program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})]
{
    func decoder<ios18>(tensor<fp16, [1, 1, ?, 1024]> attention_mask, tensor<fp16, [1, ?, 1024]> input_embeds, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_0, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_1, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_10, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_11, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_12, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_13, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_14, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_15, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_16, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_17, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_18, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_19, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_2, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_20, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_21, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_22, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_23, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_24, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_25, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_26, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_27, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_3, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_4, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_5, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_6, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_7, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_8, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_9, tensor<int32, [?]> positions, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_0, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_1, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_10, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_11, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_12, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_13, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_14, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_15, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_16, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_17, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_18, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_19, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_2, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_20, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_21, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_22, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_23, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_24, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_25, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_26, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_27, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_3, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_4, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_5, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_6, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_7, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_8, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_9) [FlexibleShapeInformation = tuple<tuple<string, dict<string, tensor<int32, [?]>>>, tuple<string, dict<string, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"attention_mask", [1, 1, 1, 1024]}, {"input_embeds", [1, 1, 1024]}, {"positions", [1]}}), ("RangeDims", {{"attention_mask", [[1, 1], [1, 1], [1, 128], [1024, 1024]]}, {"input_embeds", [[1, 1], [1, 128], [1024, 1024]]}, {"positions", [[1, 128]]}})))] {
            tensor<int32, [4]> var_127 = const()[name = string("op_127"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<int32, [1, 1, ?, 1]> var_128 = reshape(shape = var_127, x = positions)[name = string("op_128")];
            tensor<int32, [4]> var_133 = const()[name = string("op_133"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<int32, [1, 8, ?, 128]> scatter_indices = tile(reps = var_133, x = var_128)[name = string("scatter_indices")];
            int32 var_148 = const()[name = string("op_148"), val = int32(-1)];
            fp16 var_147_promoted_to_fp16 = const()[name = string("op_147_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_157_cast_fp16 = pow(x = input_embeds, y = var_147_promoted_to_fp16)[name = string("op_157_cast_fp16")];
            tensor<int32, [1]> var_159_axes_0 = const()[name = string("op_159_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_159_keep_dims_0 = const()[name = string("op_159_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_159_cast_fp16 = reduce_mean(axes = var_159_axes_0, keep_dims = var_159_keep_dims_0, x = var_157_cast_fp16)[name = string("op_159_cast_fp16")];
            fp16 var_160_to_fp16 = const()[name = string("op_160_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_161_cast_fp16 = add(x = var_159_cast_fp16, y = var_160_to_fp16)[name = string("op_161_cast_fp16")];
            fp32 norm_1_epsilon_0 = const()[name = string("norm_1_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_1_cast_fp16 = rsqrt(epsilon = norm_1_epsilon_0, x = var_161_cast_fp16)[name = string("norm_1_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_163_cast_fp16 = mul(x = input_embeds, y = norm_1_cast_fp16)[name = string("op_163_cast_fp16")];
            tensor<fp16, [1024]> layers_0_input_layernorm_weight_to_fp16 = const()[name = string("layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            tensor<fp16, [1, ?, 1024]> var_164_cast_fp16 = mul(x = var_163_cast_fp16, y = layers_0_input_layernorm_weight_to_fp16)[name = string("op_164_cast_fp16")];
            tensor<int32, [3]> var_176_shape_cast_fp16 = shape(x = var_164_cast_fp16)[name = string("op_176_shape_cast_fp16")];
            int32 gather_0 = const()[name = string("gather_0"), val = int32(1)];
            int32 gather_1_axis_0 = const()[name = string("gather_1_axis_0"), val = int32(0)];
            int32 gather_1_batch_dims_0 = const()[name = string("gather_1_batch_dims_0"), val = int32(0)];
            bool gather_1_validate_indices_0 = const()[name = string("gather_1_validate_indices_0"), val = bool(false)];
            string var_176_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_176_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")];
            uint16 select_1_to_uint16 = const()[name = string("select_1_to_uint16"), val = uint16(1)];
            tensor<int16, [3]> var_176_shape_cast_fp16_to_int16 = cast(dtype = var_176_shape_cast_fp16_to_int16_dtype_0, x = var_176_shape_cast_fp16)[name = string("cast_56")];
            int16 gather_1_cast_uint16 = gather(axis = gather_1_axis_0, batch_dims = gather_1_batch_dims_0, indices = select_1_to_uint16, validate_indices = gather_1_validate_indices_0, x = var_176_shape_cast_fp16_to_int16)[name = string("gather_1_cast_uint16")];
            string gather_1_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_1_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2176))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2099392))))[name = string("layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [2048]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2230528)))];
            tensor<fp16, [1, ?, 2048]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = var_164_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_181_cast_fp16 = reshape(shape = concat_0x, x = linear_0_cast_fp16)[name = string("op_181_cast_fp16")];
            tensor<int32, [4]> x_5_perm_0 = const()[name = string("x_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2234688))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3283328))))[name = string("layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1024]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3348928)))];
            tensor<fp16, [1, ?, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = var_164_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_186_cast_fp16 = reshape(shape = concat_1x, x = linear_1_cast_fp16)[name = string("op_186_cast_fp16")];
            tensor<int32, [4]> x_9_perm_0 = const()[name = string("x_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3351040))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4399680))))[name = string("layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_2_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = var_164_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_191_cast_fp16 = reshape(shape = concat_2x, x = linear_2_cast_fp16)[name = string("op_191_cast_fp16")];
            tensor<int32, [4]> v_1_perm_0 = const()[name = string("v_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_147_promoted_1_to_fp16 = const()[name = string("op_147_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = var_181_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 16, ?, 128]> var_195_cast_fp16 = pow(x = x_5_cast_fp16, y = var_147_promoted_1_to_fp16)[name = string("op_195_cast_fp16")];
            tensor<int32, [1]> var_197_axes_0 = const()[name = string("op_197_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_197_keep_dims_0 = const()[name = string("op_197_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_197_cast_fp16 = reduce_mean(axes = var_197_axes_0, keep_dims = var_197_keep_dims_0, x = var_195_cast_fp16)[name = string("op_197_cast_fp16")];
            fp16 var_198_to_fp16 = const()[name = string("op_198_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_199_cast_fp16 = add(x = var_197_cast_fp16, y = var_198_to_fp16)[name = string("op_199_cast_fp16")];
            fp32 norm_3_epsilon_0 = const()[name = string("norm_3_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_3_cast_fp16 = rsqrt(epsilon = norm_3_epsilon_0, x = var_199_cast_fp16)[name = string("norm_3_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_201_cast_fp16 = mul(x = x_5_cast_fp16, y = norm_3_cast_fp16)[name = string("op_201_cast_fp16")];
            tensor<fp16, [128]> layers_0_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4465280)))];
            tensor<fp16, [1, 16, ?, 128]> var_202_cast_fp16 = mul(x = var_201_cast_fp16, y = layers_0_self_attn_q_norm_weight_to_fp16)[name = string("op_202_cast_fp16")];
            fp16 var_147_promoted_2_to_fp16 = const()[name = string("op_147_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_9_cast_fp16 = transpose(perm = x_9_perm_0, x = var_186_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 8, ?, 128]> var_206_cast_fp16 = pow(x = x_9_cast_fp16, y = var_147_promoted_2_to_fp16)[name = string("op_206_cast_fp16")];
            tensor<int32, [1]> var_208_axes_0 = const()[name = string("op_208_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_208_keep_dims_0 = const()[name = string("op_208_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_208_cast_fp16 = reduce_mean(axes = var_208_axes_0, keep_dims = var_208_keep_dims_0, x = var_206_cast_fp16)[name = string("op_208_cast_fp16")];
            fp16 var_209_to_fp16 = const()[name = string("op_209_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_210_cast_fp16 = add(x = var_208_cast_fp16, y = var_209_to_fp16)[name = string("op_210_cast_fp16")];
            fp32 norm_5_epsilon_0 = const()[name = string("norm_5_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_5_cast_fp16 = rsqrt(epsilon = norm_5_epsilon_0, x = var_210_cast_fp16)[name = string("norm_5_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_212_cast_fp16 = mul(x = x_9_cast_fp16, y = norm_5_cast_fp16)[name = string("op_212_cast_fp16")];
            tensor<fp16, [128]> layers_0_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4465600)))];
            tensor<fp16, [1, 8, ?, 128]> var_213_cast_fp16 = mul(x = var_212_cast_fp16, y = layers_0_self_attn_k_norm_weight_to_fp16)[name = string("op_213_cast_fp16")];
            tensor<int32, [1]> var_217_axes_0 = const()[name = string("op_217_axes_0"), val = tensor<int32, [1]>([-1])];
            string cast_13_to_fp16_dtype_0 = const()[name = string("cast_13_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [?]> positions_to_fp16 = cast(dtype = cast_13_to_fp16_dtype_0, x = positions)[name = string("cast_55")];
            tensor<fp16, [?, 1]> var_217_cast_fp16 = expand_dims(axes = var_217_axes_0, x = positions_to_fp16)[name = string("op_217_cast_fp16")];
            tensor<fp16, [64]> layers_0_self_attn_rope_inv_freq_to_fp16 = const()[name = string("layers_0_self_attn_rope_inv_freq_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4465920)))];
            tensor<fp16, [?, 64]> freqs_1_cast_fp16 = mul(x = var_217_cast_fp16, y = layers_0_self_attn_rope_inv_freq_to_fp16)[name = string("freqs_1_cast_fp16")];
            tensor<fp16, [?, 64]> var_219_cast_fp16 = cos(x = freqs_1_cast_fp16)[name = string("op_219_cast_fp16")];
            tensor<int32, [4]> var_221 = const()[name = string("op_221"), val = tensor<int32, [4]>([1, 1, -1, 64])];
            tensor<fp16, [1, 1, ?, 64]> cos_val_1_cast_fp16 = reshape(shape = var_221, x = var_219_cast_fp16)[name = string("cos_val_1_cast_fp16")];
            tensor<fp16, [?, 64]> var_223_cast_fp16 = sin(x = freqs_1_cast_fp16)[name = string("op_223_cast_fp16")];
            tensor<int32, [4]> var_225 = const()[name = string("op_225"), val = tensor<int32, [4]>([1, 1, -1, 64])];
            tensor<fp16, [1, 1, ?, 64]> sin_val_1_cast_fp16 = reshape(shape = var_225, x = var_223_cast_fp16)[name = string("sin_val_1_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = var_202_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = var_202_cast_fp16)[name = string("x2_1_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_229_cast_fp16 = mul(x = x1_1_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_229_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_230_cast_fp16 = mul(x = x2_1_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_230_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_231_cast_fp16 = sub(x = var_229_cast_fp16, y = var_230_cast_fp16)[name = string("op_231_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_232_cast_fp16 = mul(x = x2_1_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_232_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_233_cast_fp16 = mul(x = x1_1_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_233_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_234_cast_fp16 = add(x = var_232_cast_fp16, y = var_233_cast_fp16)[name = string("op_234_cast_fp16")];
            bool q_1_interleave_0 = const()[name = string("q_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_1_cast_fp16 = concat(axis = var_148, interleave = q_1_interleave_0, values = (var_231_cast_fp16, var_234_cast_fp16))[name = string("q_1_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = var_213_cast_fp16)[name = string("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = var_213_cast_fp16)[name = string("x2_3_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_251_cast_fp16 = mul(x = x1_3_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_251_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_252_cast_fp16 = mul(x = x2_3_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_252_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_253_cast_fp16 = sub(x = var_251_cast_fp16, y = var_252_cast_fp16)[name = string("op_253_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_254_cast_fp16 = mul(x = x2_3_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_254_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_255_cast_fp16 = mul(x = x1_3_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_255_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_256_cast_fp16 = add(x = var_254_cast_fp16, y = var_255_cast_fp16)[name = string("op_256_cast_fp16")];
            bool k_1_interleave_0 = const()[name = string("k_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_1_cast_fp16 = concat(axis = var_148, interleave = k_1_interleave_0, values = (var_253_cast_fp16, var_256_cast_fp16))[name = string("k_1_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_0 = read_state(input = k_cache_0)[name = string("read_state_0")];
            int32 k_cache_3_axis_0 = const()[name = string("k_cache_3_axis_0"), val = int32(2)];
            string k_cache_3_mode_0 = const()[name = string("k_cache_3_mode_0"), val = string("update")];
            bool k_cache_3_validate_indices_0 = const()[name = string("k_cache_3_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_3_cast_fp16 = scatter_along_axis(axis = k_cache_3_axis_0, data = read_state_0, indices = scatter_indices, mode = k_cache_3_mode_0, updates = k_1_cast_fp16, validate_indices = k_cache_3_validate_indices_0)[name = string("k_cache_3_cast_fp16")];
            write_state(data = k_cache_3_cast_fp16, input = k_cache_0)[name = string("coreml_update_state_0_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_56 = read_state(input = k_cache_0)[name = string("coreml_update_state_0")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_1 = read_state(input = v_cache_0)[name = string("read_state_1")];
            int32 v_cache_3_axis_0 = const()[name = string("v_cache_3_axis_0"), val = int32(2)];
            string v_cache_3_mode_0 = const()[name = string("v_cache_3_mode_0"), val = string("update")];
            bool v_cache_3_validate_indices_0 = const()[name = string("v_cache_3_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_1_cast_fp16 = transpose(perm = v_1_perm_0, x = var_191_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_3_cast_fp16 = scatter_along_axis(axis = v_cache_3_axis_0, data = read_state_1, indices = scatter_indices, mode = v_cache_3_mode_0, updates = v_1_cast_fp16, validate_indices = v_cache_3_validate_indices_0)[name = string("v_cache_3_cast_fp16")];
            write_state(data = v_cache_3_cast_fp16, input = v_cache_0)[name = string("coreml_update_state_1_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_57 = read_state(input = v_cache_0)[name = string("coreml_update_state_1")];
            tensor<int32, [1]> var_261_axes_0 = const()[name = string("op_261_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_261_cast_fp16 = expand_dims(axes = var_261_axes_0, x = coreml_update_state_56)[name = string("op_261_cast_fp16")];
            tensor<int32, [5]> k_exp_1_reps_0 = const()[name = string("k_exp_1_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_1_cast_fp16 = tile(reps = k_exp_1_reps_0, x = var_261_cast_fp16)[name = string("k_exp_1_cast_fp16")];
            tensor<int32, [4]> var_264 = const()[name = string("op_264"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_3_cast_fp16 = reshape(shape = var_264, x = k_exp_1_cast_fp16)[name = string("k_exp_3_cast_fp16")];
            tensor<int32, [1]> var_266_axes_0 = const()[name = string("op_266_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_266_cast_fp16 = expand_dims(axes = var_266_axes_0, x = coreml_update_state_57)[name = string("op_266_cast_fp16")];
            tensor<int32, [5]> v_exp_1_reps_0 = const()[name = string("v_exp_1_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_1_cast_fp16 = tile(reps = v_exp_1_reps_0, x = var_266_cast_fp16)[name = string("v_exp_1_cast_fp16")];
            tensor<int32, [4]> var_269 = const()[name = string("op_269"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_3_cast_fp16 = reshape(shape = var_269, x = v_exp_1_cast_fp16)[name = string("v_exp_3_cast_fp16")];
            bool var_272_transpose_x_1 = const()[name = string("op_272_transpose_x_1"), val = bool(false)];
            bool var_272_transpose_y_1 = const()[name = string("op_272_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_272_cast_fp16 = matmul(transpose_x = var_272_transpose_x_1, transpose_y = var_272_transpose_y_1, x = q_1_cast_fp16, y = k_exp_3_cast_fp16)[name = string("op_272_cast_fp16")];
            fp16 var_273_to_fp16 = const()[name = string("op_273_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_1_cast_fp16 = mul(x = var_272_cast_fp16, y = var_273_to_fp16)[name = string("attn_1_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_1_cast_fp16 = add(x = attn_1_cast_fp16, y = attention_mask)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_3_cast_fp16 = softmax(axis = var_148, x = input_1_cast_fp16)[name = string("attn_3_cast_fp16")];
            bool out_1_transpose_x_0 = const()[name = string("out_1_transpose_x_0"), val = bool(false)];
            bool out_1_transpose_y_0 = const()[name = string("out_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_1_cast_fp16 = matmul(transpose_x = out_1_transpose_x_0, transpose_y = out_1_transpose_y_0, x = attn_3_cast_fp16, y = v_exp_3_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [4]> var_278_perm_0 = const()[name = string("op_278_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_5_axis_0 = const()[name = string("concat_5_axis_0"), val = int32(0)];
            bool concat_5_interleave_0 = const()[name = string("concat_5_interleave_0"), val = bool(false)];
            int32 gather_1_cast_uint16_to_int32 = cast(dtype = gather_1_cast_uint16_to_int32_dtype_0, x = gather_1_cast_uint16)[name = string("cast_54")];
