program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
{
    func main<ios18>(tensor<fp16, [1, 1, ?, 128]> attention_mask, tensor<fp16, [1, ?, 256]> cos_full, tensor<fp16, [1, ?, 256]> cos_sliding, tensor<int32, [1, ?]> input_ids, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_0, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_1, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_10, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_11, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_12, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_13, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_14, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_15, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_16, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_17, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_2, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_3, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_4, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_5, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_6, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_7, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_8, state<tensor<fp16, [1, 256, 1, 128]>> k_cache_9, tensor<fp16, [1, 1, 1, ?]> logits_mask, tensor<fp16, [1, ?, 256]> sin_full, tensor<fp16, [1, ?, 256]> sin_sliding, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_0, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_1, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_10, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_11, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_12, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_13, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_14, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_15, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_16, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_17, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_2, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_3, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_4, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_5, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_6, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_7, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_8, state<tensor<fp16, [1, 256, 1, 128]>> v_cache_9, tensor<fp16, [1, 1, 1, 128]> write_mask) [FlexibleShapeInformation = tuple<tuple<string, dict<string, tensor<int32, [?]>>>, tuple<string, dict<string, dict<string, tensor<int32, [?]>>>>>((("DefaultShapes", {{"attention_mask", [1, 1, 128, 128]}, {"cos_full", [1, 128, 256]}, {"cos_sliding", [1, 128, 256]}, {"input_ids", [1, 128]}, {"logits_mask", [1, 1, 1, 128]}, {"sin_full", [1, 128, 256]}, {"sin_sliding", [1, 128, 256]}}), ("EnumeratedShapes", {{"029ac3ee", {{"attention_mask", [1, 1, 128, 128]}, {"cos_full", [1, 128, 256]}, {"cos_sliding", [1, 128, 256]}, {"input_ids", [1, 128]}, {"k_cache_0", [1, 256, 1, 128]}, {"k_cache_1", [1, 256, 1, 128]}, {"k_cache_10", [1, 256, 1, 128]}, {"k_cache_11", [1, 256, 1, 128]}, {"k_cache_12", [1, 256, 1, 128]}, {"k_cache_13", [1, 256, 1, 128]}, {"k_cache_14", [1, 256, 1, 128]}, {"k_cache_15", [1, 256, 1, 128]}, {"k_cache_16", [1, 256, 1, 128]}, {"k_cache_17", [1, 256, 1, 128]}, {"k_cache_2", [1, 256, 1, 128]}, {"k_cache_3", [1, 256, 1, 128]}, {"k_cache_4", [1, 256, 1, 128]}, {"k_cache_5", [1, 256, 1, 128]}, {"k_cache_6", [1, 256, 1, 128]}, {"k_cache_7", [1, 256, 1, 128]}, {"k_cache_8", [1, 256, 1, 128]}, {"k_cache_9", [1, 256, 1, 128]}, {"logits_mask", [1, 1, 1, 128]}, {"sin_full", [1, 128, 256]}, {"sin_sliding", [1, 128, 256]}, {"v_cache_0", [1, 256, 1, 128]}, {"v_cache_1", [1, 256, 1, 128]}, {"v_cache_10", [1, 256, 1, 128]}, {"v_cache_11", [1, 256, 1, 128]}, {"v_cache_12", [1, 256, 1, 128]}, {"v_cache_13", [1, 256, 1, 128]}, {"v_cache_14", [1, 256, 1, 128]}, {"v_cache_15", [1, 256, 1, 128]}, {"v_cache_16", [1, 256, 1, 128]}, {"v_cache_17", [1, 256, 1, 128]}, {"v_cache_2", [1, 256, 1, 128]}, {"v_cache_3", [1, 256, 1, 128]}, {"v_cache_4", [1, 256, 1, 128]}, {"v_cache_5", [1, 256, 1, 128]}, {"v_cache_6", [1, 256, 1, 128]}, {"v_cache_7", [1, 256, 1, 128]}, {"v_cache_8", [1, 256, 1, 128]}, {"v_cache_9", [1, 256, 1, 128]}, {"write_mask", [1, 1, 1, 128]}}}, {"222597ec", {{"attention_mask", [1, 1, 1, 128]}, {"cos_full", [1, 1, 256]}, {"cos_sliding", [1, 1, 256]}, {"input_ids", [1, 1]}, {"k_cache_0", [1, 256, 1, 128]}, {"k_cache_1", [1, 256, 1, 128]}, {"k_cache_10", [1, 256, 1, 128]}, {"k_cache_11", [1, 256, 1, 128]}, {"k_cache_12", [1, 256, 1, 128]}, {"k_cache_13", [1, 256, 1, 128]}, {"k_cache_14", [1, 256, 1, 128]}, {"k_cache_15", [1, 256, 1, 128]}, {"k_cache_16", [1, 256, 1, 128]}, {"k_cache_17", [1, 256, 1, 128]}, {"k_cache_2", [1, 256, 1, 128]}, {"k_cache_3", [1, 256, 1, 128]}, {"k_cache_4", [1, 256, 1, 128]}, {"k_cache_5", [1, 256, 1, 128]}, {"k_cache_6", [1, 256, 1, 128]}, {"k_cache_7", [1, 256, 1, 128]}, {"k_cache_8", [1, 256, 1, 128]}, {"k_cache_9", [1, 256, 1, 128]}, {"logits_mask", [1, 1, 1, 1]}, {"sin_full", [1, 1, 256]}, {"sin_sliding", [1, 1, 256]}, {"v_cache_0", [1, 256, 1, 128]}, {"v_cache_1", [1, 256, 1, 128]}, {"v_cache_10", [1, 256, 1, 128]}, {"v_cache_11", [1, 256, 1, 128]}, {"v_cache_12", [1, 256, 1, 128]}, {"v_cache_13", [1, 256, 1, 128]}, {"v_cache_14", [1, 256, 1, 128]}, {"v_cache_15", [1, 256, 1, 128]}, {"v_cache_16", [1, 256, 1, 128]}, {"v_cache_17", [1, 256, 1, 128]}, {"v_cache_2", [1, 256, 1, 128]}, {"v_cache_3", [1, 256, 1, 128]}, {"v_cache_4", [1, 256, 1, 128]}, {"v_cache_5", [1, 256, 1, 128]}, {"v_cache_6", [1, 256, 1, 128]}, {"v_cache_7", [1, 256, 1, 128]}, {"v_cache_8", [1, 256, 1, 128]}, {"v_cache_9", [1, 256, 1, 128]}, {"write_mask", [1, 1, 1, 128]}}}})))] {
            int32 var_289_batch_dims_0 = const()[name = string("op_289_batch_dims_0"), val = int32(0)];
            bool var_289_validate_indices_0 = const()[name = string("op_289_validate_indices_0"), val = bool(false)];
            tensor<fp16, [262144, 640]> embed_weight_to_fp16 = const()[name = string("embed_weight_to_fp16"), val = tensor<fp16, [262144, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [1, ?]> 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, ?]> add_0 = add(x = input_ids, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, ?]> select_0 = select(a = input_ids, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 var_289_cast_fp16_axis_0 = const()[name = string("op_289_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, ?, 640]> var_289_cast_fp16 = gather(axis = var_289_cast_fp16_axis_0, batch_dims = var_289_batch_dims_0, indices = select_0, validate_indices = var_289_validate_indices_0, x = embed_weight_to_fp16)[name = string("op_289_cast_fp16")];
            fp16 const_0_to_fp16 = const()[name = string("const_0_to_fp16"), val = fp16(0x1.94cp+4)];
            tensor<fp16, [1, ?, 640]> h_btC_cast_fp16 = mul(x = var_289_cast_fp16, y = const_0_to_fp16)[name = string("h_btC_cast_fp16")];
            tensor<int32, [3]> var_295 = const()[name = string("op_295"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> x_1_axes_0 = const()[name = string("x_1_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 640, ?]> var_296_cast_fp16 = transpose(perm = var_295, x = h_btC_cast_fp16)[name = string("transpose_580")];
            tensor<fp16, [1, 640, 1, ?]> x_1_cast_fp16 = expand_dims(axes = x_1_axes_0, x = var_296_cast_fp16)[name = string("x_1_cast_fp16")];
            fp16 var_299_promoted_to_fp16 = const()[name = string("op_299_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_300_cast_fp16 = pow(x = x_1_cast_fp16, y = var_299_promoted_to_fp16)[name = string("op_300_cast_fp16")];
            tensor<int32, [1]> var_1_axes_0 = const()[name = string("var_1_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_1_keep_dims_0 = const()[name = string("var_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_1_cast_fp16 = reduce_mean(axes = var_1_axes_0, keep_dims = var_1_keep_dims_0, x = var_300_cast_fp16)[name = string("var_1_cast_fp16")];
            fp16 var_307_to_fp16 = const()[name = string("op_307_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_308_cast_fp16 = add(x = var_1_cast_fp16, y = var_307_to_fp16)[name = string("op_308_cast_fp16")];
            fp32 var_309_epsilon_0 = const()[name = string("op_309_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_309_cast_fp16 = rsqrt(epsilon = var_309_epsilon_0, x = var_308_cast_fp16)[name = string("op_309_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_1_cast_fp16 = mul(x = x_1_cast_fp16, y = var_309_cast_fp16)[name = string("x_normed_1_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_319_to_fp16 = const()[name = string("op_319_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335544448)))];
            tensor<fp16, [1, 640, 1, ?]> input_1_cast_fp16 = mul(x = x_normed_1_cast_fp16, y = var_319_to_fp16)[name = string("input_1_cast_fp16")];
            string var_331_pad_type_0 = const()[name = string("op_331_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_331_strides_0 = const()[name = string("op_331_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_331_pad_0 = const()[name = string("op_331_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_331_dilations_0 = const()[name = string("op_331_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_331_groups_0 = const()[name = string("op_331_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_0_weight_to_fp16 = const()[name = string("q_proj_0_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335545792)))];
            tensor<fp16, [1, 1024, 1, ?]> var_331_cast_fp16 = conv(dilations = var_331_dilations_0, groups = var_331_groups_0, pad = var_331_pad_0, pad_type = var_331_pad_type_0, strides = var_331_strides_0, weight = q_proj_0_weight_to_fp16, x = input_1_cast_fp16)[name = string("op_331_cast_fp16")];
            string var_342_pad_type_0 = const()[name = string("op_342_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_342_strides_0 = const()[name = string("op_342_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_342_pad_0 = const()[name = string("op_342_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_342_dilations_0 = const()[name = string("op_342_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_342_groups_0 = const()[name = string("op_342_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_0_weight_to_fp16 = const()[name = string("k_proj_0_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336856576)))];
            tensor<fp16, [1, 256, 1, ?]> var_342_cast_fp16 = conv(dilations = var_342_dilations_0, groups = var_342_groups_0, pad = var_342_pad_0, pad_type = var_342_pad_type_0, strides = var_342_strides_0, weight = k_proj_0_weight_to_fp16, x = input_1_cast_fp16)[name = string("op_342_cast_fp16")];
            string var_353_pad_type_0 = const()[name = string("op_353_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_353_strides_0 = const()[name = string("op_353_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_353_pad_0 = const()[name = string("op_353_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_353_dilations_0 = const()[name = string("op_353_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_353_groups_0 = const()[name = string("op_353_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_0_weight_to_fp16 = const()[name = string("v_proj_0_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337184320)))];
            tensor<fp16, [1, 256, 1, ?]> var_353_cast_fp16 = conv(dilations = var_353_dilations_0, groups = var_353_groups_0, pad = var_353_pad_0, pad_type = var_353_pad_type_0, strides = var_353_strides_0, weight = v_proj_0_weight_to_fp16, x = input_1_cast_fp16)[name = string("op_353_cast_fp16")];
            tensor<int32, [4]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_356_axis_0 = const()[name = string("op_356_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_356_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_356_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_356_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_356_cast_fp16_3 = split(axis = var_356_axis_0, split_sizes = tile_0, x = var_331_cast_fp16)[name = string("op_356_cast_fp16")];
            fp16 var_369_promoted_to_fp16 = const()[name = string("op_369_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_370_cast_fp16 = pow(x = var_356_cast_fp16_0, y = var_369_promoted_to_fp16)[name = string("op_370_cast_fp16")];
            tensor<int32, [1]> var_3_axes_0 = const()[name = string("var_3_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_3_keep_dims_0 = const()[name = string("var_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_3_cast_fp16 = reduce_mean(axes = var_3_axes_0, keep_dims = var_3_keep_dims_0, x = var_370_cast_fp16)[name = string("var_3_cast_fp16")];
            fp16 var_377_to_fp16 = const()[name = string("op_377_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_378_cast_fp16 = add(x = var_3_cast_fp16, y = var_377_to_fp16)[name = string("op_378_cast_fp16")];
            fp32 var_379_epsilon_0 = const()[name = string("op_379_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_379_cast_fp16 = rsqrt(epsilon = var_379_epsilon_0, x = var_378_cast_fp16)[name = string("op_379_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_3_cast_fp16 = mul(x = var_356_cast_fp16_0, y = var_379_cast_fp16)[name = string("x_normed_3_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_389_to_fp16 = const()[name = string("op_389_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337512064)))];
            tensor<fp16, [1, 256, 1, ?]> x_5_cast_fp16 = mul(x = x_normed_3_cast_fp16, y = var_389_to_fp16)[name = string("x_5_cast_fp16")];
            tensor<int32, [4]> x_first_1_begin_0 = const()[name = string("x_first_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_1_end_0 = const()[name = string("x_first_1_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_1_end_mask_0 = const()[name = string("x_first_1_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_1_cast_fp16 = slice_by_index(begin = x_first_1_begin_0, end = x_first_1_end_0, end_mask = x_first_1_end_mask_0, x = x_5_cast_fp16)[name = string("x_first_1_cast_fp16")];
            tensor<int32, [4]> x_second_1_begin_0 = const()[name = string("x_second_1_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_1_end_0 = const()[name = string("x_second_1_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_1_end_mask_0 = const()[name = string("x_second_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_1_cast_fp16 = slice_by_index(begin = x_second_1_begin_0, end = x_second_1_end_0, end_mask = x_second_1_end_mask_0, x = x_5_cast_fp16)[name = string("x_second_1_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_411_cast_fp16 = mul(x = x_second_1_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_411_cast_fp16")];
            int32 var_413 = const()[name = string("op_413"), val = int32(1)];
            bool x_rot_1_interleave_0 = const()[name = string("x_rot_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_1_cast_fp16 = concat(axis = var_413, interleave = x_rot_1_interleave_0, values = (var_411_cast_fp16, x_first_1_cast_fp16))[name = string("x_rot_1_cast_fp16")];
            tensor<int32, [3]> var_418 = const()[name = string("op_418"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> cos_bc1t_1_axes_0 = const()[name = string("cos_bc1t_1_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 256, ?]> var_419_cast_fp16 = transpose(perm = var_418, x = cos_sliding)[name = string("transpose_579")];
            tensor<fp16, [1, 256, 1, ?]> cos_bc1t_1_cast_fp16 = expand_dims(axes = cos_bc1t_1_axes_0, x = var_419_cast_fp16)[name = string("cos_bc1t_1_cast_fp16")];
            tensor<int32, [3]> var_425 = const()[name = string("op_425"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> sin_bc1t_1_axes_0 = const()[name = string("sin_bc1t_1_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 256, ?]> var_426_cast_fp16 = transpose(perm = var_425, x = sin_sliding)[name = string("transpose_578")];
            tensor<fp16, [1, 256, 1, ?]> sin_bc1t_1_cast_fp16 = expand_dims(axes = sin_bc1t_1_axes_0, x = var_426_cast_fp16)[name = string("sin_bc1t_1_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_429_cast_fp16 = mul(x = x_5_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_429_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_430_cast_fp16 = mul(x = x_rot_1_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_430_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_1_cast_fp16 = add(x = var_429_cast_fp16, y = var_430_cast_fp16)[name = string("q_h_1_cast_fp16")];
            fp16 var_433_promoted_to_fp16 = const()[name = string("op_433_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_434_cast_fp16 = pow(x = var_356_cast_fp16_1, y = var_433_promoted_to_fp16)[name = string("op_434_cast_fp16")];
            tensor<int32, [1]> var_5_axes_0 = const()[name = string("var_5_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_5_keep_dims_0 = const()[name = string("var_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_5_cast_fp16 = reduce_mean(axes = var_5_axes_0, keep_dims = var_5_keep_dims_0, x = var_434_cast_fp16)[name = string("var_5_cast_fp16")];
            fp16 var_441_to_fp16 = const()[name = string("op_441_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_442_cast_fp16 = add(x = var_5_cast_fp16, y = var_441_to_fp16)[name = string("op_442_cast_fp16")];
            fp32 var_443_epsilon_0 = const()[name = string("op_443_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_443_cast_fp16 = rsqrt(epsilon = var_443_epsilon_0, x = var_442_cast_fp16)[name = string("op_443_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_5_cast_fp16 = mul(x = var_356_cast_fp16_1, y = var_443_cast_fp16)[name = string("x_normed_5_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_9_cast_fp16 = mul(x = x_normed_5_cast_fp16, y = var_389_to_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [4]> x_first_3_begin_0 = const()[name = string("x_first_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_3_end_0 = const()[name = string("x_first_3_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_3_end_mask_0 = const()[name = string("x_first_3_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_3_cast_fp16 = slice_by_index(begin = x_first_3_begin_0, end = x_first_3_end_0, end_mask = x_first_3_end_mask_0, x = x_9_cast_fp16)[name = string("x_first_3_cast_fp16")];
            tensor<int32, [4]> x_second_3_begin_0 = const()[name = string("x_second_3_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_3_end_0 = const()[name = string("x_second_3_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_3_end_mask_0 = const()[name = string("x_second_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_3_cast_fp16 = slice_by_index(begin = x_second_3_begin_0, end = x_second_3_end_0, end_mask = x_second_3_end_mask_0, x = x_9_cast_fp16)[name = string("x_second_3_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_475_cast_fp16 = mul(x = x_second_3_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_475_cast_fp16")];
            int32 var_477 = const()[name = string("op_477"), val = int32(1)];
            bool x_rot_3_interleave_0 = const()[name = string("x_rot_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_3_cast_fp16 = concat(axis = var_477, interleave = x_rot_3_interleave_0, values = (var_475_cast_fp16, x_first_3_cast_fp16))[name = string("x_rot_3_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_493_cast_fp16 = mul(x = x_9_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_493_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_494_cast_fp16 = mul(x = x_rot_3_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_494_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_3_cast_fp16 = add(x = var_493_cast_fp16, y = var_494_cast_fp16)[name = string("q_h_3_cast_fp16")];
            fp16 var_497_promoted_to_fp16 = const()[name = string("op_497_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_498_cast_fp16 = pow(x = var_356_cast_fp16_2, y = var_497_promoted_to_fp16)[name = string("op_498_cast_fp16")];
            tensor<int32, [1]> var_7_axes_0 = const()[name = string("var_7_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_7_keep_dims_0 = const()[name = string("var_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_7_cast_fp16 = reduce_mean(axes = var_7_axes_0, keep_dims = var_7_keep_dims_0, x = var_498_cast_fp16)[name = string("var_7_cast_fp16")];
            fp16 var_505_to_fp16 = const()[name = string("op_505_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_506_cast_fp16 = add(x = var_7_cast_fp16, y = var_505_to_fp16)[name = string("op_506_cast_fp16")];
            fp32 var_507_epsilon_0 = const()[name = string("op_507_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_507_cast_fp16 = rsqrt(epsilon = var_507_epsilon_0, x = var_506_cast_fp16)[name = string("op_507_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_7_cast_fp16 = mul(x = var_356_cast_fp16_2, y = var_507_cast_fp16)[name = string("x_normed_7_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_13_cast_fp16 = mul(x = x_normed_7_cast_fp16, y = var_389_to_fp16)[name = string("x_13_cast_fp16")];
            tensor<int32, [4]> x_first_5_begin_0 = const()[name = string("x_first_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_5_end_0 = const()[name = string("x_first_5_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_5_end_mask_0 = const()[name = string("x_first_5_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_5_cast_fp16 = slice_by_index(begin = x_first_5_begin_0, end = x_first_5_end_0, end_mask = x_first_5_end_mask_0, x = x_13_cast_fp16)[name = string("x_first_5_cast_fp16")];
            tensor<int32, [4]> x_second_5_begin_0 = const()[name = string("x_second_5_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_5_end_0 = const()[name = string("x_second_5_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_5_end_mask_0 = const()[name = string("x_second_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_5_cast_fp16 = slice_by_index(begin = x_second_5_begin_0, end = x_second_5_end_0, end_mask = x_second_5_end_mask_0, x = x_13_cast_fp16)[name = string("x_second_5_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_539_cast_fp16 = mul(x = x_second_5_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_539_cast_fp16")];
            int32 var_541 = const()[name = string("op_541"), val = int32(1)];
            bool x_rot_5_interleave_0 = const()[name = string("x_rot_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_5_cast_fp16 = concat(axis = var_541, interleave = x_rot_5_interleave_0, values = (var_539_cast_fp16, x_first_5_cast_fp16))[name = string("x_rot_5_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_557_cast_fp16 = mul(x = x_13_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_557_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_558_cast_fp16 = mul(x = x_rot_5_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_558_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_5_cast_fp16 = add(x = var_557_cast_fp16, y = var_558_cast_fp16)[name = string("q_h_5_cast_fp16")];
            fp16 var_561_promoted_to_fp16 = const()[name = string("op_561_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_562_cast_fp16 = pow(x = var_356_cast_fp16_3, y = var_561_promoted_to_fp16)[name = string("op_562_cast_fp16")];
            tensor<int32, [1]> var_9_axes_0 = const()[name = string("var_9_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_9_keep_dims_0 = const()[name = string("var_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_9_cast_fp16 = reduce_mean(axes = var_9_axes_0, keep_dims = var_9_keep_dims_0, x = var_562_cast_fp16)[name = string("var_9_cast_fp16")];
            fp16 var_569_to_fp16 = const()[name = string("op_569_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_570_cast_fp16 = add(x = var_9_cast_fp16, y = var_569_to_fp16)[name = string("op_570_cast_fp16")];
            fp32 var_571_epsilon_0 = const()[name = string("op_571_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_571_cast_fp16 = rsqrt(epsilon = var_571_epsilon_0, x = var_570_cast_fp16)[name = string("op_571_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_9_cast_fp16 = mul(x = var_356_cast_fp16_3, y = var_571_cast_fp16)[name = string("x_normed_9_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_17_cast_fp16 = mul(x = x_normed_9_cast_fp16, y = var_389_to_fp16)[name = string("x_17_cast_fp16")];
            tensor<int32, [4]> x_first_7_begin_0 = const()[name = string("x_first_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_7_end_0 = const()[name = string("x_first_7_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_7_end_mask_0 = const()[name = string("x_first_7_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_7_cast_fp16 = slice_by_index(begin = x_first_7_begin_0, end = x_first_7_end_0, end_mask = x_first_7_end_mask_0, x = x_17_cast_fp16)[name = string("x_first_7_cast_fp16")];
            tensor<int32, [4]> x_second_7_begin_0 = const()[name = string("x_second_7_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_7_end_0 = const()[name = string("x_second_7_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_7_end_mask_0 = const()[name = string("x_second_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_7_cast_fp16 = slice_by_index(begin = x_second_7_begin_0, end = x_second_7_end_0, end_mask = x_second_7_end_mask_0, x = x_17_cast_fp16)[name = string("x_second_7_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_603_cast_fp16 = mul(x = x_second_7_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_603_cast_fp16")];
            int32 var_605 = const()[name = string("op_605"), val = int32(1)];
            bool x_rot_7_interleave_0 = const()[name = string("x_rot_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_7_cast_fp16 = concat(axis = var_605, interleave = x_rot_7_interleave_0, values = (var_603_cast_fp16, x_first_7_cast_fp16))[name = string("x_rot_7_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_621_cast_fp16 = mul(x = x_17_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_621_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_622_cast_fp16 = mul(x = x_rot_7_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_622_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_7_cast_fp16 = add(x = var_621_cast_fp16, y = var_622_cast_fp16)[name = string("q_h_7_cast_fp16")];
            fp16 var_625_promoted_to_fp16 = const()[name = string("op_625_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_626_cast_fp16 = pow(x = var_342_cast_fp16, y = var_625_promoted_to_fp16)[name = string("op_626_cast_fp16")];
            tensor<int32, [1]> var_11_axes_0 = const()[name = string("var_11_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_11_keep_dims_0 = const()[name = string("var_11_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_11_cast_fp16 = reduce_mean(axes = var_11_axes_0, keep_dims = var_11_keep_dims_0, x = var_626_cast_fp16)[name = string("var_11_cast_fp16")];
            fp16 var_633_to_fp16 = const()[name = string("op_633_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_634_cast_fp16 = add(x = var_11_cast_fp16, y = var_633_to_fp16)[name = string("op_634_cast_fp16")];
            fp32 var_635_epsilon_0 = const()[name = string("op_635_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_635_cast_fp16 = rsqrt(epsilon = var_635_epsilon_0, x = var_634_cast_fp16)[name = string("op_635_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_11_cast_fp16 = mul(x = var_342_cast_fp16, y = var_635_cast_fp16)[name = string("x_normed_11_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_645_to_fp16 = const()[name = string("op_645_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337512640)))];
            tensor<fp16, [1, 256, 1, ?]> x_21_cast_fp16 = mul(x = x_normed_11_cast_fp16, y = var_645_to_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> x_first_9_begin_0 = const()[name = string("x_first_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_9_end_0 = const()[name = string("x_first_9_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_9_end_mask_0 = const()[name = string("x_first_9_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_9_cast_fp16 = slice_by_index(begin = x_first_9_begin_0, end = x_first_9_end_0, end_mask = x_first_9_end_mask_0, x = x_21_cast_fp16)[name = string("x_first_9_cast_fp16")];
            tensor<int32, [4]> x_second_9_begin_0 = const()[name = string("x_second_9_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_9_end_0 = const()[name = string("x_second_9_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_9_end_mask_0 = const()[name = string("x_second_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_9_cast_fp16 = slice_by_index(begin = x_second_9_begin_0, end = x_second_9_end_0, end_mask = x_second_9_end_mask_0, x = x_21_cast_fp16)[name = string("x_second_9_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_667_cast_fp16 = mul(x = x_second_9_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_667_cast_fp16")];
            int32 var_669 = const()[name = string("op_669"), val = int32(1)];
            bool x_rot_9_interleave_0 = const()[name = string("x_rot_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_9_cast_fp16 = concat(axis = var_669, interleave = x_rot_9_interleave_0, values = (var_667_cast_fp16, x_first_9_cast_fp16))[name = string("x_rot_9_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_685_cast_fp16 = mul(x = x_21_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_685_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_686_cast_fp16 = mul(x = x_rot_9_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_686_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_1_cast_fp16 = add(x = var_685_cast_fp16, y = var_686_cast_fp16)[name = string("k_new_1_cast_fp16")];
            fp16 var_689_to_fp16 = const()[name = string("op_689_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 1, 1, 128]> one_minus_1_cast_fp16 = sub(x = var_689_to_fp16, y = write_mask)[name = string("one_minus_1_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_0 = read_state(input = k_cache_0)[name = string("read_state_0")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_3_cast_fp16 = mul(x = read_state_0, y = one_minus_1_cast_fp16)[name = string("k_cache_3_cast_fp16")];
            write_state(data = k_cache_3_cast_fp16, input = k_cache_0)[name = string("coreml_update_state_72_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_72 = read_state(input = k_cache_0)[name = string("coreml_update_state_72")];
            tensor<fp16, [1, 256, 1, 128]> var_693_cast_fp16 = mul(x = k_new_1_cast_fp16, y = write_mask)[name = string("op_693_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_5_cast_fp16 = add(x = coreml_update_state_72, y = var_693_cast_fp16)[name = string("k_cache_5_cast_fp16")];
            write_state(data = k_cache_5_cast_fp16, input = k_cache_0)[name = string("coreml_update_state_73_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_73 = read_state(input = k_cache_0)[name = string("coreml_update_state_73")];
            tensor<fp16, [1, 256, 1, 128]> read_state_1 = read_state(input = v_cache_0)[name = string("read_state_1")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_3_cast_fp16 = mul(x = read_state_1, y = one_minus_1_cast_fp16)[name = string("v_cache_3_cast_fp16")];
            write_state(data = v_cache_3_cast_fp16, input = v_cache_0)[name = string("coreml_update_state_74_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_74 = read_state(input = v_cache_0)[name = string("coreml_update_state_74")];
            tensor<fp16, [1, 256, 1, 128]> var_697_cast_fp16 = mul(x = var_353_cast_fp16, y = write_mask)[name = string("op_697_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_5_cast_fp16 = add(x = coreml_update_state_74, y = var_697_cast_fp16)[name = string("v_cache_5_cast_fp16")];
            write_state(data = v_cache_5_cast_fp16, input = v_cache_0)[name = string("coreml_update_state_75_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_75 = read_state(input = v_cache_0)[name = string("coreml_update_state_75")];
            bool var_712_transpose_x_0 = const()[name = string("op_712_transpose_x_0"), val = bool(false)];
            bool var_712_transpose_y_0 = const()[name = string("op_712_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_306_perm_0 = const()[name = string("transpose_306_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_307_perm_0 = const()[name = string("transpose_307_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_307 = transpose(perm = transpose_307_perm_0, x = coreml_update_state_73)[name = string("transpose_576")];
            tensor<fp16, [1, 1, ?, 256]> transpose_306 = transpose(perm = transpose_306_perm_0, x = q_h_1_cast_fp16)[name = string("transpose_577")];
            tensor<fp16, [1, 1, ?, 128]> var_712_cast_fp16 = matmul(transpose_x = var_712_transpose_x_0, transpose_y = var_712_transpose_y_0, x = transpose_306, y = transpose_307)[name = string("op_712_cast_fp16")];
            fp16 var_713_to_fp16 = const()[name = string("op_713_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_714_cast_fp16 = mul(x = var_712_cast_fp16, y = var_713_to_fp16)[name = string("op_714_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_1_cast_fp16 = add(x = var_714_cast_fp16, y = attention_mask)[name = string("scores_1_cast_fp16")];
            int32 var_717 = const()[name = string("op_717"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_1_cast_fp16 = softmax(axis = var_717, x = scores_1_cast_fp16)[name = string("probs_1_cast_fp16")];
            tensor<int32, [4]> var_724 = const()[name = string("op_724"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_726_transpose_x_0 = const()[name = string("op_726_transpose_x_0"), val = bool(false)];
            bool var_726_transpose_y_0 = const()[name = string("op_726_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_1_cast_fp16 = transpose(perm = var_724, x = coreml_update_state_75)[name = string("transpose_575")];
            tensor<fp16, [1, 1, ?, 256]> var_726_cast_fp16 = matmul(transpose_x = var_726_transpose_x_0, transpose_y = var_726_transpose_y_0, x = probs_1_cast_fp16, y = v_t_1_cast_fp16)[name = string("op_726_cast_fp16")];
            bool var_745_transpose_x_0 = const()[name = string("op_745_transpose_x_0"), val = bool(false)];
            bool var_745_transpose_y_0 = const()[name = string("op_745_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_311_perm_0 = const()[name = string("transpose_311_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_311 = transpose(perm = transpose_311_perm_0, x = q_h_3_cast_fp16)[name = string("transpose_574")];
            tensor<fp16, [1, 1, ?, 128]> var_745_cast_fp16 = matmul(transpose_x = var_745_transpose_x_0, transpose_y = var_745_transpose_y_0, x = transpose_311, y = transpose_307)[name = string("op_745_cast_fp16")];
            fp16 var_746_to_fp16 = const()[name = string("op_746_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_747_cast_fp16 = mul(x = var_745_cast_fp16, y = var_746_to_fp16)[name = string("op_747_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_3_cast_fp16 = add(x = var_747_cast_fp16, y = attention_mask)[name = string("scores_3_cast_fp16")];
            int32 var_750 = const()[name = string("op_750"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_3_cast_fp16 = softmax(axis = var_750, x = scores_3_cast_fp16)[name = string("probs_3_cast_fp16")];
            bool var_759_transpose_x_0 = const()[name = string("op_759_transpose_x_0"), val = bool(false)];
            bool var_759_transpose_y_0 = const()[name = string("op_759_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_759_cast_fp16 = matmul(transpose_x = var_759_transpose_x_0, transpose_y = var_759_transpose_y_0, x = probs_3_cast_fp16, y = v_t_1_cast_fp16)[name = string("op_759_cast_fp16")];
            bool var_778_transpose_x_0 = const()[name = string("op_778_transpose_x_0"), val = bool(false)];
            bool var_778_transpose_y_0 = const()[name = string("op_778_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_312_perm_0 = const()[name = string("transpose_312_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_312 = transpose(perm = transpose_312_perm_0, x = q_h_5_cast_fp16)[name = string("transpose_573")];
            tensor<fp16, [1, 1, ?, 128]> var_778_cast_fp16 = matmul(transpose_x = var_778_transpose_x_0, transpose_y = var_778_transpose_y_0, x = transpose_312, y = transpose_307)[name = string("op_778_cast_fp16")];
            fp16 var_779_to_fp16 = const()[name = string("op_779_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_780_cast_fp16 = mul(x = var_778_cast_fp16, y = var_779_to_fp16)[name = string("op_780_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_5_cast_fp16 = add(x = var_780_cast_fp16, y = attention_mask)[name = string("scores_5_cast_fp16")];
            int32 var_783 = const()[name = string("op_783"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_5_cast_fp16 = softmax(axis = var_783, x = scores_5_cast_fp16)[name = string("probs_5_cast_fp16")];
            bool var_792_transpose_x_0 = const()[name = string("op_792_transpose_x_0"), val = bool(false)];
            bool var_792_transpose_y_0 = const()[name = string("op_792_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_792_cast_fp16 = matmul(transpose_x = var_792_transpose_x_0, transpose_y = var_792_transpose_y_0, x = probs_5_cast_fp16, y = v_t_1_cast_fp16)[name = string("op_792_cast_fp16")];
            bool var_811_transpose_x_0 = const()[name = string("op_811_transpose_x_0"), val = bool(false)];
            bool var_811_transpose_y_0 = const()[name = string("op_811_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_313_perm_0 = const()[name = string("transpose_313_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_313 = transpose(perm = transpose_313_perm_0, x = q_h_7_cast_fp16)[name = string("transpose_572")];
            tensor<fp16, [1, 1, ?, 128]> var_811_cast_fp16 = matmul(transpose_x = var_811_transpose_x_0, transpose_y = var_811_transpose_y_0, x = transpose_313, y = transpose_307)[name = string("op_811_cast_fp16")];
            fp16 var_812_to_fp16 = const()[name = string("op_812_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_813_cast_fp16 = mul(x = var_811_cast_fp16, y = var_812_to_fp16)[name = string("op_813_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_7_cast_fp16 = add(x = var_813_cast_fp16, y = attention_mask)[name = string("scores_7_cast_fp16")];
            int32 var_816 = const()[name = string("op_816"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_7_cast_fp16 = softmax(axis = var_816, x = scores_7_cast_fp16)[name = string("probs_7_cast_fp16")];
            bool var_825_transpose_x_0 = const()[name = string("op_825_transpose_x_0"), val = bool(false)];
            bool var_825_transpose_y_0 = const()[name = string("op_825_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_825_cast_fp16 = matmul(transpose_x = var_825_transpose_x_0, transpose_y = var_825_transpose_y_0, x = probs_7_cast_fp16, y = v_t_1_cast_fp16)[name = string("op_825_cast_fp16")];
            bool input_3_interleave_0 = const()[name = string("input_3_interleave_0"), val = bool(false)];
            int32 const_145 = const()[name = string("const_145"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_3_cast_fp16 = concat(axis = const_145, interleave = input_3_interleave_0, values = (var_726_cast_fp16, var_759_cast_fp16, var_792_cast_fp16, var_825_cast_fp16))[name = string("input_3_cast_fp16")];
            string x_23_pad_type_0 = const()[name = string("x_23_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_23_strides_0 = const()[name = string("x_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_23_pad_0 = const()[name = string("x_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_23_dilations_0 = const()[name = string("x_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_23_groups_0 = const()[name = string("x_23_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_8_perm_0 = const()[name = string("transpose_8_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_0_weight_to_fp16 = const()[name = string("o_proj_0_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337513216)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = input_3_cast_fp16)[name = string("transpose_571")];
            tensor<fp16, [1, 640, 1, ?]> x_23_cast_fp16 = conv(dilations = x_23_dilations_0, groups = x_23_groups_0, pad = x_23_pad_0, pad_type = x_23_pad_type_0, strides = x_23_strides_0, weight = o_proj_0_weight_to_fp16, x = transpose_8_cast_fp16)[name = string("x_23_cast_fp16")];
            fp16 var_846_promoted_to_fp16 = const()[name = string("op_846_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_847_cast_fp16 = pow(x = x_23_cast_fp16, y = var_846_promoted_to_fp16)[name = string("op_847_cast_fp16")];
            tensor<int32, [1]> var_13_axes_0 = const()[name = string("var_13_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_13_keep_dims_0 = const()[name = string("var_13_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_13_cast_fp16 = reduce_mean(axes = var_13_axes_0, keep_dims = var_13_keep_dims_0, x = var_847_cast_fp16)[name = string("var_13_cast_fp16")];
            fp16 var_854_to_fp16 = const()[name = string("op_854_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_855_cast_fp16 = add(x = var_13_cast_fp16, y = var_854_to_fp16)[name = string("op_855_cast_fp16")];
            fp32 var_856_epsilon_0 = const()[name = string("op_856_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_856_cast_fp16 = rsqrt(epsilon = var_856_epsilon_0, x = var_855_cast_fp16)[name = string("op_856_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_13_cast_fp16 = mul(x = x_23_cast_fp16, y = var_856_cast_fp16)[name = string("x_normed_13_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_866_to_fp16 = const()[name = string("op_866_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338824000)))];
            tensor<fp16, [1, 640, 1, ?]> attn_3_cast_fp16 = mul(x = x_normed_13_cast_fp16, y = var_866_to_fp16)[name = string("attn_3_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_25_cast_fp16 = add(x = x_1_cast_fp16, y = attn_3_cast_fp16)[name = string("x_25_cast_fp16")];
            fp16 var_870_promoted_to_fp16 = const()[name = string("op_870_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_871_cast_fp16 = pow(x = x_25_cast_fp16, y = var_870_promoted_to_fp16)[name = string("op_871_cast_fp16")];
            tensor<int32, [1]> var_15_axes_0 = const()[name = string("var_15_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_15_keep_dims_0 = const()[name = string("var_15_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_15_cast_fp16 = reduce_mean(axes = var_15_axes_0, keep_dims = var_15_keep_dims_0, x = var_871_cast_fp16)[name = string("var_15_cast_fp16")];
            fp16 var_878_to_fp16 = const()[name = string("op_878_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_879_cast_fp16 = add(x = var_15_cast_fp16, y = var_878_to_fp16)[name = string("op_879_cast_fp16")];
            fp32 var_880_epsilon_0 = const()[name = string("op_880_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_880_cast_fp16 = rsqrt(epsilon = var_880_epsilon_0, x = var_879_cast_fp16)[name = string("op_880_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_15_cast_fp16 = mul(x = x_25_cast_fp16, y = var_880_cast_fp16)[name = string("x_normed_15_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_890_to_fp16 = const()[name = string("op_890_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338825344)))];
            tensor<fp16, [1, 640, 1, ?]> input_5_cast_fp16 = mul(x = x_normed_15_cast_fp16, y = var_890_to_fp16)[name = string("input_5_cast_fp16")];
            string gate_1_pad_type_0 = const()[name = string("gate_1_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_1_strides_0 = const()[name = string("gate_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_1_pad_0 = const()[name = string("gate_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_1_dilations_0 = const()[name = string("gate_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_1_groups_0 = const()[name = string("gate_1_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_0_weight_to_fp16 = const()[name = string("gate_proj_0_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338826688)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_1_cast_fp16 = conv(dilations = gate_1_dilations_0, groups = gate_1_groups_0, pad = gate_1_pad_0, pad_type = gate_1_pad_type_0, strides = gate_1_strides_0, weight = gate_proj_0_weight_to_fp16, x = input_5_cast_fp16)[name = string("gate_1_cast_fp16")];
            string up_1_pad_type_0 = const()[name = string("up_1_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_1_strides_0 = const()[name = string("up_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_1_pad_0 = const()[name = string("up_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_1_dilations_0 = const()[name = string("up_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_1_groups_0 = const()[name = string("up_1_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_0_weight_to_fp16 = const()[name = string("up_proj_0_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341448192)))];
            tensor<fp16, [1, 2048, 1, ?]> up_1_cast_fp16 = conv(dilations = up_1_dilations_0, groups = up_1_groups_0, pad = up_1_pad_0, pad_type = up_1_pad_type_0, strides = up_1_strides_0, weight = up_proj_0_weight_to_fp16, x = input_5_cast_fp16)[name = string("up_1_cast_fp16")];
            string var_915_mode_0 = const()[name = string("op_915_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_915_cast_fp16 = gelu(mode = var_915_mode_0, x = gate_1_cast_fp16)[name = string("op_915_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_7_cast_fp16 = mul(x = var_915_cast_fp16, y = up_1_cast_fp16)[name = string("input_7_cast_fp16")];
            string x_27_pad_type_0 = const()[name = string("x_27_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_27_strides_0 = const()[name = string("x_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_27_pad_0 = const()[name = string("x_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_27_dilations_0 = const()[name = string("x_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_27_groups_0 = const()[name = string("x_27_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_0_weight_to_fp16 = const()[name = string("down_proj_0_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344069696)))];
            tensor<fp16, [1, 640, 1, ?]> x_27_cast_fp16 = conv(dilations = x_27_dilations_0, groups = x_27_groups_0, pad = x_27_pad_0, pad_type = x_27_pad_type_0, strides = x_27_strides_0, weight = down_proj_0_weight_to_fp16, x = input_7_cast_fp16)[name = string("x_27_cast_fp16")];
            fp16 var_928_promoted_to_fp16 = const()[name = string("op_928_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_929_cast_fp16 = pow(x = x_27_cast_fp16, y = var_928_promoted_to_fp16)[name = string("op_929_cast_fp16")];
            tensor<int32, [1]> var_17_axes_0 = const()[name = string("var_17_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_17_keep_dims_0 = const()[name = string("var_17_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_17_cast_fp16 = reduce_mean(axes = var_17_axes_0, keep_dims = var_17_keep_dims_0, x = var_929_cast_fp16)[name = string("var_17_cast_fp16")];
            fp16 var_936_to_fp16 = const()[name = string("op_936_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_937_cast_fp16 = add(x = var_17_cast_fp16, y = var_936_to_fp16)[name = string("op_937_cast_fp16")];
            fp32 var_938_epsilon_0 = const()[name = string("op_938_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_938_cast_fp16 = rsqrt(epsilon = var_938_epsilon_0, x = var_937_cast_fp16)[name = string("op_938_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_17_cast_fp16 = mul(x = x_27_cast_fp16, y = var_938_cast_fp16)[name = string("x_normed_17_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_948_to_fp16 = const()[name = string("op_948_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346691200)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_1_cast_fp16 = mul(x = x_normed_17_cast_fp16, y = var_948_to_fp16)[name = string("mlp_1_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_29_cast_fp16 = add(x = x_25_cast_fp16, y = mlp_1_cast_fp16)[name = string("x_29_cast_fp16")];
            fp16 var_952_promoted_to_fp16 = const()[name = string("op_952_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_953_cast_fp16 = pow(x = x_29_cast_fp16, y = var_952_promoted_to_fp16)[name = string("op_953_cast_fp16")];
            tensor<int32, [1]> var_19_axes_0 = const()[name = string("var_19_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_19_keep_dims_0 = const()[name = string("var_19_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_19_cast_fp16 = reduce_mean(axes = var_19_axes_0, keep_dims = var_19_keep_dims_0, x = var_953_cast_fp16)[name = string("var_19_cast_fp16")];
            fp16 var_960_to_fp16 = const()[name = string("op_960_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_961_cast_fp16 = add(x = var_19_cast_fp16, y = var_960_to_fp16)[name = string("op_961_cast_fp16")];
            fp32 var_962_epsilon_0 = const()[name = string("op_962_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_962_cast_fp16 = rsqrt(epsilon = var_962_epsilon_0, x = var_961_cast_fp16)[name = string("op_962_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_19_cast_fp16 = mul(x = x_29_cast_fp16, y = var_962_cast_fp16)[name = string("x_normed_19_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_972_to_fp16 = const()[name = string("op_972_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346692544)))];
            tensor<fp16, [1, 640, 1, ?]> input_9_cast_fp16 = mul(x = x_normed_19_cast_fp16, y = var_972_to_fp16)[name = string("input_9_cast_fp16")];
            string var_984_pad_type_0 = const()[name = string("op_984_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_984_strides_0 = const()[name = string("op_984_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_984_pad_0 = const()[name = string("op_984_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_984_dilations_0 = const()[name = string("op_984_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_984_groups_0 = const()[name = string("op_984_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_1_weight_to_fp16 = const()[name = string("q_proj_1_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346693888)))];
            tensor<fp16, [1, 1024, 1, ?]> var_984_cast_fp16 = conv(dilations = var_984_dilations_0, groups = var_984_groups_0, pad = var_984_pad_0, pad_type = var_984_pad_type_0, strides = var_984_strides_0, weight = q_proj_1_weight_to_fp16, x = input_9_cast_fp16)[name = string("op_984_cast_fp16")];
            string var_995_pad_type_0 = const()[name = string("op_995_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_995_strides_0 = const()[name = string("op_995_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_995_pad_0 = const()[name = string("op_995_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_995_dilations_0 = const()[name = string("op_995_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_995_groups_0 = const()[name = string("op_995_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_1_weight_to_fp16 = const()[name = string("k_proj_1_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348004672)))];
            tensor<fp16, [1, 256, 1, ?]> var_995_cast_fp16 = conv(dilations = var_995_dilations_0, groups = var_995_groups_0, pad = var_995_pad_0, pad_type = var_995_pad_type_0, strides = var_995_strides_0, weight = k_proj_1_weight_to_fp16, x = input_9_cast_fp16)[name = string("op_995_cast_fp16")];
            string var_1006_pad_type_0 = const()[name = string("op_1006_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1006_strides_0 = const()[name = string("op_1006_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1006_pad_0 = const()[name = string("op_1006_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1006_dilations_0 = const()[name = string("op_1006_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1006_groups_0 = const()[name = string("op_1006_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_1_weight_to_fp16 = const()[name = string("v_proj_1_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348332416)))];
            tensor<fp16, [1, 256, 1, ?]> var_1006_cast_fp16 = conv(dilations = var_1006_dilations_0, groups = var_1006_groups_0, pad = var_1006_pad_0, pad_type = var_1006_pad_type_0, strides = var_1006_strides_0, weight = v_proj_1_weight_to_fp16, x = input_9_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [4]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_1009_axis_0 = const()[name = string("op_1009_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_1009_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_1009_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_1009_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_1009_cast_fp16_3 = split(axis = var_1009_axis_0, split_sizes = tile_3, x = var_984_cast_fp16)[name = string("op_1009_cast_fp16")];
            fp16 var_1022_promoted_to_fp16 = const()[name = string("op_1022_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_1023_cast_fp16 = pow(x = var_1009_cast_fp16_0, y = var_1022_promoted_to_fp16)[name = string("op_1023_cast_fp16")];
            tensor<int32, [1]> var_21_axes_0 = const()[name = string("var_21_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_21_keep_dims_0 = const()[name = string("var_21_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_21_cast_fp16 = reduce_mean(axes = var_21_axes_0, keep_dims = var_21_keep_dims_0, x = var_1023_cast_fp16)[name = string("var_21_cast_fp16")];
            fp16 var_1030_to_fp16 = const()[name = string("op_1030_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_1031_cast_fp16 = add(x = var_21_cast_fp16, y = var_1030_to_fp16)[name = string("op_1031_cast_fp16")];
            fp32 var_1032_epsilon_0 = const()[name = string("op_1032_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_1032_cast_fp16 = rsqrt(epsilon = var_1032_epsilon_0, x = var_1031_cast_fp16)[name = string("op_1032_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_21_cast_fp16 = mul(x = var_1009_cast_fp16_0, y = var_1032_cast_fp16)[name = string("x_normed_21_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_1042_to_fp16 = const()[name = string("op_1042_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348660160)))];
            tensor<fp16, [1, 256, 1, ?]> x_33_cast_fp16 = mul(x = x_normed_21_cast_fp16, y = var_1042_to_fp16)[name = string("x_33_cast_fp16")];
            tensor<int32, [4]> x_first_11_begin_0 = const()[name = string("x_first_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_11_end_0 = const()[name = string("x_first_11_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_11_end_mask_0 = const()[name = string("x_first_11_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_11_cast_fp16 = slice_by_index(begin = x_first_11_begin_0, end = x_first_11_end_0, end_mask = x_first_11_end_mask_0, x = x_33_cast_fp16)[name = string("x_first_11_cast_fp16")];
            tensor<int32, [4]> x_second_11_begin_0 = const()[name = string("x_second_11_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_11_end_0 = const()[name = string("x_second_11_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_11_end_mask_0 = const()[name = string("x_second_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_11_cast_fp16 = slice_by_index(begin = x_second_11_begin_0, end = x_second_11_end_0, end_mask = x_second_11_end_mask_0, x = x_33_cast_fp16)[name = string("x_second_11_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_1064_cast_fp16 = mul(x = x_second_11_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_1064_cast_fp16")];
            int32 var_1066 = const()[name = string("op_1066"), val = int32(1)];
            bool x_rot_11_interleave_0 = const()[name = string("x_rot_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_11_cast_fp16 = concat(axis = var_1066, interleave = x_rot_11_interleave_0, values = (var_1064_cast_fp16, x_first_11_cast_fp16))[name = string("x_rot_11_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1082_cast_fp16 = mul(x = x_33_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_1082_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1083_cast_fp16 = mul(x = x_rot_11_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_1083_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_9_cast_fp16 = add(x = var_1082_cast_fp16, y = var_1083_cast_fp16)[name = string("q_h_9_cast_fp16")];
            fp16 var_1086_promoted_to_fp16 = const()[name = string("op_1086_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_1087_cast_fp16 = pow(x = var_1009_cast_fp16_1, y = var_1086_promoted_to_fp16)[name = string("op_1087_cast_fp16")];
            tensor<int32, [1]> var_23_axes_0 = const()[name = string("var_23_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_23_keep_dims_0 = const()[name = string("var_23_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_23_cast_fp16 = reduce_mean(axes = var_23_axes_0, keep_dims = var_23_keep_dims_0, x = var_1087_cast_fp16)[name = string("var_23_cast_fp16")];
            fp16 var_1094_to_fp16 = const()[name = string("op_1094_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_1095_cast_fp16 = add(x = var_23_cast_fp16, y = var_1094_to_fp16)[name = string("op_1095_cast_fp16")];
            fp32 var_1096_epsilon_0 = const()[name = string("op_1096_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_1096_cast_fp16 = rsqrt(epsilon = var_1096_epsilon_0, x = var_1095_cast_fp16)[name = string("op_1096_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_23_cast_fp16 = mul(x = var_1009_cast_fp16_1, y = var_1096_cast_fp16)[name = string("x_normed_23_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_37_cast_fp16 = mul(x = x_normed_23_cast_fp16, y = var_1042_to_fp16)[name = string("x_37_cast_fp16")];
            tensor<int32, [4]> x_first_13_begin_0 = const()[name = string("x_first_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_13_end_0 = const()[name = string("x_first_13_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_13_end_mask_0 = const()[name = string("x_first_13_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_13_cast_fp16 = slice_by_index(begin = x_first_13_begin_0, end = x_first_13_end_0, end_mask = x_first_13_end_mask_0, x = x_37_cast_fp16)[name = string("x_first_13_cast_fp16")];
            tensor<int32, [4]> x_second_13_begin_0 = const()[name = string("x_second_13_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_13_end_0 = const()[name = string("x_second_13_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_13_end_mask_0 = const()[name = string("x_second_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_13_cast_fp16 = slice_by_index(begin = x_second_13_begin_0, end = x_second_13_end_0, end_mask = x_second_13_end_mask_0, x = x_37_cast_fp16)[name = string("x_second_13_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_1128_cast_fp16 = mul(x = x_second_13_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_1128_cast_fp16")];
            int32 var_1130 = const()[name = string("op_1130"), val = int32(1)];
            bool x_rot_13_interleave_0 = const()[name = string("x_rot_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_13_cast_fp16 = concat(axis = var_1130, interleave = x_rot_13_interleave_0, values = (var_1128_cast_fp16, x_first_13_cast_fp16))[name = string("x_rot_13_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1146_cast_fp16 = mul(x = x_37_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_1146_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1147_cast_fp16 = mul(x = x_rot_13_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_1147_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_11_cast_fp16 = add(x = var_1146_cast_fp16, y = var_1147_cast_fp16)[name = string("q_h_11_cast_fp16")];
            fp16 var_1150_promoted_to_fp16 = const()[name = string("op_1150_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_1151_cast_fp16 = pow(x = var_1009_cast_fp16_2, y = var_1150_promoted_to_fp16)[name = string("op_1151_cast_fp16")];
            tensor<int32, [1]> var_25_axes_0 = const()[name = string("var_25_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_25_keep_dims_0 = const()[name = string("var_25_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_25_cast_fp16 = reduce_mean(axes = var_25_axes_0, keep_dims = var_25_keep_dims_0, x = var_1151_cast_fp16)[name = string("var_25_cast_fp16")];
            fp16 var_1158_to_fp16 = const()[name = string("op_1158_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_1159_cast_fp16 = add(x = var_25_cast_fp16, y = var_1158_to_fp16)[name = string("op_1159_cast_fp16")];
            fp32 var_1160_epsilon_0 = const()[name = string("op_1160_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_1160_cast_fp16 = rsqrt(epsilon = var_1160_epsilon_0, x = var_1159_cast_fp16)[name = string("op_1160_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_25_cast_fp16 = mul(x = var_1009_cast_fp16_2, y = var_1160_cast_fp16)[name = string("x_normed_25_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_41_cast_fp16 = mul(x = x_normed_25_cast_fp16, y = var_1042_to_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> x_first_15_begin_0 = const()[name = string("x_first_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_15_end_0 = const()[name = string("x_first_15_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_15_end_mask_0 = const()[name = string("x_first_15_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_15_cast_fp16 = slice_by_index(begin = x_first_15_begin_0, end = x_first_15_end_0, end_mask = x_first_15_end_mask_0, x = x_41_cast_fp16)[name = string("x_first_15_cast_fp16")];
            tensor<int32, [4]> x_second_15_begin_0 = const()[name = string("x_second_15_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_15_end_0 = const()[name = string("x_second_15_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_15_end_mask_0 = const()[name = string("x_second_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_15_cast_fp16 = slice_by_index(begin = x_second_15_begin_0, end = x_second_15_end_0, end_mask = x_second_15_end_mask_0, x = x_41_cast_fp16)[name = string("x_second_15_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_1192_cast_fp16 = mul(x = x_second_15_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_1192_cast_fp16")];
            int32 var_1194 = const()[name = string("op_1194"), val = int32(1)];
            bool x_rot_15_interleave_0 = const()[name = string("x_rot_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_15_cast_fp16 = concat(axis = var_1194, interleave = x_rot_15_interleave_0, values = (var_1192_cast_fp16, x_first_15_cast_fp16))[name = string("x_rot_15_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1210_cast_fp16 = mul(x = x_41_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_1210_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1211_cast_fp16 = mul(x = x_rot_15_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_1211_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_13_cast_fp16 = add(x = var_1210_cast_fp16, y = var_1211_cast_fp16)[name = string("q_h_13_cast_fp16")];
            fp16 var_1214_promoted_to_fp16 = const()[name = string("op_1214_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_1215_cast_fp16 = pow(x = var_1009_cast_fp16_3, y = var_1214_promoted_to_fp16)[name = string("op_1215_cast_fp16")];
            tensor<int32, [1]> var_27_axes_0 = const()[name = string("var_27_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_27_keep_dims_0 = const()[name = string("var_27_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_27_cast_fp16 = reduce_mean(axes = var_27_axes_0, keep_dims = var_27_keep_dims_0, x = var_1215_cast_fp16)[name = string("var_27_cast_fp16")];
            fp16 var_1222_to_fp16 = const()[name = string("op_1222_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_1223_cast_fp16 = add(x = var_27_cast_fp16, y = var_1222_to_fp16)[name = string("op_1223_cast_fp16")];
            fp32 var_1224_epsilon_0 = const()[name = string("op_1224_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_1224_cast_fp16 = rsqrt(epsilon = var_1224_epsilon_0, x = var_1223_cast_fp16)[name = string("op_1224_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_27_cast_fp16 = mul(x = var_1009_cast_fp16_3, y = var_1224_cast_fp16)[name = string("x_normed_27_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_45_cast_fp16 = mul(x = x_normed_27_cast_fp16, y = var_1042_to_fp16)[name = string("x_45_cast_fp16")];
            tensor<int32, [4]> x_first_17_begin_0 = const()[name = string("x_first_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_17_end_0 = const()[name = string("x_first_17_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_17_end_mask_0 = const()[name = string("x_first_17_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_17_cast_fp16 = slice_by_index(begin = x_first_17_begin_0, end = x_first_17_end_0, end_mask = x_first_17_end_mask_0, x = x_45_cast_fp16)[name = string("x_first_17_cast_fp16")];
            tensor<int32, [4]> x_second_17_begin_0 = const()[name = string("x_second_17_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_17_end_0 = const()[name = string("x_second_17_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_17_end_mask_0 = const()[name = string("x_second_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_17_cast_fp16 = slice_by_index(begin = x_second_17_begin_0, end = x_second_17_end_0, end_mask = x_second_17_end_mask_0, x = x_45_cast_fp16)[name = string("x_second_17_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_1256_cast_fp16 = mul(x = x_second_17_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_1256_cast_fp16")];
            int32 var_1258 = const()[name = string("op_1258"), val = int32(1)];
            bool x_rot_17_interleave_0 = const()[name = string("x_rot_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_17_cast_fp16 = concat(axis = var_1258, interleave = x_rot_17_interleave_0, values = (var_1256_cast_fp16, x_first_17_cast_fp16))[name = string("x_rot_17_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1274_cast_fp16 = mul(x = x_45_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_1274_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1275_cast_fp16 = mul(x = x_rot_17_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_1275_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_15_cast_fp16 = add(x = var_1274_cast_fp16, y = var_1275_cast_fp16)[name = string("q_h_15_cast_fp16")];
            fp16 var_1278_promoted_to_fp16 = const()[name = string("op_1278_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_1279_cast_fp16 = pow(x = var_995_cast_fp16, y = var_1278_promoted_to_fp16)[name = string("op_1279_cast_fp16")];
            tensor<int32, [1]> var_29_axes_0 = const()[name = string("var_29_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_29_keep_dims_0 = const()[name = string("var_29_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_29_cast_fp16 = reduce_mean(axes = var_29_axes_0, keep_dims = var_29_keep_dims_0, x = var_1279_cast_fp16)[name = string("var_29_cast_fp16")];
            fp16 var_1286_to_fp16 = const()[name = string("op_1286_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_1287_cast_fp16 = add(x = var_29_cast_fp16, y = var_1286_to_fp16)[name = string("op_1287_cast_fp16")];
            fp32 var_1288_epsilon_0 = const()[name = string("op_1288_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_1288_cast_fp16 = rsqrt(epsilon = var_1288_epsilon_0, x = var_1287_cast_fp16)[name = string("op_1288_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_29_cast_fp16 = mul(x = var_995_cast_fp16, y = var_1288_cast_fp16)[name = string("x_normed_29_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_1298_to_fp16 = const()[name = string("op_1298_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348660736)))];
            tensor<fp16, [1, 256, 1, ?]> x_49_cast_fp16 = mul(x = x_normed_29_cast_fp16, y = var_1298_to_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [4]> x_first_19_begin_0 = const()[name = string("x_first_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_19_end_0 = const()[name = string("x_first_19_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_19_end_mask_0 = const()[name = string("x_first_19_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_19_cast_fp16 = slice_by_index(begin = x_first_19_begin_0, end = x_first_19_end_0, end_mask = x_first_19_end_mask_0, x = x_49_cast_fp16)[name = string("x_first_19_cast_fp16")];
            tensor<int32, [4]> x_second_19_begin_0 = const()[name = string("x_second_19_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_19_end_0 = const()[name = string("x_second_19_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_19_end_mask_0 = const()[name = string("x_second_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_19_cast_fp16 = slice_by_index(begin = x_second_19_begin_0, end = x_second_19_end_0, end_mask = x_second_19_end_mask_0, x = x_49_cast_fp16)[name = string("x_second_19_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_1320_cast_fp16 = mul(x = x_second_19_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_1320_cast_fp16")];
            int32 var_1322 = const()[name = string("op_1322"), val = int32(1)];
            bool x_rot_19_interleave_0 = const()[name = string("x_rot_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_19_cast_fp16 = concat(axis = var_1322, interleave = x_rot_19_interleave_0, values = (var_1320_cast_fp16, x_first_19_cast_fp16))[name = string("x_rot_19_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1338_cast_fp16 = mul(x = x_49_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_1338_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1339_cast_fp16 = mul(x = x_rot_19_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_1339_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_3_cast_fp16 = add(x = var_1338_cast_fp16, y = var_1339_cast_fp16)[name = string("k_new_3_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_2 = read_state(input = k_cache_1)[name = string("read_state_2")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_9_cast_fp16 = mul(x = read_state_2, y = one_minus_1_cast_fp16)[name = string("k_cache_9_cast_fp16")];
            write_state(data = k_cache_9_cast_fp16, input = k_cache_1)[name = string("coreml_update_state_76_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_76 = read_state(input = k_cache_1)[name = string("coreml_update_state_76")];
            tensor<fp16, [1, 256, 1, 128]> var_1346_cast_fp16 = mul(x = k_new_3_cast_fp16, y = write_mask)[name = string("op_1346_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_11_cast_fp16 = add(x = coreml_update_state_76, y = var_1346_cast_fp16)[name = string("k_cache_11_cast_fp16")];
            write_state(data = k_cache_11_cast_fp16, input = k_cache_1)[name = string("coreml_update_state_77_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_77 = read_state(input = k_cache_1)[name = string("coreml_update_state_77")];
            tensor<fp16, [1, 256, 1, 128]> read_state_3 = read_state(input = v_cache_1)[name = string("read_state_3")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_9_cast_fp16 = mul(x = read_state_3, y = one_minus_1_cast_fp16)[name = string("v_cache_9_cast_fp16")];
            write_state(data = v_cache_9_cast_fp16, input = v_cache_1)[name = string("coreml_update_state_78_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_78 = read_state(input = v_cache_1)[name = string("coreml_update_state_78")];
            tensor<fp16, [1, 256, 1, 128]> var_1350_cast_fp16 = mul(x = var_1006_cast_fp16, y = write_mask)[name = string("op_1350_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_11_cast_fp16 = add(x = coreml_update_state_78, y = var_1350_cast_fp16)[name = string("v_cache_11_cast_fp16")];
            write_state(data = v_cache_11_cast_fp16, input = v_cache_1)[name = string("coreml_update_state_79_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_79 = read_state(input = v_cache_1)[name = string("coreml_update_state_79")];
            bool var_1365_transpose_x_0 = const()[name = string("op_1365_transpose_x_0"), val = bool(false)];
            bool var_1365_transpose_y_0 = const()[name = string("op_1365_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_314_perm_0 = const()[name = string("transpose_314_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_315_perm_0 = const()[name = string("transpose_315_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_315 = transpose(perm = transpose_315_perm_0, x = coreml_update_state_77)[name = string("transpose_569")];
            tensor<fp16, [1, 1, ?, 256]> transpose_314 = transpose(perm = transpose_314_perm_0, x = q_h_9_cast_fp16)[name = string("transpose_570")];
            tensor<fp16, [1, 1, ?, 128]> var_1365_cast_fp16 = matmul(transpose_x = var_1365_transpose_x_0, transpose_y = var_1365_transpose_y_0, x = transpose_314, y = transpose_315)[name = string("op_1365_cast_fp16")];
            fp16 var_1366_to_fp16 = const()[name = string("op_1366_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_1367_cast_fp16 = mul(x = var_1365_cast_fp16, y = var_1366_to_fp16)[name = string("op_1367_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_9_cast_fp16 = add(x = var_1367_cast_fp16, y = attention_mask)[name = string("scores_9_cast_fp16")];
            int32 var_1370 = const()[name = string("op_1370"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_9_cast_fp16 = softmax(axis = var_1370, x = scores_9_cast_fp16)[name = string("probs_9_cast_fp16")];
            tensor<int32, [4]> var_1377 = const()[name = string("op_1377"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_1379_transpose_x_0 = const()[name = string("op_1379_transpose_x_0"), val = bool(false)];
            bool var_1379_transpose_y_0 = const()[name = string("op_1379_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_9_cast_fp16 = transpose(perm = var_1377, x = coreml_update_state_79)[name = string("transpose_568")];
            tensor<fp16, [1, 1, ?, 256]> var_1379_cast_fp16 = matmul(transpose_x = var_1379_transpose_x_0, transpose_y = var_1379_transpose_y_0, x = probs_9_cast_fp16, y = v_t_9_cast_fp16)[name = string("op_1379_cast_fp16")];
            bool var_1398_transpose_x_0 = const()[name = string("op_1398_transpose_x_0"), val = bool(false)];
            bool var_1398_transpose_y_0 = const()[name = string("op_1398_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_319_perm_0 = const()[name = string("transpose_319_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_319 = transpose(perm = transpose_319_perm_0, x = q_h_11_cast_fp16)[name = string("transpose_567")];
            tensor<fp16, [1, 1, ?, 128]> var_1398_cast_fp16 = matmul(transpose_x = var_1398_transpose_x_0, transpose_y = var_1398_transpose_y_0, x = transpose_319, y = transpose_315)[name = string("op_1398_cast_fp16")];
            fp16 var_1399_to_fp16 = const()[name = string("op_1399_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_1400_cast_fp16 = mul(x = var_1398_cast_fp16, y = var_1399_to_fp16)[name = string("op_1400_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_11_cast_fp16 = add(x = var_1400_cast_fp16, y = attention_mask)[name = string("scores_11_cast_fp16")];
            int32 var_1403 = const()[name = string("op_1403"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_11_cast_fp16 = softmax(axis = var_1403, x = scores_11_cast_fp16)[name = string("probs_11_cast_fp16")];
            bool var_1412_transpose_x_0 = const()[name = string("op_1412_transpose_x_0"), val = bool(false)];
            bool var_1412_transpose_y_0 = const()[name = string("op_1412_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_1412_cast_fp16 = matmul(transpose_x = var_1412_transpose_x_0, transpose_y = var_1412_transpose_y_0, x = probs_11_cast_fp16, y = v_t_9_cast_fp16)[name = string("op_1412_cast_fp16")];
            bool var_1431_transpose_x_0 = const()[name = string("op_1431_transpose_x_0"), val = bool(false)];
            bool var_1431_transpose_y_0 = const()[name = string("op_1431_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_320_perm_0 = const()[name = string("transpose_320_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_320 = transpose(perm = transpose_320_perm_0, x = q_h_13_cast_fp16)[name = string("transpose_566")];
            tensor<fp16, [1, 1, ?, 128]> var_1431_cast_fp16 = matmul(transpose_x = var_1431_transpose_x_0, transpose_y = var_1431_transpose_y_0, x = transpose_320, y = transpose_315)[name = string("op_1431_cast_fp16")];
            fp16 var_1432_to_fp16 = const()[name = string("op_1432_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_1433_cast_fp16 = mul(x = var_1431_cast_fp16, y = var_1432_to_fp16)[name = string("op_1433_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_13_cast_fp16 = add(x = var_1433_cast_fp16, y = attention_mask)[name = string("scores_13_cast_fp16")];
            int32 var_1436 = const()[name = string("op_1436"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_13_cast_fp16 = softmax(axis = var_1436, x = scores_13_cast_fp16)[name = string("probs_13_cast_fp16")];
            bool var_1445_transpose_x_0 = const()[name = string("op_1445_transpose_x_0"), val = bool(false)];
            bool var_1445_transpose_y_0 = const()[name = string("op_1445_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_1445_cast_fp16 = matmul(transpose_x = var_1445_transpose_x_0, transpose_y = var_1445_transpose_y_0, x = probs_13_cast_fp16, y = v_t_9_cast_fp16)[name = string("op_1445_cast_fp16")];
            bool var_1464_transpose_x_0 = const()[name = string("op_1464_transpose_x_0"), val = bool(false)];
            bool var_1464_transpose_y_0 = const()[name = string("op_1464_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_321_perm_0 = const()[name = string("transpose_321_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_321 = transpose(perm = transpose_321_perm_0, x = q_h_15_cast_fp16)[name = string("transpose_565")];
            tensor<fp16, [1, 1, ?, 128]> var_1464_cast_fp16 = matmul(transpose_x = var_1464_transpose_x_0, transpose_y = var_1464_transpose_y_0, x = transpose_321, y = transpose_315)[name = string("op_1464_cast_fp16")];
            fp16 var_1465_to_fp16 = const()[name = string("op_1465_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_1466_cast_fp16 = mul(x = var_1464_cast_fp16, y = var_1465_to_fp16)[name = string("op_1466_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_15_cast_fp16 = add(x = var_1466_cast_fp16, y = attention_mask)[name = string("scores_15_cast_fp16")];
            int32 var_1469 = const()[name = string("op_1469"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_15_cast_fp16 = softmax(axis = var_1469, x = scores_15_cast_fp16)[name = string("probs_15_cast_fp16")];
            bool var_1478_transpose_x_0 = const()[name = string("op_1478_transpose_x_0"), val = bool(false)];
            bool var_1478_transpose_y_0 = const()[name = string("op_1478_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_1478_cast_fp16 = matmul(transpose_x = var_1478_transpose_x_0, transpose_y = var_1478_transpose_y_0, x = probs_15_cast_fp16, y = v_t_9_cast_fp16)[name = string("op_1478_cast_fp16")];
            bool input_11_interleave_0 = const()[name = string("input_11_interleave_0"), val = bool(false)];
            int32 const_146 = const()[name = string("const_146"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_11_cast_fp16 = concat(axis = const_146, interleave = input_11_interleave_0, values = (var_1379_cast_fp16, var_1412_cast_fp16, var_1445_cast_fp16, var_1478_cast_fp16))[name = string("input_11_cast_fp16")];
            string x_51_pad_type_0 = const()[name = string("x_51_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_51_strides_0 = const()[name = string("x_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_51_pad_0 = const()[name = string("x_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_51_dilations_0 = const()[name = string("x_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_51_groups_0 = const()[name = string("x_51_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_17_perm_0 = const()[name = string("transpose_17_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_1_weight_to_fp16 = const()[name = string("o_proj_1_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348661312)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = input_11_cast_fp16)[name = string("transpose_564")];
            tensor<fp16, [1, 640, 1, ?]> x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = o_proj_1_weight_to_fp16, x = transpose_17_cast_fp16)[name = string("x_51_cast_fp16")];
            fp16 var_1499_promoted_to_fp16 = const()[name = string("op_1499_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_1500_cast_fp16 = pow(x = x_51_cast_fp16, y = var_1499_promoted_to_fp16)[name = string("op_1500_cast_fp16")];
            tensor<int32, [1]> var_31_axes_0 = const()[name = string("var_31_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_31_keep_dims_0 = const()[name = string("var_31_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_31_cast_fp16 = reduce_mean(axes = var_31_axes_0, keep_dims = var_31_keep_dims_0, x = var_1500_cast_fp16)[name = string("var_31_cast_fp16")];
            fp16 var_1507_to_fp16 = const()[name = string("op_1507_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_1508_cast_fp16 = add(x = var_31_cast_fp16, y = var_1507_to_fp16)[name = string("op_1508_cast_fp16")];
            fp32 var_1509_epsilon_0 = const()[name = string("op_1509_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_1509_cast_fp16 = rsqrt(epsilon = var_1509_epsilon_0, x = var_1508_cast_fp16)[name = string("op_1509_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_31_cast_fp16 = mul(x = x_51_cast_fp16, y = var_1509_cast_fp16)[name = string("x_normed_31_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_1519_to_fp16 = const()[name = string("op_1519_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349972096)))];
            tensor<fp16, [1, 640, 1, ?]> attn_7_cast_fp16 = mul(x = x_normed_31_cast_fp16, y = var_1519_to_fp16)[name = string("attn_7_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_53_cast_fp16 = add(x = x_29_cast_fp16, y = attn_7_cast_fp16)[name = string("x_53_cast_fp16")];
            fp16 var_1523_promoted_to_fp16 = const()[name = string("op_1523_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_1524_cast_fp16 = pow(x = x_53_cast_fp16, y = var_1523_promoted_to_fp16)[name = string("op_1524_cast_fp16")];
            tensor<int32, [1]> var_33_axes_0 = const()[name = string("var_33_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_33_keep_dims_0 = const()[name = string("var_33_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_33_cast_fp16 = reduce_mean(axes = var_33_axes_0, keep_dims = var_33_keep_dims_0, x = var_1524_cast_fp16)[name = string("var_33_cast_fp16")];
            fp16 var_1531_to_fp16 = const()[name = string("op_1531_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_1532_cast_fp16 = add(x = var_33_cast_fp16, y = var_1531_to_fp16)[name = string("op_1532_cast_fp16")];
            fp32 var_1533_epsilon_0 = const()[name = string("op_1533_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_1533_cast_fp16 = rsqrt(epsilon = var_1533_epsilon_0, x = var_1532_cast_fp16)[name = string("op_1533_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_33_cast_fp16 = mul(x = x_53_cast_fp16, y = var_1533_cast_fp16)[name = string("x_normed_33_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_1543_to_fp16 = const()[name = string("op_1543_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349973440)))];
            tensor<fp16, [1, 640, 1, ?]> input_13_cast_fp16 = mul(x = x_normed_33_cast_fp16, y = var_1543_to_fp16)[name = string("input_13_cast_fp16")];
            string gate_3_pad_type_0 = const()[name = string("gate_3_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_3_strides_0 = const()[name = string("gate_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_3_pad_0 = const()[name = string("gate_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_3_dilations_0 = const()[name = string("gate_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_3_groups_0 = const()[name = string("gate_3_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_1_weight_to_fp16 = const()[name = string("gate_proj_1_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349974784)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_3_cast_fp16 = conv(dilations = gate_3_dilations_0, groups = gate_3_groups_0, pad = gate_3_pad_0, pad_type = gate_3_pad_type_0, strides = gate_3_strides_0, weight = gate_proj_1_weight_to_fp16, x = input_13_cast_fp16)[name = string("gate_3_cast_fp16")];
            string up_3_pad_type_0 = const()[name = string("up_3_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_3_strides_0 = const()[name = string("up_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_3_pad_0 = const()[name = string("up_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_3_dilations_0 = const()[name = string("up_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_3_groups_0 = const()[name = string("up_3_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_1_weight_to_fp16 = const()[name = string("up_proj_1_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352596288)))];
            tensor<fp16, [1, 2048, 1, ?]> up_3_cast_fp16 = conv(dilations = up_3_dilations_0, groups = up_3_groups_0, pad = up_3_pad_0, pad_type = up_3_pad_type_0, strides = up_3_strides_0, weight = up_proj_1_weight_to_fp16, x = input_13_cast_fp16)[name = string("up_3_cast_fp16")];
            string var_1568_mode_0 = const()[name = string("op_1568_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_1568_cast_fp16 = gelu(mode = var_1568_mode_0, x = gate_3_cast_fp16)[name = string("op_1568_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_15_cast_fp16 = mul(x = var_1568_cast_fp16, y = up_3_cast_fp16)[name = string("input_15_cast_fp16")];
            string x_55_pad_type_0 = const()[name = string("x_55_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_55_strides_0 = const()[name = string("x_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_55_pad_0 = const()[name = string("x_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_55_dilations_0 = const()[name = string("x_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_55_groups_0 = const()[name = string("x_55_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_1_weight_to_fp16 = const()[name = string("down_proj_1_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355217792)))];
            tensor<fp16, [1, 640, 1, ?]> x_55_cast_fp16 = conv(dilations = x_55_dilations_0, groups = x_55_groups_0, pad = x_55_pad_0, pad_type = x_55_pad_type_0, strides = x_55_strides_0, weight = down_proj_1_weight_to_fp16, x = input_15_cast_fp16)[name = string("x_55_cast_fp16")];
            fp16 var_1581_promoted_to_fp16 = const()[name = string("op_1581_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_1582_cast_fp16 = pow(x = x_55_cast_fp16, y = var_1581_promoted_to_fp16)[name = string("op_1582_cast_fp16")];
            tensor<int32, [1]> var_35_axes_0 = const()[name = string("var_35_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_35_keep_dims_0 = const()[name = string("var_35_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_35_cast_fp16 = reduce_mean(axes = var_35_axes_0, keep_dims = var_35_keep_dims_0, x = var_1582_cast_fp16)[name = string("var_35_cast_fp16")];
            fp16 var_1589_to_fp16 = const()[name = string("op_1589_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_1590_cast_fp16 = add(x = var_35_cast_fp16, y = var_1589_to_fp16)[name = string("op_1590_cast_fp16")];
            fp32 var_1591_epsilon_0 = const()[name = string("op_1591_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_1591_cast_fp16 = rsqrt(epsilon = var_1591_epsilon_0, x = var_1590_cast_fp16)[name = string("op_1591_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_35_cast_fp16 = mul(x = x_55_cast_fp16, y = var_1591_cast_fp16)[name = string("x_normed_35_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_1601_to_fp16 = const()[name = string("op_1601_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357839296)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_3_cast_fp16 = mul(x = x_normed_35_cast_fp16, y = var_1601_to_fp16)[name = string("mlp_3_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_57_cast_fp16 = add(x = x_53_cast_fp16, y = mlp_3_cast_fp16)[name = string("x_57_cast_fp16")];
            fp16 var_1605_promoted_to_fp16 = const()[name = string("op_1605_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_1606_cast_fp16 = pow(x = x_57_cast_fp16, y = var_1605_promoted_to_fp16)[name = string("op_1606_cast_fp16")];
            tensor<int32, [1]> var_37_axes_0 = const()[name = string("var_37_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_37_keep_dims_0 = const()[name = string("var_37_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_37_cast_fp16 = reduce_mean(axes = var_37_axes_0, keep_dims = var_37_keep_dims_0, x = var_1606_cast_fp16)[name = string("var_37_cast_fp16")];
            fp16 var_1613_to_fp16 = const()[name = string("op_1613_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_1614_cast_fp16 = add(x = var_37_cast_fp16, y = var_1613_to_fp16)[name = string("op_1614_cast_fp16")];
            fp32 var_1615_epsilon_0 = const()[name = string("op_1615_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_1615_cast_fp16 = rsqrt(epsilon = var_1615_epsilon_0, x = var_1614_cast_fp16)[name = string("op_1615_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_37_cast_fp16 = mul(x = x_57_cast_fp16, y = var_1615_cast_fp16)[name = string("x_normed_37_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_1625_to_fp16 = const()[name = string("op_1625_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357840640)))];
            tensor<fp16, [1, 640, 1, ?]> input_17_cast_fp16 = mul(x = x_normed_37_cast_fp16, y = var_1625_to_fp16)[name = string("input_17_cast_fp16")];
            string var_1637_pad_type_0 = const()[name = string("op_1637_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1637_strides_0 = const()[name = string("op_1637_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1637_pad_0 = const()[name = string("op_1637_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1637_dilations_0 = const()[name = string("op_1637_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1637_groups_0 = const()[name = string("op_1637_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_2_weight_to_fp16 = const()[name = string("q_proj_2_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357841984)))];
            tensor<fp16, [1, 1024, 1, ?]> var_1637_cast_fp16 = conv(dilations = var_1637_dilations_0, groups = var_1637_groups_0, pad = var_1637_pad_0, pad_type = var_1637_pad_type_0, strides = var_1637_strides_0, weight = q_proj_2_weight_to_fp16, x = input_17_cast_fp16)[name = string("op_1637_cast_fp16")];
            string var_1648_pad_type_0 = const()[name = string("op_1648_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1648_strides_0 = const()[name = string("op_1648_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1648_pad_0 = const()[name = string("op_1648_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1648_dilations_0 = const()[name = string("op_1648_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1648_groups_0 = const()[name = string("op_1648_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_2_weight_to_fp16 = const()[name = string("k_proj_2_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359152768)))];
            tensor<fp16, [1, 256, 1, ?]> var_1648_cast_fp16 = conv(dilations = var_1648_dilations_0, groups = var_1648_groups_0, pad = var_1648_pad_0, pad_type = var_1648_pad_type_0, strides = var_1648_strides_0, weight = k_proj_2_weight_to_fp16, x = input_17_cast_fp16)[name = string("op_1648_cast_fp16")];
            string var_1659_pad_type_0 = const()[name = string("op_1659_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1659_strides_0 = const()[name = string("op_1659_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1659_pad_0 = const()[name = string("op_1659_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1659_dilations_0 = const()[name = string("op_1659_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1659_groups_0 = const()[name = string("op_1659_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_2_weight_to_fp16 = const()[name = string("v_proj_2_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359480512)))];
            tensor<fp16, [1, 256, 1, ?]> var_1659_cast_fp16 = conv(dilations = var_1659_dilations_0, groups = var_1659_groups_0, pad = var_1659_pad_0, pad_type = var_1659_pad_type_0, strides = var_1659_strides_0, weight = v_proj_2_weight_to_fp16, x = input_17_cast_fp16)[name = string("op_1659_cast_fp16")];
            tensor<int32, [4]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_1662_axis_0 = const()[name = string("op_1662_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_1662_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_1662_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_1662_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_1662_cast_fp16_3 = split(axis = var_1662_axis_0, split_sizes = tile_6, x = var_1637_cast_fp16)[name = string("op_1662_cast_fp16")];
            fp16 var_1675_promoted_to_fp16 = const()[name = string("op_1675_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_1676_cast_fp16 = pow(x = var_1662_cast_fp16_0, y = var_1675_promoted_to_fp16)[name = string("op_1676_cast_fp16")];
            tensor<int32, [1]> var_39_axes_0 = const()[name = string("var_39_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_39_keep_dims_0 = const()[name = string("var_39_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_39_cast_fp16 = reduce_mean(axes = var_39_axes_0, keep_dims = var_39_keep_dims_0, x = var_1676_cast_fp16)[name = string("var_39_cast_fp16")];
