program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}})]
{
    func main<ios18>(tensor<int32, [1, 512]> attention_mask, tensor<int32, [1, 512]> input_ids) {
            int32 embedding_batch_dims_0 = const()[name = string("embedding_batch_dims_0"), val = int32(0)];
            bool embedding_validate_indices_0 = const()[name = string("embedding_validate_indices_0"), val = bool(false)];
            tensor<fp16, [262144, 768]> p_st_0_model_embed_tokens_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [262144, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [262144, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201326720))))[name = string("p_st_0_model_embed_tokens_weight_to_fp16_quantized")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [1, 512]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(262144)];
            tensor<int32, [1, 512]> add_0 = add(x = input_ids, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 512]> select_0 = select(a = input_ids, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
            tensor<bool, [1, 512]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(262144)];
            tensor<int32, [1, 512]> add_0_1 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0_1")];
            tensor<int32, [1, 512]> select_0_1 = select(a = select_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")];
            int32 embedding_cast_fp16_axis_0 = const()[name = string("embedding_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 512, 768]> embedding_cast_fp16 = gather(axis = embedding_cast_fp16_axis_0, batch_dims = embedding_batch_dims_0, indices = select_0_1, validate_indices = embedding_validate_indices_0, x = p_st_0_model_embed_tokens_weight_to_fp16_quantized)[name = string("embedding_cast_fp16")];
            fp16 const_6_to_fp16 = const()[name = string("const_6_to_fp16"), val = fp16(0x1.bb8p+4)];
            tensor<fp16, [1, 512, 768]> mul_cast_fp16 = mul(x = embedding_cast_fp16, y = const_6_to_fp16)[name = string("mul_cast_fp16")];
            tensor<int32, [1]> unsqueeze_1_axes_0 = const()[name = string("unsqueeze_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1, 512]> unsqueeze_1 = expand_dims(axes = unsqueeze_1_axes_0, x = attention_mask)[name = string("unsqueeze_1")];
            tensor<int32, [1]> unsqueeze_2_axes_0 = const()[name = string("unsqueeze_2_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 1, 1, 512]> unsqueeze_2 = expand_dims(axes = unsqueeze_2_axes_0, x = unsqueeze_1)[name = string("unsqueeze_2")];
            fp16 const_21_to_fp16 = const()[name = string("const_21_to_fp16"), val = fp16(0x1p+0)];
            string _to_copy_3_to_fp16_dtype_0 = const()[name = string("_to_copy_3_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 1, 1, 512]> unsqueeze_2_to_fp16 = cast(dtype = _to_copy_3_to_fp16_dtype_0, x = unsqueeze_2)[name = string("cast_1")];
            tensor<fp16, [1, 1, 1, 512]> rsub_cast_fp16 = sub(x = const_21_to_fp16, y = unsqueeze_2_to_fp16)[name = string("rsub_cast_fp16")];
            fp16 const_22_to_fp16 = const()[name = string("const_22_to_fp16"), val = fp16(-inf)];
            tensor<fp16, [1, 1, 1, 512]> mul_1_cast_fp16 = mul(x = rsub_cast_fp16, y = const_22_to_fp16)[name = string("mul_1_cast_fp16")];
            tensor<int32, [4]> expand_reps_0 = const()[name = string("expand_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<fp16, [1, 1, 512, 512]> expand_cast_fp16 = tile(reps = expand_reps_0, x = mul_1_cast_fp16)[name = string("expand_cast_fp16")];
            fp16 const_280_promoted_to_fp16 = const()[name = string("const_280_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_1_cast_fp16 = pow(x = mul_cast_fp16, y = const_280_promoted_to_fp16)[name = string("pow_1_cast_fp16")];
            tensor<int32, [1]> mean_axes_0 = const()[name = string("mean_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_keep_dims_0 = const()[name = string("mean_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_cast_fp16 = reduce_mean(axes = mean_axes_0, keep_dims = mean_keep_dims_0, x = pow_1_cast_fp16)[name = string("mean_cast_fp16")];
            fp16 const_283_to_fp16 = const()[name = string("const_283_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_cast_fp16 = add(x = mean_cast_fp16, y = const_283_to_fp16)[name = string("add_cast_fp16")];
            fp32 rsqrt_epsilon_0 = const()[name = string("rsqrt_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_cast_fp16 = rsqrt(epsilon = rsqrt_epsilon_0, x = add_cast_fp16)[name = string("rsqrt_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_51_cast_fp16 = mul(x = mul_cast_fp16, y = rsqrt_cast_fp16)[name = string("mul_51_cast_fp16")];
            tensor<fp16, [768]> add_1_to_fp16 = const()[name = string("add_1_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201851072)))];
            tensor<fp16, [1, 512, 768]> mul_52_cast_fp16 = mul(x = mul_51_cast_fp16, y = add_1_to_fp16)[name = string("mul_52_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201852672))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202442560))))[name = string("p_st_0_model_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [768]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202444160)))];
            tensor<fp16, [1, 512, 768]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = mul_52_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> const_287 = const()[name = string("const_287"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_cast_fp16 = reshape(shape = const_287, x = linear_0_cast_fp16)[name = string("view_cast_fp16")];
            tensor<int32, [4]> transpose_24_perm_0 = const()[name = string("transpose_24_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202445760))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202642432))))[name = string("p_st_0_model_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [256]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202643008)))];
            tensor<fp16, [1, 512, 256]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = mul_52_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> const_290 = const()[name = string("const_290"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_1_cast_fp16 = reshape(shape = const_290, x = linear_1_cast_fp16)[name = string("view_1_cast_fp16")];
            tensor<int32, [4]> transpose_25_perm_0 = const()[name = string("transpose_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202643584))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202840256))))[name = string("p_st_0_model_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_2_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = mul_52_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> const_293 = const()[name = string("const_293"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_2_cast_fp16 = reshape(shape = const_293, x = linear_2_cast_fp16)[name = string("view_2_cast_fp16")];
            tensor<int32, [4]> transpose_26_perm_0 = const()[name = string("transpose_26_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_297_promoted_to_fp16 = const()[name = string("const_297_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_24_cast_fp16 = transpose(perm = transpose_24_perm_0, x = view_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 3, 512, 256]> pow_2_cast_fp16 = pow(x = transpose_24_cast_fp16, y = const_297_promoted_to_fp16)[name = string("pow_2_cast_fp16")];
            tensor<int32, [1]> mean_1_axes_0 = const()[name = string("mean_1_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_1_keep_dims_0 = const()[name = string("mean_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_1_cast_fp16 = reduce_mean(axes = mean_1_axes_0, keep_dims = mean_1_keep_dims_0, x = pow_2_cast_fp16)[name = string("mean_1_cast_fp16")];
            fp16 const_300_to_fp16 = const()[name = string("const_300_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_2_cast_fp16 = add(x = mean_1_cast_fp16, y = const_300_to_fp16)[name = string("add_2_cast_fp16")];
            fp32 rsqrt_1_epsilon_0 = const()[name = string("rsqrt_1_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_1_cast_fp16 = rsqrt(epsilon = rsqrt_1_epsilon_0, x = add_2_cast_fp16)[name = string("rsqrt_1_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_53_cast_fp16 = mul(x = transpose_24_cast_fp16, y = rsqrt_1_cast_fp16)[name = string("mul_53_cast_fp16")];
            tensor<fp16, [256]> add_3_to_fp16 = const()[name = string("add_3_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202840832)))];
            tensor<fp16, [1, 3, 512, 256]> mul_54_cast_fp16 = mul(x = mul_53_cast_fp16, y = add_3_to_fp16)[name = string("mul_54_cast_fp16")];
            fp16 const_305_promoted_to_fp16 = const()[name = string("const_305_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_25_cast_fp16 = transpose(perm = transpose_25_perm_0, x = view_1_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 1, 512, 256]> pow_3_cast_fp16 = pow(x = transpose_25_cast_fp16, y = const_305_promoted_to_fp16)[name = string("pow_3_cast_fp16")];
            tensor<int32, [1]> mean_2_axes_0 = const()[name = string("mean_2_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_2_keep_dims_0 = const()[name = string("mean_2_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_2_cast_fp16 = reduce_mean(axes = mean_2_axes_0, keep_dims = mean_2_keep_dims_0, x = pow_3_cast_fp16)[name = string("mean_2_cast_fp16")];
            fp16 const_308_to_fp16 = const()[name = string("const_308_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_4_cast_fp16 = add(x = mean_2_cast_fp16, y = const_308_to_fp16)[name = string("add_4_cast_fp16")];
            fp32 rsqrt_2_epsilon_0 = const()[name = string("rsqrt_2_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_2_cast_fp16 = rsqrt(epsilon = rsqrt_2_epsilon_0, x = add_4_cast_fp16)[name = string("rsqrt_2_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_55_cast_fp16 = mul(x = transpose_25_cast_fp16, y = rsqrt_2_cast_fp16)[name = string("mul_55_cast_fp16")];
            tensor<fp16, [256]> add_5_to_fp16 = const()[name = string("add_5_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202841408)))];
            tensor<fp16, [1, 1, 512, 256]> mul_56_cast_fp16 = mul(x = mul_55_cast_fp16, y = add_5_to_fp16)[name = string("mul_56_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> unsqueeze_77_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 1, 512, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202841984))), scale = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.02p-7]]]]))[name = string("unsqueeze_77_to_fp16_quantized")];
            tensor<fp16, [1, 3, 512, 256]> mul_57_cast_fp16 = mul(x = mul_54_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_57_cast_fp16")];
            tensor<int32, [4]> slice_77_begin_0 = const()[name = string("slice_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_77_end_0 = const()[name = string("slice_77_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_77_end_mask_0 = const()[name = string("slice_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_77_cast_fp16 = slice_by_index(begin = slice_77_begin_0, end = slice_77_end_0, end_mask = slice_77_end_mask_0, x = mul_54_cast_fp16)[name = string("slice_77_cast_fp16")];
            tensor<int32, [4]> slice_78_begin_0 = const()[name = string("slice_78_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_78_end_0 = const()[name = string("slice_78_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_78_end_mask_0 = const()[name = string("slice_78_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_78_cast_fp16 = slice_by_index(begin = slice_78_begin_0, end = slice_78_end_0, end_mask = slice_78_end_mask_0, x = mul_54_cast_fp16)[name = string("slice_78_cast_fp16")];
            fp16 const_320_promoted_to_fp16 = const()[name = string("const_320_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_cast_fp16 = mul(x = slice_78_cast_fp16, y = const_320_promoted_to_fp16)[name = string("neg_cast_fp16")];
            int32 const_321 = const()[name = string("const_321"), val = int32(-1)];
            bool cat_24_interleave_0 = const()[name = string("cat_24_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_24_cast_fp16 = concat(axis = const_321, interleave = cat_24_interleave_0, values = (neg_cast_fp16, slice_77_cast_fp16))[name = string("cat_24_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> unsqueeze_78_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 1, 512, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202973120))), scale = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.02p-7]]]]))[name = string("unsqueeze_78_to_fp16_quantized")];
            tensor<fp16, [1, 3, 512, 256]> mul_58_cast_fp16 = mul(x = cat_24_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_58_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_6_cast_fp16 = add(x = mul_57_cast_fp16, y = mul_58_cast_fp16)[name = string("add_6_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_59_cast_fp16 = mul(x = mul_56_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_59_cast_fp16")];
            tensor<int32, [4]> slice_79_begin_0 = const()[name = string("slice_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_79_end_0 = const()[name = string("slice_79_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_79_end_mask_0 = const()[name = string("slice_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_79_cast_fp16 = slice_by_index(begin = slice_79_begin_0, end = slice_79_end_0, end_mask = slice_79_end_mask_0, x = mul_56_cast_fp16)[name = string("slice_79_cast_fp16")];
            tensor<int32, [4]> slice_80_begin_0 = const()[name = string("slice_80_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_80_end_0 = const()[name = string("slice_80_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_80_end_mask_0 = const()[name = string("slice_80_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_80_cast_fp16 = slice_by_index(begin = slice_80_begin_0, end = slice_80_end_0, end_mask = slice_80_end_mask_0, x = mul_56_cast_fp16)[name = string("slice_80_cast_fp16")];
            fp16 const_328_promoted_to_fp16 = const()[name = string("const_328_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_1_cast_fp16 = mul(x = slice_80_cast_fp16, y = const_328_promoted_to_fp16)[name = string("neg_1_cast_fp16")];
            int32 const_329 = const()[name = string("const_329"), val = int32(-1)];
            bool cat_25_interleave_0 = const()[name = string("cat_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_25_cast_fp16 = concat(axis = const_329, interleave = cat_25_interleave_0, values = (neg_1_cast_fp16, slice_79_cast_fp16))[name = string("cat_25_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_60_cast_fp16 = mul(x = cat_25_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_60_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_7_cast_fp16 = add(x = mul_59_cast_fp16, y = mul_60_cast_fp16)[name = string("add_7_cast_fp16")];
            int32 const_330 = const()[name = string("const_330"), val = int32(-2)];
            bool cat_26_interleave_0 = const()[name = string("cat_26_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_26_cast_fp16 = concat(axis = const_330, interleave = cat_26_interleave_0, values = add_7_cast_fp16)[name = string("cat_26_cast_fp16")];
            int32 const_331 = const()[name = string("const_331"), val = int32(-2)];
            bool cat_27_interleave_0 = const()[name = string("cat_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_26_cast_fp16 = transpose(perm = transpose_26_perm_0, x = view_2_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 1, 512, 256]> cat_27_cast_fp16 = concat(axis = const_331, interleave = cat_27_interleave_0, values = transpose_26_cast_fp16)[name = string("cat_27_cast_fp16")];
            tensor<int32, [1]> unsqueeze_79_axes_0 = const()[name = string("unsqueeze_79_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_79_cast_fp16 = expand_dims(axes = unsqueeze_79_axes_0, x = cat_26_cast_fp16)[name = string("unsqueeze_79_cast_fp16")];
            tensor<int32, [5]> expand_26_reps_0 = const()[name = string("expand_26_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_26_cast_fp16 = tile(reps = expand_26_reps_0, x = unsqueeze_79_cast_fp16)[name = string("expand_26_cast_fp16")];
            tensor<int32, [4]> const_346 = const()[name = string("const_346"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_3_cast_fp16 = reshape(shape = const_346, x = expand_26_cast_fp16)[name = string("view_3_cast_fp16")];
            tensor<int32, [1]> unsqueeze_80_axes_0 = const()[name = string("unsqueeze_80_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_80_cast_fp16 = expand_dims(axes = unsqueeze_80_axes_0, x = cat_27_cast_fp16)[name = string("unsqueeze_80_cast_fp16")];
            tensor<int32, [5]> expand_27_reps_0 = const()[name = string("expand_27_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_27_cast_fp16 = tile(reps = expand_27_reps_0, x = unsqueeze_80_cast_fp16)[name = string("expand_27_cast_fp16")];
            tensor<int32, [4]> const_361 = const()[name = string("const_361"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_4_cast_fp16 = reshape(shape = const_361, x = expand_27_cast_fp16)[name = string("view_4_cast_fp16")];
            bool matmul_24_transpose_x_1 = const()[name = string("matmul_24_transpose_x_1"), val = bool(false)];
            bool matmul_24_transpose_y_1 = const()[name = string("matmul_24_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_24_cast_fp16 = matmul(transpose_x = matmul_24_transpose_x_1, transpose_y = matmul_24_transpose_y_1, x = add_6_cast_fp16, y = view_3_cast_fp16)[name = string("matmul_24_cast_fp16")];
            fp16 const_364_to_fp16 = const()[name = string("const_364_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_61_cast_fp16 = mul(x = matmul_24_cast_fp16, y = const_364_to_fp16)[name = string("mul_61_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_8_cast_fp16 = add(x = mul_61_cast_fp16, y = expand_cast_fp16)[name = string("add_8_cast_fp16")];
            int32 const_374 = const()[name = string("const_374"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_cast_fp16 = softmax(axis = const_374, x = add_8_cast_fp16)[name = string("softmax_cast_fp16")];
            bool matmul_25_transpose_x_0 = const()[name = string("matmul_25_transpose_x_0"), val = bool(false)];
            bool matmul_25_transpose_y_0 = const()[name = string("matmul_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_25_cast_fp16 = matmul(transpose_x = matmul_25_transpose_x_0, transpose_y = matmul_25_transpose_y_0, x = softmax_cast_fp16, y = view_4_cast_fp16)[name = string("matmul_25_cast_fp16")];
            tensor<int32, [4]> transpose_28_perm_0 = const()[name = string("transpose_28_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_379 = const()[name = string("const_379"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_28_cast_fp16 = transpose(perm = transpose_28_perm_0, x = matmul_25_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 512, 768]> view_5_cast_fp16 = reshape(shape = const_379, x = transpose_28_cast_fp16)[name = string("view_5_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203104256))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203694144))))[name = string("p_st_0_model_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_3_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = view_5_cast_fp16)[name = string("linear_3_cast_fp16")];
            fp16 const_381_promoted_to_fp16 = const()[name = string("const_381_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_4_cast_fp16 = pow(x = linear_3_cast_fp16, y = const_381_promoted_to_fp16)[name = string("pow_4_cast_fp16")];
            tensor<int32, [1]> mean_3_axes_0 = const()[name = string("mean_3_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_3_keep_dims_0 = const()[name = string("mean_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_3_cast_fp16 = reduce_mean(axes = mean_3_axes_0, keep_dims = mean_3_keep_dims_0, x = pow_4_cast_fp16)[name = string("mean_3_cast_fp16")];
            fp16 const_384_to_fp16 = const()[name = string("const_384_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_9_cast_fp16 = add(x = mean_3_cast_fp16, y = const_384_to_fp16)[name = string("add_9_cast_fp16")];
            fp32 rsqrt_3_epsilon_0 = const()[name = string("rsqrt_3_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_3_cast_fp16 = rsqrt(epsilon = rsqrt_3_epsilon_0, x = add_9_cast_fp16)[name = string("rsqrt_3_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_62_cast_fp16 = mul(x = linear_3_cast_fp16, y = rsqrt_3_cast_fp16)[name = string("mul_62_cast_fp16")];
            tensor<fp16, [768]> add_10_to_fp16 = const()[name = string("add_10_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203695744)))];
            tensor<fp16, [1, 512, 768]> mul_63_cast_fp16 = mul(x = mul_62_cast_fp16, y = add_10_to_fp16)[name = string("mul_63_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_11_cast_fp16 = add(x = mul_cast_fp16, y = mul_63_cast_fp16)[name = string("add_11_cast_fp16")];
            fp16 const_389_promoted_to_fp16 = const()[name = string("const_389_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_5_cast_fp16 = pow(x = add_11_cast_fp16, y = const_389_promoted_to_fp16)[name = string("pow_5_cast_fp16")];
            tensor<int32, [1]> mean_4_axes_0 = const()[name = string("mean_4_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_4_keep_dims_0 = const()[name = string("mean_4_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_4_cast_fp16 = reduce_mean(axes = mean_4_axes_0, keep_dims = mean_4_keep_dims_0, x = pow_5_cast_fp16)[name = string("mean_4_cast_fp16")];
            fp16 const_392_to_fp16 = const()[name = string("const_392_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_12_cast_fp16 = add(x = mean_4_cast_fp16, y = const_392_to_fp16)[name = string("add_12_cast_fp16")];
            fp32 rsqrt_4_epsilon_0 = const()[name = string("rsqrt_4_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_4_cast_fp16 = rsqrt(epsilon = rsqrt_4_epsilon_0, x = add_12_cast_fp16)[name = string("rsqrt_4_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_64_cast_fp16 = mul(x = add_11_cast_fp16, y = rsqrt_4_cast_fp16)[name = string("mul_64_cast_fp16")];
            tensor<fp16, [768]> add_13_to_fp16 = const()[name = string("add_13_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203697344)))];
            tensor<fp16, [1, 512, 768]> mul_65_cast_fp16 = mul(x = mul_64_cast_fp16, y = add_13_to_fp16)[name = string("mul_65_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203698944))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204583744))))[name = string("p_st_0_model_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1152]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204586112)))];
            tensor<fp16, [1, 512, 1152]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = mul_65_cast_fp16)[name = string("linear_4_cast_fp16")];
            string gelu_mode_0 = const()[name = string("gelu_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_cast_fp16 = gelu(mode = gelu_mode_0, x = linear_4_cast_fp16)[name = string("gelu_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204588480))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205473280))))[name = string("p_st_0_model_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = mul_65_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_66_cast_fp16 = mul(x = gelu_cast_fp16, y = linear_5_cast_fp16)[name = string("mul_66_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205475648))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206360448))))[name = string("p_st_0_model_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = mul_66_cast_fp16)[name = string("linear_6_cast_fp16")];
            fp16 const_397_promoted_to_fp16 = const()[name = string("const_397_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_6_cast_fp16 = pow(x = linear_6_cast_fp16, y = const_397_promoted_to_fp16)[name = string("pow_6_cast_fp16")];
            tensor<int32, [1]> mean_5_axes_0 = const()[name = string("mean_5_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_5_keep_dims_0 = const()[name = string("mean_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_5_cast_fp16 = reduce_mean(axes = mean_5_axes_0, keep_dims = mean_5_keep_dims_0, x = pow_6_cast_fp16)[name = string("mean_5_cast_fp16")];
            fp16 const_400_to_fp16 = const()[name = string("const_400_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_14_cast_fp16 = add(x = mean_5_cast_fp16, y = const_400_to_fp16)[name = string("add_14_cast_fp16")];
            fp32 rsqrt_5_epsilon_0 = const()[name = string("rsqrt_5_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_5_cast_fp16 = rsqrt(epsilon = rsqrt_5_epsilon_0, x = add_14_cast_fp16)[name = string("rsqrt_5_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_67_cast_fp16 = mul(x = linear_6_cast_fp16, y = rsqrt_5_cast_fp16)[name = string("mul_67_cast_fp16")];
            tensor<fp16, [768]> add_15_to_fp16 = const()[name = string("add_15_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206362048)))];
            tensor<fp16, [1, 512, 768]> mul_68_cast_fp16 = mul(x = mul_67_cast_fp16, y = add_15_to_fp16)[name = string("mul_68_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_16_cast_fp16 = add(x = add_11_cast_fp16, y = mul_68_cast_fp16)[name = string("add_16_cast_fp16")];
            fp16 const_405_promoted_to_fp16 = const()[name = string("const_405_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_7_cast_fp16 = pow(x = add_16_cast_fp16, y = const_405_promoted_to_fp16)[name = string("pow_7_cast_fp16")];
            tensor<int32, [1]> mean_6_axes_0 = const()[name = string("mean_6_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_6_keep_dims_0 = const()[name = string("mean_6_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_6_cast_fp16 = reduce_mean(axes = mean_6_axes_0, keep_dims = mean_6_keep_dims_0, x = pow_7_cast_fp16)[name = string("mean_6_cast_fp16")];
            fp16 const_408_to_fp16 = const()[name = string("const_408_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_17_cast_fp16 = add(x = mean_6_cast_fp16, y = const_408_to_fp16)[name = string("add_17_cast_fp16")];
            fp32 rsqrt_6_epsilon_0 = const()[name = string("rsqrt_6_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_6_cast_fp16 = rsqrt(epsilon = rsqrt_6_epsilon_0, x = add_17_cast_fp16)[name = string("rsqrt_6_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_69_cast_fp16 = mul(x = add_16_cast_fp16, y = rsqrt_6_cast_fp16)[name = string("mul_69_cast_fp16")];
            tensor<fp16, [768]> add_18_to_fp16 = const()[name = string("add_18_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206363648)))];
            tensor<fp16, [1, 512, 768]> mul_70_cast_fp16 = mul(x = mul_69_cast_fp16, y = add_18_to_fp16)[name = string("mul_70_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206365248))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206955136))))[name = string("p_st_0_model_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_7_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = mul_70_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> const_412 = const()[name = string("const_412"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_6_cast_fp16 = reshape(shape = const_412, x = linear_7_cast_fp16)[name = string("view_6_cast_fp16")];
            tensor<int32, [4]> transpose_29_perm_0 = const()[name = string("transpose_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206956736))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207153408))))[name = string("p_st_0_model_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_8_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = mul_70_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> const_415 = const()[name = string("const_415"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_7_cast_fp16 = reshape(shape = const_415, x = linear_8_cast_fp16)[name = string("view_7_cast_fp16")];
            tensor<int32, [4]> transpose_30_perm_0 = const()[name = string("transpose_30_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207153984))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207350656))))[name = string("p_st_0_model_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = mul_70_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> const_418 = const()[name = string("const_418"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_8_cast_fp16 = reshape(shape = const_418, x = linear_9_cast_fp16)[name = string("view_8_cast_fp16")];
            tensor<int32, [4]> transpose_31_perm_0 = const()[name = string("transpose_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_422_promoted_to_fp16 = const()[name = string("const_422_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_29_cast_fp16 = transpose(perm = transpose_29_perm_0, x = view_6_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 3, 512, 256]> pow_8_cast_fp16 = pow(x = transpose_29_cast_fp16, y = const_422_promoted_to_fp16)[name = string("pow_8_cast_fp16")];
            tensor<int32, [1]> mean_7_axes_0 = const()[name = string("mean_7_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_7_keep_dims_0 = const()[name = string("mean_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_7_cast_fp16 = reduce_mean(axes = mean_7_axes_0, keep_dims = mean_7_keep_dims_0, x = pow_8_cast_fp16)[name = string("mean_7_cast_fp16")];
            fp16 const_425_to_fp16 = const()[name = string("const_425_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_19_cast_fp16 = add(x = mean_7_cast_fp16, y = const_425_to_fp16)[name = string("add_19_cast_fp16")];
            fp32 rsqrt_7_epsilon_0 = const()[name = string("rsqrt_7_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_7_cast_fp16 = rsqrt(epsilon = rsqrt_7_epsilon_0, x = add_19_cast_fp16)[name = string("rsqrt_7_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_71_cast_fp16 = mul(x = transpose_29_cast_fp16, y = rsqrt_7_cast_fp16)[name = string("mul_71_cast_fp16")];
            tensor<fp16, [256]> add_20_to_fp16 = const()[name = string("add_20_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207351232)))];
            tensor<fp16, [1, 3, 512, 256]> mul_72_cast_fp16 = mul(x = mul_71_cast_fp16, y = add_20_to_fp16)[name = string("mul_72_cast_fp16")];
            fp16 const_430_promoted_to_fp16 = const()[name = string("const_430_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_30_cast_fp16 = transpose(perm = transpose_30_perm_0, x = view_7_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 1, 512, 256]> pow_9_cast_fp16 = pow(x = transpose_30_cast_fp16, y = const_430_promoted_to_fp16)[name = string("pow_9_cast_fp16")];
            tensor<int32, [1]> mean_8_axes_0 = const()[name = string("mean_8_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_8_keep_dims_0 = const()[name = string("mean_8_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_8_cast_fp16 = reduce_mean(axes = mean_8_axes_0, keep_dims = mean_8_keep_dims_0, x = pow_9_cast_fp16)[name = string("mean_8_cast_fp16")];
            fp16 const_433_to_fp16 = const()[name = string("const_433_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_21_cast_fp16 = add(x = mean_8_cast_fp16, y = const_433_to_fp16)[name = string("add_21_cast_fp16")];
            fp32 rsqrt_8_epsilon_0 = const()[name = string("rsqrt_8_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_8_cast_fp16 = rsqrt(epsilon = rsqrt_8_epsilon_0, x = add_21_cast_fp16)[name = string("rsqrt_8_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_73_cast_fp16 = mul(x = transpose_30_cast_fp16, y = rsqrt_8_cast_fp16)[name = string("mul_73_cast_fp16")];
            tensor<fp16, [256]> add_22_to_fp16 = const()[name = string("add_22_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207351808)))];
            tensor<fp16, [1, 1, 512, 256]> mul_74_cast_fp16 = mul(x = mul_73_cast_fp16, y = add_22_to_fp16)[name = string("mul_74_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_75_cast_fp16 = mul(x = mul_72_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_75_cast_fp16")];
            tensor<int32, [4]> slice_100_begin_0 = const()[name = string("slice_100_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_100_end_0 = const()[name = string("slice_100_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_100_end_mask_0 = const()[name = string("slice_100_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_100_cast_fp16 = slice_by_index(begin = slice_100_begin_0, end = slice_100_end_0, end_mask = slice_100_end_mask_0, x = mul_72_cast_fp16)[name = string("slice_100_cast_fp16")];
            tensor<int32, [4]> slice_101_begin_0 = const()[name = string("slice_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_101_end_0 = const()[name = string("slice_101_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_101_end_mask_0 = const()[name = string("slice_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_101_cast_fp16 = slice_by_index(begin = slice_101_begin_0, end = slice_101_end_0, end_mask = slice_101_end_mask_0, x = mul_72_cast_fp16)[name = string("slice_101_cast_fp16")];
            fp16 const_445_promoted_to_fp16 = const()[name = string("const_445_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_2_cast_fp16 = mul(x = slice_101_cast_fp16, y = const_445_promoted_to_fp16)[name = string("neg_2_cast_fp16")];
            int32 const_446 = const()[name = string("const_446"), val = int32(-1)];
            bool cat_28_interleave_0 = const()[name = string("cat_28_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_28_cast_fp16 = concat(axis = const_446, interleave = cat_28_interleave_0, values = (neg_2_cast_fp16, slice_100_cast_fp16))[name = string("cat_28_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_76_cast_fp16 = mul(x = cat_28_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_76_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_23_cast_fp16 = add(x = mul_75_cast_fp16, y = mul_76_cast_fp16)[name = string("add_23_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_77_cast_fp16 = mul(x = mul_74_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_77_cast_fp16")];
            tensor<int32, [4]> slice_102_begin_0 = const()[name = string("slice_102_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_102_end_0 = const()[name = string("slice_102_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_102_end_mask_0 = const()[name = string("slice_102_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_102_cast_fp16 = slice_by_index(begin = slice_102_begin_0, end = slice_102_end_0, end_mask = slice_102_end_mask_0, x = mul_74_cast_fp16)[name = string("slice_102_cast_fp16")];
            tensor<int32, [4]> slice_103_begin_0 = const()[name = string("slice_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_103_end_0 = const()[name = string("slice_103_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_103_end_mask_0 = const()[name = string("slice_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_103_cast_fp16 = slice_by_index(begin = slice_103_begin_0, end = slice_103_end_0, end_mask = slice_103_end_mask_0, x = mul_74_cast_fp16)[name = string("slice_103_cast_fp16")];
            fp16 const_453_promoted_to_fp16 = const()[name = string("const_453_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_3_cast_fp16 = mul(x = slice_103_cast_fp16, y = const_453_promoted_to_fp16)[name = string("neg_3_cast_fp16")];
            int32 const_454 = const()[name = string("const_454"), val = int32(-1)];
            bool cat_29_interleave_0 = const()[name = string("cat_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_29_cast_fp16 = concat(axis = const_454, interleave = cat_29_interleave_0, values = (neg_3_cast_fp16, slice_102_cast_fp16))[name = string("cat_29_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_78_cast_fp16 = mul(x = cat_29_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_78_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_24_cast_fp16 = add(x = mul_77_cast_fp16, y = mul_78_cast_fp16)[name = string("add_24_cast_fp16")];
            int32 const_455 = const()[name = string("const_455"), val = int32(-2)];
            bool cat_30_interleave_0 = const()[name = string("cat_30_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_30_cast_fp16 = concat(axis = const_455, interleave = cat_30_interleave_0, values = add_24_cast_fp16)[name = string("cat_30_cast_fp16")];
            int32 const_456 = const()[name = string("const_456"), val = int32(-2)];
            bool cat_31_interleave_0 = const()[name = string("cat_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_31_cast_fp16 = transpose(perm = transpose_31_perm_0, x = view_8_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 1, 512, 256]> cat_31_cast_fp16 = concat(axis = const_456, interleave = cat_31_interleave_0, values = transpose_31_cast_fp16)[name = string("cat_31_cast_fp16")];
            tensor<int32, [1]> unsqueeze_83_axes_0 = const()[name = string("unsqueeze_83_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_83_cast_fp16 = expand_dims(axes = unsqueeze_83_axes_0, x = cat_30_cast_fp16)[name = string("unsqueeze_83_cast_fp16")];
            tensor<int32, [5]> expand_28_reps_0 = const()[name = string("expand_28_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_28_cast_fp16 = tile(reps = expand_28_reps_0, x = unsqueeze_83_cast_fp16)[name = string("expand_28_cast_fp16")];
            tensor<int32, [4]> const_471 = const()[name = string("const_471"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_9_cast_fp16 = reshape(shape = const_471, x = expand_28_cast_fp16)[name = string("view_9_cast_fp16")];
            tensor<int32, [1]> unsqueeze_84_axes_0 = const()[name = string("unsqueeze_84_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_84_cast_fp16 = expand_dims(axes = unsqueeze_84_axes_0, x = cat_31_cast_fp16)[name = string("unsqueeze_84_cast_fp16")];
            tensor<int32, [5]> expand_29_reps_0 = const()[name = string("expand_29_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_29_cast_fp16 = tile(reps = expand_29_reps_0, x = unsqueeze_84_cast_fp16)[name = string("expand_29_cast_fp16")];
            tensor<int32, [4]> const_486 = const()[name = string("const_486"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_10_cast_fp16 = reshape(shape = const_486, x = expand_29_cast_fp16)[name = string("view_10_cast_fp16")];
            bool matmul_26_transpose_x_1 = const()[name = string("matmul_26_transpose_x_1"), val = bool(false)];
            bool matmul_26_transpose_y_1 = const()[name = string("matmul_26_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_26_cast_fp16 = matmul(transpose_x = matmul_26_transpose_x_1, transpose_y = matmul_26_transpose_y_1, x = add_23_cast_fp16, y = view_9_cast_fp16)[name = string("matmul_26_cast_fp16")];
            fp16 const_489_to_fp16 = const()[name = string("const_489_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_79_cast_fp16 = mul(x = matmul_26_cast_fp16, y = const_489_to_fp16)[name = string("mul_79_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_25_cast_fp16 = add(x = mul_79_cast_fp16, y = expand_cast_fp16)[name = string("add_25_cast_fp16")];
            int32 const_499 = const()[name = string("const_499"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_1_cast_fp16 = softmax(axis = const_499, x = add_25_cast_fp16)[name = string("softmax_1_cast_fp16")];
            bool matmul_27_transpose_x_0 = const()[name = string("matmul_27_transpose_x_0"), val = bool(false)];
            bool matmul_27_transpose_y_0 = const()[name = string("matmul_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_27_cast_fp16 = matmul(transpose_x = matmul_27_transpose_x_0, transpose_y = matmul_27_transpose_y_0, x = softmax_1_cast_fp16, y = view_10_cast_fp16)[name = string("matmul_27_cast_fp16")];
            tensor<int32, [4]> transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_504 = const()[name = string("const_504"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_33_cast_fp16 = transpose(perm = transpose_33_perm_0, x = matmul_27_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 512, 768]> view_11_cast_fp16 = reshape(shape = const_504, x = transpose_33_cast_fp16)[name = string("view_11_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207352384))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207942272))))[name = string("p_st_0_model_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = view_11_cast_fp16)[name = string("linear_10_cast_fp16")];
            fp16 const_506_promoted_to_fp16 = const()[name = string("const_506_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_10_cast_fp16 = pow(x = linear_10_cast_fp16, y = const_506_promoted_to_fp16)[name = string("pow_10_cast_fp16")];
            tensor<int32, [1]> mean_9_axes_0 = const()[name = string("mean_9_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_9_keep_dims_0 = const()[name = string("mean_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_9_cast_fp16 = reduce_mean(axes = mean_9_axes_0, keep_dims = mean_9_keep_dims_0, x = pow_10_cast_fp16)[name = string("mean_9_cast_fp16")];
            fp16 const_509_to_fp16 = const()[name = string("const_509_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_26_cast_fp16 = add(x = mean_9_cast_fp16, y = const_509_to_fp16)[name = string("add_26_cast_fp16")];
            fp32 rsqrt_9_epsilon_0 = const()[name = string("rsqrt_9_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_9_cast_fp16 = rsqrt(epsilon = rsqrt_9_epsilon_0, x = add_26_cast_fp16)[name = string("rsqrt_9_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_80_cast_fp16 = mul(x = linear_10_cast_fp16, y = rsqrt_9_cast_fp16)[name = string("mul_80_cast_fp16")];
            tensor<fp16, [768]> add_27_to_fp16 = const()[name = string("add_27_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207943872)))];
            tensor<fp16, [1, 512, 768]> mul_81_cast_fp16 = mul(x = mul_80_cast_fp16, y = add_27_to_fp16)[name = string("mul_81_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_28_cast_fp16 = add(x = add_16_cast_fp16, y = mul_81_cast_fp16)[name = string("add_28_cast_fp16")];
            fp16 const_514_promoted_to_fp16 = const()[name = string("const_514_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_11_cast_fp16 = pow(x = add_28_cast_fp16, y = const_514_promoted_to_fp16)[name = string("pow_11_cast_fp16")];
            tensor<int32, [1]> mean_10_axes_0 = const()[name = string("mean_10_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_10_keep_dims_0 = const()[name = string("mean_10_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_10_cast_fp16 = reduce_mean(axes = mean_10_axes_0, keep_dims = mean_10_keep_dims_0, x = pow_11_cast_fp16)[name = string("mean_10_cast_fp16")];
            fp16 const_517_to_fp16 = const()[name = string("const_517_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_29_cast_fp16 = add(x = mean_10_cast_fp16, y = const_517_to_fp16)[name = string("add_29_cast_fp16")];
            fp32 rsqrt_10_epsilon_0 = const()[name = string("rsqrt_10_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_10_cast_fp16 = rsqrt(epsilon = rsqrt_10_epsilon_0, x = add_29_cast_fp16)[name = string("rsqrt_10_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_82_cast_fp16 = mul(x = add_28_cast_fp16, y = rsqrt_10_cast_fp16)[name = string("mul_82_cast_fp16")];
            tensor<fp16, [768]> add_30_to_fp16 = const()[name = string("add_30_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207945472)))];
            tensor<fp16, [1, 512, 768]> mul_83_cast_fp16 = mul(x = mul_82_cast_fp16, y = add_30_to_fp16)[name = string("mul_83_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207947072))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208831872))))[name = string("p_st_0_model_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = mul_83_cast_fp16)[name = string("linear_11_cast_fp16")];
            string gelu_1_mode_0 = const()[name = string("gelu_1_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_1_cast_fp16 = gelu(mode = gelu_1_mode_0, x = linear_11_cast_fp16)[name = string("gelu_1_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208834240))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209719040))))[name = string("p_st_0_model_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = mul_83_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_84_cast_fp16 = mul(x = gelu_1_cast_fp16, y = linear_12_cast_fp16)[name = string("mul_84_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209721408))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210606208))))[name = string("p_st_0_model_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_13_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = mul_84_cast_fp16)[name = string("linear_13_cast_fp16")];
            fp16 const_522_promoted_to_fp16 = const()[name = string("const_522_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_12_cast_fp16 = pow(x = linear_13_cast_fp16, y = const_522_promoted_to_fp16)[name = string("pow_12_cast_fp16")];
            tensor<int32, [1]> mean_11_axes_0 = const()[name = string("mean_11_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_11_keep_dims_0 = const()[name = string("mean_11_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_11_cast_fp16 = reduce_mean(axes = mean_11_axes_0, keep_dims = mean_11_keep_dims_0, x = pow_12_cast_fp16)[name = string("mean_11_cast_fp16")];
            fp16 const_525_to_fp16 = const()[name = string("const_525_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_31_cast_fp16 = add(x = mean_11_cast_fp16, y = const_525_to_fp16)[name = string("add_31_cast_fp16")];
            fp32 rsqrt_11_epsilon_0 = const()[name = string("rsqrt_11_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_11_cast_fp16 = rsqrt(epsilon = rsqrt_11_epsilon_0, x = add_31_cast_fp16)[name = string("rsqrt_11_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_85_cast_fp16 = mul(x = linear_13_cast_fp16, y = rsqrt_11_cast_fp16)[name = string("mul_85_cast_fp16")];
            tensor<fp16, [768]> add_32_to_fp16 = const()[name = string("add_32_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210607808)))];
            tensor<fp16, [1, 512, 768]> mul_86_cast_fp16 = mul(x = mul_85_cast_fp16, y = add_32_to_fp16)[name = string("mul_86_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_33_cast_fp16 = add(x = add_28_cast_fp16, y = mul_86_cast_fp16)[name = string("add_33_cast_fp16")];
            fp16 const_530_promoted_to_fp16 = const()[name = string("const_530_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_13_cast_fp16 = pow(x = add_33_cast_fp16, y = const_530_promoted_to_fp16)[name = string("pow_13_cast_fp16")];
            tensor<int32, [1]> mean_12_axes_0 = const()[name = string("mean_12_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_12_keep_dims_0 = const()[name = string("mean_12_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_12_cast_fp16 = reduce_mean(axes = mean_12_axes_0, keep_dims = mean_12_keep_dims_0, x = pow_13_cast_fp16)[name = string("mean_12_cast_fp16")];
            fp16 const_533_to_fp16 = const()[name = string("const_533_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_34_cast_fp16 = add(x = mean_12_cast_fp16, y = const_533_to_fp16)[name = string("add_34_cast_fp16")];
            fp32 rsqrt_12_epsilon_0 = const()[name = string("rsqrt_12_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_12_cast_fp16 = rsqrt(epsilon = rsqrt_12_epsilon_0, x = add_34_cast_fp16)[name = string("rsqrt_12_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_87_cast_fp16 = mul(x = add_33_cast_fp16, y = rsqrt_12_cast_fp16)[name = string("mul_87_cast_fp16")];
            tensor<fp16, [768]> add_35_to_fp16 = const()[name = string("add_35_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210609408)))];
            tensor<fp16, [1, 512, 768]> mul_88_cast_fp16 = mul(x = mul_87_cast_fp16, y = add_35_to_fp16)[name = string("mul_88_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210611008))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211200896))))[name = string("p_st_0_model_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = mul_88_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> const_537 = const()[name = string("const_537"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_12_cast_fp16 = reshape(shape = const_537, x = linear_14_cast_fp16)[name = string("view_12_cast_fp16")];
            tensor<int32, [4]> transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211202496))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211399168))))[name = string("p_st_0_model_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = mul_88_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> const_540 = const()[name = string("const_540"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_13_cast_fp16 = reshape(shape = const_540, x = linear_15_cast_fp16)[name = string("view_13_cast_fp16")];
            tensor<int32, [4]> transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211399744))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211596416))))[name = string("p_st_0_model_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_16_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = mul_88_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> const_543 = const()[name = string("const_543"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_14_cast_fp16 = reshape(shape = const_543, x = linear_16_cast_fp16)[name = string("view_14_cast_fp16")];
            tensor<int32, [4]> transpose_36_perm_0 = const()[name = string("transpose_36_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_547_promoted_to_fp16 = const()[name = string("const_547_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_34_cast_fp16 = transpose(perm = transpose_34_perm_0, x = view_12_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 3, 512, 256]> pow_14_cast_fp16 = pow(x = transpose_34_cast_fp16, y = const_547_promoted_to_fp16)[name = string("pow_14_cast_fp16")];
            tensor<int32, [1]> mean_13_axes_0 = const()[name = string("mean_13_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_13_keep_dims_0 = const()[name = string("mean_13_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_13_cast_fp16 = reduce_mean(axes = mean_13_axes_0, keep_dims = mean_13_keep_dims_0, x = pow_14_cast_fp16)[name = string("mean_13_cast_fp16")];
            fp16 const_550_to_fp16 = const()[name = string("const_550_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_36_cast_fp16 = add(x = mean_13_cast_fp16, y = const_550_to_fp16)[name = string("add_36_cast_fp16")];
            fp32 rsqrt_13_epsilon_0 = const()[name = string("rsqrt_13_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_13_cast_fp16 = rsqrt(epsilon = rsqrt_13_epsilon_0, x = add_36_cast_fp16)[name = string("rsqrt_13_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_89_cast_fp16 = mul(x = transpose_34_cast_fp16, y = rsqrt_13_cast_fp16)[name = string("mul_89_cast_fp16")];
            tensor<fp16, [256]> add_37_to_fp16 = const()[name = string("add_37_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211596992)))];
            tensor<fp16, [1, 3, 512, 256]> mul_90_cast_fp16 = mul(x = mul_89_cast_fp16, y = add_37_to_fp16)[name = string("mul_90_cast_fp16")];
            fp16 const_555_promoted_to_fp16 = const()[name = string("const_555_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_35_cast_fp16 = transpose(perm = transpose_35_perm_0, x = view_13_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 1, 512, 256]> pow_15_cast_fp16 = pow(x = transpose_35_cast_fp16, y = const_555_promoted_to_fp16)[name = string("pow_15_cast_fp16")];
            tensor<int32, [1]> mean_14_axes_0 = const()[name = string("mean_14_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_14_keep_dims_0 = const()[name = string("mean_14_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_14_cast_fp16 = reduce_mean(axes = mean_14_axes_0, keep_dims = mean_14_keep_dims_0, x = pow_15_cast_fp16)[name = string("mean_14_cast_fp16")];
            fp16 const_558_to_fp16 = const()[name = string("const_558_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_38_cast_fp16 = add(x = mean_14_cast_fp16, y = const_558_to_fp16)[name = string("add_38_cast_fp16")];
            fp32 rsqrt_14_epsilon_0 = const()[name = string("rsqrt_14_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_14_cast_fp16 = rsqrt(epsilon = rsqrt_14_epsilon_0, x = add_38_cast_fp16)[name = string("rsqrt_14_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_91_cast_fp16 = mul(x = transpose_35_cast_fp16, y = rsqrt_14_cast_fp16)[name = string("mul_91_cast_fp16")];
            tensor<fp16, [256]> add_39_to_fp16 = const()[name = string("add_39_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211597568)))];
            tensor<fp16, [1, 1, 512, 256]> mul_92_cast_fp16 = mul(x = mul_91_cast_fp16, y = add_39_to_fp16)[name = string("mul_92_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_93_cast_fp16 = mul(x = mul_90_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_93_cast_fp16")];
            tensor<int32, [4]> slice_123_begin_0 = const()[name = string("slice_123_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_123_end_0 = const()[name = string("slice_123_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_123_end_mask_0 = const()[name = string("slice_123_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_123_cast_fp16 = slice_by_index(begin = slice_123_begin_0, end = slice_123_end_0, end_mask = slice_123_end_mask_0, x = mul_90_cast_fp16)[name = string("slice_123_cast_fp16")];
            tensor<int32, [4]> slice_124_begin_0 = const()[name = string("slice_124_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_124_end_0 = const()[name = string("slice_124_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_124_end_mask_0 = const()[name = string("slice_124_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_124_cast_fp16 = slice_by_index(begin = slice_124_begin_0, end = slice_124_end_0, end_mask = slice_124_end_mask_0, x = mul_90_cast_fp16)[name = string("slice_124_cast_fp16")];
            fp16 const_570_promoted_to_fp16 = const()[name = string("const_570_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_4_cast_fp16 = mul(x = slice_124_cast_fp16, y = const_570_promoted_to_fp16)[name = string("neg_4_cast_fp16")];
            int32 const_571 = const()[name = string("const_571"), val = int32(-1)];
            bool cat_32_interleave_0 = const()[name = string("cat_32_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_32_cast_fp16 = concat(axis = const_571, interleave = cat_32_interleave_0, values = (neg_4_cast_fp16, slice_123_cast_fp16))[name = string("cat_32_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_94_cast_fp16 = mul(x = cat_32_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_94_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_40_cast_fp16 = add(x = mul_93_cast_fp16, y = mul_94_cast_fp16)[name = string("add_40_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_95_cast_fp16 = mul(x = mul_92_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_95_cast_fp16")];
            tensor<int32, [4]> slice_125_begin_0 = const()[name = string("slice_125_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_125_end_0 = const()[name = string("slice_125_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_125_end_mask_0 = const()[name = string("slice_125_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_125_cast_fp16 = slice_by_index(begin = slice_125_begin_0, end = slice_125_end_0, end_mask = slice_125_end_mask_0, x = mul_92_cast_fp16)[name = string("slice_125_cast_fp16")];
            tensor<int32, [4]> slice_126_begin_0 = const()[name = string("slice_126_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_126_end_0 = const()[name = string("slice_126_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_126_end_mask_0 = const()[name = string("slice_126_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_126_cast_fp16 = slice_by_index(begin = slice_126_begin_0, end = slice_126_end_0, end_mask = slice_126_end_mask_0, x = mul_92_cast_fp16)[name = string("slice_126_cast_fp16")];
            fp16 const_578_promoted_to_fp16 = const()[name = string("const_578_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_5_cast_fp16 = mul(x = slice_126_cast_fp16, y = const_578_promoted_to_fp16)[name = string("neg_5_cast_fp16")];
            int32 const_579 = const()[name = string("const_579"), val = int32(-1)];
            bool cat_33_interleave_0 = const()[name = string("cat_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_33_cast_fp16 = concat(axis = const_579, interleave = cat_33_interleave_0, values = (neg_5_cast_fp16, slice_125_cast_fp16))[name = string("cat_33_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_96_cast_fp16 = mul(x = cat_33_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_96_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_41_cast_fp16 = add(x = mul_95_cast_fp16, y = mul_96_cast_fp16)[name = string("add_41_cast_fp16")];
            int32 const_580 = const()[name = string("const_580"), val = int32(-2)];
            bool cat_34_interleave_0 = const()[name = string("cat_34_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_34_cast_fp16 = concat(axis = const_580, interleave = cat_34_interleave_0, values = add_41_cast_fp16)[name = string("cat_34_cast_fp16")];
            int32 const_581 = const()[name = string("const_581"), val = int32(-2)];
            bool cat_35_interleave_0 = const()[name = string("cat_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_36_cast_fp16 = transpose(perm = transpose_36_perm_0, x = view_14_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 1, 512, 256]> cat_35_cast_fp16 = concat(axis = const_581, interleave = cat_35_interleave_0, values = transpose_36_cast_fp16)[name = string("cat_35_cast_fp16")];
            tensor<int32, [1]> unsqueeze_87_axes_0 = const()[name = string("unsqueeze_87_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_87_cast_fp16 = expand_dims(axes = unsqueeze_87_axes_0, x = cat_34_cast_fp16)[name = string("unsqueeze_87_cast_fp16")];
            tensor<int32, [5]> expand_30_reps_0 = const()[name = string("expand_30_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_30_cast_fp16 = tile(reps = expand_30_reps_0, x = unsqueeze_87_cast_fp16)[name = string("expand_30_cast_fp16")];
            tensor<int32, [4]> const_596 = const()[name = string("const_596"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_15_cast_fp16 = reshape(shape = const_596, x = expand_30_cast_fp16)[name = string("view_15_cast_fp16")];
            tensor<int32, [1]> unsqueeze_88_axes_0 = const()[name = string("unsqueeze_88_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_88_cast_fp16 = expand_dims(axes = unsqueeze_88_axes_0, x = cat_35_cast_fp16)[name = string("unsqueeze_88_cast_fp16")];
            tensor<int32, [5]> expand_31_reps_0 = const()[name = string("expand_31_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_31_cast_fp16 = tile(reps = expand_31_reps_0, x = unsqueeze_88_cast_fp16)[name = string("expand_31_cast_fp16")];
            tensor<int32, [4]> const_611 = const()[name = string("const_611"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_16_cast_fp16 = reshape(shape = const_611, x = expand_31_cast_fp16)[name = string("view_16_cast_fp16")];
            bool matmul_28_transpose_x_1 = const()[name = string("matmul_28_transpose_x_1"), val = bool(false)];
            bool matmul_28_transpose_y_1 = const()[name = string("matmul_28_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_28_cast_fp16 = matmul(transpose_x = matmul_28_transpose_x_1, transpose_y = matmul_28_transpose_y_1, x = add_40_cast_fp16, y = view_15_cast_fp16)[name = string("matmul_28_cast_fp16")];
            fp16 const_614_to_fp16 = const()[name = string("const_614_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_97_cast_fp16 = mul(x = matmul_28_cast_fp16, y = const_614_to_fp16)[name = string("mul_97_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_42_cast_fp16 = add(x = mul_97_cast_fp16, y = expand_cast_fp16)[name = string("add_42_cast_fp16")];
            int32 const_624 = const()[name = string("const_624"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_2_cast_fp16 = softmax(axis = const_624, x = add_42_cast_fp16)[name = string("softmax_2_cast_fp16")];
            bool matmul_29_transpose_x_0 = const()[name = string("matmul_29_transpose_x_0"), val = bool(false)];
            bool matmul_29_transpose_y_0 = const()[name = string("matmul_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_29_cast_fp16 = matmul(transpose_x = matmul_29_transpose_x_0, transpose_y = matmul_29_transpose_y_0, x = softmax_2_cast_fp16, y = view_16_cast_fp16)[name = string("matmul_29_cast_fp16")];
            tensor<int32, [4]> transpose_38_perm_0 = const()[name = string("transpose_38_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_629 = const()[name = string("const_629"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_38_cast_fp16 = transpose(perm = transpose_38_perm_0, x = matmul_29_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 512, 768]> view_17_cast_fp16 = reshape(shape = const_629, x = transpose_38_cast_fp16)[name = string("view_17_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211598144))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212188032))))[name = string("p_st_0_model_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_17_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = view_17_cast_fp16)[name = string("linear_17_cast_fp16")];
            fp16 const_631_promoted_to_fp16 = const()[name = string("const_631_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_16_cast_fp16 = pow(x = linear_17_cast_fp16, y = const_631_promoted_to_fp16)[name = string("pow_16_cast_fp16")];
            tensor<int32, [1]> mean_15_axes_0 = const()[name = string("mean_15_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_15_keep_dims_0 = const()[name = string("mean_15_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_15_cast_fp16 = reduce_mean(axes = mean_15_axes_0, keep_dims = mean_15_keep_dims_0, x = pow_16_cast_fp16)[name = string("mean_15_cast_fp16")];
            fp16 const_634_to_fp16 = const()[name = string("const_634_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_43_cast_fp16 = add(x = mean_15_cast_fp16, y = const_634_to_fp16)[name = string("add_43_cast_fp16")];
            fp32 rsqrt_15_epsilon_0 = const()[name = string("rsqrt_15_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_15_cast_fp16 = rsqrt(epsilon = rsqrt_15_epsilon_0, x = add_43_cast_fp16)[name = string("rsqrt_15_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_98_cast_fp16 = mul(x = linear_17_cast_fp16, y = rsqrt_15_cast_fp16)[name = string("mul_98_cast_fp16")];
            tensor<fp16, [768]> add_44_to_fp16 = const()[name = string("add_44_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212189632)))];
            tensor<fp16, [1, 512, 768]> mul_99_cast_fp16 = mul(x = mul_98_cast_fp16, y = add_44_to_fp16)[name = string("mul_99_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_45_cast_fp16 = add(x = add_33_cast_fp16, y = mul_99_cast_fp16)[name = string("add_45_cast_fp16")];
            fp16 const_639_promoted_to_fp16 = const()[name = string("const_639_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_17_cast_fp16 = pow(x = add_45_cast_fp16, y = const_639_promoted_to_fp16)[name = string("pow_17_cast_fp16")];
            tensor<int32, [1]> mean_16_axes_0 = const()[name = string("mean_16_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_16_keep_dims_0 = const()[name = string("mean_16_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_16_cast_fp16 = reduce_mean(axes = mean_16_axes_0, keep_dims = mean_16_keep_dims_0, x = pow_17_cast_fp16)[name = string("mean_16_cast_fp16")];
            fp16 const_642_to_fp16 = const()[name = string("const_642_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_46_cast_fp16 = add(x = mean_16_cast_fp16, y = const_642_to_fp16)[name = string("add_46_cast_fp16")];
            fp32 rsqrt_16_epsilon_0 = const()[name = string("rsqrt_16_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_16_cast_fp16 = rsqrt(epsilon = rsqrt_16_epsilon_0, x = add_46_cast_fp16)[name = string("rsqrt_16_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_100_cast_fp16 = mul(x = add_45_cast_fp16, y = rsqrt_16_cast_fp16)[name = string("mul_100_cast_fp16")];
            tensor<fp16, [768]> add_47_to_fp16 = const()[name = string("add_47_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212191232)))];
            tensor<fp16, [1, 512, 768]> mul_101_cast_fp16 = mul(x = mul_100_cast_fp16, y = add_47_to_fp16)[name = string("mul_101_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212192832))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213077632))))[name = string("p_st_0_model_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = mul_101_cast_fp16)[name = string("linear_18_cast_fp16")];
            string gelu_2_mode_0 = const()[name = string("gelu_2_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_2_cast_fp16 = gelu(mode = gelu_2_mode_0, x = linear_18_cast_fp16)[name = string("gelu_2_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213080000))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213964800))))[name = string("p_st_0_model_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = mul_101_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_102_cast_fp16 = mul(x = gelu_2_cast_fp16, y = linear_19_cast_fp16)[name = string("mul_102_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213967168))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214851968))))[name = string("p_st_0_model_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = mul_102_cast_fp16)[name = string("linear_20_cast_fp16")];
