x = var_880_cast_fp16)[name = string("shape_82_cast_fp16")];
            int32 concat_72_axis_0 = const()[name = string("concat_72_axis_0"), val = int32(0)];
            bool concat_72_interleave_0 = const()[name = string("concat_72_interleave_0"), val = bool(false)];
            int32 gather_73_cast_uint16_to_int32 = cast(dtype = gather_73_cast_uint16_to_int32_dtype_0, x = gather_73_cast_uint16)[name = string("cast_146")];
            tensor<int32, [5]> concat_72 = concat(axis = concat_72_axis_0, interleave = concat_72_interleave_0, values = (gather_71, gather_72, var_77, gather_73_cast_uint16_to_int32, gather_74))[name = string("concat_72")];
            tensor<int32, [5]> real_div_7 = real_div(x = concat_72, y = shape_82_cast_fp16)[name = string("real_div_7")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_105_cast_fp16 = tile(reps = real_div_7, x = var_880_cast_fp16)[name = string("hidden_states_105_cast_fp16")];
            tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_15_cast_fp16 = reshape(shape = concat_73x, x = hidden_states_105_cast_fp16)[name = string("value_states_15_cast_fp16")];
            tensor<int32, [4]> var_890_shape_cast_fp16 = shape(x = key_states_15_cast_fp16)[name = string("op_890_shape_cast_fp16")];
            int32 gather_75_axis_0 = const()[name = string("gather_75_axis_0"), val = int32(0)];
            int32 gather_75_batch_dims_0 = const()[name = string("gather_75_batch_dims_0"), val = int32(0)];
            bool gather_75_validate_indices_0 = const()[name = string("gather_75_validate_indices_0"), val = bool(false)];
            string var_890_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_890_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_75_to_uint16 = const()[name = string("select_75_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_890_shape_cast_fp16_to_uint16 = cast(dtype = var_890_shape_cast_fp16_to_uint16_dtype_0, x = var_890_shape_cast_fp16)[name = string("cast_145")];
            uint16 gather_75_cast_uint16 = gather(axis = gather_75_axis_0, batch_dims = gather_75_batch_dims_0, indices = select_75_to_uint16, validate_indices = gather_75_validate_indices_0, x = var_890_shape_cast_fp16_to_uint16)[name = string("gather_75_cast_uint16")];
            string gather_75_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_75_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_74_values0_0 = const()[name = string("concat_74_values0_0"), val = int32(1)];
            int32 concat_74_values1_0 = const()[name = string("concat_74_values1_0"), val = int32(1)];
            int32 concat_74_values2_0 = const()[name = string("concat_74_values2_0"), val = int32(0)];
            int32 concat_74_axis_0 = const()[name = string("concat_74_axis_0"), val = int32(0)];
            bool concat_74_interleave_0 = const()[name = string("concat_74_interleave_0"), val = bool(false)];
            int32 gather_75_cast_uint16_to_int32 = cast(dtype = gather_75_cast_uint16_to_int32_dtype_0, x = gather_75_cast_uint16)[name = string("cast_144")];
            tensor<int32, [4]> concat_74 = concat(axis = concat_74_axis_0, interleave = concat_74_interleave_0, values = (concat_74_values0_0, concat_74_values1_0, concat_74_values2_0, gather_75_cast_uint16_to_int32))[name = string("concat_74")];
            tensor<int32, [4]> causal_mask_9_begin_0 = const()[name = string("causal_mask_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_9_end_mask_0 = const()[name = string("causal_mask_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_9_cast_fp16 = slice_by_index(begin = causal_mask_9_begin_0, end = concat_74, end_mask = causal_mask_9_end_mask_0, x = causal_mask)[name = string("causal_mask_9_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_13_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_9_cast_fp16, key = key_states_15_cast_fp16, query = query_states_15_cast_fp16, value = value_states_15_cast_fp16)[name = string("attn_output_13_cast_fp16")];
            tensor<int32, [4]> var_896_perm_0 = const()[name = string("op_896_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_75x = const()[name = string("concat_75x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_896_cast_fp16 = transpose(perm = var_896_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, ?, 1536]> input_25_cast_fp16 = reshape(shape = concat_75x, x = var_896_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212068544))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213248256))))[name = string("model_model_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_24_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_109_cast_fp16 = add(x = hidden_states_89_cast_fp16, y = linear_24_cast_fp16)[name = string("hidden_states_109_cast_fp16")];
            fp16 var_78_promoted_7_to_fp16 = const()[name = string("op_78_promoted_7_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_905_cast_fp16 = pow(x = hidden_states_109_cast_fp16, y = var_78_promoted_7_to_fp16)[name = string("op_905_cast_fp16")];
            tensor<int32, [1]> variance_15_axes_0 = const()[name = string("variance_15_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_15_keep_dims_0 = const()[name = string("variance_15_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_15_cast_fp16 = reduce_mean(axes = variance_15_axes_0, keep_dims = variance_15_keep_dims_0, x = var_905_cast_fp16)[name = string("variance_15_cast_fp16")];
            fp16 var_908_to_fp16 = const()[name = string("op_908_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_909_cast_fp16 = add(x = variance_15_cast_fp16, y = var_908_to_fp16)[name = string("op_909_cast_fp16")];
            fp32 var_910_epsilon_0 = const()[name = string("op_910_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_910_cast_fp16 = rsqrt(epsilon = var_910_epsilon_0, x = var_909_cast_fp16)[name = string("op_910_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_113_cast_fp16 = mul(x = hidden_states_109_cast_fp16, y = var_910_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213395776)))];
            tensor<fp16, [1, ?, 1536]> input_27_cast_fp16 = mul(x = model_model_layers_3_post_attention_layernorm_weight_to_fp16, y = hidden_states_113_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213398912))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220280256))))[name = string("model_model_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_922_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_922_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221140480))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228021824))))[name = string("model_model_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_31_cast_fp16 = mul(x = var_922_cast_fp16, y = linear_26_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228882048))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235763392))))[name = string("model_model_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_119_cast_fp16 = add(x = hidden_states_109_cast_fp16, y = linear_27_cast_fp16)[name = string("hidden_states_119_cast_fp16")];
            fp16 var_78_promoted_8_to_fp16 = const()[name = string("op_78_promoted_8_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_935_cast_fp16 = pow(x = hidden_states_119_cast_fp16, y = var_78_promoted_8_to_fp16)[name = string("op_935_cast_fp16")];
            tensor<int32, [1]> variance_17_axes_0 = const()[name = string("variance_17_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_17_keep_dims_0 = const()[name = string("variance_17_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_17_cast_fp16 = reduce_mean(axes = variance_17_axes_0, keep_dims = variance_17_keep_dims_0, x = var_935_cast_fp16)[name = string("variance_17_cast_fp16")];
            fp16 var_938_to_fp16 = const()[name = string("op_938_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_939_cast_fp16 = add(x = variance_17_cast_fp16, y = var_938_to_fp16)[name = string("op_939_cast_fp16")];
            fp32 var_940_epsilon_0 = const()[name = string("op_940_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_940_cast_fp16 = rsqrt(epsilon = var_940_epsilon_0, x = var_939_cast_fp16)[name = string("op_940_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_123_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = var_940_cast_fp16)[name = string("hidden_states_123_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236623616)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_127_cast_fp16 = mul(x = model_model_layers_4_input_layernorm_weight_to_fp16, y = hidden_states_123_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236626752))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237806464))))[name = string("model_model_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237953984)))];
            tensor<fp16, [1, ?, 1536]> linear_28_cast_fp16 = linear(bias = model_model_layers_4_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237957120))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238153792))))[name = string("model_model_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_4_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_4_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238178432)))];
            tensor<fp16, [1, ?, 256]> linear_29_cast_fp16 = linear(bias = model_model_layers_4_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238179008))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238375680))))[name = string("model_model_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238400320)))];
            tensor<fp16, [1, ?, 256]> linear_30_cast_fp16 = linear(bias = model_model_layers_4_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> concat_76x = const()[name = string("concat_76x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_963_cast_fp16 = reshape(shape = concat_76x, x = linear_28_cast_fp16)[name = string("op_963_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = string("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_77x = const()[name = string("concat_77x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_966_cast_fp16 = reshape(shape = concat_77x, x = linear_29_cast_fp16)[name = string("op_966_cast_fp16")];
            tensor<int32, [4]> k_9_perm_0 = const()[name = string("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_78x = const()[name = string("concat_78x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_969_cast_fp16 = reshape(shape = concat_78x, x = linear_30_cast_fp16)[name = string("op_969_cast_fp16")];
            tensor<int32, [4]> v_state_9_perm_0 = const()[name = string("v_state_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_963_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 12, ?, 128]> var_973_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_7_cast_fp16)[name = string("op_973_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_9_cast_fp16)[name = string("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_9_cast_fp16)[name = string("x2_17_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_984_cast_fp16 = mul(x = x2_17_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_984_cast_fp16")];
            bool var_986_interleave_0 = const()[name = string("op_986_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_986_cast_fp16 = concat(axis = var_72, interleave = var_986_interleave_0, values = (var_984_cast_fp16, x1_17_cast_fp16))[name = string("op_986_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_987_cast_fp16 = mul(x = var_986_cast_fp16, y = sin_7_cast_fp16)[name = string("op_987_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_19_cast_fp16 = add(x = var_973_cast_fp16, y = var_987_cast_fp16)[name = string("query_states_19_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = var_966_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 2, ?, 128]> var_989_cast_fp16 = mul(x = k_9_cast_fp16, y = cos_7_cast_fp16)[name = string("op_989_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_9_cast_fp16)[name = string("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_9_cast_fp16)[name = string("x2_19_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_1000_cast_fp16 = mul(x = x2_19_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_1000_cast_fp16")];
            bool var_1002_interleave_0 = const()[name = string("op_1002_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_1002_cast_fp16 = concat(axis = var_72, interleave = var_1002_interleave_0, values = (var_1000_cast_fp16, x1_19_cast_fp16))[name = string("op_1002_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_1003_cast_fp16 = mul(x = var_1002_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1003_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_9_cast_fp16 = add(x = var_989_cast_fp16, y = var_1003_cast_fp16)[name = string("k_state_9_cast_fp16")];
            tensor<int32, [1]> expand_dims_48 = const()[name = string("expand_dims_48"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_49 = const()[name = string("expand_dims_49"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_51 = const()[name = string("expand_dims_51"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_81_values0_0 = const()[name = string("concat_81_values0_0"), val = tensor<int32, [1]>([4])];
            int32 concat_81_axis_0 = const()[name = string("concat_81_axis_0"), val = int32(0)];
            bool concat_81_interleave_0 = const()[name = string("concat_81_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_81 = concat(axis = concat_81_axis_0, interleave = concat_81_interleave_0, values = (concat_81_values0_0, expand_dims_48, expand_dims_49, expand_dims_2, expand_dims_51))[name = string("concat_81")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_5_stride_0 = const()[name = string("key_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_81, begin_mask = key_cache_internal_tensor_assign_5_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_5_squeeze_mask_0, stride = key_cache_internal_tensor_assign_5_stride_0, update = k_state_9_cast_fp16, x = coreml_update_state_62)[name = string("key_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_5_cast_fp16, input = key_cache)[name = string("coreml_update_state_64_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_64 = read_state(input = key_cache)[name = string("coreml_update_state_64")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_5_stride_0 = const()[name = string("value_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_9_cast_fp16 = transpose(perm = v_state_9_perm_0, x = var_969_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_81, begin_mask = value_cache_internal_tensor_assign_5_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_5_squeeze_mask_0, stride = value_cache_internal_tensor_assign_5_stride_0, update = v_state_9_cast_fp16, x = coreml_update_state_63)[name = string("value_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_5_cast_fp16, input = value_cache)[name = string("coreml_update_state_65_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_65 = read_state(input = value_cache)[name = string("coreml_update_state_65")];
            tensor<int32, [5]> var_1026_begin_0 = const()[name = string("op_1026_begin_0"), val = tensor<int32, [5]>([4, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1026_end_0 = const()[name = string("op_1026_end_0"), val = tensor<int32, [5]>([5, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_1026_end_mask_0 = const()[name = string("op_1026_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1026_squeeze_mask_0 = const()[name = string("op_1026_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_1026_cast_fp16 = slice_by_index(begin = var_1026_begin_0, end = var_1026_end_0, end_mask = var_1026_end_mask_0, squeeze_mask = var_1026_squeeze_mask_0, x = coreml_update_state_64)[name = string("op_1026_cast_fp16")];
            tensor<int32, [4]> var_1029_begin_0 = const()[name = string("op_1029_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1029_end_mask_0 = const()[name = string("op_1029_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_1029_cast_fp16 = slice_by_index(begin = var_1029_begin_0, end = concat_11, end_mask = var_1029_end_mask_0, x = var_1026_cast_fp16)[name = string("op_1029_cast_fp16")];
            tensor<int32, [5]> var_1031_begin_0 = const()[name = string("op_1031_begin_0"), val = tensor<int32, [5]>([4, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1031_end_0 = const()[name = string("op_1031_end_0"), val = tensor<int32, [5]>([5, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_1031_end_mask_0 = const()[name = string("op_1031_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1031_squeeze_mask_0 = const()[name = string("op_1031_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_1031_cast_fp16 = slice_by_index(begin = var_1031_begin_0, end = var_1031_end_0, end_mask = var_1031_end_mask_0, squeeze_mask = var_1031_squeeze_mask_0, x = coreml_update_state_65)[name = string("op_1031_cast_fp16")];
            tensor<int32, [4]> var_1034_begin_0 = const()[name = string("op_1034_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1034_end_mask_0 = const()[name = string("op_1034_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_1034_cast_fp16 = slice_by_index(begin = var_1034_begin_0, end = concat_11, end_mask = var_1034_end_mask_0, x = var_1031_cast_fp16)[name = string("op_1034_cast_fp16")];
            tensor<int32, [4]> var_1036_shape_cast_fp16 = shape(x = var_1029_cast_fp16)[name = string("op_1036_shape_cast_fp16")];
            int32 gather_85 = const()[name = string("gather_85"), val = int32(1)];
            int32 gather_86 = const()[name = string("gather_86"), val = int32(2)];
            int32 gather_87_axis_0 = const()[name = string("gather_87_axis_0"), val = int32(0)];
            int32 gather_87_batch_dims_0 = const()[name = string("gather_87_batch_dims_0"), val = int32(0)];
            bool gather_87_validate_indices_0 = const()[name = string("gather_87_validate_indices_0"), val = bool(false)];
            string var_1036_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1036_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_87_to_uint16 = const()[name = string("select_87_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1036_shape_cast_fp16_to_uint16 = cast(dtype = var_1036_shape_cast_fp16_to_uint16_dtype_0, x = var_1036_shape_cast_fp16)[name = string("cast_143")];
            uint16 gather_87_cast_uint16 = gather(axis = gather_87_axis_0, batch_dims = gather_87_batch_dims_0, indices = select_87_to_uint16, validate_indices = gather_87_validate_indices_0, x = var_1036_shape_cast_fp16_to_uint16)[name = string("gather_87_cast_uint16")];
            string gather_87_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_87_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_88 = const()[name = string("gather_88"), val = int32(128)];
            tensor<int32, [1]> var_1043_axes_0 = const()[name = string("op_1043_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_1043_cast_fp16 = expand_dims(axes = var_1043_axes_0, x = var_1029_cast_fp16)[name = string("op_1043_cast_fp16")];
            tensor<int32, [5]> shape_97_cast_fp16 = shape(x = var_1043_cast_fp16)[name = string("shape_97_cast_fp16")];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            int32 gather_87_cast_uint16_to_int32 = cast(dtype = gather_87_cast_uint16_to_int32_dtype_0, x = gather_87_cast_uint16)[name = string("cast_142")];
            tensor<int32, [5]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (gather_85, gather_86, var_77, gather_87_cast_uint16_to_int32, gather_88))[name = string("concat_89")];
            tensor<int32, [5]> real_div_8 = real_div(x = concat_89, y = shape_97_cast_fp16)[name = string("real_div_8")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_131_cast_fp16 = tile(reps = real_div_8, x = var_1043_cast_fp16)[name = string("hidden_states_131_cast_fp16")];
            tensor<int32, [4]> concat_90x = const()[name = string("concat_90x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_19_cast_fp16 = reshape(shape = concat_90x, x = hidden_states_131_cast_fp16)[name = string("key_states_19_cast_fp16")];
            tensor<int32, [4]> var_1053_shape_cast_fp16 = shape(x = var_1034_cast_fp16)[name = string("op_1053_shape_cast_fp16")];
            int32 gather_89 = const()[name = string("gather_89"), val = int32(1)];
            int32 gather_90 = const()[name = string("gather_90"), val = int32(2)];
            int32 gather_91_axis_0 = const()[name = string("gather_91_axis_0"), val = int32(0)];
            int32 gather_91_batch_dims_0 = const()[name = string("gather_91_batch_dims_0"), val = int32(0)];
            bool gather_91_validate_indices_0 = const()[name = string("gather_91_validate_indices_0"), val = bool(false)];
            string var_1053_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1053_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_91_to_uint16 = const()[name = string("select_91_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1053_shape_cast_fp16_to_uint16 = cast(dtype = var_1053_shape_cast_fp16_to_uint16_dtype_0, x = var_1053_shape_cast_fp16)[name = string("cast_141")];
            uint16 gather_91_cast_uint16 = gather(axis = gather_91_axis_0, batch_dims = gather_91_batch_dims_0, indices = select_91_to_uint16, validate_indices = gather_91_validate_indices_0, x = var_1053_shape_cast_fp16_to_uint16)[name = string("gather_91_cast_uint16")];
            string gather_91_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_91_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_92 = const()[name = string("gather_92"), val = int32(128)];
            tensor<int32, [1]> var_1060_axes_0 = const()[name = string("op_1060_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_1060_cast_fp16 = expand_dims(axes = var_1060_axes_0, x = var_1034_cast_fp16)[name = string("op_1060_cast_fp16")];
            tensor<int32, [5]> shape_102_cast_fp16 = shape(x = var_1060_cast_fp16)[name = string("shape_102_cast_fp16")];
            int32 concat_91_axis_0 = const()[name = string("concat_91_axis_0"), val = int32(0)];
            bool concat_91_interleave_0 = const()[name = string("concat_91_interleave_0"), val = bool(false)];
            int32 gather_91_cast_uint16_to_int32 = cast(dtype = gather_91_cast_uint16_to_int32_dtype_0, x = gather_91_cast_uint16)[name = string("cast_140")];
            tensor<int32, [5]> concat_91 = concat(axis = concat_91_axis_0, interleave = concat_91_interleave_0, values = (gather_89, gather_90, var_77, gather_91_cast_uint16_to_int32, gather_92))[name = string("concat_91")];
            tensor<int32, [5]> real_div_9 = real_div(x = concat_91, y = shape_102_cast_fp16)[name = string("real_div_9")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_135_cast_fp16 = tile(reps = real_div_9, x = var_1060_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            tensor<int32, [4]> concat_92x = const()[name = string("concat_92x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_19_cast_fp16 = reshape(shape = concat_92x, x = hidden_states_135_cast_fp16)[name = string("value_states_19_cast_fp16")];
            tensor<int32, [4]> var_1070_shape_cast_fp16 = shape(x = key_states_19_cast_fp16)[name = string("op_1070_shape_cast_fp16")];
            int32 gather_93_axis_0 = const()[name = string("gather_93_axis_0"), val = int32(0)];
            int32 gather_93_batch_dims_0 = const()[name = string("gather_93_batch_dims_0"), val = int32(0)];
            bool gather_93_validate_indices_0 = const()[name = string("gather_93_validate_indices_0"), val = bool(false)];
            string var_1070_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1070_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_93_to_uint16 = const()[name = string("select_93_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1070_shape_cast_fp16_to_uint16 = cast(dtype = var_1070_shape_cast_fp16_to_uint16_dtype_0, x = var_1070_shape_cast_fp16)[name = string("cast_139")];
            uint16 gather_93_cast_uint16 = gather(axis = gather_93_axis_0, batch_dims = gather_93_batch_dims_0, indices = select_93_to_uint16, validate_indices = gather_93_validate_indices_0, x = var_1070_shape_cast_fp16_to_uint16)[name = string("gather_93_cast_uint16")];
            string gather_93_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_93_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_93_values0_0 = const()[name = string("concat_93_values0_0"), val = int32(1)];
            int32 concat_93_values1_0 = const()[name = string("concat_93_values1_0"), val = int32(1)];
            int32 concat_93_values2_0 = const()[name = string("concat_93_values2_0"), val = int32(0)];
            int32 concat_93_axis_0 = const()[name = string("concat_93_axis_0"), val = int32(0)];
            bool concat_93_interleave_0 = const()[name = string("concat_93_interleave_0"), val = bool(false)];
            int32 gather_93_cast_uint16_to_int32 = cast(dtype = gather_93_cast_uint16_to_int32_dtype_0, x = gather_93_cast_uint16)[name = string("cast_138")];
            tensor<int32, [4]> concat_93 = concat(axis = concat_93_axis_0, interleave = concat_93_interleave_0, values = (concat_93_values0_0, concat_93_values1_0, concat_93_values2_0, gather_93_cast_uint16_to_int32))[name = string("concat_93")];
            tensor<int32, [4]> causal_mask_11_begin_0 = const()[name = string("causal_mask_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_11_end_mask_0 = const()[name = string("causal_mask_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_11_cast_fp16 = slice_by_index(begin = causal_mask_11_begin_0, end = concat_93, end_mask = causal_mask_11_end_mask_0, x = causal_mask)[name = string("causal_mask_11_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_17_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_11_cast_fp16, key = key_states_19_cast_fp16, query = query_states_19_cast_fp16, value = value_states_19_cast_fp16)[name = string("attn_output_17_cast_fp16")];
            tensor<int32, [4]> var_1076_perm_0 = const()[name = string("op_1076_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_94x = const()[name = string("concat_94x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_1076_cast_fp16 = transpose(perm = var_1076_perm_0, x = attn_output_17_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, ?, 1536]> input_33_cast_fp16 = reshape(shape = concat_94x, x = var_1076_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238400896))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239580608))))[name = string("model_model_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_139_cast_fp16 = add(x = hidden_states_119_cast_fp16, y = linear_31_cast_fp16)[name = string("hidden_states_139_cast_fp16")];
            fp16 var_78_promoted_9_to_fp16 = const()[name = string("op_78_promoted_9_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_1085_cast_fp16 = pow(x = hidden_states_139_cast_fp16, y = var_78_promoted_9_to_fp16)[name = string("op_1085_cast_fp16")];
            tensor<int32, [1]> variance_19_axes_0 = const()[name = string("variance_19_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_19_keep_dims_0 = const()[name = string("variance_19_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_19_cast_fp16 = reduce_mean(axes = variance_19_axes_0, keep_dims = variance_19_keep_dims_0, x = var_1085_cast_fp16)[name = string("variance_19_cast_fp16")];
            fp16 var_1088_to_fp16 = const()[name = string("op_1088_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1089_cast_fp16 = add(x = variance_19_cast_fp16, y = var_1088_to_fp16)[name = string("op_1089_cast_fp16")];
            fp32 var_1090_epsilon_0 = const()[name = string("op_1090_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1090_cast_fp16 = rsqrt(epsilon = var_1090_epsilon_0, x = var_1089_cast_fp16)[name = string("op_1090_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_143_cast_fp16 = mul(x = hidden_states_139_cast_fp16, y = var_1090_cast_fp16)[name = string("hidden_states_143_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239728128)))];
            tensor<fp16, [1, ?, 1536]> input_35_cast_fp16 = mul(x = model_model_layers_4_post_attention_layernorm_weight_to_fp16, y = hidden_states_143_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239731264))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246612608))))[name = string("model_model_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_1102_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_1102_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247472832))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254354176))))[name = string("model_model_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_39_cast_fp16 = mul(x = var_1102_cast_fp16, y = linear_33_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255214400))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262095744))))[name = string("model_model_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_34_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_149_cast_fp16 = add(x = hidden_states_139_cast_fp16, y = linear_34_cast_fp16)[name = string("hidden_states_149_cast_fp16")];
            fp16 var_78_promoted_10_to_fp16 = const()[name = string("op_78_promoted_10_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_1115_cast_fp16 = pow(x = hidden_states_149_cast_fp16, y = var_78_promoted_10_to_fp16)[name = string("op_1115_cast_fp16")];
            tensor<int32, [1]> variance_21_axes_0 = const()[name = string("variance_21_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_21_keep_dims_0 = const()[name = string("variance_21_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_21_cast_fp16 = reduce_mean(axes = variance_21_axes_0, keep_dims = variance_21_keep_dims_0, x = var_1115_cast_fp16)[name = string("variance_21_cast_fp16")];
            fp16 var_1118_to_fp16 = const()[name = string("op_1118_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1119_cast_fp16 = add(x = variance_21_cast_fp16, y = var_1118_to_fp16)[name = string("op_1119_cast_fp16")];
            fp32 var_1120_epsilon_0 = const()[name = string("op_1120_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1120_cast_fp16 = rsqrt(epsilon = var_1120_epsilon_0, x = var_1119_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_153_cast_fp16 = mul(x = hidden_states_149_cast_fp16, y = var_1120_cast_fp16)[name = string("hidden_states_153_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262955968)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_157_cast_fp16 = mul(x = model_model_layers_5_input_layernorm_weight_to_fp16, y = hidden_states_153_cast_fp16)[name = string("hidden_states_157_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262959104))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264138816))))[name = string("model_model_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264286336)))];
            tensor<fp16, [1, ?, 1536]> linear_35_cast_fp16 = linear(bias = model_model_layers_5_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_157_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264289472))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264486144))))[name = string("model_model_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_5_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_5_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264510784)))];
            tensor<fp16, [1, ?, 256]> linear_36_cast_fp16 = linear(bias = model_model_layers_5_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_157_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264511360))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264708032))))[name = string("model_model_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264732672)))];
            tensor<fp16, [1, ?, 256]> linear_37_cast_fp16 = linear(bias = model_model_layers_5_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_157_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> concat_95x = const()[name = string("concat_95x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_1143_cast_fp16 = reshape(shape = concat_95x, x = linear_35_cast_fp16)[name = string("op_1143_cast_fp16")];
            tensor<int32, [4]> q_11_perm_0 = const()[name = string("q_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_1146_cast_fp16 = reshape(shape = concat_96x, x = linear_36_cast_fp16)[name = string("op_1146_cast_fp16")];
            tensor<int32, [4]> k_11_perm_0 = const()[name = string("k_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_1149_cast_fp16 = reshape(shape = concat_97x, x = linear_37_cast_fp16)[name = string("op_1149_cast_fp16")];
            tensor<int32, [4]> v_state_11_perm_0 = const()[name = string("v_state_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_11_cast_fp16 = transpose(perm = q_11_perm_0, x = var_1143_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 12, ?, 128]> var_1153_cast_fp16 = mul(x = q_11_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1153_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_11_cast_fp16)[name = string("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_11_cast_fp16)[name = string("x2_21_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_1164_cast_fp16 = mul(x = x2_21_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_1164_cast_fp16")];
            bool var_1166_interleave_0 = const()[name = string("op_1166_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_1166_cast_fp16 = concat(axis = var_72, interleave = var_1166_interleave_0, values = (var_1164_cast_fp16, x1_21_cast_fp16))[name = string("op_1166_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_1167_cast_fp16 = mul(x = var_1166_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1167_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_23_cast_fp16 = add(x = var_1153_cast_fp16, y = var_1167_cast_fp16)[name = string("query_states_23_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_11_cast_fp16 = transpose(perm = k_11_perm_0, x = var_1146_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 2, ?, 128]> var_1169_cast_fp16 = mul(x = k_11_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1169_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_11_cast_fp16)[name = string("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_11_cast_fp16)[name = string("x2_23_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_1180_cast_fp16 = mul(x = x2_23_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_1180_cast_fp16")];
            bool var_1182_interleave_0 = const()[name = string("op_1182_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_1182_cast_fp16 = concat(axis = var_72, interleave = var_1182_interleave_0, values = (var_1180_cast_fp16, x1_23_cast_fp16))[name = string("op_1182_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_1183_cast_fp16 = mul(x = var_1182_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1183_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_11_cast_fp16 = add(x = var_1169_cast_fp16, y = var_1183_cast_fp16)[name = string("k_state_11_cast_fp16")];
            tensor<int32, [1]> expand_dims_60 = const()[name = string("expand_dims_60"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_61 = const()[name = string("expand_dims_61"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_63 = const()[name = string("expand_dims_63"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_100_values0_0 = const()[name = string("concat_100_values0_0"), val = tensor<int32, [1]>([5])];
            int32 concat_100_axis_0 = const()[name = string("concat_100_axis_0"), val = int32(0)];
            bool concat_100_interleave_0 = const()[name = string("concat_100_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_100 = concat(axis = concat_100_axis_0, interleave = concat_100_interleave_0, values = (concat_100_values0_0, expand_dims_60, expand_dims_61, expand_dims_2, expand_dims_63))[name = string("concat_100")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_6_stride_0 = const()[name = string("key_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_100, begin_mask = key_cache_internal_tensor_assign_6_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_6_squeeze_mask_0, stride = key_cache_internal_tensor_assign_6_stride_0, update = k_state_11_cast_fp16, x = coreml_update_state_64)[name = string("key_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_6_cast_fp16, input = key_cache)[name = string("coreml_update_state_66_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_66 = read_state(input = key_cache)[name = string("coreml_update_state_66")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_6_stride_0 = const()[name = string("value_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_11_cast_fp16 = transpose(perm = v_state_11_perm_0, x = var_1149_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_100, begin_mask = value_cache_internal_tensor_assign_6_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_6_squeeze_mask_0, stride = value_cache_internal_tensor_assign_6_stride_0, update = v_state_11_cast_fp16, x = coreml_update_state_65)[name = string("value_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_6_cast_fp16, input = value_cache)[name = string("coreml_update_state_67_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_67 = read_state(input = value_cache)[name = string("coreml_update_state_67")];
            tensor<int32, [5]> var_1206_begin_0 = const()[name = string("op_1206_begin_0"), val = tensor<int32, [5]>([5, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1206_end_0 = const()[name = string("op_1206_end_0"), val = tensor<int32, [5]>([6, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_1206_end_mask_0 = const()[name = string("op_1206_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1206_squeeze_mask_0 = const()[name = string("op_1206_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_1206_cast_fp16 = slice_by_index(begin = var_1206_begin_0, end = var_1206_end_0, end_mask = var_1206_end_mask_0, squeeze_mask = var_1206_squeeze_mask_0, x = coreml_update_state_66)[name = string("op_1206_cast_fp16")];
            tensor<int32, [4]> var_1209_begin_0 = const()[name = string("op_1209_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1209_end_mask_0 = const()[name = string("op_1209_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_1209_cast_fp16 = slice_by_index(begin = var_1209_begin_0, end = concat_11, end_mask = var_1209_end_mask_0, x = var_1206_cast_fp16)[name = string("op_1209_cast_fp16")];
            tensor<int32, [5]> var_1211_begin_0 = const()[name = string("op_1211_begin_0"), val = tensor<int32, [5]>([5, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1211_end_0 = const()[name = string("op_1211_end_0"), val = tensor<int32, [5]>([6, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_1211_end_mask_0 = const()[name = string("op_1211_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1211_squeeze_mask_0 = const()[name = string("op_1211_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_1211_cast_fp16 = slice_by_index(begin = var_1211_begin_0, end = var_1211_end_0, end_mask = var_1211_end_mask_0, squeeze_mask = var_1211_squeeze_mask_0, x = coreml_update_state_67)[name = string("op_1211_cast_fp16")];
            tensor<int32, [4]> var_1214_begin_0 = const()[name = string("op_1214_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1214_end_mask_0 = const()[name = string("op_1214_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_1214_cast_fp16 = slice_by_index(begin = var_1214_begin_0, end = concat_11, end_mask = var_1214_end_mask_0, x = var_1211_cast_fp16)[name = string("op_1214_cast_fp16")];
            tensor<int32, [4]> var_1216_shape_cast_fp16 = shape(x = var_1209_cast_fp16)[name = string("op_1216_shape_cast_fp16")];
            int32 gather_103 = const()[name = string("gather_103"), val = int32(1)];
            int32 gather_104 = const()[name = string("gather_104"), val = int32(2)];
            int32 gather_105_axis_0 = const()[name = string("gather_105_axis_0"), val = int32(0)];
            int32 gather_105_batch_dims_0 = const()[name = string("gather_105_batch_dims_0"), val = int32(0)];
            bool gather_105_validate_indices_0 = const()[name = string("gather_105_validate_indices_0"), val = bool(false)];
            string var_1216_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1216_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_105_to_uint16 = const()[name = string("select_105_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1216_shape_cast_fp16_to_uint16 = cast(dtype = var_1216_shape_cast_fp16_to_uint16_dtype_0, x = var_1216_shape_cast_fp16)[name = string("cast_137")];
            uint16 gather_105_cast_uint16 = gather(axis = gather_105_axis_0, batch_dims = gather_105_batch_dims_0, indices = select_105_to_uint16, validate_indices = gather_105_validate_indices_0, x = var_1216_shape_cast_fp16_to_uint16)[name = string("gather_105_cast_uint16")];
            string gather_105_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_105_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_106 = const()[name = string("gather_106"), val = int32(128)];
            tensor<int32, [1]> var_1223_axes_0 = const()[name = string("op_1223_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_1223_cast_fp16 = expand_dims(axes = var_1223_axes_0, x = var_1209_cast_fp16)[name = string("op_1223_cast_fp16")];
            tensor<int32, [5]> shape_117_cast_fp16 = shape(x = var_1223_cast_fp16)[name = string("shape_117_cast_fp16")];
            int32 concat_108_axis_0 = const()[name = string("concat_108_axis_0"), val = int32(0)];
            bool concat_108_interleave_0 = const()[name = string("concat_108_interleave_0"), val = bool(false)];
            int32 gather_105_cast_uint16_to_int32 = cast(dtype = gather_105_cast_uint16_to_int32_dtype_0, x = gather_105_cast_uint16)[name = string("cast_136")];
            tensor<int32, [5]> concat_108 = concat(axis = concat_108_axis_0, interleave = concat_108_interleave_0, values = (gather_103, gather_104, var_77, gather_105_cast_uint16_to_int32, gather_106))[name = string("concat_108")];
            tensor<int32, [5]> real_div_10 = real_div(x = concat_108, y = shape_117_cast_fp16)[name = string("real_div_10")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_161_cast_fp16 = tile(reps = real_div_10, x = var_1223_cast_fp16)[name = string("hidden_states_161_cast_fp16")];
            tensor<int32, [4]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_23_cast_fp16 = reshape(shape = concat_109x, x = hidden_states_161_cast_fp16)[name = string("key_states_23_cast_fp16")];
            tensor<int32, [4]> var_1233_shape_cast_fp16 = shape(x = var_1214_cast_fp16)[name = string("op_1233_shape_cast_fp16")];
            int32 gather_107 = const()[name = string("gather_107"), val = int32(1)];
            int32 gather_108 = const()[name = string("gather_108"), val = int32(2)];
            int32 gather_109_axis_0 = const()[name = string("gather_109_axis_0"), val = int32(0)];
            int32 gather_109_batch_dims_0 = const()[name = string("gather_109_batch_dims_0"), val = int32(0)];
            bool gather_109_validate_indices_0 = const()[name = string("gather_109_validate_indices_0"), val = bool(false)];
            string var_1233_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1233_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_109_to_uint16 = const()[name = string("select_109_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1233_shape_cast_fp16_to_uint16 = cast(dtype = var_1233_shape_cast_fp16_to_uint16_dtype_0, x = var_1233_shape_cast_fp16)[name = string("cast_135")];
            uint16 gather_109_cast_uint16 = gather(axis = gather_109_axis_0, batch_dims = gather_109_batch_dims_0, indices = select_109_to_uint16, validate_indices = gather_109_validate_indices_0, x = var_1233_shape_cast_fp16_to_uint16)[name = string("gather_109_cast_uint16")];
            string gather_109_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_109_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_110 = const()[name = string("gather_110"), val = int32(128)];
            tensor<int32, [1]> var_1240_axes_0 = const()[name = string("op_1240_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_1240_cast_fp16 = expand_dims(axes = var_1240_axes_0, x = var_1214_cast_fp16)[name = string("op_1240_cast_fp16")];
            tensor<int32, [5]> shape_122_cast_fp16 = shape(x = var_1240_cast_fp16)[name = string("shape_122_cast_fp16")];
            int32 concat_110_axis_0 = const()[name = string("concat_110_axis_0"), val = int32(0)];
            bool concat_110_interleave_0 = const()[name = string("concat_110_interleave_0"), val = bool(false)];
            int32 gather_109_cast_uint16_to_int32 = cast(dtype = gather_109_cast_uint16_to_int32_dtype_0, x = gather_109_cast_uint16)[name = string("cast_134")];
            tensor<int32, [5]> concat_110 = concat(axis = concat_110_axis_0, interleave = concat_110_interleave_0, values = (gather_107, gather_108, var_77, gather_109_cast_uint16_to_int32, gather_110))[name = string("concat_110")];
            tensor<int32, [5]> real_div_11 = real_div(x = concat_110, y = shape_122_cast_fp16)[name = string("real_div_11")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_165_cast_fp16 = tile(reps = real_div_11, x = var_1240_cast_fp16)[name = string("hidden_states_165_cast_fp16")];
            tensor<int32, [4]> concat_111x = const()[name = string("concat_111x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_23_cast_fp16 = reshape(shape = concat_111x, x = hidden_states_165_cast_fp16)[name = string("value_states_23_cast_fp16")];
            tensor<int32, [4]> var_1250_shape_cast_fp16 = shape(x = key_states_23_cast_fp16)[name = string("op_1250_shape_cast_fp16")];
            int32 gather_111_axis_0 = const()[name = string("gather_111_axis_0"), val = int32(0)];
            int32 gather_111_batch_dims_0 = const()[name = string("gather_111_batch_dims_0"), val = int32(0)];
            bool gather_111_validate_indices_0 = const()[name = string("gather_111_validate_indices_0"), val = bool(false)];
            string var_1250_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1250_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_111_to_uint16 = const()[name = string("select_111_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1250_shape_cast_fp16_to_uint16 = cast(dtype = var_1250_shape_cast_fp16_to_uint16_dtype_0, x = var_1250_shape_cast_fp16)[name = string("cast_133")];
            uint16 gather_111_cast_uint16 = gather(axis = gather_111_axis_0, batch_dims = gather_111_batch_dims_0, indices = select_111_to_uint16, validate_indices = gather_111_validate_indices_0, x = var_1250_shape_cast_fp16_to_uint16)[name = string("gather_111_cast_uint16")];
            string gather_111_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_111_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_112_values0_0 = const()[name = string("concat_112_values0_0"), val = int32(1)];
            int32 concat_112_values1_0 = const()[name = string("concat_112_values1_0"), val = int32(1)];
            int32 concat_112_values2_0 = const()[name = string("concat_112_values2_0"), val = int32(0)];
            int32 concat_112_axis_0 = const()[name = string("concat_112_axis_0"), val = int32(0)];
            bool concat_112_interleave_0 = const()[name = string("concat_112_interleave_0"), val = bool(false)];
            int32 gather_111_cast_uint16_to_int32 = cast(dtype = gather_111_cast_uint16_to_int32_dtype_0, x = gather_111_cast_uint16)[name = string("cast_132")];
            tensor<int32, [4]> concat_112 = concat(axis = concat_112_axis_0, interleave = concat_112_interleave_0, values = (concat_112_values0_0, concat_112_values1_0, concat_112_values2_0, gather_111_cast_uint16_to_int32))[name = string("concat_112")];
            tensor<int32, [4]> causal_mask_13_begin_0 = const()[name = string("causal_mask_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_13_end_mask_0 = const()[name = string("causal_mask_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_13_cast_fp16 = slice_by_index(begin = causal_mask_13_begin_0, end = concat_112, end_mask = causal_mask_13_end_mask_0, x = causal_mask)[name = string("causal_mask_13_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_21_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_13_cast_fp16, key = key_states_23_cast_fp16, query = query_states_23_cast_fp16, value = value_states_23_cast_fp16)[name = string("attn_output_21_cast_fp16")];
            tensor<int32, [4]> var_1256_perm_0 = const()[name = string("op_1256_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_113x = const()[name = string("concat_113x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_1256_cast_fp16 = transpose(perm = var_1256_perm_0, x = attn_output_21_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, ?, 1536]> input_41_cast_fp16 = reshape(shape = concat_113x, x = var_1256_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264733248))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265912960))))[name = string("model_model_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_38_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_169_cast_fp16 = add(x = hidden_states_149_cast_fp16, y = linear_38_cast_fp16)[name = string("hidden_states_169_cast_fp16")];
            fp16 var_78_promoted_11_to_fp16 = const()[name = string("op_78_promoted_11_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_1265_cast_fp16 = pow(x = hidden_states_169_cast_fp16, y = var_78_promoted_11_to_fp16)[name = string("op_1265_cast_fp16")];
            tensor<int32, [1]> variance_23_axes_0 = const()[name = string("variance_23_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_23_keep_dims_0 = const()[name = string("variance_23_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_23_cast_fp16 = reduce_mean(axes = variance_23_axes_0, keep_dims = variance_23_keep_dims_0, x = var_1265_cast_fp16)[name = string("variance_23_cast_fp16")];
            fp16 var_1268_to_fp16 = const()[name = string("op_1268_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1269_cast_fp16 = add(x = variance_23_cast_fp16, y = var_1268_to_fp16)[name = string("op_1269_cast_fp16")];
            fp32 var_1270_epsilon_0 = const()[name = string("op_1270_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1270_cast_fp16 = rsqrt(epsilon = var_1270_epsilon_0, x = var_1269_cast_fp16)[name = string("op_1270_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_173_cast_fp16 = mul(x = hidden_states_169_cast_fp16, y = var_1270_cast_fp16)[name = string("hidden_states_173_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266060480)))];
            tensor<fp16, [1, ?, 1536]> input_43_cast_fp16 = mul(x = model_model_layers_5_post_attention_layernorm_weight_to_fp16, y = hidden_states_173_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266063616))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272944960))))[name = string("model_model_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_1282_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_1282_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273805184))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280686528))))[name = string("model_model_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_47_cast_fp16 = mul(x = var_1282_cast_fp16, y = linear_40_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281546752))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288428096))))[name = string("model_model_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_41_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_179_cast_fp16 = add(x = hidden_states_169_cast_fp16, y = linear_41_cast_fp16)[name = string("hidden_states_179_cast_fp16")];
            fp16 var_78_promoted_12_to_fp16 = const()[name = string("op_78_promoted_12_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_1295_cast_fp16 = pow(x = hidden_states_179_cast_fp16, y = var_78_promoted_12_to_fp16)[name = string("op_1295_cast_fp16")];
            tensor<int32, [1]> variance_25_axes_0 = const()[name = string("variance_25_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_25_keep_dims_0 = const()[name = string("variance_25_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_25_cast_fp16 = reduce_mean(axes = variance_25_axes_0, keep_dims = variance_25_keep_dims_0, x = var_1295_cast_fp16)[name = string("variance_25_cast_fp16")];
            fp16 var_1298_to_fp16 = const()[name = string("op_1298_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1299_cast_fp16 = add(x = variance_25_cast_fp16, y = var_1298_to_fp16)[name = string("op_1299_cast_fp16")];
            fp32 var_1300_epsilon_0 = const()[name = string("op_1300_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1300_cast_fp16 = rsqrt(epsilon = var_1300_epsilon_0, x = var_1299_cast_fp16)[name = string("op_1300_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_183_cast_fp16 = mul(x = hidden_states_179_cast_fp16, y = var_1300_cast_fp16)[name = string("hidden_states_183_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289288320)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_187_cast_fp16 = mul(x = model_model_layers_6_input_layernorm_weight_to_fp16, y = hidden_states_183_cast_fp16)[name = string("hidden_states_187_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289291456))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290471168))))[name = string("model_model_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290618688)))];
            tensor<fp16, [1, ?, 1536]> linear_42_cast_fp16 = linear(bias = model_model_layers_6_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_187_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290621824))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290818496))))[name = string("model_model_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_6_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_6_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290843136)))];
            tensor<fp16, [1, ?, 256]> linear_43_cast_fp16 = linear(bias = model_model_layers_6_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_187_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290843712))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291040384))))[name = string("model_model_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291065024)))];
            tensor<fp16, [1, ?, 256]> linear_44_cast_fp16 = linear(bias = model_model_layers_6_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_187_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> concat_114x = const()[name = string("concat_114x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_1323_cast_fp16 = reshape(shape = concat_114x, x = linear_42_cast_fp16)[name = string("op_1323_cast_fp16")];
            tensor<int32, [4]> q_13_perm_0 = const()[name = string("q_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_115x = const()[name = string("concat_115x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_1326_cast_fp16 = reshape(shape = concat_115x, x = linear_43_cast_fp16)[name = string("op_1326_cast_fp16")];
            tensor<int32, [4]> k_13_perm_0 = const()[name = string("k_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_116x = const()[name = string("concat_116x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_1329_cast_fp16 = reshape(shape = concat_116x, x = linear_44_cast_fp16)[name = string("op_1329_cast_fp16")];
            tensor<int32, [4]> v_state_13_perm_0 = const()[name = string("v_state_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_13_cast_fp16 = transpose(perm = q_13_perm_0, x = var_1323_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 12, ?, 128]> var_1333_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1333_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_13_cast_fp16)[name = string("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_13_cast_fp16)[name = string("x2_25_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_1344_cast_fp16 = mul(x = x2_25_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_1344_cast_fp16")];
            bool var_1346_interleave_0 = const()[name = string("op_1346_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_1346_cast_fp16 = concat(axis = var_72, interleave = var_1346_interleave_0, values = (var_1344_cast_fp16, x1_25_cast_fp16))[name = string("op_1346_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_1347_cast_fp16 = mul(x = var_1346_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1347_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_27_cast_fp16 = add(x = var_1333_cast_fp16, y = var_1347_cast_fp16)[name = string("query_states_27_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_13_cast_fp16 = transpose(perm = k_13_perm_0, x = var_1326_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 2, ?, 128]> var_1349_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1349_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_13_cast_fp16)[name = string("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_13_cast_fp16)[name = string("x2_27_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_1360_cast_fp16 = mul(x = x2_27_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_1360_cast_fp16")];
            bool var_1362_interleave_0 = const()[name = string("op_1362_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_1362_cast_fp16 = concat(axis = var_72, interleave = var_1362_interleave_0, values = (var_1360_cast_fp16, x1_27_cast_fp16))[name = string("op_1362_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_1363_cast_fp16 = mul(x = var_1362_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1363_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_13_cast_fp16 = add(x = var_1349_cast_fp16, y = var_1363_cast_fp16)[name = string("k_state_13_cast_fp16")];
            tensor<int32, [1]> expand_dims_72 = const()[name = string("expand_dims_72"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_73 = const()[name = string("expand_dims_73"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_75 = const()[name = string("expand_dims_75"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_119_values0_0 = const()[name = string("concat_119_values0_0"), val = tensor<int32, [1]>([6])];
            int32 concat_119_axis_0 = const()[name = string("concat_119_axis_0"), val = int32(0)];
            bool concat_119_interleave_0 = const()[name = string("concat_119_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_119 = concat(axis = concat_119_axis_0, interleave = concat_119_interleave_0, values = (concat_119_values0_0, expand_dims_72, expand_dims_73, expand_dims_2, expand_dims_75))[name = string("concat_119")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_7_stride_0 = const()[name = string("key_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_119, begin_mask = key_cache_internal_tensor_assign_7_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_7_squeeze_mask_0, stride = key_cache_internal_tensor_assign_7_stride_0, update = k_state_13_cast_fp16, x = coreml_update_state_66)[name = string("key_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_7_cast_fp16, input = key_cache)[name = string("coreml_update_state_68_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_68 = read_state(input = key_cache)[name = string("coreml_update_state_68")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_7_stride_0 = const()[name = string("value_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_13_cast_fp16 = transpose(perm = v_state_13_perm_0, x = var_1329_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_119, begin_mask = value_cache_internal_tensor_assign_7_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_7_squeeze_mask_0, stride = value_cache_internal_tensor_assign_7_stride_0, update = v_state_13_cast_fp16, x = coreml_update_state_67)[name = string("value_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_7_cast_fp16, input = value_cache)[name = string("coreml_update_state_69_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_69 = read_state(input = value_cache)[name = string("coreml_update_state_69")];
            tensor<int32, [5]> var_1386_begin_0 = const()[name = string("op_1386_begin_0"), val = tensor<int32, [5]>([6, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1386_end_0 = const()[name = string("op_1386_end_0"), val = tensor<int32, [5]>([7, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_1386_end_mask_0 = const()[name = string("op_1386_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1386_squeeze_mask_0 = const()[name = string("op_1386_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_1386_cast_fp16 = slice_by_index(begin = var_1386_begin_0, end = var_1386_end_0, end_mask = var_1386_end_mask_0, squeeze_mask = var_1386_squeeze_mask_0, x = coreml_update_state_68)[name = string("op_1386_cast_fp16")];
            tensor<int32, [4]> var_1389_begin_0 = const()[name = string("op_1389_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1389_end_mask_0 = const()[name = string("op_1389_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_1389_cast_fp16 = slice_by_index(begin = var_1389_begin_0, end = concat_11, end_mask = var_1389_end_mask_0, x = var_1386_cast_fp16)[name = string("op_1389_cast_fp16")];
            tensor<int32, [5]> var_1391_begin_0 = const()[name = string("op_1391_begin_0"), val = tensor<int32, [5]>([6, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1391_end_0 = const()[name = string("op_1391_end_0"), val = tensor<int32, [5]>([7, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_1391_end_mask_0 = const()[name = string("op_1391_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1391_squeeze_mask_0 = const()[name = string("op_1391_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_1391_cast_fp16 = slice_by_index(begin = var_1391_begin_0, end = var_1391_end_0, end_mask = var_1391_end_mask_0, squeeze_mask = var_1391_squeeze_mask_0, x = coreml_update_state_69)[name = string("op_1391_cast_fp16")];
            tensor<int32, [4]> var_1394_begin_0 = const()[name = string("op_1394_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1394_end_mask_0 = const()[name = string("op_1394_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_1394_cast_fp16 = slice_by_index(begin = var_1394_begin_0, end = concat_11, end_mask = var_1394_end_mask_0, x = var_1391_cast_fp16)[name = string("op_1394_cast_fp16")];
            tensor<int32, [4]> var_1396_shape_cast_fp16 = shape(x = var_1389_cast_fp16)[name = string("op_1396_shape_cast_fp16")];
            int32 gather_121 = const()[name = string("gather_121"), val = int32(1)];
            int32 gather_122 = const()[name = string("gather_122"), val = int32(2)];
            int32 gather_123_axis_0 = const()[name = string("gather_123_axis_0"), val = int32(0)];
            int32 gather_123_batch_dims_0 = const()[name = string("gather_123_batch_dims_0"), val = int32(0)];
            bool gather_123_validate_indices_0 = const()[name = string("gather_123_validate_indices_0"), val = bool(false)];
            string var_1396_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1396_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_123_to_uint16 = const()[name = string("select_123_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1396_shape_cast_fp16_to_uint16 = cast(dtype = var_1396_shape_cast_fp16_to_uint16_dtype_0, x = var_1396_shape_cast_fp16)[name = string("cast_131")];
            uint16 gather_123_cast_uint16 = gather(axis = gather_123_axis_0, batch_dims = gather_123_batch_dims_0, indices = select_123_to_uint16, validate_indices = gather_123_validate_indices_0, x = var_1396_shape_cast_fp16_to_uint16)[name = string("gather_123_cast_uint16")];
            string gather_123_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_123_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_124 = const()[name = string("gather_124"), val = int32(128)];
            tensor<int32, [1]> var_1403_axes_0 = const()[name = string("op_1403_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_1403_cast_fp16 = expand_dims(axes = var_1403_axes_0, x = var_1389_cast_fp16)[name = string("op_1403_cast_fp16")];
            tensor<int32, [5]> shape_137_cast_fp16 = shape(x = var_1403_cast_fp16)[name = string("shape_137_cast_fp16")];
            int32 concat_127_axis_0 = const()[name = string("concat_127_axis_0"), val = int32(0)];
            bool concat_127_interleave_0 = const()[name = string("concat_127_interleave_0"), val = bool(false)];
            int32 gather_123_cast_uint16_to_int32 = cast(dtype = gather_123_cast_uint16_to_int32_dtype_0, x = gather_123_cast_uint16)[name = string("cast_130")];
            tensor<int32, [5]> concat_127 = concat(axis = concat_127_axis_0, interleave = concat_127_interleave_0, values = (gather_121, gather_122, var_77, gather_123_cast_uint16_to_int32, gather_124))[name = string("concat_127")];
            tensor<int32, [5]> real_div_12 = real_div(x = concat_127, y = shape_137_cast_fp16)[name = string("real_div_12")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_191_cast_fp16 = tile(reps = real_div_12, x = var_1403_cast_fp16)[name = string("hidden_states_191_cast_fp16")];
            tensor<int32, [4]> concat_128x = const()[name = string("concat_128x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_27_cast_fp16 = reshape(shape = concat_128x, x = hidden_states_191_cast_fp16)[name = string("key_states_27_cast_fp16")];
            tensor<int32, [4]> var_1413_shape_cast_fp16 = shape(x = var_1394_cast_fp16)[name = string("op_1413_shape_cast_fp16")];
            int32 gather_125 = const()[name = string("gather_125"), val = int32(1)];
            int32 gather_126 = const()[name = string("gather_126"), val = int32(2)];
            int32 gather_127_axis_0 = const()[name = string("gather_127_axis_0"), val = int32(0)];
            int32 gather_127_batch_dims_0 = const()[name = string("gather_127_batch_dims_0"), val = int32(0)];
            bool gather_127_validate_indices_0 = const()[name = string("gather_127_validate_indices_0"), val = bool(false)];
            string var_1413_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1413_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_127_to_uint16 = const()[name = string("select_127_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1413_shape_cast_fp16_to_uint16 = cast(dtype = var_1413_shape_cast_fp16_to_uint16_dtype_0, x = var_1413_shape_cast_fp16)[name = string("cast_129")];
            uint16 gather_127_cast_uint16 = gather(axis = gather_127_axis_0, batch_dims = gather_127_batch_dims_0, indices = select_127_to_uint16, validate_indices = gather_127_validate_indices_0, x = var_1413_shape_cast_fp16_to_uint16)[name = string("gather_127_cast_uint16")];
            string gather_127_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_127_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_128 = const()[name = string("gather_128"), val = int32(128)];
            tensor<int32, [1]> var_1420_axes_0 = const()[name = string("op_1420_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_1420_cast_fp16 = expand_dims(axes = var_1420_axes_0, x = var_1394_cast_fp16)[name = string("op_1420_cast_fp16")];
            tensor<int32, [5]> shape_142_cast_fp16 = shape(x = var_1420_cast_fp16)[name = string("shape_142_cast_fp16")];
            int32 concat_129_axis_0 = const()[name = string("concat_129_axis_0"), val = int32(0)];
            bool concat_129_interleave_0 = const()[name = string("concat_129_interleave_0"), val = bool(false)];
            int32 gather_127_cast_uint16_to_int32 = cast(dtype = gather_127_cast_uint16_to_int32_dtype_0, x = gather_127_cast_uint16)[name = string("cast_128")];
            tensor<int32, [5]> concat_129 = concat(axis = concat_129_axis_0, interleave = concat_129_interleave_0, values = (gather_125, gather_126, var_77, gather_127_cast_uint16_to_int32, gather_128))[name = string("concat_129")];
            tensor<int32, [5]> real_div_13 = real_div(x = concat_129, y = shape_142_cast_fp16)[name = string("real_div_13")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_195_cast_fp16 = tile(reps = real_div_13, x = var_1420_cast_fp16)[name = string("hidden_states_195_cast_fp16")];
            tensor<int32, [4]> concat_130x = const()[name = string("concat_130x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_27_cast_fp16 = reshape(shape = concat_130x, x = hidden_states_195_cast_fp16)[name = string("value_states_27_cast_fp16")];
            tensor<int32, [4]> var_1430_shape_cast_fp16 = shape(x = key_states_27_cast_fp16)[name = string("op_1430_shape_cast_fp16")];
            int32 gather_129_axis_0 = const()[name = string("gather_129_axis_0"), val = int32(0)];
            int32 gather_129_batch_dims_0 = const()[name = string("gather_129_batch_dims_0"), val = int32(0)];
            bool gather_129_validate_indices_0 = const()[name = string("gather_129_validate_indices_0"), val = bool(false)];
            string var_1430_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1430_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_129_to_uint16 = const()[name = string("select_129_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1430_shape_cast_fp16_to_uint16 = cast(dtype = var_1430_shape_cast_fp16_to_uint16_dtype_0, x = var_1430_shape_cast_fp16)[name = string("cast_127")];
            uint16 gather_129_cast_uint16 = gather(axis = gather_129_axis_0, batch_dims = gather_129_batch_dims_0, indices = select_129_to_uint16, validate_indices = gather_129_validate_indices_0, x = var_1430_shape_cast_fp16_to_uint16)[name = string("gather_129_cast_uint16")];
            string gather_129_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_129_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_131_values0_0 = const()[name = string("concat_131_values0_0"), val = int32(1)];
            int32 concat_131_values1_0 = const()[name = string("concat_131_values1_0"), val = int32(1)];
            int32 concat_131_values2_0 = const()[name = string("concat_131_values2_0"), val = int32(0)];
            int32 concat_131_axis_0 = const()[name = string("concat_131_axis_0"), val = int32(0)];
            bool concat_131_interleave_0 = const()[name = string("concat_131_interleave_0"), val = bool(false)];
            int32 gather_129_cast_uint16_to_int32 = cast(dtype = gather_129_cast_uint16_to_int32_dtype_0, x = gather_129_cast_uint16)[name = string("cast_126")];
            tensor<int32, [4]> concat_131 = concat(axis = concat_131_axis_0, interleave = concat_131_interleave_0, values = (concat_131_values0_0, concat_131_values1_0, concat_131_values2_0, gather_129_cast_uint16_to_int32))[name = string("concat_131")];
            tensor<int32, [4]> causal_mask_15_begin_0 = const()[name = string("causal_mask_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_15_end_mask_0 = const()[name = string("causal_mask_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_15_cast_fp16 = slice_by_index(begin = causal_mask_15_begin_0, end = concat_131, end_mask = causal_mask_15_end_mask_0, x = causal_mask)[name = string("causal_mask_15_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_25_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_15_cast_fp16, key = key_states_27_cast_fp16, query = query_states_27_cast_fp16, value = value_states_27_cast_fp16)[name = string("attn_output_25_cast_fp16")];
            tensor<int32, [4]> var_1436_perm_0 = const()[name = string("op_1436_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_1436_cast_fp16 = transpose(perm = var_1436_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, ?, 1536]> input_49_cast_fp16 = reshape(shape = concat_132x, x = var_1436_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291065600))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292245312))))[name = string("model_model_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_45_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_199_cast_fp16 = add(x = hidden_states_179_cast_fp16, y = linear_45_cast_fp16)[name = string("hidden_states_199_cast_fp16")];
            fp16 var_78_promoted_13_to_fp16 = const()[name = string("op_78_promoted_13_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_1445_cast_fp16 = pow(x = hidden_states_199_cast_fp16, y = var_78_promoted_13_to_fp16)[name = string("op_1445_cast_fp16")];
            tensor<int32, [1]> variance_27_axes_0 = const()[name = string("variance_27_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_27_keep_dims_0 = const()[name = string("variance_27_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_27_cast_fp16 = reduce_mean(axes = variance_27_axes_0, keep_dims = variance_27_keep_dims_0, x = var_1445_cast_fp16)[name = string("variance_27_cast_fp16")];
            fp16 var_1448_to_fp16 = const()[name = string("op_1448_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1449_cast_fp16 = add(x = variance_27_cast_fp16, y = var_1448_to_fp16)[name = string("op_1449_cast_fp16")];
            fp32 var_1450_epsilon_0 = const()[name = string("op_1450_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1450_cast_fp16 = rsqrt(epsilon = var_1450_epsilon_0, x = var_1449_cast_fp16)[name = string("op_1450_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_203_cast_fp16 = mul(x = hidden_states_199_cast_fp16, y = var_1450_cast_fp16)[name = string("hidden_states_203_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292392832)))];
            tensor<fp16, [1, ?, 1536]> input_51_cast_fp16 = mul(x = model_model_layers_6_post_attention_layernorm_weight_to_fp16, y = hidden_states_203_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292395968))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299277312))))[name = string("model_model_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_1462_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_1462_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300137536))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307018880))))[name = string("model_model_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_55_cast_fp16 = mul(x = var_1462_cast_fp16, y = linear_47_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307879104))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314760448))))[name = string("model_model_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_48_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_209_cast_fp16 = add(x = hidden_states_199_cast_fp16, y = linear_48_cast_fp16)[name = string("hidden_states_209_cast_fp16")];
            fp16 var_78_promoted_14_to_fp16 = const()[name = string("op_78_promoted_14_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_1475_cast_fp16 = pow(x = hidden_states_209_cast_fp16, y = var_78_promoted_14_to_fp16)[name = string("op_1475_cast_fp16")];
            tensor<int32, [1]> variance_29_axes_0 = const()[name = string("variance_29_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_29_keep_dims_0 = const()[name = string("variance_29_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_29_cast_fp16 = reduce_mean(axes = variance_29_axes_0, keep_dims = variance_29_keep_dims_0, x = var_1475_cast_fp16)[name = string("variance_29_cast_fp16")];
            fp16 var_1478_to_fp16 = const()[name = string("op_1478_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1479_cast_fp16 = add(x = variance_29_cast_fp16, y = var_1478_to_fp16)[name = string("op_1479_cast_fp16")];