            fp16 var_1683_to_fp16 = const()[name = string("op_1683_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_1684_cast_fp16 = add(x = var_39_cast_fp16, y = var_1683_to_fp16)[name = string("op_1684_cast_fp16")];
            fp32 var_1685_epsilon_0 = const()[name = string("op_1685_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_1685_cast_fp16 = rsqrt(epsilon = var_1685_epsilon_0, x = var_1684_cast_fp16)[name = string("op_1685_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_39_cast_fp16 = mul(x = var_1662_cast_fp16_0, y = var_1685_cast_fp16)[name = string("x_normed_39_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_1695_to_fp16 = const()[name = string("op_1695_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359808256)))];
            tensor<fp16, [1, 256, 1, ?]> x_61_cast_fp16 = mul(x = x_normed_39_cast_fp16, y = var_1695_to_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> x_first_21_begin_0 = const()[name = string("x_first_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_21_end_0 = const()[name = string("x_first_21_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_21_end_mask_0 = const()[name = string("x_first_21_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_21_cast_fp16 = slice_by_index(begin = x_first_21_begin_0, end = x_first_21_end_0, end_mask = x_first_21_end_mask_0, x = x_61_cast_fp16)[name = string("x_first_21_cast_fp16")];
            tensor<int32, [4]> x_second_21_begin_0 = const()[name = string("x_second_21_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_21_end_0 = const()[name = string("x_second_21_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_21_end_mask_0 = const()[name = string("x_second_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_21_cast_fp16 = slice_by_index(begin = x_second_21_begin_0, end = x_second_21_end_0, end_mask = x_second_21_end_mask_0, x = x_61_cast_fp16)[name = string("x_second_21_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_1717_cast_fp16 = mul(x = x_second_21_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_1717_cast_fp16")];
            int32 var_1719 = const()[name = string("op_1719"), val = int32(1)];
            bool x_rot_21_interleave_0 = const()[name = string("x_rot_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_21_cast_fp16 = concat(axis = var_1719, interleave = x_rot_21_interleave_0, values = (var_1717_cast_fp16, x_first_21_cast_fp16))[name = string("x_rot_21_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1735_cast_fp16 = mul(x = x_61_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_1735_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1736_cast_fp16 = mul(x = x_rot_21_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_1736_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_17_cast_fp16 = add(x = var_1735_cast_fp16, y = var_1736_cast_fp16)[name = string("q_h_17_cast_fp16")];
            fp16 var_1739_promoted_to_fp16 = const()[name = string("op_1739_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_1740_cast_fp16 = pow(x = var_1662_cast_fp16_1, y = var_1739_promoted_to_fp16)[name = string("op_1740_cast_fp16")];
            tensor<int32, [1]> var_41_axes_0 = const()[name = string("var_41_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_41_keep_dims_0 = const()[name = string("var_41_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_41_cast_fp16 = reduce_mean(axes = var_41_axes_0, keep_dims = var_41_keep_dims_0, x = var_1740_cast_fp16)[name = string("var_41_cast_fp16")];
            fp16 var_1747_to_fp16 = const()[name = string("op_1747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_1748_cast_fp16 = add(x = var_41_cast_fp16, y = var_1747_to_fp16)[name = string("op_1748_cast_fp16")];
            fp32 var_1749_epsilon_0 = const()[name = string("op_1749_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_1749_cast_fp16 = rsqrt(epsilon = var_1749_epsilon_0, x = var_1748_cast_fp16)[name = string("op_1749_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_41_cast_fp16 = mul(x = var_1662_cast_fp16_1, y = var_1749_cast_fp16)[name = string("x_normed_41_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_65_cast_fp16 = mul(x = x_normed_41_cast_fp16, y = var_1695_to_fp16)[name = string("x_65_cast_fp16")];
            tensor<int32, [4]> x_first_23_begin_0 = const()[name = string("x_first_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_23_end_0 = const()[name = string("x_first_23_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_23_end_mask_0 = const()[name = string("x_first_23_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_23_cast_fp16 = slice_by_index(begin = x_first_23_begin_0, end = x_first_23_end_0, end_mask = x_first_23_end_mask_0, x = x_65_cast_fp16)[name = string("x_first_23_cast_fp16")];
            tensor<int32, [4]> x_second_23_begin_0 = const()[name = string("x_second_23_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_23_end_0 = const()[name = string("x_second_23_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_23_end_mask_0 = const()[name = string("x_second_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_23_cast_fp16 = slice_by_index(begin = x_second_23_begin_0, end = x_second_23_end_0, end_mask = x_second_23_end_mask_0, x = x_65_cast_fp16)[name = string("x_second_23_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_1781_cast_fp16 = mul(x = x_second_23_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_1781_cast_fp16")];
            int32 var_1783 = const()[name = string("op_1783"), val = int32(1)];
            bool x_rot_23_interleave_0 = const()[name = string("x_rot_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_23_cast_fp16 = concat(axis = var_1783, interleave = x_rot_23_interleave_0, values = (var_1781_cast_fp16, x_first_23_cast_fp16))[name = string("x_rot_23_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1799_cast_fp16 = mul(x = x_65_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_1799_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1800_cast_fp16 = mul(x = x_rot_23_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_1800_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_19_cast_fp16 = add(x = var_1799_cast_fp16, y = var_1800_cast_fp16)[name = string("q_h_19_cast_fp16")];
            fp16 var_1803_promoted_to_fp16 = const()[name = string("op_1803_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_1804_cast_fp16 = pow(x = var_1662_cast_fp16_2, y = var_1803_promoted_to_fp16)[name = string("op_1804_cast_fp16")];
            tensor<int32, [1]> var_43_axes_0 = const()[name = string("var_43_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_43_keep_dims_0 = const()[name = string("var_43_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_43_cast_fp16 = reduce_mean(axes = var_43_axes_0, keep_dims = var_43_keep_dims_0, x = var_1804_cast_fp16)[name = string("var_43_cast_fp16")];
            fp16 var_1811_to_fp16 = const()[name = string("op_1811_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_1812_cast_fp16 = add(x = var_43_cast_fp16, y = var_1811_to_fp16)[name = string("op_1812_cast_fp16")];
            fp32 var_1813_epsilon_0 = const()[name = string("op_1813_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_1813_cast_fp16 = rsqrt(epsilon = var_1813_epsilon_0, x = var_1812_cast_fp16)[name = string("op_1813_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_43_cast_fp16 = mul(x = var_1662_cast_fp16_2, y = var_1813_cast_fp16)[name = string("x_normed_43_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_69_cast_fp16 = mul(x = x_normed_43_cast_fp16, y = var_1695_to_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [4]> x_first_25_begin_0 = const()[name = string("x_first_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_25_end_0 = const()[name = string("x_first_25_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_25_end_mask_0 = const()[name = string("x_first_25_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_25_cast_fp16 = slice_by_index(begin = x_first_25_begin_0, end = x_first_25_end_0, end_mask = x_first_25_end_mask_0, x = x_69_cast_fp16)[name = string("x_first_25_cast_fp16")];
            tensor<int32, [4]> x_second_25_begin_0 = const()[name = string("x_second_25_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_25_end_0 = const()[name = string("x_second_25_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_25_end_mask_0 = const()[name = string("x_second_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_25_cast_fp16 = slice_by_index(begin = x_second_25_begin_0, end = x_second_25_end_0, end_mask = x_second_25_end_mask_0, x = x_69_cast_fp16)[name = string("x_second_25_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_1845_cast_fp16 = mul(x = x_second_25_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1845_cast_fp16")];
            int32 var_1847 = const()[name = string("op_1847"), val = int32(1)];
            bool x_rot_25_interleave_0 = const()[name = string("x_rot_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_25_cast_fp16 = concat(axis = var_1847, interleave = x_rot_25_interleave_0, values = (var_1845_cast_fp16, x_first_25_cast_fp16))[name = string("x_rot_25_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1863_cast_fp16 = mul(x = x_69_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_1863_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1864_cast_fp16 = mul(x = x_rot_25_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_1864_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_21_cast_fp16 = add(x = var_1863_cast_fp16, y = var_1864_cast_fp16)[name = string("q_h_21_cast_fp16")];
            fp16 var_1867_promoted_to_fp16 = const()[name = string("op_1867_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_1868_cast_fp16 = pow(x = var_1662_cast_fp16_3, y = var_1867_promoted_to_fp16)[name = string("op_1868_cast_fp16")];
            tensor<int32, [1]> var_45_axes_0 = const()[name = string("var_45_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_45_keep_dims_0 = const()[name = string("var_45_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_45_cast_fp16 = reduce_mean(axes = var_45_axes_0, keep_dims = var_45_keep_dims_0, x = var_1868_cast_fp16)[name = string("var_45_cast_fp16")];
            fp16 var_1875_to_fp16 = const()[name = string("op_1875_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_1876_cast_fp16 = add(x = var_45_cast_fp16, y = var_1875_to_fp16)[name = string("op_1876_cast_fp16")];
            fp32 var_1877_epsilon_0 = const()[name = string("op_1877_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_1877_cast_fp16 = rsqrt(epsilon = var_1877_epsilon_0, x = var_1876_cast_fp16)[name = string("op_1877_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_45_cast_fp16 = mul(x = var_1662_cast_fp16_3, y = var_1877_cast_fp16)[name = string("x_normed_45_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_73_cast_fp16 = mul(x = x_normed_45_cast_fp16, y = var_1695_to_fp16)[name = string("x_73_cast_fp16")];
            tensor<int32, [4]> x_first_27_begin_0 = const()[name = string("x_first_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_27_end_0 = const()[name = string("x_first_27_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_27_end_mask_0 = const()[name = string("x_first_27_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_27_cast_fp16 = slice_by_index(begin = x_first_27_begin_0, end = x_first_27_end_0, end_mask = x_first_27_end_mask_0, x = x_73_cast_fp16)[name = string("x_first_27_cast_fp16")];
            tensor<int32, [4]> x_second_27_begin_0 = const()[name = string("x_second_27_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_27_end_0 = const()[name = string("x_second_27_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_27_end_mask_0 = const()[name = string("x_second_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_27_cast_fp16 = slice_by_index(begin = x_second_27_begin_0, end = x_second_27_end_0, end_mask = x_second_27_end_mask_0, x = x_73_cast_fp16)[name = string("x_second_27_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_1909_cast_fp16 = mul(x = x_second_27_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_1909_cast_fp16")];
            int32 var_1911 = const()[name = string("op_1911"), val = int32(1)];
            bool x_rot_27_interleave_0 = const()[name = string("x_rot_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_27_cast_fp16 = concat(axis = var_1911, interleave = x_rot_27_interleave_0, values = (var_1909_cast_fp16, x_first_27_cast_fp16))[name = string("x_rot_27_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1927_cast_fp16 = mul(x = x_73_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_1927_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1928_cast_fp16 = mul(x = x_rot_27_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_1928_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_23_cast_fp16 = add(x = var_1927_cast_fp16, y = var_1928_cast_fp16)[name = string("q_h_23_cast_fp16")];
            fp16 var_1931_promoted_to_fp16 = const()[name = string("op_1931_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_1932_cast_fp16 = pow(x = var_1648_cast_fp16, y = var_1931_promoted_to_fp16)[name = string("op_1932_cast_fp16")];
            tensor<int32, [1]> var_47_axes_0 = const()[name = string("var_47_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_47_keep_dims_0 = const()[name = string("var_47_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_47_cast_fp16 = reduce_mean(axes = var_47_axes_0, keep_dims = var_47_keep_dims_0, x = var_1932_cast_fp16)[name = string("var_47_cast_fp16")];
            fp16 var_1939_to_fp16 = const()[name = string("op_1939_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_1940_cast_fp16 = add(x = var_47_cast_fp16, y = var_1939_to_fp16)[name = string("op_1940_cast_fp16")];
            fp32 var_1941_epsilon_0 = const()[name = string("op_1941_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_1941_cast_fp16 = rsqrt(epsilon = var_1941_epsilon_0, x = var_1940_cast_fp16)[name = string("op_1941_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_47_cast_fp16 = mul(x = var_1648_cast_fp16, y = var_1941_cast_fp16)[name = string("x_normed_47_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_1951_to_fp16 = const()[name = string("op_1951_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359808832)))];
            tensor<fp16, [1, 256, 1, ?]> x_77_cast_fp16 = mul(x = x_normed_47_cast_fp16, y = var_1951_to_fp16)[name = string("x_77_cast_fp16")];
            tensor<int32, [4]> x_first_29_begin_0 = const()[name = string("x_first_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_29_end_0 = const()[name = string("x_first_29_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_29_end_mask_0 = const()[name = string("x_first_29_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_29_cast_fp16 = slice_by_index(begin = x_first_29_begin_0, end = x_first_29_end_0, end_mask = x_first_29_end_mask_0, x = x_77_cast_fp16)[name = string("x_first_29_cast_fp16")];
            tensor<int32, [4]> x_second_29_begin_0 = const()[name = string("x_second_29_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_29_end_0 = const()[name = string("x_second_29_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_29_end_mask_0 = const()[name = string("x_second_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_29_cast_fp16 = slice_by_index(begin = x_second_29_begin_0, end = x_second_29_end_0, end_mask = x_second_29_end_mask_0, x = x_77_cast_fp16)[name = string("x_second_29_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_1973_cast_fp16 = mul(x = x_second_29_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_1973_cast_fp16")];
            int32 var_1975 = const()[name = string("op_1975"), val = int32(1)];
            bool x_rot_29_interleave_0 = const()[name = string("x_rot_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_29_cast_fp16 = concat(axis = var_1975, interleave = x_rot_29_interleave_0, values = (var_1973_cast_fp16, x_first_29_cast_fp16))[name = string("x_rot_29_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1991_cast_fp16 = mul(x = x_77_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_1991_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_1992_cast_fp16 = mul(x = x_rot_29_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_1992_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_5_cast_fp16 = add(x = var_1991_cast_fp16, y = var_1992_cast_fp16)[name = string("k_new_5_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_4 = read_state(input = k_cache_2)[name = string("read_state_4")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_15_cast_fp16 = mul(x = read_state_4, y = one_minus_1_cast_fp16)[name = string("k_cache_15_cast_fp16")];
            write_state(data = k_cache_15_cast_fp16, input = k_cache_2)[name = string("coreml_update_state_80_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_80 = read_state(input = k_cache_2)[name = string("coreml_update_state_80")];
            tensor<fp16, [1, 256, 1, 128]> var_1999_cast_fp16 = mul(x = k_new_5_cast_fp16, y = write_mask)[name = string("op_1999_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_17_cast_fp16 = add(x = coreml_update_state_80, y = var_1999_cast_fp16)[name = string("k_cache_17_cast_fp16")];
            write_state(data = k_cache_17_cast_fp16, input = k_cache_2)[name = string("coreml_update_state_81_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_81 = read_state(input = k_cache_2)[name = string("coreml_update_state_81")];
            tensor<fp16, [1, 256, 1, 128]> read_state_5 = read_state(input = v_cache_2)[name = string("read_state_5")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_15_cast_fp16 = mul(x = read_state_5, y = one_minus_1_cast_fp16)[name = string("v_cache_15_cast_fp16")];
            write_state(data = v_cache_15_cast_fp16, input = v_cache_2)[name = string("coreml_update_state_82_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_82 = read_state(input = v_cache_2)[name = string("coreml_update_state_82")];
            tensor<fp16, [1, 256, 1, 128]> var_2003_cast_fp16 = mul(x = var_1659_cast_fp16, y = write_mask)[name = string("op_2003_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_17_cast_fp16 = add(x = coreml_update_state_82, y = var_2003_cast_fp16)[name = string("v_cache_17_cast_fp16")];
            write_state(data = v_cache_17_cast_fp16, input = v_cache_2)[name = string("coreml_update_state_83_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_83 = read_state(input = v_cache_2)[name = string("coreml_update_state_83")];
            bool var_2018_transpose_x_0 = const()[name = string("op_2018_transpose_x_0"), val = bool(false)];
            bool var_2018_transpose_y_0 = const()[name = string("op_2018_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_322_perm_0 = const()[name = string("transpose_322_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_323_perm_0 = const()[name = string("transpose_323_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_323 = transpose(perm = transpose_323_perm_0, x = coreml_update_state_81)[name = string("transpose_562")];
            tensor<fp16, [1, 1, ?, 256]> transpose_322 = transpose(perm = transpose_322_perm_0, x = q_h_17_cast_fp16)[name = string("transpose_563")];
            tensor<fp16, [1, 1, ?, 128]> var_2018_cast_fp16 = matmul(transpose_x = var_2018_transpose_x_0, transpose_y = var_2018_transpose_y_0, x = transpose_322, y = transpose_323)[name = string("op_2018_cast_fp16")];
            fp16 var_2019_to_fp16 = const()[name = string("op_2019_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_2020_cast_fp16 = mul(x = var_2018_cast_fp16, y = var_2019_to_fp16)[name = string("op_2020_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_17_cast_fp16 = add(x = var_2020_cast_fp16, y = attention_mask)[name = string("scores_17_cast_fp16")];
            int32 var_2023 = const()[name = string("op_2023"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_17_cast_fp16 = softmax(axis = var_2023, x = scores_17_cast_fp16)[name = string("probs_17_cast_fp16")];
            tensor<int32, [4]> var_2030 = const()[name = string("op_2030"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_2032_transpose_x_0 = const()[name = string("op_2032_transpose_x_0"), val = bool(false)];
            bool var_2032_transpose_y_0 = const()[name = string("op_2032_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_17_cast_fp16 = transpose(perm = var_2030, x = coreml_update_state_83)[name = string("transpose_561")];
            tensor<fp16, [1, 1, ?, 256]> var_2032_cast_fp16 = matmul(transpose_x = var_2032_transpose_x_0, transpose_y = var_2032_transpose_y_0, x = probs_17_cast_fp16, y = v_t_17_cast_fp16)[name = string("op_2032_cast_fp16")];
            bool var_2051_transpose_x_0 = const()[name = string("op_2051_transpose_x_0"), val = bool(false)];
            bool var_2051_transpose_y_0 = const()[name = string("op_2051_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_327_perm_0 = const()[name = string("transpose_327_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_327 = transpose(perm = transpose_327_perm_0, x = q_h_19_cast_fp16)[name = string("transpose_560")];
            tensor<fp16, [1, 1, ?, 128]> var_2051_cast_fp16 = matmul(transpose_x = var_2051_transpose_x_0, transpose_y = var_2051_transpose_y_0, x = transpose_327, y = transpose_323)[name = string("op_2051_cast_fp16")];
            fp16 var_2052_to_fp16 = const()[name = string("op_2052_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_2053_cast_fp16 = mul(x = var_2051_cast_fp16, y = var_2052_to_fp16)[name = string("op_2053_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_19_cast_fp16 = add(x = var_2053_cast_fp16, y = attention_mask)[name = string("scores_19_cast_fp16")];
            int32 var_2056 = const()[name = string("op_2056"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_19_cast_fp16 = softmax(axis = var_2056, x = scores_19_cast_fp16)[name = string("probs_19_cast_fp16")];
            bool var_2065_transpose_x_0 = const()[name = string("op_2065_transpose_x_0"), val = bool(false)];
            bool var_2065_transpose_y_0 = const()[name = string("op_2065_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_2065_cast_fp16 = matmul(transpose_x = var_2065_transpose_x_0, transpose_y = var_2065_transpose_y_0, x = probs_19_cast_fp16, y = v_t_17_cast_fp16)[name = string("op_2065_cast_fp16")];
            bool var_2084_transpose_x_0 = const()[name = string("op_2084_transpose_x_0"), val = bool(false)];
            bool var_2084_transpose_y_0 = const()[name = string("op_2084_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_328_perm_0 = const()[name = string("transpose_328_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_328 = transpose(perm = transpose_328_perm_0, x = q_h_21_cast_fp16)[name = string("transpose_559")];
            tensor<fp16, [1, 1, ?, 128]> var_2084_cast_fp16 = matmul(transpose_x = var_2084_transpose_x_0, transpose_y = var_2084_transpose_y_0, x = transpose_328, y = transpose_323)[name = string("op_2084_cast_fp16")];
            fp16 var_2085_to_fp16 = const()[name = string("op_2085_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_2086_cast_fp16 = mul(x = var_2084_cast_fp16, y = var_2085_to_fp16)[name = string("op_2086_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_21_cast_fp16 = add(x = var_2086_cast_fp16, y = attention_mask)[name = string("scores_21_cast_fp16")];
            int32 var_2089 = const()[name = string("op_2089"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_21_cast_fp16 = softmax(axis = var_2089, x = scores_21_cast_fp16)[name = string("probs_21_cast_fp16")];
            bool var_2098_transpose_x_0 = const()[name = string("op_2098_transpose_x_0"), val = bool(false)];
            bool var_2098_transpose_y_0 = const()[name = string("op_2098_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_2098_cast_fp16 = matmul(transpose_x = var_2098_transpose_x_0, transpose_y = var_2098_transpose_y_0, x = probs_21_cast_fp16, y = v_t_17_cast_fp16)[name = string("op_2098_cast_fp16")];
            bool var_2117_transpose_x_0 = const()[name = string("op_2117_transpose_x_0"), val = bool(false)];
            bool var_2117_transpose_y_0 = const()[name = string("op_2117_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_329_perm_0 = const()[name = string("transpose_329_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_329 = transpose(perm = transpose_329_perm_0, x = q_h_23_cast_fp16)[name = string("transpose_558")];
            tensor<fp16, [1, 1, ?, 128]> var_2117_cast_fp16 = matmul(transpose_x = var_2117_transpose_x_0, transpose_y = var_2117_transpose_y_0, x = transpose_329, y = transpose_323)[name = string("op_2117_cast_fp16")];
            fp16 var_2118_to_fp16 = const()[name = string("op_2118_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_2119_cast_fp16 = mul(x = var_2117_cast_fp16, y = var_2118_to_fp16)[name = string("op_2119_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_23_cast_fp16 = add(x = var_2119_cast_fp16, y = attention_mask)[name = string("scores_23_cast_fp16")];
            int32 var_2122 = const()[name = string("op_2122"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_23_cast_fp16 = softmax(axis = var_2122, x = scores_23_cast_fp16)[name = string("probs_23_cast_fp16")];
            bool var_2131_transpose_x_0 = const()[name = string("op_2131_transpose_x_0"), val = bool(false)];
            bool var_2131_transpose_y_0 = const()[name = string("op_2131_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_2131_cast_fp16 = matmul(transpose_x = var_2131_transpose_x_0, transpose_y = var_2131_transpose_y_0, x = probs_23_cast_fp16, y = v_t_17_cast_fp16)[name = string("op_2131_cast_fp16")];
            bool input_19_interleave_0 = const()[name = string("input_19_interleave_0"), val = bool(false)];
            int32 const_147 = const()[name = string("const_147"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_19_cast_fp16 = concat(axis = const_147, interleave = input_19_interleave_0, values = (var_2032_cast_fp16, var_2065_cast_fp16, var_2098_cast_fp16, var_2131_cast_fp16))[name = string("input_19_cast_fp16")];
            string x_79_pad_type_0 = const()[name = string("x_79_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_79_strides_0 = const()[name = string("x_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_79_pad_0 = const()[name = string("x_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_79_dilations_0 = const()[name = string("x_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_79_groups_0 = const()[name = string("x_79_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_26_perm_0 = const()[name = string("transpose_26_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_2_weight_to_fp16 = const()[name = string("o_proj_2_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359809408)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_26_cast_fp16 = transpose(perm = transpose_26_perm_0, x = input_19_cast_fp16)[name = string("transpose_557")];
            tensor<fp16, [1, 640, 1, ?]> x_79_cast_fp16 = conv(dilations = x_79_dilations_0, groups = x_79_groups_0, pad = x_79_pad_0, pad_type = x_79_pad_type_0, strides = x_79_strides_0, weight = o_proj_2_weight_to_fp16, x = transpose_26_cast_fp16)[name = string("x_79_cast_fp16")];
            fp16 var_2152_promoted_to_fp16 = const()[name = string("op_2152_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_2153_cast_fp16 = pow(x = x_79_cast_fp16, y = var_2152_promoted_to_fp16)[name = string("op_2153_cast_fp16")];
            tensor<int32, [1]> var_49_axes_0 = const()[name = string("var_49_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_49_keep_dims_0 = const()[name = string("var_49_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_49_cast_fp16 = reduce_mean(axes = var_49_axes_0, keep_dims = var_49_keep_dims_0, x = var_2153_cast_fp16)[name = string("var_49_cast_fp16")];
            fp16 var_2160_to_fp16 = const()[name = string("op_2160_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_2161_cast_fp16 = add(x = var_49_cast_fp16, y = var_2160_to_fp16)[name = string("op_2161_cast_fp16")];
            fp32 var_2162_epsilon_0 = const()[name = string("op_2162_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_2162_cast_fp16 = rsqrt(epsilon = var_2162_epsilon_0, x = var_2161_cast_fp16)[name = string("op_2162_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_49_cast_fp16 = mul(x = x_79_cast_fp16, y = var_2162_cast_fp16)[name = string("x_normed_49_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_2172_to_fp16 = const()[name = string("op_2172_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361120192)))];
            tensor<fp16, [1, 640, 1, ?]> attn_11_cast_fp16 = mul(x = x_normed_49_cast_fp16, y = var_2172_to_fp16)[name = string("attn_11_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_81_cast_fp16 = add(x = x_57_cast_fp16, y = attn_11_cast_fp16)[name = string("x_81_cast_fp16")];
            fp16 var_2176_promoted_to_fp16 = const()[name = string("op_2176_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_2177_cast_fp16 = pow(x = x_81_cast_fp16, y = var_2176_promoted_to_fp16)[name = string("op_2177_cast_fp16")];
            tensor<int32, [1]> var_51_axes_0 = const()[name = string("var_51_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_51_keep_dims_0 = const()[name = string("var_51_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_51_cast_fp16 = reduce_mean(axes = var_51_axes_0, keep_dims = var_51_keep_dims_0, x = var_2177_cast_fp16)[name = string("var_51_cast_fp16")];
            fp16 var_2184_to_fp16 = const()[name = string("op_2184_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_2185_cast_fp16 = add(x = var_51_cast_fp16, y = var_2184_to_fp16)[name = string("op_2185_cast_fp16")];
            fp32 var_2186_epsilon_0 = const()[name = string("op_2186_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_2186_cast_fp16 = rsqrt(epsilon = var_2186_epsilon_0, x = var_2185_cast_fp16)[name = string("op_2186_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_51_cast_fp16 = mul(x = x_81_cast_fp16, y = var_2186_cast_fp16)[name = string("x_normed_51_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_2196_to_fp16 = const()[name = string("op_2196_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361121536)))];
            tensor<fp16, [1, 640, 1, ?]> input_21_cast_fp16 = mul(x = x_normed_51_cast_fp16, y = var_2196_to_fp16)[name = string("input_21_cast_fp16")];
            string gate_5_pad_type_0 = const()[name = string("gate_5_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_5_strides_0 = const()[name = string("gate_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_5_pad_0 = const()[name = string("gate_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_5_dilations_0 = const()[name = string("gate_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_5_groups_0 = const()[name = string("gate_5_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_2_weight_to_fp16 = const()[name = string("gate_proj_2_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361122880)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_5_cast_fp16 = conv(dilations = gate_5_dilations_0, groups = gate_5_groups_0, pad = gate_5_pad_0, pad_type = gate_5_pad_type_0, strides = gate_5_strides_0, weight = gate_proj_2_weight_to_fp16, x = input_21_cast_fp16)[name = string("gate_5_cast_fp16")];
            string up_5_pad_type_0 = const()[name = string("up_5_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_5_strides_0 = const()[name = string("up_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_5_pad_0 = const()[name = string("up_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_5_dilations_0 = const()[name = string("up_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_5_groups_0 = const()[name = string("up_5_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_2_weight_to_fp16 = const()[name = string("up_proj_2_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363744384)))];
            tensor<fp16, [1, 2048, 1, ?]> up_5_cast_fp16 = conv(dilations = up_5_dilations_0, groups = up_5_groups_0, pad = up_5_pad_0, pad_type = up_5_pad_type_0, strides = up_5_strides_0, weight = up_proj_2_weight_to_fp16, x = input_21_cast_fp16)[name = string("up_5_cast_fp16")];
            string var_2221_mode_0 = const()[name = string("op_2221_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_2221_cast_fp16 = gelu(mode = var_2221_mode_0, x = gate_5_cast_fp16)[name = string("op_2221_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_23_cast_fp16 = mul(x = var_2221_cast_fp16, y = up_5_cast_fp16)[name = string("input_23_cast_fp16")];
            string x_83_pad_type_0 = const()[name = string("x_83_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_83_strides_0 = const()[name = string("x_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_83_pad_0 = const()[name = string("x_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_83_dilations_0 = const()[name = string("x_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_83_groups_0 = const()[name = string("x_83_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_2_weight_to_fp16 = const()[name = string("down_proj_2_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366365888)))];
            tensor<fp16, [1, 640, 1, ?]> x_83_cast_fp16 = conv(dilations = x_83_dilations_0, groups = x_83_groups_0, pad = x_83_pad_0, pad_type = x_83_pad_type_0, strides = x_83_strides_0, weight = down_proj_2_weight_to_fp16, x = input_23_cast_fp16)[name = string("x_83_cast_fp16")];
            fp16 var_2234_promoted_to_fp16 = const()[name = string("op_2234_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_2235_cast_fp16 = pow(x = x_83_cast_fp16, y = var_2234_promoted_to_fp16)[name = string("op_2235_cast_fp16")];
            tensor<int32, [1]> var_53_axes_0 = const()[name = string("var_53_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_53_keep_dims_0 = const()[name = string("var_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_53_cast_fp16 = reduce_mean(axes = var_53_axes_0, keep_dims = var_53_keep_dims_0, x = var_2235_cast_fp16)[name = string("var_53_cast_fp16")];
            fp16 var_2242_to_fp16 = const()[name = string("op_2242_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_2243_cast_fp16 = add(x = var_53_cast_fp16, y = var_2242_to_fp16)[name = string("op_2243_cast_fp16")];
            fp32 var_2244_epsilon_0 = const()[name = string("op_2244_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_2244_cast_fp16 = rsqrt(epsilon = var_2244_epsilon_0, x = var_2243_cast_fp16)[name = string("op_2244_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_53_cast_fp16 = mul(x = x_83_cast_fp16, y = var_2244_cast_fp16)[name = string("x_normed_53_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_2254_to_fp16 = const()[name = string("op_2254_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368987392)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_5_cast_fp16 = mul(x = x_normed_53_cast_fp16, y = var_2254_to_fp16)[name = string("mlp_5_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_85_cast_fp16 = add(x = x_81_cast_fp16, y = mlp_5_cast_fp16)[name = string("x_85_cast_fp16")];
            fp16 var_2258_promoted_to_fp16 = const()[name = string("op_2258_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_2259_cast_fp16 = pow(x = x_85_cast_fp16, y = var_2258_promoted_to_fp16)[name = string("op_2259_cast_fp16")];
            tensor<int32, [1]> var_55_axes_0 = const()[name = string("var_55_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_55_keep_dims_0 = const()[name = string("var_55_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_55_cast_fp16 = reduce_mean(axes = var_55_axes_0, keep_dims = var_55_keep_dims_0, x = var_2259_cast_fp16)[name = string("var_55_cast_fp16")];
            fp16 var_2266_to_fp16 = const()[name = string("op_2266_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_2267_cast_fp16 = add(x = var_55_cast_fp16, y = var_2266_to_fp16)[name = string("op_2267_cast_fp16")];
            fp32 var_2268_epsilon_0 = const()[name = string("op_2268_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_2268_cast_fp16 = rsqrt(epsilon = var_2268_epsilon_0, x = var_2267_cast_fp16)[name = string("op_2268_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_55_cast_fp16 = mul(x = x_85_cast_fp16, y = var_2268_cast_fp16)[name = string("x_normed_55_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_2278_to_fp16 = const()[name = string("op_2278_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368988736)))];
            tensor<fp16, [1, 640, 1, ?]> input_25_cast_fp16 = mul(x = x_normed_55_cast_fp16, y = var_2278_to_fp16)[name = string("input_25_cast_fp16")];
            string var_2290_pad_type_0 = const()[name = string("op_2290_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2290_strides_0 = const()[name = string("op_2290_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2290_pad_0 = const()[name = string("op_2290_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2290_dilations_0 = const()[name = string("op_2290_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2290_groups_0 = const()[name = string("op_2290_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_3_weight_to_fp16 = const()[name = string("q_proj_3_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368990080)))];
            tensor<fp16, [1, 1024, 1, ?]> var_2290_cast_fp16 = conv(dilations = var_2290_dilations_0, groups = var_2290_groups_0, pad = var_2290_pad_0, pad_type = var_2290_pad_type_0, strides = var_2290_strides_0, weight = q_proj_3_weight_to_fp16, x = input_25_cast_fp16)[name = string("op_2290_cast_fp16")];
            string var_2301_pad_type_0 = const()[name = string("op_2301_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2301_strides_0 = const()[name = string("op_2301_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2301_pad_0 = const()[name = string("op_2301_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2301_dilations_0 = const()[name = string("op_2301_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2301_groups_0 = const()[name = string("op_2301_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_3_weight_to_fp16 = const()[name = string("k_proj_3_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370300864)))];
            tensor<fp16, [1, 256, 1, ?]> var_2301_cast_fp16 = conv(dilations = var_2301_dilations_0, groups = var_2301_groups_0, pad = var_2301_pad_0, pad_type = var_2301_pad_type_0, strides = var_2301_strides_0, weight = k_proj_3_weight_to_fp16, x = input_25_cast_fp16)[name = string("op_2301_cast_fp16")];
            string var_2312_pad_type_0 = const()[name = string("op_2312_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2312_strides_0 = const()[name = string("op_2312_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2312_pad_0 = const()[name = string("op_2312_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2312_dilations_0 = const()[name = string("op_2312_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2312_groups_0 = const()[name = string("op_2312_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_3_weight_to_fp16 = const()[name = string("v_proj_3_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370628608)))];
            tensor<fp16, [1, 256, 1, ?]> var_2312_cast_fp16 = conv(dilations = var_2312_dilations_0, groups = var_2312_groups_0, pad = var_2312_pad_0, pad_type = var_2312_pad_type_0, strides = var_2312_strides_0, weight = v_proj_3_weight_to_fp16, x = input_25_cast_fp16)[name = string("op_2312_cast_fp16")];
            tensor<int32, [4]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_2315_axis_0 = const()[name = string("op_2315_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_2315_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_2315_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_2315_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_2315_cast_fp16_3 = split(axis = var_2315_axis_0, split_sizes = tile_9, x = var_2290_cast_fp16)[name = string("op_2315_cast_fp16")];
            fp16 var_2328_promoted_to_fp16 = const()[name = string("op_2328_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_2329_cast_fp16 = pow(x = var_2315_cast_fp16_0, y = var_2328_promoted_to_fp16)[name = string("op_2329_cast_fp16")];
            tensor<int32, [1]> var_57_axes_0 = const()[name = string("var_57_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_57_keep_dims_0 = const()[name = string("var_57_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_57_cast_fp16 = reduce_mean(axes = var_57_axes_0, keep_dims = var_57_keep_dims_0, x = var_2329_cast_fp16)[name = string("var_57_cast_fp16")];
            fp16 var_2336_to_fp16 = const()[name = string("op_2336_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_2337_cast_fp16 = add(x = var_57_cast_fp16, y = var_2336_to_fp16)[name = string("op_2337_cast_fp16")];
            fp32 var_2338_epsilon_0 = const()[name = string("op_2338_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_2338_cast_fp16 = rsqrt(epsilon = var_2338_epsilon_0, x = var_2337_cast_fp16)[name = string("op_2338_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_57_cast_fp16 = mul(x = var_2315_cast_fp16_0, y = var_2338_cast_fp16)[name = string("x_normed_57_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_2348_to_fp16 = const()[name = string("op_2348_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370956352)))];
            tensor<fp16, [1, 256, 1, ?]> x_89_cast_fp16 = mul(x = x_normed_57_cast_fp16, y = var_2348_to_fp16)[name = string("x_89_cast_fp16")];
            tensor<int32, [4]> x_first_31_begin_0 = const()[name = string("x_first_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_31_end_0 = const()[name = string("x_first_31_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_31_end_mask_0 = const()[name = string("x_first_31_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_31_cast_fp16 = slice_by_index(begin = x_first_31_begin_0, end = x_first_31_end_0, end_mask = x_first_31_end_mask_0, x = x_89_cast_fp16)[name = string("x_first_31_cast_fp16")];
            tensor<int32, [4]> x_second_31_begin_0 = const()[name = string("x_second_31_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_31_end_0 = const()[name = string("x_second_31_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_31_end_mask_0 = const()[name = string("x_second_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_31_cast_fp16 = slice_by_index(begin = x_second_31_begin_0, end = x_second_31_end_0, end_mask = x_second_31_end_mask_0, x = x_89_cast_fp16)[name = string("x_second_31_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_2370_cast_fp16 = mul(x = x_second_31_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_2370_cast_fp16")];
            int32 var_2372 = const()[name = string("op_2372"), val = int32(1)];
            bool x_rot_31_interleave_0 = const()[name = string("x_rot_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_31_cast_fp16 = concat(axis = var_2372, interleave = x_rot_31_interleave_0, values = (var_2370_cast_fp16, x_first_31_cast_fp16))[name = string("x_rot_31_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_2388_cast_fp16 = mul(x = x_89_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_2388_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_2389_cast_fp16 = mul(x = x_rot_31_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_2389_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_25_cast_fp16 = add(x = var_2388_cast_fp16, y = var_2389_cast_fp16)[name = string("q_h_25_cast_fp16")];
            fp16 var_2392_promoted_to_fp16 = const()[name = string("op_2392_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_2393_cast_fp16 = pow(x = var_2315_cast_fp16_1, y = var_2392_promoted_to_fp16)[name = string("op_2393_cast_fp16")];
            tensor<int32, [1]> var_59_axes_0 = const()[name = string("var_59_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_59_keep_dims_0 = const()[name = string("var_59_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_59_cast_fp16 = reduce_mean(axes = var_59_axes_0, keep_dims = var_59_keep_dims_0, x = var_2393_cast_fp16)[name = string("var_59_cast_fp16")];
            fp16 var_2400_to_fp16 = const()[name = string("op_2400_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_2401_cast_fp16 = add(x = var_59_cast_fp16, y = var_2400_to_fp16)[name = string("op_2401_cast_fp16")];
            fp32 var_2402_epsilon_0 = const()[name = string("op_2402_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_2402_cast_fp16 = rsqrt(epsilon = var_2402_epsilon_0, x = var_2401_cast_fp16)[name = string("op_2402_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_59_cast_fp16 = mul(x = var_2315_cast_fp16_1, y = var_2402_cast_fp16)[name = string("x_normed_59_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_93_cast_fp16 = mul(x = x_normed_59_cast_fp16, y = var_2348_to_fp16)[name = string("x_93_cast_fp16")];
            tensor<int32, [4]> x_first_33_begin_0 = const()[name = string("x_first_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_33_end_0 = const()[name = string("x_first_33_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_33_end_mask_0 = const()[name = string("x_first_33_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_33_cast_fp16 = slice_by_index(begin = x_first_33_begin_0, end = x_first_33_end_0, end_mask = x_first_33_end_mask_0, x = x_93_cast_fp16)[name = string("x_first_33_cast_fp16")];
            tensor<int32, [4]> x_second_33_begin_0 = const()[name = string("x_second_33_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_33_end_0 = const()[name = string("x_second_33_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_33_end_mask_0 = const()[name = string("x_second_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_33_cast_fp16 = slice_by_index(begin = x_second_33_begin_0, end = x_second_33_end_0, end_mask = x_second_33_end_mask_0, x = x_93_cast_fp16)[name = string("x_second_33_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_2434_cast_fp16 = mul(x = x_second_33_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_2434_cast_fp16")];
            int32 var_2436 = const()[name = string("op_2436"), val = int32(1)];
            bool x_rot_33_interleave_0 = const()[name = string("x_rot_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_33_cast_fp16 = concat(axis = var_2436, interleave = x_rot_33_interleave_0, values = (var_2434_cast_fp16, x_first_33_cast_fp16))[name = string("x_rot_33_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_2452_cast_fp16 = mul(x = x_93_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_2452_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_2453_cast_fp16 = mul(x = x_rot_33_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_2453_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_27_cast_fp16 = add(x = var_2452_cast_fp16, y = var_2453_cast_fp16)[name = string("q_h_27_cast_fp16")];
            fp16 var_2456_promoted_to_fp16 = const()[name = string("op_2456_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_2457_cast_fp16 = pow(x = var_2315_cast_fp16_2, y = var_2456_promoted_to_fp16)[name = string("op_2457_cast_fp16")];
            tensor<int32, [1]> var_61_axes_0 = const()[name = string("var_61_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_61_keep_dims_0 = const()[name = string("var_61_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_61_cast_fp16 = reduce_mean(axes = var_61_axes_0, keep_dims = var_61_keep_dims_0, x = var_2457_cast_fp16)[name = string("var_61_cast_fp16")];
            fp16 var_2464_to_fp16 = const()[name = string("op_2464_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_2465_cast_fp16 = add(x = var_61_cast_fp16, y = var_2464_to_fp16)[name = string("op_2465_cast_fp16")];
            fp32 var_2466_epsilon_0 = const()[name = string("op_2466_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_2466_cast_fp16 = rsqrt(epsilon = var_2466_epsilon_0, x = var_2465_cast_fp16)[name = string("op_2466_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_61_cast_fp16 = mul(x = var_2315_cast_fp16_2, y = var_2466_cast_fp16)[name = string("x_normed_61_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_97_cast_fp16 = mul(x = x_normed_61_cast_fp16, y = var_2348_to_fp16)[name = string("x_97_cast_fp16")];
            tensor<int32, [4]> x_first_35_begin_0 = const()[name = string("x_first_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_35_end_0 = const()[name = string("x_first_35_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_35_end_mask_0 = const()[name = string("x_first_35_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_35_cast_fp16 = slice_by_index(begin = x_first_35_begin_0, end = x_first_35_end_0, end_mask = x_first_35_end_mask_0, x = x_97_cast_fp16)[name = string("x_first_35_cast_fp16")];
            tensor<int32, [4]> x_second_35_begin_0 = const()[name = string("x_second_35_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_35_end_0 = const()[name = string("x_second_35_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_35_end_mask_0 = const()[name = string("x_second_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_35_cast_fp16 = slice_by_index(begin = x_second_35_begin_0, end = x_second_35_end_0, end_mask = x_second_35_end_mask_0, x = x_97_cast_fp16)[name = string("x_second_35_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_2498_cast_fp16 = mul(x = x_second_35_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_2498_cast_fp16")];
            int32 var_2500 = const()[name = string("op_2500"), val = int32(1)];
            bool x_rot_35_interleave_0 = const()[name = string("x_rot_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_35_cast_fp16 = concat(axis = var_2500, interleave = x_rot_35_interleave_0, values = (var_2498_cast_fp16, x_first_35_cast_fp16))[name = string("x_rot_35_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_2516_cast_fp16 = mul(x = x_97_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_2516_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_2517_cast_fp16 = mul(x = x_rot_35_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_2517_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_29_cast_fp16 = add(x = var_2516_cast_fp16, y = var_2517_cast_fp16)[name = string("q_h_29_cast_fp16")];
            fp16 var_2520_promoted_to_fp16 = const()[name = string("op_2520_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_2521_cast_fp16 = pow(x = var_2315_cast_fp16_3, y = var_2520_promoted_to_fp16)[name = string("op_2521_cast_fp16")];
            tensor<int32, [1]> var_63_axes_0 = const()[name = string("var_63_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_63_keep_dims_0 = const()[name = string("var_63_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_63_cast_fp16 = reduce_mean(axes = var_63_axes_0, keep_dims = var_63_keep_dims_0, x = var_2521_cast_fp16)[name = string("var_63_cast_fp16")];
            fp16 var_2528_to_fp16 = const()[name = string("op_2528_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_2529_cast_fp16 = add(x = var_63_cast_fp16, y = var_2528_to_fp16)[name = string("op_2529_cast_fp16")];
            fp32 var_2530_epsilon_0 = const()[name = string("op_2530_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_2530_cast_fp16 = rsqrt(epsilon = var_2530_epsilon_0, x = var_2529_cast_fp16)[name = string("op_2530_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_63_cast_fp16 = mul(x = var_2315_cast_fp16_3, y = var_2530_cast_fp16)[name = string("x_normed_63_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_101_cast_fp16 = mul(x = x_normed_63_cast_fp16, y = var_2348_to_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> x_first_37_begin_0 = const()[name = string("x_first_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_37_end_0 = const()[name = string("x_first_37_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_37_end_mask_0 = const()[name = string("x_first_37_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_37_cast_fp16 = slice_by_index(begin = x_first_37_begin_0, end = x_first_37_end_0, end_mask = x_first_37_end_mask_0, x = x_101_cast_fp16)[name = string("x_first_37_cast_fp16")];
            tensor<int32, [4]> x_second_37_begin_0 = const()[name = string("x_second_37_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_37_end_0 = const()[name = string("x_second_37_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_37_end_mask_0 = const()[name = string("x_second_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_37_cast_fp16 = slice_by_index(begin = x_second_37_begin_0, end = x_second_37_end_0, end_mask = x_second_37_end_mask_0, x = x_101_cast_fp16)[name = string("x_second_37_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_2562_cast_fp16 = mul(x = x_second_37_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_2562_cast_fp16")];
            int32 var_2564 = const()[name = string("op_2564"), val = int32(1)];
            bool x_rot_37_interleave_0 = const()[name = string("x_rot_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_37_cast_fp16 = concat(axis = var_2564, interleave = x_rot_37_interleave_0, values = (var_2562_cast_fp16, x_first_37_cast_fp16))[name = string("x_rot_37_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_2580_cast_fp16 = mul(x = x_101_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_2580_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_2581_cast_fp16 = mul(x = x_rot_37_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_31_cast_fp16 = add(x = var_2580_cast_fp16, y = var_2581_cast_fp16)[name = string("q_h_31_cast_fp16")];
            fp16 var_2584_promoted_to_fp16 = const()[name = string("op_2584_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_2585_cast_fp16 = pow(x = var_2301_cast_fp16, y = var_2584_promoted_to_fp16)[name = string("op_2585_cast_fp16")];
            tensor<int32, [1]> var_65_axes_0 = const()[name = string("var_65_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_65_keep_dims_0 = const()[name = string("var_65_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_65_cast_fp16 = reduce_mean(axes = var_65_axes_0, keep_dims = var_65_keep_dims_0, x = var_2585_cast_fp16)[name = string("var_65_cast_fp16")];
            fp16 var_2592_to_fp16 = const()[name = string("op_2592_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_2593_cast_fp16 = add(x = var_65_cast_fp16, y = var_2592_to_fp16)[name = string("op_2593_cast_fp16")];
            fp32 var_2594_epsilon_0 = const()[name = string("op_2594_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_2594_cast_fp16 = rsqrt(epsilon = var_2594_epsilon_0, x = var_2593_cast_fp16)[name = string("op_2594_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_65_cast_fp16 = mul(x = var_2301_cast_fp16, y = var_2594_cast_fp16)[name = string("x_normed_65_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_2604_to_fp16 = const()[name = string("op_2604_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370956928)))];
            tensor<fp16, [1, 256, 1, ?]> x_105_cast_fp16 = mul(x = x_normed_65_cast_fp16, y = var_2604_to_fp16)[name = string("x_105_cast_fp16")];
            tensor<int32, [4]> x_first_39_begin_0 = const()[name = string("x_first_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_39_end_0 = const()[name = string("x_first_39_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_39_end_mask_0 = const()[name = string("x_first_39_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_39_cast_fp16 = slice_by_index(begin = x_first_39_begin_0, end = x_first_39_end_0, end_mask = x_first_39_end_mask_0, x = x_105_cast_fp16)[name = string("x_first_39_cast_fp16")];
            tensor<int32, [4]> x_second_39_begin_0 = const()[name = string("x_second_39_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_39_end_0 = const()[name = string("x_second_39_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_39_end_mask_0 = const()[name = string("x_second_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_39_cast_fp16 = slice_by_index(begin = x_second_39_begin_0, end = x_second_39_end_0, end_mask = x_second_39_end_mask_0, x = x_105_cast_fp16)[name = string("x_second_39_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_2626_cast_fp16 = mul(x = x_second_39_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_2626_cast_fp16")];
            int32 var_2628 = const()[name = string("op_2628"), val = int32(1)];
            bool x_rot_39_interleave_0 = const()[name = string("x_rot_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_39_cast_fp16 = concat(axis = var_2628, interleave = x_rot_39_interleave_0, values = (var_2626_cast_fp16, x_first_39_cast_fp16))[name = string("x_rot_39_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_2644_cast_fp16 = mul(x = x_105_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_2644_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_2645_cast_fp16 = mul(x = x_rot_39_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_2645_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_7_cast_fp16 = add(x = var_2644_cast_fp16, y = var_2645_cast_fp16)[name = string("k_new_7_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_6 = read_state(input = k_cache_3)[name = string("read_state_6")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_21_cast_fp16 = mul(x = read_state_6, y = one_minus_1_cast_fp16)[name = string("k_cache_21_cast_fp16")];
            write_state(data = k_cache_21_cast_fp16, input = k_cache_3)[name = string("coreml_update_state_84_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_84 = read_state(input = k_cache_3)[name = string("coreml_update_state_84")];
            tensor<fp16, [1, 256, 1, 128]> var_2652_cast_fp16 = mul(x = k_new_7_cast_fp16, y = write_mask)[name = string("op_2652_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_23_cast_fp16 = add(x = coreml_update_state_84, y = var_2652_cast_fp16)[name = string("k_cache_23_cast_fp16")];
            write_state(data = k_cache_23_cast_fp16, input = k_cache_3)[name = string("coreml_update_state_85_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_85 = read_state(input = k_cache_3)[name = string("coreml_update_state_85")];
            tensor<fp16, [1, 256, 1, 128]> read_state_7 = read_state(input = v_cache_3)[name = string("read_state_7")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_21_cast_fp16 = mul(x = read_state_7, y = one_minus_1_cast_fp16)[name = string("v_cache_21_cast_fp16")];
            write_state(data = v_cache_21_cast_fp16, input = v_cache_3)[name = string("coreml_update_state_86_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_86 = read_state(input = v_cache_3)[name = string("coreml_update_state_86")];
            tensor<fp16, [1, 256, 1, 128]> var_2656_cast_fp16 = mul(x = var_2312_cast_fp16, y = write_mask)[name = string("op_2656_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_23_cast_fp16 = add(x = coreml_update_state_86, y = var_2656_cast_fp16)[name = string("v_cache_23_cast_fp16")];
            write_state(data = v_cache_23_cast_fp16, input = v_cache_3)[name = string("coreml_update_state_87_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_87 = read_state(input = v_cache_3)[name = string("coreml_update_state_87")];
            bool var_2671_transpose_x_0 = const()[name = string("op_2671_transpose_x_0"), val = bool(false)];
            bool var_2671_transpose_y_0 = const()[name = string("op_2671_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_330_perm_0 = const()[name = string("transpose_330_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_331_perm_0 = const()[name = string("transpose_331_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_331 = transpose(perm = transpose_331_perm_0, x = coreml_update_state_85)[name = string("transpose_555")];
            tensor<fp16, [1, 1, ?, 256]> transpose_330 = transpose(perm = transpose_330_perm_0, x = q_h_25_cast_fp16)[name = string("transpose_556")];
            tensor<fp16, [1, 1, ?, 128]> var_2671_cast_fp16 = matmul(transpose_x = var_2671_transpose_x_0, transpose_y = var_2671_transpose_y_0, x = transpose_330, y = transpose_331)[name = string("op_2671_cast_fp16")];
            fp16 var_2672_to_fp16 = const()[name = string("op_2672_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_2673_cast_fp16 = mul(x = var_2671_cast_fp16, y = var_2672_to_fp16)[name = string("op_2673_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_25_cast_fp16 = add(x = var_2673_cast_fp16, y = attention_mask)[name = string("scores_25_cast_fp16")];
            int32 var_2676 = const()[name = string("op_2676"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_25_cast_fp16 = softmax(axis = var_2676, x = scores_25_cast_fp16)[name = string("probs_25_cast_fp16")];
            tensor<int32, [4]> var_2683 = const()[name = string("op_2683"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_2685_transpose_x_0 = const()[name = string("op_2685_transpose_x_0"), val = bool(false)];
            bool var_2685_transpose_y_0 = const()[name = string("op_2685_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_25_cast_fp16 = transpose(perm = var_2683, x = coreml_update_state_87)[name = string("transpose_554")];
            tensor<fp16, [1, 1, ?, 256]> var_2685_cast_fp16 = matmul(transpose_x = var_2685_transpose_x_0, transpose_y = var_2685_transpose_y_0, x = probs_25_cast_fp16, y = v_t_25_cast_fp16)[name = string("op_2685_cast_fp16")];
            bool var_2704_transpose_x_0 = const()[name = string("op_2704_transpose_x_0"), val = bool(false)];
            bool var_2704_transpose_y_0 = const()[name = string("op_2704_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_335_perm_0 = const()[name = string("transpose_335_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_335 = transpose(perm = transpose_335_perm_0, x = q_h_27_cast_fp16)[name = string("transpose_553")];
            tensor<fp16, [1, 1, ?, 128]> var_2704_cast_fp16 = matmul(transpose_x = var_2704_transpose_x_0, transpose_y = var_2704_transpose_y_0, x = transpose_335, y = transpose_331)[name = string("op_2704_cast_fp16")];
            fp16 var_2705_to_fp16 = const()[name = string("op_2705_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_2706_cast_fp16 = mul(x = var_2704_cast_fp16, y = var_2705_to_fp16)[name = string("op_2706_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_27_cast_fp16 = add(x = var_2706_cast_fp16, y = attention_mask)[name = string("scores_27_cast_fp16")];
            int32 var_2709 = const()[name = string("op_2709"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_27_cast_fp16 = softmax(axis = var_2709, x = scores_27_cast_fp16)[name = string("probs_27_cast_fp16")];
            bool var_2718_transpose_x_0 = const()[name = string("op_2718_transpose_x_0"), val = bool(false)];
            bool var_2718_transpose_y_0 = const()[name = string("op_2718_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_2718_cast_fp16 = matmul(transpose_x = var_2718_transpose_x_0, transpose_y = var_2718_transpose_y_0, x = probs_27_cast_fp16, y = v_t_25_cast_fp16)[name = string("op_2718_cast_fp16")];
            bool var_2737_transpose_x_0 = const()[name = string("op_2737_transpose_x_0"), val = bool(false)];
            bool var_2737_transpose_y_0 = const()[name = string("op_2737_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_336_perm_0 = const()[name = string("transpose_336_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_336 = transpose(perm = transpose_336_perm_0, x = q_h_29_cast_fp16)[name = string("transpose_552")];
            tensor<fp16, [1, 1, ?, 128]> var_2737_cast_fp16 = matmul(transpose_x = var_2737_transpose_x_0, transpose_y = var_2737_transpose_y_0, x = transpose_336, y = transpose_331)[name = string("op_2737_cast_fp16")];
            fp16 var_2738_to_fp16 = const()[name = string("op_2738_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_2739_cast_fp16 = mul(x = var_2737_cast_fp16, y = var_2738_to_fp16)[name = string("op_2739_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_29_cast_fp16 = add(x = var_2739_cast_fp16, y = attention_mask)[name = string("scores_29_cast_fp16")];
            int32 var_2742 = const()[name = string("op_2742"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_29_cast_fp16 = softmax(axis = var_2742, x = scores_29_cast_fp16)[name = string("probs_29_cast_fp16")];
            bool var_2751_transpose_x_0 = const()[name = string("op_2751_transpose_x_0"), val = bool(false)];
            bool var_2751_transpose_y_0 = const()[name = string("op_2751_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_2751_cast_fp16 = matmul(transpose_x = var_2751_transpose_x_0, transpose_y = var_2751_transpose_y_0, x = probs_29_cast_fp16, y = v_t_25_cast_fp16)[name = string("op_2751_cast_fp16")];
            bool var_2770_transpose_x_0 = const()[name = string("op_2770_transpose_x_0"), val = bool(false)];
            bool var_2770_transpose_y_0 = const()[name = string("op_2770_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_337_perm_0 = const()[name = string("transpose_337_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_337 = transpose(perm = transpose_337_perm_0, x = q_h_31_cast_fp16)[name = string("transpose_551")];
            tensor<fp16, [1, 1, ?, 128]> var_2770_cast_fp16 = matmul(transpose_x = var_2770_transpose_x_0, transpose_y = var_2770_transpose_y_0, x = transpose_337, y = transpose_331)[name = string("op_2770_cast_fp16")];
            fp16 var_2771_to_fp16 = const()[name = string("op_2771_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_2772_cast_fp16 = mul(x = var_2770_cast_fp16, y = var_2771_to_fp16)[name = string("op_2772_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_31_cast_fp16 = add(x = var_2772_cast_fp16, y = attention_mask)[name = string("scores_31_cast_fp16")];
            int32 var_2775 = const()[name = string("op_2775"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_31_cast_fp16 = softmax(axis = var_2775, x = scores_31_cast_fp16)[name = string("probs_31_cast_fp16")];
            bool var_2784_transpose_x_0 = const()[name = string("op_2784_transpose_x_0"), val = bool(false)];
            bool var_2784_transpose_y_0 = const()[name = string("op_2784_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_2784_cast_fp16 = matmul(transpose_x = var_2784_transpose_x_0, transpose_y = var_2784_transpose_y_0, x = probs_31_cast_fp16, y = v_t_25_cast_fp16)[name = string("op_2784_cast_fp16")];
            bool input_27_interleave_0 = const()[name = string("input_27_interleave_0"), val = bool(false)];
            int32 const_148 = const()[name = string("const_148"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_27_cast_fp16 = concat(axis = const_148, interleave = input_27_interleave_0, values = (var_2685_cast_fp16, var_2718_cast_fp16, var_2751_cast_fp16, var_2784_cast_fp16))[name = string("input_27_cast_fp16")];
            string x_107_pad_type_0 = const()[name = string("x_107_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_107_strides_0 = const()[name = string("x_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_107_pad_0 = const()[name = string("x_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_107_dilations_0 = const()[name = string("x_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_107_groups_0 = const()[name = string("x_107_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_3_weight_to_fp16 = const()[name = string("o_proj_3_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370957504)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_35_cast_fp16 = transpose(perm = transpose_35_perm_0, x = input_27_cast_fp16)[name = string("transpose_550")];
            tensor<fp16, [1, 640, 1, ?]> x_107_cast_fp16 = conv(dilations = x_107_dilations_0, groups = x_107_groups_0, pad = x_107_pad_0, pad_type = x_107_pad_type_0, strides = x_107_strides_0, weight = o_proj_3_weight_to_fp16, x = transpose_35_cast_fp16)[name = string("x_107_cast_fp16")];
            fp16 var_2805_promoted_to_fp16 = const()[name = string("op_2805_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_2806_cast_fp16 = pow(x = x_107_cast_fp16, y = var_2805_promoted_to_fp16)[name = string("op_2806_cast_fp16")];
            tensor<int32, [1]> var_67_axes_0 = const()[name = string("var_67_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_67_keep_dims_0 = const()[name = string("var_67_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_67_cast_fp16 = reduce_mean(axes = var_67_axes_0, keep_dims = var_67_keep_dims_0, x = var_2806_cast_fp16)[name = string("var_67_cast_fp16")];
            fp16 var_2813_to_fp16 = const()[name = string("op_2813_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_2814_cast_fp16 = add(x = var_67_cast_fp16, y = var_2813_to_fp16)[name = string("op_2814_cast_fp16")];
            fp32 var_2815_epsilon_0 = const()[name = string("op_2815_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_2815_cast_fp16 = rsqrt(epsilon = var_2815_epsilon_0, x = var_2814_cast_fp16)[name = string("op_2815_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_67_cast_fp16 = mul(x = x_107_cast_fp16, y = var_2815_cast_fp16)[name = string("x_normed_67_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_2825_to_fp16 = const()[name = string("op_2825_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372268288)))];
            tensor<fp16, [1, 640, 1, ?]> attn_15_cast_fp16 = mul(x = x_normed_67_cast_fp16, y = var_2825_to_fp16)[name = string("attn_15_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_109_cast_fp16 = add(x = x_85_cast_fp16, y = attn_15_cast_fp16)[name = string("x_109_cast_fp16")];
            fp16 var_2829_promoted_to_fp16 = const()[name = string("op_2829_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_2830_cast_fp16 = pow(x = x_109_cast_fp16, y = var_2829_promoted_to_fp16)[name = string("op_2830_cast_fp16")];
            tensor<int32, [1]> var_69_axes_0 = const()[name = string("var_69_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_69_keep_dims_0 = const()[name = string("var_69_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_69_cast_fp16 = reduce_mean(axes = var_69_axes_0, keep_dims = var_69_keep_dims_0, x = var_2830_cast_fp16)[name = string("var_69_cast_fp16")];
            fp16 var_2837_to_fp16 = const()[name = string("op_2837_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_2838_cast_fp16 = add(x = var_69_cast_fp16, y = var_2837_to_fp16)[name = string("op_2838_cast_fp16")];
            fp32 var_2839_epsilon_0 = const()[name = string("op_2839_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_2839_cast_fp16 = rsqrt(epsilon = var_2839_epsilon_0, x = var_2838_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_69_cast_fp16 = mul(x = x_109_cast_fp16, y = var_2839_cast_fp16)[name = string("x_normed_69_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_2849_to_fp16 = const()[name = string("op_2849_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372269632)))];
            tensor<fp16, [1, 640, 1, ?]> input_29_cast_fp16 = mul(x = x_normed_69_cast_fp16, y = var_2849_to_fp16)[name = string("input_29_cast_fp16")];
            string gate_7_pad_type_0 = const()[name = string("gate_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_7_strides_0 = const()[name = string("gate_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_7_pad_0 = const()[name = string("gate_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_7_dilations_0 = const()[name = string("gate_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_7_groups_0 = const()[name = string("gate_7_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_3_weight_to_fp16 = const()[name = string("gate_proj_3_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372270976)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_7_cast_fp16 = conv(dilations = gate_7_dilations_0, groups = gate_7_groups_0, pad = gate_7_pad_0, pad_type = gate_7_pad_type_0, strides = gate_7_strides_0, weight = gate_proj_3_weight_to_fp16, x = input_29_cast_fp16)[name = string("gate_7_cast_fp16")];
            string up_7_pad_type_0 = const()[name = string("up_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_7_strides_0 = const()[name = string("up_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_7_pad_0 = const()[name = string("up_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_7_dilations_0 = const()[name = string("up_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_7_groups_0 = const()[name = string("up_7_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_3_weight_to_fp16 = const()[name = string("up_proj_3_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374892480)))];
            tensor<fp16, [1, 2048, 1, ?]> up_7_cast_fp16 = conv(dilations = up_7_dilations_0, groups = up_7_groups_0, pad = up_7_pad_0, pad_type = up_7_pad_type_0, strides = up_7_strides_0, weight = up_proj_3_weight_to_fp16, x = input_29_cast_fp16)[name = string("up_7_cast_fp16")];
            string var_2874_mode_0 = const()[name = string("op_2874_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_2874_cast_fp16 = gelu(mode = var_2874_mode_0, x = gate_7_cast_fp16)[name = string("op_2874_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_31_cast_fp16 = mul(x = var_2874_cast_fp16, y = up_7_cast_fp16)[name = string("input_31_cast_fp16")];
            string x_111_pad_type_0 = const()[name = string("x_111_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_111_strides_0 = const()[name = string("x_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_111_pad_0 = const()[name = string("x_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_111_dilations_0 = const()[name = string("x_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_111_groups_0 = const()[name = string("x_111_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_3_weight_to_fp16 = const()[name = string("down_proj_3_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377513984)))];
            tensor<fp16, [1, 640, 1, ?]> x_111_cast_fp16 = conv(dilations = x_111_dilations_0, groups = x_111_groups_0, pad = x_111_pad_0, pad_type = x_111_pad_type_0, strides = x_111_strides_0, weight = down_proj_3_weight_to_fp16, x = input_31_cast_fp16)[name = string("x_111_cast_fp16")];
            fp16 var_2887_promoted_to_fp16 = const()[name = string("op_2887_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_2888_cast_fp16 = pow(x = x_111_cast_fp16, y = var_2887_promoted_to_fp16)[name = string("op_2888_cast_fp16")];
            tensor<int32, [1]> var_71_axes_0 = const()[name = string("var_71_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_71_keep_dims_0 = const()[name = string("var_71_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_71_cast_fp16 = reduce_mean(axes = var_71_axes_0, keep_dims = var_71_keep_dims_0, x = var_2888_cast_fp16)[name = string("var_71_cast_fp16")];
            fp16 var_2895_to_fp16 = const()[name = string("op_2895_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_2896_cast_fp16 = add(x = var_71_cast_fp16, y = var_2895_to_fp16)[name = string("op_2896_cast_fp16")];
            fp32 var_2897_epsilon_0 = const()[name = string("op_2897_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_2897_cast_fp16 = rsqrt(epsilon = var_2897_epsilon_0, x = var_2896_cast_fp16)[name = string("op_2897_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_71_cast_fp16 = mul(x = x_111_cast_fp16, y = var_2897_cast_fp16)[name = string("x_normed_71_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_2907_to_fp16 = const()[name = string("op_2907_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380135488)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_7_cast_fp16 = mul(x = x_normed_71_cast_fp16, y = var_2907_to_fp16)[name = string("mlp_7_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_113_cast_fp16 = add(x = x_109_cast_fp16, y = mlp_7_cast_fp16)[name = string("x_113_cast_fp16")];
            fp16 var_2911_promoted_to_fp16 = const()[name = string("op_2911_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_2912_cast_fp16 = pow(x = x_113_cast_fp16, y = var_2911_promoted_to_fp16)[name = string("op_2912_cast_fp16")];
            tensor<int32, [1]> var_73_axes_0 = const()[name = string("var_73_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_73_keep_dims_0 = const()[name = string("var_73_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_73_cast_fp16 = reduce_mean(axes = var_73_axes_0, keep_dims = var_73_keep_dims_0, x = var_2912_cast_fp16)[name = string("var_73_cast_fp16")];
            fp16 var_2919_to_fp16 = const()[name = string("op_2919_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_2920_cast_fp16 = add(x = var_73_cast_fp16, y = var_2919_to_fp16)[name = string("op_2920_cast_fp16")];
            fp32 var_2921_epsilon_0 = const()[name = string("op_2921_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_2921_cast_fp16 = rsqrt(epsilon = var_2921_epsilon_0, x = var_2920_cast_fp16)[name = string("op_2921_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_73_cast_fp16 = mul(x = x_113_cast_fp16, y = var_2921_cast_fp16)[name = string("x_normed_73_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_2931_to_fp16 = const()[name = string("op_2931_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380136832)))];
            tensor<fp16, [1, 640, 1, ?]> input_33_cast_fp16 = mul(x = x_normed_73_cast_fp16, y = var_2931_to_fp16)[name = string("input_33_cast_fp16")];
            string var_2943_pad_type_0 = const()[name = string("op_2943_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2943_strides_0 = const()[name = string("op_2943_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2943_pad_0 = const()[name = string("op_2943_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2943_dilations_0 = const()[name = string("op_2943_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2943_groups_0 = const()[name = string("op_2943_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_4_weight_to_fp16 = const()[name = string("q_proj_4_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380138176)))];
            tensor<fp16, [1, 1024, 1, ?]> var_2943_cast_fp16 = conv(dilations = var_2943_dilations_0, groups = var_2943_groups_0, pad = var_2943_pad_0, pad_type = var_2943_pad_type_0, strides = var_2943_strides_0, weight = q_proj_4_weight_to_fp16, x = input_33_cast_fp16)[name = string("op_2943_cast_fp16")];
            string var_2954_pad_type_0 = const()[name = string("op_2954_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2954_strides_0 = const()[name = string("op_2954_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2954_pad_0 = const()[name = string("op_2954_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2954_dilations_0 = const()[name = string("op_2954_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2954_groups_0 = const()[name = string("op_2954_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_4_weight_to_fp16 = const()[name = string("k_proj_4_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381448960)))];
            tensor<fp16, [1, 256, 1, ?]> var_2954_cast_fp16 = conv(dilations = var_2954_dilations_0, groups = var_2954_groups_0, pad = var_2954_pad_0, pad_type = var_2954_pad_type_0, strides = var_2954_strides_0, weight = k_proj_4_weight_to_fp16, x = input_33_cast_fp16)[name = string("op_2954_cast_fp16")];
            string var_2965_pad_type_0 = const()[name = string("op_2965_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2965_strides_0 = const()[name = string("op_2965_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2965_pad_0 = const()[name = string("op_2965_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2965_dilations_0 = const()[name = string("op_2965_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2965_groups_0 = const()[name = string("op_2965_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_4_weight_to_fp16 = const()[name = string("v_proj_4_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381776704)))];
            tensor<fp16, [1, 256, 1, ?]> var_2965_cast_fp16 = conv(dilations = var_2965_dilations_0, groups = var_2965_groups_0, pad = var_2965_pad_0, pad_type = var_2965_pad_type_0, strides = var_2965_strides_0, weight = v_proj_4_weight_to_fp16, x = input_33_cast_fp16)[name = string("op_2965_cast_fp16")];
            tensor<int32, [4]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_2968_axis_0 = const()[name = string("op_2968_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_2968_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_2968_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_2968_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_2968_cast_fp16_3 = split(axis = var_2968_axis_0, split_sizes = tile_12, x = var_2943_cast_fp16)[name = string("op_2968_cast_fp16")];
            fp16 var_2981_promoted_to_fp16 = const()[name = string("op_2981_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_2982_cast_fp16 = pow(x = var_2968_cast_fp16_0, y = var_2981_promoted_to_fp16)[name = string("op_2982_cast_fp16")];
            tensor<int32, [1]> var_75_axes_0 = const()[name = string("var_75_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_75_keep_dims_0 = const()[name = string("var_75_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_75_cast_fp16 = reduce_mean(axes = var_75_axes_0, keep_dims = var_75_keep_dims_0, x = var_2982_cast_fp16)[name = string("var_75_cast_fp16")];
            fp16 var_2989_to_fp16 = const()[name = string("op_2989_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_2990_cast_fp16 = add(x = var_75_cast_fp16, y = var_2989_to_fp16)[name = string("op_2990_cast_fp16")];
            fp32 var_2991_epsilon_0 = const()[name = string("op_2991_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_2991_cast_fp16 = rsqrt(epsilon = var_2991_epsilon_0, x = var_2990_cast_fp16)[name = string("op_2991_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_75_cast_fp16 = mul(x = var_2968_cast_fp16_0, y = var_2991_cast_fp16)[name = string("x_normed_75_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_3001_to_fp16 = const()[name = string("op_3001_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382104448)))];
            tensor<fp16, [1, 256, 1, ?]> x_117_cast_fp16 = mul(x = x_normed_75_cast_fp16, y = var_3001_to_fp16)[name = string("x_117_cast_fp16")];
            tensor<int32, [4]> x_first_41_begin_0 = const()[name = string("x_first_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_41_end_0 = const()[name = string("x_first_41_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_41_end_mask_0 = const()[name = string("x_first_41_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_41_cast_fp16 = slice_by_index(begin = x_first_41_begin_0, end = x_first_41_end_0, end_mask = x_first_41_end_mask_0, x = x_117_cast_fp16)[name = string("x_first_41_cast_fp16")];
            tensor<int32, [4]> x_second_41_begin_0 = const()[name = string("x_second_41_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_41_end_0 = const()[name = string("x_second_41_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_41_end_mask_0 = const()[name = string("x_second_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_41_cast_fp16 = slice_by_index(begin = x_second_41_begin_0, end = x_second_41_end_0, end_mask = x_second_41_end_mask_0, x = x_117_cast_fp16)[name = string("x_second_41_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_3023_cast_fp16 = mul(x = x_second_41_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_3023_cast_fp16")];
            int32 var_3025 = const()[name = string("op_3025"), val = int32(1)];
            bool x_rot_41_interleave_0 = const()[name = string("x_rot_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_41_cast_fp16 = concat(axis = var_3025, interleave = x_rot_41_interleave_0, values = (var_3023_cast_fp16, x_first_41_cast_fp16))[name = string("x_rot_41_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3041_cast_fp16 = mul(x = x_117_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_3041_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3042_cast_fp16 = mul(x = x_rot_41_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_3042_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_33_cast_fp16 = add(x = var_3041_cast_fp16, y = var_3042_cast_fp16)[name = string("q_h_33_cast_fp16")];
            fp16 var_3045_promoted_to_fp16 = const()[name = string("op_3045_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_3046_cast_fp16 = pow(x = var_2968_cast_fp16_1, y = var_3045_promoted_to_fp16)[name = string("op_3046_cast_fp16")];
            tensor<int32, [1]> var_77_axes_0 = const()[name = string("var_77_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_77_keep_dims_0 = const()[name = string("var_77_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_77_cast_fp16 = reduce_mean(axes = var_77_axes_0, keep_dims = var_77_keep_dims_0, x = var_3046_cast_fp16)[name = string("var_77_cast_fp16")];
            fp16 var_3053_to_fp16 = const()[name = string("op_3053_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_3054_cast_fp16 = add(x = var_77_cast_fp16, y = var_3053_to_fp16)[name = string("op_3054_cast_fp16")];
            fp32 var_3055_epsilon_0 = const()[name = string("op_3055_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_3055_cast_fp16 = rsqrt(epsilon = var_3055_epsilon_0, x = var_3054_cast_fp16)[name = string("op_3055_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_77_cast_fp16 = mul(x = var_2968_cast_fp16_1, y = var_3055_cast_fp16)[name = string("x_normed_77_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_121_cast_fp16 = mul(x = x_normed_77_cast_fp16, y = var_3001_to_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [4]> x_first_43_begin_0 = const()[name = string("x_first_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_43_end_0 = const()[name = string("x_first_43_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_43_end_mask_0 = const()[name = string("x_first_43_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_43_cast_fp16 = slice_by_index(begin = x_first_43_begin_0, end = x_first_43_end_0, end_mask = x_first_43_end_mask_0, x = x_121_cast_fp16)[name = string("x_first_43_cast_fp16")];
            tensor<int32, [4]> x_second_43_begin_0 = const()[name = string("x_second_43_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_43_end_0 = const()[name = string("x_second_43_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_43_end_mask_0 = const()[name = string("x_second_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_43_cast_fp16 = slice_by_index(begin = x_second_43_begin_0, end = x_second_43_end_0, end_mask = x_second_43_end_mask_0, x = x_121_cast_fp16)[name = string("x_second_43_cast_fp16")];
            fp16 const_37_promoted_to_fp16 = const()[name = string("const_37_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_3087_cast_fp16 = mul(x = x_second_43_cast_fp16, y = const_37_promoted_to_fp16)[name = string("op_3087_cast_fp16")];
            int32 var_3089 = const()[name = string("op_3089"), val = int32(1)];
            bool x_rot_43_interleave_0 = const()[name = string("x_rot_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_43_cast_fp16 = concat(axis = var_3089, interleave = x_rot_43_interleave_0, values = (var_3087_cast_fp16, x_first_43_cast_fp16))[name = string("x_rot_43_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3105_cast_fp16 = mul(x = x_121_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_3105_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3106_cast_fp16 = mul(x = x_rot_43_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_3106_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_35_cast_fp16 = add(x = var_3105_cast_fp16, y = var_3106_cast_fp16)[name = string("q_h_35_cast_fp16")];
            fp16 var_3109_promoted_to_fp16 = const()[name = string("op_3109_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_3110_cast_fp16 = pow(x = var_2968_cast_fp16_2, y = var_3109_promoted_to_fp16)[name = string("op_3110_cast_fp16")];
            tensor<int32, [1]> var_79_axes_0 = const()[name = string("var_79_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_79_keep_dims_0 = const()[name = string("var_79_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_79_cast_fp16 = reduce_mean(axes = var_79_axes_0, keep_dims = var_79_keep_dims_0, x = var_3110_cast_fp16)[name = string("var_79_cast_fp16")];
            fp16 var_3117_to_fp16 = const()[name = string("op_3117_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_3118_cast_fp16 = add(x = var_79_cast_fp16, y = var_3117_to_fp16)[name = string("op_3118_cast_fp16")];
            fp32 var_3119_epsilon_0 = const()[name = string("op_3119_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_3119_cast_fp16 = rsqrt(epsilon = var_3119_epsilon_0, x = var_3118_cast_fp16)[name = string("op_3119_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_79_cast_fp16 = mul(x = var_2968_cast_fp16_2, y = var_3119_cast_fp16)[name = string("x_normed_79_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_125_cast_fp16 = mul(x = x_normed_79_cast_fp16, y = var_3001_to_fp16)[name = string("x_125_cast_fp16")];
            tensor<int32, [4]> x_first_45_begin_0 = const()[name = string("x_first_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_45_end_0 = const()[name = string("x_first_45_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_45_end_mask_0 = const()[name = string("x_first_45_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_45_cast_fp16 = slice_by_index(begin = x_first_45_begin_0, end = x_first_45_end_0, end_mask = x_first_45_end_mask_0, x = x_125_cast_fp16)[name = string("x_first_45_cast_fp16")];
            tensor<int32, [4]> x_second_45_begin_0 = const()[name = string("x_second_45_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_45_end_0 = const()[name = string("x_second_45_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_45_end_mask_0 = const()[name = string("x_second_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_45_cast_fp16 = slice_by_index(begin = x_second_45_begin_0, end = x_second_45_end_0, end_mask = x_second_45_end_mask_0, x = x_125_cast_fp16)[name = string("x_second_45_cast_fp16")];
            fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_3151_cast_fp16 = mul(x = x_second_45_cast_fp16, y = const_38_promoted_to_fp16)[name = string("op_3151_cast_fp16")];
            int32 var_3153 = const()[name = string("op_3153"), val = int32(1)];
            bool x_rot_45_interleave_0 = const()[name = string("x_rot_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_45_cast_fp16 = concat(axis = var_3153, interleave = x_rot_45_interleave_0, values = (var_3151_cast_fp16, x_first_45_cast_fp16))[name = string("x_rot_45_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3169_cast_fp16 = mul(x = x_125_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_3169_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3170_cast_fp16 = mul(x = x_rot_45_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_3170_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_37_cast_fp16 = add(x = var_3169_cast_fp16, y = var_3170_cast_fp16)[name = string("q_h_37_cast_fp16")];
            fp16 var_3173_promoted_to_fp16 = const()[name = string("op_3173_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_3174_cast_fp16 = pow(x = var_2968_cast_fp16_3, y = var_3173_promoted_to_fp16)[name = string("op_3174_cast_fp16")];
            tensor<int32, [1]> var_81_axes_0 = const()[name = string("var_81_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_81_keep_dims_0 = const()[name = string("var_81_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_81_cast_fp16 = reduce_mean(axes = var_81_axes_0, keep_dims = var_81_keep_dims_0, x = var_3174_cast_fp16)[name = string("var_81_cast_fp16")];
            fp16 var_3181_to_fp16 = const()[name = string("op_3181_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_3182_cast_fp16 = add(x = var_81_cast_fp16, y = var_3181_to_fp16)[name = string("op_3182_cast_fp16")];
            fp32 var_3183_epsilon_0 = const()[name = string("op_3183_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_3183_cast_fp16 = rsqrt(epsilon = var_3183_epsilon_0, x = var_3182_cast_fp16)[name = string("op_3183_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_81_cast_fp16 = mul(x = var_2968_cast_fp16_3, y = var_3183_cast_fp16)[name = string("x_normed_81_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_129_cast_fp16 = mul(x = x_normed_81_cast_fp16, y = var_3001_to_fp16)[name = string("x_129_cast_fp16")];
            tensor<int32, [4]> x_first_47_begin_0 = const()[name = string("x_first_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_47_end_0 = const()[name = string("x_first_47_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_47_end_mask_0 = const()[name = string("x_first_47_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_47_cast_fp16 = slice_by_index(begin = x_first_47_begin_0, end = x_first_47_end_0, end_mask = x_first_47_end_mask_0, x = x_129_cast_fp16)[name = string("x_first_47_cast_fp16")];
            tensor<int32, [4]> x_second_47_begin_0 = const()[name = string("x_second_47_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_47_end_0 = const()[name = string("x_second_47_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_47_end_mask_0 = const()[name = string("x_second_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_47_cast_fp16 = slice_by_index(begin = x_second_47_begin_0, end = x_second_47_end_0, end_mask = x_second_47_end_mask_0, x = x_129_cast_fp16)[name = string("x_second_47_cast_fp16")];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_3215_cast_fp16 = mul(x = x_second_47_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_3215_cast_fp16")];
            int32 var_3217 = const()[name = string("op_3217"), val = int32(1)];
            bool x_rot_47_interleave_0 = const()[name = string("x_rot_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_47_cast_fp16 = concat(axis = var_3217, interleave = x_rot_47_interleave_0, values = (var_3215_cast_fp16, x_first_47_cast_fp16))[name = string("x_rot_47_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3233_cast_fp16 = mul(x = x_129_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_3233_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3234_cast_fp16 = mul(x = x_rot_47_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_39_cast_fp16 = add(x = var_3233_cast_fp16, y = var_3234_cast_fp16)[name = string("q_h_39_cast_fp16")];
            fp16 var_3237_promoted_to_fp16 = const()[name = string("op_3237_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_3238_cast_fp16 = pow(x = var_2954_cast_fp16, y = var_3237_promoted_to_fp16)[name = string("op_3238_cast_fp16")];