            tensor<int32, [3]> concat_5 = concat(axis = concat_5_axis_0, interleave = concat_5_interleave_0, values = (gather_0, gather_1_cast_uint16_to_int32, var_148))[name = string("concat_5")];
            tensor<fp16, [1, ?, 16, 128]> var_278_cast_fp16 = transpose(perm = var_278_perm_0, x = out_1_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, ?, ?]> input_3_cast_fp16 = reshape(shape = concat_5, x = var_278_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4466112))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6563328))))[name = string("layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_19_cast_fp16 = add(x = input_embeds, y = linear_3_cast_fp16)[name = string("x_19_cast_fp16")];
            fp16 var_147_promoted_3_to_fp16 = const()[name = string("op_147_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_286_cast_fp16 = pow(x = x_19_cast_fp16, y = var_147_promoted_3_to_fp16)[name = string("op_286_cast_fp16")];
            tensor<int32, [1]> var_288_axes_0 = const()[name = string("op_288_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_288_keep_dims_0 = const()[name = string("op_288_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_288_cast_fp16 = reduce_mean(axes = var_288_axes_0, keep_dims = var_288_keep_dims_0, x = var_286_cast_fp16)[name = string("op_288_cast_fp16")];
            fp16 var_289_to_fp16 = const()[name = string("op_289_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_290_cast_fp16 = add(x = var_288_cast_fp16, y = var_289_to_fp16)[name = string("op_290_cast_fp16")];
            fp32 norm_7_epsilon_0 = const()[name = string("norm_7_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_7_cast_fp16 = rsqrt(epsilon = norm_7_epsilon_0, x = var_290_cast_fp16)[name = string("norm_7_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_292_cast_fp16 = mul(x = x_19_cast_fp16, y = norm_7_cast_fp16)[name = string("op_292_cast_fp16")];
            tensor<fp16, [1024]> layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6694464)))];
            tensor<fp16, [1, ?, 1024]> var_293_cast_fp16 = mul(x = var_292_cast_fp16, y = layers_0_post_attention_layernorm_weight_to_fp16)[name = string("op_293_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6696576))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9842368))))[name = string("layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [3072]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10039040)))];
            tensor<fp16, [1, ?, 3072]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = var_293_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_303_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_303_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10045248))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13191040))))[name = string("layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_0_mlp_up_proj_weight_to_fp16_quantized, x = var_293_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_9_cast_fp16 = mul(x = var_303_cast_fp16, y = linear_5_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13387712))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16533504))))[name = string("layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_6_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_25_cast_fp16 = add(x = x_19_cast_fp16, y = linear_6_cast_fp16)[name = string("x_25_cast_fp16")];
            int32 var_323 = const()[name = string("op_323"), val = int32(-1)];
            fp16 var_322_promoted_to_fp16 = const()[name = string("op_322_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_332_cast_fp16 = pow(x = x_25_cast_fp16, y = var_322_promoted_to_fp16)[name = string("op_332_cast_fp16")];
            tensor<int32, [1]> var_334_axes_0 = const()[name = string("op_334_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_334_keep_dims_0 = const()[name = string("op_334_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_334_cast_fp16 = reduce_mean(axes = var_334_axes_0, keep_dims = var_334_keep_dims_0, x = var_332_cast_fp16)[name = string("op_334_cast_fp16")];
            fp16 var_335_to_fp16 = const()[name = string("op_335_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_336_cast_fp16 = add(x = var_334_cast_fp16, y = var_335_to_fp16)[name = string("op_336_cast_fp16")];
            fp32 norm_9_epsilon_0 = const()[name = string("norm_9_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_9_cast_fp16 = rsqrt(epsilon = norm_9_epsilon_0, x = var_336_cast_fp16)[name = string("norm_9_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_338_cast_fp16 = mul(x = x_25_cast_fp16, y = norm_9_cast_fp16)[name = string("op_338_cast_fp16")];
            tensor<fp16, [1024]> layers_1_input_layernorm_weight_to_fp16 = const()[name = string("layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16730176)))];
            tensor<fp16, [1, ?, 1024]> var_339_cast_fp16 = mul(x = var_338_cast_fp16, y = layers_1_input_layernorm_weight_to_fp16)[name = string("op_339_cast_fp16")];
            tensor<int32, [3]> var_351_shape_cast_fp16 = shape(x = var_339_cast_fp16)[name = string("op_351_shape_cast_fp16")];
            int32 gather_6 = const()[name = string("gather_6"), val = int32(1)];
            int32 gather_7_axis_0 = const()[name = string("gather_7_axis_0"), val = int32(0)];
            int32 gather_7_batch_dims_0 = const()[name = string("gather_7_batch_dims_0"), val = int32(0)];
            bool gather_7_validate_indices_0 = const()[name = string("gather_7_validate_indices_0"), val = bool(false)];
            string var_351_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_351_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_7_to_uint16 = const()[name = string("select_7_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_351_shape_cast_fp16_to_uint16 = cast(dtype = var_351_shape_cast_fp16_to_uint16_dtype_0, x = var_351_shape_cast_fp16)[name = string("cast_53")];
            uint16 gather_7_cast_uint16 = gather(axis = gather_7_axis_0, batch_dims = gather_7_batch_dims_0, indices = select_7_to_uint16, validate_indices = gather_7_validate_indices_0, x = var_351_shape_cast_fp16_to_uint16)[name = string("gather_7_cast_uint16")];
            string gather_7_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_7_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16732288))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18829504))))[name = string("layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_7_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = var_339_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> concat_6x = const()[name = string("concat_6x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_356_cast_fp16 = reshape(shape = concat_6x, x = linear_7_cast_fp16)[name = string("op_356_cast_fp16")];
            tensor<int32, [4]> x_31_perm_0 = const()[name = string("x_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18960640))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20009280))))[name = string("layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_8_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = var_339_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> concat_7x = const()[name = string("concat_7x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_361_cast_fp16 = reshape(shape = concat_7x, x = linear_8_cast_fp16)[name = string("op_361_cast_fp16")];
            tensor<int32, [4]> x_35_perm_0 = const()[name = string("x_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20074880))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21123520))))[name = string("layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = var_339_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> concat_8x = const()[name = string("concat_8x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_366_cast_fp16 = reshape(shape = concat_8x, x = linear_9_cast_fp16)[name = string("op_366_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_322_promoted_1_to_fp16 = const()[name = string("op_322_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_31_cast_fp16 = transpose(perm = x_31_perm_0, x = var_356_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 16, ?, 128]> var_370_cast_fp16 = pow(x = x_31_cast_fp16, y = var_322_promoted_1_to_fp16)[name = string("op_370_cast_fp16")];
            tensor<int32, [1]> var_372_axes_0 = const()[name = string("op_372_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_372_keep_dims_0 = const()[name = string("op_372_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_372_cast_fp16 = reduce_mean(axes = var_372_axes_0, keep_dims = var_372_keep_dims_0, x = var_370_cast_fp16)[name = string("op_372_cast_fp16")];
            fp16 var_373_to_fp16 = const()[name = string("op_373_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_374_cast_fp16 = add(x = var_372_cast_fp16, y = var_373_to_fp16)[name = string("op_374_cast_fp16")];
            fp32 norm_11_epsilon_0 = const()[name = string("norm_11_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_11_cast_fp16 = rsqrt(epsilon = norm_11_epsilon_0, x = var_374_cast_fp16)[name = string("norm_11_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_376_cast_fp16 = mul(x = x_31_cast_fp16, y = norm_11_cast_fp16)[name = string("op_376_cast_fp16")];
            tensor<fp16, [128]> layers_1_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21189120)))];
            tensor<fp16, [1, 16, ?, 128]> var_377_cast_fp16 = mul(x = var_376_cast_fp16, y = layers_1_self_attn_q_norm_weight_to_fp16)[name = string("op_377_cast_fp16")];
            fp16 var_322_promoted_2_to_fp16 = const()[name = string("op_322_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_35_cast_fp16 = transpose(perm = x_35_perm_0, x = var_361_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [1, 8, ?, 128]> var_381_cast_fp16 = pow(x = x_35_cast_fp16, y = var_322_promoted_2_to_fp16)[name = string("op_381_cast_fp16")];
            tensor<int32, [1]> var_383_axes_0 = const()[name = string("op_383_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_383_keep_dims_0 = const()[name = string("op_383_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_383_cast_fp16 = reduce_mean(axes = var_383_axes_0, keep_dims = var_383_keep_dims_0, x = var_381_cast_fp16)[name = string("op_383_cast_fp16")];
            fp16 var_384_to_fp16 = const()[name = string("op_384_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_385_cast_fp16 = add(x = var_383_cast_fp16, y = var_384_to_fp16)[name = string("op_385_cast_fp16")];
            fp32 norm_13_epsilon_0 = const()[name = string("norm_13_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_13_cast_fp16 = rsqrt(epsilon = norm_13_epsilon_0, x = var_385_cast_fp16)[name = string("norm_13_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_387_cast_fp16 = mul(x = x_35_cast_fp16, y = norm_13_cast_fp16)[name = string("op_387_cast_fp16")];
            tensor<fp16, [128]> layers_1_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21189440)))];
            tensor<fp16, [1, 8, ?, 128]> var_388_cast_fp16 = mul(x = var_387_cast_fp16, y = layers_1_self_attn_k_norm_weight_to_fp16)[name = string("op_388_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = var_377_cast_fp16)[name = string("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = var_377_cast_fp16)[name = string("x2_5_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_404_cast_fp16 = mul(x = x1_5_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_404_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_405_cast_fp16 = mul(x = x2_5_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_405_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_406_cast_fp16 = sub(x = var_404_cast_fp16, y = var_405_cast_fp16)[name = string("op_406_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_407_cast_fp16 = mul(x = x2_5_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_407_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_408_cast_fp16 = mul(x = x1_5_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_408_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_409_cast_fp16 = add(x = var_407_cast_fp16, y = var_408_cast_fp16)[name = string("op_409_cast_fp16")];
            bool q_3_interleave_0 = const()[name = string("q_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_3_cast_fp16 = concat(axis = var_323, interleave = q_3_interleave_0, values = (var_406_cast_fp16, var_409_cast_fp16))[name = string("q_3_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = var_388_cast_fp16)[name = string("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = var_388_cast_fp16)[name = string("x2_7_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_426_cast_fp16 = mul(x = x1_7_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_426_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_427_cast_fp16 = mul(x = x2_7_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_427_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_428_cast_fp16 = sub(x = var_426_cast_fp16, y = var_427_cast_fp16)[name = string("op_428_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_429_cast_fp16 = mul(x = x2_7_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_429_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_430_cast_fp16 = mul(x = x1_7_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_430_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_431_cast_fp16 = add(x = var_429_cast_fp16, y = var_430_cast_fp16)[name = string("op_431_cast_fp16")];
            bool k_3_interleave_0 = const()[name = string("k_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_3_cast_fp16 = concat(axis = var_323, interleave = k_3_interleave_0, values = (var_428_cast_fp16, var_431_cast_fp16))[name = string("k_3_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_2 = read_state(input = k_cache_1)[name = string("read_state_2")];
            int32 k_cache_7_axis_0 = const()[name = string("k_cache_7_axis_0"), val = int32(2)];
            string k_cache_7_mode_0 = const()[name = string("k_cache_7_mode_0"), val = string("update")];
            bool k_cache_7_validate_indices_0 = const()[name = string("k_cache_7_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_7_cast_fp16 = scatter_along_axis(axis = k_cache_7_axis_0, data = read_state_2, indices = scatter_indices, mode = k_cache_7_mode_0, updates = k_3_cast_fp16, validate_indices = k_cache_7_validate_indices_0)[name = string("k_cache_7_cast_fp16")];
            write_state(data = k_cache_7_cast_fp16, input = k_cache_1)[name = string("coreml_update_state_2_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_58 = read_state(input = k_cache_1)[name = string("coreml_update_state_2")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_3 = read_state(input = v_cache_1)[name = string("read_state_3")];
            int32 v_cache_7_axis_0 = const()[name = string("v_cache_7_axis_0"), val = int32(2)];
            string v_cache_7_mode_0 = const()[name = string("v_cache_7_mode_0"), val = string("update")];
            bool v_cache_7_validate_indices_0 = const()[name = string("v_cache_7_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_366_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_7_cast_fp16 = scatter_along_axis(axis = v_cache_7_axis_0, data = read_state_3, indices = scatter_indices, mode = v_cache_7_mode_0, updates = v_3_cast_fp16, validate_indices = v_cache_7_validate_indices_0)[name = string("v_cache_7_cast_fp16")];
            write_state(data = v_cache_7_cast_fp16, input = v_cache_1)[name = string("coreml_update_state_3_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_59 = read_state(input = v_cache_1)[name = string("coreml_update_state_3")];
            tensor<int32, [1]> var_436_axes_0 = const()[name = string("op_436_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_436_cast_fp16 = expand_dims(axes = var_436_axes_0, x = coreml_update_state_58)[name = string("op_436_cast_fp16")];
            tensor<int32, [5]> k_exp_5_reps_0 = const()[name = string("k_exp_5_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_5_cast_fp16 = tile(reps = k_exp_5_reps_0, x = var_436_cast_fp16)[name = string("k_exp_5_cast_fp16")];
            tensor<int32, [4]> var_439 = const()[name = string("op_439"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_7_cast_fp16 = reshape(shape = var_439, x = k_exp_5_cast_fp16)[name = string("k_exp_7_cast_fp16")];
            tensor<int32, [1]> var_441_axes_0 = const()[name = string("op_441_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_441_cast_fp16 = expand_dims(axes = var_441_axes_0, x = coreml_update_state_59)[name = string("op_441_cast_fp16")];
            tensor<int32, [5]> v_exp_5_reps_0 = const()[name = string("v_exp_5_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_5_cast_fp16 = tile(reps = v_exp_5_reps_0, x = var_441_cast_fp16)[name = string("v_exp_5_cast_fp16")];
            tensor<int32, [4]> var_444 = const()[name = string("op_444"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_7_cast_fp16 = reshape(shape = var_444, x = v_exp_5_cast_fp16)[name = string("v_exp_7_cast_fp16")];
            bool var_447_transpose_x_1 = const()[name = string("op_447_transpose_x_1"), val = bool(false)];
            bool var_447_transpose_y_1 = const()[name = string("op_447_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_447_cast_fp16 = matmul(transpose_x = var_447_transpose_x_1, transpose_y = var_447_transpose_y_1, x = q_3_cast_fp16, y = k_exp_7_cast_fp16)[name = string("op_447_cast_fp16")];
            fp16 var_448_to_fp16 = const()[name = string("op_448_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_5_cast_fp16 = mul(x = var_447_cast_fp16, y = var_448_to_fp16)[name = string("attn_5_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_11_cast_fp16 = add(x = attn_5_cast_fp16, y = attention_mask)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_7_cast_fp16 = softmax(axis = var_323, x = input_11_cast_fp16)[name = string("attn_7_cast_fp16")];
            bool out_3_transpose_x_0 = const()[name = string("out_3_transpose_x_0"), val = bool(false)];
            bool out_3_transpose_y_0 = const()[name = string("out_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_3_cast_fp16 = matmul(transpose_x = out_3_transpose_x_0, transpose_y = out_3_transpose_y_0, x = attn_7_cast_fp16, y = v_exp_7_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [4]> var_453_perm_0 = const()[name = string("op_453_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_11_axis_0 = const()[name = string("concat_11_axis_0"), val = int32(0)];
            bool concat_11_interleave_0 = const()[name = string("concat_11_interleave_0"), val = bool(false)];
            int32 gather_7_cast_uint16_to_int32 = cast(dtype = gather_7_cast_uint16_to_int32_dtype_0, x = gather_7_cast_uint16)[name = string("cast_52")];
            tensor<int32, [3]> concat_11 = concat(axis = concat_11_axis_0, interleave = concat_11_interleave_0, values = (gather_6, gather_7_cast_uint16_to_int32, var_323))[name = string("concat_11")];
            tensor<fp16, [1, ?, 16, 128]> var_453_cast_fp16 = transpose(perm = var_453_perm_0, x = out_3_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [1, ?, ?]> input_13_cast_fp16 = reshape(shape = concat_11, x = var_453_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21189760))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23286976))))[name = string("layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_45_cast_fp16 = add(x = x_25_cast_fp16, y = linear_10_cast_fp16)[name = string("x_45_cast_fp16")];
            fp16 var_322_promoted_3_to_fp16 = const()[name = string("op_322_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_461_cast_fp16 = pow(x = x_45_cast_fp16, y = var_322_promoted_3_to_fp16)[name = string("op_461_cast_fp16")];
            tensor<int32, [1]> var_463_axes_0 = const()[name = string("op_463_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_463_keep_dims_0 = const()[name = string("op_463_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_463_cast_fp16 = reduce_mean(axes = var_463_axes_0, keep_dims = var_463_keep_dims_0, x = var_461_cast_fp16)[name = string("op_463_cast_fp16")];
            fp16 var_464_to_fp16 = const()[name = string("op_464_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_465_cast_fp16 = add(x = var_463_cast_fp16, y = var_464_to_fp16)[name = string("op_465_cast_fp16")];
            fp32 norm_15_epsilon_0 = const()[name = string("norm_15_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_15_cast_fp16 = rsqrt(epsilon = norm_15_epsilon_0, x = var_465_cast_fp16)[name = string("norm_15_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_467_cast_fp16 = mul(x = x_45_cast_fp16, y = norm_15_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<fp16, [1024]> layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23418112)))];