            fp16 const_647_promoted_to_fp16 = const()[name = string("const_647_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_18_cast_fp16 = pow(x = linear_20_cast_fp16, y = const_647_promoted_to_fp16)[name = string("pow_18_cast_fp16")];
            tensor<int32, [1]> mean_17_axes_0 = const()[name = string("mean_17_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_17_keep_dims_0 = const()[name = string("mean_17_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_17_cast_fp16 = reduce_mean(axes = mean_17_axes_0, keep_dims = mean_17_keep_dims_0, x = pow_18_cast_fp16)[name = string("mean_17_cast_fp16")];
            fp16 const_650_to_fp16 = const()[name = string("const_650_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_48_cast_fp16 = add(x = mean_17_cast_fp16, y = const_650_to_fp16)[name = string("add_48_cast_fp16")];
            fp32 rsqrt_17_epsilon_0 = const()[name = string("rsqrt_17_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_17_cast_fp16 = rsqrt(epsilon = rsqrt_17_epsilon_0, x = add_48_cast_fp16)[name = string("rsqrt_17_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_103_cast_fp16 = mul(x = linear_20_cast_fp16, y = rsqrt_17_cast_fp16)[name = string("mul_103_cast_fp16")];
            tensor<fp16, [768]> add_49_to_fp16 = const()[name = string("add_49_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214853568)))];
            tensor<fp16, [1, 512, 768]> mul_104_cast_fp16 = mul(x = mul_103_cast_fp16, y = add_49_to_fp16)[name = string("mul_104_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_50_cast_fp16 = add(x = add_45_cast_fp16, y = mul_104_cast_fp16)[name = string("add_50_cast_fp16")];
            fp16 const_655_promoted_to_fp16 = const()[name = string("const_655_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_19_cast_fp16 = pow(x = add_50_cast_fp16, y = const_655_promoted_to_fp16)[name = string("pow_19_cast_fp16")];
            tensor<int32, [1]> mean_18_axes_0 = const()[name = string("mean_18_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_18_keep_dims_0 = const()[name = string("mean_18_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_18_cast_fp16 = reduce_mean(axes = mean_18_axes_0, keep_dims = mean_18_keep_dims_0, x = pow_19_cast_fp16)[name = string("mean_18_cast_fp16")];
            fp16 const_658_to_fp16 = const()[name = string("const_658_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_51_cast_fp16 = add(x = mean_18_cast_fp16, y = const_658_to_fp16)[name = string("add_51_cast_fp16")];
            fp32 rsqrt_18_epsilon_0 = const()[name = string("rsqrt_18_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_18_cast_fp16 = rsqrt(epsilon = rsqrt_18_epsilon_0, x = add_51_cast_fp16)[name = string("rsqrt_18_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_105_cast_fp16 = mul(x = add_50_cast_fp16, y = rsqrt_18_cast_fp16)[name = string("mul_105_cast_fp16")];
            tensor<fp16, [768]> add_52_to_fp16 = const()[name = string("add_52_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214855168)))];
            tensor<fp16, [1, 512, 768]> mul_106_cast_fp16 = mul(x = mul_105_cast_fp16, y = add_52_to_fp16)[name = string("mul_106_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214856768))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215446656))))[name = string("p_st_0_model_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_21_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = mul_106_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> const_662 = const()[name = string("const_662"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_18_cast_fp16 = reshape(shape = const_662, x = linear_21_cast_fp16)[name = string("view_18_cast_fp16")];
            tensor<int32, [4]> transpose_39_perm_0 = const()[name = string("transpose_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215448256))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215644928))))[name = string("p_st_0_model_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_22_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = mul_106_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> const_665 = const()[name = string("const_665"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_19_cast_fp16 = reshape(shape = const_665, x = linear_22_cast_fp16)[name = string("view_19_cast_fp16")];
            tensor<int32, [4]> transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215645504))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215842176))))[name = string("p_st_0_model_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = mul_106_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> const_668 = const()[name = string("const_668"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_20_cast_fp16 = reshape(shape = const_668, x = linear_23_cast_fp16)[name = string("view_20_cast_fp16")];
            tensor<int32, [4]> transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_672_promoted_to_fp16 = const()[name = string("const_672_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_39_cast_fp16 = transpose(perm = transpose_39_perm_0, x = view_18_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 3, 512, 256]> pow_20_cast_fp16 = pow(x = transpose_39_cast_fp16, y = const_672_promoted_to_fp16)[name = string("pow_20_cast_fp16")];
            tensor<int32, [1]> mean_19_axes_0 = const()[name = string("mean_19_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_19_keep_dims_0 = const()[name = string("mean_19_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_19_cast_fp16 = reduce_mean(axes = mean_19_axes_0, keep_dims = mean_19_keep_dims_0, x = pow_20_cast_fp16)[name = string("mean_19_cast_fp16")];
            fp16 const_675_to_fp16 = const()[name = string("const_675_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_53_cast_fp16 = add(x = mean_19_cast_fp16, y = const_675_to_fp16)[name = string("add_53_cast_fp16")];
            fp32 rsqrt_19_epsilon_0 = const()[name = string("rsqrt_19_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_19_cast_fp16 = rsqrt(epsilon = rsqrt_19_epsilon_0, x = add_53_cast_fp16)[name = string("rsqrt_19_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_107_cast_fp16 = mul(x = transpose_39_cast_fp16, y = rsqrt_19_cast_fp16)[name = string("mul_107_cast_fp16")];
            tensor<fp16, [256]> add_54_to_fp16 = const()[name = string("add_54_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215842752)))];
            tensor<fp16, [1, 3, 512, 256]> mul_108_cast_fp16 = mul(x = mul_107_cast_fp16, y = add_54_to_fp16)[name = string("mul_108_cast_fp16")];
            fp16 const_680_promoted_to_fp16 = const()[name = string("const_680_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_40_cast_fp16 = transpose(perm = transpose_40_perm_0, x = view_19_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 1, 512, 256]> pow_21_cast_fp16 = pow(x = transpose_40_cast_fp16, y = const_680_promoted_to_fp16)[name = string("pow_21_cast_fp16")];
            tensor<int32, [1]> mean_20_axes_0 = const()[name = string("mean_20_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_20_keep_dims_0 = const()[name = string("mean_20_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_20_cast_fp16 = reduce_mean(axes = mean_20_axes_0, keep_dims = mean_20_keep_dims_0, x = pow_21_cast_fp16)[name = string("mean_20_cast_fp16")];
            fp16 const_683_to_fp16 = const()[name = string("const_683_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_55_cast_fp16 = add(x = mean_20_cast_fp16, y = const_683_to_fp16)[name = string("add_55_cast_fp16")];
            fp32 rsqrt_20_epsilon_0 = const()[name = string("rsqrt_20_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_20_cast_fp16 = rsqrt(epsilon = rsqrt_20_epsilon_0, x = add_55_cast_fp16)[name = string("rsqrt_20_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_109_cast_fp16 = mul(x = transpose_40_cast_fp16, y = rsqrt_20_cast_fp16)[name = string("mul_109_cast_fp16")];
            tensor<fp16, [256]> add_56_to_fp16 = const()[name = string("add_56_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215843328)))];
            tensor<fp16, [1, 1, 512, 256]> mul_110_cast_fp16 = mul(x = mul_109_cast_fp16, y = add_56_to_fp16)[name = string("mul_110_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_111_cast_fp16 = mul(x = mul_108_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_111_cast_fp16")];
            tensor<int32, [4]> slice_146_begin_0 = const()[name = string("slice_146_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_146_end_0 = const()[name = string("slice_146_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_146_end_mask_0 = const()[name = string("slice_146_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_146_cast_fp16 = slice_by_index(begin = slice_146_begin_0, end = slice_146_end_0, end_mask = slice_146_end_mask_0, x = mul_108_cast_fp16)[name = string("slice_146_cast_fp16")];
            tensor<int32, [4]> slice_147_begin_0 = const()[name = string("slice_147_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_147_end_0 = const()[name = string("slice_147_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_147_end_mask_0 = const()[name = string("slice_147_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_147_cast_fp16 = slice_by_index(begin = slice_147_begin_0, end = slice_147_end_0, end_mask = slice_147_end_mask_0, x = mul_108_cast_fp16)[name = string("slice_147_cast_fp16")];
            fp16 const_695_promoted_to_fp16 = const()[name = string("const_695_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_6_cast_fp16 = mul(x = slice_147_cast_fp16, y = const_695_promoted_to_fp16)[name = string("neg_6_cast_fp16")];
            int32 const_696 = const()[name = string("const_696"), val = int32(-1)];
            bool cat_36_interleave_0 = const()[name = string("cat_36_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_36_cast_fp16 = concat(axis = const_696, interleave = cat_36_interleave_0, values = (neg_6_cast_fp16, slice_146_cast_fp16))[name = string("cat_36_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_112_cast_fp16 = mul(x = cat_36_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_112_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_57_cast_fp16 = add(x = mul_111_cast_fp16, y = mul_112_cast_fp16)[name = string("add_57_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_113_cast_fp16 = mul(x = mul_110_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_113_cast_fp16")];
            tensor<int32, [4]> slice_148_begin_0 = const()[name = string("slice_148_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_148_end_0 = const()[name = string("slice_148_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_148_end_mask_0 = const()[name = string("slice_148_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_148_cast_fp16 = slice_by_index(begin = slice_148_begin_0, end = slice_148_end_0, end_mask = slice_148_end_mask_0, x = mul_110_cast_fp16)[name = string("slice_148_cast_fp16")];
            tensor<int32, [4]> slice_149_begin_0 = const()[name = string("slice_149_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_149_end_0 = const()[name = string("slice_149_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_149_end_mask_0 = const()[name = string("slice_149_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_149_cast_fp16 = slice_by_index(begin = slice_149_begin_0, end = slice_149_end_0, end_mask = slice_149_end_mask_0, x = mul_110_cast_fp16)[name = string("slice_149_cast_fp16")];
            fp16 const_703_promoted_to_fp16 = const()[name = string("const_703_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_7_cast_fp16 = mul(x = slice_149_cast_fp16, y = const_703_promoted_to_fp16)[name = string("neg_7_cast_fp16")];
            int32 const_704 = const()[name = string("const_704"), val = int32(-1)];
            bool cat_37_interleave_0 = const()[name = string("cat_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_37_cast_fp16 = concat(axis = const_704, interleave = cat_37_interleave_0, values = (neg_7_cast_fp16, slice_148_cast_fp16))[name = string("cat_37_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_114_cast_fp16 = mul(x = cat_37_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_114_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_58_cast_fp16 = add(x = mul_113_cast_fp16, y = mul_114_cast_fp16)[name = string("add_58_cast_fp16")];
            int32 const_705 = const()[name = string("const_705"), val = int32(-2)];
            bool cat_38_interleave_0 = const()[name = string("cat_38_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_38_cast_fp16 = concat(axis = const_705, interleave = cat_38_interleave_0, values = add_58_cast_fp16)[name = string("cat_38_cast_fp16")];
            int32 const_706 = const()[name = string("const_706"), val = int32(-2)];
            bool cat_39_interleave_0 = const()[name = string("cat_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_41_cast_fp16 = transpose(perm = transpose_41_perm_0, x = view_20_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 1, 512, 256]> cat_39_cast_fp16 = concat(axis = const_706, interleave = cat_39_interleave_0, values = transpose_41_cast_fp16)[name = string("cat_39_cast_fp16")];
            tensor<int32, [1]> unsqueeze_91_axes_0 = const()[name = string("unsqueeze_91_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_91_cast_fp16 = expand_dims(axes = unsqueeze_91_axes_0, x = cat_38_cast_fp16)[name = string("unsqueeze_91_cast_fp16")];
            tensor<int32, [5]> expand_32_reps_0 = const()[name = string("expand_32_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_32_cast_fp16 = tile(reps = expand_32_reps_0, x = unsqueeze_91_cast_fp16)[name = string("expand_32_cast_fp16")];
            tensor<int32, [4]> const_721 = const()[name = string("const_721"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_21_cast_fp16 = reshape(shape = const_721, x = expand_32_cast_fp16)[name = string("view_21_cast_fp16")];
            tensor<int32, [1]> unsqueeze_92_axes_0 = const()[name = string("unsqueeze_92_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_92_cast_fp16 = expand_dims(axes = unsqueeze_92_axes_0, x = cat_39_cast_fp16)[name = string("unsqueeze_92_cast_fp16")];
            tensor<int32, [5]> expand_33_reps_0 = const()[name = string("expand_33_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_33_cast_fp16 = tile(reps = expand_33_reps_0, x = unsqueeze_92_cast_fp16)[name = string("expand_33_cast_fp16")];
            tensor<int32, [4]> const_736 = const()[name = string("const_736"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_22_cast_fp16 = reshape(shape = const_736, x = expand_33_cast_fp16)[name = string("view_22_cast_fp16")];
            bool matmul_30_transpose_x_1 = const()[name = string("matmul_30_transpose_x_1"), val = bool(false)];
            bool matmul_30_transpose_y_1 = const()[name = string("matmul_30_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_30_cast_fp16 = matmul(transpose_x = matmul_30_transpose_x_1, transpose_y = matmul_30_transpose_y_1, x = add_57_cast_fp16, y = view_21_cast_fp16)[name = string("matmul_30_cast_fp16")];
            fp16 const_739_to_fp16 = const()[name = string("const_739_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_115_cast_fp16 = mul(x = matmul_30_cast_fp16, y = const_739_to_fp16)[name = string("mul_115_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_59_cast_fp16 = add(x = mul_115_cast_fp16, y = expand_cast_fp16)[name = string("add_59_cast_fp16")];
            int32 const_749 = const()[name = string("const_749"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_3_cast_fp16 = softmax(axis = const_749, x = add_59_cast_fp16)[name = string("softmax_3_cast_fp16")];
            bool matmul_31_transpose_x_0 = const()[name = string("matmul_31_transpose_x_0"), val = bool(false)];
            bool matmul_31_transpose_y_0 = const()[name = string("matmul_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_31_cast_fp16 = matmul(transpose_x = matmul_31_transpose_x_0, transpose_y = matmul_31_transpose_y_0, x = softmax_3_cast_fp16, y = view_22_cast_fp16)[name = string("matmul_31_cast_fp16")];
            tensor<int32, [4]> transpose_43_perm_0 = const()[name = string("transpose_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_754 = const()[name = string("const_754"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_43_cast_fp16 = transpose(perm = transpose_43_perm_0, x = matmul_31_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 512, 768]> view_23_cast_fp16 = reshape(shape = const_754, x = transpose_43_cast_fp16)[name = string("view_23_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215843904))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216433792))))[name = string("p_st_0_model_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = view_23_cast_fp16)[name = string("linear_24_cast_fp16")];
            fp16 const_756_promoted_to_fp16 = const()[name = string("const_756_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_22_cast_fp16 = pow(x = linear_24_cast_fp16, y = const_756_promoted_to_fp16)[name = string("pow_22_cast_fp16")];
            tensor<int32, [1]> mean_21_axes_0 = const()[name = string("mean_21_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_21_keep_dims_0 = const()[name = string("mean_21_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_21_cast_fp16 = reduce_mean(axes = mean_21_axes_0, keep_dims = mean_21_keep_dims_0, x = pow_22_cast_fp16)[name = string("mean_21_cast_fp16")];
            fp16 const_759_to_fp16 = const()[name = string("const_759_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_60_cast_fp16 = add(x = mean_21_cast_fp16, y = const_759_to_fp16)[name = string("add_60_cast_fp16")];
            fp32 rsqrt_21_epsilon_0 = const()[name = string("rsqrt_21_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_21_cast_fp16 = rsqrt(epsilon = rsqrt_21_epsilon_0, x = add_60_cast_fp16)[name = string("rsqrt_21_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_116_cast_fp16 = mul(x = linear_24_cast_fp16, y = rsqrt_21_cast_fp16)[name = string("mul_116_cast_fp16")];
            tensor<fp16, [768]> add_61_to_fp16 = const()[name = string("add_61_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216435392)))];
            tensor<fp16, [1, 512, 768]> mul_117_cast_fp16 = mul(x = mul_116_cast_fp16, y = add_61_to_fp16)[name = string("mul_117_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_62_cast_fp16 = add(x = add_50_cast_fp16, y = mul_117_cast_fp16)[name = string("add_62_cast_fp16")];
            fp16 const_764_promoted_to_fp16 = const()[name = string("const_764_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_23_cast_fp16 = pow(x = add_62_cast_fp16, y = const_764_promoted_to_fp16)[name = string("pow_23_cast_fp16")];
            tensor<int32, [1]> mean_22_axes_0 = const()[name = string("mean_22_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_22_keep_dims_0 = const()[name = string("mean_22_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_22_cast_fp16 = reduce_mean(axes = mean_22_axes_0, keep_dims = mean_22_keep_dims_0, x = pow_23_cast_fp16)[name = string("mean_22_cast_fp16")];
            fp16 const_767_to_fp16 = const()[name = string("const_767_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_63_cast_fp16 = add(x = mean_22_cast_fp16, y = const_767_to_fp16)[name = string("add_63_cast_fp16")];
            fp32 rsqrt_22_epsilon_0 = const()[name = string("rsqrt_22_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_22_cast_fp16 = rsqrt(epsilon = rsqrt_22_epsilon_0, x = add_63_cast_fp16)[name = string("rsqrt_22_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_118_cast_fp16 = mul(x = add_62_cast_fp16, y = rsqrt_22_cast_fp16)[name = string("mul_118_cast_fp16")];
            tensor<fp16, [768]> add_64_to_fp16 = const()[name = string("add_64_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216436992)))];
            tensor<fp16, [1, 512, 768]> mul_119_cast_fp16 = mul(x = mul_118_cast_fp16, y = add_64_to_fp16)[name = string("mul_119_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216438592))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217323392))))[name = string("p_st_0_model_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = mul_119_cast_fp16)[name = string("linear_25_cast_fp16")];
            string gelu_3_mode_0 = const()[name = string("gelu_3_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_3_cast_fp16 = gelu(mode = gelu_3_mode_0, x = linear_25_cast_fp16)[name = string("gelu_3_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217325760))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218210560))))[name = string("p_st_0_model_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = mul_119_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_120_cast_fp16 = mul(x = gelu_3_cast_fp16, y = linear_26_cast_fp16)[name = string("mul_120_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218212928))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219097728))))[name = string("p_st_0_model_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_27_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = mul_120_cast_fp16)[name = string("linear_27_cast_fp16")];
            fp16 const_772_promoted_to_fp16 = const()[name = string("const_772_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_24_cast_fp16 = pow(x = linear_27_cast_fp16, y = const_772_promoted_to_fp16)[name = string("pow_24_cast_fp16")];
            tensor<int32, [1]> mean_23_axes_0 = const()[name = string("mean_23_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_23_keep_dims_0 = const()[name = string("mean_23_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_23_cast_fp16 = reduce_mean(axes = mean_23_axes_0, keep_dims = mean_23_keep_dims_0, x = pow_24_cast_fp16)[name = string("mean_23_cast_fp16")];
            fp16 const_775_to_fp16 = const()[name = string("const_775_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_65_cast_fp16 = add(x = mean_23_cast_fp16, y = const_775_to_fp16)[name = string("add_65_cast_fp16")];
            fp32 rsqrt_23_epsilon_0 = const()[name = string("rsqrt_23_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_23_cast_fp16 = rsqrt(epsilon = rsqrt_23_epsilon_0, x = add_65_cast_fp16)[name = string("rsqrt_23_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_121_cast_fp16 = mul(x = linear_27_cast_fp16, y = rsqrt_23_cast_fp16)[name = string("mul_121_cast_fp16")];
            tensor<fp16, [768]> add_66_to_fp16 = const()[name = string("add_66_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219099328)))];
            tensor<fp16, [1, 512, 768]> mul_122_cast_fp16 = mul(x = mul_121_cast_fp16, y = add_66_to_fp16)[name = string("mul_122_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_67_cast_fp16 = add(x = add_62_cast_fp16, y = mul_122_cast_fp16)[name = string("add_67_cast_fp16")];
            fp16 const_780_promoted_to_fp16 = const()[name = string("const_780_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_25_cast_fp16 = pow(x = add_67_cast_fp16, y = const_780_promoted_to_fp16)[name = string("pow_25_cast_fp16")];
            tensor<int32, [1]> mean_24_axes_0 = const()[name = string("mean_24_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_24_keep_dims_0 = const()[name = string("mean_24_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_24_cast_fp16 = reduce_mean(axes = mean_24_axes_0, keep_dims = mean_24_keep_dims_0, x = pow_25_cast_fp16)[name = string("mean_24_cast_fp16")];
            fp16 const_783_to_fp16 = const()[name = string("const_783_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_68_cast_fp16 = add(x = mean_24_cast_fp16, y = const_783_to_fp16)[name = string("add_68_cast_fp16")];
            fp32 rsqrt_24_epsilon_0 = const()[name = string("rsqrt_24_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_24_cast_fp16 = rsqrt(epsilon = rsqrt_24_epsilon_0, x = add_68_cast_fp16)[name = string("rsqrt_24_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_123_cast_fp16 = mul(x = add_67_cast_fp16, y = rsqrt_24_cast_fp16)[name = string("mul_123_cast_fp16")];
            tensor<fp16, [768]> add_69_to_fp16 = const()[name = string("add_69_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219100928)))];
            tensor<fp16, [1, 512, 768]> mul_124_cast_fp16 = mul(x = mul_123_cast_fp16, y = add_69_to_fp16)[name = string("mul_124_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219102528))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219692416))))[name = string("p_st_0_model_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = mul_124_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> const_787 = const()[name = string("const_787"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_24_cast_fp16 = reshape(shape = const_787, x = linear_28_cast_fp16)[name = string("view_24_cast_fp16")];
            tensor<int32, [4]> transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219694016))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219890688))))[name = string("p_st_0_model_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = mul_124_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> const_790 = const()[name = string("const_790"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_25_cast_fp16 = reshape(shape = const_790, x = linear_29_cast_fp16)[name = string("view_25_cast_fp16")];
            tensor<int32, [4]> transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219891264))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220087936))))[name = string("p_st_0_model_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_30_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = mul_124_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> const_793 = const()[name = string("const_793"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_26_cast_fp16 = reshape(shape = const_793, x = linear_30_cast_fp16)[name = string("view_26_cast_fp16")];
            tensor<int32, [4]> transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_797_promoted_to_fp16 = const()[name = string("const_797_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_44_cast_fp16 = transpose(perm = transpose_44_perm_0, x = view_24_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 3, 512, 256]> pow_26_cast_fp16 = pow(x = transpose_44_cast_fp16, y = const_797_promoted_to_fp16)[name = string("pow_26_cast_fp16")];
            tensor<int32, [1]> mean_25_axes_0 = const()[name = string("mean_25_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_25_keep_dims_0 = const()[name = string("mean_25_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_25_cast_fp16 = reduce_mean(axes = mean_25_axes_0, keep_dims = mean_25_keep_dims_0, x = pow_26_cast_fp16)[name = string("mean_25_cast_fp16")];
            fp16 const_800_to_fp16 = const()[name = string("const_800_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_70_cast_fp16 = add(x = mean_25_cast_fp16, y = const_800_to_fp16)[name = string("add_70_cast_fp16")];
            fp32 rsqrt_25_epsilon_0 = const()[name = string("rsqrt_25_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_25_cast_fp16 = rsqrt(epsilon = rsqrt_25_epsilon_0, x = add_70_cast_fp16)[name = string("rsqrt_25_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_125_cast_fp16 = mul(x = transpose_44_cast_fp16, y = rsqrt_25_cast_fp16)[name = string("mul_125_cast_fp16")];
            tensor<fp16, [256]> add_71_to_fp16 = const()[name = string("add_71_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220088512)))];
            tensor<fp16, [1, 3, 512, 256]> mul_126_cast_fp16 = mul(x = mul_125_cast_fp16, y = add_71_to_fp16)[name = string("mul_126_cast_fp16")];
            fp16 const_805_promoted_to_fp16 = const()[name = string("const_805_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_45_cast_fp16 = transpose(perm = transpose_45_perm_0, x = view_25_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 1, 512, 256]> pow_27_cast_fp16 = pow(x = transpose_45_cast_fp16, y = const_805_promoted_to_fp16)[name = string("pow_27_cast_fp16")];
            tensor<int32, [1]> mean_26_axes_0 = const()[name = string("mean_26_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_26_keep_dims_0 = const()[name = string("mean_26_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_26_cast_fp16 = reduce_mean(axes = mean_26_axes_0, keep_dims = mean_26_keep_dims_0, x = pow_27_cast_fp16)[name = string("mean_26_cast_fp16")];
            fp16 const_808_to_fp16 = const()[name = string("const_808_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_72_cast_fp16 = add(x = mean_26_cast_fp16, y = const_808_to_fp16)[name = string("add_72_cast_fp16")];
            fp32 rsqrt_26_epsilon_0 = const()[name = string("rsqrt_26_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_26_cast_fp16 = rsqrt(epsilon = rsqrt_26_epsilon_0, x = add_72_cast_fp16)[name = string("rsqrt_26_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_127_cast_fp16 = mul(x = transpose_45_cast_fp16, y = rsqrt_26_cast_fp16)[name = string("mul_127_cast_fp16")];
            tensor<fp16, [256]> add_73_to_fp16 = const()[name = string("add_73_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220089088)))];
            tensor<fp16, [1, 1, 512, 256]> mul_128_cast_fp16 = mul(x = mul_127_cast_fp16, y = add_73_to_fp16)[name = string("mul_128_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_129_cast_fp16 = mul(x = mul_126_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_129_cast_fp16")];
            tensor<int32, [4]> slice_169_begin_0 = const()[name = string("slice_169_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_169_end_0 = const()[name = string("slice_169_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_169_end_mask_0 = const()[name = string("slice_169_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_169_cast_fp16 = slice_by_index(begin = slice_169_begin_0, end = slice_169_end_0, end_mask = slice_169_end_mask_0, x = mul_126_cast_fp16)[name = string("slice_169_cast_fp16")];
            tensor<int32, [4]> slice_170_begin_0 = const()[name = string("slice_170_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_170_end_0 = const()[name = string("slice_170_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_170_end_mask_0 = const()[name = string("slice_170_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_170_cast_fp16 = slice_by_index(begin = slice_170_begin_0, end = slice_170_end_0, end_mask = slice_170_end_mask_0, x = mul_126_cast_fp16)[name = string("slice_170_cast_fp16")];
            fp16 const_820_promoted_to_fp16 = const()[name = string("const_820_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_8_cast_fp16 = mul(x = slice_170_cast_fp16, y = const_820_promoted_to_fp16)[name = string("neg_8_cast_fp16")];
            int32 const_821 = const()[name = string("const_821"), val = int32(-1)];
            bool cat_40_interleave_0 = const()[name = string("cat_40_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_40_cast_fp16 = concat(axis = const_821, interleave = cat_40_interleave_0, values = (neg_8_cast_fp16, slice_169_cast_fp16))[name = string("cat_40_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_130_cast_fp16 = mul(x = cat_40_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_130_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_74_cast_fp16 = add(x = mul_129_cast_fp16, y = mul_130_cast_fp16)[name = string("add_74_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_131_cast_fp16 = mul(x = mul_128_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_131_cast_fp16")];
            tensor<int32, [4]> slice_171_begin_0 = const()[name = string("slice_171_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_171_end_0 = const()[name = string("slice_171_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_171_end_mask_0 = const()[name = string("slice_171_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_171_cast_fp16 = slice_by_index(begin = slice_171_begin_0, end = slice_171_end_0, end_mask = slice_171_end_mask_0, x = mul_128_cast_fp16)[name = string("slice_171_cast_fp16")];
            tensor<int32, [4]> slice_172_begin_0 = const()[name = string("slice_172_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_172_end_0 = const()[name = string("slice_172_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_172_end_mask_0 = const()[name = string("slice_172_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_172_cast_fp16 = slice_by_index(begin = slice_172_begin_0, end = slice_172_end_0, end_mask = slice_172_end_mask_0, x = mul_128_cast_fp16)[name = string("slice_172_cast_fp16")];
            fp16 const_828_promoted_to_fp16 = const()[name = string("const_828_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_9_cast_fp16 = mul(x = slice_172_cast_fp16, y = const_828_promoted_to_fp16)[name = string("neg_9_cast_fp16")];
            int32 const_829 = const()[name = string("const_829"), val = int32(-1)];
            bool cat_41_interleave_0 = const()[name = string("cat_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_41_cast_fp16 = concat(axis = const_829, interleave = cat_41_interleave_0, values = (neg_9_cast_fp16, slice_171_cast_fp16))[name = string("cat_41_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_132_cast_fp16 = mul(x = cat_41_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_132_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_75_cast_fp16 = add(x = mul_131_cast_fp16, y = mul_132_cast_fp16)[name = string("add_75_cast_fp16")];
            int32 const_830 = const()[name = string("const_830"), val = int32(-2)];
            bool cat_42_interleave_0 = const()[name = string("cat_42_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_42_cast_fp16 = concat(axis = const_830, interleave = cat_42_interleave_0, values = add_75_cast_fp16)[name = string("cat_42_cast_fp16")];
            int32 const_831 = const()[name = string("const_831"), val = int32(-2)];
            bool cat_43_interleave_0 = const()[name = string("cat_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_46_cast_fp16 = transpose(perm = transpose_46_perm_0, x = view_26_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 1, 512, 256]> cat_43_cast_fp16 = concat(axis = const_831, interleave = cat_43_interleave_0, values = transpose_46_cast_fp16)[name = string("cat_43_cast_fp16")];
            tensor<int32, [1]> unsqueeze_95_axes_0 = const()[name = string("unsqueeze_95_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_95_cast_fp16 = expand_dims(axes = unsqueeze_95_axes_0, x = cat_42_cast_fp16)[name = string("unsqueeze_95_cast_fp16")];
            tensor<int32, [5]> expand_34_reps_0 = const()[name = string("expand_34_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_34_cast_fp16 = tile(reps = expand_34_reps_0, x = unsqueeze_95_cast_fp16)[name = string("expand_34_cast_fp16")];
            tensor<int32, [4]> const_846 = const()[name = string("const_846"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_27_cast_fp16 = reshape(shape = const_846, x = expand_34_cast_fp16)[name = string("view_27_cast_fp16")];
            tensor<int32, [1]> unsqueeze_96_axes_0 = const()[name = string("unsqueeze_96_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_96_cast_fp16 = expand_dims(axes = unsqueeze_96_axes_0, x = cat_43_cast_fp16)[name = string("unsqueeze_96_cast_fp16")];
            tensor<int32, [5]> expand_35_reps_0 = const()[name = string("expand_35_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_35_cast_fp16 = tile(reps = expand_35_reps_0, x = unsqueeze_96_cast_fp16)[name = string("expand_35_cast_fp16")];
            tensor<int32, [4]> const_861 = const()[name = string("const_861"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_28_cast_fp16 = reshape(shape = const_861, x = expand_35_cast_fp16)[name = string("view_28_cast_fp16")];
            bool matmul_32_transpose_x_1 = const()[name = string("matmul_32_transpose_x_1"), val = bool(false)];
            bool matmul_32_transpose_y_1 = const()[name = string("matmul_32_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_32_cast_fp16 = matmul(transpose_x = matmul_32_transpose_x_1, transpose_y = matmul_32_transpose_y_1, x = add_74_cast_fp16, y = view_27_cast_fp16)[name = string("matmul_32_cast_fp16")];
            fp16 const_864_to_fp16 = const()[name = string("const_864_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_133_cast_fp16 = mul(x = matmul_32_cast_fp16, y = const_864_to_fp16)[name = string("mul_133_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_76_cast_fp16 = add(x = mul_133_cast_fp16, y = expand_cast_fp16)[name = string("add_76_cast_fp16")];
            int32 const_874 = const()[name = string("const_874"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_4_cast_fp16 = softmax(axis = const_874, x = add_76_cast_fp16)[name = string("softmax_4_cast_fp16")];
            bool matmul_33_transpose_x_0 = const()[name = string("matmul_33_transpose_x_0"), val = bool(false)];
            bool matmul_33_transpose_y_0 = const()[name = string("matmul_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_33_cast_fp16 = matmul(transpose_x = matmul_33_transpose_x_0, transpose_y = matmul_33_transpose_y_0, x = softmax_4_cast_fp16, y = view_28_cast_fp16)[name = string("matmul_33_cast_fp16")];
            tensor<int32, [4]> transpose_48_perm_0 = const()[name = string("transpose_48_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_879 = const()[name = string("const_879"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_48_cast_fp16 = transpose(perm = transpose_48_perm_0, x = matmul_33_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 512, 768]> view_29_cast_fp16 = reshape(shape = const_879, x = transpose_48_cast_fp16)[name = string("view_29_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220089664))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220679552))))[name = string("p_st_0_model_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_31_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = view_29_cast_fp16)[name = string("linear_31_cast_fp16")];
            fp16 const_881_promoted_to_fp16 = const()[name = string("const_881_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_28_cast_fp16 = pow(x = linear_31_cast_fp16, y = const_881_promoted_to_fp16)[name = string("pow_28_cast_fp16")];
            tensor<int32, [1]> mean_27_axes_0 = const()[name = string("mean_27_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_27_keep_dims_0 = const()[name = string("mean_27_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_27_cast_fp16 = reduce_mean(axes = mean_27_axes_0, keep_dims = mean_27_keep_dims_0, x = pow_28_cast_fp16)[name = string("mean_27_cast_fp16")];
            fp16 const_884_to_fp16 = const()[name = string("const_884_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_77_cast_fp16 = add(x = mean_27_cast_fp16, y = const_884_to_fp16)[name = string("add_77_cast_fp16")];
            fp32 rsqrt_27_epsilon_0 = const()[name = string("rsqrt_27_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_27_cast_fp16 = rsqrt(epsilon = rsqrt_27_epsilon_0, x = add_77_cast_fp16)[name = string("rsqrt_27_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_134_cast_fp16 = mul(x = linear_31_cast_fp16, y = rsqrt_27_cast_fp16)[name = string("mul_134_cast_fp16")];
            tensor<fp16, [768]> add_78_to_fp16 = const()[name = string("add_78_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220681152)))];
            tensor<fp16, [1, 512, 768]> mul_135_cast_fp16 = mul(x = mul_134_cast_fp16, y = add_78_to_fp16)[name = string("mul_135_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_79_cast_fp16 = add(x = add_67_cast_fp16, y = mul_135_cast_fp16)[name = string("add_79_cast_fp16")];
            fp16 const_889_promoted_to_fp16 = const()[name = string("const_889_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_29_cast_fp16 = pow(x = add_79_cast_fp16, y = const_889_promoted_to_fp16)[name = string("pow_29_cast_fp16")];
            tensor<int32, [1]> mean_28_axes_0 = const()[name = string("mean_28_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_28_keep_dims_0 = const()[name = string("mean_28_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_28_cast_fp16 = reduce_mean(axes = mean_28_axes_0, keep_dims = mean_28_keep_dims_0, x = pow_29_cast_fp16)[name = string("mean_28_cast_fp16")];
            fp16 const_892_to_fp16 = const()[name = string("const_892_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_80_cast_fp16 = add(x = mean_28_cast_fp16, y = const_892_to_fp16)[name = string("add_80_cast_fp16")];
            fp32 rsqrt_28_epsilon_0 = const()[name = string("rsqrt_28_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_28_cast_fp16 = rsqrt(epsilon = rsqrt_28_epsilon_0, x = add_80_cast_fp16)[name = string("rsqrt_28_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_136_cast_fp16 = mul(x = add_79_cast_fp16, y = rsqrt_28_cast_fp16)[name = string("mul_136_cast_fp16")];
            tensor<fp16, [768]> add_81_to_fp16 = const()[name = string("add_81_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220682752)))];
            tensor<fp16, [1, 512, 768]> mul_137_cast_fp16 = mul(x = mul_136_cast_fp16, y = add_81_to_fp16)[name = string("mul_137_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220684352))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221569152))))[name = string("p_st_0_model_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = mul_137_cast_fp16)[name = string("linear_32_cast_fp16")];
            string gelu_4_mode_0 = const()[name = string("gelu_4_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_4_cast_fp16 = gelu(mode = gelu_4_mode_0, x = linear_32_cast_fp16)[name = string("gelu_4_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221571520))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222456320))))[name = string("p_st_0_model_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = mul_137_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_138_cast_fp16 = mul(x = gelu_4_cast_fp16, y = linear_33_cast_fp16)[name = string("mul_138_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222458688))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223343488))))[name = string("p_st_0_model_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = mul_138_cast_fp16)[name = string("linear_34_cast_fp16")];
            fp16 const_897_promoted_to_fp16 = const()[name = string("const_897_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_30_cast_fp16 = pow(x = linear_34_cast_fp16, y = const_897_promoted_to_fp16)[name = string("pow_30_cast_fp16")];
            tensor<int32, [1]> mean_29_axes_0 = const()[name = string("mean_29_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_29_keep_dims_0 = const()[name = string("mean_29_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_29_cast_fp16 = reduce_mean(axes = mean_29_axes_0, keep_dims = mean_29_keep_dims_0, x = pow_30_cast_fp16)[name = string("mean_29_cast_fp16")];
            fp16 const_900_to_fp16 = const()[name = string("const_900_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_82_cast_fp16 = add(x = mean_29_cast_fp16, y = const_900_to_fp16)[name = string("add_82_cast_fp16")];
            fp32 rsqrt_29_epsilon_0 = const()[name = string("rsqrt_29_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_29_cast_fp16 = rsqrt(epsilon = rsqrt_29_epsilon_0, x = add_82_cast_fp16)[name = string("rsqrt_29_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_139_cast_fp16 = mul(x = linear_34_cast_fp16, y = rsqrt_29_cast_fp16)[name = string("mul_139_cast_fp16")];
            tensor<fp16, [768]> add_83_to_fp16 = const()[name = string("add_83_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223345088)))];
            tensor<fp16, [1, 512, 768]> mul_140_cast_fp16 = mul(x = mul_139_cast_fp16, y = add_83_to_fp16)[name = string("mul_140_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_84_cast_fp16 = add(x = add_79_cast_fp16, y = mul_140_cast_fp16)[name = string("add_84_cast_fp16")];
            fp16 const_905_promoted_to_fp16 = const()[name = string("const_905_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_31_cast_fp16 = pow(x = add_84_cast_fp16, y = const_905_promoted_to_fp16)[name = string("pow_31_cast_fp16")];
            tensor<int32, [1]> mean_30_axes_0 = const()[name = string("mean_30_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_30_keep_dims_0 = const()[name = string("mean_30_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_30_cast_fp16 = reduce_mean(axes = mean_30_axes_0, keep_dims = mean_30_keep_dims_0, x = pow_31_cast_fp16)[name = string("mean_30_cast_fp16")];
            fp16 const_908_to_fp16 = const()[name = string("const_908_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_85_cast_fp16 = add(x = mean_30_cast_fp16, y = const_908_to_fp16)[name = string("add_85_cast_fp16")];
            fp32 rsqrt_30_epsilon_0 = const()[name = string("rsqrt_30_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_30_cast_fp16 = rsqrt(epsilon = rsqrt_30_epsilon_0, x = add_85_cast_fp16)[name = string("rsqrt_30_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_141_cast_fp16 = mul(x = add_84_cast_fp16, y = rsqrt_30_cast_fp16)[name = string("mul_141_cast_fp16")];
            tensor<fp16, [768]> add_86_to_fp16 = const()[name = string("add_86_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223346688)))];
            tensor<fp16, [1, 512, 768]> mul_142_cast_fp16 = mul(x = mul_141_cast_fp16, y = add_86_to_fp16)[name = string("mul_142_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223348288))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223938176))))[name = string("p_st_0_model_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_35_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = mul_142_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> const_912 = const()[name = string("const_912"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_30_cast_fp16 = reshape(shape = const_912, x = linear_35_cast_fp16)[name = string("view_30_cast_fp16")];
            tensor<int32, [4]> transpose_49_perm_0 = const()[name = string("transpose_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223939776))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224136448))))[name = string("p_st_0_model_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_36_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = mul_142_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> const_915 = const()[name = string("const_915"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_31_cast_fp16 = reshape(shape = const_915, x = linear_36_cast_fp16)[name = string("view_31_cast_fp16")];
            tensor<int32, [4]> transpose_50_perm_0 = const()[name = string("transpose_50_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224137024))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224333696))))[name = string("p_st_0_model_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = mul_142_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> const_918 = const()[name = string("const_918"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_32_cast_fp16 = reshape(shape = const_918, x = linear_37_cast_fp16)[name = string("view_32_cast_fp16")];
            tensor<int32, [4]> transpose_51_perm_0 = const()[name = string("transpose_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_922_promoted_to_fp16 = const()[name = string("const_922_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_49_cast_fp16 = transpose(perm = transpose_49_perm_0, x = view_30_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 3, 512, 256]> pow_32_cast_fp16 = pow(x = transpose_49_cast_fp16, y = const_922_promoted_to_fp16)[name = string("pow_32_cast_fp16")];
            tensor<int32, [1]> mean_31_axes_0 = const()[name = string("mean_31_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_31_keep_dims_0 = const()[name = string("mean_31_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_31_cast_fp16 = reduce_mean(axes = mean_31_axes_0, keep_dims = mean_31_keep_dims_0, x = pow_32_cast_fp16)[name = string("mean_31_cast_fp16")];
            fp16 const_925_to_fp16 = const()[name = string("const_925_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_87_cast_fp16 = add(x = mean_31_cast_fp16, y = const_925_to_fp16)[name = string("add_87_cast_fp16")];
            fp32 rsqrt_31_epsilon_0 = const()[name = string("rsqrt_31_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_31_cast_fp16 = rsqrt(epsilon = rsqrt_31_epsilon_0, x = add_87_cast_fp16)[name = string("rsqrt_31_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_143_cast_fp16 = mul(x = transpose_49_cast_fp16, y = rsqrt_31_cast_fp16)[name = string("mul_143_cast_fp16")];
            tensor<fp16, [256]> add_88_to_fp16 = const()[name = string("add_88_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224334272)))];
            tensor<fp16, [1, 3, 512, 256]> mul_144_cast_fp16 = mul(x = mul_143_cast_fp16, y = add_88_to_fp16)[name = string("mul_144_cast_fp16")];
            fp16 const_930_promoted_to_fp16 = const()[name = string("const_930_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_50_cast_fp16 = transpose(perm = transpose_50_perm_0, x = view_31_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 1, 512, 256]> pow_33_cast_fp16 = pow(x = transpose_50_cast_fp16, y = const_930_promoted_to_fp16)[name = string("pow_33_cast_fp16")];
            tensor<int32, [1]> mean_32_axes_0 = const()[name = string("mean_32_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_32_keep_dims_0 = const()[name = string("mean_32_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_32_cast_fp16 = reduce_mean(axes = mean_32_axes_0, keep_dims = mean_32_keep_dims_0, x = pow_33_cast_fp16)[name = string("mean_32_cast_fp16")];
            fp16 const_933_to_fp16 = const()[name = string("const_933_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_89_cast_fp16 = add(x = mean_32_cast_fp16, y = const_933_to_fp16)[name = string("add_89_cast_fp16")];
            fp32 rsqrt_32_epsilon_0 = const()[name = string("rsqrt_32_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_32_cast_fp16 = rsqrt(epsilon = rsqrt_32_epsilon_0, x = add_89_cast_fp16)[name = string("rsqrt_32_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_145_cast_fp16 = mul(x = transpose_50_cast_fp16, y = rsqrt_32_cast_fp16)[name = string("mul_145_cast_fp16")];
            tensor<fp16, [256]> add_90_to_fp16 = const()[name = string("add_90_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224334848)))];
            tensor<fp16, [1, 1, 512, 256]> mul_146_cast_fp16 = mul(x = mul_145_cast_fp16, y = add_90_to_fp16)[name = string("mul_146_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> unsqueeze_97_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 1, 512, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224335424))), scale = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.02p-7]]]]))[name = string("unsqueeze_97_to_fp16_quantized")];
            tensor<fp16, [1, 3, 512, 256]> mul_147_cast_fp16 = mul(x = mul_144_cast_fp16, y = unsqueeze_97_to_fp16_quantized)[name = string("mul_147_cast_fp16")];
            tensor<int32, [4]> slice_192_begin_0 = const()[name = string("slice_192_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_192_end_0 = const()[name = string("slice_192_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_192_end_mask_0 = const()[name = string("slice_192_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_192_cast_fp16 = slice_by_index(begin = slice_192_begin_0, end = slice_192_end_0, end_mask = slice_192_end_mask_0, x = mul_144_cast_fp16)[name = string("slice_192_cast_fp16")];
            tensor<int32, [4]> slice_193_begin_0 = const()[name = string("slice_193_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_193_end_0 = const()[name = string("slice_193_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_193_end_mask_0 = const()[name = string("slice_193_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_193_cast_fp16 = slice_by_index(begin = slice_193_begin_0, end = slice_193_end_0, end_mask = slice_193_end_mask_0, x = mul_144_cast_fp16)[name = string("slice_193_cast_fp16")];
            fp16 const_945_promoted_to_fp16 = const()[name = string("const_945_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_10_cast_fp16 = mul(x = slice_193_cast_fp16, y = const_945_promoted_to_fp16)[name = string("neg_10_cast_fp16")];
            int32 const_946 = const()[name = string("const_946"), val = int32(-1)];
            bool cat_44_interleave_0 = const()[name = string("cat_44_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_44_cast_fp16 = concat(axis = const_946, interleave = cat_44_interleave_0, values = (neg_10_cast_fp16, slice_192_cast_fp16))[name = string("cat_44_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> unsqueeze_98_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 1, 512, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224466560))), scale = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.02p-7]]]]))[name = string("unsqueeze_98_to_fp16_quantized")];
            tensor<fp16, [1, 3, 512, 256]> mul_148_cast_fp16 = mul(x = cat_44_cast_fp16, y = unsqueeze_98_to_fp16_quantized)[name = string("mul_148_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_91_cast_fp16 = add(x = mul_147_cast_fp16, y = mul_148_cast_fp16)[name = string("add_91_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_149_cast_fp16 = mul(x = mul_146_cast_fp16, y = unsqueeze_97_to_fp16_quantized)[name = string("mul_149_cast_fp16")];
            tensor<int32, [4]> slice_194_begin_0 = const()[name = string("slice_194_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_194_end_0 = const()[name = string("slice_194_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_194_end_mask_0 = const()[name = string("slice_194_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_194_cast_fp16 = slice_by_index(begin = slice_194_begin_0, end = slice_194_end_0, end_mask = slice_194_end_mask_0, x = mul_146_cast_fp16)[name = string("slice_194_cast_fp16")];
            tensor<int32, [4]> slice_195_begin_0 = const()[name = string("slice_195_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_195_end_0 = const()[name = string("slice_195_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_195_end_mask_0 = const()[name = string("slice_195_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_195_cast_fp16 = slice_by_index(begin = slice_195_begin_0, end = slice_195_end_0, end_mask = slice_195_end_mask_0, x = mul_146_cast_fp16)[name = string("slice_195_cast_fp16")];
            fp16 const_953_promoted_to_fp16 = const()[name = string("const_953_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_11_cast_fp16 = mul(x = slice_195_cast_fp16, y = const_953_promoted_to_fp16)[name = string("neg_11_cast_fp16")];
            int32 const_954 = const()[name = string("const_954"), val = int32(-1)];
            bool cat_45_interleave_0 = const()[name = string("cat_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_45_cast_fp16 = concat(axis = const_954, interleave = cat_45_interleave_0, values = (neg_11_cast_fp16, slice_194_cast_fp16))[name = string("cat_45_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_150_cast_fp16 = mul(x = cat_45_cast_fp16, y = unsqueeze_98_to_fp16_quantized)[name = string("mul_150_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_92_cast_fp16 = add(x = mul_149_cast_fp16, y = mul_150_cast_fp16)[name = string("add_92_cast_fp16")];
            int32 const_955 = const()[name = string("const_955"), val = int32(-2)];
            bool cat_46_interleave_0 = const()[name = string("cat_46_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_46_cast_fp16 = concat(axis = const_955, interleave = cat_46_interleave_0, values = add_92_cast_fp16)[name = string("cat_46_cast_fp16")];
            int32 const_956 = const()[name = string("const_956"), val = int32(-2)];
            bool cat_47_interleave_0 = const()[name = string("cat_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_51_cast_fp16 = transpose(perm = transpose_51_perm_0, x = view_32_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 1, 512, 256]> cat_47_cast_fp16 = concat(axis = const_956, interleave = cat_47_interleave_0, values = transpose_51_cast_fp16)[name = string("cat_47_cast_fp16")];
            tensor<int32, [1]> unsqueeze_99_axes_0 = const()[name = string("unsqueeze_99_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_99_cast_fp16 = expand_dims(axes = unsqueeze_99_axes_0, x = cat_46_cast_fp16)[name = string("unsqueeze_99_cast_fp16")];
            tensor<int32, [5]> expand_36_reps_0 = const()[name = string("expand_36_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_36_cast_fp16 = tile(reps = expand_36_reps_0, x = unsqueeze_99_cast_fp16)[name = string("expand_36_cast_fp16")];
            tensor<int32, [4]> const_971 = const()[name = string("const_971"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_33_cast_fp16 = reshape(shape = const_971, x = expand_36_cast_fp16)[name = string("view_33_cast_fp16")];
            tensor<int32, [1]> unsqueeze_100_axes_0 = const()[name = string("unsqueeze_100_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_100_cast_fp16 = expand_dims(axes = unsqueeze_100_axes_0, x = cat_47_cast_fp16)[name = string("unsqueeze_100_cast_fp16")];
            tensor<int32, [5]> expand_37_reps_0 = const()[name = string("expand_37_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_37_cast_fp16 = tile(reps = expand_37_reps_0, x = unsqueeze_100_cast_fp16)[name = string("expand_37_cast_fp16")];
            tensor<int32, [4]> const_986 = const()[name = string("const_986"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_34_cast_fp16 = reshape(shape = const_986, x = expand_37_cast_fp16)[name = string("view_34_cast_fp16")];
            bool matmul_34_transpose_x_1 = const()[name = string("matmul_34_transpose_x_1"), val = bool(false)];
            bool matmul_34_transpose_y_1 = const()[name = string("matmul_34_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_34_cast_fp16 = matmul(transpose_x = matmul_34_transpose_x_1, transpose_y = matmul_34_transpose_y_1, x = add_91_cast_fp16, y = view_33_cast_fp16)[name = string("matmul_34_cast_fp16")];
            fp16 const_989_to_fp16 = const()[name = string("const_989_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_151_cast_fp16 = mul(x = matmul_34_cast_fp16, y = const_989_to_fp16)[name = string("mul_151_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_93_cast_fp16 = add(x = mul_151_cast_fp16, y = expand_cast_fp16)[name = string("add_93_cast_fp16")];
            int32 const_999 = const()[name = string("const_999"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_5_cast_fp16 = softmax(axis = const_999, x = add_93_cast_fp16)[name = string("softmax_5_cast_fp16")];
            bool matmul_35_transpose_x_0 = const()[name = string("matmul_35_transpose_x_0"), val = bool(false)];
            bool matmul_35_transpose_y_0 = const()[name = string("matmul_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_35_cast_fp16 = matmul(transpose_x = matmul_35_transpose_x_0, transpose_y = matmul_35_transpose_y_0, x = softmax_5_cast_fp16, y = view_34_cast_fp16)[name = string("matmul_35_cast_fp16")];
            tensor<int32, [4]> transpose_53_perm_0 = const()[name = string("transpose_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_1004 = const()[name = string("const_1004"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_53_cast_fp16 = transpose(perm = transpose_53_perm_0, x = matmul_35_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 512, 768]> view_35_cast_fp16 = reshape(shape = const_1004, x = transpose_53_cast_fp16)[name = string("view_35_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224597696))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225187584))))[name = string("p_st_0_model_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = view_35_cast_fp16)[name = string("linear_38_cast_fp16")];
            fp16 const_1006_promoted_to_fp16 = const()[name = string("const_1006_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_34_cast_fp16 = pow(x = linear_38_cast_fp16, y = const_1006_promoted_to_fp16)[name = string("pow_34_cast_fp16")];
            tensor<int32, [1]> mean_33_axes_0 = const()[name = string("mean_33_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_33_keep_dims_0 = const()[name = string("mean_33_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_33_cast_fp16 = reduce_mean(axes = mean_33_axes_0, keep_dims = mean_33_keep_dims_0, x = pow_34_cast_fp16)[name = string("mean_33_cast_fp16")];
            fp16 const_1009_to_fp16 = const()[name = string("const_1009_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_94_cast_fp16 = add(x = mean_33_cast_fp16, y = const_1009_to_fp16)[name = string("add_94_cast_fp16")];
            fp32 rsqrt_33_epsilon_0 = const()[name = string("rsqrt_33_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_33_cast_fp16 = rsqrt(epsilon = rsqrt_33_epsilon_0, x = add_94_cast_fp16)[name = string("rsqrt_33_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_152_cast_fp16 = mul(x = linear_38_cast_fp16, y = rsqrt_33_cast_fp16)[name = string("mul_152_cast_fp16")];
            tensor<fp16, [768]> add_95_to_fp16 = const()[name = string("add_95_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225189184)))];
            tensor<fp16, [1, 512, 768]> mul_153_cast_fp16 = mul(x = mul_152_cast_fp16, y = add_95_to_fp16)[name = string("mul_153_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_96_cast_fp16 = add(x = add_84_cast_fp16, y = mul_153_cast_fp16)[name = string("add_96_cast_fp16")];
            fp16 const_1014_promoted_to_fp16 = const()[name = string("const_1014_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_35_cast_fp16 = pow(x = add_96_cast_fp16, y = const_1014_promoted_to_fp16)[name = string("pow_35_cast_fp16")];
            tensor<int32, [1]> mean_34_axes_0 = const()[name = string("mean_34_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_34_keep_dims_0 = const()[name = string("mean_34_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_34_cast_fp16 = reduce_mean(axes = mean_34_axes_0, keep_dims = mean_34_keep_dims_0, x = pow_35_cast_fp16)[name = string("mean_34_cast_fp16")];
            fp16 const_1017_to_fp16 = const()[name = string("const_1017_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_97_cast_fp16 = add(x = mean_34_cast_fp16, y = const_1017_to_fp16)[name = string("add_97_cast_fp16")];
            fp32 rsqrt_34_epsilon_0 = const()[name = string("rsqrt_34_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_34_cast_fp16 = rsqrt(epsilon = rsqrt_34_epsilon_0, x = add_97_cast_fp16)[name = string("rsqrt_34_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_154_cast_fp16 = mul(x = add_96_cast_fp16, y = rsqrt_34_cast_fp16)[name = string("mul_154_cast_fp16")];
            tensor<fp16, [768]> add_98_to_fp16 = const()[name = string("add_98_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225190784)))];
            tensor<fp16, [1, 512, 768]> mul_155_cast_fp16 = mul(x = mul_154_cast_fp16, y = add_98_to_fp16)[name = string("mul_155_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225192384))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226077184))))[name = string("p_st_0_model_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = mul_155_cast_fp16)[name = string("linear_39_cast_fp16")];
            string gelu_5_mode_0 = const()[name = string("gelu_5_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_5_cast_fp16 = gelu(mode = gelu_5_mode_0, x = linear_39_cast_fp16)[name = string("gelu_5_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226079552))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226964352))))[name = string("p_st_0_model_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = mul_155_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_156_cast_fp16 = mul(x = gelu_5_cast_fp16, y = linear_40_cast_fp16)[name = string("mul_156_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226966720))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227851520))))[name = string("p_st_0_model_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_41_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = mul_156_cast_fp16)[name = string("linear_41_cast_fp16")];
            fp16 const_1022_promoted_to_fp16 = const()[name = string("const_1022_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_36_cast_fp16 = pow(x = linear_41_cast_fp16, y = const_1022_promoted_to_fp16)[name = string("pow_36_cast_fp16")];
            tensor<int32, [1]> mean_35_axes_0 = const()[name = string("mean_35_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_35_keep_dims_0 = const()[name = string("mean_35_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_35_cast_fp16 = reduce_mean(axes = mean_35_axes_0, keep_dims = mean_35_keep_dims_0, x = pow_36_cast_fp16)[name = string("mean_35_cast_fp16")];
            fp16 const_1025_to_fp16 = const()[name = string("const_1025_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_99_cast_fp16 = add(x = mean_35_cast_fp16, y = const_1025_to_fp16)[name = string("add_99_cast_fp16")];
            fp32 rsqrt_35_epsilon_0 = const()[name = string("rsqrt_35_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_35_cast_fp16 = rsqrt(epsilon = rsqrt_35_epsilon_0, x = add_99_cast_fp16)[name = string("rsqrt_35_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_157_cast_fp16 = mul(x = linear_41_cast_fp16, y = rsqrt_35_cast_fp16)[name = string("mul_157_cast_fp16")];
            tensor<fp16, [768]> add_100_to_fp16 = const()[name = string("add_100_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227853120)))];
            tensor<fp16, [1, 512, 768]> mul_158_cast_fp16 = mul(x = mul_157_cast_fp16, y = add_100_to_fp16)[name = string("mul_158_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_101_cast_fp16 = add(x = add_96_cast_fp16, y = mul_158_cast_fp16)[name = string("add_101_cast_fp16")];