            tensor<int32, [1]> var_83_axes_0 = const()[name = string("var_83_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_83_keep_dims_0 = const()[name = string("var_83_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_83_cast_fp16 = reduce_mean(axes = var_83_axes_0, keep_dims = var_83_keep_dims_0, x = var_3238_cast_fp16)[name = string("var_83_cast_fp16")];
            fp16 var_3245_to_fp16 = const()[name = string("op_3245_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_3246_cast_fp16 = add(x = var_83_cast_fp16, y = var_3245_to_fp16)[name = string("op_3246_cast_fp16")];
            fp32 var_3247_epsilon_0 = const()[name = string("op_3247_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_3247_cast_fp16 = rsqrt(epsilon = var_3247_epsilon_0, x = var_3246_cast_fp16)[name = string("op_3247_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_83_cast_fp16 = mul(x = var_2954_cast_fp16, y = var_3247_cast_fp16)[name = string("x_normed_83_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_3257_to_fp16 = const()[name = string("op_3257_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382105024)))];
            tensor<fp16, [1, 256, 1, ?]> x_133_cast_fp16 = mul(x = x_normed_83_cast_fp16, y = var_3257_to_fp16)[name = string("x_133_cast_fp16")];
            tensor<int32, [4]> x_first_49_begin_0 = const()[name = string("x_first_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_49_end_0 = const()[name = string("x_first_49_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_49_end_mask_0 = const()[name = string("x_first_49_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_49_cast_fp16 = slice_by_index(begin = x_first_49_begin_0, end = x_first_49_end_0, end_mask = x_first_49_end_mask_0, x = x_133_cast_fp16)[name = string("x_first_49_cast_fp16")];
            tensor<int32, [4]> x_second_49_begin_0 = const()[name = string("x_second_49_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_49_end_0 = const()[name = string("x_second_49_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_49_end_mask_0 = const()[name = string("x_second_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_49_cast_fp16 = slice_by_index(begin = x_second_49_begin_0, end = x_second_49_end_0, end_mask = x_second_49_end_mask_0, x = x_133_cast_fp16)[name = string("x_second_49_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_3279_cast_fp16 = mul(x = x_second_49_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_3279_cast_fp16")];
            int32 var_3281 = const()[name = string("op_3281"), val = int32(1)];
            bool x_rot_49_interleave_0 = const()[name = string("x_rot_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_49_cast_fp16 = concat(axis = var_3281, interleave = x_rot_49_interleave_0, values = (var_3279_cast_fp16, x_first_49_cast_fp16))[name = string("x_rot_49_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3297_cast_fp16 = mul(x = x_133_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_3297_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3298_cast_fp16 = mul(x = x_rot_49_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_3298_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_9_cast_fp16 = add(x = var_3297_cast_fp16, y = var_3298_cast_fp16)[name = string("k_new_9_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_8 = read_state(input = k_cache_4)[name = string("read_state_8")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_27_cast_fp16 = mul(x = read_state_8, y = one_minus_1_cast_fp16)[name = string("k_cache_27_cast_fp16")];
            write_state(data = k_cache_27_cast_fp16, input = k_cache_4)[name = string("coreml_update_state_88_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_88 = read_state(input = k_cache_4)[name = string("coreml_update_state_88")];
            tensor<fp16, [1, 256, 1, 128]> var_3305_cast_fp16 = mul(x = k_new_9_cast_fp16, y = write_mask)[name = string("op_3305_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_29_cast_fp16 = add(x = coreml_update_state_88, y = var_3305_cast_fp16)[name = string("k_cache_29_cast_fp16")];
            write_state(data = k_cache_29_cast_fp16, input = k_cache_4)[name = string("coreml_update_state_89_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_89 = read_state(input = k_cache_4)[name = string("coreml_update_state_89")];
            tensor<fp16, [1, 256, 1, 128]> read_state_9 = read_state(input = v_cache_4)[name = string("read_state_9")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_27_cast_fp16 = mul(x = read_state_9, y = one_minus_1_cast_fp16)[name = string("v_cache_27_cast_fp16")];
            write_state(data = v_cache_27_cast_fp16, input = v_cache_4)[name = string("coreml_update_state_90_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_90 = read_state(input = v_cache_4)[name = string("coreml_update_state_90")];
            tensor<fp16, [1, 256, 1, 128]> var_3309_cast_fp16 = mul(x = var_2965_cast_fp16, y = write_mask)[name = string("op_3309_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_29_cast_fp16 = add(x = coreml_update_state_90, y = var_3309_cast_fp16)[name = string("v_cache_29_cast_fp16")];
            write_state(data = v_cache_29_cast_fp16, input = v_cache_4)[name = string("coreml_update_state_91_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_91 = read_state(input = v_cache_4)[name = string("coreml_update_state_91")];
            bool var_3324_transpose_x_0 = const()[name = string("op_3324_transpose_x_0"), val = bool(false)];
            bool var_3324_transpose_y_0 = const()[name = string("op_3324_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_338_perm_0 = const()[name = string("transpose_338_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_339_perm_0 = const()[name = string("transpose_339_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_339 = transpose(perm = transpose_339_perm_0, x = coreml_update_state_89)[name = string("transpose_548")];
            tensor<fp16, [1, 1, ?, 256]> transpose_338 = transpose(perm = transpose_338_perm_0, x = q_h_33_cast_fp16)[name = string("transpose_549")];
            tensor<fp16, [1, 1, ?, 128]> var_3324_cast_fp16 = matmul(transpose_x = var_3324_transpose_x_0, transpose_y = var_3324_transpose_y_0, x = transpose_338, y = transpose_339)[name = string("op_3324_cast_fp16")];
            fp16 var_3325_to_fp16 = const()[name = string("op_3325_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_3326_cast_fp16 = mul(x = var_3324_cast_fp16, y = var_3325_to_fp16)[name = string("op_3326_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_33_cast_fp16 = add(x = var_3326_cast_fp16, y = attention_mask)[name = string("scores_33_cast_fp16")];
            int32 var_3329 = const()[name = string("op_3329"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_33_cast_fp16 = softmax(axis = var_3329, x = scores_33_cast_fp16)[name = string("probs_33_cast_fp16")];
            tensor<int32, [4]> var_3336 = const()[name = string("op_3336"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_3338_transpose_x_0 = const()[name = string("op_3338_transpose_x_0"), val = bool(false)];
            bool var_3338_transpose_y_0 = const()[name = string("op_3338_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_33_cast_fp16 = transpose(perm = var_3336, x = coreml_update_state_91)[name = string("transpose_547")];
            tensor<fp16, [1, 1, ?, 256]> var_3338_cast_fp16 = matmul(transpose_x = var_3338_transpose_x_0, transpose_y = var_3338_transpose_y_0, x = probs_33_cast_fp16, y = v_t_33_cast_fp16)[name = string("op_3338_cast_fp16")];
            bool var_3357_transpose_x_0 = const()[name = string("op_3357_transpose_x_0"), val = bool(false)];
            bool var_3357_transpose_y_0 = const()[name = string("op_3357_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_343_perm_0 = const()[name = string("transpose_343_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_343 = transpose(perm = transpose_343_perm_0, x = q_h_35_cast_fp16)[name = string("transpose_546")];
            tensor<fp16, [1, 1, ?, 128]> var_3357_cast_fp16 = matmul(transpose_x = var_3357_transpose_x_0, transpose_y = var_3357_transpose_y_0, x = transpose_343, y = transpose_339)[name = string("op_3357_cast_fp16")];
            fp16 var_3358_to_fp16 = const()[name = string("op_3358_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_3359_cast_fp16 = mul(x = var_3357_cast_fp16, y = var_3358_to_fp16)[name = string("op_3359_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_35_cast_fp16 = add(x = var_3359_cast_fp16, y = attention_mask)[name = string("scores_35_cast_fp16")];
            int32 var_3362 = const()[name = string("op_3362"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_35_cast_fp16 = softmax(axis = var_3362, x = scores_35_cast_fp16)[name = string("probs_35_cast_fp16")];
            bool var_3371_transpose_x_0 = const()[name = string("op_3371_transpose_x_0"), val = bool(false)];
            bool var_3371_transpose_y_0 = const()[name = string("op_3371_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_3371_cast_fp16 = matmul(transpose_x = var_3371_transpose_x_0, transpose_y = var_3371_transpose_y_0, x = probs_35_cast_fp16, y = v_t_33_cast_fp16)[name = string("op_3371_cast_fp16")];
            bool var_3390_transpose_x_0 = const()[name = string("op_3390_transpose_x_0"), val = bool(false)];
            bool var_3390_transpose_y_0 = const()[name = string("op_3390_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_344_perm_0 = const()[name = string("transpose_344_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_344 = transpose(perm = transpose_344_perm_0, x = q_h_37_cast_fp16)[name = string("transpose_545")];
            tensor<fp16, [1, 1, ?, 128]> var_3390_cast_fp16 = matmul(transpose_x = var_3390_transpose_x_0, transpose_y = var_3390_transpose_y_0, x = transpose_344, y = transpose_339)[name = string("op_3390_cast_fp16")];
            fp16 var_3391_to_fp16 = const()[name = string("op_3391_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_3392_cast_fp16 = mul(x = var_3390_cast_fp16, y = var_3391_to_fp16)[name = string("op_3392_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_37_cast_fp16 = add(x = var_3392_cast_fp16, y = attention_mask)[name = string("scores_37_cast_fp16")];
            int32 var_3395 = const()[name = string("op_3395"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_37_cast_fp16 = softmax(axis = var_3395, x = scores_37_cast_fp16)[name = string("probs_37_cast_fp16")];
            bool var_3404_transpose_x_0 = const()[name = string("op_3404_transpose_x_0"), val = bool(false)];
            bool var_3404_transpose_y_0 = const()[name = string("op_3404_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_3404_cast_fp16 = matmul(transpose_x = var_3404_transpose_x_0, transpose_y = var_3404_transpose_y_0, x = probs_37_cast_fp16, y = v_t_33_cast_fp16)[name = string("op_3404_cast_fp16")];
            bool var_3423_transpose_x_0 = const()[name = string("op_3423_transpose_x_0"), val = bool(false)];
            bool var_3423_transpose_y_0 = const()[name = string("op_3423_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_345_perm_0 = const()[name = string("transpose_345_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_345 = transpose(perm = transpose_345_perm_0, x = q_h_39_cast_fp16)[name = string("transpose_544")];
            tensor<fp16, [1, 1, ?, 128]> var_3423_cast_fp16 = matmul(transpose_x = var_3423_transpose_x_0, transpose_y = var_3423_transpose_y_0, x = transpose_345, y = transpose_339)[name = string("op_3423_cast_fp16")];
            fp16 var_3424_to_fp16 = const()[name = string("op_3424_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_3425_cast_fp16 = mul(x = var_3423_cast_fp16, y = var_3424_to_fp16)[name = string("op_3425_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_39_cast_fp16 = add(x = var_3425_cast_fp16, y = attention_mask)[name = string("scores_39_cast_fp16")];
            int32 var_3428 = const()[name = string("op_3428"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_39_cast_fp16 = softmax(axis = var_3428, x = scores_39_cast_fp16)[name = string("probs_39_cast_fp16")];
            bool var_3437_transpose_x_0 = const()[name = string("op_3437_transpose_x_0"), val = bool(false)];
            bool var_3437_transpose_y_0 = const()[name = string("op_3437_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_3437_cast_fp16 = matmul(transpose_x = var_3437_transpose_x_0, transpose_y = var_3437_transpose_y_0, x = probs_39_cast_fp16, y = v_t_33_cast_fp16)[name = string("op_3437_cast_fp16")];
            bool input_35_interleave_0 = const()[name = string("input_35_interleave_0"), val = bool(false)];
            int32 const_149 = const()[name = string("const_149"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_35_cast_fp16 = concat(axis = const_149, interleave = input_35_interleave_0, values = (var_3338_cast_fp16, var_3371_cast_fp16, var_3404_cast_fp16, var_3437_cast_fp16))[name = string("input_35_cast_fp16")];
            string x_135_pad_type_0 = const()[name = string("x_135_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_135_strides_0 = const()[name = string("x_135_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_135_pad_0 = const()[name = string("x_135_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_135_dilations_0 = const()[name = string("x_135_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_135_groups_0 = const()[name = string("x_135_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_4_weight_to_fp16 = const()[name = string("o_proj_4_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382105600)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_44_cast_fp16 = transpose(perm = transpose_44_perm_0, x = input_35_cast_fp16)[name = string("transpose_543")];
            tensor<fp16, [1, 640, 1, ?]> x_135_cast_fp16 = conv(dilations = x_135_dilations_0, groups = x_135_groups_0, pad = x_135_pad_0, pad_type = x_135_pad_type_0, strides = x_135_strides_0, weight = o_proj_4_weight_to_fp16, x = transpose_44_cast_fp16)[name = string("x_135_cast_fp16")];
            fp16 var_3458_promoted_to_fp16 = const()[name = string("op_3458_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_3459_cast_fp16 = pow(x = x_135_cast_fp16, y = var_3458_promoted_to_fp16)[name = string("op_3459_cast_fp16")];
            tensor<int32, [1]> var_85_axes_0 = const()[name = string("var_85_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_85_keep_dims_0 = const()[name = string("var_85_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_85_cast_fp16 = reduce_mean(axes = var_85_axes_0, keep_dims = var_85_keep_dims_0, x = var_3459_cast_fp16)[name = string("var_85_cast_fp16")];
            fp16 var_3466_to_fp16 = const()[name = string("op_3466_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_3467_cast_fp16 = add(x = var_85_cast_fp16, y = var_3466_to_fp16)[name = string("op_3467_cast_fp16")];
            fp32 var_3468_epsilon_0 = const()[name = string("op_3468_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_3468_cast_fp16 = rsqrt(epsilon = var_3468_epsilon_0, x = var_3467_cast_fp16)[name = string("op_3468_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_85_cast_fp16 = mul(x = x_135_cast_fp16, y = var_3468_cast_fp16)[name = string("x_normed_85_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_3478_to_fp16 = const()[name = string("op_3478_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383416384)))];
            tensor<fp16, [1, 640, 1, ?]> attn_19_cast_fp16 = mul(x = x_normed_85_cast_fp16, y = var_3478_to_fp16)[name = string("attn_19_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_137_cast_fp16 = add(x = x_113_cast_fp16, y = attn_19_cast_fp16)[name = string("x_137_cast_fp16")];
            fp16 var_3482_promoted_to_fp16 = const()[name = string("op_3482_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_3483_cast_fp16 = pow(x = x_137_cast_fp16, y = var_3482_promoted_to_fp16)[name = string("op_3483_cast_fp16")];
            tensor<int32, [1]> var_87_axes_0 = const()[name = string("var_87_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_87_keep_dims_0 = const()[name = string("var_87_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_87_cast_fp16 = reduce_mean(axes = var_87_axes_0, keep_dims = var_87_keep_dims_0, x = var_3483_cast_fp16)[name = string("var_87_cast_fp16")];
            fp16 var_3490_to_fp16 = const()[name = string("op_3490_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_3491_cast_fp16 = add(x = var_87_cast_fp16, y = var_3490_to_fp16)[name = string("op_3491_cast_fp16")];
            fp32 var_3492_epsilon_0 = const()[name = string("op_3492_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_3492_cast_fp16 = rsqrt(epsilon = var_3492_epsilon_0, x = var_3491_cast_fp16)[name = string("op_3492_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_87_cast_fp16 = mul(x = x_137_cast_fp16, y = var_3492_cast_fp16)[name = string("x_normed_87_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_3502_to_fp16 = const()[name = string("op_3502_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383417728)))];
            tensor<fp16, [1, 640, 1, ?]> input_37_cast_fp16 = mul(x = x_normed_87_cast_fp16, y = var_3502_to_fp16)[name = string("input_37_cast_fp16")];
            string gate_9_pad_type_0 = const()[name = string("gate_9_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_9_strides_0 = const()[name = string("gate_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_9_pad_0 = const()[name = string("gate_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_9_dilations_0 = const()[name = string("gate_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_9_groups_0 = const()[name = string("gate_9_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_4_weight_to_fp16 = const()[name = string("gate_proj_4_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383419072)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_9_cast_fp16 = conv(dilations = gate_9_dilations_0, groups = gate_9_groups_0, pad = gate_9_pad_0, pad_type = gate_9_pad_type_0, strides = gate_9_strides_0, weight = gate_proj_4_weight_to_fp16, x = input_37_cast_fp16)[name = string("gate_9_cast_fp16")];
            string up_9_pad_type_0 = const()[name = string("up_9_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_9_strides_0 = const()[name = string("up_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_9_pad_0 = const()[name = string("up_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_9_dilations_0 = const()[name = string("up_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_9_groups_0 = const()[name = string("up_9_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_4_weight_to_fp16 = const()[name = string("up_proj_4_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386040576)))];
            tensor<fp16, [1, 2048, 1, ?]> up_9_cast_fp16 = conv(dilations = up_9_dilations_0, groups = up_9_groups_0, pad = up_9_pad_0, pad_type = up_9_pad_type_0, strides = up_9_strides_0, weight = up_proj_4_weight_to_fp16, x = input_37_cast_fp16)[name = string("up_9_cast_fp16")];
            string var_3527_mode_0 = const()[name = string("op_3527_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_3527_cast_fp16 = gelu(mode = var_3527_mode_0, x = gate_9_cast_fp16)[name = string("op_3527_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_39_cast_fp16 = mul(x = var_3527_cast_fp16, y = up_9_cast_fp16)[name = string("input_39_cast_fp16")];
            string x_139_pad_type_0 = const()[name = string("x_139_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_139_strides_0 = const()[name = string("x_139_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_139_pad_0 = const()[name = string("x_139_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_139_dilations_0 = const()[name = string("x_139_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_139_groups_0 = const()[name = string("x_139_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_4_weight_to_fp16 = const()[name = string("down_proj_4_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388662080)))];
            tensor<fp16, [1, 640, 1, ?]> x_139_cast_fp16 = conv(dilations = x_139_dilations_0, groups = x_139_groups_0, pad = x_139_pad_0, pad_type = x_139_pad_type_0, strides = x_139_strides_0, weight = down_proj_4_weight_to_fp16, x = input_39_cast_fp16)[name = string("x_139_cast_fp16")];
            fp16 var_3540_promoted_to_fp16 = const()[name = string("op_3540_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_3541_cast_fp16 = pow(x = x_139_cast_fp16, y = var_3540_promoted_to_fp16)[name = string("op_3541_cast_fp16")];
            tensor<int32, [1]> var_89_axes_0 = const()[name = string("var_89_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_89_keep_dims_0 = const()[name = string("var_89_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_89_cast_fp16 = reduce_mean(axes = var_89_axes_0, keep_dims = var_89_keep_dims_0, x = var_3541_cast_fp16)[name = string("var_89_cast_fp16")];
            fp16 var_3548_to_fp16 = const()[name = string("op_3548_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_3549_cast_fp16 = add(x = var_89_cast_fp16, y = var_3548_to_fp16)[name = string("op_3549_cast_fp16")];
            fp32 var_3550_epsilon_0 = const()[name = string("op_3550_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_3550_cast_fp16 = rsqrt(epsilon = var_3550_epsilon_0, x = var_3549_cast_fp16)[name = string("op_3550_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_89_cast_fp16 = mul(x = x_139_cast_fp16, y = var_3550_cast_fp16)[name = string("x_normed_89_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_3560_to_fp16 = const()[name = string("op_3560_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391283584)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_9_cast_fp16 = mul(x = x_normed_89_cast_fp16, y = var_3560_to_fp16)[name = string("mlp_9_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_141_cast_fp16 = add(x = x_137_cast_fp16, y = mlp_9_cast_fp16)[name = string("x_141_cast_fp16")];
            fp16 var_3564_promoted_to_fp16 = const()[name = string("op_3564_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_3565_cast_fp16 = pow(x = x_141_cast_fp16, y = var_3564_promoted_to_fp16)[name = string("op_3565_cast_fp16")];
            tensor<int32, [1]> var_91_axes_0 = const()[name = string("var_91_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_91_keep_dims_0 = const()[name = string("var_91_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_91_cast_fp16 = reduce_mean(axes = var_91_axes_0, keep_dims = var_91_keep_dims_0, x = var_3565_cast_fp16)[name = string("var_91_cast_fp16")];
            fp16 var_3572_to_fp16 = const()[name = string("op_3572_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_3573_cast_fp16 = add(x = var_91_cast_fp16, y = var_3572_to_fp16)[name = string("op_3573_cast_fp16")];
            fp32 var_3574_epsilon_0 = const()[name = string("op_3574_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_3574_cast_fp16 = rsqrt(epsilon = var_3574_epsilon_0, x = var_3573_cast_fp16)[name = string("op_3574_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_91_cast_fp16 = mul(x = x_141_cast_fp16, y = var_3574_cast_fp16)[name = string("x_normed_91_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_3584_to_fp16 = const()[name = string("op_3584_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391284928)))];
            tensor<fp16, [1, 640, 1, ?]> input_41_cast_fp16 = mul(x = x_normed_91_cast_fp16, y = var_3584_to_fp16)[name = string("input_41_cast_fp16")];
            string var_3596_pad_type_0 = const()[name = string("op_3596_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3596_strides_0 = const()[name = string("op_3596_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3596_pad_0 = const()[name = string("op_3596_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3596_dilations_0 = const()[name = string("op_3596_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3596_groups_0 = const()[name = string("op_3596_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_5_weight_to_fp16 = const()[name = string("q_proj_5_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391286272)))];
            tensor<fp16, [1, 1024, 1, ?]> var_3596_cast_fp16 = conv(dilations = var_3596_dilations_0, groups = var_3596_groups_0, pad = var_3596_pad_0, pad_type = var_3596_pad_type_0, strides = var_3596_strides_0, weight = q_proj_5_weight_to_fp16, x = input_41_cast_fp16)[name = string("op_3596_cast_fp16")];
            string var_3607_pad_type_0 = const()[name = string("op_3607_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3607_strides_0 = const()[name = string("op_3607_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3607_pad_0 = const()[name = string("op_3607_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3607_dilations_0 = const()[name = string("op_3607_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3607_groups_0 = const()[name = string("op_3607_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_5_weight_to_fp16 = const()[name = string("k_proj_5_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392597056)))];
            tensor<fp16, [1, 256, 1, ?]> var_3607_cast_fp16 = conv(dilations = var_3607_dilations_0, groups = var_3607_groups_0, pad = var_3607_pad_0, pad_type = var_3607_pad_type_0, strides = var_3607_strides_0, weight = k_proj_5_weight_to_fp16, x = input_41_cast_fp16)[name = string("op_3607_cast_fp16")];
            string var_3618_pad_type_0 = const()[name = string("op_3618_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3618_strides_0 = const()[name = string("op_3618_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3618_pad_0 = const()[name = string("op_3618_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3618_dilations_0 = const()[name = string("op_3618_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3618_groups_0 = const()[name = string("op_3618_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_5_weight_to_fp16 = const()[name = string("v_proj_5_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392924800)))];
            tensor<fp16, [1, 256, 1, ?]> var_3618_cast_fp16 = conv(dilations = var_3618_dilations_0, groups = var_3618_groups_0, pad = var_3618_pad_0, pad_type = var_3618_pad_type_0, strides = var_3618_strides_0, weight = v_proj_5_weight_to_fp16, x = input_41_cast_fp16)[name = string("op_3618_cast_fp16")];
            tensor<int32, [4]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_3621_axis_0 = const()[name = string("op_3621_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_3621_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_3621_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_3621_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_3621_cast_fp16_3 = split(axis = var_3621_axis_0, split_sizes = tile_15, x = var_3596_cast_fp16)[name = string("op_3621_cast_fp16")];
            fp16 var_3634_promoted_to_fp16 = const()[name = string("op_3634_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_3635_cast_fp16 = pow(x = var_3621_cast_fp16_0, y = var_3634_promoted_to_fp16)[name = string("op_3635_cast_fp16")];
            tensor<int32, [1]> var_93_axes_0 = const()[name = string("var_93_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_93_keep_dims_0 = const()[name = string("var_93_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_93_cast_fp16 = reduce_mean(axes = var_93_axes_0, keep_dims = var_93_keep_dims_0, x = var_3635_cast_fp16)[name = string("var_93_cast_fp16")];
            fp16 var_3642_to_fp16 = const()[name = string("op_3642_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_3643_cast_fp16 = add(x = var_93_cast_fp16, y = var_3642_to_fp16)[name = string("op_3643_cast_fp16")];
            fp32 var_3644_epsilon_0 = const()[name = string("op_3644_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_3644_cast_fp16 = rsqrt(epsilon = var_3644_epsilon_0, x = var_3643_cast_fp16)[name = string("op_3644_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_93_cast_fp16 = mul(x = var_3621_cast_fp16_0, y = var_3644_cast_fp16)[name = string("x_normed_93_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_3654_to_fp16 = const()[name = string("op_3654_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393252544)))];
            tensor<fp16, [1, 256, 1, ?]> x_145_cast_fp16 = mul(x = x_normed_93_cast_fp16, y = var_3654_to_fp16)[name = string("x_145_cast_fp16")];
            tensor<int32, [4]> x_first_51_begin_0 = const()[name = string("x_first_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_51_end_0 = const()[name = string("x_first_51_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_51_end_mask_0 = const()[name = string("x_first_51_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_51_cast_fp16 = slice_by_index(begin = x_first_51_begin_0, end = x_first_51_end_0, end_mask = x_first_51_end_mask_0, x = x_145_cast_fp16)[name = string("x_first_51_cast_fp16")];
            tensor<int32, [4]> x_second_51_begin_0 = const()[name = string("x_second_51_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_51_end_0 = const()[name = string("x_second_51_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_51_end_mask_0 = const()[name = string("x_second_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_51_cast_fp16 = slice_by_index(begin = x_second_51_begin_0, end = x_second_51_end_0, end_mask = x_second_51_end_mask_0, x = x_145_cast_fp16)[name = string("x_second_51_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_3676_cast_fp16 = mul(x = x_second_51_cast_fp16, y = const_44_promoted_to_fp16)[name = string("op_3676_cast_fp16")];
            int32 var_3678 = const()[name = string("op_3678"), val = int32(1)];
            bool x_rot_51_interleave_0 = const()[name = string("x_rot_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_51_cast_fp16 = concat(axis = var_3678, interleave = x_rot_51_interleave_0, values = (var_3676_cast_fp16, x_first_51_cast_fp16))[name = string("x_rot_51_cast_fp16")];
            tensor<int32, [3]> var_3683 = const()[name = string("op_3683"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> cos_bc1t_51_axes_0 = const()[name = string("cos_bc1t_51_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 256, ?]> var_3684_cast_fp16 = transpose(perm = var_3683, x = cos_full)[name = string("transpose_542")];
            tensor<fp16, [1, 256, 1, ?]> cos_bc1t_51_cast_fp16 = expand_dims(axes = cos_bc1t_51_axes_0, x = var_3684_cast_fp16)[name = string("cos_bc1t_51_cast_fp16")];
            tensor<int32, [3]> var_3690 = const()[name = string("op_3690"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> sin_bc1t_51_axes_0 = const()[name = string("sin_bc1t_51_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 256, ?]> var_3691_cast_fp16 = transpose(perm = var_3690, x = sin_full)[name = string("transpose_541")];
            tensor<fp16, [1, 256, 1, ?]> sin_bc1t_51_cast_fp16 = expand_dims(axes = sin_bc1t_51_axes_0, x = var_3691_cast_fp16)[name = string("sin_bc1t_51_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3694_cast_fp16 = mul(x = x_145_cast_fp16, y = cos_bc1t_51_cast_fp16)[name = string("op_3694_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3695_cast_fp16 = mul(x = x_rot_51_cast_fp16, y = sin_bc1t_51_cast_fp16)[name = string("op_3695_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_41_cast_fp16 = add(x = var_3694_cast_fp16, y = var_3695_cast_fp16)[name = string("q_h_41_cast_fp16")];
            fp16 var_3698_promoted_to_fp16 = const()[name = string("op_3698_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_3699_cast_fp16 = pow(x = var_3621_cast_fp16_1, y = var_3698_promoted_to_fp16)[name = string("op_3699_cast_fp16")];
            tensor<int32, [1]> var_95_axes_0 = const()[name = string("var_95_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_95_keep_dims_0 = const()[name = string("var_95_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_95_cast_fp16 = reduce_mean(axes = var_95_axes_0, keep_dims = var_95_keep_dims_0, x = var_3699_cast_fp16)[name = string("var_95_cast_fp16")];
            fp16 var_3706_to_fp16 = const()[name = string("op_3706_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_3707_cast_fp16 = add(x = var_95_cast_fp16, y = var_3706_to_fp16)[name = string("op_3707_cast_fp16")];
            fp32 var_3708_epsilon_0 = const()[name = string("op_3708_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_3708_cast_fp16 = rsqrt(epsilon = var_3708_epsilon_0, x = var_3707_cast_fp16)[name = string("op_3708_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_95_cast_fp16 = mul(x = var_3621_cast_fp16_1, y = var_3708_cast_fp16)[name = string("x_normed_95_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_149_cast_fp16 = mul(x = x_normed_95_cast_fp16, y = var_3654_to_fp16)[name = string("x_149_cast_fp16")];
            tensor<int32, [4]> x_first_53_begin_0 = const()[name = string("x_first_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_53_end_0 = const()[name = string("x_first_53_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_53_end_mask_0 = const()[name = string("x_first_53_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_53_cast_fp16 = slice_by_index(begin = x_first_53_begin_0, end = x_first_53_end_0, end_mask = x_first_53_end_mask_0, x = x_149_cast_fp16)[name = string("x_first_53_cast_fp16")];
            tensor<int32, [4]> x_second_53_begin_0 = const()[name = string("x_second_53_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_53_end_0 = const()[name = string("x_second_53_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_53_end_mask_0 = const()[name = string("x_second_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_53_cast_fp16 = slice_by_index(begin = x_second_53_begin_0, end = x_second_53_end_0, end_mask = x_second_53_end_mask_0, x = x_149_cast_fp16)[name = string("x_second_53_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_3740_cast_fp16 = mul(x = x_second_53_cast_fp16, y = const_45_promoted_to_fp16)[name = string("op_3740_cast_fp16")];
            int32 var_3742 = const()[name = string("op_3742"), val = int32(1)];
            bool x_rot_53_interleave_0 = const()[name = string("x_rot_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_53_cast_fp16 = concat(axis = var_3742, interleave = x_rot_53_interleave_0, values = (var_3740_cast_fp16, x_first_53_cast_fp16))[name = string("x_rot_53_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3758_cast_fp16 = mul(x = x_149_cast_fp16, y = cos_bc1t_51_cast_fp16)[name = string("op_3758_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3759_cast_fp16 = mul(x = x_rot_53_cast_fp16, y = sin_bc1t_51_cast_fp16)[name = string("op_3759_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_43_cast_fp16 = add(x = var_3758_cast_fp16, y = var_3759_cast_fp16)[name = string("q_h_43_cast_fp16")];
            fp16 var_3762_promoted_to_fp16 = const()[name = string("op_3762_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_3763_cast_fp16 = pow(x = var_3621_cast_fp16_2, y = var_3762_promoted_to_fp16)[name = string("op_3763_cast_fp16")];
            tensor<int32, [1]> var_97_axes_0 = const()[name = string("var_97_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_97_keep_dims_0 = const()[name = string("var_97_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_97_cast_fp16 = reduce_mean(axes = var_97_axes_0, keep_dims = var_97_keep_dims_0, x = var_3763_cast_fp16)[name = string("var_97_cast_fp16")];
            fp16 var_3770_to_fp16 = const()[name = string("op_3770_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_3771_cast_fp16 = add(x = var_97_cast_fp16, y = var_3770_to_fp16)[name = string("op_3771_cast_fp16")];
            fp32 var_3772_epsilon_0 = const()[name = string("op_3772_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_3772_cast_fp16 = rsqrt(epsilon = var_3772_epsilon_0, x = var_3771_cast_fp16)[name = string("op_3772_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_97_cast_fp16 = mul(x = var_3621_cast_fp16_2, y = var_3772_cast_fp16)[name = string("x_normed_97_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_153_cast_fp16 = mul(x = x_normed_97_cast_fp16, y = var_3654_to_fp16)[name = string("x_153_cast_fp16")];
            tensor<int32, [4]> x_first_55_begin_0 = const()[name = string("x_first_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_55_end_0 = const()[name = string("x_first_55_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_55_end_mask_0 = const()[name = string("x_first_55_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_55_cast_fp16 = slice_by_index(begin = x_first_55_begin_0, end = x_first_55_end_0, end_mask = x_first_55_end_mask_0, x = x_153_cast_fp16)[name = string("x_first_55_cast_fp16")];
            tensor<int32, [4]> x_second_55_begin_0 = const()[name = string("x_second_55_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_55_end_0 = const()[name = string("x_second_55_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_55_end_mask_0 = const()[name = string("x_second_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_55_cast_fp16 = slice_by_index(begin = x_second_55_begin_0, end = x_second_55_end_0, end_mask = x_second_55_end_mask_0, x = x_153_cast_fp16)[name = string("x_second_55_cast_fp16")];
            fp16 const_46_promoted_to_fp16 = const()[name = string("const_46_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_3804_cast_fp16 = mul(x = x_second_55_cast_fp16, y = const_46_promoted_to_fp16)[name = string("op_3804_cast_fp16")];
            int32 var_3806 = const()[name = string("op_3806"), val = int32(1)];
            bool x_rot_55_interleave_0 = const()[name = string("x_rot_55_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_55_cast_fp16 = concat(axis = var_3806, interleave = x_rot_55_interleave_0, values = (var_3804_cast_fp16, x_first_55_cast_fp16))[name = string("x_rot_55_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3822_cast_fp16 = mul(x = x_153_cast_fp16, y = cos_bc1t_51_cast_fp16)[name = string("op_3822_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3823_cast_fp16 = mul(x = x_rot_55_cast_fp16, y = sin_bc1t_51_cast_fp16)[name = string("op_3823_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_45_cast_fp16 = add(x = var_3822_cast_fp16, y = var_3823_cast_fp16)[name = string("q_h_45_cast_fp16")];
            fp16 var_3826_promoted_to_fp16 = const()[name = string("op_3826_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_3827_cast_fp16 = pow(x = var_3621_cast_fp16_3, y = var_3826_promoted_to_fp16)[name = string("op_3827_cast_fp16")];
            tensor<int32, [1]> var_99_axes_0 = const()[name = string("var_99_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_99_keep_dims_0 = const()[name = string("var_99_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_99_cast_fp16 = reduce_mean(axes = var_99_axes_0, keep_dims = var_99_keep_dims_0, x = var_3827_cast_fp16)[name = string("var_99_cast_fp16")];
            fp16 var_3834_to_fp16 = const()[name = string("op_3834_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_3835_cast_fp16 = add(x = var_99_cast_fp16, y = var_3834_to_fp16)[name = string("op_3835_cast_fp16")];
            fp32 var_3836_epsilon_0 = const()[name = string("op_3836_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_3836_cast_fp16 = rsqrt(epsilon = var_3836_epsilon_0, x = var_3835_cast_fp16)[name = string("op_3836_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_99_cast_fp16 = mul(x = var_3621_cast_fp16_3, y = var_3836_cast_fp16)[name = string("x_normed_99_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_157_cast_fp16 = mul(x = x_normed_99_cast_fp16, y = var_3654_to_fp16)[name = string("x_157_cast_fp16")];
            tensor<int32, [4]> x_first_57_begin_0 = const()[name = string("x_first_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_57_end_0 = const()[name = string("x_first_57_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_57_end_mask_0 = const()[name = string("x_first_57_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_57_cast_fp16 = slice_by_index(begin = x_first_57_begin_0, end = x_first_57_end_0, end_mask = x_first_57_end_mask_0, x = x_157_cast_fp16)[name = string("x_first_57_cast_fp16")];
            tensor<int32, [4]> x_second_57_begin_0 = const()[name = string("x_second_57_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_57_end_0 = const()[name = string("x_second_57_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_57_end_mask_0 = const()[name = string("x_second_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_57_cast_fp16 = slice_by_index(begin = x_second_57_begin_0, end = x_second_57_end_0, end_mask = x_second_57_end_mask_0, x = x_157_cast_fp16)[name = string("x_second_57_cast_fp16")];
            fp16 const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_3868_cast_fp16 = mul(x = x_second_57_cast_fp16, y = const_47_promoted_to_fp16)[name = string("op_3868_cast_fp16")];
            int32 var_3870 = const()[name = string("op_3870"), val = int32(1)];
            bool x_rot_57_interleave_0 = const()[name = string("x_rot_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_57_cast_fp16 = concat(axis = var_3870, interleave = x_rot_57_interleave_0, values = (var_3868_cast_fp16, x_first_57_cast_fp16))[name = string("x_rot_57_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3886_cast_fp16 = mul(x = x_157_cast_fp16, y = cos_bc1t_51_cast_fp16)[name = string("op_3886_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3887_cast_fp16 = mul(x = x_rot_57_cast_fp16, y = sin_bc1t_51_cast_fp16)[name = string("op_3887_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_47_cast_fp16 = add(x = var_3886_cast_fp16, y = var_3887_cast_fp16)[name = string("q_h_47_cast_fp16")];
            fp16 var_3890_promoted_to_fp16 = const()[name = string("op_3890_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_3891_cast_fp16 = pow(x = var_3607_cast_fp16, y = var_3890_promoted_to_fp16)[name = string("op_3891_cast_fp16")];
            tensor<int32, [1]> var_101_axes_0 = const()[name = string("var_101_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_101_keep_dims_0 = const()[name = string("var_101_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_101_cast_fp16 = reduce_mean(axes = var_101_axes_0, keep_dims = var_101_keep_dims_0, x = var_3891_cast_fp16)[name = string("var_101_cast_fp16")];
            fp16 var_3898_to_fp16 = const()[name = string("op_3898_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_3899_cast_fp16 = add(x = var_101_cast_fp16, y = var_3898_to_fp16)[name = string("op_3899_cast_fp16")];
            fp32 var_3900_epsilon_0 = const()[name = string("op_3900_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_3900_cast_fp16 = rsqrt(epsilon = var_3900_epsilon_0, x = var_3899_cast_fp16)[name = string("op_3900_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_101_cast_fp16 = mul(x = var_3607_cast_fp16, y = var_3900_cast_fp16)[name = string("x_normed_101_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_3910_to_fp16 = const()[name = string("op_3910_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393253120)))];
            tensor<fp16, [1, 256, 1, ?]> x_161_cast_fp16 = mul(x = x_normed_101_cast_fp16, y = var_3910_to_fp16)[name = string("x_161_cast_fp16")];
            tensor<int32, [4]> x_first_59_begin_0 = const()[name = string("x_first_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_59_end_0 = const()[name = string("x_first_59_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_59_end_mask_0 = const()[name = string("x_first_59_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_59_cast_fp16 = slice_by_index(begin = x_first_59_begin_0, end = x_first_59_end_0, end_mask = x_first_59_end_mask_0, x = x_161_cast_fp16)[name = string("x_first_59_cast_fp16")];
            tensor<int32, [4]> x_second_59_begin_0 = const()[name = string("x_second_59_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_59_end_0 = const()[name = string("x_second_59_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_59_end_mask_0 = const()[name = string("x_second_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_59_cast_fp16 = slice_by_index(begin = x_second_59_begin_0, end = x_second_59_end_0, end_mask = x_second_59_end_mask_0, x = x_161_cast_fp16)[name = string("x_second_59_cast_fp16")];
            fp16 const_48_promoted_to_fp16 = const()[name = string("const_48_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_3932_cast_fp16 = mul(x = x_second_59_cast_fp16, y = const_48_promoted_to_fp16)[name = string("op_3932_cast_fp16")];
            int32 var_3934 = const()[name = string("op_3934"), val = int32(1)];
            bool x_rot_59_interleave_0 = const()[name = string("x_rot_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_59_cast_fp16 = concat(axis = var_3934, interleave = x_rot_59_interleave_0, values = (var_3932_cast_fp16, x_first_59_cast_fp16))[name = string("x_rot_59_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3950_cast_fp16 = mul(x = x_161_cast_fp16, y = cos_bc1t_51_cast_fp16)[name = string("op_3950_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_3951_cast_fp16 = mul(x = x_rot_59_cast_fp16, y = sin_bc1t_51_cast_fp16)[name = string("op_3951_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_11_cast_fp16 = add(x = var_3950_cast_fp16, y = var_3951_cast_fp16)[name = string("k_new_11_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_10 = read_state(input = k_cache_5)[name = string("read_state_10")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_33_cast_fp16 = mul(x = read_state_10, y = one_minus_1_cast_fp16)[name = string("k_cache_33_cast_fp16")];
            write_state(data = k_cache_33_cast_fp16, input = k_cache_5)[name = string("coreml_update_state_92_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_92 = read_state(input = k_cache_5)[name = string("coreml_update_state_92")];
            tensor<fp16, [1, 256, 1, 128]> var_3958_cast_fp16 = mul(x = k_new_11_cast_fp16, y = write_mask)[name = string("op_3958_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_35_cast_fp16 = add(x = coreml_update_state_92, y = var_3958_cast_fp16)[name = string("k_cache_35_cast_fp16")];
            write_state(data = k_cache_35_cast_fp16, input = k_cache_5)[name = string("coreml_update_state_93_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_93 = read_state(input = k_cache_5)[name = string("coreml_update_state_93")];
            tensor<fp16, [1, 256, 1, 128]> read_state_11 = read_state(input = v_cache_5)[name = string("read_state_11")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_33_cast_fp16 = mul(x = read_state_11, y = one_minus_1_cast_fp16)[name = string("v_cache_33_cast_fp16")];
            write_state(data = v_cache_33_cast_fp16, input = v_cache_5)[name = string("coreml_update_state_94_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_94 = read_state(input = v_cache_5)[name = string("coreml_update_state_94")];
            tensor<fp16, [1, 256, 1, 128]> var_3962_cast_fp16 = mul(x = var_3618_cast_fp16, y = write_mask)[name = string("op_3962_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_35_cast_fp16 = add(x = coreml_update_state_94, y = var_3962_cast_fp16)[name = string("v_cache_35_cast_fp16")];
            write_state(data = v_cache_35_cast_fp16, input = v_cache_5)[name = string("coreml_update_state_95_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_95 = read_state(input = v_cache_5)[name = string("coreml_update_state_95")];
            bool var_3977_transpose_x_0 = const()[name = string("op_3977_transpose_x_0"), val = bool(false)];
            bool var_3977_transpose_y_0 = const()[name = string("op_3977_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_346_perm_0 = const()[name = string("transpose_346_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_347_perm_0 = const()[name = string("transpose_347_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_347 = transpose(perm = transpose_347_perm_0, x = coreml_update_state_93)[name = string("transpose_539")];
            tensor<fp16, [1, 1, ?, 256]> transpose_346 = transpose(perm = transpose_346_perm_0, x = q_h_41_cast_fp16)[name = string("transpose_540")];
            tensor<fp16, [1, 1, ?, 128]> var_3977_cast_fp16 = matmul(transpose_x = var_3977_transpose_x_0, transpose_y = var_3977_transpose_y_0, x = transpose_346, y = transpose_347)[name = string("op_3977_cast_fp16")];
            fp16 var_3978_to_fp16 = const()[name = string("op_3978_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_3979_cast_fp16 = mul(x = var_3977_cast_fp16, y = var_3978_to_fp16)[name = string("op_3979_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_41_cast_fp16 = add(x = var_3979_cast_fp16, y = attention_mask)[name = string("scores_41_cast_fp16")];
            int32 var_3982 = const()[name = string("op_3982"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_41_cast_fp16 = softmax(axis = var_3982, x = scores_41_cast_fp16)[name = string("probs_41_cast_fp16")];
            tensor<int32, [4]> var_3989 = const()[name = string("op_3989"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_3991_transpose_x_0 = const()[name = string("op_3991_transpose_x_0"), val = bool(false)];
            bool var_3991_transpose_y_0 = const()[name = string("op_3991_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_41_cast_fp16 = transpose(perm = var_3989, x = coreml_update_state_95)[name = string("transpose_538")];
            tensor<fp16, [1, 1, ?, 256]> var_3991_cast_fp16 = matmul(transpose_x = var_3991_transpose_x_0, transpose_y = var_3991_transpose_y_0, x = probs_41_cast_fp16, y = v_t_41_cast_fp16)[name = string("op_3991_cast_fp16")];
            bool var_4010_transpose_x_0 = const()[name = string("op_4010_transpose_x_0"), val = bool(false)];
            bool var_4010_transpose_y_0 = const()[name = string("op_4010_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_351_perm_0 = const()[name = string("transpose_351_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_351 = transpose(perm = transpose_351_perm_0, x = q_h_43_cast_fp16)[name = string("transpose_537")];
            tensor<fp16, [1, 1, ?, 128]> var_4010_cast_fp16 = matmul(transpose_x = var_4010_transpose_x_0, transpose_y = var_4010_transpose_y_0, x = transpose_351, y = transpose_347)[name = string("op_4010_cast_fp16")];
            fp16 var_4011_to_fp16 = const()[name = string("op_4011_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_4012_cast_fp16 = mul(x = var_4010_cast_fp16, y = var_4011_to_fp16)[name = string("op_4012_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_43_cast_fp16 = add(x = var_4012_cast_fp16, y = attention_mask)[name = string("scores_43_cast_fp16")];
            int32 var_4015 = const()[name = string("op_4015"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_43_cast_fp16 = softmax(axis = var_4015, x = scores_43_cast_fp16)[name = string("probs_43_cast_fp16")];
            bool var_4024_transpose_x_0 = const()[name = string("op_4024_transpose_x_0"), val = bool(false)];
            bool var_4024_transpose_y_0 = const()[name = string("op_4024_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_4024_cast_fp16 = matmul(transpose_x = var_4024_transpose_x_0, transpose_y = var_4024_transpose_y_0, x = probs_43_cast_fp16, y = v_t_41_cast_fp16)[name = string("op_4024_cast_fp16")];
            bool var_4043_transpose_x_0 = const()[name = string("op_4043_transpose_x_0"), val = bool(false)];
            bool var_4043_transpose_y_0 = const()[name = string("op_4043_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_352_perm_0 = const()[name = string("transpose_352_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_352 = transpose(perm = transpose_352_perm_0, x = q_h_45_cast_fp16)[name = string("transpose_536")];
            tensor<fp16, [1, 1, ?, 128]> var_4043_cast_fp16 = matmul(transpose_x = var_4043_transpose_x_0, transpose_y = var_4043_transpose_y_0, x = transpose_352, y = transpose_347)[name = string("op_4043_cast_fp16")];
            fp16 var_4044_to_fp16 = const()[name = string("op_4044_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_4045_cast_fp16 = mul(x = var_4043_cast_fp16, y = var_4044_to_fp16)[name = string("op_4045_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_45_cast_fp16 = add(x = var_4045_cast_fp16, y = attention_mask)[name = string("scores_45_cast_fp16")];
            int32 var_4048 = const()[name = string("op_4048"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_45_cast_fp16 = softmax(axis = var_4048, x = scores_45_cast_fp16)[name = string("probs_45_cast_fp16")];
            bool var_4057_transpose_x_0 = const()[name = string("op_4057_transpose_x_0"), val = bool(false)];
            bool var_4057_transpose_y_0 = const()[name = string("op_4057_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_4057_cast_fp16 = matmul(transpose_x = var_4057_transpose_x_0, transpose_y = var_4057_transpose_y_0, x = probs_45_cast_fp16, y = v_t_41_cast_fp16)[name = string("op_4057_cast_fp16")];
            bool var_4076_transpose_x_0 = const()[name = string("op_4076_transpose_x_0"), val = bool(false)];
            bool var_4076_transpose_y_0 = const()[name = string("op_4076_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_353_perm_0 = const()[name = string("transpose_353_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_353 = transpose(perm = transpose_353_perm_0, x = q_h_47_cast_fp16)[name = string("transpose_535")];
            tensor<fp16, [1, 1, ?, 128]> var_4076_cast_fp16 = matmul(transpose_x = var_4076_transpose_x_0, transpose_y = var_4076_transpose_y_0, x = transpose_353, y = transpose_347)[name = string("op_4076_cast_fp16")];
            fp16 var_4077_to_fp16 = const()[name = string("op_4077_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_4078_cast_fp16 = mul(x = var_4076_cast_fp16, y = var_4077_to_fp16)[name = string("op_4078_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_47_cast_fp16 = add(x = var_4078_cast_fp16, y = attention_mask)[name = string("scores_47_cast_fp16")];
            int32 var_4081 = const()[name = string("op_4081"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_47_cast_fp16 = softmax(axis = var_4081, x = scores_47_cast_fp16)[name = string("probs_47_cast_fp16")];
            bool var_4090_transpose_x_0 = const()[name = string("op_4090_transpose_x_0"), val = bool(false)];
            bool var_4090_transpose_y_0 = const()[name = string("op_4090_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_4090_cast_fp16 = matmul(transpose_x = var_4090_transpose_x_0, transpose_y = var_4090_transpose_y_0, x = probs_47_cast_fp16, y = v_t_41_cast_fp16)[name = string("op_4090_cast_fp16")];
            bool input_43_interleave_0 = const()[name = string("input_43_interleave_0"), val = bool(false)];
            int32 const_150 = const()[name = string("const_150"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_43_cast_fp16 = concat(axis = const_150, interleave = input_43_interleave_0, values = (var_3991_cast_fp16, var_4024_cast_fp16, var_4057_cast_fp16, var_4090_cast_fp16))[name = string("input_43_cast_fp16")];
            string x_163_pad_type_0 = const()[name = string("x_163_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_163_strides_0 = const()[name = string("x_163_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_163_pad_0 = const()[name = string("x_163_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_163_dilations_0 = const()[name = string("x_163_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_163_groups_0 = const()[name = string("x_163_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_53_perm_0 = const()[name = string("transpose_53_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_5_weight_to_fp16 = const()[name = string("o_proj_5_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393253696)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_53_cast_fp16 = transpose(perm = transpose_53_perm_0, x = input_43_cast_fp16)[name = string("transpose_534")];
            tensor<fp16, [1, 640, 1, ?]> x_163_cast_fp16 = conv(dilations = x_163_dilations_0, groups = x_163_groups_0, pad = x_163_pad_0, pad_type = x_163_pad_type_0, strides = x_163_strides_0, weight = o_proj_5_weight_to_fp16, x = transpose_53_cast_fp16)[name = string("x_163_cast_fp16")];
            fp16 var_4111_promoted_to_fp16 = const()[name = string("op_4111_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_4112_cast_fp16 = pow(x = x_163_cast_fp16, y = var_4111_promoted_to_fp16)[name = string("op_4112_cast_fp16")];
            tensor<int32, [1]> var_103_axes_0 = const()[name = string("var_103_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_103_keep_dims_0 = const()[name = string("var_103_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_103_cast_fp16 = reduce_mean(axes = var_103_axes_0, keep_dims = var_103_keep_dims_0, x = var_4112_cast_fp16)[name = string("var_103_cast_fp16")];
            fp16 var_4119_to_fp16 = const()[name = string("op_4119_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_4120_cast_fp16 = add(x = var_103_cast_fp16, y = var_4119_to_fp16)[name = string("op_4120_cast_fp16")];
            fp32 var_4121_epsilon_0 = const()[name = string("op_4121_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_4121_cast_fp16 = rsqrt(epsilon = var_4121_epsilon_0, x = var_4120_cast_fp16)[name = string("op_4121_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_103_cast_fp16 = mul(x = x_163_cast_fp16, y = var_4121_cast_fp16)[name = string("x_normed_103_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_4131_to_fp16 = const()[name = string("op_4131_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394564480)))];
            tensor<fp16, [1, 640, 1, ?]> attn_23_cast_fp16 = mul(x = x_normed_103_cast_fp16, y = var_4131_to_fp16)[name = string("attn_23_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_165_cast_fp16 = add(x = x_141_cast_fp16, y = attn_23_cast_fp16)[name = string("x_165_cast_fp16")];
            fp16 var_4135_promoted_to_fp16 = const()[name = string("op_4135_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_4136_cast_fp16 = pow(x = x_165_cast_fp16, y = var_4135_promoted_to_fp16)[name = string("op_4136_cast_fp16")];
            tensor<int32, [1]> var_105_axes_0 = const()[name = string("var_105_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_105_keep_dims_0 = const()[name = string("var_105_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_105_cast_fp16 = reduce_mean(axes = var_105_axes_0, keep_dims = var_105_keep_dims_0, x = var_4136_cast_fp16)[name = string("var_105_cast_fp16")];
            fp16 var_4143_to_fp16 = const()[name = string("op_4143_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_4144_cast_fp16 = add(x = var_105_cast_fp16, y = var_4143_to_fp16)[name = string("op_4144_cast_fp16")];
            fp32 var_4145_epsilon_0 = const()[name = string("op_4145_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_4145_cast_fp16 = rsqrt(epsilon = var_4145_epsilon_0, x = var_4144_cast_fp16)[name = string("op_4145_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_105_cast_fp16 = mul(x = x_165_cast_fp16, y = var_4145_cast_fp16)[name = string("x_normed_105_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_4155_to_fp16 = const()[name = string("op_4155_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394565824)))];
            tensor<fp16, [1, 640, 1, ?]> input_45_cast_fp16 = mul(x = x_normed_105_cast_fp16, y = var_4155_to_fp16)[name = string("input_45_cast_fp16")];
            string gate_11_pad_type_0 = const()[name = string("gate_11_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_11_strides_0 = const()[name = string("gate_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_11_pad_0 = const()[name = string("gate_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_11_dilations_0 = const()[name = string("gate_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_11_groups_0 = const()[name = string("gate_11_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_5_weight_to_fp16 = const()[name = string("gate_proj_5_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394567168)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_11_cast_fp16 = conv(dilations = gate_11_dilations_0, groups = gate_11_groups_0, pad = gate_11_pad_0, pad_type = gate_11_pad_type_0, strides = gate_11_strides_0, weight = gate_proj_5_weight_to_fp16, x = input_45_cast_fp16)[name = string("gate_11_cast_fp16")];
            string up_11_pad_type_0 = const()[name = string("up_11_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_11_strides_0 = const()[name = string("up_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_11_pad_0 = const()[name = string("up_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_11_dilations_0 = const()[name = string("up_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_11_groups_0 = const()[name = string("up_11_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_5_weight_to_fp16 = const()[name = string("up_proj_5_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397188672)))];
            tensor<fp16, [1, 2048, 1, ?]> up_11_cast_fp16 = conv(dilations = up_11_dilations_0, groups = up_11_groups_0, pad = up_11_pad_0, pad_type = up_11_pad_type_0, strides = up_11_strides_0, weight = up_proj_5_weight_to_fp16, x = input_45_cast_fp16)[name = string("up_11_cast_fp16")];
            string var_4180_mode_0 = const()[name = string("op_4180_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_4180_cast_fp16 = gelu(mode = var_4180_mode_0, x = gate_11_cast_fp16)[name = string("op_4180_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_47_cast_fp16 = mul(x = var_4180_cast_fp16, y = up_11_cast_fp16)[name = string("input_47_cast_fp16")];
            string x_167_pad_type_0 = const()[name = string("x_167_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_167_strides_0 = const()[name = string("x_167_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_167_pad_0 = const()[name = string("x_167_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_167_dilations_0 = const()[name = string("x_167_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_167_groups_0 = const()[name = string("x_167_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_5_weight_to_fp16 = const()[name = string("down_proj_5_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399810176)))];
            tensor<fp16, [1, 640, 1, ?]> x_167_cast_fp16 = conv(dilations = x_167_dilations_0, groups = x_167_groups_0, pad = x_167_pad_0, pad_type = x_167_pad_type_0, strides = x_167_strides_0, weight = down_proj_5_weight_to_fp16, x = input_47_cast_fp16)[name = string("x_167_cast_fp16")];
            fp16 var_4193_promoted_to_fp16 = const()[name = string("op_4193_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_4194_cast_fp16 = pow(x = x_167_cast_fp16, y = var_4193_promoted_to_fp16)[name = string("op_4194_cast_fp16")];
            tensor<int32, [1]> var_107_axes_0 = const()[name = string("var_107_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_107_keep_dims_0 = const()[name = string("var_107_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_107_cast_fp16 = reduce_mean(axes = var_107_axes_0, keep_dims = var_107_keep_dims_0, x = var_4194_cast_fp16)[name = string("var_107_cast_fp16")];
            fp16 var_4201_to_fp16 = const()[name = string("op_4201_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_4202_cast_fp16 = add(x = var_107_cast_fp16, y = var_4201_to_fp16)[name = string("op_4202_cast_fp16")];
            fp32 var_4203_epsilon_0 = const()[name = string("op_4203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_4203_cast_fp16 = rsqrt(epsilon = var_4203_epsilon_0, x = var_4202_cast_fp16)[name = string("op_4203_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_107_cast_fp16 = mul(x = x_167_cast_fp16, y = var_4203_cast_fp16)[name = string("x_normed_107_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_4213_to_fp16 = const()[name = string("op_4213_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402431680)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_11_cast_fp16 = mul(x = x_normed_107_cast_fp16, y = var_4213_to_fp16)[name = string("mlp_11_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_169_cast_fp16 = add(x = x_165_cast_fp16, y = mlp_11_cast_fp16)[name = string("x_169_cast_fp16")];
            fp16 var_4217_promoted_to_fp16 = const()[name = string("op_4217_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_4218_cast_fp16 = pow(x = x_169_cast_fp16, y = var_4217_promoted_to_fp16)[name = string("op_4218_cast_fp16")];
            tensor<int32, [1]> var_109_axes_0 = const()[name = string("var_109_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_109_keep_dims_0 = const()[name = string("var_109_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_109_cast_fp16 = reduce_mean(axes = var_109_axes_0, keep_dims = var_109_keep_dims_0, x = var_4218_cast_fp16)[name = string("var_109_cast_fp16")];
            fp16 var_4225_to_fp16 = const()[name = string("op_4225_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_4226_cast_fp16 = add(x = var_109_cast_fp16, y = var_4225_to_fp16)[name = string("op_4226_cast_fp16")];
            fp32 var_4227_epsilon_0 = const()[name = string("op_4227_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_4227_cast_fp16 = rsqrt(epsilon = var_4227_epsilon_0, x = var_4226_cast_fp16)[name = string("op_4227_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_109_cast_fp16 = mul(x = x_169_cast_fp16, y = var_4227_cast_fp16)[name = string("x_normed_109_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_4237_to_fp16 = const()[name = string("op_4237_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402433024)))];
            tensor<fp16, [1, 640, 1, ?]> input_49_cast_fp16 = mul(x = x_normed_109_cast_fp16, y = var_4237_to_fp16)[name = string("input_49_cast_fp16")];
            string var_4249_pad_type_0 = const()[name = string("op_4249_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4249_strides_0 = const()[name = string("op_4249_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4249_pad_0 = const()[name = string("op_4249_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4249_dilations_0 = const()[name = string("op_4249_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4249_groups_0 = const()[name = string("op_4249_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_6_weight_to_fp16 = const()[name = string("q_proj_6_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402434368)))];
            tensor<fp16, [1, 1024, 1, ?]> var_4249_cast_fp16 = conv(dilations = var_4249_dilations_0, groups = var_4249_groups_0, pad = var_4249_pad_0, pad_type = var_4249_pad_type_0, strides = var_4249_strides_0, weight = q_proj_6_weight_to_fp16, x = input_49_cast_fp16)[name = string("op_4249_cast_fp16")];
            string var_4260_pad_type_0 = const()[name = string("op_4260_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4260_strides_0 = const()[name = string("op_4260_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4260_pad_0 = const()[name = string("op_4260_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4260_dilations_0 = const()[name = string("op_4260_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4260_groups_0 = const()[name = string("op_4260_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_6_weight_to_fp16 = const()[name = string("k_proj_6_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403745152)))];
            tensor<fp16, [1, 256, 1, ?]> var_4260_cast_fp16 = conv(dilations = var_4260_dilations_0, groups = var_4260_groups_0, pad = var_4260_pad_0, pad_type = var_4260_pad_type_0, strides = var_4260_strides_0, weight = k_proj_6_weight_to_fp16, x = input_49_cast_fp16)[name = string("op_4260_cast_fp16")];
            string var_4271_pad_type_0 = const()[name = string("op_4271_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4271_strides_0 = const()[name = string("op_4271_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4271_pad_0 = const()[name = string("op_4271_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4271_dilations_0 = const()[name = string("op_4271_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4271_groups_0 = const()[name = string("op_4271_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_6_weight_to_fp16 = const()[name = string("v_proj_6_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404072896)))];
            tensor<fp16, [1, 256, 1, ?]> var_4271_cast_fp16 = conv(dilations = var_4271_dilations_0, groups = var_4271_groups_0, pad = var_4271_pad_0, pad_type = var_4271_pad_type_0, strides = var_4271_strides_0, weight = v_proj_6_weight_to_fp16, x = input_49_cast_fp16)[name = string("op_4271_cast_fp16")];
            tensor<int32, [4]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_4274_axis_0 = const()[name = string("op_4274_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_4274_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_4274_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_4274_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_4274_cast_fp16_3 = split(axis = var_4274_axis_0, split_sizes = tile_18, x = var_4249_cast_fp16)[name = string("op_4274_cast_fp16")];
            fp16 var_4287_promoted_to_fp16 = const()[name = string("op_4287_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_4288_cast_fp16 = pow(x = var_4274_cast_fp16_0, y = var_4287_promoted_to_fp16)[name = string("op_4288_cast_fp16")];
            tensor<int32, [1]> var_111_axes_0 = const()[name = string("var_111_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_111_keep_dims_0 = const()[name = string("var_111_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_111_cast_fp16 = reduce_mean(axes = var_111_axes_0, keep_dims = var_111_keep_dims_0, x = var_4288_cast_fp16)[name = string("var_111_cast_fp16")];
            fp16 var_4295_to_fp16 = const()[name = string("op_4295_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_4296_cast_fp16 = add(x = var_111_cast_fp16, y = var_4295_to_fp16)[name = string("op_4296_cast_fp16")];
            fp32 var_4297_epsilon_0 = const()[name = string("op_4297_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_4297_cast_fp16 = rsqrt(epsilon = var_4297_epsilon_0, x = var_4296_cast_fp16)[name = string("op_4297_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_111_cast_fp16 = mul(x = var_4274_cast_fp16_0, y = var_4297_cast_fp16)[name = string("x_normed_111_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_4307_to_fp16 = const()[name = string("op_4307_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404400640)))];
            tensor<fp16, [1, 256, 1, ?]> x_173_cast_fp16 = mul(x = x_normed_111_cast_fp16, y = var_4307_to_fp16)[name = string("x_173_cast_fp16")];
            tensor<int32, [4]> x_first_61_begin_0 = const()[name = string("x_first_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_61_end_0 = const()[name = string("x_first_61_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_61_end_mask_0 = const()[name = string("x_first_61_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_61_cast_fp16 = slice_by_index(begin = x_first_61_begin_0, end = x_first_61_end_0, end_mask = x_first_61_end_mask_0, x = x_173_cast_fp16)[name = string("x_first_61_cast_fp16")];
            tensor<int32, [4]> x_second_61_begin_0 = const()[name = string("x_second_61_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_61_end_0 = const()[name = string("x_second_61_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_61_end_mask_0 = const()[name = string("x_second_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_61_cast_fp16 = slice_by_index(begin = x_second_61_begin_0, end = x_second_61_end_0, end_mask = x_second_61_end_mask_0, x = x_173_cast_fp16)[name = string("x_second_61_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_4329_cast_fp16 = mul(x = x_second_61_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_4329_cast_fp16")];
            int32 var_4331 = const()[name = string("op_4331"), val = int32(1)];
            bool x_rot_61_interleave_0 = const()[name = string("x_rot_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_61_cast_fp16 = concat(axis = var_4331, interleave = x_rot_61_interleave_0, values = (var_4329_cast_fp16, x_first_61_cast_fp16))[name = string("x_rot_61_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_4347_cast_fp16 = mul(x = x_173_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_4347_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_4348_cast_fp16 = mul(x = x_rot_61_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_4348_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_49_cast_fp16 = add(x = var_4347_cast_fp16, y = var_4348_cast_fp16)[name = string("q_h_49_cast_fp16")];
            fp16 var_4351_promoted_to_fp16 = const()[name = string("op_4351_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_4352_cast_fp16 = pow(x = var_4274_cast_fp16_1, y = var_4351_promoted_to_fp16)[name = string("op_4352_cast_fp16")];
            tensor<int32, [1]> var_113_axes_0 = const()[name = string("var_113_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_113_keep_dims_0 = const()[name = string("var_113_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_113_cast_fp16 = reduce_mean(axes = var_113_axes_0, keep_dims = var_113_keep_dims_0, x = var_4352_cast_fp16)[name = string("var_113_cast_fp16")];
            fp16 var_4359_to_fp16 = const()[name = string("op_4359_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_4360_cast_fp16 = add(x = var_113_cast_fp16, y = var_4359_to_fp16)[name = string("op_4360_cast_fp16")];
            fp32 var_4361_epsilon_0 = const()[name = string("op_4361_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_4361_cast_fp16 = rsqrt(epsilon = var_4361_epsilon_0, x = var_4360_cast_fp16)[name = string("op_4361_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_113_cast_fp16 = mul(x = var_4274_cast_fp16_1, y = var_4361_cast_fp16)[name = string("x_normed_113_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_177_cast_fp16 = mul(x = x_normed_113_cast_fp16, y = var_4307_to_fp16)[name = string("x_177_cast_fp16")];
            tensor<int32, [4]> x_first_63_begin_0 = const()[name = string("x_first_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_63_end_0 = const()[name = string("x_first_63_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_63_end_mask_0 = const()[name = string("x_first_63_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_63_cast_fp16 = slice_by_index(begin = x_first_63_begin_0, end = x_first_63_end_0, end_mask = x_first_63_end_mask_0, x = x_177_cast_fp16)[name = string("x_first_63_cast_fp16")];
            tensor<int32, [4]> x_second_63_begin_0 = const()[name = string("x_second_63_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_63_end_0 = const()[name = string("x_second_63_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_63_end_mask_0 = const()[name = string("x_second_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_63_cast_fp16 = slice_by_index(begin = x_second_63_begin_0, end = x_second_63_end_0, end_mask = x_second_63_end_mask_0, x = x_177_cast_fp16)[name = string("x_second_63_cast_fp16")];
            fp16 const_53_promoted_to_fp16 = const()[name = string("const_53_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_4393_cast_fp16 = mul(x = x_second_63_cast_fp16, y = const_53_promoted_to_fp16)[name = string("op_4393_cast_fp16")];
            int32 var_4395 = const()[name = string("op_4395"), val = int32(1)];
            bool x_rot_63_interleave_0 = const()[name = string("x_rot_63_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_63_cast_fp16 = concat(axis = var_4395, interleave = x_rot_63_interleave_0, values = (var_4393_cast_fp16, x_first_63_cast_fp16))[name = string("x_rot_63_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_4411_cast_fp16 = mul(x = x_177_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_4411_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_4412_cast_fp16 = mul(x = x_rot_63_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_4412_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_51_cast_fp16 = add(x = var_4411_cast_fp16, y = var_4412_cast_fp16)[name = string("q_h_51_cast_fp16")];
            fp16 var_4415_promoted_to_fp16 = const()[name = string("op_4415_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_4416_cast_fp16 = pow(x = var_4274_cast_fp16_2, y = var_4415_promoted_to_fp16)[name = string("op_4416_cast_fp16")];
            tensor<int32, [1]> var_115_axes_0 = const()[name = string("var_115_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_115_keep_dims_0 = const()[name = string("var_115_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_115_cast_fp16 = reduce_mean(axes = var_115_axes_0, keep_dims = var_115_keep_dims_0, x = var_4416_cast_fp16)[name = string("var_115_cast_fp16")];
            fp16 var_4423_to_fp16 = const()[name = string("op_4423_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_4424_cast_fp16 = add(x = var_115_cast_fp16, y = var_4423_to_fp16)[name = string("op_4424_cast_fp16")];
            fp32 var_4425_epsilon_0 = const()[name = string("op_4425_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_4425_cast_fp16 = rsqrt(epsilon = var_4425_epsilon_0, x = var_4424_cast_fp16)[name = string("op_4425_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_115_cast_fp16 = mul(x = var_4274_cast_fp16_2, y = var_4425_cast_fp16)[name = string("x_normed_115_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_181_cast_fp16 = mul(x = x_normed_115_cast_fp16, y = var_4307_to_fp16)[name = string("x_181_cast_fp16")];
            tensor<int32, [4]> x_first_65_begin_0 = const()[name = string("x_first_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_65_end_0 = const()[name = string("x_first_65_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_65_end_mask_0 = const()[name = string("x_first_65_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_65_cast_fp16 = slice_by_index(begin = x_first_65_begin_0, end = x_first_65_end_0, end_mask = x_first_65_end_mask_0, x = x_181_cast_fp16)[name = string("x_first_65_cast_fp16")];
            tensor<int32, [4]> x_second_65_begin_0 = const()[name = string("x_second_65_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_65_end_0 = const()[name = string("x_second_65_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_65_end_mask_0 = const()[name = string("x_second_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_65_cast_fp16 = slice_by_index(begin = x_second_65_begin_0, end = x_second_65_end_0, end_mask = x_second_65_end_mask_0, x = x_181_cast_fp16)[name = string("x_second_65_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_4457_cast_fp16 = mul(x = x_second_65_cast_fp16, y = const_54_promoted_to_fp16)[name = string("op_4457_cast_fp16")];
            int32 var_4459 = const()[name = string("op_4459"), val = int32(1)];
            bool x_rot_65_interleave_0 = const()[name = string("x_rot_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_65_cast_fp16 = concat(axis = var_4459, interleave = x_rot_65_interleave_0, values = (var_4457_cast_fp16, x_first_65_cast_fp16))[name = string("x_rot_65_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_4475_cast_fp16 = mul(x = x_181_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_4475_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_4476_cast_fp16 = mul(x = x_rot_65_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_4476_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_53_cast_fp16 = add(x = var_4475_cast_fp16, y = var_4476_cast_fp16)[name = string("q_h_53_cast_fp16")];
            fp16 var_4479_promoted_to_fp16 = const()[name = string("op_4479_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_4480_cast_fp16 = pow(x = var_4274_cast_fp16_3, y = var_4479_promoted_to_fp16)[name = string("op_4480_cast_fp16")];
            tensor<int32, [1]> var_117_axes_0 = const()[name = string("var_117_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_117_keep_dims_0 = const()[name = string("var_117_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_117_cast_fp16 = reduce_mean(axes = var_117_axes_0, keep_dims = var_117_keep_dims_0, x = var_4480_cast_fp16)[name = string("var_117_cast_fp16")];
            fp16 var_4487_to_fp16 = const()[name = string("op_4487_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_4488_cast_fp16 = add(x = var_117_cast_fp16, y = var_4487_to_fp16)[name = string("op_4488_cast_fp16")];
            fp32 var_4489_epsilon_0 = const()[name = string("op_4489_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_4489_cast_fp16 = rsqrt(epsilon = var_4489_epsilon_0, x = var_4488_cast_fp16)[name = string("op_4489_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_117_cast_fp16 = mul(x = var_4274_cast_fp16_3, y = var_4489_cast_fp16)[name = string("x_normed_117_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_185_cast_fp16 = mul(x = x_normed_117_cast_fp16, y = var_4307_to_fp16)[name = string("x_185_cast_fp16")];
            tensor<int32, [4]> x_first_67_begin_0 = const()[name = string("x_first_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_67_end_0 = const()[name = string("x_first_67_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_67_end_mask_0 = const()[name = string("x_first_67_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_67_cast_fp16 = slice_by_index(begin = x_first_67_begin_0, end = x_first_67_end_0, end_mask = x_first_67_end_mask_0, x = x_185_cast_fp16)[name = string("x_first_67_cast_fp16")];
            tensor<int32, [4]> x_second_67_begin_0 = const()[name = string("x_second_67_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_67_end_0 = const()[name = string("x_second_67_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_67_end_mask_0 = const()[name = string("x_second_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_67_cast_fp16 = slice_by_index(begin = x_second_67_begin_0, end = x_second_67_end_0, end_mask = x_second_67_end_mask_0, x = x_185_cast_fp16)[name = string("x_second_67_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_4521_cast_fp16 = mul(x = x_second_67_cast_fp16, y = const_55_promoted_to_fp16)[name = string("op_4521_cast_fp16")];
            int32 var_4523 = const()[name = string("op_4523"), val = int32(1)];
            bool x_rot_67_interleave_0 = const()[name = string("x_rot_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_67_cast_fp16 = concat(axis = var_4523, interleave = x_rot_67_interleave_0, values = (var_4521_cast_fp16, x_first_67_cast_fp16))[name = string("x_rot_67_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_4539_cast_fp16 = mul(x = x_185_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_4539_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_4540_cast_fp16 = mul(x = x_rot_67_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_4540_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_55_cast_fp16 = add(x = var_4539_cast_fp16, y = var_4540_cast_fp16)[name = string("q_h_55_cast_fp16")];
            fp16 var_4543_promoted_to_fp16 = const()[name = string("op_4543_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_4544_cast_fp16 = pow(x = var_4260_cast_fp16, y = var_4543_promoted_to_fp16)[name = string("op_4544_cast_fp16")];
            tensor<int32, [1]> var_119_axes_0 = const()[name = string("var_119_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_119_keep_dims_0 = const()[name = string("var_119_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_119_cast_fp16 = reduce_mean(axes = var_119_axes_0, keep_dims = var_119_keep_dims_0, x = var_4544_cast_fp16)[name = string("var_119_cast_fp16")];
            fp16 var_4551_to_fp16 = const()[name = string("op_4551_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_4552_cast_fp16 = add(x = var_119_cast_fp16, y = var_4551_to_fp16)[name = string("op_4552_cast_fp16")];
            fp32 var_4553_epsilon_0 = const()[name = string("op_4553_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_4553_cast_fp16 = rsqrt(epsilon = var_4553_epsilon_0, x = var_4552_cast_fp16)[name = string("op_4553_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_119_cast_fp16 = mul(x = var_4260_cast_fp16, y = var_4553_cast_fp16)[name = string("x_normed_119_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_4563_to_fp16 = const()[name = string("op_4563_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404401216)))];
            tensor<fp16, [1, 256, 1, ?]> x_189_cast_fp16 = mul(x = x_normed_119_cast_fp16, y = var_4563_to_fp16)[name = string("x_189_cast_fp16")];
            tensor<int32, [4]> x_first_69_begin_0 = const()[name = string("x_first_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_69_end_0 = const()[name = string("x_first_69_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_69_end_mask_0 = const()[name = string("x_first_69_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_69_cast_fp16 = slice_by_index(begin = x_first_69_begin_0, end = x_first_69_end_0, end_mask = x_first_69_end_mask_0, x = x_189_cast_fp16)[name = string("x_first_69_cast_fp16")];
            tensor<int32, [4]> x_second_69_begin_0 = const()[name = string("x_second_69_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_69_end_0 = const()[name = string("x_second_69_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_69_end_mask_0 = const()[name = string("x_second_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_69_cast_fp16 = slice_by_index(begin = x_second_69_begin_0, end = x_second_69_end_0, end_mask = x_second_69_end_mask_0, x = x_189_cast_fp16)[name = string("x_second_69_cast_fp16")];
            fp16 const_56_promoted_to_fp16 = const()[name = string("const_56_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_4585_cast_fp16 = mul(x = x_second_69_cast_fp16, y = const_56_promoted_to_fp16)[name = string("op_4585_cast_fp16")];
            int32 var_4587 = const()[name = string("op_4587"), val = int32(1)];
            bool x_rot_69_interleave_0 = const()[name = string("x_rot_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_69_cast_fp16 = concat(axis = var_4587, interleave = x_rot_69_interleave_0, values = (var_4585_cast_fp16, x_first_69_cast_fp16))[name = string("x_rot_69_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_4603_cast_fp16 = mul(x = x_189_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_4603_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_4604_cast_fp16 = mul(x = x_rot_69_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_4604_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_13_cast_fp16 = add(x = var_4603_cast_fp16, y = var_4604_cast_fp16)[name = string("k_new_13_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_12 = read_state(input = k_cache_6)[name = string("read_state_12")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_39_cast_fp16 = mul(x = read_state_12, y = one_minus_1_cast_fp16)[name = string("k_cache_39_cast_fp16")];
            write_state(data = k_cache_39_cast_fp16, input = k_cache_6)[name = string("coreml_update_state_96_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_96 = read_state(input = k_cache_6)[name = string("coreml_update_state_96")];
            tensor<fp16, [1, 256, 1, 128]> var_4611_cast_fp16 = mul(x = k_new_13_cast_fp16, y = write_mask)[name = string("op_4611_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_41_cast_fp16 = add(x = coreml_update_state_96, y = var_4611_cast_fp16)[name = string("k_cache_41_cast_fp16")];
            write_state(data = k_cache_41_cast_fp16, input = k_cache_6)[name = string("coreml_update_state_97_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_97 = read_state(input = k_cache_6)[name = string("coreml_update_state_97")];
            tensor<fp16, [1, 256, 1, 128]> read_state_13 = read_state(input = v_cache_6)[name = string("read_state_13")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_39_cast_fp16 = mul(x = read_state_13, y = one_minus_1_cast_fp16)[name = string("v_cache_39_cast_fp16")];
            write_state(data = v_cache_39_cast_fp16, input = v_cache_6)[name = string("coreml_update_state_98_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_98 = read_state(input = v_cache_6)[name = string("coreml_update_state_98")];
            tensor<fp16, [1, 256, 1, 128]> var_4615_cast_fp16 = mul(x = var_4271_cast_fp16, y = write_mask)[name = string("op_4615_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_41_cast_fp16 = add(x = coreml_update_state_98, y = var_4615_cast_fp16)[name = string("v_cache_41_cast_fp16")];
            write_state(data = v_cache_41_cast_fp16, input = v_cache_6)[name = string("coreml_update_state_99_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_99 = read_state(input = v_cache_6)[name = string("coreml_update_state_99")];
            bool var_4630_transpose_x_0 = const()[name = string("op_4630_transpose_x_0"), val = bool(false)];
            bool var_4630_transpose_y_0 = const()[name = string("op_4630_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_354_perm_0 = const()[name = string("transpose_354_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_355_perm_0 = const()[name = string("transpose_355_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_355 = transpose(perm = transpose_355_perm_0, x = coreml_update_state_97)[name = string("transpose_532")];
            tensor<fp16, [1, 1, ?, 256]> transpose_354 = transpose(perm = transpose_354_perm_0, x = q_h_49_cast_fp16)[name = string("transpose_533")];
            tensor<fp16, [1, 1, ?, 128]> var_4630_cast_fp16 = matmul(transpose_x = var_4630_transpose_x_0, transpose_y = var_4630_transpose_y_0, x = transpose_354, y = transpose_355)[name = string("op_4630_cast_fp16")];
            fp16 var_4631_to_fp16 = const()[name = string("op_4631_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_4632_cast_fp16 = mul(x = var_4630_cast_fp16, y = var_4631_to_fp16)[name = string("op_4632_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_49_cast_fp16 = add(x = var_4632_cast_fp16, y = attention_mask)[name = string("scores_49_cast_fp16")];
            int32 var_4635 = const()[name = string("op_4635"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_49_cast_fp16 = softmax(axis = var_4635, x = scores_49_cast_fp16)[name = string("probs_49_cast_fp16")];
            tensor<int32, [4]> var_4642 = const()[name = string("op_4642"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_4644_transpose_x_0 = const()[name = string("op_4644_transpose_x_0"), val = bool(false)];
            bool var_4644_transpose_y_0 = const()[name = string("op_4644_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_49_cast_fp16 = transpose(perm = var_4642, x = coreml_update_state_99)[name = string("transpose_531")];
            tensor<fp16, [1, 1, ?, 256]> var_4644_cast_fp16 = matmul(transpose_x = var_4644_transpose_x_0, transpose_y = var_4644_transpose_y_0, x = probs_49_cast_fp16, y = v_t_49_cast_fp16)[name = string("op_4644_cast_fp16")];
            bool var_4663_transpose_x_0 = const()[name = string("op_4663_transpose_x_0"), val = bool(false)];
            bool var_4663_transpose_y_0 = const()[name = string("op_4663_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_359_perm_0 = const()[name = string("transpose_359_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_359 = transpose(perm = transpose_359_perm_0, x = q_h_51_cast_fp16)[name = string("transpose_530")];
            tensor<fp16, [1, 1, ?, 128]> var_4663_cast_fp16 = matmul(transpose_x = var_4663_transpose_x_0, transpose_y = var_4663_transpose_y_0, x = transpose_359, y = transpose_355)[name = string("op_4663_cast_fp16")];
            fp16 var_4664_to_fp16 = const()[name = string("op_4664_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_4665_cast_fp16 = mul(x = var_4663_cast_fp16, y = var_4664_to_fp16)[name = string("op_4665_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_51_cast_fp16 = add(x = var_4665_cast_fp16, y = attention_mask)[name = string("scores_51_cast_fp16")];
            int32 var_4668 = const()[name = string("op_4668"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_51_cast_fp16 = softmax(axis = var_4668, x = scores_51_cast_fp16)[name = string("probs_51_cast_fp16")];
            bool var_4677_transpose_x_0 = const()[name = string("op_4677_transpose_x_0"), val = bool(false)];
            bool var_4677_transpose_y_0 = const()[name = string("op_4677_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_4677_cast_fp16 = matmul(transpose_x = var_4677_transpose_x_0, transpose_y = var_4677_transpose_y_0, x = probs_51_cast_fp16, y = v_t_49_cast_fp16)[name = string("op_4677_cast_fp16")];
            bool var_4696_transpose_x_0 = const()[name = string("op_4696_transpose_x_0"), val = bool(false)];
            bool var_4696_transpose_y_0 = const()[name = string("op_4696_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_360_perm_0 = const()[name = string("transpose_360_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_360 = transpose(perm = transpose_360_perm_0, x = q_h_53_cast_fp16)[name = string("transpose_529")];
            tensor<fp16, [1, 1, ?, 128]> var_4696_cast_fp16 = matmul(transpose_x = var_4696_transpose_x_0, transpose_y = var_4696_transpose_y_0, x = transpose_360, y = transpose_355)[name = string("op_4696_cast_fp16")];
            fp16 var_4697_to_fp16 = const()[name = string("op_4697_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_4698_cast_fp16 = mul(x = var_4696_cast_fp16, y = var_4697_to_fp16)[name = string("op_4698_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_53_cast_fp16 = add(x = var_4698_cast_fp16, y = attention_mask)[name = string("scores_53_cast_fp16")];
            int32 var_4701 = const()[name = string("op_4701"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_53_cast_fp16 = softmax(axis = var_4701, x = scores_53_cast_fp16)[name = string("probs_53_cast_fp16")];
            bool var_4710_transpose_x_0 = const()[name = string("op_4710_transpose_x_0"), val = bool(false)];
            bool var_4710_transpose_y_0 = const()[name = string("op_4710_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_4710_cast_fp16 = matmul(transpose_x = var_4710_transpose_x_0, transpose_y = var_4710_transpose_y_0, x = probs_53_cast_fp16, y = v_t_49_cast_fp16)[name = string("op_4710_cast_fp16")];
            bool var_4729_transpose_x_0 = const()[name = string("op_4729_transpose_x_0"), val = bool(false)];
            bool var_4729_transpose_y_0 = const()[name = string("op_4729_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_361_perm_0 = const()[name = string("transpose_361_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_361 = transpose(perm = transpose_361_perm_0, x = q_h_55_cast_fp16)[name = string("transpose_528")];
            tensor<fp16, [1, 1, ?, 128]> var_4729_cast_fp16 = matmul(transpose_x = var_4729_transpose_x_0, transpose_y = var_4729_transpose_y_0, x = transpose_361, y = transpose_355)[name = string("op_4729_cast_fp16")];