            tensor<fp16, [1, ?, 1024]> var_468_cast_fp16 = mul(x = var_467_cast_fp16, y = layers_1_post_attention_layernorm_weight_to_fp16)[name = string("op_468_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23420224))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26566016))))[name = string("layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = var_468_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_478_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_478_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26762688))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29908480))))[name = string("layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_1_mlp_up_proj_weight_to_fp16_quantized, x = var_468_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_19_cast_fp16 = mul(x = var_478_cast_fp16, y = linear_12_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30105152))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33250944))))[name = string("layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_51_cast_fp16 = add(x = x_45_cast_fp16, y = linear_13_cast_fp16)[name = string("x_51_cast_fp16")];
            int32 var_498 = const()[name = string("op_498"), val = int32(-1)];
            fp16 var_497_promoted_to_fp16 = const()[name = string("op_497_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_507_cast_fp16 = pow(x = x_51_cast_fp16, y = var_497_promoted_to_fp16)[name = string("op_507_cast_fp16")];
            tensor<int32, [1]> var_509_axes_0 = const()[name = string("op_509_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_509_keep_dims_0 = const()[name = string("op_509_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_509_cast_fp16 = reduce_mean(axes = var_509_axes_0, keep_dims = var_509_keep_dims_0, x = var_507_cast_fp16)[name = string("op_509_cast_fp16")];
            fp16 var_510_to_fp16 = const()[name = string("op_510_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_511_cast_fp16 = add(x = var_509_cast_fp16, y = var_510_to_fp16)[name = string("op_511_cast_fp16")];
            fp32 norm_17_epsilon_0 = const()[name = string("norm_17_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_17_cast_fp16 = rsqrt(epsilon = norm_17_epsilon_0, x = var_511_cast_fp16)[name = string("norm_17_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_513_cast_fp16 = mul(x = x_51_cast_fp16, y = norm_17_cast_fp16)[name = string("op_513_cast_fp16")];
            tensor<fp16, [1024]> layers_2_input_layernorm_weight_to_fp16 = const()[name = string("layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33447616)))];
            tensor<fp16, [1, ?, 1024]> var_514_cast_fp16 = mul(x = var_513_cast_fp16, y = layers_2_input_layernorm_weight_to_fp16)[name = string("op_514_cast_fp16")];
            tensor<int32, [3]> var_526_shape_cast_fp16 = shape(x = var_514_cast_fp16)[name = string("op_526_shape_cast_fp16")];
            int32 gather_12 = const()[name = string("gather_12"), val = int32(1)];
            int32 gather_13_axis_0 = const()[name = string("gather_13_axis_0"), val = int32(0)];
            int32 gather_13_batch_dims_0 = const()[name = string("gather_13_batch_dims_0"), val = int32(0)];
            bool gather_13_validate_indices_0 = const()[name = string("gather_13_validate_indices_0"), val = bool(false)];
            string var_526_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_526_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_13_to_uint16 = const()[name = string("select_13_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_526_shape_cast_fp16_to_uint16 = cast(dtype = var_526_shape_cast_fp16_to_uint16_dtype_0, x = var_526_shape_cast_fp16)[name = string("cast_51")];
            uint16 gather_13_cast_uint16 = gather(axis = gather_13_axis_0, batch_dims = gather_13_batch_dims_0, indices = select_13_to_uint16, validate_indices = gather_13_validate_indices_0, x = var_526_shape_cast_fp16_to_uint16)[name = string("gather_13_cast_uint16")];
            string gather_13_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_13_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33449728))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35546944))))[name = string("layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = var_514_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_531_cast_fp16 = reshape(shape = concat_12x, x = linear_14_cast_fp16)[name = string("op_531_cast_fp16")];
            tensor<int32, [4]> x_57_perm_0 = const()[name = string("x_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35678080))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36726720))))[name = string("layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = var_514_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_536_cast_fp16 = reshape(shape = concat_13x, x = linear_15_cast_fp16)[name = string("op_536_cast_fp16")];
            tensor<int32, [4]> x_61_perm_0 = const()[name = string("x_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36792320))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37840960))))[name = string("layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_16_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = var_514_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_541_cast_fp16 = reshape(shape = concat_14x, x = linear_16_cast_fp16)[name = string("op_541_cast_fp16")];
            tensor<int32, [4]> v_5_perm_0 = const()[name = string("v_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_497_promoted_1_to_fp16 = const()[name = string("op_497_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_57_cast_fp16 = transpose(perm = x_57_perm_0, x = var_531_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 16, ?, 128]> var_545_cast_fp16 = pow(x = x_57_cast_fp16, y = var_497_promoted_1_to_fp16)[name = string("op_545_cast_fp16")];
            tensor<int32, [1]> var_547_axes_0 = const()[name = string("op_547_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_547_keep_dims_0 = const()[name = string("op_547_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_547_cast_fp16 = reduce_mean(axes = var_547_axes_0, keep_dims = var_547_keep_dims_0, x = var_545_cast_fp16)[name = string("op_547_cast_fp16")];
            fp16 var_548_to_fp16 = const()[name = string("op_548_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_549_cast_fp16 = add(x = var_547_cast_fp16, y = var_548_to_fp16)[name = string("op_549_cast_fp16")];
            fp32 norm_19_epsilon_0 = const()[name = string("norm_19_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_19_cast_fp16 = rsqrt(epsilon = norm_19_epsilon_0, x = var_549_cast_fp16)[name = string("norm_19_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_551_cast_fp16 = mul(x = x_57_cast_fp16, y = norm_19_cast_fp16)[name = string("op_551_cast_fp16")];
            tensor<fp16, [128]> layers_2_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37906560)))];
            tensor<fp16, [1, 16, ?, 128]> var_552_cast_fp16 = mul(x = var_551_cast_fp16, y = layers_2_self_attn_q_norm_weight_to_fp16)[name = string("op_552_cast_fp16")];
            fp16 var_497_promoted_2_to_fp16 = const()[name = string("op_497_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_61_cast_fp16 = transpose(perm = x_61_perm_0, x = var_536_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 8, ?, 128]> var_556_cast_fp16 = pow(x = x_61_cast_fp16, y = var_497_promoted_2_to_fp16)[name = string("op_556_cast_fp16")];
            tensor<int32, [1]> var_558_axes_0 = const()[name = string("op_558_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_558_keep_dims_0 = const()[name = string("op_558_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_558_cast_fp16 = reduce_mean(axes = var_558_axes_0, keep_dims = var_558_keep_dims_0, x = var_556_cast_fp16)[name = string("op_558_cast_fp16")];
            fp16 var_559_to_fp16 = const()[name = string("op_559_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_560_cast_fp16 = add(x = var_558_cast_fp16, y = var_559_to_fp16)[name = string("op_560_cast_fp16")];
            fp32 norm_21_epsilon_0 = const()[name = string("norm_21_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_21_cast_fp16 = rsqrt(epsilon = norm_21_epsilon_0, x = var_560_cast_fp16)[name = string("norm_21_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_562_cast_fp16 = mul(x = x_61_cast_fp16, y = norm_21_cast_fp16)[name = string("op_562_cast_fp16")];
            tensor<fp16, [128]> layers_2_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37906880)))];
            tensor<fp16, [1, 8, ?, 128]> var_563_cast_fp16 = mul(x = var_562_cast_fp16, y = layers_2_self_attn_k_norm_weight_to_fp16)[name = string("op_563_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = var_552_cast_fp16)[name = string("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = var_552_cast_fp16)[name = string("x2_9_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_579_cast_fp16 = mul(x = x1_9_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_579_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_580_cast_fp16 = mul(x = x2_9_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_580_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_581_cast_fp16 = sub(x = var_579_cast_fp16, y = var_580_cast_fp16)[name = string("op_581_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_582_cast_fp16 = mul(x = x2_9_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_582_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_583_cast_fp16 = mul(x = x1_9_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_583_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_584_cast_fp16 = add(x = var_582_cast_fp16, y = var_583_cast_fp16)[name = string("op_584_cast_fp16")];
            bool q_5_interleave_0 = const()[name = string("q_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_5_cast_fp16 = concat(axis = var_498, interleave = q_5_interleave_0, values = (var_581_cast_fp16, var_584_cast_fp16))[name = string("q_5_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = var_563_cast_fp16)[name = string("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = var_563_cast_fp16)[name = string("x2_11_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_601_cast_fp16 = mul(x = x1_11_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_601_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_602_cast_fp16 = mul(x = x2_11_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_602_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_603_cast_fp16 = sub(x = var_601_cast_fp16, y = var_602_cast_fp16)[name = string("op_603_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_604_cast_fp16 = mul(x = x2_11_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_604_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_605_cast_fp16 = mul(x = x1_11_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_605_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_606_cast_fp16 = add(x = var_604_cast_fp16, y = var_605_cast_fp16)[name = string("op_606_cast_fp16")];
            bool k_5_interleave_0 = const()[name = string("k_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_5_cast_fp16 = concat(axis = var_498, interleave = k_5_interleave_0, values = (var_603_cast_fp16, var_606_cast_fp16))[name = string("k_5_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_4 = read_state(input = k_cache_2)[name = string("read_state_4")];
            int32 k_cache_11_axis_0 = const()[name = string("k_cache_11_axis_0"), val = int32(2)];
            string k_cache_11_mode_0 = const()[name = string("k_cache_11_mode_0"), val = string("update")];
            bool k_cache_11_validate_indices_0 = const()[name = string("k_cache_11_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_11_cast_fp16 = scatter_along_axis(axis = k_cache_11_axis_0, data = read_state_4, indices = scatter_indices, mode = k_cache_11_mode_0, updates = k_5_cast_fp16, validate_indices = k_cache_11_validate_indices_0)[name = string("k_cache_11_cast_fp16")];
            write_state(data = k_cache_11_cast_fp16, input = k_cache_2)[name = string("coreml_update_state_4_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_60 = read_state(input = k_cache_2)[name = string("coreml_update_state_4")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_5 = read_state(input = v_cache_2)[name = string("read_state_5")];
            int32 v_cache_11_axis_0 = const()[name = string("v_cache_11_axis_0"), val = int32(2)];
            string v_cache_11_mode_0 = const()[name = string("v_cache_11_mode_0"), val = string("update")];
            bool v_cache_11_validate_indices_0 = const()[name = string("v_cache_11_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_5_cast_fp16 = transpose(perm = v_5_perm_0, x = var_541_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_11_cast_fp16 = scatter_along_axis(axis = v_cache_11_axis_0, data = read_state_5, indices = scatter_indices, mode = v_cache_11_mode_0, updates = v_5_cast_fp16, validate_indices = v_cache_11_validate_indices_0)[name = string("v_cache_11_cast_fp16")];
            write_state(data = v_cache_11_cast_fp16, input = v_cache_2)[name = string("coreml_update_state_5_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_61 = read_state(input = v_cache_2)[name = string("coreml_update_state_5")];
            tensor<int32, [1]> var_611_axes_0 = const()[name = string("op_611_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_611_cast_fp16 = expand_dims(axes = var_611_axes_0, x = coreml_update_state_60)[name = string("op_611_cast_fp16")];
            tensor<int32, [5]> k_exp_9_reps_0 = const()[name = string("k_exp_9_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_9_cast_fp16 = tile(reps = k_exp_9_reps_0, x = var_611_cast_fp16)[name = string("k_exp_9_cast_fp16")];
            tensor<int32, [4]> var_614 = const()[name = string("op_614"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_11_cast_fp16 = reshape(shape = var_614, x = k_exp_9_cast_fp16)[name = string("k_exp_11_cast_fp16")];
            tensor<int32, [1]> var_616_axes_0 = const()[name = string("op_616_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_616_cast_fp16 = expand_dims(axes = var_616_axes_0, x = coreml_update_state_61)[name = string("op_616_cast_fp16")];
            tensor<int32, [5]> v_exp_9_reps_0 = const()[name = string("v_exp_9_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_9_cast_fp16 = tile(reps = v_exp_9_reps_0, x = var_616_cast_fp16)[name = string("v_exp_9_cast_fp16")];
            tensor<int32, [4]> var_619 = const()[name = string("op_619"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_11_cast_fp16 = reshape(shape = var_619, x = v_exp_9_cast_fp16)[name = string("v_exp_11_cast_fp16")];
            bool var_622_transpose_x_1 = const()[name = string("op_622_transpose_x_1"), val = bool(false)];
            bool var_622_transpose_y_1 = const()[name = string("op_622_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_622_cast_fp16 = matmul(transpose_x = var_622_transpose_x_1, transpose_y = var_622_transpose_y_1, x = q_5_cast_fp16, y = k_exp_11_cast_fp16)[name = string("op_622_cast_fp16")];
            fp16 var_623_to_fp16 = const()[name = string("op_623_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_9_cast_fp16 = mul(x = var_622_cast_fp16, y = var_623_to_fp16)[name = string("attn_9_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_21_cast_fp16 = add(x = attn_9_cast_fp16, y = attention_mask)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_11_cast_fp16 = softmax(axis = var_498, x = input_21_cast_fp16)[name = string("attn_11_cast_fp16")];
            bool out_5_transpose_x_0 = const()[name = string("out_5_transpose_x_0"), val = bool(false)];
            bool out_5_transpose_y_0 = const()[name = string("out_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_5_cast_fp16 = matmul(transpose_x = out_5_transpose_x_0, transpose_y = out_5_transpose_y_0, x = attn_11_cast_fp16, y = v_exp_11_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [4]> var_628_perm_0 = const()[name = string("op_628_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_17_axis_0 = const()[name = string("concat_17_axis_0"), val = int32(0)];
            bool concat_17_interleave_0 = const()[name = string("concat_17_interleave_0"), val = bool(false)];
            int32 gather_13_cast_uint16_to_int32 = cast(dtype = gather_13_cast_uint16_to_int32_dtype_0, x = gather_13_cast_uint16)[name = string("cast_50")];
            tensor<int32, [3]> concat_17 = concat(axis = concat_17_axis_0, interleave = concat_17_interleave_0, values = (gather_12, gather_13_cast_uint16_to_int32, var_498))[name = string("concat_17")];
            tensor<fp16, [1, ?, 16, 128]> var_628_cast_fp16 = transpose(perm = var_628_perm_0, x = out_5_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, ?, ?]> input_23_cast_fp16 = reshape(shape = concat_17, x = var_628_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37907200))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40004416))))[name = string("layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_71_cast_fp16 = add(x = x_51_cast_fp16, y = linear_17_cast_fp16)[name = string("x_71_cast_fp16")];
            fp16 var_497_promoted_3_to_fp16 = const()[name = string("op_497_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_636_cast_fp16 = pow(x = x_71_cast_fp16, y = var_497_promoted_3_to_fp16)[name = string("op_636_cast_fp16")];
            tensor<int32, [1]> var_638_axes_0 = const()[name = string("op_638_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_638_keep_dims_0 = const()[name = string("op_638_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_638_cast_fp16 = reduce_mean(axes = var_638_axes_0, keep_dims = var_638_keep_dims_0, x = var_636_cast_fp16)[name = string("op_638_cast_fp16")];
            fp16 var_639_to_fp16 = const()[name = string("op_639_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_640_cast_fp16 = add(x = var_638_cast_fp16, y = var_639_to_fp16)[name = string("op_640_cast_fp16")];
            fp32 norm_23_epsilon_0 = const()[name = string("norm_23_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_23_cast_fp16 = rsqrt(epsilon = norm_23_epsilon_0, x = var_640_cast_fp16)[name = string("norm_23_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_642_cast_fp16 = mul(x = x_71_cast_fp16, y = norm_23_cast_fp16)[name = string("op_642_cast_fp16")];
            tensor<fp16, [1024]> layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40135552)))];
            tensor<fp16, [1, ?, 1024]> var_643_cast_fp16 = mul(x = var_642_cast_fp16, y = layers_2_post_attention_layernorm_weight_to_fp16)[name = string("op_643_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40137664))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43283456))))[name = string("layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = var_643_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_653_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_653_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43480128))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46625920))))[name = string("layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_2_mlp_up_proj_weight_to_fp16_quantized, x = var_643_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_29_cast_fp16 = mul(x = var_653_cast_fp16, y = linear_19_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46822592))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49968384))))[name = string("layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_20_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_77_cast_fp16 = add(x = x_71_cast_fp16, y = linear_20_cast_fp16)[name = string("x_77_cast_fp16")];
            int32 var_673 = const()[name = string("op_673"), val = int32(-1)];
            fp16 var_672_promoted_to_fp16 = const()[name = string("op_672_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_682_cast_fp16 = pow(x = x_77_cast_fp16, y = var_672_promoted_to_fp16)[name = string("op_682_cast_fp16")];
            tensor<int32, [1]> var_684_axes_0 = const()[name = string("op_684_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_684_keep_dims_0 = const()[name = string("op_684_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_684_cast_fp16 = reduce_mean(axes = var_684_axes_0, keep_dims = var_684_keep_dims_0, x = var_682_cast_fp16)[name = string("op_684_cast_fp16")];
            fp16 var_685_to_fp16 = const()[name = string("op_685_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_686_cast_fp16 = add(x = var_684_cast_fp16, y = var_685_to_fp16)[name = string("op_686_cast_fp16")];
            fp32 norm_25_epsilon_0 = const()[name = string("norm_25_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_25_cast_fp16 = rsqrt(epsilon = norm_25_epsilon_0, x = var_686_cast_fp16)[name = string("norm_25_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_688_cast_fp16 = mul(x = x_77_cast_fp16, y = norm_25_cast_fp16)[name = string("op_688_cast_fp16")];
            tensor<fp16, [1024]> layers_3_input_layernorm_weight_to_fp16 = const()[name = string("layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50165056)))];
            tensor<fp16, [1, ?, 1024]> var_689_cast_fp16 = mul(x = var_688_cast_fp16, y = layers_3_input_layernorm_weight_to_fp16)[name = string("op_689_cast_fp16")];
            tensor<int32, [3]> var_701_shape_cast_fp16 = shape(x = var_689_cast_fp16)[name = string("op_701_shape_cast_fp16")];
            int32 gather_18 = const()[name = string("gather_18"), val = int32(1)];
            int32 gather_19_axis_0 = const()[name = string("gather_19_axis_0"), val = int32(0)];
            int32 gather_19_batch_dims_0 = const()[name = string("gather_19_batch_dims_0"), val = int32(0)];
            bool gather_19_validate_indices_0 = const()[name = string("gather_19_validate_indices_0"), val = bool(false)];