            fp16 const_1030_promoted_to_fp16 = const()[name = string("const_1030_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_37_cast_fp16 = pow(x = add_101_cast_fp16, y = const_1030_promoted_to_fp16)[name = string("pow_37_cast_fp16")];
            tensor<int32, [1]> mean_36_axes_0 = const()[name = string("mean_36_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_36_keep_dims_0 = const()[name = string("mean_36_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_36_cast_fp16 = reduce_mean(axes = mean_36_axes_0, keep_dims = mean_36_keep_dims_0, x = pow_37_cast_fp16)[name = string("mean_36_cast_fp16")];
            fp16 const_1033_to_fp16 = const()[name = string("const_1033_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_102_cast_fp16 = add(x = mean_36_cast_fp16, y = const_1033_to_fp16)[name = string("add_102_cast_fp16")];
            fp32 rsqrt_36_epsilon_0 = const()[name = string("rsqrt_36_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_36_cast_fp16 = rsqrt(epsilon = rsqrt_36_epsilon_0, x = add_102_cast_fp16)[name = string("rsqrt_36_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_159_cast_fp16 = mul(x = add_101_cast_fp16, y = rsqrt_36_cast_fp16)[name = string("mul_159_cast_fp16")];
            tensor<fp16, [768]> add_103_to_fp16 = const()[name = string("add_103_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227854720)))];
            tensor<fp16, [1, 512, 768]> mul_160_cast_fp16 = mul(x = mul_159_cast_fp16, y = add_103_to_fp16)[name = string("mul_160_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227856320))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228446208))))[name = string("p_st_0_model_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = mul_160_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> const_1037 = const()[name = string("const_1037"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_36_cast_fp16 = reshape(shape = const_1037, x = linear_42_cast_fp16)[name = string("view_36_cast_fp16")];
            tensor<int32, [4]> transpose_54_perm_0 = const()[name = string("transpose_54_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228447808))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228644480))))[name = string("p_st_0_model_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = mul_160_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> const_1040 = const()[name = string("const_1040"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_37_cast_fp16 = reshape(shape = const_1040, x = linear_43_cast_fp16)[name = string("view_37_cast_fp16")];
            tensor<int32, [4]> transpose_55_perm_0 = const()[name = string("transpose_55_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228645056))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228841728))))[name = string("p_st_0_model_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = mul_160_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> const_1043 = const()[name = string("const_1043"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_38_cast_fp16 = reshape(shape = const_1043, x = linear_44_cast_fp16)[name = string("view_38_cast_fp16")];
            tensor<int32, [4]> transpose_56_perm_0 = const()[name = string("transpose_56_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_1047_promoted_to_fp16 = const()[name = string("const_1047_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_54_cast_fp16 = transpose(perm = transpose_54_perm_0, x = view_36_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 3, 512, 256]> pow_38_cast_fp16 = pow(x = transpose_54_cast_fp16, y = const_1047_promoted_to_fp16)[name = string("pow_38_cast_fp16")];
            tensor<int32, [1]> mean_37_axes_0 = const()[name = string("mean_37_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_37_keep_dims_0 = const()[name = string("mean_37_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_37_cast_fp16 = reduce_mean(axes = mean_37_axes_0, keep_dims = mean_37_keep_dims_0, x = pow_38_cast_fp16)[name = string("mean_37_cast_fp16")];
            fp16 const_1050_to_fp16 = const()[name = string("const_1050_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_104_cast_fp16 = add(x = mean_37_cast_fp16, y = const_1050_to_fp16)[name = string("add_104_cast_fp16")];
            fp32 rsqrt_37_epsilon_0 = const()[name = string("rsqrt_37_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_37_cast_fp16 = rsqrt(epsilon = rsqrt_37_epsilon_0, x = add_104_cast_fp16)[name = string("rsqrt_37_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_161_cast_fp16 = mul(x = transpose_54_cast_fp16, y = rsqrt_37_cast_fp16)[name = string("mul_161_cast_fp16")];
            tensor<fp16, [256]> add_105_to_fp16 = const()[name = string("add_105_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228842304)))];
            tensor<fp16, [1, 3, 512, 256]> mul_162_cast_fp16 = mul(x = mul_161_cast_fp16, y = add_105_to_fp16)[name = string("mul_162_cast_fp16")];
            fp16 const_1055_promoted_to_fp16 = const()[name = string("const_1055_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_55_cast_fp16 = transpose(perm = transpose_55_perm_0, x = view_37_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 1, 512, 256]> pow_39_cast_fp16 = pow(x = transpose_55_cast_fp16, y = const_1055_promoted_to_fp16)[name = string("pow_39_cast_fp16")];
            tensor<int32, [1]> mean_38_axes_0 = const()[name = string("mean_38_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_38_keep_dims_0 = const()[name = string("mean_38_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_38_cast_fp16 = reduce_mean(axes = mean_38_axes_0, keep_dims = mean_38_keep_dims_0, x = pow_39_cast_fp16)[name = string("mean_38_cast_fp16")];
            fp16 const_1058_to_fp16 = const()[name = string("const_1058_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_106_cast_fp16 = add(x = mean_38_cast_fp16, y = const_1058_to_fp16)[name = string("add_106_cast_fp16")];
            fp32 rsqrt_38_epsilon_0 = const()[name = string("rsqrt_38_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_38_cast_fp16 = rsqrt(epsilon = rsqrt_38_epsilon_0, x = add_106_cast_fp16)[name = string("rsqrt_38_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_163_cast_fp16 = mul(x = transpose_55_cast_fp16, y = rsqrt_38_cast_fp16)[name = string("mul_163_cast_fp16")];
            tensor<fp16, [256]> add_107_to_fp16 = const()[name = string("add_107_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228842880)))];
            tensor<fp16, [1, 1, 512, 256]> mul_164_cast_fp16 = mul(x = mul_163_cast_fp16, y = add_107_to_fp16)[name = string("mul_164_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_165_cast_fp16 = mul(x = mul_162_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_165_cast_fp16")];
            tensor<int32, [4]> slice_207_begin_0 = const()[name = string("slice_207_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_207_end_0 = const()[name = string("slice_207_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_207_end_mask_0 = const()[name = string("slice_207_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_207_cast_fp16 = slice_by_index(begin = slice_207_begin_0, end = slice_207_end_0, end_mask = slice_207_end_mask_0, x = mul_162_cast_fp16)[name = string("slice_207_cast_fp16")];
            tensor<int32, [4]> slice_208_begin_0 = const()[name = string("slice_208_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_208_end_0 = const()[name = string("slice_208_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_208_end_mask_0 = const()[name = string("slice_208_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_208_cast_fp16 = slice_by_index(begin = slice_208_begin_0, end = slice_208_end_0, end_mask = slice_208_end_mask_0, x = mul_162_cast_fp16)[name = string("slice_208_cast_fp16")];
            fp16 const_1070_promoted_to_fp16 = const()[name = string("const_1070_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_12_cast_fp16 = mul(x = slice_208_cast_fp16, y = const_1070_promoted_to_fp16)[name = string("neg_12_cast_fp16")];
            int32 const_1071 = const()[name = string("const_1071"), val = int32(-1)];
            bool cat_48_interleave_0 = const()[name = string("cat_48_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_48_cast_fp16 = concat(axis = const_1071, interleave = cat_48_interleave_0, values = (neg_12_cast_fp16, slice_207_cast_fp16))[name = string("cat_48_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_166_cast_fp16 = mul(x = cat_48_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_166_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_108_cast_fp16 = add(x = mul_165_cast_fp16, y = mul_166_cast_fp16)[name = string("add_108_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_167_cast_fp16 = mul(x = mul_164_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_167_cast_fp16")];
            tensor<int32, [4]> slice_209_begin_0 = const()[name = string("slice_209_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_209_end_0 = const()[name = string("slice_209_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_209_end_mask_0 = const()[name = string("slice_209_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_209_cast_fp16 = slice_by_index(begin = slice_209_begin_0, end = slice_209_end_0, end_mask = slice_209_end_mask_0, x = mul_164_cast_fp16)[name = string("slice_209_cast_fp16")];
            tensor<int32, [4]> slice_210_begin_0 = const()[name = string("slice_210_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_210_end_0 = const()[name = string("slice_210_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_210_end_mask_0 = const()[name = string("slice_210_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_210_cast_fp16 = slice_by_index(begin = slice_210_begin_0, end = slice_210_end_0, end_mask = slice_210_end_mask_0, x = mul_164_cast_fp16)[name = string("slice_210_cast_fp16")];
            fp16 const_1078_promoted_to_fp16 = const()[name = string("const_1078_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_13_cast_fp16 = mul(x = slice_210_cast_fp16, y = const_1078_promoted_to_fp16)[name = string("neg_13_cast_fp16")];
            int32 const_1079 = const()[name = string("const_1079"), val = int32(-1)];
            bool cat_49_interleave_0 = const()[name = string("cat_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_49_cast_fp16 = concat(axis = const_1079, interleave = cat_49_interleave_0, values = (neg_13_cast_fp16, slice_209_cast_fp16))[name = string("cat_49_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_168_cast_fp16 = mul(x = cat_49_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_168_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_109_cast_fp16 = add(x = mul_167_cast_fp16, y = mul_168_cast_fp16)[name = string("add_109_cast_fp16")];
            int32 const_1080 = const()[name = string("const_1080"), val = int32(-2)];
            bool cat_50_interleave_0 = const()[name = string("cat_50_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_50_cast_fp16 = concat(axis = const_1080, interleave = cat_50_interleave_0, values = add_109_cast_fp16)[name = string("cat_50_cast_fp16")];
            int32 const_1081 = const()[name = string("const_1081"), val = int32(-2)];
            bool cat_51_interleave_0 = const()[name = string("cat_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_56_cast_fp16 = transpose(perm = transpose_56_perm_0, x = view_38_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 1, 512, 256]> cat_51_cast_fp16 = concat(axis = const_1081, interleave = cat_51_interleave_0, values = transpose_56_cast_fp16)[name = string("cat_51_cast_fp16")];
            tensor<int32, [1]> unsqueeze_103_axes_0 = const()[name = string("unsqueeze_103_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_103_cast_fp16 = expand_dims(axes = unsqueeze_103_axes_0, x = cat_50_cast_fp16)[name = string("unsqueeze_103_cast_fp16")];
            tensor<int32, [5]> expand_38_reps_0 = const()[name = string("expand_38_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_38_cast_fp16 = tile(reps = expand_38_reps_0, x = unsqueeze_103_cast_fp16)[name = string("expand_38_cast_fp16")];
            tensor<int32, [4]> const_1096 = const()[name = string("const_1096"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_39_cast_fp16 = reshape(shape = const_1096, x = expand_38_cast_fp16)[name = string("view_39_cast_fp16")];
            tensor<int32, [1]> unsqueeze_104_axes_0 = const()[name = string("unsqueeze_104_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_104_cast_fp16 = expand_dims(axes = unsqueeze_104_axes_0, x = cat_51_cast_fp16)[name = string("unsqueeze_104_cast_fp16")];
            tensor<int32, [5]> expand_39_reps_0 = const()[name = string("expand_39_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_39_cast_fp16 = tile(reps = expand_39_reps_0, x = unsqueeze_104_cast_fp16)[name = string("expand_39_cast_fp16")];
            tensor<int32, [4]> const_1111 = const()[name = string("const_1111"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_40_cast_fp16 = reshape(shape = const_1111, x = expand_39_cast_fp16)[name = string("view_40_cast_fp16")];
            bool matmul_36_transpose_x_1 = const()[name = string("matmul_36_transpose_x_1"), val = bool(false)];
            bool matmul_36_transpose_y_1 = const()[name = string("matmul_36_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_36_cast_fp16 = matmul(transpose_x = matmul_36_transpose_x_1, transpose_y = matmul_36_transpose_y_1, x = add_108_cast_fp16, y = view_39_cast_fp16)[name = string("matmul_36_cast_fp16")];
            fp16 const_1114_to_fp16 = const()[name = string("const_1114_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_169_cast_fp16 = mul(x = matmul_36_cast_fp16, y = const_1114_to_fp16)[name = string("mul_169_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_110_cast_fp16 = add(x = mul_169_cast_fp16, y = expand_cast_fp16)[name = string("add_110_cast_fp16")];
            int32 const_1124 = const()[name = string("const_1124"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_6_cast_fp16 = softmax(axis = const_1124, x = add_110_cast_fp16)[name = string("softmax_6_cast_fp16")];
            bool matmul_37_transpose_x_0 = const()[name = string("matmul_37_transpose_x_0"), val = bool(false)];
            bool matmul_37_transpose_y_0 = const()[name = string("matmul_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_37_cast_fp16 = matmul(transpose_x = matmul_37_transpose_x_0, transpose_y = matmul_37_transpose_y_0, x = softmax_6_cast_fp16, y = view_40_cast_fp16)[name = string("matmul_37_cast_fp16")];
            tensor<int32, [4]> transpose_58_perm_0 = const()[name = string("transpose_58_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_1129 = const()[name = string("const_1129"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_58_cast_fp16 = transpose(perm = transpose_58_perm_0, x = matmul_37_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 512, 768]> view_41_cast_fp16 = reshape(shape = const_1129, x = transpose_58_cast_fp16)[name = string("view_41_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228843456))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229433344))))[name = string("p_st_0_model_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_45_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = view_41_cast_fp16)[name = string("linear_45_cast_fp16")];
            fp16 const_1131_promoted_to_fp16 = const()[name = string("const_1131_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_40_cast_fp16 = pow(x = linear_45_cast_fp16, y = const_1131_promoted_to_fp16)[name = string("pow_40_cast_fp16")];
            tensor<int32, [1]> mean_39_axes_0 = const()[name = string("mean_39_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_39_keep_dims_0 = const()[name = string("mean_39_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_39_cast_fp16 = reduce_mean(axes = mean_39_axes_0, keep_dims = mean_39_keep_dims_0, x = pow_40_cast_fp16)[name = string("mean_39_cast_fp16")];
            fp16 const_1134_to_fp16 = const()[name = string("const_1134_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_111_cast_fp16 = add(x = mean_39_cast_fp16, y = const_1134_to_fp16)[name = string("add_111_cast_fp16")];
            fp32 rsqrt_39_epsilon_0 = const()[name = string("rsqrt_39_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_39_cast_fp16 = rsqrt(epsilon = rsqrt_39_epsilon_0, x = add_111_cast_fp16)[name = string("rsqrt_39_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_170_cast_fp16 = mul(x = linear_45_cast_fp16, y = rsqrt_39_cast_fp16)[name = string("mul_170_cast_fp16")];
            tensor<fp16, [768]> add_112_to_fp16 = const()[name = string("add_112_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229434944)))];
            tensor<fp16, [1, 512, 768]> mul_171_cast_fp16 = mul(x = mul_170_cast_fp16, y = add_112_to_fp16)[name = string("mul_171_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_113_cast_fp16 = add(x = add_101_cast_fp16, y = mul_171_cast_fp16)[name = string("add_113_cast_fp16")];
            fp16 const_1139_promoted_to_fp16 = const()[name = string("const_1139_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_41_cast_fp16 = pow(x = add_113_cast_fp16, y = const_1139_promoted_to_fp16)[name = string("pow_41_cast_fp16")];
            tensor<int32, [1]> mean_40_axes_0 = const()[name = string("mean_40_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_40_keep_dims_0 = const()[name = string("mean_40_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_40_cast_fp16 = reduce_mean(axes = mean_40_axes_0, keep_dims = mean_40_keep_dims_0, x = pow_41_cast_fp16)[name = string("mean_40_cast_fp16")];
            fp16 const_1142_to_fp16 = const()[name = string("const_1142_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_114_cast_fp16 = add(x = mean_40_cast_fp16, y = const_1142_to_fp16)[name = string("add_114_cast_fp16")];
            fp32 rsqrt_40_epsilon_0 = const()[name = string("rsqrt_40_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_40_cast_fp16 = rsqrt(epsilon = rsqrt_40_epsilon_0, x = add_114_cast_fp16)[name = string("rsqrt_40_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_172_cast_fp16 = mul(x = add_113_cast_fp16, y = rsqrt_40_cast_fp16)[name = string("mul_172_cast_fp16")];
            tensor<fp16, [768]> add_115_to_fp16 = const()[name = string("add_115_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229436544)))];
            tensor<fp16, [1, 512, 768]> mul_173_cast_fp16 = mul(x = mul_172_cast_fp16, y = add_115_to_fp16)[name = string("mul_173_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229438144))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230322944))))[name = string("p_st_0_model_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = mul_173_cast_fp16)[name = string("linear_46_cast_fp16")];
            string gelu_6_mode_0 = const()[name = string("gelu_6_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_6_cast_fp16 = gelu(mode = gelu_6_mode_0, x = linear_46_cast_fp16)[name = string("gelu_6_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230325312))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231210112))))[name = string("p_st_0_model_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = mul_173_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_174_cast_fp16 = mul(x = gelu_6_cast_fp16, y = linear_47_cast_fp16)[name = string("mul_174_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231212480))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232097280))))[name = string("p_st_0_model_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = mul_174_cast_fp16)[name = string("linear_48_cast_fp16")];
            fp16 const_1147_promoted_to_fp16 = const()[name = string("const_1147_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_42_cast_fp16 = pow(x = linear_48_cast_fp16, y = const_1147_promoted_to_fp16)[name = string("pow_42_cast_fp16")];
            tensor<int32, [1]> mean_41_axes_0 = const()[name = string("mean_41_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_41_keep_dims_0 = const()[name = string("mean_41_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_41_cast_fp16 = reduce_mean(axes = mean_41_axes_0, keep_dims = mean_41_keep_dims_0, x = pow_42_cast_fp16)[name = string("mean_41_cast_fp16")];
            fp16 const_1150_to_fp16 = const()[name = string("const_1150_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_116_cast_fp16 = add(x = mean_41_cast_fp16, y = const_1150_to_fp16)[name = string("add_116_cast_fp16")];
            fp32 rsqrt_41_epsilon_0 = const()[name = string("rsqrt_41_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_41_cast_fp16 = rsqrt(epsilon = rsqrt_41_epsilon_0, x = add_116_cast_fp16)[name = string("rsqrt_41_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_175_cast_fp16 = mul(x = linear_48_cast_fp16, y = rsqrt_41_cast_fp16)[name = string("mul_175_cast_fp16")];
            tensor<fp16, [768]> add_117_to_fp16 = const()[name = string("add_117_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232098880)))];
            tensor<fp16, [1, 512, 768]> mul_176_cast_fp16 = mul(x = mul_175_cast_fp16, y = add_117_to_fp16)[name = string("mul_176_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_118_cast_fp16 = add(x = add_113_cast_fp16, y = mul_176_cast_fp16)[name = string("add_118_cast_fp16")];
            fp16 const_1155_promoted_to_fp16 = const()[name = string("const_1155_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_43_cast_fp16 = pow(x = add_118_cast_fp16, y = const_1155_promoted_to_fp16)[name = string("pow_43_cast_fp16")];
            tensor<int32, [1]> mean_42_axes_0 = const()[name = string("mean_42_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_42_keep_dims_0 = const()[name = string("mean_42_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_42_cast_fp16 = reduce_mean(axes = mean_42_axes_0, keep_dims = mean_42_keep_dims_0, x = pow_43_cast_fp16)[name = string("mean_42_cast_fp16")];
            fp16 const_1158_to_fp16 = const()[name = string("const_1158_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_119_cast_fp16 = add(x = mean_42_cast_fp16, y = const_1158_to_fp16)[name = string("add_119_cast_fp16")];
            fp32 rsqrt_42_epsilon_0 = const()[name = string("rsqrt_42_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_42_cast_fp16 = rsqrt(epsilon = rsqrt_42_epsilon_0, x = add_119_cast_fp16)[name = string("rsqrt_42_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_177_cast_fp16 = mul(x = add_118_cast_fp16, y = rsqrt_42_cast_fp16)[name = string("mul_177_cast_fp16")];
            tensor<fp16, [768]> add_120_to_fp16 = const()[name = string("add_120_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232100480)))];
            tensor<fp16, [1, 512, 768]> mul_178_cast_fp16 = mul(x = mul_177_cast_fp16, y = add_120_to_fp16)[name = string("mul_178_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232102080))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232691968))))[name = string("p_st_0_model_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_49_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = mul_178_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> const_1162 = const()[name = string("const_1162"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_42_cast_fp16 = reshape(shape = const_1162, x = linear_49_cast_fp16)[name = string("view_42_cast_fp16")];
            tensor<int32, [4]> transpose_59_perm_0 = const()[name = string("transpose_59_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232693568))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232890240))))[name = string("p_st_0_model_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_50_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = mul_178_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> const_1165 = const()[name = string("const_1165"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_43_cast_fp16 = reshape(shape = const_1165, x = linear_50_cast_fp16)[name = string("view_43_cast_fp16")];
            tensor<int32, [4]> transpose_60_perm_0 = const()[name = string("transpose_60_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232890816))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233087488))))[name = string("p_st_0_model_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = mul_178_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> const_1168 = const()[name = string("const_1168"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_44_cast_fp16 = reshape(shape = const_1168, x = linear_51_cast_fp16)[name = string("view_44_cast_fp16")];
            tensor<int32, [4]> transpose_61_perm_0 = const()[name = string("transpose_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_1172_promoted_to_fp16 = const()[name = string("const_1172_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_59_cast_fp16 = transpose(perm = transpose_59_perm_0, x = view_42_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 3, 512, 256]> pow_44_cast_fp16 = pow(x = transpose_59_cast_fp16, y = const_1172_promoted_to_fp16)[name = string("pow_44_cast_fp16")];
            tensor<int32, [1]> mean_43_axes_0 = const()[name = string("mean_43_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_43_keep_dims_0 = const()[name = string("mean_43_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_43_cast_fp16 = reduce_mean(axes = mean_43_axes_0, keep_dims = mean_43_keep_dims_0, x = pow_44_cast_fp16)[name = string("mean_43_cast_fp16")];
            fp16 const_1175_to_fp16 = const()[name = string("const_1175_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_121_cast_fp16 = add(x = mean_43_cast_fp16, y = const_1175_to_fp16)[name = string("add_121_cast_fp16")];
            fp32 rsqrt_43_epsilon_0 = const()[name = string("rsqrt_43_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_43_cast_fp16 = rsqrt(epsilon = rsqrt_43_epsilon_0, x = add_121_cast_fp16)[name = string("rsqrt_43_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_179_cast_fp16 = mul(x = transpose_59_cast_fp16, y = rsqrt_43_cast_fp16)[name = string("mul_179_cast_fp16")];
            tensor<fp16, [256]> add_122_to_fp16 = const()[name = string("add_122_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233088064)))];
            tensor<fp16, [1, 3, 512, 256]> mul_180_cast_fp16 = mul(x = mul_179_cast_fp16, y = add_122_to_fp16)[name = string("mul_180_cast_fp16")];
            fp16 const_1180_promoted_to_fp16 = const()[name = string("const_1180_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_60_cast_fp16 = transpose(perm = transpose_60_perm_0, x = view_43_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 1, 512, 256]> pow_45_cast_fp16 = pow(x = transpose_60_cast_fp16, y = const_1180_promoted_to_fp16)[name = string("pow_45_cast_fp16")];
            tensor<int32, [1]> mean_44_axes_0 = const()[name = string("mean_44_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_44_keep_dims_0 = const()[name = string("mean_44_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_44_cast_fp16 = reduce_mean(axes = mean_44_axes_0, keep_dims = mean_44_keep_dims_0, x = pow_45_cast_fp16)[name = string("mean_44_cast_fp16")];
            fp16 const_1183_to_fp16 = const()[name = string("const_1183_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_123_cast_fp16 = add(x = mean_44_cast_fp16, y = const_1183_to_fp16)[name = string("add_123_cast_fp16")];
            fp32 rsqrt_44_epsilon_0 = const()[name = string("rsqrt_44_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_44_cast_fp16 = rsqrt(epsilon = rsqrt_44_epsilon_0, x = add_123_cast_fp16)[name = string("rsqrt_44_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_181_cast_fp16 = mul(x = transpose_60_cast_fp16, y = rsqrt_44_cast_fp16)[name = string("mul_181_cast_fp16")];
            tensor<fp16, [256]> add_124_to_fp16 = const()[name = string("add_124_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233088640)))];
            tensor<fp16, [1, 1, 512, 256]> mul_182_cast_fp16 = mul(x = mul_181_cast_fp16, y = add_124_to_fp16)[name = string("mul_182_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_183_cast_fp16 = mul(x = mul_180_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_183_cast_fp16")];
            tensor<int32, [4]> slice_230_begin_0 = const()[name = string("slice_230_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_230_end_0 = const()[name = string("slice_230_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_230_end_mask_0 = const()[name = string("slice_230_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_230_cast_fp16 = slice_by_index(begin = slice_230_begin_0, end = slice_230_end_0, end_mask = slice_230_end_mask_0, x = mul_180_cast_fp16)[name = string("slice_230_cast_fp16")];
            tensor<int32, [4]> slice_231_begin_0 = const()[name = string("slice_231_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_231_end_0 = const()[name = string("slice_231_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_231_end_mask_0 = const()[name = string("slice_231_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_231_cast_fp16 = slice_by_index(begin = slice_231_begin_0, end = slice_231_end_0, end_mask = slice_231_end_mask_0, x = mul_180_cast_fp16)[name = string("slice_231_cast_fp16")];
            fp16 const_1195_promoted_to_fp16 = const()[name = string("const_1195_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_14_cast_fp16 = mul(x = slice_231_cast_fp16, y = const_1195_promoted_to_fp16)[name = string("neg_14_cast_fp16")];
            int32 const_1196 = const()[name = string("const_1196"), val = int32(-1)];
            bool cat_52_interleave_0 = const()[name = string("cat_52_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_52_cast_fp16 = concat(axis = const_1196, interleave = cat_52_interleave_0, values = (neg_14_cast_fp16, slice_230_cast_fp16))[name = string("cat_52_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_184_cast_fp16 = mul(x = cat_52_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_184_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_125_cast_fp16 = add(x = mul_183_cast_fp16, y = mul_184_cast_fp16)[name = string("add_125_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_185_cast_fp16 = mul(x = mul_182_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_185_cast_fp16")];
            tensor<int32, [4]> slice_232_begin_0 = const()[name = string("slice_232_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_232_end_0 = const()[name = string("slice_232_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_232_end_mask_0 = const()[name = string("slice_232_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_232_cast_fp16 = slice_by_index(begin = slice_232_begin_0, end = slice_232_end_0, end_mask = slice_232_end_mask_0, x = mul_182_cast_fp16)[name = string("slice_232_cast_fp16")];
            tensor<int32, [4]> slice_233_begin_0 = const()[name = string("slice_233_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_233_end_0 = const()[name = string("slice_233_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_233_end_mask_0 = const()[name = string("slice_233_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_233_cast_fp16 = slice_by_index(begin = slice_233_begin_0, end = slice_233_end_0, end_mask = slice_233_end_mask_0, x = mul_182_cast_fp16)[name = string("slice_233_cast_fp16")];
            fp16 const_1203_promoted_to_fp16 = const()[name = string("const_1203_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_15_cast_fp16 = mul(x = slice_233_cast_fp16, y = const_1203_promoted_to_fp16)[name = string("neg_15_cast_fp16")];
            int32 const_1204 = const()[name = string("const_1204"), val = int32(-1)];
            bool cat_53_interleave_0 = const()[name = string("cat_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_53_cast_fp16 = concat(axis = const_1204, interleave = cat_53_interleave_0, values = (neg_15_cast_fp16, slice_232_cast_fp16))[name = string("cat_53_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_186_cast_fp16 = mul(x = cat_53_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_186_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_126_cast_fp16 = add(x = mul_185_cast_fp16, y = mul_186_cast_fp16)[name = string("add_126_cast_fp16")];
            int32 const_1205 = const()[name = string("const_1205"), val = int32(-2)];
            bool cat_54_interleave_0 = const()[name = string("cat_54_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_54_cast_fp16 = concat(axis = const_1205, interleave = cat_54_interleave_0, values = add_126_cast_fp16)[name = string("cat_54_cast_fp16")];
            int32 const_1206 = const()[name = string("const_1206"), val = int32(-2)];
            bool cat_55_interleave_0 = const()[name = string("cat_55_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_61_cast_fp16 = transpose(perm = transpose_61_perm_0, x = view_44_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 1, 512, 256]> cat_55_cast_fp16 = concat(axis = const_1206, interleave = cat_55_interleave_0, values = transpose_61_cast_fp16)[name = string("cat_55_cast_fp16")];
            tensor<int32, [1]> unsqueeze_107_axes_0 = const()[name = string("unsqueeze_107_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_107_cast_fp16 = expand_dims(axes = unsqueeze_107_axes_0, x = cat_54_cast_fp16)[name = string("unsqueeze_107_cast_fp16")];
            tensor<int32, [5]> expand_40_reps_0 = const()[name = string("expand_40_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_40_cast_fp16 = tile(reps = expand_40_reps_0, x = unsqueeze_107_cast_fp16)[name = string("expand_40_cast_fp16")];
            tensor<int32, [4]> const_1221 = const()[name = string("const_1221"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_45_cast_fp16 = reshape(shape = const_1221, x = expand_40_cast_fp16)[name = string("view_45_cast_fp16")];
            tensor<int32, [1]> unsqueeze_108_axes_0 = const()[name = string("unsqueeze_108_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_108_cast_fp16 = expand_dims(axes = unsqueeze_108_axes_0, x = cat_55_cast_fp16)[name = string("unsqueeze_108_cast_fp16")];
            tensor<int32, [5]> expand_41_reps_0 = const()[name = string("expand_41_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_41_cast_fp16 = tile(reps = expand_41_reps_0, x = unsqueeze_108_cast_fp16)[name = string("expand_41_cast_fp16")];
            tensor<int32, [4]> const_1236 = const()[name = string("const_1236"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_46_cast_fp16 = reshape(shape = const_1236, x = expand_41_cast_fp16)[name = string("view_46_cast_fp16")];
            bool matmul_38_transpose_x_1 = const()[name = string("matmul_38_transpose_x_1"), val = bool(false)];
            bool matmul_38_transpose_y_1 = const()[name = string("matmul_38_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_38_cast_fp16 = matmul(transpose_x = matmul_38_transpose_x_1, transpose_y = matmul_38_transpose_y_1, x = add_125_cast_fp16, y = view_45_cast_fp16)[name = string("matmul_38_cast_fp16")];
            fp16 const_1239_to_fp16 = const()[name = string("const_1239_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_187_cast_fp16 = mul(x = matmul_38_cast_fp16, y = const_1239_to_fp16)[name = string("mul_187_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_127_cast_fp16 = add(x = mul_187_cast_fp16, y = expand_cast_fp16)[name = string("add_127_cast_fp16")];
            int32 const_1249 = const()[name = string("const_1249"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_7_cast_fp16 = softmax(axis = const_1249, x = add_127_cast_fp16)[name = string("softmax_7_cast_fp16")];
            bool matmul_39_transpose_x_0 = const()[name = string("matmul_39_transpose_x_0"), val = bool(false)];
            bool matmul_39_transpose_y_0 = const()[name = string("matmul_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_39_cast_fp16 = matmul(transpose_x = matmul_39_transpose_x_0, transpose_y = matmul_39_transpose_y_0, x = softmax_7_cast_fp16, y = view_46_cast_fp16)[name = string("matmul_39_cast_fp16")];
            tensor<int32, [4]> transpose_63_perm_0 = const()[name = string("transpose_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_1254 = const()[name = string("const_1254"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_63_cast_fp16 = transpose(perm = transpose_63_perm_0, x = matmul_39_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 512, 768]> view_47_cast_fp16 = reshape(shape = const_1254, x = transpose_63_cast_fp16)[name = string("view_47_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233089216))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233679104))))[name = string("p_st_0_model_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = view_47_cast_fp16)[name = string("linear_52_cast_fp16")];
            fp16 const_1256_promoted_to_fp16 = const()[name = string("const_1256_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_46_cast_fp16 = pow(x = linear_52_cast_fp16, y = const_1256_promoted_to_fp16)[name = string("pow_46_cast_fp16")];
            tensor<int32, [1]> mean_45_axes_0 = const()[name = string("mean_45_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_45_keep_dims_0 = const()[name = string("mean_45_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_45_cast_fp16 = reduce_mean(axes = mean_45_axes_0, keep_dims = mean_45_keep_dims_0, x = pow_46_cast_fp16)[name = string("mean_45_cast_fp16")];
            fp16 const_1259_to_fp16 = const()[name = string("const_1259_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_128_cast_fp16 = add(x = mean_45_cast_fp16, y = const_1259_to_fp16)[name = string("add_128_cast_fp16")];
            fp32 rsqrt_45_epsilon_0 = const()[name = string("rsqrt_45_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_45_cast_fp16 = rsqrt(epsilon = rsqrt_45_epsilon_0, x = add_128_cast_fp16)[name = string("rsqrt_45_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_188_cast_fp16 = mul(x = linear_52_cast_fp16, y = rsqrt_45_cast_fp16)[name = string("mul_188_cast_fp16")];
            tensor<fp16, [768]> add_129_to_fp16 = const()[name = string("add_129_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233680704)))];
            tensor<fp16, [1, 512, 768]> mul_189_cast_fp16 = mul(x = mul_188_cast_fp16, y = add_129_to_fp16)[name = string("mul_189_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_130_cast_fp16 = add(x = add_118_cast_fp16, y = mul_189_cast_fp16)[name = string("add_130_cast_fp16")];
            fp16 const_1264_promoted_to_fp16 = const()[name = string("const_1264_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_47_cast_fp16 = pow(x = add_130_cast_fp16, y = const_1264_promoted_to_fp16)[name = string("pow_47_cast_fp16")];
            tensor<int32, [1]> mean_46_axes_0 = const()[name = string("mean_46_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_46_keep_dims_0 = const()[name = string("mean_46_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_46_cast_fp16 = reduce_mean(axes = mean_46_axes_0, keep_dims = mean_46_keep_dims_0, x = pow_47_cast_fp16)[name = string("mean_46_cast_fp16")];
            fp16 const_1267_to_fp16 = const()[name = string("const_1267_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_131_cast_fp16 = add(x = mean_46_cast_fp16, y = const_1267_to_fp16)[name = string("add_131_cast_fp16")];
            fp32 rsqrt_46_epsilon_0 = const()[name = string("rsqrt_46_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_46_cast_fp16 = rsqrt(epsilon = rsqrt_46_epsilon_0, x = add_131_cast_fp16)[name = string("rsqrt_46_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_190_cast_fp16 = mul(x = add_130_cast_fp16, y = rsqrt_46_cast_fp16)[name = string("mul_190_cast_fp16")];
            tensor<fp16, [768]> add_132_to_fp16 = const()[name = string("add_132_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233682304)))];
            tensor<fp16, [1, 512, 768]> mul_191_cast_fp16 = mul(x = mul_190_cast_fp16, y = add_132_to_fp16)[name = string("mul_191_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233683904))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234568704))))[name = string("p_st_0_model_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = mul_191_cast_fp16)[name = string("linear_53_cast_fp16")];
            string gelu_7_mode_0 = const()[name = string("gelu_7_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_7_cast_fp16 = gelu(mode = gelu_7_mode_0, x = linear_53_cast_fp16)[name = string("gelu_7_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234571072))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235455872))))[name = string("p_st_0_model_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = mul_191_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_192_cast_fp16 = mul(x = gelu_7_cast_fp16, y = linear_54_cast_fp16)[name = string("mul_192_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235458240))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236343040))))[name = string("p_st_0_model_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_55_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = mul_192_cast_fp16)[name = string("linear_55_cast_fp16")];
            fp16 const_1272_promoted_to_fp16 = const()[name = string("const_1272_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_48_cast_fp16 = pow(x = linear_55_cast_fp16, y = const_1272_promoted_to_fp16)[name = string("pow_48_cast_fp16")];
            tensor<int32, [1]> mean_47_axes_0 = const()[name = string("mean_47_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_47_keep_dims_0 = const()[name = string("mean_47_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_47_cast_fp16 = reduce_mean(axes = mean_47_axes_0, keep_dims = mean_47_keep_dims_0, x = pow_48_cast_fp16)[name = string("mean_47_cast_fp16")];
            fp16 const_1275_to_fp16 = const()[name = string("const_1275_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_133_cast_fp16 = add(x = mean_47_cast_fp16, y = const_1275_to_fp16)[name = string("add_133_cast_fp16")];
            fp32 rsqrt_47_epsilon_0 = const()[name = string("rsqrt_47_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_47_cast_fp16 = rsqrt(epsilon = rsqrt_47_epsilon_0, x = add_133_cast_fp16)[name = string("rsqrt_47_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_193_cast_fp16 = mul(x = linear_55_cast_fp16, y = rsqrt_47_cast_fp16)[name = string("mul_193_cast_fp16")];
            tensor<fp16, [768]> add_134_to_fp16 = const()[name = string("add_134_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236344640)))];
            tensor<fp16, [1, 512, 768]> mul_194_cast_fp16 = mul(x = mul_193_cast_fp16, y = add_134_to_fp16)[name = string("mul_194_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_135_cast_fp16 = add(x = add_130_cast_fp16, y = mul_194_cast_fp16)[name = string("add_135_cast_fp16")];
            fp16 const_1280_promoted_to_fp16 = const()[name = string("const_1280_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_49_cast_fp16 = pow(x = add_135_cast_fp16, y = const_1280_promoted_to_fp16)[name = string("pow_49_cast_fp16")];
            tensor<int32, [1]> mean_48_axes_0 = const()[name = string("mean_48_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_48_keep_dims_0 = const()[name = string("mean_48_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_48_cast_fp16 = reduce_mean(axes = mean_48_axes_0, keep_dims = mean_48_keep_dims_0, x = pow_49_cast_fp16)[name = string("mean_48_cast_fp16")];
            fp16 const_1283_to_fp16 = const()[name = string("const_1283_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_136_cast_fp16 = add(x = mean_48_cast_fp16, y = const_1283_to_fp16)[name = string("add_136_cast_fp16")];
            fp32 rsqrt_48_epsilon_0 = const()[name = string("rsqrt_48_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_48_cast_fp16 = rsqrt(epsilon = rsqrt_48_epsilon_0, x = add_136_cast_fp16)[name = string("rsqrt_48_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_195_cast_fp16 = mul(x = add_135_cast_fp16, y = rsqrt_48_cast_fp16)[name = string("mul_195_cast_fp16")];
            tensor<fp16, [768]> add_137_to_fp16 = const()[name = string("add_137_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236346240)))];
            tensor<fp16, [1, 512, 768]> mul_196_cast_fp16 = mul(x = mul_195_cast_fp16, y = add_137_to_fp16)[name = string("mul_196_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236347840))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236937728))))[name = string("p_st_0_model_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = mul_196_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> const_1287 = const()[name = string("const_1287"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_48_cast_fp16 = reshape(shape = const_1287, x = linear_56_cast_fp16)[name = string("view_48_cast_fp16")];
            tensor<int32, [4]> transpose_64_perm_0 = const()[name = string("transpose_64_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236939328))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237136000))))[name = string("p_st_0_model_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = mul_196_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> const_1290 = const()[name = string("const_1290"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_49_cast_fp16 = reshape(shape = const_1290, x = linear_57_cast_fp16)[name = string("view_49_cast_fp16")];
            tensor<int32, [4]> transpose_65_perm_0 = const()[name = string("transpose_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237136576))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237333248))))[name = string("p_st_0_model_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_58_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = mul_196_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> const_1293 = const()[name = string("const_1293"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_50_cast_fp16 = reshape(shape = const_1293, x = linear_58_cast_fp16)[name = string("view_50_cast_fp16")];
            tensor<int32, [4]> transpose_66_perm_0 = const()[name = string("transpose_66_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_1297_promoted_to_fp16 = const()[name = string("const_1297_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_64_cast_fp16 = transpose(perm = transpose_64_perm_0, x = view_48_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 3, 512, 256]> pow_50_cast_fp16 = pow(x = transpose_64_cast_fp16, y = const_1297_promoted_to_fp16)[name = string("pow_50_cast_fp16")];
            tensor<int32, [1]> mean_49_axes_0 = const()[name = string("mean_49_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_49_keep_dims_0 = const()[name = string("mean_49_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_49_cast_fp16 = reduce_mean(axes = mean_49_axes_0, keep_dims = mean_49_keep_dims_0, x = pow_50_cast_fp16)[name = string("mean_49_cast_fp16")];
            fp16 const_1300_to_fp16 = const()[name = string("const_1300_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_138_cast_fp16 = add(x = mean_49_cast_fp16, y = const_1300_to_fp16)[name = string("add_138_cast_fp16")];
            fp32 rsqrt_49_epsilon_0 = const()[name = string("rsqrt_49_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_49_cast_fp16 = rsqrt(epsilon = rsqrt_49_epsilon_0, x = add_138_cast_fp16)[name = string("rsqrt_49_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_197_cast_fp16 = mul(x = transpose_64_cast_fp16, y = rsqrt_49_cast_fp16)[name = string("mul_197_cast_fp16")];
            tensor<fp16, [256]> add_139_to_fp16 = const()[name = string("add_139_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237333824)))];
            tensor<fp16, [1, 3, 512, 256]> mul_198_cast_fp16 = mul(x = mul_197_cast_fp16, y = add_139_to_fp16)[name = string("mul_198_cast_fp16")];
            fp16 const_1305_promoted_to_fp16 = const()[name = string("const_1305_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_65_cast_fp16 = transpose(perm = transpose_65_perm_0, x = view_49_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 1, 512, 256]> pow_51_cast_fp16 = pow(x = transpose_65_cast_fp16, y = const_1305_promoted_to_fp16)[name = string("pow_51_cast_fp16")];
            tensor<int32, [1]> mean_50_axes_0 = const()[name = string("mean_50_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_50_keep_dims_0 = const()[name = string("mean_50_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_50_cast_fp16 = reduce_mean(axes = mean_50_axes_0, keep_dims = mean_50_keep_dims_0, x = pow_51_cast_fp16)[name = string("mean_50_cast_fp16")];
            fp16 const_1308_to_fp16 = const()[name = string("const_1308_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_140_cast_fp16 = add(x = mean_50_cast_fp16, y = const_1308_to_fp16)[name = string("add_140_cast_fp16")];
            fp32 rsqrt_50_epsilon_0 = const()[name = string("rsqrt_50_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_50_cast_fp16 = rsqrt(epsilon = rsqrt_50_epsilon_0, x = add_140_cast_fp16)[name = string("rsqrt_50_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_199_cast_fp16 = mul(x = transpose_65_cast_fp16, y = rsqrt_50_cast_fp16)[name = string("mul_199_cast_fp16")];
            tensor<fp16, [256]> add_141_to_fp16 = const()[name = string("add_141_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237334400)))];
            tensor<fp16, [1, 1, 512, 256]> mul_200_cast_fp16 = mul(x = mul_199_cast_fp16, y = add_141_to_fp16)[name = string("mul_200_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_201_cast_fp16 = mul(x = mul_198_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_201_cast_fp16")];
            tensor<int32, [4]> slice_253_begin_0 = const()[name = string("slice_253_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_253_end_0 = const()[name = string("slice_253_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_253_end_mask_0 = const()[name = string("slice_253_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_253_cast_fp16 = slice_by_index(begin = slice_253_begin_0, end = slice_253_end_0, end_mask = slice_253_end_mask_0, x = mul_198_cast_fp16)[name = string("slice_253_cast_fp16")];
            tensor<int32, [4]> slice_254_begin_0 = const()[name = string("slice_254_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_254_end_0 = const()[name = string("slice_254_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_254_end_mask_0 = const()[name = string("slice_254_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_254_cast_fp16 = slice_by_index(begin = slice_254_begin_0, end = slice_254_end_0, end_mask = slice_254_end_mask_0, x = mul_198_cast_fp16)[name = string("slice_254_cast_fp16")];
            fp16 const_1320_promoted_to_fp16 = const()[name = string("const_1320_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_16_cast_fp16 = mul(x = slice_254_cast_fp16, y = const_1320_promoted_to_fp16)[name = string("neg_16_cast_fp16")];
            int32 const_1321 = const()[name = string("const_1321"), val = int32(-1)];
            bool cat_56_interleave_0 = const()[name = string("cat_56_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_56_cast_fp16 = concat(axis = const_1321, interleave = cat_56_interleave_0, values = (neg_16_cast_fp16, slice_253_cast_fp16))[name = string("cat_56_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_202_cast_fp16 = mul(x = cat_56_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_202_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_142_cast_fp16 = add(x = mul_201_cast_fp16, y = mul_202_cast_fp16)[name = string("add_142_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_203_cast_fp16 = mul(x = mul_200_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_203_cast_fp16")];
            tensor<int32, [4]> slice_255_begin_0 = const()[name = string("slice_255_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_255_end_0 = const()[name = string("slice_255_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_255_end_mask_0 = const()[name = string("slice_255_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_255_cast_fp16 = slice_by_index(begin = slice_255_begin_0, end = slice_255_end_0, end_mask = slice_255_end_mask_0, x = mul_200_cast_fp16)[name = string("slice_255_cast_fp16")];
            tensor<int32, [4]> slice_256_begin_0 = const()[name = string("slice_256_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_256_end_0 = const()[name = string("slice_256_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_256_end_mask_0 = const()[name = string("slice_256_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_256_cast_fp16 = slice_by_index(begin = slice_256_begin_0, end = slice_256_end_0, end_mask = slice_256_end_mask_0, x = mul_200_cast_fp16)[name = string("slice_256_cast_fp16")];
            fp16 const_1328_promoted_to_fp16 = const()[name = string("const_1328_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_17_cast_fp16 = mul(x = slice_256_cast_fp16, y = const_1328_promoted_to_fp16)[name = string("neg_17_cast_fp16")];
            int32 const_1329 = const()[name = string("const_1329"), val = int32(-1)];
            bool cat_57_interleave_0 = const()[name = string("cat_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_57_cast_fp16 = concat(axis = const_1329, interleave = cat_57_interleave_0, values = (neg_17_cast_fp16, slice_255_cast_fp16))[name = string("cat_57_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_204_cast_fp16 = mul(x = cat_57_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_204_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_143_cast_fp16 = add(x = mul_203_cast_fp16, y = mul_204_cast_fp16)[name = string("add_143_cast_fp16")];
            int32 const_1330 = const()[name = string("const_1330"), val = int32(-2)];
            bool cat_58_interleave_0 = const()[name = string("cat_58_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_58_cast_fp16 = concat(axis = const_1330, interleave = cat_58_interleave_0, values = add_143_cast_fp16)[name = string("cat_58_cast_fp16")];
            int32 const_1331 = const()[name = string("const_1331"), val = int32(-2)];
            bool cat_59_interleave_0 = const()[name = string("cat_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_66_cast_fp16 = transpose(perm = transpose_66_perm_0, x = view_50_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 1, 512, 256]> cat_59_cast_fp16 = concat(axis = const_1331, interleave = cat_59_interleave_0, values = transpose_66_cast_fp16)[name = string("cat_59_cast_fp16")];
            tensor<int32, [1]> unsqueeze_111_axes_0 = const()[name = string("unsqueeze_111_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_111_cast_fp16 = expand_dims(axes = unsqueeze_111_axes_0, x = cat_58_cast_fp16)[name = string("unsqueeze_111_cast_fp16")];
            tensor<int32, [5]> expand_42_reps_0 = const()[name = string("expand_42_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_42_cast_fp16 = tile(reps = expand_42_reps_0, x = unsqueeze_111_cast_fp16)[name = string("expand_42_cast_fp16")];
            tensor<int32, [4]> const_1346 = const()[name = string("const_1346"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_51_cast_fp16 = reshape(shape = const_1346, x = expand_42_cast_fp16)[name = string("view_51_cast_fp16")];
            tensor<int32, [1]> unsqueeze_112_axes_0 = const()[name = string("unsqueeze_112_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_112_cast_fp16 = expand_dims(axes = unsqueeze_112_axes_0, x = cat_59_cast_fp16)[name = string("unsqueeze_112_cast_fp16")];
            tensor<int32, [5]> expand_43_reps_0 = const()[name = string("expand_43_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_43_cast_fp16 = tile(reps = expand_43_reps_0, x = unsqueeze_112_cast_fp16)[name = string("expand_43_cast_fp16")];
            tensor<int32, [4]> const_1361 = const()[name = string("const_1361"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_52_cast_fp16 = reshape(shape = const_1361, x = expand_43_cast_fp16)[name = string("view_52_cast_fp16")];
            bool matmul_40_transpose_x_1 = const()[name = string("matmul_40_transpose_x_1"), val = bool(false)];
            bool matmul_40_transpose_y_1 = const()[name = string("matmul_40_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_40_cast_fp16 = matmul(transpose_x = matmul_40_transpose_x_1, transpose_y = matmul_40_transpose_y_1, x = add_142_cast_fp16, y = view_51_cast_fp16)[name = string("matmul_40_cast_fp16")];
            fp16 const_1364_to_fp16 = const()[name = string("const_1364_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_205_cast_fp16 = mul(x = matmul_40_cast_fp16, y = const_1364_to_fp16)[name = string("mul_205_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_144_cast_fp16 = add(x = mul_205_cast_fp16, y = expand_cast_fp16)[name = string("add_144_cast_fp16")];
            int32 const_1374 = const()[name = string("const_1374"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_8_cast_fp16 = softmax(axis = const_1374, x = add_144_cast_fp16)[name = string("softmax_8_cast_fp16")];
            bool matmul_41_transpose_x_0 = const()[name = string("matmul_41_transpose_x_0"), val = bool(false)];
            bool matmul_41_transpose_y_0 = const()[name = string("matmul_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_41_cast_fp16 = matmul(transpose_x = matmul_41_transpose_x_0, transpose_y = matmul_41_transpose_y_0, x = softmax_8_cast_fp16, y = view_52_cast_fp16)[name = string("matmul_41_cast_fp16")];
            tensor<int32, [4]> transpose_68_perm_0 = const()[name = string("transpose_68_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_1379 = const()[name = string("const_1379"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_68_cast_fp16 = transpose(perm = transpose_68_perm_0, x = matmul_41_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 512, 768]> view_53_cast_fp16 = reshape(shape = const_1379, x = transpose_68_cast_fp16)[name = string("view_53_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237334976))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237924864))))[name = string("p_st_0_model_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_59_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = view_53_cast_fp16)[name = string("linear_59_cast_fp16")];
            fp16 const_1381_promoted_to_fp16 = const()[name = string("const_1381_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_52_cast_fp16 = pow(x = linear_59_cast_fp16, y = const_1381_promoted_to_fp16)[name = string("pow_52_cast_fp16")];
            tensor<int32, [1]> mean_51_axes_0 = const()[name = string("mean_51_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_51_keep_dims_0 = const()[name = string("mean_51_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_51_cast_fp16 = reduce_mean(axes = mean_51_axes_0, keep_dims = mean_51_keep_dims_0, x = pow_52_cast_fp16)[name = string("mean_51_cast_fp16")];
            fp16 const_1384_to_fp16 = const()[name = string("const_1384_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_145_cast_fp16 = add(x = mean_51_cast_fp16, y = const_1384_to_fp16)[name = string("add_145_cast_fp16")];
            fp32 rsqrt_51_epsilon_0 = const()[name = string("rsqrt_51_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_51_cast_fp16 = rsqrt(epsilon = rsqrt_51_epsilon_0, x = add_145_cast_fp16)[name = string("rsqrt_51_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_206_cast_fp16 = mul(x = linear_59_cast_fp16, y = rsqrt_51_cast_fp16)[name = string("mul_206_cast_fp16")];
            tensor<fp16, [768]> add_146_to_fp16 = const()[name = string("add_146_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237926464)))];
            tensor<fp16, [1, 512, 768]> mul_207_cast_fp16 = mul(x = mul_206_cast_fp16, y = add_146_to_fp16)[name = string("mul_207_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_147_cast_fp16 = add(x = add_135_cast_fp16, y = mul_207_cast_fp16)[name = string("add_147_cast_fp16")];
            fp16 const_1389_promoted_to_fp16 = const()[name = string("const_1389_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_53_cast_fp16 = pow(x = add_147_cast_fp16, y = const_1389_promoted_to_fp16)[name = string("pow_53_cast_fp16")];
            tensor<int32, [1]> mean_52_axes_0 = const()[name = string("mean_52_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_52_keep_dims_0 = const()[name = string("mean_52_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_52_cast_fp16 = reduce_mean(axes = mean_52_axes_0, keep_dims = mean_52_keep_dims_0, x = pow_53_cast_fp16)[name = string("mean_52_cast_fp16")];
            fp16 const_1392_to_fp16 = const()[name = string("const_1392_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_148_cast_fp16 = add(x = mean_52_cast_fp16, y = const_1392_to_fp16)[name = string("add_148_cast_fp16")];
            fp32 rsqrt_52_epsilon_0 = const()[name = string("rsqrt_52_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_52_cast_fp16 = rsqrt(epsilon = rsqrt_52_epsilon_0, x = add_148_cast_fp16)[name = string("rsqrt_52_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_208_cast_fp16 = mul(x = add_147_cast_fp16, y = rsqrt_52_cast_fp16)[name = string("mul_208_cast_fp16")];
            tensor<fp16, [768]> add_149_to_fp16 = const()[name = string("add_149_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237928064)))];
            tensor<fp16, [1, 512, 768]> mul_209_cast_fp16 = mul(x = mul_208_cast_fp16, y = add_149_to_fp16)[name = string("mul_209_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237929664))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238814464))))[name = string("p_st_0_model_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = mul_209_cast_fp16)[name = string("linear_60_cast_fp16")];
            string gelu_8_mode_0 = const()[name = string("gelu_8_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_8_cast_fp16 = gelu(mode = gelu_8_mode_0, x = linear_60_cast_fp16)[name = string("gelu_8_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238816832))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239701632))))[name = string("p_st_0_model_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = mul_209_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_210_cast_fp16 = mul(x = gelu_8_cast_fp16, y = linear_61_cast_fp16)[name = string("mul_210_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239704000))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240588800))))[name = string("p_st_0_model_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = mul_210_cast_fp16)[name = string("linear_62_cast_fp16")];
            fp16 const_1397_promoted_to_fp16 = const()[name = string("const_1397_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_54_cast_fp16 = pow(x = linear_62_cast_fp16, y = const_1397_promoted_to_fp16)[name = string("pow_54_cast_fp16")];
            tensor<int32, [1]> mean_53_axes_0 = const()[name = string("mean_53_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_53_keep_dims_0 = const()[name = string("mean_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_53_cast_fp16 = reduce_mean(axes = mean_53_axes_0, keep_dims = mean_53_keep_dims_0, x = pow_54_cast_fp16)[name = string("mean_53_cast_fp16")];
            fp16 const_1400_to_fp16 = const()[name = string("const_1400_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_150_cast_fp16 = add(x = mean_53_cast_fp16, y = const_1400_to_fp16)[name = string("add_150_cast_fp16")];
            fp32 rsqrt_53_epsilon_0 = const()[name = string("rsqrt_53_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_53_cast_fp16 = rsqrt(epsilon = rsqrt_53_epsilon_0, x = add_150_cast_fp16)[name = string("rsqrt_53_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_211_cast_fp16 = mul(x = linear_62_cast_fp16, y = rsqrt_53_cast_fp16)[name = string("mul_211_cast_fp16")];
            tensor<fp16, [768]> add_151_to_fp16 = const()[name = string("add_151_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240590400)))];
            tensor<fp16, [1, 512, 768]> mul_212_cast_fp16 = mul(x = mul_211_cast_fp16, y = add_151_to_fp16)[name = string("mul_212_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_152_cast_fp16 = add(x = add_147_cast_fp16, y = mul_212_cast_fp16)[name = string("add_152_cast_fp16")];
            fp16 const_1405_promoted_to_fp16 = const()[name = string("const_1405_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_55_cast_fp16 = pow(x = add_152_cast_fp16, y = const_1405_promoted_to_fp16)[name = string("pow_55_cast_fp16")];
            tensor<int32, [1]> mean_54_axes_0 = const()[name = string("mean_54_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_54_keep_dims_0 = const()[name = string("mean_54_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_54_cast_fp16 = reduce_mean(axes = mean_54_axes_0, keep_dims = mean_54_keep_dims_0, x = pow_55_cast_fp16)[name = string("mean_54_cast_fp16")];
            fp16 const_1408_to_fp16 = const()[name = string("const_1408_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_153_cast_fp16 = add(x = mean_54_cast_fp16, y = const_1408_to_fp16)[name = string("add_153_cast_fp16")];
            fp32 rsqrt_54_epsilon_0 = const()[name = string("rsqrt_54_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_54_cast_fp16 = rsqrt(epsilon = rsqrt_54_epsilon_0, x = add_153_cast_fp16)[name = string("rsqrt_54_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_213_cast_fp16 = mul(x = add_152_cast_fp16, y = rsqrt_54_cast_fp16)[name = string("mul_213_cast_fp16")];
            tensor<fp16, [768]> add_154_to_fp16 = const()[name = string("add_154_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240592000)))];