            fp16 var_4730_to_fp16 = const()[name = string("op_4730_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_4731_cast_fp16 = mul(x = var_4729_cast_fp16, y = var_4730_to_fp16)[name = string("op_4731_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_55_cast_fp16 = add(x = var_4731_cast_fp16, y = attention_mask)[name = string("scores_55_cast_fp16")];
            int32 var_4734 = const()[name = string("op_4734"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_55_cast_fp16 = softmax(axis = var_4734, x = scores_55_cast_fp16)[name = string("probs_55_cast_fp16")];
            bool var_4743_transpose_x_0 = const()[name = string("op_4743_transpose_x_0"), val = bool(false)];
            bool var_4743_transpose_y_0 = const()[name = string("op_4743_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_4743_cast_fp16 = matmul(transpose_x = var_4743_transpose_x_0, transpose_y = var_4743_transpose_y_0, x = probs_55_cast_fp16, y = v_t_49_cast_fp16)[name = string("op_4743_cast_fp16")];
            bool input_51_interleave_0 = const()[name = string("input_51_interleave_0"), val = bool(false)];
            int32 const_151 = const()[name = string("const_151"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_51_cast_fp16 = concat(axis = const_151, interleave = input_51_interleave_0, values = (var_4644_cast_fp16, var_4677_cast_fp16, var_4710_cast_fp16, var_4743_cast_fp16))[name = string("input_51_cast_fp16")];
            string x_191_pad_type_0 = const()[name = string("x_191_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_191_strides_0 = const()[name = string("x_191_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_191_pad_0 = const()[name = string("x_191_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_191_dilations_0 = const()[name = string("x_191_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_191_groups_0 = const()[name = string("x_191_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_62_perm_0 = const()[name = string("transpose_62_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_6_weight_to_fp16 = const()[name = string("o_proj_6_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404401792)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_62_cast_fp16 = transpose(perm = transpose_62_perm_0, x = input_51_cast_fp16)[name = string("transpose_527")];
            tensor<fp16, [1, 640, 1, ?]> x_191_cast_fp16 = conv(dilations = x_191_dilations_0, groups = x_191_groups_0, pad = x_191_pad_0, pad_type = x_191_pad_type_0, strides = x_191_strides_0, weight = o_proj_6_weight_to_fp16, x = transpose_62_cast_fp16)[name = string("x_191_cast_fp16")];
            fp16 var_4764_promoted_to_fp16 = const()[name = string("op_4764_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_4765_cast_fp16 = pow(x = x_191_cast_fp16, y = var_4764_promoted_to_fp16)[name = string("op_4765_cast_fp16")];
            tensor<int32, [1]> var_121_axes_0 = const()[name = string("var_121_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_121_keep_dims_0 = const()[name = string("var_121_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_121_cast_fp16 = reduce_mean(axes = var_121_axes_0, keep_dims = var_121_keep_dims_0, x = var_4765_cast_fp16)[name = string("var_121_cast_fp16")];
            fp16 var_4772_to_fp16 = const()[name = string("op_4772_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_4773_cast_fp16 = add(x = var_121_cast_fp16, y = var_4772_to_fp16)[name = string("op_4773_cast_fp16")];
            fp32 var_4774_epsilon_0 = const()[name = string("op_4774_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_4774_cast_fp16 = rsqrt(epsilon = var_4774_epsilon_0, x = var_4773_cast_fp16)[name = string("op_4774_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_121_cast_fp16 = mul(x = x_191_cast_fp16, y = var_4774_cast_fp16)[name = string("x_normed_121_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_4784_to_fp16 = const()[name = string("op_4784_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405712576)))];
            tensor<fp16, [1, 640, 1, ?]> attn_27_cast_fp16 = mul(x = x_normed_121_cast_fp16, y = var_4784_to_fp16)[name = string("attn_27_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_193_cast_fp16 = add(x = x_169_cast_fp16, y = attn_27_cast_fp16)[name = string("x_193_cast_fp16")];
            fp16 var_4788_promoted_to_fp16 = const()[name = string("op_4788_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_4789_cast_fp16 = pow(x = x_193_cast_fp16, y = var_4788_promoted_to_fp16)[name = string("op_4789_cast_fp16")];
            tensor<int32, [1]> var_123_axes_0 = const()[name = string("var_123_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_123_keep_dims_0 = const()[name = string("var_123_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_123_cast_fp16 = reduce_mean(axes = var_123_axes_0, keep_dims = var_123_keep_dims_0, x = var_4789_cast_fp16)[name = string("var_123_cast_fp16")];
            fp16 var_4796_to_fp16 = const()[name = string("op_4796_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_4797_cast_fp16 = add(x = var_123_cast_fp16, y = var_4796_to_fp16)[name = string("op_4797_cast_fp16")];
            fp32 var_4798_epsilon_0 = const()[name = string("op_4798_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_4798_cast_fp16 = rsqrt(epsilon = var_4798_epsilon_0, x = var_4797_cast_fp16)[name = string("op_4798_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_123_cast_fp16 = mul(x = x_193_cast_fp16, y = var_4798_cast_fp16)[name = string("x_normed_123_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_4808_to_fp16 = const()[name = string("op_4808_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405713920)))];
            tensor<fp16, [1, 640, 1, ?]> input_53_cast_fp16 = mul(x = x_normed_123_cast_fp16, y = var_4808_to_fp16)[name = string("input_53_cast_fp16")];
            string gate_13_pad_type_0 = const()[name = string("gate_13_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_13_strides_0 = const()[name = string("gate_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_13_pad_0 = const()[name = string("gate_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_13_dilations_0 = const()[name = string("gate_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_13_groups_0 = const()[name = string("gate_13_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_6_weight_to_fp16 = const()[name = string("gate_proj_6_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405715264)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_13_cast_fp16 = conv(dilations = gate_13_dilations_0, groups = gate_13_groups_0, pad = gate_13_pad_0, pad_type = gate_13_pad_type_0, strides = gate_13_strides_0, weight = gate_proj_6_weight_to_fp16, x = input_53_cast_fp16)[name = string("gate_13_cast_fp16")];
            string up_13_pad_type_0 = const()[name = string("up_13_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_13_strides_0 = const()[name = string("up_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_13_pad_0 = const()[name = string("up_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_13_dilations_0 = const()[name = string("up_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_13_groups_0 = const()[name = string("up_13_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_6_weight_to_fp16 = const()[name = string("up_proj_6_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408336768)))];
            tensor<fp16, [1, 2048, 1, ?]> up_13_cast_fp16 = conv(dilations = up_13_dilations_0, groups = up_13_groups_0, pad = up_13_pad_0, pad_type = up_13_pad_type_0, strides = up_13_strides_0, weight = up_proj_6_weight_to_fp16, x = input_53_cast_fp16)[name = string("up_13_cast_fp16")];
            string var_4833_mode_0 = const()[name = string("op_4833_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_4833_cast_fp16 = gelu(mode = var_4833_mode_0, x = gate_13_cast_fp16)[name = string("op_4833_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_55_cast_fp16 = mul(x = var_4833_cast_fp16, y = up_13_cast_fp16)[name = string("input_55_cast_fp16")];
            string x_195_pad_type_0 = const()[name = string("x_195_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_195_strides_0 = const()[name = string("x_195_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_195_pad_0 = const()[name = string("x_195_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_195_dilations_0 = const()[name = string("x_195_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_195_groups_0 = const()[name = string("x_195_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_6_weight_to_fp16 = const()[name = string("down_proj_6_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410958272)))];
            tensor<fp16, [1, 640, 1, ?]> x_195_cast_fp16 = conv(dilations = x_195_dilations_0, groups = x_195_groups_0, pad = x_195_pad_0, pad_type = x_195_pad_type_0, strides = x_195_strides_0, weight = down_proj_6_weight_to_fp16, x = input_55_cast_fp16)[name = string("x_195_cast_fp16")];
            fp16 var_4846_promoted_to_fp16 = const()[name = string("op_4846_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_4847_cast_fp16 = pow(x = x_195_cast_fp16, y = var_4846_promoted_to_fp16)[name = string("op_4847_cast_fp16")];
            tensor<int32, [1]> var_125_axes_0 = const()[name = string("var_125_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_125_keep_dims_0 = const()[name = string("var_125_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_125_cast_fp16 = reduce_mean(axes = var_125_axes_0, keep_dims = var_125_keep_dims_0, x = var_4847_cast_fp16)[name = string("var_125_cast_fp16")];
            fp16 var_4854_to_fp16 = const()[name = string("op_4854_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_4855_cast_fp16 = add(x = var_125_cast_fp16, y = var_4854_to_fp16)[name = string("op_4855_cast_fp16")];
            fp32 var_4856_epsilon_0 = const()[name = string("op_4856_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_4856_cast_fp16 = rsqrt(epsilon = var_4856_epsilon_0, x = var_4855_cast_fp16)[name = string("op_4856_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_125_cast_fp16 = mul(x = x_195_cast_fp16, y = var_4856_cast_fp16)[name = string("x_normed_125_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_4866_to_fp16 = const()[name = string("op_4866_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413579776)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_13_cast_fp16 = mul(x = x_normed_125_cast_fp16, y = var_4866_to_fp16)[name = string("mlp_13_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_197_cast_fp16 = add(x = x_193_cast_fp16, y = mlp_13_cast_fp16)[name = string("x_197_cast_fp16")];
            fp16 var_4870_promoted_to_fp16 = const()[name = string("op_4870_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_4871_cast_fp16 = pow(x = x_197_cast_fp16, y = var_4870_promoted_to_fp16)[name = string("op_4871_cast_fp16")];
            tensor<int32, [1]> var_127_axes_0 = const()[name = string("var_127_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_127_keep_dims_0 = const()[name = string("var_127_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_127_cast_fp16 = reduce_mean(axes = var_127_axes_0, keep_dims = var_127_keep_dims_0, x = var_4871_cast_fp16)[name = string("var_127_cast_fp16")];
            fp16 var_4878_to_fp16 = const()[name = string("op_4878_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_4879_cast_fp16 = add(x = var_127_cast_fp16, y = var_4878_to_fp16)[name = string("op_4879_cast_fp16")];
            fp32 var_4880_epsilon_0 = const()[name = string("op_4880_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_4880_cast_fp16 = rsqrt(epsilon = var_4880_epsilon_0, x = var_4879_cast_fp16)[name = string("op_4880_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_127_cast_fp16 = mul(x = x_197_cast_fp16, y = var_4880_cast_fp16)[name = string("x_normed_127_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_4890_to_fp16 = const()[name = string("op_4890_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413581120)))];
            tensor<fp16, [1, 640, 1, ?]> input_57_cast_fp16 = mul(x = x_normed_127_cast_fp16, y = var_4890_to_fp16)[name = string("input_57_cast_fp16")];
            string var_4902_pad_type_0 = const()[name = string("op_4902_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4902_strides_0 = const()[name = string("op_4902_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4902_pad_0 = const()[name = string("op_4902_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4902_dilations_0 = const()[name = string("op_4902_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4902_groups_0 = const()[name = string("op_4902_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_7_weight_to_fp16 = const()[name = string("q_proj_7_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413582464)))];
            tensor<fp16, [1, 1024, 1, ?]> var_4902_cast_fp16 = conv(dilations = var_4902_dilations_0, groups = var_4902_groups_0, pad = var_4902_pad_0, pad_type = var_4902_pad_type_0, strides = var_4902_strides_0, weight = q_proj_7_weight_to_fp16, x = input_57_cast_fp16)[name = string("op_4902_cast_fp16")];
            string var_4913_pad_type_0 = const()[name = string("op_4913_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4913_strides_0 = const()[name = string("op_4913_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4913_pad_0 = const()[name = string("op_4913_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4913_dilations_0 = const()[name = string("op_4913_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4913_groups_0 = const()[name = string("op_4913_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_7_weight_to_fp16 = const()[name = string("k_proj_7_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(414893248)))];
            tensor<fp16, [1, 256, 1, ?]> var_4913_cast_fp16 = conv(dilations = var_4913_dilations_0, groups = var_4913_groups_0, pad = var_4913_pad_0, pad_type = var_4913_pad_type_0, strides = var_4913_strides_0, weight = k_proj_7_weight_to_fp16, x = input_57_cast_fp16)[name = string("op_4913_cast_fp16")];
            string var_4924_pad_type_0 = const()[name = string("op_4924_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4924_strides_0 = const()[name = string("op_4924_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4924_pad_0 = const()[name = string("op_4924_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4924_dilations_0 = const()[name = string("op_4924_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4924_groups_0 = const()[name = string("op_4924_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_7_weight_to_fp16 = const()[name = string("v_proj_7_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415220992)))];
            tensor<fp16, [1, 256, 1, ?]> var_4924_cast_fp16 = conv(dilations = var_4924_dilations_0, groups = var_4924_groups_0, pad = var_4924_pad_0, pad_type = var_4924_pad_type_0, strides = var_4924_strides_0, weight = v_proj_7_weight_to_fp16, x = input_57_cast_fp16)[name = string("op_4924_cast_fp16")];
            tensor<int32, [4]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_4927_axis_0 = const()[name = string("op_4927_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_4927_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_4927_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_4927_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_4927_cast_fp16_3 = split(axis = var_4927_axis_0, split_sizes = tile_21, x = var_4902_cast_fp16)[name = string("op_4927_cast_fp16")];
            fp16 var_4940_promoted_to_fp16 = const()[name = string("op_4940_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_4941_cast_fp16 = pow(x = var_4927_cast_fp16_0, y = var_4940_promoted_to_fp16)[name = string("op_4941_cast_fp16")];
            tensor<int32, [1]> var_129_axes_0 = const()[name = string("var_129_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_129_keep_dims_0 = const()[name = string("var_129_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_129_cast_fp16 = reduce_mean(axes = var_129_axes_0, keep_dims = var_129_keep_dims_0, x = var_4941_cast_fp16)[name = string("var_129_cast_fp16")];
            fp16 var_4948_to_fp16 = const()[name = string("op_4948_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_4949_cast_fp16 = add(x = var_129_cast_fp16, y = var_4948_to_fp16)[name = string("op_4949_cast_fp16")];
            fp32 var_4950_epsilon_0 = const()[name = string("op_4950_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_4950_cast_fp16 = rsqrt(epsilon = var_4950_epsilon_0, x = var_4949_cast_fp16)[name = string("op_4950_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_129_cast_fp16 = mul(x = var_4927_cast_fp16_0, y = var_4950_cast_fp16)[name = string("x_normed_129_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_4960_to_fp16 = const()[name = string("op_4960_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415548736)))];
            tensor<fp16, [1, 256, 1, ?]> x_201_cast_fp16 = mul(x = x_normed_129_cast_fp16, y = var_4960_to_fp16)[name = string("x_201_cast_fp16")];
            tensor<int32, [4]> x_first_71_begin_0 = const()[name = string("x_first_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_71_end_0 = const()[name = string("x_first_71_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_71_end_mask_0 = const()[name = string("x_first_71_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_71_cast_fp16 = slice_by_index(begin = x_first_71_begin_0, end = x_first_71_end_0, end_mask = x_first_71_end_mask_0, x = x_201_cast_fp16)[name = string("x_first_71_cast_fp16")];
            tensor<int32, [4]> x_second_71_begin_0 = const()[name = string("x_second_71_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_71_end_0 = const()[name = string("x_second_71_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_71_end_mask_0 = const()[name = string("x_second_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_71_cast_fp16 = slice_by_index(begin = x_second_71_begin_0, end = x_second_71_end_0, end_mask = x_second_71_end_mask_0, x = x_201_cast_fp16)[name = string("x_second_71_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_4982_cast_fp16 = mul(x = x_second_71_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_4982_cast_fp16")];
            int32 var_4984 = const()[name = string("op_4984"), val = int32(1)];
            bool x_rot_71_interleave_0 = const()[name = string("x_rot_71_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_71_cast_fp16 = concat(axis = var_4984, interleave = x_rot_71_interleave_0, values = (var_4982_cast_fp16, x_first_71_cast_fp16))[name = string("x_rot_71_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5000_cast_fp16 = mul(x = x_201_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_5000_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5001_cast_fp16 = mul(x = x_rot_71_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_5001_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_57_cast_fp16 = add(x = var_5000_cast_fp16, y = var_5001_cast_fp16)[name = string("q_h_57_cast_fp16")];
            fp16 var_5004_promoted_to_fp16 = const()[name = string("op_5004_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_5005_cast_fp16 = pow(x = var_4927_cast_fp16_1, y = var_5004_promoted_to_fp16)[name = string("op_5005_cast_fp16")];
            tensor<int32, [1]> var_131_axes_0 = const()[name = string("var_131_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_131_keep_dims_0 = const()[name = string("var_131_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_131_cast_fp16 = reduce_mean(axes = var_131_axes_0, keep_dims = var_131_keep_dims_0, x = var_5005_cast_fp16)[name = string("var_131_cast_fp16")];
            fp16 var_5012_to_fp16 = const()[name = string("op_5012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_5013_cast_fp16 = add(x = var_131_cast_fp16, y = var_5012_to_fp16)[name = string("op_5013_cast_fp16")];
            fp32 var_5014_epsilon_0 = const()[name = string("op_5014_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_5014_cast_fp16 = rsqrt(epsilon = var_5014_epsilon_0, x = var_5013_cast_fp16)[name = string("op_5014_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_131_cast_fp16 = mul(x = var_4927_cast_fp16_1, y = var_5014_cast_fp16)[name = string("x_normed_131_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_205_cast_fp16 = mul(x = x_normed_131_cast_fp16, y = var_4960_to_fp16)[name = string("x_205_cast_fp16")];
            tensor<int32, [4]> x_first_73_begin_0 = const()[name = string("x_first_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_73_end_0 = const()[name = string("x_first_73_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_73_end_mask_0 = const()[name = string("x_first_73_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_73_cast_fp16 = slice_by_index(begin = x_first_73_begin_0, end = x_first_73_end_0, end_mask = x_first_73_end_mask_0, x = x_205_cast_fp16)[name = string("x_first_73_cast_fp16")];
            tensor<int32, [4]> x_second_73_begin_0 = const()[name = string("x_second_73_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_73_end_0 = const()[name = string("x_second_73_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_73_end_mask_0 = const()[name = string("x_second_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_73_cast_fp16 = slice_by_index(begin = x_second_73_begin_0, end = x_second_73_end_0, end_mask = x_second_73_end_mask_0, x = x_205_cast_fp16)[name = string("x_second_73_cast_fp16")];
            fp16 const_61_promoted_to_fp16 = const()[name = string("const_61_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_5046_cast_fp16 = mul(x = x_second_73_cast_fp16, y = const_61_promoted_to_fp16)[name = string("op_5046_cast_fp16")];
            int32 var_5048 = const()[name = string("op_5048"), val = int32(1)];
            bool x_rot_73_interleave_0 = const()[name = string("x_rot_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_73_cast_fp16 = concat(axis = var_5048, interleave = x_rot_73_interleave_0, values = (var_5046_cast_fp16, x_first_73_cast_fp16))[name = string("x_rot_73_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5064_cast_fp16 = mul(x = x_205_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_5064_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5065_cast_fp16 = mul(x = x_rot_73_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_5065_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_59_cast_fp16 = add(x = var_5064_cast_fp16, y = var_5065_cast_fp16)[name = string("q_h_59_cast_fp16")];
            fp16 var_5068_promoted_to_fp16 = const()[name = string("op_5068_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_5069_cast_fp16 = pow(x = var_4927_cast_fp16_2, y = var_5068_promoted_to_fp16)[name = string("op_5069_cast_fp16")];
            tensor<int32, [1]> var_133_axes_0 = const()[name = string("var_133_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_133_keep_dims_0 = const()[name = string("var_133_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_133_cast_fp16 = reduce_mean(axes = var_133_axes_0, keep_dims = var_133_keep_dims_0, x = var_5069_cast_fp16)[name = string("var_133_cast_fp16")];
            fp16 var_5076_to_fp16 = const()[name = string("op_5076_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_5077_cast_fp16 = add(x = var_133_cast_fp16, y = var_5076_to_fp16)[name = string("op_5077_cast_fp16")];
            fp32 var_5078_epsilon_0 = const()[name = string("op_5078_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_5078_cast_fp16 = rsqrt(epsilon = var_5078_epsilon_0, x = var_5077_cast_fp16)[name = string("op_5078_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_133_cast_fp16 = mul(x = var_4927_cast_fp16_2, y = var_5078_cast_fp16)[name = string("x_normed_133_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_209_cast_fp16 = mul(x = x_normed_133_cast_fp16, y = var_4960_to_fp16)[name = string("x_209_cast_fp16")];
            tensor<int32, [4]> x_first_75_begin_0 = const()[name = string("x_first_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_75_end_0 = const()[name = string("x_first_75_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_75_end_mask_0 = const()[name = string("x_first_75_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_75_cast_fp16 = slice_by_index(begin = x_first_75_begin_0, end = x_first_75_end_0, end_mask = x_first_75_end_mask_0, x = x_209_cast_fp16)[name = string("x_first_75_cast_fp16")];
            tensor<int32, [4]> x_second_75_begin_0 = const()[name = string("x_second_75_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_75_end_0 = const()[name = string("x_second_75_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_75_end_mask_0 = const()[name = string("x_second_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_75_cast_fp16 = slice_by_index(begin = x_second_75_begin_0, end = x_second_75_end_0, end_mask = x_second_75_end_mask_0, x = x_209_cast_fp16)[name = string("x_second_75_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_5110_cast_fp16 = mul(x = x_second_75_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_5110_cast_fp16")];
            int32 var_5112 = const()[name = string("op_5112"), val = int32(1)];
            bool x_rot_75_interleave_0 = const()[name = string("x_rot_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_75_cast_fp16 = concat(axis = var_5112, interleave = x_rot_75_interleave_0, values = (var_5110_cast_fp16, x_first_75_cast_fp16))[name = string("x_rot_75_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5128_cast_fp16 = mul(x = x_209_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_5128_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5129_cast_fp16 = mul(x = x_rot_75_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_5129_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_61_cast_fp16 = add(x = var_5128_cast_fp16, y = var_5129_cast_fp16)[name = string("q_h_61_cast_fp16")];
            fp16 var_5132_promoted_to_fp16 = const()[name = string("op_5132_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_5133_cast_fp16 = pow(x = var_4927_cast_fp16_3, y = var_5132_promoted_to_fp16)[name = string("op_5133_cast_fp16")];
            tensor<int32, [1]> var_135_axes_0 = const()[name = string("var_135_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_135_keep_dims_0 = const()[name = string("var_135_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_135_cast_fp16 = reduce_mean(axes = var_135_axes_0, keep_dims = var_135_keep_dims_0, x = var_5133_cast_fp16)[name = string("var_135_cast_fp16")];
            fp16 var_5140_to_fp16 = const()[name = string("op_5140_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_5141_cast_fp16 = add(x = var_135_cast_fp16, y = var_5140_to_fp16)[name = string("op_5141_cast_fp16")];
            fp32 var_5142_epsilon_0 = const()[name = string("op_5142_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_5142_cast_fp16 = rsqrt(epsilon = var_5142_epsilon_0, x = var_5141_cast_fp16)[name = string("op_5142_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_135_cast_fp16 = mul(x = var_4927_cast_fp16_3, y = var_5142_cast_fp16)[name = string("x_normed_135_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_213_cast_fp16 = mul(x = x_normed_135_cast_fp16, y = var_4960_to_fp16)[name = string("x_213_cast_fp16")];
            tensor<int32, [4]> x_first_77_begin_0 = const()[name = string("x_first_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_77_end_0 = const()[name = string("x_first_77_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_77_end_mask_0 = const()[name = string("x_first_77_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_77_cast_fp16 = slice_by_index(begin = x_first_77_begin_0, end = x_first_77_end_0, end_mask = x_first_77_end_mask_0, x = x_213_cast_fp16)[name = string("x_first_77_cast_fp16")];
            tensor<int32, [4]> x_second_77_begin_0 = const()[name = string("x_second_77_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_77_end_0 = const()[name = string("x_second_77_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_77_end_mask_0 = const()[name = string("x_second_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_77_cast_fp16 = slice_by_index(begin = x_second_77_begin_0, end = x_second_77_end_0, end_mask = x_second_77_end_mask_0, x = x_213_cast_fp16)[name = string("x_second_77_cast_fp16")];
            fp16 const_63_promoted_to_fp16 = const()[name = string("const_63_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_5174_cast_fp16 = mul(x = x_second_77_cast_fp16, y = const_63_promoted_to_fp16)[name = string("op_5174_cast_fp16")];
            int32 var_5176 = const()[name = string("op_5176"), val = int32(1)];
            bool x_rot_77_interleave_0 = const()[name = string("x_rot_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_77_cast_fp16 = concat(axis = var_5176, interleave = x_rot_77_interleave_0, values = (var_5174_cast_fp16, x_first_77_cast_fp16))[name = string("x_rot_77_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5192_cast_fp16 = mul(x = x_213_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_5192_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5193_cast_fp16 = mul(x = x_rot_77_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_5193_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_63_cast_fp16 = add(x = var_5192_cast_fp16, y = var_5193_cast_fp16)[name = string("q_h_63_cast_fp16")];
            fp16 var_5196_promoted_to_fp16 = const()[name = string("op_5196_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_5197_cast_fp16 = pow(x = var_4913_cast_fp16, y = var_5196_promoted_to_fp16)[name = string("op_5197_cast_fp16")];
            tensor<int32, [1]> var_137_axes_0 = const()[name = string("var_137_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_137_keep_dims_0 = const()[name = string("var_137_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_137_cast_fp16 = reduce_mean(axes = var_137_axes_0, keep_dims = var_137_keep_dims_0, x = var_5197_cast_fp16)[name = string("var_137_cast_fp16")];
            fp16 var_5204_to_fp16 = const()[name = string("op_5204_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_5205_cast_fp16 = add(x = var_137_cast_fp16, y = var_5204_to_fp16)[name = string("op_5205_cast_fp16")];
            fp32 var_5206_epsilon_0 = const()[name = string("op_5206_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_5206_cast_fp16 = rsqrt(epsilon = var_5206_epsilon_0, x = var_5205_cast_fp16)[name = string("op_5206_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_137_cast_fp16 = mul(x = var_4913_cast_fp16, y = var_5206_cast_fp16)[name = string("x_normed_137_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_5216_to_fp16 = const()[name = string("op_5216_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415549312)))];
            tensor<fp16, [1, 256, 1, ?]> x_217_cast_fp16 = mul(x = x_normed_137_cast_fp16, y = var_5216_to_fp16)[name = string("x_217_cast_fp16")];
            tensor<int32, [4]> x_first_79_begin_0 = const()[name = string("x_first_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_79_end_0 = const()[name = string("x_first_79_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_79_end_mask_0 = const()[name = string("x_first_79_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_79_cast_fp16 = slice_by_index(begin = x_first_79_begin_0, end = x_first_79_end_0, end_mask = x_first_79_end_mask_0, x = x_217_cast_fp16)[name = string("x_first_79_cast_fp16")];
            tensor<int32, [4]> x_second_79_begin_0 = const()[name = string("x_second_79_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_79_end_0 = const()[name = string("x_second_79_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_79_end_mask_0 = const()[name = string("x_second_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_79_cast_fp16 = slice_by_index(begin = x_second_79_begin_0, end = x_second_79_end_0, end_mask = x_second_79_end_mask_0, x = x_217_cast_fp16)[name = string("x_second_79_cast_fp16")];
            fp16 const_64_promoted_to_fp16 = const()[name = string("const_64_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_5238_cast_fp16 = mul(x = x_second_79_cast_fp16, y = const_64_promoted_to_fp16)[name = string("op_5238_cast_fp16")];
            int32 var_5240 = const()[name = string("op_5240"), val = int32(1)];
            bool x_rot_79_interleave_0 = const()[name = string("x_rot_79_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_79_cast_fp16 = concat(axis = var_5240, interleave = x_rot_79_interleave_0, values = (var_5238_cast_fp16, x_first_79_cast_fp16))[name = string("x_rot_79_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5256_cast_fp16 = mul(x = x_217_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_5256_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5257_cast_fp16 = mul(x = x_rot_79_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_5257_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_15_cast_fp16 = add(x = var_5256_cast_fp16, y = var_5257_cast_fp16)[name = string("k_new_15_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_14 = read_state(input = k_cache_7)[name = string("read_state_14")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_45_cast_fp16 = mul(x = read_state_14, y = one_minus_1_cast_fp16)[name = string("k_cache_45_cast_fp16")];
            write_state(data = k_cache_45_cast_fp16, input = k_cache_7)[name = string("coreml_update_state_100_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_100 = read_state(input = k_cache_7)[name = string("coreml_update_state_100")];
            tensor<fp16, [1, 256, 1, 128]> var_5264_cast_fp16 = mul(x = k_new_15_cast_fp16, y = write_mask)[name = string("op_5264_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_47_cast_fp16 = add(x = coreml_update_state_100, y = var_5264_cast_fp16)[name = string("k_cache_47_cast_fp16")];
            write_state(data = k_cache_47_cast_fp16, input = k_cache_7)[name = string("coreml_update_state_101_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_101 = read_state(input = k_cache_7)[name = string("coreml_update_state_101")];
            tensor<fp16, [1, 256, 1, 128]> read_state_15 = read_state(input = v_cache_7)[name = string("read_state_15")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_45_cast_fp16 = mul(x = read_state_15, y = one_minus_1_cast_fp16)[name = string("v_cache_45_cast_fp16")];
            write_state(data = v_cache_45_cast_fp16, input = v_cache_7)[name = string("coreml_update_state_102_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_102 = read_state(input = v_cache_7)[name = string("coreml_update_state_102")];
            tensor<fp16, [1, 256, 1, 128]> var_5268_cast_fp16 = mul(x = var_4924_cast_fp16, y = write_mask)[name = string("op_5268_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_47_cast_fp16 = add(x = coreml_update_state_102, y = var_5268_cast_fp16)[name = string("v_cache_47_cast_fp16")];
            write_state(data = v_cache_47_cast_fp16, input = v_cache_7)[name = string("coreml_update_state_103_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_103 = read_state(input = v_cache_7)[name = string("coreml_update_state_103")];
            bool var_5283_transpose_x_0 = const()[name = string("op_5283_transpose_x_0"), val = bool(false)];
            bool var_5283_transpose_y_0 = const()[name = string("op_5283_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_362_perm_0 = const()[name = string("transpose_362_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_363_perm_0 = const()[name = string("transpose_363_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_363 = transpose(perm = transpose_363_perm_0, x = coreml_update_state_101)[name = string("transpose_525")];
            tensor<fp16, [1, 1, ?, 256]> transpose_362 = transpose(perm = transpose_362_perm_0, x = q_h_57_cast_fp16)[name = string("transpose_526")];
            tensor<fp16, [1, 1, ?, 128]> var_5283_cast_fp16 = matmul(transpose_x = var_5283_transpose_x_0, transpose_y = var_5283_transpose_y_0, x = transpose_362, y = transpose_363)[name = string("op_5283_cast_fp16")];
            fp16 var_5284_to_fp16 = const()[name = string("op_5284_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_5285_cast_fp16 = mul(x = var_5283_cast_fp16, y = var_5284_to_fp16)[name = string("op_5285_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_57_cast_fp16 = add(x = var_5285_cast_fp16, y = attention_mask)[name = string("scores_57_cast_fp16")];
            int32 var_5288 = const()[name = string("op_5288"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_57_cast_fp16 = softmax(axis = var_5288, x = scores_57_cast_fp16)[name = string("probs_57_cast_fp16")];
            tensor<int32, [4]> var_5295 = const()[name = string("op_5295"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_5297_transpose_x_0 = const()[name = string("op_5297_transpose_x_0"), val = bool(false)];
            bool var_5297_transpose_y_0 = const()[name = string("op_5297_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_57_cast_fp16 = transpose(perm = var_5295, x = coreml_update_state_103)[name = string("transpose_524")];
            tensor<fp16, [1, 1, ?, 256]> var_5297_cast_fp16 = matmul(transpose_x = var_5297_transpose_x_0, transpose_y = var_5297_transpose_y_0, x = probs_57_cast_fp16, y = v_t_57_cast_fp16)[name = string("op_5297_cast_fp16")];
            bool var_5316_transpose_x_0 = const()[name = string("op_5316_transpose_x_0"), val = bool(false)];
            bool var_5316_transpose_y_0 = const()[name = string("op_5316_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_367_perm_0 = const()[name = string("transpose_367_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_367 = transpose(perm = transpose_367_perm_0, x = q_h_59_cast_fp16)[name = string("transpose_523")];
            tensor<fp16, [1, 1, ?, 128]> var_5316_cast_fp16 = matmul(transpose_x = var_5316_transpose_x_0, transpose_y = var_5316_transpose_y_0, x = transpose_367, y = transpose_363)[name = string("op_5316_cast_fp16")];
            fp16 var_5317_to_fp16 = const()[name = string("op_5317_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_5318_cast_fp16 = mul(x = var_5316_cast_fp16, y = var_5317_to_fp16)[name = string("op_5318_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_59_cast_fp16 = add(x = var_5318_cast_fp16, y = attention_mask)[name = string("scores_59_cast_fp16")];
            int32 var_5321 = const()[name = string("op_5321"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_59_cast_fp16 = softmax(axis = var_5321, x = scores_59_cast_fp16)[name = string("probs_59_cast_fp16")];
            bool var_5330_transpose_x_0 = const()[name = string("op_5330_transpose_x_0"), val = bool(false)];
            bool var_5330_transpose_y_0 = const()[name = string("op_5330_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_5330_cast_fp16 = matmul(transpose_x = var_5330_transpose_x_0, transpose_y = var_5330_transpose_y_0, x = probs_59_cast_fp16, y = v_t_57_cast_fp16)[name = string("op_5330_cast_fp16")];
            bool var_5349_transpose_x_0 = const()[name = string("op_5349_transpose_x_0"), val = bool(false)];
            bool var_5349_transpose_y_0 = const()[name = string("op_5349_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_368_perm_0 = const()[name = string("transpose_368_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_368 = transpose(perm = transpose_368_perm_0, x = q_h_61_cast_fp16)[name = string("transpose_522")];
            tensor<fp16, [1, 1, ?, 128]> var_5349_cast_fp16 = matmul(transpose_x = var_5349_transpose_x_0, transpose_y = var_5349_transpose_y_0, x = transpose_368, y = transpose_363)[name = string("op_5349_cast_fp16")];
            fp16 var_5350_to_fp16 = const()[name = string("op_5350_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_5351_cast_fp16 = mul(x = var_5349_cast_fp16, y = var_5350_to_fp16)[name = string("op_5351_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_61_cast_fp16 = add(x = var_5351_cast_fp16, y = attention_mask)[name = string("scores_61_cast_fp16")];
            int32 var_5354 = const()[name = string("op_5354"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_61_cast_fp16 = softmax(axis = var_5354, x = scores_61_cast_fp16)[name = string("probs_61_cast_fp16")];
            bool var_5363_transpose_x_0 = const()[name = string("op_5363_transpose_x_0"), val = bool(false)];
            bool var_5363_transpose_y_0 = const()[name = string("op_5363_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_5363_cast_fp16 = matmul(transpose_x = var_5363_transpose_x_0, transpose_y = var_5363_transpose_y_0, x = probs_61_cast_fp16, y = v_t_57_cast_fp16)[name = string("op_5363_cast_fp16")];
            bool var_5382_transpose_x_0 = const()[name = string("op_5382_transpose_x_0"), val = bool(false)];
            bool var_5382_transpose_y_0 = const()[name = string("op_5382_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_369_perm_0 = const()[name = string("transpose_369_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_369 = transpose(perm = transpose_369_perm_0, x = q_h_63_cast_fp16)[name = string("transpose_521")];
            tensor<fp16, [1, 1, ?, 128]> var_5382_cast_fp16 = matmul(transpose_x = var_5382_transpose_x_0, transpose_y = var_5382_transpose_y_0, x = transpose_369, y = transpose_363)[name = string("op_5382_cast_fp16")];
            fp16 var_5383_to_fp16 = const()[name = string("op_5383_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_5384_cast_fp16 = mul(x = var_5382_cast_fp16, y = var_5383_to_fp16)[name = string("op_5384_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_63_cast_fp16 = add(x = var_5384_cast_fp16, y = attention_mask)[name = string("scores_63_cast_fp16")];
            int32 var_5387 = const()[name = string("op_5387"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_63_cast_fp16 = softmax(axis = var_5387, x = scores_63_cast_fp16)[name = string("probs_63_cast_fp16")];
            bool var_5396_transpose_x_0 = const()[name = string("op_5396_transpose_x_0"), val = bool(false)];
            bool var_5396_transpose_y_0 = const()[name = string("op_5396_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_5396_cast_fp16 = matmul(transpose_x = var_5396_transpose_x_0, transpose_y = var_5396_transpose_y_0, x = probs_63_cast_fp16, y = v_t_57_cast_fp16)[name = string("op_5396_cast_fp16")];
            bool input_59_interleave_0 = const()[name = string("input_59_interleave_0"), val = bool(false)];
            int32 const_152 = const()[name = string("const_152"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_59_cast_fp16 = concat(axis = const_152, interleave = input_59_interleave_0, values = (var_5297_cast_fp16, var_5330_cast_fp16, var_5363_cast_fp16, var_5396_cast_fp16))[name = string("input_59_cast_fp16")];
            string x_219_pad_type_0 = const()[name = string("x_219_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_219_strides_0 = const()[name = string("x_219_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_219_pad_0 = const()[name = string("x_219_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_219_dilations_0 = const()[name = string("x_219_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_219_groups_0 = const()[name = string("x_219_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_71_perm_0 = const()[name = string("transpose_71_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_7_weight_to_fp16 = const()[name = string("o_proj_7_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415549888)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_71_cast_fp16 = transpose(perm = transpose_71_perm_0, x = input_59_cast_fp16)[name = string("transpose_520")];
            tensor<fp16, [1, 640, 1, ?]> x_219_cast_fp16 = conv(dilations = x_219_dilations_0, groups = x_219_groups_0, pad = x_219_pad_0, pad_type = x_219_pad_type_0, strides = x_219_strides_0, weight = o_proj_7_weight_to_fp16, x = transpose_71_cast_fp16)[name = string("x_219_cast_fp16")];
            fp16 var_5417_promoted_to_fp16 = const()[name = string("op_5417_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_5418_cast_fp16 = pow(x = x_219_cast_fp16, y = var_5417_promoted_to_fp16)[name = string("op_5418_cast_fp16")];
            tensor<int32, [1]> var_139_axes_0 = const()[name = string("var_139_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_139_keep_dims_0 = const()[name = string("var_139_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_139_cast_fp16 = reduce_mean(axes = var_139_axes_0, keep_dims = var_139_keep_dims_0, x = var_5418_cast_fp16)[name = string("var_139_cast_fp16")];
            fp16 var_5425_to_fp16 = const()[name = string("op_5425_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_5426_cast_fp16 = add(x = var_139_cast_fp16, y = var_5425_to_fp16)[name = string("op_5426_cast_fp16")];
            fp32 var_5427_epsilon_0 = const()[name = string("op_5427_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_5427_cast_fp16 = rsqrt(epsilon = var_5427_epsilon_0, x = var_5426_cast_fp16)[name = string("op_5427_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_139_cast_fp16 = mul(x = x_219_cast_fp16, y = var_5427_cast_fp16)[name = string("x_normed_139_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_5437_to_fp16 = const()[name = string("op_5437_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416860672)))];
            tensor<fp16, [1, 640, 1, ?]> attn_31_cast_fp16 = mul(x = x_normed_139_cast_fp16, y = var_5437_to_fp16)[name = string("attn_31_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_221_cast_fp16 = add(x = x_197_cast_fp16, y = attn_31_cast_fp16)[name = string("x_221_cast_fp16")];
            fp16 var_5441_promoted_to_fp16 = const()[name = string("op_5441_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_5442_cast_fp16 = pow(x = x_221_cast_fp16, y = var_5441_promoted_to_fp16)[name = string("op_5442_cast_fp16")];
            tensor<int32, [1]> var_141_axes_0 = const()[name = string("var_141_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_141_keep_dims_0 = const()[name = string("var_141_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_141_cast_fp16 = reduce_mean(axes = var_141_axes_0, keep_dims = var_141_keep_dims_0, x = var_5442_cast_fp16)[name = string("var_141_cast_fp16")];
            fp16 var_5449_to_fp16 = const()[name = string("op_5449_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_5450_cast_fp16 = add(x = var_141_cast_fp16, y = var_5449_to_fp16)[name = string("op_5450_cast_fp16")];
            fp32 var_5451_epsilon_0 = const()[name = string("op_5451_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_5451_cast_fp16 = rsqrt(epsilon = var_5451_epsilon_0, x = var_5450_cast_fp16)[name = string("op_5451_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_141_cast_fp16 = mul(x = x_221_cast_fp16, y = var_5451_cast_fp16)[name = string("x_normed_141_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_5461_to_fp16 = const()[name = string("op_5461_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416862016)))];
            tensor<fp16, [1, 640, 1, ?]> input_61_cast_fp16 = mul(x = x_normed_141_cast_fp16, y = var_5461_to_fp16)[name = string("input_61_cast_fp16")];
            string gate_15_pad_type_0 = const()[name = string("gate_15_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_15_strides_0 = const()[name = string("gate_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_15_pad_0 = const()[name = string("gate_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_15_dilations_0 = const()[name = string("gate_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_15_groups_0 = const()[name = string("gate_15_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_7_weight_to_fp16 = const()[name = string("gate_proj_7_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416863360)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_15_cast_fp16 = conv(dilations = gate_15_dilations_0, groups = gate_15_groups_0, pad = gate_15_pad_0, pad_type = gate_15_pad_type_0, strides = gate_15_strides_0, weight = gate_proj_7_weight_to_fp16, x = input_61_cast_fp16)[name = string("gate_15_cast_fp16")];
            string up_15_pad_type_0 = const()[name = string("up_15_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_15_strides_0 = const()[name = string("up_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_15_pad_0 = const()[name = string("up_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_15_dilations_0 = const()[name = string("up_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_15_groups_0 = const()[name = string("up_15_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_7_weight_to_fp16 = const()[name = string("up_proj_7_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419484864)))];
            tensor<fp16, [1, 2048, 1, ?]> up_15_cast_fp16 = conv(dilations = up_15_dilations_0, groups = up_15_groups_0, pad = up_15_pad_0, pad_type = up_15_pad_type_0, strides = up_15_strides_0, weight = up_proj_7_weight_to_fp16, x = input_61_cast_fp16)[name = string("up_15_cast_fp16")];
            string var_5486_mode_0 = const()[name = string("op_5486_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_5486_cast_fp16 = gelu(mode = var_5486_mode_0, x = gate_15_cast_fp16)[name = string("op_5486_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_63_cast_fp16 = mul(x = var_5486_cast_fp16, y = up_15_cast_fp16)[name = string("input_63_cast_fp16")];
            string x_223_pad_type_0 = const()[name = string("x_223_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_223_strides_0 = const()[name = string("x_223_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_223_pad_0 = const()[name = string("x_223_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_223_dilations_0 = const()[name = string("x_223_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_223_groups_0 = const()[name = string("x_223_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_7_weight_to_fp16 = const()[name = string("down_proj_7_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422106368)))];
            tensor<fp16, [1, 640, 1, ?]> x_223_cast_fp16 = conv(dilations = x_223_dilations_0, groups = x_223_groups_0, pad = x_223_pad_0, pad_type = x_223_pad_type_0, strides = x_223_strides_0, weight = down_proj_7_weight_to_fp16, x = input_63_cast_fp16)[name = string("x_223_cast_fp16")];
            fp16 var_5499_promoted_to_fp16 = const()[name = string("op_5499_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_5500_cast_fp16 = pow(x = x_223_cast_fp16, y = var_5499_promoted_to_fp16)[name = string("op_5500_cast_fp16")];
            tensor<int32, [1]> var_143_axes_0 = const()[name = string("var_143_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_143_keep_dims_0 = const()[name = string("var_143_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_143_cast_fp16 = reduce_mean(axes = var_143_axes_0, keep_dims = var_143_keep_dims_0, x = var_5500_cast_fp16)[name = string("var_143_cast_fp16")];
            fp16 var_5507_to_fp16 = const()[name = string("op_5507_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_5508_cast_fp16 = add(x = var_143_cast_fp16, y = var_5507_to_fp16)[name = string("op_5508_cast_fp16")];
            fp32 var_5509_epsilon_0 = const()[name = string("op_5509_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_5509_cast_fp16 = rsqrt(epsilon = var_5509_epsilon_0, x = var_5508_cast_fp16)[name = string("op_5509_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_143_cast_fp16 = mul(x = x_223_cast_fp16, y = var_5509_cast_fp16)[name = string("x_normed_143_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_5519_to_fp16 = const()[name = string("op_5519_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424727872)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_15_cast_fp16 = mul(x = x_normed_143_cast_fp16, y = var_5519_to_fp16)[name = string("mlp_15_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_225_cast_fp16 = add(x = x_221_cast_fp16, y = mlp_15_cast_fp16)[name = string("x_225_cast_fp16")];
            fp16 var_5523_promoted_to_fp16 = const()[name = string("op_5523_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_5524_cast_fp16 = pow(x = x_225_cast_fp16, y = var_5523_promoted_to_fp16)[name = string("op_5524_cast_fp16")];
            tensor<int32, [1]> var_145_axes_0 = const()[name = string("var_145_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_145_keep_dims_0 = const()[name = string("var_145_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_145_cast_fp16 = reduce_mean(axes = var_145_axes_0, keep_dims = var_145_keep_dims_0, x = var_5524_cast_fp16)[name = string("var_145_cast_fp16")];
            fp16 var_5531_to_fp16 = const()[name = string("op_5531_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_5532_cast_fp16 = add(x = var_145_cast_fp16, y = var_5531_to_fp16)[name = string("op_5532_cast_fp16")];
            fp32 var_5533_epsilon_0 = const()[name = string("op_5533_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_5533_cast_fp16 = rsqrt(epsilon = var_5533_epsilon_0, x = var_5532_cast_fp16)[name = string("op_5533_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_145_cast_fp16 = mul(x = x_225_cast_fp16, y = var_5533_cast_fp16)[name = string("x_normed_145_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_5543_to_fp16 = const()[name = string("op_5543_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424729216)))];
            tensor<fp16, [1, 640, 1, ?]> input_65_cast_fp16 = mul(x = x_normed_145_cast_fp16, y = var_5543_to_fp16)[name = string("input_65_cast_fp16")];
            string var_5555_pad_type_0 = const()[name = string("op_5555_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5555_strides_0 = const()[name = string("op_5555_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5555_pad_0 = const()[name = string("op_5555_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5555_dilations_0 = const()[name = string("op_5555_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5555_groups_0 = const()[name = string("op_5555_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_8_weight_to_fp16 = const()[name = string("q_proj_8_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424730560)))];
            tensor<fp16, [1, 1024, 1, ?]> var_5555_cast_fp16 = conv(dilations = var_5555_dilations_0, groups = var_5555_groups_0, pad = var_5555_pad_0, pad_type = var_5555_pad_type_0, strides = var_5555_strides_0, weight = q_proj_8_weight_to_fp16, x = input_65_cast_fp16)[name = string("op_5555_cast_fp16")];
            string var_5566_pad_type_0 = const()[name = string("op_5566_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5566_strides_0 = const()[name = string("op_5566_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5566_pad_0 = const()[name = string("op_5566_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5566_dilations_0 = const()[name = string("op_5566_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5566_groups_0 = const()[name = string("op_5566_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_8_weight_to_fp16 = const()[name = string("k_proj_8_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426041344)))];
            tensor<fp16, [1, 256, 1, ?]> var_5566_cast_fp16 = conv(dilations = var_5566_dilations_0, groups = var_5566_groups_0, pad = var_5566_pad_0, pad_type = var_5566_pad_type_0, strides = var_5566_strides_0, weight = k_proj_8_weight_to_fp16, x = input_65_cast_fp16)[name = string("op_5566_cast_fp16")];
            string var_5577_pad_type_0 = const()[name = string("op_5577_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5577_strides_0 = const()[name = string("op_5577_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5577_pad_0 = const()[name = string("op_5577_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5577_dilations_0 = const()[name = string("op_5577_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5577_groups_0 = const()[name = string("op_5577_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_8_weight_to_fp16 = const()[name = string("v_proj_8_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426369088)))];
            tensor<fp16, [1, 256, 1, ?]> var_5577_cast_fp16 = conv(dilations = var_5577_dilations_0, groups = var_5577_groups_0, pad = var_5577_pad_0, pad_type = var_5577_pad_type_0, strides = var_5577_strides_0, weight = v_proj_8_weight_to_fp16, x = input_65_cast_fp16)[name = string("op_5577_cast_fp16")];
            tensor<int32, [4]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_5580_axis_0 = const()[name = string("op_5580_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_5580_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_5580_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_5580_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_5580_cast_fp16_3 = split(axis = var_5580_axis_0, split_sizes = tile_24, x = var_5555_cast_fp16)[name = string("op_5580_cast_fp16")];
            fp16 var_5593_promoted_to_fp16 = const()[name = string("op_5593_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_5594_cast_fp16 = pow(x = var_5580_cast_fp16_0, y = var_5593_promoted_to_fp16)[name = string("op_5594_cast_fp16")];
            tensor<int32, [1]> var_147_axes_0 = const()[name = string("var_147_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_147_keep_dims_0 = const()[name = string("var_147_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_147_cast_fp16 = reduce_mean(axes = var_147_axes_0, keep_dims = var_147_keep_dims_0, x = var_5594_cast_fp16)[name = string("var_147_cast_fp16")];
            fp16 var_5601_to_fp16 = const()[name = string("op_5601_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_5602_cast_fp16 = add(x = var_147_cast_fp16, y = var_5601_to_fp16)[name = string("op_5602_cast_fp16")];
            fp32 var_5603_epsilon_0 = const()[name = string("op_5603_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_5603_cast_fp16 = rsqrt(epsilon = var_5603_epsilon_0, x = var_5602_cast_fp16)[name = string("op_5603_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_147_cast_fp16 = mul(x = var_5580_cast_fp16_0, y = var_5603_cast_fp16)[name = string("x_normed_147_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_5613_to_fp16 = const()[name = string("op_5613_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426696832)))];
            tensor<fp16, [1, 256, 1, ?]> x_229_cast_fp16 = mul(x = x_normed_147_cast_fp16, y = var_5613_to_fp16)[name = string("x_229_cast_fp16")];
            tensor<int32, [4]> x_first_81_begin_0 = const()[name = string("x_first_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_81_end_0 = const()[name = string("x_first_81_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_81_end_mask_0 = const()[name = string("x_first_81_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_81_cast_fp16 = slice_by_index(begin = x_first_81_begin_0, end = x_first_81_end_0, end_mask = x_first_81_end_mask_0, x = x_229_cast_fp16)[name = string("x_first_81_cast_fp16")];
            tensor<int32, [4]> x_second_81_begin_0 = const()[name = string("x_second_81_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_81_end_0 = const()[name = string("x_second_81_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_81_end_mask_0 = const()[name = string("x_second_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_81_cast_fp16 = slice_by_index(begin = x_second_81_begin_0, end = x_second_81_end_0, end_mask = x_second_81_end_mask_0, x = x_229_cast_fp16)[name = string("x_second_81_cast_fp16")];
            fp16 const_68_promoted_to_fp16 = const()[name = string("const_68_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_5635_cast_fp16 = mul(x = x_second_81_cast_fp16, y = const_68_promoted_to_fp16)[name = string("op_5635_cast_fp16")];
            int32 var_5637 = const()[name = string("op_5637"), val = int32(1)];
            bool x_rot_81_interleave_0 = const()[name = string("x_rot_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_81_cast_fp16 = concat(axis = var_5637, interleave = x_rot_81_interleave_0, values = (var_5635_cast_fp16, x_first_81_cast_fp16))[name = string("x_rot_81_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5653_cast_fp16 = mul(x = x_229_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_5653_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5654_cast_fp16 = mul(x = x_rot_81_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_5654_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_65_cast_fp16 = add(x = var_5653_cast_fp16, y = var_5654_cast_fp16)[name = string("q_h_65_cast_fp16")];
            fp16 var_5657_promoted_to_fp16 = const()[name = string("op_5657_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_5658_cast_fp16 = pow(x = var_5580_cast_fp16_1, y = var_5657_promoted_to_fp16)[name = string("op_5658_cast_fp16")];
            tensor<int32, [1]> var_149_axes_0 = const()[name = string("var_149_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_149_keep_dims_0 = const()[name = string("var_149_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_149_cast_fp16 = reduce_mean(axes = var_149_axes_0, keep_dims = var_149_keep_dims_0, x = var_5658_cast_fp16)[name = string("var_149_cast_fp16")];
            fp16 var_5665_to_fp16 = const()[name = string("op_5665_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_5666_cast_fp16 = add(x = var_149_cast_fp16, y = var_5665_to_fp16)[name = string("op_5666_cast_fp16")];
            fp32 var_5667_epsilon_0 = const()[name = string("op_5667_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_5667_cast_fp16 = rsqrt(epsilon = var_5667_epsilon_0, x = var_5666_cast_fp16)[name = string("op_5667_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_149_cast_fp16 = mul(x = var_5580_cast_fp16_1, y = var_5667_cast_fp16)[name = string("x_normed_149_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_233_cast_fp16 = mul(x = x_normed_149_cast_fp16, y = var_5613_to_fp16)[name = string("x_233_cast_fp16")];
            tensor<int32, [4]> x_first_83_begin_0 = const()[name = string("x_first_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_83_end_0 = const()[name = string("x_first_83_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_83_end_mask_0 = const()[name = string("x_first_83_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_83_cast_fp16 = slice_by_index(begin = x_first_83_begin_0, end = x_first_83_end_0, end_mask = x_first_83_end_mask_0, x = x_233_cast_fp16)[name = string("x_first_83_cast_fp16")];
            tensor<int32, [4]> x_second_83_begin_0 = const()[name = string("x_second_83_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_83_end_0 = const()[name = string("x_second_83_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_83_end_mask_0 = const()[name = string("x_second_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_83_cast_fp16 = slice_by_index(begin = x_second_83_begin_0, end = x_second_83_end_0, end_mask = x_second_83_end_mask_0, x = x_233_cast_fp16)[name = string("x_second_83_cast_fp16")];
            fp16 const_69_promoted_to_fp16 = const()[name = string("const_69_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_5699_cast_fp16 = mul(x = x_second_83_cast_fp16, y = const_69_promoted_to_fp16)[name = string("op_5699_cast_fp16")];
            int32 var_5701 = const()[name = string("op_5701"), val = int32(1)];
            bool x_rot_83_interleave_0 = const()[name = string("x_rot_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_83_cast_fp16 = concat(axis = var_5701, interleave = x_rot_83_interleave_0, values = (var_5699_cast_fp16, x_first_83_cast_fp16))[name = string("x_rot_83_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5717_cast_fp16 = mul(x = x_233_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_5717_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5718_cast_fp16 = mul(x = x_rot_83_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_5718_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_67_cast_fp16 = add(x = var_5717_cast_fp16, y = var_5718_cast_fp16)[name = string("q_h_67_cast_fp16")];
            fp16 var_5721_promoted_to_fp16 = const()[name = string("op_5721_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_5722_cast_fp16 = pow(x = var_5580_cast_fp16_2, y = var_5721_promoted_to_fp16)[name = string("op_5722_cast_fp16")];
            tensor<int32, [1]> var_151_axes_0 = const()[name = string("var_151_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_151_keep_dims_0 = const()[name = string("var_151_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_151_cast_fp16 = reduce_mean(axes = var_151_axes_0, keep_dims = var_151_keep_dims_0, x = var_5722_cast_fp16)[name = string("var_151_cast_fp16")];
            fp16 var_5729_to_fp16 = const()[name = string("op_5729_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_5730_cast_fp16 = add(x = var_151_cast_fp16, y = var_5729_to_fp16)[name = string("op_5730_cast_fp16")];
            fp32 var_5731_epsilon_0 = const()[name = string("op_5731_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_5731_cast_fp16 = rsqrt(epsilon = var_5731_epsilon_0, x = var_5730_cast_fp16)[name = string("op_5731_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_151_cast_fp16 = mul(x = var_5580_cast_fp16_2, y = var_5731_cast_fp16)[name = string("x_normed_151_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_237_cast_fp16 = mul(x = x_normed_151_cast_fp16, y = var_5613_to_fp16)[name = string("x_237_cast_fp16")];
            tensor<int32, [4]> x_first_85_begin_0 = const()[name = string("x_first_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_85_end_0 = const()[name = string("x_first_85_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_85_end_mask_0 = const()[name = string("x_first_85_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_85_cast_fp16 = slice_by_index(begin = x_first_85_begin_0, end = x_first_85_end_0, end_mask = x_first_85_end_mask_0, x = x_237_cast_fp16)[name = string("x_first_85_cast_fp16")];
            tensor<int32, [4]> x_second_85_begin_0 = const()[name = string("x_second_85_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_85_end_0 = const()[name = string("x_second_85_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_85_end_mask_0 = const()[name = string("x_second_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_85_cast_fp16 = slice_by_index(begin = x_second_85_begin_0, end = x_second_85_end_0, end_mask = x_second_85_end_mask_0, x = x_237_cast_fp16)[name = string("x_second_85_cast_fp16")];
            fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_5763_cast_fp16 = mul(x = x_second_85_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_5763_cast_fp16")];
            int32 var_5765 = const()[name = string("op_5765"), val = int32(1)];
            bool x_rot_85_interleave_0 = const()[name = string("x_rot_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_85_cast_fp16 = concat(axis = var_5765, interleave = x_rot_85_interleave_0, values = (var_5763_cast_fp16, x_first_85_cast_fp16))[name = string("x_rot_85_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5781_cast_fp16 = mul(x = x_237_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_5781_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5782_cast_fp16 = mul(x = x_rot_85_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_5782_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_69_cast_fp16 = add(x = var_5781_cast_fp16, y = var_5782_cast_fp16)[name = string("q_h_69_cast_fp16")];
            fp16 var_5785_promoted_to_fp16 = const()[name = string("op_5785_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_5786_cast_fp16 = pow(x = var_5580_cast_fp16_3, y = var_5785_promoted_to_fp16)[name = string("op_5786_cast_fp16")];
            tensor<int32, [1]> var_153_axes_0 = const()[name = string("var_153_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_153_keep_dims_0 = const()[name = string("var_153_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_153_cast_fp16 = reduce_mean(axes = var_153_axes_0, keep_dims = var_153_keep_dims_0, x = var_5786_cast_fp16)[name = string("var_153_cast_fp16")];
            fp16 var_5793_to_fp16 = const()[name = string("op_5793_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_5794_cast_fp16 = add(x = var_153_cast_fp16, y = var_5793_to_fp16)[name = string("op_5794_cast_fp16")];
            fp32 var_5795_epsilon_0 = const()[name = string("op_5795_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_5795_cast_fp16 = rsqrt(epsilon = var_5795_epsilon_0, x = var_5794_cast_fp16)[name = string("op_5795_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_153_cast_fp16 = mul(x = var_5580_cast_fp16_3, y = var_5795_cast_fp16)[name = string("x_normed_153_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_241_cast_fp16 = mul(x = x_normed_153_cast_fp16, y = var_5613_to_fp16)[name = string("x_241_cast_fp16")];
            tensor<int32, [4]> x_first_87_begin_0 = const()[name = string("x_first_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_87_end_0 = const()[name = string("x_first_87_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_87_end_mask_0 = const()[name = string("x_first_87_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_87_cast_fp16 = slice_by_index(begin = x_first_87_begin_0, end = x_first_87_end_0, end_mask = x_first_87_end_mask_0, x = x_241_cast_fp16)[name = string("x_first_87_cast_fp16")];
            tensor<int32, [4]> x_second_87_begin_0 = const()[name = string("x_second_87_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_87_end_0 = const()[name = string("x_second_87_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_87_end_mask_0 = const()[name = string("x_second_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_87_cast_fp16 = slice_by_index(begin = x_second_87_begin_0, end = x_second_87_end_0, end_mask = x_second_87_end_mask_0, x = x_241_cast_fp16)[name = string("x_second_87_cast_fp16")];
            fp16 const_71_promoted_to_fp16 = const()[name = string("const_71_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_5827_cast_fp16 = mul(x = x_second_87_cast_fp16, y = const_71_promoted_to_fp16)[name = string("op_5827_cast_fp16")];
            int32 var_5829 = const()[name = string("op_5829"), val = int32(1)];
            bool x_rot_87_interleave_0 = const()[name = string("x_rot_87_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_87_cast_fp16 = concat(axis = var_5829, interleave = x_rot_87_interleave_0, values = (var_5827_cast_fp16, x_first_87_cast_fp16))[name = string("x_rot_87_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5845_cast_fp16 = mul(x = x_241_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_5845_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5846_cast_fp16 = mul(x = x_rot_87_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_5846_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_71_cast_fp16 = add(x = var_5845_cast_fp16, y = var_5846_cast_fp16)[name = string("q_h_71_cast_fp16")];
            fp16 var_5849_promoted_to_fp16 = const()[name = string("op_5849_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_5850_cast_fp16 = pow(x = var_5566_cast_fp16, y = var_5849_promoted_to_fp16)[name = string("op_5850_cast_fp16")];
            tensor<int32, [1]> var_155_axes_0 = const()[name = string("var_155_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_155_keep_dims_0 = const()[name = string("var_155_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_155_cast_fp16 = reduce_mean(axes = var_155_axes_0, keep_dims = var_155_keep_dims_0, x = var_5850_cast_fp16)[name = string("var_155_cast_fp16")];
            fp16 var_5857_to_fp16 = const()[name = string("op_5857_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_5858_cast_fp16 = add(x = var_155_cast_fp16, y = var_5857_to_fp16)[name = string("op_5858_cast_fp16")];
            fp32 var_5859_epsilon_0 = const()[name = string("op_5859_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_5859_cast_fp16 = rsqrt(epsilon = var_5859_epsilon_0, x = var_5858_cast_fp16)[name = string("op_5859_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_155_cast_fp16 = mul(x = var_5566_cast_fp16, y = var_5859_cast_fp16)[name = string("x_normed_155_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_5869_to_fp16 = const()[name = string("op_5869_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426697408)))];
            tensor<fp16, [1, 256, 1, ?]> x_245_cast_fp16 = mul(x = x_normed_155_cast_fp16, y = var_5869_to_fp16)[name = string("x_245_cast_fp16")];
            tensor<int32, [4]> x_first_89_begin_0 = const()[name = string("x_first_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_89_end_0 = const()[name = string("x_first_89_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_89_end_mask_0 = const()[name = string("x_first_89_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_89_cast_fp16 = slice_by_index(begin = x_first_89_begin_0, end = x_first_89_end_0, end_mask = x_first_89_end_mask_0, x = x_245_cast_fp16)[name = string("x_first_89_cast_fp16")];
            tensor<int32, [4]> x_second_89_begin_0 = const()[name = string("x_second_89_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_89_end_0 = const()[name = string("x_second_89_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_89_end_mask_0 = const()[name = string("x_second_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_89_cast_fp16 = slice_by_index(begin = x_second_89_begin_0, end = x_second_89_end_0, end_mask = x_second_89_end_mask_0, x = x_245_cast_fp16)[name = string("x_second_89_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_5891_cast_fp16 = mul(x = x_second_89_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_5891_cast_fp16")];
            int32 var_5893 = const()[name = string("op_5893"), val = int32(1)];
            bool x_rot_89_interleave_0 = const()[name = string("x_rot_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_89_cast_fp16 = concat(axis = var_5893, interleave = x_rot_89_interleave_0, values = (var_5891_cast_fp16, x_first_89_cast_fp16))[name = string("x_rot_89_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5909_cast_fp16 = mul(x = x_245_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_5909_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_5910_cast_fp16 = mul(x = x_rot_89_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_5910_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_17_cast_fp16 = add(x = var_5909_cast_fp16, y = var_5910_cast_fp16)[name = string("k_new_17_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_16 = read_state(input = k_cache_8)[name = string("read_state_16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_51_cast_fp16 = mul(x = read_state_16, y = one_minus_1_cast_fp16)[name = string("k_cache_51_cast_fp16")];
            write_state(data = k_cache_51_cast_fp16, input = k_cache_8)[name = string("coreml_update_state_104_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_104 = read_state(input = k_cache_8)[name = string("coreml_update_state_104")];
            tensor<fp16, [1, 256, 1, 128]> var_5917_cast_fp16 = mul(x = k_new_17_cast_fp16, y = write_mask)[name = string("op_5917_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_53_cast_fp16 = add(x = coreml_update_state_104, y = var_5917_cast_fp16)[name = string("k_cache_53_cast_fp16")];
            write_state(data = k_cache_53_cast_fp16, input = k_cache_8)[name = string("coreml_update_state_105_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_105 = read_state(input = k_cache_8)[name = string("coreml_update_state_105")];
            tensor<fp16, [1, 256, 1, 128]> read_state_17 = read_state(input = v_cache_8)[name = string("read_state_17")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_51_cast_fp16 = mul(x = read_state_17, y = one_minus_1_cast_fp16)[name = string("v_cache_51_cast_fp16")];
            write_state(data = v_cache_51_cast_fp16, input = v_cache_8)[name = string("coreml_update_state_106_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_106 = read_state(input = v_cache_8)[name = string("coreml_update_state_106")];
            tensor<fp16, [1, 256, 1, 128]> var_5921_cast_fp16 = mul(x = var_5577_cast_fp16, y = write_mask)[name = string("op_5921_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_53_cast_fp16 = add(x = coreml_update_state_106, y = var_5921_cast_fp16)[name = string("v_cache_53_cast_fp16")];
            write_state(data = v_cache_53_cast_fp16, input = v_cache_8)[name = string("coreml_update_state_107_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_107 = read_state(input = v_cache_8)[name = string("coreml_update_state_107")];
            bool var_5936_transpose_x_0 = const()[name = string("op_5936_transpose_x_0"), val = bool(false)];
            bool var_5936_transpose_y_0 = const()[name = string("op_5936_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_370_perm_0 = const()[name = string("transpose_370_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_371_perm_0 = const()[name = string("transpose_371_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_371 = transpose(perm = transpose_371_perm_0, x = coreml_update_state_105)[name = string("transpose_518")];
            tensor<fp16, [1, 1, ?, 256]> transpose_370 = transpose(perm = transpose_370_perm_0, x = q_h_65_cast_fp16)[name = string("transpose_519")];
            tensor<fp16, [1, 1, ?, 128]> var_5936_cast_fp16 = matmul(transpose_x = var_5936_transpose_x_0, transpose_y = var_5936_transpose_y_0, x = transpose_370, y = transpose_371)[name = string("op_5936_cast_fp16")];
            fp16 var_5937_to_fp16 = const()[name = string("op_5937_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_5938_cast_fp16 = mul(x = var_5936_cast_fp16, y = var_5937_to_fp16)[name = string("op_5938_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_65_cast_fp16 = add(x = var_5938_cast_fp16, y = attention_mask)[name = string("scores_65_cast_fp16")];
            int32 var_5941 = const()[name = string("op_5941"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_65_cast_fp16 = softmax(axis = var_5941, x = scores_65_cast_fp16)[name = string("probs_65_cast_fp16")];
            tensor<int32, [4]> var_5948 = const()[name = string("op_5948"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_5950_transpose_x_0 = const()[name = string("op_5950_transpose_x_0"), val = bool(false)];
            bool var_5950_transpose_y_0 = const()[name = string("op_5950_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_65_cast_fp16 = transpose(perm = var_5948, x = coreml_update_state_107)[name = string("transpose_517")];
            tensor<fp16, [1, 1, ?, 256]> var_5950_cast_fp16 = matmul(transpose_x = var_5950_transpose_x_0, transpose_y = var_5950_transpose_y_0, x = probs_65_cast_fp16, y = v_t_65_cast_fp16)[name = string("op_5950_cast_fp16")];
            bool var_5969_transpose_x_0 = const()[name = string("op_5969_transpose_x_0"), val = bool(false)];
            bool var_5969_transpose_y_0 = const()[name = string("op_5969_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_375_perm_0 = const()[name = string("transpose_375_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_375 = transpose(perm = transpose_375_perm_0, x = q_h_67_cast_fp16)[name = string("transpose_516")];
            tensor<fp16, [1, 1, ?, 128]> var_5969_cast_fp16 = matmul(transpose_x = var_5969_transpose_x_0, transpose_y = var_5969_transpose_y_0, x = transpose_375, y = transpose_371)[name = string("op_5969_cast_fp16")];
            fp16 var_5970_to_fp16 = const()[name = string("op_5970_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_5971_cast_fp16 = mul(x = var_5969_cast_fp16, y = var_5970_to_fp16)[name = string("op_5971_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_67_cast_fp16 = add(x = var_5971_cast_fp16, y = attention_mask)[name = string("scores_67_cast_fp16")];
            int32 var_5974 = const()[name = string("op_5974"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_67_cast_fp16 = softmax(axis = var_5974, x = scores_67_cast_fp16)[name = string("probs_67_cast_fp16")];
            bool var_5983_transpose_x_0 = const()[name = string("op_5983_transpose_x_0"), val = bool(false)];
            bool var_5983_transpose_y_0 = const()[name = string("op_5983_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_5983_cast_fp16 = matmul(transpose_x = var_5983_transpose_x_0, transpose_y = var_5983_transpose_y_0, x = probs_67_cast_fp16, y = v_t_65_cast_fp16)[name = string("op_5983_cast_fp16")];
            bool var_6002_transpose_x_0 = const()[name = string("op_6002_transpose_x_0"), val = bool(false)];
            bool var_6002_transpose_y_0 = const()[name = string("op_6002_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_376_perm_0 = const()[name = string("transpose_376_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_376 = transpose(perm = transpose_376_perm_0, x = q_h_69_cast_fp16)[name = string("transpose_515")];
            tensor<fp16, [1, 1, ?, 128]> var_6002_cast_fp16 = matmul(transpose_x = var_6002_transpose_x_0, transpose_y = var_6002_transpose_y_0, x = transpose_376, y = transpose_371)[name = string("op_6002_cast_fp16")];
            fp16 var_6003_to_fp16 = const()[name = string("op_6003_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_6004_cast_fp16 = mul(x = var_6002_cast_fp16, y = var_6003_to_fp16)[name = string("op_6004_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_69_cast_fp16 = add(x = var_6004_cast_fp16, y = attention_mask)[name = string("scores_69_cast_fp16")];
            int32 var_6007 = const()[name = string("op_6007"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_69_cast_fp16 = softmax(axis = var_6007, x = scores_69_cast_fp16)[name = string("probs_69_cast_fp16")];
            bool var_6016_transpose_x_0 = const()[name = string("op_6016_transpose_x_0"), val = bool(false)];
            bool var_6016_transpose_y_0 = const()[name = string("op_6016_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_6016_cast_fp16 = matmul(transpose_x = var_6016_transpose_x_0, transpose_y = var_6016_transpose_y_0, x = probs_69_cast_fp16, y = v_t_65_cast_fp16)[name = string("op_6016_cast_fp16")];
            bool var_6035_transpose_x_0 = const()[name = string("op_6035_transpose_x_0"), val = bool(false)];
            bool var_6035_transpose_y_0 = const()[name = string("op_6035_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_377_perm_0 = const()[name = string("transpose_377_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_377 = transpose(perm = transpose_377_perm_0, x = q_h_71_cast_fp16)[name = string("transpose_514")];
            tensor<fp16, [1, 1, ?, 128]> var_6035_cast_fp16 = matmul(transpose_x = var_6035_transpose_x_0, transpose_y = var_6035_transpose_y_0, x = transpose_377, y = transpose_371)[name = string("op_6035_cast_fp16")];
            fp16 var_6036_to_fp16 = const()[name = string("op_6036_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_6037_cast_fp16 = mul(x = var_6035_cast_fp16, y = var_6036_to_fp16)[name = string("op_6037_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_71_cast_fp16 = add(x = var_6037_cast_fp16, y = attention_mask)[name = string("scores_71_cast_fp16")];
            int32 var_6040 = const()[name = string("op_6040"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_71_cast_fp16 = softmax(axis = var_6040, x = scores_71_cast_fp16)[name = string("probs_71_cast_fp16")];
            bool var_6049_transpose_x_0 = const()[name = string("op_6049_transpose_x_0"), val = bool(false)];
            bool var_6049_transpose_y_0 = const()[name = string("op_6049_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_6049_cast_fp16 = matmul(transpose_x = var_6049_transpose_x_0, transpose_y = var_6049_transpose_y_0, x = probs_71_cast_fp16, y = v_t_65_cast_fp16)[name = string("op_6049_cast_fp16")];
            bool input_67_interleave_0 = const()[name = string("input_67_interleave_0"), val = bool(false)];
            int32 const_153 = const()[name = string("const_153"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_67_cast_fp16 = concat(axis = const_153, interleave = input_67_interleave_0, values = (var_5950_cast_fp16, var_5983_cast_fp16, var_6016_cast_fp16, var_6049_cast_fp16))[name = string("input_67_cast_fp16")];
            string x_247_pad_type_0 = const()[name = string("x_247_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_247_strides_0 = const()[name = string("x_247_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_247_pad_0 = const()[name = string("x_247_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_247_dilations_0 = const()[name = string("x_247_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_247_groups_0 = const()[name = string("x_247_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_80_perm_0 = const()[name = string("transpose_80_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_8_weight_to_fp16 = const()[name = string("o_proj_8_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426697984)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_80_cast_fp16 = transpose(perm = transpose_80_perm_0, x = input_67_cast_fp16)[name = string("transpose_513")];
            tensor<fp16, [1, 640, 1, ?]> x_247_cast_fp16 = conv(dilations = x_247_dilations_0, groups = x_247_groups_0, pad = x_247_pad_0, pad_type = x_247_pad_type_0, strides = x_247_strides_0, weight = o_proj_8_weight_to_fp16, x = transpose_80_cast_fp16)[name = string("x_247_cast_fp16")];
            fp16 var_6070_promoted_to_fp16 = const()[name = string("op_6070_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_6071_cast_fp16 = pow(x = x_247_cast_fp16, y = var_6070_promoted_to_fp16)[name = string("op_6071_cast_fp16")];
            tensor<int32, [1]> var_157_axes_0 = const()[name = string("var_157_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_157_keep_dims_0 = const()[name = string("var_157_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_157_cast_fp16 = reduce_mean(axes = var_157_axes_0, keep_dims = var_157_keep_dims_0, x = var_6071_cast_fp16)[name = string("var_157_cast_fp16")];
            fp16 var_6078_to_fp16 = const()[name = string("op_6078_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_6079_cast_fp16 = add(x = var_157_cast_fp16, y = var_6078_to_fp16)[name = string("op_6079_cast_fp16")];
            fp32 var_6080_epsilon_0 = const()[name = string("op_6080_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_6080_cast_fp16 = rsqrt(epsilon = var_6080_epsilon_0, x = var_6079_cast_fp16)[name = string("op_6080_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_157_cast_fp16 = mul(x = x_247_cast_fp16, y = var_6080_cast_fp16)[name = string("x_normed_157_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_6090_to_fp16 = const()[name = string("op_6090_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428008768)))];
            tensor<fp16, [1, 640, 1, ?]> attn_35_cast_fp16 = mul(x = x_normed_157_cast_fp16, y = var_6090_to_fp16)[name = string("attn_35_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_249_cast_fp16 = add(x = x_225_cast_fp16, y = attn_35_cast_fp16)[name = string("x_249_cast_fp16")];
            fp16 var_6094_promoted_to_fp16 = const()[name = string("op_6094_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_6095_cast_fp16 = pow(x = x_249_cast_fp16, y = var_6094_promoted_to_fp16)[name = string("op_6095_cast_fp16")];
            tensor<int32, [1]> var_159_axes_0 = const()[name = string("var_159_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_159_keep_dims_0 = const()[name = string("var_159_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_159_cast_fp16 = reduce_mean(axes = var_159_axes_0, keep_dims = var_159_keep_dims_0, x = var_6095_cast_fp16)[name = string("var_159_cast_fp16")];
            fp16 var_6102_to_fp16 = const()[name = string("op_6102_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_6103_cast_fp16 = add(x = var_159_cast_fp16, y = var_6102_to_fp16)[name = string("op_6103_cast_fp16")];
            fp32 var_6104_epsilon_0 = const()[name = string("op_6104_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_6104_cast_fp16 = rsqrt(epsilon = var_6104_epsilon_0, x = var_6103_cast_fp16)[name = string("op_6104_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_159_cast_fp16 = mul(x = x_249_cast_fp16, y = var_6104_cast_fp16)[name = string("x_normed_159_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_6114_to_fp16 = const()[name = string("op_6114_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428010112)))];
            tensor<fp16, [1, 640, 1, ?]> input_69_cast_fp16 = mul(x = x_normed_159_cast_fp16, y = var_6114_to_fp16)[name = string("input_69_cast_fp16")];
            string gate_17_pad_type_0 = const()[name = string("gate_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_17_strides_0 = const()[name = string("gate_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_17_pad_0 = const()[name = string("gate_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_17_dilations_0 = const()[name = string("gate_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_17_groups_0 = const()[name = string("gate_17_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_8_weight_to_fp16 = const()[name = string("gate_proj_8_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428011456)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_17_cast_fp16 = conv(dilations = gate_17_dilations_0, groups = gate_17_groups_0, pad = gate_17_pad_0, pad_type = gate_17_pad_type_0, strides = gate_17_strides_0, weight = gate_proj_8_weight_to_fp16, x = input_69_cast_fp16)[name = string("gate_17_cast_fp16")];
            string up_17_pad_type_0 = const()[name = string("up_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_17_strides_0 = const()[name = string("up_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_17_pad_0 = const()[name = string("up_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_17_dilations_0 = const()[name = string("up_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_17_groups_0 = const()[name = string("up_17_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_8_weight_to_fp16 = const()[name = string("up_proj_8_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430632960)))];
            tensor<fp16, [1, 2048, 1, ?]> up_17_cast_fp16 = conv(dilations = up_17_dilations_0, groups = up_17_groups_0, pad = up_17_pad_0, pad_type = up_17_pad_type_0, strides = up_17_strides_0, weight = up_proj_8_weight_to_fp16, x = input_69_cast_fp16)[name = string("up_17_cast_fp16")];
            string var_6139_mode_0 = const()[name = string("op_6139_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_6139_cast_fp16 = gelu(mode = var_6139_mode_0, x = gate_17_cast_fp16)[name = string("op_6139_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_71_cast_fp16 = mul(x = var_6139_cast_fp16, y = up_17_cast_fp16)[name = string("input_71_cast_fp16")];
            string x_251_pad_type_0 = const()[name = string("x_251_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_251_strides_0 = const()[name = string("x_251_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_251_pad_0 = const()[name = string("x_251_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_251_dilations_0 = const()[name = string("x_251_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_251_groups_0 = const()[name = string("x_251_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_8_weight_to_fp16 = const()[name = string("down_proj_8_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433254464)))];
            tensor<fp16, [1, 640, 1, ?]> x_251_cast_fp16 = conv(dilations = x_251_dilations_0, groups = x_251_groups_0, pad = x_251_pad_0, pad_type = x_251_pad_type_0, strides = x_251_strides_0, weight = down_proj_8_weight_to_fp16, x = input_71_cast_fp16)[name = string("x_251_cast_fp16")];
            fp16 var_6152_promoted_to_fp16 = const()[name = string("op_6152_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_6153_cast_fp16 = pow(x = x_251_cast_fp16, y = var_6152_promoted_to_fp16)[name = string("op_6153_cast_fp16")];
            tensor<int32, [1]> var_161_axes_0 = const()[name = string("var_161_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_161_keep_dims_0 = const()[name = string("var_161_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_161_cast_fp16 = reduce_mean(axes = var_161_axes_0, keep_dims = var_161_keep_dims_0, x = var_6153_cast_fp16)[name = string("var_161_cast_fp16")];
            fp16 var_6160_to_fp16 = const()[name = string("op_6160_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_6161_cast_fp16 = add(x = var_161_cast_fp16, y = var_6160_to_fp16)[name = string("op_6161_cast_fp16")];
            fp32 var_6162_epsilon_0 = const()[name = string("op_6162_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_6162_cast_fp16 = rsqrt(epsilon = var_6162_epsilon_0, x = var_6161_cast_fp16)[name = string("op_6162_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_161_cast_fp16 = mul(x = x_251_cast_fp16, y = var_6162_cast_fp16)[name = string("x_normed_161_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_6172_to_fp16 = const()[name = string("op_6172_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435875968)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_17_cast_fp16 = mul(x = x_normed_161_cast_fp16, y = var_6172_to_fp16)[name = string("mlp_17_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_253_cast_fp16 = add(x = x_249_cast_fp16, y = mlp_17_cast_fp16)[name = string("x_253_cast_fp16")];
            fp16 var_6176_promoted_to_fp16 = const()[name = string("op_6176_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_6177_cast_fp16 = pow(x = x_253_cast_fp16, y = var_6176_promoted_to_fp16)[name = string("op_6177_cast_fp16")];
            tensor<int32, [1]> var_163_axes_0 = const()[name = string("var_163_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_163_keep_dims_0 = const()[name = string("var_163_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_163_cast_fp16 = reduce_mean(axes = var_163_axes_0, keep_dims = var_163_keep_dims_0, x = var_6177_cast_fp16)[name = string("var_163_cast_fp16")];
            fp16 var_6184_to_fp16 = const()[name = string("op_6184_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_6185_cast_fp16 = add(x = var_163_cast_fp16, y = var_6184_to_fp16)[name = string("op_6185_cast_fp16")];
            fp32 var_6186_epsilon_0 = const()[name = string("op_6186_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_6186_cast_fp16 = rsqrt(epsilon = var_6186_epsilon_0, x = var_6185_cast_fp16)[name = string("op_6186_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_163_cast_fp16 = mul(x = x_253_cast_fp16, y = var_6186_cast_fp16)[name = string("x_normed_163_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_6196_to_fp16 = const()[name = string("op_6196_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435877312)))];
            tensor<fp16, [1, 640, 1, ?]> input_73_cast_fp16 = mul(x = x_normed_163_cast_fp16, y = var_6196_to_fp16)[name = string("input_73_cast_fp16")];