            string var_701_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_701_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_19_to_uint16 = const()[name = string("select_19_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_701_shape_cast_fp16_to_uint16 = cast(dtype = var_701_shape_cast_fp16_to_uint16_dtype_0, x = var_701_shape_cast_fp16)[name = string("cast_49")];
            uint16 gather_19_cast_uint16 = gather(axis = gather_19_axis_0, batch_dims = gather_19_batch_dims_0, indices = select_19_to_uint16, validate_indices = gather_19_validate_indices_0, x = var_701_shape_cast_fp16_to_uint16)[name = string("gather_19_cast_uint16")];
            string gather_19_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_19_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50167168))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52264384))))[name = string("layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_21_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = var_689_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> concat_18x = const()[name = string("concat_18x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_706_cast_fp16 = reshape(shape = concat_18x, x = linear_21_cast_fp16)[name = string("op_706_cast_fp16")];
            tensor<int32, [4]> x_83_perm_0 = const()[name = string("x_83_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52395520))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53444160))))[name = string("layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_22_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = var_689_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> concat_19x = const()[name = string("concat_19x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_711_cast_fp16 = reshape(shape = concat_19x, x = linear_22_cast_fp16)[name = string("op_711_cast_fp16")];
            tensor<int32, [4]> x_87_perm_0 = const()[name = string("x_87_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53509760))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54558400))))[name = string("layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = var_689_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> concat_20x = const()[name = string("concat_20x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_716_cast_fp16 = reshape(shape = concat_20x, x = linear_23_cast_fp16)[name = string("op_716_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_672_promoted_1_to_fp16 = const()[name = string("op_672_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_83_cast_fp16 = transpose(perm = x_83_perm_0, x = var_706_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 16, ?, 128]> var_720_cast_fp16 = pow(x = x_83_cast_fp16, y = var_672_promoted_1_to_fp16)[name = string("op_720_cast_fp16")];
            tensor<int32, [1]> var_722_axes_0 = const()[name = string("op_722_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_722_keep_dims_0 = const()[name = string("op_722_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_722_cast_fp16 = reduce_mean(axes = var_722_axes_0, keep_dims = var_722_keep_dims_0, x = var_720_cast_fp16)[name = string("op_722_cast_fp16")];
            fp16 var_723_to_fp16 = const()[name = string("op_723_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_724_cast_fp16 = add(x = var_722_cast_fp16, y = var_723_to_fp16)[name = string("op_724_cast_fp16")];
            fp32 norm_27_epsilon_0 = const()[name = string("norm_27_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_27_cast_fp16 = rsqrt(epsilon = norm_27_epsilon_0, x = var_724_cast_fp16)[name = string("norm_27_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_726_cast_fp16 = mul(x = x_83_cast_fp16, y = norm_27_cast_fp16)[name = string("op_726_cast_fp16")];
            tensor<fp16, [128]> layers_3_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54624000)))];
            tensor<fp16, [1, 16, ?, 128]> var_727_cast_fp16 = mul(x = var_726_cast_fp16, y = layers_3_self_attn_q_norm_weight_to_fp16)[name = string("op_727_cast_fp16")];
            fp16 var_672_promoted_2_to_fp16 = const()[name = string("op_672_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_87_cast_fp16 = transpose(perm = x_87_perm_0, x = var_711_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 8, ?, 128]> var_731_cast_fp16 = pow(x = x_87_cast_fp16, y = var_672_promoted_2_to_fp16)[name = string("op_731_cast_fp16")];
            tensor<int32, [1]> var_733_axes_0 = const()[name = string("op_733_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_733_keep_dims_0 = const()[name = string("op_733_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_733_cast_fp16 = reduce_mean(axes = var_733_axes_0, keep_dims = var_733_keep_dims_0, x = var_731_cast_fp16)[name = string("op_733_cast_fp16")];
            fp16 var_734_to_fp16 = const()[name = string("op_734_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_735_cast_fp16 = add(x = var_733_cast_fp16, y = var_734_to_fp16)[name = string("op_735_cast_fp16")];
            fp32 norm_29_epsilon_0 = const()[name = string("norm_29_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_29_cast_fp16 = rsqrt(epsilon = norm_29_epsilon_0, x = var_735_cast_fp16)[name = string("norm_29_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_737_cast_fp16 = mul(x = x_87_cast_fp16, y = norm_29_cast_fp16)[name = string("op_737_cast_fp16")];
            tensor<fp16, [128]> layers_3_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54624320)))];
            tensor<fp16, [1, 8, ?, 128]> var_738_cast_fp16 = mul(x = var_737_cast_fp16, y = layers_3_self_attn_k_norm_weight_to_fp16)[name = string("op_738_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = var_727_cast_fp16)[name = string("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = var_727_cast_fp16)[name = string("x2_13_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_754_cast_fp16 = mul(x = x1_13_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_754_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_755_cast_fp16 = mul(x = x2_13_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_755_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_756_cast_fp16 = sub(x = var_754_cast_fp16, y = var_755_cast_fp16)[name = string("op_756_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_757_cast_fp16 = mul(x = x2_13_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_757_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_758_cast_fp16 = mul(x = x1_13_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_758_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_759_cast_fp16 = add(x = var_757_cast_fp16, y = var_758_cast_fp16)[name = string("op_759_cast_fp16")];
            bool q_7_interleave_0 = const()[name = string("q_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_7_cast_fp16 = concat(axis = var_673, interleave = q_7_interleave_0, values = (var_756_cast_fp16, var_759_cast_fp16))[name = string("q_7_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = var_738_cast_fp16)[name = string("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = var_738_cast_fp16)[name = string("x2_15_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_776_cast_fp16 = mul(x = x1_15_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_776_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_777_cast_fp16 = mul(x = x2_15_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_777_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_778_cast_fp16 = sub(x = var_776_cast_fp16, y = var_777_cast_fp16)[name = string("op_778_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_779_cast_fp16 = mul(x = x2_15_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_779_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_780_cast_fp16 = mul(x = x1_15_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_780_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_781_cast_fp16 = add(x = var_779_cast_fp16, y = var_780_cast_fp16)[name = string("op_781_cast_fp16")];
            bool k_7_interleave_0 = const()[name = string("k_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_7_cast_fp16 = concat(axis = var_673, interleave = k_7_interleave_0, values = (var_778_cast_fp16, var_781_cast_fp16))[name = string("k_7_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_6 = read_state(input = k_cache_3)[name = string("read_state_6")];
            int32 k_cache_15_axis_0 = const()[name = string("k_cache_15_axis_0"), val = int32(2)];
            string k_cache_15_mode_0 = const()[name = string("k_cache_15_mode_0"), val = string("update")];
            bool k_cache_15_validate_indices_0 = const()[name = string("k_cache_15_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_15_cast_fp16 = scatter_along_axis(axis = k_cache_15_axis_0, data = read_state_6, indices = scatter_indices, mode = k_cache_15_mode_0, updates = k_7_cast_fp16, validate_indices = k_cache_15_validate_indices_0)[name = string("k_cache_15_cast_fp16")];
            write_state(data = k_cache_15_cast_fp16, input = k_cache_3)[name = string("coreml_update_state_6_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_62 = read_state(input = k_cache_3)[name = string("coreml_update_state_6")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_7 = read_state(input = v_cache_3)[name = string("read_state_7")];
            int32 v_cache_15_axis_0 = const()[name = string("v_cache_15_axis_0"), val = int32(2)];
            string v_cache_15_mode_0 = const()[name = string("v_cache_15_mode_0"), val = string("update")];
            bool v_cache_15_validate_indices_0 = const()[name = string("v_cache_15_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_716_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_15_cast_fp16 = scatter_along_axis(axis = v_cache_15_axis_0, data = read_state_7, indices = scatter_indices, mode = v_cache_15_mode_0, updates = v_7_cast_fp16, validate_indices = v_cache_15_validate_indices_0)[name = string("v_cache_15_cast_fp16")];
            write_state(data = v_cache_15_cast_fp16, input = v_cache_3)[name = string("coreml_update_state_7_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_63 = read_state(input = v_cache_3)[name = string("coreml_update_state_7")];
            tensor<int32, [1]> var_786_axes_0 = const()[name = string("op_786_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_786_cast_fp16 = expand_dims(axes = var_786_axes_0, x = coreml_update_state_62)[name = string("op_786_cast_fp16")];
            tensor<int32, [5]> k_exp_13_reps_0 = const()[name = string("k_exp_13_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_13_cast_fp16 = tile(reps = k_exp_13_reps_0, x = var_786_cast_fp16)[name = string("k_exp_13_cast_fp16")];
            tensor<int32, [4]> var_789 = const()[name = string("op_789"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_15_cast_fp16 = reshape(shape = var_789, x = k_exp_13_cast_fp16)[name = string("k_exp_15_cast_fp16")];
            tensor<int32, [1]> var_791_axes_0 = const()[name = string("op_791_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_791_cast_fp16 = expand_dims(axes = var_791_axes_0, x = coreml_update_state_63)[name = string("op_791_cast_fp16")];
            tensor<int32, [5]> v_exp_13_reps_0 = const()[name = string("v_exp_13_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_13_cast_fp16 = tile(reps = v_exp_13_reps_0, x = var_791_cast_fp16)[name = string("v_exp_13_cast_fp16")];
            tensor<int32, [4]> var_794 = const()[name = string("op_794"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_15_cast_fp16 = reshape(shape = var_794, x = v_exp_13_cast_fp16)[name = string("v_exp_15_cast_fp16")];
            bool var_797_transpose_x_1 = const()[name = string("op_797_transpose_x_1"), val = bool(false)];
            bool var_797_transpose_y_1 = const()[name = string("op_797_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_797_cast_fp16 = matmul(transpose_x = var_797_transpose_x_1, transpose_y = var_797_transpose_y_1, x = q_7_cast_fp16, y = k_exp_15_cast_fp16)[name = string("op_797_cast_fp16")];
            fp16 var_798_to_fp16 = const()[name = string("op_798_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_13_cast_fp16 = mul(x = var_797_cast_fp16, y = var_798_to_fp16)[name = string("attn_13_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_31_cast_fp16 = add(x = attn_13_cast_fp16, y = attention_mask)[name = string("input_31_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_15_cast_fp16 = softmax(axis = var_673, x = input_31_cast_fp16)[name = string("attn_15_cast_fp16")];
            bool out_7_transpose_x_0 = const()[name = string("out_7_transpose_x_0"), val = bool(false)];
            bool out_7_transpose_y_0 = const()[name = string("out_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_7_cast_fp16 = matmul(transpose_x = out_7_transpose_x_0, transpose_y = out_7_transpose_y_0, x = attn_15_cast_fp16, y = v_exp_15_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [4]> var_803_perm_0 = const()[name = string("op_803_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_23_axis_0 = const()[name = string("concat_23_axis_0"), val = int32(0)];
            bool concat_23_interleave_0 = const()[name = string("concat_23_interleave_0"), val = bool(false)];
            int32 gather_19_cast_uint16_to_int32 = cast(dtype = gather_19_cast_uint16_to_int32_dtype_0, x = gather_19_cast_uint16)[name = string("cast_48")];
            tensor<int32, [3]> concat_23 = concat(axis = concat_23_axis_0, interleave = concat_23_interleave_0, values = (gather_18, gather_19_cast_uint16_to_int32, var_673))[name = string("concat_23")];
            tensor<fp16, [1, ?, 16, 128]> var_803_cast_fp16 = transpose(perm = var_803_perm_0, x = out_7_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, ?, ?]> input_33_cast_fp16 = reshape(shape = concat_23, x = var_803_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54624640))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56721856))))[name = string("layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_24_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_97_cast_fp16 = add(x = x_77_cast_fp16, y = linear_24_cast_fp16)[name = string("x_97_cast_fp16")];
            fp16 var_672_promoted_3_to_fp16 = const()[name = string("op_672_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_811_cast_fp16 = pow(x = x_97_cast_fp16, y = var_672_promoted_3_to_fp16)[name = string("op_811_cast_fp16")];
            tensor<int32, [1]> var_813_axes_0 = const()[name = string("op_813_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_813_keep_dims_0 = const()[name = string("op_813_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_813_cast_fp16 = reduce_mean(axes = var_813_axes_0, keep_dims = var_813_keep_dims_0, x = var_811_cast_fp16)[name = string("op_813_cast_fp16")];
            fp16 var_814_to_fp16 = const()[name = string("op_814_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_815_cast_fp16 = add(x = var_813_cast_fp16, y = var_814_to_fp16)[name = string("op_815_cast_fp16")];
            fp32 norm_31_epsilon_0 = const()[name = string("norm_31_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_31_cast_fp16 = rsqrt(epsilon = norm_31_epsilon_0, x = var_815_cast_fp16)[name = string("norm_31_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_817_cast_fp16 = mul(x = x_97_cast_fp16, y = norm_31_cast_fp16)[name = string("op_817_cast_fp16")];
            tensor<fp16, [1024]> layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56852992)))];
            tensor<fp16, [1, ?, 1024]> var_818_cast_fp16 = mul(x = var_817_cast_fp16, y = layers_3_post_attention_layernorm_weight_to_fp16)[name = string("op_818_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56855104))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60000896))))[name = string("layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = var_818_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_828_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_828_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60197568))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63343360))))[name = string("layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_3_mlp_up_proj_weight_to_fp16_quantized, x = var_818_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_39_cast_fp16 = mul(x = var_828_cast_fp16, y = linear_26_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63540032))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66685824))))[name = string("layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_103_cast_fp16 = add(x = x_97_cast_fp16, y = linear_27_cast_fp16)[name = string("x_103_cast_fp16")];
            int32 var_848 = const()[name = string("op_848"), val = int32(-1)];
            fp16 var_847_promoted_to_fp16 = const()[name = string("op_847_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_857_cast_fp16 = pow(x = x_103_cast_fp16, y = var_847_promoted_to_fp16)[name = string("op_857_cast_fp16")];
            tensor<int32, [1]> var_859_axes_0 = const()[name = string("op_859_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_859_keep_dims_0 = const()[name = string("op_859_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_859_cast_fp16 = reduce_mean(axes = var_859_axes_0, keep_dims = var_859_keep_dims_0, x = var_857_cast_fp16)[name = string("op_859_cast_fp16")];
            fp16 var_860_to_fp16 = const()[name = string("op_860_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_861_cast_fp16 = add(x = var_859_cast_fp16, y = var_860_to_fp16)[name = string("op_861_cast_fp16")];
            fp32 norm_33_epsilon_0 = const()[name = string("norm_33_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_33_cast_fp16 = rsqrt(epsilon = norm_33_epsilon_0, x = var_861_cast_fp16)[name = string("norm_33_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_863_cast_fp16 = mul(x = x_103_cast_fp16, y = norm_33_cast_fp16)[name = string("op_863_cast_fp16")];
            tensor<fp16, [1024]> layers_4_input_layernorm_weight_to_fp16 = const()[name = string("layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66882496)))];
            tensor<fp16, [1, ?, 1024]> var_864_cast_fp16 = mul(x = var_863_cast_fp16, y = layers_4_input_layernorm_weight_to_fp16)[name = string("op_864_cast_fp16")];
            tensor<int32, [3]> var_876_shape_cast_fp16 = shape(x = var_864_cast_fp16)[name = string("op_876_shape_cast_fp16")];
            int32 gather_24 = const()[name = string("gather_24"), val = int32(1)];
            int32 gather_25_axis_0 = const()[name = string("gather_25_axis_0"), val = int32(0)];
            int32 gather_25_batch_dims_0 = const()[name = string("gather_25_batch_dims_0"), val = int32(0)];
            bool gather_25_validate_indices_0 = const()[name = string("gather_25_validate_indices_0"), val = bool(false)];
            string var_876_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_876_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_25_to_uint16 = const()[name = string("select_25_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_876_shape_cast_fp16_to_uint16 = cast(dtype = var_876_shape_cast_fp16_to_uint16_dtype_0, x = var_876_shape_cast_fp16)[name = string("cast_47")];
            uint16 gather_25_cast_uint16 = gather(axis = gather_25_axis_0, batch_dims = gather_25_batch_dims_0, indices = select_25_to_uint16, validate_indices = gather_25_validate_indices_0, x = var_876_shape_cast_fp16_to_uint16)[name = string("gather_25_cast_uint16")];
            string gather_25_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_25_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66884608))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68981824))))[name = string("layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = var_864_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_881_cast_fp16 = reshape(shape = concat_24x, x = linear_28_cast_fp16)[name = string("op_881_cast_fp16")];
            tensor<int32, [4]> x_109_perm_0 = const()[name = string("x_109_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69112960))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70161600))))[name = string("layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = var_864_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_886_cast_fp16 = reshape(shape = concat_25x, x = linear_29_cast_fp16)[name = string("op_886_cast_fp16")];
            tensor<int32, [4]> x_113_perm_0 = const()[name = string("x_113_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70227200))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71275840))))[name = string("layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_30_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = var_864_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_891_cast_fp16 = reshape(shape = concat_26x, x = linear_30_cast_fp16)[name = string("op_891_cast_fp16")];
            tensor<int32, [4]> v_9_perm_0 = const()[name = string("v_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_847_promoted_1_to_fp16 = const()[name = string("op_847_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_109_cast_fp16 = transpose(perm = x_109_perm_0, x = var_881_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 16, ?, 128]> var_895_cast_fp16 = pow(x = x_109_cast_fp16, y = var_847_promoted_1_to_fp16)[name = string("op_895_cast_fp16")];