            fp32 var_1480_epsilon_0 = const()[name = string("op_1480_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1480_cast_fp16 = rsqrt(epsilon = var_1480_epsilon_0, x = var_1479_cast_fp16)[name = string("op_1480_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_213_cast_fp16 = mul(x = hidden_states_209_cast_fp16, y = var_1480_cast_fp16)[name = string("hidden_states_213_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315620672)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_217_cast_fp16 = mul(x = model_model_layers_7_input_layernorm_weight_to_fp16, y = hidden_states_213_cast_fp16)[name = string("hidden_states_217_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315623808))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316803520))))[name = string("model_model_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316951040)))];
            tensor<fp16, [1, ?, 1536]> linear_49_cast_fp16 = linear(bias = model_model_layers_7_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_217_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316954176))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317150848))))[name = string("model_model_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_7_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_7_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317175488)))];
            tensor<fp16, [1, ?, 256]> linear_50_cast_fp16 = linear(bias = model_model_layers_7_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_217_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317176064))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317372736))))[name = string("model_model_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317397376)))];
            tensor<fp16, [1, ?, 256]> linear_51_cast_fp16 = linear(bias = model_model_layers_7_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_217_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_1503_cast_fp16 = reshape(shape = concat_133x, x = linear_49_cast_fp16)[name = string("op_1503_cast_fp16")];
            tensor<int32, [4]> q_15_perm_0 = const()[name = string("q_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_1506_cast_fp16 = reshape(shape = concat_134x, x = linear_50_cast_fp16)[name = string("op_1506_cast_fp16")];
            tensor<int32, [4]> k_15_perm_0 = const()[name = string("k_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_135x = const()[name = string("concat_135x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_1509_cast_fp16 = reshape(shape = concat_135x, x = linear_51_cast_fp16)[name = string("op_1509_cast_fp16")];
            tensor<int32, [4]> v_state_15_perm_0 = const()[name = string("v_state_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_15_cast_fp16 = transpose(perm = q_15_perm_0, x = var_1503_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 12, ?, 128]> var_1513_cast_fp16 = mul(x = q_15_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1513_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_15_cast_fp16)[name = string("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_15_cast_fp16)[name = string("x2_29_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_1524_cast_fp16 = mul(x = x2_29_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_1524_cast_fp16")];
            bool var_1526_interleave_0 = const()[name = string("op_1526_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_1526_cast_fp16 = concat(axis = var_72, interleave = var_1526_interleave_0, values = (var_1524_cast_fp16, x1_29_cast_fp16))[name = string("op_1526_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_1527_cast_fp16 = mul(x = var_1526_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1527_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_31_cast_fp16 = add(x = var_1513_cast_fp16, y = var_1527_cast_fp16)[name = string("query_states_31_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_15_cast_fp16 = transpose(perm = k_15_perm_0, x = var_1506_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 2, ?, 128]> var_1529_cast_fp16 = mul(x = k_15_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1529_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_15_cast_fp16)[name = string("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_15_cast_fp16)[name = string("x2_31_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_1540_cast_fp16 = mul(x = x2_31_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_1540_cast_fp16")];
            bool var_1542_interleave_0 = const()[name = string("op_1542_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_1542_cast_fp16 = concat(axis = var_72, interleave = var_1542_interleave_0, values = (var_1540_cast_fp16, x1_31_cast_fp16))[name = string("op_1542_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_1543_cast_fp16 = mul(x = var_1542_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1543_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_15_cast_fp16 = add(x = var_1529_cast_fp16, y = var_1543_cast_fp16)[name = string("k_state_15_cast_fp16")];
            tensor<int32, [1]> expand_dims_84 = const()[name = string("expand_dims_84"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_85 = const()[name = string("expand_dims_85"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_87 = const()[name = string("expand_dims_87"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_138_values0_0 = const()[name = string("concat_138_values0_0"), val = tensor<int32, [1]>([7])];
            int32 concat_138_axis_0 = const()[name = string("concat_138_axis_0"), val = int32(0)];
            bool concat_138_interleave_0 = const()[name = string("concat_138_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_138 = concat(axis = concat_138_axis_0, interleave = concat_138_interleave_0, values = (concat_138_values0_0, expand_dims_84, expand_dims_85, expand_dims_2, expand_dims_87))[name = string("concat_138")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_8_stride_0 = const()[name = string("key_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_138, begin_mask = key_cache_internal_tensor_assign_8_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_8_squeeze_mask_0, stride = key_cache_internal_tensor_assign_8_stride_0, update = k_state_15_cast_fp16, x = coreml_update_state_68)[name = string("key_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_8_cast_fp16, input = key_cache)[name = string("coreml_update_state_70_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_70 = read_state(input = key_cache)[name = string("coreml_update_state_70")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_8_stride_0 = const()[name = string("value_cache_internal_tensor_assign_8_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_8_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_8_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_8_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_8_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_15_cast_fp16 = transpose(perm = v_state_15_perm_0, x = var_1509_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_8_cast_fp16 = slice_update(begin = concat_138, begin_mask = value_cache_internal_tensor_assign_8_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_8_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_8_squeeze_mask_0, stride = value_cache_internal_tensor_assign_8_stride_0, update = v_state_15_cast_fp16, x = coreml_update_state_69)[name = string("value_cache_internal_tensor_assign_8_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_8_cast_fp16, input = value_cache)[name = string("coreml_update_state_71_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_71 = read_state(input = value_cache)[name = string("coreml_update_state_71")];
            tensor<int32, [5]> var_1566_begin_0 = const()[name = string("op_1566_begin_0"), val = tensor<int32, [5]>([7, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1566_end_0 = const()[name = string("op_1566_end_0"), val = tensor<int32, [5]>([8, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_1566_end_mask_0 = const()[name = string("op_1566_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1566_squeeze_mask_0 = const()[name = string("op_1566_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_1566_cast_fp16 = slice_by_index(begin = var_1566_begin_0, end = var_1566_end_0, end_mask = var_1566_end_mask_0, squeeze_mask = var_1566_squeeze_mask_0, x = coreml_update_state_70)[name = string("op_1566_cast_fp16")];
            tensor<int32, [4]> var_1569_begin_0 = const()[name = string("op_1569_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1569_end_mask_0 = const()[name = string("op_1569_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_1569_cast_fp16 = slice_by_index(begin = var_1569_begin_0, end = concat_11, end_mask = var_1569_end_mask_0, x = var_1566_cast_fp16)[name = string("op_1569_cast_fp16")];
            tensor<int32, [5]> var_1571_begin_0 = const()[name = string("op_1571_begin_0"), val = tensor<int32, [5]>([7, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1571_end_0 = const()[name = string("op_1571_end_0"), val = tensor<int32, [5]>([8, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_1571_end_mask_0 = const()[name = string("op_1571_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1571_squeeze_mask_0 = const()[name = string("op_1571_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_1571_cast_fp16 = slice_by_index(begin = var_1571_begin_0, end = var_1571_end_0, end_mask = var_1571_end_mask_0, squeeze_mask = var_1571_squeeze_mask_0, x = coreml_update_state_71)[name = string("op_1571_cast_fp16")];
            tensor<int32, [4]> var_1574_begin_0 = const()[name = string("op_1574_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1574_end_mask_0 = const()[name = string("op_1574_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_1574_cast_fp16 = slice_by_index(begin = var_1574_begin_0, end = concat_11, end_mask = var_1574_end_mask_0, x = var_1571_cast_fp16)[name = string("op_1574_cast_fp16")];
            tensor<int32, [4]> var_1576_shape_cast_fp16 = shape(x = var_1569_cast_fp16)[name = string("op_1576_shape_cast_fp16")];
            int32 gather_139 = const()[name = string("gather_139"), val = int32(1)];
            int32 gather_140 = const()[name = string("gather_140"), val = int32(2)];
            int32 gather_141_axis_0 = const()[name = string("gather_141_axis_0"), val = int32(0)];
            int32 gather_141_batch_dims_0 = const()[name = string("gather_141_batch_dims_0"), val = int32(0)];
            bool gather_141_validate_indices_0 = const()[name = string("gather_141_validate_indices_0"), val = bool(false)];
            string var_1576_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1576_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_141_to_uint16 = const()[name = string("select_141_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1576_shape_cast_fp16_to_uint16 = cast(dtype = var_1576_shape_cast_fp16_to_uint16_dtype_0, x = var_1576_shape_cast_fp16)[name = string("cast_125")];
            uint16 gather_141_cast_uint16 = gather(axis = gather_141_axis_0, batch_dims = gather_141_batch_dims_0, indices = select_141_to_uint16, validate_indices = gather_141_validate_indices_0, x = var_1576_shape_cast_fp16_to_uint16)[name = string("gather_141_cast_uint16")];
            string gather_141_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_141_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_142 = const()[name = string("gather_142"), val = int32(128)];
            tensor<int32, [1]> var_1583_axes_0 = const()[name = string("op_1583_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_1583_cast_fp16 = expand_dims(axes = var_1583_axes_0, x = var_1569_cast_fp16)[name = string("op_1583_cast_fp16")];
            tensor<int32, [5]> shape_157_cast_fp16 = shape(x = var_1583_cast_fp16)[name = string("shape_157_cast_fp16")];
            int32 concat_146_axis_0 = const()[name = string("concat_146_axis_0"), val = int32(0)];
            bool concat_146_interleave_0 = const()[name = string("concat_146_interleave_0"), val = bool(false)];
            int32 gather_141_cast_uint16_to_int32 = cast(dtype = gather_141_cast_uint16_to_int32_dtype_0, x = gather_141_cast_uint16)[name = string("cast_124")];
            tensor<int32, [5]> concat_146 = concat(axis = concat_146_axis_0, interleave = concat_146_interleave_0, values = (gather_139, gather_140, var_77, gather_141_cast_uint16_to_int32, gather_142))[name = string("concat_146")];
            tensor<int32, [5]> real_div_14 = real_div(x = concat_146, y = shape_157_cast_fp16)[name = string("real_div_14")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_221_cast_fp16 = tile(reps = real_div_14, x = var_1583_cast_fp16)[name = string("hidden_states_221_cast_fp16")];
            tensor<int32, [4]> concat_147x = const()[name = string("concat_147x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_31_cast_fp16 = reshape(shape = concat_147x, x = hidden_states_221_cast_fp16)[name = string("key_states_31_cast_fp16")];
            tensor<int32, [4]> var_1593_shape_cast_fp16 = shape(x = var_1574_cast_fp16)[name = string("op_1593_shape_cast_fp16")];
            int32 gather_143 = const()[name = string("gather_143"), val = int32(1)];
            int32 gather_144 = const()[name = string("gather_144"), val = int32(2)];
            int32 gather_145_axis_0 = const()[name = string("gather_145_axis_0"), val = int32(0)];
            int32 gather_145_batch_dims_0 = const()[name = string("gather_145_batch_dims_0"), val = int32(0)];
            bool gather_145_validate_indices_0 = const()[name = string("gather_145_validate_indices_0"), val = bool(false)];
            string var_1593_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1593_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_145_to_uint16 = const()[name = string("select_145_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1593_shape_cast_fp16_to_uint16 = cast(dtype = var_1593_shape_cast_fp16_to_uint16_dtype_0, x = var_1593_shape_cast_fp16)[name = string("cast_123")];
            uint16 gather_145_cast_uint16 = gather(axis = gather_145_axis_0, batch_dims = gather_145_batch_dims_0, indices = select_145_to_uint16, validate_indices = gather_145_validate_indices_0, x = var_1593_shape_cast_fp16_to_uint16)[name = string("gather_145_cast_uint16")];
            string gather_145_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_145_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_146 = const()[name = string("gather_146"), val = int32(128)];
            tensor<int32, [1]> var_1600_axes_0 = const()[name = string("op_1600_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_1600_cast_fp16 = expand_dims(axes = var_1600_axes_0, x = var_1574_cast_fp16)[name = string("op_1600_cast_fp16")];
            tensor<int32, [5]> shape_162_cast_fp16 = shape(x = var_1600_cast_fp16)[name = string("shape_162_cast_fp16")];
            int32 concat_148_axis_0 = const()[name = string("concat_148_axis_0"), val = int32(0)];
            bool concat_148_interleave_0 = const()[name = string("concat_148_interleave_0"), val = bool(false)];
            int32 gather_145_cast_uint16_to_int32 = cast(dtype = gather_145_cast_uint16_to_int32_dtype_0, x = gather_145_cast_uint16)[name = string("cast_122")];
            tensor<int32, [5]> concat_148 = concat(axis = concat_148_axis_0, interleave = concat_148_interleave_0, values = (gather_143, gather_144, var_77, gather_145_cast_uint16_to_int32, gather_146))[name = string("concat_148")];
            tensor<int32, [5]> real_div_15 = real_div(x = concat_148, y = shape_162_cast_fp16)[name = string("real_div_15")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_225_cast_fp16 = tile(reps = real_div_15, x = var_1600_cast_fp16)[name = string("hidden_states_225_cast_fp16")];
            tensor<int32, [4]> concat_149x = const()[name = string("concat_149x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_31_cast_fp16 = reshape(shape = concat_149x, x = hidden_states_225_cast_fp16)[name = string("value_states_31_cast_fp16")];
            tensor<int32, [4]> var_1610_shape_cast_fp16 = shape(x = key_states_31_cast_fp16)[name = string("op_1610_shape_cast_fp16")];
            int32 gather_147_axis_0 = const()[name = string("gather_147_axis_0"), val = int32(0)];
            int32 gather_147_batch_dims_0 = const()[name = string("gather_147_batch_dims_0"), val = int32(0)];
            bool gather_147_validate_indices_0 = const()[name = string("gather_147_validate_indices_0"), val = bool(false)];
            string var_1610_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1610_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_147_to_uint16 = const()[name = string("select_147_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1610_shape_cast_fp16_to_uint16 = cast(dtype = var_1610_shape_cast_fp16_to_uint16_dtype_0, x = var_1610_shape_cast_fp16)[name = string("cast_121")];
            uint16 gather_147_cast_uint16 = gather(axis = gather_147_axis_0, batch_dims = gather_147_batch_dims_0, indices = select_147_to_uint16, validate_indices = gather_147_validate_indices_0, x = var_1610_shape_cast_fp16_to_uint16)[name = string("gather_147_cast_uint16")];
            string gather_147_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_147_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_150_values0_0 = const()[name = string("concat_150_values0_0"), val = int32(1)];
            int32 concat_150_values1_0 = const()[name = string("concat_150_values1_0"), val = int32(1)];
            int32 concat_150_values2_0 = const()[name = string("concat_150_values2_0"), val = int32(0)];
            int32 concat_150_axis_0 = const()[name = string("concat_150_axis_0"), val = int32(0)];
            bool concat_150_interleave_0 = const()[name = string("concat_150_interleave_0"), val = bool(false)];
            int32 gather_147_cast_uint16_to_int32 = cast(dtype = gather_147_cast_uint16_to_int32_dtype_0, x = gather_147_cast_uint16)[name = string("cast_120")];
            tensor<int32, [4]> concat_150 = concat(axis = concat_150_axis_0, interleave = concat_150_interleave_0, values = (concat_150_values0_0, concat_150_values1_0, concat_150_values2_0, gather_147_cast_uint16_to_int32))[name = string("concat_150")];
            tensor<int32, [4]> causal_mask_17_begin_0 = const()[name = string("causal_mask_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_17_end_mask_0 = const()[name = string("causal_mask_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_17_cast_fp16 = slice_by_index(begin = causal_mask_17_begin_0, end = concat_150, end_mask = causal_mask_17_end_mask_0, x = causal_mask)[name = string("causal_mask_17_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_29_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_17_cast_fp16, key = key_states_31_cast_fp16, query = query_states_31_cast_fp16, value = value_states_31_cast_fp16)[name = string("attn_output_29_cast_fp16")];
            tensor<int32, [4]> var_1616_perm_0 = const()[name = string("op_1616_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_151x = const()[name = string("concat_151x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_1616_cast_fp16 = transpose(perm = var_1616_perm_0, x = attn_output_29_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, ?, 1536]> input_57_cast_fp16 = reshape(shape = concat_151x, x = var_1616_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317397952))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318577664))))[name = string("model_model_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_52_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = input_57_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_229_cast_fp16 = add(x = hidden_states_209_cast_fp16, y = linear_52_cast_fp16)[name = string("hidden_states_229_cast_fp16")];
            fp16 var_78_promoted_15_to_fp16 = const()[name = string("op_78_promoted_15_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_1625_cast_fp16 = pow(x = hidden_states_229_cast_fp16, y = var_78_promoted_15_to_fp16)[name = string("op_1625_cast_fp16")];
            tensor<int32, [1]> variance_31_axes_0 = const()[name = string("variance_31_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_31_keep_dims_0 = const()[name = string("variance_31_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_31_cast_fp16 = reduce_mean(axes = variance_31_axes_0, keep_dims = variance_31_keep_dims_0, x = var_1625_cast_fp16)[name = string("variance_31_cast_fp16")];
            fp16 var_1628_to_fp16 = const()[name = string("op_1628_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1629_cast_fp16 = add(x = variance_31_cast_fp16, y = var_1628_to_fp16)[name = string("op_1629_cast_fp16")];
            fp32 var_1630_epsilon_0 = const()[name = string("op_1630_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1630_cast_fp16 = rsqrt(epsilon = var_1630_epsilon_0, x = var_1629_cast_fp16)[name = string("op_1630_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_233_cast_fp16 = mul(x = hidden_states_229_cast_fp16, y = var_1630_cast_fp16)[name = string("hidden_states_233_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318725184)))];
            tensor<fp16, [1, ?, 1536]> input_59_cast_fp16 = mul(x = model_model_layers_7_post_attention_layernorm_weight_to_fp16, y = hidden_states_233_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318728320))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325609664))))[name = string("model_model_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_1642_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_1642_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326469888))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333351232))))[name = string("model_model_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_63_cast_fp16 = mul(x = var_1642_cast_fp16, y = linear_54_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334211456))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341092800))))[name = string("model_model_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_239_cast_fp16 = add(x = hidden_states_229_cast_fp16, y = linear_55_cast_fp16)[name = string("hidden_states_239_cast_fp16")];
            fp16 var_78_promoted_16_to_fp16 = const()[name = string("op_78_promoted_16_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_1655_cast_fp16 = pow(x = hidden_states_239_cast_fp16, y = var_78_promoted_16_to_fp16)[name = string("op_1655_cast_fp16")];
            tensor<int32, [1]> variance_33_axes_0 = const()[name = string("variance_33_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_33_keep_dims_0 = const()[name = string("variance_33_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_33_cast_fp16 = reduce_mean(axes = variance_33_axes_0, keep_dims = variance_33_keep_dims_0, x = var_1655_cast_fp16)[name = string("variance_33_cast_fp16")];
            fp16 var_1658_to_fp16 = const()[name = string("op_1658_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1659_cast_fp16 = add(x = variance_33_cast_fp16, y = var_1658_to_fp16)[name = string("op_1659_cast_fp16")];
            fp32 var_1660_epsilon_0 = const()[name = string("op_1660_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1660_cast_fp16 = rsqrt(epsilon = var_1660_epsilon_0, x = var_1659_cast_fp16)[name = string("op_1660_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_243_cast_fp16 = mul(x = hidden_states_239_cast_fp16, y = var_1660_cast_fp16)[name = string("hidden_states_243_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341953024)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_247_cast_fp16 = mul(x = model_model_layers_8_input_layernorm_weight_to_fp16, y = hidden_states_243_cast_fp16)[name = string("hidden_states_247_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341956160))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343135872))))[name = string("model_model_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343283392)))];
            tensor<fp16, [1, ?, 1536]> linear_56_cast_fp16 = linear(bias = model_model_layers_8_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_247_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343286528))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343483200))))[name = string("model_model_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_8_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_8_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343507840)))];
            tensor<fp16, [1, ?, 256]> linear_57_cast_fp16 = linear(bias = model_model_layers_8_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_247_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343508416))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343705088))))[name = string("model_model_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343729728)))];
            tensor<fp16, [1, ?, 256]> linear_58_cast_fp16 = linear(bias = model_model_layers_8_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_247_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> concat_152x = const()[name = string("concat_152x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_1683_cast_fp16 = reshape(shape = concat_152x, x = linear_56_cast_fp16)[name = string("op_1683_cast_fp16")];
            tensor<int32, [4]> q_17_perm_0 = const()[name = string("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_153x = const()[name = string("concat_153x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_1686_cast_fp16 = reshape(shape = concat_153x, x = linear_57_cast_fp16)[name = string("op_1686_cast_fp16")];
            tensor<int32, [4]> k_17_perm_0 = const()[name = string("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_154x = const()[name = string("concat_154x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_1689_cast_fp16 = reshape(shape = concat_154x, x = linear_58_cast_fp16)[name = string("op_1689_cast_fp16")];
            tensor<int32, [4]> v_state_17_perm_0 = const()[name = string("v_state_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_17_cast_fp16 = transpose(perm = q_17_perm_0, x = var_1683_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 12, ?, 128]> var_1693_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1693_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_17_cast_fp16)[name = string("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_17_cast_fp16)[name = string("x2_33_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_1704_cast_fp16 = mul(x = x2_33_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_1704_cast_fp16")];
            bool var_1706_interleave_0 = const()[name = string("op_1706_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_1706_cast_fp16 = concat(axis = var_72, interleave = var_1706_interleave_0, values = (var_1704_cast_fp16, x1_33_cast_fp16))[name = string("op_1706_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_1707_cast_fp16 = mul(x = var_1706_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1707_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_35_cast_fp16 = add(x = var_1693_cast_fp16, y = var_1707_cast_fp16)[name = string("query_states_35_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_17_cast_fp16 = transpose(perm = k_17_perm_0, x = var_1686_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2, ?, 128]> var_1709_cast_fp16 = mul(x = k_17_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1709_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_17_cast_fp16)[name = string("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_17_cast_fp16)[name = string("x2_35_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_1720_cast_fp16 = mul(x = x2_35_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_1720_cast_fp16")];
            bool var_1722_interleave_0 = const()[name = string("op_1722_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_1722_cast_fp16 = concat(axis = var_72, interleave = var_1722_interleave_0, values = (var_1720_cast_fp16, x1_35_cast_fp16))[name = string("op_1722_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_1723_cast_fp16 = mul(x = var_1722_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1723_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_17_cast_fp16 = add(x = var_1709_cast_fp16, y = var_1723_cast_fp16)[name = string("k_state_17_cast_fp16")];
            tensor<int32, [1]> expand_dims_96 = const()[name = string("expand_dims_96"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_97 = const()[name = string("expand_dims_97"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_99 = const()[name = string("expand_dims_99"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_157_values0_0 = const()[name = string("concat_157_values0_0"), val = tensor<int32, [1]>([8])];
            int32 concat_157_axis_0 = const()[name = string("concat_157_axis_0"), val = int32(0)];
            bool concat_157_interleave_0 = const()[name = string("concat_157_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_157 = concat(axis = concat_157_axis_0, interleave = concat_157_interleave_0, values = (concat_157_values0_0, expand_dims_96, expand_dims_97, expand_dims_2, expand_dims_99))[name = string("concat_157")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_9_stride_0 = const()[name = string("key_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_157, begin_mask = key_cache_internal_tensor_assign_9_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_9_squeeze_mask_0, stride = key_cache_internal_tensor_assign_9_stride_0, update = k_state_17_cast_fp16, x = coreml_update_state_70)[name = string("key_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_9_cast_fp16, input = key_cache)[name = string("coreml_update_state_72_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_72 = read_state(input = key_cache)[name = string("coreml_update_state_72")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_9_stride_0 = const()[name = string("value_cache_internal_tensor_assign_9_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_9_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_9_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_9_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_9_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_17_cast_fp16 = transpose(perm = v_state_17_perm_0, x = var_1689_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_9_cast_fp16 = slice_update(begin = concat_157, begin_mask = value_cache_internal_tensor_assign_9_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_9_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_9_squeeze_mask_0, stride = value_cache_internal_tensor_assign_9_stride_0, update = v_state_17_cast_fp16, x = coreml_update_state_71)[name = string("value_cache_internal_tensor_assign_9_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_9_cast_fp16, input = value_cache)[name = string("coreml_update_state_73_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_73 = read_state(input = value_cache)[name = string("coreml_update_state_73")];
            tensor<int32, [5]> var_1746_begin_0 = const()[name = string("op_1746_begin_0"), val = tensor<int32, [5]>([8, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1746_end_0 = const()[name = string("op_1746_end_0"), val = tensor<int32, [5]>([9, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_1746_end_mask_0 = const()[name = string("op_1746_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1746_squeeze_mask_0 = const()[name = string("op_1746_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_1746_cast_fp16 = slice_by_index(begin = var_1746_begin_0, end = var_1746_end_0, end_mask = var_1746_end_mask_0, squeeze_mask = var_1746_squeeze_mask_0, x = coreml_update_state_72)[name = string("op_1746_cast_fp16")];
            tensor<int32, [4]> var_1749_begin_0 = const()[name = string("op_1749_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1749_end_mask_0 = const()[name = string("op_1749_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_1749_cast_fp16 = slice_by_index(begin = var_1749_begin_0, end = concat_11, end_mask = var_1749_end_mask_0, x = var_1746_cast_fp16)[name = string("op_1749_cast_fp16")];
            tensor<int32, [5]> var_1751_begin_0 = const()[name = string("op_1751_begin_0"), val = tensor<int32, [5]>([8, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1751_end_0 = const()[name = string("op_1751_end_0"), val = tensor<int32, [5]>([9, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_1751_end_mask_0 = const()[name = string("op_1751_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1751_squeeze_mask_0 = const()[name = string("op_1751_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_1751_cast_fp16 = slice_by_index(begin = var_1751_begin_0, end = var_1751_end_0, end_mask = var_1751_end_mask_0, squeeze_mask = var_1751_squeeze_mask_0, x = coreml_update_state_73)[name = string("op_1751_cast_fp16")];
            tensor<int32, [4]> var_1754_begin_0 = const()[name = string("op_1754_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1754_end_mask_0 = const()[name = string("op_1754_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_1754_cast_fp16 = slice_by_index(begin = var_1754_begin_0, end = concat_11, end_mask = var_1754_end_mask_0, x = var_1751_cast_fp16)[name = string("op_1754_cast_fp16")];
            tensor<int32, [4]> var_1756_shape_cast_fp16 = shape(x = var_1749_cast_fp16)[name = string("op_1756_shape_cast_fp16")];
            int32 gather_157 = const()[name = string("gather_157"), val = int32(1)];
            int32 gather_158 = const()[name = string("gather_158"), val = int32(2)];
            int32 gather_159_axis_0 = const()[name = string("gather_159_axis_0"), val = int32(0)];
            int32 gather_159_batch_dims_0 = const()[name = string("gather_159_batch_dims_0"), val = int32(0)];
            bool gather_159_validate_indices_0 = const()[name = string("gather_159_validate_indices_0"), val = bool(false)];
            string var_1756_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1756_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_159_to_uint16 = const()[name = string("select_159_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1756_shape_cast_fp16_to_uint16 = cast(dtype = var_1756_shape_cast_fp16_to_uint16_dtype_0, x = var_1756_shape_cast_fp16)[name = string("cast_119")];
            uint16 gather_159_cast_uint16 = gather(axis = gather_159_axis_0, batch_dims = gather_159_batch_dims_0, indices = select_159_to_uint16, validate_indices = gather_159_validate_indices_0, x = var_1756_shape_cast_fp16_to_uint16)[name = string("gather_159_cast_uint16")];
            string gather_159_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_159_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_160 = const()[name = string("gather_160"), val = int32(128)];
            tensor<int32, [1]> var_1763_axes_0 = const()[name = string("op_1763_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_1763_cast_fp16 = expand_dims(axes = var_1763_axes_0, x = var_1749_cast_fp16)[name = string("op_1763_cast_fp16")];
            tensor<int32, [5]> shape_177_cast_fp16 = shape(x = var_1763_cast_fp16)[name = string("shape_177_cast_fp16")];
            int32 concat_165_axis_0 = const()[name = string("concat_165_axis_0"), val = int32(0)];
            bool concat_165_interleave_0 = const()[name = string("concat_165_interleave_0"), val = bool(false)];
            int32 gather_159_cast_uint16_to_int32 = cast(dtype = gather_159_cast_uint16_to_int32_dtype_0, x = gather_159_cast_uint16)[name = string("cast_118")];
            tensor<int32, [5]> concat_165 = concat(axis = concat_165_axis_0, interleave = concat_165_interleave_0, values = (gather_157, gather_158, var_77, gather_159_cast_uint16_to_int32, gather_160))[name = string("concat_165")];
            tensor<int32, [5]> real_div_16 = real_div(x = concat_165, y = shape_177_cast_fp16)[name = string("real_div_16")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_251_cast_fp16 = tile(reps = real_div_16, x = var_1763_cast_fp16)[name = string("hidden_states_251_cast_fp16")];
            tensor<int32, [4]> concat_166x = const()[name = string("concat_166x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_35_cast_fp16 = reshape(shape = concat_166x, x = hidden_states_251_cast_fp16)[name = string("key_states_35_cast_fp16")];
            tensor<int32, [4]> var_1773_shape_cast_fp16 = shape(x = var_1754_cast_fp16)[name = string("op_1773_shape_cast_fp16")];
            int32 gather_161 = const()[name = string("gather_161"), val = int32(1)];
            int32 gather_162 = const()[name = string("gather_162"), val = int32(2)];
            int32 gather_163_axis_0 = const()[name = string("gather_163_axis_0"), val = int32(0)];
            int32 gather_163_batch_dims_0 = const()[name = string("gather_163_batch_dims_0"), val = int32(0)];
            bool gather_163_validate_indices_0 = const()[name = string("gather_163_validate_indices_0"), val = bool(false)];
            string var_1773_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1773_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_163_to_uint16 = const()[name = string("select_163_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1773_shape_cast_fp16_to_uint16 = cast(dtype = var_1773_shape_cast_fp16_to_uint16_dtype_0, x = var_1773_shape_cast_fp16)[name = string("cast_117")];
            uint16 gather_163_cast_uint16 = gather(axis = gather_163_axis_0, batch_dims = gather_163_batch_dims_0, indices = select_163_to_uint16, validate_indices = gather_163_validate_indices_0, x = var_1773_shape_cast_fp16_to_uint16)[name = string("gather_163_cast_uint16")];
            string gather_163_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_163_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_164 = const()[name = string("gather_164"), val = int32(128)];
            tensor<int32, [1]> var_1780_axes_0 = const()[name = string("op_1780_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_1780_cast_fp16 = expand_dims(axes = var_1780_axes_0, x = var_1754_cast_fp16)[name = string("op_1780_cast_fp16")];
            tensor<int32, [5]> shape_182_cast_fp16 = shape(x = var_1780_cast_fp16)[name = string("shape_182_cast_fp16")];
            int32 concat_167_axis_0 = const()[name = string("concat_167_axis_0"), val = int32(0)];
            bool concat_167_interleave_0 = const()[name = string("concat_167_interleave_0"), val = bool(false)];
            int32 gather_163_cast_uint16_to_int32 = cast(dtype = gather_163_cast_uint16_to_int32_dtype_0, x = gather_163_cast_uint16)[name = string("cast_116")];
            tensor<int32, [5]> concat_167 = concat(axis = concat_167_axis_0, interleave = concat_167_interleave_0, values = (gather_161, gather_162, var_77, gather_163_cast_uint16_to_int32, gather_164))[name = string("concat_167")];
            tensor<int32, [5]> real_div_17 = real_div(x = concat_167, y = shape_182_cast_fp16)[name = string("real_div_17")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_255_cast_fp16 = tile(reps = real_div_17, x = var_1780_cast_fp16)[name = string("hidden_states_255_cast_fp16")];
            tensor<int32, [4]> concat_168x = const()[name = string("concat_168x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_35_cast_fp16 = reshape(shape = concat_168x, x = hidden_states_255_cast_fp16)[name = string("value_states_35_cast_fp16")];
            tensor<int32, [4]> var_1790_shape_cast_fp16 = shape(x = key_states_35_cast_fp16)[name = string("op_1790_shape_cast_fp16")];
            int32 gather_165_axis_0 = const()[name = string("gather_165_axis_0"), val = int32(0)];
            int32 gather_165_batch_dims_0 = const()[name = string("gather_165_batch_dims_0"), val = int32(0)];
            bool gather_165_validate_indices_0 = const()[name = string("gather_165_validate_indices_0"), val = bool(false)];
            string var_1790_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1790_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_165_to_uint16 = const()[name = string("select_165_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1790_shape_cast_fp16_to_uint16 = cast(dtype = var_1790_shape_cast_fp16_to_uint16_dtype_0, x = var_1790_shape_cast_fp16)[name = string("cast_115")];
            uint16 gather_165_cast_uint16 = gather(axis = gather_165_axis_0, batch_dims = gather_165_batch_dims_0, indices = select_165_to_uint16, validate_indices = gather_165_validate_indices_0, x = var_1790_shape_cast_fp16_to_uint16)[name = string("gather_165_cast_uint16")];
            string gather_165_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_165_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_169_values0_0 = const()[name = string("concat_169_values0_0"), val = int32(1)];
            int32 concat_169_values1_0 = const()[name = string("concat_169_values1_0"), val = int32(1)];
            int32 concat_169_values2_0 = const()[name = string("concat_169_values2_0"), val = int32(0)];
            int32 concat_169_axis_0 = const()[name = string("concat_169_axis_0"), val = int32(0)];
            bool concat_169_interleave_0 = const()[name = string("concat_169_interleave_0"), val = bool(false)];
            int32 gather_165_cast_uint16_to_int32 = cast(dtype = gather_165_cast_uint16_to_int32_dtype_0, x = gather_165_cast_uint16)[name = string("cast_114")];
            tensor<int32, [4]> concat_169 = concat(axis = concat_169_axis_0, interleave = concat_169_interleave_0, values = (concat_169_values0_0, concat_169_values1_0, concat_169_values2_0, gather_165_cast_uint16_to_int32))[name = string("concat_169")];
            tensor<int32, [4]> causal_mask_19_begin_0 = const()[name = string("causal_mask_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_19_end_mask_0 = const()[name = string("causal_mask_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_19_cast_fp16 = slice_by_index(begin = causal_mask_19_begin_0, end = concat_169, end_mask = causal_mask_19_end_mask_0, x = causal_mask)[name = string("causal_mask_19_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_33_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_19_cast_fp16, key = key_states_35_cast_fp16, query = query_states_35_cast_fp16, value = value_states_35_cast_fp16)[name = string("attn_output_33_cast_fp16")];
            tensor<int32, [4]> var_1796_perm_0 = const()[name = string("op_1796_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_170x = const()[name = string("concat_170x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_1796_cast_fp16 = transpose(perm = var_1796_perm_0, x = attn_output_33_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, ?, 1536]> input_65_cast_fp16 = reshape(shape = concat_170x, x = var_1796_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343730304))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344910016))))[name = string("model_model_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_259_cast_fp16 = add(x = hidden_states_239_cast_fp16, y = linear_59_cast_fp16)[name = string("hidden_states_259_cast_fp16")];
            fp16 var_78_promoted_17_to_fp16 = const()[name = string("op_78_promoted_17_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_1805_cast_fp16 = pow(x = hidden_states_259_cast_fp16, y = var_78_promoted_17_to_fp16)[name = string("op_1805_cast_fp16")];
            tensor<int32, [1]> variance_35_axes_0 = const()[name = string("variance_35_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_35_keep_dims_0 = const()[name = string("variance_35_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_35_cast_fp16 = reduce_mean(axes = variance_35_axes_0, keep_dims = variance_35_keep_dims_0, x = var_1805_cast_fp16)[name = string("variance_35_cast_fp16")];
            fp16 var_1808_to_fp16 = const()[name = string("op_1808_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1809_cast_fp16 = add(x = variance_35_cast_fp16, y = var_1808_to_fp16)[name = string("op_1809_cast_fp16")];
            fp32 var_1810_epsilon_0 = const()[name = string("op_1810_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1810_cast_fp16 = rsqrt(epsilon = var_1810_epsilon_0, x = var_1809_cast_fp16)[name = string("op_1810_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_263_cast_fp16 = mul(x = hidden_states_259_cast_fp16, y = var_1810_cast_fp16)[name = string("hidden_states_263_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345057536)))];
            tensor<fp16, [1, ?, 1536]> input_67_cast_fp16 = mul(x = model_model_layers_8_post_attention_layernorm_weight_to_fp16, y = hidden_states_263_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345060672))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351942016))))[name = string("model_model_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_1822_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_1822_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352802240))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359683584))))[name = string("model_model_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_71_cast_fp16 = mul(x = var_1822_cast_fp16, y = linear_61_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360543808))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367425152))))[name = string("model_model_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_62_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_269_cast_fp16 = add(x = hidden_states_259_cast_fp16, y = linear_62_cast_fp16)[name = string("hidden_states_269_cast_fp16")];
            fp16 var_78_promoted_18_to_fp16 = const()[name = string("op_78_promoted_18_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_1835_cast_fp16 = pow(x = hidden_states_269_cast_fp16, y = var_78_promoted_18_to_fp16)[name = string("op_1835_cast_fp16")];
            tensor<int32, [1]> variance_37_axes_0 = const()[name = string("variance_37_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_37_keep_dims_0 = const()[name = string("variance_37_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_37_cast_fp16 = reduce_mean(axes = variance_37_axes_0, keep_dims = variance_37_keep_dims_0, x = var_1835_cast_fp16)[name = string("variance_37_cast_fp16")];
            fp16 var_1838_to_fp16 = const()[name = string("op_1838_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1839_cast_fp16 = add(x = variance_37_cast_fp16, y = var_1838_to_fp16)[name = string("op_1839_cast_fp16")];
            fp32 var_1840_epsilon_0 = const()[name = string("op_1840_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1840_cast_fp16 = rsqrt(epsilon = var_1840_epsilon_0, x = var_1839_cast_fp16)[name = string("op_1840_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_273_cast_fp16 = mul(x = hidden_states_269_cast_fp16, y = var_1840_cast_fp16)[name = string("hidden_states_273_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368285376)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_277_cast_fp16 = mul(x = model_model_layers_9_input_layernorm_weight_to_fp16, y = hidden_states_273_cast_fp16)[name = string("hidden_states_277_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368288512))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369468224))))[name = string("model_model_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369615744)))];
            tensor<fp16, [1, ?, 1536]> linear_63_cast_fp16 = linear(bias = model_model_layers_9_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_277_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369618880))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369815552))))[name = string("model_model_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_9_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_9_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369840192)))];
            tensor<fp16, [1, ?, 256]> linear_64_cast_fp16 = linear(bias = model_model_layers_9_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_277_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369840768))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370037440))))[name = string("model_model_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370062080)))];
            tensor<fp16, [1, ?, 256]> linear_65_cast_fp16 = linear(bias = model_model_layers_9_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_277_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> concat_171x = const()[name = string("concat_171x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_1863_cast_fp16 = reshape(shape = concat_171x, x = linear_63_cast_fp16)[name = string("op_1863_cast_fp16")];
            tensor<int32, [4]> q_19_perm_0 = const()[name = string("q_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_172x = const()[name = string("concat_172x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_1866_cast_fp16 = reshape(shape = concat_172x, x = linear_64_cast_fp16)[name = string("op_1866_cast_fp16")];
            tensor<int32, [4]> k_19_perm_0 = const()[name = string("k_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_173x = const()[name = string("concat_173x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_1869_cast_fp16 = reshape(shape = concat_173x, x = linear_65_cast_fp16)[name = string("op_1869_cast_fp16")];
            tensor<int32, [4]> v_state_19_perm_0 = const()[name = string("v_state_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_19_cast_fp16 = transpose(perm = q_19_perm_0, x = var_1863_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 12, ?, 128]> var_1873_cast_fp16 = mul(x = q_19_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1873_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_19_cast_fp16)[name = string("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_19_cast_fp16)[name = string("x2_37_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_1884_cast_fp16 = mul(x = x2_37_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_1884_cast_fp16")];
            bool var_1886_interleave_0 = const()[name = string("op_1886_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_1886_cast_fp16 = concat(axis = var_72, interleave = var_1886_interleave_0, values = (var_1884_cast_fp16, x1_37_cast_fp16))[name = string("op_1886_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_1887_cast_fp16 = mul(x = var_1886_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1887_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_39_cast_fp16 = add(x = var_1873_cast_fp16, y = var_1887_cast_fp16)[name = string("query_states_39_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_19_cast_fp16 = transpose(perm = k_19_perm_0, x = var_1866_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 2, ?, 128]> var_1889_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_7_cast_fp16)[name = string("op_1889_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_19_cast_fp16)[name = string("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_19_cast_fp16)[name = string("x2_39_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_1900_cast_fp16 = mul(x = x2_39_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1900_cast_fp16")];
            bool var_1902_interleave_0 = const()[name = string("op_1902_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_1902_cast_fp16 = concat(axis = var_72, interleave = var_1902_interleave_0, values = (var_1900_cast_fp16, x1_39_cast_fp16))[name = string("op_1902_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_1903_cast_fp16 = mul(x = var_1902_cast_fp16, y = sin_7_cast_fp16)[name = string("op_1903_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_19_cast_fp16 = add(x = var_1889_cast_fp16, y = var_1903_cast_fp16)[name = string("k_state_19_cast_fp16")];
            tensor<int32, [1]> expand_dims_108 = const()[name = string("expand_dims_108"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_109 = const()[name = string("expand_dims_109"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_111 = const()[name = string("expand_dims_111"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_176_values0_0 = const()[name = string("concat_176_values0_0"), val = tensor<int32, [1]>([9])];
            int32 concat_176_axis_0 = const()[name = string("concat_176_axis_0"), val = int32(0)];
            bool concat_176_interleave_0 = const()[name = string("concat_176_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_176 = concat(axis = concat_176_axis_0, interleave = concat_176_interleave_0, values = (concat_176_values0_0, expand_dims_108, expand_dims_109, expand_dims_2, expand_dims_111))[name = string("concat_176")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_10_stride_0 = const()[name = string("key_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_176, begin_mask = key_cache_internal_tensor_assign_10_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_10_squeeze_mask_0, stride = key_cache_internal_tensor_assign_10_stride_0, update = k_state_19_cast_fp16, x = coreml_update_state_72)[name = string("key_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_10_cast_fp16, input = key_cache)[name = string("coreml_update_state_74_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_74 = read_state(input = key_cache)[name = string("coreml_update_state_74")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_10_stride_0 = const()[name = string("value_cache_internal_tensor_assign_10_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_10_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_10_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_10_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_10_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_19_cast_fp16 = transpose(perm = v_state_19_perm_0, x = var_1869_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_10_cast_fp16 = slice_update(begin = concat_176, begin_mask = value_cache_internal_tensor_assign_10_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_10_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_10_squeeze_mask_0, stride = value_cache_internal_tensor_assign_10_stride_0, update = v_state_19_cast_fp16, x = coreml_update_state_73)[name = string("value_cache_internal_tensor_assign_10_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_10_cast_fp16, input = value_cache)[name = string("coreml_update_state_75_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_75 = read_state(input = value_cache)[name = string("coreml_update_state_75")];
            tensor<int32, [5]> var_1926_begin_0 = const()[name = string("op_1926_begin_0"), val = tensor<int32, [5]>([9, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1926_end_0 = const()[name = string("op_1926_end_0"), val = tensor<int32, [5]>([10, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_1926_end_mask_0 = const()[name = string("op_1926_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1926_squeeze_mask_0 = const()[name = string("op_1926_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_1926_cast_fp16 = slice_by_index(begin = var_1926_begin_0, end = var_1926_end_0, end_mask = var_1926_end_mask_0, squeeze_mask = var_1926_squeeze_mask_0, x = coreml_update_state_74)[name = string("op_1926_cast_fp16")];
            tensor<int32, [4]> var_1929_begin_0 = const()[name = string("op_1929_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1929_end_mask_0 = const()[name = string("op_1929_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_1929_cast_fp16 = slice_by_index(begin = var_1929_begin_0, end = concat_11, end_mask = var_1929_end_mask_0, x = var_1926_cast_fp16)[name = string("op_1929_cast_fp16")];
            tensor<int32, [5]> var_1931_begin_0 = const()[name = string("op_1931_begin_0"), val = tensor<int32, [5]>([9, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1931_end_0 = const()[name = string("op_1931_end_0"), val = tensor<int32, [5]>([10, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_1931_end_mask_0 = const()[name = string("op_1931_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1931_squeeze_mask_0 = const()[name = string("op_1931_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_1931_cast_fp16 = slice_by_index(begin = var_1931_begin_0, end = var_1931_end_0, end_mask = var_1931_end_mask_0, squeeze_mask = var_1931_squeeze_mask_0, x = coreml_update_state_75)[name = string("op_1931_cast_fp16")];
            tensor<int32, [4]> var_1934_begin_0 = const()[name = string("op_1934_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_1934_end_mask_0 = const()[name = string("op_1934_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_1934_cast_fp16 = slice_by_index(begin = var_1934_begin_0, end = concat_11, end_mask = var_1934_end_mask_0, x = var_1931_cast_fp16)[name = string("op_1934_cast_fp16")];
            tensor<int32, [4]> var_1936_shape_cast_fp16 = shape(x = var_1929_cast_fp16)[name = string("op_1936_shape_cast_fp16")];
            int32 gather_175 = const()[name = string("gather_175"), val = int32(1)];
            int32 gather_176 = const()[name = string("gather_176"), val = int32(2)];
            int32 gather_177_axis_0 = const()[name = string("gather_177_axis_0"), val = int32(0)];
            int32 gather_177_batch_dims_0 = const()[name = string("gather_177_batch_dims_0"), val = int32(0)];
            bool gather_177_validate_indices_0 = const()[name = string("gather_177_validate_indices_0"), val = bool(false)];
            string var_1936_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1936_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_177_to_uint16 = const()[name = string("select_177_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1936_shape_cast_fp16_to_uint16 = cast(dtype = var_1936_shape_cast_fp16_to_uint16_dtype_0, x = var_1936_shape_cast_fp16)[name = string("cast_113")];
            uint16 gather_177_cast_uint16 = gather(axis = gather_177_axis_0, batch_dims = gather_177_batch_dims_0, indices = select_177_to_uint16, validate_indices = gather_177_validate_indices_0, x = var_1936_shape_cast_fp16_to_uint16)[name = string("gather_177_cast_uint16")];
            string gather_177_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_177_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_178 = const()[name = string("gather_178"), val = int32(128)];
            tensor<int32, [1]> var_1943_axes_0 = const()[name = string("op_1943_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_1943_cast_fp16 = expand_dims(axes = var_1943_axes_0, x = var_1929_cast_fp16)[name = string("op_1943_cast_fp16")];
            tensor<int32, [5]> shape_197_cast_fp16 = shape(x = var_1943_cast_fp16)[name = string("shape_197_cast_fp16")];
            int32 concat_184_axis_0 = const()[name = string("concat_184_axis_0"), val = int32(0)];
            bool concat_184_interleave_0 = const()[name = string("concat_184_interleave_0"), val = bool(false)];
            int32 gather_177_cast_uint16_to_int32 = cast(dtype = gather_177_cast_uint16_to_int32_dtype_0, x = gather_177_cast_uint16)[name = string("cast_112")];
            tensor<int32, [5]> concat_184 = concat(axis = concat_184_axis_0, interleave = concat_184_interleave_0, values = (gather_175, gather_176, var_77, gather_177_cast_uint16_to_int32, gather_178))[name = string("concat_184")];
            tensor<int32, [5]> real_div_18 = real_div(x = concat_184, y = shape_197_cast_fp16)[name = string("real_div_18")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_281_cast_fp16 = tile(reps = real_div_18, x = var_1943_cast_fp16)[name = string("hidden_states_281_cast_fp16")];
            tensor<int32, [4]> concat_185x = const()[name = string("concat_185x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_39_cast_fp16 = reshape(shape = concat_185x, x = hidden_states_281_cast_fp16)[name = string("key_states_39_cast_fp16")];
            tensor<int32, [4]> var_1953_shape_cast_fp16 = shape(x = var_1934_cast_fp16)[name = string("op_1953_shape_cast_fp16")];
            int32 gather_179 = const()[name = string("gather_179"), val = int32(1)];
            int32 gather_180 = const()[name = string("gather_180"), val = int32(2)];
            int32 gather_181_axis_0 = const()[name = string("gather_181_axis_0"), val = int32(0)];
            int32 gather_181_batch_dims_0 = const()[name = string("gather_181_batch_dims_0"), val = int32(0)];
            bool gather_181_validate_indices_0 = const()[name = string("gather_181_validate_indices_0"), val = bool(false)];
            string var_1953_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1953_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_181_to_uint16 = const()[name = string("select_181_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1953_shape_cast_fp16_to_uint16 = cast(dtype = var_1953_shape_cast_fp16_to_uint16_dtype_0, x = var_1953_shape_cast_fp16)[name = string("cast_111")];
            uint16 gather_181_cast_uint16 = gather(axis = gather_181_axis_0, batch_dims = gather_181_batch_dims_0, indices = select_181_to_uint16, validate_indices = gather_181_validate_indices_0, x = var_1953_shape_cast_fp16_to_uint16)[name = string("gather_181_cast_uint16")];
            string gather_181_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_181_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_182 = const()[name = string("gather_182"), val = int32(128)];
            tensor<int32, [1]> var_1960_axes_0 = const()[name = string("op_1960_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_1960_cast_fp16 = expand_dims(axes = var_1960_axes_0, x = var_1934_cast_fp16)[name = string("op_1960_cast_fp16")];
            tensor<int32, [5]> shape_202_cast_fp16 = shape(x = var_1960_cast_fp16)[name = string("shape_202_cast_fp16")];
            int32 concat_186_axis_0 = const()[name = string("concat_186_axis_0"), val = int32(0)];
            bool concat_186_interleave_0 = const()[name = string("concat_186_interleave_0"), val = bool(false)];
            int32 gather_181_cast_uint16_to_int32 = cast(dtype = gather_181_cast_uint16_to_int32_dtype_0, x = gather_181_cast_uint16)[name = string("cast_110")];
            tensor<int32, [5]> concat_186 = concat(axis = concat_186_axis_0, interleave = concat_186_interleave_0, values = (gather_179, gather_180, var_77, gather_181_cast_uint16_to_int32, gather_182))[name = string("concat_186")];
            tensor<int32, [5]> real_div_19 = real_div(x = concat_186, y = shape_202_cast_fp16)[name = string("real_div_19")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_285_cast_fp16 = tile(reps = real_div_19, x = var_1960_cast_fp16)[name = string("hidden_states_285_cast_fp16")];
            tensor<int32, [4]> concat_187x = const()[name = string("concat_187x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_39_cast_fp16 = reshape(shape = concat_187x, x = hidden_states_285_cast_fp16)[name = string("value_states_39_cast_fp16")];
            tensor<int32, [4]> var_1970_shape_cast_fp16 = shape(x = key_states_39_cast_fp16)[name = string("op_1970_shape_cast_fp16")];
            int32 gather_183_axis_0 = const()[name = string("gather_183_axis_0"), val = int32(0)];
            int32 gather_183_batch_dims_0 = const()[name = string("gather_183_batch_dims_0"), val = int32(0)];
            bool gather_183_validate_indices_0 = const()[name = string("gather_183_validate_indices_0"), val = bool(false)];
            string var_1970_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1970_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_183_to_uint16 = const()[name = string("select_183_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_1970_shape_cast_fp16_to_uint16 = cast(dtype = var_1970_shape_cast_fp16_to_uint16_dtype_0, x = var_1970_shape_cast_fp16)[name = string("cast_109")];
            uint16 gather_183_cast_uint16 = gather(axis = gather_183_axis_0, batch_dims = gather_183_batch_dims_0, indices = select_183_to_uint16, validate_indices = gather_183_validate_indices_0, x = var_1970_shape_cast_fp16_to_uint16)[name = string("gather_183_cast_uint16")];
            string gather_183_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_183_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_188_values0_0 = const()[name = string("concat_188_values0_0"), val = int32(1)];
            int32 concat_188_values1_0 = const()[name = string("concat_188_values1_0"), val = int32(1)];
            int32 concat_188_values2_0 = const()[name = string("concat_188_values2_0"), val = int32(0)];
            int32 concat_188_axis_0 = const()[name = string("concat_188_axis_0"), val = int32(0)];
            bool concat_188_interleave_0 = const()[name = string("concat_188_interleave_0"), val = bool(false)];
            int32 gather_183_cast_uint16_to_int32 = cast(dtype = gather_183_cast_uint16_to_int32_dtype_0, x = gather_183_cast_uint16)[name = string("cast_108")];
            tensor<int32, [4]> concat_188 = concat(axis = concat_188_axis_0, interleave = concat_188_interleave_0, values = (concat_188_values0_0, concat_188_values1_0, concat_188_values2_0, gather_183_cast_uint16_to_int32))[name = string("concat_188")];
            tensor<int32, [4]> causal_mask_21_begin_0 = const()[name = string("causal_mask_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_21_end_mask_0 = const()[name = string("causal_mask_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_21_cast_fp16 = slice_by_index(begin = causal_mask_21_begin_0, end = concat_188, end_mask = causal_mask_21_end_mask_0, x = causal_mask)[name = string("causal_mask_21_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_37_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_21_cast_fp16, key = key_states_39_cast_fp16, query = query_states_39_cast_fp16, value = value_states_39_cast_fp16)[name = string("attn_output_37_cast_fp16")];
            tensor<int32, [4]> var_1976_perm_0 = const()[name = string("op_1976_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_189x = const()[name = string("concat_189x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_1976_cast_fp16 = transpose(perm = var_1976_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, ?, 1536]> input_73_cast_fp16 = reshape(shape = concat_189x, x = var_1976_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370062656))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371242368))))[name = string("model_model_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_66_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_289_cast_fp16 = add(x = hidden_states_269_cast_fp16, y = linear_66_cast_fp16)[name = string("hidden_states_289_cast_fp16")];
            fp16 var_78_promoted_19_to_fp16 = const()[name = string("op_78_promoted_19_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_1985_cast_fp16 = pow(x = hidden_states_289_cast_fp16, y = var_78_promoted_19_to_fp16)[name = string("op_1985_cast_fp16")];
            tensor<int32, [1]> variance_39_axes_0 = const()[name = string("variance_39_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_39_keep_dims_0 = const()[name = string("variance_39_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_39_cast_fp16 = reduce_mean(axes = variance_39_axes_0, keep_dims = variance_39_keep_dims_0, x = var_1985_cast_fp16)[name = string("variance_39_cast_fp16")];
            fp16 var_1988_to_fp16 = const()[name = string("op_1988_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1989_cast_fp16 = add(x = variance_39_cast_fp16, y = var_1988_to_fp16)[name = string("op_1989_cast_fp16")];
            fp32 var_1990_epsilon_0 = const()[name = string("op_1990_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_1990_cast_fp16 = rsqrt(epsilon = var_1990_epsilon_0, x = var_1989_cast_fp16)[name = string("op_1990_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_293_cast_fp16 = mul(x = hidden_states_289_cast_fp16, y = var_1990_cast_fp16)[name = string("hidden_states_293_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371389888)))];
            tensor<fp16, [1, ?, 1536]> input_75_cast_fp16 = mul(x = model_model_layers_9_post_attention_layernorm_weight_to_fp16, y = hidden_states_293_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371393024))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378274368))))[name = string("model_model_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_2002_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_2002_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379134592))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386015936))))[name = string("model_model_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_79_cast_fp16 = mul(x = var_2002_cast_fp16, y = linear_68_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386876160))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393757504))))[name = string("model_model_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_69_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_299_cast_fp16 = add(x = hidden_states_289_cast_fp16, y = linear_69_cast_fp16)[name = string("hidden_states_299_cast_fp16")];
            fp16 var_78_promoted_20_to_fp16 = const()[name = string("op_78_promoted_20_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_2015_cast_fp16 = pow(x = hidden_states_299_cast_fp16, y = var_78_promoted_20_to_fp16)[name = string("op_2015_cast_fp16")];
            tensor<int32, [1]> variance_41_axes_0 = const()[name = string("variance_41_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_41_keep_dims_0 = const()[name = string("variance_41_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_41_cast_fp16 = reduce_mean(axes = variance_41_axes_0, keep_dims = variance_41_keep_dims_0, x = var_2015_cast_fp16)[name = string("variance_41_cast_fp16")];
            fp16 var_2018_to_fp16 = const()[name = string("op_2018_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2019_cast_fp16 = add(x = variance_41_cast_fp16, y = var_2018_to_fp16)[name = string("op_2019_cast_fp16")];
            fp32 var_2020_epsilon_0 = const()[name = string("op_2020_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2020_cast_fp16 = rsqrt(epsilon = var_2020_epsilon_0, x = var_2019_cast_fp16)[name = string("op_2020_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_303_cast_fp16 = mul(x = hidden_states_299_cast_fp16, y = var_2020_cast_fp16)[name = string("hidden_states_303_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394617728)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_307_cast_fp16 = mul(x = model_model_layers_10_input_layernorm_weight_to_fp16, y = hidden_states_303_cast_fp16)[name = string("hidden_states_307_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394620864))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395800576))))[name = string("model_model_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395948096)))];
            tensor<fp16, [1, ?, 1536]> linear_70_cast_fp16 = linear(bias = model_model_layers_10_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_307_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395951232))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396147904))))[name = string("model_model_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_10_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_10_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396172544)))];
            tensor<fp16, [1, ?, 256]> linear_71_cast_fp16 = linear(bias = model_model_layers_10_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_307_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396173120))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396369792))))[name = string("model_model_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396394432)))];
            tensor<fp16, [1, ?, 256]> linear_72_cast_fp16 = linear(bias = model_model_layers_10_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_307_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> concat_190x = const()[name = string("concat_190x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_2043_cast_fp16 = reshape(shape = concat_190x, x = linear_70_cast_fp16)[name = string("op_2043_cast_fp16")];
            tensor<int32, [4]> q_21_perm_0 = const()[name = string("q_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_191x = const()[name = string("concat_191x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_2046_cast_fp16 = reshape(shape = concat_191x, x = linear_71_cast_fp16)[name = string("op_2046_cast_fp16")];
            tensor<int32, [4]> k_21_perm_0 = const()[name = string("k_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_192x = const()[name = string("concat_192x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_2049_cast_fp16 = reshape(shape = concat_192x, x = linear_72_cast_fp16)[name = string("op_2049_cast_fp16")];
            tensor<int32, [4]> v_state_21_perm_0 = const()[name = string("v_state_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = var_2043_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 12, ?, 128]> var_2053_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2053_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_21_cast_fp16)[name = string("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_21_cast_fp16)[name = string("x2_41_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_2064_cast_fp16 = mul(x = x2_41_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_2064_cast_fp16")];
            bool var_2066_interleave_0 = const()[name = string("op_2066_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_2066_cast_fp16 = concat(axis = var_72, interleave = var_2066_interleave_0, values = (var_2064_cast_fp16, x1_41_cast_fp16))[name = string("op_2066_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_2067_cast_fp16 = mul(x = var_2066_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2067_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_43_cast_fp16 = add(x = var_2053_cast_fp16, y = var_2067_cast_fp16)[name = string("query_states_43_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = var_2046_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 2, ?, 128]> var_2069_cast_fp16 = mul(x = k_21_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2069_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_21_cast_fp16)[name = string("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_21_cast_fp16)[name = string("x2_43_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_2080_cast_fp16 = mul(x = x2_43_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_2080_cast_fp16")];
            bool var_2082_interleave_0 = const()[name = string("op_2082_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_2082_cast_fp16 = concat(axis = var_72, interleave = var_2082_interleave_0, values = (var_2080_cast_fp16, x1_43_cast_fp16))[name = string("op_2082_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_2083_cast_fp16 = mul(x = var_2082_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2083_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_21_cast_fp16 = add(x = var_2069_cast_fp16, y = var_2083_cast_fp16)[name = string("k_state_21_cast_fp16")];
            tensor<int32, [1]> expand_dims_120 = const()[name = string("expand_dims_120"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_121 = const()[name = string("expand_dims_121"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_123 = const()[name = string("expand_dims_123"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_195_values0_0 = const()[name = string("concat_195_values0_0"), val = tensor<int32, [1]>([10])];
            int32 concat_195_axis_0 = const()[name = string("concat_195_axis_0"), val = int32(0)];