            tensor<fp16, [1, 512, 768]> mul_214_cast_fp16 = mul(x = mul_213_cast_fp16, y = add_154_to_fp16)[name = string("mul_214_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240593600))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241183488))))[name = string("p_st_0_model_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_63_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = mul_214_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> const_1412 = const()[name = string("const_1412"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_54_cast_fp16 = reshape(shape = const_1412, x = linear_63_cast_fp16)[name = string("view_54_cast_fp16")];
            tensor<int32, [4]> transpose_69_perm_0 = const()[name = string("transpose_69_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241185088))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241381760))))[name = string("p_st_0_model_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = mul_214_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> const_1415 = const()[name = string("const_1415"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_55_cast_fp16 = reshape(shape = const_1415, x = linear_64_cast_fp16)[name = string("view_55_cast_fp16")];
            tensor<int32, [4]> transpose_70_perm_0 = const()[name = string("transpose_70_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241382336))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241579008))))[name = string("p_st_0_model_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = mul_214_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> const_1418 = const()[name = string("const_1418"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_56_cast_fp16 = reshape(shape = const_1418, x = linear_65_cast_fp16)[name = string("view_56_cast_fp16")];
            tensor<int32, [4]> transpose_71_perm_0 = const()[name = string("transpose_71_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_1422_promoted_to_fp16 = const()[name = string("const_1422_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_69_cast_fp16 = transpose(perm = transpose_69_perm_0, x = view_54_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 3, 512, 256]> pow_56_cast_fp16 = pow(x = transpose_69_cast_fp16, y = const_1422_promoted_to_fp16)[name = string("pow_56_cast_fp16")];
            tensor<int32, [1]> mean_55_axes_0 = const()[name = string("mean_55_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_55_keep_dims_0 = const()[name = string("mean_55_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_55_cast_fp16 = reduce_mean(axes = mean_55_axes_0, keep_dims = mean_55_keep_dims_0, x = pow_56_cast_fp16)[name = string("mean_55_cast_fp16")];
            fp16 const_1425_to_fp16 = const()[name = string("const_1425_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_155_cast_fp16 = add(x = mean_55_cast_fp16, y = const_1425_to_fp16)[name = string("add_155_cast_fp16")];
            fp32 rsqrt_55_epsilon_0 = const()[name = string("rsqrt_55_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_55_cast_fp16 = rsqrt(epsilon = rsqrt_55_epsilon_0, x = add_155_cast_fp16)[name = string("rsqrt_55_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_215_cast_fp16 = mul(x = transpose_69_cast_fp16, y = rsqrt_55_cast_fp16)[name = string("mul_215_cast_fp16")];
            tensor<fp16, [256]> add_156_to_fp16 = const()[name = string("add_156_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241579584)))];
            tensor<fp16, [1, 3, 512, 256]> mul_216_cast_fp16 = mul(x = mul_215_cast_fp16, y = add_156_to_fp16)[name = string("mul_216_cast_fp16")];
            fp16 const_1430_promoted_to_fp16 = const()[name = string("const_1430_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_70_cast_fp16 = transpose(perm = transpose_70_perm_0, x = view_55_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 1, 512, 256]> pow_57_cast_fp16 = pow(x = transpose_70_cast_fp16, y = const_1430_promoted_to_fp16)[name = string("pow_57_cast_fp16")];
            tensor<int32, [1]> mean_56_axes_0 = const()[name = string("mean_56_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_56_keep_dims_0 = const()[name = string("mean_56_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_56_cast_fp16 = reduce_mean(axes = mean_56_axes_0, keep_dims = mean_56_keep_dims_0, x = pow_57_cast_fp16)[name = string("mean_56_cast_fp16")];
            fp16 const_1433_to_fp16 = const()[name = string("const_1433_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_157_cast_fp16 = add(x = mean_56_cast_fp16, y = const_1433_to_fp16)[name = string("add_157_cast_fp16")];
            fp32 rsqrt_56_epsilon_0 = const()[name = string("rsqrt_56_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_56_cast_fp16 = rsqrt(epsilon = rsqrt_56_epsilon_0, x = add_157_cast_fp16)[name = string("rsqrt_56_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_217_cast_fp16 = mul(x = transpose_70_cast_fp16, y = rsqrt_56_cast_fp16)[name = string("mul_217_cast_fp16")];
            tensor<fp16, [256]> add_158_to_fp16 = const()[name = string("add_158_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241580160)))];
            tensor<fp16, [1, 1, 512, 256]> mul_218_cast_fp16 = mul(x = mul_217_cast_fp16, y = add_158_to_fp16)[name = string("mul_218_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_219_cast_fp16 = mul(x = mul_216_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_219_cast_fp16")];
            tensor<int32, [4]> slice_276_begin_0 = const()[name = string("slice_276_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_276_end_0 = const()[name = string("slice_276_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_276_end_mask_0 = const()[name = string("slice_276_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_276_cast_fp16 = slice_by_index(begin = slice_276_begin_0, end = slice_276_end_0, end_mask = slice_276_end_mask_0, x = mul_216_cast_fp16)[name = string("slice_276_cast_fp16")];
            tensor<int32, [4]> slice_277_begin_0 = const()[name = string("slice_277_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_277_end_0 = const()[name = string("slice_277_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_277_end_mask_0 = const()[name = string("slice_277_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_277_cast_fp16 = slice_by_index(begin = slice_277_begin_0, end = slice_277_end_0, end_mask = slice_277_end_mask_0, x = mul_216_cast_fp16)[name = string("slice_277_cast_fp16")];
            fp16 const_1445_promoted_to_fp16 = const()[name = string("const_1445_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_18_cast_fp16 = mul(x = slice_277_cast_fp16, y = const_1445_promoted_to_fp16)[name = string("neg_18_cast_fp16")];
            int32 const_1446 = const()[name = string("const_1446"), val = int32(-1)];
            bool cat_60_interleave_0 = const()[name = string("cat_60_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_60_cast_fp16 = concat(axis = const_1446, interleave = cat_60_interleave_0, values = (neg_18_cast_fp16, slice_276_cast_fp16))[name = string("cat_60_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_220_cast_fp16 = mul(x = cat_60_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_220_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_159_cast_fp16 = add(x = mul_219_cast_fp16, y = mul_220_cast_fp16)[name = string("add_159_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_221_cast_fp16 = mul(x = mul_218_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_221_cast_fp16")];
            tensor<int32, [4]> slice_278_begin_0 = const()[name = string("slice_278_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_278_end_0 = const()[name = string("slice_278_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_278_end_mask_0 = const()[name = string("slice_278_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_278_cast_fp16 = slice_by_index(begin = slice_278_begin_0, end = slice_278_end_0, end_mask = slice_278_end_mask_0, x = mul_218_cast_fp16)[name = string("slice_278_cast_fp16")];
            tensor<int32, [4]> slice_279_begin_0 = const()[name = string("slice_279_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_279_end_0 = const()[name = string("slice_279_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_279_end_mask_0 = const()[name = string("slice_279_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_279_cast_fp16 = slice_by_index(begin = slice_279_begin_0, end = slice_279_end_0, end_mask = slice_279_end_mask_0, x = mul_218_cast_fp16)[name = string("slice_279_cast_fp16")];
            fp16 const_1453_promoted_to_fp16 = const()[name = string("const_1453_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_19_cast_fp16 = mul(x = slice_279_cast_fp16, y = const_1453_promoted_to_fp16)[name = string("neg_19_cast_fp16")];
            int32 const_1454 = const()[name = string("const_1454"), val = int32(-1)];
            bool cat_61_interleave_0 = const()[name = string("cat_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_61_cast_fp16 = concat(axis = const_1454, interleave = cat_61_interleave_0, values = (neg_19_cast_fp16, slice_278_cast_fp16))[name = string("cat_61_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_222_cast_fp16 = mul(x = cat_61_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_222_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_160_cast_fp16 = add(x = mul_221_cast_fp16, y = mul_222_cast_fp16)[name = string("add_160_cast_fp16")];
            int32 const_1455 = const()[name = string("const_1455"), val = int32(-2)];
            bool cat_62_interleave_0 = const()[name = string("cat_62_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_62_cast_fp16 = concat(axis = const_1455, interleave = cat_62_interleave_0, values = add_160_cast_fp16)[name = string("cat_62_cast_fp16")];
            int32 const_1456 = const()[name = string("const_1456"), val = int32(-2)];
            bool cat_63_interleave_0 = const()[name = string("cat_63_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_71_cast_fp16 = transpose(perm = transpose_71_perm_0, x = view_56_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 1, 512, 256]> cat_63_cast_fp16 = concat(axis = const_1456, interleave = cat_63_interleave_0, values = transpose_71_cast_fp16)[name = string("cat_63_cast_fp16")];
            tensor<int32, [1]> unsqueeze_115_axes_0 = const()[name = string("unsqueeze_115_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_115_cast_fp16 = expand_dims(axes = unsqueeze_115_axes_0, x = cat_62_cast_fp16)[name = string("unsqueeze_115_cast_fp16")];
            tensor<int32, [5]> expand_44_reps_0 = const()[name = string("expand_44_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_44_cast_fp16 = tile(reps = expand_44_reps_0, x = unsqueeze_115_cast_fp16)[name = string("expand_44_cast_fp16")];
            tensor<int32, [4]> const_1471 = const()[name = string("const_1471"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_57_cast_fp16 = reshape(shape = const_1471, x = expand_44_cast_fp16)[name = string("view_57_cast_fp16")];
            tensor<int32, [1]> unsqueeze_116_axes_0 = const()[name = string("unsqueeze_116_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_116_cast_fp16 = expand_dims(axes = unsqueeze_116_axes_0, x = cat_63_cast_fp16)[name = string("unsqueeze_116_cast_fp16")];
            tensor<int32, [5]> expand_45_reps_0 = const()[name = string("expand_45_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_45_cast_fp16 = tile(reps = expand_45_reps_0, x = unsqueeze_116_cast_fp16)[name = string("expand_45_cast_fp16")];
            tensor<int32, [4]> const_1486 = const()[name = string("const_1486"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_58_cast_fp16 = reshape(shape = const_1486, x = expand_45_cast_fp16)[name = string("view_58_cast_fp16")];
            bool matmul_42_transpose_x_1 = const()[name = string("matmul_42_transpose_x_1"), val = bool(false)];
            bool matmul_42_transpose_y_1 = const()[name = string("matmul_42_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_42_cast_fp16 = matmul(transpose_x = matmul_42_transpose_x_1, transpose_y = matmul_42_transpose_y_1, x = add_159_cast_fp16, y = view_57_cast_fp16)[name = string("matmul_42_cast_fp16")];
            fp16 const_1489_to_fp16 = const()[name = string("const_1489_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_223_cast_fp16 = mul(x = matmul_42_cast_fp16, y = const_1489_to_fp16)[name = string("mul_223_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_161_cast_fp16 = add(x = mul_223_cast_fp16, y = expand_cast_fp16)[name = string("add_161_cast_fp16")];
            int32 const_1499 = const()[name = string("const_1499"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_9_cast_fp16 = softmax(axis = const_1499, x = add_161_cast_fp16)[name = string("softmax_9_cast_fp16")];
            bool matmul_43_transpose_x_0 = const()[name = string("matmul_43_transpose_x_0"), val = bool(false)];
            bool matmul_43_transpose_y_0 = const()[name = string("matmul_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_43_cast_fp16 = matmul(transpose_x = matmul_43_transpose_x_0, transpose_y = matmul_43_transpose_y_0, x = softmax_9_cast_fp16, y = view_58_cast_fp16)[name = string("matmul_43_cast_fp16")];
            tensor<int32, [4]> transpose_73_perm_0 = const()[name = string("transpose_73_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_1504 = const()[name = string("const_1504"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_73_cast_fp16 = transpose(perm = transpose_73_perm_0, x = matmul_43_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 512, 768]> view_59_cast_fp16 = reshape(shape = const_1504, x = transpose_73_cast_fp16)[name = string("view_59_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241580736))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242170624))))[name = string("p_st_0_model_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = view_59_cast_fp16)[name = string("linear_66_cast_fp16")];
            fp16 const_1506_promoted_to_fp16 = const()[name = string("const_1506_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_58_cast_fp16 = pow(x = linear_66_cast_fp16, y = const_1506_promoted_to_fp16)[name = string("pow_58_cast_fp16")];
            tensor<int32, [1]> mean_57_axes_0 = const()[name = string("mean_57_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_57_keep_dims_0 = const()[name = string("mean_57_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_57_cast_fp16 = reduce_mean(axes = mean_57_axes_0, keep_dims = mean_57_keep_dims_0, x = pow_58_cast_fp16)[name = string("mean_57_cast_fp16")];
            fp16 const_1509_to_fp16 = const()[name = string("const_1509_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_162_cast_fp16 = add(x = mean_57_cast_fp16, y = const_1509_to_fp16)[name = string("add_162_cast_fp16")];
            fp32 rsqrt_57_epsilon_0 = const()[name = string("rsqrt_57_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_57_cast_fp16 = rsqrt(epsilon = rsqrt_57_epsilon_0, x = add_162_cast_fp16)[name = string("rsqrt_57_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_224_cast_fp16 = mul(x = linear_66_cast_fp16, y = rsqrt_57_cast_fp16)[name = string("mul_224_cast_fp16")];
            tensor<fp16, [768]> add_163_to_fp16 = const()[name = string("add_163_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242172224)))];
            tensor<fp16, [1, 512, 768]> mul_225_cast_fp16 = mul(x = mul_224_cast_fp16, y = add_163_to_fp16)[name = string("mul_225_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_164_cast_fp16 = add(x = add_152_cast_fp16, y = mul_225_cast_fp16)[name = string("add_164_cast_fp16")];
            fp16 const_1514_promoted_to_fp16 = const()[name = string("const_1514_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_59_cast_fp16 = pow(x = add_164_cast_fp16, y = const_1514_promoted_to_fp16)[name = string("pow_59_cast_fp16")];
            tensor<int32, [1]> mean_58_axes_0 = const()[name = string("mean_58_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_58_keep_dims_0 = const()[name = string("mean_58_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_58_cast_fp16 = reduce_mean(axes = mean_58_axes_0, keep_dims = mean_58_keep_dims_0, x = pow_59_cast_fp16)[name = string("mean_58_cast_fp16")];
            fp16 const_1517_to_fp16 = const()[name = string("const_1517_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_165_cast_fp16 = add(x = mean_58_cast_fp16, y = const_1517_to_fp16)[name = string("add_165_cast_fp16")];
            fp32 rsqrt_58_epsilon_0 = const()[name = string("rsqrt_58_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_58_cast_fp16 = rsqrt(epsilon = rsqrt_58_epsilon_0, x = add_165_cast_fp16)[name = string("rsqrt_58_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_226_cast_fp16 = mul(x = add_164_cast_fp16, y = rsqrt_58_cast_fp16)[name = string("mul_226_cast_fp16")];
            tensor<fp16, [768]> add_166_to_fp16 = const()[name = string("add_166_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242173824)))];
            tensor<fp16, [1, 512, 768]> mul_227_cast_fp16 = mul(x = mul_226_cast_fp16, y = add_166_to_fp16)[name = string("mul_227_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242175424))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243060224))))[name = string("p_st_0_model_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = mul_227_cast_fp16)[name = string("linear_67_cast_fp16")];
            string gelu_9_mode_0 = const()[name = string("gelu_9_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_9_cast_fp16 = gelu(mode = gelu_9_mode_0, x = linear_67_cast_fp16)[name = string("gelu_9_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243062592))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243947392))))[name = string("p_st_0_model_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = mul_227_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_228_cast_fp16 = mul(x = gelu_9_cast_fp16, y = linear_68_cast_fp16)[name = string("mul_228_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243949760))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244834560))))[name = string("p_st_0_model_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_69_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = mul_228_cast_fp16)[name = string("linear_69_cast_fp16")];
            fp16 const_1522_promoted_to_fp16 = const()[name = string("const_1522_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_60_cast_fp16 = pow(x = linear_69_cast_fp16, y = const_1522_promoted_to_fp16)[name = string("pow_60_cast_fp16")];
            tensor<int32, [1]> mean_59_axes_0 = const()[name = string("mean_59_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_59_keep_dims_0 = const()[name = string("mean_59_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_59_cast_fp16 = reduce_mean(axes = mean_59_axes_0, keep_dims = mean_59_keep_dims_0, x = pow_60_cast_fp16)[name = string("mean_59_cast_fp16")];
            fp16 const_1525_to_fp16 = const()[name = string("const_1525_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_167_cast_fp16 = add(x = mean_59_cast_fp16, y = const_1525_to_fp16)[name = string("add_167_cast_fp16")];
            fp32 rsqrt_59_epsilon_0 = const()[name = string("rsqrt_59_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_59_cast_fp16 = rsqrt(epsilon = rsqrt_59_epsilon_0, x = add_167_cast_fp16)[name = string("rsqrt_59_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_229_cast_fp16 = mul(x = linear_69_cast_fp16, y = rsqrt_59_cast_fp16)[name = string("mul_229_cast_fp16")];
            tensor<fp16, [768]> add_168_to_fp16 = const()[name = string("add_168_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244836160)))];
            tensor<fp16, [1, 512, 768]> mul_230_cast_fp16 = mul(x = mul_229_cast_fp16, y = add_168_to_fp16)[name = string("mul_230_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_169_cast_fp16 = add(x = add_164_cast_fp16, y = mul_230_cast_fp16)[name = string("add_169_cast_fp16")];
            fp16 const_1530_promoted_to_fp16 = const()[name = string("const_1530_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_61_cast_fp16 = pow(x = add_169_cast_fp16, y = const_1530_promoted_to_fp16)[name = string("pow_61_cast_fp16")];
            tensor<int32, [1]> mean_60_axes_0 = const()[name = string("mean_60_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_60_keep_dims_0 = const()[name = string("mean_60_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_60_cast_fp16 = reduce_mean(axes = mean_60_axes_0, keep_dims = mean_60_keep_dims_0, x = pow_61_cast_fp16)[name = string("mean_60_cast_fp16")];
            fp16 const_1533_to_fp16 = const()[name = string("const_1533_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_170_cast_fp16 = add(x = mean_60_cast_fp16, y = const_1533_to_fp16)[name = string("add_170_cast_fp16")];
            fp32 rsqrt_60_epsilon_0 = const()[name = string("rsqrt_60_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_60_cast_fp16 = rsqrt(epsilon = rsqrt_60_epsilon_0, x = add_170_cast_fp16)[name = string("rsqrt_60_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_231_cast_fp16 = mul(x = add_169_cast_fp16, y = rsqrt_60_cast_fp16)[name = string("mul_231_cast_fp16")];
            tensor<fp16, [768]> add_171_to_fp16 = const()[name = string("add_171_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244837760)))];
            tensor<fp16, [1, 512, 768]> mul_232_cast_fp16 = mul(x = mul_231_cast_fp16, y = add_171_to_fp16)[name = string("mul_232_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244839360))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245429248))))[name = string("p_st_0_model_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = mul_232_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [4]> const_1537 = const()[name = string("const_1537"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_60_cast_fp16 = reshape(shape = const_1537, x = linear_70_cast_fp16)[name = string("view_60_cast_fp16")];
            tensor<int32, [4]> transpose_74_perm_0 = const()[name = string("transpose_74_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245430848))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245627520))))[name = string("p_st_0_model_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = mul_232_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> const_1540 = const()[name = string("const_1540"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_61_cast_fp16 = reshape(shape = const_1540, x = linear_71_cast_fp16)[name = string("view_61_cast_fp16")];
            tensor<int32, [4]> transpose_75_perm_0 = const()[name = string("transpose_75_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245628096))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245824768))))[name = string("p_st_0_model_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_72_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = mul_232_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> const_1543 = const()[name = string("const_1543"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_62_cast_fp16 = reshape(shape = const_1543, x = linear_72_cast_fp16)[name = string("view_62_cast_fp16")];
            tensor<int32, [4]> transpose_76_perm_0 = const()[name = string("transpose_76_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_1547_promoted_to_fp16 = const()[name = string("const_1547_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_74_cast_fp16 = transpose(perm = transpose_74_perm_0, x = view_60_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 3, 512, 256]> pow_62_cast_fp16 = pow(x = transpose_74_cast_fp16, y = const_1547_promoted_to_fp16)[name = string("pow_62_cast_fp16")];
            tensor<int32, [1]> mean_61_axes_0 = const()[name = string("mean_61_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_61_keep_dims_0 = const()[name = string("mean_61_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_61_cast_fp16 = reduce_mean(axes = mean_61_axes_0, keep_dims = mean_61_keep_dims_0, x = pow_62_cast_fp16)[name = string("mean_61_cast_fp16")];
            fp16 const_1550_to_fp16 = const()[name = string("const_1550_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_172_cast_fp16 = add(x = mean_61_cast_fp16, y = const_1550_to_fp16)[name = string("add_172_cast_fp16")];
            fp32 rsqrt_61_epsilon_0 = const()[name = string("rsqrt_61_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_61_cast_fp16 = rsqrt(epsilon = rsqrt_61_epsilon_0, x = add_172_cast_fp16)[name = string("rsqrt_61_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_233_cast_fp16 = mul(x = transpose_74_cast_fp16, y = rsqrt_61_cast_fp16)[name = string("mul_233_cast_fp16")];
            tensor<fp16, [256]> add_173_to_fp16 = const()[name = string("add_173_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245825344)))];
            tensor<fp16, [1, 3, 512, 256]> mul_234_cast_fp16 = mul(x = mul_233_cast_fp16, y = add_173_to_fp16)[name = string("mul_234_cast_fp16")];
            fp16 const_1555_promoted_to_fp16 = const()[name = string("const_1555_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_75_cast_fp16 = transpose(perm = transpose_75_perm_0, x = view_61_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 1, 512, 256]> pow_63_cast_fp16 = pow(x = transpose_75_cast_fp16, y = const_1555_promoted_to_fp16)[name = string("pow_63_cast_fp16")];
            tensor<int32, [1]> mean_62_axes_0 = const()[name = string("mean_62_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_62_keep_dims_0 = const()[name = string("mean_62_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_62_cast_fp16 = reduce_mean(axes = mean_62_axes_0, keep_dims = mean_62_keep_dims_0, x = pow_63_cast_fp16)[name = string("mean_62_cast_fp16")];
            fp16 const_1558_to_fp16 = const()[name = string("const_1558_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_174_cast_fp16 = add(x = mean_62_cast_fp16, y = const_1558_to_fp16)[name = string("add_174_cast_fp16")];
            fp32 rsqrt_62_epsilon_0 = const()[name = string("rsqrt_62_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_62_cast_fp16 = rsqrt(epsilon = rsqrt_62_epsilon_0, x = add_174_cast_fp16)[name = string("rsqrt_62_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_235_cast_fp16 = mul(x = transpose_75_cast_fp16, y = rsqrt_62_cast_fp16)[name = string("mul_235_cast_fp16")];
            tensor<fp16, [256]> add_175_to_fp16 = const()[name = string("add_175_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245825920)))];
            tensor<fp16, [1, 1, 512, 256]> mul_236_cast_fp16 = mul(x = mul_235_cast_fp16, y = add_175_to_fp16)[name = string("mul_236_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_237_cast_fp16 = mul(x = mul_234_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_237_cast_fp16")];
            tensor<int32, [4]> slice_299_begin_0 = const()[name = string("slice_299_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_299_end_0 = const()[name = string("slice_299_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_299_end_mask_0 = const()[name = string("slice_299_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_299_cast_fp16 = slice_by_index(begin = slice_299_begin_0, end = slice_299_end_0, end_mask = slice_299_end_mask_0, x = mul_234_cast_fp16)[name = string("slice_299_cast_fp16")];
            tensor<int32, [4]> slice_300_begin_0 = const()[name = string("slice_300_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_300_end_0 = const()[name = string("slice_300_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_300_end_mask_0 = const()[name = string("slice_300_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_300_cast_fp16 = slice_by_index(begin = slice_300_begin_0, end = slice_300_end_0, end_mask = slice_300_end_mask_0, x = mul_234_cast_fp16)[name = string("slice_300_cast_fp16")];
            fp16 const_1570_promoted_to_fp16 = const()[name = string("const_1570_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_20_cast_fp16 = mul(x = slice_300_cast_fp16, y = const_1570_promoted_to_fp16)[name = string("neg_20_cast_fp16")];
            int32 const_1571 = const()[name = string("const_1571"), val = int32(-1)];
            bool cat_64_interleave_0 = const()[name = string("cat_64_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_64_cast_fp16 = concat(axis = const_1571, interleave = cat_64_interleave_0, values = (neg_20_cast_fp16, slice_299_cast_fp16))[name = string("cat_64_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_238_cast_fp16 = mul(x = cat_64_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_238_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_176_cast_fp16 = add(x = mul_237_cast_fp16, y = mul_238_cast_fp16)[name = string("add_176_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_239_cast_fp16 = mul(x = mul_236_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_239_cast_fp16")];
            tensor<int32, [4]> slice_301_begin_0 = const()[name = string("slice_301_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_301_end_0 = const()[name = string("slice_301_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_301_end_mask_0 = const()[name = string("slice_301_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_301_cast_fp16 = slice_by_index(begin = slice_301_begin_0, end = slice_301_end_0, end_mask = slice_301_end_mask_0, x = mul_236_cast_fp16)[name = string("slice_301_cast_fp16")];
            tensor<int32, [4]> slice_302_begin_0 = const()[name = string("slice_302_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_302_end_0 = const()[name = string("slice_302_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_302_end_mask_0 = const()[name = string("slice_302_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_302_cast_fp16 = slice_by_index(begin = slice_302_begin_0, end = slice_302_end_0, end_mask = slice_302_end_mask_0, x = mul_236_cast_fp16)[name = string("slice_302_cast_fp16")];
            fp16 const_1578_promoted_to_fp16 = const()[name = string("const_1578_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_21_cast_fp16 = mul(x = slice_302_cast_fp16, y = const_1578_promoted_to_fp16)[name = string("neg_21_cast_fp16")];
            int32 const_1579 = const()[name = string("const_1579"), val = int32(-1)];
            bool cat_65_interleave_0 = const()[name = string("cat_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_65_cast_fp16 = concat(axis = const_1579, interleave = cat_65_interleave_0, values = (neg_21_cast_fp16, slice_301_cast_fp16))[name = string("cat_65_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_240_cast_fp16 = mul(x = cat_65_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_240_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_177_cast_fp16 = add(x = mul_239_cast_fp16, y = mul_240_cast_fp16)[name = string("add_177_cast_fp16")];
            int32 const_1580 = const()[name = string("const_1580"), val = int32(-2)];
            bool cat_66_interleave_0 = const()[name = string("cat_66_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_66_cast_fp16 = concat(axis = const_1580, interleave = cat_66_interleave_0, values = add_177_cast_fp16)[name = string("cat_66_cast_fp16")];
            int32 const_1581 = const()[name = string("const_1581"), val = int32(-2)];
            bool cat_67_interleave_0 = const()[name = string("cat_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_76_cast_fp16 = transpose(perm = transpose_76_perm_0, x = view_62_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 1, 512, 256]> cat_67_cast_fp16 = concat(axis = const_1581, interleave = cat_67_interleave_0, values = transpose_76_cast_fp16)[name = string("cat_67_cast_fp16")];
            tensor<int32, [1]> unsqueeze_119_axes_0 = const()[name = string("unsqueeze_119_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_119_cast_fp16 = expand_dims(axes = unsqueeze_119_axes_0, x = cat_66_cast_fp16)[name = string("unsqueeze_119_cast_fp16")];
            tensor<int32, [5]> expand_46_reps_0 = const()[name = string("expand_46_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_46_cast_fp16 = tile(reps = expand_46_reps_0, x = unsqueeze_119_cast_fp16)[name = string("expand_46_cast_fp16")];
            tensor<int32, [4]> const_1596 = const()[name = string("const_1596"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_63_cast_fp16 = reshape(shape = const_1596, x = expand_46_cast_fp16)[name = string("view_63_cast_fp16")];
            tensor<int32, [1]> unsqueeze_120_axes_0 = const()[name = string("unsqueeze_120_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_120_cast_fp16 = expand_dims(axes = unsqueeze_120_axes_0, x = cat_67_cast_fp16)[name = string("unsqueeze_120_cast_fp16")];
            tensor<int32, [5]> expand_47_reps_0 = const()[name = string("expand_47_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_47_cast_fp16 = tile(reps = expand_47_reps_0, x = unsqueeze_120_cast_fp16)[name = string("expand_47_cast_fp16")];
            tensor<int32, [4]> const_1611 = const()[name = string("const_1611"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_64_cast_fp16 = reshape(shape = const_1611, x = expand_47_cast_fp16)[name = string("view_64_cast_fp16")];
            bool matmul_44_transpose_x_1 = const()[name = string("matmul_44_transpose_x_1"), val = bool(false)];
            bool matmul_44_transpose_y_1 = const()[name = string("matmul_44_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_44_cast_fp16 = matmul(transpose_x = matmul_44_transpose_x_1, transpose_y = matmul_44_transpose_y_1, x = add_176_cast_fp16, y = view_63_cast_fp16)[name = string("matmul_44_cast_fp16")];
            fp16 const_1614_to_fp16 = const()[name = string("const_1614_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_241_cast_fp16 = mul(x = matmul_44_cast_fp16, y = const_1614_to_fp16)[name = string("mul_241_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_178_cast_fp16 = add(x = mul_241_cast_fp16, y = expand_cast_fp16)[name = string("add_178_cast_fp16")];
            int32 const_1624 = const()[name = string("const_1624"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_10_cast_fp16 = softmax(axis = const_1624, x = add_178_cast_fp16)[name = string("softmax_10_cast_fp16")];
            bool matmul_45_transpose_x_0 = const()[name = string("matmul_45_transpose_x_0"), val = bool(false)];
            bool matmul_45_transpose_y_0 = const()[name = string("matmul_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_45_cast_fp16 = matmul(transpose_x = matmul_45_transpose_x_0, transpose_y = matmul_45_transpose_y_0, x = softmax_10_cast_fp16, y = view_64_cast_fp16)[name = string("matmul_45_cast_fp16")];
            tensor<int32, [4]> transpose_78_perm_0 = const()[name = string("transpose_78_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_1629 = const()[name = string("const_1629"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_78_cast_fp16 = transpose(perm = transpose_78_perm_0, x = matmul_45_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 512, 768]> view_65_cast_fp16 = reshape(shape = const_1629, x = transpose_78_cast_fp16)[name = string("view_65_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245826496))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246416384))))[name = string("p_st_0_model_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_73_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = view_65_cast_fp16)[name = string("linear_73_cast_fp16")];
            fp16 const_1631_promoted_to_fp16 = const()[name = string("const_1631_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_64_cast_fp16 = pow(x = linear_73_cast_fp16, y = const_1631_promoted_to_fp16)[name = string("pow_64_cast_fp16")];
            tensor<int32, [1]> mean_63_axes_0 = const()[name = string("mean_63_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_63_keep_dims_0 = const()[name = string("mean_63_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_63_cast_fp16 = reduce_mean(axes = mean_63_axes_0, keep_dims = mean_63_keep_dims_0, x = pow_64_cast_fp16)[name = string("mean_63_cast_fp16")];
            fp16 const_1634_to_fp16 = const()[name = string("const_1634_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_179_cast_fp16 = add(x = mean_63_cast_fp16, y = const_1634_to_fp16)[name = string("add_179_cast_fp16")];
            fp32 rsqrt_63_epsilon_0 = const()[name = string("rsqrt_63_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_63_cast_fp16 = rsqrt(epsilon = rsqrt_63_epsilon_0, x = add_179_cast_fp16)[name = string("rsqrt_63_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_242_cast_fp16 = mul(x = linear_73_cast_fp16, y = rsqrt_63_cast_fp16)[name = string("mul_242_cast_fp16")];
            tensor<fp16, [768]> add_180_to_fp16 = const()[name = string("add_180_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246417984)))];
            tensor<fp16, [1, 512, 768]> mul_243_cast_fp16 = mul(x = mul_242_cast_fp16, y = add_180_to_fp16)[name = string("mul_243_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_181_cast_fp16 = add(x = add_169_cast_fp16, y = mul_243_cast_fp16)[name = string("add_181_cast_fp16")];
            fp16 const_1639_promoted_to_fp16 = const()[name = string("const_1639_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_65_cast_fp16 = pow(x = add_181_cast_fp16, y = const_1639_promoted_to_fp16)[name = string("pow_65_cast_fp16")];
            tensor<int32, [1]> mean_64_axes_0 = const()[name = string("mean_64_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_64_keep_dims_0 = const()[name = string("mean_64_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_64_cast_fp16 = reduce_mean(axes = mean_64_axes_0, keep_dims = mean_64_keep_dims_0, x = pow_65_cast_fp16)[name = string("mean_64_cast_fp16")];
            fp16 const_1642_to_fp16 = const()[name = string("const_1642_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_182_cast_fp16 = add(x = mean_64_cast_fp16, y = const_1642_to_fp16)[name = string("add_182_cast_fp16")];
            fp32 rsqrt_64_epsilon_0 = const()[name = string("rsqrt_64_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_64_cast_fp16 = rsqrt(epsilon = rsqrt_64_epsilon_0, x = add_182_cast_fp16)[name = string("rsqrt_64_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_244_cast_fp16 = mul(x = add_181_cast_fp16, y = rsqrt_64_cast_fp16)[name = string("mul_244_cast_fp16")];
            tensor<fp16, [768]> add_183_to_fp16 = const()[name = string("add_183_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246419584)))];
            tensor<fp16, [1, 512, 768]> mul_245_cast_fp16 = mul(x = mul_244_cast_fp16, y = add_183_to_fp16)[name = string("mul_245_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246421184))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247305984))))[name = string("p_st_0_model_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = mul_245_cast_fp16)[name = string("linear_74_cast_fp16")];
            string gelu_10_mode_0 = const()[name = string("gelu_10_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_10_cast_fp16 = gelu(mode = gelu_10_mode_0, x = linear_74_cast_fp16)[name = string("gelu_10_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247308352))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248193152))))[name = string("p_st_0_model_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = mul_245_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_246_cast_fp16 = mul(x = gelu_10_cast_fp16, y = linear_75_cast_fp16)[name = string("mul_246_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248195520))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249080320))))[name = string("p_st_0_model_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = mul_246_cast_fp16)[name = string("linear_76_cast_fp16")];
            fp16 const_1647_promoted_to_fp16 = const()[name = string("const_1647_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_66_cast_fp16 = pow(x = linear_76_cast_fp16, y = const_1647_promoted_to_fp16)[name = string("pow_66_cast_fp16")];
            tensor<int32, [1]> mean_65_axes_0 = const()[name = string("mean_65_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_65_keep_dims_0 = const()[name = string("mean_65_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_65_cast_fp16 = reduce_mean(axes = mean_65_axes_0, keep_dims = mean_65_keep_dims_0, x = pow_66_cast_fp16)[name = string("mean_65_cast_fp16")];
            fp16 const_1650_to_fp16 = const()[name = string("const_1650_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_184_cast_fp16 = add(x = mean_65_cast_fp16, y = const_1650_to_fp16)[name = string("add_184_cast_fp16")];
            fp32 rsqrt_65_epsilon_0 = const()[name = string("rsqrt_65_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_65_cast_fp16 = rsqrt(epsilon = rsqrt_65_epsilon_0, x = add_184_cast_fp16)[name = string("rsqrt_65_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_247_cast_fp16 = mul(x = linear_76_cast_fp16, y = rsqrt_65_cast_fp16)[name = string("mul_247_cast_fp16")];
            tensor<fp16, [768]> add_185_to_fp16 = const()[name = string("add_185_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249081920)))];
            tensor<fp16, [1, 512, 768]> mul_248_cast_fp16 = mul(x = mul_247_cast_fp16, y = add_185_to_fp16)[name = string("mul_248_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_186_cast_fp16 = add(x = add_181_cast_fp16, y = mul_248_cast_fp16)[name = string("add_186_cast_fp16")];
            fp16 const_1655_promoted_to_fp16 = const()[name = string("const_1655_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_67_cast_fp16 = pow(x = add_186_cast_fp16, y = const_1655_promoted_to_fp16)[name = string("pow_67_cast_fp16")];
            tensor<int32, [1]> mean_66_axes_0 = const()[name = string("mean_66_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_66_keep_dims_0 = const()[name = string("mean_66_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_66_cast_fp16 = reduce_mean(axes = mean_66_axes_0, keep_dims = mean_66_keep_dims_0, x = pow_67_cast_fp16)[name = string("mean_66_cast_fp16")];
            fp16 const_1658_to_fp16 = const()[name = string("const_1658_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_187_cast_fp16 = add(x = mean_66_cast_fp16, y = const_1658_to_fp16)[name = string("add_187_cast_fp16")];
            fp32 rsqrt_66_epsilon_0 = const()[name = string("rsqrt_66_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_66_cast_fp16 = rsqrt(epsilon = rsqrt_66_epsilon_0, x = add_187_cast_fp16)[name = string("rsqrt_66_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_249_cast_fp16 = mul(x = add_186_cast_fp16, y = rsqrt_66_cast_fp16)[name = string("mul_249_cast_fp16")];
            tensor<fp16, [768]> add_188_to_fp16 = const()[name = string("add_188_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249083520)))];
            tensor<fp16, [1, 512, 768]> mul_250_cast_fp16 = mul(x = mul_249_cast_fp16, y = add_188_to_fp16)[name = string("mul_250_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249085120))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249675008))))[name = string("p_st_0_model_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_77_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = mul_250_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> const_1662 = const()[name = string("const_1662"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_66_cast_fp16 = reshape(shape = const_1662, x = linear_77_cast_fp16)[name = string("view_66_cast_fp16")];
            tensor<int32, [4]> transpose_79_perm_0 = const()[name = string("transpose_79_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249676608))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249873280))))[name = string("p_st_0_model_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_78_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = mul_250_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> const_1665 = const()[name = string("const_1665"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_67_cast_fp16 = reshape(shape = const_1665, x = linear_78_cast_fp16)[name = string("view_67_cast_fp16")];
            tensor<int32, [4]> transpose_80_perm_0 = const()[name = string("transpose_80_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249873856))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250070528))))[name = string("p_st_0_model_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = mul_250_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> const_1668 = const()[name = string("const_1668"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_68_cast_fp16 = reshape(shape = const_1668, x = linear_79_cast_fp16)[name = string("view_68_cast_fp16")];
            tensor<int32, [4]> transpose_81_perm_0 = const()[name = string("transpose_81_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_1672_promoted_to_fp16 = const()[name = string("const_1672_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_79_cast_fp16 = transpose(perm = transpose_79_perm_0, x = view_66_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 3, 512, 256]> pow_68_cast_fp16 = pow(x = transpose_79_cast_fp16, y = const_1672_promoted_to_fp16)[name = string("pow_68_cast_fp16")];
            tensor<int32, [1]> mean_67_axes_0 = const()[name = string("mean_67_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_67_keep_dims_0 = const()[name = string("mean_67_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_67_cast_fp16 = reduce_mean(axes = mean_67_axes_0, keep_dims = mean_67_keep_dims_0, x = pow_68_cast_fp16)[name = string("mean_67_cast_fp16")];
            fp16 const_1675_to_fp16 = const()[name = string("const_1675_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_189_cast_fp16 = add(x = mean_67_cast_fp16, y = const_1675_to_fp16)[name = string("add_189_cast_fp16")];
            fp32 rsqrt_67_epsilon_0 = const()[name = string("rsqrt_67_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_67_cast_fp16 = rsqrt(epsilon = rsqrt_67_epsilon_0, x = add_189_cast_fp16)[name = string("rsqrt_67_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_251_cast_fp16 = mul(x = transpose_79_cast_fp16, y = rsqrt_67_cast_fp16)[name = string("mul_251_cast_fp16")];
            tensor<fp16, [256]> add_190_to_fp16 = const()[name = string("add_190_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250071104)))];
            tensor<fp16, [1, 3, 512, 256]> mul_252_cast_fp16 = mul(x = mul_251_cast_fp16, y = add_190_to_fp16)[name = string("mul_252_cast_fp16")];
            fp16 const_1680_promoted_to_fp16 = const()[name = string("const_1680_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_80_cast_fp16 = transpose(perm = transpose_80_perm_0, x = view_67_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 1, 512, 256]> pow_69_cast_fp16 = pow(x = transpose_80_cast_fp16, y = const_1680_promoted_to_fp16)[name = string("pow_69_cast_fp16")];
            tensor<int32, [1]> mean_68_axes_0 = const()[name = string("mean_68_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_68_keep_dims_0 = const()[name = string("mean_68_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_68_cast_fp16 = reduce_mean(axes = mean_68_axes_0, keep_dims = mean_68_keep_dims_0, x = pow_69_cast_fp16)[name = string("mean_68_cast_fp16")];
            fp16 const_1683_to_fp16 = const()[name = string("const_1683_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_191_cast_fp16 = add(x = mean_68_cast_fp16, y = const_1683_to_fp16)[name = string("add_191_cast_fp16")];
            fp32 rsqrt_68_epsilon_0 = const()[name = string("rsqrt_68_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_68_cast_fp16 = rsqrt(epsilon = rsqrt_68_epsilon_0, x = add_191_cast_fp16)[name = string("rsqrt_68_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_253_cast_fp16 = mul(x = transpose_80_cast_fp16, y = rsqrt_68_cast_fp16)[name = string("mul_253_cast_fp16")];
            tensor<fp16, [256]> add_192_to_fp16 = const()[name = string("add_192_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250071680)))];
            tensor<fp16, [1, 1, 512, 256]> mul_254_cast_fp16 = mul(x = mul_253_cast_fp16, y = add_192_to_fp16)[name = string("mul_254_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_255_cast_fp16 = mul(x = mul_252_cast_fp16, y = unsqueeze_97_to_fp16_quantized)[name = string("mul_255_cast_fp16")];
            tensor<int32, [4]> slice_322_begin_0 = const()[name = string("slice_322_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_322_end_0 = const()[name = string("slice_322_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_322_end_mask_0 = const()[name = string("slice_322_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_322_cast_fp16 = slice_by_index(begin = slice_322_begin_0, end = slice_322_end_0, end_mask = slice_322_end_mask_0, x = mul_252_cast_fp16)[name = string("slice_322_cast_fp16")];
            tensor<int32, [4]> slice_323_begin_0 = const()[name = string("slice_323_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_323_end_0 = const()[name = string("slice_323_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_323_end_mask_0 = const()[name = string("slice_323_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_323_cast_fp16 = slice_by_index(begin = slice_323_begin_0, end = slice_323_end_0, end_mask = slice_323_end_mask_0, x = mul_252_cast_fp16)[name = string("slice_323_cast_fp16")];
            fp16 const_1695_promoted_to_fp16 = const()[name = string("const_1695_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_22_cast_fp16 = mul(x = slice_323_cast_fp16, y = const_1695_promoted_to_fp16)[name = string("neg_22_cast_fp16")];
            int32 const_1696 = const()[name = string("const_1696"), val = int32(-1)];
            bool cat_68_interleave_0 = const()[name = string("cat_68_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_68_cast_fp16 = concat(axis = const_1696, interleave = cat_68_interleave_0, values = (neg_22_cast_fp16, slice_322_cast_fp16))[name = string("cat_68_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_256_cast_fp16 = mul(x = cat_68_cast_fp16, y = unsqueeze_98_to_fp16_quantized)[name = string("mul_256_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_193_cast_fp16 = add(x = mul_255_cast_fp16, y = mul_256_cast_fp16)[name = string("add_193_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_257_cast_fp16 = mul(x = mul_254_cast_fp16, y = unsqueeze_97_to_fp16_quantized)[name = string("mul_257_cast_fp16")];
            tensor<int32, [4]> slice_324_begin_0 = const()[name = string("slice_324_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_324_end_0 = const()[name = string("slice_324_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_324_end_mask_0 = const()[name = string("slice_324_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_324_cast_fp16 = slice_by_index(begin = slice_324_begin_0, end = slice_324_end_0, end_mask = slice_324_end_mask_0, x = mul_254_cast_fp16)[name = string("slice_324_cast_fp16")];
            tensor<int32, [4]> slice_325_begin_0 = const()[name = string("slice_325_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_325_end_0 = const()[name = string("slice_325_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_325_end_mask_0 = const()[name = string("slice_325_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_325_cast_fp16 = slice_by_index(begin = slice_325_begin_0, end = slice_325_end_0, end_mask = slice_325_end_mask_0, x = mul_254_cast_fp16)[name = string("slice_325_cast_fp16")];
            fp16 const_1703_promoted_to_fp16 = const()[name = string("const_1703_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_23_cast_fp16 = mul(x = slice_325_cast_fp16, y = const_1703_promoted_to_fp16)[name = string("neg_23_cast_fp16")];
            int32 const_1704 = const()[name = string("const_1704"), val = int32(-1)];
            bool cat_69_interleave_0 = const()[name = string("cat_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_69_cast_fp16 = concat(axis = const_1704, interleave = cat_69_interleave_0, values = (neg_23_cast_fp16, slice_324_cast_fp16))[name = string("cat_69_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_258_cast_fp16 = mul(x = cat_69_cast_fp16, y = unsqueeze_98_to_fp16_quantized)[name = string("mul_258_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_194_cast_fp16 = add(x = mul_257_cast_fp16, y = mul_258_cast_fp16)[name = string("add_194_cast_fp16")];
            int32 const_1705 = const()[name = string("const_1705"), val = int32(-2)];
            bool cat_70_interleave_0 = const()[name = string("cat_70_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_70_cast_fp16 = concat(axis = const_1705, interleave = cat_70_interleave_0, values = add_194_cast_fp16)[name = string("cat_70_cast_fp16")];
            int32 const_1706 = const()[name = string("const_1706"), val = int32(-2)];
            bool cat_71_interleave_0 = const()[name = string("cat_71_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_81_cast_fp16 = transpose(perm = transpose_81_perm_0, x = view_68_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 1, 512, 256]> cat_71_cast_fp16 = concat(axis = const_1706, interleave = cat_71_interleave_0, values = transpose_81_cast_fp16)[name = string("cat_71_cast_fp16")];
            tensor<int32, [1]> unsqueeze_123_axes_0 = const()[name = string("unsqueeze_123_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_123_cast_fp16 = expand_dims(axes = unsqueeze_123_axes_0, x = cat_70_cast_fp16)[name = string("unsqueeze_123_cast_fp16")];
            tensor<int32, [5]> expand_48_reps_0 = const()[name = string("expand_48_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_48_cast_fp16 = tile(reps = expand_48_reps_0, x = unsqueeze_123_cast_fp16)[name = string("expand_48_cast_fp16")];
            tensor<int32, [4]> const_1721 = const()[name = string("const_1721"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_69_cast_fp16 = reshape(shape = const_1721, x = expand_48_cast_fp16)[name = string("view_69_cast_fp16")];
            tensor<int32, [1]> unsqueeze_124_axes_0 = const()[name = string("unsqueeze_124_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_124_cast_fp16 = expand_dims(axes = unsqueeze_124_axes_0, x = cat_71_cast_fp16)[name = string("unsqueeze_124_cast_fp16")];
            tensor<int32, [5]> expand_49_reps_0 = const()[name = string("expand_49_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_49_cast_fp16 = tile(reps = expand_49_reps_0, x = unsqueeze_124_cast_fp16)[name = string("expand_49_cast_fp16")];
            tensor<int32, [4]> const_1736 = const()[name = string("const_1736"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_70_cast_fp16 = reshape(shape = const_1736, x = expand_49_cast_fp16)[name = string("view_70_cast_fp16")];
            bool matmul_46_transpose_x_1 = const()[name = string("matmul_46_transpose_x_1"), val = bool(false)];
            bool matmul_46_transpose_y_1 = const()[name = string("matmul_46_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_46_cast_fp16 = matmul(transpose_x = matmul_46_transpose_x_1, transpose_y = matmul_46_transpose_y_1, x = add_193_cast_fp16, y = view_69_cast_fp16)[name = string("matmul_46_cast_fp16")];
            fp16 const_1739_to_fp16 = const()[name = string("const_1739_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_259_cast_fp16 = mul(x = matmul_46_cast_fp16, y = const_1739_to_fp16)[name = string("mul_259_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_195_cast_fp16 = add(x = mul_259_cast_fp16, y = expand_cast_fp16)[name = string("add_195_cast_fp16")];
            int32 const_1749 = const()[name = string("const_1749"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_11_cast_fp16 = softmax(axis = const_1749, x = add_195_cast_fp16)[name = string("softmax_11_cast_fp16")];
            bool matmul_47_transpose_x_0 = const()[name = string("matmul_47_transpose_x_0"), val = bool(false)];
            bool matmul_47_transpose_y_0 = const()[name = string("matmul_47_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_47_cast_fp16 = matmul(transpose_x = matmul_47_transpose_x_0, transpose_y = matmul_47_transpose_y_0, x = softmax_11_cast_fp16, y = view_70_cast_fp16)[name = string("matmul_47_cast_fp16")];
            tensor<int32, [4]> transpose_83_perm_0 = const()[name = string("transpose_83_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_1754 = const()[name = string("const_1754"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_83_cast_fp16 = transpose(perm = transpose_83_perm_0, x = matmul_47_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 512, 768]> view_71_cast_fp16 = reshape(shape = const_1754, x = transpose_83_cast_fp16)[name = string("view_71_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250072256))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250662144))))[name = string("p_st_0_model_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = view_71_cast_fp16)[name = string("linear_80_cast_fp16")];
            fp16 const_1756_promoted_to_fp16 = const()[name = string("const_1756_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_70_cast_fp16 = pow(x = linear_80_cast_fp16, y = const_1756_promoted_to_fp16)[name = string("pow_70_cast_fp16")];
            tensor<int32, [1]> mean_69_axes_0 = const()[name = string("mean_69_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_69_keep_dims_0 = const()[name = string("mean_69_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_69_cast_fp16 = reduce_mean(axes = mean_69_axes_0, keep_dims = mean_69_keep_dims_0, x = pow_70_cast_fp16)[name = string("mean_69_cast_fp16")];
            fp16 const_1759_to_fp16 = const()[name = string("const_1759_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_196_cast_fp16 = add(x = mean_69_cast_fp16, y = const_1759_to_fp16)[name = string("add_196_cast_fp16")];
            fp32 rsqrt_69_epsilon_0 = const()[name = string("rsqrt_69_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_69_cast_fp16 = rsqrt(epsilon = rsqrt_69_epsilon_0, x = add_196_cast_fp16)[name = string("rsqrt_69_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_260_cast_fp16 = mul(x = linear_80_cast_fp16, y = rsqrt_69_cast_fp16)[name = string("mul_260_cast_fp16")];
            tensor<fp16, [768]> add_197_to_fp16 = const()[name = string("add_197_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250663744)))];
            tensor<fp16, [1, 512, 768]> mul_261_cast_fp16 = mul(x = mul_260_cast_fp16, y = add_197_to_fp16)[name = string("mul_261_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_198_cast_fp16 = add(x = add_186_cast_fp16, y = mul_261_cast_fp16)[name = string("add_198_cast_fp16")];
            fp16 const_1764_promoted_to_fp16 = const()[name = string("const_1764_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_71_cast_fp16 = pow(x = add_198_cast_fp16, y = const_1764_promoted_to_fp16)[name = string("pow_71_cast_fp16")];
            tensor<int32, [1]> mean_70_axes_0 = const()[name = string("mean_70_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_70_keep_dims_0 = const()[name = string("mean_70_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_70_cast_fp16 = reduce_mean(axes = mean_70_axes_0, keep_dims = mean_70_keep_dims_0, x = pow_71_cast_fp16)[name = string("mean_70_cast_fp16")];
            fp16 const_1767_to_fp16 = const()[name = string("const_1767_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_199_cast_fp16 = add(x = mean_70_cast_fp16, y = const_1767_to_fp16)[name = string("add_199_cast_fp16")];
            fp32 rsqrt_70_epsilon_0 = const()[name = string("rsqrt_70_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_70_cast_fp16 = rsqrt(epsilon = rsqrt_70_epsilon_0, x = add_199_cast_fp16)[name = string("rsqrt_70_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_262_cast_fp16 = mul(x = add_198_cast_fp16, y = rsqrt_70_cast_fp16)[name = string("mul_262_cast_fp16")];
            tensor<fp16, [768]> add_200_to_fp16 = const()[name = string("add_200_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250665344)))];
            tensor<fp16, [1, 512, 768]> mul_263_cast_fp16 = mul(x = mul_262_cast_fp16, y = add_200_to_fp16)[name = string("mul_263_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250666944))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251551744))))[name = string("p_st_0_model_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = mul_263_cast_fp16)[name = string("linear_81_cast_fp16")];
            string gelu_11_mode_0 = const()[name = string("gelu_11_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_11_cast_fp16 = gelu(mode = gelu_11_mode_0, x = linear_81_cast_fp16)[name = string("gelu_11_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251554112))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252438912))))[name = string("p_st_0_model_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = mul_263_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_264_cast_fp16 = mul(x = gelu_11_cast_fp16, y = linear_82_cast_fp16)[name = string("mul_264_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252441280))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253326080))))[name = string("p_st_0_model_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_83_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = mul_264_cast_fp16)[name = string("linear_83_cast_fp16")];
            fp16 const_1772_promoted_to_fp16 = const()[name = string("const_1772_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_72_cast_fp16 = pow(x = linear_83_cast_fp16, y = const_1772_promoted_to_fp16)[name = string("pow_72_cast_fp16")];
            tensor<int32, [1]> mean_71_axes_0 = const()[name = string("mean_71_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_71_keep_dims_0 = const()[name = string("mean_71_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_71_cast_fp16 = reduce_mean(axes = mean_71_axes_0, keep_dims = mean_71_keep_dims_0, x = pow_72_cast_fp16)[name = string("mean_71_cast_fp16")];
            fp16 const_1775_to_fp16 = const()[name = string("const_1775_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_201_cast_fp16 = add(x = mean_71_cast_fp16, y = const_1775_to_fp16)[name = string("add_201_cast_fp16")];
            fp32 rsqrt_71_epsilon_0 = const()[name = string("rsqrt_71_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_71_cast_fp16 = rsqrt(epsilon = rsqrt_71_epsilon_0, x = add_201_cast_fp16)[name = string("rsqrt_71_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_265_cast_fp16 = mul(x = linear_83_cast_fp16, y = rsqrt_71_cast_fp16)[name = string("mul_265_cast_fp16")];
            tensor<fp16, [768]> add_202_to_fp16 = const()[name = string("add_202_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253327680)))];
            tensor<fp16, [1, 512, 768]> mul_266_cast_fp16 = mul(x = mul_265_cast_fp16, y = add_202_to_fp16)[name = string("mul_266_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_203_cast_fp16 = add(x = add_198_cast_fp16, y = mul_266_cast_fp16)[name = string("add_203_cast_fp16")];
            fp16 const_1780_promoted_to_fp16 = const()[name = string("const_1780_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_73_cast_fp16 = pow(x = add_203_cast_fp16, y = const_1780_promoted_to_fp16)[name = string("pow_73_cast_fp16")];
            tensor<int32, [1]> mean_72_axes_0 = const()[name = string("mean_72_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_72_keep_dims_0 = const()[name = string("mean_72_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_72_cast_fp16 = reduce_mean(axes = mean_72_axes_0, keep_dims = mean_72_keep_dims_0, x = pow_73_cast_fp16)[name = string("mean_72_cast_fp16")];
            fp16 const_1783_to_fp16 = const()[name = string("const_1783_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_204_cast_fp16 = add(x = mean_72_cast_fp16, y = const_1783_to_fp16)[name = string("add_204_cast_fp16")];
            fp32 rsqrt_72_epsilon_0 = const()[name = string("rsqrt_72_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_72_cast_fp16 = rsqrt(epsilon = rsqrt_72_epsilon_0, x = add_204_cast_fp16)[name = string("rsqrt_72_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_267_cast_fp16 = mul(x = add_203_cast_fp16, y = rsqrt_72_cast_fp16)[name = string("mul_267_cast_fp16")];
            tensor<fp16, [768]> add_205_to_fp16 = const()[name = string("add_205_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253329280)))];
            tensor<fp16, [1, 512, 768]> mul_268_cast_fp16 = mul(x = mul_267_cast_fp16, y = add_205_to_fp16)[name = string("mul_268_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253330880))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253920768))))[name = string("p_st_0_model_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = mul_268_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> const_1787 = const()[name = string("const_1787"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_72_cast_fp16 = reshape(shape = const_1787, x = linear_84_cast_fp16)[name = string("view_72_cast_fp16")];
            tensor<int32, [4]> transpose_84_perm_0 = const()[name = string("transpose_84_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253922368))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254119040))))[name = string("p_st_0_model_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = mul_268_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> const_1790 = const()[name = string("const_1790"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_73_cast_fp16 = reshape(shape = const_1790, x = linear_85_cast_fp16)[name = string("view_73_cast_fp16")];
            tensor<int32, [4]> transpose_85_perm_0 = const()[name = string("transpose_85_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254119616))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254316288))))[name = string("p_st_0_model_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_86_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = mul_268_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> const_1793 = const()[name = string("const_1793"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_74_cast_fp16 = reshape(shape = const_1793, x = linear_86_cast_fp16)[name = string("view_74_cast_fp16")];