            string var_6208_pad_type_0 = const()[name = string("op_6208_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6208_strides_0 = const()[name = string("op_6208_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6208_pad_0 = const()[name = string("op_6208_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6208_dilations_0 = const()[name = string("op_6208_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6208_groups_0 = const()[name = string("op_6208_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_9_weight_to_fp16 = const()[name = string("q_proj_9_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435878656)))];
            tensor<fp16, [1, 1024, 1, ?]> var_6208_cast_fp16 = conv(dilations = var_6208_dilations_0, groups = var_6208_groups_0, pad = var_6208_pad_0, pad_type = var_6208_pad_type_0, strides = var_6208_strides_0, weight = q_proj_9_weight_to_fp16, x = input_73_cast_fp16)[name = string("op_6208_cast_fp16")];
            string var_6219_pad_type_0 = const()[name = string("op_6219_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6219_strides_0 = const()[name = string("op_6219_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6219_pad_0 = const()[name = string("op_6219_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6219_dilations_0 = const()[name = string("op_6219_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6219_groups_0 = const()[name = string("op_6219_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_9_weight_to_fp16 = const()[name = string("k_proj_9_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437189440)))];
            tensor<fp16, [1, 256, 1, ?]> var_6219_cast_fp16 = conv(dilations = var_6219_dilations_0, groups = var_6219_groups_0, pad = var_6219_pad_0, pad_type = var_6219_pad_type_0, strides = var_6219_strides_0, weight = k_proj_9_weight_to_fp16, x = input_73_cast_fp16)[name = string("op_6219_cast_fp16")];
            string var_6230_pad_type_0 = const()[name = string("op_6230_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6230_strides_0 = const()[name = string("op_6230_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6230_pad_0 = const()[name = string("op_6230_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6230_dilations_0 = const()[name = string("op_6230_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6230_groups_0 = const()[name = string("op_6230_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_9_weight_to_fp16 = const()[name = string("v_proj_9_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437517184)))];
            tensor<fp16, [1, 256, 1, ?]> var_6230_cast_fp16 = conv(dilations = var_6230_dilations_0, groups = var_6230_groups_0, pad = var_6230_pad_0, pad_type = var_6230_pad_type_0, strides = var_6230_strides_0, weight = v_proj_9_weight_to_fp16, x = input_73_cast_fp16)[name = string("op_6230_cast_fp16")];
            tensor<int32, [4]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_6233_axis_0 = const()[name = string("op_6233_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_6233_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_6233_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_6233_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_6233_cast_fp16_3 = split(axis = var_6233_axis_0, split_sizes = tile_27, x = var_6208_cast_fp16)[name = string("op_6233_cast_fp16")];
            fp16 var_6246_promoted_to_fp16 = const()[name = string("op_6246_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_6247_cast_fp16 = pow(x = var_6233_cast_fp16_0, y = var_6246_promoted_to_fp16)[name = string("op_6247_cast_fp16")];
            tensor<int32, [1]> var_165_axes_0 = const()[name = string("var_165_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_165_keep_dims_0 = const()[name = string("var_165_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_165_cast_fp16 = reduce_mean(axes = var_165_axes_0, keep_dims = var_165_keep_dims_0, x = var_6247_cast_fp16)[name = string("var_165_cast_fp16")];
            fp16 var_6254_to_fp16 = const()[name = string("op_6254_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_6255_cast_fp16 = add(x = var_165_cast_fp16, y = var_6254_to_fp16)[name = string("op_6255_cast_fp16")];
            fp32 var_6256_epsilon_0 = const()[name = string("op_6256_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_6256_cast_fp16 = rsqrt(epsilon = var_6256_epsilon_0, x = var_6255_cast_fp16)[name = string("op_6256_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_165_cast_fp16 = mul(x = var_6233_cast_fp16_0, y = var_6256_cast_fp16)[name = string("x_normed_165_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_6266_to_fp16 = const()[name = string("op_6266_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437844928)))];
            tensor<fp16, [1, 256, 1, ?]> x_257_cast_fp16 = mul(x = x_normed_165_cast_fp16, y = var_6266_to_fp16)[name = string("x_257_cast_fp16")];
            tensor<int32, [4]> x_first_91_begin_0 = const()[name = string("x_first_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_91_end_0 = const()[name = string("x_first_91_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_91_end_mask_0 = const()[name = string("x_first_91_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_91_cast_fp16 = slice_by_index(begin = x_first_91_begin_0, end = x_first_91_end_0, end_mask = x_first_91_end_mask_0, x = x_257_cast_fp16)[name = string("x_first_91_cast_fp16")];
            tensor<int32, [4]> x_second_91_begin_0 = const()[name = string("x_second_91_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_91_end_0 = const()[name = string("x_second_91_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_91_end_mask_0 = const()[name = string("x_second_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_91_cast_fp16 = slice_by_index(begin = x_second_91_begin_0, end = x_second_91_end_0, end_mask = x_second_91_end_mask_0, x = x_257_cast_fp16)[name = string("x_second_91_cast_fp16")];
            fp16 const_76_promoted_to_fp16 = const()[name = string("const_76_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_6288_cast_fp16 = mul(x = x_second_91_cast_fp16, y = const_76_promoted_to_fp16)[name = string("op_6288_cast_fp16")];
            int32 var_6290 = const()[name = string("op_6290"), val = int32(1)];
            bool x_rot_91_interleave_0 = const()[name = string("x_rot_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_91_cast_fp16 = concat(axis = var_6290, interleave = x_rot_91_interleave_0, values = (var_6288_cast_fp16, x_first_91_cast_fp16))[name = string("x_rot_91_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_6306_cast_fp16 = mul(x = x_257_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_6306_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_6307_cast_fp16 = mul(x = x_rot_91_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_6307_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_73_cast_fp16 = add(x = var_6306_cast_fp16, y = var_6307_cast_fp16)[name = string("q_h_73_cast_fp16")];
            fp16 var_6310_promoted_to_fp16 = const()[name = string("op_6310_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_6311_cast_fp16 = pow(x = var_6233_cast_fp16_1, y = var_6310_promoted_to_fp16)[name = string("op_6311_cast_fp16")];
            tensor<int32, [1]> var_167_axes_0 = const()[name = string("var_167_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_167_keep_dims_0 = const()[name = string("var_167_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_167_cast_fp16 = reduce_mean(axes = var_167_axes_0, keep_dims = var_167_keep_dims_0, x = var_6311_cast_fp16)[name = string("var_167_cast_fp16")];
            fp16 var_6318_to_fp16 = const()[name = string("op_6318_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_6319_cast_fp16 = add(x = var_167_cast_fp16, y = var_6318_to_fp16)[name = string("op_6319_cast_fp16")];
            fp32 var_6320_epsilon_0 = const()[name = string("op_6320_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_6320_cast_fp16 = rsqrt(epsilon = var_6320_epsilon_0, x = var_6319_cast_fp16)[name = string("op_6320_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_167_cast_fp16 = mul(x = var_6233_cast_fp16_1, y = var_6320_cast_fp16)[name = string("x_normed_167_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_261_cast_fp16 = mul(x = x_normed_167_cast_fp16, y = var_6266_to_fp16)[name = string("x_261_cast_fp16")];
            tensor<int32, [4]> x_first_93_begin_0 = const()[name = string("x_first_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_93_end_0 = const()[name = string("x_first_93_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_93_end_mask_0 = const()[name = string("x_first_93_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_93_cast_fp16 = slice_by_index(begin = x_first_93_begin_0, end = x_first_93_end_0, end_mask = x_first_93_end_mask_0, x = x_261_cast_fp16)[name = string("x_first_93_cast_fp16")];
            tensor<int32, [4]> x_second_93_begin_0 = const()[name = string("x_second_93_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_93_end_0 = const()[name = string("x_second_93_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_93_end_mask_0 = const()[name = string("x_second_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_93_cast_fp16 = slice_by_index(begin = x_second_93_begin_0, end = x_second_93_end_0, end_mask = x_second_93_end_mask_0, x = x_261_cast_fp16)[name = string("x_second_93_cast_fp16")];
            fp16 const_77_promoted_to_fp16 = const()[name = string("const_77_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_6352_cast_fp16 = mul(x = x_second_93_cast_fp16, y = const_77_promoted_to_fp16)[name = string("op_6352_cast_fp16")];
            int32 var_6354 = const()[name = string("op_6354"), val = int32(1)];
            bool x_rot_93_interleave_0 = const()[name = string("x_rot_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_93_cast_fp16 = concat(axis = var_6354, interleave = x_rot_93_interleave_0, values = (var_6352_cast_fp16, x_first_93_cast_fp16))[name = string("x_rot_93_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_6370_cast_fp16 = mul(x = x_261_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_6370_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_6371_cast_fp16 = mul(x = x_rot_93_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_6371_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_75_cast_fp16 = add(x = var_6370_cast_fp16, y = var_6371_cast_fp16)[name = string("q_h_75_cast_fp16")];
            fp16 var_6374_promoted_to_fp16 = const()[name = string("op_6374_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_6375_cast_fp16 = pow(x = var_6233_cast_fp16_2, y = var_6374_promoted_to_fp16)[name = string("op_6375_cast_fp16")];
            tensor<int32, [1]> var_169_axes_0 = const()[name = string("var_169_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_169_keep_dims_0 = const()[name = string("var_169_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_169_cast_fp16 = reduce_mean(axes = var_169_axes_0, keep_dims = var_169_keep_dims_0, x = var_6375_cast_fp16)[name = string("var_169_cast_fp16")];
            fp16 var_6382_to_fp16 = const()[name = string("op_6382_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_6383_cast_fp16 = add(x = var_169_cast_fp16, y = var_6382_to_fp16)[name = string("op_6383_cast_fp16")];
            fp32 var_6384_epsilon_0 = const()[name = string("op_6384_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_6384_cast_fp16 = rsqrt(epsilon = var_6384_epsilon_0, x = var_6383_cast_fp16)[name = string("op_6384_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_169_cast_fp16 = mul(x = var_6233_cast_fp16_2, y = var_6384_cast_fp16)[name = string("x_normed_169_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_265_cast_fp16 = mul(x = x_normed_169_cast_fp16, y = var_6266_to_fp16)[name = string("x_265_cast_fp16")];
            tensor<int32, [4]> x_first_95_begin_0 = const()[name = string("x_first_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_95_end_0 = const()[name = string("x_first_95_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_95_end_mask_0 = const()[name = string("x_first_95_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_95_cast_fp16 = slice_by_index(begin = x_first_95_begin_0, end = x_first_95_end_0, end_mask = x_first_95_end_mask_0, x = x_265_cast_fp16)[name = string("x_first_95_cast_fp16")];
            tensor<int32, [4]> x_second_95_begin_0 = const()[name = string("x_second_95_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_95_end_0 = const()[name = string("x_second_95_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_95_end_mask_0 = const()[name = string("x_second_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_95_cast_fp16 = slice_by_index(begin = x_second_95_begin_0, end = x_second_95_end_0, end_mask = x_second_95_end_mask_0, x = x_265_cast_fp16)[name = string("x_second_95_cast_fp16")];
            fp16 const_78_promoted_to_fp16 = const()[name = string("const_78_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_6416_cast_fp16 = mul(x = x_second_95_cast_fp16, y = const_78_promoted_to_fp16)[name = string("op_6416_cast_fp16")];
            int32 var_6418 = const()[name = string("op_6418"), val = int32(1)];
            bool x_rot_95_interleave_0 = const()[name = string("x_rot_95_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_95_cast_fp16 = concat(axis = var_6418, interleave = x_rot_95_interleave_0, values = (var_6416_cast_fp16, x_first_95_cast_fp16))[name = string("x_rot_95_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_6434_cast_fp16 = mul(x = x_265_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_6434_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_6435_cast_fp16 = mul(x = x_rot_95_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_6435_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_77_cast_fp16 = add(x = var_6434_cast_fp16, y = var_6435_cast_fp16)[name = string("q_h_77_cast_fp16")];
            fp16 var_6438_promoted_to_fp16 = const()[name = string("op_6438_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_6439_cast_fp16 = pow(x = var_6233_cast_fp16_3, y = var_6438_promoted_to_fp16)[name = string("op_6439_cast_fp16")];
            tensor<int32, [1]> var_171_axes_0 = const()[name = string("var_171_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_171_keep_dims_0 = const()[name = string("var_171_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_171_cast_fp16 = reduce_mean(axes = var_171_axes_0, keep_dims = var_171_keep_dims_0, x = var_6439_cast_fp16)[name = string("var_171_cast_fp16")];
            fp16 var_6446_to_fp16 = const()[name = string("op_6446_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_6447_cast_fp16 = add(x = var_171_cast_fp16, y = var_6446_to_fp16)[name = string("op_6447_cast_fp16")];
            fp32 var_6448_epsilon_0 = const()[name = string("op_6448_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_6448_cast_fp16 = rsqrt(epsilon = var_6448_epsilon_0, x = var_6447_cast_fp16)[name = string("op_6448_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_171_cast_fp16 = mul(x = var_6233_cast_fp16_3, y = var_6448_cast_fp16)[name = string("x_normed_171_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_269_cast_fp16 = mul(x = x_normed_171_cast_fp16, y = var_6266_to_fp16)[name = string("x_269_cast_fp16")];
            tensor<int32, [4]> x_first_97_begin_0 = const()[name = string("x_first_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_97_end_0 = const()[name = string("x_first_97_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_97_end_mask_0 = const()[name = string("x_first_97_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_97_cast_fp16 = slice_by_index(begin = x_first_97_begin_0, end = x_first_97_end_0, end_mask = x_first_97_end_mask_0, x = x_269_cast_fp16)[name = string("x_first_97_cast_fp16")];
            tensor<int32, [4]> x_second_97_begin_0 = const()[name = string("x_second_97_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_97_end_0 = const()[name = string("x_second_97_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_97_end_mask_0 = const()[name = string("x_second_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_97_cast_fp16 = slice_by_index(begin = x_second_97_begin_0, end = x_second_97_end_0, end_mask = x_second_97_end_mask_0, x = x_269_cast_fp16)[name = string("x_second_97_cast_fp16")];
            fp16 const_79_promoted_to_fp16 = const()[name = string("const_79_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_6480_cast_fp16 = mul(x = x_second_97_cast_fp16, y = const_79_promoted_to_fp16)[name = string("op_6480_cast_fp16")];
            int32 var_6482 = const()[name = string("op_6482"), val = int32(1)];
            bool x_rot_97_interleave_0 = const()[name = string("x_rot_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_97_cast_fp16 = concat(axis = var_6482, interleave = x_rot_97_interleave_0, values = (var_6480_cast_fp16, x_first_97_cast_fp16))[name = string("x_rot_97_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_6498_cast_fp16 = mul(x = x_269_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_6498_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_6499_cast_fp16 = mul(x = x_rot_97_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_6499_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_79_cast_fp16 = add(x = var_6498_cast_fp16, y = var_6499_cast_fp16)[name = string("q_h_79_cast_fp16")];
            fp16 var_6502_promoted_to_fp16 = const()[name = string("op_6502_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_6503_cast_fp16 = pow(x = var_6219_cast_fp16, y = var_6502_promoted_to_fp16)[name = string("op_6503_cast_fp16")];
            tensor<int32, [1]> var_173_axes_0 = const()[name = string("var_173_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_173_keep_dims_0 = const()[name = string("var_173_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_173_cast_fp16 = reduce_mean(axes = var_173_axes_0, keep_dims = var_173_keep_dims_0, x = var_6503_cast_fp16)[name = string("var_173_cast_fp16")];
            fp16 var_6510_to_fp16 = const()[name = string("op_6510_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_6511_cast_fp16 = add(x = var_173_cast_fp16, y = var_6510_to_fp16)[name = string("op_6511_cast_fp16")];
            fp32 var_6512_epsilon_0 = const()[name = string("op_6512_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_6512_cast_fp16 = rsqrt(epsilon = var_6512_epsilon_0, x = var_6511_cast_fp16)[name = string("op_6512_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_173_cast_fp16 = mul(x = var_6219_cast_fp16, y = var_6512_cast_fp16)[name = string("x_normed_173_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_6522_to_fp16 = const()[name = string("op_6522_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437845504)))];
            tensor<fp16, [1, 256, 1, ?]> x_273_cast_fp16 = mul(x = x_normed_173_cast_fp16, y = var_6522_to_fp16)[name = string("x_273_cast_fp16")];
            tensor<int32, [4]> x_first_99_begin_0 = const()[name = string("x_first_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_99_end_0 = const()[name = string("x_first_99_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_99_end_mask_0 = const()[name = string("x_first_99_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_99_cast_fp16 = slice_by_index(begin = x_first_99_begin_0, end = x_first_99_end_0, end_mask = x_first_99_end_mask_0, x = x_273_cast_fp16)[name = string("x_first_99_cast_fp16")];
            tensor<int32, [4]> x_second_99_begin_0 = const()[name = string("x_second_99_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_99_end_0 = const()[name = string("x_second_99_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_99_end_mask_0 = const()[name = string("x_second_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_99_cast_fp16 = slice_by_index(begin = x_second_99_begin_0, end = x_second_99_end_0, end_mask = x_second_99_end_mask_0, x = x_273_cast_fp16)[name = string("x_second_99_cast_fp16")];
            fp16 const_80_promoted_to_fp16 = const()[name = string("const_80_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_6544_cast_fp16 = mul(x = x_second_99_cast_fp16, y = const_80_promoted_to_fp16)[name = string("op_6544_cast_fp16")];
            int32 var_6546 = const()[name = string("op_6546"), val = int32(1)];
            bool x_rot_99_interleave_0 = const()[name = string("x_rot_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_99_cast_fp16 = concat(axis = var_6546, interleave = x_rot_99_interleave_0, values = (var_6544_cast_fp16, x_first_99_cast_fp16))[name = string("x_rot_99_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_6562_cast_fp16 = mul(x = x_273_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_6562_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_6563_cast_fp16 = mul(x = x_rot_99_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_6563_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_19_cast_fp16 = add(x = var_6562_cast_fp16, y = var_6563_cast_fp16)[name = string("k_new_19_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_18 = read_state(input = k_cache_9)[name = string("read_state_18")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_57_cast_fp16 = mul(x = read_state_18, y = one_minus_1_cast_fp16)[name = string("k_cache_57_cast_fp16")];
            write_state(data = k_cache_57_cast_fp16, input = k_cache_9)[name = string("coreml_update_state_108_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_108 = read_state(input = k_cache_9)[name = string("coreml_update_state_108")];
            tensor<fp16, [1, 256, 1, 128]> var_6570_cast_fp16 = mul(x = k_new_19_cast_fp16, y = write_mask)[name = string("op_6570_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_59_cast_fp16 = add(x = coreml_update_state_108, y = var_6570_cast_fp16)[name = string("k_cache_59_cast_fp16")];
            write_state(data = k_cache_59_cast_fp16, input = k_cache_9)[name = string("coreml_update_state_109_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_109 = read_state(input = k_cache_9)[name = string("coreml_update_state_109")];
            tensor<fp16, [1, 256, 1, 128]> read_state_19 = read_state(input = v_cache_9)[name = string("read_state_19")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_57_cast_fp16 = mul(x = read_state_19, y = one_minus_1_cast_fp16)[name = string("v_cache_57_cast_fp16")];
            write_state(data = v_cache_57_cast_fp16, input = v_cache_9)[name = string("coreml_update_state_110_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_110 = read_state(input = v_cache_9)[name = string("coreml_update_state_110")];
            tensor<fp16, [1, 256, 1, 128]> var_6574_cast_fp16 = mul(x = var_6230_cast_fp16, y = write_mask)[name = string("op_6574_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_59_cast_fp16 = add(x = coreml_update_state_110, y = var_6574_cast_fp16)[name = string("v_cache_59_cast_fp16")];
            write_state(data = v_cache_59_cast_fp16, input = v_cache_9)[name = string("coreml_update_state_111_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_111 = read_state(input = v_cache_9)[name = string("coreml_update_state_111")];
            bool var_6589_transpose_x_0 = const()[name = string("op_6589_transpose_x_0"), val = bool(false)];
            bool var_6589_transpose_y_0 = const()[name = string("op_6589_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_378_perm_0 = const()[name = string("transpose_378_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_379_perm_0 = const()[name = string("transpose_379_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_379 = transpose(perm = transpose_379_perm_0, x = coreml_update_state_109)[name = string("transpose_511")];
            tensor<fp16, [1, 1, ?, 256]> transpose_378 = transpose(perm = transpose_378_perm_0, x = q_h_73_cast_fp16)[name = string("transpose_512")];
            tensor<fp16, [1, 1, ?, 128]> var_6589_cast_fp16 = matmul(transpose_x = var_6589_transpose_x_0, transpose_y = var_6589_transpose_y_0, x = transpose_378, y = transpose_379)[name = string("op_6589_cast_fp16")];
            fp16 var_6590_to_fp16 = const()[name = string("op_6590_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_6591_cast_fp16 = mul(x = var_6589_cast_fp16, y = var_6590_to_fp16)[name = string("op_6591_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_73_cast_fp16 = add(x = var_6591_cast_fp16, y = attention_mask)[name = string("scores_73_cast_fp16")];
            int32 var_6594 = const()[name = string("op_6594"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_73_cast_fp16 = softmax(axis = var_6594, x = scores_73_cast_fp16)[name = string("probs_73_cast_fp16")];
            tensor<int32, [4]> var_6601 = const()[name = string("op_6601"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_6603_transpose_x_0 = const()[name = string("op_6603_transpose_x_0"), val = bool(false)];
            bool var_6603_transpose_y_0 = const()[name = string("op_6603_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_73_cast_fp16 = transpose(perm = var_6601, x = coreml_update_state_111)[name = string("transpose_510")];
            tensor<fp16, [1, 1, ?, 256]> var_6603_cast_fp16 = matmul(transpose_x = var_6603_transpose_x_0, transpose_y = var_6603_transpose_y_0, x = probs_73_cast_fp16, y = v_t_73_cast_fp16)[name = string("op_6603_cast_fp16")];
            bool var_6622_transpose_x_0 = const()[name = string("op_6622_transpose_x_0"), val = bool(false)];
            bool var_6622_transpose_y_0 = const()[name = string("op_6622_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_383_perm_0 = const()[name = string("transpose_383_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_383 = transpose(perm = transpose_383_perm_0, x = q_h_75_cast_fp16)[name = string("transpose_509")];
            tensor<fp16, [1, 1, ?, 128]> var_6622_cast_fp16 = matmul(transpose_x = var_6622_transpose_x_0, transpose_y = var_6622_transpose_y_0, x = transpose_383, y = transpose_379)[name = string("op_6622_cast_fp16")];
            fp16 var_6623_to_fp16 = const()[name = string("op_6623_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_6624_cast_fp16 = mul(x = var_6622_cast_fp16, y = var_6623_to_fp16)[name = string("op_6624_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_75_cast_fp16 = add(x = var_6624_cast_fp16, y = attention_mask)[name = string("scores_75_cast_fp16")];
            int32 var_6627 = const()[name = string("op_6627"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_75_cast_fp16 = softmax(axis = var_6627, x = scores_75_cast_fp16)[name = string("probs_75_cast_fp16")];
            bool var_6636_transpose_x_0 = const()[name = string("op_6636_transpose_x_0"), val = bool(false)];
            bool var_6636_transpose_y_0 = const()[name = string("op_6636_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_6636_cast_fp16 = matmul(transpose_x = var_6636_transpose_x_0, transpose_y = var_6636_transpose_y_0, x = probs_75_cast_fp16, y = v_t_73_cast_fp16)[name = string("op_6636_cast_fp16")];
            bool var_6655_transpose_x_0 = const()[name = string("op_6655_transpose_x_0"), val = bool(false)];
            bool var_6655_transpose_y_0 = const()[name = string("op_6655_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_384_perm_0 = const()[name = string("transpose_384_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_384 = transpose(perm = transpose_384_perm_0, x = q_h_77_cast_fp16)[name = string("transpose_508")];
            tensor<fp16, [1, 1, ?, 128]> var_6655_cast_fp16 = matmul(transpose_x = var_6655_transpose_x_0, transpose_y = var_6655_transpose_y_0, x = transpose_384, y = transpose_379)[name = string("op_6655_cast_fp16")];
            fp16 var_6656_to_fp16 = const()[name = string("op_6656_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_6657_cast_fp16 = mul(x = var_6655_cast_fp16, y = var_6656_to_fp16)[name = string("op_6657_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_77_cast_fp16 = add(x = var_6657_cast_fp16, y = attention_mask)[name = string("scores_77_cast_fp16")];
            int32 var_6660 = const()[name = string("op_6660"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_77_cast_fp16 = softmax(axis = var_6660, x = scores_77_cast_fp16)[name = string("probs_77_cast_fp16")];
            bool var_6669_transpose_x_0 = const()[name = string("op_6669_transpose_x_0"), val = bool(false)];
            bool var_6669_transpose_y_0 = const()[name = string("op_6669_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_6669_cast_fp16 = matmul(transpose_x = var_6669_transpose_x_0, transpose_y = var_6669_transpose_y_0, x = probs_77_cast_fp16, y = v_t_73_cast_fp16)[name = string("op_6669_cast_fp16")];
            bool var_6688_transpose_x_0 = const()[name = string("op_6688_transpose_x_0"), val = bool(false)];
            bool var_6688_transpose_y_0 = const()[name = string("op_6688_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_385_perm_0 = const()[name = string("transpose_385_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_385 = transpose(perm = transpose_385_perm_0, x = q_h_79_cast_fp16)[name = string("transpose_507")];
            tensor<fp16, [1, 1, ?, 128]> var_6688_cast_fp16 = matmul(transpose_x = var_6688_transpose_x_0, transpose_y = var_6688_transpose_y_0, x = transpose_385, y = transpose_379)[name = string("op_6688_cast_fp16")];
            fp16 var_6689_to_fp16 = const()[name = string("op_6689_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_6690_cast_fp16 = mul(x = var_6688_cast_fp16, y = var_6689_to_fp16)[name = string("op_6690_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_79_cast_fp16 = add(x = var_6690_cast_fp16, y = attention_mask)[name = string("scores_79_cast_fp16")];
            int32 var_6693 = const()[name = string("op_6693"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_79_cast_fp16 = softmax(axis = var_6693, x = scores_79_cast_fp16)[name = string("probs_79_cast_fp16")];
            bool var_6702_transpose_x_0 = const()[name = string("op_6702_transpose_x_0"), val = bool(false)];
            bool var_6702_transpose_y_0 = const()[name = string("op_6702_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_6702_cast_fp16 = matmul(transpose_x = var_6702_transpose_x_0, transpose_y = var_6702_transpose_y_0, x = probs_79_cast_fp16, y = v_t_73_cast_fp16)[name = string("op_6702_cast_fp16")];
            bool input_75_interleave_0 = const()[name = string("input_75_interleave_0"), val = bool(false)];
            int32 const_154 = const()[name = string("const_154"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_75_cast_fp16 = concat(axis = const_154, interleave = input_75_interleave_0, values = (var_6603_cast_fp16, var_6636_cast_fp16, var_6669_cast_fp16, var_6702_cast_fp16))[name = string("input_75_cast_fp16")];
            string x_275_pad_type_0 = const()[name = string("x_275_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_275_strides_0 = const()[name = string("x_275_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_275_pad_0 = const()[name = string("x_275_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_275_dilations_0 = const()[name = string("x_275_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_275_groups_0 = const()[name = string("x_275_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_89_perm_0 = const()[name = string("transpose_89_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_9_weight_to_fp16 = const()[name = string("o_proj_9_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437846080)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_89_cast_fp16 = transpose(perm = transpose_89_perm_0, x = input_75_cast_fp16)[name = string("transpose_506")];
            tensor<fp16, [1, 640, 1, ?]> x_275_cast_fp16 = conv(dilations = x_275_dilations_0, groups = x_275_groups_0, pad = x_275_pad_0, pad_type = x_275_pad_type_0, strides = x_275_strides_0, weight = o_proj_9_weight_to_fp16, x = transpose_89_cast_fp16)[name = string("x_275_cast_fp16")];
            fp16 var_6723_promoted_to_fp16 = const()[name = string("op_6723_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_6724_cast_fp16 = pow(x = x_275_cast_fp16, y = var_6723_promoted_to_fp16)[name = string("op_6724_cast_fp16")];
            tensor<int32, [1]> var_175_axes_0 = const()[name = string("var_175_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_175_keep_dims_0 = const()[name = string("var_175_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_175_cast_fp16 = reduce_mean(axes = var_175_axes_0, keep_dims = var_175_keep_dims_0, x = var_6724_cast_fp16)[name = string("var_175_cast_fp16")];
            fp16 var_6731_to_fp16 = const()[name = string("op_6731_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_6732_cast_fp16 = add(x = var_175_cast_fp16, y = var_6731_to_fp16)[name = string("op_6732_cast_fp16")];
            fp32 var_6733_epsilon_0 = const()[name = string("op_6733_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_6733_cast_fp16 = rsqrt(epsilon = var_6733_epsilon_0, x = var_6732_cast_fp16)[name = string("op_6733_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_175_cast_fp16 = mul(x = x_275_cast_fp16, y = var_6733_cast_fp16)[name = string("x_normed_175_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_6743_to_fp16 = const()[name = string("op_6743_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439156864)))];
            tensor<fp16, [1, 640, 1, ?]> attn_39_cast_fp16 = mul(x = x_normed_175_cast_fp16, y = var_6743_to_fp16)[name = string("attn_39_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_277_cast_fp16 = add(x = x_253_cast_fp16, y = attn_39_cast_fp16)[name = string("x_277_cast_fp16")];
            fp16 var_6747_promoted_to_fp16 = const()[name = string("op_6747_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_6748_cast_fp16 = pow(x = x_277_cast_fp16, y = var_6747_promoted_to_fp16)[name = string("op_6748_cast_fp16")];
            tensor<int32, [1]> var_177_axes_0 = const()[name = string("var_177_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_177_keep_dims_0 = const()[name = string("var_177_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_177_cast_fp16 = reduce_mean(axes = var_177_axes_0, keep_dims = var_177_keep_dims_0, x = var_6748_cast_fp16)[name = string("var_177_cast_fp16")];
            fp16 var_6755_to_fp16 = const()[name = string("op_6755_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_6756_cast_fp16 = add(x = var_177_cast_fp16, y = var_6755_to_fp16)[name = string("op_6756_cast_fp16")];
            fp32 var_6757_epsilon_0 = const()[name = string("op_6757_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_6757_cast_fp16 = rsqrt(epsilon = var_6757_epsilon_0, x = var_6756_cast_fp16)[name = string("op_6757_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_177_cast_fp16 = mul(x = x_277_cast_fp16, y = var_6757_cast_fp16)[name = string("x_normed_177_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_6767_to_fp16 = const()[name = string("op_6767_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439158208)))];
            tensor<fp16, [1, 640, 1, ?]> input_77_cast_fp16 = mul(x = x_normed_177_cast_fp16, y = var_6767_to_fp16)[name = string("input_77_cast_fp16")];
            string gate_19_pad_type_0 = const()[name = string("gate_19_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_19_strides_0 = const()[name = string("gate_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_19_pad_0 = const()[name = string("gate_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_19_dilations_0 = const()[name = string("gate_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_19_groups_0 = const()[name = string("gate_19_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_9_weight_to_fp16 = const()[name = string("gate_proj_9_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439159552)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_19_cast_fp16 = conv(dilations = gate_19_dilations_0, groups = gate_19_groups_0, pad = gate_19_pad_0, pad_type = gate_19_pad_type_0, strides = gate_19_strides_0, weight = gate_proj_9_weight_to_fp16, x = input_77_cast_fp16)[name = string("gate_19_cast_fp16")];
            string up_19_pad_type_0 = const()[name = string("up_19_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_19_strides_0 = const()[name = string("up_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_19_pad_0 = const()[name = string("up_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_19_dilations_0 = const()[name = string("up_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_19_groups_0 = const()[name = string("up_19_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_9_weight_to_fp16 = const()[name = string("up_proj_9_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441781056)))];
            tensor<fp16, [1, 2048, 1, ?]> up_19_cast_fp16 = conv(dilations = up_19_dilations_0, groups = up_19_groups_0, pad = up_19_pad_0, pad_type = up_19_pad_type_0, strides = up_19_strides_0, weight = up_proj_9_weight_to_fp16, x = input_77_cast_fp16)[name = string("up_19_cast_fp16")];
            string var_6792_mode_0 = const()[name = string("op_6792_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_6792_cast_fp16 = gelu(mode = var_6792_mode_0, x = gate_19_cast_fp16)[name = string("op_6792_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_79_cast_fp16 = mul(x = var_6792_cast_fp16, y = up_19_cast_fp16)[name = string("input_79_cast_fp16")];
            string x_279_pad_type_0 = const()[name = string("x_279_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_279_strides_0 = const()[name = string("x_279_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_279_pad_0 = const()[name = string("x_279_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_279_dilations_0 = const()[name = string("x_279_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_279_groups_0 = const()[name = string("x_279_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_9_weight_to_fp16 = const()[name = string("down_proj_9_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444402560)))];
            tensor<fp16, [1, 640, 1, ?]> x_279_cast_fp16 = conv(dilations = x_279_dilations_0, groups = x_279_groups_0, pad = x_279_pad_0, pad_type = x_279_pad_type_0, strides = x_279_strides_0, weight = down_proj_9_weight_to_fp16, x = input_79_cast_fp16)[name = string("x_279_cast_fp16")];
            fp16 var_6805_promoted_to_fp16 = const()[name = string("op_6805_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_6806_cast_fp16 = pow(x = x_279_cast_fp16, y = var_6805_promoted_to_fp16)[name = string("op_6806_cast_fp16")];
            tensor<int32, [1]> var_179_axes_0 = const()[name = string("var_179_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_179_keep_dims_0 = const()[name = string("var_179_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_179_cast_fp16 = reduce_mean(axes = var_179_axes_0, keep_dims = var_179_keep_dims_0, x = var_6806_cast_fp16)[name = string("var_179_cast_fp16")];
            fp16 var_6813_to_fp16 = const()[name = string("op_6813_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_6814_cast_fp16 = add(x = var_179_cast_fp16, y = var_6813_to_fp16)[name = string("op_6814_cast_fp16")];
            fp32 var_6815_epsilon_0 = const()[name = string("op_6815_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_6815_cast_fp16 = rsqrt(epsilon = var_6815_epsilon_0, x = var_6814_cast_fp16)[name = string("op_6815_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_179_cast_fp16 = mul(x = x_279_cast_fp16, y = var_6815_cast_fp16)[name = string("x_normed_179_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_6825_to_fp16 = const()[name = string("op_6825_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447024064)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_19_cast_fp16 = mul(x = x_normed_179_cast_fp16, y = var_6825_to_fp16)[name = string("mlp_19_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_281_cast_fp16 = add(x = x_277_cast_fp16, y = mlp_19_cast_fp16)[name = string("x_281_cast_fp16")];
            fp16 var_6829_promoted_to_fp16 = const()[name = string("op_6829_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_6830_cast_fp16 = pow(x = x_281_cast_fp16, y = var_6829_promoted_to_fp16)[name = string("op_6830_cast_fp16")];
            tensor<int32, [1]> var_181_axes_0 = const()[name = string("var_181_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_181_keep_dims_0 = const()[name = string("var_181_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_181_cast_fp16 = reduce_mean(axes = var_181_axes_0, keep_dims = var_181_keep_dims_0, x = var_6830_cast_fp16)[name = string("var_181_cast_fp16")];
            fp16 var_6837_to_fp16 = const()[name = string("op_6837_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_6838_cast_fp16 = add(x = var_181_cast_fp16, y = var_6837_to_fp16)[name = string("op_6838_cast_fp16")];
            fp32 var_6839_epsilon_0 = const()[name = string("op_6839_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_6839_cast_fp16 = rsqrt(epsilon = var_6839_epsilon_0, x = var_6838_cast_fp16)[name = string("op_6839_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_181_cast_fp16 = mul(x = x_281_cast_fp16, y = var_6839_cast_fp16)[name = string("x_normed_181_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_6849_to_fp16 = const()[name = string("op_6849_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447025408)))];
            tensor<fp16, [1, 640, 1, ?]> input_81_cast_fp16 = mul(x = x_normed_181_cast_fp16, y = var_6849_to_fp16)[name = string("input_81_cast_fp16")];
            string var_6861_pad_type_0 = const()[name = string("op_6861_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6861_strides_0 = const()[name = string("op_6861_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6861_pad_0 = const()[name = string("op_6861_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6861_dilations_0 = const()[name = string("op_6861_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6861_groups_0 = const()[name = string("op_6861_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_10_weight_to_fp16 = const()[name = string("q_proj_10_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447026752)))];
            tensor<fp16, [1, 1024, 1, ?]> var_6861_cast_fp16 = conv(dilations = var_6861_dilations_0, groups = var_6861_groups_0, pad = var_6861_pad_0, pad_type = var_6861_pad_type_0, strides = var_6861_strides_0, weight = q_proj_10_weight_to_fp16, x = input_81_cast_fp16)[name = string("op_6861_cast_fp16")];
            string var_6872_pad_type_0 = const()[name = string("op_6872_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6872_strides_0 = const()[name = string("op_6872_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6872_pad_0 = const()[name = string("op_6872_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6872_dilations_0 = const()[name = string("op_6872_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6872_groups_0 = const()[name = string("op_6872_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_10_weight_to_fp16 = const()[name = string("k_proj_10_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448337536)))];
            tensor<fp16, [1, 256, 1, ?]> var_6872_cast_fp16 = conv(dilations = var_6872_dilations_0, groups = var_6872_groups_0, pad = var_6872_pad_0, pad_type = var_6872_pad_type_0, strides = var_6872_strides_0, weight = k_proj_10_weight_to_fp16, x = input_81_cast_fp16)[name = string("op_6872_cast_fp16")];
            string var_6883_pad_type_0 = const()[name = string("op_6883_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6883_strides_0 = const()[name = string("op_6883_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6883_pad_0 = const()[name = string("op_6883_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6883_dilations_0 = const()[name = string("op_6883_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6883_groups_0 = const()[name = string("op_6883_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_10_weight_to_fp16 = const()[name = string("v_proj_10_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448665280)))];
            tensor<fp16, [1, 256, 1, ?]> var_6883_cast_fp16 = conv(dilations = var_6883_dilations_0, groups = var_6883_groups_0, pad = var_6883_pad_0, pad_type = var_6883_pad_type_0, strides = var_6883_strides_0, weight = v_proj_10_weight_to_fp16, x = input_81_cast_fp16)[name = string("op_6883_cast_fp16")];
            tensor<int32, [4]> tile_30 = const()[name = string("tile_30"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_6886_axis_0 = const()[name = string("op_6886_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_6886_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_6886_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_6886_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_6886_cast_fp16_3 = split(axis = var_6886_axis_0, split_sizes = tile_30, x = var_6861_cast_fp16)[name = string("op_6886_cast_fp16")];
            fp16 var_6899_promoted_to_fp16 = const()[name = string("op_6899_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_6900_cast_fp16 = pow(x = var_6886_cast_fp16_0, y = var_6899_promoted_to_fp16)[name = string("op_6900_cast_fp16")];
            tensor<int32, [1]> var_183_axes_0 = const()[name = string("var_183_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_183_keep_dims_0 = const()[name = string("var_183_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_183_cast_fp16 = reduce_mean(axes = var_183_axes_0, keep_dims = var_183_keep_dims_0, x = var_6900_cast_fp16)[name = string("var_183_cast_fp16")];
            fp16 var_6907_to_fp16 = const()[name = string("op_6907_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_6908_cast_fp16 = add(x = var_183_cast_fp16, y = var_6907_to_fp16)[name = string("op_6908_cast_fp16")];
            fp32 var_6909_epsilon_0 = const()[name = string("op_6909_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_6909_cast_fp16 = rsqrt(epsilon = var_6909_epsilon_0, x = var_6908_cast_fp16)[name = string("op_6909_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_183_cast_fp16 = mul(x = var_6886_cast_fp16_0, y = var_6909_cast_fp16)[name = string("x_normed_183_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_6919_to_fp16 = const()[name = string("op_6919_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448993024)))];
            tensor<fp16, [1, 256, 1, ?]> x_285_cast_fp16 = mul(x = x_normed_183_cast_fp16, y = var_6919_to_fp16)[name = string("x_285_cast_fp16")];
            tensor<int32, [4]> x_first_101_begin_0 = const()[name = string("x_first_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_101_end_0 = const()[name = string("x_first_101_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_101_end_mask_0 = const()[name = string("x_first_101_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_101_cast_fp16 = slice_by_index(begin = x_first_101_begin_0, end = x_first_101_end_0, end_mask = x_first_101_end_mask_0, x = x_285_cast_fp16)[name = string("x_first_101_cast_fp16")];
            tensor<int32, [4]> x_second_101_begin_0 = const()[name = string("x_second_101_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_101_end_0 = const()[name = string("x_second_101_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_101_end_mask_0 = const()[name = string("x_second_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_101_cast_fp16 = slice_by_index(begin = x_second_101_begin_0, end = x_second_101_end_0, end_mask = x_second_101_end_mask_0, x = x_285_cast_fp16)[name = string("x_second_101_cast_fp16")];
            fp16 const_84_promoted_to_fp16 = const()[name = string("const_84_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_6941_cast_fp16 = mul(x = x_second_101_cast_fp16, y = const_84_promoted_to_fp16)[name = string("op_6941_cast_fp16")];
            int32 var_6943 = const()[name = string("op_6943"), val = int32(1)];
            bool x_rot_101_interleave_0 = const()[name = string("x_rot_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_101_cast_fp16 = concat(axis = var_6943, interleave = x_rot_101_interleave_0, values = (var_6941_cast_fp16, x_first_101_cast_fp16))[name = string("x_rot_101_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_6959_cast_fp16 = mul(x = x_285_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_6959_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_6960_cast_fp16 = mul(x = x_rot_101_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_6960_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_81_cast_fp16 = add(x = var_6959_cast_fp16, y = var_6960_cast_fp16)[name = string("q_h_81_cast_fp16")];
            fp16 var_6963_promoted_to_fp16 = const()[name = string("op_6963_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_6964_cast_fp16 = pow(x = var_6886_cast_fp16_1, y = var_6963_promoted_to_fp16)[name = string("op_6964_cast_fp16")];
            tensor<int32, [1]> var_185_axes_0 = const()[name = string("var_185_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_185_keep_dims_0 = const()[name = string("var_185_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_185_cast_fp16 = reduce_mean(axes = var_185_axes_0, keep_dims = var_185_keep_dims_0, x = var_6964_cast_fp16)[name = string("var_185_cast_fp16")];
            fp16 var_6971_to_fp16 = const()[name = string("op_6971_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_6972_cast_fp16 = add(x = var_185_cast_fp16, y = var_6971_to_fp16)[name = string("op_6972_cast_fp16")];
            fp32 var_6973_epsilon_0 = const()[name = string("op_6973_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_6973_cast_fp16 = rsqrt(epsilon = var_6973_epsilon_0, x = var_6972_cast_fp16)[name = string("op_6973_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_185_cast_fp16 = mul(x = var_6886_cast_fp16_1, y = var_6973_cast_fp16)[name = string("x_normed_185_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_289_cast_fp16 = mul(x = x_normed_185_cast_fp16, y = var_6919_to_fp16)[name = string("x_289_cast_fp16")];
            tensor<int32, [4]> x_first_103_begin_0 = const()[name = string("x_first_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_103_end_0 = const()[name = string("x_first_103_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_103_end_mask_0 = const()[name = string("x_first_103_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_103_cast_fp16 = slice_by_index(begin = x_first_103_begin_0, end = x_first_103_end_0, end_mask = x_first_103_end_mask_0, x = x_289_cast_fp16)[name = string("x_first_103_cast_fp16")];
            tensor<int32, [4]> x_second_103_begin_0 = const()[name = string("x_second_103_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_103_end_0 = const()[name = string("x_second_103_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_103_end_mask_0 = const()[name = string("x_second_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_103_cast_fp16 = slice_by_index(begin = x_second_103_begin_0, end = x_second_103_end_0, end_mask = x_second_103_end_mask_0, x = x_289_cast_fp16)[name = string("x_second_103_cast_fp16")];
            fp16 const_85_promoted_to_fp16 = const()[name = string("const_85_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_7005_cast_fp16 = mul(x = x_second_103_cast_fp16, y = const_85_promoted_to_fp16)[name = string("op_7005_cast_fp16")];
            int32 var_7007 = const()[name = string("op_7007"), val = int32(1)];
            bool x_rot_103_interleave_0 = const()[name = string("x_rot_103_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_103_cast_fp16 = concat(axis = var_7007, interleave = x_rot_103_interleave_0, values = (var_7005_cast_fp16, x_first_103_cast_fp16))[name = string("x_rot_103_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7023_cast_fp16 = mul(x = x_289_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_7023_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7024_cast_fp16 = mul(x = x_rot_103_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_7024_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_83_cast_fp16 = add(x = var_7023_cast_fp16, y = var_7024_cast_fp16)[name = string("q_h_83_cast_fp16")];
            fp16 var_7027_promoted_to_fp16 = const()[name = string("op_7027_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_7028_cast_fp16 = pow(x = var_6886_cast_fp16_2, y = var_7027_promoted_to_fp16)[name = string("op_7028_cast_fp16")];
            tensor<int32, [1]> var_187_axes_0 = const()[name = string("var_187_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_187_keep_dims_0 = const()[name = string("var_187_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_187_cast_fp16 = reduce_mean(axes = var_187_axes_0, keep_dims = var_187_keep_dims_0, x = var_7028_cast_fp16)[name = string("var_187_cast_fp16")];
            fp16 var_7035_to_fp16 = const()[name = string("op_7035_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_7036_cast_fp16 = add(x = var_187_cast_fp16, y = var_7035_to_fp16)[name = string("op_7036_cast_fp16")];
            fp32 var_7037_epsilon_0 = const()[name = string("op_7037_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_7037_cast_fp16 = rsqrt(epsilon = var_7037_epsilon_0, x = var_7036_cast_fp16)[name = string("op_7037_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_187_cast_fp16 = mul(x = var_6886_cast_fp16_2, y = var_7037_cast_fp16)[name = string("x_normed_187_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_293_cast_fp16 = mul(x = x_normed_187_cast_fp16, y = var_6919_to_fp16)[name = string("x_293_cast_fp16")];
            tensor<int32, [4]> x_first_105_begin_0 = const()[name = string("x_first_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_105_end_0 = const()[name = string("x_first_105_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_105_end_mask_0 = const()[name = string("x_first_105_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_105_cast_fp16 = slice_by_index(begin = x_first_105_begin_0, end = x_first_105_end_0, end_mask = x_first_105_end_mask_0, x = x_293_cast_fp16)[name = string("x_first_105_cast_fp16")];
            tensor<int32, [4]> x_second_105_begin_0 = const()[name = string("x_second_105_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_105_end_0 = const()[name = string("x_second_105_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_105_end_mask_0 = const()[name = string("x_second_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_105_cast_fp16 = slice_by_index(begin = x_second_105_begin_0, end = x_second_105_end_0, end_mask = x_second_105_end_mask_0, x = x_293_cast_fp16)[name = string("x_second_105_cast_fp16")];
            fp16 const_86_promoted_to_fp16 = const()[name = string("const_86_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_7069_cast_fp16 = mul(x = x_second_105_cast_fp16, y = const_86_promoted_to_fp16)[name = string("op_7069_cast_fp16")];
            int32 var_7071 = const()[name = string("op_7071"), val = int32(1)];
            bool x_rot_105_interleave_0 = const()[name = string("x_rot_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_105_cast_fp16 = concat(axis = var_7071, interleave = x_rot_105_interleave_0, values = (var_7069_cast_fp16, x_first_105_cast_fp16))[name = string("x_rot_105_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7087_cast_fp16 = mul(x = x_293_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_7087_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7088_cast_fp16 = mul(x = x_rot_105_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_7088_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_85_cast_fp16 = add(x = var_7087_cast_fp16, y = var_7088_cast_fp16)[name = string("q_h_85_cast_fp16")];
            fp16 var_7091_promoted_to_fp16 = const()[name = string("op_7091_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_7092_cast_fp16 = pow(x = var_6886_cast_fp16_3, y = var_7091_promoted_to_fp16)[name = string("op_7092_cast_fp16")];
            tensor<int32, [1]> var_189_axes_0 = const()[name = string("var_189_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_189_keep_dims_0 = const()[name = string("var_189_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_189_cast_fp16 = reduce_mean(axes = var_189_axes_0, keep_dims = var_189_keep_dims_0, x = var_7092_cast_fp16)[name = string("var_189_cast_fp16")];
            fp16 var_7099_to_fp16 = const()[name = string("op_7099_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_7100_cast_fp16 = add(x = var_189_cast_fp16, y = var_7099_to_fp16)[name = string("op_7100_cast_fp16")];
            fp32 var_7101_epsilon_0 = const()[name = string("op_7101_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_7101_cast_fp16 = rsqrt(epsilon = var_7101_epsilon_0, x = var_7100_cast_fp16)[name = string("op_7101_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_189_cast_fp16 = mul(x = var_6886_cast_fp16_3, y = var_7101_cast_fp16)[name = string("x_normed_189_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_297_cast_fp16 = mul(x = x_normed_189_cast_fp16, y = var_6919_to_fp16)[name = string("x_297_cast_fp16")];
            tensor<int32, [4]> x_first_107_begin_0 = const()[name = string("x_first_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_107_end_0 = const()[name = string("x_first_107_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_107_end_mask_0 = const()[name = string("x_first_107_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_107_cast_fp16 = slice_by_index(begin = x_first_107_begin_0, end = x_first_107_end_0, end_mask = x_first_107_end_mask_0, x = x_297_cast_fp16)[name = string("x_first_107_cast_fp16")];
            tensor<int32, [4]> x_second_107_begin_0 = const()[name = string("x_second_107_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_107_end_0 = const()[name = string("x_second_107_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_107_end_mask_0 = const()[name = string("x_second_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_107_cast_fp16 = slice_by_index(begin = x_second_107_begin_0, end = x_second_107_end_0, end_mask = x_second_107_end_mask_0, x = x_297_cast_fp16)[name = string("x_second_107_cast_fp16")];
            fp16 const_87_promoted_to_fp16 = const()[name = string("const_87_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_7133_cast_fp16 = mul(x = x_second_107_cast_fp16, y = const_87_promoted_to_fp16)[name = string("op_7133_cast_fp16")];
            int32 var_7135 = const()[name = string("op_7135"), val = int32(1)];
            bool x_rot_107_interleave_0 = const()[name = string("x_rot_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_107_cast_fp16 = concat(axis = var_7135, interleave = x_rot_107_interleave_0, values = (var_7133_cast_fp16, x_first_107_cast_fp16))[name = string("x_rot_107_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7151_cast_fp16 = mul(x = x_297_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_7151_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7152_cast_fp16 = mul(x = x_rot_107_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_7152_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_87_cast_fp16 = add(x = var_7151_cast_fp16, y = var_7152_cast_fp16)[name = string("q_h_87_cast_fp16")];
            fp16 var_7155_promoted_to_fp16 = const()[name = string("op_7155_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_7156_cast_fp16 = pow(x = var_6872_cast_fp16, y = var_7155_promoted_to_fp16)[name = string("op_7156_cast_fp16")];
            tensor<int32, [1]> var_191_axes_0 = const()[name = string("var_191_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_191_keep_dims_0 = const()[name = string("var_191_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_191_cast_fp16 = reduce_mean(axes = var_191_axes_0, keep_dims = var_191_keep_dims_0, x = var_7156_cast_fp16)[name = string("var_191_cast_fp16")];
            fp16 var_7163_to_fp16 = const()[name = string("op_7163_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_7164_cast_fp16 = add(x = var_191_cast_fp16, y = var_7163_to_fp16)[name = string("op_7164_cast_fp16")];
            fp32 var_7165_epsilon_0 = const()[name = string("op_7165_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_7165_cast_fp16 = rsqrt(epsilon = var_7165_epsilon_0, x = var_7164_cast_fp16)[name = string("op_7165_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_191_cast_fp16 = mul(x = var_6872_cast_fp16, y = var_7165_cast_fp16)[name = string("x_normed_191_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_7175_to_fp16 = const()[name = string("op_7175_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448993600)))];
            tensor<fp16, [1, 256, 1, ?]> x_301_cast_fp16 = mul(x = x_normed_191_cast_fp16, y = var_7175_to_fp16)[name = string("x_301_cast_fp16")];
            tensor<int32, [4]> x_first_109_begin_0 = const()[name = string("x_first_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_109_end_0 = const()[name = string("x_first_109_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_109_end_mask_0 = const()[name = string("x_first_109_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_109_cast_fp16 = slice_by_index(begin = x_first_109_begin_0, end = x_first_109_end_0, end_mask = x_first_109_end_mask_0, x = x_301_cast_fp16)[name = string("x_first_109_cast_fp16")];
            tensor<int32, [4]> x_second_109_begin_0 = const()[name = string("x_second_109_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_109_end_0 = const()[name = string("x_second_109_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_109_end_mask_0 = const()[name = string("x_second_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_109_cast_fp16 = slice_by_index(begin = x_second_109_begin_0, end = x_second_109_end_0, end_mask = x_second_109_end_mask_0, x = x_301_cast_fp16)[name = string("x_second_109_cast_fp16")];
            fp16 const_88_promoted_to_fp16 = const()[name = string("const_88_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_7197_cast_fp16 = mul(x = x_second_109_cast_fp16, y = const_88_promoted_to_fp16)[name = string("op_7197_cast_fp16")];
            int32 var_7199 = const()[name = string("op_7199"), val = int32(1)];
            bool x_rot_109_interleave_0 = const()[name = string("x_rot_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_109_cast_fp16 = concat(axis = var_7199, interleave = x_rot_109_interleave_0, values = (var_7197_cast_fp16, x_first_109_cast_fp16))[name = string("x_rot_109_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7215_cast_fp16 = mul(x = x_301_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_7215_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7216_cast_fp16 = mul(x = x_rot_109_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_7216_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_21_cast_fp16 = add(x = var_7215_cast_fp16, y = var_7216_cast_fp16)[name = string("k_new_21_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_20 = read_state(input = k_cache_10)[name = string("read_state_20")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_63_cast_fp16 = mul(x = read_state_20, y = one_minus_1_cast_fp16)[name = string("k_cache_63_cast_fp16")];
            write_state(data = k_cache_63_cast_fp16, input = k_cache_10)[name = string("coreml_update_state_112_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_112 = read_state(input = k_cache_10)[name = string("coreml_update_state_112")];
            tensor<fp16, [1, 256, 1, 128]> var_7223_cast_fp16 = mul(x = k_new_21_cast_fp16, y = write_mask)[name = string("op_7223_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_65_cast_fp16 = add(x = coreml_update_state_112, y = var_7223_cast_fp16)[name = string("k_cache_65_cast_fp16")];
            write_state(data = k_cache_65_cast_fp16, input = k_cache_10)[name = string("coreml_update_state_113_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_113 = read_state(input = k_cache_10)[name = string("coreml_update_state_113")];
            tensor<fp16, [1, 256, 1, 128]> read_state_21 = read_state(input = v_cache_10)[name = string("read_state_21")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_63_cast_fp16 = mul(x = read_state_21, y = one_minus_1_cast_fp16)[name = string("v_cache_63_cast_fp16")];
            write_state(data = v_cache_63_cast_fp16, input = v_cache_10)[name = string("coreml_update_state_114_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_114 = read_state(input = v_cache_10)[name = string("coreml_update_state_114")];
            tensor<fp16, [1, 256, 1, 128]> var_7227_cast_fp16 = mul(x = var_6883_cast_fp16, y = write_mask)[name = string("op_7227_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_65_cast_fp16 = add(x = coreml_update_state_114, y = var_7227_cast_fp16)[name = string("v_cache_65_cast_fp16")];
            write_state(data = v_cache_65_cast_fp16, input = v_cache_10)[name = string("coreml_update_state_115_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_115 = read_state(input = v_cache_10)[name = string("coreml_update_state_115")];
            bool var_7242_transpose_x_0 = const()[name = string("op_7242_transpose_x_0"), val = bool(false)];
            bool var_7242_transpose_y_0 = const()[name = string("op_7242_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_386_perm_0 = const()[name = string("transpose_386_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_387_perm_0 = const()[name = string("transpose_387_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_387 = transpose(perm = transpose_387_perm_0, x = coreml_update_state_113)[name = string("transpose_504")];
            tensor<fp16, [1, 1, ?, 256]> transpose_386 = transpose(perm = transpose_386_perm_0, x = q_h_81_cast_fp16)[name = string("transpose_505")];
            tensor<fp16, [1, 1, ?, 128]> var_7242_cast_fp16 = matmul(transpose_x = var_7242_transpose_x_0, transpose_y = var_7242_transpose_y_0, x = transpose_386, y = transpose_387)[name = string("op_7242_cast_fp16")];
            fp16 var_7243_to_fp16 = const()[name = string("op_7243_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_7244_cast_fp16 = mul(x = var_7242_cast_fp16, y = var_7243_to_fp16)[name = string("op_7244_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_81_cast_fp16 = add(x = var_7244_cast_fp16, y = attention_mask)[name = string("scores_81_cast_fp16")];
            int32 var_7247 = const()[name = string("op_7247"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_81_cast_fp16 = softmax(axis = var_7247, x = scores_81_cast_fp16)[name = string("probs_81_cast_fp16")];
            tensor<int32, [4]> var_7254 = const()[name = string("op_7254"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_7256_transpose_x_0 = const()[name = string("op_7256_transpose_x_0"), val = bool(false)];
            bool var_7256_transpose_y_0 = const()[name = string("op_7256_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_81_cast_fp16 = transpose(perm = var_7254, x = coreml_update_state_115)[name = string("transpose_503")];
            tensor<fp16, [1, 1, ?, 256]> var_7256_cast_fp16 = matmul(transpose_x = var_7256_transpose_x_0, transpose_y = var_7256_transpose_y_0, x = probs_81_cast_fp16, y = v_t_81_cast_fp16)[name = string("op_7256_cast_fp16")];
            bool var_7275_transpose_x_0 = const()[name = string("op_7275_transpose_x_0"), val = bool(false)];
            bool var_7275_transpose_y_0 = const()[name = string("op_7275_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_391_perm_0 = const()[name = string("transpose_391_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_391 = transpose(perm = transpose_391_perm_0, x = q_h_83_cast_fp16)[name = string("transpose_502")];
            tensor<fp16, [1, 1, ?, 128]> var_7275_cast_fp16 = matmul(transpose_x = var_7275_transpose_x_0, transpose_y = var_7275_transpose_y_0, x = transpose_391, y = transpose_387)[name = string("op_7275_cast_fp16")];
            fp16 var_7276_to_fp16 = const()[name = string("op_7276_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_7277_cast_fp16 = mul(x = var_7275_cast_fp16, y = var_7276_to_fp16)[name = string("op_7277_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_83_cast_fp16 = add(x = var_7277_cast_fp16, y = attention_mask)[name = string("scores_83_cast_fp16")];
            int32 var_7280 = const()[name = string("op_7280"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_83_cast_fp16 = softmax(axis = var_7280, x = scores_83_cast_fp16)[name = string("probs_83_cast_fp16")];
            bool var_7289_transpose_x_0 = const()[name = string("op_7289_transpose_x_0"), val = bool(false)];
            bool var_7289_transpose_y_0 = const()[name = string("op_7289_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_7289_cast_fp16 = matmul(transpose_x = var_7289_transpose_x_0, transpose_y = var_7289_transpose_y_0, x = probs_83_cast_fp16, y = v_t_81_cast_fp16)[name = string("op_7289_cast_fp16")];
            bool var_7308_transpose_x_0 = const()[name = string("op_7308_transpose_x_0"), val = bool(false)];
            bool var_7308_transpose_y_0 = const()[name = string("op_7308_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_392_perm_0 = const()[name = string("transpose_392_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_392 = transpose(perm = transpose_392_perm_0, x = q_h_85_cast_fp16)[name = string("transpose_501")];
            tensor<fp16, [1, 1, ?, 128]> var_7308_cast_fp16 = matmul(transpose_x = var_7308_transpose_x_0, transpose_y = var_7308_transpose_y_0, x = transpose_392, y = transpose_387)[name = string("op_7308_cast_fp16")];
            fp16 var_7309_to_fp16 = const()[name = string("op_7309_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_7310_cast_fp16 = mul(x = var_7308_cast_fp16, y = var_7309_to_fp16)[name = string("op_7310_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_85_cast_fp16 = add(x = var_7310_cast_fp16, y = attention_mask)[name = string("scores_85_cast_fp16")];
            int32 var_7313 = const()[name = string("op_7313"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_85_cast_fp16 = softmax(axis = var_7313, x = scores_85_cast_fp16)[name = string("probs_85_cast_fp16")];
            bool var_7322_transpose_x_0 = const()[name = string("op_7322_transpose_x_0"), val = bool(false)];
            bool var_7322_transpose_y_0 = const()[name = string("op_7322_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_7322_cast_fp16 = matmul(transpose_x = var_7322_transpose_x_0, transpose_y = var_7322_transpose_y_0, x = probs_85_cast_fp16, y = v_t_81_cast_fp16)[name = string("op_7322_cast_fp16")];
            bool var_7341_transpose_x_0 = const()[name = string("op_7341_transpose_x_0"), val = bool(false)];
            bool var_7341_transpose_y_0 = const()[name = string("op_7341_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_393_perm_0 = const()[name = string("transpose_393_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_393 = transpose(perm = transpose_393_perm_0, x = q_h_87_cast_fp16)[name = string("transpose_500")];
            tensor<fp16, [1, 1, ?, 128]> var_7341_cast_fp16 = matmul(transpose_x = var_7341_transpose_x_0, transpose_y = var_7341_transpose_y_0, x = transpose_393, y = transpose_387)[name = string("op_7341_cast_fp16")];
            fp16 var_7342_to_fp16 = const()[name = string("op_7342_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_7343_cast_fp16 = mul(x = var_7341_cast_fp16, y = var_7342_to_fp16)[name = string("op_7343_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_87_cast_fp16 = add(x = var_7343_cast_fp16, y = attention_mask)[name = string("scores_87_cast_fp16")];
            int32 var_7346 = const()[name = string("op_7346"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_87_cast_fp16 = softmax(axis = var_7346, x = scores_87_cast_fp16)[name = string("probs_87_cast_fp16")];
            bool var_7355_transpose_x_0 = const()[name = string("op_7355_transpose_x_0"), val = bool(false)];
            bool var_7355_transpose_y_0 = const()[name = string("op_7355_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_7355_cast_fp16 = matmul(transpose_x = var_7355_transpose_x_0, transpose_y = var_7355_transpose_y_0, x = probs_87_cast_fp16, y = v_t_81_cast_fp16)[name = string("op_7355_cast_fp16")];
            bool input_83_interleave_0 = const()[name = string("input_83_interleave_0"), val = bool(false)];
            int32 const_155 = const()[name = string("const_155"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_83_cast_fp16 = concat(axis = const_155, interleave = input_83_interleave_0, values = (var_7256_cast_fp16, var_7289_cast_fp16, var_7322_cast_fp16, var_7355_cast_fp16))[name = string("input_83_cast_fp16")];
            string x_303_pad_type_0 = const()[name = string("x_303_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_303_strides_0 = const()[name = string("x_303_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_303_pad_0 = const()[name = string("x_303_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_303_dilations_0 = const()[name = string("x_303_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_303_groups_0 = const()[name = string("x_303_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_10_weight_to_fp16 = const()[name = string("o_proj_10_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448994176)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_98_cast_fp16 = transpose(perm = transpose_98_perm_0, x = input_83_cast_fp16)[name = string("transpose_499")];
            tensor<fp16, [1, 640, 1, ?]> x_303_cast_fp16 = conv(dilations = x_303_dilations_0, groups = x_303_groups_0, pad = x_303_pad_0, pad_type = x_303_pad_type_0, strides = x_303_strides_0, weight = o_proj_10_weight_to_fp16, x = transpose_98_cast_fp16)[name = string("x_303_cast_fp16")];
            fp16 var_7376_promoted_to_fp16 = const()[name = string("op_7376_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_7377_cast_fp16 = pow(x = x_303_cast_fp16, y = var_7376_promoted_to_fp16)[name = string("op_7377_cast_fp16")];
            tensor<int32, [1]> var_193_axes_0 = const()[name = string("var_193_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_193_keep_dims_0 = const()[name = string("var_193_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_193_cast_fp16 = reduce_mean(axes = var_193_axes_0, keep_dims = var_193_keep_dims_0, x = var_7377_cast_fp16)[name = string("var_193_cast_fp16")];
            fp16 var_7384_to_fp16 = const()[name = string("op_7384_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_7385_cast_fp16 = add(x = var_193_cast_fp16, y = var_7384_to_fp16)[name = string("op_7385_cast_fp16")];
            fp32 var_7386_epsilon_0 = const()[name = string("op_7386_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_7386_cast_fp16 = rsqrt(epsilon = var_7386_epsilon_0, x = var_7385_cast_fp16)[name = string("op_7386_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_193_cast_fp16 = mul(x = x_303_cast_fp16, y = var_7386_cast_fp16)[name = string("x_normed_193_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_7396_to_fp16 = const()[name = string("op_7396_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450304960)))];
            tensor<fp16, [1, 640, 1, ?]> attn_43_cast_fp16 = mul(x = x_normed_193_cast_fp16, y = var_7396_to_fp16)[name = string("attn_43_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_305_cast_fp16 = add(x = x_281_cast_fp16, y = attn_43_cast_fp16)[name = string("x_305_cast_fp16")];
            fp16 var_7400_promoted_to_fp16 = const()[name = string("op_7400_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_7401_cast_fp16 = pow(x = x_305_cast_fp16, y = var_7400_promoted_to_fp16)[name = string("op_7401_cast_fp16")];
            tensor<int32, [1]> var_195_axes_0 = const()[name = string("var_195_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_195_keep_dims_0 = const()[name = string("var_195_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_195_cast_fp16 = reduce_mean(axes = var_195_axes_0, keep_dims = var_195_keep_dims_0, x = var_7401_cast_fp16)[name = string("var_195_cast_fp16")];
            fp16 var_7408_to_fp16 = const()[name = string("op_7408_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_7409_cast_fp16 = add(x = var_195_cast_fp16, y = var_7408_to_fp16)[name = string("op_7409_cast_fp16")];
            fp32 var_7410_epsilon_0 = const()[name = string("op_7410_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_7410_cast_fp16 = rsqrt(epsilon = var_7410_epsilon_0, x = var_7409_cast_fp16)[name = string("op_7410_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_195_cast_fp16 = mul(x = x_305_cast_fp16, y = var_7410_cast_fp16)[name = string("x_normed_195_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_7420_to_fp16 = const()[name = string("op_7420_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450306304)))];
            tensor<fp16, [1, 640, 1, ?]> input_85_cast_fp16 = mul(x = x_normed_195_cast_fp16, y = var_7420_to_fp16)[name = string("input_85_cast_fp16")];
            string gate_21_pad_type_0 = const()[name = string("gate_21_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_21_strides_0 = const()[name = string("gate_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_21_pad_0 = const()[name = string("gate_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_21_dilations_0 = const()[name = string("gate_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_21_groups_0 = const()[name = string("gate_21_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_10_weight_to_fp16 = const()[name = string("gate_proj_10_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450307648)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_21_cast_fp16 = conv(dilations = gate_21_dilations_0, groups = gate_21_groups_0, pad = gate_21_pad_0, pad_type = gate_21_pad_type_0, strides = gate_21_strides_0, weight = gate_proj_10_weight_to_fp16, x = input_85_cast_fp16)[name = string("gate_21_cast_fp16")];
            string up_21_pad_type_0 = const()[name = string("up_21_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_21_strides_0 = const()[name = string("up_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_21_pad_0 = const()[name = string("up_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_21_dilations_0 = const()[name = string("up_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_21_groups_0 = const()[name = string("up_21_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_10_weight_to_fp16 = const()[name = string("up_proj_10_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452929152)))];
            tensor<fp16, [1, 2048, 1, ?]> up_21_cast_fp16 = conv(dilations = up_21_dilations_0, groups = up_21_groups_0, pad = up_21_pad_0, pad_type = up_21_pad_type_0, strides = up_21_strides_0, weight = up_proj_10_weight_to_fp16, x = input_85_cast_fp16)[name = string("up_21_cast_fp16")];
            string var_7445_mode_0 = const()[name = string("op_7445_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_7445_cast_fp16 = gelu(mode = var_7445_mode_0, x = gate_21_cast_fp16)[name = string("op_7445_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_87_cast_fp16 = mul(x = var_7445_cast_fp16, y = up_21_cast_fp16)[name = string("input_87_cast_fp16")];
            string x_307_pad_type_0 = const()[name = string("x_307_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_307_strides_0 = const()[name = string("x_307_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_307_pad_0 = const()[name = string("x_307_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_307_dilations_0 = const()[name = string("x_307_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_307_groups_0 = const()[name = string("x_307_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_10_weight_to_fp16 = const()[name = string("down_proj_10_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455550656)))];
            tensor<fp16, [1, 640, 1, ?]> x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = down_proj_10_weight_to_fp16, x = input_87_cast_fp16)[name = string("x_307_cast_fp16")];
            fp16 var_7458_promoted_to_fp16 = const()[name = string("op_7458_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_7459_cast_fp16 = pow(x = x_307_cast_fp16, y = var_7458_promoted_to_fp16)[name = string("op_7459_cast_fp16")];
            tensor<int32, [1]> var_197_axes_0 = const()[name = string("var_197_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_197_keep_dims_0 = const()[name = string("var_197_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_197_cast_fp16 = reduce_mean(axes = var_197_axes_0, keep_dims = var_197_keep_dims_0, x = var_7459_cast_fp16)[name = string("var_197_cast_fp16")];
            fp16 var_7466_to_fp16 = const()[name = string("op_7466_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_7467_cast_fp16 = add(x = var_197_cast_fp16, y = var_7466_to_fp16)[name = string("op_7467_cast_fp16")];
            fp32 var_7468_epsilon_0 = const()[name = string("op_7468_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_7468_cast_fp16 = rsqrt(epsilon = var_7468_epsilon_0, x = var_7467_cast_fp16)[name = string("op_7468_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_197_cast_fp16 = mul(x = x_307_cast_fp16, y = var_7468_cast_fp16)[name = string("x_normed_197_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_7478_to_fp16 = const()[name = string("op_7478_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458172160)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_21_cast_fp16 = mul(x = x_normed_197_cast_fp16, y = var_7478_to_fp16)[name = string("mlp_21_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_309_cast_fp16 = add(x = x_305_cast_fp16, y = mlp_21_cast_fp16)[name = string("x_309_cast_fp16")];
            fp16 var_7482_promoted_to_fp16 = const()[name = string("op_7482_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_7483_cast_fp16 = pow(x = x_309_cast_fp16, y = var_7482_promoted_to_fp16)[name = string("op_7483_cast_fp16")];
            tensor<int32, [1]> var_199_axes_0 = const()[name = string("var_199_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_199_keep_dims_0 = const()[name = string("var_199_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_199_cast_fp16 = reduce_mean(axes = var_199_axes_0, keep_dims = var_199_keep_dims_0, x = var_7483_cast_fp16)[name = string("var_199_cast_fp16")];
            fp16 var_7490_to_fp16 = const()[name = string("op_7490_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_7491_cast_fp16 = add(x = var_199_cast_fp16, y = var_7490_to_fp16)[name = string("op_7491_cast_fp16")];
            fp32 var_7492_epsilon_0 = const()[name = string("op_7492_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_7492_cast_fp16 = rsqrt(epsilon = var_7492_epsilon_0, x = var_7491_cast_fp16)[name = string("op_7492_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_199_cast_fp16 = mul(x = x_309_cast_fp16, y = var_7492_cast_fp16)[name = string("x_normed_199_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_7502_to_fp16 = const()[name = string("op_7502_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458173504)))];
            tensor<fp16, [1, 640, 1, ?]> input_89_cast_fp16 = mul(x = x_normed_199_cast_fp16, y = var_7502_to_fp16)[name = string("input_89_cast_fp16")];
            string var_7514_pad_type_0 = const()[name = string("op_7514_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_7514_strides_0 = const()[name = string("op_7514_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7514_pad_0 = const()[name = string("op_7514_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7514_dilations_0 = const()[name = string("op_7514_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_7514_groups_0 = const()[name = string("op_7514_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_11_weight_to_fp16 = const()[name = string("q_proj_11_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458174848)))];
            tensor<fp16, [1, 1024, 1, ?]> var_7514_cast_fp16 = conv(dilations = var_7514_dilations_0, groups = var_7514_groups_0, pad = var_7514_pad_0, pad_type = var_7514_pad_type_0, strides = var_7514_strides_0, weight = q_proj_11_weight_to_fp16, x = input_89_cast_fp16)[name = string("op_7514_cast_fp16")];
            string var_7525_pad_type_0 = const()[name = string("op_7525_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_7525_strides_0 = const()[name = string("op_7525_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7525_pad_0 = const()[name = string("op_7525_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7525_dilations_0 = const()[name = string("op_7525_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_7525_groups_0 = const()[name = string("op_7525_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_11_weight_to_fp16 = const()[name = string("k_proj_11_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459485632)))];
            tensor<fp16, [1, 256, 1, ?]> var_7525_cast_fp16 = conv(dilations = var_7525_dilations_0, groups = var_7525_groups_0, pad = var_7525_pad_0, pad_type = var_7525_pad_type_0, strides = var_7525_strides_0, weight = k_proj_11_weight_to_fp16, x = input_89_cast_fp16)[name = string("op_7525_cast_fp16")];
            string var_7536_pad_type_0 = const()[name = string("op_7536_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_7536_strides_0 = const()[name = string("op_7536_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_7536_pad_0 = const()[name = string("op_7536_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_7536_dilations_0 = const()[name = string("op_7536_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_7536_groups_0 = const()[name = string("op_7536_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_11_weight_to_fp16 = const()[name = string("v_proj_11_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459813376)))];
            tensor<fp16, [1, 256, 1, ?]> var_7536_cast_fp16 = conv(dilations = var_7536_dilations_0, groups = var_7536_groups_0, pad = var_7536_pad_0, pad_type = var_7536_pad_type_0, strides = var_7536_strides_0, weight = v_proj_11_weight_to_fp16, x = input_89_cast_fp16)[name = string("op_7536_cast_fp16")];
            tensor<int32, [4]> tile_33 = const()[name = string("tile_33"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_7539_axis_0 = const()[name = string("op_7539_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_7539_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_7539_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_7539_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_7539_cast_fp16_3 = split(axis = var_7539_axis_0, split_sizes = tile_33, x = var_7514_cast_fp16)[name = string("op_7539_cast_fp16")];
            fp16 var_7552_promoted_to_fp16 = const()[name = string("op_7552_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_7553_cast_fp16 = pow(x = var_7539_cast_fp16_0, y = var_7552_promoted_to_fp16)[name = string("op_7553_cast_fp16")];
            tensor<int32, [1]> var_201_axes_0 = const()[name = string("var_201_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_201_keep_dims_0 = const()[name = string("var_201_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_201_cast_fp16 = reduce_mean(axes = var_201_axes_0, keep_dims = var_201_keep_dims_0, x = var_7553_cast_fp16)[name = string("var_201_cast_fp16")];
            fp16 var_7560_to_fp16 = const()[name = string("op_7560_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_7561_cast_fp16 = add(x = var_201_cast_fp16, y = var_7560_to_fp16)[name = string("op_7561_cast_fp16")];
            fp32 var_7562_epsilon_0 = const()[name = string("op_7562_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_7562_cast_fp16 = rsqrt(epsilon = var_7562_epsilon_0, x = var_7561_cast_fp16)[name = string("op_7562_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_201_cast_fp16 = mul(x = var_7539_cast_fp16_0, y = var_7562_cast_fp16)[name = string("x_normed_201_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_7572_to_fp16 = const()[name = string("op_7572_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460141120)))];
            tensor<fp16, [1, 256, 1, ?]> x_313_cast_fp16 = mul(x = x_normed_201_cast_fp16, y = var_7572_to_fp16)[name = string("x_313_cast_fp16")];
            tensor<int32, [4]> x_first_111_begin_0 = const()[name = string("x_first_111_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_111_end_0 = const()[name = string("x_first_111_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_111_end_mask_0 = const()[name = string("x_first_111_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_111_cast_fp16 = slice_by_index(begin = x_first_111_begin_0, end = x_first_111_end_0, end_mask = x_first_111_end_mask_0, x = x_313_cast_fp16)[name = string("x_first_111_cast_fp16")];
            tensor<int32, [4]> x_second_111_begin_0 = const()[name = string("x_second_111_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_111_end_0 = const()[name = string("x_second_111_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_111_end_mask_0 = const()[name = string("x_second_111_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_111_cast_fp16 = slice_by_index(begin = x_second_111_begin_0, end = x_second_111_end_0, end_mask = x_second_111_end_mask_0, x = x_313_cast_fp16)[name = string("x_second_111_cast_fp16")];
            fp16 const_92_promoted_to_fp16 = const()[name = string("const_92_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_7594_cast_fp16 = mul(x = x_second_111_cast_fp16, y = const_92_promoted_to_fp16)[name = string("op_7594_cast_fp16")];
            int32 var_7596 = const()[name = string("op_7596"), val = int32(1)];
            bool x_rot_111_interleave_0 = const()[name = string("x_rot_111_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_111_cast_fp16 = concat(axis = var_7596, interleave = x_rot_111_interleave_0, values = (var_7594_cast_fp16, x_first_111_cast_fp16))[name = string("x_rot_111_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7612_cast_fp16 = mul(x = x_313_cast_fp16, y = cos_bc1t_51_cast_fp16)[name = string("op_7612_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7613_cast_fp16 = mul(x = x_rot_111_cast_fp16, y = sin_bc1t_51_cast_fp16)[name = string("op_7613_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_89_cast_fp16 = add(x = var_7612_cast_fp16, y = var_7613_cast_fp16)[name = string("q_h_89_cast_fp16")];
            fp16 var_7616_promoted_to_fp16 = const()[name = string("op_7616_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_7617_cast_fp16 = pow(x = var_7539_cast_fp16_1, y = var_7616_promoted_to_fp16)[name = string("op_7617_cast_fp16")];
            tensor<int32, [1]> var_203_axes_0 = const()[name = string("var_203_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_203_keep_dims_0 = const()[name = string("var_203_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_203_cast_fp16 = reduce_mean(axes = var_203_axes_0, keep_dims = var_203_keep_dims_0, x = var_7617_cast_fp16)[name = string("var_203_cast_fp16")];
            fp16 var_7624_to_fp16 = const()[name = string("op_7624_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_7625_cast_fp16 = add(x = var_203_cast_fp16, y = var_7624_to_fp16)[name = string("op_7625_cast_fp16")];
            fp32 var_7626_epsilon_0 = const()[name = string("op_7626_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_7626_cast_fp16 = rsqrt(epsilon = var_7626_epsilon_0, x = var_7625_cast_fp16)[name = string("op_7626_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_203_cast_fp16 = mul(x = var_7539_cast_fp16_1, y = var_7626_cast_fp16)[name = string("x_normed_203_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_317_cast_fp16 = mul(x = x_normed_203_cast_fp16, y = var_7572_to_fp16)[name = string("x_317_cast_fp16")];
            tensor<int32, [4]> x_first_113_begin_0 = const()[name = string("x_first_113_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_113_end_0 = const()[name = string("x_first_113_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_113_end_mask_0 = const()[name = string("x_first_113_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_113_cast_fp16 = slice_by_index(begin = x_first_113_begin_0, end = x_first_113_end_0, end_mask = x_first_113_end_mask_0, x = x_317_cast_fp16)[name = string("x_first_113_cast_fp16")];
            tensor<int32, [4]> x_second_113_begin_0 = const()[name = string("x_second_113_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_113_end_0 = const()[name = string("x_second_113_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_113_end_mask_0 = const()[name = string("x_second_113_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_113_cast_fp16 = slice_by_index(begin = x_second_113_begin_0, end = x_second_113_end_0, end_mask = x_second_113_end_mask_0, x = x_317_cast_fp16)[name = string("x_second_113_cast_fp16")];
            fp16 const_93_promoted_to_fp16 = const()[name = string("const_93_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_7658_cast_fp16 = mul(x = x_second_113_cast_fp16, y = const_93_promoted_to_fp16)[name = string("op_7658_cast_fp16")];
            int32 var_7660 = const()[name = string("op_7660"), val = int32(1)];
            bool x_rot_113_interleave_0 = const()[name = string("x_rot_113_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_113_cast_fp16 = concat(axis = var_7660, interleave = x_rot_113_interleave_0, values = (var_7658_cast_fp16, x_first_113_cast_fp16))[name = string("x_rot_113_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7676_cast_fp16 = mul(x = x_317_cast_fp16, y = cos_bc1t_51_cast_fp16)[name = string("op_7676_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7677_cast_fp16 = mul(x = x_rot_113_cast_fp16, y = sin_bc1t_51_cast_fp16)[name = string("op_7677_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_91_cast_fp16 = add(x = var_7676_cast_fp16, y = var_7677_cast_fp16)[name = string("q_h_91_cast_fp16")];
            fp16 var_7680_promoted_to_fp16 = const()[name = string("op_7680_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_7681_cast_fp16 = pow(x = var_7539_cast_fp16_2, y = var_7680_promoted_to_fp16)[name = string("op_7681_cast_fp16")];
            tensor<int32, [1]> var_205_axes_0 = const()[name = string("var_205_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_205_keep_dims_0 = const()[name = string("var_205_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_205_cast_fp16 = reduce_mean(axes = var_205_axes_0, keep_dims = var_205_keep_dims_0, x = var_7681_cast_fp16)[name = string("var_205_cast_fp16")];
            fp16 var_7688_to_fp16 = const()[name = string("op_7688_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_7689_cast_fp16 = add(x = var_205_cast_fp16, y = var_7688_to_fp16)[name = string("op_7689_cast_fp16")];