            bool concat_195_interleave_0 = const()[name = string("concat_195_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_195 = concat(axis = concat_195_axis_0, interleave = concat_195_interleave_0, values = (concat_195_values0_0, expand_dims_120, expand_dims_121, expand_dims_2, expand_dims_123))[name = string("concat_195")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_11_stride_0 = const()[name = string("key_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_195, begin_mask = key_cache_internal_tensor_assign_11_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_11_squeeze_mask_0, stride = key_cache_internal_tensor_assign_11_stride_0, update = k_state_21_cast_fp16, x = coreml_update_state_74)[name = string("key_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_11_cast_fp16, input = key_cache)[name = string("coreml_update_state_76_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_76 = read_state(input = key_cache)[name = string("coreml_update_state_76")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_11_stride_0 = const()[name = string("value_cache_internal_tensor_assign_11_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_11_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_11_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_11_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_11_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_21_cast_fp16 = transpose(perm = v_state_21_perm_0, x = var_2049_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_11_cast_fp16 = slice_update(begin = concat_195, begin_mask = value_cache_internal_tensor_assign_11_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_11_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_11_squeeze_mask_0, stride = value_cache_internal_tensor_assign_11_stride_0, update = v_state_21_cast_fp16, x = coreml_update_state_75)[name = string("value_cache_internal_tensor_assign_11_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_11_cast_fp16, input = value_cache)[name = string("coreml_update_state_77_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_77 = read_state(input = value_cache)[name = string("coreml_update_state_77")];
            tensor<int32, [5]> var_2106_begin_0 = const()[name = string("op_2106_begin_0"), val = tensor<int32, [5]>([10, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2106_end_0 = const()[name = string("op_2106_end_0"), val = tensor<int32, [5]>([11, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_2106_end_mask_0 = const()[name = string("op_2106_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2106_squeeze_mask_0 = const()[name = string("op_2106_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_2106_cast_fp16 = slice_by_index(begin = var_2106_begin_0, end = var_2106_end_0, end_mask = var_2106_end_mask_0, squeeze_mask = var_2106_squeeze_mask_0, x = coreml_update_state_76)[name = string("op_2106_cast_fp16")];
            tensor<int32, [4]> var_2109_begin_0 = const()[name = string("op_2109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2109_end_mask_0 = const()[name = string("op_2109_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_2109_cast_fp16 = slice_by_index(begin = var_2109_begin_0, end = concat_11, end_mask = var_2109_end_mask_0, x = var_2106_cast_fp16)[name = string("op_2109_cast_fp16")];
            tensor<int32, [5]> var_2111_begin_0 = const()[name = string("op_2111_begin_0"), val = tensor<int32, [5]>([10, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2111_end_0 = const()[name = string("op_2111_end_0"), val = tensor<int32, [5]>([11, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_2111_end_mask_0 = const()[name = string("op_2111_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2111_squeeze_mask_0 = const()[name = string("op_2111_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_2111_cast_fp16 = slice_by_index(begin = var_2111_begin_0, end = var_2111_end_0, end_mask = var_2111_end_mask_0, squeeze_mask = var_2111_squeeze_mask_0, x = coreml_update_state_77)[name = string("op_2111_cast_fp16")];
            tensor<int32, [4]> var_2114_begin_0 = const()[name = string("op_2114_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2114_end_mask_0 = const()[name = string("op_2114_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_2114_cast_fp16 = slice_by_index(begin = var_2114_begin_0, end = concat_11, end_mask = var_2114_end_mask_0, x = var_2111_cast_fp16)[name = string("op_2114_cast_fp16")];
            tensor<int32, [4]> var_2116_shape_cast_fp16 = shape(x = var_2109_cast_fp16)[name = string("op_2116_shape_cast_fp16")];
            int32 gather_193 = const()[name = string("gather_193"), val = int32(1)];
            int32 gather_194 = const()[name = string("gather_194"), val = int32(2)];
            int32 gather_195_axis_0 = const()[name = string("gather_195_axis_0"), val = int32(0)];
            int32 gather_195_batch_dims_0 = const()[name = string("gather_195_batch_dims_0"), val = int32(0)];
            bool gather_195_validate_indices_0 = const()[name = string("gather_195_validate_indices_0"), val = bool(false)];
            string var_2116_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2116_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_195_to_uint16 = const()[name = string("select_195_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2116_shape_cast_fp16_to_uint16 = cast(dtype = var_2116_shape_cast_fp16_to_uint16_dtype_0, x = var_2116_shape_cast_fp16)[name = string("cast_107")];
            uint16 gather_195_cast_uint16 = gather(axis = gather_195_axis_0, batch_dims = gather_195_batch_dims_0, indices = select_195_to_uint16, validate_indices = gather_195_validate_indices_0, x = var_2116_shape_cast_fp16_to_uint16)[name = string("gather_195_cast_uint16")];
            string gather_195_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_195_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_196 = const()[name = string("gather_196"), val = int32(128)];
            tensor<int32, [1]> var_2123_axes_0 = const()[name = string("op_2123_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_2123_cast_fp16 = expand_dims(axes = var_2123_axes_0, x = var_2109_cast_fp16)[name = string("op_2123_cast_fp16")];
            tensor<int32, [5]> shape_217_cast_fp16 = shape(x = var_2123_cast_fp16)[name = string("shape_217_cast_fp16")];
            int32 concat_203_axis_0 = const()[name = string("concat_203_axis_0"), val = int32(0)];
            bool concat_203_interleave_0 = const()[name = string("concat_203_interleave_0"), val = bool(false)];
            int32 gather_195_cast_uint16_to_int32 = cast(dtype = gather_195_cast_uint16_to_int32_dtype_0, x = gather_195_cast_uint16)[name = string("cast_106")];
            tensor<int32, [5]> concat_203 = concat(axis = concat_203_axis_0, interleave = concat_203_interleave_0, values = (gather_193, gather_194, var_77, gather_195_cast_uint16_to_int32, gather_196))[name = string("concat_203")];
            tensor<int32, [5]> real_div_20 = real_div(x = concat_203, y = shape_217_cast_fp16)[name = string("real_div_20")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_311_cast_fp16 = tile(reps = real_div_20, x = var_2123_cast_fp16)[name = string("hidden_states_311_cast_fp16")];
            tensor<int32, [4]> concat_204x = const()[name = string("concat_204x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_43_cast_fp16 = reshape(shape = concat_204x, x = hidden_states_311_cast_fp16)[name = string("key_states_43_cast_fp16")];
            tensor<int32, [4]> var_2133_shape_cast_fp16 = shape(x = var_2114_cast_fp16)[name = string("op_2133_shape_cast_fp16")];
            int32 gather_197 = const()[name = string("gather_197"), val = int32(1)];
            int32 gather_198 = const()[name = string("gather_198"), val = int32(2)];
            int32 gather_199_axis_0 = const()[name = string("gather_199_axis_0"), val = int32(0)];
            int32 gather_199_batch_dims_0 = const()[name = string("gather_199_batch_dims_0"), val = int32(0)];
            bool gather_199_validate_indices_0 = const()[name = string("gather_199_validate_indices_0"), val = bool(false)];
            string var_2133_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2133_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_199_to_uint16 = const()[name = string("select_199_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2133_shape_cast_fp16_to_uint16 = cast(dtype = var_2133_shape_cast_fp16_to_uint16_dtype_0, x = var_2133_shape_cast_fp16)[name = string("cast_105")];
            uint16 gather_199_cast_uint16 = gather(axis = gather_199_axis_0, batch_dims = gather_199_batch_dims_0, indices = select_199_to_uint16, validate_indices = gather_199_validate_indices_0, x = var_2133_shape_cast_fp16_to_uint16)[name = string("gather_199_cast_uint16")];
            string gather_199_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_199_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_200 = const()[name = string("gather_200"), val = int32(128)];
            tensor<int32, [1]> var_2140_axes_0 = const()[name = string("op_2140_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_2140_cast_fp16 = expand_dims(axes = var_2140_axes_0, x = var_2114_cast_fp16)[name = string("op_2140_cast_fp16")];
            tensor<int32, [5]> shape_222_cast_fp16 = shape(x = var_2140_cast_fp16)[name = string("shape_222_cast_fp16")];
            int32 concat_205_axis_0 = const()[name = string("concat_205_axis_0"), val = int32(0)];
            bool concat_205_interleave_0 = const()[name = string("concat_205_interleave_0"), val = bool(false)];
            int32 gather_199_cast_uint16_to_int32 = cast(dtype = gather_199_cast_uint16_to_int32_dtype_0, x = gather_199_cast_uint16)[name = string("cast_104")];
            tensor<int32, [5]> concat_205 = concat(axis = concat_205_axis_0, interleave = concat_205_interleave_0, values = (gather_197, gather_198, var_77, gather_199_cast_uint16_to_int32, gather_200))[name = string("concat_205")];
            tensor<int32, [5]> real_div_21 = real_div(x = concat_205, y = shape_222_cast_fp16)[name = string("real_div_21")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_315_cast_fp16 = tile(reps = real_div_21, x = var_2140_cast_fp16)[name = string("hidden_states_315_cast_fp16")];
            tensor<int32, [4]> concat_206x = const()[name = string("concat_206x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_43_cast_fp16 = reshape(shape = concat_206x, x = hidden_states_315_cast_fp16)[name = string("value_states_43_cast_fp16")];
            tensor<int32, [4]> var_2150_shape_cast_fp16 = shape(x = key_states_43_cast_fp16)[name = string("op_2150_shape_cast_fp16")];
            int32 gather_201_axis_0 = const()[name = string("gather_201_axis_0"), val = int32(0)];
            int32 gather_201_batch_dims_0 = const()[name = string("gather_201_batch_dims_0"), val = int32(0)];
            bool gather_201_validate_indices_0 = const()[name = string("gather_201_validate_indices_0"), val = bool(false)];
            string var_2150_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2150_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_201_to_uint16 = const()[name = string("select_201_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2150_shape_cast_fp16_to_uint16 = cast(dtype = var_2150_shape_cast_fp16_to_uint16_dtype_0, x = var_2150_shape_cast_fp16)[name = string("cast_103")];
            uint16 gather_201_cast_uint16 = gather(axis = gather_201_axis_0, batch_dims = gather_201_batch_dims_0, indices = select_201_to_uint16, validate_indices = gather_201_validate_indices_0, x = var_2150_shape_cast_fp16_to_uint16)[name = string("gather_201_cast_uint16")];
            string gather_201_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_201_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_207_values0_0 = const()[name = string("concat_207_values0_0"), val = int32(1)];
            int32 concat_207_values1_0 = const()[name = string("concat_207_values1_0"), val = int32(1)];
            int32 concat_207_values2_0 = const()[name = string("concat_207_values2_0"), val = int32(0)];
            int32 concat_207_axis_0 = const()[name = string("concat_207_axis_0"), val = int32(0)];
            bool concat_207_interleave_0 = const()[name = string("concat_207_interleave_0"), val = bool(false)];
            int32 gather_201_cast_uint16_to_int32 = cast(dtype = gather_201_cast_uint16_to_int32_dtype_0, x = gather_201_cast_uint16)[name = string("cast_102")];
            tensor<int32, [4]> concat_207 = concat(axis = concat_207_axis_0, interleave = concat_207_interleave_0, values = (concat_207_values0_0, concat_207_values1_0, concat_207_values2_0, gather_201_cast_uint16_to_int32))[name = string("concat_207")];
            tensor<int32, [4]> causal_mask_23_begin_0 = const()[name = string("causal_mask_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_23_end_mask_0 = const()[name = string("causal_mask_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_23_cast_fp16 = slice_by_index(begin = causal_mask_23_begin_0, end = concat_207, end_mask = causal_mask_23_end_mask_0, x = causal_mask)[name = string("causal_mask_23_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_41_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_23_cast_fp16, key = key_states_43_cast_fp16, query = query_states_43_cast_fp16, value = value_states_43_cast_fp16)[name = string("attn_output_41_cast_fp16")];
            tensor<int32, [4]> var_2156_perm_0 = const()[name = string("op_2156_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_208x = const()[name = string("concat_208x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_2156_cast_fp16 = transpose(perm = var_2156_perm_0, x = attn_output_41_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, ?, 1536]> input_81_cast_fp16 = reshape(shape = concat_208x, x = var_2156_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396395008))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397574720))))[name = string("model_model_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_73_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_319_cast_fp16 = add(x = hidden_states_299_cast_fp16, y = linear_73_cast_fp16)[name = string("hidden_states_319_cast_fp16")];
            fp16 var_78_promoted_21_to_fp16 = const()[name = string("op_78_promoted_21_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_2165_cast_fp16 = pow(x = hidden_states_319_cast_fp16, y = var_78_promoted_21_to_fp16)[name = string("op_2165_cast_fp16")];
            tensor<int32, [1]> variance_43_axes_0 = const()[name = string("variance_43_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_43_keep_dims_0 = const()[name = string("variance_43_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_43_cast_fp16 = reduce_mean(axes = variance_43_axes_0, keep_dims = variance_43_keep_dims_0, x = var_2165_cast_fp16)[name = string("variance_43_cast_fp16")];
            fp16 var_2168_to_fp16 = const()[name = string("op_2168_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2169_cast_fp16 = add(x = variance_43_cast_fp16, y = var_2168_to_fp16)[name = string("op_2169_cast_fp16")];
            fp32 var_2170_epsilon_0 = const()[name = string("op_2170_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2170_cast_fp16 = rsqrt(epsilon = var_2170_epsilon_0, x = var_2169_cast_fp16)[name = string("op_2170_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_323_cast_fp16 = mul(x = hidden_states_319_cast_fp16, y = var_2170_cast_fp16)[name = string("hidden_states_323_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397722240)))];
            tensor<fp16, [1, ?, 1536]> input_83_cast_fp16 = mul(x = model_model_layers_10_post_attention_layernorm_weight_to_fp16, y = hidden_states_323_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397725376))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404606720))))[name = string("model_model_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_2182_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_2182_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405466944))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412348288))))[name = string("model_model_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_87_cast_fp16 = mul(x = var_2182_cast_fp16, y = linear_75_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413208512))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420089856))))[name = string("model_model_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_76_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_329_cast_fp16 = add(x = hidden_states_319_cast_fp16, y = linear_76_cast_fp16)[name = string("hidden_states_329_cast_fp16")];
            fp16 var_78_promoted_22_to_fp16 = const()[name = string("op_78_promoted_22_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_2195_cast_fp16 = pow(x = hidden_states_329_cast_fp16, y = var_78_promoted_22_to_fp16)[name = string("op_2195_cast_fp16")];
            tensor<int32, [1]> variance_45_axes_0 = const()[name = string("variance_45_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_45_keep_dims_0 = const()[name = string("variance_45_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_45_cast_fp16 = reduce_mean(axes = variance_45_axes_0, keep_dims = variance_45_keep_dims_0, x = var_2195_cast_fp16)[name = string("variance_45_cast_fp16")];
            fp16 var_2198_to_fp16 = const()[name = string("op_2198_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2199_cast_fp16 = add(x = variance_45_cast_fp16, y = var_2198_to_fp16)[name = string("op_2199_cast_fp16")];
            fp32 var_2200_epsilon_0 = const()[name = string("op_2200_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2200_cast_fp16 = rsqrt(epsilon = var_2200_epsilon_0, x = var_2199_cast_fp16)[name = string("op_2200_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_333_cast_fp16 = mul(x = hidden_states_329_cast_fp16, y = var_2200_cast_fp16)[name = string("hidden_states_333_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420950080)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_337_cast_fp16 = mul(x = model_model_layers_11_input_layernorm_weight_to_fp16, y = hidden_states_333_cast_fp16)[name = string("hidden_states_337_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420953216))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422132928))))[name = string("model_model_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422280448)))];
            tensor<fp16, [1, ?, 1536]> linear_77_cast_fp16 = linear(bias = model_model_layers_11_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_337_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422283584))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422480256))))[name = string("model_model_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_11_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_11_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422504896)))];
            tensor<fp16, [1, ?, 256]> linear_78_cast_fp16 = linear(bias = model_model_layers_11_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_337_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422505472))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422702144))))[name = string("model_model_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422726784)))];
            tensor<fp16, [1, ?, 256]> linear_79_cast_fp16 = linear(bias = model_model_layers_11_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_337_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> concat_209x = const()[name = string("concat_209x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_2223_cast_fp16 = reshape(shape = concat_209x, x = linear_77_cast_fp16)[name = string("op_2223_cast_fp16")];
            tensor<int32, [4]> q_23_perm_0 = const()[name = string("q_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_210x = const()[name = string("concat_210x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_2226_cast_fp16 = reshape(shape = concat_210x, x = linear_78_cast_fp16)[name = string("op_2226_cast_fp16")];
            tensor<int32, [4]> k_23_perm_0 = const()[name = string("k_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_211x = const()[name = string("concat_211x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_2229_cast_fp16 = reshape(shape = concat_211x, x = linear_79_cast_fp16)[name = string("op_2229_cast_fp16")];
            tensor<int32, [4]> v_state_23_perm_0 = const()[name = string("v_state_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_23_cast_fp16 = transpose(perm = q_23_perm_0, x = var_2223_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 12, ?, 128]> var_2233_cast_fp16 = mul(x = q_23_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2233_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_23_cast_fp16)[name = string("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_23_cast_fp16)[name = string("x2_45_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_2244_cast_fp16 = mul(x = x2_45_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_2244_cast_fp16")];
            bool var_2246_interleave_0 = const()[name = string("op_2246_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_2246_cast_fp16 = concat(axis = var_72, interleave = var_2246_interleave_0, values = (var_2244_cast_fp16, x1_45_cast_fp16))[name = string("op_2246_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_2247_cast_fp16 = mul(x = var_2246_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2247_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_47_cast_fp16 = add(x = var_2233_cast_fp16, y = var_2247_cast_fp16)[name = string("query_states_47_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_23_cast_fp16 = transpose(perm = k_23_perm_0, x = var_2226_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 2, ?, 128]> var_2249_cast_fp16 = mul(x = k_23_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2249_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_23_cast_fp16)[name = string("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_23_cast_fp16)[name = string("x2_47_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_2260_cast_fp16 = mul(x = x2_47_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_2260_cast_fp16")];
            bool var_2262_interleave_0 = const()[name = string("op_2262_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_2262_cast_fp16 = concat(axis = var_72, interleave = var_2262_interleave_0, values = (var_2260_cast_fp16, x1_47_cast_fp16))[name = string("op_2262_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_2263_cast_fp16 = mul(x = var_2262_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2263_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_23_cast_fp16 = add(x = var_2249_cast_fp16, y = var_2263_cast_fp16)[name = string("k_state_23_cast_fp16")];
            tensor<int32, [1]> expand_dims_132 = const()[name = string("expand_dims_132"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_133 = const()[name = string("expand_dims_133"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_135 = const()[name = string("expand_dims_135"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_214_values0_0 = const()[name = string("concat_214_values0_0"), val = tensor<int32, [1]>([11])];
            int32 concat_214_axis_0 = const()[name = string("concat_214_axis_0"), val = int32(0)];
            bool concat_214_interleave_0 = const()[name = string("concat_214_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_214 = concat(axis = concat_214_axis_0, interleave = concat_214_interleave_0, values = (concat_214_values0_0, expand_dims_132, expand_dims_133, expand_dims_2, expand_dims_135))[name = string("concat_214")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_12_stride_0 = const()[name = string("key_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_214, begin_mask = key_cache_internal_tensor_assign_12_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_12_squeeze_mask_0, stride = key_cache_internal_tensor_assign_12_stride_0, update = k_state_23_cast_fp16, x = coreml_update_state_76)[name = string("key_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_12_cast_fp16, input = key_cache)[name = string("coreml_update_state_78_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_78 = read_state(input = key_cache)[name = string("coreml_update_state_78")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_12_stride_0 = const()[name = string("value_cache_internal_tensor_assign_12_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_12_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_12_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_12_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_12_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_23_cast_fp16 = transpose(perm = v_state_23_perm_0, x = var_2229_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_12_cast_fp16 = slice_update(begin = concat_214, begin_mask = value_cache_internal_tensor_assign_12_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_12_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_12_squeeze_mask_0, stride = value_cache_internal_tensor_assign_12_stride_0, update = v_state_23_cast_fp16, x = coreml_update_state_77)[name = string("value_cache_internal_tensor_assign_12_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_12_cast_fp16, input = value_cache)[name = string("coreml_update_state_79_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_79 = read_state(input = value_cache)[name = string("coreml_update_state_79")];
            tensor<int32, [5]> var_2286_begin_0 = const()[name = string("op_2286_begin_0"), val = tensor<int32, [5]>([11, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2286_end_0 = const()[name = string("op_2286_end_0"), val = tensor<int32, [5]>([12, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_2286_end_mask_0 = const()[name = string("op_2286_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2286_squeeze_mask_0 = const()[name = string("op_2286_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_2286_cast_fp16 = slice_by_index(begin = var_2286_begin_0, end = var_2286_end_0, end_mask = var_2286_end_mask_0, squeeze_mask = var_2286_squeeze_mask_0, x = coreml_update_state_78)[name = string("op_2286_cast_fp16")];
            tensor<int32, [4]> var_2289_begin_0 = const()[name = string("op_2289_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2289_end_mask_0 = const()[name = string("op_2289_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_2289_cast_fp16 = slice_by_index(begin = var_2289_begin_0, end = concat_11, end_mask = var_2289_end_mask_0, x = var_2286_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [5]> var_2291_begin_0 = const()[name = string("op_2291_begin_0"), val = tensor<int32, [5]>([11, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2291_end_0 = const()[name = string("op_2291_end_0"), val = tensor<int32, [5]>([12, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_2291_end_mask_0 = const()[name = string("op_2291_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2291_squeeze_mask_0 = const()[name = string("op_2291_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_2291_cast_fp16 = slice_by_index(begin = var_2291_begin_0, end = var_2291_end_0, end_mask = var_2291_end_mask_0, squeeze_mask = var_2291_squeeze_mask_0, x = coreml_update_state_79)[name = string("op_2291_cast_fp16")];
            tensor<int32, [4]> var_2294_begin_0 = const()[name = string("op_2294_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2294_end_mask_0 = const()[name = string("op_2294_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_2294_cast_fp16 = slice_by_index(begin = var_2294_begin_0, end = concat_11, end_mask = var_2294_end_mask_0, x = var_2291_cast_fp16)[name = string("op_2294_cast_fp16")];
            tensor<int32, [4]> var_2296_shape_cast_fp16 = shape(x = var_2289_cast_fp16)[name = string("op_2296_shape_cast_fp16")];
            int32 gather_211 = const()[name = string("gather_211"), val = int32(1)];
            int32 gather_212 = const()[name = string("gather_212"), val = int32(2)];
            int32 gather_213_axis_0 = const()[name = string("gather_213_axis_0"), val = int32(0)];
            int32 gather_213_batch_dims_0 = const()[name = string("gather_213_batch_dims_0"), val = int32(0)];
            bool gather_213_validate_indices_0 = const()[name = string("gather_213_validate_indices_0"), val = bool(false)];
            string var_2296_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2296_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_213_to_uint16 = const()[name = string("select_213_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2296_shape_cast_fp16_to_uint16 = cast(dtype = var_2296_shape_cast_fp16_to_uint16_dtype_0, x = var_2296_shape_cast_fp16)[name = string("cast_101")];
            uint16 gather_213_cast_uint16 = gather(axis = gather_213_axis_0, batch_dims = gather_213_batch_dims_0, indices = select_213_to_uint16, validate_indices = gather_213_validate_indices_0, x = var_2296_shape_cast_fp16_to_uint16)[name = string("gather_213_cast_uint16")];
            string gather_213_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_213_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_214 = const()[name = string("gather_214"), val = int32(128)];
            tensor<int32, [1]> var_2303_axes_0 = const()[name = string("op_2303_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_2303_cast_fp16 = expand_dims(axes = var_2303_axes_0, x = var_2289_cast_fp16)[name = string("op_2303_cast_fp16")];
            tensor<int32, [5]> shape_237_cast_fp16 = shape(x = var_2303_cast_fp16)[name = string("shape_237_cast_fp16")];
            int32 concat_222_axis_0 = const()[name = string("concat_222_axis_0"), val = int32(0)];
            bool concat_222_interleave_0 = const()[name = string("concat_222_interleave_0"), val = bool(false)];
            int32 gather_213_cast_uint16_to_int32 = cast(dtype = gather_213_cast_uint16_to_int32_dtype_0, x = gather_213_cast_uint16)[name = string("cast_100")];
            tensor<int32, [5]> concat_222 = concat(axis = concat_222_axis_0, interleave = concat_222_interleave_0, values = (gather_211, gather_212, var_77, gather_213_cast_uint16_to_int32, gather_214))[name = string("concat_222")];
            tensor<int32, [5]> real_div_22 = real_div(x = concat_222, y = shape_237_cast_fp16)[name = string("real_div_22")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_341_cast_fp16 = tile(reps = real_div_22, x = var_2303_cast_fp16)[name = string("hidden_states_341_cast_fp16")];
            tensor<int32, [4]> concat_223x = const()[name = string("concat_223x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_47_cast_fp16 = reshape(shape = concat_223x, x = hidden_states_341_cast_fp16)[name = string("key_states_47_cast_fp16")];
            tensor<int32, [4]> var_2313_shape_cast_fp16 = shape(x = var_2294_cast_fp16)[name = string("op_2313_shape_cast_fp16")];
            int32 gather_215 = const()[name = string("gather_215"), val = int32(1)];
            int32 gather_216 = const()[name = string("gather_216"), val = int32(2)];
            int32 gather_217_axis_0 = const()[name = string("gather_217_axis_0"), val = int32(0)];
            int32 gather_217_batch_dims_0 = const()[name = string("gather_217_batch_dims_0"), val = int32(0)];
            bool gather_217_validate_indices_0 = const()[name = string("gather_217_validate_indices_0"), val = bool(false)];
            string var_2313_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2313_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_217_to_uint16 = const()[name = string("select_217_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2313_shape_cast_fp16_to_uint16 = cast(dtype = var_2313_shape_cast_fp16_to_uint16_dtype_0, x = var_2313_shape_cast_fp16)[name = string("cast_99")];
            uint16 gather_217_cast_uint16 = gather(axis = gather_217_axis_0, batch_dims = gather_217_batch_dims_0, indices = select_217_to_uint16, validate_indices = gather_217_validate_indices_0, x = var_2313_shape_cast_fp16_to_uint16)[name = string("gather_217_cast_uint16")];
            string gather_217_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_217_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_218 = const()[name = string("gather_218"), val = int32(128)];
            tensor<int32, [1]> var_2320_axes_0 = const()[name = string("op_2320_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_2320_cast_fp16 = expand_dims(axes = var_2320_axes_0, x = var_2294_cast_fp16)[name = string("op_2320_cast_fp16")];
            tensor<int32, [5]> shape_242_cast_fp16 = shape(x = var_2320_cast_fp16)[name = string("shape_242_cast_fp16")];
            int32 concat_224_axis_0 = const()[name = string("concat_224_axis_0"), val = int32(0)];
            bool concat_224_interleave_0 = const()[name = string("concat_224_interleave_0"), val = bool(false)];
            int32 gather_217_cast_uint16_to_int32 = cast(dtype = gather_217_cast_uint16_to_int32_dtype_0, x = gather_217_cast_uint16)[name = string("cast_98")];
            tensor<int32, [5]> concat_224 = concat(axis = concat_224_axis_0, interleave = concat_224_interleave_0, values = (gather_215, gather_216, var_77, gather_217_cast_uint16_to_int32, gather_218))[name = string("concat_224")];
            tensor<int32, [5]> real_div_23 = real_div(x = concat_224, y = shape_242_cast_fp16)[name = string("real_div_23")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_345_cast_fp16 = tile(reps = real_div_23, x = var_2320_cast_fp16)[name = string("hidden_states_345_cast_fp16")];
            tensor<int32, [4]> concat_225x = const()[name = string("concat_225x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_47_cast_fp16 = reshape(shape = concat_225x, x = hidden_states_345_cast_fp16)[name = string("value_states_47_cast_fp16")];
            tensor<int32, [4]> var_2330_shape_cast_fp16 = shape(x = key_states_47_cast_fp16)[name = string("op_2330_shape_cast_fp16")];
            int32 gather_219_axis_0 = const()[name = string("gather_219_axis_0"), val = int32(0)];
            int32 gather_219_batch_dims_0 = const()[name = string("gather_219_batch_dims_0"), val = int32(0)];
            bool gather_219_validate_indices_0 = const()[name = string("gather_219_validate_indices_0"), val = bool(false)];
            string var_2330_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2330_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_219_to_uint16 = const()[name = string("select_219_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2330_shape_cast_fp16_to_uint16 = cast(dtype = var_2330_shape_cast_fp16_to_uint16_dtype_0, x = var_2330_shape_cast_fp16)[name = string("cast_97")];
            uint16 gather_219_cast_uint16 = gather(axis = gather_219_axis_0, batch_dims = gather_219_batch_dims_0, indices = select_219_to_uint16, validate_indices = gather_219_validate_indices_0, x = var_2330_shape_cast_fp16_to_uint16)[name = string("gather_219_cast_uint16")];
            string gather_219_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_219_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_226_values0_0 = const()[name = string("concat_226_values0_0"), val = int32(1)];
            int32 concat_226_values1_0 = const()[name = string("concat_226_values1_0"), val = int32(1)];
            int32 concat_226_values2_0 = const()[name = string("concat_226_values2_0"), val = int32(0)];
            int32 concat_226_axis_0 = const()[name = string("concat_226_axis_0"), val = int32(0)];
            bool concat_226_interleave_0 = const()[name = string("concat_226_interleave_0"), val = bool(false)];
            int32 gather_219_cast_uint16_to_int32 = cast(dtype = gather_219_cast_uint16_to_int32_dtype_0, x = gather_219_cast_uint16)[name = string("cast_96")];
            tensor<int32, [4]> concat_226 = concat(axis = concat_226_axis_0, interleave = concat_226_interleave_0, values = (concat_226_values0_0, concat_226_values1_0, concat_226_values2_0, gather_219_cast_uint16_to_int32))[name = string("concat_226")];
            tensor<int32, [4]> causal_mask_25_begin_0 = const()[name = string("causal_mask_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_25_end_mask_0 = const()[name = string("causal_mask_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_25_cast_fp16 = slice_by_index(begin = causal_mask_25_begin_0, end = concat_226, end_mask = causal_mask_25_end_mask_0, x = causal_mask)[name = string("causal_mask_25_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_45_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_25_cast_fp16, key = key_states_47_cast_fp16, query = query_states_47_cast_fp16, value = value_states_47_cast_fp16)[name = string("attn_output_45_cast_fp16")];
            tensor<int32, [4]> var_2336_perm_0 = const()[name = string("op_2336_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_227x = const()[name = string("concat_227x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_2336_cast_fp16 = transpose(perm = var_2336_perm_0, x = attn_output_45_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, ?, 1536]> input_89_cast_fp16 = reshape(shape = concat_227x, x = var_2336_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422727360))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423907072))))[name = string("model_model_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_80_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_349_cast_fp16 = add(x = hidden_states_329_cast_fp16, y = linear_80_cast_fp16)[name = string("hidden_states_349_cast_fp16")];
            fp16 var_78_promoted_23_to_fp16 = const()[name = string("op_78_promoted_23_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_2345_cast_fp16 = pow(x = hidden_states_349_cast_fp16, y = var_78_promoted_23_to_fp16)[name = string("op_2345_cast_fp16")];
            tensor<int32, [1]> variance_47_axes_0 = const()[name = string("variance_47_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_47_keep_dims_0 = const()[name = string("variance_47_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_47_cast_fp16 = reduce_mean(axes = variance_47_axes_0, keep_dims = variance_47_keep_dims_0, x = var_2345_cast_fp16)[name = string("variance_47_cast_fp16")];
            fp16 var_2348_to_fp16 = const()[name = string("op_2348_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2349_cast_fp16 = add(x = variance_47_cast_fp16, y = var_2348_to_fp16)[name = string("op_2349_cast_fp16")];
            fp32 var_2350_epsilon_0 = const()[name = string("op_2350_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2350_cast_fp16 = rsqrt(epsilon = var_2350_epsilon_0, x = var_2349_cast_fp16)[name = string("op_2350_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_353_cast_fp16 = mul(x = hidden_states_349_cast_fp16, y = var_2350_cast_fp16)[name = string("hidden_states_353_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424054592)))];
            tensor<fp16, [1, ?, 1536]> input_91_cast_fp16 = mul(x = model_model_layers_11_post_attention_layernorm_weight_to_fp16, y = hidden_states_353_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424057728))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430939072))))[name = string("model_model_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_2362_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_2362_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431799296))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(438680640))))[name = string("model_model_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_95_cast_fp16 = mul(x = var_2362_cast_fp16, y = linear_82_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439540864))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446422208))))[name = string("model_model_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_359_cast_fp16 = add(x = hidden_states_349_cast_fp16, y = linear_83_cast_fp16)[name = string("hidden_states_359_cast_fp16")];
            fp16 var_78_promoted_24_to_fp16 = const()[name = string("op_78_promoted_24_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_2375_cast_fp16 = pow(x = hidden_states_359_cast_fp16, y = var_78_promoted_24_to_fp16)[name = string("op_2375_cast_fp16")];
            tensor<int32, [1]> variance_49_axes_0 = const()[name = string("variance_49_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_49_keep_dims_0 = const()[name = string("variance_49_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_49_cast_fp16 = reduce_mean(axes = variance_49_axes_0, keep_dims = variance_49_keep_dims_0, x = var_2375_cast_fp16)[name = string("variance_49_cast_fp16")];
            fp16 var_2378_to_fp16 = const()[name = string("op_2378_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2379_cast_fp16 = add(x = variance_49_cast_fp16, y = var_2378_to_fp16)[name = string("op_2379_cast_fp16")];
            fp32 var_2380_epsilon_0 = const()[name = string("op_2380_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2380_cast_fp16 = rsqrt(epsilon = var_2380_epsilon_0, x = var_2379_cast_fp16)[name = string("op_2380_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_363_cast_fp16 = mul(x = hidden_states_359_cast_fp16, y = var_2380_cast_fp16)[name = string("hidden_states_363_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447282432)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_367_cast_fp16 = mul(x = model_model_layers_12_input_layernorm_weight_to_fp16, y = hidden_states_363_cast_fp16)[name = string("hidden_states_367_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447285568))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448465280))))[name = string("model_model_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448612800)))];
            tensor<fp16, [1, ?, 1536]> linear_84_cast_fp16 = linear(bias = model_model_layers_12_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_367_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448615936))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448812608))))[name = string("model_model_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_12_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_12_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448837248)))];
            tensor<fp16, [1, ?, 256]> linear_85_cast_fp16 = linear(bias = model_model_layers_12_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_367_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448837824))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449034496))))[name = string("model_model_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449059136)))];
            tensor<fp16, [1, ?, 256]> linear_86_cast_fp16 = linear(bias = model_model_layers_12_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_367_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> concat_228x = const()[name = string("concat_228x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_2403_cast_fp16 = reshape(shape = concat_228x, x = linear_84_cast_fp16)[name = string("op_2403_cast_fp16")];
            tensor<int32, [4]> q_25_perm_0 = const()[name = string("q_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_229x = const()[name = string("concat_229x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_2406_cast_fp16 = reshape(shape = concat_229x, x = linear_85_cast_fp16)[name = string("op_2406_cast_fp16")];
            tensor<int32, [4]> k_25_perm_0 = const()[name = string("k_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_230x = const()[name = string("concat_230x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_2409_cast_fp16 = reshape(shape = concat_230x, x = linear_86_cast_fp16)[name = string("op_2409_cast_fp16")];
            tensor<int32, [4]> v_state_25_perm_0 = const()[name = string("v_state_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_25_cast_fp16 = transpose(perm = q_25_perm_0, x = var_2403_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 12, ?, 128]> var_2413_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2413_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_25_cast_fp16)[name = string("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_25_cast_fp16)[name = string("x2_49_cast_fp16")];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_2424_cast_fp16 = mul(x = x2_49_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_2424_cast_fp16")];
            bool var_2426_interleave_0 = const()[name = string("op_2426_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_2426_cast_fp16 = concat(axis = var_72, interleave = var_2426_interleave_0, values = (var_2424_cast_fp16, x1_49_cast_fp16))[name = string("op_2426_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_2427_cast_fp16 = mul(x = var_2426_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2427_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_51_cast_fp16 = add(x = var_2413_cast_fp16, y = var_2427_cast_fp16)[name = string("query_states_51_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = var_2406_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 2, ?, 128]> var_2429_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2429_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_25_cast_fp16)[name = string("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_25_cast_fp16)[name = string("x2_51_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_2440_cast_fp16 = mul(x = x2_51_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_2440_cast_fp16")];
            bool var_2442_interleave_0 = const()[name = string("op_2442_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_2442_cast_fp16 = concat(axis = var_72, interleave = var_2442_interleave_0, values = (var_2440_cast_fp16, x1_51_cast_fp16))[name = string("op_2442_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_2443_cast_fp16 = mul(x = var_2442_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2443_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_25_cast_fp16 = add(x = var_2429_cast_fp16, y = var_2443_cast_fp16)[name = string("k_state_25_cast_fp16")];
            tensor<int32, [1]> expand_dims_144 = const()[name = string("expand_dims_144"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_145 = const()[name = string("expand_dims_145"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_147 = const()[name = string("expand_dims_147"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_233_values0_0 = const()[name = string("concat_233_values0_0"), val = tensor<int32, [1]>([12])];
            int32 concat_233_axis_0 = const()[name = string("concat_233_axis_0"), val = int32(0)];
            bool concat_233_interleave_0 = const()[name = string("concat_233_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_233 = concat(axis = concat_233_axis_0, interleave = concat_233_interleave_0, values = (concat_233_values0_0, expand_dims_144, expand_dims_145, expand_dims_2, expand_dims_147))[name = string("concat_233")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_13_stride_0 = const()[name = string("key_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_233, begin_mask = key_cache_internal_tensor_assign_13_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_13_squeeze_mask_0, stride = key_cache_internal_tensor_assign_13_stride_0, update = k_state_25_cast_fp16, x = coreml_update_state_78)[name = string("key_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_13_cast_fp16, input = key_cache)[name = string("coreml_update_state_80_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_80 = read_state(input = key_cache)[name = string("coreml_update_state_80")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_13_stride_0 = const()[name = string("value_cache_internal_tensor_assign_13_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_13_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_13_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_13_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_13_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_25_cast_fp16 = transpose(perm = v_state_25_perm_0, x = var_2409_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_13_cast_fp16 = slice_update(begin = concat_233, begin_mask = value_cache_internal_tensor_assign_13_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_13_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_13_squeeze_mask_0, stride = value_cache_internal_tensor_assign_13_stride_0, update = v_state_25_cast_fp16, x = coreml_update_state_79)[name = string("value_cache_internal_tensor_assign_13_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_13_cast_fp16, input = value_cache)[name = string("coreml_update_state_81_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_81 = read_state(input = value_cache)[name = string("coreml_update_state_81")];
            tensor<int32, [5]> var_2466_begin_0 = const()[name = string("op_2466_begin_0"), val = tensor<int32, [5]>([12, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2466_end_0 = const()[name = string("op_2466_end_0"), val = tensor<int32, [5]>([13, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_2466_end_mask_0 = const()[name = string("op_2466_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2466_squeeze_mask_0 = const()[name = string("op_2466_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_2466_cast_fp16 = slice_by_index(begin = var_2466_begin_0, end = var_2466_end_0, end_mask = var_2466_end_mask_0, squeeze_mask = var_2466_squeeze_mask_0, x = coreml_update_state_80)[name = string("op_2466_cast_fp16")];
            tensor<int32, [4]> var_2469_begin_0 = const()[name = string("op_2469_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2469_end_mask_0 = const()[name = string("op_2469_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_2469_cast_fp16 = slice_by_index(begin = var_2469_begin_0, end = concat_11, end_mask = var_2469_end_mask_0, x = var_2466_cast_fp16)[name = string("op_2469_cast_fp16")];
            tensor<int32, [5]> var_2471_begin_0 = const()[name = string("op_2471_begin_0"), val = tensor<int32, [5]>([12, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2471_end_0 = const()[name = string("op_2471_end_0"), val = tensor<int32, [5]>([13, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_2471_end_mask_0 = const()[name = string("op_2471_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2471_squeeze_mask_0 = const()[name = string("op_2471_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_2471_cast_fp16 = slice_by_index(begin = var_2471_begin_0, end = var_2471_end_0, end_mask = var_2471_end_mask_0, squeeze_mask = var_2471_squeeze_mask_0, x = coreml_update_state_81)[name = string("op_2471_cast_fp16")];
            tensor<int32, [4]> var_2474_begin_0 = const()[name = string("op_2474_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2474_end_mask_0 = const()[name = string("op_2474_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_2474_cast_fp16 = slice_by_index(begin = var_2474_begin_0, end = concat_11, end_mask = var_2474_end_mask_0, x = var_2471_cast_fp16)[name = string("op_2474_cast_fp16")];
            tensor<int32, [4]> var_2476_shape_cast_fp16 = shape(x = var_2469_cast_fp16)[name = string("op_2476_shape_cast_fp16")];
            int32 gather_229 = const()[name = string("gather_229"), val = int32(1)];
            int32 gather_230 = const()[name = string("gather_230"), val = int32(2)];
            int32 gather_231_axis_0 = const()[name = string("gather_231_axis_0"), val = int32(0)];
            int32 gather_231_batch_dims_0 = const()[name = string("gather_231_batch_dims_0"), val = int32(0)];
            bool gather_231_validate_indices_0 = const()[name = string("gather_231_validate_indices_0"), val = bool(false)];
            string var_2476_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2476_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_231_to_uint16 = const()[name = string("select_231_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2476_shape_cast_fp16_to_uint16 = cast(dtype = var_2476_shape_cast_fp16_to_uint16_dtype_0, x = var_2476_shape_cast_fp16)[name = string("cast_95")];
            uint16 gather_231_cast_uint16 = gather(axis = gather_231_axis_0, batch_dims = gather_231_batch_dims_0, indices = select_231_to_uint16, validate_indices = gather_231_validate_indices_0, x = var_2476_shape_cast_fp16_to_uint16)[name = string("gather_231_cast_uint16")];
            string gather_231_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_231_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_232 = const()[name = string("gather_232"), val = int32(128)];
            tensor<int32, [1]> var_2483_axes_0 = const()[name = string("op_2483_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_2483_cast_fp16 = expand_dims(axes = var_2483_axes_0, x = var_2469_cast_fp16)[name = string("op_2483_cast_fp16")];
            tensor<int32, [5]> shape_257_cast_fp16 = shape(x = var_2483_cast_fp16)[name = string("shape_257_cast_fp16")];
            int32 concat_241_axis_0 = const()[name = string("concat_241_axis_0"), val = int32(0)];
            bool concat_241_interleave_0 = const()[name = string("concat_241_interleave_0"), val = bool(false)];
            int32 gather_231_cast_uint16_to_int32 = cast(dtype = gather_231_cast_uint16_to_int32_dtype_0, x = gather_231_cast_uint16)[name = string("cast_94")];
            tensor<int32, [5]> concat_241 = concat(axis = concat_241_axis_0, interleave = concat_241_interleave_0, values = (gather_229, gather_230, var_77, gather_231_cast_uint16_to_int32, gather_232))[name = string("concat_241")];
            tensor<int32, [5]> real_div_24 = real_div(x = concat_241, y = shape_257_cast_fp16)[name = string("real_div_24")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_371_cast_fp16 = tile(reps = real_div_24, x = var_2483_cast_fp16)[name = string("hidden_states_371_cast_fp16")];
            tensor<int32, [4]> concat_242x = const()[name = string("concat_242x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_51_cast_fp16 = reshape(shape = concat_242x, x = hidden_states_371_cast_fp16)[name = string("key_states_51_cast_fp16")];
            tensor<int32, [4]> var_2493_shape_cast_fp16 = shape(x = var_2474_cast_fp16)[name = string("op_2493_shape_cast_fp16")];
            int32 gather_233 = const()[name = string("gather_233"), val = int32(1)];
            int32 gather_234 = const()[name = string("gather_234"), val = int32(2)];
            int32 gather_235_axis_0 = const()[name = string("gather_235_axis_0"), val = int32(0)];
            int32 gather_235_batch_dims_0 = const()[name = string("gather_235_batch_dims_0"), val = int32(0)];
            bool gather_235_validate_indices_0 = const()[name = string("gather_235_validate_indices_0"), val = bool(false)];
            string var_2493_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2493_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_235_to_uint16 = const()[name = string("select_235_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2493_shape_cast_fp16_to_uint16 = cast(dtype = var_2493_shape_cast_fp16_to_uint16_dtype_0, x = var_2493_shape_cast_fp16)[name = string("cast_93")];
            uint16 gather_235_cast_uint16 = gather(axis = gather_235_axis_0, batch_dims = gather_235_batch_dims_0, indices = select_235_to_uint16, validate_indices = gather_235_validate_indices_0, x = var_2493_shape_cast_fp16_to_uint16)[name = string("gather_235_cast_uint16")];
            string gather_235_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_235_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_236 = const()[name = string("gather_236"), val = int32(128)];
            tensor<int32, [1]> var_2500_axes_0 = const()[name = string("op_2500_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_2500_cast_fp16 = expand_dims(axes = var_2500_axes_0, x = var_2474_cast_fp16)[name = string("op_2500_cast_fp16")];
            tensor<int32, [5]> shape_262_cast_fp16 = shape(x = var_2500_cast_fp16)[name = string("shape_262_cast_fp16")];
            int32 concat_243_axis_0 = const()[name = string("concat_243_axis_0"), val = int32(0)];
            bool concat_243_interleave_0 = const()[name = string("concat_243_interleave_0"), val = bool(false)];
            int32 gather_235_cast_uint16_to_int32 = cast(dtype = gather_235_cast_uint16_to_int32_dtype_0, x = gather_235_cast_uint16)[name = string("cast_92")];
            tensor<int32, [5]> concat_243 = concat(axis = concat_243_axis_0, interleave = concat_243_interleave_0, values = (gather_233, gather_234, var_77, gather_235_cast_uint16_to_int32, gather_236))[name = string("concat_243")];
            tensor<int32, [5]> real_div_25 = real_div(x = concat_243, y = shape_262_cast_fp16)[name = string("real_div_25")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_375_cast_fp16 = tile(reps = real_div_25, x = var_2500_cast_fp16)[name = string("hidden_states_375_cast_fp16")];
            tensor<int32, [4]> concat_244x = const()[name = string("concat_244x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_51_cast_fp16 = reshape(shape = concat_244x, x = hidden_states_375_cast_fp16)[name = string("value_states_51_cast_fp16")];
            tensor<int32, [4]> var_2510_shape_cast_fp16 = shape(x = key_states_51_cast_fp16)[name = string("op_2510_shape_cast_fp16")];
            int32 gather_237_axis_0 = const()[name = string("gather_237_axis_0"), val = int32(0)];
            int32 gather_237_batch_dims_0 = const()[name = string("gather_237_batch_dims_0"), val = int32(0)];
            bool gather_237_validate_indices_0 = const()[name = string("gather_237_validate_indices_0"), val = bool(false)];
            string var_2510_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2510_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_237_to_uint16 = const()[name = string("select_237_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2510_shape_cast_fp16_to_uint16 = cast(dtype = var_2510_shape_cast_fp16_to_uint16_dtype_0, x = var_2510_shape_cast_fp16)[name = string("cast_91")];
            uint16 gather_237_cast_uint16 = gather(axis = gather_237_axis_0, batch_dims = gather_237_batch_dims_0, indices = select_237_to_uint16, validate_indices = gather_237_validate_indices_0, x = var_2510_shape_cast_fp16_to_uint16)[name = string("gather_237_cast_uint16")];
            string gather_237_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_237_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_245_values0_0 = const()[name = string("concat_245_values0_0"), val = int32(1)];
            int32 concat_245_values1_0 = const()[name = string("concat_245_values1_0"), val = int32(1)];
            int32 concat_245_values2_0 = const()[name = string("concat_245_values2_0"), val = int32(0)];
            int32 concat_245_axis_0 = const()[name = string("concat_245_axis_0"), val = int32(0)];
            bool concat_245_interleave_0 = const()[name = string("concat_245_interleave_0"), val = bool(false)];
            int32 gather_237_cast_uint16_to_int32 = cast(dtype = gather_237_cast_uint16_to_int32_dtype_0, x = gather_237_cast_uint16)[name = string("cast_90")];
            tensor<int32, [4]> concat_245 = concat(axis = concat_245_axis_0, interleave = concat_245_interleave_0, values = (concat_245_values0_0, concat_245_values1_0, concat_245_values2_0, gather_237_cast_uint16_to_int32))[name = string("concat_245")];
            tensor<int32, [4]> causal_mask_27_begin_0 = const()[name = string("causal_mask_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_27_end_mask_0 = const()[name = string("causal_mask_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_27_cast_fp16 = slice_by_index(begin = causal_mask_27_begin_0, end = concat_245, end_mask = causal_mask_27_end_mask_0, x = causal_mask)[name = string("causal_mask_27_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_49_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_27_cast_fp16, key = key_states_51_cast_fp16, query = query_states_51_cast_fp16, value = value_states_51_cast_fp16)[name = string("attn_output_49_cast_fp16")];
            tensor<int32, [4]> var_2516_perm_0 = const()[name = string("op_2516_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_246x = const()[name = string("concat_246x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_2516_cast_fp16 = transpose(perm = var_2516_perm_0, x = attn_output_49_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, ?, 1536]> input_97_cast_fp16 = reshape(shape = concat_246x, x = var_2516_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449059712))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450239424))))[name = string("model_model_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_379_cast_fp16 = add(x = hidden_states_359_cast_fp16, y = linear_87_cast_fp16)[name = string("hidden_states_379_cast_fp16")];
            fp16 var_78_promoted_25_to_fp16 = const()[name = string("op_78_promoted_25_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_2525_cast_fp16 = pow(x = hidden_states_379_cast_fp16, y = var_78_promoted_25_to_fp16)[name = string("op_2525_cast_fp16")];
            tensor<int32, [1]> variance_51_axes_0 = const()[name = string("variance_51_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_51_keep_dims_0 = const()[name = string("variance_51_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_51_cast_fp16 = reduce_mean(axes = variance_51_axes_0, keep_dims = variance_51_keep_dims_0, x = var_2525_cast_fp16)[name = string("variance_51_cast_fp16")];
            fp16 var_2528_to_fp16 = const()[name = string("op_2528_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2529_cast_fp16 = add(x = variance_51_cast_fp16, y = var_2528_to_fp16)[name = string("op_2529_cast_fp16")];
            fp32 var_2530_epsilon_0 = const()[name = string("op_2530_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2530_cast_fp16 = rsqrt(epsilon = var_2530_epsilon_0, x = var_2529_cast_fp16)[name = string("op_2530_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_383_cast_fp16 = mul(x = hidden_states_379_cast_fp16, y = var_2530_cast_fp16)[name = string("hidden_states_383_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450386944)))];
            tensor<fp16, [1, ?, 1536]> input_99_cast_fp16 = mul(x = model_model_layers_12_post_attention_layernorm_weight_to_fp16, y = hidden_states_383_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450390080))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457271424))))[name = string("model_model_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_2542_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_2542_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458131648))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465012992))))[name = string("model_model_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_103_cast_fp16 = mul(x = var_2542_cast_fp16, y = linear_89_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465873216))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(472754560))))[name = string("model_model_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_90_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_389_cast_fp16 = add(x = hidden_states_379_cast_fp16, y = linear_90_cast_fp16)[name = string("hidden_states_389_cast_fp16")];
            fp16 var_78_promoted_26_to_fp16 = const()[name = string("op_78_promoted_26_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_2555_cast_fp16 = pow(x = hidden_states_389_cast_fp16, y = var_78_promoted_26_to_fp16)[name = string("op_2555_cast_fp16")];
            tensor<int32, [1]> variance_53_axes_0 = const()[name = string("variance_53_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_53_keep_dims_0 = const()[name = string("variance_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_53_cast_fp16 = reduce_mean(axes = variance_53_axes_0, keep_dims = variance_53_keep_dims_0, x = var_2555_cast_fp16)[name = string("variance_53_cast_fp16")];
            fp16 var_2558_to_fp16 = const()[name = string("op_2558_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2559_cast_fp16 = add(x = variance_53_cast_fp16, y = var_2558_to_fp16)[name = string("op_2559_cast_fp16")];
            fp32 var_2560_epsilon_0 = const()[name = string("op_2560_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2560_cast_fp16 = rsqrt(epsilon = var_2560_epsilon_0, x = var_2559_cast_fp16)[name = string("op_2560_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_393_cast_fp16 = mul(x = hidden_states_389_cast_fp16, y = var_2560_cast_fp16)[name = string("hidden_states_393_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473614784)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_397_cast_fp16 = mul(x = model_model_layers_13_input_layernorm_weight_to_fp16, y = hidden_states_393_cast_fp16)[name = string("hidden_states_397_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473617920))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474797632))))[name = string("model_model_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474945152)))];
            tensor<fp16, [1, ?, 1536]> linear_91_cast_fp16 = linear(bias = model_model_layers_13_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_397_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474948288))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475144960))))[name = string("model_model_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_13_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_13_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475169600)))];
            tensor<fp16, [1, ?, 256]> linear_92_cast_fp16 = linear(bias = model_model_layers_13_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_397_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475170176))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475366848))))[name = string("model_model_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475391488)))];
            tensor<fp16, [1, ?, 256]> linear_93_cast_fp16 = linear(bias = model_model_layers_13_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_397_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> concat_247x = const()[name = string("concat_247x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_2583_cast_fp16 = reshape(shape = concat_247x, x = linear_91_cast_fp16)[name = string("op_2583_cast_fp16")];
            tensor<int32, [4]> q_27_perm_0 = const()[name = string("q_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_248x = const()[name = string("concat_248x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_2586_cast_fp16 = reshape(shape = concat_248x, x = linear_92_cast_fp16)[name = string("op_2586_cast_fp16")];
            tensor<int32, [4]> k_27_perm_0 = const()[name = string("k_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_249x = const()[name = string("concat_249x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_2589_cast_fp16 = reshape(shape = concat_249x, x = linear_93_cast_fp16)[name = string("op_2589_cast_fp16")];
            tensor<int32, [4]> v_state_27_perm_0 = const()[name = string("v_state_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_27_cast_fp16 = transpose(perm = q_27_perm_0, x = var_2583_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 12, ?, 128]> var_2593_cast_fp16 = mul(x = q_27_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2593_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_27_cast_fp16)[name = string("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_27_cast_fp16)[name = string("x2_53_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_2604_cast_fp16 = mul(x = x2_53_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_2604_cast_fp16")];
            bool var_2606_interleave_0 = const()[name = string("op_2606_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_2606_cast_fp16 = concat(axis = var_72, interleave = var_2606_interleave_0, values = (var_2604_cast_fp16, x1_53_cast_fp16))[name = string("op_2606_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_2607_cast_fp16 = mul(x = var_2606_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2607_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_55_cast_fp16 = add(x = var_2593_cast_fp16, y = var_2607_cast_fp16)[name = string("query_states_55_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_27_cast_fp16 = transpose(perm = k_27_perm_0, x = var_2586_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 2, ?, 128]> var_2609_cast_fp16 = mul(x = k_27_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2609_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_27_cast_fp16)[name = string("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_27_cast_fp16)[name = string("x2_55_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_2620_cast_fp16 = mul(x = x2_55_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_2620_cast_fp16")];
            bool var_2622_interleave_0 = const()[name = string("op_2622_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_2622_cast_fp16 = concat(axis = var_72, interleave = var_2622_interleave_0, values = (var_2620_cast_fp16, x1_55_cast_fp16))[name = string("op_2622_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_2623_cast_fp16 = mul(x = var_2622_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2623_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_27_cast_fp16 = add(x = var_2609_cast_fp16, y = var_2623_cast_fp16)[name = string("k_state_27_cast_fp16")];
            tensor<int32, [1]> expand_dims_156 = const()[name = string("expand_dims_156"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_157 = const()[name = string("expand_dims_157"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_159 = const()[name = string("expand_dims_159"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_252_values0_0 = const()[name = string("concat_252_values0_0"), val = tensor<int32, [1]>([13])];
            int32 concat_252_axis_0 = const()[name = string("concat_252_axis_0"), val = int32(0)];
            bool concat_252_interleave_0 = const()[name = string("concat_252_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_252 = concat(axis = concat_252_axis_0, interleave = concat_252_interleave_0, values = (concat_252_values0_0, expand_dims_156, expand_dims_157, expand_dims_2, expand_dims_159))[name = string("concat_252")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_14_stride_0 = const()[name = string("key_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_252, begin_mask = key_cache_internal_tensor_assign_14_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_14_squeeze_mask_0, stride = key_cache_internal_tensor_assign_14_stride_0, update = k_state_27_cast_fp16, x = coreml_update_state_80)[name = string("key_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_14_cast_fp16, input = key_cache)[name = string("coreml_update_state_82_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_82 = read_state(input = key_cache)[name = string("coreml_update_state_82")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_14_stride_0 = const()[name = string("value_cache_internal_tensor_assign_14_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_14_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_14_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_14_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_14_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_27_cast_fp16 = transpose(perm = v_state_27_perm_0, x = var_2589_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_14_cast_fp16 = slice_update(begin = concat_252, begin_mask = value_cache_internal_tensor_assign_14_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_14_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_14_squeeze_mask_0, stride = value_cache_internal_tensor_assign_14_stride_0, update = v_state_27_cast_fp16, x = coreml_update_state_81)[name = string("value_cache_internal_tensor_assign_14_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_14_cast_fp16, input = value_cache)[name = string("coreml_update_state_83_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_83 = read_state(input = value_cache)[name = string("coreml_update_state_83")];
            tensor<int32, [5]> var_2646_begin_0 = const()[name = string("op_2646_begin_0"), val = tensor<int32, [5]>([13, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2646_end_0 = const()[name = string("op_2646_end_0"), val = tensor<int32, [5]>([14, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_2646_end_mask_0 = const()[name = string("op_2646_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2646_squeeze_mask_0 = const()[name = string("op_2646_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_2646_cast_fp16 = slice_by_index(begin = var_2646_begin_0, end = var_2646_end_0, end_mask = var_2646_end_mask_0, squeeze_mask = var_2646_squeeze_mask_0, x = coreml_update_state_82)[name = string("op_2646_cast_fp16")];
            tensor<int32, [4]> var_2649_begin_0 = const()[name = string("op_2649_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2649_end_mask_0 = const()[name = string("op_2649_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_2649_cast_fp16 = slice_by_index(begin = var_2649_begin_0, end = concat_11, end_mask = var_2649_end_mask_0, x = var_2646_cast_fp16)[name = string("op_2649_cast_fp16")];
            tensor<int32, [5]> var_2651_begin_0 = const()[name = string("op_2651_begin_0"), val = tensor<int32, [5]>([13, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2651_end_0 = const()[name = string("op_2651_end_0"), val = tensor<int32, [5]>([14, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_2651_end_mask_0 = const()[name = string("op_2651_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2651_squeeze_mask_0 = const()[name = string("op_2651_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_2651_cast_fp16 = slice_by_index(begin = var_2651_begin_0, end = var_2651_end_0, end_mask = var_2651_end_mask_0, squeeze_mask = var_2651_squeeze_mask_0, x = coreml_update_state_83)[name = string("op_2651_cast_fp16")];
            tensor<int32, [4]> var_2654_begin_0 = const()[name = string("op_2654_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2654_end_mask_0 = const()[name = string("op_2654_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_2654_cast_fp16 = slice_by_index(begin = var_2654_begin_0, end = concat_11, end_mask = var_2654_end_mask_0, x = var_2651_cast_fp16)[name = string("op_2654_cast_fp16")];
            tensor<int32, [4]> var_2656_shape_cast_fp16 = shape(x = var_2649_cast_fp16)[name = string("op_2656_shape_cast_fp16")];
            int32 gather_247 = const()[name = string("gather_247"), val = int32(1)];
            int32 gather_248 = const()[name = string("gather_248"), val = int32(2)];
            int32 gather_249_axis_0 = const()[name = string("gather_249_axis_0"), val = int32(0)];
            int32 gather_249_batch_dims_0 = const()[name = string("gather_249_batch_dims_0"), val = int32(0)];
            bool gather_249_validate_indices_0 = const()[name = string("gather_249_validate_indices_0"), val = bool(false)];
            string var_2656_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2656_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_249_to_uint16 = const()[name = string("select_249_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2656_shape_cast_fp16_to_uint16 = cast(dtype = var_2656_shape_cast_fp16_to_uint16_dtype_0, x = var_2656_shape_cast_fp16)[name = string("cast_89")];
            uint16 gather_249_cast_uint16 = gather(axis = gather_249_axis_0, batch_dims = gather_249_batch_dims_0, indices = select_249_to_uint16, validate_indices = gather_249_validate_indices_0, x = var_2656_shape_cast_fp16_to_uint16)[name = string("gather_249_cast_uint16")];