            tensor<int32, [4]> transpose_86_perm_0 = const()[name = string("transpose_86_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_1797_promoted_to_fp16 = const()[name = string("const_1797_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_84_cast_fp16 = transpose(perm = transpose_84_perm_0, x = view_72_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 3, 512, 256]> pow_74_cast_fp16 = pow(x = transpose_84_cast_fp16, y = const_1797_promoted_to_fp16)[name = string("pow_74_cast_fp16")];
            tensor<int32, [1]> mean_73_axes_0 = const()[name = string("mean_73_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_73_keep_dims_0 = const()[name = string("mean_73_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_73_cast_fp16 = reduce_mean(axes = mean_73_axes_0, keep_dims = mean_73_keep_dims_0, x = pow_74_cast_fp16)[name = string("mean_73_cast_fp16")];
            fp16 const_1800_to_fp16 = const()[name = string("const_1800_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_206_cast_fp16 = add(x = mean_73_cast_fp16, y = const_1800_to_fp16)[name = string("add_206_cast_fp16")];
            fp32 rsqrt_73_epsilon_0 = const()[name = string("rsqrt_73_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_73_cast_fp16 = rsqrt(epsilon = rsqrt_73_epsilon_0, x = add_206_cast_fp16)[name = string("rsqrt_73_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_269_cast_fp16 = mul(x = transpose_84_cast_fp16, y = rsqrt_73_cast_fp16)[name = string("mul_269_cast_fp16")];
            tensor<fp16, [256]> add_207_to_fp16 = const()[name = string("add_207_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254316864)))];
            tensor<fp16, [1, 3, 512, 256]> mul_270_cast_fp16 = mul(x = mul_269_cast_fp16, y = add_207_to_fp16)[name = string("mul_270_cast_fp16")];
            fp16 const_1805_promoted_to_fp16 = const()[name = string("const_1805_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_85_cast_fp16 = transpose(perm = transpose_85_perm_0, x = view_73_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 1, 512, 256]> pow_75_cast_fp16 = pow(x = transpose_85_cast_fp16, y = const_1805_promoted_to_fp16)[name = string("pow_75_cast_fp16")];
            tensor<int32, [1]> mean_74_axes_0 = const()[name = string("mean_74_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_74_keep_dims_0 = const()[name = string("mean_74_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_74_cast_fp16 = reduce_mean(axes = mean_74_axes_0, keep_dims = mean_74_keep_dims_0, x = pow_75_cast_fp16)[name = string("mean_74_cast_fp16")];
            fp16 const_1808_to_fp16 = const()[name = string("const_1808_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_208_cast_fp16 = add(x = mean_74_cast_fp16, y = const_1808_to_fp16)[name = string("add_208_cast_fp16")];
            fp32 rsqrt_74_epsilon_0 = const()[name = string("rsqrt_74_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_74_cast_fp16 = rsqrt(epsilon = rsqrt_74_epsilon_0, x = add_208_cast_fp16)[name = string("rsqrt_74_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_271_cast_fp16 = mul(x = transpose_85_cast_fp16, y = rsqrt_74_cast_fp16)[name = string("mul_271_cast_fp16")];
            tensor<fp16, [256]> add_209_to_fp16 = const()[name = string("add_209_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254317440)))];
            tensor<fp16, [1, 1, 512, 256]> mul_272_cast_fp16 = mul(x = mul_271_cast_fp16, y = add_209_to_fp16)[name = string("mul_272_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_273_cast_fp16 = mul(x = mul_270_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_273_cast_fp16")];
            tensor<int32, [4]> slice_337_begin_0 = const()[name = string("slice_337_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_337_end_0 = const()[name = string("slice_337_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_337_end_mask_0 = const()[name = string("slice_337_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_337_cast_fp16 = slice_by_index(begin = slice_337_begin_0, end = slice_337_end_0, end_mask = slice_337_end_mask_0, x = mul_270_cast_fp16)[name = string("slice_337_cast_fp16")];
            tensor<int32, [4]> slice_338_begin_0 = const()[name = string("slice_338_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_338_end_0 = const()[name = string("slice_338_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_338_end_mask_0 = const()[name = string("slice_338_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_338_cast_fp16 = slice_by_index(begin = slice_338_begin_0, end = slice_338_end_0, end_mask = slice_338_end_mask_0, x = mul_270_cast_fp16)[name = string("slice_338_cast_fp16")];
            fp16 const_1820_promoted_to_fp16 = const()[name = string("const_1820_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_24_cast_fp16 = mul(x = slice_338_cast_fp16, y = const_1820_promoted_to_fp16)[name = string("neg_24_cast_fp16")];
            int32 const_1821 = const()[name = string("const_1821"), val = int32(-1)];
            bool cat_72_interleave_0 = const()[name = string("cat_72_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_72_cast_fp16 = concat(axis = const_1821, interleave = cat_72_interleave_0, values = (neg_24_cast_fp16, slice_337_cast_fp16))[name = string("cat_72_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_274_cast_fp16 = mul(x = cat_72_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_274_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_210_cast_fp16 = add(x = mul_273_cast_fp16, y = mul_274_cast_fp16)[name = string("add_210_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_275_cast_fp16 = mul(x = mul_272_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_275_cast_fp16")];
            tensor<int32, [4]> slice_339_begin_0 = const()[name = string("slice_339_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_339_end_0 = const()[name = string("slice_339_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_339_end_mask_0 = const()[name = string("slice_339_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_339_cast_fp16 = slice_by_index(begin = slice_339_begin_0, end = slice_339_end_0, end_mask = slice_339_end_mask_0, x = mul_272_cast_fp16)[name = string("slice_339_cast_fp16")];
            tensor<int32, [4]> slice_340_begin_0 = const()[name = string("slice_340_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_340_end_0 = const()[name = string("slice_340_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_340_end_mask_0 = const()[name = string("slice_340_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_340_cast_fp16 = slice_by_index(begin = slice_340_begin_0, end = slice_340_end_0, end_mask = slice_340_end_mask_0, x = mul_272_cast_fp16)[name = string("slice_340_cast_fp16")];
            fp16 const_1828_promoted_to_fp16 = const()[name = string("const_1828_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_25_cast_fp16 = mul(x = slice_340_cast_fp16, y = const_1828_promoted_to_fp16)[name = string("neg_25_cast_fp16")];
            int32 const_1829 = const()[name = string("const_1829"), val = int32(-1)];
            bool cat_73_interleave_0 = const()[name = string("cat_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_73_cast_fp16 = concat(axis = const_1829, interleave = cat_73_interleave_0, values = (neg_25_cast_fp16, slice_339_cast_fp16))[name = string("cat_73_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_276_cast_fp16 = mul(x = cat_73_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_276_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_211_cast_fp16 = add(x = mul_275_cast_fp16, y = mul_276_cast_fp16)[name = string("add_211_cast_fp16")];
            int32 const_1830 = const()[name = string("const_1830"), val = int32(-2)];
            bool cat_74_interleave_0 = const()[name = string("cat_74_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_74_cast_fp16 = concat(axis = const_1830, interleave = cat_74_interleave_0, values = add_211_cast_fp16)[name = string("cat_74_cast_fp16")];
            int32 const_1831 = const()[name = string("const_1831"), val = int32(-2)];
            bool cat_75_interleave_0 = const()[name = string("cat_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_86_cast_fp16 = transpose(perm = transpose_86_perm_0, x = view_74_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 1, 512, 256]> cat_75_cast_fp16 = concat(axis = const_1831, interleave = cat_75_interleave_0, values = transpose_86_cast_fp16)[name = string("cat_75_cast_fp16")];
            tensor<int32, [1]> unsqueeze_127_axes_0 = const()[name = string("unsqueeze_127_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_127_cast_fp16 = expand_dims(axes = unsqueeze_127_axes_0, x = cat_74_cast_fp16)[name = string("unsqueeze_127_cast_fp16")];
            tensor<int32, [5]> expand_50_reps_0 = const()[name = string("expand_50_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_50_cast_fp16 = tile(reps = expand_50_reps_0, x = unsqueeze_127_cast_fp16)[name = string("expand_50_cast_fp16")];
            tensor<int32, [4]> const_1846 = const()[name = string("const_1846"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_75_cast_fp16 = reshape(shape = const_1846, x = expand_50_cast_fp16)[name = string("view_75_cast_fp16")];
            tensor<int32, [1]> unsqueeze_128_axes_0 = const()[name = string("unsqueeze_128_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_128_cast_fp16 = expand_dims(axes = unsqueeze_128_axes_0, x = cat_75_cast_fp16)[name = string("unsqueeze_128_cast_fp16")];
            tensor<int32, [5]> expand_51_reps_0 = const()[name = string("expand_51_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_51_cast_fp16 = tile(reps = expand_51_reps_0, x = unsqueeze_128_cast_fp16)[name = string("expand_51_cast_fp16")];
            tensor<int32, [4]> const_1861 = const()[name = string("const_1861"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_76_cast_fp16 = reshape(shape = const_1861, x = expand_51_cast_fp16)[name = string("view_76_cast_fp16")];
            bool matmul_48_transpose_x_1 = const()[name = string("matmul_48_transpose_x_1"), val = bool(false)];
            bool matmul_48_transpose_y_1 = const()[name = string("matmul_48_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_48_cast_fp16 = matmul(transpose_x = matmul_48_transpose_x_1, transpose_y = matmul_48_transpose_y_1, x = add_210_cast_fp16, y = view_75_cast_fp16)[name = string("matmul_48_cast_fp16")];
            fp16 const_1864_to_fp16 = const()[name = string("const_1864_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_277_cast_fp16 = mul(x = matmul_48_cast_fp16, y = const_1864_to_fp16)[name = string("mul_277_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_212_cast_fp16 = add(x = mul_277_cast_fp16, y = expand_cast_fp16)[name = string("add_212_cast_fp16")];
            int32 const_1874 = const()[name = string("const_1874"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_12_cast_fp16 = softmax(axis = const_1874, x = add_212_cast_fp16)[name = string("softmax_12_cast_fp16")];
            bool matmul_49_transpose_x_0 = const()[name = string("matmul_49_transpose_x_0"), val = bool(false)];
            bool matmul_49_transpose_y_0 = const()[name = string("matmul_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_49_cast_fp16 = matmul(transpose_x = matmul_49_transpose_x_0, transpose_y = matmul_49_transpose_y_0, x = softmax_12_cast_fp16, y = view_76_cast_fp16)[name = string("matmul_49_cast_fp16")];
            tensor<int32, [4]> transpose_88_perm_0 = const()[name = string("transpose_88_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_1879 = const()[name = string("const_1879"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_88_cast_fp16 = transpose(perm = transpose_88_perm_0, x = matmul_49_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 512, 768]> view_77_cast_fp16 = reshape(shape = const_1879, x = transpose_88_cast_fp16)[name = string("view_77_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254318016))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254907904))))[name = string("p_st_0_model_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_87_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = view_77_cast_fp16)[name = string("linear_87_cast_fp16")];
            fp16 const_1881_promoted_to_fp16 = const()[name = string("const_1881_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_76_cast_fp16 = pow(x = linear_87_cast_fp16, y = const_1881_promoted_to_fp16)[name = string("pow_76_cast_fp16")];
            tensor<int32, [1]> mean_75_axes_0 = const()[name = string("mean_75_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_75_keep_dims_0 = const()[name = string("mean_75_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_75_cast_fp16 = reduce_mean(axes = mean_75_axes_0, keep_dims = mean_75_keep_dims_0, x = pow_76_cast_fp16)[name = string("mean_75_cast_fp16")];
            fp16 const_1884_to_fp16 = const()[name = string("const_1884_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_213_cast_fp16 = add(x = mean_75_cast_fp16, y = const_1884_to_fp16)[name = string("add_213_cast_fp16")];
            fp32 rsqrt_75_epsilon_0 = const()[name = string("rsqrt_75_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_75_cast_fp16 = rsqrt(epsilon = rsqrt_75_epsilon_0, x = add_213_cast_fp16)[name = string("rsqrt_75_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_278_cast_fp16 = mul(x = linear_87_cast_fp16, y = rsqrt_75_cast_fp16)[name = string("mul_278_cast_fp16")];
            tensor<fp16, [768]> add_214_to_fp16 = const()[name = string("add_214_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254909504)))];
            tensor<fp16, [1, 512, 768]> mul_279_cast_fp16 = mul(x = mul_278_cast_fp16, y = add_214_to_fp16)[name = string("mul_279_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_215_cast_fp16 = add(x = add_203_cast_fp16, y = mul_279_cast_fp16)[name = string("add_215_cast_fp16")];
            fp16 const_1889_promoted_to_fp16 = const()[name = string("const_1889_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_77_cast_fp16 = pow(x = add_215_cast_fp16, y = const_1889_promoted_to_fp16)[name = string("pow_77_cast_fp16")];
            tensor<int32, [1]> mean_76_axes_0 = const()[name = string("mean_76_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_76_keep_dims_0 = const()[name = string("mean_76_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_76_cast_fp16 = reduce_mean(axes = mean_76_axes_0, keep_dims = mean_76_keep_dims_0, x = pow_77_cast_fp16)[name = string("mean_76_cast_fp16")];
            fp16 const_1892_to_fp16 = const()[name = string("const_1892_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_216_cast_fp16 = add(x = mean_76_cast_fp16, y = const_1892_to_fp16)[name = string("add_216_cast_fp16")];
            fp32 rsqrt_76_epsilon_0 = const()[name = string("rsqrt_76_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_76_cast_fp16 = rsqrt(epsilon = rsqrt_76_epsilon_0, x = add_216_cast_fp16)[name = string("rsqrt_76_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_280_cast_fp16 = mul(x = add_215_cast_fp16, y = rsqrt_76_cast_fp16)[name = string("mul_280_cast_fp16")];
            tensor<fp16, [768]> add_217_to_fp16 = const()[name = string("add_217_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254911104)))];
            tensor<fp16, [1, 512, 768]> mul_281_cast_fp16 = mul(x = mul_280_cast_fp16, y = add_217_to_fp16)[name = string("mul_281_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254912704))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255797504))))[name = string("p_st_0_model_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = mul_281_cast_fp16)[name = string("linear_88_cast_fp16")];
            string gelu_12_mode_0 = const()[name = string("gelu_12_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_12_cast_fp16 = gelu(mode = gelu_12_mode_0, x = linear_88_cast_fp16)[name = string("gelu_12_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255799872))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256684672))))[name = string("p_st_0_model_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = mul_281_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_282_cast_fp16 = mul(x = gelu_12_cast_fp16, y = linear_89_cast_fp16)[name = string("mul_282_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256687040))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257571840))))[name = string("p_st_0_model_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = mul_282_cast_fp16)[name = string("linear_90_cast_fp16")];
            fp16 const_1897_promoted_to_fp16 = const()[name = string("const_1897_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_78_cast_fp16 = pow(x = linear_90_cast_fp16, y = const_1897_promoted_to_fp16)[name = string("pow_78_cast_fp16")];
            tensor<int32, [1]> mean_77_axes_0 = const()[name = string("mean_77_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_77_keep_dims_0 = const()[name = string("mean_77_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_77_cast_fp16 = reduce_mean(axes = mean_77_axes_0, keep_dims = mean_77_keep_dims_0, x = pow_78_cast_fp16)[name = string("mean_77_cast_fp16")];
            fp16 const_1900_to_fp16 = const()[name = string("const_1900_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_218_cast_fp16 = add(x = mean_77_cast_fp16, y = const_1900_to_fp16)[name = string("add_218_cast_fp16")];
            fp32 rsqrt_77_epsilon_0 = const()[name = string("rsqrt_77_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_77_cast_fp16 = rsqrt(epsilon = rsqrt_77_epsilon_0, x = add_218_cast_fp16)[name = string("rsqrt_77_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_283_cast_fp16 = mul(x = linear_90_cast_fp16, y = rsqrt_77_cast_fp16)[name = string("mul_283_cast_fp16")];
            tensor<fp16, [768]> add_219_to_fp16 = const()[name = string("add_219_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257573440)))];
            tensor<fp16, [1, 512, 768]> mul_284_cast_fp16 = mul(x = mul_283_cast_fp16, y = add_219_to_fp16)[name = string("mul_284_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_220_cast_fp16 = add(x = add_215_cast_fp16, y = mul_284_cast_fp16)[name = string("add_220_cast_fp16")];
            fp16 const_1905_promoted_to_fp16 = const()[name = string("const_1905_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_79_cast_fp16 = pow(x = add_220_cast_fp16, y = const_1905_promoted_to_fp16)[name = string("pow_79_cast_fp16")];
            tensor<int32, [1]> mean_78_axes_0 = const()[name = string("mean_78_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_78_keep_dims_0 = const()[name = string("mean_78_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_78_cast_fp16 = reduce_mean(axes = mean_78_axes_0, keep_dims = mean_78_keep_dims_0, x = pow_79_cast_fp16)[name = string("mean_78_cast_fp16")];
            fp16 const_1908_to_fp16 = const()[name = string("const_1908_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_221_cast_fp16 = add(x = mean_78_cast_fp16, y = const_1908_to_fp16)[name = string("add_221_cast_fp16")];
            fp32 rsqrt_78_epsilon_0 = const()[name = string("rsqrt_78_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_78_cast_fp16 = rsqrt(epsilon = rsqrt_78_epsilon_0, x = add_221_cast_fp16)[name = string("rsqrt_78_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_285_cast_fp16 = mul(x = add_220_cast_fp16, y = rsqrt_78_cast_fp16)[name = string("mul_285_cast_fp16")];
            tensor<fp16, [768]> add_222_to_fp16 = const()[name = string("add_222_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257575040)))];
            tensor<fp16, [1, 512, 768]> mul_286_cast_fp16 = mul(x = mul_285_cast_fp16, y = add_222_to_fp16)[name = string("mul_286_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257576640))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258166528))))[name = string("p_st_0_model_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_91_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = mul_286_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> const_1912 = const()[name = string("const_1912"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_78_cast_fp16 = reshape(shape = const_1912, x = linear_91_cast_fp16)[name = string("view_78_cast_fp16")];
            tensor<int32, [4]> transpose_89_perm_0 = const()[name = string("transpose_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258168128))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258364800))))[name = string("p_st_0_model_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_92_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = mul_286_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> const_1915 = const()[name = string("const_1915"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_79_cast_fp16 = reshape(shape = const_1915, x = linear_92_cast_fp16)[name = string("view_79_cast_fp16")];
            tensor<int32, [4]> transpose_90_perm_0 = const()[name = string("transpose_90_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258365376))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258562048))))[name = string("p_st_0_model_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = mul_286_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> const_1918 = const()[name = string("const_1918"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_80_cast_fp16 = reshape(shape = const_1918, x = linear_93_cast_fp16)[name = string("view_80_cast_fp16")];
            tensor<int32, [4]> transpose_91_perm_0 = const()[name = string("transpose_91_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_1922_promoted_to_fp16 = const()[name = string("const_1922_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_89_cast_fp16 = transpose(perm = transpose_89_perm_0, x = view_78_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 3, 512, 256]> pow_80_cast_fp16 = pow(x = transpose_89_cast_fp16, y = const_1922_promoted_to_fp16)[name = string("pow_80_cast_fp16")];
            tensor<int32, [1]> mean_79_axes_0 = const()[name = string("mean_79_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_79_keep_dims_0 = const()[name = string("mean_79_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_79_cast_fp16 = reduce_mean(axes = mean_79_axes_0, keep_dims = mean_79_keep_dims_0, x = pow_80_cast_fp16)[name = string("mean_79_cast_fp16")];
            fp16 const_1925_to_fp16 = const()[name = string("const_1925_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_223_cast_fp16 = add(x = mean_79_cast_fp16, y = const_1925_to_fp16)[name = string("add_223_cast_fp16")];
            fp32 rsqrt_79_epsilon_0 = const()[name = string("rsqrt_79_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_79_cast_fp16 = rsqrt(epsilon = rsqrt_79_epsilon_0, x = add_223_cast_fp16)[name = string("rsqrt_79_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_287_cast_fp16 = mul(x = transpose_89_cast_fp16, y = rsqrt_79_cast_fp16)[name = string("mul_287_cast_fp16")];
            tensor<fp16, [256]> add_224_to_fp16 = const()[name = string("add_224_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258562624)))];
            tensor<fp16, [1, 3, 512, 256]> mul_288_cast_fp16 = mul(x = mul_287_cast_fp16, y = add_224_to_fp16)[name = string("mul_288_cast_fp16")];
            fp16 const_1930_promoted_to_fp16 = const()[name = string("const_1930_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_90_cast_fp16 = transpose(perm = transpose_90_perm_0, x = view_79_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 1, 512, 256]> pow_81_cast_fp16 = pow(x = transpose_90_cast_fp16, y = const_1930_promoted_to_fp16)[name = string("pow_81_cast_fp16")];
            tensor<int32, [1]> mean_80_axes_0 = const()[name = string("mean_80_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_80_keep_dims_0 = const()[name = string("mean_80_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_80_cast_fp16 = reduce_mean(axes = mean_80_axes_0, keep_dims = mean_80_keep_dims_0, x = pow_81_cast_fp16)[name = string("mean_80_cast_fp16")];
            fp16 const_1933_to_fp16 = const()[name = string("const_1933_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_225_cast_fp16 = add(x = mean_80_cast_fp16, y = const_1933_to_fp16)[name = string("add_225_cast_fp16")];
            fp32 rsqrt_80_epsilon_0 = const()[name = string("rsqrt_80_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_80_cast_fp16 = rsqrt(epsilon = rsqrt_80_epsilon_0, x = add_225_cast_fp16)[name = string("rsqrt_80_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_289_cast_fp16 = mul(x = transpose_90_cast_fp16, y = rsqrt_80_cast_fp16)[name = string("mul_289_cast_fp16")];
            tensor<fp16, [256]> add_226_to_fp16 = const()[name = string("add_226_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258563200)))];
            tensor<fp16, [1, 1, 512, 256]> mul_290_cast_fp16 = mul(x = mul_289_cast_fp16, y = add_226_to_fp16)[name = string("mul_290_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_291_cast_fp16 = mul(x = mul_288_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_291_cast_fp16")];
            tensor<int32, [4]> slice_360_begin_0 = const()[name = string("slice_360_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_360_end_0 = const()[name = string("slice_360_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_360_end_mask_0 = const()[name = string("slice_360_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_360_cast_fp16 = slice_by_index(begin = slice_360_begin_0, end = slice_360_end_0, end_mask = slice_360_end_mask_0, x = mul_288_cast_fp16)[name = string("slice_360_cast_fp16")];
            tensor<int32, [4]> slice_361_begin_0 = const()[name = string("slice_361_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_361_end_0 = const()[name = string("slice_361_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_361_end_mask_0 = const()[name = string("slice_361_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_361_cast_fp16 = slice_by_index(begin = slice_361_begin_0, end = slice_361_end_0, end_mask = slice_361_end_mask_0, x = mul_288_cast_fp16)[name = string("slice_361_cast_fp16")];
            fp16 const_1945_promoted_to_fp16 = const()[name = string("const_1945_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_26_cast_fp16 = mul(x = slice_361_cast_fp16, y = const_1945_promoted_to_fp16)[name = string("neg_26_cast_fp16")];
            int32 const_1946 = const()[name = string("const_1946"), val = int32(-1)];
            bool cat_76_interleave_0 = const()[name = string("cat_76_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_76_cast_fp16 = concat(axis = const_1946, interleave = cat_76_interleave_0, values = (neg_26_cast_fp16, slice_360_cast_fp16))[name = string("cat_76_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_292_cast_fp16 = mul(x = cat_76_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_292_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_227_cast_fp16 = add(x = mul_291_cast_fp16, y = mul_292_cast_fp16)[name = string("add_227_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_293_cast_fp16 = mul(x = mul_290_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_293_cast_fp16")];
            tensor<int32, [4]> slice_362_begin_0 = const()[name = string("slice_362_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_362_end_0 = const()[name = string("slice_362_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_362_end_mask_0 = const()[name = string("slice_362_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_362_cast_fp16 = slice_by_index(begin = slice_362_begin_0, end = slice_362_end_0, end_mask = slice_362_end_mask_0, x = mul_290_cast_fp16)[name = string("slice_362_cast_fp16")];
            tensor<int32, [4]> slice_363_begin_0 = const()[name = string("slice_363_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_363_end_0 = const()[name = string("slice_363_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_363_end_mask_0 = const()[name = string("slice_363_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_363_cast_fp16 = slice_by_index(begin = slice_363_begin_0, end = slice_363_end_0, end_mask = slice_363_end_mask_0, x = mul_290_cast_fp16)[name = string("slice_363_cast_fp16")];
            fp16 const_1953_promoted_to_fp16 = const()[name = string("const_1953_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_27_cast_fp16 = mul(x = slice_363_cast_fp16, y = const_1953_promoted_to_fp16)[name = string("neg_27_cast_fp16")];
            int32 const_1954 = const()[name = string("const_1954"), val = int32(-1)];
            bool cat_77_interleave_0 = const()[name = string("cat_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_77_cast_fp16 = concat(axis = const_1954, interleave = cat_77_interleave_0, values = (neg_27_cast_fp16, slice_362_cast_fp16))[name = string("cat_77_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_294_cast_fp16 = mul(x = cat_77_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_294_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_228_cast_fp16 = add(x = mul_293_cast_fp16, y = mul_294_cast_fp16)[name = string("add_228_cast_fp16")];
            int32 const_1955 = const()[name = string("const_1955"), val = int32(-2)];
            bool cat_78_interleave_0 = const()[name = string("cat_78_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_78_cast_fp16 = concat(axis = const_1955, interleave = cat_78_interleave_0, values = add_228_cast_fp16)[name = string("cat_78_cast_fp16")];
            int32 const_1956 = const()[name = string("const_1956"), val = int32(-2)];
            bool cat_79_interleave_0 = const()[name = string("cat_79_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_91_cast_fp16 = transpose(perm = transpose_91_perm_0, x = view_80_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 1, 512, 256]> cat_79_cast_fp16 = concat(axis = const_1956, interleave = cat_79_interleave_0, values = transpose_91_cast_fp16)[name = string("cat_79_cast_fp16")];
            tensor<int32, [1]> unsqueeze_131_axes_0 = const()[name = string("unsqueeze_131_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_131_cast_fp16 = expand_dims(axes = unsqueeze_131_axes_0, x = cat_78_cast_fp16)[name = string("unsqueeze_131_cast_fp16")];
            tensor<int32, [5]> expand_52_reps_0 = const()[name = string("expand_52_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_52_cast_fp16 = tile(reps = expand_52_reps_0, x = unsqueeze_131_cast_fp16)[name = string("expand_52_cast_fp16")];
            tensor<int32, [4]> const_1971 = const()[name = string("const_1971"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_81_cast_fp16 = reshape(shape = const_1971, x = expand_52_cast_fp16)[name = string("view_81_cast_fp16")];
            tensor<int32, [1]> unsqueeze_132_axes_0 = const()[name = string("unsqueeze_132_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_132_cast_fp16 = expand_dims(axes = unsqueeze_132_axes_0, x = cat_79_cast_fp16)[name = string("unsqueeze_132_cast_fp16")];
            tensor<int32, [5]> expand_53_reps_0 = const()[name = string("expand_53_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_53_cast_fp16 = tile(reps = expand_53_reps_0, x = unsqueeze_132_cast_fp16)[name = string("expand_53_cast_fp16")];
            tensor<int32, [4]> const_1986 = const()[name = string("const_1986"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_82_cast_fp16 = reshape(shape = const_1986, x = expand_53_cast_fp16)[name = string("view_82_cast_fp16")];
            bool matmul_50_transpose_x_1 = const()[name = string("matmul_50_transpose_x_1"), val = bool(false)];
            bool matmul_50_transpose_y_1 = const()[name = string("matmul_50_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_50_cast_fp16 = matmul(transpose_x = matmul_50_transpose_x_1, transpose_y = matmul_50_transpose_y_1, x = add_227_cast_fp16, y = view_81_cast_fp16)[name = string("matmul_50_cast_fp16")];
            fp16 const_1989_to_fp16 = const()[name = string("const_1989_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_295_cast_fp16 = mul(x = matmul_50_cast_fp16, y = const_1989_to_fp16)[name = string("mul_295_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_229_cast_fp16 = add(x = mul_295_cast_fp16, y = expand_cast_fp16)[name = string("add_229_cast_fp16")];
            int32 const_1999 = const()[name = string("const_1999"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_13_cast_fp16 = softmax(axis = const_1999, x = add_229_cast_fp16)[name = string("softmax_13_cast_fp16")];
            bool matmul_51_transpose_x_0 = const()[name = string("matmul_51_transpose_x_0"), val = bool(false)];
            bool matmul_51_transpose_y_0 = const()[name = string("matmul_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_51_cast_fp16 = matmul(transpose_x = matmul_51_transpose_x_0, transpose_y = matmul_51_transpose_y_0, x = softmax_13_cast_fp16, y = view_82_cast_fp16)[name = string("matmul_51_cast_fp16")];
            tensor<int32, [4]> transpose_93_perm_0 = const()[name = string("transpose_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_2004 = const()[name = string("const_2004"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_93_cast_fp16 = transpose(perm = transpose_93_perm_0, x = matmul_51_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 512, 768]> view_83_cast_fp16 = reshape(shape = const_2004, x = transpose_93_cast_fp16)[name = string("view_83_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258563776))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259153664))))[name = string("p_st_0_model_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = view_83_cast_fp16)[name = string("linear_94_cast_fp16")];
            fp16 const_2006_promoted_to_fp16 = const()[name = string("const_2006_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_82_cast_fp16 = pow(x = linear_94_cast_fp16, y = const_2006_promoted_to_fp16)[name = string("pow_82_cast_fp16")];
            tensor<int32, [1]> mean_81_axes_0 = const()[name = string("mean_81_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_81_keep_dims_0 = const()[name = string("mean_81_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_81_cast_fp16 = reduce_mean(axes = mean_81_axes_0, keep_dims = mean_81_keep_dims_0, x = pow_82_cast_fp16)[name = string("mean_81_cast_fp16")];
            fp16 const_2009_to_fp16 = const()[name = string("const_2009_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_230_cast_fp16 = add(x = mean_81_cast_fp16, y = const_2009_to_fp16)[name = string("add_230_cast_fp16")];
            fp32 rsqrt_81_epsilon_0 = const()[name = string("rsqrt_81_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_81_cast_fp16 = rsqrt(epsilon = rsqrt_81_epsilon_0, x = add_230_cast_fp16)[name = string("rsqrt_81_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_296_cast_fp16 = mul(x = linear_94_cast_fp16, y = rsqrt_81_cast_fp16)[name = string("mul_296_cast_fp16")];
            tensor<fp16, [768]> add_231_to_fp16 = const()[name = string("add_231_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259155264)))];
            tensor<fp16, [1, 512, 768]> mul_297_cast_fp16 = mul(x = mul_296_cast_fp16, y = add_231_to_fp16)[name = string("mul_297_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_232_cast_fp16 = add(x = add_220_cast_fp16, y = mul_297_cast_fp16)[name = string("add_232_cast_fp16")];
            fp16 const_2014_promoted_to_fp16 = const()[name = string("const_2014_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_83_cast_fp16 = pow(x = add_232_cast_fp16, y = const_2014_promoted_to_fp16)[name = string("pow_83_cast_fp16")];
            tensor<int32, [1]> mean_82_axes_0 = const()[name = string("mean_82_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_82_keep_dims_0 = const()[name = string("mean_82_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_82_cast_fp16 = reduce_mean(axes = mean_82_axes_0, keep_dims = mean_82_keep_dims_0, x = pow_83_cast_fp16)[name = string("mean_82_cast_fp16")];
            fp16 const_2017_to_fp16 = const()[name = string("const_2017_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_233_cast_fp16 = add(x = mean_82_cast_fp16, y = const_2017_to_fp16)[name = string("add_233_cast_fp16")];
            fp32 rsqrt_82_epsilon_0 = const()[name = string("rsqrt_82_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_82_cast_fp16 = rsqrt(epsilon = rsqrt_82_epsilon_0, x = add_233_cast_fp16)[name = string("rsqrt_82_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_298_cast_fp16 = mul(x = add_232_cast_fp16, y = rsqrt_82_cast_fp16)[name = string("mul_298_cast_fp16")];
            tensor<fp16, [768]> add_234_to_fp16 = const()[name = string("add_234_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259156864)))];
            tensor<fp16, [1, 512, 768]> mul_299_cast_fp16 = mul(x = mul_298_cast_fp16, y = add_234_to_fp16)[name = string("mul_299_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259158464))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260043264))))[name = string("p_st_0_model_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = mul_299_cast_fp16)[name = string("linear_95_cast_fp16")];
            string gelu_13_mode_0 = const()[name = string("gelu_13_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_13_cast_fp16 = gelu(mode = gelu_13_mode_0, x = linear_95_cast_fp16)[name = string("gelu_13_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260045632))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260930432))))[name = string("p_st_0_model_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = mul_299_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_300_cast_fp16 = mul(x = gelu_13_cast_fp16, y = linear_96_cast_fp16)[name = string("mul_300_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260932800))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261817600))))[name = string("p_st_0_model_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_97_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = mul_300_cast_fp16)[name = string("linear_97_cast_fp16")];
            fp16 const_2022_promoted_to_fp16 = const()[name = string("const_2022_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_84_cast_fp16 = pow(x = linear_97_cast_fp16, y = const_2022_promoted_to_fp16)[name = string("pow_84_cast_fp16")];
            tensor<int32, [1]> mean_83_axes_0 = const()[name = string("mean_83_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_83_keep_dims_0 = const()[name = string("mean_83_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_83_cast_fp16 = reduce_mean(axes = mean_83_axes_0, keep_dims = mean_83_keep_dims_0, x = pow_84_cast_fp16)[name = string("mean_83_cast_fp16")];
            fp16 const_2025_to_fp16 = const()[name = string("const_2025_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_235_cast_fp16 = add(x = mean_83_cast_fp16, y = const_2025_to_fp16)[name = string("add_235_cast_fp16")];
            fp32 rsqrt_83_epsilon_0 = const()[name = string("rsqrt_83_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_83_cast_fp16 = rsqrt(epsilon = rsqrt_83_epsilon_0, x = add_235_cast_fp16)[name = string("rsqrt_83_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_301_cast_fp16 = mul(x = linear_97_cast_fp16, y = rsqrt_83_cast_fp16)[name = string("mul_301_cast_fp16")];
            tensor<fp16, [768]> add_236_to_fp16 = const()[name = string("add_236_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261819200)))];
            tensor<fp16, [1, 512, 768]> mul_302_cast_fp16 = mul(x = mul_301_cast_fp16, y = add_236_to_fp16)[name = string("mul_302_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_237_cast_fp16 = add(x = add_232_cast_fp16, y = mul_302_cast_fp16)[name = string("add_237_cast_fp16")];
            fp16 const_2030_promoted_to_fp16 = const()[name = string("const_2030_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_85_cast_fp16 = pow(x = add_237_cast_fp16, y = const_2030_promoted_to_fp16)[name = string("pow_85_cast_fp16")];
            tensor<int32, [1]> mean_84_axes_0 = const()[name = string("mean_84_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_84_keep_dims_0 = const()[name = string("mean_84_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_84_cast_fp16 = reduce_mean(axes = mean_84_axes_0, keep_dims = mean_84_keep_dims_0, x = pow_85_cast_fp16)[name = string("mean_84_cast_fp16")];
            fp16 const_2033_to_fp16 = const()[name = string("const_2033_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_238_cast_fp16 = add(x = mean_84_cast_fp16, y = const_2033_to_fp16)[name = string("add_238_cast_fp16")];
            fp32 rsqrt_84_epsilon_0 = const()[name = string("rsqrt_84_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_84_cast_fp16 = rsqrt(epsilon = rsqrt_84_epsilon_0, x = add_238_cast_fp16)[name = string("rsqrt_84_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_303_cast_fp16 = mul(x = add_237_cast_fp16, y = rsqrt_84_cast_fp16)[name = string("mul_303_cast_fp16")];
            tensor<fp16, [768]> add_239_to_fp16 = const()[name = string("add_239_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261820800)))];
            tensor<fp16, [1, 512, 768]> mul_304_cast_fp16 = mul(x = mul_303_cast_fp16, y = add_239_to_fp16)[name = string("mul_304_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261822400))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262412288))))[name = string("p_st_0_model_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_98_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = mul_304_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [4]> const_2037 = const()[name = string("const_2037"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_84_cast_fp16 = reshape(shape = const_2037, x = linear_98_cast_fp16)[name = string("view_84_cast_fp16")];
            tensor<int32, [4]> transpose_94_perm_0 = const()[name = string("transpose_94_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262413888))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262610560))))[name = string("p_st_0_model_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_99_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = mul_304_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> const_2040 = const()[name = string("const_2040"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_85_cast_fp16 = reshape(shape = const_2040, x = linear_99_cast_fp16)[name = string("view_85_cast_fp16")];
            tensor<int32, [4]> transpose_95_perm_0 = const()[name = string("transpose_95_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262611136))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262807808))))[name = string("p_st_0_model_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_100_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = mul_304_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> const_2043 = const()[name = string("const_2043"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_86_cast_fp16 = reshape(shape = const_2043, x = linear_100_cast_fp16)[name = string("view_86_cast_fp16")];
            tensor<int32, [4]> transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_2047_promoted_to_fp16 = const()[name = string("const_2047_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_94_cast_fp16 = transpose(perm = transpose_94_perm_0, x = view_84_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 3, 512, 256]> pow_86_cast_fp16 = pow(x = transpose_94_cast_fp16, y = const_2047_promoted_to_fp16)[name = string("pow_86_cast_fp16")];
            tensor<int32, [1]> mean_85_axes_0 = const()[name = string("mean_85_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_85_keep_dims_0 = const()[name = string("mean_85_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_85_cast_fp16 = reduce_mean(axes = mean_85_axes_0, keep_dims = mean_85_keep_dims_0, x = pow_86_cast_fp16)[name = string("mean_85_cast_fp16")];
            fp16 const_2050_to_fp16 = const()[name = string("const_2050_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_240_cast_fp16 = add(x = mean_85_cast_fp16, y = const_2050_to_fp16)[name = string("add_240_cast_fp16")];
            fp32 rsqrt_85_epsilon_0 = const()[name = string("rsqrt_85_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_85_cast_fp16 = rsqrt(epsilon = rsqrt_85_epsilon_0, x = add_240_cast_fp16)[name = string("rsqrt_85_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_305_cast_fp16 = mul(x = transpose_94_cast_fp16, y = rsqrt_85_cast_fp16)[name = string("mul_305_cast_fp16")];
            tensor<fp16, [256]> add_241_to_fp16 = const()[name = string("add_241_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262808384)))];
            tensor<fp16, [1, 3, 512, 256]> mul_306_cast_fp16 = mul(x = mul_305_cast_fp16, y = add_241_to_fp16)[name = string("mul_306_cast_fp16")];
            fp16 const_2055_promoted_to_fp16 = const()[name = string("const_2055_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_95_cast_fp16 = transpose(perm = transpose_95_perm_0, x = view_85_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 1, 512, 256]> pow_87_cast_fp16 = pow(x = transpose_95_cast_fp16, y = const_2055_promoted_to_fp16)[name = string("pow_87_cast_fp16")];
            tensor<int32, [1]> mean_86_axes_0 = const()[name = string("mean_86_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_86_keep_dims_0 = const()[name = string("mean_86_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_86_cast_fp16 = reduce_mean(axes = mean_86_axes_0, keep_dims = mean_86_keep_dims_0, x = pow_87_cast_fp16)[name = string("mean_86_cast_fp16")];
            fp16 const_2058_to_fp16 = const()[name = string("const_2058_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_242_cast_fp16 = add(x = mean_86_cast_fp16, y = const_2058_to_fp16)[name = string("add_242_cast_fp16")];
            fp32 rsqrt_86_epsilon_0 = const()[name = string("rsqrt_86_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_86_cast_fp16 = rsqrt(epsilon = rsqrt_86_epsilon_0, x = add_242_cast_fp16)[name = string("rsqrt_86_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_307_cast_fp16 = mul(x = transpose_95_cast_fp16, y = rsqrt_86_cast_fp16)[name = string("mul_307_cast_fp16")];
            tensor<fp16, [256]> add_243_to_fp16 = const()[name = string("add_243_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262808960)))];
            tensor<fp16, [1, 1, 512, 256]> mul_308_cast_fp16 = mul(x = mul_307_cast_fp16, y = add_243_to_fp16)[name = string("mul_308_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_309_cast_fp16 = mul(x = mul_306_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_309_cast_fp16")];
            tensor<int32, [4]> slice_383_begin_0 = const()[name = string("slice_383_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_383_end_0 = const()[name = string("slice_383_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_383_end_mask_0 = const()[name = string("slice_383_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_383_cast_fp16 = slice_by_index(begin = slice_383_begin_0, end = slice_383_end_0, end_mask = slice_383_end_mask_0, x = mul_306_cast_fp16)[name = string("slice_383_cast_fp16")];
            tensor<int32, [4]> slice_384_begin_0 = const()[name = string("slice_384_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_384_end_0 = const()[name = string("slice_384_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_384_end_mask_0 = const()[name = string("slice_384_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_384_cast_fp16 = slice_by_index(begin = slice_384_begin_0, end = slice_384_end_0, end_mask = slice_384_end_mask_0, x = mul_306_cast_fp16)[name = string("slice_384_cast_fp16")];
            fp16 const_2070_promoted_to_fp16 = const()[name = string("const_2070_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_28_cast_fp16 = mul(x = slice_384_cast_fp16, y = const_2070_promoted_to_fp16)[name = string("neg_28_cast_fp16")];
            int32 const_2071 = const()[name = string("const_2071"), val = int32(-1)];
            bool cat_80_interleave_0 = const()[name = string("cat_80_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_80_cast_fp16 = concat(axis = const_2071, interleave = cat_80_interleave_0, values = (neg_28_cast_fp16, slice_383_cast_fp16))[name = string("cat_80_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_310_cast_fp16 = mul(x = cat_80_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_310_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_244_cast_fp16 = add(x = mul_309_cast_fp16, y = mul_310_cast_fp16)[name = string("add_244_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_311_cast_fp16 = mul(x = mul_308_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_311_cast_fp16")];
            tensor<int32, [4]> slice_385_begin_0 = const()[name = string("slice_385_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_385_end_0 = const()[name = string("slice_385_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_385_end_mask_0 = const()[name = string("slice_385_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_385_cast_fp16 = slice_by_index(begin = slice_385_begin_0, end = slice_385_end_0, end_mask = slice_385_end_mask_0, x = mul_308_cast_fp16)[name = string("slice_385_cast_fp16")];
            tensor<int32, [4]> slice_386_begin_0 = const()[name = string("slice_386_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_386_end_0 = const()[name = string("slice_386_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_386_end_mask_0 = const()[name = string("slice_386_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_386_cast_fp16 = slice_by_index(begin = slice_386_begin_0, end = slice_386_end_0, end_mask = slice_386_end_mask_0, x = mul_308_cast_fp16)[name = string("slice_386_cast_fp16")];
            fp16 const_2078_promoted_to_fp16 = const()[name = string("const_2078_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_29_cast_fp16 = mul(x = slice_386_cast_fp16, y = const_2078_promoted_to_fp16)[name = string("neg_29_cast_fp16")];
            int32 const_2079 = const()[name = string("const_2079"), val = int32(-1)];
            bool cat_81_interleave_0 = const()[name = string("cat_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_81_cast_fp16 = concat(axis = const_2079, interleave = cat_81_interleave_0, values = (neg_29_cast_fp16, slice_385_cast_fp16))[name = string("cat_81_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_312_cast_fp16 = mul(x = cat_81_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_312_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_245_cast_fp16 = add(x = mul_311_cast_fp16, y = mul_312_cast_fp16)[name = string("add_245_cast_fp16")];
            int32 const_2080 = const()[name = string("const_2080"), val = int32(-2)];
            bool cat_82_interleave_0 = const()[name = string("cat_82_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_82_cast_fp16 = concat(axis = const_2080, interleave = cat_82_interleave_0, values = add_245_cast_fp16)[name = string("cat_82_cast_fp16")];
            int32 const_2081 = const()[name = string("const_2081"), val = int32(-2)];
            bool cat_83_interleave_0 = const()[name = string("cat_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_96_cast_fp16 = transpose(perm = transpose_96_perm_0, x = view_86_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 1, 512, 256]> cat_83_cast_fp16 = concat(axis = const_2081, interleave = cat_83_interleave_0, values = transpose_96_cast_fp16)[name = string("cat_83_cast_fp16")];
            tensor<int32, [1]> unsqueeze_135_axes_0 = const()[name = string("unsqueeze_135_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_135_cast_fp16 = expand_dims(axes = unsqueeze_135_axes_0, x = cat_82_cast_fp16)[name = string("unsqueeze_135_cast_fp16")];
            tensor<int32, [5]> expand_54_reps_0 = const()[name = string("expand_54_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_54_cast_fp16 = tile(reps = expand_54_reps_0, x = unsqueeze_135_cast_fp16)[name = string("expand_54_cast_fp16")];
            tensor<int32, [4]> const_2096 = const()[name = string("const_2096"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_87_cast_fp16 = reshape(shape = const_2096, x = expand_54_cast_fp16)[name = string("view_87_cast_fp16")];
            tensor<int32, [1]> unsqueeze_136_axes_0 = const()[name = string("unsqueeze_136_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_136_cast_fp16 = expand_dims(axes = unsqueeze_136_axes_0, x = cat_83_cast_fp16)[name = string("unsqueeze_136_cast_fp16")];
            tensor<int32, [5]> expand_55_reps_0 = const()[name = string("expand_55_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_55_cast_fp16 = tile(reps = expand_55_reps_0, x = unsqueeze_136_cast_fp16)[name = string("expand_55_cast_fp16")];
            tensor<int32, [4]> const_2111 = const()[name = string("const_2111"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_88_cast_fp16 = reshape(shape = const_2111, x = expand_55_cast_fp16)[name = string("view_88_cast_fp16")];
            bool matmul_52_transpose_x_1 = const()[name = string("matmul_52_transpose_x_1"), val = bool(false)];
            bool matmul_52_transpose_y_1 = const()[name = string("matmul_52_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_52_cast_fp16 = matmul(transpose_x = matmul_52_transpose_x_1, transpose_y = matmul_52_transpose_y_1, x = add_244_cast_fp16, y = view_87_cast_fp16)[name = string("matmul_52_cast_fp16")];
            fp16 const_2114_to_fp16 = const()[name = string("const_2114_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_313_cast_fp16 = mul(x = matmul_52_cast_fp16, y = const_2114_to_fp16)[name = string("mul_313_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_246_cast_fp16 = add(x = mul_313_cast_fp16, y = expand_cast_fp16)[name = string("add_246_cast_fp16")];
            int32 const_2124 = const()[name = string("const_2124"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_14_cast_fp16 = softmax(axis = const_2124, x = add_246_cast_fp16)[name = string("softmax_14_cast_fp16")];
            bool matmul_53_transpose_x_0 = const()[name = string("matmul_53_transpose_x_0"), val = bool(false)];
            bool matmul_53_transpose_y_0 = const()[name = string("matmul_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_53_cast_fp16 = matmul(transpose_x = matmul_53_transpose_x_0, transpose_y = matmul_53_transpose_y_0, x = softmax_14_cast_fp16, y = view_88_cast_fp16)[name = string("matmul_53_cast_fp16")];
            tensor<int32, [4]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_2129 = const()[name = string("const_2129"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_98_cast_fp16 = transpose(perm = transpose_98_perm_0, x = matmul_53_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 512, 768]> view_89_cast_fp16 = reshape(shape = const_2129, x = transpose_98_cast_fp16)[name = string("view_89_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262809536))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263399424))))[name = string("p_st_0_model_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_101_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = view_89_cast_fp16)[name = string("linear_101_cast_fp16")];
            fp16 const_2131_promoted_to_fp16 = const()[name = string("const_2131_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_88_cast_fp16 = pow(x = linear_101_cast_fp16, y = const_2131_promoted_to_fp16)[name = string("pow_88_cast_fp16")];
            tensor<int32, [1]> mean_87_axes_0 = const()[name = string("mean_87_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_87_keep_dims_0 = const()[name = string("mean_87_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_87_cast_fp16 = reduce_mean(axes = mean_87_axes_0, keep_dims = mean_87_keep_dims_0, x = pow_88_cast_fp16)[name = string("mean_87_cast_fp16")];
            fp16 const_2134_to_fp16 = const()[name = string("const_2134_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_247_cast_fp16 = add(x = mean_87_cast_fp16, y = const_2134_to_fp16)[name = string("add_247_cast_fp16")];
            fp32 rsqrt_87_epsilon_0 = const()[name = string("rsqrt_87_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_87_cast_fp16 = rsqrt(epsilon = rsqrt_87_epsilon_0, x = add_247_cast_fp16)[name = string("rsqrt_87_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_314_cast_fp16 = mul(x = linear_101_cast_fp16, y = rsqrt_87_cast_fp16)[name = string("mul_314_cast_fp16")];
            tensor<fp16, [768]> add_248_to_fp16 = const()[name = string("add_248_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263401024)))];
            tensor<fp16, [1, 512, 768]> mul_315_cast_fp16 = mul(x = mul_314_cast_fp16, y = add_248_to_fp16)[name = string("mul_315_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_249_cast_fp16 = add(x = add_237_cast_fp16, y = mul_315_cast_fp16)[name = string("add_249_cast_fp16")];
            fp16 const_2139_promoted_to_fp16 = const()[name = string("const_2139_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_89_cast_fp16 = pow(x = add_249_cast_fp16, y = const_2139_promoted_to_fp16)[name = string("pow_89_cast_fp16")];
            tensor<int32, [1]> mean_88_axes_0 = const()[name = string("mean_88_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_88_keep_dims_0 = const()[name = string("mean_88_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_88_cast_fp16 = reduce_mean(axes = mean_88_axes_0, keep_dims = mean_88_keep_dims_0, x = pow_89_cast_fp16)[name = string("mean_88_cast_fp16")];
            fp16 const_2142_to_fp16 = const()[name = string("const_2142_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_250_cast_fp16 = add(x = mean_88_cast_fp16, y = const_2142_to_fp16)[name = string("add_250_cast_fp16")];
            fp32 rsqrt_88_epsilon_0 = const()[name = string("rsqrt_88_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_88_cast_fp16 = rsqrt(epsilon = rsqrt_88_epsilon_0, x = add_250_cast_fp16)[name = string("rsqrt_88_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_316_cast_fp16 = mul(x = add_249_cast_fp16, y = rsqrt_88_cast_fp16)[name = string("mul_316_cast_fp16")];
            tensor<fp16, [768]> add_251_to_fp16 = const()[name = string("add_251_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263402624)))];
            tensor<fp16, [1, 512, 768]> mul_317_cast_fp16 = mul(x = mul_316_cast_fp16, y = add_251_to_fp16)[name = string("mul_317_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263404224))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264289024))))[name = string("p_st_0_model_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = mul_317_cast_fp16)[name = string("linear_102_cast_fp16")];
            string gelu_14_mode_0 = const()[name = string("gelu_14_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_14_cast_fp16 = gelu(mode = gelu_14_mode_0, x = linear_102_cast_fp16)[name = string("gelu_14_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264291392))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265176192))))[name = string("p_st_0_model_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = mul_317_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_318_cast_fp16 = mul(x = gelu_14_cast_fp16, y = linear_103_cast_fp16)[name = string("mul_318_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265178560))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266063360))))[name = string("p_st_0_model_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_104_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = mul_318_cast_fp16)[name = string("linear_104_cast_fp16")];
            fp16 const_2147_promoted_to_fp16 = const()[name = string("const_2147_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_90_cast_fp16 = pow(x = linear_104_cast_fp16, y = const_2147_promoted_to_fp16)[name = string("pow_90_cast_fp16")];
            tensor<int32, [1]> mean_89_axes_0 = const()[name = string("mean_89_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_89_keep_dims_0 = const()[name = string("mean_89_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_89_cast_fp16 = reduce_mean(axes = mean_89_axes_0, keep_dims = mean_89_keep_dims_0, x = pow_90_cast_fp16)[name = string("mean_89_cast_fp16")];
            fp16 const_2150_to_fp16 = const()[name = string("const_2150_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_252_cast_fp16 = add(x = mean_89_cast_fp16, y = const_2150_to_fp16)[name = string("add_252_cast_fp16")];
            fp32 rsqrt_89_epsilon_0 = const()[name = string("rsqrt_89_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_89_cast_fp16 = rsqrt(epsilon = rsqrt_89_epsilon_0, x = add_252_cast_fp16)[name = string("rsqrt_89_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_319_cast_fp16 = mul(x = linear_104_cast_fp16, y = rsqrt_89_cast_fp16)[name = string("mul_319_cast_fp16")];
            tensor<fp16, [768]> add_253_to_fp16 = const()[name = string("add_253_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266064960)))];
            tensor<fp16, [1, 512, 768]> mul_320_cast_fp16 = mul(x = mul_319_cast_fp16, y = add_253_to_fp16)[name = string("mul_320_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_254_cast_fp16 = add(x = add_249_cast_fp16, y = mul_320_cast_fp16)[name = string("add_254_cast_fp16")];
            fp16 const_2155_promoted_to_fp16 = const()[name = string("const_2155_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_91_cast_fp16 = pow(x = add_254_cast_fp16, y = const_2155_promoted_to_fp16)[name = string("pow_91_cast_fp16")];
            tensor<int32, [1]> mean_90_axes_0 = const()[name = string("mean_90_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_90_keep_dims_0 = const()[name = string("mean_90_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_90_cast_fp16 = reduce_mean(axes = mean_90_axes_0, keep_dims = mean_90_keep_dims_0, x = pow_91_cast_fp16)[name = string("mean_90_cast_fp16")];
            fp16 const_2158_to_fp16 = const()[name = string("const_2158_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_255_cast_fp16 = add(x = mean_90_cast_fp16, y = const_2158_to_fp16)[name = string("add_255_cast_fp16")];
            fp32 rsqrt_90_epsilon_0 = const()[name = string("rsqrt_90_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_90_cast_fp16 = rsqrt(epsilon = rsqrt_90_epsilon_0, x = add_255_cast_fp16)[name = string("rsqrt_90_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_321_cast_fp16 = mul(x = add_254_cast_fp16, y = rsqrt_90_cast_fp16)[name = string("mul_321_cast_fp16")];
            tensor<fp16, [768]> add_256_to_fp16 = const()[name = string("add_256_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266066560)))];
            tensor<fp16, [1, 512, 768]> mul_322_cast_fp16 = mul(x = mul_321_cast_fp16, y = add_256_to_fp16)[name = string("mul_322_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266068160))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266658048))))[name = string("p_st_0_model_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_105_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = mul_322_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [4]> const_2162 = const()[name = string("const_2162"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_90_cast_fp16 = reshape(shape = const_2162, x = linear_105_cast_fp16)[name = string("view_90_cast_fp16")];
            tensor<int32, [4]> transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266659648))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266856320))))[name = string("p_st_0_model_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_106_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = mul_322_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> const_2165 = const()[name = string("const_2165"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_91_cast_fp16 = reshape(shape = const_2165, x = linear_106_cast_fp16)[name = string("view_91_cast_fp16")];
            tensor<int32, [4]> transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266856896))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267053568))))[name = string("p_st_0_model_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_107_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = mul_322_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> const_2168 = const()[name = string("const_2168"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_92_cast_fp16 = reshape(shape = const_2168, x = linear_107_cast_fp16)[name = string("view_92_cast_fp16")];
            tensor<int32, [4]> transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_2172_promoted_to_fp16 = const()[name = string("const_2172_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_99_cast_fp16 = transpose(perm = transpose_99_perm_0, x = view_90_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 3, 512, 256]> pow_92_cast_fp16 = pow(x = transpose_99_cast_fp16, y = const_2172_promoted_to_fp16)[name = string("pow_92_cast_fp16")];
            tensor<int32, [1]> mean_91_axes_0 = const()[name = string("mean_91_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_91_keep_dims_0 = const()[name = string("mean_91_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_91_cast_fp16 = reduce_mean(axes = mean_91_axes_0, keep_dims = mean_91_keep_dims_0, x = pow_92_cast_fp16)[name = string("mean_91_cast_fp16")];
            fp16 const_2175_to_fp16 = const()[name = string("const_2175_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_257_cast_fp16 = add(x = mean_91_cast_fp16, y = const_2175_to_fp16)[name = string("add_257_cast_fp16")];
            fp32 rsqrt_91_epsilon_0 = const()[name = string("rsqrt_91_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_91_cast_fp16 = rsqrt(epsilon = rsqrt_91_epsilon_0, x = add_257_cast_fp16)[name = string("rsqrt_91_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_323_cast_fp16 = mul(x = transpose_99_cast_fp16, y = rsqrt_91_cast_fp16)[name = string("mul_323_cast_fp16")];
            tensor<fp16, [256]> add_258_to_fp16 = const()[name = string("add_258_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267054144)))];