            fp32 var_7690_epsilon_0 = const()[name = string("op_7690_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_7690_cast_fp16 = rsqrt(epsilon = var_7690_epsilon_0, x = var_7689_cast_fp16)[name = string("op_7690_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_205_cast_fp16 = mul(x = var_7539_cast_fp16_2, y = var_7690_cast_fp16)[name = string("x_normed_205_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_321_cast_fp16 = mul(x = x_normed_205_cast_fp16, y = var_7572_to_fp16)[name = string("x_321_cast_fp16")];
            tensor<int32, [4]> x_first_115_begin_0 = const()[name = string("x_first_115_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_115_end_0 = const()[name = string("x_first_115_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_115_end_mask_0 = const()[name = string("x_first_115_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_115_cast_fp16 = slice_by_index(begin = x_first_115_begin_0, end = x_first_115_end_0, end_mask = x_first_115_end_mask_0, x = x_321_cast_fp16)[name = string("x_first_115_cast_fp16")];
            tensor<int32, [4]> x_second_115_begin_0 = const()[name = string("x_second_115_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_115_end_0 = const()[name = string("x_second_115_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_115_end_mask_0 = const()[name = string("x_second_115_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_115_cast_fp16 = slice_by_index(begin = x_second_115_begin_0, end = x_second_115_end_0, end_mask = x_second_115_end_mask_0, x = x_321_cast_fp16)[name = string("x_second_115_cast_fp16")];
            fp16 const_94_promoted_to_fp16 = const()[name = string("const_94_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_7722_cast_fp16 = mul(x = x_second_115_cast_fp16, y = const_94_promoted_to_fp16)[name = string("op_7722_cast_fp16")];
            int32 var_7724 = const()[name = string("op_7724"), val = int32(1)];
            bool x_rot_115_interleave_0 = const()[name = string("x_rot_115_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_115_cast_fp16 = concat(axis = var_7724, interleave = x_rot_115_interleave_0, values = (var_7722_cast_fp16, x_first_115_cast_fp16))[name = string("x_rot_115_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7740_cast_fp16 = mul(x = x_321_cast_fp16, y = cos_bc1t_51_cast_fp16)[name = string("op_7740_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7741_cast_fp16 = mul(x = x_rot_115_cast_fp16, y = sin_bc1t_51_cast_fp16)[name = string("op_7741_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_93_cast_fp16 = add(x = var_7740_cast_fp16, y = var_7741_cast_fp16)[name = string("q_h_93_cast_fp16")];
            fp16 var_7744_promoted_to_fp16 = const()[name = string("op_7744_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_7745_cast_fp16 = pow(x = var_7539_cast_fp16_3, y = var_7744_promoted_to_fp16)[name = string("op_7745_cast_fp16")];
            tensor<int32, [1]> var_207_axes_0 = const()[name = string("var_207_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_207_keep_dims_0 = const()[name = string("var_207_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_207_cast_fp16 = reduce_mean(axes = var_207_axes_0, keep_dims = var_207_keep_dims_0, x = var_7745_cast_fp16)[name = string("var_207_cast_fp16")];
            fp16 var_7752_to_fp16 = const()[name = string("op_7752_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_7753_cast_fp16 = add(x = var_207_cast_fp16, y = var_7752_to_fp16)[name = string("op_7753_cast_fp16")];
            fp32 var_7754_epsilon_0 = const()[name = string("op_7754_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_7754_cast_fp16 = rsqrt(epsilon = var_7754_epsilon_0, x = var_7753_cast_fp16)[name = string("op_7754_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_207_cast_fp16 = mul(x = var_7539_cast_fp16_3, y = var_7754_cast_fp16)[name = string("x_normed_207_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_325_cast_fp16 = mul(x = x_normed_207_cast_fp16, y = var_7572_to_fp16)[name = string("x_325_cast_fp16")];
            tensor<int32, [4]> x_first_117_begin_0 = const()[name = string("x_first_117_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_117_end_0 = const()[name = string("x_first_117_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_117_end_mask_0 = const()[name = string("x_first_117_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_117_cast_fp16 = slice_by_index(begin = x_first_117_begin_0, end = x_first_117_end_0, end_mask = x_first_117_end_mask_0, x = x_325_cast_fp16)[name = string("x_first_117_cast_fp16")];
            tensor<int32, [4]> x_second_117_begin_0 = const()[name = string("x_second_117_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_117_end_0 = const()[name = string("x_second_117_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_117_end_mask_0 = const()[name = string("x_second_117_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_117_cast_fp16 = slice_by_index(begin = x_second_117_begin_0, end = x_second_117_end_0, end_mask = x_second_117_end_mask_0, x = x_325_cast_fp16)[name = string("x_second_117_cast_fp16")];
            fp16 const_95_promoted_to_fp16 = const()[name = string("const_95_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_7786_cast_fp16 = mul(x = x_second_117_cast_fp16, y = const_95_promoted_to_fp16)[name = string("op_7786_cast_fp16")];
            int32 var_7788 = const()[name = string("op_7788"), val = int32(1)];
            bool x_rot_117_interleave_0 = const()[name = string("x_rot_117_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_117_cast_fp16 = concat(axis = var_7788, interleave = x_rot_117_interleave_0, values = (var_7786_cast_fp16, x_first_117_cast_fp16))[name = string("x_rot_117_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7804_cast_fp16 = mul(x = x_325_cast_fp16, y = cos_bc1t_51_cast_fp16)[name = string("op_7804_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7805_cast_fp16 = mul(x = x_rot_117_cast_fp16, y = sin_bc1t_51_cast_fp16)[name = string("op_7805_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_95_cast_fp16 = add(x = var_7804_cast_fp16, y = var_7805_cast_fp16)[name = string("q_h_95_cast_fp16")];
            fp16 var_7808_promoted_to_fp16 = const()[name = string("op_7808_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_7809_cast_fp16 = pow(x = var_7525_cast_fp16, y = var_7808_promoted_to_fp16)[name = string("op_7809_cast_fp16")];
            tensor<int32, [1]> var_209_axes_0 = const()[name = string("var_209_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_209_keep_dims_0 = const()[name = string("var_209_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_209_cast_fp16 = reduce_mean(axes = var_209_axes_0, keep_dims = var_209_keep_dims_0, x = var_7809_cast_fp16)[name = string("var_209_cast_fp16")];
            fp16 var_7816_to_fp16 = const()[name = string("op_7816_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_7817_cast_fp16 = add(x = var_209_cast_fp16, y = var_7816_to_fp16)[name = string("op_7817_cast_fp16")];
            fp32 var_7818_epsilon_0 = const()[name = string("op_7818_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_7818_cast_fp16 = rsqrt(epsilon = var_7818_epsilon_0, x = var_7817_cast_fp16)[name = string("op_7818_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_209_cast_fp16 = mul(x = var_7525_cast_fp16, y = var_7818_cast_fp16)[name = string("x_normed_209_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_7828_to_fp16 = const()[name = string("op_7828_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460141696)))];
            tensor<fp16, [1, 256, 1, ?]> x_329_cast_fp16 = mul(x = x_normed_209_cast_fp16, y = var_7828_to_fp16)[name = string("x_329_cast_fp16")];
            tensor<int32, [4]> x_first_119_begin_0 = const()[name = string("x_first_119_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_119_end_0 = const()[name = string("x_first_119_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_119_end_mask_0 = const()[name = string("x_first_119_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_119_cast_fp16 = slice_by_index(begin = x_first_119_begin_0, end = x_first_119_end_0, end_mask = x_first_119_end_mask_0, x = x_329_cast_fp16)[name = string("x_first_119_cast_fp16")];
            tensor<int32, [4]> x_second_119_begin_0 = const()[name = string("x_second_119_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_119_end_0 = const()[name = string("x_second_119_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_119_end_mask_0 = const()[name = string("x_second_119_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_119_cast_fp16 = slice_by_index(begin = x_second_119_begin_0, end = x_second_119_end_0, end_mask = x_second_119_end_mask_0, x = x_329_cast_fp16)[name = string("x_second_119_cast_fp16")];
            fp16 const_96_promoted_to_fp16 = const()[name = string("const_96_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_7850_cast_fp16 = mul(x = x_second_119_cast_fp16, y = const_96_promoted_to_fp16)[name = string("op_7850_cast_fp16")];
            int32 var_7852 = const()[name = string("op_7852"), val = int32(1)];
            bool x_rot_119_interleave_0 = const()[name = string("x_rot_119_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_119_cast_fp16 = concat(axis = var_7852, interleave = x_rot_119_interleave_0, values = (var_7850_cast_fp16, x_first_119_cast_fp16))[name = string("x_rot_119_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7868_cast_fp16 = mul(x = x_329_cast_fp16, y = cos_bc1t_51_cast_fp16)[name = string("op_7868_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_7869_cast_fp16 = mul(x = x_rot_119_cast_fp16, y = sin_bc1t_51_cast_fp16)[name = string("op_7869_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_23_cast_fp16 = add(x = var_7868_cast_fp16, y = var_7869_cast_fp16)[name = string("k_new_23_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_22 = read_state(input = k_cache_11)[name = string("read_state_22")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_69_cast_fp16 = mul(x = read_state_22, y = one_minus_1_cast_fp16)[name = string("k_cache_69_cast_fp16")];
            write_state(data = k_cache_69_cast_fp16, input = k_cache_11)[name = string("coreml_update_state_116_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_116 = read_state(input = k_cache_11)[name = string("coreml_update_state_116")];
            tensor<fp16, [1, 256, 1, 128]> var_7876_cast_fp16 = mul(x = k_new_23_cast_fp16, y = write_mask)[name = string("op_7876_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_71_cast_fp16 = add(x = coreml_update_state_116, y = var_7876_cast_fp16)[name = string("k_cache_71_cast_fp16")];
            write_state(data = k_cache_71_cast_fp16, input = k_cache_11)[name = string("coreml_update_state_117_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_117 = read_state(input = k_cache_11)[name = string("coreml_update_state_117")];
            tensor<fp16, [1, 256, 1, 128]> read_state_23 = read_state(input = v_cache_11)[name = string("read_state_23")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_69_cast_fp16 = mul(x = read_state_23, y = one_minus_1_cast_fp16)[name = string("v_cache_69_cast_fp16")];
            write_state(data = v_cache_69_cast_fp16, input = v_cache_11)[name = string("coreml_update_state_118_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_118 = read_state(input = v_cache_11)[name = string("coreml_update_state_118")];
            tensor<fp16, [1, 256, 1, 128]> var_7880_cast_fp16 = mul(x = var_7536_cast_fp16, y = write_mask)[name = string("op_7880_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_71_cast_fp16 = add(x = coreml_update_state_118, y = var_7880_cast_fp16)[name = string("v_cache_71_cast_fp16")];
            write_state(data = v_cache_71_cast_fp16, input = v_cache_11)[name = string("coreml_update_state_119_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_119 = read_state(input = v_cache_11)[name = string("coreml_update_state_119")];
            bool var_7895_transpose_x_0 = const()[name = string("op_7895_transpose_x_0"), val = bool(false)];
            bool var_7895_transpose_y_0 = const()[name = string("op_7895_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_394_perm_0 = const()[name = string("transpose_394_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_395_perm_0 = const()[name = string("transpose_395_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_395 = transpose(perm = transpose_395_perm_0, x = coreml_update_state_117)[name = string("transpose_497")];
            tensor<fp16, [1, 1, ?, 256]> transpose_394 = transpose(perm = transpose_394_perm_0, x = q_h_89_cast_fp16)[name = string("transpose_498")];
            tensor<fp16, [1, 1, ?, 128]> var_7895_cast_fp16 = matmul(transpose_x = var_7895_transpose_x_0, transpose_y = var_7895_transpose_y_0, x = transpose_394, y = transpose_395)[name = string("op_7895_cast_fp16")];
            fp16 var_7896_to_fp16 = const()[name = string("op_7896_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_7897_cast_fp16 = mul(x = var_7895_cast_fp16, y = var_7896_to_fp16)[name = string("op_7897_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_89_cast_fp16 = add(x = var_7897_cast_fp16, y = attention_mask)[name = string("scores_89_cast_fp16")];
            int32 var_7900 = const()[name = string("op_7900"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_89_cast_fp16 = softmax(axis = var_7900, x = scores_89_cast_fp16)[name = string("probs_89_cast_fp16")];
            tensor<int32, [4]> var_7907 = const()[name = string("op_7907"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_7909_transpose_x_0 = const()[name = string("op_7909_transpose_x_0"), val = bool(false)];
            bool var_7909_transpose_y_0 = const()[name = string("op_7909_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_89_cast_fp16 = transpose(perm = var_7907, x = coreml_update_state_119)[name = string("transpose_496")];
            tensor<fp16, [1, 1, ?, 256]> var_7909_cast_fp16 = matmul(transpose_x = var_7909_transpose_x_0, transpose_y = var_7909_transpose_y_0, x = probs_89_cast_fp16, y = v_t_89_cast_fp16)[name = string("op_7909_cast_fp16")];
            bool var_7928_transpose_x_0 = const()[name = string("op_7928_transpose_x_0"), val = bool(false)];
            bool var_7928_transpose_y_0 = const()[name = string("op_7928_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_399_perm_0 = const()[name = string("transpose_399_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_399 = transpose(perm = transpose_399_perm_0, x = q_h_91_cast_fp16)[name = string("transpose_495")];
            tensor<fp16, [1, 1, ?, 128]> var_7928_cast_fp16 = matmul(transpose_x = var_7928_transpose_x_0, transpose_y = var_7928_transpose_y_0, x = transpose_399, y = transpose_395)[name = string("op_7928_cast_fp16")];
            fp16 var_7929_to_fp16 = const()[name = string("op_7929_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_7930_cast_fp16 = mul(x = var_7928_cast_fp16, y = var_7929_to_fp16)[name = string("op_7930_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_91_cast_fp16 = add(x = var_7930_cast_fp16, y = attention_mask)[name = string("scores_91_cast_fp16")];
            int32 var_7933 = const()[name = string("op_7933"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_91_cast_fp16 = softmax(axis = var_7933, x = scores_91_cast_fp16)[name = string("probs_91_cast_fp16")];
            bool var_7942_transpose_x_0 = const()[name = string("op_7942_transpose_x_0"), val = bool(false)];
            bool var_7942_transpose_y_0 = const()[name = string("op_7942_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_7942_cast_fp16 = matmul(transpose_x = var_7942_transpose_x_0, transpose_y = var_7942_transpose_y_0, x = probs_91_cast_fp16, y = v_t_89_cast_fp16)[name = string("op_7942_cast_fp16")];
            bool var_7961_transpose_x_0 = const()[name = string("op_7961_transpose_x_0"), val = bool(false)];
            bool var_7961_transpose_y_0 = const()[name = string("op_7961_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_400_perm_0 = const()[name = string("transpose_400_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_400 = transpose(perm = transpose_400_perm_0, x = q_h_93_cast_fp16)[name = string("transpose_494")];
            tensor<fp16, [1, 1, ?, 128]> var_7961_cast_fp16 = matmul(transpose_x = var_7961_transpose_x_0, transpose_y = var_7961_transpose_y_0, x = transpose_400, y = transpose_395)[name = string("op_7961_cast_fp16")];
            fp16 var_7962_to_fp16 = const()[name = string("op_7962_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_7963_cast_fp16 = mul(x = var_7961_cast_fp16, y = var_7962_to_fp16)[name = string("op_7963_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_93_cast_fp16 = add(x = var_7963_cast_fp16, y = attention_mask)[name = string("scores_93_cast_fp16")];
            int32 var_7966 = const()[name = string("op_7966"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_93_cast_fp16 = softmax(axis = var_7966, x = scores_93_cast_fp16)[name = string("probs_93_cast_fp16")];
            bool var_7975_transpose_x_0 = const()[name = string("op_7975_transpose_x_0"), val = bool(false)];
            bool var_7975_transpose_y_0 = const()[name = string("op_7975_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_7975_cast_fp16 = matmul(transpose_x = var_7975_transpose_x_0, transpose_y = var_7975_transpose_y_0, x = probs_93_cast_fp16, y = v_t_89_cast_fp16)[name = string("op_7975_cast_fp16")];
            bool var_7994_transpose_x_0 = const()[name = string("op_7994_transpose_x_0"), val = bool(false)];
            bool var_7994_transpose_y_0 = const()[name = string("op_7994_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_401_perm_0 = const()[name = string("transpose_401_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_401 = transpose(perm = transpose_401_perm_0, x = q_h_95_cast_fp16)[name = string("transpose_493")];
            tensor<fp16, [1, 1, ?, 128]> var_7994_cast_fp16 = matmul(transpose_x = var_7994_transpose_x_0, transpose_y = var_7994_transpose_y_0, x = transpose_401, y = transpose_395)[name = string("op_7994_cast_fp16")];
            fp16 var_7995_to_fp16 = const()[name = string("op_7995_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_7996_cast_fp16 = mul(x = var_7994_cast_fp16, y = var_7995_to_fp16)[name = string("op_7996_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_95_cast_fp16 = add(x = var_7996_cast_fp16, y = attention_mask)[name = string("scores_95_cast_fp16")];
            int32 var_7999 = const()[name = string("op_7999"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_95_cast_fp16 = softmax(axis = var_7999, x = scores_95_cast_fp16)[name = string("probs_95_cast_fp16")];
            bool var_8008_transpose_x_0 = const()[name = string("op_8008_transpose_x_0"), val = bool(false)];
            bool var_8008_transpose_y_0 = const()[name = string("op_8008_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_8008_cast_fp16 = matmul(transpose_x = var_8008_transpose_x_0, transpose_y = var_8008_transpose_y_0, x = probs_95_cast_fp16, y = v_t_89_cast_fp16)[name = string("op_8008_cast_fp16")];
            bool input_91_interleave_0 = const()[name = string("input_91_interleave_0"), val = bool(false)];
            int32 const_156 = const()[name = string("const_156"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_91_cast_fp16 = concat(axis = const_156, interleave = input_91_interleave_0, values = (var_7909_cast_fp16, var_7942_cast_fp16, var_7975_cast_fp16, var_8008_cast_fp16))[name = string("input_91_cast_fp16")];
            string x_331_pad_type_0 = const()[name = string("x_331_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_331_strides_0 = const()[name = string("x_331_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_331_pad_0 = const()[name = string("x_331_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_331_dilations_0 = const()[name = string("x_331_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_331_groups_0 = const()[name = string("x_331_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_11_weight_to_fp16 = const()[name = string("o_proj_11_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460142272)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_107_cast_fp16 = transpose(perm = transpose_107_perm_0, x = input_91_cast_fp16)[name = string("transpose_492")];
            tensor<fp16, [1, 640, 1, ?]> x_331_cast_fp16 = conv(dilations = x_331_dilations_0, groups = x_331_groups_0, pad = x_331_pad_0, pad_type = x_331_pad_type_0, strides = x_331_strides_0, weight = o_proj_11_weight_to_fp16, x = transpose_107_cast_fp16)[name = string("x_331_cast_fp16")];
            fp16 var_8029_promoted_to_fp16 = const()[name = string("op_8029_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_8030_cast_fp16 = pow(x = x_331_cast_fp16, y = var_8029_promoted_to_fp16)[name = string("op_8030_cast_fp16")];
            tensor<int32, [1]> var_211_axes_0 = const()[name = string("var_211_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_211_keep_dims_0 = const()[name = string("var_211_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_211_cast_fp16 = reduce_mean(axes = var_211_axes_0, keep_dims = var_211_keep_dims_0, x = var_8030_cast_fp16)[name = string("var_211_cast_fp16")];
            fp16 var_8037_to_fp16 = const()[name = string("op_8037_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_8038_cast_fp16 = add(x = var_211_cast_fp16, y = var_8037_to_fp16)[name = string("op_8038_cast_fp16")];
            fp32 var_8039_epsilon_0 = const()[name = string("op_8039_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_8039_cast_fp16 = rsqrt(epsilon = var_8039_epsilon_0, x = var_8038_cast_fp16)[name = string("op_8039_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_211_cast_fp16 = mul(x = x_331_cast_fp16, y = var_8039_cast_fp16)[name = string("x_normed_211_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_8049_to_fp16 = const()[name = string("op_8049_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461453056)))];
            tensor<fp16, [1, 640, 1, ?]> attn_47_cast_fp16 = mul(x = x_normed_211_cast_fp16, y = var_8049_to_fp16)[name = string("attn_47_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_333_cast_fp16 = add(x = x_309_cast_fp16, y = attn_47_cast_fp16)[name = string("x_333_cast_fp16")];
            fp16 var_8053_promoted_to_fp16 = const()[name = string("op_8053_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_8054_cast_fp16 = pow(x = x_333_cast_fp16, y = var_8053_promoted_to_fp16)[name = string("op_8054_cast_fp16")];
            tensor<int32, [1]> var_213_axes_0 = const()[name = string("var_213_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_213_keep_dims_0 = const()[name = string("var_213_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_213_cast_fp16 = reduce_mean(axes = var_213_axes_0, keep_dims = var_213_keep_dims_0, x = var_8054_cast_fp16)[name = string("var_213_cast_fp16")];
            fp16 var_8061_to_fp16 = const()[name = string("op_8061_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_8062_cast_fp16 = add(x = var_213_cast_fp16, y = var_8061_to_fp16)[name = string("op_8062_cast_fp16")];
            fp32 var_8063_epsilon_0 = const()[name = string("op_8063_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_8063_cast_fp16 = rsqrt(epsilon = var_8063_epsilon_0, x = var_8062_cast_fp16)[name = string("op_8063_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_213_cast_fp16 = mul(x = x_333_cast_fp16, y = var_8063_cast_fp16)[name = string("x_normed_213_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_8073_to_fp16 = const()[name = string("op_8073_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461454400)))];
            tensor<fp16, [1, 640, 1, ?]> input_93_cast_fp16 = mul(x = x_normed_213_cast_fp16, y = var_8073_to_fp16)[name = string("input_93_cast_fp16")];
            string gate_23_pad_type_0 = const()[name = string("gate_23_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_23_strides_0 = const()[name = string("gate_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_23_pad_0 = const()[name = string("gate_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_23_dilations_0 = const()[name = string("gate_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_23_groups_0 = const()[name = string("gate_23_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_11_weight_to_fp16 = const()[name = string("gate_proj_11_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461455744)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_23_cast_fp16 = conv(dilations = gate_23_dilations_0, groups = gate_23_groups_0, pad = gate_23_pad_0, pad_type = gate_23_pad_type_0, strides = gate_23_strides_0, weight = gate_proj_11_weight_to_fp16, x = input_93_cast_fp16)[name = string("gate_23_cast_fp16")];
            string up_23_pad_type_0 = const()[name = string("up_23_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_23_strides_0 = const()[name = string("up_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_23_pad_0 = const()[name = string("up_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_23_dilations_0 = const()[name = string("up_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_23_groups_0 = const()[name = string("up_23_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_11_weight_to_fp16 = const()[name = string("up_proj_11_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464077248)))];
            tensor<fp16, [1, 2048, 1, ?]> up_23_cast_fp16 = conv(dilations = up_23_dilations_0, groups = up_23_groups_0, pad = up_23_pad_0, pad_type = up_23_pad_type_0, strides = up_23_strides_0, weight = up_proj_11_weight_to_fp16, x = input_93_cast_fp16)[name = string("up_23_cast_fp16")];
            string var_8098_mode_0 = const()[name = string("op_8098_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_8098_cast_fp16 = gelu(mode = var_8098_mode_0, x = gate_23_cast_fp16)[name = string("op_8098_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_95_cast_fp16 = mul(x = var_8098_cast_fp16, y = up_23_cast_fp16)[name = string("input_95_cast_fp16")];
            string x_335_pad_type_0 = const()[name = string("x_335_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_335_strides_0 = const()[name = string("x_335_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_335_pad_0 = const()[name = string("x_335_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_335_dilations_0 = const()[name = string("x_335_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_335_groups_0 = const()[name = string("x_335_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_11_weight_to_fp16 = const()[name = string("down_proj_11_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466698752)))];
            tensor<fp16, [1, 640, 1, ?]> x_335_cast_fp16 = conv(dilations = x_335_dilations_0, groups = x_335_groups_0, pad = x_335_pad_0, pad_type = x_335_pad_type_0, strides = x_335_strides_0, weight = down_proj_11_weight_to_fp16, x = input_95_cast_fp16)[name = string("x_335_cast_fp16")];
            fp16 var_8111_promoted_to_fp16 = const()[name = string("op_8111_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_8112_cast_fp16 = pow(x = x_335_cast_fp16, y = var_8111_promoted_to_fp16)[name = string("op_8112_cast_fp16")];
            tensor<int32, [1]> var_215_axes_0 = const()[name = string("var_215_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_215_keep_dims_0 = const()[name = string("var_215_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_215_cast_fp16 = reduce_mean(axes = var_215_axes_0, keep_dims = var_215_keep_dims_0, x = var_8112_cast_fp16)[name = string("var_215_cast_fp16")];
            fp16 var_8119_to_fp16 = const()[name = string("op_8119_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_8120_cast_fp16 = add(x = var_215_cast_fp16, y = var_8119_to_fp16)[name = string("op_8120_cast_fp16")];
            fp32 var_8121_epsilon_0 = const()[name = string("op_8121_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_8121_cast_fp16 = rsqrt(epsilon = var_8121_epsilon_0, x = var_8120_cast_fp16)[name = string("op_8121_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_215_cast_fp16 = mul(x = x_335_cast_fp16, y = var_8121_cast_fp16)[name = string("x_normed_215_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_8131_to_fp16 = const()[name = string("op_8131_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469320256)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_23_cast_fp16 = mul(x = x_normed_215_cast_fp16, y = var_8131_to_fp16)[name = string("mlp_23_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_337_cast_fp16 = add(x = x_333_cast_fp16, y = mlp_23_cast_fp16)[name = string("x_337_cast_fp16")];
            fp16 var_8135_promoted_to_fp16 = const()[name = string("op_8135_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_8136_cast_fp16 = pow(x = x_337_cast_fp16, y = var_8135_promoted_to_fp16)[name = string("op_8136_cast_fp16")];
            tensor<int32, [1]> var_217_axes_0 = const()[name = string("var_217_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_217_keep_dims_0 = const()[name = string("var_217_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_217_cast_fp16 = reduce_mean(axes = var_217_axes_0, keep_dims = var_217_keep_dims_0, x = var_8136_cast_fp16)[name = string("var_217_cast_fp16")];
            fp16 var_8143_to_fp16 = const()[name = string("op_8143_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_8144_cast_fp16 = add(x = var_217_cast_fp16, y = var_8143_to_fp16)[name = string("op_8144_cast_fp16")];
            fp32 var_8145_epsilon_0 = const()[name = string("op_8145_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_8145_cast_fp16 = rsqrt(epsilon = var_8145_epsilon_0, x = var_8144_cast_fp16)[name = string("op_8145_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_217_cast_fp16 = mul(x = x_337_cast_fp16, y = var_8145_cast_fp16)[name = string("x_normed_217_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_8155_to_fp16 = const()[name = string("op_8155_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469321600)))];
            tensor<fp16, [1, 640, 1, ?]> input_97_cast_fp16 = mul(x = x_normed_217_cast_fp16, y = var_8155_to_fp16)[name = string("input_97_cast_fp16")];
            string var_8167_pad_type_0 = const()[name = string("op_8167_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_8167_strides_0 = const()[name = string("op_8167_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8167_pad_0 = const()[name = string("op_8167_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8167_dilations_0 = const()[name = string("op_8167_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_8167_groups_0 = const()[name = string("op_8167_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_12_weight_to_fp16 = const()[name = string("q_proj_12_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469322944)))];
            tensor<fp16, [1, 1024, 1, ?]> var_8167_cast_fp16 = conv(dilations = var_8167_dilations_0, groups = var_8167_groups_0, pad = var_8167_pad_0, pad_type = var_8167_pad_type_0, strides = var_8167_strides_0, weight = q_proj_12_weight_to_fp16, x = input_97_cast_fp16)[name = string("op_8167_cast_fp16")];
            string var_8178_pad_type_0 = const()[name = string("op_8178_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_8178_strides_0 = const()[name = string("op_8178_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8178_pad_0 = const()[name = string("op_8178_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8178_dilations_0 = const()[name = string("op_8178_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_8178_groups_0 = const()[name = string("op_8178_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_12_weight_to_fp16 = const()[name = string("k_proj_12_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470633728)))];
            tensor<fp16, [1, 256, 1, ?]> var_8178_cast_fp16 = conv(dilations = var_8178_dilations_0, groups = var_8178_groups_0, pad = var_8178_pad_0, pad_type = var_8178_pad_type_0, strides = var_8178_strides_0, weight = k_proj_12_weight_to_fp16, x = input_97_cast_fp16)[name = string("op_8178_cast_fp16")];
            string var_8189_pad_type_0 = const()[name = string("op_8189_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_8189_strides_0 = const()[name = string("op_8189_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8189_pad_0 = const()[name = string("op_8189_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8189_dilations_0 = const()[name = string("op_8189_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_8189_groups_0 = const()[name = string("op_8189_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_12_weight_to_fp16 = const()[name = string("v_proj_12_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470961472)))];
            tensor<fp16, [1, 256, 1, ?]> var_8189_cast_fp16 = conv(dilations = var_8189_dilations_0, groups = var_8189_groups_0, pad = var_8189_pad_0, pad_type = var_8189_pad_type_0, strides = var_8189_strides_0, weight = v_proj_12_weight_to_fp16, x = input_97_cast_fp16)[name = string("op_8189_cast_fp16")];
            tensor<int32, [4]> tile_36 = const()[name = string("tile_36"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_8192_axis_0 = const()[name = string("op_8192_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_8192_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_8192_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_8192_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_8192_cast_fp16_3 = split(axis = var_8192_axis_0, split_sizes = tile_36, x = var_8167_cast_fp16)[name = string("op_8192_cast_fp16")];
            fp16 var_8205_promoted_to_fp16 = const()[name = string("op_8205_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_8206_cast_fp16 = pow(x = var_8192_cast_fp16_0, y = var_8205_promoted_to_fp16)[name = string("op_8206_cast_fp16")];
            tensor<int32, [1]> var_219_axes_0 = const()[name = string("var_219_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_219_keep_dims_0 = const()[name = string("var_219_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_219_cast_fp16 = reduce_mean(axes = var_219_axes_0, keep_dims = var_219_keep_dims_0, x = var_8206_cast_fp16)[name = string("var_219_cast_fp16")];
            fp16 var_8213_to_fp16 = const()[name = string("op_8213_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_8214_cast_fp16 = add(x = var_219_cast_fp16, y = var_8213_to_fp16)[name = string("op_8214_cast_fp16")];
            fp32 var_8215_epsilon_0 = const()[name = string("op_8215_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_8215_cast_fp16 = rsqrt(epsilon = var_8215_epsilon_0, x = var_8214_cast_fp16)[name = string("op_8215_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_219_cast_fp16 = mul(x = var_8192_cast_fp16_0, y = var_8215_cast_fp16)[name = string("x_normed_219_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_8225_to_fp16 = const()[name = string("op_8225_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471289216)))];
            tensor<fp16, [1, 256, 1, ?]> x_341_cast_fp16 = mul(x = x_normed_219_cast_fp16, y = var_8225_to_fp16)[name = string("x_341_cast_fp16")];
            tensor<int32, [4]> x_first_121_begin_0 = const()[name = string("x_first_121_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_121_end_0 = const()[name = string("x_first_121_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_121_end_mask_0 = const()[name = string("x_first_121_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_121_cast_fp16 = slice_by_index(begin = x_first_121_begin_0, end = x_first_121_end_0, end_mask = x_first_121_end_mask_0, x = x_341_cast_fp16)[name = string("x_first_121_cast_fp16")];
            tensor<int32, [4]> x_second_121_begin_0 = const()[name = string("x_second_121_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_121_end_0 = const()[name = string("x_second_121_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_121_end_mask_0 = const()[name = string("x_second_121_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_121_cast_fp16 = slice_by_index(begin = x_second_121_begin_0, end = x_second_121_end_0, end_mask = x_second_121_end_mask_0, x = x_341_cast_fp16)[name = string("x_second_121_cast_fp16")];
            fp16 const_100_promoted_to_fp16 = const()[name = string("const_100_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_8247_cast_fp16 = mul(x = x_second_121_cast_fp16, y = const_100_promoted_to_fp16)[name = string("op_8247_cast_fp16")];
            int32 var_8249 = const()[name = string("op_8249"), val = int32(1)];
            bool x_rot_121_interleave_0 = const()[name = string("x_rot_121_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_121_cast_fp16 = concat(axis = var_8249, interleave = x_rot_121_interleave_0, values = (var_8247_cast_fp16, x_first_121_cast_fp16))[name = string("x_rot_121_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_8265_cast_fp16 = mul(x = x_341_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_8265_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_8266_cast_fp16 = mul(x = x_rot_121_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_8266_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_97_cast_fp16 = add(x = var_8265_cast_fp16, y = var_8266_cast_fp16)[name = string("q_h_97_cast_fp16")];
            fp16 var_8269_promoted_to_fp16 = const()[name = string("op_8269_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_8270_cast_fp16 = pow(x = var_8192_cast_fp16_1, y = var_8269_promoted_to_fp16)[name = string("op_8270_cast_fp16")];
            tensor<int32, [1]> var_221_axes_0 = const()[name = string("var_221_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_221_keep_dims_0 = const()[name = string("var_221_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_221_cast_fp16 = reduce_mean(axes = var_221_axes_0, keep_dims = var_221_keep_dims_0, x = var_8270_cast_fp16)[name = string("var_221_cast_fp16")];
            fp16 var_8277_to_fp16 = const()[name = string("op_8277_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_8278_cast_fp16 = add(x = var_221_cast_fp16, y = var_8277_to_fp16)[name = string("op_8278_cast_fp16")];
            fp32 var_8279_epsilon_0 = const()[name = string("op_8279_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_8279_cast_fp16 = rsqrt(epsilon = var_8279_epsilon_0, x = var_8278_cast_fp16)[name = string("op_8279_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_221_cast_fp16 = mul(x = var_8192_cast_fp16_1, y = var_8279_cast_fp16)[name = string("x_normed_221_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_345_cast_fp16 = mul(x = x_normed_221_cast_fp16, y = var_8225_to_fp16)[name = string("x_345_cast_fp16")];
            tensor<int32, [4]> x_first_123_begin_0 = const()[name = string("x_first_123_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_123_end_0 = const()[name = string("x_first_123_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_123_end_mask_0 = const()[name = string("x_first_123_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_123_cast_fp16 = slice_by_index(begin = x_first_123_begin_0, end = x_first_123_end_0, end_mask = x_first_123_end_mask_0, x = x_345_cast_fp16)[name = string("x_first_123_cast_fp16")];
            tensor<int32, [4]> x_second_123_begin_0 = const()[name = string("x_second_123_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_123_end_0 = const()[name = string("x_second_123_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_123_end_mask_0 = const()[name = string("x_second_123_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_123_cast_fp16 = slice_by_index(begin = x_second_123_begin_0, end = x_second_123_end_0, end_mask = x_second_123_end_mask_0, x = x_345_cast_fp16)[name = string("x_second_123_cast_fp16")];
            fp16 const_101_promoted_to_fp16 = const()[name = string("const_101_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_8311_cast_fp16 = mul(x = x_second_123_cast_fp16, y = const_101_promoted_to_fp16)[name = string("op_8311_cast_fp16")];
            int32 var_8313 = const()[name = string("op_8313"), val = int32(1)];
            bool x_rot_123_interleave_0 = const()[name = string("x_rot_123_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_123_cast_fp16 = concat(axis = var_8313, interleave = x_rot_123_interleave_0, values = (var_8311_cast_fp16, x_first_123_cast_fp16))[name = string("x_rot_123_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_8329_cast_fp16 = mul(x = x_345_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_8329_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_8330_cast_fp16 = mul(x = x_rot_123_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_8330_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_99_cast_fp16 = add(x = var_8329_cast_fp16, y = var_8330_cast_fp16)[name = string("q_h_99_cast_fp16")];
            fp16 var_8333_promoted_to_fp16 = const()[name = string("op_8333_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_8334_cast_fp16 = pow(x = var_8192_cast_fp16_2, y = var_8333_promoted_to_fp16)[name = string("op_8334_cast_fp16")];
            tensor<int32, [1]> var_223_axes_0 = const()[name = string("var_223_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_223_keep_dims_0 = const()[name = string("var_223_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_223_cast_fp16 = reduce_mean(axes = var_223_axes_0, keep_dims = var_223_keep_dims_0, x = var_8334_cast_fp16)[name = string("var_223_cast_fp16")];
            fp16 var_8341_to_fp16 = const()[name = string("op_8341_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_8342_cast_fp16 = add(x = var_223_cast_fp16, y = var_8341_to_fp16)[name = string("op_8342_cast_fp16")];
            fp32 var_8343_epsilon_0 = const()[name = string("op_8343_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_8343_cast_fp16 = rsqrt(epsilon = var_8343_epsilon_0, x = var_8342_cast_fp16)[name = string("op_8343_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_223_cast_fp16 = mul(x = var_8192_cast_fp16_2, y = var_8343_cast_fp16)[name = string("x_normed_223_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_349_cast_fp16 = mul(x = x_normed_223_cast_fp16, y = var_8225_to_fp16)[name = string("x_349_cast_fp16")];
            tensor<int32, [4]> x_first_125_begin_0 = const()[name = string("x_first_125_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_125_end_0 = const()[name = string("x_first_125_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_125_end_mask_0 = const()[name = string("x_first_125_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_125_cast_fp16 = slice_by_index(begin = x_first_125_begin_0, end = x_first_125_end_0, end_mask = x_first_125_end_mask_0, x = x_349_cast_fp16)[name = string("x_first_125_cast_fp16")];
            tensor<int32, [4]> x_second_125_begin_0 = const()[name = string("x_second_125_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_125_end_0 = const()[name = string("x_second_125_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_125_end_mask_0 = const()[name = string("x_second_125_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_125_cast_fp16 = slice_by_index(begin = x_second_125_begin_0, end = x_second_125_end_0, end_mask = x_second_125_end_mask_0, x = x_349_cast_fp16)[name = string("x_second_125_cast_fp16")];
            fp16 const_102_promoted_to_fp16 = const()[name = string("const_102_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_8375_cast_fp16 = mul(x = x_second_125_cast_fp16, y = const_102_promoted_to_fp16)[name = string("op_8375_cast_fp16")];
            int32 var_8377 = const()[name = string("op_8377"), val = int32(1)];
            bool x_rot_125_interleave_0 = const()[name = string("x_rot_125_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_125_cast_fp16 = concat(axis = var_8377, interleave = x_rot_125_interleave_0, values = (var_8375_cast_fp16, x_first_125_cast_fp16))[name = string("x_rot_125_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_8393_cast_fp16 = mul(x = x_349_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_8393_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_8394_cast_fp16 = mul(x = x_rot_125_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_8394_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_101_cast_fp16 = add(x = var_8393_cast_fp16, y = var_8394_cast_fp16)[name = string("q_h_101_cast_fp16")];
            fp16 var_8397_promoted_to_fp16 = const()[name = string("op_8397_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_8398_cast_fp16 = pow(x = var_8192_cast_fp16_3, y = var_8397_promoted_to_fp16)[name = string("op_8398_cast_fp16")];
            tensor<int32, [1]> var_225_axes_0 = const()[name = string("var_225_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_225_keep_dims_0 = const()[name = string("var_225_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_225_cast_fp16 = reduce_mean(axes = var_225_axes_0, keep_dims = var_225_keep_dims_0, x = var_8398_cast_fp16)[name = string("var_225_cast_fp16")];
            fp16 var_8405_to_fp16 = const()[name = string("op_8405_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_8406_cast_fp16 = add(x = var_225_cast_fp16, y = var_8405_to_fp16)[name = string("op_8406_cast_fp16")];
            fp32 var_8407_epsilon_0 = const()[name = string("op_8407_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_8407_cast_fp16 = rsqrt(epsilon = var_8407_epsilon_0, x = var_8406_cast_fp16)[name = string("op_8407_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_225_cast_fp16 = mul(x = var_8192_cast_fp16_3, y = var_8407_cast_fp16)[name = string("x_normed_225_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_353_cast_fp16 = mul(x = x_normed_225_cast_fp16, y = var_8225_to_fp16)[name = string("x_353_cast_fp16")];
            tensor<int32, [4]> x_first_127_begin_0 = const()[name = string("x_first_127_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_127_end_0 = const()[name = string("x_first_127_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_127_end_mask_0 = const()[name = string("x_first_127_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_127_cast_fp16 = slice_by_index(begin = x_first_127_begin_0, end = x_first_127_end_0, end_mask = x_first_127_end_mask_0, x = x_353_cast_fp16)[name = string("x_first_127_cast_fp16")];
            tensor<int32, [4]> x_second_127_begin_0 = const()[name = string("x_second_127_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_127_end_0 = const()[name = string("x_second_127_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_127_end_mask_0 = const()[name = string("x_second_127_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_127_cast_fp16 = slice_by_index(begin = x_second_127_begin_0, end = x_second_127_end_0, end_mask = x_second_127_end_mask_0, x = x_353_cast_fp16)[name = string("x_second_127_cast_fp16")];
            fp16 const_103_promoted_to_fp16 = const()[name = string("const_103_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_8439_cast_fp16 = mul(x = x_second_127_cast_fp16, y = const_103_promoted_to_fp16)[name = string("op_8439_cast_fp16")];
            int32 var_8441 = const()[name = string("op_8441"), val = int32(1)];
            bool x_rot_127_interleave_0 = const()[name = string("x_rot_127_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_127_cast_fp16 = concat(axis = var_8441, interleave = x_rot_127_interleave_0, values = (var_8439_cast_fp16, x_first_127_cast_fp16))[name = string("x_rot_127_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_8457_cast_fp16 = mul(x = x_353_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_8457_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_8458_cast_fp16 = mul(x = x_rot_127_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_8458_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_103_cast_fp16 = add(x = var_8457_cast_fp16, y = var_8458_cast_fp16)[name = string("q_h_103_cast_fp16")];
            fp16 var_8461_promoted_to_fp16 = const()[name = string("op_8461_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_8462_cast_fp16 = pow(x = var_8178_cast_fp16, y = var_8461_promoted_to_fp16)[name = string("op_8462_cast_fp16")];
            tensor<int32, [1]> var_227_axes_0 = const()[name = string("var_227_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_227_keep_dims_0 = const()[name = string("var_227_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_227_cast_fp16 = reduce_mean(axes = var_227_axes_0, keep_dims = var_227_keep_dims_0, x = var_8462_cast_fp16)[name = string("var_227_cast_fp16")];
            fp16 var_8469_to_fp16 = const()[name = string("op_8469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_8470_cast_fp16 = add(x = var_227_cast_fp16, y = var_8469_to_fp16)[name = string("op_8470_cast_fp16")];
            fp32 var_8471_epsilon_0 = const()[name = string("op_8471_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_8471_cast_fp16 = rsqrt(epsilon = var_8471_epsilon_0, x = var_8470_cast_fp16)[name = string("op_8471_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_227_cast_fp16 = mul(x = var_8178_cast_fp16, y = var_8471_cast_fp16)[name = string("x_normed_227_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_8481_to_fp16 = const()[name = string("op_8481_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471289792)))];
            tensor<fp16, [1, 256, 1, ?]> x_357_cast_fp16 = mul(x = x_normed_227_cast_fp16, y = var_8481_to_fp16)[name = string("x_357_cast_fp16")];
            tensor<int32, [4]> x_first_129_begin_0 = const()[name = string("x_first_129_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_129_end_0 = const()[name = string("x_first_129_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_129_end_mask_0 = const()[name = string("x_first_129_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_129_cast_fp16 = slice_by_index(begin = x_first_129_begin_0, end = x_first_129_end_0, end_mask = x_first_129_end_mask_0, x = x_357_cast_fp16)[name = string("x_first_129_cast_fp16")];
            tensor<int32, [4]> x_second_129_begin_0 = const()[name = string("x_second_129_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_129_end_0 = const()[name = string("x_second_129_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_129_end_mask_0 = const()[name = string("x_second_129_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_129_cast_fp16 = slice_by_index(begin = x_second_129_begin_0, end = x_second_129_end_0, end_mask = x_second_129_end_mask_0, x = x_357_cast_fp16)[name = string("x_second_129_cast_fp16")];
            fp16 const_104_promoted_to_fp16 = const()[name = string("const_104_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_8503_cast_fp16 = mul(x = x_second_129_cast_fp16, y = const_104_promoted_to_fp16)[name = string("op_8503_cast_fp16")];
            int32 var_8505 = const()[name = string("op_8505"), val = int32(1)];
            bool x_rot_129_interleave_0 = const()[name = string("x_rot_129_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_129_cast_fp16 = concat(axis = var_8505, interleave = x_rot_129_interleave_0, values = (var_8503_cast_fp16, x_first_129_cast_fp16))[name = string("x_rot_129_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_8521_cast_fp16 = mul(x = x_357_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_8521_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_8522_cast_fp16 = mul(x = x_rot_129_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_8522_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_25_cast_fp16 = add(x = var_8521_cast_fp16, y = var_8522_cast_fp16)[name = string("k_new_25_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_24 = read_state(input = k_cache_12)[name = string("read_state_24")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_75_cast_fp16 = mul(x = read_state_24, y = one_minus_1_cast_fp16)[name = string("k_cache_75_cast_fp16")];
            write_state(data = k_cache_75_cast_fp16, input = k_cache_12)[name = string("coreml_update_state_120_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_120 = read_state(input = k_cache_12)[name = string("coreml_update_state_120")];
            tensor<fp16, [1, 256, 1, 128]> var_8529_cast_fp16 = mul(x = k_new_25_cast_fp16, y = write_mask)[name = string("op_8529_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_77_cast_fp16 = add(x = coreml_update_state_120, y = var_8529_cast_fp16)[name = string("k_cache_77_cast_fp16")];
            write_state(data = k_cache_77_cast_fp16, input = k_cache_12)[name = string("coreml_update_state_121_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_121 = read_state(input = k_cache_12)[name = string("coreml_update_state_121")];
            tensor<fp16, [1, 256, 1, 128]> read_state_25 = read_state(input = v_cache_12)[name = string("read_state_25")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_75_cast_fp16 = mul(x = read_state_25, y = one_minus_1_cast_fp16)[name = string("v_cache_75_cast_fp16")];
            write_state(data = v_cache_75_cast_fp16, input = v_cache_12)[name = string("coreml_update_state_122_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_122 = read_state(input = v_cache_12)[name = string("coreml_update_state_122")];
            tensor<fp16, [1, 256, 1, 128]> var_8533_cast_fp16 = mul(x = var_8189_cast_fp16, y = write_mask)[name = string("op_8533_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_77_cast_fp16 = add(x = coreml_update_state_122, y = var_8533_cast_fp16)[name = string("v_cache_77_cast_fp16")];
            write_state(data = v_cache_77_cast_fp16, input = v_cache_12)[name = string("coreml_update_state_123_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_123 = read_state(input = v_cache_12)[name = string("coreml_update_state_123")];
            bool var_8548_transpose_x_0 = const()[name = string("op_8548_transpose_x_0"), val = bool(false)];
            bool var_8548_transpose_y_0 = const()[name = string("op_8548_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_402_perm_0 = const()[name = string("transpose_402_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_403_perm_0 = const()[name = string("transpose_403_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_403 = transpose(perm = transpose_403_perm_0, x = coreml_update_state_121)[name = string("transpose_490")];
            tensor<fp16, [1, 1, ?, 256]> transpose_402 = transpose(perm = transpose_402_perm_0, x = q_h_97_cast_fp16)[name = string("transpose_491")];
            tensor<fp16, [1, 1, ?, 128]> var_8548_cast_fp16 = matmul(transpose_x = var_8548_transpose_x_0, transpose_y = var_8548_transpose_y_0, x = transpose_402, y = transpose_403)[name = string("op_8548_cast_fp16")];
            fp16 var_8549_to_fp16 = const()[name = string("op_8549_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_8550_cast_fp16 = mul(x = var_8548_cast_fp16, y = var_8549_to_fp16)[name = string("op_8550_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_97_cast_fp16 = add(x = var_8550_cast_fp16, y = attention_mask)[name = string("scores_97_cast_fp16")];
            int32 var_8553 = const()[name = string("op_8553"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_97_cast_fp16 = softmax(axis = var_8553, x = scores_97_cast_fp16)[name = string("probs_97_cast_fp16")];
            tensor<int32, [4]> var_8560 = const()[name = string("op_8560"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_8562_transpose_x_0 = const()[name = string("op_8562_transpose_x_0"), val = bool(false)];
            bool var_8562_transpose_y_0 = const()[name = string("op_8562_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_97_cast_fp16 = transpose(perm = var_8560, x = coreml_update_state_123)[name = string("transpose_489")];
            tensor<fp16, [1, 1, ?, 256]> var_8562_cast_fp16 = matmul(transpose_x = var_8562_transpose_x_0, transpose_y = var_8562_transpose_y_0, x = probs_97_cast_fp16, y = v_t_97_cast_fp16)[name = string("op_8562_cast_fp16")];
            bool var_8581_transpose_x_0 = const()[name = string("op_8581_transpose_x_0"), val = bool(false)];
            bool var_8581_transpose_y_0 = const()[name = string("op_8581_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_407_perm_0 = const()[name = string("transpose_407_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_407 = transpose(perm = transpose_407_perm_0, x = q_h_99_cast_fp16)[name = string("transpose_488")];
            tensor<fp16, [1, 1, ?, 128]> var_8581_cast_fp16 = matmul(transpose_x = var_8581_transpose_x_0, transpose_y = var_8581_transpose_y_0, x = transpose_407, y = transpose_403)[name = string("op_8581_cast_fp16")];
            fp16 var_8582_to_fp16 = const()[name = string("op_8582_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_8583_cast_fp16 = mul(x = var_8581_cast_fp16, y = var_8582_to_fp16)[name = string("op_8583_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_99_cast_fp16 = add(x = var_8583_cast_fp16, y = attention_mask)[name = string("scores_99_cast_fp16")];
            int32 var_8586 = const()[name = string("op_8586"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_99_cast_fp16 = softmax(axis = var_8586, x = scores_99_cast_fp16)[name = string("probs_99_cast_fp16")];
            bool var_8595_transpose_x_0 = const()[name = string("op_8595_transpose_x_0"), val = bool(false)];
            bool var_8595_transpose_y_0 = const()[name = string("op_8595_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_8595_cast_fp16 = matmul(transpose_x = var_8595_transpose_x_0, transpose_y = var_8595_transpose_y_0, x = probs_99_cast_fp16, y = v_t_97_cast_fp16)[name = string("op_8595_cast_fp16")];
            bool var_8614_transpose_x_0 = const()[name = string("op_8614_transpose_x_0"), val = bool(false)];
            bool var_8614_transpose_y_0 = const()[name = string("op_8614_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_408_perm_0 = const()[name = string("transpose_408_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_408 = transpose(perm = transpose_408_perm_0, x = q_h_101_cast_fp16)[name = string("transpose_487")];
            tensor<fp16, [1, 1, ?, 128]> var_8614_cast_fp16 = matmul(transpose_x = var_8614_transpose_x_0, transpose_y = var_8614_transpose_y_0, x = transpose_408, y = transpose_403)[name = string("op_8614_cast_fp16")];
            fp16 var_8615_to_fp16 = const()[name = string("op_8615_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_8616_cast_fp16 = mul(x = var_8614_cast_fp16, y = var_8615_to_fp16)[name = string("op_8616_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_101_cast_fp16 = add(x = var_8616_cast_fp16, y = attention_mask)[name = string("scores_101_cast_fp16")];
            int32 var_8619 = const()[name = string("op_8619"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_101_cast_fp16 = softmax(axis = var_8619, x = scores_101_cast_fp16)[name = string("probs_101_cast_fp16")];
            bool var_8628_transpose_x_0 = const()[name = string("op_8628_transpose_x_0"), val = bool(false)];
            bool var_8628_transpose_y_0 = const()[name = string("op_8628_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_8628_cast_fp16 = matmul(transpose_x = var_8628_transpose_x_0, transpose_y = var_8628_transpose_y_0, x = probs_101_cast_fp16, y = v_t_97_cast_fp16)[name = string("op_8628_cast_fp16")];
            bool var_8647_transpose_x_0 = const()[name = string("op_8647_transpose_x_0"), val = bool(false)];
            bool var_8647_transpose_y_0 = const()[name = string("op_8647_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_409_perm_0 = const()[name = string("transpose_409_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_409 = transpose(perm = transpose_409_perm_0, x = q_h_103_cast_fp16)[name = string("transpose_486")];
            tensor<fp16, [1, 1, ?, 128]> var_8647_cast_fp16 = matmul(transpose_x = var_8647_transpose_x_0, transpose_y = var_8647_transpose_y_0, x = transpose_409, y = transpose_403)[name = string("op_8647_cast_fp16")];
            fp16 var_8648_to_fp16 = const()[name = string("op_8648_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_8649_cast_fp16 = mul(x = var_8647_cast_fp16, y = var_8648_to_fp16)[name = string("op_8649_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_103_cast_fp16 = add(x = var_8649_cast_fp16, y = attention_mask)[name = string("scores_103_cast_fp16")];
            int32 var_8652 = const()[name = string("op_8652"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_103_cast_fp16 = softmax(axis = var_8652, x = scores_103_cast_fp16)[name = string("probs_103_cast_fp16")];
            bool var_8661_transpose_x_0 = const()[name = string("op_8661_transpose_x_0"), val = bool(false)];
            bool var_8661_transpose_y_0 = const()[name = string("op_8661_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_8661_cast_fp16 = matmul(transpose_x = var_8661_transpose_x_0, transpose_y = var_8661_transpose_y_0, x = probs_103_cast_fp16, y = v_t_97_cast_fp16)[name = string("op_8661_cast_fp16")];
            bool input_99_interleave_0 = const()[name = string("input_99_interleave_0"), val = bool(false)];
            int32 const_157 = const()[name = string("const_157"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_99_cast_fp16 = concat(axis = const_157, interleave = input_99_interleave_0, values = (var_8562_cast_fp16, var_8595_cast_fp16, var_8628_cast_fp16, var_8661_cast_fp16))[name = string("input_99_cast_fp16")];
            string x_359_pad_type_0 = const()[name = string("x_359_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_359_strides_0 = const()[name = string("x_359_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_359_pad_0 = const()[name = string("x_359_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_359_dilations_0 = const()[name = string("x_359_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_359_groups_0 = const()[name = string("x_359_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_12_weight_to_fp16 = const()[name = string("o_proj_12_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471290368)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_116_cast_fp16 = transpose(perm = transpose_116_perm_0, x = input_99_cast_fp16)[name = string("transpose_485")];
            tensor<fp16, [1, 640, 1, ?]> x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = o_proj_12_weight_to_fp16, x = transpose_116_cast_fp16)[name = string("x_359_cast_fp16")];
            fp16 var_8682_promoted_to_fp16 = const()[name = string("op_8682_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_8683_cast_fp16 = pow(x = x_359_cast_fp16, y = var_8682_promoted_to_fp16)[name = string("op_8683_cast_fp16")];
            tensor<int32, [1]> var_229_axes_0 = const()[name = string("var_229_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_229_keep_dims_0 = const()[name = string("var_229_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_229_cast_fp16 = reduce_mean(axes = var_229_axes_0, keep_dims = var_229_keep_dims_0, x = var_8683_cast_fp16)[name = string("var_229_cast_fp16")];
            fp16 var_8690_to_fp16 = const()[name = string("op_8690_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_8691_cast_fp16 = add(x = var_229_cast_fp16, y = var_8690_to_fp16)[name = string("op_8691_cast_fp16")];
            fp32 var_8692_epsilon_0 = const()[name = string("op_8692_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_8692_cast_fp16 = rsqrt(epsilon = var_8692_epsilon_0, x = var_8691_cast_fp16)[name = string("op_8692_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_229_cast_fp16 = mul(x = x_359_cast_fp16, y = var_8692_cast_fp16)[name = string("x_normed_229_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_8702_to_fp16 = const()[name = string("op_8702_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(472601152)))];
            tensor<fp16, [1, 640, 1, ?]> attn_51_cast_fp16 = mul(x = x_normed_229_cast_fp16, y = var_8702_to_fp16)[name = string("attn_51_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_361_cast_fp16 = add(x = x_337_cast_fp16, y = attn_51_cast_fp16)[name = string("x_361_cast_fp16")];
            fp16 var_8706_promoted_to_fp16 = const()[name = string("op_8706_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_8707_cast_fp16 = pow(x = x_361_cast_fp16, y = var_8706_promoted_to_fp16)[name = string("op_8707_cast_fp16")];
            tensor<int32, [1]> var_231_axes_0 = const()[name = string("var_231_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_231_keep_dims_0 = const()[name = string("var_231_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_231_cast_fp16 = reduce_mean(axes = var_231_axes_0, keep_dims = var_231_keep_dims_0, x = var_8707_cast_fp16)[name = string("var_231_cast_fp16")];
            fp16 var_8714_to_fp16 = const()[name = string("op_8714_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_8715_cast_fp16 = add(x = var_231_cast_fp16, y = var_8714_to_fp16)[name = string("op_8715_cast_fp16")];
            fp32 var_8716_epsilon_0 = const()[name = string("op_8716_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_8716_cast_fp16 = rsqrt(epsilon = var_8716_epsilon_0, x = var_8715_cast_fp16)[name = string("op_8716_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_231_cast_fp16 = mul(x = x_361_cast_fp16, y = var_8716_cast_fp16)[name = string("x_normed_231_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_8726_to_fp16 = const()[name = string("op_8726_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(472602496)))];
            tensor<fp16, [1, 640, 1, ?]> input_101_cast_fp16 = mul(x = x_normed_231_cast_fp16, y = var_8726_to_fp16)[name = string("input_101_cast_fp16")];
            string gate_25_pad_type_0 = const()[name = string("gate_25_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_25_strides_0 = const()[name = string("gate_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_25_pad_0 = const()[name = string("gate_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_25_dilations_0 = const()[name = string("gate_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_25_groups_0 = const()[name = string("gate_25_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_12_weight_to_fp16 = const()[name = string("gate_proj_12_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(472603840)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_25_cast_fp16 = conv(dilations = gate_25_dilations_0, groups = gate_25_groups_0, pad = gate_25_pad_0, pad_type = gate_25_pad_type_0, strides = gate_25_strides_0, weight = gate_proj_12_weight_to_fp16, x = input_101_cast_fp16)[name = string("gate_25_cast_fp16")];
            string up_25_pad_type_0 = const()[name = string("up_25_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_25_strides_0 = const()[name = string("up_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_25_pad_0 = const()[name = string("up_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_25_dilations_0 = const()[name = string("up_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_25_groups_0 = const()[name = string("up_25_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_12_weight_to_fp16 = const()[name = string("up_proj_12_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475225344)))];
            tensor<fp16, [1, 2048, 1, ?]> up_25_cast_fp16 = conv(dilations = up_25_dilations_0, groups = up_25_groups_0, pad = up_25_pad_0, pad_type = up_25_pad_type_0, strides = up_25_strides_0, weight = up_proj_12_weight_to_fp16, x = input_101_cast_fp16)[name = string("up_25_cast_fp16")];
            string var_8751_mode_0 = const()[name = string("op_8751_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_8751_cast_fp16 = gelu(mode = var_8751_mode_0, x = gate_25_cast_fp16)[name = string("op_8751_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_103_cast_fp16 = mul(x = var_8751_cast_fp16, y = up_25_cast_fp16)[name = string("input_103_cast_fp16")];
            string x_363_pad_type_0 = const()[name = string("x_363_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_363_strides_0 = const()[name = string("x_363_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_363_pad_0 = const()[name = string("x_363_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_363_dilations_0 = const()[name = string("x_363_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_363_groups_0 = const()[name = string("x_363_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_12_weight_to_fp16 = const()[name = string("down_proj_12_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477846848)))];
            tensor<fp16, [1, 640, 1, ?]> x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = down_proj_12_weight_to_fp16, x = input_103_cast_fp16)[name = string("x_363_cast_fp16")];
            fp16 var_8764_promoted_to_fp16 = const()[name = string("op_8764_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_8765_cast_fp16 = pow(x = x_363_cast_fp16, y = var_8764_promoted_to_fp16)[name = string("op_8765_cast_fp16")];
            tensor<int32, [1]> var_233_axes_0 = const()[name = string("var_233_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_233_keep_dims_0 = const()[name = string("var_233_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_233_cast_fp16 = reduce_mean(axes = var_233_axes_0, keep_dims = var_233_keep_dims_0, x = var_8765_cast_fp16)[name = string("var_233_cast_fp16")];
            fp16 var_8772_to_fp16 = const()[name = string("op_8772_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_8773_cast_fp16 = add(x = var_233_cast_fp16, y = var_8772_to_fp16)[name = string("op_8773_cast_fp16")];
            fp32 var_8774_epsilon_0 = const()[name = string("op_8774_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_8774_cast_fp16 = rsqrt(epsilon = var_8774_epsilon_0, x = var_8773_cast_fp16)[name = string("op_8774_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_233_cast_fp16 = mul(x = x_363_cast_fp16, y = var_8774_cast_fp16)[name = string("x_normed_233_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_8784_to_fp16 = const()[name = string("op_8784_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480468352)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_25_cast_fp16 = mul(x = x_normed_233_cast_fp16, y = var_8784_to_fp16)[name = string("mlp_25_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_365_cast_fp16 = add(x = x_361_cast_fp16, y = mlp_25_cast_fp16)[name = string("x_365_cast_fp16")];
            fp16 var_8788_promoted_to_fp16 = const()[name = string("op_8788_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_8789_cast_fp16 = pow(x = x_365_cast_fp16, y = var_8788_promoted_to_fp16)[name = string("op_8789_cast_fp16")];
            tensor<int32, [1]> var_235_axes_0 = const()[name = string("var_235_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_235_keep_dims_0 = const()[name = string("var_235_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_235_cast_fp16 = reduce_mean(axes = var_235_axes_0, keep_dims = var_235_keep_dims_0, x = var_8789_cast_fp16)[name = string("var_235_cast_fp16")];
            fp16 var_8796_to_fp16 = const()[name = string("op_8796_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_8797_cast_fp16 = add(x = var_235_cast_fp16, y = var_8796_to_fp16)[name = string("op_8797_cast_fp16")];
            fp32 var_8798_epsilon_0 = const()[name = string("op_8798_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_8798_cast_fp16 = rsqrt(epsilon = var_8798_epsilon_0, x = var_8797_cast_fp16)[name = string("op_8798_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_235_cast_fp16 = mul(x = x_365_cast_fp16, y = var_8798_cast_fp16)[name = string("x_normed_235_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_8808_to_fp16 = const()[name = string("op_8808_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480469696)))];
            tensor<fp16, [1, 640, 1, ?]> input_105_cast_fp16 = mul(x = x_normed_235_cast_fp16, y = var_8808_to_fp16)[name = string("input_105_cast_fp16")];
            string var_8820_pad_type_0 = const()[name = string("op_8820_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_8820_strides_0 = const()[name = string("op_8820_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8820_pad_0 = const()[name = string("op_8820_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8820_dilations_0 = const()[name = string("op_8820_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_8820_groups_0 = const()[name = string("op_8820_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_13_weight_to_fp16 = const()[name = string("q_proj_13_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480471040)))];
            tensor<fp16, [1, 1024, 1, ?]> var_8820_cast_fp16 = conv(dilations = var_8820_dilations_0, groups = var_8820_groups_0, pad = var_8820_pad_0, pad_type = var_8820_pad_type_0, strides = var_8820_strides_0, weight = q_proj_13_weight_to_fp16, x = input_105_cast_fp16)[name = string("op_8820_cast_fp16")];
            string var_8831_pad_type_0 = const()[name = string("op_8831_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_8831_strides_0 = const()[name = string("op_8831_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8831_pad_0 = const()[name = string("op_8831_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8831_dilations_0 = const()[name = string("op_8831_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_8831_groups_0 = const()[name = string("op_8831_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_13_weight_to_fp16 = const()[name = string("k_proj_13_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481781824)))];
            tensor<fp16, [1, 256, 1, ?]> var_8831_cast_fp16 = conv(dilations = var_8831_dilations_0, groups = var_8831_groups_0, pad = var_8831_pad_0, pad_type = var_8831_pad_type_0, strides = var_8831_strides_0, weight = k_proj_13_weight_to_fp16, x = input_105_cast_fp16)[name = string("op_8831_cast_fp16")];
            string var_8842_pad_type_0 = const()[name = string("op_8842_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_8842_strides_0 = const()[name = string("op_8842_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_8842_pad_0 = const()[name = string("op_8842_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_8842_dilations_0 = const()[name = string("op_8842_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_8842_groups_0 = const()[name = string("op_8842_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_13_weight_to_fp16 = const()[name = string("v_proj_13_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482109568)))];
            tensor<fp16, [1, 256, 1, ?]> var_8842_cast_fp16 = conv(dilations = var_8842_dilations_0, groups = var_8842_groups_0, pad = var_8842_pad_0, pad_type = var_8842_pad_type_0, strides = var_8842_strides_0, weight = v_proj_13_weight_to_fp16, x = input_105_cast_fp16)[name = string("op_8842_cast_fp16")];
            tensor<int32, [4]> tile_39 = const()[name = string("tile_39"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_8845_axis_0 = const()[name = string("op_8845_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_8845_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_8845_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_8845_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_8845_cast_fp16_3 = split(axis = var_8845_axis_0, split_sizes = tile_39, x = var_8820_cast_fp16)[name = string("op_8845_cast_fp16")];
            fp16 var_8858_promoted_to_fp16 = const()[name = string("op_8858_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_8859_cast_fp16 = pow(x = var_8845_cast_fp16_0, y = var_8858_promoted_to_fp16)[name = string("op_8859_cast_fp16")];
            tensor<int32, [1]> var_237_axes_0 = const()[name = string("var_237_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_237_keep_dims_0 = const()[name = string("var_237_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_237_cast_fp16 = reduce_mean(axes = var_237_axes_0, keep_dims = var_237_keep_dims_0, x = var_8859_cast_fp16)[name = string("var_237_cast_fp16")];
            fp16 var_8866_to_fp16 = const()[name = string("op_8866_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_8867_cast_fp16 = add(x = var_237_cast_fp16, y = var_8866_to_fp16)[name = string("op_8867_cast_fp16")];
            fp32 var_8868_epsilon_0 = const()[name = string("op_8868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_8868_cast_fp16 = rsqrt(epsilon = var_8868_epsilon_0, x = var_8867_cast_fp16)[name = string("op_8868_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_237_cast_fp16 = mul(x = var_8845_cast_fp16_0, y = var_8868_cast_fp16)[name = string("x_normed_237_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_8878_to_fp16 = const()[name = string("op_8878_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482437312)))];
            tensor<fp16, [1, 256, 1, ?]> x_369_cast_fp16 = mul(x = x_normed_237_cast_fp16, y = var_8878_to_fp16)[name = string("x_369_cast_fp16")];
            tensor<int32, [4]> x_first_131_begin_0 = const()[name = string("x_first_131_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_131_end_0 = const()[name = string("x_first_131_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_131_end_mask_0 = const()[name = string("x_first_131_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_131_cast_fp16 = slice_by_index(begin = x_first_131_begin_0, end = x_first_131_end_0, end_mask = x_first_131_end_mask_0, x = x_369_cast_fp16)[name = string("x_first_131_cast_fp16")];
            tensor<int32, [4]> x_second_131_begin_0 = const()[name = string("x_second_131_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_131_end_0 = const()[name = string("x_second_131_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_131_end_mask_0 = const()[name = string("x_second_131_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_131_cast_fp16 = slice_by_index(begin = x_second_131_begin_0, end = x_second_131_end_0, end_mask = x_second_131_end_mask_0, x = x_369_cast_fp16)[name = string("x_second_131_cast_fp16")];
            fp16 const_108_promoted_to_fp16 = const()[name = string("const_108_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_8900_cast_fp16 = mul(x = x_second_131_cast_fp16, y = const_108_promoted_to_fp16)[name = string("op_8900_cast_fp16")];
            int32 var_8902 = const()[name = string("op_8902"), val = int32(1)];
            bool x_rot_131_interleave_0 = const()[name = string("x_rot_131_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_131_cast_fp16 = concat(axis = var_8902, interleave = x_rot_131_interleave_0, values = (var_8900_cast_fp16, x_first_131_cast_fp16))[name = string("x_rot_131_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_8918_cast_fp16 = mul(x = x_369_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_8918_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_8919_cast_fp16 = mul(x = x_rot_131_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_8919_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_105_cast_fp16 = add(x = var_8918_cast_fp16, y = var_8919_cast_fp16)[name = string("q_h_105_cast_fp16")];
            fp16 var_8922_promoted_to_fp16 = const()[name = string("op_8922_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_8923_cast_fp16 = pow(x = var_8845_cast_fp16_1, y = var_8922_promoted_to_fp16)[name = string("op_8923_cast_fp16")];
            tensor<int32, [1]> var_239_axes_0 = const()[name = string("var_239_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_239_keep_dims_0 = const()[name = string("var_239_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_239_cast_fp16 = reduce_mean(axes = var_239_axes_0, keep_dims = var_239_keep_dims_0, x = var_8923_cast_fp16)[name = string("var_239_cast_fp16")];
            fp16 var_8930_to_fp16 = const()[name = string("op_8930_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_8931_cast_fp16 = add(x = var_239_cast_fp16, y = var_8930_to_fp16)[name = string("op_8931_cast_fp16")];
            fp32 var_8932_epsilon_0 = const()[name = string("op_8932_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_8932_cast_fp16 = rsqrt(epsilon = var_8932_epsilon_0, x = var_8931_cast_fp16)[name = string("op_8932_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_239_cast_fp16 = mul(x = var_8845_cast_fp16_1, y = var_8932_cast_fp16)[name = string("x_normed_239_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_373_cast_fp16 = mul(x = x_normed_239_cast_fp16, y = var_8878_to_fp16)[name = string("x_373_cast_fp16")];
            tensor<int32, [4]> x_first_133_begin_0 = const()[name = string("x_first_133_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_133_end_0 = const()[name = string("x_first_133_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_133_end_mask_0 = const()[name = string("x_first_133_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_133_cast_fp16 = slice_by_index(begin = x_first_133_begin_0, end = x_first_133_end_0, end_mask = x_first_133_end_mask_0, x = x_373_cast_fp16)[name = string("x_first_133_cast_fp16")];
            tensor<int32, [4]> x_second_133_begin_0 = const()[name = string("x_second_133_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_133_end_0 = const()[name = string("x_second_133_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_133_end_mask_0 = const()[name = string("x_second_133_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_133_cast_fp16 = slice_by_index(begin = x_second_133_begin_0, end = x_second_133_end_0, end_mask = x_second_133_end_mask_0, x = x_373_cast_fp16)[name = string("x_second_133_cast_fp16")];
            fp16 const_109_promoted_to_fp16 = const()[name = string("const_109_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_8964_cast_fp16 = mul(x = x_second_133_cast_fp16, y = const_109_promoted_to_fp16)[name = string("op_8964_cast_fp16")];
            int32 var_8966 = const()[name = string("op_8966"), val = int32(1)];
            bool x_rot_133_interleave_0 = const()[name = string("x_rot_133_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_133_cast_fp16 = concat(axis = var_8966, interleave = x_rot_133_interleave_0, values = (var_8964_cast_fp16, x_first_133_cast_fp16))[name = string("x_rot_133_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_8982_cast_fp16 = mul(x = x_373_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_8982_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_8983_cast_fp16 = mul(x = x_rot_133_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_8983_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_107_cast_fp16 = add(x = var_8982_cast_fp16, y = var_8983_cast_fp16)[name = string("q_h_107_cast_fp16")];
            fp16 var_8986_promoted_to_fp16 = const()[name = string("op_8986_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_8987_cast_fp16 = pow(x = var_8845_cast_fp16_2, y = var_8986_promoted_to_fp16)[name = string("op_8987_cast_fp16")];
            tensor<int32, [1]> var_241_axes_0 = const()[name = string("var_241_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_241_keep_dims_0 = const()[name = string("var_241_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_241_cast_fp16 = reduce_mean(axes = var_241_axes_0, keep_dims = var_241_keep_dims_0, x = var_8987_cast_fp16)[name = string("var_241_cast_fp16")];
            fp16 var_8994_to_fp16 = const()[name = string("op_8994_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_8995_cast_fp16 = add(x = var_241_cast_fp16, y = var_8994_to_fp16)[name = string("op_8995_cast_fp16")];
            fp32 var_8996_epsilon_0 = const()[name = string("op_8996_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_8996_cast_fp16 = rsqrt(epsilon = var_8996_epsilon_0, x = var_8995_cast_fp16)[name = string("op_8996_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_241_cast_fp16 = mul(x = var_8845_cast_fp16_2, y = var_8996_cast_fp16)[name = string("x_normed_241_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_377_cast_fp16 = mul(x = x_normed_241_cast_fp16, y = var_8878_to_fp16)[name = string("x_377_cast_fp16")];
            tensor<int32, [4]> x_first_135_begin_0 = const()[name = string("x_first_135_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_135_end_0 = const()[name = string("x_first_135_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_135_end_mask_0 = const()[name = string("x_first_135_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_135_cast_fp16 = slice_by_index(begin = x_first_135_begin_0, end = x_first_135_end_0, end_mask = x_first_135_end_mask_0, x = x_377_cast_fp16)[name = string("x_first_135_cast_fp16")];
            tensor<int32, [4]> x_second_135_begin_0 = const()[name = string("x_second_135_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_135_end_0 = const()[name = string("x_second_135_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_135_end_mask_0 = const()[name = string("x_second_135_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_135_cast_fp16 = slice_by_index(begin = x_second_135_begin_0, end = x_second_135_end_0, end_mask = x_second_135_end_mask_0, x = x_377_cast_fp16)[name = string("x_second_135_cast_fp16")];
            fp16 const_110_promoted_to_fp16 = const()[name = string("const_110_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_9028_cast_fp16 = mul(x = x_second_135_cast_fp16, y = const_110_promoted_to_fp16)[name = string("op_9028_cast_fp16")];
            int32 var_9030 = const()[name = string("op_9030"), val = int32(1)];
            bool x_rot_135_interleave_0 = const()[name = string("x_rot_135_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_135_cast_fp16 = concat(axis = var_9030, interleave = x_rot_135_interleave_0, values = (var_9028_cast_fp16, x_first_135_cast_fp16))[name = string("x_rot_135_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_9046_cast_fp16 = mul(x = x_377_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_9046_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_9047_cast_fp16 = mul(x = x_rot_135_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_9047_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_109_cast_fp16 = add(x = var_9046_cast_fp16, y = var_9047_cast_fp16)[name = string("q_h_109_cast_fp16")];
            fp16 var_9050_promoted_to_fp16 = const()[name = string("op_9050_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_9051_cast_fp16 = pow(x = var_8845_cast_fp16_3, y = var_9050_promoted_to_fp16)[name = string("op_9051_cast_fp16")];
            tensor<int32, [1]> var_243_axes_0 = const()[name = string("var_243_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_243_keep_dims_0 = const()[name = string("var_243_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_243_cast_fp16 = reduce_mean(axes = var_243_axes_0, keep_dims = var_243_keep_dims_0, x = var_9051_cast_fp16)[name = string("var_243_cast_fp16")];
            fp16 var_9058_to_fp16 = const()[name = string("op_9058_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_9059_cast_fp16 = add(x = var_243_cast_fp16, y = var_9058_to_fp16)[name = string("op_9059_cast_fp16")];
            fp32 var_9060_epsilon_0 = const()[name = string("op_9060_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_9060_cast_fp16 = rsqrt(epsilon = var_9060_epsilon_0, x = var_9059_cast_fp16)[name = string("op_9060_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_243_cast_fp16 = mul(x = var_8845_cast_fp16_3, y = var_9060_cast_fp16)[name = string("x_normed_243_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_381_cast_fp16 = mul(x = x_normed_243_cast_fp16, y = var_8878_to_fp16)[name = string("x_381_cast_fp16")];
            tensor<int32, [4]> x_first_137_begin_0 = const()[name = string("x_first_137_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_137_end_0 = const()[name = string("x_first_137_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_137_end_mask_0 = const()[name = string("x_first_137_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_137_cast_fp16 = slice_by_index(begin = x_first_137_begin_0, end = x_first_137_end_0, end_mask = x_first_137_end_mask_0, x = x_381_cast_fp16)[name = string("x_first_137_cast_fp16")];
            tensor<int32, [4]> x_second_137_begin_0 = const()[name = string("x_second_137_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_137_end_0 = const()[name = string("x_second_137_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_137_end_mask_0 = const()[name = string("x_second_137_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_137_cast_fp16 = slice_by_index(begin = x_second_137_begin_0, end = x_second_137_end_0, end_mask = x_second_137_end_mask_0, x = x_381_cast_fp16)[name = string("x_second_137_cast_fp16")];
            fp16 const_111_promoted_to_fp16 = const()[name = string("const_111_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_9092_cast_fp16 = mul(x = x_second_137_cast_fp16, y = const_111_promoted_to_fp16)[name = string("op_9092_cast_fp16")];
            int32 var_9094 = const()[name = string("op_9094"), val = int32(1)];
            bool x_rot_137_interleave_0 = const()[name = string("x_rot_137_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_137_cast_fp16 = concat(axis = var_9094, interleave = x_rot_137_interleave_0, values = (var_9092_cast_fp16, x_first_137_cast_fp16))[name = string("x_rot_137_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_9110_cast_fp16 = mul(x = x_381_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_9110_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_9111_cast_fp16 = mul(x = x_rot_137_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_9111_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_111_cast_fp16 = add(x = var_9110_cast_fp16, y = var_9111_cast_fp16)[name = string("q_h_111_cast_fp16")];
            fp16 var_9114_promoted_to_fp16 = const()[name = string("op_9114_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_9115_cast_fp16 = pow(x = var_8831_cast_fp16, y = var_9114_promoted_to_fp16)[name = string("op_9115_cast_fp16")];
            tensor<int32, [1]> var_245_axes_0 = const()[name = string("var_245_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_245_keep_dims_0 = const()[name = string("var_245_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_245_cast_fp16 = reduce_mean(axes = var_245_axes_0, keep_dims = var_245_keep_dims_0, x = var_9115_cast_fp16)[name = string("var_245_cast_fp16")];
            fp16 var_9122_to_fp16 = const()[name = string("op_9122_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_9123_cast_fp16 = add(x = var_245_cast_fp16, y = var_9122_to_fp16)[name = string("op_9123_cast_fp16")];
            fp32 var_9124_epsilon_0 = const()[name = string("op_9124_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_9124_cast_fp16 = rsqrt(epsilon = var_9124_epsilon_0, x = var_9123_cast_fp16)[name = string("op_9124_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_245_cast_fp16 = mul(x = var_8831_cast_fp16, y = var_9124_cast_fp16)[name = string("x_normed_245_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_9134_to_fp16 = const()[name = string("op_9134_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482437888)))];
            tensor<fp16, [1, 256, 1, ?]> x_385_cast_fp16 = mul(x = x_normed_245_cast_fp16, y = var_9134_to_fp16)[name = string("x_385_cast_fp16")];
            tensor<int32, [4]> x_first_139_begin_0 = const()[name = string("x_first_139_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_139_end_0 = const()[name = string("x_first_139_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_139_end_mask_0 = const()[name = string("x_first_139_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_139_cast_fp16 = slice_by_index(begin = x_first_139_begin_0, end = x_first_139_end_0, end_mask = x_first_139_end_mask_0, x = x_385_cast_fp16)[name = string("x_first_139_cast_fp16")];
            tensor<int32, [4]> x_second_139_begin_0 = const()[name = string("x_second_139_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_139_end_0 = const()[name = string("x_second_139_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_139_end_mask_0 = const()[name = string("x_second_139_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_139_cast_fp16 = slice_by_index(begin = x_second_139_begin_0, end = x_second_139_end_0, end_mask = x_second_139_end_mask_0, x = x_385_cast_fp16)[name = string("x_second_139_cast_fp16")];
            fp16 const_112_promoted_to_fp16 = const()[name = string("const_112_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_9156_cast_fp16 = mul(x = x_second_139_cast_fp16, y = const_112_promoted_to_fp16)[name = string("op_9156_cast_fp16")];
            int32 var_9158 = const()[name = string("op_9158"), val = int32(1)];
            bool x_rot_139_interleave_0 = const()[name = string("x_rot_139_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_139_cast_fp16 = concat(axis = var_9158, interleave = x_rot_139_interleave_0, values = (var_9156_cast_fp16, x_first_139_cast_fp16))[name = string("x_rot_139_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_9174_cast_fp16 = mul(x = x_385_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_9174_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_9175_cast_fp16 = mul(x = x_rot_139_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_9175_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_27_cast_fp16 = add(x = var_9174_cast_fp16, y = var_9175_cast_fp16)[name = string("k_new_27_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_26 = read_state(input = k_cache_13)[name = string("read_state_26")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_81_cast_fp16 = mul(x = read_state_26, y = one_minus_1_cast_fp16)[name = string("k_cache_81_cast_fp16")];
            write_state(data = k_cache_81_cast_fp16, input = k_cache_13)[name = string("coreml_update_state_124_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_124 = read_state(input = k_cache_13)[name = string("coreml_update_state_124")];
            tensor<fp16, [1, 256, 1, 128]> var_9182_cast_fp16 = mul(x = k_new_27_cast_fp16, y = write_mask)[name = string("op_9182_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_83_cast_fp16 = add(x = coreml_update_state_124, y = var_9182_cast_fp16)[name = string("k_cache_83_cast_fp16")];
            write_state(data = k_cache_83_cast_fp16, input = k_cache_13)[name = string("coreml_update_state_125_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_125 = read_state(input = k_cache_13)[name = string("coreml_update_state_125")];
            tensor<fp16, [1, 256, 1, 128]> read_state_27 = read_state(input = v_cache_13)[name = string("read_state_27")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_81_cast_fp16 = mul(x = read_state_27, y = one_minus_1_cast_fp16)[name = string("v_cache_81_cast_fp16")];
            write_state(data = v_cache_81_cast_fp16, input = v_cache_13)[name = string("coreml_update_state_126_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_126 = read_state(input = v_cache_13)[name = string("coreml_update_state_126")];
            tensor<fp16, [1, 256, 1, 128]> var_9186_cast_fp16 = mul(x = var_8842_cast_fp16, y = write_mask)[name = string("op_9186_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_83_cast_fp16 = add(x = coreml_update_state_126, y = var_9186_cast_fp16)[name = string("v_cache_83_cast_fp16")];