            tensor<int32, [1]> var_897_axes_0 = const()[name = string("op_897_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_897_keep_dims_0 = const()[name = string("op_897_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_897_cast_fp16 = reduce_mean(axes = var_897_axes_0, keep_dims = var_897_keep_dims_0, x = var_895_cast_fp16)[name = string("op_897_cast_fp16")];
            fp16 var_898_to_fp16 = const()[name = string("op_898_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_899_cast_fp16 = add(x = var_897_cast_fp16, y = var_898_to_fp16)[name = string("op_899_cast_fp16")];
            fp32 norm_35_epsilon_0 = const()[name = string("norm_35_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_35_cast_fp16 = rsqrt(epsilon = norm_35_epsilon_0, x = var_899_cast_fp16)[name = string("norm_35_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_901_cast_fp16 = mul(x = x_109_cast_fp16, y = norm_35_cast_fp16)[name = string("op_901_cast_fp16")];
            tensor<fp16, [128]> layers_4_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71341440)))];
            tensor<fp16, [1, 16, ?, 128]> var_902_cast_fp16 = mul(x = var_901_cast_fp16, y = layers_4_self_attn_q_norm_weight_to_fp16)[name = string("op_902_cast_fp16")];
            fp16 var_847_promoted_2_to_fp16 = const()[name = string("op_847_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_113_cast_fp16 = transpose(perm = x_113_perm_0, x = var_886_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 8, ?, 128]> var_906_cast_fp16 = pow(x = x_113_cast_fp16, y = var_847_promoted_2_to_fp16)[name = string("op_906_cast_fp16")];
            tensor<int32, [1]> var_908_axes_0 = const()[name = string("op_908_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_908_keep_dims_0 = const()[name = string("op_908_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_908_cast_fp16 = reduce_mean(axes = var_908_axes_0, keep_dims = var_908_keep_dims_0, x = var_906_cast_fp16)[name = string("op_908_cast_fp16")];
            fp16 var_909_to_fp16 = const()[name = string("op_909_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_910_cast_fp16 = add(x = var_908_cast_fp16, y = var_909_to_fp16)[name = string("op_910_cast_fp16")];
            fp32 norm_37_epsilon_0 = const()[name = string("norm_37_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_37_cast_fp16 = rsqrt(epsilon = norm_37_epsilon_0, x = var_910_cast_fp16)[name = string("norm_37_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_912_cast_fp16 = mul(x = x_113_cast_fp16, y = norm_37_cast_fp16)[name = string("op_912_cast_fp16")];
            tensor<fp16, [128]> layers_4_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71341760)))];
            tensor<fp16, [1, 8, ?, 128]> var_913_cast_fp16 = mul(x = var_912_cast_fp16, y = layers_4_self_attn_k_norm_weight_to_fp16)[name = string("op_913_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = var_902_cast_fp16)[name = string("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = var_902_cast_fp16)[name = string("x2_17_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_929_cast_fp16 = mul(x = x1_17_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_929_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_930_cast_fp16 = mul(x = x2_17_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_930_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_931_cast_fp16 = sub(x = var_929_cast_fp16, y = var_930_cast_fp16)[name = string("op_931_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_932_cast_fp16 = mul(x = x2_17_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_932_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_933_cast_fp16 = mul(x = x1_17_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_933_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_934_cast_fp16 = add(x = var_932_cast_fp16, y = var_933_cast_fp16)[name = string("op_934_cast_fp16")];
            bool q_9_interleave_0 = const()[name = string("q_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_9_cast_fp16 = concat(axis = var_848, interleave = q_9_interleave_0, values = (var_931_cast_fp16, var_934_cast_fp16))[name = string("q_9_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = var_913_cast_fp16)[name = string("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = var_913_cast_fp16)[name = string("x2_19_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_951_cast_fp16 = mul(x = x1_19_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_951_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_952_cast_fp16 = mul(x = x2_19_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_952_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_953_cast_fp16 = sub(x = var_951_cast_fp16, y = var_952_cast_fp16)[name = string("op_953_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_954_cast_fp16 = mul(x = x2_19_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_954_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_955_cast_fp16 = mul(x = x1_19_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_955_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_956_cast_fp16 = add(x = var_954_cast_fp16, y = var_955_cast_fp16)[name = string("op_956_cast_fp16")];
            bool k_9_interleave_0 = const()[name = string("k_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_9_cast_fp16 = concat(axis = var_848, interleave = k_9_interleave_0, values = (var_953_cast_fp16, var_956_cast_fp16))[name = string("k_9_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_8 = read_state(input = k_cache_4)[name = string("read_state_8")];
            int32 k_cache_19_axis_0 = const()[name = string("k_cache_19_axis_0"), val = int32(2)];
            string k_cache_19_mode_0 = const()[name = string("k_cache_19_mode_0"), val = string("update")];
            bool k_cache_19_validate_indices_0 = const()[name = string("k_cache_19_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_19_cast_fp16 = scatter_along_axis(axis = k_cache_19_axis_0, data = read_state_8, indices = scatter_indices, mode = k_cache_19_mode_0, updates = k_9_cast_fp16, validate_indices = k_cache_19_validate_indices_0)[name = string("k_cache_19_cast_fp16")];
            write_state(data = k_cache_19_cast_fp16, input = k_cache_4)[name = string("coreml_update_state_8_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_64 = read_state(input = k_cache_4)[name = string("coreml_update_state_8")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_9 = read_state(input = v_cache_4)[name = string("read_state_9")];
            int32 v_cache_19_axis_0 = const()[name = string("v_cache_19_axis_0"), val = int32(2)];
            string v_cache_19_mode_0 = const()[name = string("v_cache_19_mode_0"), val = string("update")];
            bool v_cache_19_validate_indices_0 = const()[name = string("v_cache_19_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = var_891_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_19_cast_fp16 = scatter_along_axis(axis = v_cache_19_axis_0, data = read_state_9, indices = scatter_indices, mode = v_cache_19_mode_0, updates = v_9_cast_fp16, validate_indices = v_cache_19_validate_indices_0)[name = string("v_cache_19_cast_fp16")];
            write_state(data = v_cache_19_cast_fp16, input = v_cache_4)[name = string("coreml_update_state_9_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_65 = read_state(input = v_cache_4)[name = string("coreml_update_state_9")];
            tensor<int32, [1]> var_961_axes_0 = const()[name = string("op_961_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_961_cast_fp16 = expand_dims(axes = var_961_axes_0, x = coreml_update_state_64)[name = string("op_961_cast_fp16")];
            tensor<int32, [5]> k_exp_17_reps_0 = const()[name = string("k_exp_17_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_17_cast_fp16 = tile(reps = k_exp_17_reps_0, x = var_961_cast_fp16)[name = string("k_exp_17_cast_fp16")];
            tensor<int32, [4]> var_964 = const()[name = string("op_964"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_19_cast_fp16 = reshape(shape = var_964, x = k_exp_17_cast_fp16)[name = string("k_exp_19_cast_fp16")];
            tensor<int32, [1]> var_966_axes_0 = const()[name = string("op_966_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_966_cast_fp16 = expand_dims(axes = var_966_axes_0, x = coreml_update_state_65)[name = string("op_966_cast_fp16")];
            tensor<int32, [5]> v_exp_17_reps_0 = const()[name = string("v_exp_17_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_17_cast_fp16 = tile(reps = v_exp_17_reps_0, x = var_966_cast_fp16)[name = string("v_exp_17_cast_fp16")];
            tensor<int32, [4]> var_969 = const()[name = string("op_969"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_19_cast_fp16 = reshape(shape = var_969, x = v_exp_17_cast_fp16)[name = string("v_exp_19_cast_fp16")];
            bool var_972_transpose_x_1 = const()[name = string("op_972_transpose_x_1"), val = bool(false)];
            bool var_972_transpose_y_1 = const()[name = string("op_972_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_972_cast_fp16 = matmul(transpose_x = var_972_transpose_x_1, transpose_y = var_972_transpose_y_1, x = q_9_cast_fp16, y = k_exp_19_cast_fp16)[name = string("op_972_cast_fp16")];
            fp16 var_973_to_fp16 = const()[name = string("op_973_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_17_cast_fp16 = mul(x = var_972_cast_fp16, y = var_973_to_fp16)[name = string("attn_17_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_41_cast_fp16 = add(x = attn_17_cast_fp16, y = attention_mask)[name = string("input_41_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_19_cast_fp16 = softmax(axis = var_848, x = input_41_cast_fp16)[name = string("attn_19_cast_fp16")];
            bool out_9_transpose_x_0 = const()[name = string("out_9_transpose_x_0"), val = bool(false)];
            bool out_9_transpose_y_0 = const()[name = string("out_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_9_cast_fp16 = matmul(transpose_x = out_9_transpose_x_0, transpose_y = out_9_transpose_y_0, x = attn_19_cast_fp16, y = v_exp_19_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [4]> var_978_perm_0 = const()[name = string("op_978_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_29_axis_0 = const()[name = string("concat_29_axis_0"), val = int32(0)];
            bool concat_29_interleave_0 = const()[name = string("concat_29_interleave_0"), val = bool(false)];
            int32 gather_25_cast_uint16_to_int32 = cast(dtype = gather_25_cast_uint16_to_int32_dtype_0, x = gather_25_cast_uint16)[name = string("cast_46")];
            tensor<int32, [3]> concat_29 = concat(axis = concat_29_axis_0, interleave = concat_29_interleave_0, values = (gather_24, gather_25_cast_uint16_to_int32, var_848))[name = string("concat_29")];
            tensor<fp16, [1, ?, 16, 128]> var_978_cast_fp16 = transpose(perm = var_978_perm_0, x = out_9_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, ?, ?]> input_43_cast_fp16 = reshape(shape = concat_29, x = var_978_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71342080))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73439296))))[name = string("layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_123_cast_fp16 = add(x = x_103_cast_fp16, y = linear_31_cast_fp16)[name = string("x_123_cast_fp16")];
            fp16 var_847_promoted_3_to_fp16 = const()[name = string("op_847_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_986_cast_fp16 = pow(x = x_123_cast_fp16, y = var_847_promoted_3_to_fp16)[name = string("op_986_cast_fp16")];
            tensor<int32, [1]> var_988_axes_0 = const()[name = string("op_988_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_988_keep_dims_0 = const()[name = string("op_988_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_988_cast_fp16 = reduce_mean(axes = var_988_axes_0, keep_dims = var_988_keep_dims_0, x = var_986_cast_fp16)[name = string("op_988_cast_fp16")];
            fp16 var_989_to_fp16 = const()[name = string("op_989_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_990_cast_fp16 = add(x = var_988_cast_fp16, y = var_989_to_fp16)[name = string("op_990_cast_fp16")];
            fp32 norm_39_epsilon_0 = const()[name = string("norm_39_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_39_cast_fp16 = rsqrt(epsilon = norm_39_epsilon_0, x = var_990_cast_fp16)[name = string("norm_39_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_992_cast_fp16 = mul(x = x_123_cast_fp16, y = norm_39_cast_fp16)[name = string("op_992_cast_fp16")];
            tensor<fp16, [1024]> layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73570432)))];
            tensor<fp16, [1, ?, 1024]> var_993_cast_fp16 = mul(x = var_992_cast_fp16, y = layers_4_post_attention_layernorm_weight_to_fp16)[name = string("op_993_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73572544))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76718336))))[name = string("layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = var_993_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_1003_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_1003_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76915008))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80060800))))[name = string("layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_4_mlp_up_proj_weight_to_fp16_quantized, x = var_993_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_49_cast_fp16 = mul(x = var_1003_cast_fp16, y = linear_33_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80257472))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83403264))))[name = string("layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_34_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_129_cast_fp16 = add(x = x_123_cast_fp16, y = linear_34_cast_fp16)[name = string("x_129_cast_fp16")];
            int32 var_1023 = const()[name = string("op_1023"), val = int32(-1)];
            fp16 var_1022_promoted_to_fp16 = const()[name = string("op_1022_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_1032_cast_fp16 = pow(x = x_129_cast_fp16, y = var_1022_promoted_to_fp16)[name = string("op_1032_cast_fp16")];
            tensor<int32, [1]> var_1034_axes_0 = const()[name = string("op_1034_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1034_keep_dims_0 = const()[name = string("op_1034_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_1034_cast_fp16 = reduce_mean(axes = var_1034_axes_0, keep_dims = var_1034_keep_dims_0, x = var_1032_cast_fp16)[name = string("op_1034_cast_fp16")];
            fp16 var_1035_to_fp16 = const()[name = string("op_1035_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1036_cast_fp16 = add(x = var_1034_cast_fp16, y = var_1035_to_fp16)[name = string("op_1036_cast_fp16")];
            fp32 norm_41_epsilon_0 = const()[name = string("norm_41_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_41_cast_fp16 = rsqrt(epsilon = norm_41_epsilon_0, x = var_1036_cast_fp16)[name = string("norm_41_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_1038_cast_fp16 = mul(x = x_129_cast_fp16, y = norm_41_cast_fp16)[name = string("op_1038_cast_fp16")];
            tensor<fp16, [1024]> layers_5_input_layernorm_weight_to_fp16 = const()[name = string("layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83599936)))];
            tensor<fp16, [1, ?, 1024]> var_1039_cast_fp16 = mul(x = var_1038_cast_fp16, y = layers_5_input_layernorm_weight_to_fp16)[name = string("op_1039_cast_fp16")];
            tensor<int32, [3]> var_1051_shape_cast_fp16 = shape(x = var_1039_cast_fp16)[name = string("op_1051_shape_cast_fp16")];
            int32 gather_30 = const()[name = string("gather_30"), val = int32(1)];
            int32 gather_31_axis_0 = const()[name = string("gather_31_axis_0"), val = int32(0)];
            int32 gather_31_batch_dims_0 = const()[name = string("gather_31_batch_dims_0"), val = int32(0)];
            bool gather_31_validate_indices_0 = const()[name = string("gather_31_validate_indices_0"), val = bool(false)];
            string var_1051_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1051_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_31_to_uint16 = const()[name = string("select_31_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_1051_shape_cast_fp16_to_uint16 = cast(dtype = var_1051_shape_cast_fp16_to_uint16_dtype_0, x = var_1051_shape_cast_fp16)[name = string("cast_45")];
            uint16 gather_31_cast_uint16 = gather(axis = gather_31_axis_0, batch_dims = gather_31_batch_dims_0, indices = select_31_to_uint16, validate_indices = gather_31_validate_indices_0, x = var_1051_shape_cast_fp16_to_uint16)[name = string("gather_31_cast_uint16")];
            string gather_31_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_31_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83602048))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85699264))))[name = string("layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_35_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = var_1039_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> concat_30x = const()[name = string("concat_30x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_1056_cast_fp16 = reshape(shape = concat_30x, x = linear_35_cast_fp16)[name = string("op_1056_cast_fp16")];
            tensor<int32, [4]> x_135_perm_0 = const()[name = string("x_135_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85830400))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86879040))))[name = string("layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_36_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = var_1039_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> concat_31x = const()[name = string("concat_31x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_1061_cast_fp16 = reshape(shape = concat_31x, x = linear_36_cast_fp16)[name = string("op_1061_cast_fp16")];
            tensor<int32, [4]> x_139_perm_0 = const()[name = string("x_139_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86944640))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87993280))))[name = string("layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = var_1039_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> concat_32x = const()[name = string("concat_32x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_1066_cast_fp16 = reshape(shape = concat_32x, x = linear_37_cast_fp16)[name = string("op_1066_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1022_promoted_1_to_fp16 = const()[name = string("op_1022_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_135_cast_fp16 = transpose(perm = x_135_perm_0, x = var_1056_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 16, ?, 128]> var_1070_cast_fp16 = pow(x = x_135_cast_fp16, y = var_1022_promoted_1_to_fp16)[name = string("op_1070_cast_fp16")];
            tensor<int32, [1]> var_1072_axes_0 = const()[name = string("op_1072_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1072_keep_dims_0 = const()[name = string("op_1072_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_1072_cast_fp16 = reduce_mean(axes = var_1072_axes_0, keep_dims = var_1072_keep_dims_0, x = var_1070_cast_fp16)[name = string("op_1072_cast_fp16")];
            fp16 var_1073_to_fp16 = const()[name = string("op_1073_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_1074_cast_fp16 = add(x = var_1072_cast_fp16, y = var_1073_to_fp16)[name = string("op_1074_cast_fp16")];
            fp32 norm_43_epsilon_0 = const()[name = string("norm_43_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_43_cast_fp16 = rsqrt(epsilon = norm_43_epsilon_0, x = var_1074_cast_fp16)[name = string("norm_43_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_1076_cast_fp16 = mul(x = x_135_cast_fp16, y = norm_43_cast_fp16)[name = string("op_1076_cast_fp16")];
            tensor<fp16, [128]> layers_5_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88058880)))];
            tensor<fp16, [1, 16, ?, 128]> var_1077_cast_fp16 = mul(x = var_1076_cast_fp16, y = layers_5_self_attn_q_norm_weight_to_fp16)[name = string("op_1077_cast_fp16")];
            fp16 var_1022_promoted_2_to_fp16 = const()[name = string("op_1022_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_139_cast_fp16 = transpose(perm = x_139_perm_0, x = var_1061_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 8, ?, 128]> var_1081_cast_fp16 = pow(x = x_139_cast_fp16, y = var_1022_promoted_2_to_fp16)[name = string("op_1081_cast_fp16")];
            tensor<int32, [1]> var_1083_axes_0 = const()[name = string("op_1083_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1083_keep_dims_0 = const()[name = string("op_1083_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_1083_cast_fp16 = reduce_mean(axes = var_1083_axes_0, keep_dims = var_1083_keep_dims_0, x = var_1081_cast_fp16)[name = string("op_1083_cast_fp16")];
            fp16 var_1084_to_fp16 = const()[name = string("op_1084_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_1085_cast_fp16 = add(x = var_1083_cast_fp16, y = var_1084_to_fp16)[name = string("op_1085_cast_fp16")];