            string gather_249_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_249_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_250 = const()[name = string("gather_250"), val = int32(128)];
            tensor<int32, [1]> var_2663_axes_0 = const()[name = string("op_2663_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_2663_cast_fp16 = expand_dims(axes = var_2663_axes_0, x = var_2649_cast_fp16)[name = string("op_2663_cast_fp16")];
            tensor<int32, [5]> shape_277_cast_fp16 = shape(x = var_2663_cast_fp16)[name = string("shape_277_cast_fp16")];
            int32 concat_260_axis_0 = const()[name = string("concat_260_axis_0"), val = int32(0)];
            bool concat_260_interleave_0 = const()[name = string("concat_260_interleave_0"), val = bool(false)];
            int32 gather_249_cast_uint16_to_int32 = cast(dtype = gather_249_cast_uint16_to_int32_dtype_0, x = gather_249_cast_uint16)[name = string("cast_88")];
            tensor<int32, [5]> concat_260 = concat(axis = concat_260_axis_0, interleave = concat_260_interleave_0, values = (gather_247, gather_248, var_77, gather_249_cast_uint16_to_int32, gather_250))[name = string("concat_260")];
            tensor<int32, [5]> real_div_26 = real_div(x = concat_260, y = shape_277_cast_fp16)[name = string("real_div_26")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_401_cast_fp16 = tile(reps = real_div_26, x = var_2663_cast_fp16)[name = string("hidden_states_401_cast_fp16")];
            tensor<int32, [4]> concat_261x = const()[name = string("concat_261x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_55_cast_fp16 = reshape(shape = concat_261x, x = hidden_states_401_cast_fp16)[name = string("key_states_55_cast_fp16")];
            tensor<int32, [4]> var_2673_shape_cast_fp16 = shape(x = var_2654_cast_fp16)[name = string("op_2673_shape_cast_fp16")];
            int32 gather_251 = const()[name = string("gather_251"), val = int32(1)];
            int32 gather_252 = const()[name = string("gather_252"), val = int32(2)];
            int32 gather_253_axis_0 = const()[name = string("gather_253_axis_0"), val = int32(0)];
            int32 gather_253_batch_dims_0 = const()[name = string("gather_253_batch_dims_0"), val = int32(0)];
            bool gather_253_validate_indices_0 = const()[name = string("gather_253_validate_indices_0"), val = bool(false)];
            string var_2673_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2673_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_253_to_uint16 = const()[name = string("select_253_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2673_shape_cast_fp16_to_uint16 = cast(dtype = var_2673_shape_cast_fp16_to_uint16_dtype_0, x = var_2673_shape_cast_fp16)[name = string("cast_87")];
            uint16 gather_253_cast_uint16 = gather(axis = gather_253_axis_0, batch_dims = gather_253_batch_dims_0, indices = select_253_to_uint16, validate_indices = gather_253_validate_indices_0, x = var_2673_shape_cast_fp16_to_uint16)[name = string("gather_253_cast_uint16")];
            string gather_253_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_253_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_254 = const()[name = string("gather_254"), val = int32(128)];
            tensor<int32, [1]> var_2680_axes_0 = const()[name = string("op_2680_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_2680_cast_fp16 = expand_dims(axes = var_2680_axes_0, x = var_2654_cast_fp16)[name = string("op_2680_cast_fp16")];
            tensor<int32, [5]> shape_282_cast_fp16 = shape(x = var_2680_cast_fp16)[name = string("shape_282_cast_fp16")];
            int32 concat_262_axis_0 = const()[name = string("concat_262_axis_0"), val = int32(0)];
            bool concat_262_interleave_0 = const()[name = string("concat_262_interleave_0"), val = bool(false)];
            int32 gather_253_cast_uint16_to_int32 = cast(dtype = gather_253_cast_uint16_to_int32_dtype_0, x = gather_253_cast_uint16)[name = string("cast_86")];
            tensor<int32, [5]> concat_262 = concat(axis = concat_262_axis_0, interleave = concat_262_interleave_0, values = (gather_251, gather_252, var_77, gather_253_cast_uint16_to_int32, gather_254))[name = string("concat_262")];
            tensor<int32, [5]> real_div_27 = real_div(x = concat_262, y = shape_282_cast_fp16)[name = string("real_div_27")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_405_cast_fp16 = tile(reps = real_div_27, x = var_2680_cast_fp16)[name = string("hidden_states_405_cast_fp16")];
            tensor<int32, [4]> concat_263x = const()[name = string("concat_263x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_55_cast_fp16 = reshape(shape = concat_263x, x = hidden_states_405_cast_fp16)[name = string("value_states_55_cast_fp16")];
            tensor<int32, [4]> var_2690_shape_cast_fp16 = shape(x = key_states_55_cast_fp16)[name = string("op_2690_shape_cast_fp16")];
            int32 gather_255_axis_0 = const()[name = string("gather_255_axis_0"), val = int32(0)];
            int32 gather_255_batch_dims_0 = const()[name = string("gather_255_batch_dims_0"), val = int32(0)];
            bool gather_255_validate_indices_0 = const()[name = string("gather_255_validate_indices_0"), val = bool(false)];
            string var_2690_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2690_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_255_to_uint16 = const()[name = string("select_255_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2690_shape_cast_fp16_to_uint16 = cast(dtype = var_2690_shape_cast_fp16_to_uint16_dtype_0, x = var_2690_shape_cast_fp16)[name = string("cast_85")];
            uint16 gather_255_cast_uint16 = gather(axis = gather_255_axis_0, batch_dims = gather_255_batch_dims_0, indices = select_255_to_uint16, validate_indices = gather_255_validate_indices_0, x = var_2690_shape_cast_fp16_to_uint16)[name = string("gather_255_cast_uint16")];
            string gather_255_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_255_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_264_values0_0 = const()[name = string("concat_264_values0_0"), val = int32(1)];
            int32 concat_264_values1_0 = const()[name = string("concat_264_values1_0"), val = int32(1)];
            int32 concat_264_values2_0 = const()[name = string("concat_264_values2_0"), val = int32(0)];
            int32 concat_264_axis_0 = const()[name = string("concat_264_axis_0"), val = int32(0)];
            bool concat_264_interleave_0 = const()[name = string("concat_264_interleave_0"), val = bool(false)];
            int32 gather_255_cast_uint16_to_int32 = cast(dtype = gather_255_cast_uint16_to_int32_dtype_0, x = gather_255_cast_uint16)[name = string("cast_84")];
            tensor<int32, [4]> concat_264 = concat(axis = concat_264_axis_0, interleave = concat_264_interleave_0, values = (concat_264_values0_0, concat_264_values1_0, concat_264_values2_0, gather_255_cast_uint16_to_int32))[name = string("concat_264")];
            tensor<int32, [4]> causal_mask_29_begin_0 = const()[name = string("causal_mask_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_29_end_mask_0 = const()[name = string("causal_mask_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_29_cast_fp16 = slice_by_index(begin = causal_mask_29_begin_0, end = concat_264, end_mask = causal_mask_29_end_mask_0, x = causal_mask)[name = string("causal_mask_29_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_53_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_29_cast_fp16, key = key_states_55_cast_fp16, query = query_states_55_cast_fp16, value = value_states_55_cast_fp16)[name = string("attn_output_53_cast_fp16")];
            tensor<int32, [4]> var_2696_perm_0 = const()[name = string("op_2696_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_265x = const()[name = string("concat_265x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_2696_cast_fp16 = transpose(perm = var_2696_perm_0, x = attn_output_53_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, ?, 1536]> input_105_cast_fp16 = reshape(shape = concat_265x, x = var_2696_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475392064))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476571776))))[name = string("model_model_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_94_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_409_cast_fp16 = add(x = hidden_states_389_cast_fp16, y = linear_94_cast_fp16)[name = string("hidden_states_409_cast_fp16")];
            fp16 var_78_promoted_27_to_fp16 = const()[name = string("op_78_promoted_27_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_2705_cast_fp16 = pow(x = hidden_states_409_cast_fp16, y = var_78_promoted_27_to_fp16)[name = string("op_2705_cast_fp16")];
            tensor<int32, [1]> variance_55_axes_0 = const()[name = string("variance_55_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_55_keep_dims_0 = const()[name = string("variance_55_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_55_cast_fp16 = reduce_mean(axes = variance_55_axes_0, keep_dims = variance_55_keep_dims_0, x = var_2705_cast_fp16)[name = string("variance_55_cast_fp16")];
            fp16 var_2708_to_fp16 = const()[name = string("op_2708_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2709_cast_fp16 = add(x = variance_55_cast_fp16, y = var_2708_to_fp16)[name = string("op_2709_cast_fp16")];
            fp32 var_2710_epsilon_0 = const()[name = string("op_2710_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2710_cast_fp16 = rsqrt(epsilon = var_2710_epsilon_0, x = var_2709_cast_fp16)[name = string("op_2710_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_413_cast_fp16 = mul(x = hidden_states_409_cast_fp16, y = var_2710_cast_fp16)[name = string("hidden_states_413_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476719296)))];
            tensor<fp16, [1, ?, 1536]> input_107_cast_fp16 = mul(x = model_model_layers_13_post_attention_layernorm_weight_to_fp16, y = hidden_states_413_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(476722432))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483603776))))[name = string("model_model_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_2722_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_2722_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484464000))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491345344))))[name = string("model_model_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_111_cast_fp16 = mul(x = var_2722_cast_fp16, y = linear_96_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492205568))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(499086912))))[name = string("model_model_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_97_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_419_cast_fp16 = add(x = hidden_states_409_cast_fp16, y = linear_97_cast_fp16)[name = string("hidden_states_419_cast_fp16")];
            fp16 var_78_promoted_28_to_fp16 = const()[name = string("op_78_promoted_28_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_2735_cast_fp16 = pow(x = hidden_states_419_cast_fp16, y = var_78_promoted_28_to_fp16)[name = string("op_2735_cast_fp16")];
            tensor<int32, [1]> variance_57_axes_0 = const()[name = string("variance_57_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_57_keep_dims_0 = const()[name = string("variance_57_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_57_cast_fp16 = reduce_mean(axes = variance_57_axes_0, keep_dims = variance_57_keep_dims_0, x = var_2735_cast_fp16)[name = string("variance_57_cast_fp16")];
            fp16 var_2738_to_fp16 = const()[name = string("op_2738_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2739_cast_fp16 = add(x = variance_57_cast_fp16, y = var_2738_to_fp16)[name = string("op_2739_cast_fp16")];
            fp32 var_2740_epsilon_0 = const()[name = string("op_2740_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2740_cast_fp16 = rsqrt(epsilon = var_2740_epsilon_0, x = var_2739_cast_fp16)[name = string("op_2740_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_423_cast_fp16 = mul(x = hidden_states_419_cast_fp16, y = var_2740_cast_fp16)[name = string("hidden_states_423_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(499947136)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_427_cast_fp16 = mul(x = model_model_layers_14_input_layernorm_weight_to_fp16, y = hidden_states_423_cast_fp16)[name = string("hidden_states_427_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(499950272))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501129984))))[name = string("model_model_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501277504)))];
            tensor<fp16, [1, ?, 1536]> linear_98_cast_fp16 = linear(bias = model_model_layers_14_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_427_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501280640))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501477312))))[name = string("model_model_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_14_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_14_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501501952)))];
            tensor<fp16, [1, ?, 256]> linear_99_cast_fp16 = linear(bias = model_model_layers_14_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_427_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501502528))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501699200))))[name = string("model_model_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501723840)))];
            tensor<fp16, [1, ?, 256]> linear_100_cast_fp16 = linear(bias = model_model_layers_14_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_427_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> concat_266x = const()[name = string("concat_266x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_2763_cast_fp16 = reshape(shape = concat_266x, x = linear_98_cast_fp16)[name = string("op_2763_cast_fp16")];
            tensor<int32, [4]> q_29_perm_0 = const()[name = string("q_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_267x = const()[name = string("concat_267x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_2766_cast_fp16 = reshape(shape = concat_267x, x = linear_99_cast_fp16)[name = string("op_2766_cast_fp16")];
            tensor<int32, [4]> k_29_perm_0 = const()[name = string("k_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_268x = const()[name = string("concat_268x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_2769_cast_fp16 = reshape(shape = concat_268x, x = linear_100_cast_fp16)[name = string("op_2769_cast_fp16")];
            tensor<int32, [4]> v_state_29_perm_0 = const()[name = string("v_state_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_29_cast_fp16 = transpose(perm = q_29_perm_0, x = var_2763_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 12, ?, 128]> var_2773_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2773_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_29_cast_fp16)[name = string("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_29_cast_fp16)[name = string("x2_57_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_2784_cast_fp16 = mul(x = x2_57_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_2784_cast_fp16")];
            bool var_2786_interleave_0 = const()[name = string("op_2786_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_2786_cast_fp16 = concat(axis = var_72, interleave = var_2786_interleave_0, values = (var_2784_cast_fp16, x1_57_cast_fp16))[name = string("op_2786_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_2787_cast_fp16 = mul(x = var_2786_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2787_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_59_cast_fp16 = add(x = var_2773_cast_fp16, y = var_2787_cast_fp16)[name = string("query_states_59_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_29_cast_fp16 = transpose(perm = k_29_perm_0, x = var_2766_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 2, ?, 128]> var_2789_cast_fp16 = mul(x = k_29_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2789_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_29_cast_fp16)[name = string("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_29_cast_fp16)[name = string("x2_59_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_2800_cast_fp16 = mul(x = x2_59_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_2800_cast_fp16")];
            bool var_2802_interleave_0 = const()[name = string("op_2802_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_2802_cast_fp16 = concat(axis = var_72, interleave = var_2802_interleave_0, values = (var_2800_cast_fp16, x1_59_cast_fp16))[name = string("op_2802_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_2803_cast_fp16 = mul(x = var_2802_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2803_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_29_cast_fp16 = add(x = var_2789_cast_fp16, y = var_2803_cast_fp16)[name = string("k_state_29_cast_fp16")];
            tensor<int32, [1]> expand_dims_168 = const()[name = string("expand_dims_168"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_169 = const()[name = string("expand_dims_169"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_171 = const()[name = string("expand_dims_171"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_271_values0_0 = const()[name = string("concat_271_values0_0"), val = tensor<int32, [1]>([14])];
            int32 concat_271_axis_0 = const()[name = string("concat_271_axis_0"), val = int32(0)];
            bool concat_271_interleave_0 = const()[name = string("concat_271_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_271 = concat(axis = concat_271_axis_0, interleave = concat_271_interleave_0, values = (concat_271_values0_0, expand_dims_168, expand_dims_169, expand_dims_2, expand_dims_171))[name = string("concat_271")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_15_stride_0 = const()[name = string("key_cache_internal_tensor_assign_15_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_15_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_15_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_15_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_15_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_15_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_15_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_15_cast_fp16 = slice_update(begin = concat_271, begin_mask = key_cache_internal_tensor_assign_15_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_15_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_15_squeeze_mask_0, stride = key_cache_internal_tensor_assign_15_stride_0, update = k_state_29_cast_fp16, x = coreml_update_state_82)[name = string("key_cache_internal_tensor_assign_15_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_15_cast_fp16, input = key_cache)[name = string("coreml_update_state_84_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_84 = read_state(input = key_cache)[name = string("coreml_update_state_84")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_15_stride_0 = const()[name = string("value_cache_internal_tensor_assign_15_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_15_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_15_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_15_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_15_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_15_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_15_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_29_cast_fp16 = transpose(perm = v_state_29_perm_0, x = var_2769_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_15_cast_fp16 = slice_update(begin = concat_271, begin_mask = value_cache_internal_tensor_assign_15_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_15_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_15_squeeze_mask_0, stride = value_cache_internal_tensor_assign_15_stride_0, update = v_state_29_cast_fp16, x = coreml_update_state_83)[name = string("value_cache_internal_tensor_assign_15_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_15_cast_fp16, input = value_cache)[name = string("coreml_update_state_85_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_85 = read_state(input = value_cache)[name = string("coreml_update_state_85")];
            tensor<int32, [5]> var_2826_begin_0 = const()[name = string("op_2826_begin_0"), val = tensor<int32, [5]>([14, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2826_end_0 = const()[name = string("op_2826_end_0"), val = tensor<int32, [5]>([15, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_2826_end_mask_0 = const()[name = string("op_2826_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2826_squeeze_mask_0 = const()[name = string("op_2826_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_2826_cast_fp16 = slice_by_index(begin = var_2826_begin_0, end = var_2826_end_0, end_mask = var_2826_end_mask_0, squeeze_mask = var_2826_squeeze_mask_0, x = coreml_update_state_84)[name = string("op_2826_cast_fp16")];
            tensor<int32, [4]> var_2829_begin_0 = const()[name = string("op_2829_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2829_end_mask_0 = const()[name = string("op_2829_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_2829_cast_fp16 = slice_by_index(begin = var_2829_begin_0, end = concat_11, end_mask = var_2829_end_mask_0, x = var_2826_cast_fp16)[name = string("op_2829_cast_fp16")];
            tensor<int32, [5]> var_2831_begin_0 = const()[name = string("op_2831_begin_0"), val = tensor<int32, [5]>([14, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2831_end_0 = const()[name = string("op_2831_end_0"), val = tensor<int32, [5]>([15, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_2831_end_mask_0 = const()[name = string("op_2831_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2831_squeeze_mask_0 = const()[name = string("op_2831_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_2831_cast_fp16 = slice_by_index(begin = var_2831_begin_0, end = var_2831_end_0, end_mask = var_2831_end_mask_0, squeeze_mask = var_2831_squeeze_mask_0, x = coreml_update_state_85)[name = string("op_2831_cast_fp16")];
            tensor<int32, [4]> var_2834_begin_0 = const()[name = string("op_2834_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_2834_end_mask_0 = const()[name = string("op_2834_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_2834_cast_fp16 = slice_by_index(begin = var_2834_begin_0, end = concat_11, end_mask = var_2834_end_mask_0, x = var_2831_cast_fp16)[name = string("op_2834_cast_fp16")];
            tensor<int32, [4]> var_2836_shape_cast_fp16 = shape(x = var_2829_cast_fp16)[name = string("op_2836_shape_cast_fp16")];
            int32 gather_265 = const()[name = string("gather_265"), val = int32(1)];
            int32 gather_266 = const()[name = string("gather_266"), val = int32(2)];
            int32 gather_267_axis_0 = const()[name = string("gather_267_axis_0"), val = int32(0)];
            int32 gather_267_batch_dims_0 = const()[name = string("gather_267_batch_dims_0"), val = int32(0)];
            bool gather_267_validate_indices_0 = const()[name = string("gather_267_validate_indices_0"), val = bool(false)];
            string var_2836_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2836_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_267_to_uint16 = const()[name = string("select_267_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2836_shape_cast_fp16_to_uint16 = cast(dtype = var_2836_shape_cast_fp16_to_uint16_dtype_0, x = var_2836_shape_cast_fp16)[name = string("cast_83")];
            uint16 gather_267_cast_uint16 = gather(axis = gather_267_axis_0, batch_dims = gather_267_batch_dims_0, indices = select_267_to_uint16, validate_indices = gather_267_validate_indices_0, x = var_2836_shape_cast_fp16_to_uint16)[name = string("gather_267_cast_uint16")];
            string gather_267_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_267_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_268 = const()[name = string("gather_268"), val = int32(128)];
            tensor<int32, [1]> var_2843_axes_0 = const()[name = string("op_2843_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_2843_cast_fp16 = expand_dims(axes = var_2843_axes_0, x = var_2829_cast_fp16)[name = string("op_2843_cast_fp16")];
            tensor<int32, [5]> shape_297_cast_fp16 = shape(x = var_2843_cast_fp16)[name = string("shape_297_cast_fp16")];
            int32 concat_279_axis_0 = const()[name = string("concat_279_axis_0"), val = int32(0)];
            bool concat_279_interleave_0 = const()[name = string("concat_279_interleave_0"), val = bool(false)];
            int32 gather_267_cast_uint16_to_int32 = cast(dtype = gather_267_cast_uint16_to_int32_dtype_0, x = gather_267_cast_uint16)[name = string("cast_82")];
            tensor<int32, [5]> concat_279 = concat(axis = concat_279_axis_0, interleave = concat_279_interleave_0, values = (gather_265, gather_266, var_77, gather_267_cast_uint16_to_int32, gather_268))[name = string("concat_279")];
            tensor<int32, [5]> real_div_28 = real_div(x = concat_279, y = shape_297_cast_fp16)[name = string("real_div_28")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_431_cast_fp16 = tile(reps = real_div_28, x = var_2843_cast_fp16)[name = string("hidden_states_431_cast_fp16")];
            tensor<int32, [4]> concat_280x = const()[name = string("concat_280x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_59_cast_fp16 = reshape(shape = concat_280x, x = hidden_states_431_cast_fp16)[name = string("key_states_59_cast_fp16")];
            tensor<int32, [4]> var_2853_shape_cast_fp16 = shape(x = var_2834_cast_fp16)[name = string("op_2853_shape_cast_fp16")];
            int32 gather_269 = const()[name = string("gather_269"), val = int32(1)];
            int32 gather_270 = const()[name = string("gather_270"), val = int32(2)];
            int32 gather_271_axis_0 = const()[name = string("gather_271_axis_0"), val = int32(0)];
            int32 gather_271_batch_dims_0 = const()[name = string("gather_271_batch_dims_0"), val = int32(0)];
            bool gather_271_validate_indices_0 = const()[name = string("gather_271_validate_indices_0"), val = bool(false)];
            string var_2853_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2853_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_271_to_uint16 = const()[name = string("select_271_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2853_shape_cast_fp16_to_uint16 = cast(dtype = var_2853_shape_cast_fp16_to_uint16_dtype_0, x = var_2853_shape_cast_fp16)[name = string("cast_81")];
            uint16 gather_271_cast_uint16 = gather(axis = gather_271_axis_0, batch_dims = gather_271_batch_dims_0, indices = select_271_to_uint16, validate_indices = gather_271_validate_indices_0, x = var_2853_shape_cast_fp16_to_uint16)[name = string("gather_271_cast_uint16")];
            string gather_271_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_271_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_272 = const()[name = string("gather_272"), val = int32(128)];
            tensor<int32, [1]> var_2860_axes_0 = const()[name = string("op_2860_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_2860_cast_fp16 = expand_dims(axes = var_2860_axes_0, x = var_2834_cast_fp16)[name = string("op_2860_cast_fp16")];
            tensor<int32, [5]> shape_302_cast_fp16 = shape(x = var_2860_cast_fp16)[name = string("shape_302_cast_fp16")];
            int32 concat_281_axis_0 = const()[name = string("concat_281_axis_0"), val = int32(0)];
            bool concat_281_interleave_0 = const()[name = string("concat_281_interleave_0"), val = bool(false)];
            int32 gather_271_cast_uint16_to_int32 = cast(dtype = gather_271_cast_uint16_to_int32_dtype_0, x = gather_271_cast_uint16)[name = string("cast_80")];
            tensor<int32, [5]> concat_281 = concat(axis = concat_281_axis_0, interleave = concat_281_interleave_0, values = (gather_269, gather_270, var_77, gather_271_cast_uint16_to_int32, gather_272))[name = string("concat_281")];
            tensor<int32, [5]> real_div_29 = real_div(x = concat_281, y = shape_302_cast_fp16)[name = string("real_div_29")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_435_cast_fp16 = tile(reps = real_div_29, x = var_2860_cast_fp16)[name = string("hidden_states_435_cast_fp16")];
            tensor<int32, [4]> concat_282x = const()[name = string("concat_282x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_59_cast_fp16 = reshape(shape = concat_282x, x = hidden_states_435_cast_fp16)[name = string("value_states_59_cast_fp16")];
            tensor<int32, [4]> var_2870_shape_cast_fp16 = shape(x = key_states_59_cast_fp16)[name = string("op_2870_shape_cast_fp16")];
            int32 gather_273_axis_0 = const()[name = string("gather_273_axis_0"), val = int32(0)];
            int32 gather_273_batch_dims_0 = const()[name = string("gather_273_batch_dims_0"), val = int32(0)];
            bool gather_273_validate_indices_0 = const()[name = string("gather_273_validate_indices_0"), val = bool(false)];
            string var_2870_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2870_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_273_to_uint16 = const()[name = string("select_273_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_2870_shape_cast_fp16_to_uint16 = cast(dtype = var_2870_shape_cast_fp16_to_uint16_dtype_0, x = var_2870_shape_cast_fp16)[name = string("cast_79")];
            uint16 gather_273_cast_uint16 = gather(axis = gather_273_axis_0, batch_dims = gather_273_batch_dims_0, indices = select_273_to_uint16, validate_indices = gather_273_validate_indices_0, x = var_2870_shape_cast_fp16_to_uint16)[name = string("gather_273_cast_uint16")];
            string gather_273_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_273_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_283_values0_0 = const()[name = string("concat_283_values0_0"), val = int32(1)];
            int32 concat_283_values1_0 = const()[name = string("concat_283_values1_0"), val = int32(1)];
            int32 concat_283_values2_0 = const()[name = string("concat_283_values2_0"), val = int32(0)];
            int32 concat_283_axis_0 = const()[name = string("concat_283_axis_0"), val = int32(0)];
            bool concat_283_interleave_0 = const()[name = string("concat_283_interleave_0"), val = bool(false)];
            int32 gather_273_cast_uint16_to_int32 = cast(dtype = gather_273_cast_uint16_to_int32_dtype_0, x = gather_273_cast_uint16)[name = string("cast_78")];
            tensor<int32, [4]> concat_283 = concat(axis = concat_283_axis_0, interleave = concat_283_interleave_0, values = (concat_283_values0_0, concat_283_values1_0, concat_283_values2_0, gather_273_cast_uint16_to_int32))[name = string("concat_283")];
            tensor<int32, [4]> causal_mask_31_begin_0 = const()[name = string("causal_mask_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_31_end_mask_0 = const()[name = string("causal_mask_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_31_cast_fp16 = slice_by_index(begin = causal_mask_31_begin_0, end = concat_283, end_mask = causal_mask_31_end_mask_0, x = causal_mask)[name = string("causal_mask_31_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_57_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_31_cast_fp16, key = key_states_59_cast_fp16, query = query_states_59_cast_fp16, value = value_states_59_cast_fp16)[name = string("attn_output_57_cast_fp16")];
            tensor<int32, [4]> var_2876_perm_0 = const()[name = string("op_2876_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_284x = const()[name = string("concat_284x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_2876_cast_fp16 = transpose(perm = var_2876_perm_0, x = attn_output_57_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, ?, 1536]> input_113_cast_fp16 = reshape(shape = concat_284x, x = var_2876_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501724416))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502904128))))[name = string("model_model_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_101_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_439_cast_fp16 = add(x = hidden_states_419_cast_fp16, y = linear_101_cast_fp16)[name = string("hidden_states_439_cast_fp16")];
            fp16 var_78_promoted_29_to_fp16 = const()[name = string("op_78_promoted_29_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_2885_cast_fp16 = pow(x = hidden_states_439_cast_fp16, y = var_78_promoted_29_to_fp16)[name = string("op_2885_cast_fp16")];
            tensor<int32, [1]> variance_59_axes_0 = const()[name = string("variance_59_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_59_keep_dims_0 = const()[name = string("variance_59_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_59_cast_fp16 = reduce_mean(axes = variance_59_axes_0, keep_dims = variance_59_keep_dims_0, x = var_2885_cast_fp16)[name = string("variance_59_cast_fp16")];
            fp16 var_2888_to_fp16 = const()[name = string("op_2888_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2889_cast_fp16 = add(x = variance_59_cast_fp16, y = var_2888_to_fp16)[name = string("op_2889_cast_fp16")];
            fp32 var_2890_epsilon_0 = const()[name = string("op_2890_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2890_cast_fp16 = rsqrt(epsilon = var_2890_epsilon_0, x = var_2889_cast_fp16)[name = string("op_2890_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_443_cast_fp16 = mul(x = hidden_states_439_cast_fp16, y = var_2890_cast_fp16)[name = string("hidden_states_443_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503051648)))];
            tensor<fp16, [1, ?, 1536]> input_115_cast_fp16 = mul(x = model_model_layers_14_post_attention_layernorm_weight_to_fp16, y = hidden_states_443_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503054784))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(509936128))))[name = string("model_model_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_2902_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_2902_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510796352))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517677696))))[name = string("model_model_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_119_cast_fp16 = mul(x = var_2902_cast_fp16, y = linear_103_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518537920))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525419264))))[name = string("model_model_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_104_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_449_cast_fp16 = add(x = hidden_states_439_cast_fp16, y = linear_104_cast_fp16)[name = string("hidden_states_449_cast_fp16")];
            fp16 var_78_promoted_30_to_fp16 = const()[name = string("op_78_promoted_30_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_2915_cast_fp16 = pow(x = hidden_states_449_cast_fp16, y = var_78_promoted_30_to_fp16)[name = string("op_2915_cast_fp16")];
            tensor<int32, [1]> variance_61_axes_0 = const()[name = string("variance_61_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_61_keep_dims_0 = const()[name = string("variance_61_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_61_cast_fp16 = reduce_mean(axes = variance_61_axes_0, keep_dims = variance_61_keep_dims_0, x = var_2915_cast_fp16)[name = string("variance_61_cast_fp16")];
            fp16 var_2918_to_fp16 = const()[name = string("op_2918_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2919_cast_fp16 = add(x = variance_61_cast_fp16, y = var_2918_to_fp16)[name = string("op_2919_cast_fp16")];
            fp32 var_2920_epsilon_0 = const()[name = string("op_2920_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_2920_cast_fp16 = rsqrt(epsilon = var_2920_epsilon_0, x = var_2919_cast_fp16)[name = string("op_2920_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_453_cast_fp16 = mul(x = hidden_states_449_cast_fp16, y = var_2920_cast_fp16)[name = string("hidden_states_453_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526279488)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_457_cast_fp16 = mul(x = model_model_layers_15_input_layernorm_weight_to_fp16, y = hidden_states_453_cast_fp16)[name = string("hidden_states_457_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526282624))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527462336))))[name = string("model_model_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527609856)))];
            tensor<fp16, [1, ?, 1536]> linear_105_cast_fp16 = linear(bias = model_model_layers_15_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_457_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527612992))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527809664))))[name = string("model_model_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_15_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_15_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527834304)))];
            tensor<fp16, [1, ?, 256]> linear_106_cast_fp16 = linear(bias = model_model_layers_15_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_457_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527834880))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528031552))))[name = string("model_model_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528056192)))];
            tensor<fp16, [1, ?, 256]> linear_107_cast_fp16 = linear(bias = model_model_layers_15_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_457_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> concat_285x = const()[name = string("concat_285x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_2943_cast_fp16 = reshape(shape = concat_285x, x = linear_105_cast_fp16)[name = string("op_2943_cast_fp16")];
            tensor<int32, [4]> q_31_perm_0 = const()[name = string("q_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_286x = const()[name = string("concat_286x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_2946_cast_fp16 = reshape(shape = concat_286x, x = linear_106_cast_fp16)[name = string("op_2946_cast_fp16")];
            tensor<int32, [4]> k_31_perm_0 = const()[name = string("k_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_287x = const()[name = string("concat_287x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_2949_cast_fp16 = reshape(shape = concat_287x, x = linear_107_cast_fp16)[name = string("op_2949_cast_fp16")];
            tensor<int32, [4]> v_state_31_perm_0 = const()[name = string("v_state_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_31_cast_fp16 = transpose(perm = q_31_perm_0, x = var_2943_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 12, ?, 128]> var_2953_cast_fp16 = mul(x = q_31_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2953_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_31_cast_fp16)[name = string("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_31_cast_fp16)[name = string("x2_61_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_2964_cast_fp16 = mul(x = x2_61_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_2964_cast_fp16")];
            bool var_2966_interleave_0 = const()[name = string("op_2966_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_2966_cast_fp16 = concat(axis = var_72, interleave = var_2966_interleave_0, values = (var_2964_cast_fp16, x1_61_cast_fp16))[name = string("op_2966_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_2967_cast_fp16 = mul(x = var_2966_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2967_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_63_cast_fp16 = add(x = var_2953_cast_fp16, y = var_2967_cast_fp16)[name = string("query_states_63_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_31_cast_fp16 = transpose(perm = k_31_perm_0, x = var_2946_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 2, ?, 128]> var_2969_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_7_cast_fp16)[name = string("op_2969_cast_fp16")];
            tensor<int32, [4]> x1_63_begin_0 = const()[name = string("x1_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_63_end_0 = const()[name = string("x1_63_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_63_end_mask_0 = const()[name = string("x1_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_63_cast_fp16 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = k_31_cast_fp16)[name = string("x1_63_cast_fp16")];
            tensor<int32, [4]> x2_63_begin_0 = const()[name = string("x2_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_63_end_0 = const()[name = string("x2_63_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_63_end_mask_0 = const()[name = string("x2_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_63_cast_fp16 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = k_31_cast_fp16)[name = string("x2_63_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_2980_cast_fp16 = mul(x = x2_63_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_2980_cast_fp16")];
            bool var_2982_interleave_0 = const()[name = string("op_2982_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_2982_cast_fp16 = concat(axis = var_72, interleave = var_2982_interleave_0, values = (var_2980_cast_fp16, x1_63_cast_fp16))[name = string("op_2982_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_2983_cast_fp16 = mul(x = var_2982_cast_fp16, y = sin_7_cast_fp16)[name = string("op_2983_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_31_cast_fp16 = add(x = var_2969_cast_fp16, y = var_2983_cast_fp16)[name = string("k_state_31_cast_fp16")];
            tensor<int32, [1]> expand_dims_180 = const()[name = string("expand_dims_180"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_181 = const()[name = string("expand_dims_181"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_183 = const()[name = string("expand_dims_183"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_290_values0_0 = const()[name = string("concat_290_values0_0"), val = tensor<int32, [1]>([15])];
            int32 concat_290_axis_0 = const()[name = string("concat_290_axis_0"), val = int32(0)];
            bool concat_290_interleave_0 = const()[name = string("concat_290_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_290 = concat(axis = concat_290_axis_0, interleave = concat_290_interleave_0, values = (concat_290_values0_0, expand_dims_180, expand_dims_181, expand_dims_2, expand_dims_183))[name = string("concat_290")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_16_stride_0 = const()[name = string("key_cache_internal_tensor_assign_16_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_16_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_16_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_16_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_16_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_16_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_16_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_16_cast_fp16 = slice_update(begin = concat_290, begin_mask = key_cache_internal_tensor_assign_16_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_16_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_16_squeeze_mask_0, stride = key_cache_internal_tensor_assign_16_stride_0, update = k_state_31_cast_fp16, x = coreml_update_state_84)[name = string("key_cache_internal_tensor_assign_16_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_16_cast_fp16, input = key_cache)[name = string("coreml_update_state_86_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_86 = read_state(input = key_cache)[name = string("coreml_update_state_86")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_16_stride_0 = const()[name = string("value_cache_internal_tensor_assign_16_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_16_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_16_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_16_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_16_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_16_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_16_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_31_cast_fp16 = transpose(perm = v_state_31_perm_0, x = var_2949_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_16_cast_fp16 = slice_update(begin = concat_290, begin_mask = value_cache_internal_tensor_assign_16_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_16_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_16_squeeze_mask_0, stride = value_cache_internal_tensor_assign_16_stride_0, update = v_state_31_cast_fp16, x = coreml_update_state_85)[name = string("value_cache_internal_tensor_assign_16_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_16_cast_fp16, input = value_cache)[name = string("coreml_update_state_87_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_87 = read_state(input = value_cache)[name = string("coreml_update_state_87")];
            tensor<int32, [5]> var_3006_begin_0 = const()[name = string("op_3006_begin_0"), val = tensor<int32, [5]>([15, 0, 0, 0, 0])];
            tensor<int32, [5]> var_3006_end_0 = const()[name = string("op_3006_end_0"), val = tensor<int32, [5]>([16, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_3006_end_mask_0 = const()[name = string("op_3006_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_3006_squeeze_mask_0 = const()[name = string("op_3006_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_3006_cast_fp16 = slice_by_index(begin = var_3006_begin_0, end = var_3006_end_0, end_mask = var_3006_end_mask_0, squeeze_mask = var_3006_squeeze_mask_0, x = coreml_update_state_86)[name = string("op_3006_cast_fp16")];
            tensor<int32, [4]> var_3009_begin_0 = const()[name = string("op_3009_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_3009_end_mask_0 = const()[name = string("op_3009_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_3009_cast_fp16 = slice_by_index(begin = var_3009_begin_0, end = concat_11, end_mask = var_3009_end_mask_0, x = var_3006_cast_fp16)[name = string("op_3009_cast_fp16")];
            tensor<int32, [5]> var_3011_begin_0 = const()[name = string("op_3011_begin_0"), val = tensor<int32, [5]>([15, 0, 0, 0, 0])];
            tensor<int32, [5]> var_3011_end_0 = const()[name = string("op_3011_end_0"), val = tensor<int32, [5]>([16, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_3011_end_mask_0 = const()[name = string("op_3011_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_3011_squeeze_mask_0 = const()[name = string("op_3011_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_3011_cast_fp16 = slice_by_index(begin = var_3011_begin_0, end = var_3011_end_0, end_mask = var_3011_end_mask_0, squeeze_mask = var_3011_squeeze_mask_0, x = coreml_update_state_87)[name = string("op_3011_cast_fp16")];
            tensor<int32, [4]> var_3014_begin_0 = const()[name = string("op_3014_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_3014_end_mask_0 = const()[name = string("op_3014_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_3014_cast_fp16 = slice_by_index(begin = var_3014_begin_0, end = concat_11, end_mask = var_3014_end_mask_0, x = var_3011_cast_fp16)[name = string("op_3014_cast_fp16")];
            tensor<int32, [4]> var_3016_shape_cast_fp16 = shape(x = var_3009_cast_fp16)[name = string("op_3016_shape_cast_fp16")];
            int32 gather_283 = const()[name = string("gather_283"), val = int32(1)];
            int32 gather_284 = const()[name = string("gather_284"), val = int32(2)];
            int32 gather_285_axis_0 = const()[name = string("gather_285_axis_0"), val = int32(0)];
            int32 gather_285_batch_dims_0 = const()[name = string("gather_285_batch_dims_0"), val = int32(0)];
            bool gather_285_validate_indices_0 = const()[name = string("gather_285_validate_indices_0"), val = bool(false)];
            string var_3016_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3016_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_285_to_uint16 = const()[name = string("select_285_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3016_shape_cast_fp16_to_uint16 = cast(dtype = var_3016_shape_cast_fp16_to_uint16_dtype_0, x = var_3016_shape_cast_fp16)[name = string("cast_77")];
            uint16 gather_285_cast_uint16 = gather(axis = gather_285_axis_0, batch_dims = gather_285_batch_dims_0, indices = select_285_to_uint16, validate_indices = gather_285_validate_indices_0, x = var_3016_shape_cast_fp16_to_uint16)[name = string("gather_285_cast_uint16")];
            string gather_285_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_285_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_286 = const()[name = string("gather_286"), val = int32(128)];
            tensor<int32, [1]> var_3023_axes_0 = const()[name = string("op_3023_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_3023_cast_fp16 = expand_dims(axes = var_3023_axes_0, x = var_3009_cast_fp16)[name = string("op_3023_cast_fp16")];
            tensor<int32, [5]> shape_317_cast_fp16 = shape(x = var_3023_cast_fp16)[name = string("shape_317_cast_fp16")];
            int32 concat_298_axis_0 = const()[name = string("concat_298_axis_0"), val = int32(0)];
            bool concat_298_interleave_0 = const()[name = string("concat_298_interleave_0"), val = bool(false)];
            int32 gather_285_cast_uint16_to_int32 = cast(dtype = gather_285_cast_uint16_to_int32_dtype_0, x = gather_285_cast_uint16)[name = string("cast_76")];
            tensor<int32, [5]> concat_298 = concat(axis = concat_298_axis_0, interleave = concat_298_interleave_0, values = (gather_283, gather_284, var_77, gather_285_cast_uint16_to_int32, gather_286))[name = string("concat_298")];
            tensor<int32, [5]> real_div_30 = real_div(x = concat_298, y = shape_317_cast_fp16)[name = string("real_div_30")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_461_cast_fp16 = tile(reps = real_div_30, x = var_3023_cast_fp16)[name = string("hidden_states_461_cast_fp16")];
            tensor<int32, [4]> concat_299x = const()[name = string("concat_299x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_63_cast_fp16 = reshape(shape = concat_299x, x = hidden_states_461_cast_fp16)[name = string("key_states_63_cast_fp16")];
            tensor<int32, [4]> var_3033_shape_cast_fp16 = shape(x = var_3014_cast_fp16)[name = string("op_3033_shape_cast_fp16")];
            int32 gather_287 = const()[name = string("gather_287"), val = int32(1)];
            int32 gather_288 = const()[name = string("gather_288"), val = int32(2)];
            int32 gather_289_axis_0 = const()[name = string("gather_289_axis_0"), val = int32(0)];
            int32 gather_289_batch_dims_0 = const()[name = string("gather_289_batch_dims_0"), val = int32(0)];
            bool gather_289_validate_indices_0 = const()[name = string("gather_289_validate_indices_0"), val = bool(false)];
            string var_3033_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3033_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_289_to_uint16 = const()[name = string("select_289_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3033_shape_cast_fp16_to_uint16 = cast(dtype = var_3033_shape_cast_fp16_to_uint16_dtype_0, x = var_3033_shape_cast_fp16)[name = string("cast_75")];
            uint16 gather_289_cast_uint16 = gather(axis = gather_289_axis_0, batch_dims = gather_289_batch_dims_0, indices = select_289_to_uint16, validate_indices = gather_289_validate_indices_0, x = var_3033_shape_cast_fp16_to_uint16)[name = string("gather_289_cast_uint16")];
            string gather_289_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_289_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_290 = const()[name = string("gather_290"), val = int32(128)];
            tensor<int32, [1]> var_3040_axes_0 = const()[name = string("op_3040_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_3040_cast_fp16 = expand_dims(axes = var_3040_axes_0, x = var_3014_cast_fp16)[name = string("op_3040_cast_fp16")];
            tensor<int32, [5]> shape_322_cast_fp16 = shape(x = var_3040_cast_fp16)[name = string("shape_322_cast_fp16")];
            int32 concat_300_axis_0 = const()[name = string("concat_300_axis_0"), val = int32(0)];
            bool concat_300_interleave_0 = const()[name = string("concat_300_interleave_0"), val = bool(false)];
            int32 gather_289_cast_uint16_to_int32 = cast(dtype = gather_289_cast_uint16_to_int32_dtype_0, x = gather_289_cast_uint16)[name = string("cast_74")];
            tensor<int32, [5]> concat_300 = concat(axis = concat_300_axis_0, interleave = concat_300_interleave_0, values = (gather_287, gather_288, var_77, gather_289_cast_uint16_to_int32, gather_290))[name = string("concat_300")];
            tensor<int32, [5]> real_div_31 = real_div(x = concat_300, y = shape_322_cast_fp16)[name = string("real_div_31")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_465_cast_fp16 = tile(reps = real_div_31, x = var_3040_cast_fp16)[name = string("hidden_states_465_cast_fp16")];
            tensor<int32, [4]> concat_301x = const()[name = string("concat_301x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_63_cast_fp16 = reshape(shape = concat_301x, x = hidden_states_465_cast_fp16)[name = string("value_states_63_cast_fp16")];
            tensor<int32, [4]> var_3050_shape_cast_fp16 = shape(x = key_states_63_cast_fp16)[name = string("op_3050_shape_cast_fp16")];
            int32 gather_291_axis_0 = const()[name = string("gather_291_axis_0"), val = int32(0)];
            int32 gather_291_batch_dims_0 = const()[name = string("gather_291_batch_dims_0"), val = int32(0)];
            bool gather_291_validate_indices_0 = const()[name = string("gather_291_validate_indices_0"), val = bool(false)];
            string var_3050_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3050_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_291_to_uint16 = const()[name = string("select_291_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3050_shape_cast_fp16_to_uint16 = cast(dtype = var_3050_shape_cast_fp16_to_uint16_dtype_0, x = var_3050_shape_cast_fp16)[name = string("cast_73")];
            uint16 gather_291_cast_uint16 = gather(axis = gather_291_axis_0, batch_dims = gather_291_batch_dims_0, indices = select_291_to_uint16, validate_indices = gather_291_validate_indices_0, x = var_3050_shape_cast_fp16_to_uint16)[name = string("gather_291_cast_uint16")];
            string gather_291_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_291_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_302_values0_0 = const()[name = string("concat_302_values0_0"), val = int32(1)];
            int32 concat_302_values1_0 = const()[name = string("concat_302_values1_0"), val = int32(1)];
            int32 concat_302_values2_0 = const()[name = string("concat_302_values2_0"), val = int32(0)];
            int32 concat_302_axis_0 = const()[name = string("concat_302_axis_0"), val = int32(0)];
            bool concat_302_interleave_0 = const()[name = string("concat_302_interleave_0"), val = bool(false)];
            int32 gather_291_cast_uint16_to_int32 = cast(dtype = gather_291_cast_uint16_to_int32_dtype_0, x = gather_291_cast_uint16)[name = string("cast_72")];
            tensor<int32, [4]> concat_302 = concat(axis = concat_302_axis_0, interleave = concat_302_interleave_0, values = (concat_302_values0_0, concat_302_values1_0, concat_302_values2_0, gather_291_cast_uint16_to_int32))[name = string("concat_302")];
            tensor<int32, [4]> causal_mask_33_begin_0 = const()[name = string("causal_mask_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_33_end_mask_0 = const()[name = string("causal_mask_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_33_cast_fp16 = slice_by_index(begin = causal_mask_33_begin_0, end = concat_302, end_mask = causal_mask_33_end_mask_0, x = causal_mask)[name = string("causal_mask_33_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_61_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_33_cast_fp16, key = key_states_63_cast_fp16, query = query_states_63_cast_fp16, value = value_states_63_cast_fp16)[name = string("attn_output_61_cast_fp16")];
            tensor<int32, [4]> var_3056_perm_0 = const()[name = string("op_3056_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_303x = const()[name = string("concat_303x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_3056_cast_fp16 = transpose(perm = var_3056_perm_0, x = attn_output_61_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, ?, 1536]> input_121_cast_fp16 = reshape(shape = concat_303x, x = var_3056_cast_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528056768))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529236480))))[name = string("model_model_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_108_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_469_cast_fp16 = add(x = hidden_states_449_cast_fp16, y = linear_108_cast_fp16)[name = string("hidden_states_469_cast_fp16")];
            fp16 var_78_promoted_31_to_fp16 = const()[name = string("op_78_promoted_31_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_3065_cast_fp16 = pow(x = hidden_states_469_cast_fp16, y = var_78_promoted_31_to_fp16)[name = string("op_3065_cast_fp16")];
            tensor<int32, [1]> variance_63_axes_0 = const()[name = string("variance_63_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_63_keep_dims_0 = const()[name = string("variance_63_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_63_cast_fp16 = reduce_mean(axes = variance_63_axes_0, keep_dims = variance_63_keep_dims_0, x = var_3065_cast_fp16)[name = string("variance_63_cast_fp16")];
            fp16 var_3068_to_fp16 = const()[name = string("op_3068_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3069_cast_fp16 = add(x = variance_63_cast_fp16, y = var_3068_to_fp16)[name = string("op_3069_cast_fp16")];
            fp32 var_3070_epsilon_0 = const()[name = string("op_3070_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_3070_cast_fp16 = rsqrt(epsilon = var_3070_epsilon_0, x = var_3069_cast_fp16)[name = string("op_3070_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_473_cast_fp16 = mul(x = hidden_states_469_cast_fp16, y = var_3070_cast_fp16)[name = string("hidden_states_473_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529384000)))];
            tensor<fp16, [1, ?, 1536]> input_123_cast_fp16 = mul(x = model_model_layers_15_post_attention_layernorm_weight_to_fp16, y = hidden_states_473_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529387136))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536268480))))[name = string("model_model_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_3082_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_3082_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537128704))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544010048))))[name = string("model_model_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_127_cast_fp16 = mul(x = var_3082_cast_fp16, y = linear_110_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544870272))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551751616))))[name = string("model_model_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_111_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_479_cast_fp16 = add(x = hidden_states_469_cast_fp16, y = linear_111_cast_fp16)[name = string("hidden_states_479_cast_fp16")];
            fp16 var_78_promoted_32_to_fp16 = const()[name = string("op_78_promoted_32_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_3095_cast_fp16 = pow(x = hidden_states_479_cast_fp16, y = var_78_promoted_32_to_fp16)[name = string("op_3095_cast_fp16")];
            tensor<int32, [1]> variance_65_axes_0 = const()[name = string("variance_65_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_65_keep_dims_0 = const()[name = string("variance_65_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_65_cast_fp16 = reduce_mean(axes = variance_65_axes_0, keep_dims = variance_65_keep_dims_0, x = var_3095_cast_fp16)[name = string("variance_65_cast_fp16")];
            fp16 var_3098_to_fp16 = const()[name = string("op_3098_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3099_cast_fp16 = add(x = variance_65_cast_fp16, y = var_3098_to_fp16)[name = string("op_3099_cast_fp16")];
            fp32 var_3100_epsilon_0 = const()[name = string("op_3100_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_3100_cast_fp16 = rsqrt(epsilon = var_3100_epsilon_0, x = var_3099_cast_fp16)[name = string("op_3100_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_483_cast_fp16 = mul(x = hidden_states_479_cast_fp16, y = var_3100_cast_fp16)[name = string("hidden_states_483_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552611840)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_487_cast_fp16 = mul(x = model_model_layers_16_input_layernorm_weight_to_fp16, y = hidden_states_483_cast_fp16)[name = string("hidden_states_487_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552614976))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553794688))))[name = string("model_model_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553942208)))];
            tensor<fp16, [1, ?, 1536]> linear_112_cast_fp16 = linear(bias = model_model_layers_16_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_487_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553945344))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554142016))))[name = string("model_model_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_16_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_16_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554166656)))];
            tensor<fp16, [1, ?, 256]> linear_113_cast_fp16 = linear(bias = model_model_layers_16_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_487_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554167232))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554363904))))[name = string("model_model_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554388544)))];
            tensor<fp16, [1, ?, 256]> linear_114_cast_fp16 = linear(bias = model_model_layers_16_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_487_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [4]> concat_304x = const()[name = string("concat_304x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_3123_cast_fp16 = reshape(shape = concat_304x, x = linear_112_cast_fp16)[name = string("op_3123_cast_fp16")];
            tensor<int32, [4]> q_33_perm_0 = const()[name = string("q_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_305x = const()[name = string("concat_305x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_3126_cast_fp16 = reshape(shape = concat_305x, x = linear_113_cast_fp16)[name = string("op_3126_cast_fp16")];
            tensor<int32, [4]> k_33_perm_0 = const()[name = string("k_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_306x = const()[name = string("concat_306x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_3129_cast_fp16 = reshape(shape = concat_306x, x = linear_114_cast_fp16)[name = string("op_3129_cast_fp16")];
            tensor<int32, [4]> v_state_33_perm_0 = const()[name = string("v_state_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = var_3123_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 12, ?, 128]> var_3133_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_7_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<int32, [4]> x1_65_begin_0 = const()[name = string("x1_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_65_end_0 = const()[name = string("x1_65_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_65_end_mask_0 = const()[name = string("x1_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_65_cast_fp16 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = q_33_cast_fp16)[name = string("x1_65_cast_fp16")];
            tensor<int32, [4]> x2_65_begin_0 = const()[name = string("x2_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_65_end_0 = const()[name = string("x2_65_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_65_end_mask_0 = const()[name = string("x2_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_65_cast_fp16 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = q_33_cast_fp16)[name = string("x2_65_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_3144_cast_fp16 = mul(x = x2_65_cast_fp16, y = const_35_promoted_to_fp16)[name = string("op_3144_cast_fp16")];
            bool var_3146_interleave_0 = const()[name = string("op_3146_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_3146_cast_fp16 = concat(axis = var_72, interleave = var_3146_interleave_0, values = (var_3144_cast_fp16, x1_65_cast_fp16))[name = string("op_3146_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_3147_cast_fp16 = mul(x = var_3146_cast_fp16, y = sin_7_cast_fp16)[name = string("op_3147_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_67_cast_fp16 = add(x = var_3133_cast_fp16, y = var_3147_cast_fp16)[name = string("query_states_67_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = var_3126_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 2, ?, 128]> var_3149_cast_fp16 = mul(x = k_33_cast_fp16, y = cos_7_cast_fp16)[name = string("op_3149_cast_fp16")];
            tensor<int32, [4]> x1_67_begin_0 = const()[name = string("x1_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_67_end_0 = const()[name = string("x1_67_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_67_end_mask_0 = const()[name = string("x1_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_67_cast_fp16 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = k_33_cast_fp16)[name = string("x1_67_cast_fp16")];
            tensor<int32, [4]> x2_67_begin_0 = const()[name = string("x2_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_67_end_0 = const()[name = string("x2_67_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_67_end_mask_0 = const()[name = string("x2_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_67_cast_fp16 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = k_33_cast_fp16)[name = string("x2_67_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_3160_cast_fp16 = mul(x = x2_67_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_3160_cast_fp16")];
            bool var_3162_interleave_0 = const()[name = string("op_3162_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_3162_cast_fp16 = concat(axis = var_72, interleave = var_3162_interleave_0, values = (var_3160_cast_fp16, x1_67_cast_fp16))[name = string("op_3162_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_3163_cast_fp16 = mul(x = var_3162_cast_fp16, y = sin_7_cast_fp16)[name = string("op_3163_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_33_cast_fp16 = add(x = var_3149_cast_fp16, y = var_3163_cast_fp16)[name = string("k_state_33_cast_fp16")];
            tensor<int32, [1]> expand_dims_192 = const()[name = string("expand_dims_192"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_193 = const()[name = string("expand_dims_193"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_195 = const()[name = string("expand_dims_195"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_309_values0_0 = const()[name = string("concat_309_values0_0"), val = tensor<int32, [1]>([16])];
            int32 concat_309_axis_0 = const()[name = string("concat_309_axis_0"), val = int32(0)];
            bool concat_309_interleave_0 = const()[name = string("concat_309_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_309 = concat(axis = concat_309_axis_0, interleave = concat_309_interleave_0, values = (concat_309_values0_0, expand_dims_192, expand_dims_193, expand_dims_2, expand_dims_195))[name = string("concat_309")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_17_stride_0 = const()[name = string("key_cache_internal_tensor_assign_17_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_17_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_17_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_17_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_17_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_17_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_17_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_17_cast_fp16 = slice_update(begin = concat_309, begin_mask = key_cache_internal_tensor_assign_17_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_17_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_17_squeeze_mask_0, stride = key_cache_internal_tensor_assign_17_stride_0, update = k_state_33_cast_fp16, x = coreml_update_state_86)[name = string("key_cache_internal_tensor_assign_17_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_17_cast_fp16, input = key_cache)[name = string("coreml_update_state_88_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_88 = read_state(input = key_cache)[name = string("coreml_update_state_88")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_17_stride_0 = const()[name = string("value_cache_internal_tensor_assign_17_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_17_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_17_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_17_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_17_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_17_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_17_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_33_cast_fp16 = transpose(perm = v_state_33_perm_0, x = var_3129_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_17_cast_fp16 = slice_update(begin = concat_309, begin_mask = value_cache_internal_tensor_assign_17_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_17_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_17_squeeze_mask_0, stride = value_cache_internal_tensor_assign_17_stride_0, update = v_state_33_cast_fp16, x = coreml_update_state_87)[name = string("value_cache_internal_tensor_assign_17_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_17_cast_fp16, input = value_cache)[name = string("coreml_update_state_89_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_89 = read_state(input = value_cache)[name = string("coreml_update_state_89")];
            tensor<int32, [5]> var_3186_begin_0 = const()[name = string("op_3186_begin_0"), val = tensor<int32, [5]>([16, 0, 0, 0, 0])];
            tensor<int32, [5]> var_3186_end_0 = const()[name = string("op_3186_end_0"), val = tensor<int32, [5]>([17, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_3186_end_mask_0 = const()[name = string("op_3186_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_3186_squeeze_mask_0 = const()[name = string("op_3186_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_3186_cast_fp16 = slice_by_index(begin = var_3186_begin_0, end = var_3186_end_0, end_mask = var_3186_end_mask_0, squeeze_mask = var_3186_squeeze_mask_0, x = coreml_update_state_88)[name = string("op_3186_cast_fp16")];
            tensor<int32, [4]> var_3189_begin_0 = const()[name = string("op_3189_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_3189_end_mask_0 = const()[name = string("op_3189_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_3189_cast_fp16 = slice_by_index(begin = var_3189_begin_0, end = concat_11, end_mask = var_3189_end_mask_0, x = var_3186_cast_fp16)[name = string("op_3189_cast_fp16")];
            tensor<int32, [5]> var_3191_begin_0 = const()[name = string("op_3191_begin_0"), val = tensor<int32, [5]>([16, 0, 0, 0, 0])];
            tensor<int32, [5]> var_3191_end_0 = const()[name = string("op_3191_end_0"), val = tensor<int32, [5]>([17, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_3191_end_mask_0 = const()[name = string("op_3191_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_3191_squeeze_mask_0 = const()[name = string("op_3191_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_3191_cast_fp16 = slice_by_index(begin = var_3191_begin_0, end = var_3191_end_0, end_mask = var_3191_end_mask_0, squeeze_mask = var_3191_squeeze_mask_0, x = coreml_update_state_89)[name = string("op_3191_cast_fp16")];
            tensor<int32, [4]> var_3194_begin_0 = const()[name = string("op_3194_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_3194_end_mask_0 = const()[name = string("op_3194_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_3194_cast_fp16 = slice_by_index(begin = var_3194_begin_0, end = concat_11, end_mask = var_3194_end_mask_0, x = var_3191_cast_fp16)[name = string("op_3194_cast_fp16")];
            tensor<int32, [4]> var_3196_shape_cast_fp16 = shape(x = var_3189_cast_fp16)[name = string("op_3196_shape_cast_fp16")];
            int32 gather_301 = const()[name = string("gather_301"), val = int32(1)];
            int32 gather_302 = const()[name = string("gather_302"), val = int32(2)];
            int32 gather_303_axis_0 = const()[name = string("gather_303_axis_0"), val = int32(0)];
            int32 gather_303_batch_dims_0 = const()[name = string("gather_303_batch_dims_0"), val = int32(0)];
            bool gather_303_validate_indices_0 = const()[name = string("gather_303_validate_indices_0"), val = bool(false)];
            string var_3196_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3196_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_303_to_uint16 = const()[name = string("select_303_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3196_shape_cast_fp16_to_uint16 = cast(dtype = var_3196_shape_cast_fp16_to_uint16_dtype_0, x = var_3196_shape_cast_fp16)[name = string("cast_71")];
            uint16 gather_303_cast_uint16 = gather(axis = gather_303_axis_0, batch_dims = gather_303_batch_dims_0, indices = select_303_to_uint16, validate_indices = gather_303_validate_indices_0, x = var_3196_shape_cast_fp16_to_uint16)[name = string("gather_303_cast_uint16")];
            string gather_303_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_303_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_304 = const()[name = string("gather_304"), val = int32(128)];
            tensor<int32, [1]> var_3203_axes_0 = const()[name = string("op_3203_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_3203_cast_fp16 = expand_dims(axes = var_3203_axes_0, x = var_3189_cast_fp16)[name = string("op_3203_cast_fp16")];
            tensor<int32, [5]> shape_337_cast_fp16 = shape(x = var_3203_cast_fp16)[name = string("shape_337_cast_fp16")];
            int32 concat_317_axis_0 = const()[name = string("concat_317_axis_0"), val = int32(0)];
            bool concat_317_interleave_0 = const()[name = string("concat_317_interleave_0"), val = bool(false)];
            int32 gather_303_cast_uint16_to_int32 = cast(dtype = gather_303_cast_uint16_to_int32_dtype_0, x = gather_303_cast_uint16)[name = string("cast_70")];
            tensor<int32, [5]> concat_317 = concat(axis = concat_317_axis_0, interleave = concat_317_interleave_0, values = (gather_301, gather_302, var_77, gather_303_cast_uint16_to_int32, gather_304))[name = string("concat_317")];
            tensor<int32, [5]> real_div_32 = real_div(x = concat_317, y = shape_337_cast_fp16)[name = string("real_div_32")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_491_cast_fp16 = tile(reps = real_div_32, x = var_3203_cast_fp16)[name = string("hidden_states_491_cast_fp16")];