            write_state(data = v_cache_83_cast_fp16, input = v_cache_13)[name = string("coreml_update_state_127_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_127 = read_state(input = v_cache_13)[name = string("coreml_update_state_127")];
            bool var_9201_transpose_x_0 = const()[name = string("op_9201_transpose_x_0"), val = bool(false)];
            bool var_9201_transpose_y_0 = const()[name = string("op_9201_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_410_perm_0 = const()[name = string("transpose_410_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_411_perm_0 = const()[name = string("transpose_411_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_411 = transpose(perm = transpose_411_perm_0, x = coreml_update_state_125)[name = string("transpose_483")];
            tensor<fp16, [1, 1, ?, 256]> transpose_410 = transpose(perm = transpose_410_perm_0, x = q_h_105_cast_fp16)[name = string("transpose_484")];
            tensor<fp16, [1, 1, ?, 128]> var_9201_cast_fp16 = matmul(transpose_x = var_9201_transpose_x_0, transpose_y = var_9201_transpose_y_0, x = transpose_410, y = transpose_411)[name = string("op_9201_cast_fp16")];
            fp16 var_9202_to_fp16 = const()[name = string("op_9202_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_9203_cast_fp16 = mul(x = var_9201_cast_fp16, y = var_9202_to_fp16)[name = string("op_9203_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_105_cast_fp16 = add(x = var_9203_cast_fp16, y = attention_mask)[name = string("scores_105_cast_fp16")];
            int32 var_9206 = const()[name = string("op_9206"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_105_cast_fp16 = softmax(axis = var_9206, x = scores_105_cast_fp16)[name = string("probs_105_cast_fp16")];
            tensor<int32, [4]> var_9213 = const()[name = string("op_9213"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_9215_transpose_x_0 = const()[name = string("op_9215_transpose_x_0"), val = bool(false)];
            bool var_9215_transpose_y_0 = const()[name = string("op_9215_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_105_cast_fp16 = transpose(perm = var_9213, x = coreml_update_state_127)[name = string("transpose_482")];
            tensor<fp16, [1, 1, ?, 256]> var_9215_cast_fp16 = matmul(transpose_x = var_9215_transpose_x_0, transpose_y = var_9215_transpose_y_0, x = probs_105_cast_fp16, y = v_t_105_cast_fp16)[name = string("op_9215_cast_fp16")];
            bool var_9234_transpose_x_0 = const()[name = string("op_9234_transpose_x_0"), val = bool(false)];
            bool var_9234_transpose_y_0 = const()[name = string("op_9234_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_415_perm_0 = const()[name = string("transpose_415_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_415 = transpose(perm = transpose_415_perm_0, x = q_h_107_cast_fp16)[name = string("transpose_481")];
            tensor<fp16, [1, 1, ?, 128]> var_9234_cast_fp16 = matmul(transpose_x = var_9234_transpose_x_0, transpose_y = var_9234_transpose_y_0, x = transpose_415, y = transpose_411)[name = string("op_9234_cast_fp16")];
            fp16 var_9235_to_fp16 = const()[name = string("op_9235_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_9236_cast_fp16 = mul(x = var_9234_cast_fp16, y = var_9235_to_fp16)[name = string("op_9236_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_107_cast_fp16 = add(x = var_9236_cast_fp16, y = attention_mask)[name = string("scores_107_cast_fp16")];
            int32 var_9239 = const()[name = string("op_9239"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_107_cast_fp16 = softmax(axis = var_9239, x = scores_107_cast_fp16)[name = string("probs_107_cast_fp16")];
            bool var_9248_transpose_x_0 = const()[name = string("op_9248_transpose_x_0"), val = bool(false)];
            bool var_9248_transpose_y_0 = const()[name = string("op_9248_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_9248_cast_fp16 = matmul(transpose_x = var_9248_transpose_x_0, transpose_y = var_9248_transpose_y_0, x = probs_107_cast_fp16, y = v_t_105_cast_fp16)[name = string("op_9248_cast_fp16")];
            bool var_9267_transpose_x_0 = const()[name = string("op_9267_transpose_x_0"), val = bool(false)];
            bool var_9267_transpose_y_0 = const()[name = string("op_9267_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_416_perm_0 = const()[name = string("transpose_416_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_416 = transpose(perm = transpose_416_perm_0, x = q_h_109_cast_fp16)[name = string("transpose_480")];
            tensor<fp16, [1, 1, ?, 128]> var_9267_cast_fp16 = matmul(transpose_x = var_9267_transpose_x_0, transpose_y = var_9267_transpose_y_0, x = transpose_416, y = transpose_411)[name = string("op_9267_cast_fp16")];
            fp16 var_9268_to_fp16 = const()[name = string("op_9268_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_9269_cast_fp16 = mul(x = var_9267_cast_fp16, y = var_9268_to_fp16)[name = string("op_9269_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_109_cast_fp16 = add(x = var_9269_cast_fp16, y = attention_mask)[name = string("scores_109_cast_fp16")];
            int32 var_9272 = const()[name = string("op_9272"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_109_cast_fp16 = softmax(axis = var_9272, x = scores_109_cast_fp16)[name = string("probs_109_cast_fp16")];
            bool var_9281_transpose_x_0 = const()[name = string("op_9281_transpose_x_0"), val = bool(false)];
            bool var_9281_transpose_y_0 = const()[name = string("op_9281_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_9281_cast_fp16 = matmul(transpose_x = var_9281_transpose_x_0, transpose_y = var_9281_transpose_y_0, x = probs_109_cast_fp16, y = v_t_105_cast_fp16)[name = string("op_9281_cast_fp16")];
            bool var_9300_transpose_x_0 = const()[name = string("op_9300_transpose_x_0"), val = bool(false)];
            bool var_9300_transpose_y_0 = const()[name = string("op_9300_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_417_perm_0 = const()[name = string("transpose_417_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_417 = transpose(perm = transpose_417_perm_0, x = q_h_111_cast_fp16)[name = string("transpose_479")];
            tensor<fp16, [1, 1, ?, 128]> var_9300_cast_fp16 = matmul(transpose_x = var_9300_transpose_x_0, transpose_y = var_9300_transpose_y_0, x = transpose_417, y = transpose_411)[name = string("op_9300_cast_fp16")];
            fp16 var_9301_to_fp16 = const()[name = string("op_9301_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_9302_cast_fp16 = mul(x = var_9300_cast_fp16, y = var_9301_to_fp16)[name = string("op_9302_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_111_cast_fp16 = add(x = var_9302_cast_fp16, y = attention_mask)[name = string("scores_111_cast_fp16")];
            int32 var_9305 = const()[name = string("op_9305"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_111_cast_fp16 = softmax(axis = var_9305, x = scores_111_cast_fp16)[name = string("probs_111_cast_fp16")];
            bool var_9314_transpose_x_0 = const()[name = string("op_9314_transpose_x_0"), val = bool(false)];
            bool var_9314_transpose_y_0 = const()[name = string("op_9314_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_9314_cast_fp16 = matmul(transpose_x = var_9314_transpose_x_0, transpose_y = var_9314_transpose_y_0, x = probs_111_cast_fp16, y = v_t_105_cast_fp16)[name = string("op_9314_cast_fp16")];
            bool input_107_interleave_0 = const()[name = string("input_107_interleave_0"), val = bool(false)];
            int32 const_158 = const()[name = string("const_158"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_107_cast_fp16 = concat(axis = const_158, interleave = input_107_interleave_0, values = (var_9215_cast_fp16, var_9248_cast_fp16, var_9281_cast_fp16, var_9314_cast_fp16))[name = string("input_107_cast_fp16")];
            string x_387_pad_type_0 = const()[name = string("x_387_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_387_strides_0 = const()[name = string("x_387_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_387_pad_0 = const()[name = string("x_387_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_387_dilations_0 = const()[name = string("x_387_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_387_groups_0 = const()[name = string("x_387_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_125_perm_0 = const()[name = string("transpose_125_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_13_weight_to_fp16 = const()[name = string("o_proj_13_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482438464)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_125_cast_fp16 = transpose(perm = transpose_125_perm_0, x = input_107_cast_fp16)[name = string("transpose_478")];
            tensor<fp16, [1, 640, 1, ?]> x_387_cast_fp16 = conv(dilations = x_387_dilations_0, groups = x_387_groups_0, pad = x_387_pad_0, pad_type = x_387_pad_type_0, strides = x_387_strides_0, weight = o_proj_13_weight_to_fp16, x = transpose_125_cast_fp16)[name = string("x_387_cast_fp16")];
            fp16 var_9335_promoted_to_fp16 = const()[name = string("op_9335_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_9336_cast_fp16 = pow(x = x_387_cast_fp16, y = var_9335_promoted_to_fp16)[name = string("op_9336_cast_fp16")];
            tensor<int32, [1]> var_247_axes_0 = const()[name = string("var_247_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_247_keep_dims_0 = const()[name = string("var_247_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_247_cast_fp16 = reduce_mean(axes = var_247_axes_0, keep_dims = var_247_keep_dims_0, x = var_9336_cast_fp16)[name = string("var_247_cast_fp16")];
            fp16 var_9343_to_fp16 = const()[name = string("op_9343_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_9344_cast_fp16 = add(x = var_247_cast_fp16, y = var_9343_to_fp16)[name = string("op_9344_cast_fp16")];
            fp32 var_9345_epsilon_0 = const()[name = string("op_9345_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_9345_cast_fp16 = rsqrt(epsilon = var_9345_epsilon_0, x = var_9344_cast_fp16)[name = string("op_9345_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_247_cast_fp16 = mul(x = x_387_cast_fp16, y = var_9345_cast_fp16)[name = string("x_normed_247_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_9355_to_fp16 = const()[name = string("op_9355_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483749248)))];
            tensor<fp16, [1, 640, 1, ?]> attn_55_cast_fp16 = mul(x = x_normed_247_cast_fp16, y = var_9355_to_fp16)[name = string("attn_55_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_389_cast_fp16 = add(x = x_365_cast_fp16, y = attn_55_cast_fp16)[name = string("x_389_cast_fp16")];
            fp16 var_9359_promoted_to_fp16 = const()[name = string("op_9359_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_9360_cast_fp16 = pow(x = x_389_cast_fp16, y = var_9359_promoted_to_fp16)[name = string("op_9360_cast_fp16")];
            tensor<int32, [1]> var_249_axes_0 = const()[name = string("var_249_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_249_keep_dims_0 = const()[name = string("var_249_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_249_cast_fp16 = reduce_mean(axes = var_249_axes_0, keep_dims = var_249_keep_dims_0, x = var_9360_cast_fp16)[name = string("var_249_cast_fp16")];
            fp16 var_9367_to_fp16 = const()[name = string("op_9367_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_9368_cast_fp16 = add(x = var_249_cast_fp16, y = var_9367_to_fp16)[name = string("op_9368_cast_fp16")];
            fp32 var_9369_epsilon_0 = const()[name = string("op_9369_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_9369_cast_fp16 = rsqrt(epsilon = var_9369_epsilon_0, x = var_9368_cast_fp16)[name = string("op_9369_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_249_cast_fp16 = mul(x = x_389_cast_fp16, y = var_9369_cast_fp16)[name = string("x_normed_249_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_9379_to_fp16 = const()[name = string("op_9379_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483750592)))];
            tensor<fp16, [1, 640, 1, ?]> input_109_cast_fp16 = mul(x = x_normed_249_cast_fp16, y = var_9379_to_fp16)[name = string("input_109_cast_fp16")];
            string gate_27_pad_type_0 = const()[name = string("gate_27_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_27_strides_0 = const()[name = string("gate_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_27_pad_0 = const()[name = string("gate_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_27_dilations_0 = const()[name = string("gate_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_27_groups_0 = const()[name = string("gate_27_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_13_weight_to_fp16 = const()[name = string("gate_proj_13_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483751936)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_27_cast_fp16 = conv(dilations = gate_27_dilations_0, groups = gate_27_groups_0, pad = gate_27_pad_0, pad_type = gate_27_pad_type_0, strides = gate_27_strides_0, weight = gate_proj_13_weight_to_fp16, x = input_109_cast_fp16)[name = string("gate_27_cast_fp16")];
            string up_27_pad_type_0 = const()[name = string("up_27_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_27_strides_0 = const()[name = string("up_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_27_pad_0 = const()[name = string("up_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_27_dilations_0 = const()[name = string("up_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_27_groups_0 = const()[name = string("up_27_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_13_weight_to_fp16 = const()[name = string("up_proj_13_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486373440)))];
            tensor<fp16, [1, 2048, 1, ?]> up_27_cast_fp16 = conv(dilations = up_27_dilations_0, groups = up_27_groups_0, pad = up_27_pad_0, pad_type = up_27_pad_type_0, strides = up_27_strides_0, weight = up_proj_13_weight_to_fp16, x = input_109_cast_fp16)[name = string("up_27_cast_fp16")];
            string var_9404_mode_0 = const()[name = string("op_9404_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_9404_cast_fp16 = gelu(mode = var_9404_mode_0, x = gate_27_cast_fp16)[name = string("op_9404_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_111_cast_fp16 = mul(x = var_9404_cast_fp16, y = up_27_cast_fp16)[name = string("input_111_cast_fp16")];
            string x_391_pad_type_0 = const()[name = string("x_391_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_391_strides_0 = const()[name = string("x_391_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_391_pad_0 = const()[name = string("x_391_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_391_dilations_0 = const()[name = string("x_391_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_391_groups_0 = const()[name = string("x_391_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_13_weight_to_fp16 = const()[name = string("down_proj_13_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488994944)))];
            tensor<fp16, [1, 640, 1, ?]> x_391_cast_fp16 = conv(dilations = x_391_dilations_0, groups = x_391_groups_0, pad = x_391_pad_0, pad_type = x_391_pad_type_0, strides = x_391_strides_0, weight = down_proj_13_weight_to_fp16, x = input_111_cast_fp16)[name = string("x_391_cast_fp16")];
            fp16 var_9417_promoted_to_fp16 = const()[name = string("op_9417_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_9418_cast_fp16 = pow(x = x_391_cast_fp16, y = var_9417_promoted_to_fp16)[name = string("op_9418_cast_fp16")];
            tensor<int32, [1]> var_251_axes_0 = const()[name = string("var_251_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_251_keep_dims_0 = const()[name = string("var_251_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_251_cast_fp16 = reduce_mean(axes = var_251_axes_0, keep_dims = var_251_keep_dims_0, x = var_9418_cast_fp16)[name = string("var_251_cast_fp16")];
            fp16 var_9425_to_fp16 = const()[name = string("op_9425_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_9426_cast_fp16 = add(x = var_251_cast_fp16, y = var_9425_to_fp16)[name = string("op_9426_cast_fp16")];
            fp32 var_9427_epsilon_0 = const()[name = string("op_9427_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_9427_cast_fp16 = rsqrt(epsilon = var_9427_epsilon_0, x = var_9426_cast_fp16)[name = string("op_9427_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_251_cast_fp16 = mul(x = x_391_cast_fp16, y = var_9427_cast_fp16)[name = string("x_normed_251_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_9437_to_fp16 = const()[name = string("op_9437_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491616448)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_27_cast_fp16 = mul(x = x_normed_251_cast_fp16, y = var_9437_to_fp16)[name = string("mlp_27_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_393_cast_fp16 = add(x = x_389_cast_fp16, y = mlp_27_cast_fp16)[name = string("x_393_cast_fp16")];
            fp16 var_9441_promoted_to_fp16 = const()[name = string("op_9441_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_9442_cast_fp16 = pow(x = x_393_cast_fp16, y = var_9441_promoted_to_fp16)[name = string("op_9442_cast_fp16")];
            tensor<int32, [1]> var_253_axes_0 = const()[name = string("var_253_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_253_keep_dims_0 = const()[name = string("var_253_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_253_cast_fp16 = reduce_mean(axes = var_253_axes_0, keep_dims = var_253_keep_dims_0, x = var_9442_cast_fp16)[name = string("var_253_cast_fp16")];
            fp16 var_9449_to_fp16 = const()[name = string("op_9449_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_9450_cast_fp16 = add(x = var_253_cast_fp16, y = var_9449_to_fp16)[name = string("op_9450_cast_fp16")];
            fp32 var_9451_epsilon_0 = const()[name = string("op_9451_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_9451_cast_fp16 = rsqrt(epsilon = var_9451_epsilon_0, x = var_9450_cast_fp16)[name = string("op_9451_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_253_cast_fp16 = mul(x = x_393_cast_fp16, y = var_9451_cast_fp16)[name = string("x_normed_253_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_9461_to_fp16 = const()[name = string("op_9461_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491617792)))];
            tensor<fp16, [1, 640, 1, ?]> input_113_cast_fp16 = mul(x = x_normed_253_cast_fp16, y = var_9461_to_fp16)[name = string("input_113_cast_fp16")];
            string var_9473_pad_type_0 = const()[name = string("op_9473_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_9473_strides_0 = const()[name = string("op_9473_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9473_pad_0 = const()[name = string("op_9473_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9473_dilations_0 = const()[name = string("op_9473_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_9473_groups_0 = const()[name = string("op_9473_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_14_weight_to_fp16 = const()[name = string("q_proj_14_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491619136)))];
            tensor<fp16, [1, 1024, 1, ?]> var_9473_cast_fp16 = conv(dilations = var_9473_dilations_0, groups = var_9473_groups_0, pad = var_9473_pad_0, pad_type = var_9473_pad_type_0, strides = var_9473_strides_0, weight = q_proj_14_weight_to_fp16, x = input_113_cast_fp16)[name = string("op_9473_cast_fp16")];
            string var_9484_pad_type_0 = const()[name = string("op_9484_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_9484_strides_0 = const()[name = string("op_9484_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9484_pad_0 = const()[name = string("op_9484_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9484_dilations_0 = const()[name = string("op_9484_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_9484_groups_0 = const()[name = string("op_9484_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_14_weight_to_fp16 = const()[name = string("k_proj_14_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492929920)))];
            tensor<fp16, [1, 256, 1, ?]> var_9484_cast_fp16 = conv(dilations = var_9484_dilations_0, groups = var_9484_groups_0, pad = var_9484_pad_0, pad_type = var_9484_pad_type_0, strides = var_9484_strides_0, weight = k_proj_14_weight_to_fp16, x = input_113_cast_fp16)[name = string("op_9484_cast_fp16")];
            string var_9495_pad_type_0 = const()[name = string("op_9495_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_9495_strides_0 = const()[name = string("op_9495_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_9495_pad_0 = const()[name = string("op_9495_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_9495_dilations_0 = const()[name = string("op_9495_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_9495_groups_0 = const()[name = string("op_9495_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_14_weight_to_fp16 = const()[name = string("v_proj_14_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493257664)))];
            tensor<fp16, [1, 256, 1, ?]> var_9495_cast_fp16 = conv(dilations = var_9495_dilations_0, groups = var_9495_groups_0, pad = var_9495_pad_0, pad_type = var_9495_pad_type_0, strides = var_9495_strides_0, weight = v_proj_14_weight_to_fp16, x = input_113_cast_fp16)[name = string("op_9495_cast_fp16")];
            tensor<int32, [4]> tile_42 = const()[name = string("tile_42"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_9498_axis_0 = const()[name = string("op_9498_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_9498_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_9498_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_9498_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_9498_cast_fp16_3 = split(axis = var_9498_axis_0, split_sizes = tile_42, x = var_9473_cast_fp16)[name = string("op_9498_cast_fp16")];
            fp16 var_9511_promoted_to_fp16 = const()[name = string("op_9511_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_9512_cast_fp16 = pow(x = var_9498_cast_fp16_0, y = var_9511_promoted_to_fp16)[name = string("op_9512_cast_fp16")];
            tensor<int32, [1]> var_255_axes_0 = const()[name = string("var_255_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_255_keep_dims_0 = const()[name = string("var_255_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_255_cast_fp16 = reduce_mean(axes = var_255_axes_0, keep_dims = var_255_keep_dims_0, x = var_9512_cast_fp16)[name = string("var_255_cast_fp16")];
            fp16 var_9519_to_fp16 = const()[name = string("op_9519_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_9520_cast_fp16 = add(x = var_255_cast_fp16, y = var_9519_to_fp16)[name = string("op_9520_cast_fp16")];
            fp32 var_9521_epsilon_0 = const()[name = string("op_9521_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_9521_cast_fp16 = rsqrt(epsilon = var_9521_epsilon_0, x = var_9520_cast_fp16)[name = string("op_9521_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_255_cast_fp16 = mul(x = var_9498_cast_fp16_0, y = var_9521_cast_fp16)[name = string("x_normed_255_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_9531_to_fp16 = const()[name = string("op_9531_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493585408)))];
            tensor<fp16, [1, 256, 1, ?]> x_397_cast_fp16 = mul(x = x_normed_255_cast_fp16, y = var_9531_to_fp16)[name = string("x_397_cast_fp16")];
            tensor<int32, [4]> x_first_141_begin_0 = const()[name = string("x_first_141_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_141_end_0 = const()[name = string("x_first_141_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_141_end_mask_0 = const()[name = string("x_first_141_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_141_cast_fp16 = slice_by_index(begin = x_first_141_begin_0, end = x_first_141_end_0, end_mask = x_first_141_end_mask_0, x = x_397_cast_fp16)[name = string("x_first_141_cast_fp16")];
            tensor<int32, [4]> x_second_141_begin_0 = const()[name = string("x_second_141_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_141_end_0 = const()[name = string("x_second_141_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_141_end_mask_0 = const()[name = string("x_second_141_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_141_cast_fp16 = slice_by_index(begin = x_second_141_begin_0, end = x_second_141_end_0, end_mask = x_second_141_end_mask_0, x = x_397_cast_fp16)[name = string("x_second_141_cast_fp16")];
            fp16 const_116_promoted_to_fp16 = const()[name = string("const_116_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_9553_cast_fp16 = mul(x = x_second_141_cast_fp16, y = const_116_promoted_to_fp16)[name = string("op_9553_cast_fp16")];
            int32 var_9555 = const()[name = string("op_9555"), val = int32(1)];
            bool x_rot_141_interleave_0 = const()[name = string("x_rot_141_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_141_cast_fp16 = concat(axis = var_9555, interleave = x_rot_141_interleave_0, values = (var_9553_cast_fp16, x_first_141_cast_fp16))[name = string("x_rot_141_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_9571_cast_fp16 = mul(x = x_397_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_9571_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_9572_cast_fp16 = mul(x = x_rot_141_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_9572_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_113_cast_fp16 = add(x = var_9571_cast_fp16, y = var_9572_cast_fp16)[name = string("q_h_113_cast_fp16")];
            fp16 var_9575_promoted_to_fp16 = const()[name = string("op_9575_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_9576_cast_fp16 = pow(x = var_9498_cast_fp16_1, y = var_9575_promoted_to_fp16)[name = string("op_9576_cast_fp16")];
            tensor<int32, [1]> var_257_axes_0 = const()[name = string("var_257_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_257_keep_dims_0 = const()[name = string("var_257_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_257_cast_fp16 = reduce_mean(axes = var_257_axes_0, keep_dims = var_257_keep_dims_0, x = var_9576_cast_fp16)[name = string("var_257_cast_fp16")];
            fp16 var_9583_to_fp16 = const()[name = string("op_9583_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_9584_cast_fp16 = add(x = var_257_cast_fp16, y = var_9583_to_fp16)[name = string("op_9584_cast_fp16")];
            fp32 var_9585_epsilon_0 = const()[name = string("op_9585_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_9585_cast_fp16 = rsqrt(epsilon = var_9585_epsilon_0, x = var_9584_cast_fp16)[name = string("op_9585_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_257_cast_fp16 = mul(x = var_9498_cast_fp16_1, y = var_9585_cast_fp16)[name = string("x_normed_257_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_401_cast_fp16 = mul(x = x_normed_257_cast_fp16, y = var_9531_to_fp16)[name = string("x_401_cast_fp16")];
            tensor<int32, [4]> x_first_143_begin_0 = const()[name = string("x_first_143_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_143_end_0 = const()[name = string("x_first_143_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_143_end_mask_0 = const()[name = string("x_first_143_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_143_cast_fp16 = slice_by_index(begin = x_first_143_begin_0, end = x_first_143_end_0, end_mask = x_first_143_end_mask_0, x = x_401_cast_fp16)[name = string("x_first_143_cast_fp16")];
            tensor<int32, [4]> x_second_143_begin_0 = const()[name = string("x_second_143_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_143_end_0 = const()[name = string("x_second_143_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_143_end_mask_0 = const()[name = string("x_second_143_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_143_cast_fp16 = slice_by_index(begin = x_second_143_begin_0, end = x_second_143_end_0, end_mask = x_second_143_end_mask_0, x = x_401_cast_fp16)[name = string("x_second_143_cast_fp16")];
            fp16 const_117_promoted_to_fp16 = const()[name = string("const_117_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_9617_cast_fp16 = mul(x = x_second_143_cast_fp16, y = const_117_promoted_to_fp16)[name = string("op_9617_cast_fp16")];
            int32 var_9619 = const()[name = string("op_9619"), val = int32(1)];
            bool x_rot_143_interleave_0 = const()[name = string("x_rot_143_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_143_cast_fp16 = concat(axis = var_9619, interleave = x_rot_143_interleave_0, values = (var_9617_cast_fp16, x_first_143_cast_fp16))[name = string("x_rot_143_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_9635_cast_fp16 = mul(x = x_401_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_9635_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_9636_cast_fp16 = mul(x = x_rot_143_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_9636_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_115_cast_fp16 = add(x = var_9635_cast_fp16, y = var_9636_cast_fp16)[name = string("q_h_115_cast_fp16")];
            fp16 var_9639_promoted_to_fp16 = const()[name = string("op_9639_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_9640_cast_fp16 = pow(x = var_9498_cast_fp16_2, y = var_9639_promoted_to_fp16)[name = string("op_9640_cast_fp16")];
            tensor<int32, [1]> var_259_axes_0 = const()[name = string("var_259_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_259_keep_dims_0 = const()[name = string("var_259_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_259_cast_fp16 = reduce_mean(axes = var_259_axes_0, keep_dims = var_259_keep_dims_0, x = var_9640_cast_fp16)[name = string("var_259_cast_fp16")];
            fp16 var_9647_to_fp16 = const()[name = string("op_9647_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_9648_cast_fp16 = add(x = var_259_cast_fp16, y = var_9647_to_fp16)[name = string("op_9648_cast_fp16")];
            fp32 var_9649_epsilon_0 = const()[name = string("op_9649_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_9649_cast_fp16 = rsqrt(epsilon = var_9649_epsilon_0, x = var_9648_cast_fp16)[name = string("op_9649_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_259_cast_fp16 = mul(x = var_9498_cast_fp16_2, y = var_9649_cast_fp16)[name = string("x_normed_259_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_405_cast_fp16 = mul(x = x_normed_259_cast_fp16, y = var_9531_to_fp16)[name = string("x_405_cast_fp16")];
            tensor<int32, [4]> x_first_145_begin_0 = const()[name = string("x_first_145_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_145_end_0 = const()[name = string("x_first_145_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_145_end_mask_0 = const()[name = string("x_first_145_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_145_cast_fp16 = slice_by_index(begin = x_first_145_begin_0, end = x_first_145_end_0, end_mask = x_first_145_end_mask_0, x = x_405_cast_fp16)[name = string("x_first_145_cast_fp16")];
            tensor<int32, [4]> x_second_145_begin_0 = const()[name = string("x_second_145_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_145_end_0 = const()[name = string("x_second_145_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_145_end_mask_0 = const()[name = string("x_second_145_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_145_cast_fp16 = slice_by_index(begin = x_second_145_begin_0, end = x_second_145_end_0, end_mask = x_second_145_end_mask_0, x = x_405_cast_fp16)[name = string("x_second_145_cast_fp16")];
            fp16 const_118_promoted_to_fp16 = const()[name = string("const_118_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_9681_cast_fp16 = mul(x = x_second_145_cast_fp16, y = const_118_promoted_to_fp16)[name = string("op_9681_cast_fp16")];
            int32 var_9683 = const()[name = string("op_9683"), val = int32(1)];
            bool x_rot_145_interleave_0 = const()[name = string("x_rot_145_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_145_cast_fp16 = concat(axis = var_9683, interleave = x_rot_145_interleave_0, values = (var_9681_cast_fp16, x_first_145_cast_fp16))[name = string("x_rot_145_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_9699_cast_fp16 = mul(x = x_405_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_9699_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_9700_cast_fp16 = mul(x = x_rot_145_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_9700_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_117_cast_fp16 = add(x = var_9699_cast_fp16, y = var_9700_cast_fp16)[name = string("q_h_117_cast_fp16")];
            fp16 var_9703_promoted_to_fp16 = const()[name = string("op_9703_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_9704_cast_fp16 = pow(x = var_9498_cast_fp16_3, y = var_9703_promoted_to_fp16)[name = string("op_9704_cast_fp16")];
            tensor<int32, [1]> var_261_axes_0 = const()[name = string("var_261_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_261_keep_dims_0 = const()[name = string("var_261_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_261_cast_fp16 = reduce_mean(axes = var_261_axes_0, keep_dims = var_261_keep_dims_0, x = var_9704_cast_fp16)[name = string("var_261_cast_fp16")];
            fp16 var_9711_to_fp16 = const()[name = string("op_9711_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_9712_cast_fp16 = add(x = var_261_cast_fp16, y = var_9711_to_fp16)[name = string("op_9712_cast_fp16")];
            fp32 var_9713_epsilon_0 = const()[name = string("op_9713_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_9713_cast_fp16 = rsqrt(epsilon = var_9713_epsilon_0, x = var_9712_cast_fp16)[name = string("op_9713_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_261_cast_fp16 = mul(x = var_9498_cast_fp16_3, y = var_9713_cast_fp16)[name = string("x_normed_261_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_409_cast_fp16 = mul(x = x_normed_261_cast_fp16, y = var_9531_to_fp16)[name = string("x_409_cast_fp16")];
            tensor<int32, [4]> x_first_147_begin_0 = const()[name = string("x_first_147_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_147_end_0 = const()[name = string("x_first_147_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_147_end_mask_0 = const()[name = string("x_first_147_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_147_cast_fp16 = slice_by_index(begin = x_first_147_begin_0, end = x_first_147_end_0, end_mask = x_first_147_end_mask_0, x = x_409_cast_fp16)[name = string("x_first_147_cast_fp16")];
            tensor<int32, [4]> x_second_147_begin_0 = const()[name = string("x_second_147_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_147_end_0 = const()[name = string("x_second_147_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_147_end_mask_0 = const()[name = string("x_second_147_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_147_cast_fp16 = slice_by_index(begin = x_second_147_begin_0, end = x_second_147_end_0, end_mask = x_second_147_end_mask_0, x = x_409_cast_fp16)[name = string("x_second_147_cast_fp16")];
            fp16 const_119_promoted_to_fp16 = const()[name = string("const_119_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_9745_cast_fp16 = mul(x = x_second_147_cast_fp16, y = const_119_promoted_to_fp16)[name = string("op_9745_cast_fp16")];
            int32 var_9747 = const()[name = string("op_9747"), val = int32(1)];
            bool x_rot_147_interleave_0 = const()[name = string("x_rot_147_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_147_cast_fp16 = concat(axis = var_9747, interleave = x_rot_147_interleave_0, values = (var_9745_cast_fp16, x_first_147_cast_fp16))[name = string("x_rot_147_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_9763_cast_fp16 = mul(x = x_409_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_9763_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_9764_cast_fp16 = mul(x = x_rot_147_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_9764_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_119_cast_fp16 = add(x = var_9763_cast_fp16, y = var_9764_cast_fp16)[name = string("q_h_119_cast_fp16")];
            fp16 var_9767_promoted_to_fp16 = const()[name = string("op_9767_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_9768_cast_fp16 = pow(x = var_9484_cast_fp16, y = var_9767_promoted_to_fp16)[name = string("op_9768_cast_fp16")];
            tensor<int32, [1]> var_263_axes_0 = const()[name = string("var_263_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_263_keep_dims_0 = const()[name = string("var_263_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_263_cast_fp16 = reduce_mean(axes = var_263_axes_0, keep_dims = var_263_keep_dims_0, x = var_9768_cast_fp16)[name = string("var_263_cast_fp16")];
            fp16 var_9775_to_fp16 = const()[name = string("op_9775_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_9776_cast_fp16 = add(x = var_263_cast_fp16, y = var_9775_to_fp16)[name = string("op_9776_cast_fp16")];
            fp32 var_9777_epsilon_0 = const()[name = string("op_9777_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_9777_cast_fp16 = rsqrt(epsilon = var_9777_epsilon_0, x = var_9776_cast_fp16)[name = string("op_9777_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_263_cast_fp16 = mul(x = var_9484_cast_fp16, y = var_9777_cast_fp16)[name = string("x_normed_263_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_9787_to_fp16 = const()[name = string("op_9787_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493585984)))];
            tensor<fp16, [1, 256, 1, ?]> x_413_cast_fp16 = mul(x = x_normed_263_cast_fp16, y = var_9787_to_fp16)[name = string("x_413_cast_fp16")];
            tensor<int32, [4]> x_first_149_begin_0 = const()[name = string("x_first_149_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_149_end_0 = const()[name = string("x_first_149_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_149_end_mask_0 = const()[name = string("x_first_149_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_149_cast_fp16 = slice_by_index(begin = x_first_149_begin_0, end = x_first_149_end_0, end_mask = x_first_149_end_mask_0, x = x_413_cast_fp16)[name = string("x_first_149_cast_fp16")];
            tensor<int32, [4]> x_second_149_begin_0 = const()[name = string("x_second_149_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_149_end_0 = const()[name = string("x_second_149_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_149_end_mask_0 = const()[name = string("x_second_149_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_149_cast_fp16 = slice_by_index(begin = x_second_149_begin_0, end = x_second_149_end_0, end_mask = x_second_149_end_mask_0, x = x_413_cast_fp16)[name = string("x_second_149_cast_fp16")];
            fp16 const_120_promoted_to_fp16 = const()[name = string("const_120_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_9809_cast_fp16 = mul(x = x_second_149_cast_fp16, y = const_120_promoted_to_fp16)[name = string("op_9809_cast_fp16")];
            int32 var_9811 = const()[name = string("op_9811"), val = int32(1)];
            bool x_rot_149_interleave_0 = const()[name = string("x_rot_149_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_149_cast_fp16 = concat(axis = var_9811, interleave = x_rot_149_interleave_0, values = (var_9809_cast_fp16, x_first_149_cast_fp16))[name = string("x_rot_149_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_9827_cast_fp16 = mul(x = x_413_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_9827_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_9828_cast_fp16 = mul(x = x_rot_149_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_9828_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_29_cast_fp16 = add(x = var_9827_cast_fp16, y = var_9828_cast_fp16)[name = string("k_new_29_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_28 = read_state(input = k_cache_14)[name = string("read_state_28")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_87_cast_fp16 = mul(x = read_state_28, y = one_minus_1_cast_fp16)[name = string("k_cache_87_cast_fp16")];
            write_state(data = k_cache_87_cast_fp16, input = k_cache_14)[name = string("coreml_update_state_128_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_128 = read_state(input = k_cache_14)[name = string("coreml_update_state_128")];
            tensor<fp16, [1, 256, 1, 128]> var_9835_cast_fp16 = mul(x = k_new_29_cast_fp16, y = write_mask)[name = string("op_9835_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_89_cast_fp16 = add(x = coreml_update_state_128, y = var_9835_cast_fp16)[name = string("k_cache_89_cast_fp16")];
            write_state(data = k_cache_89_cast_fp16, input = k_cache_14)[name = string("coreml_update_state_129_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_129 = read_state(input = k_cache_14)[name = string("coreml_update_state_129")];
            tensor<fp16, [1, 256, 1, 128]> read_state_29 = read_state(input = v_cache_14)[name = string("read_state_29")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_87_cast_fp16 = mul(x = read_state_29, y = one_minus_1_cast_fp16)[name = string("v_cache_87_cast_fp16")];
            write_state(data = v_cache_87_cast_fp16, input = v_cache_14)[name = string("coreml_update_state_130_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_130 = read_state(input = v_cache_14)[name = string("coreml_update_state_130")];
            tensor<fp16, [1, 256, 1, 128]> var_9839_cast_fp16 = mul(x = var_9495_cast_fp16, y = write_mask)[name = string("op_9839_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_89_cast_fp16 = add(x = coreml_update_state_130, y = var_9839_cast_fp16)[name = string("v_cache_89_cast_fp16")];
            write_state(data = v_cache_89_cast_fp16, input = v_cache_14)[name = string("coreml_update_state_131_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_131 = read_state(input = v_cache_14)[name = string("coreml_update_state_131")];
            bool var_9854_transpose_x_0 = const()[name = string("op_9854_transpose_x_0"), val = bool(false)];
            bool var_9854_transpose_y_0 = const()[name = string("op_9854_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_418_perm_0 = const()[name = string("transpose_418_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_419_perm_0 = const()[name = string("transpose_419_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_419 = transpose(perm = transpose_419_perm_0, x = coreml_update_state_129)[name = string("transpose_476")];
            tensor<fp16, [1, 1, ?, 256]> transpose_418 = transpose(perm = transpose_418_perm_0, x = q_h_113_cast_fp16)[name = string("transpose_477")];
            tensor<fp16, [1, 1, ?, 128]> var_9854_cast_fp16 = matmul(transpose_x = var_9854_transpose_x_0, transpose_y = var_9854_transpose_y_0, x = transpose_418, y = transpose_419)[name = string("op_9854_cast_fp16")];
            fp16 var_9855_to_fp16 = const()[name = string("op_9855_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_9856_cast_fp16 = mul(x = var_9854_cast_fp16, y = var_9855_to_fp16)[name = string("op_9856_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_113_cast_fp16 = add(x = var_9856_cast_fp16, y = attention_mask)[name = string("scores_113_cast_fp16")];
            int32 var_9859 = const()[name = string("op_9859"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_113_cast_fp16 = softmax(axis = var_9859, x = scores_113_cast_fp16)[name = string("probs_113_cast_fp16")];
            tensor<int32, [4]> var_9866 = const()[name = string("op_9866"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_9868_transpose_x_0 = const()[name = string("op_9868_transpose_x_0"), val = bool(false)];
            bool var_9868_transpose_y_0 = const()[name = string("op_9868_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_113_cast_fp16 = transpose(perm = var_9866, x = coreml_update_state_131)[name = string("transpose_475")];
            tensor<fp16, [1, 1, ?, 256]> var_9868_cast_fp16 = matmul(transpose_x = var_9868_transpose_x_0, transpose_y = var_9868_transpose_y_0, x = probs_113_cast_fp16, y = v_t_113_cast_fp16)[name = string("op_9868_cast_fp16")];
            bool var_9887_transpose_x_0 = const()[name = string("op_9887_transpose_x_0"), val = bool(false)];
            bool var_9887_transpose_y_0 = const()[name = string("op_9887_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_423_perm_0 = const()[name = string("transpose_423_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_423 = transpose(perm = transpose_423_perm_0, x = q_h_115_cast_fp16)[name = string("transpose_474")];
            tensor<fp16, [1, 1, ?, 128]> var_9887_cast_fp16 = matmul(transpose_x = var_9887_transpose_x_0, transpose_y = var_9887_transpose_y_0, x = transpose_423, y = transpose_419)[name = string("op_9887_cast_fp16")];
            fp16 var_9888_to_fp16 = const()[name = string("op_9888_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_9889_cast_fp16 = mul(x = var_9887_cast_fp16, y = var_9888_to_fp16)[name = string("op_9889_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_115_cast_fp16 = add(x = var_9889_cast_fp16, y = attention_mask)[name = string("scores_115_cast_fp16")];
            int32 var_9892 = const()[name = string("op_9892"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_115_cast_fp16 = softmax(axis = var_9892, x = scores_115_cast_fp16)[name = string("probs_115_cast_fp16")];
            bool var_9901_transpose_x_0 = const()[name = string("op_9901_transpose_x_0"), val = bool(false)];
            bool var_9901_transpose_y_0 = const()[name = string("op_9901_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_9901_cast_fp16 = matmul(transpose_x = var_9901_transpose_x_0, transpose_y = var_9901_transpose_y_0, x = probs_115_cast_fp16, y = v_t_113_cast_fp16)[name = string("op_9901_cast_fp16")];
            bool var_9920_transpose_x_0 = const()[name = string("op_9920_transpose_x_0"), val = bool(false)];
            bool var_9920_transpose_y_0 = const()[name = string("op_9920_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_424_perm_0 = const()[name = string("transpose_424_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_424 = transpose(perm = transpose_424_perm_0, x = q_h_117_cast_fp16)[name = string("transpose_473")];
            tensor<fp16, [1, 1, ?, 128]> var_9920_cast_fp16 = matmul(transpose_x = var_9920_transpose_x_0, transpose_y = var_9920_transpose_y_0, x = transpose_424, y = transpose_419)[name = string("op_9920_cast_fp16")];
            fp16 var_9921_to_fp16 = const()[name = string("op_9921_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_9922_cast_fp16 = mul(x = var_9920_cast_fp16, y = var_9921_to_fp16)[name = string("op_9922_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_117_cast_fp16 = add(x = var_9922_cast_fp16, y = attention_mask)[name = string("scores_117_cast_fp16")];
            int32 var_9925 = const()[name = string("op_9925"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_117_cast_fp16 = softmax(axis = var_9925, x = scores_117_cast_fp16)[name = string("probs_117_cast_fp16")];
            bool var_9934_transpose_x_0 = const()[name = string("op_9934_transpose_x_0"), val = bool(false)];
            bool var_9934_transpose_y_0 = const()[name = string("op_9934_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_9934_cast_fp16 = matmul(transpose_x = var_9934_transpose_x_0, transpose_y = var_9934_transpose_y_0, x = probs_117_cast_fp16, y = v_t_113_cast_fp16)[name = string("op_9934_cast_fp16")];
            bool var_9953_transpose_x_0 = const()[name = string("op_9953_transpose_x_0"), val = bool(false)];
            bool var_9953_transpose_y_0 = const()[name = string("op_9953_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_425_perm_0 = const()[name = string("transpose_425_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_425 = transpose(perm = transpose_425_perm_0, x = q_h_119_cast_fp16)[name = string("transpose_472")];
            tensor<fp16, [1, 1, ?, 128]> var_9953_cast_fp16 = matmul(transpose_x = var_9953_transpose_x_0, transpose_y = var_9953_transpose_y_0, x = transpose_425, y = transpose_419)[name = string("op_9953_cast_fp16")];
            fp16 var_9954_to_fp16 = const()[name = string("op_9954_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_9955_cast_fp16 = mul(x = var_9953_cast_fp16, y = var_9954_to_fp16)[name = string("op_9955_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_119_cast_fp16 = add(x = var_9955_cast_fp16, y = attention_mask)[name = string("scores_119_cast_fp16")];
            int32 var_9958 = const()[name = string("op_9958"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_119_cast_fp16 = softmax(axis = var_9958, x = scores_119_cast_fp16)[name = string("probs_119_cast_fp16")];
            bool var_9967_transpose_x_0 = const()[name = string("op_9967_transpose_x_0"), val = bool(false)];
            bool var_9967_transpose_y_0 = const()[name = string("op_9967_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_9967_cast_fp16 = matmul(transpose_x = var_9967_transpose_x_0, transpose_y = var_9967_transpose_y_0, x = probs_119_cast_fp16, y = v_t_113_cast_fp16)[name = string("op_9967_cast_fp16")];
            bool input_115_interleave_0 = const()[name = string("input_115_interleave_0"), val = bool(false)];
            int32 const_159 = const()[name = string("const_159"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_115_cast_fp16 = concat(axis = const_159, interleave = input_115_interleave_0, values = (var_9868_cast_fp16, var_9901_cast_fp16, var_9934_cast_fp16, var_9967_cast_fp16))[name = string("input_115_cast_fp16")];
            string x_415_pad_type_0 = const()[name = string("x_415_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_415_strides_0 = const()[name = string("x_415_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_415_pad_0 = const()[name = string("x_415_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_415_dilations_0 = const()[name = string("x_415_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_415_groups_0 = const()[name = string("x_415_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_134_perm_0 = const()[name = string("transpose_134_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_14_weight_to_fp16 = const()[name = string("o_proj_14_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493586560)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_134_cast_fp16 = transpose(perm = transpose_134_perm_0, x = input_115_cast_fp16)[name = string("transpose_471")];
            tensor<fp16, [1, 640, 1, ?]> x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = o_proj_14_weight_to_fp16, x = transpose_134_cast_fp16)[name = string("x_415_cast_fp16")];
            fp16 var_9988_promoted_to_fp16 = const()[name = string("op_9988_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_9989_cast_fp16 = pow(x = x_415_cast_fp16, y = var_9988_promoted_to_fp16)[name = string("op_9989_cast_fp16")];
            tensor<int32, [1]> var_265_axes_0 = const()[name = string("var_265_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_265_keep_dims_0 = const()[name = string("var_265_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_265_cast_fp16 = reduce_mean(axes = var_265_axes_0, keep_dims = var_265_keep_dims_0, x = var_9989_cast_fp16)[name = string("var_265_cast_fp16")];
            fp16 var_9996_to_fp16 = const()[name = string("op_9996_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_9997_cast_fp16 = add(x = var_265_cast_fp16, y = var_9996_to_fp16)[name = string("op_9997_cast_fp16")];
            fp32 var_9998_epsilon_0 = const()[name = string("op_9998_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_9998_cast_fp16 = rsqrt(epsilon = var_9998_epsilon_0, x = var_9997_cast_fp16)[name = string("op_9998_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_265_cast_fp16 = mul(x = x_415_cast_fp16, y = var_9998_cast_fp16)[name = string("x_normed_265_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_10008_to_fp16 = const()[name = string("op_10008_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494897344)))];
            tensor<fp16, [1, 640, 1, ?]> attn_59_cast_fp16 = mul(x = x_normed_265_cast_fp16, y = var_10008_to_fp16)[name = string("attn_59_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_417_cast_fp16 = add(x = x_393_cast_fp16, y = attn_59_cast_fp16)[name = string("x_417_cast_fp16")];
            fp16 var_10012_promoted_to_fp16 = const()[name = string("op_10012_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_10013_cast_fp16 = pow(x = x_417_cast_fp16, y = var_10012_promoted_to_fp16)[name = string("op_10013_cast_fp16")];
            tensor<int32, [1]> var_267_axes_0 = const()[name = string("var_267_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_267_keep_dims_0 = const()[name = string("var_267_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_267_cast_fp16 = reduce_mean(axes = var_267_axes_0, keep_dims = var_267_keep_dims_0, x = var_10013_cast_fp16)[name = string("var_267_cast_fp16")];
            fp16 var_10020_to_fp16 = const()[name = string("op_10020_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_10021_cast_fp16 = add(x = var_267_cast_fp16, y = var_10020_to_fp16)[name = string("op_10021_cast_fp16")];
            fp32 var_10022_epsilon_0 = const()[name = string("op_10022_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_10022_cast_fp16 = rsqrt(epsilon = var_10022_epsilon_0, x = var_10021_cast_fp16)[name = string("op_10022_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_267_cast_fp16 = mul(x = x_417_cast_fp16, y = var_10022_cast_fp16)[name = string("x_normed_267_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_10032_to_fp16 = const()[name = string("op_10032_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494898688)))];
            tensor<fp16, [1, 640, 1, ?]> input_117_cast_fp16 = mul(x = x_normed_267_cast_fp16, y = var_10032_to_fp16)[name = string("input_117_cast_fp16")];
            string gate_29_pad_type_0 = const()[name = string("gate_29_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_29_strides_0 = const()[name = string("gate_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_29_pad_0 = const()[name = string("gate_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_29_dilations_0 = const()[name = string("gate_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_29_groups_0 = const()[name = string("gate_29_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_14_weight_to_fp16 = const()[name = string("gate_proj_14_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494900032)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_29_cast_fp16 = conv(dilations = gate_29_dilations_0, groups = gate_29_groups_0, pad = gate_29_pad_0, pad_type = gate_29_pad_type_0, strides = gate_29_strides_0, weight = gate_proj_14_weight_to_fp16, x = input_117_cast_fp16)[name = string("gate_29_cast_fp16")];
            string up_29_pad_type_0 = const()[name = string("up_29_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_29_strides_0 = const()[name = string("up_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_29_pad_0 = const()[name = string("up_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_29_dilations_0 = const()[name = string("up_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_29_groups_0 = const()[name = string("up_29_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_14_weight_to_fp16 = const()[name = string("up_proj_14_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(497521536)))];
            tensor<fp16, [1, 2048, 1, ?]> up_29_cast_fp16 = conv(dilations = up_29_dilations_0, groups = up_29_groups_0, pad = up_29_pad_0, pad_type = up_29_pad_type_0, strides = up_29_strides_0, weight = up_proj_14_weight_to_fp16, x = input_117_cast_fp16)[name = string("up_29_cast_fp16")];
            string var_10057_mode_0 = const()[name = string("op_10057_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_10057_cast_fp16 = gelu(mode = var_10057_mode_0, x = gate_29_cast_fp16)[name = string("op_10057_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_119_cast_fp16 = mul(x = var_10057_cast_fp16, y = up_29_cast_fp16)[name = string("input_119_cast_fp16")];
            string x_419_pad_type_0 = const()[name = string("x_419_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_419_strides_0 = const()[name = string("x_419_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_419_pad_0 = const()[name = string("x_419_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_419_dilations_0 = const()[name = string("x_419_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_419_groups_0 = const()[name = string("x_419_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_14_weight_to_fp16 = const()[name = string("down_proj_14_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500143040)))];
            tensor<fp16, [1, 640, 1, ?]> x_419_cast_fp16 = conv(dilations = x_419_dilations_0, groups = x_419_groups_0, pad = x_419_pad_0, pad_type = x_419_pad_type_0, strides = x_419_strides_0, weight = down_proj_14_weight_to_fp16, x = input_119_cast_fp16)[name = string("x_419_cast_fp16")];
            fp16 var_10070_promoted_to_fp16 = const()[name = string("op_10070_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_10071_cast_fp16 = pow(x = x_419_cast_fp16, y = var_10070_promoted_to_fp16)[name = string("op_10071_cast_fp16")];
            tensor<int32, [1]> var_269_axes_0 = const()[name = string("var_269_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_269_keep_dims_0 = const()[name = string("var_269_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_269_cast_fp16 = reduce_mean(axes = var_269_axes_0, keep_dims = var_269_keep_dims_0, x = var_10071_cast_fp16)[name = string("var_269_cast_fp16")];
            fp16 var_10078_to_fp16 = const()[name = string("op_10078_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_10079_cast_fp16 = add(x = var_269_cast_fp16, y = var_10078_to_fp16)[name = string("op_10079_cast_fp16")];
            fp32 var_10080_epsilon_0 = const()[name = string("op_10080_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_10080_cast_fp16 = rsqrt(epsilon = var_10080_epsilon_0, x = var_10079_cast_fp16)[name = string("op_10080_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_269_cast_fp16 = mul(x = x_419_cast_fp16, y = var_10080_cast_fp16)[name = string("x_normed_269_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_10090_to_fp16 = const()[name = string("op_10090_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502764544)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_29_cast_fp16 = mul(x = x_normed_269_cast_fp16, y = var_10090_to_fp16)[name = string("mlp_29_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_421_cast_fp16 = add(x = x_417_cast_fp16, y = mlp_29_cast_fp16)[name = string("x_421_cast_fp16")];
            fp16 var_10094_promoted_to_fp16 = const()[name = string("op_10094_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_10095_cast_fp16 = pow(x = x_421_cast_fp16, y = var_10094_promoted_to_fp16)[name = string("op_10095_cast_fp16")];
            tensor<int32, [1]> var_271_axes_0 = const()[name = string("var_271_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_271_keep_dims_0 = const()[name = string("var_271_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_271_cast_fp16 = reduce_mean(axes = var_271_axes_0, keep_dims = var_271_keep_dims_0, x = var_10095_cast_fp16)[name = string("var_271_cast_fp16")];
            fp16 var_10102_to_fp16 = const()[name = string("op_10102_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_10103_cast_fp16 = add(x = var_271_cast_fp16, y = var_10102_to_fp16)[name = string("op_10103_cast_fp16")];
            fp32 var_10104_epsilon_0 = const()[name = string("op_10104_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_10104_cast_fp16 = rsqrt(epsilon = var_10104_epsilon_0, x = var_10103_cast_fp16)[name = string("op_10104_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_271_cast_fp16 = mul(x = x_421_cast_fp16, y = var_10104_cast_fp16)[name = string("x_normed_271_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_10114_to_fp16 = const()[name = string("op_10114_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502765888)))];
            tensor<fp16, [1, 640, 1, ?]> input_121_cast_fp16 = mul(x = x_normed_271_cast_fp16, y = var_10114_to_fp16)[name = string("input_121_cast_fp16")];
            string var_10126_pad_type_0 = const()[name = string("op_10126_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_10126_strides_0 = const()[name = string("op_10126_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10126_pad_0 = const()[name = string("op_10126_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10126_dilations_0 = const()[name = string("op_10126_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_10126_groups_0 = const()[name = string("op_10126_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_15_weight_to_fp16 = const()[name = string("q_proj_15_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502767232)))];
            tensor<fp16, [1, 1024, 1, ?]> var_10126_cast_fp16 = conv(dilations = var_10126_dilations_0, groups = var_10126_groups_0, pad = var_10126_pad_0, pad_type = var_10126_pad_type_0, strides = var_10126_strides_0, weight = q_proj_15_weight_to_fp16, x = input_121_cast_fp16)[name = string("op_10126_cast_fp16")];
            string var_10137_pad_type_0 = const()[name = string("op_10137_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_10137_strides_0 = const()[name = string("op_10137_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10137_pad_0 = const()[name = string("op_10137_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10137_dilations_0 = const()[name = string("op_10137_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_10137_groups_0 = const()[name = string("op_10137_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_15_weight_to_fp16 = const()[name = string("k_proj_15_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(504078016)))];
            tensor<fp16, [1, 256, 1, ?]> var_10137_cast_fp16 = conv(dilations = var_10137_dilations_0, groups = var_10137_groups_0, pad = var_10137_pad_0, pad_type = var_10137_pad_type_0, strides = var_10137_strides_0, weight = k_proj_15_weight_to_fp16, x = input_121_cast_fp16)[name = string("op_10137_cast_fp16")];
            string var_10148_pad_type_0 = const()[name = string("op_10148_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_10148_strides_0 = const()[name = string("op_10148_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10148_pad_0 = const()[name = string("op_10148_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10148_dilations_0 = const()[name = string("op_10148_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_10148_groups_0 = const()[name = string("op_10148_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_15_weight_to_fp16 = const()[name = string("v_proj_15_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(504405760)))];
            tensor<fp16, [1, 256, 1, ?]> var_10148_cast_fp16 = conv(dilations = var_10148_dilations_0, groups = var_10148_groups_0, pad = var_10148_pad_0, pad_type = var_10148_pad_type_0, strides = var_10148_strides_0, weight = v_proj_15_weight_to_fp16, x = input_121_cast_fp16)[name = string("op_10148_cast_fp16")];
            tensor<int32, [4]> tile_45 = const()[name = string("tile_45"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_10151_axis_0 = const()[name = string("op_10151_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_10151_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_10151_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_10151_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_10151_cast_fp16_3 = split(axis = var_10151_axis_0, split_sizes = tile_45, x = var_10126_cast_fp16)[name = string("op_10151_cast_fp16")];
            fp16 var_10164_promoted_to_fp16 = const()[name = string("op_10164_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_10165_cast_fp16 = pow(x = var_10151_cast_fp16_0, y = var_10164_promoted_to_fp16)[name = string("op_10165_cast_fp16")];
            tensor<int32, [1]> var_273_axes_0 = const()[name = string("var_273_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_273_keep_dims_0 = const()[name = string("var_273_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_273_cast_fp16 = reduce_mean(axes = var_273_axes_0, keep_dims = var_273_keep_dims_0, x = var_10165_cast_fp16)[name = string("var_273_cast_fp16")];
            fp16 var_10172_to_fp16 = const()[name = string("op_10172_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_10173_cast_fp16 = add(x = var_273_cast_fp16, y = var_10172_to_fp16)[name = string("op_10173_cast_fp16")];
            fp32 var_10174_epsilon_0 = const()[name = string("op_10174_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_10174_cast_fp16 = rsqrt(epsilon = var_10174_epsilon_0, x = var_10173_cast_fp16)[name = string("op_10174_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_273_cast_fp16 = mul(x = var_10151_cast_fp16_0, y = var_10174_cast_fp16)[name = string("x_normed_273_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_10184_to_fp16 = const()[name = string("op_10184_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(504733504)))];
            tensor<fp16, [1, 256, 1, ?]> x_425_cast_fp16 = mul(x = x_normed_273_cast_fp16, y = var_10184_to_fp16)[name = string("x_425_cast_fp16")];
            tensor<int32, [4]> x_first_151_begin_0 = const()[name = string("x_first_151_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_151_end_0 = const()[name = string("x_first_151_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_151_end_mask_0 = const()[name = string("x_first_151_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_151_cast_fp16 = slice_by_index(begin = x_first_151_begin_0, end = x_first_151_end_0, end_mask = x_first_151_end_mask_0, x = x_425_cast_fp16)[name = string("x_first_151_cast_fp16")];
            tensor<int32, [4]> x_second_151_begin_0 = const()[name = string("x_second_151_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_151_end_0 = const()[name = string("x_second_151_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_151_end_mask_0 = const()[name = string("x_second_151_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_151_cast_fp16 = slice_by_index(begin = x_second_151_begin_0, end = x_second_151_end_0, end_mask = x_second_151_end_mask_0, x = x_425_cast_fp16)[name = string("x_second_151_cast_fp16")];
            fp16 const_124_promoted_to_fp16 = const()[name = string("const_124_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_10206_cast_fp16 = mul(x = x_second_151_cast_fp16, y = const_124_promoted_to_fp16)[name = string("op_10206_cast_fp16")];
            int32 var_10208 = const()[name = string("op_10208"), val = int32(1)];
            bool x_rot_151_interleave_0 = const()[name = string("x_rot_151_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_151_cast_fp16 = concat(axis = var_10208, interleave = x_rot_151_interleave_0, values = (var_10206_cast_fp16, x_first_151_cast_fp16))[name = string("x_rot_151_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_10224_cast_fp16 = mul(x = x_425_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_10224_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_10225_cast_fp16 = mul(x = x_rot_151_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_10225_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_121_cast_fp16 = add(x = var_10224_cast_fp16, y = var_10225_cast_fp16)[name = string("q_h_121_cast_fp16")];
            fp16 var_10228_promoted_to_fp16 = const()[name = string("op_10228_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_10229_cast_fp16 = pow(x = var_10151_cast_fp16_1, y = var_10228_promoted_to_fp16)[name = string("op_10229_cast_fp16")];
            tensor<int32, [1]> var_275_axes_0 = const()[name = string("var_275_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_275_keep_dims_0 = const()[name = string("var_275_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_275_cast_fp16 = reduce_mean(axes = var_275_axes_0, keep_dims = var_275_keep_dims_0, x = var_10229_cast_fp16)[name = string("var_275_cast_fp16")];
            fp16 var_10236_to_fp16 = const()[name = string("op_10236_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_10237_cast_fp16 = add(x = var_275_cast_fp16, y = var_10236_to_fp16)[name = string("op_10237_cast_fp16")];
            fp32 var_10238_epsilon_0 = const()[name = string("op_10238_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_10238_cast_fp16 = rsqrt(epsilon = var_10238_epsilon_0, x = var_10237_cast_fp16)[name = string("op_10238_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_275_cast_fp16 = mul(x = var_10151_cast_fp16_1, y = var_10238_cast_fp16)[name = string("x_normed_275_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_429_cast_fp16 = mul(x = x_normed_275_cast_fp16, y = var_10184_to_fp16)[name = string("x_429_cast_fp16")];
            tensor<int32, [4]> x_first_153_begin_0 = const()[name = string("x_first_153_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_153_end_0 = const()[name = string("x_first_153_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_153_end_mask_0 = const()[name = string("x_first_153_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_153_cast_fp16 = slice_by_index(begin = x_first_153_begin_0, end = x_first_153_end_0, end_mask = x_first_153_end_mask_0, x = x_429_cast_fp16)[name = string("x_first_153_cast_fp16")];
            tensor<int32, [4]> x_second_153_begin_0 = const()[name = string("x_second_153_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_153_end_0 = const()[name = string("x_second_153_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_153_end_mask_0 = const()[name = string("x_second_153_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_153_cast_fp16 = slice_by_index(begin = x_second_153_begin_0, end = x_second_153_end_0, end_mask = x_second_153_end_mask_0, x = x_429_cast_fp16)[name = string("x_second_153_cast_fp16")];
            fp16 const_125_promoted_to_fp16 = const()[name = string("const_125_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_10270_cast_fp16 = mul(x = x_second_153_cast_fp16, y = const_125_promoted_to_fp16)[name = string("op_10270_cast_fp16")];
            int32 var_10272 = const()[name = string("op_10272"), val = int32(1)];
            bool x_rot_153_interleave_0 = const()[name = string("x_rot_153_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_153_cast_fp16 = concat(axis = var_10272, interleave = x_rot_153_interleave_0, values = (var_10270_cast_fp16, x_first_153_cast_fp16))[name = string("x_rot_153_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_10288_cast_fp16 = mul(x = x_429_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_10288_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_10289_cast_fp16 = mul(x = x_rot_153_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_10289_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_123_cast_fp16 = add(x = var_10288_cast_fp16, y = var_10289_cast_fp16)[name = string("q_h_123_cast_fp16")];
            fp16 var_10292_promoted_to_fp16 = const()[name = string("op_10292_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_10293_cast_fp16 = pow(x = var_10151_cast_fp16_2, y = var_10292_promoted_to_fp16)[name = string("op_10293_cast_fp16")];
            tensor<int32, [1]> var_277_axes_0 = const()[name = string("var_277_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_277_keep_dims_0 = const()[name = string("var_277_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_277_cast_fp16 = reduce_mean(axes = var_277_axes_0, keep_dims = var_277_keep_dims_0, x = var_10293_cast_fp16)[name = string("var_277_cast_fp16")];
            fp16 var_10300_to_fp16 = const()[name = string("op_10300_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_10301_cast_fp16 = add(x = var_277_cast_fp16, y = var_10300_to_fp16)[name = string("op_10301_cast_fp16")];
            fp32 var_10302_epsilon_0 = const()[name = string("op_10302_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_10302_cast_fp16 = rsqrt(epsilon = var_10302_epsilon_0, x = var_10301_cast_fp16)[name = string("op_10302_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_277_cast_fp16 = mul(x = var_10151_cast_fp16_2, y = var_10302_cast_fp16)[name = string("x_normed_277_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_433_cast_fp16 = mul(x = x_normed_277_cast_fp16, y = var_10184_to_fp16)[name = string("x_433_cast_fp16")];
            tensor<int32, [4]> x_first_155_begin_0 = const()[name = string("x_first_155_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_155_end_0 = const()[name = string("x_first_155_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_155_end_mask_0 = const()[name = string("x_first_155_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_155_cast_fp16 = slice_by_index(begin = x_first_155_begin_0, end = x_first_155_end_0, end_mask = x_first_155_end_mask_0, x = x_433_cast_fp16)[name = string("x_first_155_cast_fp16")];
            tensor<int32, [4]> x_second_155_begin_0 = const()[name = string("x_second_155_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_155_end_0 = const()[name = string("x_second_155_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_155_end_mask_0 = const()[name = string("x_second_155_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_155_cast_fp16 = slice_by_index(begin = x_second_155_begin_0, end = x_second_155_end_0, end_mask = x_second_155_end_mask_0, x = x_433_cast_fp16)[name = string("x_second_155_cast_fp16")];
            fp16 const_126_promoted_to_fp16 = const()[name = string("const_126_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_10334_cast_fp16 = mul(x = x_second_155_cast_fp16, y = const_126_promoted_to_fp16)[name = string("op_10334_cast_fp16")];
            int32 var_10336 = const()[name = string("op_10336"), val = int32(1)];
            bool x_rot_155_interleave_0 = const()[name = string("x_rot_155_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_155_cast_fp16 = concat(axis = var_10336, interleave = x_rot_155_interleave_0, values = (var_10334_cast_fp16, x_first_155_cast_fp16))[name = string("x_rot_155_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_10352_cast_fp16 = mul(x = x_433_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_10352_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_10353_cast_fp16 = mul(x = x_rot_155_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_10353_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_125_cast_fp16 = add(x = var_10352_cast_fp16, y = var_10353_cast_fp16)[name = string("q_h_125_cast_fp16")];
            fp16 var_10356_promoted_to_fp16 = const()[name = string("op_10356_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_10357_cast_fp16 = pow(x = var_10151_cast_fp16_3, y = var_10356_promoted_to_fp16)[name = string("op_10357_cast_fp16")];
            tensor<int32, [1]> var_279_axes_0 = const()[name = string("var_279_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_279_keep_dims_0 = const()[name = string("var_279_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_279_cast_fp16 = reduce_mean(axes = var_279_axes_0, keep_dims = var_279_keep_dims_0, x = var_10357_cast_fp16)[name = string("var_279_cast_fp16")];
            fp16 var_10364_to_fp16 = const()[name = string("op_10364_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_10365_cast_fp16 = add(x = var_279_cast_fp16, y = var_10364_to_fp16)[name = string("op_10365_cast_fp16")];
            fp32 var_10366_epsilon_0 = const()[name = string("op_10366_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_10366_cast_fp16 = rsqrt(epsilon = var_10366_epsilon_0, x = var_10365_cast_fp16)[name = string("op_10366_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_279_cast_fp16 = mul(x = var_10151_cast_fp16_3, y = var_10366_cast_fp16)[name = string("x_normed_279_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_437_cast_fp16 = mul(x = x_normed_279_cast_fp16, y = var_10184_to_fp16)[name = string("x_437_cast_fp16")];
            tensor<int32, [4]> x_first_157_begin_0 = const()[name = string("x_first_157_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_157_end_0 = const()[name = string("x_first_157_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_157_end_mask_0 = const()[name = string("x_first_157_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_157_cast_fp16 = slice_by_index(begin = x_first_157_begin_0, end = x_first_157_end_0, end_mask = x_first_157_end_mask_0, x = x_437_cast_fp16)[name = string("x_first_157_cast_fp16")];
            tensor<int32, [4]> x_second_157_begin_0 = const()[name = string("x_second_157_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_157_end_0 = const()[name = string("x_second_157_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_157_end_mask_0 = const()[name = string("x_second_157_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_157_cast_fp16 = slice_by_index(begin = x_second_157_begin_0, end = x_second_157_end_0, end_mask = x_second_157_end_mask_0, x = x_437_cast_fp16)[name = string("x_second_157_cast_fp16")];
            fp16 const_127_promoted_to_fp16 = const()[name = string("const_127_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_10398_cast_fp16 = mul(x = x_second_157_cast_fp16, y = const_127_promoted_to_fp16)[name = string("op_10398_cast_fp16")];
            int32 var_10400 = const()[name = string("op_10400"), val = int32(1)];
            bool x_rot_157_interleave_0 = const()[name = string("x_rot_157_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_157_cast_fp16 = concat(axis = var_10400, interleave = x_rot_157_interleave_0, values = (var_10398_cast_fp16, x_first_157_cast_fp16))[name = string("x_rot_157_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_10416_cast_fp16 = mul(x = x_437_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_10416_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_10417_cast_fp16 = mul(x = x_rot_157_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_10417_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_127_cast_fp16 = add(x = var_10416_cast_fp16, y = var_10417_cast_fp16)[name = string("q_h_127_cast_fp16")];
            fp16 var_10420_promoted_to_fp16 = const()[name = string("op_10420_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_10421_cast_fp16 = pow(x = var_10137_cast_fp16, y = var_10420_promoted_to_fp16)[name = string("op_10421_cast_fp16")];
            tensor<int32, [1]> var_281_axes_0 = const()[name = string("var_281_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_281_keep_dims_0 = const()[name = string("var_281_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_281_cast_fp16 = reduce_mean(axes = var_281_axes_0, keep_dims = var_281_keep_dims_0, x = var_10421_cast_fp16)[name = string("var_281_cast_fp16")];
            fp16 var_10428_to_fp16 = const()[name = string("op_10428_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_10429_cast_fp16 = add(x = var_281_cast_fp16, y = var_10428_to_fp16)[name = string("op_10429_cast_fp16")];
            fp32 var_10430_epsilon_0 = const()[name = string("op_10430_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_10430_cast_fp16 = rsqrt(epsilon = var_10430_epsilon_0, x = var_10429_cast_fp16)[name = string("op_10430_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_281_cast_fp16 = mul(x = var_10137_cast_fp16, y = var_10430_cast_fp16)[name = string("x_normed_281_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_10440_to_fp16 = const()[name = string("op_10440_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(504734080)))];
            tensor<fp16, [1, 256, 1, ?]> x_441_cast_fp16 = mul(x = x_normed_281_cast_fp16, y = var_10440_to_fp16)[name = string("x_441_cast_fp16")];
            tensor<int32, [4]> x_first_159_begin_0 = const()[name = string("x_first_159_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_159_end_0 = const()[name = string("x_first_159_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_159_end_mask_0 = const()[name = string("x_first_159_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_159_cast_fp16 = slice_by_index(begin = x_first_159_begin_0, end = x_first_159_end_0, end_mask = x_first_159_end_mask_0, x = x_441_cast_fp16)[name = string("x_first_159_cast_fp16")];
            tensor<int32, [4]> x_second_159_begin_0 = const()[name = string("x_second_159_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_159_end_0 = const()[name = string("x_second_159_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_159_end_mask_0 = const()[name = string("x_second_159_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_159_cast_fp16 = slice_by_index(begin = x_second_159_begin_0, end = x_second_159_end_0, end_mask = x_second_159_end_mask_0, x = x_441_cast_fp16)[name = string("x_second_159_cast_fp16")];
            fp16 const_128_promoted_to_fp16 = const()[name = string("const_128_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_10462_cast_fp16 = mul(x = x_second_159_cast_fp16, y = const_128_promoted_to_fp16)[name = string("op_10462_cast_fp16")];
            int32 var_10464 = const()[name = string("op_10464"), val = int32(1)];
            bool x_rot_159_interleave_0 = const()[name = string("x_rot_159_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_159_cast_fp16 = concat(axis = var_10464, interleave = x_rot_159_interleave_0, values = (var_10462_cast_fp16, x_first_159_cast_fp16))[name = string("x_rot_159_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_10480_cast_fp16 = mul(x = x_441_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_10480_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_10481_cast_fp16 = mul(x = x_rot_159_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_10481_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_31_cast_fp16 = add(x = var_10480_cast_fp16, y = var_10481_cast_fp16)[name = string("k_new_31_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_30 = read_state(input = k_cache_15)[name = string("read_state_30")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_93_cast_fp16 = mul(x = read_state_30, y = one_minus_1_cast_fp16)[name = string("k_cache_93_cast_fp16")];
            write_state(data = k_cache_93_cast_fp16, input = k_cache_15)[name = string("coreml_update_state_132_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_132 = read_state(input = k_cache_15)[name = string("coreml_update_state_132")];
            tensor<fp16, [1, 256, 1, 128]> var_10488_cast_fp16 = mul(x = k_new_31_cast_fp16, y = write_mask)[name = string("op_10488_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_95_cast_fp16 = add(x = coreml_update_state_132, y = var_10488_cast_fp16)[name = string("k_cache_95_cast_fp16")];
            write_state(data = k_cache_95_cast_fp16, input = k_cache_15)[name = string("coreml_update_state_133_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_133 = read_state(input = k_cache_15)[name = string("coreml_update_state_133")];
            tensor<fp16, [1, 256, 1, 128]> read_state_31 = read_state(input = v_cache_15)[name = string("read_state_31")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_93_cast_fp16 = mul(x = read_state_31, y = one_minus_1_cast_fp16)[name = string("v_cache_93_cast_fp16")];
            write_state(data = v_cache_93_cast_fp16, input = v_cache_15)[name = string("coreml_update_state_134_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_134 = read_state(input = v_cache_15)[name = string("coreml_update_state_134")];
            tensor<fp16, [1, 256, 1, 128]> var_10492_cast_fp16 = mul(x = var_10148_cast_fp16, y = write_mask)[name = string("op_10492_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_95_cast_fp16 = add(x = coreml_update_state_134, y = var_10492_cast_fp16)[name = string("v_cache_95_cast_fp16")];
            write_state(data = v_cache_95_cast_fp16, input = v_cache_15)[name = string("coreml_update_state_135_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_135 = read_state(input = v_cache_15)[name = string("coreml_update_state_135")];
            bool var_10507_transpose_x_0 = const()[name = string("op_10507_transpose_x_0"), val = bool(false)];
            bool var_10507_transpose_y_0 = const()[name = string("op_10507_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_426_perm_0 = const()[name = string("transpose_426_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_427_perm_0 = const()[name = string("transpose_427_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_427 = transpose(perm = transpose_427_perm_0, x = coreml_update_state_133)[name = string("transpose_469")];
            tensor<fp16, [1, 1, ?, 256]> transpose_426 = transpose(perm = transpose_426_perm_0, x = q_h_121_cast_fp16)[name = string("transpose_470")];
            tensor<fp16, [1, 1, ?, 128]> var_10507_cast_fp16 = matmul(transpose_x = var_10507_transpose_x_0, transpose_y = var_10507_transpose_y_0, x = transpose_426, y = transpose_427)[name = string("op_10507_cast_fp16")];
            fp16 var_10508_to_fp16 = const()[name = string("op_10508_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_10509_cast_fp16 = mul(x = var_10507_cast_fp16, y = var_10508_to_fp16)[name = string("op_10509_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_121_cast_fp16 = add(x = var_10509_cast_fp16, y = attention_mask)[name = string("scores_121_cast_fp16")];
            int32 var_10512 = const()[name = string("op_10512"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_121_cast_fp16 = softmax(axis = var_10512, x = scores_121_cast_fp16)[name = string("probs_121_cast_fp16")];
            tensor<int32, [4]> var_10519 = const()[name = string("op_10519"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_10521_transpose_x_0 = const()[name = string("op_10521_transpose_x_0"), val = bool(false)];
            bool var_10521_transpose_y_0 = const()[name = string("op_10521_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_121_cast_fp16 = transpose(perm = var_10519, x = coreml_update_state_135)[name = string("transpose_468")];
            tensor<fp16, [1, 1, ?, 256]> var_10521_cast_fp16 = matmul(transpose_x = var_10521_transpose_x_0, transpose_y = var_10521_transpose_y_0, x = probs_121_cast_fp16, y = v_t_121_cast_fp16)[name = string("op_10521_cast_fp16")];
            bool var_10540_transpose_x_0 = const()[name = string("op_10540_transpose_x_0"), val = bool(false)];
            bool var_10540_transpose_y_0 = const()[name = string("op_10540_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_431_perm_0 = const()[name = string("transpose_431_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_431 = transpose(perm = transpose_431_perm_0, x = q_h_123_cast_fp16)[name = string("transpose_467")];
            tensor<fp16, [1, 1, ?, 128]> var_10540_cast_fp16 = matmul(transpose_x = var_10540_transpose_x_0, transpose_y = var_10540_transpose_y_0, x = transpose_431, y = transpose_427)[name = string("op_10540_cast_fp16")];
            fp16 var_10541_to_fp16 = const()[name = string("op_10541_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_10542_cast_fp16 = mul(x = var_10540_cast_fp16, y = var_10541_to_fp16)[name = string("op_10542_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_123_cast_fp16 = add(x = var_10542_cast_fp16, y = attention_mask)[name = string("scores_123_cast_fp16")];
            int32 var_10545 = const()[name = string("op_10545"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_123_cast_fp16 = softmax(axis = var_10545, x = scores_123_cast_fp16)[name = string("probs_123_cast_fp16")];
            bool var_10554_transpose_x_0 = const()[name = string("op_10554_transpose_x_0"), val = bool(false)];
            bool var_10554_transpose_y_0 = const()[name = string("op_10554_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_10554_cast_fp16 = matmul(transpose_x = var_10554_transpose_x_0, transpose_y = var_10554_transpose_y_0, x = probs_123_cast_fp16, y = v_t_121_cast_fp16)[name = string("op_10554_cast_fp16")];
            bool var_10573_transpose_x_0 = const()[name = string("op_10573_transpose_x_0"), val = bool(false)];
            bool var_10573_transpose_y_0 = const()[name = string("op_10573_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_432_perm_0 = const()[name = string("transpose_432_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_432 = transpose(perm = transpose_432_perm_0, x = q_h_125_cast_fp16)[name = string("transpose_466")];
            tensor<fp16, [1, 1, ?, 128]> var_10573_cast_fp16 = matmul(transpose_x = var_10573_transpose_x_0, transpose_y = var_10573_transpose_y_0, x = transpose_432, y = transpose_427)[name = string("op_10573_cast_fp16")];
            fp16 var_10574_to_fp16 = const()[name = string("op_10574_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_10575_cast_fp16 = mul(x = var_10573_cast_fp16, y = var_10574_to_fp16)[name = string("op_10575_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_125_cast_fp16 = add(x = var_10575_cast_fp16, y = attention_mask)[name = string("scores_125_cast_fp16")];
            int32 var_10578 = const()[name = string("op_10578"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_125_cast_fp16 = softmax(axis = var_10578, x = scores_125_cast_fp16)[name = string("probs_125_cast_fp16")];
            bool var_10587_transpose_x_0 = const()[name = string("op_10587_transpose_x_0"), val = bool(false)];
            bool var_10587_transpose_y_0 = const()[name = string("op_10587_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_10587_cast_fp16 = matmul(transpose_x = var_10587_transpose_x_0, transpose_y = var_10587_transpose_y_0, x = probs_125_cast_fp16, y = v_t_121_cast_fp16)[name = string("op_10587_cast_fp16")];
            bool var_10606_transpose_x_0 = const()[name = string("op_10606_transpose_x_0"), val = bool(false)];
            bool var_10606_transpose_y_0 = const()[name = string("op_10606_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_433_perm_0 = const()[name = string("transpose_433_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_433 = transpose(perm = transpose_433_perm_0, x = q_h_127_cast_fp16)[name = string("transpose_465")];
            tensor<fp16, [1, 1, ?, 128]> var_10606_cast_fp16 = matmul(transpose_x = var_10606_transpose_x_0, transpose_y = var_10606_transpose_y_0, x = transpose_433, y = transpose_427)[name = string("op_10606_cast_fp16")];
            fp16 var_10607_to_fp16 = const()[name = string("op_10607_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_10608_cast_fp16 = mul(x = var_10606_cast_fp16, y = var_10607_to_fp16)[name = string("op_10608_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_127_cast_fp16 = add(x = var_10608_cast_fp16, y = attention_mask)[name = string("scores_127_cast_fp16")];
            int32 var_10611 = const()[name = string("op_10611"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_127_cast_fp16 = softmax(axis = var_10611, x = scores_127_cast_fp16)[name = string("probs_127_cast_fp16")];
            bool var_10620_transpose_x_0 = const()[name = string("op_10620_transpose_x_0"), val = bool(false)];
            bool var_10620_transpose_y_0 = const()[name = string("op_10620_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_10620_cast_fp16 = matmul(transpose_x = var_10620_transpose_x_0, transpose_y = var_10620_transpose_y_0, x = probs_127_cast_fp16, y = v_t_121_cast_fp16)[name = string("op_10620_cast_fp16")];
            bool input_123_interleave_0 = const()[name = string("input_123_interleave_0"), val = bool(false)];
            int32 const_160 = const()[name = string("const_160"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_123_cast_fp16 = concat(axis = const_160, interleave = input_123_interleave_0, values = (var_10521_cast_fp16, var_10554_cast_fp16, var_10587_cast_fp16, var_10620_cast_fp16))[name = string("input_123_cast_fp16")];
            string x_443_pad_type_0 = const()[name = string("x_443_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_443_strides_0 = const()[name = string("x_443_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_443_pad_0 = const()[name = string("x_443_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_443_dilations_0 = const()[name = string("x_443_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_443_groups_0 = const()[name = string("x_443_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_143_perm_0 = const()[name = string("transpose_143_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_15_weight_to_fp16 = const()[name = string("o_proj_15_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(504734656)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_143_cast_fp16 = transpose(perm = transpose_143_perm_0, x = input_123_cast_fp16)[name = string("transpose_464")];
            tensor<fp16, [1, 640, 1, ?]> x_443_cast_fp16 = conv(dilations = x_443_dilations_0, groups = x_443_groups_0, pad = x_443_pad_0, pad_type = x_443_pad_type_0, strides = x_443_strides_0, weight = o_proj_15_weight_to_fp16, x = transpose_143_cast_fp16)[name = string("x_443_cast_fp16")];
            fp16 var_10641_promoted_to_fp16 = const()[name = string("op_10641_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_10642_cast_fp16 = pow(x = x_443_cast_fp16, y = var_10641_promoted_to_fp16)[name = string("op_10642_cast_fp16")];
            tensor<int32, [1]> var_283_axes_0 = const()[name = string("var_283_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_283_keep_dims_0 = const()[name = string("var_283_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_283_cast_fp16 = reduce_mean(axes = var_283_axes_0, keep_dims = var_283_keep_dims_0, x = var_10642_cast_fp16)[name = string("var_283_cast_fp16")];
            fp16 var_10649_to_fp16 = const()[name = string("op_10649_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_10650_cast_fp16 = add(x = var_283_cast_fp16, y = var_10649_to_fp16)[name = string("op_10650_cast_fp16")];
            fp32 var_10651_epsilon_0 = const()[name = string("op_10651_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_10651_cast_fp16 = rsqrt(epsilon = var_10651_epsilon_0, x = var_10650_cast_fp16)[name = string("op_10651_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_283_cast_fp16 = mul(x = x_443_cast_fp16, y = var_10651_cast_fp16)[name = string("x_normed_283_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_10661_to_fp16 = const()[name = string("op_10661_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506045440)))];