            tensor<fp16, [1, 3, 512, 256]> mul_324_cast_fp16 = mul(x = mul_323_cast_fp16, y = add_258_to_fp16)[name = string("mul_324_cast_fp16")];
            fp16 const_2180_promoted_to_fp16 = const()[name = string("const_2180_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_100_cast_fp16 = transpose(perm = transpose_100_perm_0, x = view_91_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 1, 512, 256]> pow_93_cast_fp16 = pow(x = transpose_100_cast_fp16, y = const_2180_promoted_to_fp16)[name = string("pow_93_cast_fp16")];
            tensor<int32, [1]> mean_92_axes_0 = const()[name = string("mean_92_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_92_keep_dims_0 = const()[name = string("mean_92_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_92_cast_fp16 = reduce_mean(axes = mean_92_axes_0, keep_dims = mean_92_keep_dims_0, x = pow_93_cast_fp16)[name = string("mean_92_cast_fp16")];
            fp16 const_2183_to_fp16 = const()[name = string("const_2183_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_259_cast_fp16 = add(x = mean_92_cast_fp16, y = const_2183_to_fp16)[name = string("add_259_cast_fp16")];
            fp32 rsqrt_92_epsilon_0 = const()[name = string("rsqrt_92_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_92_cast_fp16 = rsqrt(epsilon = rsqrt_92_epsilon_0, x = add_259_cast_fp16)[name = string("rsqrt_92_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_325_cast_fp16 = mul(x = transpose_100_cast_fp16, y = rsqrt_92_cast_fp16)[name = string("mul_325_cast_fp16")];
            tensor<fp16, [256]> add_260_to_fp16 = const()[name = string("add_260_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267054720)))];
            tensor<fp16, [1, 1, 512, 256]> mul_326_cast_fp16 = mul(x = mul_325_cast_fp16, y = add_260_to_fp16)[name = string("mul_326_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_327_cast_fp16 = mul(x = mul_324_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_327_cast_fp16")];
            tensor<int32, [4]> slice_406_begin_0 = const()[name = string("slice_406_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_406_end_0 = const()[name = string("slice_406_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_406_end_mask_0 = const()[name = string("slice_406_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_406_cast_fp16 = slice_by_index(begin = slice_406_begin_0, end = slice_406_end_0, end_mask = slice_406_end_mask_0, x = mul_324_cast_fp16)[name = string("slice_406_cast_fp16")];
            tensor<int32, [4]> slice_407_begin_0 = const()[name = string("slice_407_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_407_end_0 = const()[name = string("slice_407_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_407_end_mask_0 = const()[name = string("slice_407_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_407_cast_fp16 = slice_by_index(begin = slice_407_begin_0, end = slice_407_end_0, end_mask = slice_407_end_mask_0, x = mul_324_cast_fp16)[name = string("slice_407_cast_fp16")];
            fp16 const_2195_promoted_to_fp16 = const()[name = string("const_2195_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_30_cast_fp16 = mul(x = slice_407_cast_fp16, y = const_2195_promoted_to_fp16)[name = string("neg_30_cast_fp16")];
            int32 const_2196 = const()[name = string("const_2196"), val = int32(-1)];
            bool cat_84_interleave_0 = const()[name = string("cat_84_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_84_cast_fp16 = concat(axis = const_2196, interleave = cat_84_interleave_0, values = (neg_30_cast_fp16, slice_406_cast_fp16))[name = string("cat_84_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_328_cast_fp16 = mul(x = cat_84_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_328_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_261_cast_fp16 = add(x = mul_327_cast_fp16, y = mul_328_cast_fp16)[name = string("add_261_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_329_cast_fp16 = mul(x = mul_326_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_329_cast_fp16")];
            tensor<int32, [4]> slice_408_begin_0 = const()[name = string("slice_408_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_408_end_0 = const()[name = string("slice_408_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_408_end_mask_0 = const()[name = string("slice_408_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_408_cast_fp16 = slice_by_index(begin = slice_408_begin_0, end = slice_408_end_0, end_mask = slice_408_end_mask_0, x = mul_326_cast_fp16)[name = string("slice_408_cast_fp16")];
            tensor<int32, [4]> slice_409_begin_0 = const()[name = string("slice_409_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_409_end_0 = const()[name = string("slice_409_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_409_end_mask_0 = const()[name = string("slice_409_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_409_cast_fp16 = slice_by_index(begin = slice_409_begin_0, end = slice_409_end_0, end_mask = slice_409_end_mask_0, x = mul_326_cast_fp16)[name = string("slice_409_cast_fp16")];
            fp16 const_2203_promoted_to_fp16 = const()[name = string("const_2203_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_31_cast_fp16 = mul(x = slice_409_cast_fp16, y = const_2203_promoted_to_fp16)[name = string("neg_31_cast_fp16")];
            int32 const_2204 = const()[name = string("const_2204"), val = int32(-1)];
            bool cat_85_interleave_0 = const()[name = string("cat_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_85_cast_fp16 = concat(axis = const_2204, interleave = cat_85_interleave_0, values = (neg_31_cast_fp16, slice_408_cast_fp16))[name = string("cat_85_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_330_cast_fp16 = mul(x = cat_85_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_330_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_262_cast_fp16 = add(x = mul_329_cast_fp16, y = mul_330_cast_fp16)[name = string("add_262_cast_fp16")];
            int32 const_2205 = const()[name = string("const_2205"), val = int32(-2)];
            bool cat_86_interleave_0 = const()[name = string("cat_86_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_86_cast_fp16 = concat(axis = const_2205, interleave = cat_86_interleave_0, values = add_262_cast_fp16)[name = string("cat_86_cast_fp16")];
            int32 const_2206 = const()[name = string("const_2206"), val = int32(-2)];
            bool cat_87_interleave_0 = const()[name = string("cat_87_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_101_cast_fp16 = transpose(perm = transpose_101_perm_0, x = view_92_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 1, 512, 256]> cat_87_cast_fp16 = concat(axis = const_2206, interleave = cat_87_interleave_0, values = transpose_101_cast_fp16)[name = string("cat_87_cast_fp16")];
            tensor<int32, [1]> unsqueeze_139_axes_0 = const()[name = string("unsqueeze_139_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_139_cast_fp16 = expand_dims(axes = unsqueeze_139_axes_0, x = cat_86_cast_fp16)[name = string("unsqueeze_139_cast_fp16")];
            tensor<int32, [5]> expand_56_reps_0 = const()[name = string("expand_56_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_56_cast_fp16 = tile(reps = expand_56_reps_0, x = unsqueeze_139_cast_fp16)[name = string("expand_56_cast_fp16")];
            tensor<int32, [4]> const_2221 = const()[name = string("const_2221"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_93_cast_fp16 = reshape(shape = const_2221, x = expand_56_cast_fp16)[name = string("view_93_cast_fp16")];
            tensor<int32, [1]> unsqueeze_140_axes_0 = const()[name = string("unsqueeze_140_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_140_cast_fp16 = expand_dims(axes = unsqueeze_140_axes_0, x = cat_87_cast_fp16)[name = string("unsqueeze_140_cast_fp16")];
            tensor<int32, [5]> expand_57_reps_0 = const()[name = string("expand_57_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_57_cast_fp16 = tile(reps = expand_57_reps_0, x = unsqueeze_140_cast_fp16)[name = string("expand_57_cast_fp16")];
            tensor<int32, [4]> const_2236 = const()[name = string("const_2236"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_94_cast_fp16 = reshape(shape = const_2236, x = expand_57_cast_fp16)[name = string("view_94_cast_fp16")];
            bool matmul_54_transpose_x_1 = const()[name = string("matmul_54_transpose_x_1"), val = bool(false)];
            bool matmul_54_transpose_y_1 = const()[name = string("matmul_54_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_54_cast_fp16 = matmul(transpose_x = matmul_54_transpose_x_1, transpose_y = matmul_54_transpose_y_1, x = add_261_cast_fp16, y = view_93_cast_fp16)[name = string("matmul_54_cast_fp16")];
            fp16 const_2239_to_fp16 = const()[name = string("const_2239_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_331_cast_fp16 = mul(x = matmul_54_cast_fp16, y = const_2239_to_fp16)[name = string("mul_331_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_263_cast_fp16 = add(x = mul_331_cast_fp16, y = expand_cast_fp16)[name = string("add_263_cast_fp16")];
            int32 const_2249 = const()[name = string("const_2249"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_15_cast_fp16 = softmax(axis = const_2249, x = add_263_cast_fp16)[name = string("softmax_15_cast_fp16")];
            bool matmul_55_transpose_x_0 = const()[name = string("matmul_55_transpose_x_0"), val = bool(false)];
            bool matmul_55_transpose_y_0 = const()[name = string("matmul_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_55_cast_fp16 = matmul(transpose_x = matmul_55_transpose_x_0, transpose_y = matmul_55_transpose_y_0, x = softmax_15_cast_fp16, y = view_94_cast_fp16)[name = string("matmul_55_cast_fp16")];
            tensor<int32, [4]> transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_2254 = const()[name = string("const_2254"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_103_cast_fp16 = transpose(perm = transpose_103_perm_0, x = matmul_55_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 512, 768]> view_95_cast_fp16 = reshape(shape = const_2254, x = transpose_103_cast_fp16)[name = string("view_95_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267055296))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267645184))))[name = string("p_st_0_model_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_108_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = view_95_cast_fp16)[name = string("linear_108_cast_fp16")];
            fp16 const_2256_promoted_to_fp16 = const()[name = string("const_2256_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_94_cast_fp16 = pow(x = linear_108_cast_fp16, y = const_2256_promoted_to_fp16)[name = string("pow_94_cast_fp16")];
            tensor<int32, [1]> mean_93_axes_0 = const()[name = string("mean_93_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_93_keep_dims_0 = const()[name = string("mean_93_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_93_cast_fp16 = reduce_mean(axes = mean_93_axes_0, keep_dims = mean_93_keep_dims_0, x = pow_94_cast_fp16)[name = string("mean_93_cast_fp16")];
            fp16 const_2259_to_fp16 = const()[name = string("const_2259_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_264_cast_fp16 = add(x = mean_93_cast_fp16, y = const_2259_to_fp16)[name = string("add_264_cast_fp16")];
            fp32 rsqrt_93_epsilon_0 = const()[name = string("rsqrt_93_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_93_cast_fp16 = rsqrt(epsilon = rsqrt_93_epsilon_0, x = add_264_cast_fp16)[name = string("rsqrt_93_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_332_cast_fp16 = mul(x = linear_108_cast_fp16, y = rsqrt_93_cast_fp16)[name = string("mul_332_cast_fp16")];
            tensor<fp16, [768]> add_265_to_fp16 = const()[name = string("add_265_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267646784)))];
            tensor<fp16, [1, 512, 768]> mul_333_cast_fp16 = mul(x = mul_332_cast_fp16, y = add_265_to_fp16)[name = string("mul_333_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_266_cast_fp16 = add(x = add_254_cast_fp16, y = mul_333_cast_fp16)[name = string("add_266_cast_fp16")];
            fp16 const_2264_promoted_to_fp16 = const()[name = string("const_2264_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_95_cast_fp16 = pow(x = add_266_cast_fp16, y = const_2264_promoted_to_fp16)[name = string("pow_95_cast_fp16")];
            tensor<int32, [1]> mean_94_axes_0 = const()[name = string("mean_94_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_94_keep_dims_0 = const()[name = string("mean_94_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_94_cast_fp16 = reduce_mean(axes = mean_94_axes_0, keep_dims = mean_94_keep_dims_0, x = pow_95_cast_fp16)[name = string("mean_94_cast_fp16")];
            fp16 const_2267_to_fp16 = const()[name = string("const_2267_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_267_cast_fp16 = add(x = mean_94_cast_fp16, y = const_2267_to_fp16)[name = string("add_267_cast_fp16")];
            fp32 rsqrt_94_epsilon_0 = const()[name = string("rsqrt_94_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_94_cast_fp16 = rsqrt(epsilon = rsqrt_94_epsilon_0, x = add_267_cast_fp16)[name = string("rsqrt_94_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_334_cast_fp16 = mul(x = add_266_cast_fp16, y = rsqrt_94_cast_fp16)[name = string("mul_334_cast_fp16")];
            tensor<fp16, [768]> add_268_to_fp16 = const()[name = string("add_268_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267648384)))];
            tensor<fp16, [1, 512, 768]> mul_335_cast_fp16 = mul(x = mul_334_cast_fp16, y = add_268_to_fp16)[name = string("mul_335_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267649984))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268534784))))[name = string("p_st_0_model_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = mul_335_cast_fp16)[name = string("linear_109_cast_fp16")];
            string gelu_15_mode_0 = const()[name = string("gelu_15_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_15_cast_fp16 = gelu(mode = gelu_15_mode_0, x = linear_109_cast_fp16)[name = string("gelu_15_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268537152))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269421952))))[name = string("p_st_0_model_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = mul_335_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_336_cast_fp16 = mul(x = gelu_15_cast_fp16, y = linear_110_cast_fp16)[name = string("mul_336_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269424320))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270309120))))[name = string("p_st_0_model_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_111_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = mul_336_cast_fp16)[name = string("linear_111_cast_fp16")];
            fp16 const_2272_promoted_to_fp16 = const()[name = string("const_2272_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_96_cast_fp16 = pow(x = linear_111_cast_fp16, y = const_2272_promoted_to_fp16)[name = string("pow_96_cast_fp16")];
            tensor<int32, [1]> mean_95_axes_0 = const()[name = string("mean_95_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_95_keep_dims_0 = const()[name = string("mean_95_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_95_cast_fp16 = reduce_mean(axes = mean_95_axes_0, keep_dims = mean_95_keep_dims_0, x = pow_96_cast_fp16)[name = string("mean_95_cast_fp16")];
            fp16 const_2275_to_fp16 = const()[name = string("const_2275_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_269_cast_fp16 = add(x = mean_95_cast_fp16, y = const_2275_to_fp16)[name = string("add_269_cast_fp16")];
            fp32 rsqrt_95_epsilon_0 = const()[name = string("rsqrt_95_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_95_cast_fp16 = rsqrt(epsilon = rsqrt_95_epsilon_0, x = add_269_cast_fp16)[name = string("rsqrt_95_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_337_cast_fp16 = mul(x = linear_111_cast_fp16, y = rsqrt_95_cast_fp16)[name = string("mul_337_cast_fp16")];
            tensor<fp16, [768]> add_270_to_fp16 = const()[name = string("add_270_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270310720)))];
            tensor<fp16, [1, 512, 768]> mul_338_cast_fp16 = mul(x = mul_337_cast_fp16, y = add_270_to_fp16)[name = string("mul_338_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_271_cast_fp16 = add(x = add_266_cast_fp16, y = mul_338_cast_fp16)[name = string("add_271_cast_fp16")];
            fp16 const_2280_promoted_to_fp16 = const()[name = string("const_2280_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_97_cast_fp16 = pow(x = add_271_cast_fp16, y = const_2280_promoted_to_fp16)[name = string("pow_97_cast_fp16")];
            tensor<int32, [1]> mean_96_axes_0 = const()[name = string("mean_96_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_96_keep_dims_0 = const()[name = string("mean_96_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_96_cast_fp16 = reduce_mean(axes = mean_96_axes_0, keep_dims = mean_96_keep_dims_0, x = pow_97_cast_fp16)[name = string("mean_96_cast_fp16")];
            fp16 const_2283_to_fp16 = const()[name = string("const_2283_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_272_cast_fp16 = add(x = mean_96_cast_fp16, y = const_2283_to_fp16)[name = string("add_272_cast_fp16")];
            fp32 rsqrt_96_epsilon_0 = const()[name = string("rsqrt_96_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_96_cast_fp16 = rsqrt(epsilon = rsqrt_96_epsilon_0, x = add_272_cast_fp16)[name = string("rsqrt_96_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_339_cast_fp16 = mul(x = add_271_cast_fp16, y = rsqrt_96_cast_fp16)[name = string("mul_339_cast_fp16")];
            tensor<fp16, [768]> add_273_to_fp16 = const()[name = string("add_273_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270312320)))];
            tensor<fp16, [1, 512, 768]> mul_340_cast_fp16 = mul(x = mul_339_cast_fp16, y = add_273_to_fp16)[name = string("mul_340_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270313920))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270903808))))[name = string("p_st_0_model_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_112_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = mul_340_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [4]> const_2287 = const()[name = string("const_2287"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_96_cast_fp16 = reshape(shape = const_2287, x = linear_112_cast_fp16)[name = string("view_96_cast_fp16")];
            tensor<int32, [4]> transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270905408))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271102080))))[name = string("p_st_0_model_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_113_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = mul_340_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [4]> const_2290 = const()[name = string("const_2290"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_97_cast_fp16 = reshape(shape = const_2290, x = linear_113_cast_fp16)[name = string("view_97_cast_fp16")];
            tensor<int32, [4]> transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271102656))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271299328))))[name = string("p_st_0_model_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_114_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = mul_340_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [4]> const_2293 = const()[name = string("const_2293"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_98_cast_fp16 = reshape(shape = const_2293, x = linear_114_cast_fp16)[name = string("view_98_cast_fp16")];
            tensor<int32, [4]> transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_2297_promoted_to_fp16 = const()[name = string("const_2297_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_104_cast_fp16 = transpose(perm = transpose_104_perm_0, x = view_96_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 3, 512, 256]> pow_98_cast_fp16 = pow(x = transpose_104_cast_fp16, y = const_2297_promoted_to_fp16)[name = string("pow_98_cast_fp16")];
            tensor<int32, [1]> mean_97_axes_0 = const()[name = string("mean_97_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_97_keep_dims_0 = const()[name = string("mean_97_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_97_cast_fp16 = reduce_mean(axes = mean_97_axes_0, keep_dims = mean_97_keep_dims_0, x = pow_98_cast_fp16)[name = string("mean_97_cast_fp16")];
            fp16 const_2300_to_fp16 = const()[name = string("const_2300_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_274_cast_fp16 = add(x = mean_97_cast_fp16, y = const_2300_to_fp16)[name = string("add_274_cast_fp16")];
            fp32 rsqrt_97_epsilon_0 = const()[name = string("rsqrt_97_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_97_cast_fp16 = rsqrt(epsilon = rsqrt_97_epsilon_0, x = add_274_cast_fp16)[name = string("rsqrt_97_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_341_cast_fp16 = mul(x = transpose_104_cast_fp16, y = rsqrt_97_cast_fp16)[name = string("mul_341_cast_fp16")];
            tensor<fp16, [256]> add_275_to_fp16 = const()[name = string("add_275_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271299904)))];
            tensor<fp16, [1, 3, 512, 256]> mul_342_cast_fp16 = mul(x = mul_341_cast_fp16, y = add_275_to_fp16)[name = string("mul_342_cast_fp16")];
            fp16 const_2305_promoted_to_fp16 = const()[name = string("const_2305_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_105_cast_fp16 = transpose(perm = transpose_105_perm_0, x = view_97_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 1, 512, 256]> pow_99_cast_fp16 = pow(x = transpose_105_cast_fp16, y = const_2305_promoted_to_fp16)[name = string("pow_99_cast_fp16")];
            tensor<int32, [1]> mean_98_axes_0 = const()[name = string("mean_98_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_98_keep_dims_0 = const()[name = string("mean_98_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_98_cast_fp16 = reduce_mean(axes = mean_98_axes_0, keep_dims = mean_98_keep_dims_0, x = pow_99_cast_fp16)[name = string("mean_98_cast_fp16")];
            fp16 const_2308_to_fp16 = const()[name = string("const_2308_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_276_cast_fp16 = add(x = mean_98_cast_fp16, y = const_2308_to_fp16)[name = string("add_276_cast_fp16")];
            fp32 rsqrt_98_epsilon_0 = const()[name = string("rsqrt_98_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_98_cast_fp16 = rsqrt(epsilon = rsqrt_98_epsilon_0, x = add_276_cast_fp16)[name = string("rsqrt_98_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_343_cast_fp16 = mul(x = transpose_105_cast_fp16, y = rsqrt_98_cast_fp16)[name = string("mul_343_cast_fp16")];
            tensor<fp16, [256]> add_277_to_fp16 = const()[name = string("add_277_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271300480)))];
            tensor<fp16, [1, 1, 512, 256]> mul_344_cast_fp16 = mul(x = mul_343_cast_fp16, y = add_277_to_fp16)[name = string("mul_344_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_345_cast_fp16 = mul(x = mul_342_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_345_cast_fp16")];
            tensor<int32, [4]> slice_429_begin_0 = const()[name = string("slice_429_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_429_end_0 = const()[name = string("slice_429_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_429_end_mask_0 = const()[name = string("slice_429_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_429_cast_fp16 = slice_by_index(begin = slice_429_begin_0, end = slice_429_end_0, end_mask = slice_429_end_mask_0, x = mul_342_cast_fp16)[name = string("slice_429_cast_fp16")];
            tensor<int32, [4]> slice_430_begin_0 = const()[name = string("slice_430_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_430_end_0 = const()[name = string("slice_430_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_430_end_mask_0 = const()[name = string("slice_430_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_430_cast_fp16 = slice_by_index(begin = slice_430_begin_0, end = slice_430_end_0, end_mask = slice_430_end_mask_0, x = mul_342_cast_fp16)[name = string("slice_430_cast_fp16")];
            fp16 const_2320_promoted_to_fp16 = const()[name = string("const_2320_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_32_cast_fp16 = mul(x = slice_430_cast_fp16, y = const_2320_promoted_to_fp16)[name = string("neg_32_cast_fp16")];
            int32 const_2321 = const()[name = string("const_2321"), val = int32(-1)];
            bool cat_88_interleave_0 = const()[name = string("cat_88_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_88_cast_fp16 = concat(axis = const_2321, interleave = cat_88_interleave_0, values = (neg_32_cast_fp16, slice_429_cast_fp16))[name = string("cat_88_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_346_cast_fp16 = mul(x = cat_88_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_346_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_278_cast_fp16 = add(x = mul_345_cast_fp16, y = mul_346_cast_fp16)[name = string("add_278_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_347_cast_fp16 = mul(x = mul_344_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_347_cast_fp16")];
            tensor<int32, [4]> slice_431_begin_0 = const()[name = string("slice_431_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_431_end_0 = const()[name = string("slice_431_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_431_end_mask_0 = const()[name = string("slice_431_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_431_cast_fp16 = slice_by_index(begin = slice_431_begin_0, end = slice_431_end_0, end_mask = slice_431_end_mask_0, x = mul_344_cast_fp16)[name = string("slice_431_cast_fp16")];
            tensor<int32, [4]> slice_432_begin_0 = const()[name = string("slice_432_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_432_end_0 = const()[name = string("slice_432_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_432_end_mask_0 = const()[name = string("slice_432_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_432_cast_fp16 = slice_by_index(begin = slice_432_begin_0, end = slice_432_end_0, end_mask = slice_432_end_mask_0, x = mul_344_cast_fp16)[name = string("slice_432_cast_fp16")];
            fp16 const_2328_promoted_to_fp16 = const()[name = string("const_2328_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_33_cast_fp16 = mul(x = slice_432_cast_fp16, y = const_2328_promoted_to_fp16)[name = string("neg_33_cast_fp16")];
            int32 const_2329 = const()[name = string("const_2329"), val = int32(-1)];
            bool cat_89_interleave_0 = const()[name = string("cat_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_89_cast_fp16 = concat(axis = const_2329, interleave = cat_89_interleave_0, values = (neg_33_cast_fp16, slice_431_cast_fp16))[name = string("cat_89_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_348_cast_fp16 = mul(x = cat_89_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_348_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_279_cast_fp16 = add(x = mul_347_cast_fp16, y = mul_348_cast_fp16)[name = string("add_279_cast_fp16")];
            int32 const_2330 = const()[name = string("const_2330"), val = int32(-2)];
            bool cat_90_interleave_0 = const()[name = string("cat_90_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_90_cast_fp16 = concat(axis = const_2330, interleave = cat_90_interleave_0, values = add_279_cast_fp16)[name = string("cat_90_cast_fp16")];
            int32 const_2331 = const()[name = string("const_2331"), val = int32(-2)];
            bool cat_91_interleave_0 = const()[name = string("cat_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_106_cast_fp16 = transpose(perm = transpose_106_perm_0, x = view_98_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 1, 512, 256]> cat_91_cast_fp16 = concat(axis = const_2331, interleave = cat_91_interleave_0, values = transpose_106_cast_fp16)[name = string("cat_91_cast_fp16")];
            tensor<int32, [1]> unsqueeze_143_axes_0 = const()[name = string("unsqueeze_143_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_143_cast_fp16 = expand_dims(axes = unsqueeze_143_axes_0, x = cat_90_cast_fp16)[name = string("unsqueeze_143_cast_fp16")];
            tensor<int32, [5]> expand_58_reps_0 = const()[name = string("expand_58_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_58_cast_fp16 = tile(reps = expand_58_reps_0, x = unsqueeze_143_cast_fp16)[name = string("expand_58_cast_fp16")];
            tensor<int32, [4]> const_2346 = const()[name = string("const_2346"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_99_cast_fp16 = reshape(shape = const_2346, x = expand_58_cast_fp16)[name = string("view_99_cast_fp16")];
            tensor<int32, [1]> unsqueeze_144_axes_0 = const()[name = string("unsqueeze_144_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_144_cast_fp16 = expand_dims(axes = unsqueeze_144_axes_0, x = cat_91_cast_fp16)[name = string("unsqueeze_144_cast_fp16")];
            tensor<int32, [5]> expand_59_reps_0 = const()[name = string("expand_59_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_59_cast_fp16 = tile(reps = expand_59_reps_0, x = unsqueeze_144_cast_fp16)[name = string("expand_59_cast_fp16")];
            tensor<int32, [4]> const_2361 = const()[name = string("const_2361"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_100_cast_fp16 = reshape(shape = const_2361, x = expand_59_cast_fp16)[name = string("view_100_cast_fp16")];
            bool matmul_56_transpose_x_1 = const()[name = string("matmul_56_transpose_x_1"), val = bool(false)];
            bool matmul_56_transpose_y_1 = const()[name = string("matmul_56_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_56_cast_fp16 = matmul(transpose_x = matmul_56_transpose_x_1, transpose_y = matmul_56_transpose_y_1, x = add_278_cast_fp16, y = view_99_cast_fp16)[name = string("matmul_56_cast_fp16")];
            fp16 const_2364_to_fp16 = const()[name = string("const_2364_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_349_cast_fp16 = mul(x = matmul_56_cast_fp16, y = const_2364_to_fp16)[name = string("mul_349_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_280_cast_fp16 = add(x = mul_349_cast_fp16, y = expand_cast_fp16)[name = string("add_280_cast_fp16")];
            int32 const_2374 = const()[name = string("const_2374"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_16_cast_fp16 = softmax(axis = const_2374, x = add_280_cast_fp16)[name = string("softmax_16_cast_fp16")];
            bool matmul_57_transpose_x_0 = const()[name = string("matmul_57_transpose_x_0"), val = bool(false)];
            bool matmul_57_transpose_y_0 = const()[name = string("matmul_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_57_cast_fp16 = matmul(transpose_x = matmul_57_transpose_x_0, transpose_y = matmul_57_transpose_y_0, x = softmax_16_cast_fp16, y = view_100_cast_fp16)[name = string("matmul_57_cast_fp16")];
            tensor<int32, [4]> transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_2379 = const()[name = string("const_2379"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_108_cast_fp16 = transpose(perm = transpose_108_perm_0, x = matmul_57_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 512, 768]> view_101_cast_fp16 = reshape(shape = const_2379, x = transpose_108_cast_fp16)[name = string("view_101_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271301056))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271890944))))[name = string("p_st_0_model_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_115_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = view_101_cast_fp16)[name = string("linear_115_cast_fp16")];
            fp16 const_2381_promoted_to_fp16 = const()[name = string("const_2381_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_100_cast_fp16 = pow(x = linear_115_cast_fp16, y = const_2381_promoted_to_fp16)[name = string("pow_100_cast_fp16")];
            tensor<int32, [1]> mean_99_axes_0 = const()[name = string("mean_99_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_99_keep_dims_0 = const()[name = string("mean_99_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_99_cast_fp16 = reduce_mean(axes = mean_99_axes_0, keep_dims = mean_99_keep_dims_0, x = pow_100_cast_fp16)[name = string("mean_99_cast_fp16")];
            fp16 const_2384_to_fp16 = const()[name = string("const_2384_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_281_cast_fp16 = add(x = mean_99_cast_fp16, y = const_2384_to_fp16)[name = string("add_281_cast_fp16")];
            fp32 rsqrt_99_epsilon_0 = const()[name = string("rsqrt_99_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_99_cast_fp16 = rsqrt(epsilon = rsqrt_99_epsilon_0, x = add_281_cast_fp16)[name = string("rsqrt_99_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_350_cast_fp16 = mul(x = linear_115_cast_fp16, y = rsqrt_99_cast_fp16)[name = string("mul_350_cast_fp16")];
            tensor<fp16, [768]> add_282_to_fp16 = const()[name = string("add_282_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271892544)))];
            tensor<fp16, [1, 512, 768]> mul_351_cast_fp16 = mul(x = mul_350_cast_fp16, y = add_282_to_fp16)[name = string("mul_351_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_283_cast_fp16 = add(x = add_271_cast_fp16, y = mul_351_cast_fp16)[name = string("add_283_cast_fp16")];
            fp16 const_2389_promoted_to_fp16 = const()[name = string("const_2389_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_101_cast_fp16 = pow(x = add_283_cast_fp16, y = const_2389_promoted_to_fp16)[name = string("pow_101_cast_fp16")];
            tensor<int32, [1]> mean_100_axes_0 = const()[name = string("mean_100_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_100_keep_dims_0 = const()[name = string("mean_100_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_100_cast_fp16 = reduce_mean(axes = mean_100_axes_0, keep_dims = mean_100_keep_dims_0, x = pow_101_cast_fp16)[name = string("mean_100_cast_fp16")];
            fp16 const_2392_to_fp16 = const()[name = string("const_2392_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_284_cast_fp16 = add(x = mean_100_cast_fp16, y = const_2392_to_fp16)[name = string("add_284_cast_fp16")];
            fp32 rsqrt_100_epsilon_0 = const()[name = string("rsqrt_100_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_100_cast_fp16 = rsqrt(epsilon = rsqrt_100_epsilon_0, x = add_284_cast_fp16)[name = string("rsqrt_100_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_352_cast_fp16 = mul(x = add_283_cast_fp16, y = rsqrt_100_cast_fp16)[name = string("mul_352_cast_fp16")];
            tensor<fp16, [768]> add_285_to_fp16 = const()[name = string("add_285_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271894144)))];
            tensor<fp16, [1, 512, 768]> mul_353_cast_fp16 = mul(x = mul_352_cast_fp16, y = add_285_to_fp16)[name = string("mul_353_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271895744))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272780544))))[name = string("p_st_0_model_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = mul_353_cast_fp16)[name = string("linear_116_cast_fp16")];
            string gelu_16_mode_0 = const()[name = string("gelu_16_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_16_cast_fp16 = gelu(mode = gelu_16_mode_0, x = linear_116_cast_fp16)[name = string("gelu_16_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272782912))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273667712))))[name = string("p_st_0_model_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = mul_353_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_354_cast_fp16 = mul(x = gelu_16_cast_fp16, y = linear_117_cast_fp16)[name = string("mul_354_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273670080))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274554880))))[name = string("p_st_0_model_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_118_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = mul_354_cast_fp16)[name = string("linear_118_cast_fp16")];
            fp16 const_2397_promoted_to_fp16 = const()[name = string("const_2397_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_102_cast_fp16 = pow(x = linear_118_cast_fp16, y = const_2397_promoted_to_fp16)[name = string("pow_102_cast_fp16")];
            tensor<int32, [1]> mean_101_axes_0 = const()[name = string("mean_101_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_101_keep_dims_0 = const()[name = string("mean_101_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_101_cast_fp16 = reduce_mean(axes = mean_101_axes_0, keep_dims = mean_101_keep_dims_0, x = pow_102_cast_fp16)[name = string("mean_101_cast_fp16")];
            fp16 const_2400_to_fp16 = const()[name = string("const_2400_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_286_cast_fp16 = add(x = mean_101_cast_fp16, y = const_2400_to_fp16)[name = string("add_286_cast_fp16")];
            fp32 rsqrt_101_epsilon_0 = const()[name = string("rsqrt_101_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_101_cast_fp16 = rsqrt(epsilon = rsqrt_101_epsilon_0, x = add_286_cast_fp16)[name = string("rsqrt_101_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_355_cast_fp16 = mul(x = linear_118_cast_fp16, y = rsqrt_101_cast_fp16)[name = string("mul_355_cast_fp16")];
            tensor<fp16, [768]> add_287_to_fp16 = const()[name = string("add_287_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274556480)))];
            tensor<fp16, [1, 512, 768]> mul_356_cast_fp16 = mul(x = mul_355_cast_fp16, y = add_287_to_fp16)[name = string("mul_356_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_288_cast_fp16 = add(x = add_283_cast_fp16, y = mul_356_cast_fp16)[name = string("add_288_cast_fp16")];
            fp16 const_2405_promoted_to_fp16 = const()[name = string("const_2405_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_103_cast_fp16 = pow(x = add_288_cast_fp16, y = const_2405_promoted_to_fp16)[name = string("pow_103_cast_fp16")];
            tensor<int32, [1]> mean_102_axes_0 = const()[name = string("mean_102_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_102_keep_dims_0 = const()[name = string("mean_102_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_102_cast_fp16 = reduce_mean(axes = mean_102_axes_0, keep_dims = mean_102_keep_dims_0, x = pow_103_cast_fp16)[name = string("mean_102_cast_fp16")];
            fp16 const_2408_to_fp16 = const()[name = string("const_2408_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_289_cast_fp16 = add(x = mean_102_cast_fp16, y = const_2408_to_fp16)[name = string("add_289_cast_fp16")];
            fp32 rsqrt_102_epsilon_0 = const()[name = string("rsqrt_102_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_102_cast_fp16 = rsqrt(epsilon = rsqrt_102_epsilon_0, x = add_289_cast_fp16)[name = string("rsqrt_102_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_357_cast_fp16 = mul(x = add_288_cast_fp16, y = rsqrt_102_cast_fp16)[name = string("mul_357_cast_fp16")];
            tensor<fp16, [768]> add_290_to_fp16 = const()[name = string("add_290_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274558080)))];
            tensor<fp16, [1, 512, 768]> mul_358_cast_fp16 = mul(x = mul_357_cast_fp16, y = add_290_to_fp16)[name = string("mul_358_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274559680))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275149568))))[name = string("p_st_0_model_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_119_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = mul_358_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> const_2412 = const()[name = string("const_2412"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_102_cast_fp16 = reshape(shape = const_2412, x = linear_119_cast_fp16)[name = string("view_102_cast_fp16")];
            tensor<int32, [4]> transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275151168))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275347840))))[name = string("p_st_0_model_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_120_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = mul_358_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> const_2415 = const()[name = string("const_2415"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_103_cast_fp16 = reshape(shape = const_2415, x = linear_120_cast_fp16)[name = string("view_103_cast_fp16")];
            tensor<int32, [4]> transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275348416))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275545088))))[name = string("p_st_0_model_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_121_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = mul_358_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> const_2418 = const()[name = string("const_2418"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_104_cast_fp16 = reshape(shape = const_2418, x = linear_121_cast_fp16)[name = string("view_104_cast_fp16")];
            tensor<int32, [4]> transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_2422_promoted_to_fp16 = const()[name = string("const_2422_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_109_cast_fp16 = transpose(perm = transpose_109_perm_0, x = view_102_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 3, 512, 256]> pow_104_cast_fp16 = pow(x = transpose_109_cast_fp16, y = const_2422_promoted_to_fp16)[name = string("pow_104_cast_fp16")];
            tensor<int32, [1]> mean_103_axes_0 = const()[name = string("mean_103_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_103_keep_dims_0 = const()[name = string("mean_103_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_103_cast_fp16 = reduce_mean(axes = mean_103_axes_0, keep_dims = mean_103_keep_dims_0, x = pow_104_cast_fp16)[name = string("mean_103_cast_fp16")];
            fp16 const_2425_to_fp16 = const()[name = string("const_2425_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_291_cast_fp16 = add(x = mean_103_cast_fp16, y = const_2425_to_fp16)[name = string("add_291_cast_fp16")];
            fp32 rsqrt_103_epsilon_0 = const()[name = string("rsqrt_103_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_103_cast_fp16 = rsqrt(epsilon = rsqrt_103_epsilon_0, x = add_291_cast_fp16)[name = string("rsqrt_103_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_359_cast_fp16 = mul(x = transpose_109_cast_fp16, y = rsqrt_103_cast_fp16)[name = string("mul_359_cast_fp16")];
            tensor<fp16, [256]> add_292_to_fp16 = const()[name = string("add_292_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275545664)))];
            tensor<fp16, [1, 3, 512, 256]> mul_360_cast_fp16 = mul(x = mul_359_cast_fp16, y = add_292_to_fp16)[name = string("mul_360_cast_fp16")];
            fp16 const_2430_promoted_to_fp16 = const()[name = string("const_2430_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_110_cast_fp16 = transpose(perm = transpose_110_perm_0, x = view_103_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 1, 512, 256]> pow_105_cast_fp16 = pow(x = transpose_110_cast_fp16, y = const_2430_promoted_to_fp16)[name = string("pow_105_cast_fp16")];
            tensor<int32, [1]> mean_104_axes_0 = const()[name = string("mean_104_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_104_keep_dims_0 = const()[name = string("mean_104_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_104_cast_fp16 = reduce_mean(axes = mean_104_axes_0, keep_dims = mean_104_keep_dims_0, x = pow_105_cast_fp16)[name = string("mean_104_cast_fp16")];
            fp16 const_2433_to_fp16 = const()[name = string("const_2433_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_293_cast_fp16 = add(x = mean_104_cast_fp16, y = const_2433_to_fp16)[name = string("add_293_cast_fp16")];
            fp32 rsqrt_104_epsilon_0 = const()[name = string("rsqrt_104_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_104_cast_fp16 = rsqrt(epsilon = rsqrt_104_epsilon_0, x = add_293_cast_fp16)[name = string("rsqrt_104_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_361_cast_fp16 = mul(x = transpose_110_cast_fp16, y = rsqrt_104_cast_fp16)[name = string("mul_361_cast_fp16")];
            tensor<fp16, [256]> add_294_to_fp16 = const()[name = string("add_294_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275546240)))];
            tensor<fp16, [1, 1, 512, 256]> mul_362_cast_fp16 = mul(x = mul_361_cast_fp16, y = add_294_to_fp16)[name = string("mul_362_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_363_cast_fp16 = mul(x = mul_360_cast_fp16, y = unsqueeze_97_to_fp16_quantized)[name = string("mul_363_cast_fp16")];
            tensor<int32, [4]> slice_452_begin_0 = const()[name = string("slice_452_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_452_end_0 = const()[name = string("slice_452_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_452_end_mask_0 = const()[name = string("slice_452_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_452_cast_fp16 = slice_by_index(begin = slice_452_begin_0, end = slice_452_end_0, end_mask = slice_452_end_mask_0, x = mul_360_cast_fp16)[name = string("slice_452_cast_fp16")];
            tensor<int32, [4]> slice_453_begin_0 = const()[name = string("slice_453_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_453_end_0 = const()[name = string("slice_453_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_453_end_mask_0 = const()[name = string("slice_453_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_453_cast_fp16 = slice_by_index(begin = slice_453_begin_0, end = slice_453_end_0, end_mask = slice_453_end_mask_0, x = mul_360_cast_fp16)[name = string("slice_453_cast_fp16")];
            fp16 const_2445_promoted_to_fp16 = const()[name = string("const_2445_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_34_cast_fp16 = mul(x = slice_453_cast_fp16, y = const_2445_promoted_to_fp16)[name = string("neg_34_cast_fp16")];
            int32 const_2446 = const()[name = string("const_2446"), val = int32(-1)];
            bool cat_92_interleave_0 = const()[name = string("cat_92_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_92_cast_fp16 = concat(axis = const_2446, interleave = cat_92_interleave_0, values = (neg_34_cast_fp16, slice_452_cast_fp16))[name = string("cat_92_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_364_cast_fp16 = mul(x = cat_92_cast_fp16, y = unsqueeze_98_to_fp16_quantized)[name = string("mul_364_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_295_cast_fp16 = add(x = mul_363_cast_fp16, y = mul_364_cast_fp16)[name = string("add_295_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_365_cast_fp16 = mul(x = mul_362_cast_fp16, y = unsqueeze_97_to_fp16_quantized)[name = string("mul_365_cast_fp16")];
            tensor<int32, [4]> slice_454_begin_0 = const()[name = string("slice_454_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_454_end_0 = const()[name = string("slice_454_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_454_end_mask_0 = const()[name = string("slice_454_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_454_cast_fp16 = slice_by_index(begin = slice_454_begin_0, end = slice_454_end_0, end_mask = slice_454_end_mask_0, x = mul_362_cast_fp16)[name = string("slice_454_cast_fp16")];
            tensor<int32, [4]> slice_455_begin_0 = const()[name = string("slice_455_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_455_end_0 = const()[name = string("slice_455_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_455_end_mask_0 = const()[name = string("slice_455_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_455_cast_fp16 = slice_by_index(begin = slice_455_begin_0, end = slice_455_end_0, end_mask = slice_455_end_mask_0, x = mul_362_cast_fp16)[name = string("slice_455_cast_fp16")];
            fp16 const_2453_promoted_to_fp16 = const()[name = string("const_2453_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_35_cast_fp16 = mul(x = slice_455_cast_fp16, y = const_2453_promoted_to_fp16)[name = string("neg_35_cast_fp16")];
            int32 const_2454 = const()[name = string("const_2454"), val = int32(-1)];
            bool cat_93_interleave_0 = const()[name = string("cat_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_93_cast_fp16 = concat(axis = const_2454, interleave = cat_93_interleave_0, values = (neg_35_cast_fp16, slice_454_cast_fp16))[name = string("cat_93_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_366_cast_fp16 = mul(x = cat_93_cast_fp16, y = unsqueeze_98_to_fp16_quantized)[name = string("mul_366_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_296_cast_fp16 = add(x = mul_365_cast_fp16, y = mul_366_cast_fp16)[name = string("add_296_cast_fp16")];
            int32 const_2455 = const()[name = string("const_2455"), val = int32(-2)];
            bool cat_94_interleave_0 = const()[name = string("cat_94_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_94_cast_fp16 = concat(axis = const_2455, interleave = cat_94_interleave_0, values = add_296_cast_fp16)[name = string("cat_94_cast_fp16")];
            int32 const_2456 = const()[name = string("const_2456"), val = int32(-2)];
            bool cat_95_interleave_0 = const()[name = string("cat_95_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_111_cast_fp16 = transpose(perm = transpose_111_perm_0, x = view_104_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 1, 512, 256]> cat_95_cast_fp16 = concat(axis = const_2456, interleave = cat_95_interleave_0, values = transpose_111_cast_fp16)[name = string("cat_95_cast_fp16")];
            tensor<int32, [1]> unsqueeze_147_axes_0 = const()[name = string("unsqueeze_147_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_147_cast_fp16 = expand_dims(axes = unsqueeze_147_axes_0, x = cat_94_cast_fp16)[name = string("unsqueeze_147_cast_fp16")];
            tensor<int32, [5]> expand_60_reps_0 = const()[name = string("expand_60_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_60_cast_fp16 = tile(reps = expand_60_reps_0, x = unsqueeze_147_cast_fp16)[name = string("expand_60_cast_fp16")];
            tensor<int32, [4]> const_2471 = const()[name = string("const_2471"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_105_cast_fp16 = reshape(shape = const_2471, x = expand_60_cast_fp16)[name = string("view_105_cast_fp16")];
            tensor<int32, [1]> unsqueeze_148_axes_0 = const()[name = string("unsqueeze_148_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_148_cast_fp16 = expand_dims(axes = unsqueeze_148_axes_0, x = cat_95_cast_fp16)[name = string("unsqueeze_148_cast_fp16")];
            tensor<int32, [5]> expand_61_reps_0 = const()[name = string("expand_61_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_61_cast_fp16 = tile(reps = expand_61_reps_0, x = unsqueeze_148_cast_fp16)[name = string("expand_61_cast_fp16")];
            tensor<int32, [4]> const_2486 = const()[name = string("const_2486"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_106_cast_fp16 = reshape(shape = const_2486, x = expand_61_cast_fp16)[name = string("view_106_cast_fp16")];
            bool matmul_58_transpose_x_1 = const()[name = string("matmul_58_transpose_x_1"), val = bool(false)];
            bool matmul_58_transpose_y_1 = const()[name = string("matmul_58_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_58_cast_fp16 = matmul(transpose_x = matmul_58_transpose_x_1, transpose_y = matmul_58_transpose_y_1, x = add_295_cast_fp16, y = view_105_cast_fp16)[name = string("matmul_58_cast_fp16")];
            fp16 const_2489_to_fp16 = const()[name = string("const_2489_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_367_cast_fp16 = mul(x = matmul_58_cast_fp16, y = const_2489_to_fp16)[name = string("mul_367_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_297_cast_fp16 = add(x = mul_367_cast_fp16, y = expand_cast_fp16)[name = string("add_297_cast_fp16")];
            int32 const_2499 = const()[name = string("const_2499"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_17_cast_fp16 = softmax(axis = const_2499, x = add_297_cast_fp16)[name = string("softmax_17_cast_fp16")];
            bool matmul_59_transpose_x_0 = const()[name = string("matmul_59_transpose_x_0"), val = bool(false)];
            bool matmul_59_transpose_y_0 = const()[name = string("matmul_59_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_59_cast_fp16 = matmul(transpose_x = matmul_59_transpose_x_0, transpose_y = matmul_59_transpose_y_0, x = softmax_17_cast_fp16, y = view_106_cast_fp16)[name = string("matmul_59_cast_fp16")];
            tensor<int32, [4]> transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_2504 = const()[name = string("const_2504"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_113_cast_fp16 = transpose(perm = transpose_113_perm_0, x = matmul_59_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 512, 768]> view_107_cast_fp16 = reshape(shape = const_2504, x = transpose_113_cast_fp16)[name = string("view_107_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275546816))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276136704))))[name = string("p_st_0_model_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_122_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = view_107_cast_fp16)[name = string("linear_122_cast_fp16")];
            fp16 const_2506_promoted_to_fp16 = const()[name = string("const_2506_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_106_cast_fp16 = pow(x = linear_122_cast_fp16, y = const_2506_promoted_to_fp16)[name = string("pow_106_cast_fp16")];
            tensor<int32, [1]> mean_105_axes_0 = const()[name = string("mean_105_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_105_keep_dims_0 = const()[name = string("mean_105_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_105_cast_fp16 = reduce_mean(axes = mean_105_axes_0, keep_dims = mean_105_keep_dims_0, x = pow_106_cast_fp16)[name = string("mean_105_cast_fp16")];
            fp16 const_2509_to_fp16 = const()[name = string("const_2509_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_298_cast_fp16 = add(x = mean_105_cast_fp16, y = const_2509_to_fp16)[name = string("add_298_cast_fp16")];
            fp32 rsqrt_105_epsilon_0 = const()[name = string("rsqrt_105_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_105_cast_fp16 = rsqrt(epsilon = rsqrt_105_epsilon_0, x = add_298_cast_fp16)[name = string("rsqrt_105_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_368_cast_fp16 = mul(x = linear_122_cast_fp16, y = rsqrt_105_cast_fp16)[name = string("mul_368_cast_fp16")];
            tensor<fp16, [768]> add_299_to_fp16 = const()[name = string("add_299_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276138304)))];
            tensor<fp16, [1, 512, 768]> mul_369_cast_fp16 = mul(x = mul_368_cast_fp16, y = add_299_to_fp16)[name = string("mul_369_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_300_cast_fp16 = add(x = add_288_cast_fp16, y = mul_369_cast_fp16)[name = string("add_300_cast_fp16")];
            fp16 const_2514_promoted_to_fp16 = const()[name = string("const_2514_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_107_cast_fp16 = pow(x = add_300_cast_fp16, y = const_2514_promoted_to_fp16)[name = string("pow_107_cast_fp16")];
            tensor<int32, [1]> mean_106_axes_0 = const()[name = string("mean_106_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_106_keep_dims_0 = const()[name = string("mean_106_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_106_cast_fp16 = reduce_mean(axes = mean_106_axes_0, keep_dims = mean_106_keep_dims_0, x = pow_107_cast_fp16)[name = string("mean_106_cast_fp16")];
            fp16 const_2517_to_fp16 = const()[name = string("const_2517_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_301_cast_fp16 = add(x = mean_106_cast_fp16, y = const_2517_to_fp16)[name = string("add_301_cast_fp16")];
            fp32 rsqrt_106_epsilon_0 = const()[name = string("rsqrt_106_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_106_cast_fp16 = rsqrt(epsilon = rsqrt_106_epsilon_0, x = add_301_cast_fp16)[name = string("rsqrt_106_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_370_cast_fp16 = mul(x = add_300_cast_fp16, y = rsqrt_106_cast_fp16)[name = string("mul_370_cast_fp16")];
            tensor<fp16, [768]> add_302_to_fp16 = const()[name = string("add_302_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276139904)))];
            tensor<fp16, [1, 512, 768]> mul_371_cast_fp16 = mul(x = mul_370_cast_fp16, y = add_302_to_fp16)[name = string("mul_371_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276141504))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277026304))))[name = string("p_st_0_model_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = mul_371_cast_fp16)[name = string("linear_123_cast_fp16")];
            string gelu_17_mode_0 = const()[name = string("gelu_17_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_17_cast_fp16 = gelu(mode = gelu_17_mode_0, x = linear_123_cast_fp16)[name = string("gelu_17_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277028672))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277913472))))[name = string("p_st_0_model_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_124_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = mul_371_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_372_cast_fp16 = mul(x = gelu_17_cast_fp16, y = linear_124_cast_fp16)[name = string("mul_372_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277915840))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278800640))))[name = string("p_st_0_model_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_125_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = mul_372_cast_fp16)[name = string("linear_125_cast_fp16")];
            fp16 const_2522_promoted_to_fp16 = const()[name = string("const_2522_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_108_cast_fp16 = pow(x = linear_125_cast_fp16, y = const_2522_promoted_to_fp16)[name = string("pow_108_cast_fp16")];
            tensor<int32, [1]> mean_107_axes_0 = const()[name = string("mean_107_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_107_keep_dims_0 = const()[name = string("mean_107_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_107_cast_fp16 = reduce_mean(axes = mean_107_axes_0, keep_dims = mean_107_keep_dims_0, x = pow_108_cast_fp16)[name = string("mean_107_cast_fp16")];
            fp16 const_2525_to_fp16 = const()[name = string("const_2525_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_303_cast_fp16 = add(x = mean_107_cast_fp16, y = const_2525_to_fp16)[name = string("add_303_cast_fp16")];
            fp32 rsqrt_107_epsilon_0 = const()[name = string("rsqrt_107_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_107_cast_fp16 = rsqrt(epsilon = rsqrt_107_epsilon_0, x = add_303_cast_fp16)[name = string("rsqrt_107_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_373_cast_fp16 = mul(x = linear_125_cast_fp16, y = rsqrt_107_cast_fp16)[name = string("mul_373_cast_fp16")];
            tensor<fp16, [768]> add_304_to_fp16 = const()[name = string("add_304_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278802240)))];
            tensor<fp16, [1, 512, 768]> mul_374_cast_fp16 = mul(x = mul_373_cast_fp16, y = add_304_to_fp16)[name = string("mul_374_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_305_cast_fp16 = add(x = add_300_cast_fp16, y = mul_374_cast_fp16)[name = string("add_305_cast_fp16")];
            fp16 const_2530_promoted_to_fp16 = const()[name = string("const_2530_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_109_cast_fp16 = pow(x = add_305_cast_fp16, y = const_2530_promoted_to_fp16)[name = string("pow_109_cast_fp16")];
            tensor<int32, [1]> mean_108_axes_0 = const()[name = string("mean_108_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_108_keep_dims_0 = const()[name = string("mean_108_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_108_cast_fp16 = reduce_mean(axes = mean_108_axes_0, keep_dims = mean_108_keep_dims_0, x = pow_109_cast_fp16)[name = string("mean_108_cast_fp16")];
            fp16 const_2533_to_fp16 = const()[name = string("const_2533_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_306_cast_fp16 = add(x = mean_108_cast_fp16, y = const_2533_to_fp16)[name = string("add_306_cast_fp16")];
            fp32 rsqrt_108_epsilon_0 = const()[name = string("rsqrt_108_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_108_cast_fp16 = rsqrt(epsilon = rsqrt_108_epsilon_0, x = add_306_cast_fp16)[name = string("rsqrt_108_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_375_cast_fp16 = mul(x = add_305_cast_fp16, y = rsqrt_108_cast_fp16)[name = string("mul_375_cast_fp16")];
            tensor<fp16, [768]> add_307_to_fp16 = const()[name = string("add_307_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278803840)))];
            tensor<fp16, [1, 512, 768]> mul_376_cast_fp16 = mul(x = mul_375_cast_fp16, y = add_307_to_fp16)[name = string("mul_376_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278805440))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279395328))))[name = string("p_st_0_model_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_126_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = mul_376_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<int32, [4]> const_2537 = const()[name = string("const_2537"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_108_cast_fp16 = reshape(shape = const_2537, x = linear_126_cast_fp16)[name = string("view_108_cast_fp16")];
            tensor<int32, [4]> transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279396928))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279593600))))[name = string("p_st_0_model_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = mul_376_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> const_2540 = const()[name = string("const_2540"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_109_cast_fp16 = reshape(shape = const_2540, x = linear_127_cast_fp16)[name = string("view_109_cast_fp16")];
            tensor<int32, [4]> transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279594176))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279790848))))[name = string("p_st_0_model_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_128_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = mul_376_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> const_2543 = const()[name = string("const_2543"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_110_cast_fp16 = reshape(shape = const_2543, x = linear_128_cast_fp16)[name = string("view_110_cast_fp16")];
            tensor<int32, [4]> transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_2547_promoted_to_fp16 = const()[name = string("const_2547_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_114_cast_fp16 = transpose(perm = transpose_114_perm_0, x = view_108_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 3, 512, 256]> pow_110_cast_fp16 = pow(x = transpose_114_cast_fp16, y = const_2547_promoted_to_fp16)[name = string("pow_110_cast_fp16")];
            tensor<int32, [1]> mean_109_axes_0 = const()[name = string("mean_109_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_109_keep_dims_0 = const()[name = string("mean_109_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_109_cast_fp16 = reduce_mean(axes = mean_109_axes_0, keep_dims = mean_109_keep_dims_0, x = pow_110_cast_fp16)[name = string("mean_109_cast_fp16")];
            fp16 const_2550_to_fp16 = const()[name = string("const_2550_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_308_cast_fp16 = add(x = mean_109_cast_fp16, y = const_2550_to_fp16)[name = string("add_308_cast_fp16")];
            fp32 rsqrt_109_epsilon_0 = const()[name = string("rsqrt_109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_109_cast_fp16 = rsqrt(epsilon = rsqrt_109_epsilon_0, x = add_308_cast_fp16)[name = string("rsqrt_109_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_377_cast_fp16 = mul(x = transpose_114_cast_fp16, y = rsqrt_109_cast_fp16)[name = string("mul_377_cast_fp16")];
            tensor<fp16, [256]> add_309_to_fp16 = const()[name = string("add_309_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279791424)))];
            tensor<fp16, [1, 3, 512, 256]> mul_378_cast_fp16 = mul(x = mul_377_cast_fp16, y = add_309_to_fp16)[name = string("mul_378_cast_fp16")];
            fp16 const_2555_promoted_to_fp16 = const()[name = string("const_2555_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_115_cast_fp16 = transpose(perm = transpose_115_perm_0, x = view_109_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 1, 512, 256]> pow_111_cast_fp16 = pow(x = transpose_115_cast_fp16, y = const_2555_promoted_to_fp16)[name = string("pow_111_cast_fp16")];
            tensor<int32, [1]> mean_110_axes_0 = const()[name = string("mean_110_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_110_keep_dims_0 = const()[name = string("mean_110_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_110_cast_fp16 = reduce_mean(axes = mean_110_axes_0, keep_dims = mean_110_keep_dims_0, x = pow_111_cast_fp16)[name = string("mean_110_cast_fp16")];
            fp16 const_2558_to_fp16 = const()[name = string("const_2558_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_310_cast_fp16 = add(x = mean_110_cast_fp16, y = const_2558_to_fp16)[name = string("add_310_cast_fp16")];
            fp32 rsqrt_110_epsilon_0 = const()[name = string("rsqrt_110_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_110_cast_fp16 = rsqrt(epsilon = rsqrt_110_epsilon_0, x = add_310_cast_fp16)[name = string("rsqrt_110_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_379_cast_fp16 = mul(x = transpose_115_cast_fp16, y = rsqrt_110_cast_fp16)[name = string("mul_379_cast_fp16")];