            fp32 norm_45_epsilon_0 = const()[name = string("norm_45_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_45_cast_fp16 = rsqrt(epsilon = norm_45_epsilon_0, x = var_1085_cast_fp16)[name = string("norm_45_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_1087_cast_fp16 = mul(x = x_139_cast_fp16, y = norm_45_cast_fp16)[name = string("op_1087_cast_fp16")];
            tensor<fp16, [128]> layers_5_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88059200)))];
            tensor<fp16, [1, 8, ?, 128]> var_1088_cast_fp16 = mul(x = var_1087_cast_fp16, y = layers_5_self_attn_k_norm_weight_to_fp16)[name = string("op_1088_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = var_1077_cast_fp16)[name = string("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = var_1077_cast_fp16)[name = string("x2_21_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1104_cast_fp16 = mul(x = x1_21_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1104_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1105_cast_fp16 = mul(x = x2_21_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1105_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1106_cast_fp16 = sub(x = var_1104_cast_fp16, y = var_1105_cast_fp16)[name = string("op_1106_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1107_cast_fp16 = mul(x = x2_21_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1107_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1108_cast_fp16 = mul(x = x1_21_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1108_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1109_cast_fp16 = add(x = var_1107_cast_fp16, y = var_1108_cast_fp16)[name = string("op_1109_cast_fp16")];
            bool q_11_interleave_0 = const()[name = string("q_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_11_cast_fp16 = concat(axis = var_1023, interleave = q_11_interleave_0, values = (var_1106_cast_fp16, var_1109_cast_fp16))[name = string("q_11_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = var_1088_cast_fp16)[name = string("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = var_1088_cast_fp16)[name = string("x2_23_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1126_cast_fp16 = mul(x = x1_23_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1126_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1127_cast_fp16 = mul(x = x2_23_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1127_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1128_cast_fp16 = sub(x = var_1126_cast_fp16, y = var_1127_cast_fp16)[name = string("op_1128_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1129_cast_fp16 = mul(x = x2_23_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1129_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1130_cast_fp16 = mul(x = x1_23_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1130_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1131_cast_fp16 = add(x = var_1129_cast_fp16, y = var_1130_cast_fp16)[name = string("op_1131_cast_fp16")];
            bool k_11_interleave_0 = const()[name = string("k_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_11_cast_fp16 = concat(axis = var_1023, interleave = k_11_interleave_0, values = (var_1128_cast_fp16, var_1131_cast_fp16))[name = string("k_11_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_10 = read_state(input = k_cache_5)[name = string("read_state_10")];
            int32 k_cache_23_axis_0 = const()[name = string("k_cache_23_axis_0"), val = int32(2)];
            string k_cache_23_mode_0 = const()[name = string("k_cache_23_mode_0"), val = string("update")];
            bool k_cache_23_validate_indices_0 = const()[name = string("k_cache_23_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_23_cast_fp16 = scatter_along_axis(axis = k_cache_23_axis_0, data = read_state_10, indices = scatter_indices, mode = k_cache_23_mode_0, updates = k_11_cast_fp16, validate_indices = k_cache_23_validate_indices_0)[name = string("k_cache_23_cast_fp16")];
            write_state(data = k_cache_23_cast_fp16, input = k_cache_5)[name = string("coreml_update_state_10_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_66 = read_state(input = k_cache_5)[name = string("coreml_update_state_10")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_11 = read_state(input = v_cache_5)[name = string("read_state_11")];
            int32 v_cache_23_axis_0 = const()[name = string("v_cache_23_axis_0"), val = int32(2)];
            string v_cache_23_mode_0 = const()[name = string("v_cache_23_mode_0"), val = string("update")];
            bool v_cache_23_validate_indices_0 = const()[name = string("v_cache_23_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = var_1066_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_23_cast_fp16 = scatter_along_axis(axis = v_cache_23_axis_0, data = read_state_11, indices = scatter_indices, mode = v_cache_23_mode_0, updates = v_11_cast_fp16, validate_indices = v_cache_23_validate_indices_0)[name = string("v_cache_23_cast_fp16")];
            write_state(data = v_cache_23_cast_fp16, input = v_cache_5)[name = string("coreml_update_state_11_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_67 = read_state(input = v_cache_5)[name = string("coreml_update_state_11")];
            tensor<int32, [1]> var_1136_axes_0 = const()[name = string("op_1136_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1136_cast_fp16 = expand_dims(axes = var_1136_axes_0, x = coreml_update_state_66)[name = string("op_1136_cast_fp16")];
            tensor<int32, [5]> k_exp_21_reps_0 = const()[name = string("k_exp_21_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_21_cast_fp16 = tile(reps = k_exp_21_reps_0, x = var_1136_cast_fp16)[name = string("k_exp_21_cast_fp16")];
            tensor<int32, [4]> var_1139 = const()[name = string("op_1139"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_23_cast_fp16 = reshape(shape = var_1139, x = k_exp_21_cast_fp16)[name = string("k_exp_23_cast_fp16")];
            tensor<int32, [1]> var_1141_axes_0 = const()[name = string("op_1141_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1141_cast_fp16 = expand_dims(axes = var_1141_axes_0, x = coreml_update_state_67)[name = string("op_1141_cast_fp16")];
            tensor<int32, [5]> v_exp_21_reps_0 = const()[name = string("v_exp_21_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_21_cast_fp16 = tile(reps = v_exp_21_reps_0, x = var_1141_cast_fp16)[name = string("v_exp_21_cast_fp16")];
            tensor<int32, [4]> var_1144 = const()[name = string("op_1144"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_23_cast_fp16 = reshape(shape = var_1144, x = v_exp_21_cast_fp16)[name = string("v_exp_23_cast_fp16")];
            bool var_1147_transpose_x_1 = const()[name = string("op_1147_transpose_x_1"), val = bool(false)];
            bool var_1147_transpose_y_1 = const()[name = string("op_1147_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_1147_cast_fp16 = matmul(transpose_x = var_1147_transpose_x_1, transpose_y = var_1147_transpose_y_1, x = q_11_cast_fp16, y = k_exp_23_cast_fp16)[name = string("op_1147_cast_fp16")];
            fp16 var_1148_to_fp16 = const()[name = string("op_1148_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_21_cast_fp16 = mul(x = var_1147_cast_fp16, y = var_1148_to_fp16)[name = string("attn_21_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_51_cast_fp16 = add(x = attn_21_cast_fp16, y = attention_mask)[name = string("input_51_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_23_cast_fp16 = softmax(axis = var_1023, x = input_51_cast_fp16)[name = string("attn_23_cast_fp16")];
            bool out_11_transpose_x_0 = const()[name = string("out_11_transpose_x_0"), val = bool(false)];
            bool out_11_transpose_y_0 = const()[name = string("out_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_11_cast_fp16 = matmul(transpose_x = out_11_transpose_x_0, transpose_y = out_11_transpose_y_0, x = attn_23_cast_fp16, y = v_exp_23_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [4]> var_1153_perm_0 = const()[name = string("op_1153_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_35_axis_0 = const()[name = string("concat_35_axis_0"), val = int32(0)];
            bool concat_35_interleave_0 = const()[name = string("concat_35_interleave_0"), val = bool(false)];
            int32 gather_31_cast_uint16_to_int32 = cast(dtype = gather_31_cast_uint16_to_int32_dtype_0, x = gather_31_cast_uint16)[name = string("cast_44")];
            tensor<int32, [3]> concat_35 = concat(axis = concat_35_axis_0, interleave = concat_35_interleave_0, values = (gather_30, gather_31_cast_uint16_to_int32, var_1023))[name = string("concat_35")];
            tensor<fp16, [1, ?, 16, 128]> var_1153_cast_fp16 = transpose(perm = var_1153_perm_0, x = out_11_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, ?, ?]> input_53_cast_fp16 = reshape(shape = concat_35, x = var_1153_cast_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88059520))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90156736))))[name = string("layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_38_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_149_cast_fp16 = add(x = x_129_cast_fp16, y = linear_38_cast_fp16)[name = string("x_149_cast_fp16")];
            fp16 var_1022_promoted_3_to_fp16 = const()[name = string("op_1022_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_1161_cast_fp16 = pow(x = x_149_cast_fp16, y = var_1022_promoted_3_to_fp16)[name = string("op_1161_cast_fp16")];
            tensor<int32, [1]> var_1163_axes_0 = const()[name = string("op_1163_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1163_keep_dims_0 = const()[name = string("op_1163_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_1163_cast_fp16 = reduce_mean(axes = var_1163_axes_0, keep_dims = var_1163_keep_dims_0, x = var_1161_cast_fp16)[name = string("op_1163_cast_fp16")];
            fp16 var_1164_to_fp16 = const()[name = string("op_1164_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1165_cast_fp16 = add(x = var_1163_cast_fp16, y = var_1164_to_fp16)[name = string("op_1165_cast_fp16")];
            fp32 norm_47_epsilon_0 = const()[name = string("norm_47_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_47_cast_fp16 = rsqrt(epsilon = norm_47_epsilon_0, x = var_1165_cast_fp16)[name = string("norm_47_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_1167_cast_fp16 = mul(x = x_149_cast_fp16, y = norm_47_cast_fp16)[name = string("op_1167_cast_fp16")];
            tensor<fp16, [1024]> layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90287872)))];
            tensor<fp16, [1, ?, 1024]> var_1168_cast_fp16 = mul(x = var_1167_cast_fp16, y = layers_5_post_attention_layernorm_weight_to_fp16)[name = string("op_1168_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90289984))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93435776))))[name = string("layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = var_1168_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_1178_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_1178_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93632448))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96778240))))[name = string("layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_5_mlp_up_proj_weight_to_fp16_quantized, x = var_1168_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_59_cast_fp16 = mul(x = var_1178_cast_fp16, y = linear_40_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96974912))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100120704))))[name = string("layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_155_cast_fp16 = add(x = x_149_cast_fp16, y = linear_41_cast_fp16)[name = string("x_155_cast_fp16")];
            int32 var_1198 = const()[name = string("op_1198"), val = int32(-1)];
            fp16 var_1197_promoted_to_fp16 = const()[name = string("op_1197_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_1207_cast_fp16 = pow(x = x_155_cast_fp16, y = var_1197_promoted_to_fp16)[name = string("op_1207_cast_fp16")];
            tensor<int32, [1]> var_1209_axes_0 = const()[name = string("op_1209_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1209_keep_dims_0 = const()[name = string("op_1209_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_1209_cast_fp16 = reduce_mean(axes = var_1209_axes_0, keep_dims = var_1209_keep_dims_0, x = var_1207_cast_fp16)[name = string("op_1209_cast_fp16")];
            fp16 var_1210_to_fp16 = const()[name = string("op_1210_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1211_cast_fp16 = add(x = var_1209_cast_fp16, y = var_1210_to_fp16)[name = string("op_1211_cast_fp16")];
            fp32 norm_49_epsilon_0 = const()[name = string("norm_49_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_49_cast_fp16 = rsqrt(epsilon = norm_49_epsilon_0, x = var_1211_cast_fp16)[name = string("norm_49_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_1213_cast_fp16 = mul(x = x_155_cast_fp16, y = norm_49_cast_fp16)[name = string("op_1213_cast_fp16")];
            tensor<fp16, [1024]> layers_6_input_layernorm_weight_to_fp16 = const()[name = string("layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100317376)))];
            tensor<fp16, [1, ?, 1024]> var_1214_cast_fp16 = mul(x = var_1213_cast_fp16, y = layers_6_input_layernorm_weight_to_fp16)[name = string("op_1214_cast_fp16")];
            tensor<int32, [3]> var_1226_shape_cast_fp16 = shape(x = var_1214_cast_fp16)[name = string("op_1226_shape_cast_fp16")];
            int32 gather_36 = const()[name = string("gather_36"), val = int32(1)];
            int32 gather_37_axis_0 = const()[name = string("gather_37_axis_0"), val = int32(0)];
            int32 gather_37_batch_dims_0 = const()[name = string("gather_37_batch_dims_0"), val = int32(0)];
            bool gather_37_validate_indices_0 = const()[name = string("gather_37_validate_indices_0"), val = bool(false)];
            string var_1226_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1226_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_37_to_uint16 = const()[name = string("select_37_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_1226_shape_cast_fp16_to_uint16 = cast(dtype = var_1226_shape_cast_fp16_to_uint16_dtype_0, x = var_1226_shape_cast_fp16)[name = string("cast_43")];
            uint16 gather_37_cast_uint16 = gather(axis = gather_37_axis_0, batch_dims = gather_37_batch_dims_0, indices = select_37_to_uint16, validate_indices = gather_37_validate_indices_0, x = var_1226_shape_cast_fp16_to_uint16)[name = string("gather_37_cast_uint16")];
            string gather_37_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_37_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100319488))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102416704))))[name = string("layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = var_1214_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_1231_cast_fp16 = reshape(shape = concat_36x, x = linear_42_cast_fp16)[name = string("op_1231_cast_fp16")];
            tensor<int32, [4]> x_161_perm_0 = const()[name = string("x_161_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102547840))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103596480))))[name = string("layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = var_1214_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_1236_cast_fp16 = reshape(shape = concat_37x, x = linear_43_cast_fp16)[name = string("op_1236_cast_fp16")];
            tensor<int32, [4]> x_165_perm_0 = const()[name = string("x_165_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103662080))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104710720))))[name = string("layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = var_1214_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_1241_cast_fp16 = reshape(shape = concat_38x, x = linear_44_cast_fp16)[name = string("op_1241_cast_fp16")];
            tensor<int32, [4]> v_13_perm_0 = const()[name = string("v_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1197_promoted_1_to_fp16 = const()[name = string("op_1197_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_161_cast_fp16 = transpose(perm = x_161_perm_0, x = var_1231_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 16, ?, 128]> var_1245_cast_fp16 = pow(x = x_161_cast_fp16, y = var_1197_promoted_1_to_fp16)[name = string("op_1245_cast_fp16")];
            tensor<int32, [1]> var_1247_axes_0 = const()[name = string("op_1247_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1247_keep_dims_0 = const()[name = string("op_1247_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_1247_cast_fp16 = reduce_mean(axes = var_1247_axes_0, keep_dims = var_1247_keep_dims_0, x = var_1245_cast_fp16)[name = string("op_1247_cast_fp16")];
            fp16 var_1248_to_fp16 = const()[name = string("op_1248_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_1249_cast_fp16 = add(x = var_1247_cast_fp16, y = var_1248_to_fp16)[name = string("op_1249_cast_fp16")];
            fp32 norm_51_epsilon_0 = const()[name = string("norm_51_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_51_cast_fp16 = rsqrt(epsilon = norm_51_epsilon_0, x = var_1249_cast_fp16)[name = string("norm_51_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_1251_cast_fp16 = mul(x = x_161_cast_fp16, y = norm_51_cast_fp16)[name = string("op_1251_cast_fp16")];
            tensor<fp16, [128]> layers_6_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104776320)))];
            tensor<fp16, [1, 16, ?, 128]> var_1252_cast_fp16 = mul(x = var_1251_cast_fp16, y = layers_6_self_attn_q_norm_weight_to_fp16)[name = string("op_1252_cast_fp16")];
            fp16 var_1197_promoted_2_to_fp16 = const()[name = string("op_1197_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_165_cast_fp16 = transpose(perm = x_165_perm_0, x = var_1236_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 8, ?, 128]> var_1256_cast_fp16 = pow(x = x_165_cast_fp16, y = var_1197_promoted_2_to_fp16)[name = string("op_1256_cast_fp16")];
            tensor<int32, [1]> var_1258_axes_0 = const()[name = string("op_1258_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1258_keep_dims_0 = const()[name = string("op_1258_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_1258_cast_fp16 = reduce_mean(axes = var_1258_axes_0, keep_dims = var_1258_keep_dims_0, x = var_1256_cast_fp16)[name = string("op_1258_cast_fp16")];
            fp16 var_1259_to_fp16 = const()[name = string("op_1259_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_1260_cast_fp16 = add(x = var_1258_cast_fp16, y = var_1259_to_fp16)[name = string("op_1260_cast_fp16")];
            fp32 norm_53_epsilon_0 = const()[name = string("norm_53_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_53_cast_fp16 = rsqrt(epsilon = norm_53_epsilon_0, x = var_1260_cast_fp16)[name = string("norm_53_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_1262_cast_fp16 = mul(x = x_165_cast_fp16, y = norm_53_cast_fp16)[name = string("op_1262_cast_fp16")];
            tensor<fp16, [128]> layers_6_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104776640)))];
            tensor<fp16, [1, 8, ?, 128]> var_1263_cast_fp16 = mul(x = var_1262_cast_fp16, y = layers_6_self_attn_k_norm_weight_to_fp16)[name = string("op_1263_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = var_1252_cast_fp16)[name = string("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = var_1252_cast_fp16)[name = string("x2_25_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1279_cast_fp16 = mul(x = x1_25_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1279_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1280_cast_fp16 = mul(x = x2_25_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1280_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1281_cast_fp16 = sub(x = var_1279_cast_fp16, y = var_1280_cast_fp16)[name = string("op_1281_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1282_cast_fp16 = mul(x = x2_25_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1282_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1283_cast_fp16 = mul(x = x1_25_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1283_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1284_cast_fp16 = add(x = var_1282_cast_fp16, y = var_1283_cast_fp16)[name = string("op_1284_cast_fp16")];
            bool q_13_interleave_0 = const()[name = string("q_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_13_cast_fp16 = concat(axis = var_1198, interleave = q_13_interleave_0, values = (var_1281_cast_fp16, var_1284_cast_fp16))[name = string("q_13_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = var_1263_cast_fp16)[name = string("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = var_1263_cast_fp16)[name = string("x2_27_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1301_cast_fp16 = mul(x = x1_27_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1301_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1302_cast_fp16 = mul(x = x2_27_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1302_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1303_cast_fp16 = sub(x = var_1301_cast_fp16, y = var_1302_cast_fp16)[name = string("op_1303_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1304_cast_fp16 = mul(x = x2_27_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1304_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1305_cast_fp16 = mul(x = x1_27_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1305_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1306_cast_fp16 = add(x = var_1304_cast_fp16, y = var_1305_cast_fp16)[name = string("op_1306_cast_fp16")];