            tensor<fp16, [1, 640, 1, ?]> attn_63_cast_fp16 = mul(x = x_normed_283_cast_fp16, y = var_10661_to_fp16)[name = string("attn_63_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_445_cast_fp16 = add(x = x_421_cast_fp16, y = attn_63_cast_fp16)[name = string("x_445_cast_fp16")];
            fp16 var_10665_promoted_to_fp16 = const()[name = string("op_10665_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_10666_cast_fp16 = pow(x = x_445_cast_fp16, y = var_10665_promoted_to_fp16)[name = string("op_10666_cast_fp16")];
            tensor<int32, [1]> var_285_axes_0 = const()[name = string("var_285_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_285_keep_dims_0 = const()[name = string("var_285_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_285_cast_fp16 = reduce_mean(axes = var_285_axes_0, keep_dims = var_285_keep_dims_0, x = var_10666_cast_fp16)[name = string("var_285_cast_fp16")];
            fp16 var_10673_to_fp16 = const()[name = string("op_10673_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_10674_cast_fp16 = add(x = var_285_cast_fp16, y = var_10673_to_fp16)[name = string("op_10674_cast_fp16")];
            fp32 var_10675_epsilon_0 = const()[name = string("op_10675_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_10675_cast_fp16 = rsqrt(epsilon = var_10675_epsilon_0, x = var_10674_cast_fp16)[name = string("op_10675_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_285_cast_fp16 = mul(x = x_445_cast_fp16, y = var_10675_cast_fp16)[name = string("x_normed_285_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_10685_to_fp16 = const()[name = string("op_10685_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506046784)))];
            tensor<fp16, [1, 640, 1, ?]> input_125_cast_fp16 = mul(x = x_normed_285_cast_fp16, y = var_10685_to_fp16)[name = string("input_125_cast_fp16")];
            string gate_31_pad_type_0 = const()[name = string("gate_31_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_31_strides_0 = const()[name = string("gate_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_31_pad_0 = const()[name = string("gate_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_31_dilations_0 = const()[name = string("gate_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_31_groups_0 = const()[name = string("gate_31_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_15_weight_to_fp16 = const()[name = string("gate_proj_15_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506048128)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_31_cast_fp16 = conv(dilations = gate_31_dilations_0, groups = gate_31_groups_0, pad = gate_31_pad_0, pad_type = gate_31_pad_type_0, strides = gate_31_strides_0, weight = gate_proj_15_weight_to_fp16, x = input_125_cast_fp16)[name = string("gate_31_cast_fp16")];
            string up_31_pad_type_0 = const()[name = string("up_31_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_31_strides_0 = const()[name = string("up_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_31_pad_0 = const()[name = string("up_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_31_dilations_0 = const()[name = string("up_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_31_groups_0 = const()[name = string("up_31_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_15_weight_to_fp16 = const()[name = string("up_proj_15_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508669632)))];
            tensor<fp16, [1, 2048, 1, ?]> up_31_cast_fp16 = conv(dilations = up_31_dilations_0, groups = up_31_groups_0, pad = up_31_pad_0, pad_type = up_31_pad_type_0, strides = up_31_strides_0, weight = up_proj_15_weight_to_fp16, x = input_125_cast_fp16)[name = string("up_31_cast_fp16")];
            string var_10710_mode_0 = const()[name = string("op_10710_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_10710_cast_fp16 = gelu(mode = var_10710_mode_0, x = gate_31_cast_fp16)[name = string("op_10710_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_127_cast_fp16 = mul(x = var_10710_cast_fp16, y = up_31_cast_fp16)[name = string("input_127_cast_fp16")];
            string x_447_pad_type_0 = const()[name = string("x_447_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_447_strides_0 = const()[name = string("x_447_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_447_pad_0 = const()[name = string("x_447_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_447_dilations_0 = const()[name = string("x_447_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_447_groups_0 = const()[name = string("x_447_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_15_weight_to_fp16 = const()[name = string("down_proj_15_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511291136)))];
            tensor<fp16, [1, 640, 1, ?]> x_447_cast_fp16 = conv(dilations = x_447_dilations_0, groups = x_447_groups_0, pad = x_447_pad_0, pad_type = x_447_pad_type_0, strides = x_447_strides_0, weight = down_proj_15_weight_to_fp16, x = input_127_cast_fp16)[name = string("x_447_cast_fp16")];
            fp16 var_10723_promoted_to_fp16 = const()[name = string("op_10723_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_10724_cast_fp16 = pow(x = x_447_cast_fp16, y = var_10723_promoted_to_fp16)[name = string("op_10724_cast_fp16")];
            tensor<int32, [1]> var_287_axes_0 = const()[name = string("var_287_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_287_keep_dims_0 = const()[name = string("var_287_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_287_cast_fp16 = reduce_mean(axes = var_287_axes_0, keep_dims = var_287_keep_dims_0, x = var_10724_cast_fp16)[name = string("var_287_cast_fp16")];
            fp16 var_10731_to_fp16 = const()[name = string("op_10731_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_10732_cast_fp16 = add(x = var_287_cast_fp16, y = var_10731_to_fp16)[name = string("op_10732_cast_fp16")];
            fp32 var_10733_epsilon_0 = const()[name = string("op_10733_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_10733_cast_fp16 = rsqrt(epsilon = var_10733_epsilon_0, x = var_10732_cast_fp16)[name = string("op_10733_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_287_cast_fp16 = mul(x = x_447_cast_fp16, y = var_10733_cast_fp16)[name = string("x_normed_287_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_10743_to_fp16 = const()[name = string("op_10743_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513912640)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_31_cast_fp16 = mul(x = x_normed_287_cast_fp16, y = var_10743_to_fp16)[name = string("mlp_31_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_449_cast_fp16 = add(x = x_445_cast_fp16, y = mlp_31_cast_fp16)[name = string("x_449_cast_fp16")];
            fp16 var_10747_promoted_to_fp16 = const()[name = string("op_10747_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_10748_cast_fp16 = pow(x = x_449_cast_fp16, y = var_10747_promoted_to_fp16)[name = string("op_10748_cast_fp16")];
            tensor<int32, [1]> var_289_axes_0 = const()[name = string("var_289_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_289_keep_dims_0 = const()[name = string("var_289_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_289_cast_fp16_0 = reduce_mean(axes = var_289_axes_0, keep_dims = var_289_keep_dims_0, x = var_10748_cast_fp16)[name = string("var_289_cast_fp16")];
            fp16 var_10755_to_fp16 = const()[name = string("op_10755_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_10756_cast_fp16 = add(x = var_289_cast_fp16_0, y = var_10755_to_fp16)[name = string("op_10756_cast_fp16")];
            fp32 var_10757_epsilon_0 = const()[name = string("op_10757_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_10757_cast_fp16 = rsqrt(epsilon = var_10757_epsilon_0, x = var_10756_cast_fp16)[name = string("op_10757_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_289_cast_fp16 = mul(x = x_449_cast_fp16, y = var_10757_cast_fp16)[name = string("x_normed_289_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_10767_to_fp16 = const()[name = string("op_10767_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513913984)))];
            tensor<fp16, [1, 640, 1, ?]> input_129_cast_fp16 = mul(x = x_normed_289_cast_fp16, y = var_10767_to_fp16)[name = string("input_129_cast_fp16")];
            string var_10779_pad_type_0 = const()[name = string("op_10779_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_10779_strides_0 = const()[name = string("op_10779_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10779_pad_0 = const()[name = string("op_10779_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10779_dilations_0 = const()[name = string("op_10779_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_10779_groups_0 = const()[name = string("op_10779_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_16_weight_to_fp16 = const()[name = string("q_proj_16_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513915328)))];
            tensor<fp16, [1, 1024, 1, ?]> var_10779_cast_fp16 = conv(dilations = var_10779_dilations_0, groups = var_10779_groups_0, pad = var_10779_pad_0, pad_type = var_10779_pad_type_0, strides = var_10779_strides_0, weight = q_proj_16_weight_to_fp16, x = input_129_cast_fp16)[name = string("op_10779_cast_fp16")];
            string var_10790_pad_type_0 = const()[name = string("op_10790_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_10790_strides_0 = const()[name = string("op_10790_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10790_pad_0 = const()[name = string("op_10790_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10790_dilations_0 = const()[name = string("op_10790_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_10790_groups_0 = const()[name = string("op_10790_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_16_weight_to_fp16 = const()[name = string("k_proj_16_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515226112)))];
            tensor<fp16, [1, 256, 1, ?]> var_10790_cast_fp16 = conv(dilations = var_10790_dilations_0, groups = var_10790_groups_0, pad = var_10790_pad_0, pad_type = var_10790_pad_type_0, strides = var_10790_strides_0, weight = k_proj_16_weight_to_fp16, x = input_129_cast_fp16)[name = string("op_10790_cast_fp16")];
            string var_10801_pad_type_0 = const()[name = string("op_10801_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_10801_strides_0 = const()[name = string("op_10801_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_10801_pad_0 = const()[name = string("op_10801_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_10801_dilations_0 = const()[name = string("op_10801_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_10801_groups_0 = const()[name = string("op_10801_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_16_weight_to_fp16 = const()[name = string("v_proj_16_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515553856)))];
            tensor<fp16, [1, 256, 1, ?]> var_10801_cast_fp16 = conv(dilations = var_10801_dilations_0, groups = var_10801_groups_0, pad = var_10801_pad_0, pad_type = var_10801_pad_type_0, strides = var_10801_strides_0, weight = v_proj_16_weight_to_fp16, x = input_129_cast_fp16)[name = string("op_10801_cast_fp16")];
            tensor<int32, [4]> tile_48 = const()[name = string("tile_48"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_10804_axis_0 = const()[name = string("op_10804_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_10804_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_10804_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_10804_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_10804_cast_fp16_3 = split(axis = var_10804_axis_0, split_sizes = tile_48, x = var_10779_cast_fp16)[name = string("op_10804_cast_fp16")];
            fp16 var_10817_promoted_to_fp16 = const()[name = string("op_10817_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_10818_cast_fp16 = pow(x = var_10804_cast_fp16_0, y = var_10817_promoted_to_fp16)[name = string("op_10818_cast_fp16")];
            tensor<int32, [1]> var_291_axes_0 = const()[name = string("var_291_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_291_keep_dims_0 = const()[name = string("var_291_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_291_cast_fp16 = reduce_mean(axes = var_291_axes_0, keep_dims = var_291_keep_dims_0, x = var_10818_cast_fp16)[name = string("var_291_cast_fp16")];
            fp16 var_10825_to_fp16 = const()[name = string("op_10825_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_10826_cast_fp16 = add(x = var_291_cast_fp16, y = var_10825_to_fp16)[name = string("op_10826_cast_fp16")];
            fp32 var_10827_epsilon_0 = const()[name = string("op_10827_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_10827_cast_fp16 = rsqrt(epsilon = var_10827_epsilon_0, x = var_10826_cast_fp16)[name = string("op_10827_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_291_cast_fp16 = mul(x = var_10804_cast_fp16_0, y = var_10827_cast_fp16)[name = string("x_normed_291_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_10837_to_fp16 = const()[name = string("op_10837_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515881600)))];
            tensor<fp16, [1, 256, 1, ?]> x_453_cast_fp16 = mul(x = x_normed_291_cast_fp16, y = var_10837_to_fp16)[name = string("x_453_cast_fp16")];
            tensor<int32, [4]> x_first_161_begin_0 = const()[name = string("x_first_161_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_161_end_0 = const()[name = string("x_first_161_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_161_end_mask_0 = const()[name = string("x_first_161_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_161_cast_fp16 = slice_by_index(begin = x_first_161_begin_0, end = x_first_161_end_0, end_mask = x_first_161_end_mask_0, x = x_453_cast_fp16)[name = string("x_first_161_cast_fp16")];
            tensor<int32, [4]> x_second_161_begin_0 = const()[name = string("x_second_161_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_161_end_0 = const()[name = string("x_second_161_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_161_end_mask_0 = const()[name = string("x_second_161_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_161_cast_fp16 = slice_by_index(begin = x_second_161_begin_0, end = x_second_161_end_0, end_mask = x_second_161_end_mask_0, x = x_453_cast_fp16)[name = string("x_second_161_cast_fp16")];
            fp16 const_132_promoted_to_fp16 = const()[name = string("const_132_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_10859_cast_fp16 = mul(x = x_second_161_cast_fp16, y = const_132_promoted_to_fp16)[name = string("op_10859_cast_fp16")];
            int32 var_10861 = const()[name = string("op_10861"), val = int32(1)];
            bool x_rot_161_interleave_0 = const()[name = string("x_rot_161_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_161_cast_fp16 = concat(axis = var_10861, interleave = x_rot_161_interleave_0, values = (var_10859_cast_fp16, x_first_161_cast_fp16))[name = string("x_rot_161_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_10877_cast_fp16 = mul(x = x_453_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_10877_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_10878_cast_fp16 = mul(x = x_rot_161_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_10878_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_129_cast_fp16 = add(x = var_10877_cast_fp16, y = var_10878_cast_fp16)[name = string("q_h_129_cast_fp16")];
            fp16 var_10881_promoted_to_fp16 = const()[name = string("op_10881_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_10882_cast_fp16 = pow(x = var_10804_cast_fp16_1, y = var_10881_promoted_to_fp16)[name = string("op_10882_cast_fp16")];
            tensor<int32, [1]> var_293_axes_0 = const()[name = string("var_293_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_293_keep_dims_0 = const()[name = string("var_293_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_293_cast_fp16 = reduce_mean(axes = var_293_axes_0, keep_dims = var_293_keep_dims_0, x = var_10882_cast_fp16)[name = string("var_293_cast_fp16")];
            fp16 var_10889_to_fp16 = const()[name = string("op_10889_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_10890_cast_fp16 = add(x = var_293_cast_fp16, y = var_10889_to_fp16)[name = string("op_10890_cast_fp16")];
            fp32 var_10891_epsilon_0 = const()[name = string("op_10891_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_10891_cast_fp16 = rsqrt(epsilon = var_10891_epsilon_0, x = var_10890_cast_fp16)[name = string("op_10891_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_293_cast_fp16 = mul(x = var_10804_cast_fp16_1, y = var_10891_cast_fp16)[name = string("x_normed_293_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_457_cast_fp16 = mul(x = x_normed_293_cast_fp16, y = var_10837_to_fp16)[name = string("x_457_cast_fp16")];
            tensor<int32, [4]> x_first_163_begin_0 = const()[name = string("x_first_163_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_163_end_0 = const()[name = string("x_first_163_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_163_end_mask_0 = const()[name = string("x_first_163_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_163_cast_fp16 = slice_by_index(begin = x_first_163_begin_0, end = x_first_163_end_0, end_mask = x_first_163_end_mask_0, x = x_457_cast_fp16)[name = string("x_first_163_cast_fp16")];
            tensor<int32, [4]> x_second_163_begin_0 = const()[name = string("x_second_163_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_163_end_0 = const()[name = string("x_second_163_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_163_end_mask_0 = const()[name = string("x_second_163_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_163_cast_fp16 = slice_by_index(begin = x_second_163_begin_0, end = x_second_163_end_0, end_mask = x_second_163_end_mask_0, x = x_457_cast_fp16)[name = string("x_second_163_cast_fp16")];
            fp16 const_133_promoted_to_fp16 = const()[name = string("const_133_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_10923_cast_fp16 = mul(x = x_second_163_cast_fp16, y = const_133_promoted_to_fp16)[name = string("op_10923_cast_fp16")];
            int32 var_10925 = const()[name = string("op_10925"), val = int32(1)];
            bool x_rot_163_interleave_0 = const()[name = string("x_rot_163_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_163_cast_fp16 = concat(axis = var_10925, interleave = x_rot_163_interleave_0, values = (var_10923_cast_fp16, x_first_163_cast_fp16))[name = string("x_rot_163_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_10941_cast_fp16 = mul(x = x_457_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_10941_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_10942_cast_fp16 = mul(x = x_rot_163_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_10942_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_131_cast_fp16 = add(x = var_10941_cast_fp16, y = var_10942_cast_fp16)[name = string("q_h_131_cast_fp16")];
            fp16 var_10945_promoted_to_fp16 = const()[name = string("op_10945_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_10946_cast_fp16 = pow(x = var_10804_cast_fp16_2, y = var_10945_promoted_to_fp16)[name = string("op_10946_cast_fp16")];
            tensor<int32, [1]> var_295_axes_0 = const()[name = string("var_295_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_295_keep_dims_0 = const()[name = string("var_295_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_295_cast_fp16 = reduce_mean(axes = var_295_axes_0, keep_dims = var_295_keep_dims_0, x = var_10946_cast_fp16)[name = string("var_295_cast_fp16")];
            fp16 var_10953_to_fp16 = const()[name = string("op_10953_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_10954_cast_fp16 = add(x = var_295_cast_fp16, y = var_10953_to_fp16)[name = string("op_10954_cast_fp16")];
            fp32 var_10955_epsilon_0 = const()[name = string("op_10955_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_10955_cast_fp16 = rsqrt(epsilon = var_10955_epsilon_0, x = var_10954_cast_fp16)[name = string("op_10955_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_295_cast_fp16 = mul(x = var_10804_cast_fp16_2, y = var_10955_cast_fp16)[name = string("x_normed_295_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_461_cast_fp16 = mul(x = x_normed_295_cast_fp16, y = var_10837_to_fp16)[name = string("x_461_cast_fp16")];
            tensor<int32, [4]> x_first_165_begin_0 = const()[name = string("x_first_165_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_165_end_0 = const()[name = string("x_first_165_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_165_end_mask_0 = const()[name = string("x_first_165_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_165_cast_fp16 = slice_by_index(begin = x_first_165_begin_0, end = x_first_165_end_0, end_mask = x_first_165_end_mask_0, x = x_461_cast_fp16)[name = string("x_first_165_cast_fp16")];
            tensor<int32, [4]> x_second_165_begin_0 = const()[name = string("x_second_165_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_165_end_0 = const()[name = string("x_second_165_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_165_end_mask_0 = const()[name = string("x_second_165_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_165_cast_fp16 = slice_by_index(begin = x_second_165_begin_0, end = x_second_165_end_0, end_mask = x_second_165_end_mask_0, x = x_461_cast_fp16)[name = string("x_second_165_cast_fp16")];
            fp16 const_134_promoted_to_fp16 = const()[name = string("const_134_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_10987_cast_fp16 = mul(x = x_second_165_cast_fp16, y = const_134_promoted_to_fp16)[name = string("op_10987_cast_fp16")];
            int32 var_10989 = const()[name = string("op_10989"), val = int32(1)];
            bool x_rot_165_interleave_0 = const()[name = string("x_rot_165_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_165_cast_fp16 = concat(axis = var_10989, interleave = x_rot_165_interleave_0, values = (var_10987_cast_fp16, x_first_165_cast_fp16))[name = string("x_rot_165_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_11005_cast_fp16 = mul(x = x_461_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_11005_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_11006_cast_fp16 = mul(x = x_rot_165_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_11006_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_133_cast_fp16 = add(x = var_11005_cast_fp16, y = var_11006_cast_fp16)[name = string("q_h_133_cast_fp16")];
            fp16 var_11009_promoted_to_fp16 = const()[name = string("op_11009_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_11010_cast_fp16 = pow(x = var_10804_cast_fp16_3, y = var_11009_promoted_to_fp16)[name = string("op_11010_cast_fp16")];
            tensor<int32, [1]> var_297_axes_0 = const()[name = string("var_297_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_297_keep_dims_0 = const()[name = string("var_297_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_297_cast_fp16 = reduce_mean(axes = var_297_axes_0, keep_dims = var_297_keep_dims_0, x = var_11010_cast_fp16)[name = string("var_297_cast_fp16")];
            fp16 var_11017_to_fp16 = const()[name = string("op_11017_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_11018_cast_fp16 = add(x = var_297_cast_fp16, y = var_11017_to_fp16)[name = string("op_11018_cast_fp16")];
            fp32 var_11019_epsilon_0 = const()[name = string("op_11019_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_11019_cast_fp16 = rsqrt(epsilon = var_11019_epsilon_0, x = var_11018_cast_fp16)[name = string("op_11019_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_297_cast_fp16 = mul(x = var_10804_cast_fp16_3, y = var_11019_cast_fp16)[name = string("x_normed_297_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_465_cast_fp16 = mul(x = x_normed_297_cast_fp16, y = var_10837_to_fp16)[name = string("x_465_cast_fp16")];
            tensor<int32, [4]> x_first_167_begin_0 = const()[name = string("x_first_167_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_167_end_0 = const()[name = string("x_first_167_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_167_end_mask_0 = const()[name = string("x_first_167_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_167_cast_fp16 = slice_by_index(begin = x_first_167_begin_0, end = x_first_167_end_0, end_mask = x_first_167_end_mask_0, x = x_465_cast_fp16)[name = string("x_first_167_cast_fp16")];
            tensor<int32, [4]> x_second_167_begin_0 = const()[name = string("x_second_167_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_167_end_0 = const()[name = string("x_second_167_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_167_end_mask_0 = const()[name = string("x_second_167_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_167_cast_fp16 = slice_by_index(begin = x_second_167_begin_0, end = x_second_167_end_0, end_mask = x_second_167_end_mask_0, x = x_465_cast_fp16)[name = string("x_second_167_cast_fp16")];
            fp16 const_135_promoted_to_fp16 = const()[name = string("const_135_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_11051_cast_fp16 = mul(x = x_second_167_cast_fp16, y = const_135_promoted_to_fp16)[name = string("op_11051_cast_fp16")];
            int32 var_11053 = const()[name = string("op_11053"), val = int32(1)];
            bool x_rot_167_interleave_0 = const()[name = string("x_rot_167_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_167_cast_fp16 = concat(axis = var_11053, interleave = x_rot_167_interleave_0, values = (var_11051_cast_fp16, x_first_167_cast_fp16))[name = string("x_rot_167_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_11069_cast_fp16 = mul(x = x_465_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_11069_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_11070_cast_fp16 = mul(x = x_rot_167_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_11070_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_135_cast_fp16 = add(x = var_11069_cast_fp16, y = var_11070_cast_fp16)[name = string("q_h_135_cast_fp16")];
            fp16 var_11073_promoted_to_fp16 = const()[name = string("op_11073_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_11074_cast_fp16 = pow(x = var_10790_cast_fp16, y = var_11073_promoted_to_fp16)[name = string("op_11074_cast_fp16")];
            tensor<int32, [1]> var_299_axes_0 = const()[name = string("var_299_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_299_keep_dims_0 = const()[name = string("var_299_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_299_cast_fp16 = reduce_mean(axes = var_299_axes_0, keep_dims = var_299_keep_dims_0, x = var_11074_cast_fp16)[name = string("var_299_cast_fp16")];
            fp16 var_11081_to_fp16 = const()[name = string("op_11081_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_11082_cast_fp16 = add(x = var_299_cast_fp16, y = var_11081_to_fp16)[name = string("op_11082_cast_fp16")];
            fp32 var_11083_epsilon_0 = const()[name = string("op_11083_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_11083_cast_fp16 = rsqrt(epsilon = var_11083_epsilon_0, x = var_11082_cast_fp16)[name = string("op_11083_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_299_cast_fp16 = mul(x = var_10790_cast_fp16, y = var_11083_cast_fp16)[name = string("x_normed_299_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_11093_to_fp16 = const()[name = string("op_11093_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515882176)))];
            tensor<fp16, [1, 256, 1, ?]> x_469_cast_fp16 = mul(x = x_normed_299_cast_fp16, y = var_11093_to_fp16)[name = string("x_469_cast_fp16")];
            tensor<int32, [4]> x_first_169_begin_0 = const()[name = string("x_first_169_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_169_end_0 = const()[name = string("x_first_169_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_169_end_mask_0 = const()[name = string("x_first_169_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_169_cast_fp16 = slice_by_index(begin = x_first_169_begin_0, end = x_first_169_end_0, end_mask = x_first_169_end_mask_0, x = x_469_cast_fp16)[name = string("x_first_169_cast_fp16")];
            tensor<int32, [4]> x_second_169_begin_0 = const()[name = string("x_second_169_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_169_end_0 = const()[name = string("x_second_169_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_169_end_mask_0 = const()[name = string("x_second_169_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_169_cast_fp16 = slice_by_index(begin = x_second_169_begin_0, end = x_second_169_end_0, end_mask = x_second_169_end_mask_0, x = x_469_cast_fp16)[name = string("x_second_169_cast_fp16")];
            fp16 const_136_promoted_to_fp16 = const()[name = string("const_136_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_11115_cast_fp16 = mul(x = x_second_169_cast_fp16, y = const_136_promoted_to_fp16)[name = string("op_11115_cast_fp16")];
            int32 var_11117 = const()[name = string("op_11117"), val = int32(1)];
            bool x_rot_169_interleave_0 = const()[name = string("x_rot_169_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_169_cast_fp16 = concat(axis = var_11117, interleave = x_rot_169_interleave_0, values = (var_11115_cast_fp16, x_first_169_cast_fp16))[name = string("x_rot_169_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_11133_cast_fp16 = mul(x = x_469_cast_fp16, y = cos_bc1t_1_cast_fp16)[name = string("op_11133_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_11134_cast_fp16 = mul(x = x_rot_169_cast_fp16, y = sin_bc1t_1_cast_fp16)[name = string("op_11134_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_33_cast_fp16 = add(x = var_11133_cast_fp16, y = var_11134_cast_fp16)[name = string("k_new_33_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_32 = read_state(input = k_cache_16)[name = string("read_state_32")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_99_cast_fp16 = mul(x = read_state_32, y = one_minus_1_cast_fp16)[name = string("k_cache_99_cast_fp16")];
            write_state(data = k_cache_99_cast_fp16, input = k_cache_16)[name = string("coreml_update_state_136_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_136 = read_state(input = k_cache_16)[name = string("coreml_update_state_136")];
            tensor<fp16, [1, 256, 1, 128]> var_11141_cast_fp16 = mul(x = k_new_33_cast_fp16, y = write_mask)[name = string("op_11141_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_101_cast_fp16 = add(x = coreml_update_state_136, y = var_11141_cast_fp16)[name = string("k_cache_101_cast_fp16")];
            write_state(data = k_cache_101_cast_fp16, input = k_cache_16)[name = string("coreml_update_state_137_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_137 = read_state(input = k_cache_16)[name = string("coreml_update_state_137")];
            tensor<fp16, [1, 256, 1, 128]> read_state_33 = read_state(input = v_cache_16)[name = string("read_state_33")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_99_cast_fp16 = mul(x = read_state_33, y = one_minus_1_cast_fp16)[name = string("v_cache_99_cast_fp16")];
            write_state(data = v_cache_99_cast_fp16, input = v_cache_16)[name = string("coreml_update_state_138_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_138 = read_state(input = v_cache_16)[name = string("coreml_update_state_138")];
            tensor<fp16, [1, 256, 1, 128]> var_11145_cast_fp16 = mul(x = var_10801_cast_fp16, y = write_mask)[name = string("op_11145_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_101_cast_fp16 = add(x = coreml_update_state_138, y = var_11145_cast_fp16)[name = string("v_cache_101_cast_fp16")];
            write_state(data = v_cache_101_cast_fp16, input = v_cache_16)[name = string("coreml_update_state_139_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_139 = read_state(input = v_cache_16)[name = string("coreml_update_state_139")];
            bool var_11160_transpose_x_0 = const()[name = string("op_11160_transpose_x_0"), val = bool(false)];
            bool var_11160_transpose_y_0 = const()[name = string("op_11160_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_434_perm_0 = const()[name = string("transpose_434_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_435_perm_0 = const()[name = string("transpose_435_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_435 = transpose(perm = transpose_435_perm_0, x = coreml_update_state_137)[name = string("transpose_462")];
            tensor<fp16, [1, 1, ?, 256]> transpose_434 = transpose(perm = transpose_434_perm_0, x = q_h_129_cast_fp16)[name = string("transpose_463")];
            tensor<fp16, [1, 1, ?, 128]> var_11160_cast_fp16 = matmul(transpose_x = var_11160_transpose_x_0, transpose_y = var_11160_transpose_y_0, x = transpose_434, y = transpose_435)[name = string("op_11160_cast_fp16")];
            fp16 var_11161_to_fp16 = const()[name = string("op_11161_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_11162_cast_fp16 = mul(x = var_11160_cast_fp16, y = var_11161_to_fp16)[name = string("op_11162_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_129_cast_fp16 = add(x = var_11162_cast_fp16, y = attention_mask)[name = string("scores_129_cast_fp16")];
            int32 var_11165 = const()[name = string("op_11165"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_129_cast_fp16 = softmax(axis = var_11165, x = scores_129_cast_fp16)[name = string("probs_129_cast_fp16")];
            tensor<int32, [4]> var_11172 = const()[name = string("op_11172"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_11174_transpose_x_0 = const()[name = string("op_11174_transpose_x_0"), val = bool(false)];
            bool var_11174_transpose_y_0 = const()[name = string("op_11174_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_129_cast_fp16 = transpose(perm = var_11172, x = coreml_update_state_139)[name = string("transpose_461")];
            tensor<fp16, [1, 1, ?, 256]> var_11174_cast_fp16 = matmul(transpose_x = var_11174_transpose_x_0, transpose_y = var_11174_transpose_y_0, x = probs_129_cast_fp16, y = v_t_129_cast_fp16)[name = string("op_11174_cast_fp16")];
            bool var_11193_transpose_x_0 = const()[name = string("op_11193_transpose_x_0"), val = bool(false)];
            bool var_11193_transpose_y_0 = const()[name = string("op_11193_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_439_perm_0 = const()[name = string("transpose_439_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_439 = transpose(perm = transpose_439_perm_0, x = q_h_131_cast_fp16)[name = string("transpose_460")];
            tensor<fp16, [1, 1, ?, 128]> var_11193_cast_fp16 = matmul(transpose_x = var_11193_transpose_x_0, transpose_y = var_11193_transpose_y_0, x = transpose_439, y = transpose_435)[name = string("op_11193_cast_fp16")];
            fp16 var_11194_to_fp16 = const()[name = string("op_11194_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_11195_cast_fp16 = mul(x = var_11193_cast_fp16, y = var_11194_to_fp16)[name = string("op_11195_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_131_cast_fp16 = add(x = var_11195_cast_fp16, y = attention_mask)[name = string("scores_131_cast_fp16")];
            int32 var_11198 = const()[name = string("op_11198"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_131_cast_fp16 = softmax(axis = var_11198, x = scores_131_cast_fp16)[name = string("probs_131_cast_fp16")];
            bool var_11207_transpose_x_0 = const()[name = string("op_11207_transpose_x_0"), val = bool(false)];
            bool var_11207_transpose_y_0 = const()[name = string("op_11207_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_11207_cast_fp16 = matmul(transpose_x = var_11207_transpose_x_0, transpose_y = var_11207_transpose_y_0, x = probs_131_cast_fp16, y = v_t_129_cast_fp16)[name = string("op_11207_cast_fp16")];
            bool var_11226_transpose_x_0 = const()[name = string("op_11226_transpose_x_0"), val = bool(false)];
            bool var_11226_transpose_y_0 = const()[name = string("op_11226_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_440_perm_0 = const()[name = string("transpose_440_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_440 = transpose(perm = transpose_440_perm_0, x = q_h_133_cast_fp16)[name = string("transpose_459")];
            tensor<fp16, [1, 1, ?, 128]> var_11226_cast_fp16 = matmul(transpose_x = var_11226_transpose_x_0, transpose_y = var_11226_transpose_y_0, x = transpose_440, y = transpose_435)[name = string("op_11226_cast_fp16")];
            fp16 var_11227_to_fp16 = const()[name = string("op_11227_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_11228_cast_fp16 = mul(x = var_11226_cast_fp16, y = var_11227_to_fp16)[name = string("op_11228_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_133_cast_fp16 = add(x = var_11228_cast_fp16, y = attention_mask)[name = string("scores_133_cast_fp16")];
            int32 var_11231 = const()[name = string("op_11231"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_133_cast_fp16 = softmax(axis = var_11231, x = scores_133_cast_fp16)[name = string("probs_133_cast_fp16")];
            bool var_11240_transpose_x_0 = const()[name = string("op_11240_transpose_x_0"), val = bool(false)];
            bool var_11240_transpose_y_0 = const()[name = string("op_11240_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_11240_cast_fp16 = matmul(transpose_x = var_11240_transpose_x_0, transpose_y = var_11240_transpose_y_0, x = probs_133_cast_fp16, y = v_t_129_cast_fp16)[name = string("op_11240_cast_fp16")];
            bool var_11259_transpose_x_0 = const()[name = string("op_11259_transpose_x_0"), val = bool(false)];
            bool var_11259_transpose_y_0 = const()[name = string("op_11259_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_441_perm_0 = const()[name = string("transpose_441_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_441 = transpose(perm = transpose_441_perm_0, x = q_h_135_cast_fp16)[name = string("transpose_458")];
            tensor<fp16, [1, 1, ?, 128]> var_11259_cast_fp16 = matmul(transpose_x = var_11259_transpose_x_0, transpose_y = var_11259_transpose_y_0, x = transpose_441, y = transpose_435)[name = string("op_11259_cast_fp16")];
            fp16 var_11260_to_fp16 = const()[name = string("op_11260_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_11261_cast_fp16 = mul(x = var_11259_cast_fp16, y = var_11260_to_fp16)[name = string("op_11261_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_135_cast_fp16 = add(x = var_11261_cast_fp16, y = attention_mask)[name = string("scores_135_cast_fp16")];
            int32 var_11264 = const()[name = string("op_11264"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_135_cast_fp16 = softmax(axis = var_11264, x = scores_135_cast_fp16)[name = string("probs_135_cast_fp16")];
            bool var_11273_transpose_x_0 = const()[name = string("op_11273_transpose_x_0"), val = bool(false)];
            bool var_11273_transpose_y_0 = const()[name = string("op_11273_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_11273_cast_fp16 = matmul(transpose_x = var_11273_transpose_x_0, transpose_y = var_11273_transpose_y_0, x = probs_135_cast_fp16, y = v_t_129_cast_fp16)[name = string("op_11273_cast_fp16")];
            bool input_131_interleave_0 = const()[name = string("input_131_interleave_0"), val = bool(false)];
            int32 const_161 = const()[name = string("const_161"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_131_cast_fp16 = concat(axis = const_161, interleave = input_131_interleave_0, values = (var_11174_cast_fp16, var_11207_cast_fp16, var_11240_cast_fp16, var_11273_cast_fp16))[name = string("input_131_cast_fp16")];
            string x_471_pad_type_0 = const()[name = string("x_471_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_471_strides_0 = const()[name = string("x_471_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_471_pad_0 = const()[name = string("x_471_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_471_dilations_0 = const()[name = string("x_471_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_471_groups_0 = const()[name = string("x_471_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_152_perm_0 = const()[name = string("transpose_152_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_16_weight_to_fp16 = const()[name = string("o_proj_16_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515882752)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_152_cast_fp16 = transpose(perm = transpose_152_perm_0, x = input_131_cast_fp16)[name = string("transpose_457")];
            tensor<fp16, [1, 640, 1, ?]> x_471_cast_fp16 = conv(dilations = x_471_dilations_0, groups = x_471_groups_0, pad = x_471_pad_0, pad_type = x_471_pad_type_0, strides = x_471_strides_0, weight = o_proj_16_weight_to_fp16, x = transpose_152_cast_fp16)[name = string("x_471_cast_fp16")];
            fp16 var_11294_promoted_to_fp16 = const()[name = string("op_11294_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_11295_cast_fp16 = pow(x = x_471_cast_fp16, y = var_11294_promoted_to_fp16)[name = string("op_11295_cast_fp16")];
            tensor<int32, [1]> var_301_axes_0 = const()[name = string("var_301_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_301_keep_dims_0 = const()[name = string("var_301_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_301_cast_fp16 = reduce_mean(axes = var_301_axes_0, keep_dims = var_301_keep_dims_0, x = var_11295_cast_fp16)[name = string("var_301_cast_fp16")];
            fp16 var_11302_to_fp16 = const()[name = string("op_11302_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_11303_cast_fp16 = add(x = var_301_cast_fp16, y = var_11302_to_fp16)[name = string("op_11303_cast_fp16")];
            fp32 var_11304_epsilon_0 = const()[name = string("op_11304_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_11304_cast_fp16 = rsqrt(epsilon = var_11304_epsilon_0, x = var_11303_cast_fp16)[name = string("op_11304_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_301_cast_fp16 = mul(x = x_471_cast_fp16, y = var_11304_cast_fp16)[name = string("x_normed_301_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_11314_to_fp16 = const()[name = string("op_11314_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517193536)))];
            tensor<fp16, [1, 640, 1, ?]> attn_67_cast_fp16 = mul(x = x_normed_301_cast_fp16, y = var_11314_to_fp16)[name = string("attn_67_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_473_cast_fp16 = add(x = x_449_cast_fp16, y = attn_67_cast_fp16)[name = string("x_473_cast_fp16")];
            fp16 var_11318_promoted_to_fp16 = const()[name = string("op_11318_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_11319_cast_fp16 = pow(x = x_473_cast_fp16, y = var_11318_promoted_to_fp16)[name = string("op_11319_cast_fp16")];
            tensor<int32, [1]> var_303_axes_0 = const()[name = string("var_303_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_303_keep_dims_0 = const()[name = string("var_303_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_303_cast_fp16 = reduce_mean(axes = var_303_axes_0, keep_dims = var_303_keep_dims_0, x = var_11319_cast_fp16)[name = string("var_303_cast_fp16")];
            fp16 var_11326_to_fp16 = const()[name = string("op_11326_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_11327_cast_fp16 = add(x = var_303_cast_fp16, y = var_11326_to_fp16)[name = string("op_11327_cast_fp16")];
            fp32 var_11328_epsilon_0 = const()[name = string("op_11328_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_11328_cast_fp16 = rsqrt(epsilon = var_11328_epsilon_0, x = var_11327_cast_fp16)[name = string("op_11328_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_303_cast_fp16 = mul(x = x_473_cast_fp16, y = var_11328_cast_fp16)[name = string("x_normed_303_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_11338_to_fp16 = const()[name = string("op_11338_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517194880)))];
            tensor<fp16, [1, 640, 1, ?]> input_133_cast_fp16 = mul(x = x_normed_303_cast_fp16, y = var_11338_to_fp16)[name = string("input_133_cast_fp16")];
            string gate_33_pad_type_0 = const()[name = string("gate_33_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_33_strides_0 = const()[name = string("gate_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_33_pad_0 = const()[name = string("gate_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_33_dilations_0 = const()[name = string("gate_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_33_groups_0 = const()[name = string("gate_33_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_16_weight_to_fp16 = const()[name = string("gate_proj_16_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517196224)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_33_cast_fp16 = conv(dilations = gate_33_dilations_0, groups = gate_33_groups_0, pad = gate_33_pad_0, pad_type = gate_33_pad_type_0, strides = gate_33_strides_0, weight = gate_proj_16_weight_to_fp16, x = input_133_cast_fp16)[name = string("gate_33_cast_fp16")];
            string up_33_pad_type_0 = const()[name = string("up_33_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_33_strides_0 = const()[name = string("up_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_33_pad_0 = const()[name = string("up_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_33_dilations_0 = const()[name = string("up_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_33_groups_0 = const()[name = string("up_33_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_16_weight_to_fp16 = const()[name = string("up_proj_16_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519817728)))];
            tensor<fp16, [1, 2048, 1, ?]> up_33_cast_fp16 = conv(dilations = up_33_dilations_0, groups = up_33_groups_0, pad = up_33_pad_0, pad_type = up_33_pad_type_0, strides = up_33_strides_0, weight = up_proj_16_weight_to_fp16, x = input_133_cast_fp16)[name = string("up_33_cast_fp16")];
            string var_11363_mode_0 = const()[name = string("op_11363_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_11363_cast_fp16 = gelu(mode = var_11363_mode_0, x = gate_33_cast_fp16)[name = string("op_11363_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_135_cast_fp16 = mul(x = var_11363_cast_fp16, y = up_33_cast_fp16)[name = string("input_135_cast_fp16")];
            string x_475_pad_type_0 = const()[name = string("x_475_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_475_strides_0 = const()[name = string("x_475_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_475_pad_0 = const()[name = string("x_475_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_475_dilations_0 = const()[name = string("x_475_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_475_groups_0 = const()[name = string("x_475_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_16_weight_to_fp16 = const()[name = string("down_proj_16_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522439232)))];
            tensor<fp16, [1, 640, 1, ?]> x_475_cast_fp16 = conv(dilations = x_475_dilations_0, groups = x_475_groups_0, pad = x_475_pad_0, pad_type = x_475_pad_type_0, strides = x_475_strides_0, weight = down_proj_16_weight_to_fp16, x = input_135_cast_fp16)[name = string("x_475_cast_fp16")];
            fp16 var_11376_promoted_to_fp16 = const()[name = string("op_11376_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_11377_cast_fp16 = pow(x = x_475_cast_fp16, y = var_11376_promoted_to_fp16)[name = string("op_11377_cast_fp16")];
            tensor<int32, [1]> var_305_axes_0 = const()[name = string("var_305_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_305_keep_dims_0 = const()[name = string("var_305_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_305_cast_fp16 = reduce_mean(axes = var_305_axes_0, keep_dims = var_305_keep_dims_0, x = var_11377_cast_fp16)[name = string("var_305_cast_fp16")];
            fp16 var_11384_to_fp16 = const()[name = string("op_11384_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_11385_cast_fp16 = add(x = var_305_cast_fp16, y = var_11384_to_fp16)[name = string("op_11385_cast_fp16")];
            fp32 var_11386_epsilon_0 = const()[name = string("op_11386_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_11386_cast_fp16 = rsqrt(epsilon = var_11386_epsilon_0, x = var_11385_cast_fp16)[name = string("op_11386_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_305_cast_fp16 = mul(x = x_475_cast_fp16, y = var_11386_cast_fp16)[name = string("x_normed_305_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_11396_to_fp16 = const()[name = string("op_11396_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525060736)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_33_cast_fp16 = mul(x = x_normed_305_cast_fp16, y = var_11396_to_fp16)[name = string("mlp_33_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_477_cast_fp16 = add(x = x_473_cast_fp16, y = mlp_33_cast_fp16)[name = string("x_477_cast_fp16")];
            fp16 var_11400_promoted_to_fp16 = const()[name = string("op_11400_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_11401_cast_fp16 = pow(x = x_477_cast_fp16, y = var_11400_promoted_to_fp16)[name = string("op_11401_cast_fp16")];
            tensor<int32, [1]> var_307_axes_0 = const()[name = string("var_307_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_307_keep_dims_0 = const()[name = string("var_307_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_307_cast_fp16 = reduce_mean(axes = var_307_axes_0, keep_dims = var_307_keep_dims_0, x = var_11401_cast_fp16)[name = string("var_307_cast_fp16")];
            fp16 var_11408_to_fp16 = const()[name = string("op_11408_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_11409_cast_fp16 = add(x = var_307_cast_fp16, y = var_11408_to_fp16)[name = string("op_11409_cast_fp16")];
            fp32 var_11410_epsilon_0 = const()[name = string("op_11410_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_11410_cast_fp16 = rsqrt(epsilon = var_11410_epsilon_0, x = var_11409_cast_fp16)[name = string("op_11410_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_307_cast_fp16 = mul(x = x_477_cast_fp16, y = var_11410_cast_fp16)[name = string("x_normed_307_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_11420_to_fp16 = const()[name = string("op_11420_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525062080)))];
            tensor<fp16, [1, 640, 1, ?]> input_137_cast_fp16 = mul(x = x_normed_307_cast_fp16, y = var_11420_to_fp16)[name = string("input_137_cast_fp16")];
            string var_11432_pad_type_0 = const()[name = string("op_11432_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_11432_strides_0 = const()[name = string("op_11432_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11432_pad_0 = const()[name = string("op_11432_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11432_dilations_0 = const()[name = string("op_11432_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_11432_groups_0 = const()[name = string("op_11432_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 640, 1, 1]> q_proj_17_weight_to_fp16 = const()[name = string("q_proj_17_weight_to_fp16"), val = tensor<fp16, [1024, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525063424)))];
            tensor<fp16, [1, 1024, 1, ?]> var_11432_cast_fp16 = conv(dilations = var_11432_dilations_0, groups = var_11432_groups_0, pad = var_11432_pad_0, pad_type = var_11432_pad_type_0, strides = var_11432_strides_0, weight = q_proj_17_weight_to_fp16, x = input_137_cast_fp16)[name = string("op_11432_cast_fp16")];
            string var_11443_pad_type_0 = const()[name = string("op_11443_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_11443_strides_0 = const()[name = string("op_11443_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11443_pad_0 = const()[name = string("op_11443_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11443_dilations_0 = const()[name = string("op_11443_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_11443_groups_0 = const()[name = string("op_11443_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> k_proj_17_weight_to_fp16 = const()[name = string("k_proj_17_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526374208)))];
            tensor<fp16, [1, 256, 1, ?]> var_11443_cast_fp16 = conv(dilations = var_11443_dilations_0, groups = var_11443_groups_0, pad = var_11443_pad_0, pad_type = var_11443_pad_type_0, strides = var_11443_strides_0, weight = k_proj_17_weight_to_fp16, x = input_137_cast_fp16)[name = string("op_11443_cast_fp16")];
            string var_11454_pad_type_0 = const()[name = string("op_11454_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_11454_strides_0 = const()[name = string("op_11454_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_11454_pad_0 = const()[name = string("op_11454_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_11454_dilations_0 = const()[name = string("op_11454_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_11454_groups_0 = const()[name = string("op_11454_groups_0"), val = int32(1)];
            tensor<fp16, [256, 640, 1, 1]> v_proj_17_weight_to_fp16 = const()[name = string("v_proj_17_weight_to_fp16"), val = tensor<fp16, [256, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526701952)))];
            tensor<fp16, [1, 256, 1, ?]> var_11454_cast_fp16 = conv(dilations = var_11454_dilations_0, groups = var_11454_groups_0, pad = var_11454_pad_0, pad_type = var_11454_pad_type_0, strides = var_11454_strides_0, weight = v_proj_17_weight_to_fp16, x = input_137_cast_fp16)[name = string("op_11454_cast_fp16")];
            tensor<int32, [4]> tile_51 = const()[name = string("tile_51"), val = tensor<int32, [4]>([256, 256, 256, 256])];
            int32 var_11457_axis_0 = const()[name = string("op_11457_axis_0"), val = int32(1)];
            tensor<fp16, [1, 256, 1, ?]> var_11457_cast_fp16_0, tensor<fp16, [1, 256, 1, ?]> var_11457_cast_fp16_1, tensor<fp16, [1, 256, 1, ?]> var_11457_cast_fp16_2, tensor<fp16, [1, 256, 1, ?]> var_11457_cast_fp16_3 = split(axis = var_11457_axis_0, split_sizes = tile_51, x = var_11432_cast_fp16)[name = string("op_11457_cast_fp16")];
            fp16 var_11470_promoted_to_fp16 = const()[name = string("op_11470_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_11471_cast_fp16 = pow(x = var_11457_cast_fp16_0, y = var_11470_promoted_to_fp16)[name = string("op_11471_cast_fp16")];
            tensor<int32, [1]> var_309_axes_0 = const()[name = string("var_309_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_309_keep_dims_0 = const()[name = string("var_309_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_309_cast_fp16_0 = reduce_mean(axes = var_309_axes_0, keep_dims = var_309_keep_dims_0, x = var_11471_cast_fp16)[name = string("var_309_cast_fp16")];
            fp16 var_11478_to_fp16 = const()[name = string("op_11478_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_11479_cast_fp16 = add(x = var_309_cast_fp16_0, y = var_11478_to_fp16)[name = string("op_11479_cast_fp16")];
            fp32 var_11480_epsilon_0 = const()[name = string("op_11480_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_11480_cast_fp16 = rsqrt(epsilon = var_11480_epsilon_0, x = var_11479_cast_fp16)[name = string("op_11480_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_309_cast_fp16 = mul(x = var_11457_cast_fp16_0, y = var_11480_cast_fp16)[name = string("x_normed_309_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_11490_to_fp16 = const()[name = string("op_11490_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527029696)))];
            tensor<fp16, [1, 256, 1, ?]> x_481_cast_fp16 = mul(x = x_normed_309_cast_fp16, y = var_11490_to_fp16)[name = string("x_481_cast_fp16")];
            tensor<int32, [4]> x_first_171_begin_0 = const()[name = string("x_first_171_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_171_end_0 = const()[name = string("x_first_171_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_171_end_mask_0 = const()[name = string("x_first_171_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_171_cast_fp16 = slice_by_index(begin = x_first_171_begin_0, end = x_first_171_end_0, end_mask = x_first_171_end_mask_0, x = x_481_cast_fp16)[name = string("x_first_171_cast_fp16")];
            tensor<int32, [4]> x_second_171_begin_0 = const()[name = string("x_second_171_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_171_end_0 = const()[name = string("x_second_171_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_171_end_mask_0 = const()[name = string("x_second_171_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_171_cast_fp16 = slice_by_index(begin = x_second_171_begin_0, end = x_second_171_end_0, end_mask = x_second_171_end_mask_0, x = x_481_cast_fp16)[name = string("x_second_171_cast_fp16")];
            fp16 const_140_promoted_to_fp16 = const()[name = string("const_140_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_11512_cast_fp16 = mul(x = x_second_171_cast_fp16, y = const_140_promoted_to_fp16)[name = string("op_11512_cast_fp16")];
            int32 var_11514 = const()[name = string("op_11514"), val = int32(1)];
            bool x_rot_171_interleave_0 = const()[name = string("x_rot_171_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_171_cast_fp16 = concat(axis = var_11514, interleave = x_rot_171_interleave_0, values = (var_11512_cast_fp16, x_first_171_cast_fp16))[name = string("x_rot_171_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_11530_cast_fp16 = mul(x = x_481_cast_fp16, y = cos_bc1t_51_cast_fp16)[name = string("op_11530_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_11531_cast_fp16 = mul(x = x_rot_171_cast_fp16, y = sin_bc1t_51_cast_fp16)[name = string("op_11531_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_137_cast_fp16 = add(x = var_11530_cast_fp16, y = var_11531_cast_fp16)[name = string("q_h_137_cast_fp16")];
            fp16 var_11534_promoted_to_fp16 = const()[name = string("op_11534_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_11535_cast_fp16 = pow(x = var_11457_cast_fp16_1, y = var_11534_promoted_to_fp16)[name = string("op_11535_cast_fp16")];
            tensor<int32, [1]> var_311_axes_0 = const()[name = string("var_311_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_311_keep_dims_0 = const()[name = string("var_311_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_311_cast_fp16 = reduce_mean(axes = var_311_axes_0, keep_dims = var_311_keep_dims_0, x = var_11535_cast_fp16)[name = string("var_311_cast_fp16")];
            fp16 var_11542_to_fp16 = const()[name = string("op_11542_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_11543_cast_fp16 = add(x = var_311_cast_fp16, y = var_11542_to_fp16)[name = string("op_11543_cast_fp16")];
            fp32 var_11544_epsilon_0 = const()[name = string("op_11544_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_11544_cast_fp16 = rsqrt(epsilon = var_11544_epsilon_0, x = var_11543_cast_fp16)[name = string("op_11544_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_311_cast_fp16 = mul(x = var_11457_cast_fp16_1, y = var_11544_cast_fp16)[name = string("x_normed_311_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_485_cast_fp16 = mul(x = x_normed_311_cast_fp16, y = var_11490_to_fp16)[name = string("x_485_cast_fp16")];
            tensor<int32, [4]> x_first_173_begin_0 = const()[name = string("x_first_173_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_173_end_0 = const()[name = string("x_first_173_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_173_end_mask_0 = const()[name = string("x_first_173_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_173_cast_fp16 = slice_by_index(begin = x_first_173_begin_0, end = x_first_173_end_0, end_mask = x_first_173_end_mask_0, x = x_485_cast_fp16)[name = string("x_first_173_cast_fp16")];
            tensor<int32, [4]> x_second_173_begin_0 = const()[name = string("x_second_173_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_173_end_0 = const()[name = string("x_second_173_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_173_end_mask_0 = const()[name = string("x_second_173_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_173_cast_fp16 = slice_by_index(begin = x_second_173_begin_0, end = x_second_173_end_0, end_mask = x_second_173_end_mask_0, x = x_485_cast_fp16)[name = string("x_second_173_cast_fp16")];
            fp16 const_141_promoted_to_fp16 = const()[name = string("const_141_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_11576_cast_fp16 = mul(x = x_second_173_cast_fp16, y = const_141_promoted_to_fp16)[name = string("op_11576_cast_fp16")];
            int32 var_11578 = const()[name = string("op_11578"), val = int32(1)];
            bool x_rot_173_interleave_0 = const()[name = string("x_rot_173_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_173_cast_fp16 = concat(axis = var_11578, interleave = x_rot_173_interleave_0, values = (var_11576_cast_fp16, x_first_173_cast_fp16))[name = string("x_rot_173_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_11594_cast_fp16 = mul(x = x_485_cast_fp16, y = cos_bc1t_51_cast_fp16)[name = string("op_11594_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_11595_cast_fp16 = mul(x = x_rot_173_cast_fp16, y = sin_bc1t_51_cast_fp16)[name = string("op_11595_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_139_cast_fp16 = add(x = var_11594_cast_fp16, y = var_11595_cast_fp16)[name = string("q_h_139_cast_fp16")];
            fp16 var_11598_promoted_to_fp16 = const()[name = string("op_11598_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_11599_cast_fp16 = pow(x = var_11457_cast_fp16_2, y = var_11598_promoted_to_fp16)[name = string("op_11599_cast_fp16")];
            tensor<int32, [1]> var_313_axes_0 = const()[name = string("var_313_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_313_keep_dims_0 = const()[name = string("var_313_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_313_cast_fp16 = reduce_mean(axes = var_313_axes_0, keep_dims = var_313_keep_dims_0, x = var_11599_cast_fp16)[name = string("var_313_cast_fp16")];
            fp16 var_11606_to_fp16 = const()[name = string("op_11606_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_11607_cast_fp16 = add(x = var_313_cast_fp16, y = var_11606_to_fp16)[name = string("op_11607_cast_fp16")];
            fp32 var_11608_epsilon_0 = const()[name = string("op_11608_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_11608_cast_fp16 = rsqrt(epsilon = var_11608_epsilon_0, x = var_11607_cast_fp16)[name = string("op_11608_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_313_cast_fp16 = mul(x = var_11457_cast_fp16_2, y = var_11608_cast_fp16)[name = string("x_normed_313_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_489_cast_fp16 = mul(x = x_normed_313_cast_fp16, y = var_11490_to_fp16)[name = string("x_489_cast_fp16")];
            tensor<int32, [4]> x_first_175_begin_0 = const()[name = string("x_first_175_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_175_end_0 = const()[name = string("x_first_175_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_175_end_mask_0 = const()[name = string("x_first_175_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_175_cast_fp16 = slice_by_index(begin = x_first_175_begin_0, end = x_first_175_end_0, end_mask = x_first_175_end_mask_0, x = x_489_cast_fp16)[name = string("x_first_175_cast_fp16")];
            tensor<int32, [4]> x_second_175_begin_0 = const()[name = string("x_second_175_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_175_end_0 = const()[name = string("x_second_175_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_175_end_mask_0 = const()[name = string("x_second_175_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_175_cast_fp16 = slice_by_index(begin = x_second_175_begin_0, end = x_second_175_end_0, end_mask = x_second_175_end_mask_0, x = x_489_cast_fp16)[name = string("x_second_175_cast_fp16")];
            fp16 const_142_promoted_to_fp16 = const()[name = string("const_142_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_11640_cast_fp16 = mul(x = x_second_175_cast_fp16, y = const_142_promoted_to_fp16)[name = string("op_11640_cast_fp16")];
            int32 var_11642 = const()[name = string("op_11642"), val = int32(1)];
            bool x_rot_175_interleave_0 = const()[name = string("x_rot_175_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_175_cast_fp16 = concat(axis = var_11642, interleave = x_rot_175_interleave_0, values = (var_11640_cast_fp16, x_first_175_cast_fp16))[name = string("x_rot_175_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_11658_cast_fp16 = mul(x = x_489_cast_fp16, y = cos_bc1t_51_cast_fp16)[name = string("op_11658_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_11659_cast_fp16 = mul(x = x_rot_175_cast_fp16, y = sin_bc1t_51_cast_fp16)[name = string("op_11659_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_141_cast_fp16 = add(x = var_11658_cast_fp16, y = var_11659_cast_fp16)[name = string("q_h_141_cast_fp16")];
            fp16 var_11662_promoted_to_fp16 = const()[name = string("op_11662_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_11663_cast_fp16 = pow(x = var_11457_cast_fp16_3, y = var_11662_promoted_to_fp16)[name = string("op_11663_cast_fp16")];
            tensor<int32, [1]> var_315_axes_0 = const()[name = string("var_315_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_315_keep_dims_0 = const()[name = string("var_315_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_315_cast_fp16 = reduce_mean(axes = var_315_axes_0, keep_dims = var_315_keep_dims_0, x = var_11663_cast_fp16)[name = string("var_315_cast_fp16")];
            fp16 var_11670_to_fp16 = const()[name = string("op_11670_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_11671_cast_fp16 = add(x = var_315_cast_fp16, y = var_11670_to_fp16)[name = string("op_11671_cast_fp16")];
            fp32 var_11672_epsilon_0 = const()[name = string("op_11672_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_11672_cast_fp16 = rsqrt(epsilon = var_11672_epsilon_0, x = var_11671_cast_fp16)[name = string("op_11672_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_315_cast_fp16 = mul(x = var_11457_cast_fp16_3, y = var_11672_cast_fp16)[name = string("x_normed_315_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_493_cast_fp16 = mul(x = x_normed_315_cast_fp16, y = var_11490_to_fp16)[name = string("x_493_cast_fp16")];
            tensor<int32, [4]> x_first_177_begin_0 = const()[name = string("x_first_177_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_177_end_0 = const()[name = string("x_first_177_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_177_end_mask_0 = const()[name = string("x_first_177_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_177_cast_fp16 = slice_by_index(begin = x_first_177_begin_0, end = x_first_177_end_0, end_mask = x_first_177_end_mask_0, x = x_493_cast_fp16)[name = string("x_first_177_cast_fp16")];
            tensor<int32, [4]> x_second_177_begin_0 = const()[name = string("x_second_177_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_177_end_0 = const()[name = string("x_second_177_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_177_end_mask_0 = const()[name = string("x_second_177_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_177_cast_fp16 = slice_by_index(begin = x_second_177_begin_0, end = x_second_177_end_0, end_mask = x_second_177_end_mask_0, x = x_493_cast_fp16)[name = string("x_second_177_cast_fp16")];
            fp16 const_143_promoted_to_fp16 = const()[name = string("const_143_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_11704_cast_fp16 = mul(x = x_second_177_cast_fp16, y = const_143_promoted_to_fp16)[name = string("op_11704_cast_fp16")];
            int32 var_11706 = const()[name = string("op_11706"), val = int32(1)];
            bool x_rot_177_interleave_0 = const()[name = string("x_rot_177_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_177_cast_fp16 = concat(axis = var_11706, interleave = x_rot_177_interleave_0, values = (var_11704_cast_fp16, x_first_177_cast_fp16))[name = string("x_rot_177_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_11722_cast_fp16 = mul(x = x_493_cast_fp16, y = cos_bc1t_51_cast_fp16)[name = string("op_11722_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_11723_cast_fp16 = mul(x = x_rot_177_cast_fp16, y = sin_bc1t_51_cast_fp16)[name = string("op_11723_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> q_h_cast_fp16 = add(x = var_11722_cast_fp16, y = var_11723_cast_fp16)[name = string("q_h_cast_fp16")];
            fp16 var_11726_promoted_to_fp16 = const()[name = string("op_11726_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 1, ?]> var_11727_cast_fp16 = pow(x = var_11443_cast_fp16, y = var_11726_promoted_to_fp16)[name = string("op_11727_cast_fp16")];
            tensor<int32, [1]> var_317_axes_0 = const()[name = string("var_317_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_317_keep_dims_0 = const()[name = string("var_317_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_317_cast_fp16 = reduce_mean(axes = var_317_axes_0, keep_dims = var_317_keep_dims_0, x = var_11727_cast_fp16)[name = string("var_317_cast_fp16")];
            fp16 var_11734_to_fp16 = const()[name = string("op_11734_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_11735_cast_fp16 = add(x = var_317_cast_fp16, y = var_11734_to_fp16)[name = string("op_11735_cast_fp16")];
            fp32 var_11736_epsilon_0 = const()[name = string("op_11736_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_11736_cast_fp16 = rsqrt(epsilon = var_11736_epsilon_0, x = var_11735_cast_fp16)[name = string("op_11736_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> x_normed_317_cast_fp16 = mul(x = var_11443_cast_fp16, y = var_11736_cast_fp16)[name = string("x_normed_317_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_11746_to_fp16 = const()[name = string("op_11746_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527030272)))];
            tensor<fp16, [1, 256, 1, ?]> x_497_cast_fp16 = mul(x = x_normed_317_cast_fp16, y = var_11746_to_fp16)[name = string("x_497_cast_fp16")];
            tensor<int32, [4]> x_first_begin_0 = const()[name = string("x_first_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_first_end_0 = const()[name = string("x_first_end_0"), val = tensor<int32, [4]>([1, 128, 1, 0])];
            tensor<bool, [4]> x_first_end_mask_0 = const()[name = string("x_first_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_first_cast_fp16 = slice_by_index(begin = x_first_begin_0, end = x_first_end_0, end_mask = x_first_end_mask_0, x = x_497_cast_fp16)[name = string("x_first_cast_fp16")];
            tensor<int32, [4]> x_second_begin_0 = const()[name = string("x_second_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> x_second_end_0 = const()[name = string("x_second_end_0"), val = tensor<int32, [4]>([1, 256, 1, 0])];
            tensor<bool, [4]> x_second_end_mask_0 = const()[name = string("x_second_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, ?]> x_second_cast_fp16 = slice_by_index(begin = x_second_begin_0, end = x_second_end_0, end_mask = x_second_end_mask_0, x = x_497_cast_fp16)[name = string("x_second_cast_fp16")];
            fp16 const_144_promoted_to_fp16 = const()[name = string("const_144_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 128, 1, ?]> var_11768_cast_fp16 = mul(x = x_second_cast_fp16, y = const_144_promoted_to_fp16)[name = string("op_11768_cast_fp16")];
            int32 var_11770 = const()[name = string("op_11770"), val = int32(1)];
            bool x_rot_interleave_0 = const()[name = string("x_rot_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 1, ?]> x_rot_cast_fp16 = concat(axis = var_11770, interleave = x_rot_interleave_0, values = (var_11768_cast_fp16, x_first_cast_fp16))[name = string("x_rot_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_11786_cast_fp16 = mul(x = x_497_cast_fp16, y = cos_bc1t_51_cast_fp16)[name = string("op_11786_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> var_11787_cast_fp16 = mul(x = x_rot_cast_fp16, y = sin_bc1t_51_cast_fp16)[name = string("op_11787_cast_fp16")];
            tensor<fp16, [1, 256, 1, ?]> k_new_cast_fp16 = add(x = var_11786_cast_fp16, y = var_11787_cast_fp16)[name = string("k_new_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> read_state_34 = read_state(input = k_cache_17)[name = string("read_state_34")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_105_cast_fp16 = mul(x = read_state_34, y = one_minus_1_cast_fp16)[name = string("k_cache_105_cast_fp16")];
            write_state(data = k_cache_105_cast_fp16, input = k_cache_17)[name = string("coreml_update_state_140_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_140 = read_state(input = k_cache_17)[name = string("coreml_update_state_140")];
            tensor<fp16, [1, 256, 1, 128]> var_11794_cast_fp16 = mul(x = k_new_cast_fp16, y = write_mask)[name = string("op_11794_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> k_cache_cast_fp16 = add(x = coreml_update_state_140, y = var_11794_cast_fp16)[name = string("k_cache_cast_fp16")];
            write_state(data = k_cache_cast_fp16, input = k_cache_17)[name = string("coreml_update_state_141_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_141 = read_state(input = k_cache_17)[name = string("coreml_update_state_141")];
            tensor<fp16, [1, 256, 1, 128]> read_state_35 = read_state(input = v_cache_17)[name = string("read_state_35")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_105_cast_fp16 = mul(x = read_state_35, y = one_minus_1_cast_fp16)[name = string("v_cache_105_cast_fp16")];
            write_state(data = v_cache_105_cast_fp16, input = v_cache_17)[name = string("coreml_update_state_142_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_142 = read_state(input = v_cache_17)[name = string("coreml_update_state_142")];
            tensor<fp16, [1, 256, 1, 128]> var_11798_cast_fp16 = mul(x = var_11454_cast_fp16, y = write_mask)[name = string("op_11798_cast_fp16")];
            tensor<fp16, [1, 256, 1, 128]> v_cache_cast_fp16 = add(x = coreml_update_state_142, y = var_11798_cast_fp16)[name = string("v_cache_cast_fp16")];
            write_state(data = v_cache_cast_fp16, input = v_cache_17)[name = string("coreml_update_state_143_write_state")];
            tensor<fp16, [1, 256, 1, 128]> coreml_update_state_143 = read_state(input = v_cache_17)[name = string("coreml_update_state_143")];
            bool var_11813_transpose_x_0 = const()[name = string("op_11813_transpose_x_0"), val = bool(false)];
            bool var_11813_transpose_y_0 = const()[name = string("op_11813_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_442_perm_0 = const()[name = string("transpose_442_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<int32, [4]> transpose_443_perm_0 = const()[name = string("transpose_443_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 256, 128]> transpose_443 = transpose(perm = transpose_443_perm_0, x = coreml_update_state_141)[name = string("transpose_455")];
            tensor<fp16, [1, 1, ?, 256]> transpose_442 = transpose(perm = transpose_442_perm_0, x = q_h_137_cast_fp16)[name = string("transpose_456")];
            tensor<fp16, [1, 1, ?, 128]> var_11813_cast_fp16 = matmul(transpose_x = var_11813_transpose_x_0, transpose_y = var_11813_transpose_y_0, x = transpose_442, y = transpose_443)[name = string("op_11813_cast_fp16")];
            fp16 var_11814_to_fp16 = const()[name = string("op_11814_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_11815_cast_fp16 = mul(x = var_11813_cast_fp16, y = var_11814_to_fp16)[name = string("op_11815_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_137_cast_fp16 = add(x = var_11815_cast_fp16, y = attention_mask)[name = string("scores_137_cast_fp16")];
            int32 var_11818 = const()[name = string("op_11818"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_137_cast_fp16 = softmax(axis = var_11818, x = scores_137_cast_fp16)[name = string("probs_137_cast_fp16")];
            tensor<int32, [4]> var_11825 = const()[name = string("op_11825"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            bool var_11827_transpose_x_0 = const()[name = string("op_11827_transpose_x_0"), val = bool(false)];
            bool var_11827_transpose_y_0 = const()[name = string("op_11827_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 256]> v_t_137_cast_fp16 = transpose(perm = var_11825, x = coreml_update_state_143)[name = string("transpose_454")];
            tensor<fp16, [1, 1, ?, 256]> var_11827_cast_fp16 = matmul(transpose_x = var_11827_transpose_x_0, transpose_y = var_11827_transpose_y_0, x = probs_137_cast_fp16, y = v_t_137_cast_fp16)[name = string("op_11827_cast_fp16")];
            bool var_11846_transpose_x_0 = const()[name = string("op_11846_transpose_x_0"), val = bool(false)];
            bool var_11846_transpose_y_0 = const()[name = string("op_11846_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_447_perm_0 = const()[name = string("transpose_447_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_447 = transpose(perm = transpose_447_perm_0, x = q_h_139_cast_fp16)[name = string("transpose_453")];
            tensor<fp16, [1, 1, ?, 128]> var_11846_cast_fp16 = matmul(transpose_x = var_11846_transpose_x_0, transpose_y = var_11846_transpose_y_0, x = transpose_447, y = transpose_443)[name = string("op_11846_cast_fp16")];
            fp16 var_11847_to_fp16 = const()[name = string("op_11847_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_11848_cast_fp16 = mul(x = var_11846_cast_fp16, y = var_11847_to_fp16)[name = string("op_11848_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_139_cast_fp16 = add(x = var_11848_cast_fp16, y = attention_mask)[name = string("scores_139_cast_fp16")];
            int32 var_11851 = const()[name = string("op_11851"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_139_cast_fp16 = softmax(axis = var_11851, x = scores_139_cast_fp16)[name = string("probs_139_cast_fp16")];
            bool var_11860_transpose_x_0 = const()[name = string("op_11860_transpose_x_0"), val = bool(false)];
            bool var_11860_transpose_y_0 = const()[name = string("op_11860_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_11860_cast_fp16 = matmul(transpose_x = var_11860_transpose_x_0, transpose_y = var_11860_transpose_y_0, x = probs_139_cast_fp16, y = v_t_137_cast_fp16)[name = string("op_11860_cast_fp16")];
            bool var_11879_transpose_x_0 = const()[name = string("op_11879_transpose_x_0"), val = bool(false)];
            bool var_11879_transpose_y_0 = const()[name = string("op_11879_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_448_perm_0 = const()[name = string("transpose_448_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_448 = transpose(perm = transpose_448_perm_0, x = q_h_141_cast_fp16)[name = string("transpose_452")];
            tensor<fp16, [1, 1, ?, 128]> var_11879_cast_fp16 = matmul(transpose_x = var_11879_transpose_x_0, transpose_y = var_11879_transpose_y_0, x = transpose_448, y = transpose_443)[name = string("op_11879_cast_fp16")];
            fp16 var_11880_to_fp16 = const()[name = string("op_11880_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_11881_cast_fp16 = mul(x = var_11879_cast_fp16, y = var_11880_to_fp16)[name = string("op_11881_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_141_cast_fp16 = add(x = var_11881_cast_fp16, y = attention_mask)[name = string("scores_141_cast_fp16")];
            int32 var_11884 = const()[name = string("op_11884"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_141_cast_fp16 = softmax(axis = var_11884, x = scores_141_cast_fp16)[name = string("probs_141_cast_fp16")];
            bool var_11893_transpose_x_0 = const()[name = string("op_11893_transpose_x_0"), val = bool(false)];
            bool var_11893_transpose_y_0 = const()[name = string("op_11893_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_11893_cast_fp16 = matmul(transpose_x = var_11893_transpose_x_0, transpose_y = var_11893_transpose_y_0, x = probs_141_cast_fp16, y = v_t_137_cast_fp16)[name = string("op_11893_cast_fp16")];
            bool var_11912_transpose_x_0 = const()[name = string("op_11912_transpose_x_0"), val = bool(false)];
            bool var_11912_transpose_y_0 = const()[name = string("op_11912_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_449_perm_0 = const()[name = string("transpose_449_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 1, ?, 256]> transpose_449 = transpose(perm = transpose_449_perm_0, x = q_h_cast_fp16)[name = string("transpose_451")];
            tensor<fp16, [1, 1, ?, 128]> var_11912_cast_fp16 = matmul(transpose_x = var_11912_transpose_x_0, transpose_y = var_11912_transpose_y_0, x = transpose_449, y = transpose_443)[name = string("op_11912_cast_fp16")];
            fp16 var_11913_to_fp16 = const()[name = string("op_11913_to_fp16"), val = fp16(0x1p-4)];
            tensor<fp16, [1, 1, ?, 128]> var_11914_cast_fp16 = mul(x = var_11912_cast_fp16, y = var_11913_to_fp16)[name = string("op_11914_cast_fp16")];
            tensor<fp16, [1, 1, ?, 128]> scores_cast_fp16 = add(x = var_11914_cast_fp16, y = attention_mask)[name = string("scores_cast_fp16")];
            int32 var_11917 = const()[name = string("op_11917"), val = int32(-1)];
            tensor<fp16, [1, 1, ?, 128]> probs_cast_fp16 = softmax(axis = var_11917, x = scores_cast_fp16)[name = string("probs_cast_fp16")];
            bool var_11926_transpose_x_0 = const()[name = string("op_11926_transpose_x_0"), val = bool(false)];
            bool var_11926_transpose_y_0 = const()[name = string("op_11926_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, ?, 256]> var_11926_cast_fp16 = matmul(transpose_x = var_11926_transpose_x_0, transpose_y = var_11926_transpose_y_0, x = probs_cast_fp16, y = v_t_137_cast_fp16)[name = string("op_11926_cast_fp16")];
            bool input_139_interleave_0 = const()[name = string("input_139_interleave_0"), val = bool(false)];
            int32 const_162 = const()[name = string("const_162"), val = int32(3)];
            tensor<fp16, [1, 1, ?, 1024]> input_139_cast_fp16 = concat(axis = const_162, interleave = input_139_interleave_0, values = (var_11827_cast_fp16, var_11860_cast_fp16, var_11893_cast_fp16, var_11926_cast_fp16))[name = string("input_139_cast_fp16")];
            string x_499_pad_type_0 = const()[name = string("x_499_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_499_strides_0 = const()[name = string("x_499_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_499_pad_0 = const()[name = string("x_499_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_499_dilations_0 = const()[name = string("x_499_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_499_groups_0 = const()[name = string("x_499_groups_0"), val = int32(1)];
            tensor<int32, [4]> transpose_161_perm_0 = const()[name = string("transpose_161_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [640, 1024, 1, 1]> o_proj_17_weight_to_fp16 = const()[name = string("o_proj_17_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527030848)))];
            tensor<fp16, [1, 1024, 1, ?]> transpose_161_cast_fp16 = transpose(perm = transpose_161_perm_0, x = input_139_cast_fp16)[name = string("transpose_450")];
            tensor<fp16, [1, 640, 1, ?]> x_499_cast_fp16 = conv(dilations = x_499_dilations_0, groups = x_499_groups_0, pad = x_499_pad_0, pad_type = x_499_pad_type_0, strides = x_499_strides_0, weight = o_proj_17_weight_to_fp16, x = transpose_161_cast_fp16)[name = string("x_499_cast_fp16")];
            fp16 var_11947_promoted_to_fp16 = const()[name = string("op_11947_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_11948_cast_fp16 = pow(x = x_499_cast_fp16, y = var_11947_promoted_to_fp16)[name = string("op_11948_cast_fp16")];
            tensor<int32, [1]> var_319_axes_0 = const()[name = string("var_319_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_319_keep_dims_0 = const()[name = string("var_319_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_319_cast_fp16 = reduce_mean(axes = var_319_axes_0, keep_dims = var_319_keep_dims_0, x = var_11948_cast_fp16)[name = string("var_319_cast_fp16")];
            fp16 var_11955_to_fp16 = const()[name = string("op_11955_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_11956_cast_fp16 = add(x = var_319_cast_fp16, y = var_11955_to_fp16)[name = string("op_11956_cast_fp16")];
            fp32 var_11957_epsilon_0 = const()[name = string("op_11957_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_11957_cast_fp16 = rsqrt(epsilon = var_11957_epsilon_0, x = var_11956_cast_fp16)[name = string("op_11957_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_319_cast_fp16 = mul(x = x_499_cast_fp16, y = var_11957_cast_fp16)[name = string("x_normed_319_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_11967_to_fp16 = const()[name = string("op_11967_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528341632)))];
            tensor<fp16, [1, 640, 1, ?]> attn_cast_fp16 = mul(x = x_normed_319_cast_fp16, y = var_11967_to_fp16)[name = string("attn_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_501_cast_fp16 = add(x = x_477_cast_fp16, y = attn_cast_fp16)[name = string("x_501_cast_fp16")];
            fp16 var_11971_promoted_to_fp16 = const()[name = string("op_11971_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_11972_cast_fp16 = pow(x = x_501_cast_fp16, y = var_11971_promoted_to_fp16)[name = string("op_11972_cast_fp16")];
            tensor<int32, [1]> var_321_axes_0 = const()[name = string("var_321_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_321_keep_dims_0 = const()[name = string("var_321_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_321_cast_fp16 = reduce_mean(axes = var_321_axes_0, keep_dims = var_321_keep_dims_0, x = var_11972_cast_fp16)[name = string("var_321_cast_fp16")];
            fp16 var_11979_to_fp16 = const()[name = string("op_11979_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_11980_cast_fp16 = add(x = var_321_cast_fp16, y = var_11979_to_fp16)[name = string("op_11980_cast_fp16")];
            fp32 var_11981_epsilon_0 = const()[name = string("op_11981_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_11981_cast_fp16 = rsqrt(epsilon = var_11981_epsilon_0, x = var_11980_cast_fp16)[name = string("op_11981_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_321_cast_fp16 = mul(x = x_501_cast_fp16, y = var_11981_cast_fp16)[name = string("x_normed_321_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_11991_to_fp16 = const()[name = string("op_11991_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528342976)))];
            tensor<fp16, [1, 640, 1, ?]> input_141_cast_fp16 = mul(x = x_normed_321_cast_fp16, y = var_11991_to_fp16)[name = string("input_141_cast_fp16")];
            string gate_pad_type_0 = const()[name = string("gate_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> gate_strides_0 = const()[name = string("gate_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> gate_pad_0 = const()[name = string("gate_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> gate_dilations_0 = const()[name = string("gate_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 gate_groups_0 = const()[name = string("gate_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> gate_proj_17_weight_to_fp16 = const()[name = string("gate_proj_17_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528344320)))];
            tensor<fp16, [1, 2048, 1, ?]> gate_cast_fp16 = conv(dilations = gate_dilations_0, groups = gate_groups_0, pad = gate_pad_0, pad_type = gate_pad_type_0, strides = gate_strides_0, weight = gate_proj_17_weight_to_fp16, x = input_141_cast_fp16)[name = string("gate_cast_fp16")];
            string up_pad_type_0 = const()[name = string("up_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_strides_0 = const()[name = string("up_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_pad_0 = const()[name = string("up_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_dilations_0 = const()[name = string("up_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_groups_0 = const()[name = string("up_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 640, 1, 1]> up_proj_17_weight_to_fp16 = const()[name = string("up_proj_17_weight_to_fp16"), val = tensor<fp16, [2048, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(530965824)))];
            tensor<fp16, [1, 2048, 1, ?]> up_cast_fp16 = conv(dilations = up_dilations_0, groups = up_groups_0, pad = up_pad_0, pad_type = up_pad_type_0, strides = up_strides_0, weight = up_proj_17_weight_to_fp16, x = input_141_cast_fp16)[name = string("up_cast_fp16")];
            string var_12016_mode_0 = const()[name = string("op_12016_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2048, 1, ?]> var_12016_cast_fp16 = gelu(mode = var_12016_mode_0, x = gate_cast_fp16)[name = string("op_12016_cast_fp16")];
            tensor<fp16, [1, 2048, 1, ?]> input_143_cast_fp16 = mul(x = var_12016_cast_fp16, y = up_cast_fp16)[name = string("input_143_cast_fp16")];
            string x_503_pad_type_0 = const()[name = string("x_503_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_503_strides_0 = const()[name = string("x_503_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_503_pad_0 = const()[name = string("x_503_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_503_dilations_0 = const()[name = string("x_503_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_503_groups_0 = const()[name = string("x_503_groups_0"), val = int32(1)];
            tensor<fp16, [640, 2048, 1, 1]> down_proj_17_weight_to_fp16 = const()[name = string("down_proj_17_weight_to_fp16"), val = tensor<fp16, [640, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533587328)))];
            tensor<fp16, [1, 640, 1, ?]> x_503_cast_fp16 = conv(dilations = x_503_dilations_0, groups = x_503_groups_0, pad = x_503_pad_0, pad_type = x_503_pad_type_0, strides = x_503_strides_0, weight = down_proj_17_weight_to_fp16, x = input_143_cast_fp16)[name = string("x_503_cast_fp16")];
            fp16 var_12029_promoted_to_fp16 = const()[name = string("op_12029_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_12030_cast_fp16 = pow(x = x_503_cast_fp16, y = var_12029_promoted_to_fp16)[name = string("op_12030_cast_fp16")];
            tensor<int32, [1]> var_323_axes_0 = const()[name = string("var_323_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_323_keep_dims_0 = const()[name = string("var_323_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_323_cast_fp16 = reduce_mean(axes = var_323_axes_0, keep_dims = var_323_keep_dims_0, x = var_12030_cast_fp16)[name = string("var_323_cast_fp16")];
            fp16 var_12037_to_fp16 = const()[name = string("op_12037_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_12038_cast_fp16 = add(x = var_323_cast_fp16, y = var_12037_to_fp16)[name = string("op_12038_cast_fp16")];
            fp32 var_12039_epsilon_0 = const()[name = string("op_12039_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_12039_cast_fp16 = rsqrt(epsilon = var_12039_epsilon_0, x = var_12038_cast_fp16)[name = string("op_12039_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_323_cast_fp16 = mul(x = x_503_cast_fp16, y = var_12039_cast_fp16)[name = string("x_normed_323_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_12049_to_fp16 = const()[name = string("op_12049_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536208832)))];
            tensor<fp16, [1, 640, 1, ?]> mlp_cast_fp16 = mul(x = x_normed_323_cast_fp16, y = var_12049_to_fp16)[name = string("mlp_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_cast_fp16 = add(x = x_501_cast_fp16, y = mlp_cast_fp16)[name = string("x_cast_fp16")];
            fp16 var_12053_promoted_to_fp16 = const()[name = string("op_12053_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640, 1, ?]> var_12054_cast_fp16 = pow(x = x_cast_fp16, y = var_12053_promoted_to_fp16)[name = string("op_12054_cast_fp16")];
            tensor<int32, [1]> var_axes_0 = const()[name = string("var_axes_0"), val = tensor<int32, [1]>([1])];
            bool var_keep_dims_0 = const()[name = string("var_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1, ?]> var_cast_fp16 = reduce_mean(axes = var_axes_0, keep_dims = var_keep_dims_0, x = var_12054_cast_fp16)[name = string("var_cast_fp16")];
            fp16 var_12061_to_fp16 = const()[name = string("op_12061_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1, ?]> var_12062_cast_fp16 = add(x = var_cast_fp16, y = var_12061_to_fp16)[name = string("op_12062_cast_fp16")];
            fp32 var_12063_epsilon_0 = const()[name = string("op_12063_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, ?]> var_12063_cast_fp16 = rsqrt(epsilon = var_12063_epsilon_0, x = var_12062_cast_fp16)[name = string("op_12063_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> x_normed_cast_fp16 = mul(x = x_cast_fp16, y = var_12063_cast_fp16)[name = string("x_normed_cast_fp16")];
            tensor<fp16, [1, 640, 1, 1]> var_12073_to_fp16 = const()[name = string("op_12073_to_fp16"), val = tensor<fp16, [1, 640, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536210176)))];
            tensor<fp16, [1, 640, 1, ?]> h_cast_fp16 = mul(x = x_normed_cast_fp16, y = var_12073_to_fp16)[name = string("h_cast_fp16")];
            tensor<fp16, [1, 640, 1, ?]> var_12075_cast_fp16 = mul(x = h_cast_fp16, y = logits_mask)[name = string("op_12075_cast_fp16")];
            tensor<int32, [1]> selected_axes_0 = const()[name = string("selected_axes_0"), val = tensor<int32, [1]>([3])];
            bool selected_keep_dims_0 = const()[name = string("selected_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 640, 1, 1]> selected_cast_fp16 = reduce_sum(axes = selected_axes_0, keep_dims = selected_keep_dims_0, x = var_12075_cast_fp16)[name = string("selected_cast_fp16")];
            tensor<int32, [1]> var_12082_axes_0 = const()[name = string("op_12082_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 640, 1]> var_12082_cast_fp16 = squeeze(axes = var_12082_axes_0, x = selected_cast_fp16)[name = string("op_12082_cast_fp16")];
            tensor<int32, [1]> input_axes_0 = const()[name = string("input_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 640]> input_cast_fp16 = squeeze(axes = input_axes_0, x = var_12082_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [262144]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [262144]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536211520)))];
            tensor<fp16, [1, 262144]> logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_weight_to_fp16, x = input_cast_fp16)[name = string("linear_0_cast_fp16")];
        } -> (logits);
}