            tensor<int32, [4]> concat_318x = const()[name = string("concat_318x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_67_cast_fp16 = reshape(shape = concat_318x, x = hidden_states_491_cast_fp16)[name = string("key_states_67_cast_fp16")];
            tensor<int32, [4]> var_3213_shape_cast_fp16 = shape(x = var_3194_cast_fp16)[name = string("op_3213_shape_cast_fp16")];
            int32 gather_305 = const()[name = string("gather_305"), val = int32(1)];
            int32 gather_306 = const()[name = string("gather_306"), val = int32(2)];
            int32 gather_307_axis_0 = const()[name = string("gather_307_axis_0"), val = int32(0)];
            int32 gather_307_batch_dims_0 = const()[name = string("gather_307_batch_dims_0"), val = int32(0)];
            bool gather_307_validate_indices_0 = const()[name = string("gather_307_validate_indices_0"), val = bool(false)];
            string var_3213_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3213_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_307_to_uint16 = const()[name = string("select_307_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3213_shape_cast_fp16_to_uint16 = cast(dtype = var_3213_shape_cast_fp16_to_uint16_dtype_0, x = var_3213_shape_cast_fp16)[name = string("cast_69")];
            uint16 gather_307_cast_uint16 = gather(axis = gather_307_axis_0, batch_dims = gather_307_batch_dims_0, indices = select_307_to_uint16, validate_indices = gather_307_validate_indices_0, x = var_3213_shape_cast_fp16_to_uint16)[name = string("gather_307_cast_uint16")];
            string gather_307_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_307_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_308 = const()[name = string("gather_308"), val = int32(128)];
            tensor<int32, [1]> var_3220_axes_0 = const()[name = string("op_3220_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_3220_cast_fp16 = expand_dims(axes = var_3220_axes_0, x = var_3194_cast_fp16)[name = string("op_3220_cast_fp16")];
            tensor<int32, [5]> shape_342_cast_fp16 = shape(x = var_3220_cast_fp16)[name = string("shape_342_cast_fp16")];
            int32 concat_319_axis_0 = const()[name = string("concat_319_axis_0"), val = int32(0)];
            bool concat_319_interleave_0 = const()[name = string("concat_319_interleave_0"), val = bool(false)];
            int32 gather_307_cast_uint16_to_int32 = cast(dtype = gather_307_cast_uint16_to_int32_dtype_0, x = gather_307_cast_uint16)[name = string("cast_68")];
            tensor<int32, [5]> concat_319 = concat(axis = concat_319_axis_0, interleave = concat_319_interleave_0, values = (gather_305, gather_306, var_77, gather_307_cast_uint16_to_int32, gather_308))[name = string("concat_319")];
            tensor<int32, [5]> real_div_33 = real_div(x = concat_319, y = shape_342_cast_fp16)[name = string("real_div_33")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_495_cast_fp16 = tile(reps = real_div_33, x = var_3220_cast_fp16)[name = string("hidden_states_495_cast_fp16")];
            tensor<int32, [4]> concat_320x = const()[name = string("concat_320x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_67_cast_fp16 = reshape(shape = concat_320x, x = hidden_states_495_cast_fp16)[name = string("value_states_67_cast_fp16")];
            tensor<int32, [4]> var_3230_shape_cast_fp16 = shape(x = key_states_67_cast_fp16)[name = string("op_3230_shape_cast_fp16")];
            int32 gather_309_axis_0 = const()[name = string("gather_309_axis_0"), val = int32(0)];
            int32 gather_309_batch_dims_0 = const()[name = string("gather_309_batch_dims_0"), val = int32(0)];
            bool gather_309_validate_indices_0 = const()[name = string("gather_309_validate_indices_0"), val = bool(false)];
            string var_3230_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3230_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_309_to_uint16 = const()[name = string("select_309_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3230_shape_cast_fp16_to_uint16 = cast(dtype = var_3230_shape_cast_fp16_to_uint16_dtype_0, x = var_3230_shape_cast_fp16)[name = string("cast_67")];
            uint16 gather_309_cast_uint16 = gather(axis = gather_309_axis_0, batch_dims = gather_309_batch_dims_0, indices = select_309_to_uint16, validate_indices = gather_309_validate_indices_0, x = var_3230_shape_cast_fp16_to_uint16)[name = string("gather_309_cast_uint16")];
            string gather_309_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_309_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_321_values0_0 = const()[name = string("concat_321_values0_0"), val = int32(1)];
            int32 concat_321_values1_0 = const()[name = string("concat_321_values1_0"), val = int32(1)];
            int32 concat_321_values2_0 = const()[name = string("concat_321_values2_0"), val = int32(0)];
            int32 concat_321_axis_0 = const()[name = string("concat_321_axis_0"), val = int32(0)];
            bool concat_321_interleave_0 = const()[name = string("concat_321_interleave_0"), val = bool(false)];
            int32 gather_309_cast_uint16_to_int32 = cast(dtype = gather_309_cast_uint16_to_int32_dtype_0, x = gather_309_cast_uint16)[name = string("cast_66")];
            tensor<int32, [4]> concat_321 = concat(axis = concat_321_axis_0, interleave = concat_321_interleave_0, values = (concat_321_values0_0, concat_321_values1_0, concat_321_values2_0, gather_309_cast_uint16_to_int32))[name = string("concat_321")];
            tensor<int32, [4]> causal_mask_35_begin_0 = const()[name = string("causal_mask_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_35_end_mask_0 = const()[name = string("causal_mask_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_35_cast_fp16 = slice_by_index(begin = causal_mask_35_begin_0, end = concat_321, end_mask = causal_mask_35_end_mask_0, x = causal_mask)[name = string("causal_mask_35_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_65_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_35_cast_fp16, key = key_states_67_cast_fp16, query = query_states_67_cast_fp16, value = value_states_67_cast_fp16)[name = string("attn_output_65_cast_fp16")];
            tensor<int32, [4]> var_3236_perm_0 = const()[name = string("op_3236_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_322x = const()[name = string("concat_322x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_3236_cast_fp16 = transpose(perm = var_3236_perm_0, x = attn_output_65_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, ?, 1536]> input_129_cast_fp16 = reshape(shape = concat_322x, x = var_3236_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554389120))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555568832))))[name = string("model_model_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_115_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = input_129_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_499_cast_fp16 = add(x = hidden_states_479_cast_fp16, y = linear_115_cast_fp16)[name = string("hidden_states_499_cast_fp16")];
            fp16 var_78_promoted_33_to_fp16 = const()[name = string("op_78_promoted_33_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_3245_cast_fp16 = pow(x = hidden_states_499_cast_fp16, y = var_78_promoted_33_to_fp16)[name = string("op_3245_cast_fp16")];
            tensor<int32, [1]> variance_67_axes_0 = const()[name = string("variance_67_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_67_keep_dims_0 = const()[name = string("variance_67_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_67_cast_fp16 = reduce_mean(axes = variance_67_axes_0, keep_dims = variance_67_keep_dims_0, x = var_3245_cast_fp16)[name = string("variance_67_cast_fp16")];
            fp16 var_3248_to_fp16 = const()[name = string("op_3248_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3249_cast_fp16 = add(x = variance_67_cast_fp16, y = var_3248_to_fp16)[name = string("op_3249_cast_fp16")];
            fp32 var_3250_epsilon_0 = const()[name = string("op_3250_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_3250_cast_fp16 = rsqrt(epsilon = var_3250_epsilon_0, x = var_3249_cast_fp16)[name = string("op_3250_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_503_cast_fp16 = mul(x = hidden_states_499_cast_fp16, y = var_3250_cast_fp16)[name = string("hidden_states_503_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555716352)))];
            tensor<fp16, [1, ?, 1536]> input_131_cast_fp16 = mul(x = model_model_layers_16_post_attention_layernorm_weight_to_fp16, y = hidden_states_503_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555719488))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562600832))))[name = string("model_model_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_3262_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_3262_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563461056))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570342400))))[name = string("model_model_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_135_cast_fp16 = mul(x = var_3262_cast_fp16, y = linear_117_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571202624))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578083968))))[name = string("model_model_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_118_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_509_cast_fp16 = add(x = hidden_states_499_cast_fp16, y = linear_118_cast_fp16)[name = string("hidden_states_509_cast_fp16")];
            fp16 var_78_promoted_34_to_fp16 = const()[name = string("op_78_promoted_34_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_3275_cast_fp16 = pow(x = hidden_states_509_cast_fp16, y = var_78_promoted_34_to_fp16)[name = string("op_3275_cast_fp16")];
            tensor<int32, [1]> variance_69_axes_0 = const()[name = string("variance_69_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_69_keep_dims_0 = const()[name = string("variance_69_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_69_cast_fp16 = reduce_mean(axes = variance_69_axes_0, keep_dims = variance_69_keep_dims_0, x = var_3275_cast_fp16)[name = string("variance_69_cast_fp16")];
            fp16 var_3278_to_fp16 = const()[name = string("op_3278_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3279_cast_fp16 = add(x = variance_69_cast_fp16, y = var_3278_to_fp16)[name = string("op_3279_cast_fp16")];
            fp32 var_3280_epsilon_0 = const()[name = string("op_3280_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_3280_cast_fp16 = rsqrt(epsilon = var_3280_epsilon_0, x = var_3279_cast_fp16)[name = string("op_3280_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_513_cast_fp16 = mul(x = hidden_states_509_cast_fp16, y = var_3280_cast_fp16)[name = string("hidden_states_513_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578944192)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_517_cast_fp16 = mul(x = model_model_layers_17_input_layernorm_weight_to_fp16, y = hidden_states_513_cast_fp16)[name = string("hidden_states_517_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578947328))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(580127040))))[name = string("model_model_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(580274560)))];
            tensor<fp16, [1, ?, 1536]> linear_119_cast_fp16 = linear(bias = model_model_layers_17_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_517_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(580277696))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(580474368))))[name = string("model_model_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_17_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_17_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(580499008)))];
            tensor<fp16, [1, ?, 256]> linear_120_cast_fp16 = linear(bias = model_model_layers_17_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_517_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(580499584))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(580696256))))[name = string("model_model_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(580720896)))];
            tensor<fp16, [1, ?, 256]> linear_121_cast_fp16 = linear(bias = model_model_layers_17_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_517_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> concat_323x = const()[name = string("concat_323x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_3303_cast_fp16 = reshape(shape = concat_323x, x = linear_119_cast_fp16)[name = string("op_3303_cast_fp16")];
            tensor<int32, [4]> q_35_perm_0 = const()[name = string("q_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_324x = const()[name = string("concat_324x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_3306_cast_fp16 = reshape(shape = concat_324x, x = linear_120_cast_fp16)[name = string("op_3306_cast_fp16")];
            tensor<int32, [4]> k_35_perm_0 = const()[name = string("k_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_325x = const()[name = string("concat_325x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_3309_cast_fp16 = reshape(shape = concat_325x, x = linear_121_cast_fp16)[name = string("op_3309_cast_fp16")];
            tensor<int32, [4]> v_state_35_perm_0 = const()[name = string("v_state_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_35_cast_fp16 = transpose(perm = q_35_perm_0, x = var_3303_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 12, ?, 128]> var_3313_cast_fp16 = mul(x = q_35_cast_fp16, y = cos_7_cast_fp16)[name = string("op_3313_cast_fp16")];
            tensor<int32, [4]> x1_69_begin_0 = const()[name = string("x1_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_69_end_0 = const()[name = string("x1_69_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_69_end_mask_0 = const()[name = string("x1_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_69_cast_fp16 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = q_35_cast_fp16)[name = string("x1_69_cast_fp16")];
            tensor<int32, [4]> x2_69_begin_0 = const()[name = string("x2_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_69_end_0 = const()[name = string("x2_69_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_69_end_mask_0 = const()[name = string("x2_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_69_cast_fp16 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = q_35_cast_fp16)[name = string("x2_69_cast_fp16")];
            fp16 const_37_promoted_to_fp16 = const()[name = string("const_37_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_3324_cast_fp16 = mul(x = x2_69_cast_fp16, y = const_37_promoted_to_fp16)[name = string("op_3324_cast_fp16")];
            bool var_3326_interleave_0 = const()[name = string("op_3326_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_3326_cast_fp16 = concat(axis = var_72, interleave = var_3326_interleave_0, values = (var_3324_cast_fp16, x1_69_cast_fp16))[name = string("op_3326_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_3327_cast_fp16 = mul(x = var_3326_cast_fp16, y = sin_7_cast_fp16)[name = string("op_3327_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_71_cast_fp16 = add(x = var_3313_cast_fp16, y = var_3327_cast_fp16)[name = string("query_states_71_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_35_cast_fp16 = transpose(perm = k_35_perm_0, x = var_3306_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 2, ?, 128]> var_3329_cast_fp16 = mul(x = k_35_cast_fp16, y = cos_7_cast_fp16)[name = string("op_3329_cast_fp16")];
            tensor<int32, [4]> x1_71_begin_0 = const()[name = string("x1_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_71_end_0 = const()[name = string("x1_71_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_71_end_mask_0 = const()[name = string("x1_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_71_cast_fp16 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = k_35_cast_fp16)[name = string("x1_71_cast_fp16")];
            tensor<int32, [4]> x2_71_begin_0 = const()[name = string("x2_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_71_end_0 = const()[name = string("x2_71_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_71_end_mask_0 = const()[name = string("x2_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_71_cast_fp16 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = k_35_cast_fp16)[name = string("x2_71_cast_fp16")];
            fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_3340_cast_fp16 = mul(x = x2_71_cast_fp16, y = const_38_promoted_to_fp16)[name = string("op_3340_cast_fp16")];
            bool var_3342_interleave_0 = const()[name = string("op_3342_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_3342_cast_fp16 = concat(axis = var_72, interleave = var_3342_interleave_0, values = (var_3340_cast_fp16, x1_71_cast_fp16))[name = string("op_3342_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_3343_cast_fp16 = mul(x = var_3342_cast_fp16, y = sin_7_cast_fp16)[name = string("op_3343_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_35_cast_fp16 = add(x = var_3329_cast_fp16, y = var_3343_cast_fp16)[name = string("k_state_35_cast_fp16")];
            tensor<int32, [1]> expand_dims_204 = const()[name = string("expand_dims_204"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_205 = const()[name = string("expand_dims_205"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_207 = const()[name = string("expand_dims_207"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_328_values0_0 = const()[name = string("concat_328_values0_0"), val = tensor<int32, [1]>([17])];
            int32 concat_328_axis_0 = const()[name = string("concat_328_axis_0"), val = int32(0)];
            bool concat_328_interleave_0 = const()[name = string("concat_328_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_328 = concat(axis = concat_328_axis_0, interleave = concat_328_interleave_0, values = (concat_328_values0_0, expand_dims_204, expand_dims_205, expand_dims_2, expand_dims_207))[name = string("concat_328")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_18_stride_0 = const()[name = string("key_cache_internal_tensor_assign_18_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_18_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_18_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_18_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_18_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_18_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_18_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_18_cast_fp16 = slice_update(begin = concat_328, begin_mask = key_cache_internal_tensor_assign_18_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_18_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_18_squeeze_mask_0, stride = key_cache_internal_tensor_assign_18_stride_0, update = k_state_35_cast_fp16, x = coreml_update_state_88)[name = string("key_cache_internal_tensor_assign_18_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_18_cast_fp16, input = key_cache)[name = string("coreml_update_state_90_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_90 = read_state(input = key_cache)[name = string("coreml_update_state_90")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_18_stride_0 = const()[name = string("value_cache_internal_tensor_assign_18_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_18_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_18_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_18_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_18_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_18_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_18_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_35_cast_fp16 = transpose(perm = v_state_35_perm_0, x = var_3309_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_18_cast_fp16 = slice_update(begin = concat_328, begin_mask = value_cache_internal_tensor_assign_18_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_18_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_18_squeeze_mask_0, stride = value_cache_internal_tensor_assign_18_stride_0, update = v_state_35_cast_fp16, x = coreml_update_state_89)[name = string("value_cache_internal_tensor_assign_18_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_18_cast_fp16, input = value_cache)[name = string("coreml_update_state_91_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_91 = read_state(input = value_cache)[name = string("coreml_update_state_91")];
            tensor<int32, [5]> var_3366_begin_0 = const()[name = string("op_3366_begin_0"), val = tensor<int32, [5]>([17, 0, 0, 0, 0])];
            tensor<int32, [5]> var_3366_end_0 = const()[name = string("op_3366_end_0"), val = tensor<int32, [5]>([18, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_3366_end_mask_0 = const()[name = string("op_3366_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_3366_squeeze_mask_0 = const()[name = string("op_3366_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_3366_cast_fp16 = slice_by_index(begin = var_3366_begin_0, end = var_3366_end_0, end_mask = var_3366_end_mask_0, squeeze_mask = var_3366_squeeze_mask_0, x = coreml_update_state_90)[name = string("op_3366_cast_fp16")];
            tensor<int32, [4]> var_3369_begin_0 = const()[name = string("op_3369_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_3369_end_mask_0 = const()[name = string("op_3369_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_3369_cast_fp16 = slice_by_index(begin = var_3369_begin_0, end = concat_11, end_mask = var_3369_end_mask_0, x = var_3366_cast_fp16)[name = string("op_3369_cast_fp16")];
            tensor<int32, [5]> var_3371_begin_0 = const()[name = string("op_3371_begin_0"), val = tensor<int32, [5]>([17, 0, 0, 0, 0])];
            tensor<int32, [5]> var_3371_end_0 = const()[name = string("op_3371_end_0"), val = tensor<int32, [5]>([18, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_3371_end_mask_0 = const()[name = string("op_3371_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_3371_squeeze_mask_0 = const()[name = string("op_3371_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_3371_cast_fp16 = slice_by_index(begin = var_3371_begin_0, end = var_3371_end_0, end_mask = var_3371_end_mask_0, squeeze_mask = var_3371_squeeze_mask_0, x = coreml_update_state_91)[name = string("op_3371_cast_fp16")];
            tensor<int32, [4]> var_3374_begin_0 = const()[name = string("op_3374_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_3374_end_mask_0 = const()[name = string("op_3374_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_3374_cast_fp16 = slice_by_index(begin = var_3374_begin_0, end = concat_11, end_mask = var_3374_end_mask_0, x = var_3371_cast_fp16)[name = string("op_3374_cast_fp16")];
            tensor<int32, [4]> var_3376_shape_cast_fp16 = shape(x = var_3369_cast_fp16)[name = string("op_3376_shape_cast_fp16")];
            int32 gather_319 = const()[name = string("gather_319"), val = int32(1)];
            int32 gather_320 = const()[name = string("gather_320"), val = int32(2)];
            int32 gather_321_axis_0 = const()[name = string("gather_321_axis_0"), val = int32(0)];
            int32 gather_321_batch_dims_0 = const()[name = string("gather_321_batch_dims_0"), val = int32(0)];
            bool gather_321_validate_indices_0 = const()[name = string("gather_321_validate_indices_0"), val = bool(false)];
            string var_3376_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3376_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_321_to_uint16 = const()[name = string("select_321_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3376_shape_cast_fp16_to_uint16 = cast(dtype = var_3376_shape_cast_fp16_to_uint16_dtype_0, x = var_3376_shape_cast_fp16)[name = string("cast_65")];
            uint16 gather_321_cast_uint16 = gather(axis = gather_321_axis_0, batch_dims = gather_321_batch_dims_0, indices = select_321_to_uint16, validate_indices = gather_321_validate_indices_0, x = var_3376_shape_cast_fp16_to_uint16)[name = string("gather_321_cast_uint16")];
            string gather_321_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_321_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_322 = const()[name = string("gather_322"), val = int32(128)];
            tensor<int32, [1]> var_3383_axes_0 = const()[name = string("op_3383_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_3383_cast_fp16 = expand_dims(axes = var_3383_axes_0, x = var_3369_cast_fp16)[name = string("op_3383_cast_fp16")];
            tensor<int32, [5]> shape_357_cast_fp16 = shape(x = var_3383_cast_fp16)[name = string("shape_357_cast_fp16")];
            int32 concat_336_axis_0 = const()[name = string("concat_336_axis_0"), val = int32(0)];
            bool concat_336_interleave_0 = const()[name = string("concat_336_interleave_0"), val = bool(false)];
            int32 gather_321_cast_uint16_to_int32 = cast(dtype = gather_321_cast_uint16_to_int32_dtype_0, x = gather_321_cast_uint16)[name = string("cast_64")];
            tensor<int32, [5]> concat_336 = concat(axis = concat_336_axis_0, interleave = concat_336_interleave_0, values = (gather_319, gather_320, var_77, gather_321_cast_uint16_to_int32, gather_322))[name = string("concat_336")];
            tensor<int32, [5]> real_div_34 = real_div(x = concat_336, y = shape_357_cast_fp16)[name = string("real_div_34")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_521_cast_fp16 = tile(reps = real_div_34, x = var_3383_cast_fp16)[name = string("hidden_states_521_cast_fp16")];
            tensor<int32, [4]> concat_337x = const()[name = string("concat_337x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_71_cast_fp16 = reshape(shape = concat_337x, x = hidden_states_521_cast_fp16)[name = string("key_states_71_cast_fp16")];
            tensor<int32, [4]> var_3393_shape_cast_fp16 = shape(x = var_3374_cast_fp16)[name = string("op_3393_shape_cast_fp16")];
            int32 gather_323 = const()[name = string("gather_323"), val = int32(1)];
            int32 gather_324 = const()[name = string("gather_324"), val = int32(2)];
            int32 gather_325_axis_0 = const()[name = string("gather_325_axis_0"), val = int32(0)];
            int32 gather_325_batch_dims_0 = const()[name = string("gather_325_batch_dims_0"), val = int32(0)];
            bool gather_325_validate_indices_0 = const()[name = string("gather_325_validate_indices_0"), val = bool(false)];
            string var_3393_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3393_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_325_to_uint16 = const()[name = string("select_325_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3393_shape_cast_fp16_to_uint16 = cast(dtype = var_3393_shape_cast_fp16_to_uint16_dtype_0, x = var_3393_shape_cast_fp16)[name = string("cast_63")];
            uint16 gather_325_cast_uint16 = gather(axis = gather_325_axis_0, batch_dims = gather_325_batch_dims_0, indices = select_325_to_uint16, validate_indices = gather_325_validate_indices_0, x = var_3393_shape_cast_fp16_to_uint16)[name = string("gather_325_cast_uint16")];
            string gather_325_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_325_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_326 = const()[name = string("gather_326"), val = int32(128)];
            tensor<int32, [1]> var_3400_axes_0 = const()[name = string("op_3400_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_3400_cast_fp16 = expand_dims(axes = var_3400_axes_0, x = var_3374_cast_fp16)[name = string("op_3400_cast_fp16")];
            tensor<int32, [5]> shape_362_cast_fp16 = shape(x = var_3400_cast_fp16)[name = string("shape_362_cast_fp16")];
            int32 concat_338_axis_0 = const()[name = string("concat_338_axis_0"), val = int32(0)];
            bool concat_338_interleave_0 = const()[name = string("concat_338_interleave_0"), val = bool(false)];
            int32 gather_325_cast_uint16_to_int32 = cast(dtype = gather_325_cast_uint16_to_int32_dtype_0, x = gather_325_cast_uint16)[name = string("cast_62")];
            tensor<int32, [5]> concat_338 = concat(axis = concat_338_axis_0, interleave = concat_338_interleave_0, values = (gather_323, gather_324, var_77, gather_325_cast_uint16_to_int32, gather_326))[name = string("concat_338")];
            tensor<int32, [5]> real_div_35 = real_div(x = concat_338, y = shape_362_cast_fp16)[name = string("real_div_35")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_525_cast_fp16 = tile(reps = real_div_35, x = var_3400_cast_fp16)[name = string("hidden_states_525_cast_fp16")];
            tensor<int32, [4]> concat_339x = const()[name = string("concat_339x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_71_cast_fp16 = reshape(shape = concat_339x, x = hidden_states_525_cast_fp16)[name = string("value_states_71_cast_fp16")];
            tensor<int32, [4]> var_3410_shape_cast_fp16 = shape(x = key_states_71_cast_fp16)[name = string("op_3410_shape_cast_fp16")];
            int32 gather_327_axis_0 = const()[name = string("gather_327_axis_0"), val = int32(0)];
            int32 gather_327_batch_dims_0 = const()[name = string("gather_327_batch_dims_0"), val = int32(0)];
            bool gather_327_validate_indices_0 = const()[name = string("gather_327_validate_indices_0"), val = bool(false)];
            string var_3410_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3410_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_327_to_uint16 = const()[name = string("select_327_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3410_shape_cast_fp16_to_uint16 = cast(dtype = var_3410_shape_cast_fp16_to_uint16_dtype_0, x = var_3410_shape_cast_fp16)[name = string("cast_61")];
            uint16 gather_327_cast_uint16 = gather(axis = gather_327_axis_0, batch_dims = gather_327_batch_dims_0, indices = select_327_to_uint16, validate_indices = gather_327_validate_indices_0, x = var_3410_shape_cast_fp16_to_uint16)[name = string("gather_327_cast_uint16")];
            string gather_327_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_327_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_340_values0_0 = const()[name = string("concat_340_values0_0"), val = int32(1)];
            int32 concat_340_values1_0 = const()[name = string("concat_340_values1_0"), val = int32(1)];
            int32 concat_340_values2_0 = const()[name = string("concat_340_values2_0"), val = int32(0)];
            int32 concat_340_axis_0 = const()[name = string("concat_340_axis_0"), val = int32(0)];
            bool concat_340_interleave_0 = const()[name = string("concat_340_interleave_0"), val = bool(false)];
            int32 gather_327_cast_uint16_to_int32 = cast(dtype = gather_327_cast_uint16_to_int32_dtype_0, x = gather_327_cast_uint16)[name = string("cast_60")];
            tensor<int32, [4]> concat_340 = concat(axis = concat_340_axis_0, interleave = concat_340_interleave_0, values = (concat_340_values0_0, concat_340_values1_0, concat_340_values2_0, gather_327_cast_uint16_to_int32))[name = string("concat_340")];
            tensor<int32, [4]> causal_mask_37_begin_0 = const()[name = string("causal_mask_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_37_end_mask_0 = const()[name = string("causal_mask_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_37_cast_fp16 = slice_by_index(begin = causal_mask_37_begin_0, end = concat_340, end_mask = causal_mask_37_end_mask_0, x = causal_mask)[name = string("causal_mask_37_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_69_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_37_cast_fp16, key = key_states_71_cast_fp16, query = query_states_71_cast_fp16, value = value_states_71_cast_fp16)[name = string("attn_output_69_cast_fp16")];
            tensor<int32, [4]> var_3416_perm_0 = const()[name = string("op_3416_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_341x = const()[name = string("concat_341x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_3416_cast_fp16 = transpose(perm = var_3416_perm_0, x = attn_output_69_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, ?, 1536]> input_137_cast_fp16 = reshape(shape = concat_341x, x = var_3416_cast_fp16)[name = string("input_137_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(580721472))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(581901184))))[name = string("model_model_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_122_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_529_cast_fp16 = add(x = hidden_states_509_cast_fp16, y = linear_122_cast_fp16)[name = string("hidden_states_529_cast_fp16")];
            fp16 var_78_promoted_35_to_fp16 = const()[name = string("op_78_promoted_35_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_3425_cast_fp16 = pow(x = hidden_states_529_cast_fp16, y = var_78_promoted_35_to_fp16)[name = string("op_3425_cast_fp16")];
            tensor<int32, [1]> variance_71_axes_0 = const()[name = string("variance_71_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_71_keep_dims_0 = const()[name = string("variance_71_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_71_cast_fp16 = reduce_mean(axes = variance_71_axes_0, keep_dims = variance_71_keep_dims_0, x = var_3425_cast_fp16)[name = string("variance_71_cast_fp16")];
            fp16 var_3428_to_fp16 = const()[name = string("op_3428_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3429_cast_fp16 = add(x = variance_71_cast_fp16, y = var_3428_to_fp16)[name = string("op_3429_cast_fp16")];
            fp32 var_3430_epsilon_0 = const()[name = string("op_3430_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_3430_cast_fp16 = rsqrt(epsilon = var_3430_epsilon_0, x = var_3429_cast_fp16)[name = string("op_3430_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_533_cast_fp16 = mul(x = hidden_states_529_cast_fp16, y = var_3430_cast_fp16)[name = string("hidden_states_533_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(582048704)))];
            tensor<fp16, [1, ?, 1536]> input_139_cast_fp16 = mul(x = model_model_layers_17_post_attention_layernorm_weight_to_fp16, y = hidden_states_533_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(582051840))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588933184))))[name = string("model_model_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_3442_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_3442_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589793408))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(596674752))))[name = string("model_model_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_124_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_143_cast_fp16 = mul(x = var_3442_cast_fp16, y = linear_124_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597534976))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604416320))))[name = string("model_model_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_125_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_539_cast_fp16 = add(x = hidden_states_529_cast_fp16, y = linear_125_cast_fp16)[name = string("hidden_states_539_cast_fp16")];
            fp16 var_78_promoted_36_to_fp16 = const()[name = string("op_78_promoted_36_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_3455_cast_fp16 = pow(x = hidden_states_539_cast_fp16, y = var_78_promoted_36_to_fp16)[name = string("op_3455_cast_fp16")];
            tensor<int32, [1]> variance_73_axes_0 = const()[name = string("variance_73_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_73_keep_dims_0 = const()[name = string("variance_73_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_73_cast_fp16 = reduce_mean(axes = variance_73_axes_0, keep_dims = variance_73_keep_dims_0, x = var_3455_cast_fp16)[name = string("variance_73_cast_fp16")];
            fp16 var_3458_to_fp16 = const()[name = string("op_3458_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3459_cast_fp16 = add(x = variance_73_cast_fp16, y = var_3458_to_fp16)[name = string("op_3459_cast_fp16")];
            fp32 var_3460_epsilon_0 = const()[name = string("op_3460_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_3460_cast_fp16 = rsqrt(epsilon = var_3460_epsilon_0, x = var_3459_cast_fp16)[name = string("op_3460_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_543_cast_fp16 = mul(x = hidden_states_539_cast_fp16, y = var_3460_cast_fp16)[name = string("hidden_states_543_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(605276544)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_547_cast_fp16 = mul(x = model_model_layers_18_input_layernorm_weight_to_fp16, y = hidden_states_543_cast_fp16)[name = string("hidden_states_547_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(605279680))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606459392))))[name = string("model_model_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606606912)))];
            tensor<fp16, [1, ?, 1536]> linear_126_cast_fp16 = linear(bias = model_model_layers_18_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_547_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606610048))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606806720))))[name = string("model_model_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_18_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_18_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606831360)))];
            tensor<fp16, [1, ?, 256]> linear_127_cast_fp16 = linear(bias = model_model_layers_18_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_547_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606831936))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607028608))))[name = string("model_model_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607053248)))];
            tensor<fp16, [1, ?, 256]> linear_128_cast_fp16 = linear(bias = model_model_layers_18_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_547_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> concat_342x = const()[name = string("concat_342x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_3483_cast_fp16 = reshape(shape = concat_342x, x = linear_126_cast_fp16)[name = string("op_3483_cast_fp16")];
            tensor<int32, [4]> q_37_perm_0 = const()[name = string("q_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_343x = const()[name = string("concat_343x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_3486_cast_fp16 = reshape(shape = concat_343x, x = linear_127_cast_fp16)[name = string("op_3486_cast_fp16")];
            tensor<int32, [4]> k_37_perm_0 = const()[name = string("k_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_344x = const()[name = string("concat_344x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_3489_cast_fp16 = reshape(shape = concat_344x, x = linear_128_cast_fp16)[name = string("op_3489_cast_fp16")];
            tensor<int32, [4]> v_state_37_perm_0 = const()[name = string("v_state_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_37_cast_fp16 = transpose(perm = q_37_perm_0, x = var_3483_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 12, ?, 128]> var_3493_cast_fp16 = mul(x = q_37_cast_fp16, y = cos_7_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [4]> x1_73_begin_0 = const()[name = string("x1_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_73_end_0 = const()[name = string("x1_73_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_73_end_mask_0 = const()[name = string("x1_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_73_cast_fp16 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = q_37_cast_fp16)[name = string("x1_73_cast_fp16")];
            tensor<int32, [4]> x2_73_begin_0 = const()[name = string("x2_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_73_end_0 = const()[name = string("x2_73_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_73_end_mask_0 = const()[name = string("x2_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_73_cast_fp16 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = q_37_cast_fp16)[name = string("x2_73_cast_fp16")];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_3504_cast_fp16 = mul(x = x2_73_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_3504_cast_fp16")];
            bool var_3506_interleave_0 = const()[name = string("op_3506_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_3506_cast_fp16 = concat(axis = var_72, interleave = var_3506_interleave_0, values = (var_3504_cast_fp16, x1_73_cast_fp16))[name = string("op_3506_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_3507_cast_fp16 = mul(x = var_3506_cast_fp16, y = sin_7_cast_fp16)[name = string("op_3507_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_75_cast_fp16 = add(x = var_3493_cast_fp16, y = var_3507_cast_fp16)[name = string("query_states_75_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_37_cast_fp16 = transpose(perm = k_37_perm_0, x = var_3486_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 2, ?, 128]> var_3509_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_7_cast_fp16)[name = string("op_3509_cast_fp16")];
            tensor<int32, [4]> x1_75_begin_0 = const()[name = string("x1_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_75_end_0 = const()[name = string("x1_75_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_75_end_mask_0 = const()[name = string("x1_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_75_cast_fp16 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = k_37_cast_fp16)[name = string("x1_75_cast_fp16")];
            tensor<int32, [4]> x2_75_begin_0 = const()[name = string("x2_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_75_end_0 = const()[name = string("x2_75_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_75_end_mask_0 = const()[name = string("x2_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_75_cast_fp16 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = k_37_cast_fp16)[name = string("x2_75_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_3520_cast_fp16 = mul(x = x2_75_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_3520_cast_fp16")];
            bool var_3522_interleave_0 = const()[name = string("op_3522_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_3522_cast_fp16 = concat(axis = var_72, interleave = var_3522_interleave_0, values = (var_3520_cast_fp16, x1_75_cast_fp16))[name = string("op_3522_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_3523_cast_fp16 = mul(x = var_3522_cast_fp16, y = sin_7_cast_fp16)[name = string("op_3523_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_37_cast_fp16 = add(x = var_3509_cast_fp16, y = var_3523_cast_fp16)[name = string("k_state_37_cast_fp16")];
            tensor<int32, [1]> expand_dims_216 = const()[name = string("expand_dims_216"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_217 = const()[name = string("expand_dims_217"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_219 = const()[name = string("expand_dims_219"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_347_values0_0 = const()[name = string("concat_347_values0_0"), val = tensor<int32, [1]>([18])];
            int32 concat_347_axis_0 = const()[name = string("concat_347_axis_0"), val = int32(0)];
            bool concat_347_interleave_0 = const()[name = string("concat_347_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_347 = concat(axis = concat_347_axis_0, interleave = concat_347_interleave_0, values = (concat_347_values0_0, expand_dims_216, expand_dims_217, expand_dims_2, expand_dims_219))[name = string("concat_347")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_19_stride_0 = const()[name = string("key_cache_internal_tensor_assign_19_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_19_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_19_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_19_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_19_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_19_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_19_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_19_cast_fp16 = slice_update(begin = concat_347, begin_mask = key_cache_internal_tensor_assign_19_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_19_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_19_squeeze_mask_0, stride = key_cache_internal_tensor_assign_19_stride_0, update = k_state_37_cast_fp16, x = coreml_update_state_90)[name = string("key_cache_internal_tensor_assign_19_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_19_cast_fp16, input = key_cache)[name = string("coreml_update_state_92_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_92 = read_state(input = key_cache)[name = string("coreml_update_state_92")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_19_stride_0 = const()[name = string("value_cache_internal_tensor_assign_19_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_19_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_19_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_19_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_19_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_19_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_19_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_37_cast_fp16 = transpose(perm = v_state_37_perm_0, x = var_3489_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_19_cast_fp16 = slice_update(begin = concat_347, begin_mask = value_cache_internal_tensor_assign_19_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_19_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_19_squeeze_mask_0, stride = value_cache_internal_tensor_assign_19_stride_0, update = v_state_37_cast_fp16, x = coreml_update_state_91)[name = string("value_cache_internal_tensor_assign_19_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_19_cast_fp16, input = value_cache)[name = string("coreml_update_state_93_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_93 = read_state(input = value_cache)[name = string("coreml_update_state_93")];
            tensor<int32, [5]> var_3546_begin_0 = const()[name = string("op_3546_begin_0"), val = tensor<int32, [5]>([18, 0, 0, 0, 0])];
            tensor<int32, [5]> var_3546_end_0 = const()[name = string("op_3546_end_0"), val = tensor<int32, [5]>([19, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_3546_end_mask_0 = const()[name = string("op_3546_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_3546_squeeze_mask_0 = const()[name = string("op_3546_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_3546_cast_fp16 = slice_by_index(begin = var_3546_begin_0, end = var_3546_end_0, end_mask = var_3546_end_mask_0, squeeze_mask = var_3546_squeeze_mask_0, x = coreml_update_state_92)[name = string("op_3546_cast_fp16")];
            tensor<int32, [4]> var_3549_begin_0 = const()[name = string("op_3549_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_3549_end_mask_0 = const()[name = string("op_3549_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_3549_cast_fp16 = slice_by_index(begin = var_3549_begin_0, end = concat_11, end_mask = var_3549_end_mask_0, x = var_3546_cast_fp16)[name = string("op_3549_cast_fp16")];
            tensor<int32, [5]> var_3551_begin_0 = const()[name = string("op_3551_begin_0"), val = tensor<int32, [5]>([18, 0, 0, 0, 0])];
            tensor<int32, [5]> var_3551_end_0 = const()[name = string("op_3551_end_0"), val = tensor<int32, [5]>([19, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_3551_end_mask_0 = const()[name = string("op_3551_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_3551_squeeze_mask_0 = const()[name = string("op_3551_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_3551_cast_fp16 = slice_by_index(begin = var_3551_begin_0, end = var_3551_end_0, end_mask = var_3551_end_mask_0, squeeze_mask = var_3551_squeeze_mask_0, x = coreml_update_state_93)[name = string("op_3551_cast_fp16")];
            tensor<int32, [4]> var_3554_begin_0 = const()[name = string("op_3554_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_3554_end_mask_0 = const()[name = string("op_3554_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_3554_cast_fp16 = slice_by_index(begin = var_3554_begin_0, end = concat_11, end_mask = var_3554_end_mask_0, x = var_3551_cast_fp16)[name = string("op_3554_cast_fp16")];
            tensor<int32, [4]> var_3556_shape_cast_fp16 = shape(x = var_3549_cast_fp16)[name = string("op_3556_shape_cast_fp16")];
            int32 gather_337 = const()[name = string("gather_337"), val = int32(1)];
            int32 gather_338 = const()[name = string("gather_338"), val = int32(2)];
            int32 gather_339_axis_0 = const()[name = string("gather_339_axis_0"), val = int32(0)];
            int32 gather_339_batch_dims_0 = const()[name = string("gather_339_batch_dims_0"), val = int32(0)];
            bool gather_339_validate_indices_0 = const()[name = string("gather_339_validate_indices_0"), val = bool(false)];
            string var_3556_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3556_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_339_to_uint16 = const()[name = string("select_339_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3556_shape_cast_fp16_to_uint16 = cast(dtype = var_3556_shape_cast_fp16_to_uint16_dtype_0, x = var_3556_shape_cast_fp16)[name = string("cast_59")];
            uint16 gather_339_cast_uint16 = gather(axis = gather_339_axis_0, batch_dims = gather_339_batch_dims_0, indices = select_339_to_uint16, validate_indices = gather_339_validate_indices_0, x = var_3556_shape_cast_fp16_to_uint16)[name = string("gather_339_cast_uint16")];
            string gather_339_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_339_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_340 = const()[name = string("gather_340"), val = int32(128)];
            tensor<int32, [1]> var_3563_axes_0 = const()[name = string("op_3563_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_3563_cast_fp16 = expand_dims(axes = var_3563_axes_0, x = var_3549_cast_fp16)[name = string("op_3563_cast_fp16")];
            tensor<int32, [5]> shape_377_cast_fp16 = shape(x = var_3563_cast_fp16)[name = string("shape_377_cast_fp16")];
            int32 concat_355_axis_0 = const()[name = string("concat_355_axis_0"), val = int32(0)];
            bool concat_355_interleave_0 = const()[name = string("concat_355_interleave_0"), val = bool(false)];
            int32 gather_339_cast_uint16_to_int32 = cast(dtype = gather_339_cast_uint16_to_int32_dtype_0, x = gather_339_cast_uint16)[name = string("cast_58")];
            tensor<int32, [5]> concat_355 = concat(axis = concat_355_axis_0, interleave = concat_355_interleave_0, values = (gather_337, gather_338, var_77, gather_339_cast_uint16_to_int32, gather_340))[name = string("concat_355")];
            tensor<int32, [5]> real_div_36 = real_div(x = concat_355, y = shape_377_cast_fp16)[name = string("real_div_36")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_551_cast_fp16 = tile(reps = real_div_36, x = var_3563_cast_fp16)[name = string("hidden_states_551_cast_fp16")];
            tensor<int32, [4]> concat_356x = const()[name = string("concat_356x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_75_cast_fp16 = reshape(shape = concat_356x, x = hidden_states_551_cast_fp16)[name = string("key_states_75_cast_fp16")];
            tensor<int32, [4]> var_3573_shape_cast_fp16 = shape(x = var_3554_cast_fp16)[name = string("op_3573_shape_cast_fp16")];
            int32 gather_341 = const()[name = string("gather_341"), val = int32(1)];
            int32 gather_342 = const()[name = string("gather_342"), val = int32(2)];
            int32 gather_343_axis_0 = const()[name = string("gather_343_axis_0"), val = int32(0)];
            int32 gather_343_batch_dims_0 = const()[name = string("gather_343_batch_dims_0"), val = int32(0)];
            bool gather_343_validate_indices_0 = const()[name = string("gather_343_validate_indices_0"), val = bool(false)];
            string var_3573_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3573_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_343_to_uint16 = const()[name = string("select_343_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3573_shape_cast_fp16_to_uint16 = cast(dtype = var_3573_shape_cast_fp16_to_uint16_dtype_0, x = var_3573_shape_cast_fp16)[name = string("cast_57")];
            uint16 gather_343_cast_uint16 = gather(axis = gather_343_axis_0, batch_dims = gather_343_batch_dims_0, indices = select_343_to_uint16, validate_indices = gather_343_validate_indices_0, x = var_3573_shape_cast_fp16_to_uint16)[name = string("gather_343_cast_uint16")];
            string gather_343_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_343_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_344 = const()[name = string("gather_344"), val = int32(128)];
            tensor<int32, [1]> var_3580_axes_0 = const()[name = string("op_3580_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_3580_cast_fp16 = expand_dims(axes = var_3580_axes_0, x = var_3554_cast_fp16)[name = string("op_3580_cast_fp16")];
            tensor<int32, [5]> shape_382_cast_fp16 = shape(x = var_3580_cast_fp16)[name = string("shape_382_cast_fp16")];
            int32 concat_357_axis_0 = const()[name = string("concat_357_axis_0"), val = int32(0)];
            bool concat_357_interleave_0 = const()[name = string("concat_357_interleave_0"), val = bool(false)];
            int32 gather_343_cast_uint16_to_int32 = cast(dtype = gather_343_cast_uint16_to_int32_dtype_0, x = gather_343_cast_uint16)[name = string("cast_56")];
            tensor<int32, [5]> concat_357 = concat(axis = concat_357_axis_0, interleave = concat_357_interleave_0, values = (gather_341, gather_342, var_77, gather_343_cast_uint16_to_int32, gather_344))[name = string("concat_357")];
            tensor<int32, [5]> real_div_37 = real_div(x = concat_357, y = shape_382_cast_fp16)[name = string("real_div_37")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_555_cast_fp16 = tile(reps = real_div_37, x = var_3580_cast_fp16)[name = string("hidden_states_555_cast_fp16")];
            tensor<int32, [4]> concat_358x = const()[name = string("concat_358x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_75_cast_fp16 = reshape(shape = concat_358x, x = hidden_states_555_cast_fp16)[name = string("value_states_75_cast_fp16")];
            tensor<int32, [4]> var_3590_shape_cast_fp16 = shape(x = key_states_75_cast_fp16)[name = string("op_3590_shape_cast_fp16")];
            int32 gather_345_axis_0 = const()[name = string("gather_345_axis_0"), val = int32(0)];
            int32 gather_345_batch_dims_0 = const()[name = string("gather_345_batch_dims_0"), val = int32(0)];
            bool gather_345_validate_indices_0 = const()[name = string("gather_345_validate_indices_0"), val = bool(false)];
            string var_3590_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3590_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_345_to_uint16 = const()[name = string("select_345_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3590_shape_cast_fp16_to_uint16 = cast(dtype = var_3590_shape_cast_fp16_to_uint16_dtype_0, x = var_3590_shape_cast_fp16)[name = string("cast_55")];
            uint16 gather_345_cast_uint16 = gather(axis = gather_345_axis_0, batch_dims = gather_345_batch_dims_0, indices = select_345_to_uint16, validate_indices = gather_345_validate_indices_0, x = var_3590_shape_cast_fp16_to_uint16)[name = string("gather_345_cast_uint16")];
            string gather_345_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_345_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_359_values0_0 = const()[name = string("concat_359_values0_0"), val = int32(1)];
            int32 concat_359_values1_0 = const()[name = string("concat_359_values1_0"), val = int32(1)];
            int32 concat_359_values2_0 = const()[name = string("concat_359_values2_0"), val = int32(0)];
            int32 concat_359_axis_0 = const()[name = string("concat_359_axis_0"), val = int32(0)];
            bool concat_359_interleave_0 = const()[name = string("concat_359_interleave_0"), val = bool(false)];
            int32 gather_345_cast_uint16_to_int32 = cast(dtype = gather_345_cast_uint16_to_int32_dtype_0, x = gather_345_cast_uint16)[name = string("cast_54")];
            tensor<int32, [4]> concat_359 = concat(axis = concat_359_axis_0, interleave = concat_359_interleave_0, values = (concat_359_values0_0, concat_359_values1_0, concat_359_values2_0, gather_345_cast_uint16_to_int32))[name = string("concat_359")];
            tensor<int32, [4]> causal_mask_39_begin_0 = const()[name = string("causal_mask_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_39_end_mask_0 = const()[name = string("causal_mask_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_39_cast_fp16 = slice_by_index(begin = causal_mask_39_begin_0, end = concat_359, end_mask = causal_mask_39_end_mask_0, x = causal_mask)[name = string("causal_mask_39_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_73_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_39_cast_fp16, key = key_states_75_cast_fp16, query = query_states_75_cast_fp16, value = value_states_75_cast_fp16)[name = string("attn_output_73_cast_fp16")];
            tensor<int32, [4]> var_3596_perm_0 = const()[name = string("op_3596_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_360x = const()[name = string("concat_360x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_3596_cast_fp16 = transpose(perm = var_3596_perm_0, x = attn_output_73_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, ?, 1536]> input_145_cast_fp16 = reshape(shape = concat_360x, x = var_3596_cast_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607053824))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608233536))))[name = string("model_model_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_129_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_559_cast_fp16 = add(x = hidden_states_539_cast_fp16, y = linear_129_cast_fp16)[name = string("hidden_states_559_cast_fp16")];
            fp16 var_78_promoted_37_to_fp16 = const()[name = string("op_78_promoted_37_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_3605_cast_fp16 = pow(x = hidden_states_559_cast_fp16, y = var_78_promoted_37_to_fp16)[name = string("op_3605_cast_fp16")];
            tensor<int32, [1]> variance_75_axes_0 = const()[name = string("variance_75_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_75_keep_dims_0 = const()[name = string("variance_75_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_75_cast_fp16 = reduce_mean(axes = variance_75_axes_0, keep_dims = variance_75_keep_dims_0, x = var_3605_cast_fp16)[name = string("variance_75_cast_fp16")];
            fp16 var_3608_to_fp16 = const()[name = string("op_3608_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3609_cast_fp16 = add(x = variance_75_cast_fp16, y = var_3608_to_fp16)[name = string("op_3609_cast_fp16")];
            fp32 var_3610_epsilon_0 = const()[name = string("op_3610_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_3610_cast_fp16 = rsqrt(epsilon = var_3610_epsilon_0, x = var_3609_cast_fp16)[name = string("op_3610_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_563_cast_fp16 = mul(x = hidden_states_559_cast_fp16, y = var_3610_cast_fp16)[name = string("hidden_states_563_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608381056)))];
            tensor<fp16, [1, ?, 1536]> input_147_cast_fp16 = mul(x = model_model_layers_18_post_attention_layernorm_weight_to_fp16, y = hidden_states_563_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608384192))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(615265536))))[name = string("model_model_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_3622_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_3622_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616125760))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(623007104))))[name = string("model_model_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_131_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_151_cast_fp16 = mul(x = var_3622_cast_fp16, y = linear_131_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(623867328))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630748672))))[name = string("model_model_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_132_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_569_cast_fp16 = add(x = hidden_states_559_cast_fp16, y = linear_132_cast_fp16)[name = string("hidden_states_569_cast_fp16")];
            fp16 var_78_promoted_38_to_fp16 = const()[name = string("op_78_promoted_38_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_3635_cast_fp16 = pow(x = hidden_states_569_cast_fp16, y = var_78_promoted_38_to_fp16)[name = string("op_3635_cast_fp16")];
            tensor<int32, [1]> variance_77_axes_0 = const()[name = string("variance_77_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_77_keep_dims_0 = const()[name = string("variance_77_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_77_cast_fp16 = reduce_mean(axes = variance_77_axes_0, keep_dims = variance_77_keep_dims_0, x = var_3635_cast_fp16)[name = string("variance_77_cast_fp16")];
            fp16 var_3638_to_fp16 = const()[name = string("op_3638_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3639_cast_fp16 = add(x = variance_77_cast_fp16, y = var_3638_to_fp16)[name = string("op_3639_cast_fp16")];
            fp32 var_3640_epsilon_0 = const()[name = string("op_3640_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_3640_cast_fp16 = rsqrt(epsilon = var_3640_epsilon_0, x = var_3639_cast_fp16)[name = string("op_3640_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_573_cast_fp16 = mul(x = hidden_states_569_cast_fp16, y = var_3640_cast_fp16)[name = string("hidden_states_573_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631608896)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_577_cast_fp16 = mul(x = model_model_layers_19_input_layernorm_weight_to_fp16, y = hidden_states_573_cast_fp16)[name = string("hidden_states_577_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631612032))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632791744))))[name = string("model_model_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632939264)))];
            tensor<fp16, [1, ?, 1536]> linear_133_cast_fp16 = linear(bias = model_model_layers_19_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_577_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632942400))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633139072))))[name = string("model_model_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_19_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_19_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633163712)))];
            tensor<fp16, [1, ?, 256]> linear_134_cast_fp16 = linear(bias = model_model_layers_19_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_577_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633164288))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633360960))))[name = string("model_model_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633385600)))];
            tensor<fp16, [1, ?, 256]> linear_135_cast_fp16 = linear(bias = model_model_layers_19_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_577_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<int32, [4]> concat_361x = const()[name = string("concat_361x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_3663_cast_fp16 = reshape(shape = concat_361x, x = linear_133_cast_fp16)[name = string("op_3663_cast_fp16")];
            tensor<int32, [4]> q_39_perm_0 = const()[name = string("q_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_362x = const()[name = string("concat_362x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_3666_cast_fp16 = reshape(shape = concat_362x, x = linear_134_cast_fp16)[name = string("op_3666_cast_fp16")];
            tensor<int32, [4]> k_39_perm_0 = const()[name = string("k_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_363x = const()[name = string("concat_363x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_3669_cast_fp16 = reshape(shape = concat_363x, x = linear_135_cast_fp16)[name = string("op_3669_cast_fp16")];
            tensor<int32, [4]> v_state_39_perm_0 = const()[name = string("v_state_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_39_cast_fp16 = transpose(perm = q_39_perm_0, x = var_3663_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 12, ?, 128]> var_3673_cast_fp16 = mul(x = q_39_cast_fp16, y = cos_7_cast_fp16)[name = string("op_3673_cast_fp16")];
            tensor<int32, [4]> x1_77_begin_0 = const()[name = string("x1_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_77_end_0 = const()[name = string("x1_77_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_77_end_mask_0 = const()[name = string("x1_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_77_cast_fp16 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = q_39_cast_fp16)[name = string("x1_77_cast_fp16")];
            tensor<int32, [4]> x2_77_begin_0 = const()[name = string("x2_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_77_end_0 = const()[name = string("x2_77_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_77_end_mask_0 = const()[name = string("x2_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_77_cast_fp16 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = q_39_cast_fp16)[name = string("x2_77_cast_fp16")];
            fp16 const_41_promoted_to_fp16 = const()[name = string("const_41_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_3684_cast_fp16 = mul(x = x2_77_cast_fp16, y = const_41_promoted_to_fp16)[name = string("op_3684_cast_fp16")];
            bool var_3686_interleave_0 = const()[name = string("op_3686_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_3686_cast_fp16 = concat(axis = var_72, interleave = var_3686_interleave_0, values = (var_3684_cast_fp16, x1_77_cast_fp16))[name = string("op_3686_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_3687_cast_fp16 = mul(x = var_3686_cast_fp16, y = sin_7_cast_fp16)[name = string("op_3687_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_79_cast_fp16 = add(x = var_3673_cast_fp16, y = var_3687_cast_fp16)[name = string("query_states_79_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_39_cast_fp16 = transpose(perm = k_39_perm_0, x = var_3666_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 2, ?, 128]> var_3689_cast_fp16 = mul(x = k_39_cast_fp16, y = cos_7_cast_fp16)[name = string("op_3689_cast_fp16")];
            tensor<int32, [4]> x1_79_begin_0 = const()[name = string("x1_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_79_end_0 = const()[name = string("x1_79_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_79_end_mask_0 = const()[name = string("x1_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_79_cast_fp16 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = k_39_cast_fp16)[name = string("x1_79_cast_fp16")];
            tensor<int32, [4]> x2_79_begin_0 = const()[name = string("x2_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_79_end_0 = const()[name = string("x2_79_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_79_end_mask_0 = const()[name = string("x2_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_79_cast_fp16 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = k_39_cast_fp16)[name = string("x2_79_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_3700_cast_fp16 = mul(x = x2_79_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_3700_cast_fp16")];
            bool var_3702_interleave_0 = const()[name = string("op_3702_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_3702_cast_fp16 = concat(axis = var_72, interleave = var_3702_interleave_0, values = (var_3700_cast_fp16, x1_79_cast_fp16))[name = string("op_3702_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_3703_cast_fp16 = mul(x = var_3702_cast_fp16, y = sin_7_cast_fp16)[name = string("op_3703_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_39_cast_fp16 = add(x = var_3689_cast_fp16, y = var_3703_cast_fp16)[name = string("k_state_39_cast_fp16")];
            tensor<int32, [1]> expand_dims_228 = const()[name = string("expand_dims_228"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_229 = const()[name = string("expand_dims_229"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_231 = const()[name = string("expand_dims_231"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_366_values0_0 = const()[name = string("concat_366_values0_0"), val = tensor<int32, [1]>([19])];
            int32 concat_366_axis_0 = const()[name = string("concat_366_axis_0"), val = int32(0)];
            bool concat_366_interleave_0 = const()[name = string("concat_366_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_366 = concat(axis = concat_366_axis_0, interleave = concat_366_interleave_0, values = (concat_366_values0_0, expand_dims_228, expand_dims_229, expand_dims_2, expand_dims_231))[name = string("concat_366")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_20_stride_0 = const()[name = string("key_cache_internal_tensor_assign_20_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_20_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_20_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_20_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_20_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_20_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_20_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_20_cast_fp16 = slice_update(begin = concat_366, begin_mask = key_cache_internal_tensor_assign_20_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_20_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_20_squeeze_mask_0, stride = key_cache_internal_tensor_assign_20_stride_0, update = k_state_39_cast_fp16, x = coreml_update_state_92)[name = string("key_cache_internal_tensor_assign_20_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_20_cast_fp16, input = key_cache)[name = string("coreml_update_state_94_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_94 = read_state(input = key_cache)[name = string("coreml_update_state_94")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_20_stride_0 = const()[name = string("value_cache_internal_tensor_assign_20_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_20_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_20_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_20_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_20_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_20_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_20_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_39_cast_fp16 = transpose(perm = v_state_39_perm_0, x = var_3669_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_20_cast_fp16 = slice_update(begin = concat_366, begin_mask = value_cache_internal_tensor_assign_20_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_20_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_20_squeeze_mask_0, stride = value_cache_internal_tensor_assign_20_stride_0, update = v_state_39_cast_fp16, x = coreml_update_state_93)[name = string("value_cache_internal_tensor_assign_20_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_20_cast_fp16, input = value_cache)[name = string("coreml_update_state_95_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_95 = read_state(input = value_cache)[name = string("coreml_update_state_95")];
            tensor<int32, [5]> var_3726_begin_0 = const()[name = string("op_3726_begin_0"), val = tensor<int32, [5]>([19, 0, 0, 0, 0])];
            tensor<int32, [5]> var_3726_end_0 = const()[name = string("op_3726_end_0"), val = tensor<int32, [5]>([20, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_3726_end_mask_0 = const()[name = string("op_3726_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_3726_squeeze_mask_0 = const()[name = string("op_3726_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_3726_cast_fp16 = slice_by_index(begin = var_3726_begin_0, end = var_3726_end_0, end_mask = var_3726_end_mask_0, squeeze_mask = var_3726_squeeze_mask_0, x = coreml_update_state_94)[name = string("op_3726_cast_fp16")];
            tensor<int32, [4]> var_3729_begin_0 = const()[name = string("op_3729_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_3729_end_mask_0 = const()[name = string("op_3729_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_3729_cast_fp16 = slice_by_index(begin = var_3729_begin_0, end = concat_11, end_mask = var_3729_end_mask_0, x = var_3726_cast_fp16)[name = string("op_3729_cast_fp16")];
            tensor<int32, [5]> var_3731_begin_0 = const()[name = string("op_3731_begin_0"), val = tensor<int32, [5]>([19, 0, 0, 0, 0])];
            tensor<int32, [5]> var_3731_end_0 = const()[name = string("op_3731_end_0"), val = tensor<int32, [5]>([20, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_3731_end_mask_0 = const()[name = string("op_3731_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_3731_squeeze_mask_0 = const()[name = string("op_3731_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_3731_cast_fp16 = slice_by_index(begin = var_3731_begin_0, end = var_3731_end_0, end_mask = var_3731_end_mask_0, squeeze_mask = var_3731_squeeze_mask_0, x = coreml_update_state_95)[name = string("op_3731_cast_fp16")];
            tensor<int32, [4]> var_3734_begin_0 = const()[name = string("op_3734_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_3734_end_mask_0 = const()[name = string("op_3734_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_3734_cast_fp16 = slice_by_index(begin = var_3734_begin_0, end = concat_11, end_mask = var_3734_end_mask_0, x = var_3731_cast_fp16)[name = string("op_3734_cast_fp16")];
            tensor<int32, [4]> var_3736_shape_cast_fp16 = shape(x = var_3729_cast_fp16)[name = string("op_3736_shape_cast_fp16")];
            int32 gather_355 = const()[name = string("gather_355"), val = int32(1)];
            int32 gather_356 = const()[name = string("gather_356"), val = int32(2)];
            int32 gather_357_axis_0 = const()[name = string("gather_357_axis_0"), val = int32(0)];
            int32 gather_357_batch_dims_0 = const()[name = string("gather_357_batch_dims_0"), val = int32(0)];
            bool gather_357_validate_indices_0 = const()[name = string("gather_357_validate_indices_0"), val = bool(false)];
            string var_3736_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3736_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_357_to_uint16 = const()[name = string("select_357_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3736_shape_cast_fp16_to_uint16 = cast(dtype = var_3736_shape_cast_fp16_to_uint16_dtype_0, x = var_3736_shape_cast_fp16)[name = string("cast_53")];
            uint16 gather_357_cast_uint16 = gather(axis = gather_357_axis_0, batch_dims = gather_357_batch_dims_0, indices = select_357_to_uint16, validate_indices = gather_357_validate_indices_0, x = var_3736_shape_cast_fp16_to_uint16)[name = string("gather_357_cast_uint16")];
            string gather_357_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_357_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_358 = const()[name = string("gather_358"), val = int32(128)];
            tensor<int32, [1]> var_3743_axes_0 = const()[name = string("op_3743_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_3743_cast_fp16 = expand_dims(axes = var_3743_axes_0, x = var_3729_cast_fp16)[name = string("op_3743_cast_fp16")];