            bool k_13_interleave_0 = const()[name = string("k_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_13_cast_fp16 = concat(axis = var_1198, interleave = k_13_interleave_0, values = (var_1303_cast_fp16, var_1306_cast_fp16))[name = string("k_13_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_12 = read_state(input = k_cache_6)[name = string("read_state_12")];
            int32 k_cache_27_axis_0 = const()[name = string("k_cache_27_axis_0"), val = int32(2)];
            string k_cache_27_mode_0 = const()[name = string("k_cache_27_mode_0"), val = string("update")];
            bool k_cache_27_validate_indices_0 = const()[name = string("k_cache_27_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_27_cast_fp16 = scatter_along_axis(axis = k_cache_27_axis_0, data = read_state_12, indices = scatter_indices, mode = k_cache_27_mode_0, updates = k_13_cast_fp16, validate_indices = k_cache_27_validate_indices_0)[name = string("k_cache_27_cast_fp16")];
            write_state(data = k_cache_27_cast_fp16, input = k_cache_6)[name = string("coreml_update_state_12_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_68 = read_state(input = k_cache_6)[name = string("coreml_update_state_12")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_13 = read_state(input = v_cache_6)[name = string("read_state_13")];
            int32 v_cache_27_axis_0 = const()[name = string("v_cache_27_axis_0"), val = int32(2)];
            string v_cache_27_mode_0 = const()[name = string("v_cache_27_mode_0"), val = string("update")];
            bool v_cache_27_validate_indices_0 = const()[name = string("v_cache_27_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_13_cast_fp16 = transpose(perm = v_13_perm_0, x = var_1241_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_27_cast_fp16 = scatter_along_axis(axis = v_cache_27_axis_0, data = read_state_13, indices = scatter_indices, mode = v_cache_27_mode_0, updates = v_13_cast_fp16, validate_indices = v_cache_27_validate_indices_0)[name = string("v_cache_27_cast_fp16")];
            write_state(data = v_cache_27_cast_fp16, input = v_cache_6)[name = string("coreml_update_state_13_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_69 = read_state(input = v_cache_6)[name = string("coreml_update_state_13")];
            tensor<int32, [1]> var_1311_axes_0 = const()[name = string("op_1311_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1311_cast_fp16 = expand_dims(axes = var_1311_axes_0, x = coreml_update_state_68)[name = string("op_1311_cast_fp16")];
            tensor<int32, [5]> k_exp_25_reps_0 = const()[name = string("k_exp_25_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_25_cast_fp16 = tile(reps = k_exp_25_reps_0, x = var_1311_cast_fp16)[name = string("k_exp_25_cast_fp16")];
            tensor<int32, [4]> var_1314 = const()[name = string("op_1314"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_27_cast_fp16 = reshape(shape = var_1314, x = k_exp_25_cast_fp16)[name = string("k_exp_27_cast_fp16")];
            tensor<int32, [1]> var_1316_axes_0 = const()[name = string("op_1316_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1316_cast_fp16 = expand_dims(axes = var_1316_axes_0, x = coreml_update_state_69)[name = string("op_1316_cast_fp16")];
            tensor<int32, [5]> v_exp_25_reps_0 = const()[name = string("v_exp_25_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_25_cast_fp16 = tile(reps = v_exp_25_reps_0, x = var_1316_cast_fp16)[name = string("v_exp_25_cast_fp16")];
            tensor<int32, [4]> var_1319 = const()[name = string("op_1319"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_27_cast_fp16 = reshape(shape = var_1319, x = v_exp_25_cast_fp16)[name = string("v_exp_27_cast_fp16")];
            bool var_1322_transpose_x_1 = const()[name = string("op_1322_transpose_x_1"), val = bool(false)];
            bool var_1322_transpose_y_1 = const()[name = string("op_1322_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_1322_cast_fp16 = matmul(transpose_x = var_1322_transpose_x_1, transpose_y = var_1322_transpose_y_1, x = q_13_cast_fp16, y = k_exp_27_cast_fp16)[name = string("op_1322_cast_fp16")];
            fp16 var_1323_to_fp16 = const()[name = string("op_1323_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_25_cast_fp16 = mul(x = var_1322_cast_fp16, y = var_1323_to_fp16)[name = string("attn_25_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_61_cast_fp16 = add(x = attn_25_cast_fp16, y = attention_mask)[name = string("input_61_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_27_cast_fp16 = softmax(axis = var_1198, x = input_61_cast_fp16)[name = string("attn_27_cast_fp16")];
            bool out_13_transpose_x_0 = const()[name = string("out_13_transpose_x_0"), val = bool(false)];
            bool out_13_transpose_y_0 = const()[name = string("out_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_13_cast_fp16 = matmul(transpose_x = out_13_transpose_x_0, transpose_y = out_13_transpose_y_0, x = attn_27_cast_fp16, y = v_exp_27_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [4]> var_1328_perm_0 = const()[name = string("op_1328_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_41_axis_0 = const()[name = string("concat_41_axis_0"), val = int32(0)];
            bool concat_41_interleave_0 = const()[name = string("concat_41_interleave_0"), val = bool(false)];
            int32 gather_37_cast_uint16_to_int32 = cast(dtype = gather_37_cast_uint16_to_int32_dtype_0, x = gather_37_cast_uint16)[name = string("cast_42")];
            tensor<int32, [3]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (gather_36, gather_37_cast_uint16_to_int32, var_1198))[name = string("concat_41")];
            tensor<fp16, [1, ?, 16, 128]> var_1328_cast_fp16 = transpose(perm = var_1328_perm_0, x = out_13_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, ?, ?]> input_63_cast_fp16 = reshape(shape = concat_41, x = var_1328_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104776960))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106874176))))[name = string("layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_175_cast_fp16 = add(x = x_155_cast_fp16, y = linear_45_cast_fp16)[name = string("x_175_cast_fp16")];
            fp16 var_1197_promoted_3_to_fp16 = const()[name = string("op_1197_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_1336_cast_fp16 = pow(x = x_175_cast_fp16, y = var_1197_promoted_3_to_fp16)[name = string("op_1336_cast_fp16")];
            tensor<int32, [1]> var_1338_axes_0 = const()[name = string("op_1338_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1338_keep_dims_0 = const()[name = string("op_1338_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_1338_cast_fp16 = reduce_mean(axes = var_1338_axes_0, keep_dims = var_1338_keep_dims_0, x = var_1336_cast_fp16)[name = string("op_1338_cast_fp16")];
            fp16 var_1339_to_fp16 = const()[name = string("op_1339_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1340_cast_fp16 = add(x = var_1338_cast_fp16, y = var_1339_to_fp16)[name = string("op_1340_cast_fp16")];
            fp32 norm_55_epsilon_0 = const()[name = string("norm_55_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_55_cast_fp16 = rsqrt(epsilon = norm_55_epsilon_0, x = var_1340_cast_fp16)[name = string("norm_55_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_1342_cast_fp16 = mul(x = x_175_cast_fp16, y = norm_55_cast_fp16)[name = string("op_1342_cast_fp16")];
            tensor<fp16, [1024]> layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107005312)))];
            tensor<fp16, [1, ?, 1024]> var_1343_cast_fp16 = mul(x = var_1342_cast_fp16, y = layers_6_post_attention_layernorm_weight_to_fp16)[name = string("op_1343_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107007424))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110153216))))[name = string("layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = var_1343_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_1353_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_1353_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110349888))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113495680))))[name = string("layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_6_mlp_up_proj_weight_to_fp16_quantized, x = var_1343_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_69_cast_fp16 = mul(x = var_1353_cast_fp16, y = linear_47_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113692352))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116838144))))[name = string("layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_48_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_181_cast_fp16 = add(x = x_175_cast_fp16, y = linear_48_cast_fp16)[name = string("x_181_cast_fp16")];
            int32 var_1373 = const()[name = string("op_1373"), val = int32(-1)];
            fp16 var_1372_promoted_to_fp16 = const()[name = string("op_1372_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_1382_cast_fp16 = pow(x = x_181_cast_fp16, y = var_1372_promoted_to_fp16)[name = string("op_1382_cast_fp16")];
            tensor<int32, [1]> var_1384_axes_0 = const()[name = string("op_1384_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1384_keep_dims_0 = const()[name = string("op_1384_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_1384_cast_fp16 = reduce_mean(axes = var_1384_axes_0, keep_dims = var_1384_keep_dims_0, x = var_1382_cast_fp16)[name = string("op_1384_cast_fp16")];
            fp16 var_1385_to_fp16 = const()[name = string("op_1385_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1386_cast_fp16 = add(x = var_1384_cast_fp16, y = var_1385_to_fp16)[name = string("op_1386_cast_fp16")];
            fp32 norm_57_epsilon_0 = const()[name = string("norm_57_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_57_cast_fp16 = rsqrt(epsilon = norm_57_epsilon_0, x = var_1386_cast_fp16)[name = string("norm_57_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_1388_cast_fp16 = mul(x = x_181_cast_fp16, y = norm_57_cast_fp16)[name = string("op_1388_cast_fp16")];
            tensor<fp16, [1024]> layers_7_input_layernorm_weight_to_fp16 = const()[name = string("layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117034816)))];
            tensor<fp16, [1, ?, 1024]> var_1389_cast_fp16 = mul(x = var_1388_cast_fp16, y = layers_7_input_layernorm_weight_to_fp16)[name = string("op_1389_cast_fp16")];
            tensor<int32, [3]> var_1401_shape_cast_fp16 = shape(x = var_1389_cast_fp16)[name = string("op_1401_shape_cast_fp16")];
            int32 gather_42 = const()[name = string("gather_42"), val = int32(1)];
            int32 gather_43_axis_0 = const()[name = string("gather_43_axis_0"), val = int32(0)];
            int32 gather_43_batch_dims_0 = const()[name = string("gather_43_batch_dims_0"), val = int32(0)];
            bool gather_43_validate_indices_0 = const()[name = string("gather_43_validate_indices_0"), val = bool(false)];
            string var_1401_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1401_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_43_to_uint16 = const()[name = string("select_43_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_1401_shape_cast_fp16_to_uint16 = cast(dtype = var_1401_shape_cast_fp16_to_uint16_dtype_0, x = var_1401_shape_cast_fp16)[name = string("cast_41")];
            uint16 gather_43_cast_uint16 = gather(axis = gather_43_axis_0, batch_dims = gather_43_batch_dims_0, indices = select_43_to_uint16, validate_indices = gather_43_validate_indices_0, x = var_1401_shape_cast_fp16_to_uint16)[name = string("gather_43_cast_uint16")];
            string gather_43_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_43_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117036928))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119134144))))[name = string("layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_49_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = var_1389_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> concat_42x = const()[name = string("concat_42x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_1406_cast_fp16 = reshape(shape = concat_42x, x = linear_49_cast_fp16)[name = string("op_1406_cast_fp16")];
            tensor<int32, [4]> x_187_perm_0 = const()[name = string("x_187_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119265280))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120313920))))[name = string("layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_50_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = var_1389_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> concat_43x = const()[name = string("concat_43x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_1411_cast_fp16 = reshape(shape = concat_43x, x = linear_50_cast_fp16)[name = string("op_1411_cast_fp16")];
            tensor<int32, [4]> x_191_perm_0 = const()[name = string("x_191_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120379520))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121428160))))[name = string("layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = var_1389_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> concat_44x = const()[name = string("concat_44x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_1416_cast_fp16 = reshape(shape = concat_44x, x = linear_51_cast_fp16)[name = string("op_1416_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1372_promoted_1_to_fp16 = const()[name = string("op_1372_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_187_cast_fp16 = transpose(perm = x_187_perm_0, x = var_1406_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 16, ?, 128]> var_1420_cast_fp16 = pow(x = x_187_cast_fp16, y = var_1372_promoted_1_to_fp16)[name = string("op_1420_cast_fp16")];
            tensor<int32, [1]> var_1422_axes_0 = const()[name = string("op_1422_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1422_keep_dims_0 = const()[name = string("op_1422_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_1422_cast_fp16 = reduce_mean(axes = var_1422_axes_0, keep_dims = var_1422_keep_dims_0, x = var_1420_cast_fp16)[name = string("op_1422_cast_fp16")];
            fp16 var_1423_to_fp16 = const()[name = string("op_1423_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_1424_cast_fp16 = add(x = var_1422_cast_fp16, y = var_1423_to_fp16)[name = string("op_1424_cast_fp16")];
            fp32 norm_59_epsilon_0 = const()[name = string("norm_59_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_59_cast_fp16 = rsqrt(epsilon = norm_59_epsilon_0, x = var_1424_cast_fp16)[name = string("norm_59_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_1426_cast_fp16 = mul(x = x_187_cast_fp16, y = norm_59_cast_fp16)[name = string("op_1426_cast_fp16")];
            tensor<fp16, [128]> layers_7_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121493760)))];
            tensor<fp16, [1, 16, ?, 128]> var_1427_cast_fp16 = mul(x = var_1426_cast_fp16, y = layers_7_self_attn_q_norm_weight_to_fp16)[name = string("op_1427_cast_fp16")];
            fp16 var_1372_promoted_2_to_fp16 = const()[name = string("op_1372_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_191_cast_fp16 = transpose(perm = x_191_perm_0, x = var_1411_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 8, ?, 128]> var_1431_cast_fp16 = pow(x = x_191_cast_fp16, y = var_1372_promoted_2_to_fp16)[name = string("op_1431_cast_fp16")];
            tensor<int32, [1]> var_1433_axes_0 = const()[name = string("op_1433_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1433_keep_dims_0 = const()[name = string("op_1433_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_1433_cast_fp16 = reduce_mean(axes = var_1433_axes_0, keep_dims = var_1433_keep_dims_0, x = var_1431_cast_fp16)[name = string("op_1433_cast_fp16")];
            fp16 var_1434_to_fp16 = const()[name = string("op_1434_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_1435_cast_fp16 = add(x = var_1433_cast_fp16, y = var_1434_to_fp16)[name = string("op_1435_cast_fp16")];
            fp32 norm_61_epsilon_0 = const()[name = string("norm_61_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_61_cast_fp16 = rsqrt(epsilon = norm_61_epsilon_0, x = var_1435_cast_fp16)[name = string("norm_61_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_1437_cast_fp16 = mul(x = x_191_cast_fp16, y = norm_61_cast_fp16)[name = string("op_1437_cast_fp16")];
            tensor<fp16, [128]> layers_7_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121494080)))];
            tensor<fp16, [1, 8, ?, 128]> var_1438_cast_fp16 = mul(x = var_1437_cast_fp16, y = layers_7_self_attn_k_norm_weight_to_fp16)[name = string("op_1438_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = var_1427_cast_fp16)[name = string("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = var_1427_cast_fp16)[name = string("x2_29_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1454_cast_fp16 = mul(x = x1_29_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1454_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1455_cast_fp16 = mul(x = x2_29_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1455_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1456_cast_fp16 = sub(x = var_1454_cast_fp16, y = var_1455_cast_fp16)[name = string("op_1456_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1457_cast_fp16 = mul(x = x2_29_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1457_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1458_cast_fp16 = mul(x = x1_29_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1458_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1459_cast_fp16 = add(x = var_1457_cast_fp16, y = var_1458_cast_fp16)[name = string("op_1459_cast_fp16")];
            bool q_15_interleave_0 = const()[name = string("q_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_15_cast_fp16 = concat(axis = var_1373, interleave = q_15_interleave_0, values = (var_1456_cast_fp16, var_1459_cast_fp16))[name = string("q_15_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = var_1438_cast_fp16)[name = string("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = var_1438_cast_fp16)[name = string("x2_31_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1476_cast_fp16 = mul(x = x1_31_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1476_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1477_cast_fp16 = mul(x = x2_31_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1477_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1478_cast_fp16 = sub(x = var_1476_cast_fp16, y = var_1477_cast_fp16)[name = string("op_1478_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1479_cast_fp16 = mul(x = x2_31_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1479_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1480_cast_fp16 = mul(x = x1_31_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1480_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1481_cast_fp16 = add(x = var_1479_cast_fp16, y = var_1480_cast_fp16)[name = string("op_1481_cast_fp16")];
            bool k_15_interleave_0 = const()[name = string("k_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_15_cast_fp16 = concat(axis = var_1373, interleave = k_15_interleave_0, values = (var_1478_cast_fp16, var_1481_cast_fp16))[name = string("k_15_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_14 = read_state(input = k_cache_7)[name = string("read_state_14")];
            int32 k_cache_31_axis_0 = const()[name = string("k_cache_31_axis_0"), val = int32(2)];
            string k_cache_31_mode_0 = const()[name = string("k_cache_31_mode_0"), val = string("update")];
            bool k_cache_31_validate_indices_0 = const()[name = string("k_cache_31_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_31_cast_fp16 = scatter_along_axis(axis = k_cache_31_axis_0, data = read_state_14, indices = scatter_indices, mode = k_cache_31_mode_0, updates = k_15_cast_fp16, validate_indices = k_cache_31_validate_indices_0)[name = string("k_cache_31_cast_fp16")];
            write_state(data = k_cache_31_cast_fp16, input = k_cache_7)[name = string("coreml_update_state_14_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_70 = read_state(input = k_cache_7)[name = string("coreml_update_state_14")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_15 = read_state(input = v_cache_7)[name = string("read_state_15")];
            int32 v_cache_31_axis_0 = const()[name = string("v_cache_31_axis_0"), val = int32(2)];
            string v_cache_31_mode_0 = const()[name = string("v_cache_31_mode_0"), val = string("update")];
            bool v_cache_31_validate_indices_0 = const()[name = string("v_cache_31_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_1416_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_31_cast_fp16 = scatter_along_axis(axis = v_cache_31_axis_0, data = read_state_15, indices = scatter_indices, mode = v_cache_31_mode_0, updates = v_15_cast_fp16, validate_indices = v_cache_31_validate_indices_0)[name = string("v_cache_31_cast_fp16")];