            tensor<fp16, [256]> add_311_to_fp16 = const()[name = string("add_311_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279792000)))];
            tensor<fp16, [1, 1, 512, 256]> mul_380_cast_fp16 = mul(x = mul_379_cast_fp16, y = add_311_to_fp16)[name = string("mul_380_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_381_cast_fp16 = mul(x = mul_378_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_381_cast_fp16")];
            tensor<int32, [4]> slice_467_begin_0 = const()[name = string("slice_467_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_467_end_0 = const()[name = string("slice_467_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_467_end_mask_0 = const()[name = string("slice_467_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_467_cast_fp16 = slice_by_index(begin = slice_467_begin_0, end = slice_467_end_0, end_mask = slice_467_end_mask_0, x = mul_378_cast_fp16)[name = string("slice_467_cast_fp16")];
            tensor<int32, [4]> slice_468_begin_0 = const()[name = string("slice_468_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_468_end_0 = const()[name = string("slice_468_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_468_end_mask_0 = const()[name = string("slice_468_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_468_cast_fp16 = slice_by_index(begin = slice_468_begin_0, end = slice_468_end_0, end_mask = slice_468_end_mask_0, x = mul_378_cast_fp16)[name = string("slice_468_cast_fp16")];
            fp16 const_2570_promoted_to_fp16 = const()[name = string("const_2570_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_36_cast_fp16 = mul(x = slice_468_cast_fp16, y = const_2570_promoted_to_fp16)[name = string("neg_36_cast_fp16")];
            int32 const_2571 = const()[name = string("const_2571"), val = int32(-1)];
            bool cat_96_interleave_0 = const()[name = string("cat_96_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_96_cast_fp16 = concat(axis = const_2571, interleave = cat_96_interleave_0, values = (neg_36_cast_fp16, slice_467_cast_fp16))[name = string("cat_96_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_382_cast_fp16 = mul(x = cat_96_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_382_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_312_cast_fp16 = add(x = mul_381_cast_fp16, y = mul_382_cast_fp16)[name = string("add_312_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_383_cast_fp16 = mul(x = mul_380_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_383_cast_fp16")];
            tensor<int32, [4]> slice_469_begin_0 = const()[name = string("slice_469_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_469_end_0 = const()[name = string("slice_469_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_469_end_mask_0 = const()[name = string("slice_469_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_469_cast_fp16 = slice_by_index(begin = slice_469_begin_0, end = slice_469_end_0, end_mask = slice_469_end_mask_0, x = mul_380_cast_fp16)[name = string("slice_469_cast_fp16")];
            tensor<int32, [4]> slice_470_begin_0 = const()[name = string("slice_470_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_470_end_0 = const()[name = string("slice_470_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_470_end_mask_0 = const()[name = string("slice_470_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_470_cast_fp16 = slice_by_index(begin = slice_470_begin_0, end = slice_470_end_0, end_mask = slice_470_end_mask_0, x = mul_380_cast_fp16)[name = string("slice_470_cast_fp16")];
            fp16 const_2578_promoted_to_fp16 = const()[name = string("const_2578_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_37_cast_fp16 = mul(x = slice_470_cast_fp16, y = const_2578_promoted_to_fp16)[name = string("neg_37_cast_fp16")];
            int32 const_2579 = const()[name = string("const_2579"), val = int32(-1)];
            bool cat_97_interleave_0 = const()[name = string("cat_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_97_cast_fp16 = concat(axis = const_2579, interleave = cat_97_interleave_0, values = (neg_37_cast_fp16, slice_469_cast_fp16))[name = string("cat_97_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_384_cast_fp16 = mul(x = cat_97_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_384_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_313_cast_fp16 = add(x = mul_383_cast_fp16, y = mul_384_cast_fp16)[name = string("add_313_cast_fp16")];
            int32 const_2580 = const()[name = string("const_2580"), val = int32(-2)];
            bool cat_98_interleave_0 = const()[name = string("cat_98_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_98_cast_fp16 = concat(axis = const_2580, interleave = cat_98_interleave_0, values = add_313_cast_fp16)[name = string("cat_98_cast_fp16")];
            int32 const_2581 = const()[name = string("const_2581"), val = int32(-2)];
            bool cat_99_interleave_0 = const()[name = string("cat_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_116_cast_fp16 = transpose(perm = transpose_116_perm_0, x = view_110_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 1, 512, 256]> cat_99_cast_fp16 = concat(axis = const_2581, interleave = cat_99_interleave_0, values = transpose_116_cast_fp16)[name = string("cat_99_cast_fp16")];
            tensor<int32, [1]> unsqueeze_151_axes_0 = const()[name = string("unsqueeze_151_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_151_cast_fp16 = expand_dims(axes = unsqueeze_151_axes_0, x = cat_98_cast_fp16)[name = string("unsqueeze_151_cast_fp16")];
            tensor<int32, [5]> expand_62_reps_0 = const()[name = string("expand_62_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_62_cast_fp16 = tile(reps = expand_62_reps_0, x = unsqueeze_151_cast_fp16)[name = string("expand_62_cast_fp16")];
            tensor<int32, [4]> const_2596 = const()[name = string("const_2596"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_111_cast_fp16 = reshape(shape = const_2596, x = expand_62_cast_fp16)[name = string("view_111_cast_fp16")];
            tensor<int32, [1]> unsqueeze_152_axes_0 = const()[name = string("unsqueeze_152_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_152_cast_fp16 = expand_dims(axes = unsqueeze_152_axes_0, x = cat_99_cast_fp16)[name = string("unsqueeze_152_cast_fp16")];
            tensor<int32, [5]> expand_63_reps_0 = const()[name = string("expand_63_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_63_cast_fp16 = tile(reps = expand_63_reps_0, x = unsqueeze_152_cast_fp16)[name = string("expand_63_cast_fp16")];
            tensor<int32, [4]> const_2611 = const()[name = string("const_2611"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_112_cast_fp16 = reshape(shape = const_2611, x = expand_63_cast_fp16)[name = string("view_112_cast_fp16")];
            bool matmul_60_transpose_x_1 = const()[name = string("matmul_60_transpose_x_1"), val = bool(false)];
            bool matmul_60_transpose_y_1 = const()[name = string("matmul_60_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_60_cast_fp16 = matmul(transpose_x = matmul_60_transpose_x_1, transpose_y = matmul_60_transpose_y_1, x = add_312_cast_fp16, y = view_111_cast_fp16)[name = string("matmul_60_cast_fp16")];
            fp16 const_2614_to_fp16 = const()[name = string("const_2614_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_385_cast_fp16 = mul(x = matmul_60_cast_fp16, y = const_2614_to_fp16)[name = string("mul_385_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_314_cast_fp16 = add(x = mul_385_cast_fp16, y = expand_cast_fp16)[name = string("add_314_cast_fp16")];
            int32 const_2624 = const()[name = string("const_2624"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_18_cast_fp16 = softmax(axis = const_2624, x = add_314_cast_fp16)[name = string("softmax_18_cast_fp16")];
            bool matmul_61_transpose_x_0 = const()[name = string("matmul_61_transpose_x_0"), val = bool(false)];
            bool matmul_61_transpose_y_0 = const()[name = string("matmul_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_61_cast_fp16 = matmul(transpose_x = matmul_61_transpose_x_0, transpose_y = matmul_61_transpose_y_0, x = softmax_18_cast_fp16, y = view_112_cast_fp16)[name = string("matmul_61_cast_fp16")];
            tensor<int32, [4]> transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_2629 = const()[name = string("const_2629"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_118_cast_fp16 = transpose(perm = transpose_118_perm_0, x = matmul_61_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 512, 768]> view_113_cast_fp16 = reshape(shape = const_2629, x = transpose_118_cast_fp16)[name = string("view_113_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279792576))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280382464))))[name = string("p_st_0_model_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_129_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = view_113_cast_fp16)[name = string("linear_129_cast_fp16")];
            fp16 const_2631_promoted_to_fp16 = const()[name = string("const_2631_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_112_cast_fp16 = pow(x = linear_129_cast_fp16, y = const_2631_promoted_to_fp16)[name = string("pow_112_cast_fp16")];
            tensor<int32, [1]> mean_111_axes_0 = const()[name = string("mean_111_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_111_keep_dims_0 = const()[name = string("mean_111_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_111_cast_fp16 = reduce_mean(axes = mean_111_axes_0, keep_dims = mean_111_keep_dims_0, x = pow_112_cast_fp16)[name = string("mean_111_cast_fp16")];
            fp16 const_2634_to_fp16 = const()[name = string("const_2634_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_315_cast_fp16 = add(x = mean_111_cast_fp16, y = const_2634_to_fp16)[name = string("add_315_cast_fp16")];
            fp32 rsqrt_111_epsilon_0 = const()[name = string("rsqrt_111_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_111_cast_fp16 = rsqrt(epsilon = rsqrt_111_epsilon_0, x = add_315_cast_fp16)[name = string("rsqrt_111_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_386_cast_fp16 = mul(x = linear_129_cast_fp16, y = rsqrt_111_cast_fp16)[name = string("mul_386_cast_fp16")];
            tensor<fp16, [768]> add_316_to_fp16 = const()[name = string("add_316_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280384064)))];
            tensor<fp16, [1, 512, 768]> mul_387_cast_fp16 = mul(x = mul_386_cast_fp16, y = add_316_to_fp16)[name = string("mul_387_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_317_cast_fp16 = add(x = add_305_cast_fp16, y = mul_387_cast_fp16)[name = string("add_317_cast_fp16")];
            fp16 const_2639_promoted_to_fp16 = const()[name = string("const_2639_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_113_cast_fp16 = pow(x = add_317_cast_fp16, y = const_2639_promoted_to_fp16)[name = string("pow_113_cast_fp16")];
            tensor<int32, [1]> mean_112_axes_0 = const()[name = string("mean_112_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_112_keep_dims_0 = const()[name = string("mean_112_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_112_cast_fp16 = reduce_mean(axes = mean_112_axes_0, keep_dims = mean_112_keep_dims_0, x = pow_113_cast_fp16)[name = string("mean_112_cast_fp16")];
            fp16 const_2642_to_fp16 = const()[name = string("const_2642_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_318_cast_fp16 = add(x = mean_112_cast_fp16, y = const_2642_to_fp16)[name = string("add_318_cast_fp16")];
            fp32 rsqrt_112_epsilon_0 = const()[name = string("rsqrt_112_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_112_cast_fp16 = rsqrt(epsilon = rsqrt_112_epsilon_0, x = add_318_cast_fp16)[name = string("rsqrt_112_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_388_cast_fp16 = mul(x = add_317_cast_fp16, y = rsqrt_112_cast_fp16)[name = string("mul_388_cast_fp16")];
            tensor<fp16, [768]> add_319_to_fp16 = const()[name = string("add_319_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280385664)))];
            tensor<fp16, [1, 512, 768]> mul_389_cast_fp16 = mul(x = mul_388_cast_fp16, y = add_319_to_fp16)[name = string("mul_389_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280387264))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281272064))))[name = string("p_st_0_model_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = mul_389_cast_fp16)[name = string("linear_130_cast_fp16")];
            string gelu_18_mode_0 = const()[name = string("gelu_18_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_18_cast_fp16 = gelu(mode = gelu_18_mode_0, x = linear_130_cast_fp16)[name = string("gelu_18_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281274432))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282159232))))[name = string("p_st_0_model_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_131_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = mul_389_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_390_cast_fp16 = mul(x = gelu_18_cast_fp16, y = linear_131_cast_fp16)[name = string("mul_390_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282161600))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283046400))))[name = string("p_st_0_model_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_132_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = mul_390_cast_fp16)[name = string("linear_132_cast_fp16")];
            fp16 const_2647_promoted_to_fp16 = const()[name = string("const_2647_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_114_cast_fp16 = pow(x = linear_132_cast_fp16, y = const_2647_promoted_to_fp16)[name = string("pow_114_cast_fp16")];
            tensor<int32, [1]> mean_113_axes_0 = const()[name = string("mean_113_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_113_keep_dims_0 = const()[name = string("mean_113_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_113_cast_fp16 = reduce_mean(axes = mean_113_axes_0, keep_dims = mean_113_keep_dims_0, x = pow_114_cast_fp16)[name = string("mean_113_cast_fp16")];
            fp16 const_2650_to_fp16 = const()[name = string("const_2650_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_320_cast_fp16 = add(x = mean_113_cast_fp16, y = const_2650_to_fp16)[name = string("add_320_cast_fp16")];
            fp32 rsqrt_113_epsilon_0 = const()[name = string("rsqrt_113_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_113_cast_fp16 = rsqrt(epsilon = rsqrt_113_epsilon_0, x = add_320_cast_fp16)[name = string("rsqrt_113_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_391_cast_fp16 = mul(x = linear_132_cast_fp16, y = rsqrt_113_cast_fp16)[name = string("mul_391_cast_fp16")];
            tensor<fp16, [768]> add_321_to_fp16 = const()[name = string("add_321_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283048000)))];
            tensor<fp16, [1, 512, 768]> mul_392_cast_fp16 = mul(x = mul_391_cast_fp16, y = add_321_to_fp16)[name = string("mul_392_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_322_cast_fp16 = add(x = add_317_cast_fp16, y = mul_392_cast_fp16)[name = string("add_322_cast_fp16")];
            fp16 const_2655_promoted_to_fp16 = const()[name = string("const_2655_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_115_cast_fp16 = pow(x = add_322_cast_fp16, y = const_2655_promoted_to_fp16)[name = string("pow_115_cast_fp16")];
            tensor<int32, [1]> mean_114_axes_0 = const()[name = string("mean_114_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_114_keep_dims_0 = const()[name = string("mean_114_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_114_cast_fp16 = reduce_mean(axes = mean_114_axes_0, keep_dims = mean_114_keep_dims_0, x = pow_115_cast_fp16)[name = string("mean_114_cast_fp16")];
            fp16 const_2658_to_fp16 = const()[name = string("const_2658_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_323_cast_fp16 = add(x = mean_114_cast_fp16, y = const_2658_to_fp16)[name = string("add_323_cast_fp16")];
            fp32 rsqrt_114_epsilon_0 = const()[name = string("rsqrt_114_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_114_cast_fp16 = rsqrt(epsilon = rsqrt_114_epsilon_0, x = add_323_cast_fp16)[name = string("rsqrt_114_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_393_cast_fp16 = mul(x = add_322_cast_fp16, y = rsqrt_114_cast_fp16)[name = string("mul_393_cast_fp16")];
            tensor<fp16, [768]> add_324_to_fp16 = const()[name = string("add_324_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283049600)))];
            tensor<fp16, [1, 512, 768]> mul_394_cast_fp16 = mul(x = mul_393_cast_fp16, y = add_324_to_fp16)[name = string("mul_394_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283051200))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283641088))))[name = string("p_st_0_model_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_133_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = mul_394_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<int32, [4]> const_2662 = const()[name = string("const_2662"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_114_cast_fp16 = reshape(shape = const_2662, x = linear_133_cast_fp16)[name = string("view_114_cast_fp16")];
            tensor<int32, [4]> transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283642688))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283839360))))[name = string("p_st_0_model_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_134_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = mul_394_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<int32, [4]> const_2665 = const()[name = string("const_2665"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_115_cast_fp16 = reshape(shape = const_2665, x = linear_134_cast_fp16)[name = string("view_115_cast_fp16")];
            tensor<int32, [4]> transpose_120_perm_0 = const()[name = string("transpose_120_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283839936))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284036608))))[name = string("p_st_0_model_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_135_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = mul_394_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<int32, [4]> const_2668 = const()[name = string("const_2668"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_116_cast_fp16 = reshape(shape = const_2668, x = linear_135_cast_fp16)[name = string("view_116_cast_fp16")];
            tensor<int32, [4]> transpose_121_perm_0 = const()[name = string("transpose_121_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_2672_promoted_to_fp16 = const()[name = string("const_2672_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_119_cast_fp16 = transpose(perm = transpose_119_perm_0, x = view_114_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 3, 512, 256]> pow_116_cast_fp16 = pow(x = transpose_119_cast_fp16, y = const_2672_promoted_to_fp16)[name = string("pow_116_cast_fp16")];
            tensor<int32, [1]> mean_115_axes_0 = const()[name = string("mean_115_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_115_keep_dims_0 = const()[name = string("mean_115_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_115_cast_fp16 = reduce_mean(axes = mean_115_axes_0, keep_dims = mean_115_keep_dims_0, x = pow_116_cast_fp16)[name = string("mean_115_cast_fp16")];
            fp16 const_2675_to_fp16 = const()[name = string("const_2675_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_325_cast_fp16 = add(x = mean_115_cast_fp16, y = const_2675_to_fp16)[name = string("add_325_cast_fp16")];
            fp32 rsqrt_115_epsilon_0 = const()[name = string("rsqrt_115_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_115_cast_fp16 = rsqrt(epsilon = rsqrt_115_epsilon_0, x = add_325_cast_fp16)[name = string("rsqrt_115_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_395_cast_fp16 = mul(x = transpose_119_cast_fp16, y = rsqrt_115_cast_fp16)[name = string("mul_395_cast_fp16")];
            tensor<fp16, [256]> add_326_to_fp16 = const()[name = string("add_326_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284037184)))];
            tensor<fp16, [1, 3, 512, 256]> mul_396_cast_fp16 = mul(x = mul_395_cast_fp16, y = add_326_to_fp16)[name = string("mul_396_cast_fp16")];
            fp16 const_2680_promoted_to_fp16 = const()[name = string("const_2680_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_120_cast_fp16 = transpose(perm = transpose_120_perm_0, x = view_115_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 1, 512, 256]> pow_117_cast_fp16 = pow(x = transpose_120_cast_fp16, y = const_2680_promoted_to_fp16)[name = string("pow_117_cast_fp16")];
            tensor<int32, [1]> mean_116_axes_0 = const()[name = string("mean_116_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_116_keep_dims_0 = const()[name = string("mean_116_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_116_cast_fp16 = reduce_mean(axes = mean_116_axes_0, keep_dims = mean_116_keep_dims_0, x = pow_117_cast_fp16)[name = string("mean_116_cast_fp16")];
            fp16 const_2683_to_fp16 = const()[name = string("const_2683_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_327_cast_fp16 = add(x = mean_116_cast_fp16, y = const_2683_to_fp16)[name = string("add_327_cast_fp16")];
            fp32 rsqrt_116_epsilon_0 = const()[name = string("rsqrt_116_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_116_cast_fp16 = rsqrt(epsilon = rsqrt_116_epsilon_0, x = add_327_cast_fp16)[name = string("rsqrt_116_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_397_cast_fp16 = mul(x = transpose_120_cast_fp16, y = rsqrt_116_cast_fp16)[name = string("mul_397_cast_fp16")];
            tensor<fp16, [256]> add_328_to_fp16 = const()[name = string("add_328_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284037760)))];
            tensor<fp16, [1, 1, 512, 256]> mul_398_cast_fp16 = mul(x = mul_397_cast_fp16, y = add_328_to_fp16)[name = string("mul_398_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_399_cast_fp16 = mul(x = mul_396_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_399_cast_fp16")];
            tensor<int32, [4]> slice_490_begin_0 = const()[name = string("slice_490_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_490_end_0 = const()[name = string("slice_490_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_490_end_mask_0 = const()[name = string("slice_490_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_490_cast_fp16 = slice_by_index(begin = slice_490_begin_0, end = slice_490_end_0, end_mask = slice_490_end_mask_0, x = mul_396_cast_fp16)[name = string("slice_490_cast_fp16")];
            tensor<int32, [4]> slice_491_begin_0 = const()[name = string("slice_491_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_491_end_0 = const()[name = string("slice_491_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_491_end_mask_0 = const()[name = string("slice_491_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_491_cast_fp16 = slice_by_index(begin = slice_491_begin_0, end = slice_491_end_0, end_mask = slice_491_end_mask_0, x = mul_396_cast_fp16)[name = string("slice_491_cast_fp16")];
            fp16 const_2695_promoted_to_fp16 = const()[name = string("const_2695_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_38_cast_fp16 = mul(x = slice_491_cast_fp16, y = const_2695_promoted_to_fp16)[name = string("neg_38_cast_fp16")];
            int32 const_2696 = const()[name = string("const_2696"), val = int32(-1)];
            bool cat_100_interleave_0 = const()[name = string("cat_100_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_100_cast_fp16 = concat(axis = const_2696, interleave = cat_100_interleave_0, values = (neg_38_cast_fp16, slice_490_cast_fp16))[name = string("cat_100_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_400_cast_fp16 = mul(x = cat_100_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_400_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_329_cast_fp16 = add(x = mul_399_cast_fp16, y = mul_400_cast_fp16)[name = string("add_329_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_401_cast_fp16 = mul(x = mul_398_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_401_cast_fp16")];
            tensor<int32, [4]> slice_492_begin_0 = const()[name = string("slice_492_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_492_end_0 = const()[name = string("slice_492_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_492_end_mask_0 = const()[name = string("slice_492_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_492_cast_fp16 = slice_by_index(begin = slice_492_begin_0, end = slice_492_end_0, end_mask = slice_492_end_mask_0, x = mul_398_cast_fp16)[name = string("slice_492_cast_fp16")];
            tensor<int32, [4]> slice_493_begin_0 = const()[name = string("slice_493_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_493_end_0 = const()[name = string("slice_493_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_493_end_mask_0 = const()[name = string("slice_493_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_493_cast_fp16 = slice_by_index(begin = slice_493_begin_0, end = slice_493_end_0, end_mask = slice_493_end_mask_0, x = mul_398_cast_fp16)[name = string("slice_493_cast_fp16")];
            fp16 const_2703_promoted_to_fp16 = const()[name = string("const_2703_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_39_cast_fp16 = mul(x = slice_493_cast_fp16, y = const_2703_promoted_to_fp16)[name = string("neg_39_cast_fp16")];
            int32 const_2704 = const()[name = string("const_2704"), val = int32(-1)];
            bool cat_101_interleave_0 = const()[name = string("cat_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_101_cast_fp16 = concat(axis = const_2704, interleave = cat_101_interleave_0, values = (neg_39_cast_fp16, slice_492_cast_fp16))[name = string("cat_101_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_402_cast_fp16 = mul(x = cat_101_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_402_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_330_cast_fp16 = add(x = mul_401_cast_fp16, y = mul_402_cast_fp16)[name = string("add_330_cast_fp16")];
            int32 const_2705 = const()[name = string("const_2705"), val = int32(-2)];
            bool cat_102_interleave_0 = const()[name = string("cat_102_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_102_cast_fp16 = concat(axis = const_2705, interleave = cat_102_interleave_0, values = add_330_cast_fp16)[name = string("cat_102_cast_fp16")];
            int32 const_2706 = const()[name = string("const_2706"), val = int32(-2)];
            bool cat_103_interleave_0 = const()[name = string("cat_103_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_121_cast_fp16 = transpose(perm = transpose_121_perm_0, x = view_116_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 1, 512, 256]> cat_103_cast_fp16 = concat(axis = const_2706, interleave = cat_103_interleave_0, values = transpose_121_cast_fp16)[name = string("cat_103_cast_fp16")];
            tensor<int32, [1]> unsqueeze_155_axes_0 = const()[name = string("unsqueeze_155_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_155_cast_fp16 = expand_dims(axes = unsqueeze_155_axes_0, x = cat_102_cast_fp16)[name = string("unsqueeze_155_cast_fp16")];
            tensor<int32, [5]> expand_64_reps_0 = const()[name = string("expand_64_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_64_cast_fp16 = tile(reps = expand_64_reps_0, x = unsqueeze_155_cast_fp16)[name = string("expand_64_cast_fp16")];
            tensor<int32, [4]> const_2721 = const()[name = string("const_2721"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_117_cast_fp16 = reshape(shape = const_2721, x = expand_64_cast_fp16)[name = string("view_117_cast_fp16")];
            tensor<int32, [1]> unsqueeze_156_axes_0 = const()[name = string("unsqueeze_156_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_156_cast_fp16 = expand_dims(axes = unsqueeze_156_axes_0, x = cat_103_cast_fp16)[name = string("unsqueeze_156_cast_fp16")];
            tensor<int32, [5]> expand_65_reps_0 = const()[name = string("expand_65_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_65_cast_fp16 = tile(reps = expand_65_reps_0, x = unsqueeze_156_cast_fp16)[name = string("expand_65_cast_fp16")];
            tensor<int32, [4]> const_2736 = const()[name = string("const_2736"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_118_cast_fp16 = reshape(shape = const_2736, x = expand_65_cast_fp16)[name = string("view_118_cast_fp16")];
            bool matmul_62_transpose_x_1 = const()[name = string("matmul_62_transpose_x_1"), val = bool(false)];
            bool matmul_62_transpose_y_1 = const()[name = string("matmul_62_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_62_cast_fp16 = matmul(transpose_x = matmul_62_transpose_x_1, transpose_y = matmul_62_transpose_y_1, x = add_329_cast_fp16, y = view_117_cast_fp16)[name = string("matmul_62_cast_fp16")];
            fp16 const_2739_to_fp16 = const()[name = string("const_2739_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_403_cast_fp16 = mul(x = matmul_62_cast_fp16, y = const_2739_to_fp16)[name = string("mul_403_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_331_cast_fp16 = add(x = mul_403_cast_fp16, y = expand_cast_fp16)[name = string("add_331_cast_fp16")];
            int32 const_2749 = const()[name = string("const_2749"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_19_cast_fp16 = softmax(axis = const_2749, x = add_331_cast_fp16)[name = string("softmax_19_cast_fp16")];
            bool matmul_63_transpose_x_0 = const()[name = string("matmul_63_transpose_x_0"), val = bool(false)];
            bool matmul_63_transpose_y_0 = const()[name = string("matmul_63_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_63_cast_fp16 = matmul(transpose_x = matmul_63_transpose_x_0, transpose_y = matmul_63_transpose_y_0, x = softmax_19_cast_fp16, y = view_118_cast_fp16)[name = string("matmul_63_cast_fp16")];
            tensor<int32, [4]> transpose_123_perm_0 = const()[name = string("transpose_123_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_2754 = const()[name = string("const_2754"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_123_cast_fp16 = transpose(perm = transpose_123_perm_0, x = matmul_63_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 512, 768]> view_119_cast_fp16 = reshape(shape = const_2754, x = transpose_123_cast_fp16)[name = string("view_119_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284038336))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284628224))))[name = string("p_st_0_model_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_136_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = view_119_cast_fp16)[name = string("linear_136_cast_fp16")];
            fp16 const_2756_promoted_to_fp16 = const()[name = string("const_2756_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_118_cast_fp16 = pow(x = linear_136_cast_fp16, y = const_2756_promoted_to_fp16)[name = string("pow_118_cast_fp16")];
            tensor<int32, [1]> mean_117_axes_0 = const()[name = string("mean_117_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_117_keep_dims_0 = const()[name = string("mean_117_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_117_cast_fp16 = reduce_mean(axes = mean_117_axes_0, keep_dims = mean_117_keep_dims_0, x = pow_118_cast_fp16)[name = string("mean_117_cast_fp16")];
            fp16 const_2759_to_fp16 = const()[name = string("const_2759_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_332_cast_fp16 = add(x = mean_117_cast_fp16, y = const_2759_to_fp16)[name = string("add_332_cast_fp16")];
            fp32 rsqrt_117_epsilon_0 = const()[name = string("rsqrt_117_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_117_cast_fp16 = rsqrt(epsilon = rsqrt_117_epsilon_0, x = add_332_cast_fp16)[name = string("rsqrt_117_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_404_cast_fp16 = mul(x = linear_136_cast_fp16, y = rsqrt_117_cast_fp16)[name = string("mul_404_cast_fp16")];
            tensor<fp16, [768]> add_333_to_fp16 = const()[name = string("add_333_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284629824)))];
            tensor<fp16, [1, 512, 768]> mul_405_cast_fp16 = mul(x = mul_404_cast_fp16, y = add_333_to_fp16)[name = string("mul_405_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_334_cast_fp16 = add(x = add_322_cast_fp16, y = mul_405_cast_fp16)[name = string("add_334_cast_fp16")];
            fp16 const_2764_promoted_to_fp16 = const()[name = string("const_2764_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_119_cast_fp16 = pow(x = add_334_cast_fp16, y = const_2764_promoted_to_fp16)[name = string("pow_119_cast_fp16")];
            tensor<int32, [1]> mean_118_axes_0 = const()[name = string("mean_118_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_118_keep_dims_0 = const()[name = string("mean_118_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_118_cast_fp16 = reduce_mean(axes = mean_118_axes_0, keep_dims = mean_118_keep_dims_0, x = pow_119_cast_fp16)[name = string("mean_118_cast_fp16")];
            fp16 const_2767_to_fp16 = const()[name = string("const_2767_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_335_cast_fp16 = add(x = mean_118_cast_fp16, y = const_2767_to_fp16)[name = string("add_335_cast_fp16")];
            fp32 rsqrt_118_epsilon_0 = const()[name = string("rsqrt_118_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_118_cast_fp16 = rsqrt(epsilon = rsqrt_118_epsilon_0, x = add_335_cast_fp16)[name = string("rsqrt_118_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_406_cast_fp16 = mul(x = add_334_cast_fp16, y = rsqrt_118_cast_fp16)[name = string("mul_406_cast_fp16")];
            tensor<fp16, [768]> add_336_to_fp16 = const()[name = string("add_336_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284631424)))];
            tensor<fp16, [1, 512, 768]> mul_407_cast_fp16 = mul(x = mul_406_cast_fp16, y = add_336_to_fp16)[name = string("mul_407_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284633024))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285517824))))[name = string("p_st_0_model_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = mul_407_cast_fp16)[name = string("linear_137_cast_fp16")];
            string gelu_19_mode_0 = const()[name = string("gelu_19_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_19_cast_fp16 = gelu(mode = gelu_19_mode_0, x = linear_137_cast_fp16)[name = string("gelu_19_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285520192))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286404992))))[name = string("p_st_0_model_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = mul_407_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_408_cast_fp16 = mul(x = gelu_19_cast_fp16, y = linear_138_cast_fp16)[name = string("mul_408_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286407360))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287292160))))[name = string("p_st_0_model_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_139_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = mul_408_cast_fp16)[name = string("linear_139_cast_fp16")];
            fp16 const_2772_promoted_to_fp16 = const()[name = string("const_2772_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_120_cast_fp16 = pow(x = linear_139_cast_fp16, y = const_2772_promoted_to_fp16)[name = string("pow_120_cast_fp16")];
            tensor<int32, [1]> mean_119_axes_0 = const()[name = string("mean_119_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_119_keep_dims_0 = const()[name = string("mean_119_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_119_cast_fp16 = reduce_mean(axes = mean_119_axes_0, keep_dims = mean_119_keep_dims_0, x = pow_120_cast_fp16)[name = string("mean_119_cast_fp16")];
            fp16 const_2775_to_fp16 = const()[name = string("const_2775_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_337_cast_fp16 = add(x = mean_119_cast_fp16, y = const_2775_to_fp16)[name = string("add_337_cast_fp16")];
            fp32 rsqrt_119_epsilon_0 = const()[name = string("rsqrt_119_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_119_cast_fp16 = rsqrt(epsilon = rsqrt_119_epsilon_0, x = add_337_cast_fp16)[name = string("rsqrt_119_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_409_cast_fp16 = mul(x = linear_139_cast_fp16, y = rsqrt_119_cast_fp16)[name = string("mul_409_cast_fp16")];
            tensor<fp16, [768]> add_338_to_fp16 = const()[name = string("add_338_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287293760)))];
            tensor<fp16, [1, 512, 768]> mul_410_cast_fp16 = mul(x = mul_409_cast_fp16, y = add_338_to_fp16)[name = string("mul_410_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_339_cast_fp16 = add(x = add_334_cast_fp16, y = mul_410_cast_fp16)[name = string("add_339_cast_fp16")];
            fp16 const_2780_promoted_to_fp16 = const()[name = string("const_2780_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_121_cast_fp16 = pow(x = add_339_cast_fp16, y = const_2780_promoted_to_fp16)[name = string("pow_121_cast_fp16")];
            tensor<int32, [1]> mean_120_axes_0 = const()[name = string("mean_120_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_120_keep_dims_0 = const()[name = string("mean_120_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_120_cast_fp16 = reduce_mean(axes = mean_120_axes_0, keep_dims = mean_120_keep_dims_0, x = pow_121_cast_fp16)[name = string("mean_120_cast_fp16")];
            fp16 const_2783_to_fp16 = const()[name = string("const_2783_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_340_cast_fp16 = add(x = mean_120_cast_fp16, y = const_2783_to_fp16)[name = string("add_340_cast_fp16")];
            fp32 rsqrt_120_epsilon_0 = const()[name = string("rsqrt_120_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_120_cast_fp16 = rsqrt(epsilon = rsqrt_120_epsilon_0, x = add_340_cast_fp16)[name = string("rsqrt_120_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_411_cast_fp16 = mul(x = add_339_cast_fp16, y = rsqrt_120_cast_fp16)[name = string("mul_411_cast_fp16")];
            tensor<fp16, [768]> add_341_to_fp16 = const()[name = string("add_341_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287295360)))];
            tensor<fp16, [1, 512, 768]> mul_412_cast_fp16 = mul(x = mul_411_cast_fp16, y = add_341_to_fp16)[name = string("mul_412_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287296960))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287886848))))[name = string("p_st_0_model_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_140_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = mul_412_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> const_2787 = const()[name = string("const_2787"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_120_cast_fp16 = reshape(shape = const_2787, x = linear_140_cast_fp16)[name = string("view_120_cast_fp16")];
            tensor<int32, [4]> transpose_124_perm_0 = const()[name = string("transpose_124_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287888448))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288085120))))[name = string("p_st_0_model_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_141_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = mul_412_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<int32, [4]> const_2790 = const()[name = string("const_2790"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_121_cast_fp16 = reshape(shape = const_2790, x = linear_141_cast_fp16)[name = string("view_121_cast_fp16")];
            tensor<int32, [4]> transpose_125_perm_0 = const()[name = string("transpose_125_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288085696))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288282368))))[name = string("p_st_0_model_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_142_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = mul_412_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<int32, [4]> const_2793 = const()[name = string("const_2793"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_122_cast_fp16 = reshape(shape = const_2793, x = linear_142_cast_fp16)[name = string("view_122_cast_fp16")];
            tensor<int32, [4]> transpose_126_perm_0 = const()[name = string("transpose_126_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_2797_promoted_to_fp16 = const()[name = string("const_2797_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_124_cast_fp16 = transpose(perm = transpose_124_perm_0, x = view_120_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 3, 512, 256]> pow_122_cast_fp16 = pow(x = transpose_124_cast_fp16, y = const_2797_promoted_to_fp16)[name = string("pow_122_cast_fp16")];
            tensor<int32, [1]> mean_121_axes_0 = const()[name = string("mean_121_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_121_keep_dims_0 = const()[name = string("mean_121_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_121_cast_fp16 = reduce_mean(axes = mean_121_axes_0, keep_dims = mean_121_keep_dims_0, x = pow_122_cast_fp16)[name = string("mean_121_cast_fp16")];
            fp16 const_2800_to_fp16 = const()[name = string("const_2800_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_342_cast_fp16 = add(x = mean_121_cast_fp16, y = const_2800_to_fp16)[name = string("add_342_cast_fp16")];
            fp32 rsqrt_121_epsilon_0 = const()[name = string("rsqrt_121_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_121_cast_fp16 = rsqrt(epsilon = rsqrt_121_epsilon_0, x = add_342_cast_fp16)[name = string("rsqrt_121_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_413_cast_fp16 = mul(x = transpose_124_cast_fp16, y = rsqrt_121_cast_fp16)[name = string("mul_413_cast_fp16")];
            tensor<fp16, [256]> add_343_to_fp16 = const()[name = string("add_343_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288282944)))];
            tensor<fp16, [1, 3, 512, 256]> mul_414_cast_fp16 = mul(x = mul_413_cast_fp16, y = add_343_to_fp16)[name = string("mul_414_cast_fp16")];
            fp16 const_2805_promoted_to_fp16 = const()[name = string("const_2805_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_125_cast_fp16 = transpose(perm = transpose_125_perm_0, x = view_121_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 1, 512, 256]> pow_123_cast_fp16 = pow(x = transpose_125_cast_fp16, y = const_2805_promoted_to_fp16)[name = string("pow_123_cast_fp16")];
            tensor<int32, [1]> mean_122_axes_0 = const()[name = string("mean_122_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_122_keep_dims_0 = const()[name = string("mean_122_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_122_cast_fp16 = reduce_mean(axes = mean_122_axes_0, keep_dims = mean_122_keep_dims_0, x = pow_123_cast_fp16)[name = string("mean_122_cast_fp16")];
            fp16 const_2808_to_fp16 = const()[name = string("const_2808_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_344_cast_fp16 = add(x = mean_122_cast_fp16, y = const_2808_to_fp16)[name = string("add_344_cast_fp16")];
            fp32 rsqrt_122_epsilon_0 = const()[name = string("rsqrt_122_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_122_cast_fp16 = rsqrt(epsilon = rsqrt_122_epsilon_0, x = add_344_cast_fp16)[name = string("rsqrt_122_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_415_cast_fp16 = mul(x = transpose_125_cast_fp16, y = rsqrt_122_cast_fp16)[name = string("mul_415_cast_fp16")];
            tensor<fp16, [256]> add_345_to_fp16 = const()[name = string("add_345_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288283520)))];
            tensor<fp16, [1, 1, 512, 256]> mul_416_cast_fp16 = mul(x = mul_415_cast_fp16, y = add_345_to_fp16)[name = string("mul_416_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_417_cast_fp16 = mul(x = mul_414_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_417_cast_fp16")];
            tensor<int32, [4]> slice_513_begin_0 = const()[name = string("slice_513_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_513_end_0 = const()[name = string("slice_513_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_513_end_mask_0 = const()[name = string("slice_513_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_513_cast_fp16 = slice_by_index(begin = slice_513_begin_0, end = slice_513_end_0, end_mask = slice_513_end_mask_0, x = mul_414_cast_fp16)[name = string("slice_513_cast_fp16")];
            tensor<int32, [4]> slice_514_begin_0 = const()[name = string("slice_514_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_514_end_0 = const()[name = string("slice_514_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_514_end_mask_0 = const()[name = string("slice_514_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_514_cast_fp16 = slice_by_index(begin = slice_514_begin_0, end = slice_514_end_0, end_mask = slice_514_end_mask_0, x = mul_414_cast_fp16)[name = string("slice_514_cast_fp16")];
            fp16 const_2820_promoted_to_fp16 = const()[name = string("const_2820_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_40_cast_fp16 = mul(x = slice_514_cast_fp16, y = const_2820_promoted_to_fp16)[name = string("neg_40_cast_fp16")];
            int32 const_2821 = const()[name = string("const_2821"), val = int32(-1)];
            bool cat_104_interleave_0 = const()[name = string("cat_104_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_104_cast_fp16 = concat(axis = const_2821, interleave = cat_104_interleave_0, values = (neg_40_cast_fp16, slice_513_cast_fp16))[name = string("cat_104_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_418_cast_fp16 = mul(x = cat_104_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_418_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_346_cast_fp16 = add(x = mul_417_cast_fp16, y = mul_418_cast_fp16)[name = string("add_346_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_419_cast_fp16 = mul(x = mul_416_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_419_cast_fp16")];
            tensor<int32, [4]> slice_515_begin_0 = const()[name = string("slice_515_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_515_end_0 = const()[name = string("slice_515_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_515_end_mask_0 = const()[name = string("slice_515_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_515_cast_fp16 = slice_by_index(begin = slice_515_begin_0, end = slice_515_end_0, end_mask = slice_515_end_mask_0, x = mul_416_cast_fp16)[name = string("slice_515_cast_fp16")];
            tensor<int32, [4]> slice_516_begin_0 = const()[name = string("slice_516_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_516_end_0 = const()[name = string("slice_516_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_516_end_mask_0 = const()[name = string("slice_516_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_516_cast_fp16 = slice_by_index(begin = slice_516_begin_0, end = slice_516_end_0, end_mask = slice_516_end_mask_0, x = mul_416_cast_fp16)[name = string("slice_516_cast_fp16")];
            fp16 const_2828_promoted_to_fp16 = const()[name = string("const_2828_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_41_cast_fp16 = mul(x = slice_516_cast_fp16, y = const_2828_promoted_to_fp16)[name = string("neg_41_cast_fp16")];
            int32 const_2829 = const()[name = string("const_2829"), val = int32(-1)];
            bool cat_105_interleave_0 = const()[name = string("cat_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_105_cast_fp16 = concat(axis = const_2829, interleave = cat_105_interleave_0, values = (neg_41_cast_fp16, slice_515_cast_fp16))[name = string("cat_105_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_420_cast_fp16 = mul(x = cat_105_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_420_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_347_cast_fp16 = add(x = mul_419_cast_fp16, y = mul_420_cast_fp16)[name = string("add_347_cast_fp16")];
            int32 const_2830 = const()[name = string("const_2830"), val = int32(-2)];
            bool cat_106_interleave_0 = const()[name = string("cat_106_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_106_cast_fp16 = concat(axis = const_2830, interleave = cat_106_interleave_0, values = add_347_cast_fp16)[name = string("cat_106_cast_fp16")];
            int32 const_2831 = const()[name = string("const_2831"), val = int32(-2)];
            bool cat_107_interleave_0 = const()[name = string("cat_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_126_cast_fp16 = transpose(perm = transpose_126_perm_0, x = view_122_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 1, 512, 256]> cat_107_cast_fp16 = concat(axis = const_2831, interleave = cat_107_interleave_0, values = transpose_126_cast_fp16)[name = string("cat_107_cast_fp16")];
            tensor<int32, [1]> unsqueeze_159_axes_0 = const()[name = string("unsqueeze_159_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_159_cast_fp16 = expand_dims(axes = unsqueeze_159_axes_0, x = cat_106_cast_fp16)[name = string("unsqueeze_159_cast_fp16")];
            tensor<int32, [5]> expand_66_reps_0 = const()[name = string("expand_66_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_66_cast_fp16 = tile(reps = expand_66_reps_0, x = unsqueeze_159_cast_fp16)[name = string("expand_66_cast_fp16")];
            tensor<int32, [4]> const_2846 = const()[name = string("const_2846"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_123_cast_fp16 = reshape(shape = const_2846, x = expand_66_cast_fp16)[name = string("view_123_cast_fp16")];
            tensor<int32, [1]> unsqueeze_160_axes_0 = const()[name = string("unsqueeze_160_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_160_cast_fp16 = expand_dims(axes = unsqueeze_160_axes_0, x = cat_107_cast_fp16)[name = string("unsqueeze_160_cast_fp16")];
            tensor<int32, [5]> expand_67_reps_0 = const()[name = string("expand_67_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_67_cast_fp16 = tile(reps = expand_67_reps_0, x = unsqueeze_160_cast_fp16)[name = string("expand_67_cast_fp16")];
            tensor<int32, [4]> const_2861 = const()[name = string("const_2861"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_124_cast_fp16 = reshape(shape = const_2861, x = expand_67_cast_fp16)[name = string("view_124_cast_fp16")];
            bool matmul_64_transpose_x_1 = const()[name = string("matmul_64_transpose_x_1"), val = bool(false)];
            bool matmul_64_transpose_y_1 = const()[name = string("matmul_64_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_64_cast_fp16 = matmul(transpose_x = matmul_64_transpose_x_1, transpose_y = matmul_64_transpose_y_1, x = add_346_cast_fp16, y = view_123_cast_fp16)[name = string("matmul_64_cast_fp16")];
            fp16 const_2864_to_fp16 = const()[name = string("const_2864_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_421_cast_fp16 = mul(x = matmul_64_cast_fp16, y = const_2864_to_fp16)[name = string("mul_421_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_348_cast_fp16 = add(x = mul_421_cast_fp16, y = expand_cast_fp16)[name = string("add_348_cast_fp16")];
            int32 const_2874 = const()[name = string("const_2874"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_20_cast_fp16 = softmax(axis = const_2874, x = add_348_cast_fp16)[name = string("softmax_20_cast_fp16")];
            bool matmul_65_transpose_x_0 = const()[name = string("matmul_65_transpose_x_0"), val = bool(false)];
            bool matmul_65_transpose_y_0 = const()[name = string("matmul_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_65_cast_fp16 = matmul(transpose_x = matmul_65_transpose_x_0, transpose_y = matmul_65_transpose_y_0, x = softmax_20_cast_fp16, y = view_124_cast_fp16)[name = string("matmul_65_cast_fp16")];
            tensor<int32, [4]> transpose_128_perm_0 = const()[name = string("transpose_128_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_2879 = const()[name = string("const_2879"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_128_cast_fp16 = transpose(perm = transpose_128_perm_0, x = matmul_65_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 512, 768]> view_125_cast_fp16 = reshape(shape = const_2879, x = transpose_128_cast_fp16)[name = string("view_125_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288284096))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288873984))))[name = string("p_st_0_model_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_143_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = view_125_cast_fp16)[name = string("linear_143_cast_fp16")];
            fp16 const_2881_promoted_to_fp16 = const()[name = string("const_2881_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_124_cast_fp16 = pow(x = linear_143_cast_fp16, y = const_2881_promoted_to_fp16)[name = string("pow_124_cast_fp16")];
            tensor<int32, [1]> mean_123_axes_0 = const()[name = string("mean_123_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_123_keep_dims_0 = const()[name = string("mean_123_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_123_cast_fp16 = reduce_mean(axes = mean_123_axes_0, keep_dims = mean_123_keep_dims_0, x = pow_124_cast_fp16)[name = string("mean_123_cast_fp16")];
            fp16 const_2884_to_fp16 = const()[name = string("const_2884_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_349_cast_fp16 = add(x = mean_123_cast_fp16, y = const_2884_to_fp16)[name = string("add_349_cast_fp16")];
            fp32 rsqrt_123_epsilon_0 = const()[name = string("rsqrt_123_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_123_cast_fp16 = rsqrt(epsilon = rsqrt_123_epsilon_0, x = add_349_cast_fp16)[name = string("rsqrt_123_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_422_cast_fp16 = mul(x = linear_143_cast_fp16, y = rsqrt_123_cast_fp16)[name = string("mul_422_cast_fp16")];
            tensor<fp16, [768]> add_350_to_fp16 = const()[name = string("add_350_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288875584)))];
            tensor<fp16, [1, 512, 768]> mul_423_cast_fp16 = mul(x = mul_422_cast_fp16, y = add_350_to_fp16)[name = string("mul_423_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_351_cast_fp16 = add(x = add_339_cast_fp16, y = mul_423_cast_fp16)[name = string("add_351_cast_fp16")];
            fp16 const_2889_promoted_to_fp16 = const()[name = string("const_2889_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_125_cast_fp16 = pow(x = add_351_cast_fp16, y = const_2889_promoted_to_fp16)[name = string("pow_125_cast_fp16")];
            tensor<int32, [1]> mean_124_axes_0 = const()[name = string("mean_124_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_124_keep_dims_0 = const()[name = string("mean_124_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_124_cast_fp16 = reduce_mean(axes = mean_124_axes_0, keep_dims = mean_124_keep_dims_0, x = pow_125_cast_fp16)[name = string("mean_124_cast_fp16")];
            fp16 const_2892_to_fp16 = const()[name = string("const_2892_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_352_cast_fp16 = add(x = mean_124_cast_fp16, y = const_2892_to_fp16)[name = string("add_352_cast_fp16")];
            fp32 rsqrt_124_epsilon_0 = const()[name = string("rsqrt_124_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_124_cast_fp16 = rsqrt(epsilon = rsqrt_124_epsilon_0, x = add_352_cast_fp16)[name = string("rsqrt_124_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_424_cast_fp16 = mul(x = add_351_cast_fp16, y = rsqrt_124_cast_fp16)[name = string("mul_424_cast_fp16")];
            tensor<fp16, [768]> add_353_to_fp16 = const()[name = string("add_353_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288877184)))];
            tensor<fp16, [1, 512, 768]> mul_425_cast_fp16 = mul(x = mul_424_cast_fp16, y = add_353_to_fp16)[name = string("mul_425_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288878784))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289763584))))[name = string("p_st_0_model_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_144_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = mul_425_cast_fp16)[name = string("linear_144_cast_fp16")];
            string gelu_20_mode_0 = const()[name = string("gelu_20_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_20_cast_fp16 = gelu(mode = gelu_20_mode_0, x = linear_144_cast_fp16)[name = string("gelu_20_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289765952))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290650752))))[name = string("p_st_0_model_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = mul_425_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_426_cast_fp16 = mul(x = gelu_20_cast_fp16, y = linear_145_cast_fp16)[name = string("mul_426_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290653120))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291537920))))[name = string("p_st_0_model_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_146_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = mul_426_cast_fp16)[name = string("linear_146_cast_fp16")];
            fp16 const_2897_promoted_to_fp16 = const()[name = string("const_2897_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_126_cast_fp16 = pow(x = linear_146_cast_fp16, y = const_2897_promoted_to_fp16)[name = string("pow_126_cast_fp16")];
            tensor<int32, [1]> mean_125_axes_0 = const()[name = string("mean_125_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_125_keep_dims_0 = const()[name = string("mean_125_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_125_cast_fp16 = reduce_mean(axes = mean_125_axes_0, keep_dims = mean_125_keep_dims_0, x = pow_126_cast_fp16)[name = string("mean_125_cast_fp16")];
            fp16 const_2900_to_fp16 = const()[name = string("const_2900_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_354_cast_fp16 = add(x = mean_125_cast_fp16, y = const_2900_to_fp16)[name = string("add_354_cast_fp16")];
            fp32 rsqrt_125_epsilon_0 = const()[name = string("rsqrt_125_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_125_cast_fp16 = rsqrt(epsilon = rsqrt_125_epsilon_0, x = add_354_cast_fp16)[name = string("rsqrt_125_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_427_cast_fp16 = mul(x = linear_146_cast_fp16, y = rsqrt_125_cast_fp16)[name = string("mul_427_cast_fp16")];
            tensor<fp16, [768]> add_355_to_fp16 = const()[name = string("add_355_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291539520)))];
            tensor<fp16, [1, 512, 768]> mul_428_cast_fp16 = mul(x = mul_427_cast_fp16, y = add_355_to_fp16)[name = string("mul_428_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_356_cast_fp16 = add(x = add_351_cast_fp16, y = mul_428_cast_fp16)[name = string("add_356_cast_fp16")];
            fp16 const_2905_promoted_to_fp16 = const()[name = string("const_2905_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_127_cast_fp16 = pow(x = add_356_cast_fp16, y = const_2905_promoted_to_fp16)[name = string("pow_127_cast_fp16")];
            tensor<int32, [1]> mean_126_axes_0 = const()[name = string("mean_126_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_126_keep_dims_0 = const()[name = string("mean_126_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_126_cast_fp16 = reduce_mean(axes = mean_126_axes_0, keep_dims = mean_126_keep_dims_0, x = pow_127_cast_fp16)[name = string("mean_126_cast_fp16")];
            fp16 const_2908_to_fp16 = const()[name = string("const_2908_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_357_cast_fp16 = add(x = mean_126_cast_fp16, y = const_2908_to_fp16)[name = string("add_357_cast_fp16")];
            fp32 rsqrt_126_epsilon_0 = const()[name = string("rsqrt_126_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_126_cast_fp16 = rsqrt(epsilon = rsqrt_126_epsilon_0, x = add_357_cast_fp16)[name = string("rsqrt_126_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_429_cast_fp16 = mul(x = add_356_cast_fp16, y = rsqrt_126_cast_fp16)[name = string("mul_429_cast_fp16")];
            tensor<fp16, [768]> add_358_to_fp16 = const()[name = string("add_358_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291541120)))];
            tensor<fp16, [1, 512, 768]> mul_430_cast_fp16 = mul(x = mul_429_cast_fp16, y = add_358_to_fp16)[name = string("mul_430_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291542720))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292132608))))[name = string("p_st_0_model_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_147_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = mul_430_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> const_2912 = const()[name = string("const_2912"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_126_cast_fp16 = reshape(shape = const_2912, x = linear_147_cast_fp16)[name = string("view_126_cast_fp16")];
            tensor<int32, [4]> transpose_129_perm_0 = const()[name = string("transpose_129_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292134208))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292330880))))[name = string("p_st_0_model_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_148_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = mul_430_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> const_2915 = const()[name = string("const_2915"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_127_cast_fp16 = reshape(shape = const_2915, x = linear_148_cast_fp16)[name = string("view_127_cast_fp16")];
            tensor<int32, [4]> transpose_130_perm_0 = const()[name = string("transpose_130_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292331456))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292528128))))[name = string("p_st_0_model_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_149_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = mul_430_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> const_2918 = const()[name = string("const_2918"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_128_cast_fp16 = reshape(shape = const_2918, x = linear_149_cast_fp16)[name = string("view_128_cast_fp16")];
            tensor<int32, [4]> transpose_131_perm_0 = const()[name = string("transpose_131_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_2922_promoted_to_fp16 = const()[name = string("const_2922_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_129_cast_fp16 = transpose(perm = transpose_129_perm_0, x = view_126_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 3, 512, 256]> pow_128_cast_fp16 = pow(x = transpose_129_cast_fp16, y = const_2922_promoted_to_fp16)[name = string("pow_128_cast_fp16")];
            tensor<int32, [1]> mean_127_axes_0 = const()[name = string("mean_127_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_127_keep_dims_0 = const()[name = string("mean_127_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_127_cast_fp16 = reduce_mean(axes = mean_127_axes_0, keep_dims = mean_127_keep_dims_0, x = pow_128_cast_fp16)[name = string("mean_127_cast_fp16")];
            fp16 const_2925_to_fp16 = const()[name = string("const_2925_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_359_cast_fp16 = add(x = mean_127_cast_fp16, y = const_2925_to_fp16)[name = string("add_359_cast_fp16")];
            fp32 rsqrt_127_epsilon_0 = const()[name = string("rsqrt_127_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_127_cast_fp16 = rsqrt(epsilon = rsqrt_127_epsilon_0, x = add_359_cast_fp16)[name = string("rsqrt_127_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_431_cast_fp16 = mul(x = transpose_129_cast_fp16, y = rsqrt_127_cast_fp16)[name = string("mul_431_cast_fp16")];
            tensor<fp16, [256]> add_360_to_fp16 = const()[name = string("add_360_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292528704)))];
            tensor<fp16, [1, 3, 512, 256]> mul_432_cast_fp16 = mul(x = mul_431_cast_fp16, y = add_360_to_fp16)[name = string("mul_432_cast_fp16")];
            fp16 const_2930_promoted_to_fp16 = const()[name = string("const_2930_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_130_cast_fp16 = transpose(perm = transpose_130_perm_0, x = view_127_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 1, 512, 256]> pow_129_cast_fp16 = pow(x = transpose_130_cast_fp16, y = const_2930_promoted_to_fp16)[name = string("pow_129_cast_fp16")];
            tensor<int32, [1]> mean_128_axes_0 = const()[name = string("mean_128_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_128_keep_dims_0 = const()[name = string("mean_128_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_128_cast_fp16 = reduce_mean(axes = mean_128_axes_0, keep_dims = mean_128_keep_dims_0, x = pow_129_cast_fp16)[name = string("mean_128_cast_fp16")];
            fp16 const_2933_to_fp16 = const()[name = string("const_2933_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_361_cast_fp16 = add(x = mean_128_cast_fp16, y = const_2933_to_fp16)[name = string("add_361_cast_fp16")];
            fp32 rsqrt_128_epsilon_0 = const()[name = string("rsqrt_128_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_128_cast_fp16 = rsqrt(epsilon = rsqrt_128_epsilon_0, x = add_361_cast_fp16)[name = string("rsqrt_128_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_433_cast_fp16 = mul(x = transpose_130_cast_fp16, y = rsqrt_128_cast_fp16)[name = string("mul_433_cast_fp16")];
            tensor<fp16, [256]> add_362_to_fp16 = const()[name = string("add_362_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292529280)))];
            tensor<fp16, [1, 1, 512, 256]> mul_434_cast_fp16 = mul(x = mul_433_cast_fp16, y = add_362_to_fp16)[name = string("mul_434_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_435_cast_fp16 = mul(x = mul_432_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_435_cast_fp16")];