            tensor<int32, [5]> shape_397_cast_fp16 = shape(x = var_3743_cast_fp16)[name = string("shape_397_cast_fp16")];
            int32 concat_374_axis_0 = const()[name = string("concat_374_axis_0"), val = int32(0)];
            bool concat_374_interleave_0 = const()[name = string("concat_374_interleave_0"), val = bool(false)];
            int32 gather_357_cast_uint16_to_int32 = cast(dtype = gather_357_cast_uint16_to_int32_dtype_0, x = gather_357_cast_uint16)[name = string("cast_52")];
            tensor<int32, [5]> concat_374 = concat(axis = concat_374_axis_0, interleave = concat_374_interleave_0, values = (gather_355, gather_356, var_77, gather_357_cast_uint16_to_int32, gather_358))[name = string("concat_374")];
            tensor<int32, [5]> real_div_38 = real_div(x = concat_374, y = shape_397_cast_fp16)[name = string("real_div_38")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_581_cast_fp16 = tile(reps = real_div_38, x = var_3743_cast_fp16)[name = string("hidden_states_581_cast_fp16")];
            tensor<int32, [4]> concat_375x = const()[name = string("concat_375x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_79_cast_fp16 = reshape(shape = concat_375x, x = hidden_states_581_cast_fp16)[name = string("key_states_79_cast_fp16")];
            tensor<int32, [4]> var_3753_shape_cast_fp16 = shape(x = var_3734_cast_fp16)[name = string("op_3753_shape_cast_fp16")];
            int32 gather_359 = const()[name = string("gather_359"), val = int32(1)];
            int32 gather_360 = const()[name = string("gather_360"), val = int32(2)];
            int32 gather_361_axis_0 = const()[name = string("gather_361_axis_0"), val = int32(0)];
            int32 gather_361_batch_dims_0 = const()[name = string("gather_361_batch_dims_0"), val = int32(0)];
            bool gather_361_validate_indices_0 = const()[name = string("gather_361_validate_indices_0"), val = bool(false)];
            string var_3753_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3753_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_361_to_uint16 = const()[name = string("select_361_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3753_shape_cast_fp16_to_uint16 = cast(dtype = var_3753_shape_cast_fp16_to_uint16_dtype_0, x = var_3753_shape_cast_fp16)[name = string("cast_51")];
            uint16 gather_361_cast_uint16 = gather(axis = gather_361_axis_0, batch_dims = gather_361_batch_dims_0, indices = select_361_to_uint16, validate_indices = gather_361_validate_indices_0, x = var_3753_shape_cast_fp16_to_uint16)[name = string("gather_361_cast_uint16")];
            string gather_361_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_361_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_362 = const()[name = string("gather_362"), val = int32(128)];
            tensor<int32, [1]> var_3760_axes_0 = const()[name = string("op_3760_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_3760_cast_fp16 = expand_dims(axes = var_3760_axes_0, x = var_3734_cast_fp16)[name = string("op_3760_cast_fp16")];
            tensor<int32, [5]> shape_402_cast_fp16 = shape(x = var_3760_cast_fp16)[name = string("shape_402_cast_fp16")];
            int32 concat_376_axis_0 = const()[name = string("concat_376_axis_0"), val = int32(0)];
            bool concat_376_interleave_0 = const()[name = string("concat_376_interleave_0"), val = bool(false)];
            int32 gather_361_cast_uint16_to_int32 = cast(dtype = gather_361_cast_uint16_to_int32_dtype_0, x = gather_361_cast_uint16)[name = string("cast_50")];
            tensor<int32, [5]> concat_376 = concat(axis = concat_376_axis_0, interleave = concat_376_interleave_0, values = (gather_359, gather_360, var_77, gather_361_cast_uint16_to_int32, gather_362))[name = string("concat_376")];
            tensor<int32, [5]> real_div_39 = real_div(x = concat_376, y = shape_402_cast_fp16)[name = string("real_div_39")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_585_cast_fp16 = tile(reps = real_div_39, x = var_3760_cast_fp16)[name = string("hidden_states_585_cast_fp16")];
            tensor<int32, [4]> concat_377x = const()[name = string("concat_377x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_79_cast_fp16 = reshape(shape = concat_377x, x = hidden_states_585_cast_fp16)[name = string("value_states_79_cast_fp16")];
            tensor<int32, [4]> var_3770_shape_cast_fp16 = shape(x = key_states_79_cast_fp16)[name = string("op_3770_shape_cast_fp16")];
            int32 gather_363_axis_0 = const()[name = string("gather_363_axis_0"), val = int32(0)];
            int32 gather_363_batch_dims_0 = const()[name = string("gather_363_batch_dims_0"), val = int32(0)];
            bool gather_363_validate_indices_0 = const()[name = string("gather_363_validate_indices_0"), val = bool(false)];
            string var_3770_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3770_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_363_to_uint16 = const()[name = string("select_363_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3770_shape_cast_fp16_to_uint16 = cast(dtype = var_3770_shape_cast_fp16_to_uint16_dtype_0, x = var_3770_shape_cast_fp16)[name = string("cast_49")];
            uint16 gather_363_cast_uint16 = gather(axis = gather_363_axis_0, batch_dims = gather_363_batch_dims_0, indices = select_363_to_uint16, validate_indices = gather_363_validate_indices_0, x = var_3770_shape_cast_fp16_to_uint16)[name = string("gather_363_cast_uint16")];
            string gather_363_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_363_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_378_values0_0 = const()[name = string("concat_378_values0_0"), val = int32(1)];
            int32 concat_378_values1_0 = const()[name = string("concat_378_values1_0"), val = int32(1)];
            int32 concat_378_values2_0 = const()[name = string("concat_378_values2_0"), val = int32(0)];
            int32 concat_378_axis_0 = const()[name = string("concat_378_axis_0"), val = int32(0)];
            bool concat_378_interleave_0 = const()[name = string("concat_378_interleave_0"), val = bool(false)];
            int32 gather_363_cast_uint16_to_int32 = cast(dtype = gather_363_cast_uint16_to_int32_dtype_0, x = gather_363_cast_uint16)[name = string("cast_48")];
            tensor<int32, [4]> concat_378 = concat(axis = concat_378_axis_0, interleave = concat_378_interleave_0, values = (concat_378_values0_0, concat_378_values1_0, concat_378_values2_0, gather_363_cast_uint16_to_int32))[name = string("concat_378")];
            tensor<int32, [4]> causal_mask_41_begin_0 = const()[name = string("causal_mask_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_41_end_mask_0 = const()[name = string("causal_mask_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_41_cast_fp16 = slice_by_index(begin = causal_mask_41_begin_0, end = concat_378, end_mask = causal_mask_41_end_mask_0, x = causal_mask)[name = string("causal_mask_41_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_77_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_41_cast_fp16, key = key_states_79_cast_fp16, query = query_states_79_cast_fp16, value = value_states_79_cast_fp16)[name = string("attn_output_77_cast_fp16")];
            tensor<int32, [4]> var_3776_perm_0 = const()[name = string("op_3776_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_379x = const()[name = string("concat_379x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_3776_cast_fp16 = transpose(perm = var_3776_perm_0, x = attn_output_77_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, ?, 1536]> input_153_cast_fp16 = reshape(shape = concat_379x, x = var_3776_cast_fp16)[name = string("input_153_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633386176))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(634565888))))[name = string("model_model_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_136_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_589_cast_fp16 = add(x = hidden_states_569_cast_fp16, y = linear_136_cast_fp16)[name = string("hidden_states_589_cast_fp16")];
            fp16 var_78_promoted_39_to_fp16 = const()[name = string("op_78_promoted_39_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_3785_cast_fp16 = pow(x = hidden_states_589_cast_fp16, y = var_78_promoted_39_to_fp16)[name = string("op_3785_cast_fp16")];
            tensor<int32, [1]> variance_79_axes_0 = const()[name = string("variance_79_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_79_keep_dims_0 = const()[name = string("variance_79_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_79_cast_fp16 = reduce_mean(axes = variance_79_axes_0, keep_dims = variance_79_keep_dims_0, x = var_3785_cast_fp16)[name = string("variance_79_cast_fp16")];
            fp16 var_3788_to_fp16 = const()[name = string("op_3788_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3789_cast_fp16 = add(x = variance_79_cast_fp16, y = var_3788_to_fp16)[name = string("op_3789_cast_fp16")];
            fp32 var_3790_epsilon_0 = const()[name = string("op_3790_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_3790_cast_fp16 = rsqrt(epsilon = var_3790_epsilon_0, x = var_3789_cast_fp16)[name = string("op_3790_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_593_cast_fp16 = mul(x = hidden_states_589_cast_fp16, y = var_3790_cast_fp16)[name = string("hidden_states_593_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(634713408)))];
            tensor<fp16, [1, ?, 1536]> input_155_cast_fp16 = mul(x = model_model_layers_19_post_attention_layernorm_weight_to_fp16, y = hidden_states_593_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(634716544))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(641597888))))[name = string("model_model_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_3802_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_3802_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642458112))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(649339456))))[name = string("model_model_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_159_cast_fp16 = mul(x = var_3802_cast_fp16, y = linear_138_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(650199680))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657081024))))[name = string("model_model_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_139_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_599_cast_fp16 = add(x = hidden_states_589_cast_fp16, y = linear_139_cast_fp16)[name = string("hidden_states_599_cast_fp16")];
            fp16 var_78_promoted_40_to_fp16 = const()[name = string("op_78_promoted_40_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_3815_cast_fp16 = pow(x = hidden_states_599_cast_fp16, y = var_78_promoted_40_to_fp16)[name = string("op_3815_cast_fp16")];
            tensor<int32, [1]> variance_81_axes_0 = const()[name = string("variance_81_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_81_keep_dims_0 = const()[name = string("variance_81_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_81_cast_fp16 = reduce_mean(axes = variance_81_axes_0, keep_dims = variance_81_keep_dims_0, x = var_3815_cast_fp16)[name = string("variance_81_cast_fp16")];
            fp16 var_3818_to_fp16 = const()[name = string("op_3818_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3819_cast_fp16 = add(x = variance_81_cast_fp16, y = var_3818_to_fp16)[name = string("op_3819_cast_fp16")];
            fp32 var_3820_epsilon_0 = const()[name = string("op_3820_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_3820_cast_fp16 = rsqrt(epsilon = var_3820_epsilon_0, x = var_3819_cast_fp16)[name = string("op_3820_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_603_cast_fp16 = mul(x = hidden_states_599_cast_fp16, y = var_3820_cast_fp16)[name = string("hidden_states_603_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657941248)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_607_cast_fp16 = mul(x = model_model_layers_20_input_layernorm_weight_to_fp16, y = hidden_states_603_cast_fp16)[name = string("hidden_states_607_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657944384))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659124096))))[name = string("model_model_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659271616)))];
            tensor<fp16, [1, ?, 1536]> linear_140_cast_fp16 = linear(bias = model_model_layers_20_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_607_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659274752))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659471424))))[name = string("model_model_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_20_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_20_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659496064)))];
            tensor<fp16, [1, ?, 256]> linear_141_cast_fp16 = linear(bias = model_model_layers_20_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_607_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659496640))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659693312))))[name = string("model_model_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659717952)))];
            tensor<fp16, [1, ?, 256]> linear_142_cast_fp16 = linear(bias = model_model_layers_20_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_607_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<int32, [4]> concat_380x = const()[name = string("concat_380x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_3843_cast_fp16 = reshape(shape = concat_380x, x = linear_140_cast_fp16)[name = string("op_3843_cast_fp16")];
            tensor<int32, [4]> q_41_perm_0 = const()[name = string("q_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_381x = const()[name = string("concat_381x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_3846_cast_fp16 = reshape(shape = concat_381x, x = linear_141_cast_fp16)[name = string("op_3846_cast_fp16")];
            tensor<int32, [4]> k_41_perm_0 = const()[name = string("k_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_382x = const()[name = string("concat_382x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_3849_cast_fp16 = reshape(shape = concat_382x, x = linear_142_cast_fp16)[name = string("op_3849_cast_fp16")];
            tensor<int32, [4]> v_state_41_perm_0 = const()[name = string("v_state_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_41_cast_fp16 = transpose(perm = q_41_perm_0, x = var_3843_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 12, ?, 128]> var_3853_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_7_cast_fp16)[name = string("op_3853_cast_fp16")];
            tensor<int32, [4]> x1_81_begin_0 = const()[name = string("x1_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_81_end_0 = const()[name = string("x1_81_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_81_end_mask_0 = const()[name = string("x1_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_81_cast_fp16 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = q_41_cast_fp16)[name = string("x1_81_cast_fp16")];
            tensor<int32, [4]> x2_81_begin_0 = const()[name = string("x2_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_81_end_0 = const()[name = string("x2_81_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_81_end_mask_0 = const()[name = string("x2_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_81_cast_fp16 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = q_41_cast_fp16)[name = string("x2_81_cast_fp16")];
            fp16 const_43_promoted_to_fp16 = const()[name = string("const_43_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_3864_cast_fp16 = mul(x = x2_81_cast_fp16, y = const_43_promoted_to_fp16)[name = string("op_3864_cast_fp16")];
            bool var_3866_interleave_0 = const()[name = string("op_3866_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_3866_cast_fp16 = concat(axis = var_72, interleave = var_3866_interleave_0, values = (var_3864_cast_fp16, x1_81_cast_fp16))[name = string("op_3866_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_3867_cast_fp16 = mul(x = var_3866_cast_fp16, y = sin_7_cast_fp16)[name = string("op_3867_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_83_cast_fp16 = add(x = var_3853_cast_fp16, y = var_3867_cast_fp16)[name = string("query_states_83_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_41_cast_fp16 = transpose(perm = k_41_perm_0, x = var_3846_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 2, ?, 128]> var_3869_cast_fp16 = mul(x = k_41_cast_fp16, y = cos_7_cast_fp16)[name = string("op_3869_cast_fp16")];
            tensor<int32, [4]> x1_83_begin_0 = const()[name = string("x1_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_83_end_0 = const()[name = string("x1_83_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_83_end_mask_0 = const()[name = string("x1_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_83_cast_fp16 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = k_41_cast_fp16)[name = string("x1_83_cast_fp16")];
            tensor<int32, [4]> x2_83_begin_0 = const()[name = string("x2_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_83_end_0 = const()[name = string("x2_83_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_83_end_mask_0 = const()[name = string("x2_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_83_cast_fp16 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = k_41_cast_fp16)[name = string("x2_83_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_3880_cast_fp16 = mul(x = x2_83_cast_fp16, y = const_44_promoted_to_fp16)[name = string("op_3880_cast_fp16")];
            bool var_3882_interleave_0 = const()[name = string("op_3882_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_3882_cast_fp16 = concat(axis = var_72, interleave = var_3882_interleave_0, values = (var_3880_cast_fp16, x1_83_cast_fp16))[name = string("op_3882_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_3883_cast_fp16 = mul(x = var_3882_cast_fp16, y = sin_7_cast_fp16)[name = string("op_3883_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_41_cast_fp16 = add(x = var_3869_cast_fp16, y = var_3883_cast_fp16)[name = string("k_state_41_cast_fp16")];
            tensor<int32, [1]> expand_dims_240 = const()[name = string("expand_dims_240"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_241 = const()[name = string("expand_dims_241"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_243 = const()[name = string("expand_dims_243"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_385_values0_0 = const()[name = string("concat_385_values0_0"), val = tensor<int32, [1]>([20])];
            int32 concat_385_axis_0 = const()[name = string("concat_385_axis_0"), val = int32(0)];
            bool concat_385_interleave_0 = const()[name = string("concat_385_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_385 = concat(axis = concat_385_axis_0, interleave = concat_385_interleave_0, values = (concat_385_values0_0, expand_dims_240, expand_dims_241, expand_dims_2, expand_dims_243))[name = string("concat_385")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_21_stride_0 = const()[name = string("key_cache_internal_tensor_assign_21_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_21_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_21_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_21_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_21_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_21_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_21_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_21_cast_fp16 = slice_update(begin = concat_385, begin_mask = key_cache_internal_tensor_assign_21_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_21_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_21_squeeze_mask_0, stride = key_cache_internal_tensor_assign_21_stride_0, update = k_state_41_cast_fp16, x = coreml_update_state_94)[name = string("key_cache_internal_tensor_assign_21_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_21_cast_fp16, input = key_cache)[name = string("coreml_update_state_96_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_96 = read_state(input = key_cache)[name = string("coreml_update_state_96")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_21_stride_0 = const()[name = string("value_cache_internal_tensor_assign_21_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_21_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_21_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_21_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_21_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_21_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_21_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_41_cast_fp16 = transpose(perm = v_state_41_perm_0, x = var_3849_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_21_cast_fp16 = slice_update(begin = concat_385, begin_mask = value_cache_internal_tensor_assign_21_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_21_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_21_squeeze_mask_0, stride = value_cache_internal_tensor_assign_21_stride_0, update = v_state_41_cast_fp16, x = coreml_update_state_95)[name = string("value_cache_internal_tensor_assign_21_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_21_cast_fp16, input = value_cache)[name = string("coreml_update_state_97_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_97 = read_state(input = value_cache)[name = string("coreml_update_state_97")];
            tensor<int32, [5]> var_3906_begin_0 = const()[name = string("op_3906_begin_0"), val = tensor<int32, [5]>([20, 0, 0, 0, 0])];
            tensor<int32, [5]> var_3906_end_0 = const()[name = string("op_3906_end_0"), val = tensor<int32, [5]>([21, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_3906_end_mask_0 = const()[name = string("op_3906_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_3906_squeeze_mask_0 = const()[name = string("op_3906_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_3906_cast_fp16 = slice_by_index(begin = var_3906_begin_0, end = var_3906_end_0, end_mask = var_3906_end_mask_0, squeeze_mask = var_3906_squeeze_mask_0, x = coreml_update_state_96)[name = string("op_3906_cast_fp16")];
            tensor<int32, [4]> var_3909_begin_0 = const()[name = string("op_3909_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_3909_end_mask_0 = const()[name = string("op_3909_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_3909_cast_fp16 = slice_by_index(begin = var_3909_begin_0, end = concat_11, end_mask = var_3909_end_mask_0, x = var_3906_cast_fp16)[name = string("op_3909_cast_fp16")];
            tensor<int32, [5]> var_3911_begin_0 = const()[name = string("op_3911_begin_0"), val = tensor<int32, [5]>([20, 0, 0, 0, 0])];
            tensor<int32, [5]> var_3911_end_0 = const()[name = string("op_3911_end_0"), val = tensor<int32, [5]>([21, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_3911_end_mask_0 = const()[name = string("op_3911_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_3911_squeeze_mask_0 = const()[name = string("op_3911_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_3911_cast_fp16 = slice_by_index(begin = var_3911_begin_0, end = var_3911_end_0, end_mask = var_3911_end_mask_0, squeeze_mask = var_3911_squeeze_mask_0, x = coreml_update_state_97)[name = string("op_3911_cast_fp16")];
            tensor<int32, [4]> var_3914_begin_0 = const()[name = string("op_3914_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_3914_end_mask_0 = const()[name = string("op_3914_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_3914_cast_fp16 = slice_by_index(begin = var_3914_begin_0, end = concat_11, end_mask = var_3914_end_mask_0, x = var_3911_cast_fp16)[name = string("op_3914_cast_fp16")];
            tensor<int32, [4]> var_3916_shape_cast_fp16 = shape(x = var_3909_cast_fp16)[name = string("op_3916_shape_cast_fp16")];
            int32 gather_373 = const()[name = string("gather_373"), val = int32(1)];
            int32 gather_374 = const()[name = string("gather_374"), val = int32(2)];
            int32 gather_375_axis_0 = const()[name = string("gather_375_axis_0"), val = int32(0)];
            int32 gather_375_batch_dims_0 = const()[name = string("gather_375_batch_dims_0"), val = int32(0)];
            bool gather_375_validate_indices_0 = const()[name = string("gather_375_validate_indices_0"), val = bool(false)];
            string var_3916_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3916_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_375_to_uint16 = const()[name = string("select_375_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3916_shape_cast_fp16_to_uint16 = cast(dtype = var_3916_shape_cast_fp16_to_uint16_dtype_0, x = var_3916_shape_cast_fp16)[name = string("cast_47")];
            uint16 gather_375_cast_uint16 = gather(axis = gather_375_axis_0, batch_dims = gather_375_batch_dims_0, indices = select_375_to_uint16, validate_indices = gather_375_validate_indices_0, x = var_3916_shape_cast_fp16_to_uint16)[name = string("gather_375_cast_uint16")];
            string gather_375_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_375_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_376 = const()[name = string("gather_376"), val = int32(128)];
            tensor<int32, [1]> var_3923_axes_0 = const()[name = string("op_3923_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_3923_cast_fp16 = expand_dims(axes = var_3923_axes_0, x = var_3909_cast_fp16)[name = string("op_3923_cast_fp16")];
            tensor<int32, [5]> shape_417_cast_fp16 = shape(x = var_3923_cast_fp16)[name = string("shape_417_cast_fp16")];
            int32 concat_393_axis_0 = const()[name = string("concat_393_axis_0"), val = int32(0)];
            bool concat_393_interleave_0 = const()[name = string("concat_393_interleave_0"), val = bool(false)];
            int32 gather_375_cast_uint16_to_int32 = cast(dtype = gather_375_cast_uint16_to_int32_dtype_0, x = gather_375_cast_uint16)[name = string("cast_46")];
            tensor<int32, [5]> concat_393 = concat(axis = concat_393_axis_0, interleave = concat_393_interleave_0, values = (gather_373, gather_374, var_77, gather_375_cast_uint16_to_int32, gather_376))[name = string("concat_393")];
            tensor<int32, [5]> real_div_40 = real_div(x = concat_393, y = shape_417_cast_fp16)[name = string("real_div_40")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_611_cast_fp16 = tile(reps = real_div_40, x = var_3923_cast_fp16)[name = string("hidden_states_611_cast_fp16")];
            tensor<int32, [4]> concat_394x = const()[name = string("concat_394x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_83_cast_fp16 = reshape(shape = concat_394x, x = hidden_states_611_cast_fp16)[name = string("key_states_83_cast_fp16")];
            tensor<int32, [4]> var_3933_shape_cast_fp16 = shape(x = var_3914_cast_fp16)[name = string("op_3933_shape_cast_fp16")];
            int32 gather_377 = const()[name = string("gather_377"), val = int32(1)];
            int32 gather_378 = const()[name = string("gather_378"), val = int32(2)];
            int32 gather_379_axis_0 = const()[name = string("gather_379_axis_0"), val = int32(0)];
            int32 gather_379_batch_dims_0 = const()[name = string("gather_379_batch_dims_0"), val = int32(0)];
            bool gather_379_validate_indices_0 = const()[name = string("gather_379_validate_indices_0"), val = bool(false)];
            string var_3933_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3933_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_379_to_uint16 = const()[name = string("select_379_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3933_shape_cast_fp16_to_uint16 = cast(dtype = var_3933_shape_cast_fp16_to_uint16_dtype_0, x = var_3933_shape_cast_fp16)[name = string("cast_45")];
            uint16 gather_379_cast_uint16 = gather(axis = gather_379_axis_0, batch_dims = gather_379_batch_dims_0, indices = select_379_to_uint16, validate_indices = gather_379_validate_indices_0, x = var_3933_shape_cast_fp16_to_uint16)[name = string("gather_379_cast_uint16")];
            string gather_379_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_379_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_380 = const()[name = string("gather_380"), val = int32(128)];
            tensor<int32, [1]> var_3940_axes_0 = const()[name = string("op_3940_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_3940_cast_fp16 = expand_dims(axes = var_3940_axes_0, x = var_3914_cast_fp16)[name = string("op_3940_cast_fp16")];
            tensor<int32, [5]> shape_422_cast_fp16 = shape(x = var_3940_cast_fp16)[name = string("shape_422_cast_fp16")];
            int32 concat_395_axis_0 = const()[name = string("concat_395_axis_0"), val = int32(0)];
            bool concat_395_interleave_0 = const()[name = string("concat_395_interleave_0"), val = bool(false)];
            int32 gather_379_cast_uint16_to_int32 = cast(dtype = gather_379_cast_uint16_to_int32_dtype_0, x = gather_379_cast_uint16)[name = string("cast_44")];
            tensor<int32, [5]> concat_395 = concat(axis = concat_395_axis_0, interleave = concat_395_interleave_0, values = (gather_377, gather_378, var_77, gather_379_cast_uint16_to_int32, gather_380))[name = string("concat_395")];
            tensor<int32, [5]> real_div_41 = real_div(x = concat_395, y = shape_422_cast_fp16)[name = string("real_div_41")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_615_cast_fp16 = tile(reps = real_div_41, x = var_3940_cast_fp16)[name = string("hidden_states_615_cast_fp16")];
            tensor<int32, [4]> concat_396x = const()[name = string("concat_396x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_83_cast_fp16 = reshape(shape = concat_396x, x = hidden_states_615_cast_fp16)[name = string("value_states_83_cast_fp16")];
            tensor<int32, [4]> var_3950_shape_cast_fp16 = shape(x = key_states_83_cast_fp16)[name = string("op_3950_shape_cast_fp16")];
            int32 gather_381_axis_0 = const()[name = string("gather_381_axis_0"), val = int32(0)];
            int32 gather_381_batch_dims_0 = const()[name = string("gather_381_batch_dims_0"), val = int32(0)];
            bool gather_381_validate_indices_0 = const()[name = string("gather_381_validate_indices_0"), val = bool(false)];
            string var_3950_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3950_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_381_to_uint16 = const()[name = string("select_381_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_3950_shape_cast_fp16_to_uint16 = cast(dtype = var_3950_shape_cast_fp16_to_uint16_dtype_0, x = var_3950_shape_cast_fp16)[name = string("cast_43")];
            uint16 gather_381_cast_uint16 = gather(axis = gather_381_axis_0, batch_dims = gather_381_batch_dims_0, indices = select_381_to_uint16, validate_indices = gather_381_validate_indices_0, x = var_3950_shape_cast_fp16_to_uint16)[name = string("gather_381_cast_uint16")];
            string gather_381_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_381_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_397_values0_0 = const()[name = string("concat_397_values0_0"), val = int32(1)];
            int32 concat_397_values1_0 = const()[name = string("concat_397_values1_0"), val = int32(1)];
            int32 concat_397_values2_0 = const()[name = string("concat_397_values2_0"), val = int32(0)];
            int32 concat_397_axis_0 = const()[name = string("concat_397_axis_0"), val = int32(0)];
            bool concat_397_interleave_0 = const()[name = string("concat_397_interleave_0"), val = bool(false)];
            int32 gather_381_cast_uint16_to_int32 = cast(dtype = gather_381_cast_uint16_to_int32_dtype_0, x = gather_381_cast_uint16)[name = string("cast_42")];
            tensor<int32, [4]> concat_397 = concat(axis = concat_397_axis_0, interleave = concat_397_interleave_0, values = (concat_397_values0_0, concat_397_values1_0, concat_397_values2_0, gather_381_cast_uint16_to_int32))[name = string("concat_397")];
            tensor<int32, [4]> causal_mask_43_begin_0 = const()[name = string("causal_mask_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_43_end_mask_0 = const()[name = string("causal_mask_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_43_cast_fp16 = slice_by_index(begin = causal_mask_43_begin_0, end = concat_397, end_mask = causal_mask_43_end_mask_0, x = causal_mask)[name = string("causal_mask_43_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_81_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_43_cast_fp16, key = key_states_83_cast_fp16, query = query_states_83_cast_fp16, value = value_states_83_cast_fp16)[name = string("attn_output_81_cast_fp16")];
            tensor<int32, [4]> var_3956_perm_0 = const()[name = string("op_3956_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_398x = const()[name = string("concat_398x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_3956_cast_fp16 = transpose(perm = var_3956_perm_0, x = attn_output_81_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, ?, 1536]> input_161_cast_fp16 = reshape(shape = concat_398x, x = var_3956_cast_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659718528))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660898240))))[name = string("model_model_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_143_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_619_cast_fp16 = add(x = hidden_states_599_cast_fp16, y = linear_143_cast_fp16)[name = string("hidden_states_619_cast_fp16")];
            fp16 var_78_promoted_41_to_fp16 = const()[name = string("op_78_promoted_41_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_3965_cast_fp16 = pow(x = hidden_states_619_cast_fp16, y = var_78_promoted_41_to_fp16)[name = string("op_3965_cast_fp16")];
            tensor<int32, [1]> variance_83_axes_0 = const()[name = string("variance_83_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_83_keep_dims_0 = const()[name = string("variance_83_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_83_cast_fp16 = reduce_mean(axes = variance_83_axes_0, keep_dims = variance_83_keep_dims_0, x = var_3965_cast_fp16)[name = string("variance_83_cast_fp16")];
            fp16 var_3968_to_fp16 = const()[name = string("op_3968_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3969_cast_fp16 = add(x = variance_83_cast_fp16, y = var_3968_to_fp16)[name = string("op_3969_cast_fp16")];
            fp32 var_3970_epsilon_0 = const()[name = string("op_3970_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_3970_cast_fp16 = rsqrt(epsilon = var_3970_epsilon_0, x = var_3969_cast_fp16)[name = string("op_3970_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_623_cast_fp16 = mul(x = hidden_states_619_cast_fp16, y = var_3970_cast_fp16)[name = string("hidden_states_623_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661045760)))];
            tensor<fp16, [1, ?, 1536]> input_163_cast_fp16 = mul(x = model_model_layers_20_post_attention_layernorm_weight_to_fp16, y = hidden_states_623_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661048896))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667930240))))[name = string("model_model_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_144_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_3982_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_3982_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(668790464))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675671808))))[name = string("model_model_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_167_cast_fp16 = mul(x = var_3982_cast_fp16, y = linear_145_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676532032))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683413376))))[name = string("model_model_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_146_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_629_cast_fp16 = add(x = hidden_states_619_cast_fp16, y = linear_146_cast_fp16)[name = string("hidden_states_629_cast_fp16")];
            fp16 var_78_promoted_42_to_fp16 = const()[name = string("op_78_promoted_42_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_3995_cast_fp16 = pow(x = hidden_states_629_cast_fp16, y = var_78_promoted_42_to_fp16)[name = string("op_3995_cast_fp16")];
            tensor<int32, [1]> variance_85_axes_0 = const()[name = string("variance_85_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_85_keep_dims_0 = const()[name = string("variance_85_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_85_cast_fp16 = reduce_mean(axes = variance_85_axes_0, keep_dims = variance_85_keep_dims_0, x = var_3995_cast_fp16)[name = string("variance_85_cast_fp16")];
            fp16 var_3998_to_fp16 = const()[name = string("op_3998_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3999_cast_fp16 = add(x = variance_85_cast_fp16, y = var_3998_to_fp16)[name = string("op_3999_cast_fp16")];
            fp32 var_4000_epsilon_0 = const()[name = string("op_4000_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_4000_cast_fp16 = rsqrt(epsilon = var_4000_epsilon_0, x = var_3999_cast_fp16)[name = string("op_4000_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_633_cast_fp16 = mul(x = hidden_states_629_cast_fp16, y = var_4000_cast_fp16)[name = string("hidden_states_633_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(684273600)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_637_cast_fp16 = mul(x = model_model_layers_21_input_layernorm_weight_to_fp16, y = hidden_states_633_cast_fp16)[name = string("hidden_states_637_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(684276736))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(685456448))))[name = string("model_model_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(685603968)))];
            tensor<fp16, [1, ?, 1536]> linear_147_cast_fp16 = linear(bias = model_model_layers_21_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_637_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(685607104))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(685803776))))[name = string("model_model_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_21_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_21_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(685828416)))];
            tensor<fp16, [1, ?, 256]> linear_148_cast_fp16 = linear(bias = model_model_layers_21_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_637_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(685828992))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686025664))))[name = string("model_model_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686050304)))];
            tensor<fp16, [1, ?, 256]> linear_149_cast_fp16 = linear(bias = model_model_layers_21_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_637_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> concat_399x = const()[name = string("concat_399x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_4023_cast_fp16 = reshape(shape = concat_399x, x = linear_147_cast_fp16)[name = string("op_4023_cast_fp16")];
            tensor<int32, [4]> q_43_perm_0 = const()[name = string("q_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_400x = const()[name = string("concat_400x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_4026_cast_fp16 = reshape(shape = concat_400x, x = linear_148_cast_fp16)[name = string("op_4026_cast_fp16")];
            tensor<int32, [4]> k_43_perm_0 = const()[name = string("k_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_401x = const()[name = string("concat_401x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_4029_cast_fp16 = reshape(shape = concat_401x, x = linear_149_cast_fp16)[name = string("op_4029_cast_fp16")];
            tensor<int32, [4]> v_state_43_perm_0 = const()[name = string("v_state_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_43_cast_fp16 = transpose(perm = q_43_perm_0, x = var_4023_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 12, ?, 128]> var_4033_cast_fp16 = mul(x = q_43_cast_fp16, y = cos_7_cast_fp16)[name = string("op_4033_cast_fp16")];
            tensor<int32, [4]> x1_85_begin_0 = const()[name = string("x1_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_85_end_0 = const()[name = string("x1_85_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_85_end_mask_0 = const()[name = string("x1_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_85_cast_fp16 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = q_43_cast_fp16)[name = string("x1_85_cast_fp16")];
            tensor<int32, [4]> x2_85_begin_0 = const()[name = string("x2_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_85_end_0 = const()[name = string("x2_85_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_85_end_mask_0 = const()[name = string("x2_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_85_cast_fp16 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = q_43_cast_fp16)[name = string("x2_85_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_4044_cast_fp16 = mul(x = x2_85_cast_fp16, y = const_45_promoted_to_fp16)[name = string("op_4044_cast_fp16")];
            bool var_4046_interleave_0 = const()[name = string("op_4046_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_4046_cast_fp16 = concat(axis = var_72, interleave = var_4046_interleave_0, values = (var_4044_cast_fp16, x1_85_cast_fp16))[name = string("op_4046_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_4047_cast_fp16 = mul(x = var_4046_cast_fp16, y = sin_7_cast_fp16)[name = string("op_4047_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_87_cast_fp16 = add(x = var_4033_cast_fp16, y = var_4047_cast_fp16)[name = string("query_states_87_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_43_cast_fp16 = transpose(perm = k_43_perm_0, x = var_4026_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 2, ?, 128]> var_4049_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_7_cast_fp16)[name = string("op_4049_cast_fp16")];
            tensor<int32, [4]> x1_87_begin_0 = const()[name = string("x1_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_87_end_0 = const()[name = string("x1_87_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_87_end_mask_0 = const()[name = string("x1_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_87_cast_fp16 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = k_43_cast_fp16)[name = string("x1_87_cast_fp16")];
            tensor<int32, [4]> x2_87_begin_0 = const()[name = string("x2_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_87_end_0 = const()[name = string("x2_87_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_87_end_mask_0 = const()[name = string("x2_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_87_cast_fp16 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = k_43_cast_fp16)[name = string("x2_87_cast_fp16")];
            fp16 const_46_promoted_to_fp16 = const()[name = string("const_46_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_4060_cast_fp16 = mul(x = x2_87_cast_fp16, y = const_46_promoted_to_fp16)[name = string("op_4060_cast_fp16")];
            bool var_4062_interleave_0 = const()[name = string("op_4062_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_4062_cast_fp16 = concat(axis = var_72, interleave = var_4062_interleave_0, values = (var_4060_cast_fp16, x1_87_cast_fp16))[name = string("op_4062_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_4063_cast_fp16 = mul(x = var_4062_cast_fp16, y = sin_7_cast_fp16)[name = string("op_4063_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_43_cast_fp16 = add(x = var_4049_cast_fp16, y = var_4063_cast_fp16)[name = string("k_state_43_cast_fp16")];
            tensor<int32, [1]> expand_dims_252 = const()[name = string("expand_dims_252"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_253 = const()[name = string("expand_dims_253"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_255 = const()[name = string("expand_dims_255"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_404_values0_0 = const()[name = string("concat_404_values0_0"), val = tensor<int32, [1]>([21])];
            int32 concat_404_axis_0 = const()[name = string("concat_404_axis_0"), val = int32(0)];
            bool concat_404_interleave_0 = const()[name = string("concat_404_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_404 = concat(axis = concat_404_axis_0, interleave = concat_404_interleave_0, values = (concat_404_values0_0, expand_dims_252, expand_dims_253, expand_dims_2, expand_dims_255))[name = string("concat_404")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_22_stride_0 = const()[name = string("key_cache_internal_tensor_assign_22_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_22_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_22_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_22_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_22_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_22_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_22_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_22_cast_fp16 = slice_update(begin = concat_404, begin_mask = key_cache_internal_tensor_assign_22_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_22_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_22_squeeze_mask_0, stride = key_cache_internal_tensor_assign_22_stride_0, update = k_state_43_cast_fp16, x = coreml_update_state_96)[name = string("key_cache_internal_tensor_assign_22_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_22_cast_fp16, input = key_cache)[name = string("coreml_update_state_98_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_98 = read_state(input = key_cache)[name = string("coreml_update_state_98")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_22_stride_0 = const()[name = string("value_cache_internal_tensor_assign_22_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_22_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_22_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_22_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_22_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_22_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_22_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_43_cast_fp16 = transpose(perm = v_state_43_perm_0, x = var_4029_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_22_cast_fp16 = slice_update(begin = concat_404, begin_mask = value_cache_internal_tensor_assign_22_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_22_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_22_squeeze_mask_0, stride = value_cache_internal_tensor_assign_22_stride_0, update = v_state_43_cast_fp16, x = coreml_update_state_97)[name = string("value_cache_internal_tensor_assign_22_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_22_cast_fp16, input = value_cache)[name = string("coreml_update_state_99_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_99 = read_state(input = value_cache)[name = string("coreml_update_state_99")];
            tensor<int32, [5]> var_4086_begin_0 = const()[name = string("op_4086_begin_0"), val = tensor<int32, [5]>([21, 0, 0, 0, 0])];
            tensor<int32, [5]> var_4086_end_0 = const()[name = string("op_4086_end_0"), val = tensor<int32, [5]>([22, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_4086_end_mask_0 = const()[name = string("op_4086_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_4086_squeeze_mask_0 = const()[name = string("op_4086_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_4086_cast_fp16 = slice_by_index(begin = var_4086_begin_0, end = var_4086_end_0, end_mask = var_4086_end_mask_0, squeeze_mask = var_4086_squeeze_mask_0, x = coreml_update_state_98)[name = string("op_4086_cast_fp16")];
            tensor<int32, [4]> var_4089_begin_0 = const()[name = string("op_4089_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_4089_end_mask_0 = const()[name = string("op_4089_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_4089_cast_fp16 = slice_by_index(begin = var_4089_begin_0, end = concat_11, end_mask = var_4089_end_mask_0, x = var_4086_cast_fp16)[name = string("op_4089_cast_fp16")];
            tensor<int32, [5]> var_4091_begin_0 = const()[name = string("op_4091_begin_0"), val = tensor<int32, [5]>([21, 0, 0, 0, 0])];
            tensor<int32, [5]> var_4091_end_0 = const()[name = string("op_4091_end_0"), val = tensor<int32, [5]>([22, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_4091_end_mask_0 = const()[name = string("op_4091_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_4091_squeeze_mask_0 = const()[name = string("op_4091_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_4091_cast_fp16 = slice_by_index(begin = var_4091_begin_0, end = var_4091_end_0, end_mask = var_4091_end_mask_0, squeeze_mask = var_4091_squeeze_mask_0, x = coreml_update_state_99)[name = string("op_4091_cast_fp16")];
            tensor<int32, [4]> var_4094_begin_0 = const()[name = string("op_4094_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_4094_end_mask_0 = const()[name = string("op_4094_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_4094_cast_fp16 = slice_by_index(begin = var_4094_begin_0, end = concat_11, end_mask = var_4094_end_mask_0, x = var_4091_cast_fp16)[name = string("op_4094_cast_fp16")];
            tensor<int32, [4]> var_4096_shape_cast_fp16 = shape(x = var_4089_cast_fp16)[name = string("op_4096_shape_cast_fp16")];
            int32 gather_391 = const()[name = string("gather_391"), val = int32(1)];
            int32 gather_392 = const()[name = string("gather_392"), val = int32(2)];
            int32 gather_393_axis_0 = const()[name = string("gather_393_axis_0"), val = int32(0)];
            int32 gather_393_batch_dims_0 = const()[name = string("gather_393_batch_dims_0"), val = int32(0)];
            bool gather_393_validate_indices_0 = const()[name = string("gather_393_validate_indices_0"), val = bool(false)];
            string var_4096_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4096_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_393_to_uint16 = const()[name = string("select_393_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_4096_shape_cast_fp16_to_uint16 = cast(dtype = var_4096_shape_cast_fp16_to_uint16_dtype_0, x = var_4096_shape_cast_fp16)[name = string("cast_41")];
            uint16 gather_393_cast_uint16 = gather(axis = gather_393_axis_0, batch_dims = gather_393_batch_dims_0, indices = select_393_to_uint16, validate_indices = gather_393_validate_indices_0, x = var_4096_shape_cast_fp16_to_uint16)[name = string("gather_393_cast_uint16")];
            string gather_393_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_393_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_394 = const()[name = string("gather_394"), val = int32(128)];
            tensor<int32, [1]> var_4103_axes_0 = const()[name = string("op_4103_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_4103_cast_fp16 = expand_dims(axes = var_4103_axes_0, x = var_4089_cast_fp16)[name = string("op_4103_cast_fp16")];
            tensor<int32, [5]> shape_437_cast_fp16 = shape(x = var_4103_cast_fp16)[name = string("shape_437_cast_fp16")];
            int32 concat_412_axis_0 = const()[name = string("concat_412_axis_0"), val = int32(0)];
            bool concat_412_interleave_0 = const()[name = string("concat_412_interleave_0"), val = bool(false)];
            int32 gather_393_cast_uint16_to_int32 = cast(dtype = gather_393_cast_uint16_to_int32_dtype_0, x = gather_393_cast_uint16)[name = string("cast_40")];
            tensor<int32, [5]> concat_412 = concat(axis = concat_412_axis_0, interleave = concat_412_interleave_0, values = (gather_391, gather_392, var_77, gather_393_cast_uint16_to_int32, gather_394))[name = string("concat_412")];
            tensor<int32, [5]> real_div_42 = real_div(x = concat_412, y = shape_437_cast_fp16)[name = string("real_div_42")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_641_cast_fp16 = tile(reps = real_div_42, x = var_4103_cast_fp16)[name = string("hidden_states_641_cast_fp16")];
            tensor<int32, [4]> concat_413x = const()[name = string("concat_413x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_87_cast_fp16 = reshape(shape = concat_413x, x = hidden_states_641_cast_fp16)[name = string("key_states_87_cast_fp16")];
            tensor<int32, [4]> var_4113_shape_cast_fp16 = shape(x = var_4094_cast_fp16)[name = string("op_4113_shape_cast_fp16")];
            int32 gather_395 = const()[name = string("gather_395"), val = int32(1)];
            int32 gather_396 = const()[name = string("gather_396"), val = int32(2)];
            int32 gather_397_axis_0 = const()[name = string("gather_397_axis_0"), val = int32(0)];
            int32 gather_397_batch_dims_0 = const()[name = string("gather_397_batch_dims_0"), val = int32(0)];
            bool gather_397_validate_indices_0 = const()[name = string("gather_397_validate_indices_0"), val = bool(false)];
            string var_4113_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4113_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_397_to_uint16 = const()[name = string("select_397_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_4113_shape_cast_fp16_to_uint16 = cast(dtype = var_4113_shape_cast_fp16_to_uint16_dtype_0, x = var_4113_shape_cast_fp16)[name = string("cast_39")];
            uint16 gather_397_cast_uint16 = gather(axis = gather_397_axis_0, batch_dims = gather_397_batch_dims_0, indices = select_397_to_uint16, validate_indices = gather_397_validate_indices_0, x = var_4113_shape_cast_fp16_to_uint16)[name = string("gather_397_cast_uint16")];
            string gather_397_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_397_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_398 = const()[name = string("gather_398"), val = int32(128)];
            tensor<int32, [1]> var_4120_axes_0 = const()[name = string("op_4120_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_4120_cast_fp16 = expand_dims(axes = var_4120_axes_0, x = var_4094_cast_fp16)[name = string("op_4120_cast_fp16")];
            tensor<int32, [5]> shape_442_cast_fp16 = shape(x = var_4120_cast_fp16)[name = string("shape_442_cast_fp16")];
            int32 concat_414_axis_0 = const()[name = string("concat_414_axis_0"), val = int32(0)];
            bool concat_414_interleave_0 = const()[name = string("concat_414_interleave_0"), val = bool(false)];
            int32 gather_397_cast_uint16_to_int32 = cast(dtype = gather_397_cast_uint16_to_int32_dtype_0, x = gather_397_cast_uint16)[name = string("cast_38")];
            tensor<int32, [5]> concat_414 = concat(axis = concat_414_axis_0, interleave = concat_414_interleave_0, values = (gather_395, gather_396, var_77, gather_397_cast_uint16_to_int32, gather_398))[name = string("concat_414")];
            tensor<int32, [5]> real_div_43 = real_div(x = concat_414, y = shape_442_cast_fp16)[name = string("real_div_43")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_645_cast_fp16 = tile(reps = real_div_43, x = var_4120_cast_fp16)[name = string("hidden_states_645_cast_fp16")];
            tensor<int32, [4]> concat_415x = const()[name = string("concat_415x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_87_cast_fp16 = reshape(shape = concat_415x, x = hidden_states_645_cast_fp16)[name = string("value_states_87_cast_fp16")];
            tensor<int32, [4]> var_4130_shape_cast_fp16 = shape(x = key_states_87_cast_fp16)[name = string("op_4130_shape_cast_fp16")];
            int32 gather_399_axis_0 = const()[name = string("gather_399_axis_0"), val = int32(0)];
            int32 gather_399_batch_dims_0 = const()[name = string("gather_399_batch_dims_0"), val = int32(0)];
            bool gather_399_validate_indices_0 = const()[name = string("gather_399_validate_indices_0"), val = bool(false)];
            string var_4130_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4130_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_399_to_uint16 = const()[name = string("select_399_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_4130_shape_cast_fp16_to_uint16 = cast(dtype = var_4130_shape_cast_fp16_to_uint16_dtype_0, x = var_4130_shape_cast_fp16)[name = string("cast_37")];
            uint16 gather_399_cast_uint16 = gather(axis = gather_399_axis_0, batch_dims = gather_399_batch_dims_0, indices = select_399_to_uint16, validate_indices = gather_399_validate_indices_0, x = var_4130_shape_cast_fp16_to_uint16)[name = string("gather_399_cast_uint16")];
            string gather_399_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_399_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_416_values0_0 = const()[name = string("concat_416_values0_0"), val = int32(1)];
            int32 concat_416_values1_0 = const()[name = string("concat_416_values1_0"), val = int32(1)];
            int32 concat_416_values2_0 = const()[name = string("concat_416_values2_0"), val = int32(0)];
            int32 concat_416_axis_0 = const()[name = string("concat_416_axis_0"), val = int32(0)];
            bool concat_416_interleave_0 = const()[name = string("concat_416_interleave_0"), val = bool(false)];
            int32 gather_399_cast_uint16_to_int32 = cast(dtype = gather_399_cast_uint16_to_int32_dtype_0, x = gather_399_cast_uint16)[name = string("cast_36")];
            tensor<int32, [4]> concat_416 = concat(axis = concat_416_axis_0, interleave = concat_416_interleave_0, values = (concat_416_values0_0, concat_416_values1_0, concat_416_values2_0, gather_399_cast_uint16_to_int32))[name = string("concat_416")];
            tensor<int32, [4]> causal_mask_45_begin_0 = const()[name = string("causal_mask_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_45_end_mask_0 = const()[name = string("causal_mask_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_45_cast_fp16 = slice_by_index(begin = causal_mask_45_begin_0, end = concat_416, end_mask = causal_mask_45_end_mask_0, x = causal_mask)[name = string("causal_mask_45_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_85_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_45_cast_fp16, key = key_states_87_cast_fp16, query = query_states_87_cast_fp16, value = value_states_87_cast_fp16)[name = string("attn_output_85_cast_fp16")];
            tensor<int32, [4]> var_4136_perm_0 = const()[name = string("op_4136_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_417x = const()[name = string("concat_417x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_4136_cast_fp16 = transpose(perm = var_4136_perm_0, x = attn_output_85_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, ?, 1536]> input_169_cast_fp16 = reshape(shape = concat_417x, x = var_4136_cast_fp16)[name = string("input_169_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686050880))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687230592))))[name = string("model_model_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_150_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_649_cast_fp16 = add(x = hidden_states_629_cast_fp16, y = linear_150_cast_fp16)[name = string("hidden_states_649_cast_fp16")];
            fp16 var_78_promoted_43_to_fp16 = const()[name = string("op_78_promoted_43_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_4145_cast_fp16 = pow(x = hidden_states_649_cast_fp16, y = var_78_promoted_43_to_fp16)[name = string("op_4145_cast_fp16")];
            tensor<int32, [1]> variance_87_axes_0 = const()[name = string("variance_87_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_87_keep_dims_0 = const()[name = string("variance_87_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_87_cast_fp16 = reduce_mean(axes = variance_87_axes_0, keep_dims = variance_87_keep_dims_0, x = var_4145_cast_fp16)[name = string("variance_87_cast_fp16")];
            fp16 var_4148_to_fp16 = const()[name = string("op_4148_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4149_cast_fp16 = add(x = variance_87_cast_fp16, y = var_4148_to_fp16)[name = string("op_4149_cast_fp16")];
            fp32 var_4150_epsilon_0 = const()[name = string("op_4150_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_4150_cast_fp16 = rsqrt(epsilon = var_4150_epsilon_0, x = var_4149_cast_fp16)[name = string("op_4150_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_653_cast_fp16 = mul(x = hidden_states_649_cast_fp16, y = var_4150_cast_fp16)[name = string("hidden_states_653_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687378112)))];
            tensor<fp16, [1, ?, 1536]> input_171_cast_fp16 = mul(x = model_model_layers_21_post_attention_layernorm_weight_to_fp16, y = hidden_states_653_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687381248))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694262592))))[name = string("model_model_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_4162_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_4162_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(695122816))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(702004160))))[name = string("model_model_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_175_cast_fp16 = mul(x = var_4162_cast_fp16, y = linear_152_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(702864384))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(709745728))))[name = string("model_model_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_153_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_659_cast_fp16 = add(x = hidden_states_649_cast_fp16, y = linear_153_cast_fp16)[name = string("hidden_states_659_cast_fp16")];
            fp16 var_78_promoted_44_to_fp16 = const()[name = string("op_78_promoted_44_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_4175_cast_fp16 = pow(x = hidden_states_659_cast_fp16, y = var_78_promoted_44_to_fp16)[name = string("op_4175_cast_fp16")];
            tensor<int32, [1]> variance_89_axes_0 = const()[name = string("variance_89_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_89_keep_dims_0 = const()[name = string("variance_89_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_89_cast_fp16 = reduce_mean(axes = variance_89_axes_0, keep_dims = variance_89_keep_dims_0, x = var_4175_cast_fp16)[name = string("variance_89_cast_fp16")];
            fp16 var_4178_to_fp16 = const()[name = string("op_4178_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4179_cast_fp16 = add(x = variance_89_cast_fp16, y = var_4178_to_fp16)[name = string("op_4179_cast_fp16")];
            fp32 var_4180_epsilon_0 = const()[name = string("op_4180_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_4180_cast_fp16 = rsqrt(epsilon = var_4180_epsilon_0, x = var_4179_cast_fp16)[name = string("op_4180_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_663_cast_fp16 = mul(x = hidden_states_659_cast_fp16, y = var_4180_cast_fp16)[name = string("hidden_states_663_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(710605952)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_667_cast_fp16 = mul(x = model_model_layers_22_input_layernorm_weight_to_fp16, y = hidden_states_663_cast_fp16)[name = string("hidden_states_667_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(710609088))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711788800))))[name = string("model_model_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711936320)))];
            tensor<fp16, [1, ?, 1536]> linear_154_cast_fp16 = linear(bias = model_model_layers_22_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_667_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711939456))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(712136128))))[name = string("model_model_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_22_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_22_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(712160768)))];
            tensor<fp16, [1, ?, 256]> linear_155_cast_fp16 = linear(bias = model_model_layers_22_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_667_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(712161344))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(712358016))))[name = string("model_model_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(712382656)))];
            tensor<fp16, [1, ?, 256]> linear_156_cast_fp16 = linear(bias = model_model_layers_22_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_667_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> concat_418x = const()[name = string("concat_418x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_4203_cast_fp16 = reshape(shape = concat_418x, x = linear_154_cast_fp16)[name = string("op_4203_cast_fp16")];
            tensor<int32, [4]> q_45_perm_0 = const()[name = string("q_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_419x = const()[name = string("concat_419x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_4206_cast_fp16 = reshape(shape = concat_419x, x = linear_155_cast_fp16)[name = string("op_4206_cast_fp16")];
            tensor<int32, [4]> k_45_perm_0 = const()[name = string("k_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_420x = const()[name = string("concat_420x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_4209_cast_fp16 = reshape(shape = concat_420x, x = linear_156_cast_fp16)[name = string("op_4209_cast_fp16")];
            tensor<int32, [4]> v_state_45_perm_0 = const()[name = string("v_state_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = var_4203_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 12, ?, 128]> var_4213_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_7_cast_fp16)[name = string("op_4213_cast_fp16")];
            tensor<int32, [4]> x1_89_begin_0 = const()[name = string("x1_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_89_end_0 = const()[name = string("x1_89_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_89_end_mask_0 = const()[name = string("x1_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_89_cast_fp16 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = q_45_cast_fp16)[name = string("x1_89_cast_fp16")];
            tensor<int32, [4]> x2_89_begin_0 = const()[name = string("x2_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_89_end_0 = const()[name = string("x2_89_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_89_end_mask_0 = const()[name = string("x2_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_89_cast_fp16 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = q_45_cast_fp16)[name = string("x2_89_cast_fp16")];
            fp16 const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_4224_cast_fp16 = mul(x = x2_89_cast_fp16, y = const_47_promoted_to_fp16)[name = string("op_4224_cast_fp16")];
            bool var_4226_interleave_0 = const()[name = string("op_4226_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_4226_cast_fp16 = concat(axis = var_72, interleave = var_4226_interleave_0, values = (var_4224_cast_fp16, x1_89_cast_fp16))[name = string("op_4226_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_4227_cast_fp16 = mul(x = var_4226_cast_fp16, y = sin_7_cast_fp16)[name = string("op_4227_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_91_cast_fp16 = add(x = var_4213_cast_fp16, y = var_4227_cast_fp16)[name = string("query_states_91_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = var_4206_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 2, ?, 128]> var_4229_cast_fp16 = mul(x = k_45_cast_fp16, y = cos_7_cast_fp16)[name = string("op_4229_cast_fp16")];
            tensor<int32, [4]> x1_91_begin_0 = const()[name = string("x1_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_91_end_0 = const()[name = string("x1_91_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_91_end_mask_0 = const()[name = string("x1_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_91_cast_fp16 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = k_45_cast_fp16)[name = string("x1_91_cast_fp16")];
            tensor<int32, [4]> x2_91_begin_0 = const()[name = string("x2_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_91_end_0 = const()[name = string("x2_91_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_91_end_mask_0 = const()[name = string("x2_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_91_cast_fp16 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = k_45_cast_fp16)[name = string("x2_91_cast_fp16")];
            fp16 const_48_promoted_to_fp16 = const()[name = string("const_48_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_4240_cast_fp16 = mul(x = x2_91_cast_fp16, y = const_48_promoted_to_fp16)[name = string("op_4240_cast_fp16")];
            bool var_4242_interleave_0 = const()[name = string("op_4242_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_4242_cast_fp16 = concat(axis = var_72, interleave = var_4242_interleave_0, values = (var_4240_cast_fp16, x1_91_cast_fp16))[name = string("op_4242_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_4243_cast_fp16 = mul(x = var_4242_cast_fp16, y = sin_7_cast_fp16)[name = string("op_4243_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_45_cast_fp16 = add(x = var_4229_cast_fp16, y = var_4243_cast_fp16)[name = string("k_state_45_cast_fp16")];
            tensor<int32, [1]> expand_dims_264 = const()[name = string("expand_dims_264"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_265 = const()[name = string("expand_dims_265"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_267 = const()[name = string("expand_dims_267"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_423_values0_0 = const()[name = string("concat_423_values0_0"), val = tensor<int32, [1]>([22])];
            int32 concat_423_axis_0 = const()[name = string("concat_423_axis_0"), val = int32(0)];
            bool concat_423_interleave_0 = const()[name = string("concat_423_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_423 = concat(axis = concat_423_axis_0, interleave = concat_423_interleave_0, values = (concat_423_values0_0, expand_dims_264, expand_dims_265, expand_dims_2, expand_dims_267))[name = string("concat_423")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_23_stride_0 = const()[name = string("key_cache_internal_tensor_assign_23_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_23_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_23_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_23_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_23_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_23_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_23_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_23_cast_fp16 = slice_update(begin = concat_423, begin_mask = key_cache_internal_tensor_assign_23_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_23_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_23_squeeze_mask_0, stride = key_cache_internal_tensor_assign_23_stride_0, update = k_state_45_cast_fp16, x = coreml_update_state_98)[name = string("key_cache_internal_tensor_assign_23_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_23_cast_fp16, input = key_cache)[name = string("coreml_update_state_100_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_100 = read_state(input = key_cache)[name = string("coreml_update_state_100")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_23_stride_0 = const()[name = string("value_cache_internal_tensor_assign_23_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_23_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_23_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_23_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_23_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_23_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_23_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_45_cast_fp16 = transpose(perm = v_state_45_perm_0, x = var_4209_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_23_cast_fp16 = slice_update(begin = concat_423, begin_mask = value_cache_internal_tensor_assign_23_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_23_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_23_squeeze_mask_0, stride = value_cache_internal_tensor_assign_23_stride_0, update = v_state_45_cast_fp16, x = coreml_update_state_99)[name = string("value_cache_internal_tensor_assign_23_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_23_cast_fp16, input = value_cache)[name = string("coreml_update_state_101_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_101 = read_state(input = value_cache)[name = string("coreml_update_state_101")];
            tensor<int32, [5]> var_4266_begin_0 = const()[name = string("op_4266_begin_0"), val = tensor<int32, [5]>([22, 0, 0, 0, 0])];
            tensor<int32, [5]> var_4266_end_0 = const()[name = string("op_4266_end_0"), val = tensor<int32, [5]>([23, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_4266_end_mask_0 = const()[name = string("op_4266_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_4266_squeeze_mask_0 = const()[name = string("op_4266_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_4266_cast_fp16 = slice_by_index(begin = var_4266_begin_0, end = var_4266_end_0, end_mask = var_4266_end_mask_0, squeeze_mask = var_4266_squeeze_mask_0, x = coreml_update_state_100)[name = string("op_4266_cast_fp16")];
            tensor<int32, [4]> var_4269_begin_0 = const()[name = string("op_4269_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_4269_end_mask_0 = const()[name = string("op_4269_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_4269_cast_fp16 = slice_by_index(begin = var_4269_begin_0, end = concat_11, end_mask = var_4269_end_mask_0, x = var_4266_cast_fp16)[name = string("op_4269_cast_fp16")];
            tensor<int32, [5]> var_4271_begin_0 = const()[name = string("op_4271_begin_0"), val = tensor<int32, [5]>([22, 0, 0, 0, 0])];
            tensor<int32, [5]> var_4271_end_0 = const()[name = string("op_4271_end_0"), val = tensor<int32, [5]>([23, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_4271_end_mask_0 = const()[name = string("op_4271_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_4271_squeeze_mask_0 = const()[name = string("op_4271_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_4271_cast_fp16 = slice_by_index(begin = var_4271_begin_0, end = var_4271_end_0, end_mask = var_4271_end_mask_0, squeeze_mask = var_4271_squeeze_mask_0, x = coreml_update_state_101)[name = string("op_4271_cast_fp16")];
            tensor<int32, [4]> var_4274_begin_0 = const()[name = string("op_4274_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_4274_end_mask_0 = const()[name = string("op_4274_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_4274_cast_fp16 = slice_by_index(begin = var_4274_begin_0, end = concat_11, end_mask = var_4274_end_mask_0, x = var_4271_cast_fp16)[name = string("op_4274_cast_fp16")];
            tensor<int32, [4]> var_4276_shape_cast_fp16 = shape(x = var_4269_cast_fp16)[name = string("op_4276_shape_cast_fp16")];
            int32 gather_409 = const()[name = string("gather_409"), val = int32(1)];
            int32 gather_410 = const()[name = string("gather_410"), val = int32(2)];
            int32 gather_411_axis_0 = const()[name = string("gather_411_axis_0"), val = int32(0)];
            int32 gather_411_batch_dims_0 = const()[name = string("gather_411_batch_dims_0"), val = int32(0)];
            bool gather_411_validate_indices_0 = const()[name = string("gather_411_validate_indices_0"), val = bool(false)];