            tensor<int32, [4]> slice_536_begin_0 = const()[name = string("slice_536_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_536_end_0 = const()[name = string("slice_536_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_536_end_mask_0 = const()[name = string("slice_536_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_536_cast_fp16 = slice_by_index(begin = slice_536_begin_0, end = slice_536_end_0, end_mask = slice_536_end_mask_0, x = mul_432_cast_fp16)[name = string("slice_536_cast_fp16")];
            tensor<int32, [4]> slice_537_begin_0 = const()[name = string("slice_537_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_537_end_0 = const()[name = string("slice_537_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_537_end_mask_0 = const()[name = string("slice_537_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_537_cast_fp16 = slice_by_index(begin = slice_537_begin_0, end = slice_537_end_0, end_mask = slice_537_end_mask_0, x = mul_432_cast_fp16)[name = string("slice_537_cast_fp16")];
            fp16 const_2945_promoted_to_fp16 = const()[name = string("const_2945_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_42_cast_fp16 = mul(x = slice_537_cast_fp16, y = const_2945_promoted_to_fp16)[name = string("neg_42_cast_fp16")];
            int32 const_2946 = const()[name = string("const_2946"), val = int32(-1)];
            bool cat_108_interleave_0 = const()[name = string("cat_108_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_108_cast_fp16 = concat(axis = const_2946, interleave = cat_108_interleave_0, values = (neg_42_cast_fp16, slice_536_cast_fp16))[name = string("cat_108_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_436_cast_fp16 = mul(x = cat_108_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_436_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_363_cast_fp16 = add(x = mul_435_cast_fp16, y = mul_436_cast_fp16)[name = string("add_363_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_437_cast_fp16 = mul(x = mul_434_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_437_cast_fp16")];
            tensor<int32, [4]> slice_538_begin_0 = const()[name = string("slice_538_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_538_end_0 = const()[name = string("slice_538_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_538_end_mask_0 = const()[name = string("slice_538_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_538_cast_fp16 = slice_by_index(begin = slice_538_begin_0, end = slice_538_end_0, end_mask = slice_538_end_mask_0, x = mul_434_cast_fp16)[name = string("slice_538_cast_fp16")];
            tensor<int32, [4]> slice_539_begin_0 = const()[name = string("slice_539_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_539_end_0 = const()[name = string("slice_539_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_539_end_mask_0 = const()[name = string("slice_539_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_539_cast_fp16 = slice_by_index(begin = slice_539_begin_0, end = slice_539_end_0, end_mask = slice_539_end_mask_0, x = mul_434_cast_fp16)[name = string("slice_539_cast_fp16")];
            fp16 const_2953_promoted_to_fp16 = const()[name = string("const_2953_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_43_cast_fp16 = mul(x = slice_539_cast_fp16, y = const_2953_promoted_to_fp16)[name = string("neg_43_cast_fp16")];
            int32 const_2954 = const()[name = string("const_2954"), val = int32(-1)];
            bool cat_109_interleave_0 = const()[name = string("cat_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_109_cast_fp16 = concat(axis = const_2954, interleave = cat_109_interleave_0, values = (neg_43_cast_fp16, slice_538_cast_fp16))[name = string("cat_109_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_438_cast_fp16 = mul(x = cat_109_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_438_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_364_cast_fp16 = add(x = mul_437_cast_fp16, y = mul_438_cast_fp16)[name = string("add_364_cast_fp16")];
            int32 const_2955 = const()[name = string("const_2955"), val = int32(-2)];
            bool cat_110_interleave_0 = const()[name = string("cat_110_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_110_cast_fp16 = concat(axis = const_2955, interleave = cat_110_interleave_0, values = add_364_cast_fp16)[name = string("cat_110_cast_fp16")];
            int32 const_2956 = const()[name = string("const_2956"), val = int32(-2)];
            bool cat_111_interleave_0 = const()[name = string("cat_111_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_131_cast_fp16 = transpose(perm = transpose_131_perm_0, x = view_128_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 1, 512, 256]> cat_111_cast_fp16 = concat(axis = const_2956, interleave = cat_111_interleave_0, values = transpose_131_cast_fp16)[name = string("cat_111_cast_fp16")];
            tensor<int32, [1]> unsqueeze_163_axes_0 = const()[name = string("unsqueeze_163_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_163_cast_fp16 = expand_dims(axes = unsqueeze_163_axes_0, x = cat_110_cast_fp16)[name = string("unsqueeze_163_cast_fp16")];
            tensor<int32, [5]> expand_68_reps_0 = const()[name = string("expand_68_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_68_cast_fp16 = tile(reps = expand_68_reps_0, x = unsqueeze_163_cast_fp16)[name = string("expand_68_cast_fp16")];
            tensor<int32, [4]> const_2971 = const()[name = string("const_2971"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_129_cast_fp16 = reshape(shape = const_2971, x = expand_68_cast_fp16)[name = string("view_129_cast_fp16")];
            tensor<int32, [1]> unsqueeze_164_axes_0 = const()[name = string("unsqueeze_164_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_164_cast_fp16 = expand_dims(axes = unsqueeze_164_axes_0, x = cat_111_cast_fp16)[name = string("unsqueeze_164_cast_fp16")];
            tensor<int32, [5]> expand_69_reps_0 = const()[name = string("expand_69_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_69_cast_fp16 = tile(reps = expand_69_reps_0, x = unsqueeze_164_cast_fp16)[name = string("expand_69_cast_fp16")];
            tensor<int32, [4]> const_2986 = const()[name = string("const_2986"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_130_cast_fp16 = reshape(shape = const_2986, x = expand_69_cast_fp16)[name = string("view_130_cast_fp16")];
            bool matmul_66_transpose_x_1 = const()[name = string("matmul_66_transpose_x_1"), val = bool(false)];
            bool matmul_66_transpose_y_1 = const()[name = string("matmul_66_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_66_cast_fp16 = matmul(transpose_x = matmul_66_transpose_x_1, transpose_y = matmul_66_transpose_y_1, x = add_363_cast_fp16, y = view_129_cast_fp16)[name = string("matmul_66_cast_fp16")];
            fp16 const_2989_to_fp16 = const()[name = string("const_2989_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_439_cast_fp16 = mul(x = matmul_66_cast_fp16, y = const_2989_to_fp16)[name = string("mul_439_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_365_cast_fp16 = add(x = mul_439_cast_fp16, y = expand_cast_fp16)[name = string("add_365_cast_fp16")];
            int32 const_2999 = const()[name = string("const_2999"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_21_cast_fp16 = softmax(axis = const_2999, x = add_365_cast_fp16)[name = string("softmax_21_cast_fp16")];
            bool matmul_67_transpose_x_0 = const()[name = string("matmul_67_transpose_x_0"), val = bool(false)];
            bool matmul_67_transpose_y_0 = const()[name = string("matmul_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_67_cast_fp16 = matmul(transpose_x = matmul_67_transpose_x_0, transpose_y = matmul_67_transpose_y_0, x = softmax_21_cast_fp16, y = view_130_cast_fp16)[name = string("matmul_67_cast_fp16")];
            tensor<int32, [4]> transpose_133_perm_0 = const()[name = string("transpose_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_3004 = const()[name = string("const_3004"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_133_cast_fp16 = transpose(perm = transpose_133_perm_0, x = matmul_67_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 512, 768]> view_131_cast_fp16 = reshape(shape = const_3004, x = transpose_133_cast_fp16)[name = string("view_131_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292529856))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293119744))))[name = string("p_st_0_model_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_150_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = view_131_cast_fp16)[name = string("linear_150_cast_fp16")];
            fp16 const_3006_promoted_to_fp16 = const()[name = string("const_3006_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_130_cast_fp16 = pow(x = linear_150_cast_fp16, y = const_3006_promoted_to_fp16)[name = string("pow_130_cast_fp16")];
            tensor<int32, [1]> mean_129_axes_0 = const()[name = string("mean_129_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_129_keep_dims_0 = const()[name = string("mean_129_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_129_cast_fp16 = reduce_mean(axes = mean_129_axes_0, keep_dims = mean_129_keep_dims_0, x = pow_130_cast_fp16)[name = string("mean_129_cast_fp16")];
            fp16 const_3009_to_fp16 = const()[name = string("const_3009_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_366_cast_fp16 = add(x = mean_129_cast_fp16, y = const_3009_to_fp16)[name = string("add_366_cast_fp16")];
            fp32 rsqrt_129_epsilon_0 = const()[name = string("rsqrt_129_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_129_cast_fp16 = rsqrt(epsilon = rsqrt_129_epsilon_0, x = add_366_cast_fp16)[name = string("rsqrt_129_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_440_cast_fp16 = mul(x = linear_150_cast_fp16, y = rsqrt_129_cast_fp16)[name = string("mul_440_cast_fp16")];
            tensor<fp16, [768]> add_367_to_fp16 = const()[name = string("add_367_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293121344)))];
            tensor<fp16, [1, 512, 768]> mul_441_cast_fp16 = mul(x = mul_440_cast_fp16, y = add_367_to_fp16)[name = string("mul_441_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_368_cast_fp16 = add(x = add_356_cast_fp16, y = mul_441_cast_fp16)[name = string("add_368_cast_fp16")];
            fp16 const_3014_promoted_to_fp16 = const()[name = string("const_3014_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_131_cast_fp16 = pow(x = add_368_cast_fp16, y = const_3014_promoted_to_fp16)[name = string("pow_131_cast_fp16")];
            tensor<int32, [1]> mean_130_axes_0 = const()[name = string("mean_130_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_130_keep_dims_0 = const()[name = string("mean_130_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_130_cast_fp16 = reduce_mean(axes = mean_130_axes_0, keep_dims = mean_130_keep_dims_0, x = pow_131_cast_fp16)[name = string("mean_130_cast_fp16")];
            fp16 const_3017_to_fp16 = const()[name = string("const_3017_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_369_cast_fp16 = add(x = mean_130_cast_fp16, y = const_3017_to_fp16)[name = string("add_369_cast_fp16")];
            fp32 rsqrt_130_epsilon_0 = const()[name = string("rsqrt_130_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_130_cast_fp16 = rsqrt(epsilon = rsqrt_130_epsilon_0, x = add_369_cast_fp16)[name = string("rsqrt_130_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_442_cast_fp16 = mul(x = add_368_cast_fp16, y = rsqrt_130_cast_fp16)[name = string("mul_442_cast_fp16")];
            tensor<fp16, [768]> add_370_to_fp16 = const()[name = string("add_370_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293122944)))];
            tensor<fp16, [1, 512, 768]> mul_443_cast_fp16 = mul(x = mul_442_cast_fp16, y = add_370_to_fp16)[name = string("mul_443_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293124544))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294009344))))[name = string("p_st_0_model_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = mul_443_cast_fp16)[name = string("linear_151_cast_fp16")];
            string gelu_21_mode_0 = const()[name = string("gelu_21_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_21_cast_fp16 = gelu(mode = gelu_21_mode_0, x = linear_151_cast_fp16)[name = string("gelu_21_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294011712))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294896512))))[name = string("p_st_0_model_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = mul_443_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_444_cast_fp16 = mul(x = gelu_21_cast_fp16, y = linear_152_cast_fp16)[name = string("mul_444_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294898880))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295783680))))[name = string("p_st_0_model_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_153_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = mul_444_cast_fp16)[name = string("linear_153_cast_fp16")];
            fp16 const_3022_promoted_to_fp16 = const()[name = string("const_3022_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_132_cast_fp16 = pow(x = linear_153_cast_fp16, y = const_3022_promoted_to_fp16)[name = string("pow_132_cast_fp16")];
            tensor<int32, [1]> mean_131_axes_0 = const()[name = string("mean_131_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_131_keep_dims_0 = const()[name = string("mean_131_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_131_cast_fp16 = reduce_mean(axes = mean_131_axes_0, keep_dims = mean_131_keep_dims_0, x = pow_132_cast_fp16)[name = string("mean_131_cast_fp16")];
            fp16 const_3025_to_fp16 = const()[name = string("const_3025_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_371_cast_fp16 = add(x = mean_131_cast_fp16, y = const_3025_to_fp16)[name = string("add_371_cast_fp16")];
            fp32 rsqrt_131_epsilon_0 = const()[name = string("rsqrt_131_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_131_cast_fp16 = rsqrt(epsilon = rsqrt_131_epsilon_0, x = add_371_cast_fp16)[name = string("rsqrt_131_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_445_cast_fp16 = mul(x = linear_153_cast_fp16, y = rsqrt_131_cast_fp16)[name = string("mul_445_cast_fp16")];
            tensor<fp16, [768]> add_372_to_fp16 = const()[name = string("add_372_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295785280)))];
            tensor<fp16, [1, 512, 768]> mul_446_cast_fp16 = mul(x = mul_445_cast_fp16, y = add_372_to_fp16)[name = string("mul_446_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_373_cast_fp16 = add(x = add_368_cast_fp16, y = mul_446_cast_fp16)[name = string("add_373_cast_fp16")];
            fp16 const_3030_promoted_to_fp16 = const()[name = string("const_3030_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_133_cast_fp16 = pow(x = add_373_cast_fp16, y = const_3030_promoted_to_fp16)[name = string("pow_133_cast_fp16")];
            tensor<int32, [1]> mean_132_axes_0 = const()[name = string("mean_132_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_132_keep_dims_0 = const()[name = string("mean_132_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_132_cast_fp16 = reduce_mean(axes = mean_132_axes_0, keep_dims = mean_132_keep_dims_0, x = pow_133_cast_fp16)[name = string("mean_132_cast_fp16")];
            fp16 const_3033_to_fp16 = const()[name = string("const_3033_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_374_cast_fp16 = add(x = mean_132_cast_fp16, y = const_3033_to_fp16)[name = string("add_374_cast_fp16")];
            fp32 rsqrt_132_epsilon_0 = const()[name = string("rsqrt_132_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_132_cast_fp16 = rsqrt(epsilon = rsqrt_132_epsilon_0, x = add_374_cast_fp16)[name = string("rsqrt_132_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_447_cast_fp16 = mul(x = add_373_cast_fp16, y = rsqrt_132_cast_fp16)[name = string("mul_447_cast_fp16")];
            tensor<fp16, [768]> add_375_to_fp16 = const()[name = string("add_375_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295786880)))];
            tensor<fp16, [1, 512, 768]> mul_448_cast_fp16 = mul(x = mul_447_cast_fp16, y = add_375_to_fp16)[name = string("mul_448_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295788480))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296378368))))[name = string("p_st_0_model_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_154_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = mul_448_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> const_3037 = const()[name = string("const_3037"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_132_cast_fp16 = reshape(shape = const_3037, x = linear_154_cast_fp16)[name = string("view_132_cast_fp16")];
            tensor<int32, [4]> transpose_134_perm_0 = const()[name = string("transpose_134_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296379968))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296576640))))[name = string("p_st_0_model_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_155_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = mul_448_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> const_3040 = const()[name = string("const_3040"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_133_cast_fp16 = reshape(shape = const_3040, x = linear_155_cast_fp16)[name = string("view_133_cast_fp16")];
            tensor<int32, [4]> transpose_135_perm_0 = const()[name = string("transpose_135_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296577216))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296773888))))[name = string("p_st_0_model_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_156_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = mul_448_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> const_3043 = const()[name = string("const_3043"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_134_cast_fp16 = reshape(shape = const_3043, x = linear_156_cast_fp16)[name = string("view_134_cast_fp16")];
            tensor<int32, [4]> transpose_136_perm_0 = const()[name = string("transpose_136_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_3047_promoted_to_fp16 = const()[name = string("const_3047_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_134_cast_fp16 = transpose(perm = transpose_134_perm_0, x = view_132_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 3, 512, 256]> pow_134_cast_fp16 = pow(x = transpose_134_cast_fp16, y = const_3047_promoted_to_fp16)[name = string("pow_134_cast_fp16")];
            tensor<int32, [1]> mean_133_axes_0 = const()[name = string("mean_133_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_133_keep_dims_0 = const()[name = string("mean_133_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_133_cast_fp16 = reduce_mean(axes = mean_133_axes_0, keep_dims = mean_133_keep_dims_0, x = pow_134_cast_fp16)[name = string("mean_133_cast_fp16")];
            fp16 const_3050_to_fp16 = const()[name = string("const_3050_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_376_cast_fp16 = add(x = mean_133_cast_fp16, y = const_3050_to_fp16)[name = string("add_376_cast_fp16")];
            fp32 rsqrt_133_epsilon_0 = const()[name = string("rsqrt_133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_133_cast_fp16 = rsqrt(epsilon = rsqrt_133_epsilon_0, x = add_376_cast_fp16)[name = string("rsqrt_133_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_449_cast_fp16 = mul(x = transpose_134_cast_fp16, y = rsqrt_133_cast_fp16)[name = string("mul_449_cast_fp16")];
            tensor<fp16, [256]> add_377_to_fp16 = const()[name = string("add_377_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296774464)))];
            tensor<fp16, [1, 3, 512, 256]> mul_450_cast_fp16 = mul(x = mul_449_cast_fp16, y = add_377_to_fp16)[name = string("mul_450_cast_fp16")];
            fp16 const_3055_promoted_to_fp16 = const()[name = string("const_3055_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_135_cast_fp16 = transpose(perm = transpose_135_perm_0, x = view_133_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 1, 512, 256]> pow_135_cast_fp16 = pow(x = transpose_135_cast_fp16, y = const_3055_promoted_to_fp16)[name = string("pow_135_cast_fp16")];
            tensor<int32, [1]> mean_134_axes_0 = const()[name = string("mean_134_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_134_keep_dims_0 = const()[name = string("mean_134_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_134_cast_fp16 = reduce_mean(axes = mean_134_axes_0, keep_dims = mean_134_keep_dims_0, x = pow_135_cast_fp16)[name = string("mean_134_cast_fp16")];
            fp16 const_3058_to_fp16 = const()[name = string("const_3058_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_378_cast_fp16 = add(x = mean_134_cast_fp16, y = const_3058_to_fp16)[name = string("add_378_cast_fp16")];
            fp32 rsqrt_134_epsilon_0 = const()[name = string("rsqrt_134_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_134_cast_fp16 = rsqrt(epsilon = rsqrt_134_epsilon_0, x = add_378_cast_fp16)[name = string("rsqrt_134_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_451_cast_fp16 = mul(x = transpose_135_cast_fp16, y = rsqrt_134_cast_fp16)[name = string("mul_451_cast_fp16")];
            tensor<fp16, [256]> add_379_to_fp16 = const()[name = string("add_379_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296775040)))];
            tensor<fp16, [1, 1, 512, 256]> mul_452_cast_fp16 = mul(x = mul_451_cast_fp16, y = add_379_to_fp16)[name = string("mul_452_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_453_cast_fp16 = mul(x = mul_450_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_453_cast_fp16")];
            tensor<int32, [4]> slice_559_begin_0 = const()[name = string("slice_559_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_559_end_0 = const()[name = string("slice_559_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_559_end_mask_0 = const()[name = string("slice_559_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_559_cast_fp16 = slice_by_index(begin = slice_559_begin_0, end = slice_559_end_0, end_mask = slice_559_end_mask_0, x = mul_450_cast_fp16)[name = string("slice_559_cast_fp16")];
            tensor<int32, [4]> slice_560_begin_0 = const()[name = string("slice_560_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_560_end_0 = const()[name = string("slice_560_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_560_end_mask_0 = const()[name = string("slice_560_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_560_cast_fp16 = slice_by_index(begin = slice_560_begin_0, end = slice_560_end_0, end_mask = slice_560_end_mask_0, x = mul_450_cast_fp16)[name = string("slice_560_cast_fp16")];
            fp16 const_3070_promoted_to_fp16 = const()[name = string("const_3070_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_44_cast_fp16 = mul(x = slice_560_cast_fp16, y = const_3070_promoted_to_fp16)[name = string("neg_44_cast_fp16")];
            int32 const_3071 = const()[name = string("const_3071"), val = int32(-1)];
            bool cat_112_interleave_0 = const()[name = string("cat_112_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_112_cast_fp16 = concat(axis = const_3071, interleave = cat_112_interleave_0, values = (neg_44_cast_fp16, slice_559_cast_fp16))[name = string("cat_112_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_454_cast_fp16 = mul(x = cat_112_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_454_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_380_cast_fp16 = add(x = mul_453_cast_fp16, y = mul_454_cast_fp16)[name = string("add_380_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_455_cast_fp16 = mul(x = mul_452_cast_fp16, y = unsqueeze_77_to_fp16_quantized)[name = string("mul_455_cast_fp16")];
            tensor<int32, [4]> slice_561_begin_0 = const()[name = string("slice_561_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_561_end_0 = const()[name = string("slice_561_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_561_end_mask_0 = const()[name = string("slice_561_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_561_cast_fp16 = slice_by_index(begin = slice_561_begin_0, end = slice_561_end_0, end_mask = slice_561_end_mask_0, x = mul_452_cast_fp16)[name = string("slice_561_cast_fp16")];
            tensor<int32, [4]> slice_562_begin_0 = const()[name = string("slice_562_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_562_end_0 = const()[name = string("slice_562_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_562_end_mask_0 = const()[name = string("slice_562_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_562_cast_fp16 = slice_by_index(begin = slice_562_begin_0, end = slice_562_end_0, end_mask = slice_562_end_mask_0, x = mul_452_cast_fp16)[name = string("slice_562_cast_fp16")];
            fp16 const_3078_promoted_to_fp16 = const()[name = string("const_3078_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_45_cast_fp16 = mul(x = slice_562_cast_fp16, y = const_3078_promoted_to_fp16)[name = string("neg_45_cast_fp16")];
            int32 const_3079 = const()[name = string("const_3079"), val = int32(-1)];
            bool cat_113_interleave_0 = const()[name = string("cat_113_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_113_cast_fp16 = concat(axis = const_3079, interleave = cat_113_interleave_0, values = (neg_45_cast_fp16, slice_561_cast_fp16))[name = string("cat_113_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_456_cast_fp16 = mul(x = cat_113_cast_fp16, y = unsqueeze_78_to_fp16_quantized)[name = string("mul_456_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_381_cast_fp16 = add(x = mul_455_cast_fp16, y = mul_456_cast_fp16)[name = string("add_381_cast_fp16")];
            int32 const_3080 = const()[name = string("const_3080"), val = int32(-2)];
            bool cat_114_interleave_0 = const()[name = string("cat_114_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_114_cast_fp16 = concat(axis = const_3080, interleave = cat_114_interleave_0, values = add_381_cast_fp16)[name = string("cat_114_cast_fp16")];
            int32 const_3081 = const()[name = string("const_3081"), val = int32(-2)];
            bool cat_115_interleave_0 = const()[name = string("cat_115_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_136_cast_fp16 = transpose(perm = transpose_136_perm_0, x = view_134_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 1, 512, 256]> cat_115_cast_fp16 = concat(axis = const_3081, interleave = cat_115_interleave_0, values = transpose_136_cast_fp16)[name = string("cat_115_cast_fp16")];
            tensor<int32, [1]> unsqueeze_167_axes_0 = const()[name = string("unsqueeze_167_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_167_cast_fp16 = expand_dims(axes = unsqueeze_167_axes_0, x = cat_114_cast_fp16)[name = string("unsqueeze_167_cast_fp16")];
            tensor<int32, [5]> expand_70_reps_0 = const()[name = string("expand_70_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_70_cast_fp16 = tile(reps = expand_70_reps_0, x = unsqueeze_167_cast_fp16)[name = string("expand_70_cast_fp16")];
            tensor<int32, [4]> const_3096 = const()[name = string("const_3096"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_135_cast_fp16 = reshape(shape = const_3096, x = expand_70_cast_fp16)[name = string("view_135_cast_fp16")];
            tensor<int32, [1]> unsqueeze_168_axes_0 = const()[name = string("unsqueeze_168_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_168_cast_fp16 = expand_dims(axes = unsqueeze_168_axes_0, x = cat_115_cast_fp16)[name = string("unsqueeze_168_cast_fp16")];
            tensor<int32, [5]> expand_71_reps_0 = const()[name = string("expand_71_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_71_cast_fp16 = tile(reps = expand_71_reps_0, x = unsqueeze_168_cast_fp16)[name = string("expand_71_cast_fp16")];
            tensor<int32, [4]> const_3111 = const()[name = string("const_3111"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_136_cast_fp16 = reshape(shape = const_3111, x = expand_71_cast_fp16)[name = string("view_136_cast_fp16")];
            bool matmul_68_transpose_x_1 = const()[name = string("matmul_68_transpose_x_1"), val = bool(false)];
            bool matmul_68_transpose_y_1 = const()[name = string("matmul_68_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_68_cast_fp16 = matmul(transpose_x = matmul_68_transpose_x_1, transpose_y = matmul_68_transpose_y_1, x = add_380_cast_fp16, y = view_135_cast_fp16)[name = string("matmul_68_cast_fp16")];
            fp16 const_3114_to_fp16 = const()[name = string("const_3114_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_457_cast_fp16 = mul(x = matmul_68_cast_fp16, y = const_3114_to_fp16)[name = string("mul_457_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_382_cast_fp16 = add(x = mul_457_cast_fp16, y = expand_cast_fp16)[name = string("add_382_cast_fp16")];
            int32 const_3124 = const()[name = string("const_3124"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_22_cast_fp16 = softmax(axis = const_3124, x = add_382_cast_fp16)[name = string("softmax_22_cast_fp16")];
            bool matmul_69_transpose_x_0 = const()[name = string("matmul_69_transpose_x_0"), val = bool(false)];
            bool matmul_69_transpose_y_0 = const()[name = string("matmul_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_69_cast_fp16 = matmul(transpose_x = matmul_69_transpose_x_0, transpose_y = matmul_69_transpose_y_0, x = softmax_22_cast_fp16, y = view_136_cast_fp16)[name = string("matmul_69_cast_fp16")];
            tensor<int32, [4]> transpose_138_perm_0 = const()[name = string("transpose_138_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_3129 = const()[name = string("const_3129"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_138_cast_fp16 = transpose(perm = transpose_138_perm_0, x = matmul_69_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 512, 768]> view_137_cast_fp16 = reshape(shape = const_3129, x = transpose_138_cast_fp16)[name = string("view_137_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296775616))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297365504))))[name = string("p_st_0_model_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_157_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = view_137_cast_fp16)[name = string("linear_157_cast_fp16")];
            fp16 const_3131_promoted_to_fp16 = const()[name = string("const_3131_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_136_cast_fp16 = pow(x = linear_157_cast_fp16, y = const_3131_promoted_to_fp16)[name = string("pow_136_cast_fp16")];
            tensor<int32, [1]> mean_135_axes_0 = const()[name = string("mean_135_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_135_keep_dims_0 = const()[name = string("mean_135_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_135_cast_fp16 = reduce_mean(axes = mean_135_axes_0, keep_dims = mean_135_keep_dims_0, x = pow_136_cast_fp16)[name = string("mean_135_cast_fp16")];
            fp16 const_3134_to_fp16 = const()[name = string("const_3134_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_383_cast_fp16 = add(x = mean_135_cast_fp16, y = const_3134_to_fp16)[name = string("add_383_cast_fp16")];
            fp32 rsqrt_135_epsilon_0 = const()[name = string("rsqrt_135_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_135_cast_fp16 = rsqrt(epsilon = rsqrt_135_epsilon_0, x = add_383_cast_fp16)[name = string("rsqrt_135_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_458_cast_fp16 = mul(x = linear_157_cast_fp16, y = rsqrt_135_cast_fp16)[name = string("mul_458_cast_fp16")];
            tensor<fp16, [768]> add_384_to_fp16 = const()[name = string("add_384_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297367104)))];
            tensor<fp16, [1, 512, 768]> mul_459_cast_fp16 = mul(x = mul_458_cast_fp16, y = add_384_to_fp16)[name = string("mul_459_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_385_cast_fp16 = add(x = add_373_cast_fp16, y = mul_459_cast_fp16)[name = string("add_385_cast_fp16")];
            fp16 const_3139_promoted_to_fp16 = const()[name = string("const_3139_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_137_cast_fp16 = pow(x = add_385_cast_fp16, y = const_3139_promoted_to_fp16)[name = string("pow_137_cast_fp16")];
            tensor<int32, [1]> mean_136_axes_0 = const()[name = string("mean_136_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_136_keep_dims_0 = const()[name = string("mean_136_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_136_cast_fp16 = reduce_mean(axes = mean_136_axes_0, keep_dims = mean_136_keep_dims_0, x = pow_137_cast_fp16)[name = string("mean_136_cast_fp16")];
            fp16 const_3142_to_fp16 = const()[name = string("const_3142_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_386_cast_fp16 = add(x = mean_136_cast_fp16, y = const_3142_to_fp16)[name = string("add_386_cast_fp16")];
            fp32 rsqrt_136_epsilon_0 = const()[name = string("rsqrt_136_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_136_cast_fp16 = rsqrt(epsilon = rsqrt_136_epsilon_0, x = add_386_cast_fp16)[name = string("rsqrt_136_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_460_cast_fp16 = mul(x = add_385_cast_fp16, y = rsqrt_136_cast_fp16)[name = string("mul_460_cast_fp16")];
            tensor<fp16, [768]> add_387_to_fp16 = const()[name = string("add_387_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297368704)))];
            tensor<fp16, [1, 512, 768]> mul_461_cast_fp16 = mul(x = mul_460_cast_fp16, y = add_387_to_fp16)[name = string("mul_461_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297370304))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298255104))))[name = string("p_st_0_model_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = mul_461_cast_fp16)[name = string("linear_158_cast_fp16")];
            string gelu_22_mode_0 = const()[name = string("gelu_22_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_22_cast_fp16 = gelu(mode = gelu_22_mode_0, x = linear_158_cast_fp16)[name = string("gelu_22_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298257472))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299142272))))[name = string("p_st_0_model_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = mul_461_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_462_cast_fp16 = mul(x = gelu_22_cast_fp16, y = linear_159_cast_fp16)[name = string("mul_462_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299144640))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300029440))))[name = string("p_st_0_model_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_160_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = mul_462_cast_fp16)[name = string("linear_160_cast_fp16")];
            fp16 const_3147_promoted_to_fp16 = const()[name = string("const_3147_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_138_cast_fp16 = pow(x = linear_160_cast_fp16, y = const_3147_promoted_to_fp16)[name = string("pow_138_cast_fp16")];
            tensor<int32, [1]> mean_137_axes_0 = const()[name = string("mean_137_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_137_keep_dims_0 = const()[name = string("mean_137_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_137_cast_fp16 = reduce_mean(axes = mean_137_axes_0, keep_dims = mean_137_keep_dims_0, x = pow_138_cast_fp16)[name = string("mean_137_cast_fp16")];
            fp16 const_3150_to_fp16 = const()[name = string("const_3150_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_388_cast_fp16 = add(x = mean_137_cast_fp16, y = const_3150_to_fp16)[name = string("add_388_cast_fp16")];
            fp32 rsqrt_137_epsilon_0 = const()[name = string("rsqrt_137_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_137_cast_fp16 = rsqrt(epsilon = rsqrt_137_epsilon_0, x = add_388_cast_fp16)[name = string("rsqrt_137_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_463_cast_fp16 = mul(x = linear_160_cast_fp16, y = rsqrt_137_cast_fp16)[name = string("mul_463_cast_fp16")];
            tensor<fp16, [768]> add_389_to_fp16 = const()[name = string("add_389_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300031040)))];
            tensor<fp16, [1, 512, 768]> mul_464_cast_fp16 = mul(x = mul_463_cast_fp16, y = add_389_to_fp16)[name = string("mul_464_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_390_cast_fp16 = add(x = add_385_cast_fp16, y = mul_464_cast_fp16)[name = string("add_390_cast_fp16")];
            fp16 const_3155_promoted_to_fp16 = const()[name = string("const_3155_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_139_cast_fp16 = pow(x = add_390_cast_fp16, y = const_3155_promoted_to_fp16)[name = string("pow_139_cast_fp16")];
            tensor<int32, [1]> mean_138_axes_0 = const()[name = string("mean_138_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_138_keep_dims_0 = const()[name = string("mean_138_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_138_cast_fp16 = reduce_mean(axes = mean_138_axes_0, keep_dims = mean_138_keep_dims_0, x = pow_139_cast_fp16)[name = string("mean_138_cast_fp16")];
            fp16 const_3158_to_fp16 = const()[name = string("const_3158_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_391_cast_fp16 = add(x = mean_138_cast_fp16, y = const_3158_to_fp16)[name = string("add_391_cast_fp16")];
            fp32 rsqrt_138_epsilon_0 = const()[name = string("rsqrt_138_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_138_cast_fp16 = rsqrt(epsilon = rsqrt_138_epsilon_0, x = add_391_cast_fp16)[name = string("rsqrt_138_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_465_cast_fp16 = mul(x = add_390_cast_fp16, y = rsqrt_138_cast_fp16)[name = string("mul_465_cast_fp16")];
            tensor<fp16, [768]> add_392_to_fp16 = const()[name = string("add_392_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300032640)))];
            tensor<fp16, [1, 512, 768]> mul_466_cast_fp16 = mul(x = mul_465_cast_fp16, y = add_392_to_fp16)[name = string("mul_466_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300034240))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300624128))))[name = string("p_st_0_model_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_161_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = mul_466_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<int32, [4]> const_3162 = const()[name = string("const_3162"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 3, 256]> view_138_cast_fp16 = reshape(shape = const_3162, x = linear_161_cast_fp16)[name = string("view_138_cast_fp16")];
            tensor<int32, [4]> transpose_139_perm_0 = const()[name = string("transpose_139_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300625728))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300822400))))[name = string("p_st_0_model_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_162_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = mul_466_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [4]> const_3165 = const()[name = string("const_3165"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_139_cast_fp16 = reshape(shape = const_3165, x = linear_162_cast_fp16)[name = string("view_139_cast_fp16")];
            tensor<int32, [4]> transpose_140_perm_0 = const()[name = string("transpose_140_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [256, 768]> p_st_0_model_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300822976))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301019648))))[name = string("p_st_0_model_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 256]> linear_163_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = p_st_0_model_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = mul_466_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> const_3168 = const()[name = string("const_3168"), val = tensor<int32, [4]>([1, 512, -1, 256])];
            tensor<fp16, [1, 512, 1, 256]> view_140_cast_fp16 = reshape(shape = const_3168, x = linear_163_cast_fp16)[name = string("view_140_cast_fp16")];
            tensor<int32, [4]> transpose_141_perm_0 = const()[name = string("transpose_141_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 const_3172_promoted_to_fp16 = const()[name = string("const_3172_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 3, 512, 256]> transpose_139_cast_fp16 = transpose(perm = transpose_139_perm_0, x = view_138_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 3, 512, 256]> pow_140_cast_fp16 = pow(x = transpose_139_cast_fp16, y = const_3172_promoted_to_fp16)[name = string("pow_140_cast_fp16")];
            tensor<int32, [1]> mean_139_axes_0 = const()[name = string("mean_139_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_139_keep_dims_0 = const()[name = string("mean_139_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 1]> mean_139_cast_fp16 = reduce_mean(axes = mean_139_axes_0, keep_dims = mean_139_keep_dims_0, x = pow_140_cast_fp16)[name = string("mean_139_cast_fp16")];
            fp16 const_3175_to_fp16 = const()[name = string("const_3175_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 3, 512, 1]> add_393_cast_fp16 = add(x = mean_139_cast_fp16, y = const_3175_to_fp16)[name = string("add_393_cast_fp16")];
            fp32 rsqrt_139_epsilon_0 = const()[name = string("rsqrt_139_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 3, 512, 1]> rsqrt_139_cast_fp16 = rsqrt(epsilon = rsqrt_139_epsilon_0, x = add_393_cast_fp16)[name = string("rsqrt_139_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_467_cast_fp16 = mul(x = transpose_139_cast_fp16, y = rsqrt_139_cast_fp16)[name = string("mul_467_cast_fp16")];
            tensor<fp16, [256]> add_394_to_fp16 = const()[name = string("add_394_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301020224)))];
            tensor<fp16, [1, 3, 512, 256]> mul_468_cast_fp16 = mul(x = mul_467_cast_fp16, y = add_394_to_fp16)[name = string("mul_468_cast_fp16")];
            fp16 const_3180_promoted_to_fp16 = const()[name = string("const_3180_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 512, 256]> transpose_140_cast_fp16 = transpose(perm = transpose_140_perm_0, x = view_139_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 1, 512, 256]> pow_141_cast_fp16 = pow(x = transpose_140_cast_fp16, y = const_3180_promoted_to_fp16)[name = string("pow_141_cast_fp16")];
            tensor<int32, [1]> mean_140_axes_0 = const()[name = string("mean_140_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_140_keep_dims_0 = const()[name = string("mean_140_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> mean_140_cast_fp16 = reduce_mean(axes = mean_140_axes_0, keep_dims = mean_140_keep_dims_0, x = pow_141_cast_fp16)[name = string("mean_140_cast_fp16")];
            fp16 const_3183_to_fp16 = const()[name = string("const_3183_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> add_395_cast_fp16 = add(x = mean_140_cast_fp16, y = const_3183_to_fp16)[name = string("add_395_cast_fp16")];
            fp32 rsqrt_140_epsilon_0 = const()[name = string("rsqrt_140_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> rsqrt_140_cast_fp16 = rsqrt(epsilon = rsqrt_140_epsilon_0, x = add_395_cast_fp16)[name = string("rsqrt_140_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_469_cast_fp16 = mul(x = transpose_140_cast_fp16, y = rsqrt_140_cast_fp16)[name = string("mul_469_cast_fp16")];
            tensor<fp16, [256]> add_396_to_fp16 = const()[name = string("add_396_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301020800)))];
            tensor<fp16, [1, 1, 512, 256]> mul_470_cast_fp16 = mul(x = mul_469_cast_fp16, y = add_396_to_fp16)[name = string("mul_470_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_471_cast_fp16 = mul(x = mul_468_cast_fp16, y = unsqueeze_97_to_fp16_quantized)[name = string("mul_471_cast_fp16")];
            tensor<int32, [4]> slice_582_begin_0 = const()[name = string("slice_582_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_582_end_0 = const()[name = string("slice_582_end_0"), val = tensor<int32, [4]>([1, 3, 512, 128])];
            tensor<bool, [4]> slice_582_end_mask_0 = const()[name = string("slice_582_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 3, 512, 128]> slice_582_cast_fp16 = slice_by_index(begin = slice_582_begin_0, end = slice_582_end_0, end_mask = slice_582_end_mask_0, x = mul_468_cast_fp16)[name = string("slice_582_cast_fp16")];
            tensor<int32, [4]> slice_583_begin_0 = const()[name = string("slice_583_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_583_end_0 = const()[name = string("slice_583_end_0"), val = tensor<int32, [4]>([1, 3, 512, 1])];
            tensor<bool, [4]> slice_583_end_mask_0 = const()[name = string("slice_583_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 3, 512, 128]> slice_583_cast_fp16 = slice_by_index(begin = slice_583_begin_0, end = slice_583_end_0, end_mask = slice_583_end_mask_0, x = mul_468_cast_fp16)[name = string("slice_583_cast_fp16")];
            fp16 const_3195_promoted_to_fp16 = const()[name = string("const_3195_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 3, 512, 128]> neg_46_cast_fp16 = mul(x = slice_583_cast_fp16, y = const_3195_promoted_to_fp16)[name = string("neg_46_cast_fp16")];
            int32 const_3196 = const()[name = string("const_3196"), val = int32(-1)];
            bool cat_116_interleave_0 = const()[name = string("cat_116_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> cat_116_cast_fp16 = concat(axis = const_3196, interleave = cat_116_interleave_0, values = (neg_46_cast_fp16, slice_582_cast_fp16))[name = string("cat_116_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> mul_472_cast_fp16 = mul(x = cat_116_cast_fp16, y = unsqueeze_98_to_fp16_quantized)[name = string("mul_472_cast_fp16")];
            tensor<fp16, [1, 3, 512, 256]> add_397_cast_fp16 = add(x = mul_471_cast_fp16, y = mul_472_cast_fp16)[name = string("add_397_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_473_cast_fp16 = mul(x = mul_470_cast_fp16, y = unsqueeze_97_to_fp16_quantized)[name = string("mul_473_cast_fp16")];
            tensor<int32, [4]> slice_584_begin_0 = const()[name = string("slice_584_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> slice_584_end_0 = const()[name = string("slice_584_end_0"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<bool, [4]> slice_584_end_mask_0 = const()[name = string("slice_584_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 128]> slice_584_cast_fp16 = slice_by_index(begin = slice_584_begin_0, end = slice_584_end_0, end_mask = slice_584_end_mask_0, x = mul_470_cast_fp16)[name = string("slice_584_cast_fp16")];
            tensor<int32, [4]> slice_585_begin_0 = const()[name = string("slice_585_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> slice_585_end_0 = const()[name = string("slice_585_end_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<bool, [4]> slice_585_end_mask_0 = const()[name = string("slice_585_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 128]> slice_585_cast_fp16 = slice_by_index(begin = slice_585_begin_0, end = slice_585_end_0, end_mask = slice_585_end_mask_0, x = mul_470_cast_fp16)[name = string("slice_585_cast_fp16")];
            fp16 const_3203_promoted_to_fp16 = const()[name = string("const_3203_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 128]> neg_47_cast_fp16 = mul(x = slice_585_cast_fp16, y = const_3203_promoted_to_fp16)[name = string("neg_47_cast_fp16")];
            int32 const_3204 = const()[name = string("const_3204"), val = int32(-1)];
            bool cat_117_interleave_0 = const()[name = string("cat_117_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_117_cast_fp16 = concat(axis = const_3204, interleave = cat_117_interleave_0, values = (neg_47_cast_fp16, slice_584_cast_fp16))[name = string("cat_117_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> mul_474_cast_fp16 = mul(x = cat_117_cast_fp16, y = unsqueeze_98_to_fp16_quantized)[name = string("mul_474_cast_fp16")];
            tensor<fp16, [1, 1, 512, 256]> add_398_cast_fp16 = add(x = mul_473_cast_fp16, y = mul_474_cast_fp16)[name = string("add_398_cast_fp16")];
            int32 const_3205 = const()[name = string("const_3205"), val = int32(-2)];
            bool cat_118_interleave_0 = const()[name = string("cat_118_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> cat_118_cast_fp16 = concat(axis = const_3205, interleave = cat_118_interleave_0, values = add_398_cast_fp16)[name = string("cat_118_cast_fp16")];
            int32 const_3206 = const()[name = string("const_3206"), val = int32(-2)];
            bool cat_119_interleave_0 = const()[name = string("cat_119_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 256]> transpose_141_cast_fp16 = transpose(perm = transpose_141_perm_0, x = view_140_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 1, 512, 256]> cat_119_cast_fp16 = concat(axis = const_3206, interleave = cat_119_interleave_0, values = transpose_141_cast_fp16)[name = string("cat_119_cast_fp16")];
            tensor<int32, [1]> unsqueeze_171_axes_0 = const()[name = string("unsqueeze_171_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_171_cast_fp16 = expand_dims(axes = unsqueeze_171_axes_0, x = cat_118_cast_fp16)[name = string("unsqueeze_171_cast_fp16")];
            tensor<int32, [5]> expand_72_reps_0 = const()[name = string("expand_72_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_72_cast_fp16 = tile(reps = expand_72_reps_0, x = unsqueeze_171_cast_fp16)[name = string("expand_72_cast_fp16")];
            tensor<int32, [4]> const_3221 = const()[name = string("const_3221"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_141_cast_fp16 = reshape(shape = const_3221, x = expand_72_cast_fp16)[name = string("view_141_cast_fp16")];
            tensor<int32, [1]> unsqueeze_172_axes_0 = const()[name = string("unsqueeze_172_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512, 256]> unsqueeze_172_cast_fp16 = expand_dims(axes = unsqueeze_172_axes_0, x = cat_119_cast_fp16)[name = string("unsqueeze_172_cast_fp16")];
            tensor<int32, [5]> expand_73_reps_0 = const()[name = string("expand_73_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 1, 3, 512, 256]> expand_73_cast_fp16 = tile(reps = expand_73_reps_0, x = unsqueeze_172_cast_fp16)[name = string("expand_73_cast_fp16")];
            tensor<int32, [4]> const_3236 = const()[name = string("const_3236"), val = tensor<int32, [4]>([1, 3, 512, 256])];
            tensor<fp16, [1, 3, 512, 256]> view_142_cast_fp16 = reshape(shape = const_3236, x = expand_73_cast_fp16)[name = string("view_142_cast_fp16")];
            bool matmul_70_transpose_x_1 = const()[name = string("matmul_70_transpose_x_1"), val = bool(false)];
            bool matmul_70_transpose_y_1 = const()[name = string("matmul_70_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 3, 512, 512]> matmul_70_cast_fp16 = matmul(transpose_x = matmul_70_transpose_x_1, transpose_y = matmul_70_transpose_y_1, x = add_397_cast_fp16, y = view_141_cast_fp16)[name = string("matmul_70_cast_fp16")];
            fp16 const_3239_to_fp16 = const()[name = string("const_3239_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 3, 512, 512]> mul_475_cast_fp16 = mul(x = matmul_70_cast_fp16, y = const_3239_to_fp16)[name = string("mul_475_cast_fp16")];
            tensor<fp16, [1, 3, 512, 512]> add_399_cast_fp16 = add(x = mul_475_cast_fp16, y = expand_cast_fp16)[name = string("add_399_cast_fp16")];
            int32 const_3249 = const()[name = string("const_3249"), val = int32(-1)];
            tensor<fp16, [1, 3, 512, 512]> softmax_23_cast_fp16 = softmax(axis = const_3249, x = add_399_cast_fp16)[name = string("softmax_23_cast_fp16")];
            bool matmul_71_transpose_x_0 = const()[name = string("matmul_71_transpose_x_0"), val = bool(false)];
            bool matmul_71_transpose_y_0 = const()[name = string("matmul_71_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 3, 512, 256]> matmul_71_cast_fp16 = matmul(transpose_x = matmul_71_transpose_x_0, transpose_y = matmul_71_transpose_y_0, x = softmax_23_cast_fp16, y = view_142_cast_fp16)[name = string("matmul_71_cast_fp16")];
            tensor<int32, [4]> transpose_143_perm_0 = const()[name = string("transpose_143_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> const_3254 = const()[name = string("const_3254"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 3, 256]> transpose_143_cast_fp16 = transpose(perm = transpose_143_perm_0, x = matmul_71_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 512, 768]> view_143_cast_fp16 = reshape(shape = const_3254, x = transpose_143_cast_fp16)[name = string("view_143_cast_fp16")];
            tensor<fp16, [768, 768]> p_st_0_model_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301021376))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301611264))))[name = string("p_st_0_model_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_164_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = view_143_cast_fp16)[name = string("linear_164_cast_fp16")];
            fp16 const_3256_promoted_to_fp16 = const()[name = string("const_3256_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_142_cast_fp16 = pow(x = linear_164_cast_fp16, y = const_3256_promoted_to_fp16)[name = string("pow_142_cast_fp16")];
            tensor<int32, [1]> mean_141_axes_0 = const()[name = string("mean_141_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_141_keep_dims_0 = const()[name = string("mean_141_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_141_cast_fp16 = reduce_mean(axes = mean_141_axes_0, keep_dims = mean_141_keep_dims_0, x = pow_142_cast_fp16)[name = string("mean_141_cast_fp16")];
            fp16 const_3259_to_fp16 = const()[name = string("const_3259_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_400_cast_fp16 = add(x = mean_141_cast_fp16, y = const_3259_to_fp16)[name = string("add_400_cast_fp16")];
            fp32 rsqrt_141_epsilon_0 = const()[name = string("rsqrt_141_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_141_cast_fp16 = rsqrt(epsilon = rsqrt_141_epsilon_0, x = add_400_cast_fp16)[name = string("rsqrt_141_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_476_cast_fp16 = mul(x = linear_164_cast_fp16, y = rsqrt_141_cast_fp16)[name = string("mul_476_cast_fp16")];
            tensor<fp16, [768]> add_401_to_fp16 = const()[name = string("add_401_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301612864)))];
            tensor<fp16, [1, 512, 768]> mul_477_cast_fp16 = mul(x = mul_476_cast_fp16, y = add_401_to_fp16)[name = string("mul_477_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_402_cast_fp16 = add(x = add_390_cast_fp16, y = mul_477_cast_fp16)[name = string("add_402_cast_fp16")];
            fp16 const_3264_promoted_to_fp16 = const()[name = string("const_3264_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_143_cast_fp16 = pow(x = add_402_cast_fp16, y = const_3264_promoted_to_fp16)[name = string("pow_143_cast_fp16")];
            tensor<int32, [1]> mean_142_axes_0 = const()[name = string("mean_142_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_142_keep_dims_0 = const()[name = string("mean_142_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_142_cast_fp16 = reduce_mean(axes = mean_142_axes_0, keep_dims = mean_142_keep_dims_0, x = pow_143_cast_fp16)[name = string("mean_142_cast_fp16")];
            fp16 const_3267_to_fp16 = const()[name = string("const_3267_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_403_cast_fp16 = add(x = mean_142_cast_fp16, y = const_3267_to_fp16)[name = string("add_403_cast_fp16")];
            fp32 rsqrt_142_epsilon_0 = const()[name = string("rsqrt_142_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_142_cast_fp16 = rsqrt(epsilon = rsqrt_142_epsilon_0, x = add_403_cast_fp16)[name = string("rsqrt_142_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_478_cast_fp16 = mul(x = add_402_cast_fp16, y = rsqrt_142_cast_fp16)[name = string("mul_478_cast_fp16")];
            tensor<fp16, [768]> add_404_to_fp16 = const()[name = string("add_404_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301614464)))];
            tensor<fp16, [1, 512, 768]> mul_479_cast_fp16 = mul(x = mul_478_cast_fp16, y = add_404_to_fp16)[name = string("mul_479_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301616064))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302500864))))[name = string("p_st_0_model_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = mul_479_cast_fp16)[name = string("linear_165_cast_fp16")];
            string gelu_23_mode_0 = const()[name = string("gelu_23_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 512, 1152]> gelu_23_cast_fp16 = gelu(mode = gelu_23_mode_0, x = linear_165_cast_fp16)[name = string("gelu_23_cast_fp16")];
            tensor<fp16, [1152, 768]> p_st_0_model_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1152, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302503232))), scale = tensor<fp16, [1152, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303388032))))[name = string("p_st_0_model_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1152]> linear_166_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = p_st_0_model_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = mul_479_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 512, 1152]> mul_480_cast_fp16 = mul(x = gelu_23_cast_fp16, y = linear_166_cast_fp16)[name = string("mul_480_cast_fp16")];
            tensor<fp16, [768, 1152]> p_st_0_model_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303390400))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304275200))))[name = string("p_st_0_model_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 768]> linear_167_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_0_model_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = mul_480_cast_fp16)[name = string("linear_167_cast_fp16")];
            fp16 const_3272_promoted_to_fp16 = const()[name = string("const_3272_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_144_cast_fp16 = pow(x = linear_167_cast_fp16, y = const_3272_promoted_to_fp16)[name = string("pow_144_cast_fp16")];
            tensor<int32, [1]> mean_143_axes_0 = const()[name = string("mean_143_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_143_keep_dims_0 = const()[name = string("mean_143_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_143_cast_fp16 = reduce_mean(axes = mean_143_axes_0, keep_dims = mean_143_keep_dims_0, x = pow_144_cast_fp16)[name = string("mean_143_cast_fp16")];
            fp16 const_3275_to_fp16 = const()[name = string("const_3275_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_405_cast_fp16 = add(x = mean_143_cast_fp16, y = const_3275_to_fp16)[name = string("add_405_cast_fp16")];
            fp32 rsqrt_143_epsilon_0 = const()[name = string("rsqrt_143_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_143_cast_fp16 = rsqrt(epsilon = rsqrt_143_epsilon_0, x = add_405_cast_fp16)[name = string("rsqrt_143_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_481_cast_fp16 = mul(x = linear_167_cast_fp16, y = rsqrt_143_cast_fp16)[name = string("mul_481_cast_fp16")];
            tensor<fp16, [768]> add_406_to_fp16 = const()[name = string("add_406_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304276800)))];
            tensor<fp16, [1, 512, 768]> mul_482_cast_fp16 = mul(x = mul_481_cast_fp16, y = add_406_to_fp16)[name = string("mul_482_cast_fp16")];
            tensor<fp16, [1, 512, 768]> add_407_cast_fp16 = add(x = add_402_cast_fp16, y = mul_482_cast_fp16)[name = string("add_407_cast_fp16")];
            fp16 const_3280_promoted_to_fp16 = const()[name = string("const_3280_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 768]> pow_145_cast_fp16 = pow(x = add_407_cast_fp16, y = const_3280_promoted_to_fp16)[name = string("pow_145_cast_fp16")];
            tensor<int32, [1]> mean_144_axes_0 = const()[name = string("mean_144_axes_0"), val = tensor<int32, [1]>([-1])];
            bool mean_144_keep_dims_0 = const()[name = string("mean_144_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> mean_144_cast_fp16 = reduce_mean(axes = mean_144_axes_0, keep_dims = mean_144_keep_dims_0, x = pow_145_cast_fp16)[name = string("mean_144_cast_fp16")];
            fp16 const_3283_to_fp16 = const()[name = string("const_3283_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> add_408_cast_fp16 = add(x = mean_144_cast_fp16, y = const_3283_to_fp16)[name = string("add_408_cast_fp16")];
            fp32 rsqrt_144_epsilon_0 = const()[name = string("rsqrt_144_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> rsqrt_144_cast_fp16 = rsqrt(epsilon = rsqrt_144_epsilon_0, x = add_408_cast_fp16)[name = string("rsqrt_144_cast_fp16")];
            tensor<fp16, [1, 512, 768]> mul_483_cast_fp16 = mul(x = add_407_cast_fp16, y = rsqrt_144_cast_fp16)[name = string("mul_483_cast_fp16")];
            tensor<fp16, [768]> add_409_to_fp16 = const()[name = string("add_409_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304278400)))];
            tensor<fp16, [1, 512, 768]> mul_484_cast_fp16 = mul(x = mul_483_cast_fp16, y = add_409_to_fp16)[name = string("mul_484_cast_fp16")];
            tensor<int32, [1]> unsqueeze_173_axes_0 = const()[name = string("unsqueeze_173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [1, 512, 1]> unsqueeze_173 = expand_dims(axes = unsqueeze_173_axes_0, x = attention_mask)[name = string("unsqueeze_173")];
            tensor<int32, [3]> expand_74_reps_0 = const()[name = string("expand_74_reps_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<int32, [1, 512, 768]> expand_74 = tile(reps = expand_74_reps_0, x = unsqueeze_173)[name = string("expand_74")];
            string _to_copy_507_to_fp16_dtype_0 = const()[name = string("_to_copy_507_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 512, 768]> expand_74_to_fp16 = cast(dtype = _to_copy_507_to_fp16_dtype_0, x = expand_74)[name = string("cast_0")];
            tensor<fp16, [1, 512, 768]> mul_485_cast_fp16 = mul(x = mul_484_cast_fp16, y = expand_74_to_fp16)[name = string("mul_485_cast_fp16")];
            tensor<int32, [1]> sum_1_axes_0 = const()[name = string("sum_1_axes_0"), val = tensor<int32, [1]>([1])];
            bool sum_1_keep_dims_0 = const()[name = string("sum_1_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 768]> sum_1_cast_fp16 = reduce_sum(axes = sum_1_axes_0, keep_dims = sum_1_keep_dims_0, x = mul_485_cast_fp16)[name = string("sum_1_cast_fp16")];
            tensor<int32, [1]> sum_2_axes_0 = const()[name = string("sum_2_axes_0"), val = tensor<int32, [1]>([1])];
            bool sum_2_keep_dims_0 = const()[name = string("sum_2_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 768]> sum_2_cast_fp16 = reduce_sum(axes = sum_2_axes_0, keep_dims = sum_2_keep_dims_0, x = expand_74_to_fp16)[name = string("sum_2_cast_fp16")];
            fp16 const_3293_to_fp16 = const()[name = string("const_3293_to_fp16"), val = fp16(0x1p-24)];
            fp16 const_3294_to_fp16 = const()[name = string("const_3294_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 768]> clip_0_cast_fp16 = clip(alpha = const_3293_to_fp16, beta = const_3294_to_fp16, x = sum_2_cast_fp16)[name = string("clip_0_cast_fp16")];
            tensor<fp16, [1, 768]> div_cast_fp16 = real_div(x = sum_1_cast_fp16, y = clip_0_cast_fp16)[name = string("div_cast_fp16")];
            int32 const_3295 = const()[name = string("const_3295"), val = int32(-1)];
            bool cat_120_interleave_0 = const()[name = string("cat_120_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 768]> cat_120_cast_fp16 = concat(axis = const_3295, interleave = cat_120_interleave_0, values = div_cast_fp16)[name = string("cat_120_cast_fp16")];
            tensor<fp16, [3072, 768]> p_st_2_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304280000))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306639360))))[name = string("p_st_2_linear_weight_to_fp16_quantized")];
            tensor<fp16, [3072]> linear_168_bias_0_to_fp16 = const()[name = string("linear_168_bias_0_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306645568)))];
            tensor<fp16, [1, 3072]> linear_168_cast_fp16 = linear(bias = linear_168_bias_0_to_fp16, weight = p_st_2_linear_weight_to_fp16_quantized, x = cat_120_cast_fp16)[name = string("linear_168_cast_fp16")];
            tensor<fp16, [768, 3072]> p_st_3_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306651776))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309011136))))[name = string("p_st_3_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 768]> linear_169_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = p_st_3_linear_weight_to_fp16_quantized, x = linear_168_cast_fp16)[name = string("linear_169_cast_fp16")];
            tensor<int32, [1]> const_3297 = const()[name = string("const_3297"), val = tensor<int32, [1]>([1])];
            bool const_3298 = const()[name = string("const_3298"), val = bool(true)];
            tensor<fp16, [1, 1]> linalg_vector_norm_cast_fp16 = reduce_l2_norm(axes = const_3297, keep_dims = const_3298, x = linear_169_cast_fp16)[name = string("linalg_vector_norm_cast_fp16")];
            fp16 const_3299_to_fp16 = const()[name = string("const_3299_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 1]> clamp_min_cast_fp16 = maximum(x = linalg_vector_norm_cast_fp16, y = const_3299_to_fp16)[name = string("clamp_min_cast_fp16")];
            tensor<int32, [2]> expand_75_reps_0 = const()[name = string("expand_75_reps_0"), val = tensor<int32, [2]>([1, 768])];
            tensor<fp16, [1, 768]> expand_75_cast_fp16 = tile(reps = expand_75_reps_0, x = clamp_min_cast_fp16)[name = string("expand_75_cast_fp16")];
            tensor<fp16, [1, 768]> embedding = real_div(x = linear_169_cast_fp16, y = expand_75_cast_fp16)[name = string("div_1_cast_fp16")];
        } -> (embedding);
}