            string var_4276_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4276_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_411_to_uint16 = const()[name = string("select_411_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_4276_shape_cast_fp16_to_uint16 = cast(dtype = var_4276_shape_cast_fp16_to_uint16_dtype_0, x = var_4276_shape_cast_fp16)[name = string("cast_35")];
            uint16 gather_411_cast_uint16 = gather(axis = gather_411_axis_0, batch_dims = gather_411_batch_dims_0, indices = select_411_to_uint16, validate_indices = gather_411_validate_indices_0, x = var_4276_shape_cast_fp16_to_uint16)[name = string("gather_411_cast_uint16")];
            string gather_411_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_411_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_412 = const()[name = string("gather_412"), val = int32(128)];
            tensor<int32, [1]> var_4283_axes_0 = const()[name = string("op_4283_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_4283_cast_fp16 = expand_dims(axes = var_4283_axes_0, x = var_4269_cast_fp16)[name = string("op_4283_cast_fp16")];
            tensor<int32, [5]> shape_457_cast_fp16 = shape(x = var_4283_cast_fp16)[name = string("shape_457_cast_fp16")];
            int32 concat_431_axis_0 = const()[name = string("concat_431_axis_0"), val = int32(0)];
            bool concat_431_interleave_0 = const()[name = string("concat_431_interleave_0"), val = bool(false)];
            int32 gather_411_cast_uint16_to_int32 = cast(dtype = gather_411_cast_uint16_to_int32_dtype_0, x = gather_411_cast_uint16)[name = string("cast_34")];
            tensor<int32, [5]> concat_431 = concat(axis = concat_431_axis_0, interleave = concat_431_interleave_0, values = (gather_409, gather_410, var_77, gather_411_cast_uint16_to_int32, gather_412))[name = string("concat_431")];
            tensor<int32, [5]> real_div_44 = real_div(x = concat_431, y = shape_457_cast_fp16)[name = string("real_div_44")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_671_cast_fp16 = tile(reps = real_div_44, x = var_4283_cast_fp16)[name = string("hidden_states_671_cast_fp16")];
            tensor<int32, [4]> concat_432x = const()[name = string("concat_432x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_91_cast_fp16 = reshape(shape = concat_432x, x = hidden_states_671_cast_fp16)[name = string("key_states_91_cast_fp16")];
            tensor<int32, [4]> var_4293_shape_cast_fp16 = shape(x = var_4274_cast_fp16)[name = string("op_4293_shape_cast_fp16")];
            int32 gather_413 = const()[name = string("gather_413"), val = int32(1)];
            int32 gather_414 = const()[name = string("gather_414"), val = int32(2)];
            int32 gather_415_axis_0 = const()[name = string("gather_415_axis_0"), val = int32(0)];
            int32 gather_415_batch_dims_0 = const()[name = string("gather_415_batch_dims_0"), val = int32(0)];
            bool gather_415_validate_indices_0 = const()[name = string("gather_415_validate_indices_0"), val = bool(false)];
            string var_4293_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4293_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_415_to_uint16 = const()[name = string("select_415_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_4293_shape_cast_fp16_to_uint16 = cast(dtype = var_4293_shape_cast_fp16_to_uint16_dtype_0, x = var_4293_shape_cast_fp16)[name = string("cast_33")];
            uint16 gather_415_cast_uint16 = gather(axis = gather_415_axis_0, batch_dims = gather_415_batch_dims_0, indices = select_415_to_uint16, validate_indices = gather_415_validate_indices_0, x = var_4293_shape_cast_fp16_to_uint16)[name = string("gather_415_cast_uint16")];
            string gather_415_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_415_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_416 = const()[name = string("gather_416"), val = int32(128)];
            tensor<int32, [1]> var_4300_axes_0 = const()[name = string("op_4300_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_4300_cast_fp16 = expand_dims(axes = var_4300_axes_0, x = var_4274_cast_fp16)[name = string("op_4300_cast_fp16")];
            tensor<int32, [5]> shape_462_cast_fp16 = shape(x = var_4300_cast_fp16)[name = string("shape_462_cast_fp16")];
            int32 concat_433_axis_0 = const()[name = string("concat_433_axis_0"), val = int32(0)];
            bool concat_433_interleave_0 = const()[name = string("concat_433_interleave_0"), val = bool(false)];
            int32 gather_415_cast_uint16_to_int32 = cast(dtype = gather_415_cast_uint16_to_int32_dtype_0, x = gather_415_cast_uint16)[name = string("cast_32")];
            tensor<int32, [5]> concat_433 = concat(axis = concat_433_axis_0, interleave = concat_433_interleave_0, values = (gather_413, gather_414, var_77, gather_415_cast_uint16_to_int32, gather_416))[name = string("concat_433")];
            tensor<int32, [5]> real_div_45 = real_div(x = concat_433, y = shape_462_cast_fp16)[name = string("real_div_45")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_675_cast_fp16 = tile(reps = real_div_45, x = var_4300_cast_fp16)[name = string("hidden_states_675_cast_fp16")];
            tensor<int32, [4]> concat_434x = const()[name = string("concat_434x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_91_cast_fp16 = reshape(shape = concat_434x, x = hidden_states_675_cast_fp16)[name = string("value_states_91_cast_fp16")];
            tensor<int32, [4]> var_4310_shape_cast_fp16 = shape(x = key_states_91_cast_fp16)[name = string("op_4310_shape_cast_fp16")];
            int32 gather_417_axis_0 = const()[name = string("gather_417_axis_0"), val = int32(0)];
            int32 gather_417_batch_dims_0 = const()[name = string("gather_417_batch_dims_0"), val = int32(0)];
            bool gather_417_validate_indices_0 = const()[name = string("gather_417_validate_indices_0"), val = bool(false)];
            string var_4310_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4310_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_417_to_uint16 = const()[name = string("select_417_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_4310_shape_cast_fp16_to_uint16 = cast(dtype = var_4310_shape_cast_fp16_to_uint16_dtype_0, x = var_4310_shape_cast_fp16)[name = string("cast_31")];
            uint16 gather_417_cast_uint16 = gather(axis = gather_417_axis_0, batch_dims = gather_417_batch_dims_0, indices = select_417_to_uint16, validate_indices = gather_417_validate_indices_0, x = var_4310_shape_cast_fp16_to_uint16)[name = string("gather_417_cast_uint16")];
            string gather_417_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_417_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_435_values0_0 = const()[name = string("concat_435_values0_0"), val = int32(1)];
            int32 concat_435_values1_0 = const()[name = string("concat_435_values1_0"), val = int32(1)];
            int32 concat_435_values2_0 = const()[name = string("concat_435_values2_0"), val = int32(0)];
            int32 concat_435_axis_0 = const()[name = string("concat_435_axis_0"), val = int32(0)];
            bool concat_435_interleave_0 = const()[name = string("concat_435_interleave_0"), val = bool(false)];
            int32 gather_417_cast_uint16_to_int32 = cast(dtype = gather_417_cast_uint16_to_int32_dtype_0, x = gather_417_cast_uint16)[name = string("cast_30")];
            tensor<int32, [4]> concat_435 = concat(axis = concat_435_axis_0, interleave = concat_435_interleave_0, values = (concat_435_values0_0, concat_435_values1_0, concat_435_values2_0, gather_417_cast_uint16_to_int32))[name = string("concat_435")];
            tensor<int32, [4]> causal_mask_47_begin_0 = const()[name = string("causal_mask_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_47_end_mask_0 = const()[name = string("causal_mask_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_47_cast_fp16 = slice_by_index(begin = causal_mask_47_begin_0, end = concat_435, end_mask = causal_mask_47_end_mask_0, x = causal_mask)[name = string("causal_mask_47_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_89_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_47_cast_fp16, key = key_states_91_cast_fp16, query = query_states_91_cast_fp16, value = value_states_91_cast_fp16)[name = string("attn_output_89_cast_fp16")];
            tensor<int32, [4]> var_4316_perm_0 = const()[name = string("op_4316_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_436x = const()[name = string("concat_436x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_4316_cast_fp16 = transpose(perm = var_4316_perm_0, x = attn_output_89_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, ?, 1536]> input_177_cast_fp16 = reshape(shape = concat_436x, x = var_4316_cast_fp16)[name = string("input_177_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(712383232))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(713562944))))[name = string("model_model_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_157_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = input_177_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_679_cast_fp16 = add(x = hidden_states_659_cast_fp16, y = linear_157_cast_fp16)[name = string("hidden_states_679_cast_fp16")];
            fp16 var_78_promoted_45_to_fp16 = const()[name = string("op_78_promoted_45_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_4325_cast_fp16 = pow(x = hidden_states_679_cast_fp16, y = var_78_promoted_45_to_fp16)[name = string("op_4325_cast_fp16")];
            tensor<int32, [1]> variance_91_axes_0 = const()[name = string("variance_91_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_91_keep_dims_0 = const()[name = string("variance_91_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_91_cast_fp16 = reduce_mean(axes = variance_91_axes_0, keep_dims = variance_91_keep_dims_0, x = var_4325_cast_fp16)[name = string("variance_91_cast_fp16")];
            fp16 var_4328_to_fp16 = const()[name = string("op_4328_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4329_cast_fp16 = add(x = variance_91_cast_fp16, y = var_4328_to_fp16)[name = string("op_4329_cast_fp16")];
            fp32 var_4330_epsilon_0 = const()[name = string("op_4330_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_4330_cast_fp16 = rsqrt(epsilon = var_4330_epsilon_0, x = var_4329_cast_fp16)[name = string("op_4330_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_683_cast_fp16 = mul(x = hidden_states_679_cast_fp16, y = var_4330_cast_fp16)[name = string("hidden_states_683_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(713710464)))];
            tensor<fp16, [1, ?, 1536]> input_179_cast_fp16 = mul(x = model_model_layers_22_post_attention_layernorm_weight_to_fp16, y = hidden_states_683_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(713713600))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(720594944))))[name = string("model_model_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_4342_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_4342_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(721455168))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(728336512))))[name = string("model_model_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_183_cast_fp16 = mul(x = var_4342_cast_fp16, y = linear_159_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729196736))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736078080))))[name = string("model_model_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_160_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_689_cast_fp16 = add(x = hidden_states_679_cast_fp16, y = linear_160_cast_fp16)[name = string("hidden_states_689_cast_fp16")];
            fp16 var_78_promoted_46_to_fp16 = const()[name = string("op_78_promoted_46_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_4355_cast_fp16 = pow(x = hidden_states_689_cast_fp16, y = var_78_promoted_46_to_fp16)[name = string("op_4355_cast_fp16")];
            tensor<int32, [1]> variance_93_axes_0 = const()[name = string("variance_93_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_93_keep_dims_0 = const()[name = string("variance_93_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_93_cast_fp16 = reduce_mean(axes = variance_93_axes_0, keep_dims = variance_93_keep_dims_0, x = var_4355_cast_fp16)[name = string("variance_93_cast_fp16")];
            fp16 var_4358_to_fp16 = const()[name = string("op_4358_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4359_cast_fp16 = add(x = variance_93_cast_fp16, y = var_4358_to_fp16)[name = string("op_4359_cast_fp16")];
            fp32 var_4360_epsilon_0 = const()[name = string("op_4360_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_4360_cast_fp16 = rsqrt(epsilon = var_4360_epsilon_0, x = var_4359_cast_fp16)[name = string("op_4360_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_693_cast_fp16 = mul(x = hidden_states_689_cast_fp16, y = var_4360_cast_fp16)[name = string("hidden_states_693_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736938304)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_697_cast_fp16 = mul(x = model_model_layers_23_input_layernorm_weight_to_fp16, y = hidden_states_693_cast_fp16)[name = string("hidden_states_697_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736941440))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738121152))))[name = string("model_model_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738268672)))];
            tensor<fp16, [1, ?, 1536]> linear_161_cast_fp16 = linear(bias = model_model_layers_23_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_697_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738271808))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738468480))))[name = string("model_model_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_23_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_23_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738493120)))];
            tensor<fp16, [1, ?, 256]> linear_162_cast_fp16 = linear(bias = model_model_layers_23_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_697_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738493696))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738690368))))[name = string("model_model_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738715008)))];
            tensor<fp16, [1, ?, 256]> linear_163_cast_fp16 = linear(bias = model_model_layers_23_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_697_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> concat_437x = const()[name = string("concat_437x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_4383_cast_fp16 = reshape(shape = concat_437x, x = linear_161_cast_fp16)[name = string("op_4383_cast_fp16")];
            tensor<int32, [4]> q_47_perm_0 = const()[name = string("q_47_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_438x = const()[name = string("concat_438x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_4386_cast_fp16 = reshape(shape = concat_438x, x = linear_162_cast_fp16)[name = string("op_4386_cast_fp16")];
            tensor<int32, [4]> k_47_perm_0 = const()[name = string("k_47_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_439x = const()[name = string("concat_439x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_4389_cast_fp16 = reshape(shape = concat_439x, x = linear_163_cast_fp16)[name = string("op_4389_cast_fp16")];
            tensor<int32, [4]> v_state_47_perm_0 = const()[name = string("v_state_47_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_47_cast_fp16 = transpose(perm = q_47_perm_0, x = var_4383_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 12, ?, 128]> var_4393_cast_fp16 = mul(x = q_47_cast_fp16, y = cos_7_cast_fp16)[name = string("op_4393_cast_fp16")];
            tensor<int32, [4]> x1_93_begin_0 = const()[name = string("x1_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_93_end_0 = const()[name = string("x1_93_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_93_end_mask_0 = const()[name = string("x1_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_93_cast_fp16 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = q_47_cast_fp16)[name = string("x1_93_cast_fp16")];
            tensor<int32, [4]> x2_93_begin_0 = const()[name = string("x2_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_93_end_0 = const()[name = string("x2_93_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_93_end_mask_0 = const()[name = string("x2_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_93_cast_fp16 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = q_47_cast_fp16)[name = string("x2_93_cast_fp16")];
            fp16 const_49_promoted_to_fp16 = const()[name = string("const_49_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_4404_cast_fp16 = mul(x = x2_93_cast_fp16, y = const_49_promoted_to_fp16)[name = string("op_4404_cast_fp16")];
            bool var_4406_interleave_0 = const()[name = string("op_4406_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_4406_cast_fp16 = concat(axis = var_72, interleave = var_4406_interleave_0, values = (var_4404_cast_fp16, x1_93_cast_fp16))[name = string("op_4406_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_4407_cast_fp16 = mul(x = var_4406_cast_fp16, y = sin_7_cast_fp16)[name = string("op_4407_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_95_cast_fp16 = add(x = var_4393_cast_fp16, y = var_4407_cast_fp16)[name = string("query_states_95_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_47_cast_fp16 = transpose(perm = k_47_perm_0, x = var_4386_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2, ?, 128]> var_4409_cast_fp16 = mul(x = k_47_cast_fp16, y = cos_7_cast_fp16)[name = string("op_4409_cast_fp16")];
            tensor<int32, [4]> x1_95_begin_0 = const()[name = string("x1_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_95_end_0 = const()[name = string("x1_95_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_95_end_mask_0 = const()[name = string("x1_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_95_cast_fp16 = slice_by_index(begin = x1_95_begin_0, end = x1_95_end_0, end_mask = x1_95_end_mask_0, x = k_47_cast_fp16)[name = string("x1_95_cast_fp16")];
            tensor<int32, [4]> x2_95_begin_0 = const()[name = string("x2_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_95_end_0 = const()[name = string("x2_95_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_95_end_mask_0 = const()[name = string("x2_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_95_cast_fp16 = slice_by_index(begin = x2_95_begin_0, end = x2_95_end_0, end_mask = x2_95_end_mask_0, x = k_47_cast_fp16)[name = string("x2_95_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_4420_cast_fp16 = mul(x = x2_95_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_4420_cast_fp16")];
            bool var_4422_interleave_0 = const()[name = string("op_4422_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_4422_cast_fp16 = concat(axis = var_72, interleave = var_4422_interleave_0, values = (var_4420_cast_fp16, x1_95_cast_fp16))[name = string("op_4422_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_4423_cast_fp16 = mul(x = var_4422_cast_fp16, y = sin_7_cast_fp16)[name = string("op_4423_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_47_cast_fp16 = add(x = var_4409_cast_fp16, y = var_4423_cast_fp16)[name = string("k_state_47_cast_fp16")];
            tensor<int32, [1]> expand_dims_276 = const()[name = string("expand_dims_276"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_277 = const()[name = string("expand_dims_277"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_279 = const()[name = string("expand_dims_279"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_442_values0_0 = const()[name = string("concat_442_values0_0"), val = tensor<int32, [1]>([23])];
            int32 concat_442_axis_0 = const()[name = string("concat_442_axis_0"), val = int32(0)];
            bool concat_442_interleave_0 = const()[name = string("concat_442_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_442 = concat(axis = concat_442_axis_0, interleave = concat_442_interleave_0, values = (concat_442_values0_0, expand_dims_276, expand_dims_277, expand_dims_2, expand_dims_279))[name = string("concat_442")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_24_stride_0 = const()[name = string("key_cache_internal_tensor_assign_24_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_24_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_24_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_24_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_24_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_24_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_24_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_24_cast_fp16 = slice_update(begin = concat_442, begin_mask = key_cache_internal_tensor_assign_24_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_24_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_24_squeeze_mask_0, stride = key_cache_internal_tensor_assign_24_stride_0, update = k_state_47_cast_fp16, x = coreml_update_state_100)[name = string("key_cache_internal_tensor_assign_24_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_24_cast_fp16, input = key_cache)[name = string("coreml_update_state_102_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_102 = read_state(input = key_cache)[name = string("coreml_update_state_102")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_24_stride_0 = const()[name = string("value_cache_internal_tensor_assign_24_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_24_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_24_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_24_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_24_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_24_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_24_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_47_cast_fp16 = transpose(perm = v_state_47_perm_0, x = var_4389_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_24_cast_fp16 = slice_update(begin = concat_442, begin_mask = value_cache_internal_tensor_assign_24_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_24_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_24_squeeze_mask_0, stride = value_cache_internal_tensor_assign_24_stride_0, update = v_state_47_cast_fp16, x = coreml_update_state_101)[name = string("value_cache_internal_tensor_assign_24_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_24_cast_fp16, input = value_cache)[name = string("coreml_update_state_103_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_103 = read_state(input = value_cache)[name = string("coreml_update_state_103")];
            tensor<int32, [5]> var_4446_begin_0 = const()[name = string("op_4446_begin_0"), val = tensor<int32, [5]>([23, 0, 0, 0, 0])];
            tensor<int32, [5]> var_4446_end_0 = const()[name = string("op_4446_end_0"), val = tensor<int32, [5]>([24, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_4446_end_mask_0 = const()[name = string("op_4446_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_4446_squeeze_mask_0 = const()[name = string("op_4446_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_4446_cast_fp16 = slice_by_index(begin = var_4446_begin_0, end = var_4446_end_0, end_mask = var_4446_end_mask_0, squeeze_mask = var_4446_squeeze_mask_0, x = coreml_update_state_102)[name = string("op_4446_cast_fp16")];
            tensor<int32, [4]> var_4449_begin_0 = const()[name = string("op_4449_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_4449_end_mask_0 = const()[name = string("op_4449_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_4449_cast_fp16 = slice_by_index(begin = var_4449_begin_0, end = concat_11, end_mask = var_4449_end_mask_0, x = var_4446_cast_fp16)[name = string("op_4449_cast_fp16")];
            tensor<int32, [5]> var_4451_begin_0 = const()[name = string("op_4451_begin_0"), val = tensor<int32, [5]>([23, 0, 0, 0, 0])];
            tensor<int32, [5]> var_4451_end_0 = const()[name = string("op_4451_end_0"), val = tensor<int32, [5]>([24, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_4451_end_mask_0 = const()[name = string("op_4451_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_4451_squeeze_mask_0 = const()[name = string("op_4451_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_4451_cast_fp16 = slice_by_index(begin = var_4451_begin_0, end = var_4451_end_0, end_mask = var_4451_end_mask_0, squeeze_mask = var_4451_squeeze_mask_0, x = coreml_update_state_103)[name = string("op_4451_cast_fp16")];
            tensor<int32, [4]> var_4454_begin_0 = const()[name = string("op_4454_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_4454_end_mask_0 = const()[name = string("op_4454_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_4454_cast_fp16 = slice_by_index(begin = var_4454_begin_0, end = concat_11, end_mask = var_4454_end_mask_0, x = var_4451_cast_fp16)[name = string("op_4454_cast_fp16")];
            tensor<int32, [4]> var_4456_shape_cast_fp16 = shape(x = var_4449_cast_fp16)[name = string("op_4456_shape_cast_fp16")];
            int32 gather_427 = const()[name = string("gather_427"), val = int32(1)];
            int32 gather_428 = const()[name = string("gather_428"), val = int32(2)];
            int32 gather_429_axis_0 = const()[name = string("gather_429_axis_0"), val = int32(0)];
            int32 gather_429_batch_dims_0 = const()[name = string("gather_429_batch_dims_0"), val = int32(0)];
            bool gather_429_validate_indices_0 = const()[name = string("gather_429_validate_indices_0"), val = bool(false)];
            string var_4456_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4456_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_429_to_uint16 = const()[name = string("select_429_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_4456_shape_cast_fp16_to_uint16 = cast(dtype = var_4456_shape_cast_fp16_to_uint16_dtype_0, x = var_4456_shape_cast_fp16)[name = string("cast_29")];
            uint16 gather_429_cast_uint16 = gather(axis = gather_429_axis_0, batch_dims = gather_429_batch_dims_0, indices = select_429_to_uint16, validate_indices = gather_429_validate_indices_0, x = var_4456_shape_cast_fp16_to_uint16)[name = string("gather_429_cast_uint16")];
            string gather_429_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_429_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_430 = const()[name = string("gather_430"), val = int32(128)];
            tensor<int32, [1]> var_4463_axes_0 = const()[name = string("op_4463_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_4463_cast_fp16 = expand_dims(axes = var_4463_axes_0, x = var_4449_cast_fp16)[name = string("op_4463_cast_fp16")];
            tensor<int32, [5]> shape_477_cast_fp16 = shape(x = var_4463_cast_fp16)[name = string("shape_477_cast_fp16")];
            int32 concat_450_axis_0 = const()[name = string("concat_450_axis_0"), val = int32(0)];
            bool concat_450_interleave_0 = const()[name = string("concat_450_interleave_0"), val = bool(false)];
            int32 gather_429_cast_uint16_to_int32 = cast(dtype = gather_429_cast_uint16_to_int32_dtype_0, x = gather_429_cast_uint16)[name = string("cast_28")];
            tensor<int32, [5]> concat_450 = concat(axis = concat_450_axis_0, interleave = concat_450_interleave_0, values = (gather_427, gather_428, var_77, gather_429_cast_uint16_to_int32, gather_430))[name = string("concat_450")];
            tensor<int32, [5]> real_div_46 = real_div(x = concat_450, y = shape_477_cast_fp16)[name = string("real_div_46")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_701_cast_fp16 = tile(reps = real_div_46, x = var_4463_cast_fp16)[name = string("hidden_states_701_cast_fp16")];
            tensor<int32, [4]> concat_451x = const()[name = string("concat_451x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_95_cast_fp16 = reshape(shape = concat_451x, x = hidden_states_701_cast_fp16)[name = string("key_states_95_cast_fp16")];
            tensor<int32, [4]> var_4473_shape_cast_fp16 = shape(x = var_4454_cast_fp16)[name = string("op_4473_shape_cast_fp16")];
            int32 gather_431 = const()[name = string("gather_431"), val = int32(1)];
            int32 gather_432 = const()[name = string("gather_432"), val = int32(2)];
            int32 gather_433_axis_0 = const()[name = string("gather_433_axis_0"), val = int32(0)];
            int32 gather_433_batch_dims_0 = const()[name = string("gather_433_batch_dims_0"), val = int32(0)];
            bool gather_433_validate_indices_0 = const()[name = string("gather_433_validate_indices_0"), val = bool(false)];
            string var_4473_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4473_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_433_to_uint16 = const()[name = string("select_433_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_4473_shape_cast_fp16_to_uint16 = cast(dtype = var_4473_shape_cast_fp16_to_uint16_dtype_0, x = var_4473_shape_cast_fp16)[name = string("cast_27")];
            uint16 gather_433_cast_uint16 = gather(axis = gather_433_axis_0, batch_dims = gather_433_batch_dims_0, indices = select_433_to_uint16, validate_indices = gather_433_validate_indices_0, x = var_4473_shape_cast_fp16_to_uint16)[name = string("gather_433_cast_uint16")];
            string gather_433_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_433_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_434 = const()[name = string("gather_434"), val = int32(128)];
            tensor<int32, [1]> var_4480_axes_0 = const()[name = string("op_4480_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_4480_cast_fp16 = expand_dims(axes = var_4480_axes_0, x = var_4454_cast_fp16)[name = string("op_4480_cast_fp16")];
            tensor<int32, [5]> shape_482_cast_fp16 = shape(x = var_4480_cast_fp16)[name = string("shape_482_cast_fp16")];
            int32 concat_452_axis_0 = const()[name = string("concat_452_axis_0"), val = int32(0)];
            bool concat_452_interleave_0 = const()[name = string("concat_452_interleave_0"), val = bool(false)];
            int32 gather_433_cast_uint16_to_int32 = cast(dtype = gather_433_cast_uint16_to_int32_dtype_0, x = gather_433_cast_uint16)[name = string("cast_26")];
            tensor<int32, [5]> concat_452 = concat(axis = concat_452_axis_0, interleave = concat_452_interleave_0, values = (gather_431, gather_432, var_77, gather_433_cast_uint16_to_int32, gather_434))[name = string("concat_452")];
            tensor<int32, [5]> real_div_47 = real_div(x = concat_452, y = shape_482_cast_fp16)[name = string("real_div_47")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_705_cast_fp16 = tile(reps = real_div_47, x = var_4480_cast_fp16)[name = string("hidden_states_705_cast_fp16")];
            tensor<int32, [4]> concat_453x = const()[name = string("concat_453x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_95_cast_fp16 = reshape(shape = concat_453x, x = hidden_states_705_cast_fp16)[name = string("value_states_95_cast_fp16")];
            tensor<int32, [4]> var_4490_shape_cast_fp16 = shape(x = key_states_95_cast_fp16)[name = string("op_4490_shape_cast_fp16")];
            int32 gather_435_axis_0 = const()[name = string("gather_435_axis_0"), val = int32(0)];
            int32 gather_435_batch_dims_0 = const()[name = string("gather_435_batch_dims_0"), val = int32(0)];
            bool gather_435_validate_indices_0 = const()[name = string("gather_435_validate_indices_0"), val = bool(false)];
            string var_4490_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4490_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_435_to_uint16 = const()[name = string("select_435_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_4490_shape_cast_fp16_to_uint16 = cast(dtype = var_4490_shape_cast_fp16_to_uint16_dtype_0, x = var_4490_shape_cast_fp16)[name = string("cast_25")];
            uint16 gather_435_cast_uint16 = gather(axis = gather_435_axis_0, batch_dims = gather_435_batch_dims_0, indices = select_435_to_uint16, validate_indices = gather_435_validate_indices_0, x = var_4490_shape_cast_fp16_to_uint16)[name = string("gather_435_cast_uint16")];
            string gather_435_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_435_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_454_values0_0 = const()[name = string("concat_454_values0_0"), val = int32(1)];
            int32 concat_454_values1_0 = const()[name = string("concat_454_values1_0"), val = int32(1)];
            int32 concat_454_values2_0 = const()[name = string("concat_454_values2_0"), val = int32(0)];
            int32 concat_454_axis_0 = const()[name = string("concat_454_axis_0"), val = int32(0)];
            bool concat_454_interleave_0 = const()[name = string("concat_454_interleave_0"), val = bool(false)];
            int32 gather_435_cast_uint16_to_int32 = cast(dtype = gather_435_cast_uint16_to_int32_dtype_0, x = gather_435_cast_uint16)[name = string("cast_24")];
            tensor<int32, [4]> concat_454 = concat(axis = concat_454_axis_0, interleave = concat_454_interleave_0, values = (concat_454_values0_0, concat_454_values1_0, concat_454_values2_0, gather_435_cast_uint16_to_int32))[name = string("concat_454")];
            tensor<int32, [4]> causal_mask_49_begin_0 = const()[name = string("causal_mask_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_49_end_mask_0 = const()[name = string("causal_mask_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_49_cast_fp16 = slice_by_index(begin = causal_mask_49_begin_0, end = concat_454, end_mask = causal_mask_49_end_mask_0, x = causal_mask)[name = string("causal_mask_49_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_93_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_49_cast_fp16, key = key_states_95_cast_fp16, query = query_states_95_cast_fp16, value = value_states_95_cast_fp16)[name = string("attn_output_93_cast_fp16")];
            tensor<int32, [4]> var_4496_perm_0 = const()[name = string("op_4496_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_455x = const()[name = string("concat_455x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_4496_cast_fp16 = transpose(perm = var_4496_perm_0, x = attn_output_93_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, ?, 1536]> input_185_cast_fp16 = reshape(shape = concat_455x, x = var_4496_cast_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738715584))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(739895296))))[name = string("model_model_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_164_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_709_cast_fp16 = add(x = hidden_states_689_cast_fp16, y = linear_164_cast_fp16)[name = string("hidden_states_709_cast_fp16")];
            fp16 var_78_promoted_47_to_fp16 = const()[name = string("op_78_promoted_47_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_4505_cast_fp16 = pow(x = hidden_states_709_cast_fp16, y = var_78_promoted_47_to_fp16)[name = string("op_4505_cast_fp16")];
            tensor<int32, [1]> variance_95_axes_0 = const()[name = string("variance_95_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_95_keep_dims_0 = const()[name = string("variance_95_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_95_cast_fp16 = reduce_mean(axes = variance_95_axes_0, keep_dims = variance_95_keep_dims_0, x = var_4505_cast_fp16)[name = string("variance_95_cast_fp16")];
            fp16 var_4508_to_fp16 = const()[name = string("op_4508_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4509_cast_fp16 = add(x = variance_95_cast_fp16, y = var_4508_to_fp16)[name = string("op_4509_cast_fp16")];
            fp32 var_4510_epsilon_0 = const()[name = string("op_4510_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_4510_cast_fp16 = rsqrt(epsilon = var_4510_epsilon_0, x = var_4509_cast_fp16)[name = string("op_4510_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_713_cast_fp16 = mul(x = hidden_states_709_cast_fp16, y = var_4510_cast_fp16)[name = string("hidden_states_713_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740042816)))];
            tensor<fp16, [1, ?, 1536]> input_187_cast_fp16 = mul(x = model_model_layers_23_post_attention_layernorm_weight_to_fp16, y = hidden_states_713_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740045952))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(746927296))))[name = string("model_model_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_4522_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_4522_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(747787520))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(754668864))))[name = string("model_model_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_166_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_191_cast_fp16 = mul(x = var_4522_cast_fp16, y = linear_166_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(755529088))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(762410432))))[name = string("model_model_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_167_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_719_cast_fp16 = add(x = hidden_states_709_cast_fp16, y = linear_167_cast_fp16)[name = string("hidden_states_719_cast_fp16")];
            fp16 var_78_promoted_48_to_fp16 = const()[name = string("op_78_promoted_48_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_4535_cast_fp16 = pow(x = hidden_states_719_cast_fp16, y = var_78_promoted_48_to_fp16)[name = string("op_4535_cast_fp16")];
            tensor<int32, [1]> variance_97_axes_0 = const()[name = string("variance_97_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_97_keep_dims_0 = const()[name = string("variance_97_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_97_cast_fp16 = reduce_mean(axes = variance_97_axes_0, keep_dims = variance_97_keep_dims_0, x = var_4535_cast_fp16)[name = string("variance_97_cast_fp16")];
            fp16 var_4538_to_fp16 = const()[name = string("op_4538_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4539_cast_fp16 = add(x = variance_97_cast_fp16, y = var_4538_to_fp16)[name = string("op_4539_cast_fp16")];
            fp32 var_4540_epsilon_0 = const()[name = string("op_4540_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_4540_cast_fp16 = rsqrt(epsilon = var_4540_epsilon_0, x = var_4539_cast_fp16)[name = string("op_4540_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_723_cast_fp16 = mul(x = hidden_states_719_cast_fp16, y = var_4540_cast_fp16)[name = string("hidden_states_723_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_24_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_24_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(763270656)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_727_cast_fp16 = mul(x = model_model_layers_24_input_layernorm_weight_to_fp16, y = hidden_states_723_cast_fp16)[name = string("hidden_states_727_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_24_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(763273792))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(764453504))))[name = string("model_model_layers_24_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_24_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_24_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(764601024)))];
            tensor<fp16, [1, ?, 1536]> linear_168_cast_fp16 = linear(bias = model_model_layers_24_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_24_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_727_cast_fp16)[name = string("linear_168_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_24_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(764604160))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(764800832))))[name = string("model_model_layers_24_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_24_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_24_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(764825472)))];
            tensor<fp16, [1, ?, 256]> linear_169_cast_fp16 = linear(bias = model_model_layers_24_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_24_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_727_cast_fp16)[name = string("linear_169_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_24_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(764826048))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(765022720))))[name = string("model_model_layers_24_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_24_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_24_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(765047360)))];
            tensor<fp16, [1, ?, 256]> linear_170_cast_fp16 = linear(bias = model_model_layers_24_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_24_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_727_cast_fp16)[name = string("linear_170_cast_fp16")];
            tensor<int32, [4]> concat_456x = const()[name = string("concat_456x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_4563_cast_fp16 = reshape(shape = concat_456x, x = linear_168_cast_fp16)[name = string("op_4563_cast_fp16")];
            tensor<int32, [4]> q_49_perm_0 = const()[name = string("q_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_457x = const()[name = string("concat_457x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_4566_cast_fp16 = reshape(shape = concat_457x, x = linear_169_cast_fp16)[name = string("op_4566_cast_fp16")];
            tensor<int32, [4]> k_49_perm_0 = const()[name = string("k_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_458x = const()[name = string("concat_458x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_4569_cast_fp16 = reshape(shape = concat_458x, x = linear_170_cast_fp16)[name = string("op_4569_cast_fp16")];
            tensor<int32, [4]> v_state_49_perm_0 = const()[name = string("v_state_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_49_cast_fp16 = transpose(perm = q_49_perm_0, x = var_4563_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 12, ?, 128]> var_4573_cast_fp16 = mul(x = q_49_cast_fp16, y = cos_7_cast_fp16)[name = string("op_4573_cast_fp16")];
            tensor<int32, [4]> x1_97_begin_0 = const()[name = string("x1_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_97_end_0 = const()[name = string("x1_97_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_97_end_mask_0 = const()[name = string("x1_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_97_cast_fp16 = slice_by_index(begin = x1_97_begin_0, end = x1_97_end_0, end_mask = x1_97_end_mask_0, x = q_49_cast_fp16)[name = string("x1_97_cast_fp16")];
            tensor<int32, [4]> x2_97_begin_0 = const()[name = string("x2_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_97_end_0 = const()[name = string("x2_97_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_97_end_mask_0 = const()[name = string("x2_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_97_cast_fp16 = slice_by_index(begin = x2_97_begin_0, end = x2_97_end_0, end_mask = x2_97_end_mask_0, x = q_49_cast_fp16)[name = string("x2_97_cast_fp16")];
            fp16 const_51_promoted_to_fp16 = const()[name = string("const_51_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_4584_cast_fp16 = mul(x = x2_97_cast_fp16, y = const_51_promoted_to_fp16)[name = string("op_4584_cast_fp16")];
            bool var_4586_interleave_0 = const()[name = string("op_4586_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_4586_cast_fp16 = concat(axis = var_72, interleave = var_4586_interleave_0, values = (var_4584_cast_fp16, x1_97_cast_fp16))[name = string("op_4586_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_4587_cast_fp16 = mul(x = var_4586_cast_fp16, y = sin_7_cast_fp16)[name = string("op_4587_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_99_cast_fp16 = add(x = var_4573_cast_fp16, y = var_4587_cast_fp16)[name = string("query_states_99_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_49_cast_fp16 = transpose(perm = k_49_perm_0, x = var_4566_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 2, ?, 128]> var_4589_cast_fp16 = mul(x = k_49_cast_fp16, y = cos_7_cast_fp16)[name = string("op_4589_cast_fp16")];
            tensor<int32, [4]> x1_99_begin_0 = const()[name = string("x1_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_99_end_0 = const()[name = string("x1_99_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_99_end_mask_0 = const()[name = string("x1_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_99_cast_fp16 = slice_by_index(begin = x1_99_begin_0, end = x1_99_end_0, end_mask = x1_99_end_mask_0, x = k_49_cast_fp16)[name = string("x1_99_cast_fp16")];
            tensor<int32, [4]> x2_99_begin_0 = const()[name = string("x2_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_99_end_0 = const()[name = string("x2_99_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_99_end_mask_0 = const()[name = string("x2_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_99_cast_fp16 = slice_by_index(begin = x2_99_begin_0, end = x2_99_end_0, end_mask = x2_99_end_mask_0, x = k_49_cast_fp16)[name = string("x2_99_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_4600_cast_fp16 = mul(x = x2_99_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_4600_cast_fp16")];
            bool var_4602_interleave_0 = const()[name = string("op_4602_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_4602_cast_fp16 = concat(axis = var_72, interleave = var_4602_interleave_0, values = (var_4600_cast_fp16, x1_99_cast_fp16))[name = string("op_4602_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_4603_cast_fp16 = mul(x = var_4602_cast_fp16, y = sin_7_cast_fp16)[name = string("op_4603_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_49_cast_fp16 = add(x = var_4589_cast_fp16, y = var_4603_cast_fp16)[name = string("k_state_49_cast_fp16")];
            tensor<int32, [1]> expand_dims_288 = const()[name = string("expand_dims_288"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_289 = const()[name = string("expand_dims_289"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_291 = const()[name = string("expand_dims_291"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_461_values0_0 = const()[name = string("concat_461_values0_0"), val = tensor<int32, [1]>([24])];
            int32 concat_461_axis_0 = const()[name = string("concat_461_axis_0"), val = int32(0)];
            bool concat_461_interleave_0 = const()[name = string("concat_461_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_461 = concat(axis = concat_461_axis_0, interleave = concat_461_interleave_0, values = (concat_461_values0_0, expand_dims_288, expand_dims_289, expand_dims_2, expand_dims_291))[name = string("concat_461")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_25_stride_0 = const()[name = string("key_cache_internal_tensor_assign_25_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_25_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_25_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_25_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_25_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_25_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_25_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_25_cast_fp16 = slice_update(begin = concat_461, begin_mask = key_cache_internal_tensor_assign_25_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_25_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_25_squeeze_mask_0, stride = key_cache_internal_tensor_assign_25_stride_0, update = k_state_49_cast_fp16, x = coreml_update_state_102)[name = string("key_cache_internal_tensor_assign_25_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_25_cast_fp16, input = key_cache)[name = string("coreml_update_state_104_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_104 = read_state(input = key_cache)[name = string("coreml_update_state_104")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_25_stride_0 = const()[name = string("value_cache_internal_tensor_assign_25_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_25_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_25_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_25_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_25_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_25_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_25_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_49_cast_fp16 = transpose(perm = v_state_49_perm_0, x = var_4569_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_25_cast_fp16 = slice_update(begin = concat_461, begin_mask = value_cache_internal_tensor_assign_25_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_25_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_25_squeeze_mask_0, stride = value_cache_internal_tensor_assign_25_stride_0, update = v_state_49_cast_fp16, x = coreml_update_state_103)[name = string("value_cache_internal_tensor_assign_25_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_25_cast_fp16, input = value_cache)[name = string("coreml_update_state_105_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_105 = read_state(input = value_cache)[name = string("coreml_update_state_105")];
            tensor<int32, [5]> var_4626_begin_0 = const()[name = string("op_4626_begin_0"), val = tensor<int32, [5]>([24, 0, 0, 0, 0])];
            tensor<int32, [5]> var_4626_end_0 = const()[name = string("op_4626_end_0"), val = tensor<int32, [5]>([25, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_4626_end_mask_0 = const()[name = string("op_4626_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_4626_squeeze_mask_0 = const()[name = string("op_4626_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_4626_cast_fp16 = slice_by_index(begin = var_4626_begin_0, end = var_4626_end_0, end_mask = var_4626_end_mask_0, squeeze_mask = var_4626_squeeze_mask_0, x = coreml_update_state_104)[name = string("op_4626_cast_fp16")];
            tensor<int32, [4]> var_4629_begin_0 = const()[name = string("op_4629_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_4629_end_mask_0 = const()[name = string("op_4629_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_4629_cast_fp16 = slice_by_index(begin = var_4629_begin_0, end = concat_11, end_mask = var_4629_end_mask_0, x = var_4626_cast_fp16)[name = string("op_4629_cast_fp16")];
            tensor<int32, [5]> var_4631_begin_0 = const()[name = string("op_4631_begin_0"), val = tensor<int32, [5]>([24, 0, 0, 0, 0])];
            tensor<int32, [5]> var_4631_end_0 = const()[name = string("op_4631_end_0"), val = tensor<int32, [5]>([25, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_4631_end_mask_0 = const()[name = string("op_4631_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_4631_squeeze_mask_0 = const()[name = string("op_4631_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_4631_cast_fp16 = slice_by_index(begin = var_4631_begin_0, end = var_4631_end_0, end_mask = var_4631_end_mask_0, squeeze_mask = var_4631_squeeze_mask_0, x = coreml_update_state_105)[name = string("op_4631_cast_fp16")];
            tensor<int32, [4]> var_4634_begin_0 = const()[name = string("op_4634_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_4634_end_mask_0 = const()[name = string("op_4634_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_4634_cast_fp16 = slice_by_index(begin = var_4634_begin_0, end = concat_11, end_mask = var_4634_end_mask_0, x = var_4631_cast_fp16)[name = string("op_4634_cast_fp16")];
            tensor<int32, [4]> var_4636_shape_cast_fp16 = shape(x = var_4629_cast_fp16)[name = string("op_4636_shape_cast_fp16")];
            int32 gather_445 = const()[name = string("gather_445"), val = int32(1)];
            int32 gather_446 = const()[name = string("gather_446"), val = int32(2)];
            int32 gather_447_axis_0 = const()[name = string("gather_447_axis_0"), val = int32(0)];
            int32 gather_447_batch_dims_0 = const()[name = string("gather_447_batch_dims_0"), val = int32(0)];
            bool gather_447_validate_indices_0 = const()[name = string("gather_447_validate_indices_0"), val = bool(false)];
            string var_4636_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4636_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_447_to_uint16 = const()[name = string("select_447_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_4636_shape_cast_fp16_to_uint16 = cast(dtype = var_4636_shape_cast_fp16_to_uint16_dtype_0, x = var_4636_shape_cast_fp16)[name = string("cast_23")];
            uint16 gather_447_cast_uint16 = gather(axis = gather_447_axis_0, batch_dims = gather_447_batch_dims_0, indices = select_447_to_uint16, validate_indices = gather_447_validate_indices_0, x = var_4636_shape_cast_fp16_to_uint16)[name = string("gather_447_cast_uint16")];
            string gather_447_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_447_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_448 = const()[name = string("gather_448"), val = int32(128)];
            tensor<int32, [1]> var_4643_axes_0 = const()[name = string("op_4643_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_4643_cast_fp16 = expand_dims(axes = var_4643_axes_0, x = var_4629_cast_fp16)[name = string("op_4643_cast_fp16")];
            tensor<int32, [5]> shape_497_cast_fp16 = shape(x = var_4643_cast_fp16)[name = string("shape_497_cast_fp16")];
            int32 concat_469_axis_0 = const()[name = string("concat_469_axis_0"), val = int32(0)];
            bool concat_469_interleave_0 = const()[name = string("concat_469_interleave_0"), val = bool(false)];
            int32 gather_447_cast_uint16_to_int32 = cast(dtype = gather_447_cast_uint16_to_int32_dtype_0, x = gather_447_cast_uint16)[name = string("cast_22")];
            tensor<int32, [5]> concat_469 = concat(axis = concat_469_axis_0, interleave = concat_469_interleave_0, values = (gather_445, gather_446, var_77, gather_447_cast_uint16_to_int32, gather_448))[name = string("concat_469")];
            tensor<int32, [5]> real_div_48 = real_div(x = concat_469, y = shape_497_cast_fp16)[name = string("real_div_48")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_731_cast_fp16 = tile(reps = real_div_48, x = var_4643_cast_fp16)[name = string("hidden_states_731_cast_fp16")];
            tensor<int32, [4]> concat_470x = const()[name = string("concat_470x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_99_cast_fp16 = reshape(shape = concat_470x, x = hidden_states_731_cast_fp16)[name = string("key_states_99_cast_fp16")];
            tensor<int32, [4]> var_4653_shape_cast_fp16 = shape(x = var_4634_cast_fp16)[name = string("op_4653_shape_cast_fp16")];
            int32 gather_449 = const()[name = string("gather_449"), val = int32(1)];
            int32 gather_450 = const()[name = string("gather_450"), val = int32(2)];
            int32 gather_451_axis_0 = const()[name = string("gather_451_axis_0"), val = int32(0)];
            int32 gather_451_batch_dims_0 = const()[name = string("gather_451_batch_dims_0"), val = int32(0)];
            bool gather_451_validate_indices_0 = const()[name = string("gather_451_validate_indices_0"), val = bool(false)];
            string var_4653_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4653_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_451_to_uint16 = const()[name = string("select_451_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_4653_shape_cast_fp16_to_uint16 = cast(dtype = var_4653_shape_cast_fp16_to_uint16_dtype_0, x = var_4653_shape_cast_fp16)[name = string("cast_21")];
            uint16 gather_451_cast_uint16 = gather(axis = gather_451_axis_0, batch_dims = gather_451_batch_dims_0, indices = select_451_to_uint16, validate_indices = gather_451_validate_indices_0, x = var_4653_shape_cast_fp16_to_uint16)[name = string("gather_451_cast_uint16")];
            string gather_451_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_451_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_452 = const()[name = string("gather_452"), val = int32(128)];
            tensor<int32, [1]> var_4660_axes_0 = const()[name = string("op_4660_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_4660_cast_fp16 = expand_dims(axes = var_4660_axes_0, x = var_4634_cast_fp16)[name = string("op_4660_cast_fp16")];
            tensor<int32, [5]> shape_502_cast_fp16 = shape(x = var_4660_cast_fp16)[name = string("shape_502_cast_fp16")];
            int32 concat_471_axis_0 = const()[name = string("concat_471_axis_0"), val = int32(0)];
            bool concat_471_interleave_0 = const()[name = string("concat_471_interleave_0"), val = bool(false)];
            int32 gather_451_cast_uint16_to_int32 = cast(dtype = gather_451_cast_uint16_to_int32_dtype_0, x = gather_451_cast_uint16)[name = string("cast_20")];
            tensor<int32, [5]> concat_471 = concat(axis = concat_471_axis_0, interleave = concat_471_interleave_0, values = (gather_449, gather_450, var_77, gather_451_cast_uint16_to_int32, gather_452))[name = string("concat_471")];
            tensor<int32, [5]> real_div_49 = real_div(x = concat_471, y = shape_502_cast_fp16)[name = string("real_div_49")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_735_cast_fp16 = tile(reps = real_div_49, x = var_4660_cast_fp16)[name = string("hidden_states_735_cast_fp16")];
            tensor<int32, [4]> concat_472x = const()[name = string("concat_472x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_99_cast_fp16 = reshape(shape = concat_472x, x = hidden_states_735_cast_fp16)[name = string("value_states_99_cast_fp16")];
            tensor<int32, [4]> var_4670_shape_cast_fp16 = shape(x = key_states_99_cast_fp16)[name = string("op_4670_shape_cast_fp16")];
            int32 gather_453_axis_0 = const()[name = string("gather_453_axis_0"), val = int32(0)];
            int32 gather_453_batch_dims_0 = const()[name = string("gather_453_batch_dims_0"), val = int32(0)];
            bool gather_453_validate_indices_0 = const()[name = string("gather_453_validate_indices_0"), val = bool(false)];
            string var_4670_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4670_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_453_to_uint16 = const()[name = string("select_453_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_4670_shape_cast_fp16_to_uint16 = cast(dtype = var_4670_shape_cast_fp16_to_uint16_dtype_0, x = var_4670_shape_cast_fp16)[name = string("cast_19")];
            uint16 gather_453_cast_uint16 = gather(axis = gather_453_axis_0, batch_dims = gather_453_batch_dims_0, indices = select_453_to_uint16, validate_indices = gather_453_validate_indices_0, x = var_4670_shape_cast_fp16_to_uint16)[name = string("gather_453_cast_uint16")];
            string gather_453_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_453_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_473_values0_0 = const()[name = string("concat_473_values0_0"), val = int32(1)];
            int32 concat_473_values1_0 = const()[name = string("concat_473_values1_0"), val = int32(1)];
            int32 concat_473_values2_0 = const()[name = string("concat_473_values2_0"), val = int32(0)];
            int32 concat_473_axis_0 = const()[name = string("concat_473_axis_0"), val = int32(0)];
            bool concat_473_interleave_0 = const()[name = string("concat_473_interleave_0"), val = bool(false)];
            int32 gather_453_cast_uint16_to_int32 = cast(dtype = gather_453_cast_uint16_to_int32_dtype_0, x = gather_453_cast_uint16)[name = string("cast_18")];
            tensor<int32, [4]> concat_473 = concat(axis = concat_473_axis_0, interleave = concat_473_interleave_0, values = (concat_473_values0_0, concat_473_values1_0, concat_473_values2_0, gather_453_cast_uint16_to_int32))[name = string("concat_473")];
            tensor<int32, [4]> causal_mask_51_begin_0 = const()[name = string("causal_mask_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_51_end_mask_0 = const()[name = string("causal_mask_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_51_cast_fp16 = slice_by_index(begin = causal_mask_51_begin_0, end = concat_473, end_mask = causal_mask_51_end_mask_0, x = causal_mask)[name = string("causal_mask_51_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_97_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_51_cast_fp16, key = key_states_99_cast_fp16, query = query_states_99_cast_fp16, value = value_states_99_cast_fp16)[name = string("attn_output_97_cast_fp16")];
            tensor<int32, [4]> var_4676_perm_0 = const()[name = string("op_4676_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_474x = const()[name = string("concat_474x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_4676_cast_fp16 = transpose(perm = var_4676_perm_0, x = attn_output_97_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, ?, 1536]> input_193_cast_fp16 = reshape(shape = concat_474x, x = var_4676_cast_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_24_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(765047936))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(766227648))))[name = string("model_model_layers_24_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_171_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_24_self_attn_o_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_171_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_739_cast_fp16 = add(x = hidden_states_719_cast_fp16, y = linear_171_cast_fp16)[name = string("hidden_states_739_cast_fp16")];
            fp16 var_78_promoted_49_to_fp16 = const()[name = string("op_78_promoted_49_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_4685_cast_fp16 = pow(x = hidden_states_739_cast_fp16, y = var_78_promoted_49_to_fp16)[name = string("op_4685_cast_fp16")];
            tensor<int32, [1]> variance_99_axes_0 = const()[name = string("variance_99_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_99_keep_dims_0 = const()[name = string("variance_99_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_99_cast_fp16 = reduce_mean(axes = variance_99_axes_0, keep_dims = variance_99_keep_dims_0, x = var_4685_cast_fp16)[name = string("variance_99_cast_fp16")];
            fp16 var_4688_to_fp16 = const()[name = string("op_4688_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4689_cast_fp16 = add(x = variance_99_cast_fp16, y = var_4688_to_fp16)[name = string("op_4689_cast_fp16")];
            fp32 var_4690_epsilon_0 = const()[name = string("op_4690_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_4690_cast_fp16 = rsqrt(epsilon = var_4690_epsilon_0, x = var_4689_cast_fp16)[name = string("op_4690_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_743_cast_fp16 = mul(x = hidden_states_739_cast_fp16, y = var_4690_cast_fp16)[name = string("hidden_states_743_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_24_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_24_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(766375168)))];
            tensor<fp16, [1, ?, 1536]> input_195_cast_fp16 = mul(x = model_model_layers_24_post_attention_layernorm_weight_to_fp16, y = hidden_states_743_cast_fp16)[name = string("input_195_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_24_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(766378304))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(773259648))))[name = string("model_model_layers_24_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_172_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_24_mlp_gate_proj_weight_to_fp16_quantized, x = input_195_cast_fp16)[name = string("linear_172_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_4702_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("op_4702_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_24_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(774119872))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(781001216))))[name = string("model_model_layers_24_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_173_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_24_mlp_up_proj_weight_to_fp16_quantized, x = input_195_cast_fp16)[name = string("linear_173_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_199_cast_fp16 = mul(x = var_4702_cast_fp16, y = linear_173_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_24_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(781861440))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788742784))))[name = string("model_model_layers_24_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_174_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_24_mlp_down_proj_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_174_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_749_cast_fp16 = add(x = hidden_states_739_cast_fp16, y = linear_174_cast_fp16)[name = string("hidden_states_749_cast_fp16")];
            fp16 var_78_promoted_50_to_fp16 = const()[name = string("op_78_promoted_50_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_4715_cast_fp16 = pow(x = hidden_states_749_cast_fp16, y = var_78_promoted_50_to_fp16)[name = string("op_4715_cast_fp16")];
            tensor<int32, [1]> variance_101_axes_0 = const()[name = string("variance_101_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_101_keep_dims_0 = const()[name = string("variance_101_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_101_cast_fp16 = reduce_mean(axes = variance_101_axes_0, keep_dims = variance_101_keep_dims_0, x = var_4715_cast_fp16)[name = string("variance_101_cast_fp16")];
            fp16 var_4718_to_fp16 = const()[name = string("op_4718_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4719_cast_fp16 = add(x = variance_101_cast_fp16, y = var_4718_to_fp16)[name = string("op_4719_cast_fp16")];
            fp32 var_4720_epsilon_0 = const()[name = string("op_4720_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_4720_cast_fp16 = rsqrt(epsilon = var_4720_epsilon_0, x = var_4719_cast_fp16)[name = string("op_4720_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_753_cast_fp16 = mul(x = hidden_states_749_cast_fp16, y = var_4720_cast_fp16)[name = string("hidden_states_753_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_25_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_25_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(789603008)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_757_cast_fp16 = mul(x = model_model_layers_25_input_layernorm_weight_to_fp16, y = hidden_states_753_cast_fp16)[name = string("hidden_states_757_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_25_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(789606144))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(790785856))))[name = string("model_model_layers_25_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_25_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_25_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(790933376)))];
            tensor<fp16, [1, ?, 1536]> linear_175_cast_fp16 = linear(bias = model_model_layers_25_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_25_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_757_cast_fp16)[name = string("linear_175_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_25_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(790936512))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791133184))))[name = string("model_model_layers_25_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_25_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_25_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791157824)))];
            tensor<fp16, [1, ?, 256]> linear_176_cast_fp16 = linear(bias = model_model_layers_25_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_25_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_757_cast_fp16)[name = string("linear_176_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_25_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791158400))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791355072))))[name = string("model_model_layers_25_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_25_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_25_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791379712)))];
            tensor<fp16, [1, ?, 256]> linear_177_cast_fp16 = linear(bias = model_model_layers_25_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_25_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_757_cast_fp16)[name = string("linear_177_cast_fp16")];
            tensor<int32, [4]> concat_475x = const()[name = string("concat_475x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_4743_cast_fp16 = reshape(shape = concat_475x, x = linear_175_cast_fp16)[name = string("op_4743_cast_fp16")];
            tensor<int32, [4]> q_51_perm_0 = const()[name = string("q_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_476x = const()[name = string("concat_476x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_4746_cast_fp16 = reshape(shape = concat_476x, x = linear_176_cast_fp16)[name = string("op_4746_cast_fp16")];
            tensor<int32, [4]> k_51_perm_0 = const()[name = string("k_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_477x = const()[name = string("concat_477x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_4749_cast_fp16 = reshape(shape = concat_477x, x = linear_177_cast_fp16)[name = string("op_4749_cast_fp16")];
            tensor<int32, [4]> v_state_51_perm_0 = const()[name = string("v_state_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_51_cast_fp16 = transpose(perm = q_51_perm_0, x = var_4743_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 12, ?, 128]> var_4753_cast_fp16 = mul(x = q_51_cast_fp16, y = cos_7_cast_fp16)[name = string("op_4753_cast_fp16")];
            tensor<int32, [4]> x1_101_begin_0 = const()[name = string("x1_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_101_end_0 = const()[name = string("x1_101_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_101_end_mask_0 = const()[name = string("x1_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_101_cast_fp16 = slice_by_index(begin = x1_101_begin_0, end = x1_101_end_0, end_mask = x1_101_end_mask_0, x = q_51_cast_fp16)[name = string("x1_101_cast_fp16")];
            tensor<int32, [4]> x2_101_begin_0 = const()[name = string("x2_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_101_end_0 = const()[name = string("x2_101_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_101_end_mask_0 = const()[name = string("x2_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_101_cast_fp16 = slice_by_index(begin = x2_101_begin_0, end = x2_101_end_0, end_mask = x2_101_end_mask_0, x = q_51_cast_fp16)[name = string("x2_101_cast_fp16")];
            fp16 const_53_promoted_to_fp16 = const()[name = string("const_53_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_4764_cast_fp16 = mul(x = x2_101_cast_fp16, y = const_53_promoted_to_fp16)[name = string("op_4764_cast_fp16")];
            bool var_4766_interleave_0 = const()[name = string("op_4766_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_4766_cast_fp16 = concat(axis = var_72, interleave = var_4766_interleave_0, values = (var_4764_cast_fp16, x1_101_cast_fp16))[name = string("op_4766_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_4767_cast_fp16 = mul(x = var_4766_cast_fp16, y = sin_7_cast_fp16)[name = string("op_4767_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_103_cast_fp16 = add(x = var_4753_cast_fp16, y = var_4767_cast_fp16)[name = string("query_states_103_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_51_cast_fp16 = transpose(perm = k_51_perm_0, x = var_4746_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 2, ?, 128]> var_4769_cast_fp16 = mul(x = k_51_cast_fp16, y = cos_7_cast_fp16)[name = string("op_4769_cast_fp16")];
            tensor<int32, [4]> x1_103_begin_0 = const()[name = string("x1_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_103_end_0 = const()[name = string("x1_103_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_103_end_mask_0 = const()[name = string("x1_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_103_cast_fp16 = slice_by_index(begin = x1_103_begin_0, end = x1_103_end_0, end_mask = x1_103_end_mask_0, x = k_51_cast_fp16)[name = string("x1_103_cast_fp16")];
            tensor<int32, [4]> x2_103_begin_0 = const()[name = string("x2_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_103_end_0 = const()[name = string("x2_103_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_103_end_mask_0 = const()[name = string("x2_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_103_cast_fp16 = slice_by_index(begin = x2_103_begin_0, end = x2_103_end_0, end_mask = x2_103_end_mask_0, x = k_51_cast_fp16)[name = string("x2_103_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_4780_cast_fp16 = mul(x = x2_103_cast_fp16, y = const_54_promoted_to_fp16)[name = string("op_4780_cast_fp16")];
            bool var_4782_interleave_0 = const()[name = string("op_4782_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_4782_cast_fp16 = concat(axis = var_72, interleave = var_4782_interleave_0, values = (var_4780_cast_fp16, x1_103_cast_fp16))[name = string("op_4782_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_4783_cast_fp16 = mul(x = var_4782_cast_fp16, y = sin_7_cast_fp16)[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_51_cast_fp16 = add(x = var_4769_cast_fp16, y = var_4783_cast_fp16)[name = string("k_state_51_cast_fp16")];
            tensor<int32, [1]> expand_dims_300 = const()[name = string("expand_dims_300"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_301 = const()[name = string("expand_dims_301"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_303 = const()[name = string("expand_dims_303"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_480_values0_0 = const()[name = string("concat_480_values0_0"), val = tensor<int32, [1]>([25])];
            int32 concat_480_axis_0 = const()[name = string("concat_480_axis_0"), val = int32(0)];
            bool concat_480_interleave_0 = const()[name = string("concat_480_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_480 = concat(axis = concat_480_axis_0, interleave = concat_480_interleave_0, values = (concat_480_values0_0, expand_dims_300, expand_dims_301, expand_dims_2, expand_dims_303))[name = string("concat_480")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_26_stride_0 = const()[name = string("key_cache_internal_tensor_assign_26_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_26_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_26_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_26_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_26_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_26_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_26_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_26_cast_fp16 = slice_update(begin = concat_480, begin_mask = key_cache_internal_tensor_assign_26_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_26_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_26_squeeze_mask_0, stride = key_cache_internal_tensor_assign_26_stride_0, update = k_state_51_cast_fp16, x = coreml_update_state_104)[name = string("key_cache_internal_tensor_assign_26_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_26_cast_fp16, input = key_cache)[name = string("coreml_update_state_106_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_106 = read_state(input = key_cache)[name = string("coreml_update_state_106")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_26_stride_0 = const()[name = string("value_cache_internal_tensor_assign_26_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_26_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_26_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_26_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_26_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_26_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_26_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_51_cast_fp16 = transpose(perm = v_state_51_perm_0, x = var_4749_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_26_cast_fp16 = slice_update(begin = concat_480, begin_mask = value_cache_internal_tensor_assign_26_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_26_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_26_squeeze_mask_0, stride = value_cache_internal_tensor_assign_26_stride_0, update = v_state_51_cast_fp16, x = coreml_update_state_105)[name = string("value_cache_internal_tensor_assign_26_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_26_cast_fp16, input = value_cache)[name = string("coreml_update_state_107_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_107 = read_state(input = value_cache)[name = string("coreml_update_state_107")];
            tensor<int32, [5]> var_4806_begin_0 = const()[name = string("op_4806_begin_0"), val = tensor<int32, [5]>([25, 0, 0, 0, 0])];
            tensor<int32, [5]> var_4806_end_0 = const()[name = string("op_4806_end_0"), val = tensor<int32, [5]>([26, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_4806_end_mask_0 = const()[name = string("op_4806_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_4806_squeeze_mask_0 = const()[name = string("op_4806_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_4806_cast_fp16 = slice_by_index(begin = var_4806_begin_0, end = var_4806_end_0, end_mask = var_4806_end_mask_0, squeeze_mask = var_4806_squeeze_mask_0, x = coreml_update_state_106)[name = string("op_4806_cast_fp16")];
            tensor<int32, [4]> var_4809_begin_0 = const()[name = string("op_4809_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_4809_end_mask_0 = const()[name = string("op_4809_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_4809_cast_fp16 = slice_by_index(begin = var_4809_begin_0, end = concat_11, end_mask = var_4809_end_mask_0, x = var_4806_cast_fp16)[name = string("op_4809_cast_fp16")];
            tensor<int32, [5]> var_4811_begin_0 = const()[name = string("op_4811_begin_0"), val = tensor<int32, [5]>([25, 0, 0, 0, 0])];
            tensor<int32, [5]> var_4811_end_0 = const()[name = string("op_4811_end_0"), val = tensor<int32, [5]>([26, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_4811_end_mask_0 = const()[name = string("op_4811_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_4811_squeeze_mask_0 = const()[name = string("op_4811_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_4811_cast_fp16 = slice_by_index(begin = var_4811_begin_0, end = var_4811_end_0, end_mask = var_4811_end_mask_0, squeeze_mask = var_4811_squeeze_mask_0, x = coreml_update_state_107)[name = string("op_4811_cast_fp16")];
            tensor<int32, [4]> var_4814_begin_0 = const()[name = string("op_4814_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_4814_end_mask_0 = const()[name = string("op_4814_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_4814_cast_fp16 = slice_by_index(begin = var_4814_begin_0, end = concat_11, end_mask = var_4814_end_mask_0, x = var_4811_cast_fp16)[name = string("op_4814_cast_fp16")];
            tensor<int32, [4]> var_4816_shape_cast_fp16 = shape(x = var_4809_cast_fp16)[name = string("op_4816_shape_cast_fp16")];
            int32 gather_463 = const()[name = string("gather_463"), val = int32(1)];
            int32 gather_464 = const()[name = string("gather_464"), val = int32(2)];
            int32 gather_465_axis_0 = const()[name = string("gather_465_axis_0"), val = int32(0)];
            int32 gather_465_batch_dims_0 = const()[name = string("gather_465_batch_dims_0"), val = int32(0)];
            bool gather_465_validate_indices_0 = const()[name = string("gather_465_validate_indices_0"), val = bool(false)];
            string var_4816_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4816_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_465_to_uint16 = const()[name = string("select_465_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_4816_shape_cast_fp16_to_uint16 = cast(dtype = var_4816_shape_cast_fp16_to_uint16_dtype_0, x = var_4816_shape_cast_fp16)[name = string("cast_17")];
            uint16 gather_465_cast_uint16 = gather(axis = gather_465_axis_0, batch_dims = gather_465_batch_dims_0, indices = select_465_to_uint16, validate_indices = gather_465_validate_indices_0, x = var_4816_shape_cast_fp16_to_uint16)[name = string("gather_465_cast_uint16")];
            string gather_465_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_465_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_466 = const()[name = string("gather_466"), val = int32(128)];
            tensor<int32, [1]> var_4823_axes_0 = const()[name = string("op_4823_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_4823_cast_fp16 = expand_dims(axes = var_4823_axes_0, x = var_4809_cast_fp16)[name = string("op_4823_cast_fp16")];
            tensor<int32, [5]> shape_517_cast_fp16 = shape(x = var_4823_cast_fp16)[name = string("shape_517_cast_fp16")];
            int32 concat_488_axis_0 = const()[name = string("concat_488_axis_0"), val = int32(0)];
            bool concat_488_interleave_0 = const()[name = string("concat_488_interleave_0"), val = bool(false)];
            int32 gather_465_cast_uint16_to_int32 = cast(dtype = gather_465_cast_uint16_to_int32_dtype_0, x = gather_465_cast_uint16)[name = string("cast_16")];
            tensor<int32, [5]> concat_488 = concat(axis = concat_488_axis_0, interleave = concat_488_interleave_0, values = (gather_463, gather_464, var_77, gather_465_cast_uint16_to_int32, gather_466))[name = string("concat_488")];
            tensor<int32, [5]> real_div_50 = real_div(x = concat_488, y = shape_517_cast_fp16)[name = string("real_div_50")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_761_cast_fp16 = tile(reps = real_div_50, x = var_4823_cast_fp16)[name = string("hidden_states_761_cast_fp16")];
            tensor<int32, [4]> concat_489x = const()[name = string("concat_489x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_103_cast_fp16 = reshape(shape = concat_489x, x = hidden_states_761_cast_fp16)[name = string("key_states_103_cast_fp16")];
            tensor<int32, [4]> var_4833_shape_cast_fp16 = shape(x = var_4814_cast_fp16)[name = string("op_4833_shape_cast_fp16")];
            int32 gather_467 = const()[name = string("gather_467"), val = int32(1)];
            int32 gather_468 = const()[name = string("gather_468"), val = int32(2)];
            int32 gather_469_axis_0 = const()[name = string("gather_469_axis_0"), val = int32(0)];
            int32 gather_469_batch_dims_0 = const()[name = string("gather_469_batch_dims_0"), val = int32(0)];
            bool gather_469_validate_indices_0 = const()[name = string("gather_469_validate_indices_0"), val = bool(false)];
            string var_4833_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4833_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_469_to_uint16 = const()[name = string("select_469_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_4833_shape_cast_fp16_to_uint16 = cast(dtype = var_4833_shape_cast_fp16_to_uint16_dtype_0, x = var_4833_shape_cast_fp16)[name = string("cast_15")];
            uint16 gather_469_cast_uint16 = gather(axis = gather_469_axis_0, batch_dims = gather_469_batch_dims_0, indices = select_469_to_uint16, validate_indices = gather_469_validate_indices_0, x = var_4833_shape_cast_fp16_to_uint16)[name = string("gather_469_cast_uint16")];
            string gather_469_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_469_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_470 = const()[name = string("gather_470"), val = int32(128)];
            tensor<int32, [1]> var_4840_axes_0 = const()[name = string("op_4840_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_4840_cast_fp16 = expand_dims(axes = var_4840_axes_0, x = var_4814_cast_fp16)[name = string("op_4840_cast_fp16")];
            tensor<int32, [5]> shape_522_cast_fp16 = shape(x = var_4840_cast_fp16)[name = string("shape_522_cast_fp16")];
            int32 concat_490_axis_0 = const()[name = string("concat_490_axis_0"), val = int32(0)];
            bool concat_490_interleave_0 = const()[name = string("concat_490_interleave_0"), val = bool(false)];
            int32 gather_469_cast_uint16_to_int32 = cast(dtype = gather_469_cast_uint16_to_int32_dtype_0, x = gather_469_cast_uint16)[name = string("cast_14")];
            tensor<int32, [5]> concat_490 = concat(axis = concat_490_axis_0, interleave = concat_490_interleave_0, values = (gather_467, gather_468, var_77, gather_469_cast_uint16_to_int32, gather_470))[name = string("concat_490")];
            tensor<int32, [5]> real_div_51 = real_div(x = concat_490, y = shape_522_cast_fp16)[name = string("real_div_51")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_765_cast_fp16 = tile(reps = real_div_51, x = var_4840_cast_fp16)[name = string("hidden_states_765_cast_fp16")];
            tensor<int32, [4]> concat_491x = const()[name = string("concat_491x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_103_cast_fp16 = reshape(shape = concat_491x, x = hidden_states_765_cast_fp16)[name = string("value_states_103_cast_fp16")];
            tensor<int32, [4]> var_4850_shape_cast_fp16 = shape(x = key_states_103_cast_fp16)[name = string("op_4850_shape_cast_fp16")];
            int32 gather_471_axis_0 = const()[name = string("gather_471_axis_0"), val = int32(0)];
            int32 gather_471_batch_dims_0 = const()[name = string("gather_471_batch_dims_0"), val = int32(0)];
            bool gather_471_validate_indices_0 = const()[name = string("gather_471_validate_indices_0"), val = bool(false)];
            string var_4850_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4850_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_471_to_uint16 = const()[name = string("select_471_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_4850_shape_cast_fp16_to_uint16 = cast(dtype = var_4850_shape_cast_fp16_to_uint16_dtype_0, x = var_4850_shape_cast_fp16)[name = string("cast_13")];
            uint16 gather_471_cast_uint16 = gather(axis = gather_471_axis_0, batch_dims = gather_471_batch_dims_0, indices = select_471_to_uint16, validate_indices = gather_471_validate_indices_0, x = var_4850_shape_cast_fp16_to_uint16)[name = string("gather_471_cast_uint16")];
            string gather_471_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_471_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_492_values0_0 = const()[name = string("concat_492_values0_0"), val = int32(1)];
            int32 concat_492_values1_0 = const()[name = string("concat_492_values1_0"), val = int32(1)];
            int32 concat_492_values2_0 = const()[name = string("concat_492_values2_0"), val = int32(0)];
            int32 concat_492_axis_0 = const()[name = string("concat_492_axis_0"), val = int32(0)];
            bool concat_492_interleave_0 = const()[name = string("concat_492_interleave_0"), val = bool(false)];
            int32 gather_471_cast_uint16_to_int32 = cast(dtype = gather_471_cast_uint16_to_int32_dtype_0, x = gather_471_cast_uint16)[name = string("cast_12")];
            tensor<int32, [4]> concat_492 = concat(axis = concat_492_axis_0, interleave = concat_492_interleave_0, values = (concat_492_values0_0, concat_492_values1_0, concat_492_values2_0, gather_471_cast_uint16_to_int32))[name = string("concat_492")];
            tensor<int32, [4]> causal_mask_53_begin_0 = const()[name = string("causal_mask_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_53_end_mask_0 = const()[name = string("causal_mask_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_53_cast_fp16 = slice_by_index(begin = causal_mask_53_begin_0, end = concat_492, end_mask = causal_mask_53_end_mask_0, x = causal_mask)[name = string("causal_mask_53_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_101_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_53_cast_fp16, key = key_states_103_cast_fp16, query = query_states_103_cast_fp16, value = value_states_103_cast_fp16)[name = string("attn_output_101_cast_fp16")];
            tensor<int32, [4]> var_4856_perm_0 = const()[name = string("op_4856_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_493x = const()[name = string("concat_493x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_4856_cast_fp16 = transpose(perm = var_4856_perm_0, x = attn_output_101_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, ?, 1536]> input_201_cast_fp16 = reshape(shape = concat_493x, x = var_4856_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_25_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(791380288))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792560000))))[name = string("model_model_layers_25_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_178_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_25_self_attn_o_proj_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("linear_178_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_769_cast_fp16 = add(x = hidden_states_749_cast_fp16, y = linear_178_cast_fp16)[name = string("hidden_states_769_cast_fp16")];
            fp16 var_78_promoted_51_to_fp16 = const()[name = string("op_78_promoted_51_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_4865_cast_fp16 = pow(x = hidden_states_769_cast_fp16, y = var_78_promoted_51_to_fp16)[name = string("op_4865_cast_fp16")];
            tensor<int32, [1]> variance_103_axes_0 = const()[name = string("variance_103_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_103_keep_dims_0 = const()[name = string("variance_103_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_103_cast_fp16 = reduce_mean(axes = variance_103_axes_0, keep_dims = variance_103_keep_dims_0, x = var_4865_cast_fp16)[name = string("variance_103_cast_fp16")];
            fp16 var_4868_to_fp16 = const()[name = string("op_4868_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4869_cast_fp16 = add(x = variance_103_cast_fp16, y = var_4868_to_fp16)[name = string("op_4869_cast_fp16")];
            fp32 var_4870_epsilon_0 = const()[name = string("op_4870_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_4870_cast_fp16 = rsqrt(epsilon = var_4870_epsilon_0, x = var_4869_cast_fp16)[name = string("op_4870_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_773_cast_fp16 = mul(x = hidden_states_769_cast_fp16, y = var_4870_cast_fp16)[name = string("hidden_states_773_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_25_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_25_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792707520)))];
            tensor<fp16, [1, ?, 1536]> input_203_cast_fp16 = mul(x = model_model_layers_25_post_attention_layernorm_weight_to_fp16, y = hidden_states_773_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_25_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792710656))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(799592000))))[name = string("model_model_layers_25_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_179_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_25_mlp_gate_proj_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_179_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_4882_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("op_4882_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_25_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(800452224))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(807333568))))[name = string("model_model_layers_25_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_180_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_25_mlp_up_proj_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_180_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_207_cast_fp16 = mul(x = var_4882_cast_fp16, y = linear_180_cast_fp16)[name = string("input_207_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_25_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(808193792))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(815075136))))[name = string("model_model_layers_25_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_181_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_25_mlp_down_proj_weight_to_fp16_quantized, x = input_207_cast_fp16)[name = string("linear_181_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_779_cast_fp16 = add(x = hidden_states_769_cast_fp16, y = linear_181_cast_fp16)[name = string("hidden_states_779_cast_fp16")];
            fp16 var_78_promoted_52_to_fp16 = const()[name = string("op_78_promoted_52_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_4895_cast_fp16 = pow(x = hidden_states_779_cast_fp16, y = var_78_promoted_52_to_fp16)[name = string("op_4895_cast_fp16")];
            tensor<int32, [1]> variance_105_axes_0 = const()[name = string("variance_105_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_105_keep_dims_0 = const()[name = string("variance_105_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_105_cast_fp16 = reduce_mean(axes = variance_105_axes_0, keep_dims = variance_105_keep_dims_0, x = var_4895_cast_fp16)[name = string("variance_105_cast_fp16")];
            fp16 var_4898_to_fp16 = const()[name = string("op_4898_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4899_cast_fp16 = add(x = variance_105_cast_fp16, y = var_4898_to_fp16)[name = string("op_4899_cast_fp16")];
            fp32 var_4900_epsilon_0 = const()[name = string("op_4900_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_4900_cast_fp16 = rsqrt(epsilon = var_4900_epsilon_0, x = var_4899_cast_fp16)[name = string("op_4900_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_783_cast_fp16 = mul(x = hidden_states_779_cast_fp16, y = var_4900_cast_fp16)[name = string("hidden_states_783_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_26_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_26_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(815935360)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_787_cast_fp16 = mul(x = model_model_layers_26_input_layernorm_weight_to_fp16, y = hidden_states_783_cast_fp16)[name = string("hidden_states_787_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_26_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(815938496))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(817118208))))[name = string("model_model_layers_26_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_26_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_26_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(817265728)))];
            tensor<fp16, [1, ?, 1536]> linear_182_cast_fp16 = linear(bias = model_model_layers_26_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_26_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_787_cast_fp16)[name = string("linear_182_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_26_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(817268864))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(817465536))))[name = string("model_model_layers_26_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_26_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_26_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(817490176)))];
            tensor<fp16, [1, ?, 256]> linear_183_cast_fp16 = linear(bias = model_model_layers_26_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_26_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_787_cast_fp16)[name = string("linear_183_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_26_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(817490752))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(817687424))))[name = string("model_model_layers_26_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_26_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_26_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(817712064)))];
            tensor<fp16, [1, ?, 256]> linear_184_cast_fp16 = linear(bias = model_model_layers_26_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_26_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_787_cast_fp16)[name = string("linear_184_cast_fp16")];
            tensor<int32, [4]> concat_494x = const()[name = string("concat_494x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_4923_cast_fp16 = reshape(shape = concat_494x, x = linear_182_cast_fp16)[name = string("op_4923_cast_fp16")];
            tensor<int32, [4]> q_53_perm_0 = const()[name = string("q_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_495x = const()[name = string("concat_495x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_4926_cast_fp16 = reshape(shape = concat_495x, x = linear_183_cast_fp16)[name = string("op_4926_cast_fp16")];
            tensor<int32, [4]> k_53_perm_0 = const()[name = string("k_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_496x = const()[name = string("concat_496x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_4929_cast_fp16 = reshape(shape = concat_496x, x = linear_184_cast_fp16)[name = string("op_4929_cast_fp16")];
            tensor<int32, [4]> v_state_53_perm_0 = const()[name = string("v_state_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_53_cast_fp16 = transpose(perm = q_53_perm_0, x = var_4923_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 12, ?, 128]> var_4933_cast_fp16 = mul(x = q_53_cast_fp16, y = cos_7_cast_fp16)[name = string("op_4933_cast_fp16")];
            tensor<int32, [4]> x1_105_begin_0 = const()[name = string("x1_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_105_end_0 = const()[name = string("x1_105_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_105_end_mask_0 = const()[name = string("x1_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_105_cast_fp16 = slice_by_index(begin = x1_105_begin_0, end = x1_105_end_0, end_mask = x1_105_end_mask_0, x = q_53_cast_fp16)[name = string("x1_105_cast_fp16")];
            tensor<int32, [4]> x2_105_begin_0 = const()[name = string("x2_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_105_end_0 = const()[name = string("x2_105_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_105_end_mask_0 = const()[name = string("x2_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_105_cast_fp16 = slice_by_index(begin = x2_105_begin_0, end = x2_105_end_0, end_mask = x2_105_end_mask_0, x = q_53_cast_fp16)[name = string("x2_105_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_4944_cast_fp16 = mul(x = x2_105_cast_fp16, y = const_55_promoted_to_fp16)[name = string("op_4944_cast_fp16")];
            bool var_4946_interleave_0 = const()[name = string("op_4946_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_4946_cast_fp16 = concat(axis = var_72, interleave = var_4946_interleave_0, values = (var_4944_cast_fp16, x1_105_cast_fp16))[name = string("op_4946_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_4947_cast_fp16 = mul(x = var_4946_cast_fp16, y = sin_7_cast_fp16)[name = string("op_4947_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_107_cast_fp16 = add(x = var_4933_cast_fp16, y = var_4947_cast_fp16)[name = string("query_states_107_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_53_cast_fp16 = transpose(perm = k_53_perm_0, x = var_4926_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 2, ?, 128]> var_4949_cast_fp16 = mul(x = k_53_cast_fp16, y = cos_7_cast_fp16)[name = string("op_4949_cast_fp16")];
            tensor<int32, [4]> x1_107_begin_0 = const()[name = string("x1_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_107_end_0 = const()[name = string("x1_107_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_107_end_mask_0 = const()[name = string("x1_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_107_cast_fp16 = slice_by_index(begin = x1_107_begin_0, end = x1_107_end_0, end_mask = x1_107_end_mask_0, x = k_53_cast_fp16)[name = string("x1_107_cast_fp16")];
            tensor<int32, [4]> x2_107_begin_0 = const()[name = string("x2_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_107_end_0 = const()[name = string("x2_107_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_107_end_mask_0 = const()[name = string("x2_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_107_cast_fp16 = slice_by_index(begin = x2_107_begin_0, end = x2_107_end_0, end_mask = x2_107_end_mask_0, x = k_53_cast_fp16)[name = string("x2_107_cast_fp16")];
            fp16 const_56_promoted_to_fp16 = const()[name = string("const_56_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_4960_cast_fp16 = mul(x = x2_107_cast_fp16, y = const_56_promoted_to_fp16)[name = string("op_4960_cast_fp16")];
            bool var_4962_interleave_0 = const()[name = string("op_4962_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_4962_cast_fp16 = concat(axis = var_72, interleave = var_4962_interleave_0, values = (var_4960_cast_fp16, x1_107_cast_fp16))[name = string("op_4962_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_4963_cast_fp16 = mul(x = var_4962_cast_fp16, y = sin_7_cast_fp16)[name = string("op_4963_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_53_cast_fp16 = add(x = var_4949_cast_fp16, y = var_4963_cast_fp16)[name = string("k_state_53_cast_fp16")];
            tensor<int32, [1]> expand_dims_312 = const()[name = string("expand_dims_312"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_313 = const()[name = string("expand_dims_313"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_315 = const()[name = string("expand_dims_315"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_499_values0_0 = const()[name = string("concat_499_values0_0"), val = tensor<int32, [1]>([26])];
            int32 concat_499_axis_0 = const()[name = string("concat_499_axis_0"), val = int32(0)];
            bool concat_499_interleave_0 = const()[name = string("concat_499_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_499 = concat(axis = concat_499_axis_0, interleave = concat_499_interleave_0, values = (concat_499_values0_0, expand_dims_312, expand_dims_313, expand_dims_2, expand_dims_315))[name = string("concat_499")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_27_stride_0 = const()[name = string("key_cache_internal_tensor_assign_27_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_27_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_27_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_27_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_27_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_27_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_27_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_27_cast_fp16 = slice_update(begin = concat_499, begin_mask = key_cache_internal_tensor_assign_27_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_27_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_27_squeeze_mask_0, stride = key_cache_internal_tensor_assign_27_stride_0, update = k_state_53_cast_fp16, x = coreml_update_state_106)[name = string("key_cache_internal_tensor_assign_27_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_27_cast_fp16, input = key_cache)[name = string("coreml_update_state_108_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_108 = read_state(input = key_cache)[name = string("coreml_update_state_108")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_27_stride_0 = const()[name = string("value_cache_internal_tensor_assign_27_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_27_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_27_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_27_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_27_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_27_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_27_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_53_cast_fp16 = transpose(perm = v_state_53_perm_0, x = var_4929_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_27_cast_fp16 = slice_update(begin = concat_499, begin_mask = value_cache_internal_tensor_assign_27_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_27_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_27_squeeze_mask_0, stride = value_cache_internal_tensor_assign_27_stride_0, update = v_state_53_cast_fp16, x = coreml_update_state_107)[name = string("value_cache_internal_tensor_assign_27_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_27_cast_fp16, input = value_cache)[name = string("coreml_update_state_109_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_109 = read_state(input = value_cache)[name = string("coreml_update_state_109")];
            tensor<int32, [5]> var_4986_begin_0 = const()[name = string("op_4986_begin_0"), val = tensor<int32, [5]>([26, 0, 0, 0, 0])];
            tensor<int32, [5]> var_4986_end_0 = const()[name = string("op_4986_end_0"), val = tensor<int32, [5]>([27, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_4986_end_mask_0 = const()[name = string("op_4986_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_4986_squeeze_mask_0 = const()[name = string("op_4986_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_4986_cast_fp16 = slice_by_index(begin = var_4986_begin_0, end = var_4986_end_0, end_mask = var_4986_end_mask_0, squeeze_mask = var_4986_squeeze_mask_0, x = coreml_update_state_108)[name = string("op_4986_cast_fp16")];
            tensor<int32, [4]> var_4989_begin_0 = const()[name = string("op_4989_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_4989_end_mask_0 = const()[name = string("op_4989_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_4989_cast_fp16 = slice_by_index(begin = var_4989_begin_0, end = concat_11, end_mask = var_4989_end_mask_0, x = var_4986_cast_fp16)[name = string("op_4989_cast_fp16")];
            tensor<int32, [5]> var_4991_begin_0 = const()[name = string("op_4991_begin_0"), val = tensor<int32, [5]>([26, 0, 0, 0, 0])];
            tensor<int32, [5]> var_4991_end_0 = const()[name = string("op_4991_end_0"), val = tensor<int32, [5]>([27, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_4991_end_mask_0 = const()[name = string("op_4991_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_4991_squeeze_mask_0 = const()[name = string("op_4991_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_4991_cast_fp16 = slice_by_index(begin = var_4991_begin_0, end = var_4991_end_0, end_mask = var_4991_end_mask_0, squeeze_mask = var_4991_squeeze_mask_0, x = coreml_update_state_109)[name = string("op_4991_cast_fp16")];
            tensor<int32, [4]> var_4994_begin_0 = const()[name = string("op_4994_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_4994_end_mask_0 = const()[name = string("op_4994_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_4994_cast_fp16 = slice_by_index(begin = var_4994_begin_0, end = concat_11, end_mask = var_4994_end_mask_0, x = var_4991_cast_fp16)[name = string("op_4994_cast_fp16")];
            tensor<int32, [4]> var_4996_shape_cast_fp16 = shape(x = var_4989_cast_fp16)[name = string("op_4996_shape_cast_fp16")];
            int32 gather_481 = const()[name = string("gather_481"), val = int32(1)];
            int32 gather_482 = const()[name = string("gather_482"), val = int32(2)];
            int32 gather_483_axis_0 = const()[name = string("gather_483_axis_0"), val = int32(0)];
            int32 gather_483_batch_dims_0 = const()[name = string("gather_483_batch_dims_0"), val = int32(0)];
            bool gather_483_validate_indices_0 = const()[name = string("gather_483_validate_indices_0"), val = bool(false)];
            string var_4996_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4996_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_483_to_uint16 = const()[name = string("select_483_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_4996_shape_cast_fp16_to_uint16 = cast(dtype = var_4996_shape_cast_fp16_to_uint16_dtype_0, x = var_4996_shape_cast_fp16)[name = string("cast_11")];
            uint16 gather_483_cast_uint16 = gather(axis = gather_483_axis_0, batch_dims = gather_483_batch_dims_0, indices = select_483_to_uint16, validate_indices = gather_483_validate_indices_0, x = var_4996_shape_cast_fp16_to_uint16)[name = string("gather_483_cast_uint16")];
            string gather_483_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_483_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_484 = const()[name = string("gather_484"), val = int32(128)];
            tensor<int32, [1]> var_5003_axes_0 = const()[name = string("op_5003_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_5003_cast_fp16 = expand_dims(axes = var_5003_axes_0, x = var_4989_cast_fp16)[name = string("op_5003_cast_fp16")];
            tensor<int32, [5]> shape_537_cast_fp16 = shape(x = var_5003_cast_fp16)[name = string("shape_537_cast_fp16")];
            int32 concat_507_axis_0 = const()[name = string("concat_507_axis_0"), val = int32(0)];
            bool concat_507_interleave_0 = const()[name = string("concat_507_interleave_0"), val = bool(false)];
            int32 gather_483_cast_uint16_to_int32 = cast(dtype = gather_483_cast_uint16_to_int32_dtype_0, x = gather_483_cast_uint16)[name = string("cast_10")];
            tensor<int32, [5]> concat_507 = concat(axis = concat_507_axis_0, interleave = concat_507_interleave_0, values = (gather_481, gather_482, var_77, gather_483_cast_uint16_to_int32, gather_484))[name = string("concat_507")];
            tensor<int32, [5]> real_div_52 = real_div(x = concat_507, y = shape_537_cast_fp16)[name = string("real_div_52")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_791_cast_fp16 = tile(reps = real_div_52, x = var_5003_cast_fp16)[name = string("hidden_states_791_cast_fp16")];
            tensor<int32, [4]> concat_508x = const()[name = string("concat_508x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_107_cast_fp16 = reshape(shape = concat_508x, x = hidden_states_791_cast_fp16)[name = string("key_states_107_cast_fp16")];
            tensor<int32, [4]> var_5013_shape_cast_fp16 = shape(x = var_4994_cast_fp16)[name = string("op_5013_shape_cast_fp16")];
            int32 gather_485 = const()[name = string("gather_485"), val = int32(1)];
            int32 gather_486 = const()[name = string("gather_486"), val = int32(2)];
            int32 gather_487_axis_0 = const()[name = string("gather_487_axis_0"), val = int32(0)];
            int32 gather_487_batch_dims_0 = const()[name = string("gather_487_batch_dims_0"), val = int32(0)];
            bool gather_487_validate_indices_0 = const()[name = string("gather_487_validate_indices_0"), val = bool(false)];
            string var_5013_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_5013_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_487_to_uint16 = const()[name = string("select_487_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_5013_shape_cast_fp16_to_uint16 = cast(dtype = var_5013_shape_cast_fp16_to_uint16_dtype_0, x = var_5013_shape_cast_fp16)[name = string("cast_9")];
            uint16 gather_487_cast_uint16 = gather(axis = gather_487_axis_0, batch_dims = gather_487_batch_dims_0, indices = select_487_to_uint16, validate_indices = gather_487_validate_indices_0, x = var_5013_shape_cast_fp16_to_uint16)[name = string("gather_487_cast_uint16")];
            string gather_487_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_487_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_488 = const()[name = string("gather_488"), val = int32(128)];
            tensor<int32, [1]> var_5020_axes_0 = const()[name = string("op_5020_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_5020_cast_fp16 = expand_dims(axes = var_5020_axes_0, x = var_4994_cast_fp16)[name = string("op_5020_cast_fp16")];
            tensor<int32, [5]> shape_542_cast_fp16 = shape(x = var_5020_cast_fp16)[name = string("shape_542_cast_fp16")];
            int32 concat_509_axis_0 = const()[name = string("concat_509_axis_0"), val = int32(0)];
            bool concat_509_interleave_0 = const()[name = string("concat_509_interleave_0"), val = bool(false)];
            int32 gather_487_cast_uint16_to_int32 = cast(dtype = gather_487_cast_uint16_to_int32_dtype_0, x = gather_487_cast_uint16)[name = string("cast_8")];
            tensor<int32, [5]> concat_509 = concat(axis = concat_509_axis_0, interleave = concat_509_interleave_0, values = (gather_485, gather_486, var_77, gather_487_cast_uint16_to_int32, gather_488))[name = string("concat_509")];
            tensor<int32, [5]> real_div_53 = real_div(x = concat_509, y = shape_542_cast_fp16)[name = string("real_div_53")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_795_cast_fp16 = tile(reps = real_div_53, x = var_5020_cast_fp16)[name = string("hidden_states_795_cast_fp16")];
            tensor<int32, [4]> concat_510x = const()[name = string("concat_510x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_107_cast_fp16 = reshape(shape = concat_510x, x = hidden_states_795_cast_fp16)[name = string("value_states_107_cast_fp16")];
            tensor<int32, [4]> var_5030_shape_cast_fp16 = shape(x = key_states_107_cast_fp16)[name = string("op_5030_shape_cast_fp16")];
            int32 gather_489_axis_0 = const()[name = string("gather_489_axis_0"), val = int32(0)];
            int32 gather_489_batch_dims_0 = const()[name = string("gather_489_batch_dims_0"), val = int32(0)];
            bool gather_489_validate_indices_0 = const()[name = string("gather_489_validate_indices_0"), val = bool(false)];
            string var_5030_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_5030_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_489_to_uint16 = const()[name = string("select_489_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_5030_shape_cast_fp16_to_uint16 = cast(dtype = var_5030_shape_cast_fp16_to_uint16_dtype_0, x = var_5030_shape_cast_fp16)[name = string("cast_7")];
            uint16 gather_489_cast_uint16 = gather(axis = gather_489_axis_0, batch_dims = gather_489_batch_dims_0, indices = select_489_to_uint16, validate_indices = gather_489_validate_indices_0, x = var_5030_shape_cast_fp16_to_uint16)[name = string("gather_489_cast_uint16")];
            string gather_489_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_489_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_511_values0_0 = const()[name = string("concat_511_values0_0"), val = int32(1)];
            int32 concat_511_values1_0 = const()[name = string("concat_511_values1_0"), val = int32(1)];
            int32 concat_511_values2_0 = const()[name = string("concat_511_values2_0"), val = int32(0)];
            int32 concat_511_axis_0 = const()[name = string("concat_511_axis_0"), val = int32(0)];
            bool concat_511_interleave_0 = const()[name = string("concat_511_interleave_0"), val = bool(false)];
            int32 gather_489_cast_uint16_to_int32 = cast(dtype = gather_489_cast_uint16_to_int32_dtype_0, x = gather_489_cast_uint16)[name = string("cast_6")];
            tensor<int32, [4]> concat_511 = concat(axis = concat_511_axis_0, interleave = concat_511_interleave_0, values = (concat_511_values0_0, concat_511_values1_0, concat_511_values2_0, gather_489_cast_uint16_to_int32))[name = string("concat_511")];
            tensor<int32, [4]> causal_mask_55_begin_0 = const()[name = string("causal_mask_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_55_end_mask_0 = const()[name = string("causal_mask_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_55_cast_fp16 = slice_by_index(begin = causal_mask_55_begin_0, end = concat_511, end_mask = causal_mask_55_end_mask_0, x = causal_mask)[name = string("causal_mask_55_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_105_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_55_cast_fp16, key = key_states_107_cast_fp16, query = query_states_107_cast_fp16, value = value_states_107_cast_fp16)[name = string("attn_output_105_cast_fp16")];
            tensor<int32, [4]> var_5036_perm_0 = const()[name = string("op_5036_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_512x = const()[name = string("concat_512x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_5036_cast_fp16 = transpose(perm = var_5036_perm_0, x = attn_output_105_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, ?, 1536]> input_209_cast_fp16 = reshape(shape = concat_512x, x = var_5036_cast_fp16)[name = string("input_209_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_26_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(817712640))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(818892352))))[name = string("model_model_layers_26_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_185_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_26_self_attn_o_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_185_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_799_cast_fp16 = add(x = hidden_states_779_cast_fp16, y = linear_185_cast_fp16)[name = string("hidden_states_799_cast_fp16")];
            fp16 var_78_promoted_53_to_fp16 = const()[name = string("op_78_promoted_53_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_5045_cast_fp16 = pow(x = hidden_states_799_cast_fp16, y = var_78_promoted_53_to_fp16)[name = string("op_5045_cast_fp16")];
            tensor<int32, [1]> variance_107_axes_0 = const()[name = string("variance_107_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_107_keep_dims_0 = const()[name = string("variance_107_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_107_cast_fp16 = reduce_mean(axes = variance_107_axes_0, keep_dims = variance_107_keep_dims_0, x = var_5045_cast_fp16)[name = string("variance_107_cast_fp16")];
            fp16 var_5048_to_fp16 = const()[name = string("op_5048_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_5049_cast_fp16 = add(x = variance_107_cast_fp16, y = var_5048_to_fp16)[name = string("op_5049_cast_fp16")];
            fp32 var_5050_epsilon_0 = const()[name = string("op_5050_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_5050_cast_fp16 = rsqrt(epsilon = var_5050_epsilon_0, x = var_5049_cast_fp16)[name = string("op_5050_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_803_cast_fp16 = mul(x = hidden_states_799_cast_fp16, y = var_5050_cast_fp16)[name = string("hidden_states_803_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_26_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_26_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819039872)))];
            tensor<fp16, [1, ?, 1536]> input_211_cast_fp16 = mul(x = model_model_layers_26_post_attention_layernorm_weight_to_fp16, y = hidden_states_803_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_26_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819043008))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(825924352))))[name = string("model_model_layers_26_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_186_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_26_mlp_gate_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_186_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_5062_cast_fp16 = silu(x = linear_186_cast_fp16)[name = string("op_5062_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_26_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(826784576))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(833665920))))[name = string("model_model_layers_26_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_187_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_26_mlp_up_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_187_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_215_cast_fp16 = mul(x = var_5062_cast_fp16, y = linear_187_cast_fp16)[name = string("input_215_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_26_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(834526144))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(841407488))))[name = string("model_model_layers_26_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_188_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_26_mlp_down_proj_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_188_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_809_cast_fp16 = add(x = hidden_states_799_cast_fp16, y = linear_188_cast_fp16)[name = string("hidden_states_809_cast_fp16")];
            fp16 var_78_promoted_54_to_fp16 = const()[name = string("op_78_promoted_54_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_5075_cast_fp16 = pow(x = hidden_states_809_cast_fp16, y = var_78_promoted_54_to_fp16)[name = string("op_5075_cast_fp16")];
            tensor<int32, [1]> variance_109_axes_0 = const()[name = string("variance_109_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_109_keep_dims_0 = const()[name = string("variance_109_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_109_cast_fp16 = reduce_mean(axes = variance_109_axes_0, keep_dims = variance_109_keep_dims_0, x = var_5075_cast_fp16)[name = string("variance_109_cast_fp16")];
            fp16 var_5078_to_fp16 = const()[name = string("op_5078_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_5079_cast_fp16 = add(x = variance_109_cast_fp16, y = var_5078_to_fp16)[name = string("op_5079_cast_fp16")];
            fp32 var_5080_epsilon_0 = const()[name = string("op_5080_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_5080_cast_fp16 = rsqrt(epsilon = var_5080_epsilon_0, x = var_5079_cast_fp16)[name = string("op_5080_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_813_cast_fp16 = mul(x = hidden_states_809_cast_fp16, y = var_5080_cast_fp16)[name = string("hidden_states_813_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_27_input_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_27_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(842267712)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_817_cast_fp16 = mul(x = model_model_layers_27_input_layernorm_weight_to_fp16, y = hidden_states_813_cast_fp16)[name = string("hidden_states_817_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_27_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(842270848))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(843450560))))[name = string("model_model_layers_27_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1536]> model_model_layers_27_self_attn_q_proj_bias_to_fp16 = const()[name = string("model_model_layers_27_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(843598080)))];
            tensor<fp16, [1, ?, 1536]> linear_189_cast_fp16 = linear(bias = model_model_layers_27_self_attn_q_proj_bias_to_fp16, weight = model_model_layers_27_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_817_cast_fp16)[name = string("linear_189_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_27_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(843601216))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(843797888))))[name = string("model_model_layers_27_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_27_self_attn_k_proj_bias_to_fp16 = const()[name = string("model_model_layers_27_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(843822528)))];
            tensor<fp16, [1, ?, 256]> linear_190_cast_fp16 = linear(bias = model_model_layers_27_self_attn_k_proj_bias_to_fp16, weight = model_model_layers_27_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_817_cast_fp16)[name = string("linear_190_cast_fp16")];
            tensor<fp16, [256, 1536]> model_model_layers_27_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(843823104))), scale = tensor<fp16, [256, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(844019776))))[name = string("model_model_layers_27_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> model_model_layers_27_self_attn_v_proj_bias_to_fp16 = const()[name = string("model_model_layers_27_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(844044416)))];
            tensor<fp16, [1, ?, 256]> linear_191_cast_fp16 = linear(bias = model_model_layers_27_self_attn_v_proj_bias_to_fp16, weight = model_model_layers_27_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_817_cast_fp16)[name = string("linear_191_cast_fp16")];
            tensor<int32, [4]> concat_513x = const()[name = string("concat_513x"), val = tensor<int32, [4]>([1, -1, 12, 128])];
            tensor<fp16, [1, ?, 12, 128]> var_5103_cast_fp16 = reshape(shape = concat_513x, x = linear_189_cast_fp16)[name = string("op_5103_cast_fp16")];
            tensor<int32, [4]> q_perm_0 = const()[name = string("q_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_514x = const()[name = string("concat_514x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_5106_cast_fp16 = reshape(shape = concat_514x, x = linear_190_cast_fp16)[name = string("op_5106_cast_fp16")];
            tensor<int32, [4]> k_perm_0 = const()[name = string("k_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> concat_515x = const()[name = string("concat_515x"), val = tensor<int32, [4]>([1, -1, 2, 128])];
            tensor<fp16, [1, ?, 2, 128]> var_5109_cast_fp16 = reshape(shape = concat_515x, x = linear_191_cast_fp16)[name = string("op_5109_cast_fp16")];
            tensor<int32, [4]> v_state_perm_0 = const()[name = string("v_state_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, ?, 128]> q_cast_fp16 = transpose(perm = q_perm_0, x = var_5103_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 12, ?, 128]> var_5113_cast_fp16 = mul(x = q_cast_fp16, y = cos_7_cast_fp16)[name = string("op_5113_cast_fp16")];
            tensor<int32, [4]> x1_109_begin_0 = const()[name = string("x1_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_109_end_0 = const()[name = string("x1_109_end_0"), val = tensor<int32, [4]>([1, 12, 0, 64])];
            tensor<bool, [4]> x1_109_end_mask_0 = const()[name = string("x1_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, ?, 64]> x1_109_cast_fp16 = slice_by_index(begin = x1_109_begin_0, end = x1_109_end_0, end_mask = x1_109_end_mask_0, x = q_cast_fp16)[name = string("x1_109_cast_fp16")];
            tensor<int32, [4]> x2_109_begin_0 = const()[name = string("x2_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_109_end_0 = const()[name = string("x2_109_end_0"), val = tensor<int32, [4]>([1, 12, 0, 128])];
            tensor<bool, [4]> x2_109_end_mask_0 = const()[name = string("x2_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, ?, 64]> x2_109_cast_fp16 = slice_by_index(begin = x2_109_begin_0, end = x2_109_end_0, end_mask = x2_109_end_mask_0, x = q_cast_fp16)[name = string("x2_109_cast_fp16")];
            fp16 const_57_promoted_to_fp16 = const()[name = string("const_57_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 12, ?, 64]> var_5124_cast_fp16 = mul(x = x2_109_cast_fp16, y = const_57_promoted_to_fp16)[name = string("op_5124_cast_fp16")];
            bool var_5126_interleave_0 = const()[name = string("op_5126_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, ?, 128]> var_5126_cast_fp16 = concat(axis = var_72, interleave = var_5126_interleave_0, values = (var_5124_cast_fp16, x1_109_cast_fp16))[name = string("op_5126_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> var_5127_cast_fp16 = mul(x = var_5126_cast_fp16, y = sin_7_cast_fp16)[name = string("op_5127_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> query_states_cast_fp16 = add(x = var_5113_cast_fp16, y = var_5127_cast_fp16)[name = string("query_states_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_cast_fp16 = transpose(perm = k_perm_0, x = var_5106_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 2, ?, 128]> var_5129_cast_fp16 = mul(x = k_cast_fp16, y = cos_7_cast_fp16)[name = string("op_5129_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([1, 2, 0, 64])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, ?, 64]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_cast_fp16)[name = string("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([1, 2, 0, 128])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, ?, 64]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_cast_fp16)[name = string("x2_cast_fp16")];
            fp16 const_58_promoted_to_fp16 = const()[name = string("const_58_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, ?, 64]> var_5140_cast_fp16 = mul(x = x2_cast_fp16, y = const_58_promoted_to_fp16)[name = string("op_5140_cast_fp16")];
            bool var_5142_interleave_0 = const()[name = string("op_5142_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, ?, 128]> var_5142_cast_fp16 = concat(axis = var_72, interleave = var_5142_interleave_0, values = (var_5140_cast_fp16, x1_cast_fp16))[name = string("op_5142_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> var_5143_cast_fp16 = mul(x = var_5142_cast_fp16, y = sin_7_cast_fp16)[name = string("op_5143_cast_fp16")];
            tensor<fp16, [1, 2, ?, 128]> k_state_cast_fp16 = add(x = var_5129_cast_fp16, y = var_5143_cast_fp16)[name = string("k_state_cast_fp16")];
            tensor<int32, [1]> expand_dims_324 = const()[name = string("expand_dims_324"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_325 = const()[name = string("expand_dims_325"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> expand_dims_327 = const()[name = string("expand_dims_327"), val = tensor<int32, [1]>([0])];
            tensor<int32, [1]> concat_518_values0_0 = const()[name = string("concat_518_values0_0"), val = tensor<int32, [1]>([27])];
            int32 concat_518_axis_0 = const()[name = string("concat_518_axis_0"), val = int32(0)];
            bool concat_518_interleave_0 = const()[name = string("concat_518_interleave_0"), val = bool(false)];
            tensor<int32, [5]> concat_518 = concat(axis = concat_518_axis_0, interleave = concat_518_interleave_0, values = (concat_518_values0_0, expand_dims_324, expand_dims_325, expand_dims_2, expand_dims_327))[name = string("concat_518")];
            tensor<int32, [5]> key_cache_internal_tensor_assign_28_stride_0 = const()[name = string("key_cache_internal_tensor_assign_28_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_28_begin_mask_0 = const()[name = string("key_cache_internal_tensor_assign_28_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_28_end_mask_0 = const()[name = string("key_cache_internal_tensor_assign_28_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> key_cache_internal_tensor_assign_28_squeeze_mask_0 = const()[name = string("key_cache_internal_tensor_assign_28_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [28, 1, 2, 2048, 128]> key_cache_internal_tensor_assign_28_cast_fp16 = slice_update(begin = concat_518, begin_mask = key_cache_internal_tensor_assign_28_begin_mask_0, end = concat_6, end_mask = key_cache_internal_tensor_assign_28_end_mask_0, squeeze_mask = key_cache_internal_tensor_assign_28_squeeze_mask_0, stride = key_cache_internal_tensor_assign_28_stride_0, update = k_state_cast_fp16, x = coreml_update_state_108)[name = string("key_cache_internal_tensor_assign_28_cast_fp16")];
            write_state(data = key_cache_internal_tensor_assign_28_cast_fp16, input = key_cache)[name = string("coreml_update_state_110_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_110 = read_state(input = key_cache)[name = string("coreml_update_state_110")];
            tensor<int32, [5]> value_cache_internal_tensor_assign_28_stride_0 = const()[name = string("value_cache_internal_tensor_assign_28_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_28_begin_mask_0 = const()[name = string("value_cache_internal_tensor_assign_28_begin_mask_0"), val = tensor<bool, [5]>([false, false, false, false, false])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_28_end_mask_0 = const()[name = string("value_cache_internal_tensor_assign_28_end_mask_0"), val = tensor<bool, [5]>([false, true, false, false, true])];
            tensor<bool, [5]> value_cache_internal_tensor_assign_28_squeeze_mask_0 = const()[name = string("value_cache_internal_tensor_assign_28_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, ?, 128]> v_state_cast_fp16 = transpose(perm = v_state_perm_0, x = var_5109_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [28, 1, 2, 2048, 128]> value_cache_internal_tensor_assign_28_cast_fp16 = slice_update(begin = concat_518, begin_mask = value_cache_internal_tensor_assign_28_begin_mask_0, end = concat_6, end_mask = value_cache_internal_tensor_assign_28_end_mask_0, squeeze_mask = value_cache_internal_tensor_assign_28_squeeze_mask_0, stride = value_cache_internal_tensor_assign_28_stride_0, update = v_state_cast_fp16, x = coreml_update_state_109)[name = string("value_cache_internal_tensor_assign_28_cast_fp16")];
            write_state(data = value_cache_internal_tensor_assign_28_cast_fp16, input = value_cache)[name = string("coreml_update_state_111_write_state")];
            tensor<fp16, [28, 1, 2, 2048, 128]> coreml_update_state_111 = read_state(input = value_cache)[name = string("coreml_update_state_111")];
            tensor<int32, [5]> var_5166_begin_0 = const()[name = string("op_5166_begin_0"), val = tensor<int32, [5]>([27, 0, 0, 0, 0])];
            tensor<int32, [5]> var_5166_end_0 = const()[name = string("op_5166_end_0"), val = tensor<int32, [5]>([28, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_5166_end_mask_0 = const()[name = string("op_5166_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_5166_squeeze_mask_0 = const()[name = string("op_5166_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_5166_cast_fp16 = slice_by_index(begin = var_5166_begin_0, end = var_5166_end_0, end_mask = var_5166_end_mask_0, squeeze_mask = var_5166_squeeze_mask_0, x = coreml_update_state_110)[name = string("op_5166_cast_fp16")];
            tensor<int32, [4]> var_5169_begin_0 = const()[name = string("op_5169_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_5169_end_mask_0 = const()[name = string("op_5169_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_5169_cast_fp16 = slice_by_index(begin = var_5169_begin_0, end = concat_11, end_mask = var_5169_end_mask_0, x = var_5166_cast_fp16)[name = string("op_5169_cast_fp16")];
            tensor<int32, [5]> var_5171_begin_0 = const()[name = string("op_5171_begin_0"), val = tensor<int32, [5]>([27, 0, 0, 0, 0])];
            tensor<int32, [5]> var_5171_end_0 = const()[name = string("op_5171_end_0"), val = tensor<int32, [5]>([28, 1, 2, 2048, 128])];
            tensor<bool, [5]> var_5171_end_mask_0 = const()[name = string("op_5171_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_5171_squeeze_mask_0 = const()[name = string("op_5171_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 2, 2048, 128]> var_5171_cast_fp16 = slice_by_index(begin = var_5171_begin_0, end = var_5171_end_0, end_mask = var_5171_end_mask_0, squeeze_mask = var_5171_squeeze_mask_0, x = coreml_update_state_111)[name = string("op_5171_cast_fp16")];
            tensor<int32, [4]> var_5174_begin_0 = const()[name = string("op_5174_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> var_5174_end_mask_0 = const()[name = string("op_5174_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 2, ?, 128]> var_5174_cast_fp16 = slice_by_index(begin = var_5174_begin_0, end = concat_11, end_mask = var_5174_end_mask_0, x = var_5171_cast_fp16)[name = string("op_5174_cast_fp16")];
            tensor<int32, [4]> var_5176_shape_cast_fp16 = shape(x = var_5169_cast_fp16)[name = string("op_5176_shape_cast_fp16")];
            int32 gather_499 = const()[name = string("gather_499"), val = int32(1)];
            int32 gather_500 = const()[name = string("gather_500"), val = int32(2)];
            int32 gather_501_axis_0 = const()[name = string("gather_501_axis_0"), val = int32(0)];
            int32 gather_501_batch_dims_0 = const()[name = string("gather_501_batch_dims_0"), val = int32(0)];
            bool gather_501_validate_indices_0 = const()[name = string("gather_501_validate_indices_0"), val = bool(false)];
            string var_5176_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_5176_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_501_to_uint16 = const()[name = string("select_501_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_5176_shape_cast_fp16_to_uint16 = cast(dtype = var_5176_shape_cast_fp16_to_uint16_dtype_0, x = var_5176_shape_cast_fp16)[name = string("cast_5")];
            uint16 gather_501_cast_uint16 = gather(axis = gather_501_axis_0, batch_dims = gather_501_batch_dims_0, indices = select_501_to_uint16, validate_indices = gather_501_validate_indices_0, x = var_5176_shape_cast_fp16_to_uint16)[name = string("gather_501_cast_uint16")];
            string gather_501_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_501_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_502 = const()[name = string("gather_502"), val = int32(128)];
            tensor<int32, [1]> var_5183_axes_0 = const()[name = string("op_5183_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_5183_cast_fp16 = expand_dims(axes = var_5183_axes_0, x = var_5169_cast_fp16)[name = string("op_5183_cast_fp16")];
            tensor<int32, [5]> shape_557_cast_fp16 = shape(x = var_5183_cast_fp16)[name = string("shape_557_cast_fp16")];
            int32 concat_526_axis_0 = const()[name = string("concat_526_axis_0"), val = int32(0)];
            bool concat_526_interleave_0 = const()[name = string("concat_526_interleave_0"), val = bool(false)];
            int32 gather_501_cast_uint16_to_int32 = cast(dtype = gather_501_cast_uint16_to_int32_dtype_0, x = gather_501_cast_uint16)[name = string("cast_4")];
            tensor<int32, [5]> concat_526 = concat(axis = concat_526_axis_0, interleave = concat_526_interleave_0, values = (gather_499, gather_500, var_77, gather_501_cast_uint16_to_int32, gather_502))[name = string("concat_526")];
            tensor<int32, [5]> real_div_54 = real_div(x = concat_526, y = shape_557_cast_fp16)[name = string("real_div_54")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_821_cast_fp16 = tile(reps = real_div_54, x = var_5183_cast_fp16)[name = string("hidden_states_821_cast_fp16")];
            tensor<int32, [4]> concat_527x = const()[name = string("concat_527x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> key_states_cast_fp16 = reshape(shape = concat_527x, x = hidden_states_821_cast_fp16)[name = string("key_states_cast_fp16")];
            tensor<int32, [4]> var_5193_shape_cast_fp16 = shape(x = var_5174_cast_fp16)[name = string("op_5193_shape_cast_fp16")];
            int32 gather_503 = const()[name = string("gather_503"), val = int32(1)];
            int32 gather_504 = const()[name = string("gather_504"), val = int32(2)];
            int32 gather_505_axis_0 = const()[name = string("gather_505_axis_0"), val = int32(0)];
            int32 gather_505_batch_dims_0 = const()[name = string("gather_505_batch_dims_0"), val = int32(0)];
            bool gather_505_validate_indices_0 = const()[name = string("gather_505_validate_indices_0"), val = bool(false)];
            string var_5193_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_5193_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_505_to_uint16 = const()[name = string("select_505_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_5193_shape_cast_fp16_to_uint16 = cast(dtype = var_5193_shape_cast_fp16_to_uint16_dtype_0, x = var_5193_shape_cast_fp16)[name = string("cast_3")];
            uint16 gather_505_cast_uint16 = gather(axis = gather_505_axis_0, batch_dims = gather_505_batch_dims_0, indices = select_505_to_uint16, validate_indices = gather_505_validate_indices_0, x = var_5193_shape_cast_fp16_to_uint16)[name = string("gather_505_cast_uint16")];
            string gather_505_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_505_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 gather_506 = const()[name = string("gather_506"), val = int32(128)];
            tensor<int32, [1]> var_5200_axes_0 = const()[name = string("op_5200_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, ?, 128]> var_5200_cast_fp16 = expand_dims(axes = var_5200_axes_0, x = var_5174_cast_fp16)[name = string("op_5200_cast_fp16")];
            tensor<int32, [5]> shape_562_cast_fp16 = shape(x = var_5200_cast_fp16)[name = string("shape_562_cast_fp16")];
            int32 concat_528_axis_0 = const()[name = string("concat_528_axis_0"), val = int32(0)];
            bool concat_528_interleave_0 = const()[name = string("concat_528_interleave_0"), val = bool(false)];
            int32 gather_505_cast_uint16_to_int32 = cast(dtype = gather_505_cast_uint16_to_int32_dtype_0, x = gather_505_cast_uint16)[name = string("cast_2")];
            tensor<int32, [5]> concat_528 = concat(axis = concat_528_axis_0, interleave = concat_528_interleave_0, values = (gather_503, gather_504, var_77, gather_505_cast_uint16_to_int32, gather_506))[name = string("concat_528")];
            tensor<int32, [5]> real_div_55 = real_div(x = concat_528, y = shape_562_cast_fp16)[name = string("real_div_55")];
            tensor<fp16, [?, ?, ?, ?, ?]> hidden_states_825_cast_fp16 = tile(reps = real_div_55, x = var_5200_cast_fp16)[name = string("hidden_states_825_cast_fp16")];
            tensor<int32, [4]> concat_529x = const()[name = string("concat_529x"), val = tensor<int32, [4]>([1, 12, -1, 128])];
            tensor<fp16, [1, 12, ?, 128]> value_states_cast_fp16 = reshape(shape = concat_529x, x = hidden_states_825_cast_fp16)[name = string("value_states_cast_fp16")];
            tensor<int32, [4]> var_5210_shape_cast_fp16 = shape(x = key_states_cast_fp16)[name = string("op_5210_shape_cast_fp16")];
            int32 gather_507_axis_0 = const()[name = string("gather_507_axis_0"), val = int32(0)];
            int32 gather_507_batch_dims_0 = const()[name = string("gather_507_batch_dims_0"), val = int32(0)];
            bool gather_507_validate_indices_0 = const()[name = string("gather_507_validate_indices_0"), val = bool(false)];
            string var_5210_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_5210_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_507_to_uint16 = const()[name = string("select_507_to_uint16"), val = uint16(2)];
            tensor<uint16, [4]> var_5210_shape_cast_fp16_to_uint16 = cast(dtype = var_5210_shape_cast_fp16_to_uint16_dtype_0, x = var_5210_shape_cast_fp16)[name = string("cast_1")];
            uint16 gather_507_cast_uint16 = gather(axis = gather_507_axis_0, batch_dims = gather_507_batch_dims_0, indices = select_507_to_uint16, validate_indices = gather_507_validate_indices_0, x = var_5210_shape_cast_fp16_to_uint16)[name = string("gather_507_cast_uint16")];
            string gather_507_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_507_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            int32 concat_530_values0_0 = const()[name = string("concat_530_values0_0"), val = int32(1)];
            int32 concat_530_values1_0 = const()[name = string("concat_530_values1_0"), val = int32(1)];
            int32 concat_530_values2_0 = const()[name = string("concat_530_values2_0"), val = int32(0)];
            int32 concat_530_axis_0 = const()[name = string("concat_530_axis_0"), val = int32(0)];
            bool concat_530_interleave_0 = const()[name = string("concat_530_interleave_0"), val = bool(false)];
            int32 gather_507_cast_uint16_to_int32 = cast(dtype = gather_507_cast_uint16_to_int32_dtype_0, x = gather_507_cast_uint16)[name = string("cast_0")];
            tensor<int32, [4]> concat_530 = concat(axis = concat_530_axis_0, interleave = concat_530_interleave_0, values = (concat_530_values0_0, concat_530_values1_0, concat_530_values2_0, gather_507_cast_uint16_to_int32))[name = string("concat_530")];
            tensor<int32, [4]> causal_mask_begin_0 = const()[name = string("causal_mask_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<bool, [4]> causal_mask_end_mask_0 = const()[name = string("causal_mask_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, ?, ?]> causal_mask_cast_fp16 = slice_by_index(begin = causal_mask_begin_0, end = concat_530, end_mask = causal_mask_end_mask_0, x = causal_mask)[name = string("causal_mask_cast_fp16")];
            tensor<fp16, [1, 12, ?, 128]> attn_output_109_cast_fp16 = scaled_dot_product_attention(attn_mask = causal_mask_cast_fp16, key = key_states_cast_fp16, query = query_states_cast_fp16, value = value_states_cast_fp16)[name = string("attn_output_109_cast_fp16")];
            tensor<int32, [4]> var_5216_perm_0 = const()[name = string("op_5216_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_531x = const()[name = string("concat_531x"), val = tensor<int32, [3]>([1, -1, 1536])];
            tensor<fp16, [1, ?, 12, 128]> var_5216_cast_fp16 = transpose(perm = var_5216_perm_0, x = attn_output_109_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, ?, 1536]> input_217_cast_fp16 = reshape(shape = concat_531x, x = var_5216_cast_fp16)[name = string("input_217_cast_fp16")];
            tensor<fp16, [1536, 1536]> model_model_layers_27_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(844044992))), scale = tensor<fp16, [1536, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(845224704))))[name = string("model_model_layers_27_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_192_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_27_self_attn_o_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_192_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_829_cast_fp16 = add(x = hidden_states_809_cast_fp16, y = linear_192_cast_fp16)[name = string("hidden_states_829_cast_fp16")];
            fp16 var_78_promoted_55_to_fp16 = const()[name = string("op_78_promoted_55_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_5225_cast_fp16 = pow(x = hidden_states_829_cast_fp16, y = var_78_promoted_55_to_fp16)[name = string("op_5225_cast_fp16")];
            tensor<int32, [1]> variance_111_axes_0 = const()[name = string("variance_111_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_111_keep_dims_0 = const()[name = string("variance_111_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_111_cast_fp16 = reduce_mean(axes = variance_111_axes_0, keep_dims = variance_111_keep_dims_0, x = var_5225_cast_fp16)[name = string("variance_111_cast_fp16")];
            fp16 var_5228_to_fp16 = const()[name = string("op_5228_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_5229_cast_fp16 = add(x = variance_111_cast_fp16, y = var_5228_to_fp16)[name = string("op_5229_cast_fp16")];
            fp32 var_5230_epsilon_0 = const()[name = string("op_5230_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_5230_cast_fp16 = rsqrt(epsilon = var_5230_epsilon_0, x = var_5229_cast_fp16)[name = string("op_5230_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_833_cast_fp16 = mul(x = hidden_states_829_cast_fp16, y = var_5230_cast_fp16)[name = string("hidden_states_833_cast_fp16")];
            tensor<fp16, [1536]> model_model_layers_27_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_model_layers_27_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(845372224)))];
            tensor<fp16, [1, ?, 1536]> input_219_cast_fp16 = mul(x = model_model_layers_27_post_attention_layernorm_weight_to_fp16, y = hidden_states_833_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_27_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(845375360))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(852256704))))[name = string("model_model_layers_27_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_193_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_27_mlp_gate_proj_weight_to_fp16_quantized, x = input_219_cast_fp16)[name = string("linear_193_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> var_5242_cast_fp16 = silu(x = linear_193_cast_fp16)[name = string("op_5242_cast_fp16")];
            tensor<fp16, [8960, 1536]> model_model_layers_27_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [8960, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(853116928))), scale = tensor<fp16, [8960, 48]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(859998272))))[name = string("model_model_layers_27_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 8960]> linear_194_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_model_layers_27_mlp_up_proj_weight_to_fp16_quantized, x = input_219_cast_fp16)[name = string("linear_194_cast_fp16")];
            tensor<fp16, [1, ?, 8960]> input_223_cast_fp16 = mul(x = var_5242_cast_fp16, y = linear_194_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [1536, 8960]> model_model_layers_27_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1536, 8960]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(860858496))), scale = tensor<fp16, [1536, 280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(867739840))))[name = string("model_model_layers_27_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1536]> linear_195_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = model_model_layers_27_mlp_down_proj_weight_to_fp16_quantized, x = input_223_cast_fp16)[name = string("linear_195_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_839_cast_fp16 = add(x = hidden_states_829_cast_fp16, y = linear_195_cast_fp16)[name = string("hidden_states_839_cast_fp16")];
            fp16 var_78_promoted_56_to_fp16 = const()[name = string("op_78_promoted_56_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1536]> var_5251_cast_fp16 = pow(x = hidden_states_839_cast_fp16, y = var_78_promoted_56_to_fp16)[name = string("op_5251_cast_fp16")];
            tensor<int32, [1]> variance_axes_0 = const()[name = string("variance_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_keep_dims_0 = const()[name = string("variance_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> variance_cast_fp16 = reduce_mean(axes = variance_axes_0, keep_dims = variance_keep_dims_0, x = var_5251_cast_fp16)[name = string("variance_cast_fp16")];
            fp16 var_5254_to_fp16 = const()[name = string("op_5254_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_5255_cast_fp16 = add(x = variance_cast_fp16, y = var_5254_to_fp16)[name = string("op_5255_cast_fp16")];
            fp32 var_5256_epsilon_0 = const()[name = string("op_5256_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> var_5256_cast_fp16 = rsqrt(epsilon = var_5256_epsilon_0, x = var_5255_cast_fp16)[name = string("op_5256_cast_fp16")];
            tensor<fp16, [1, ?, 1536]> hidden_states_843_cast_fp16 = mul(x = hidden_states_839_cast_fp16, y = var_5256_cast_fp16)[name = string("hidden_states_843_cast_fp16")];
            tensor<fp16, [1536]> model_model_norm_weight_to_fp16 = const()[name = string("model_model_norm_weight_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(868600064)))];
            tensor<fp16, [1, ?, 1536]> hidden_states_cast_fp16 = mul(x = model_model_norm_weight_to_fp16, y = hidden_states_843_cast_fp16)[name = string("hidden_states_cast_fp16")];
            tensor<fp16, [151936]> linear_196_bias_0_to_fp16 = const()[name = string("linear_196_bias_0_to_fp16"), val = tensor<fp16, [151936]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(868603200)))];
            tensor<fp16, [1, ?, 151936]> logits = linear(bias = linear_196_bias_0_to_fp16, weight = model_model_embed_tokens_weight_to_fp16_quantized, x = hidden_states_cast_fp16)[name = string("linear_196_cast_fp16")];
        } -> (logits);
}