            write_state(data = v_cache_31_cast_fp16, input = v_cache_7)[name = string("coreml_update_state_15_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_71 = read_state(input = v_cache_7)[name = string("coreml_update_state_15")];
            tensor<int32, [1]> var_1486_axes_0 = const()[name = string("op_1486_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1486_cast_fp16 = expand_dims(axes = var_1486_axes_0, x = coreml_update_state_70)[name = string("op_1486_cast_fp16")];
            tensor<int32, [5]> k_exp_29_reps_0 = const()[name = string("k_exp_29_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_29_cast_fp16 = tile(reps = k_exp_29_reps_0, x = var_1486_cast_fp16)[name = string("k_exp_29_cast_fp16")];
            tensor<int32, [4]> var_1489 = const()[name = string("op_1489"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_31_cast_fp16 = reshape(shape = var_1489, x = k_exp_29_cast_fp16)[name = string("k_exp_31_cast_fp16")];
            tensor<int32, [1]> var_1491_axes_0 = const()[name = string("op_1491_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1491_cast_fp16 = expand_dims(axes = var_1491_axes_0, x = coreml_update_state_71)[name = string("op_1491_cast_fp16")];
            tensor<int32, [5]> v_exp_29_reps_0 = const()[name = string("v_exp_29_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_29_cast_fp16 = tile(reps = v_exp_29_reps_0, x = var_1491_cast_fp16)[name = string("v_exp_29_cast_fp16")];
            tensor<int32, [4]> var_1494 = const()[name = string("op_1494"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_31_cast_fp16 = reshape(shape = var_1494, x = v_exp_29_cast_fp16)[name = string("v_exp_31_cast_fp16")];
            bool var_1497_transpose_x_1 = const()[name = string("op_1497_transpose_x_1"), val = bool(false)];
            bool var_1497_transpose_y_1 = const()[name = string("op_1497_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_1497_cast_fp16 = matmul(transpose_x = var_1497_transpose_x_1, transpose_y = var_1497_transpose_y_1, x = q_15_cast_fp16, y = k_exp_31_cast_fp16)[name = string("op_1497_cast_fp16")];
            fp16 var_1498_to_fp16 = const()[name = string("op_1498_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_29_cast_fp16 = mul(x = var_1497_cast_fp16, y = var_1498_to_fp16)[name = string("attn_29_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_71_cast_fp16 = add(x = attn_29_cast_fp16, y = attention_mask)[name = string("input_71_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_31_cast_fp16 = softmax(axis = var_1373, x = input_71_cast_fp16)[name = string("attn_31_cast_fp16")];
            bool out_15_transpose_x_0 = const()[name = string("out_15_transpose_x_0"), val = bool(false)];
            bool out_15_transpose_y_0 = const()[name = string("out_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_15_cast_fp16 = matmul(transpose_x = out_15_transpose_x_0, transpose_y = out_15_transpose_y_0, x = attn_31_cast_fp16, y = v_exp_31_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [4]> var_1503_perm_0 = const()[name = string("op_1503_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_47_axis_0 = const()[name = string("concat_47_axis_0"), val = int32(0)];
            bool concat_47_interleave_0 = const()[name = string("concat_47_interleave_0"), val = bool(false)];
            int32 gather_43_cast_uint16_to_int32 = cast(dtype = gather_43_cast_uint16_to_int32_dtype_0, x = gather_43_cast_uint16)[name = string("cast_40")];
            tensor<int32, [3]> concat_47 = concat(axis = concat_47_axis_0, interleave = concat_47_interleave_0, values = (gather_42, gather_43_cast_uint16_to_int32, var_1373))[name = string("concat_47")];
            tensor<fp16, [1, ?, 16, 128]> var_1503_cast_fp16 = transpose(perm = var_1503_perm_0, x = out_15_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, ?, ?]> input_73_cast_fp16 = reshape(shape = concat_47, x = var_1503_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121494400))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123591616))))[name = string("layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_52_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_201_cast_fp16 = add(x = x_181_cast_fp16, y = linear_52_cast_fp16)[name = string("x_201_cast_fp16")];
            fp16 var_1372_promoted_3_to_fp16 = const()[name = string("op_1372_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_1511_cast_fp16 = pow(x = x_201_cast_fp16, y = var_1372_promoted_3_to_fp16)[name = string("op_1511_cast_fp16")];
            tensor<int32, [1]> var_1513_axes_0 = const()[name = string("op_1513_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1513_keep_dims_0 = const()[name = string("op_1513_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_1513_cast_fp16 = reduce_mean(axes = var_1513_axes_0, keep_dims = var_1513_keep_dims_0, x = var_1511_cast_fp16)[name = string("op_1513_cast_fp16")];
            fp16 var_1514_to_fp16 = const()[name = string("op_1514_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1515_cast_fp16 = add(x = var_1513_cast_fp16, y = var_1514_to_fp16)[name = string("op_1515_cast_fp16")];
            fp32 norm_63_epsilon_0 = const()[name = string("norm_63_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_63_cast_fp16 = rsqrt(epsilon = norm_63_epsilon_0, x = var_1515_cast_fp16)[name = string("norm_63_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_1517_cast_fp16 = mul(x = x_201_cast_fp16, y = norm_63_cast_fp16)[name = string("op_1517_cast_fp16")];
            tensor<fp16, [1024]> layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123722752)))];
            tensor<fp16, [1, ?, 1024]> var_1518_cast_fp16 = mul(x = var_1517_cast_fp16, y = layers_7_post_attention_layernorm_weight_to_fp16)[name = string("op_1518_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123724864))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126870656))))[name = string("layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = var_1518_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_1528_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_1528_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127067328))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130213120))))[name = string("layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_7_mlp_up_proj_weight_to_fp16_quantized, x = var_1518_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_79_cast_fp16 = mul(x = var_1528_cast_fp16, y = linear_54_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130409792))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133555584))))[name = string("layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_207_cast_fp16 = add(x = x_201_cast_fp16, y = linear_55_cast_fp16)[name = string("x_207_cast_fp16")];
            int32 var_1548 = const()[name = string("op_1548"), val = int32(-1)];
            fp16 var_1547_promoted_to_fp16 = const()[name = string("op_1547_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_1557_cast_fp16 = pow(x = x_207_cast_fp16, y = var_1547_promoted_to_fp16)[name = string("op_1557_cast_fp16")];
            tensor<int32, [1]> var_1559_axes_0 = const()[name = string("op_1559_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1559_keep_dims_0 = const()[name = string("op_1559_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_1559_cast_fp16 = reduce_mean(axes = var_1559_axes_0, keep_dims = var_1559_keep_dims_0, x = var_1557_cast_fp16)[name = string("op_1559_cast_fp16")];
            fp16 var_1560_to_fp16 = const()[name = string("op_1560_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1561_cast_fp16 = add(x = var_1559_cast_fp16, y = var_1560_to_fp16)[name = string("op_1561_cast_fp16")];
            fp32 norm_65_epsilon_0 = const()[name = string("norm_65_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_65_cast_fp16 = rsqrt(epsilon = norm_65_epsilon_0, x = var_1561_cast_fp16)[name = string("norm_65_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_1563_cast_fp16 = mul(x = x_207_cast_fp16, y = norm_65_cast_fp16)[name = string("op_1563_cast_fp16")];
            tensor<fp16, [1024]> layers_8_input_layernorm_weight_to_fp16 = const()[name = string("layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133752256)))];
            tensor<fp16, [1, ?, 1024]> var_1564_cast_fp16 = mul(x = var_1563_cast_fp16, y = layers_8_input_layernorm_weight_to_fp16)[name = string("op_1564_cast_fp16")];
            tensor<int32, [3]> var_1576_shape_cast_fp16 = shape(x = var_1564_cast_fp16)[name = string("op_1576_shape_cast_fp16")];
            int32 gather_48 = const()[name = string("gather_48"), val = int32(1)];
            int32 gather_49_axis_0 = const()[name = string("gather_49_axis_0"), val = int32(0)];
            int32 gather_49_batch_dims_0 = const()[name = string("gather_49_batch_dims_0"), val = int32(0)];
            bool gather_49_validate_indices_0 = const()[name = string("gather_49_validate_indices_0"), val = bool(false)];
            string var_1576_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1576_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_49_to_uint16 = const()[name = string("select_49_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_1576_shape_cast_fp16_to_uint16 = cast(dtype = var_1576_shape_cast_fp16_to_uint16_dtype_0, x = var_1576_shape_cast_fp16)[name = string("cast_39")];
            uint16 gather_49_cast_uint16 = gather(axis = gather_49_axis_0, batch_dims = gather_49_batch_dims_0, indices = select_49_to_uint16, validate_indices = gather_49_validate_indices_0, x = var_1576_shape_cast_fp16_to_uint16)[name = string("gather_49_cast_uint16")];
            string gather_49_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_49_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133754368))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135851584))))[name = string("layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = var_1564_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> concat_48x = const()[name = string("concat_48x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_1581_cast_fp16 = reshape(shape = concat_48x, x = linear_56_cast_fp16)[name = string("op_1581_cast_fp16")];
            tensor<int32, [4]> x_213_perm_0 = const()[name = string("x_213_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135982720))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137031360))))[name = string("layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = var_1564_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> concat_49x = const()[name = string("concat_49x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_1586_cast_fp16 = reshape(shape = concat_49x, x = linear_57_cast_fp16)[name = string("op_1586_cast_fp16")];
            tensor<int32, [4]> x_217_perm_0 = const()[name = string("x_217_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137096960))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138145600))))[name = string("layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_58_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = var_1564_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> concat_50x = const()[name = string("concat_50x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_1591_cast_fp16 = reshape(shape = concat_50x, x = linear_58_cast_fp16)[name = string("op_1591_cast_fp16")];
            tensor<int32, [4]> v_17_perm_0 = const()[name = string("v_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1547_promoted_1_to_fp16 = const()[name = string("op_1547_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_213_cast_fp16 = transpose(perm = x_213_perm_0, x = var_1581_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 16, ?, 128]> var_1595_cast_fp16 = pow(x = x_213_cast_fp16, y = var_1547_promoted_1_to_fp16)[name = string("op_1595_cast_fp16")];
            tensor<int32, [1]> var_1597_axes_0 = const()[name = string("op_1597_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1597_keep_dims_0 = const()[name = string("op_1597_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_1597_cast_fp16 = reduce_mean(axes = var_1597_axes_0, keep_dims = var_1597_keep_dims_0, x = var_1595_cast_fp16)[name = string("op_1597_cast_fp16")];
            fp16 var_1598_to_fp16 = const()[name = string("op_1598_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_1599_cast_fp16 = add(x = var_1597_cast_fp16, y = var_1598_to_fp16)[name = string("op_1599_cast_fp16")];
            fp32 norm_67_epsilon_0 = const()[name = string("norm_67_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_67_cast_fp16 = rsqrt(epsilon = norm_67_epsilon_0, x = var_1599_cast_fp16)[name = string("norm_67_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_1601_cast_fp16 = mul(x = x_213_cast_fp16, y = norm_67_cast_fp16)[name = string("op_1601_cast_fp16")];
            tensor<fp16, [128]> layers_8_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138211200)))];
            tensor<fp16, [1, 16, ?, 128]> var_1602_cast_fp16 = mul(x = var_1601_cast_fp16, y = layers_8_self_attn_q_norm_weight_to_fp16)[name = string("op_1602_cast_fp16")];
            fp16 var_1547_promoted_2_to_fp16 = const()[name = string("op_1547_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_217_cast_fp16 = transpose(perm = x_217_perm_0, x = var_1586_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 8, ?, 128]> var_1606_cast_fp16 = pow(x = x_217_cast_fp16, y = var_1547_promoted_2_to_fp16)[name = string("op_1606_cast_fp16")];
            tensor<int32, [1]> var_1608_axes_0 = const()[name = string("op_1608_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1608_keep_dims_0 = const()[name = string("op_1608_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_1608_cast_fp16 = reduce_mean(axes = var_1608_axes_0, keep_dims = var_1608_keep_dims_0, x = var_1606_cast_fp16)[name = string("op_1608_cast_fp16")];
            fp16 var_1609_to_fp16 = const()[name = string("op_1609_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_1610_cast_fp16 = add(x = var_1608_cast_fp16, y = var_1609_to_fp16)[name = string("op_1610_cast_fp16")];
            fp32 norm_69_epsilon_0 = const()[name = string("norm_69_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_69_cast_fp16 = rsqrt(epsilon = norm_69_epsilon_0, x = var_1610_cast_fp16)[name = string("norm_69_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_1612_cast_fp16 = mul(x = x_217_cast_fp16, y = norm_69_cast_fp16)[name = string("op_1612_cast_fp16")];
            tensor<fp16, [128]> layers_8_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138211520)))];
            tensor<fp16, [1, 8, ?, 128]> var_1613_cast_fp16 = mul(x = var_1612_cast_fp16, y = layers_8_self_attn_k_norm_weight_to_fp16)[name = string("op_1613_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = var_1602_cast_fp16)[name = string("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = var_1602_cast_fp16)[name = string("x2_33_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1629_cast_fp16 = mul(x = x1_33_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1629_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1630_cast_fp16 = mul(x = x2_33_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1630_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1631_cast_fp16 = sub(x = var_1629_cast_fp16, y = var_1630_cast_fp16)[name = string("op_1631_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1632_cast_fp16 = mul(x = x2_33_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1632_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1633_cast_fp16 = mul(x = x1_33_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1633_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1634_cast_fp16 = add(x = var_1632_cast_fp16, y = var_1633_cast_fp16)[name = string("op_1634_cast_fp16")];
            bool q_17_interleave_0 = const()[name = string("q_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_17_cast_fp16 = concat(axis = var_1548, interleave = q_17_interleave_0, values = (var_1631_cast_fp16, var_1634_cast_fp16))[name = string("q_17_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = var_1613_cast_fp16)[name = string("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = var_1613_cast_fp16)[name = string("x2_35_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1651_cast_fp16 = mul(x = x1_35_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1651_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1652_cast_fp16 = mul(x = x2_35_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1652_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1653_cast_fp16 = sub(x = var_1651_cast_fp16, y = var_1652_cast_fp16)[name = string("op_1653_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1654_cast_fp16 = mul(x = x2_35_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1654_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1655_cast_fp16 = mul(x = x1_35_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1655_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1656_cast_fp16 = add(x = var_1654_cast_fp16, y = var_1655_cast_fp16)[name = string("op_1656_cast_fp16")];
            bool k_17_interleave_0 = const()[name = string("k_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_17_cast_fp16 = concat(axis = var_1548, interleave = k_17_interleave_0, values = (var_1653_cast_fp16, var_1656_cast_fp16))[name = string("k_17_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_16 = read_state(input = k_cache_8)[name = string("read_state_16")];
            int32 k_cache_35_axis_0 = const()[name = string("k_cache_35_axis_0"), val = int32(2)];
            string k_cache_35_mode_0 = const()[name = string("k_cache_35_mode_0"), val = string("update")];
            bool k_cache_35_validate_indices_0 = const()[name = string("k_cache_35_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_35_cast_fp16 = scatter_along_axis(axis = k_cache_35_axis_0, data = read_state_16, indices = scatter_indices, mode = k_cache_35_mode_0, updates = k_17_cast_fp16, validate_indices = k_cache_35_validate_indices_0)[name = string("k_cache_35_cast_fp16")];
            write_state(data = k_cache_35_cast_fp16, input = k_cache_8)[name = string("coreml_update_state_16_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_72 = read_state(input = k_cache_8)[name = string("coreml_update_state_16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_17 = read_state(input = v_cache_8)[name = string("read_state_17")];
            int32 v_cache_35_axis_0 = const()[name = string("v_cache_35_axis_0"), val = int32(2)];
            string v_cache_35_mode_0 = const()[name = string("v_cache_35_mode_0"), val = string("update")];
            bool v_cache_35_validate_indices_0 = const()[name = string("v_cache_35_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_17_cast_fp16 = transpose(perm = v_17_perm_0, x = var_1591_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_35_cast_fp16 = scatter_along_axis(axis = v_cache_35_axis_0, data = read_state_17, indices = scatter_indices, mode = v_cache_35_mode_0, updates = v_17_cast_fp16, validate_indices = v_cache_35_validate_indices_0)[name = string("v_cache_35_cast_fp16")];
            write_state(data = v_cache_35_cast_fp16, input = v_cache_8)[name = string("coreml_update_state_17_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_73 = read_state(input = v_cache_8)[name = string("coreml_update_state_17")];
            tensor<int32, [1]> var_1661_axes_0 = const()[name = string("op_1661_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1661_cast_fp16 = expand_dims(axes = var_1661_axes_0, x = coreml_update_state_72)[name = string("op_1661_cast_fp16")];
            tensor<int32, [5]> k_exp_33_reps_0 = const()[name = string("k_exp_33_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_33_cast_fp16 = tile(reps = k_exp_33_reps_0, x = var_1661_cast_fp16)[name = string("k_exp_33_cast_fp16")];
            tensor<int32, [4]> var_1664 = const()[name = string("op_1664"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_35_cast_fp16 = reshape(shape = var_1664, x = k_exp_33_cast_fp16)[name = string("k_exp_35_cast_fp16")];
            tensor<int32, [1]> var_1666_axes_0 = const()[name = string("op_1666_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1666_cast_fp16 = expand_dims(axes = var_1666_axes_0, x = coreml_update_state_73)[name = string("op_1666_cast_fp16")];
            tensor<int32, [5]> v_exp_33_reps_0 = const()[name = string("v_exp_33_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_33_cast_fp16 = tile(reps = v_exp_33_reps_0, x = var_1666_cast_fp16)[name = string("v_exp_33_cast_fp16")];
            tensor<int32, [4]> var_1669 = const()[name = string("op_1669"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_35_cast_fp16 = reshape(shape = var_1669, x = v_exp_33_cast_fp16)[name = string("v_exp_35_cast_fp16")];
            bool var_1672_transpose_x_1 = const()[name = string("op_1672_transpose_x_1"), val = bool(false)];
            bool var_1672_transpose_y_1 = const()[name = string("op_1672_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_1672_cast_fp16 = matmul(transpose_x = var_1672_transpose_x_1, transpose_y = var_1672_transpose_y_1, x = q_17_cast_fp16, y = k_exp_35_cast_fp16)[name = string("op_1672_cast_fp16")];
            fp16 var_1673_to_fp16 = const()[name = string("op_1673_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_33_cast_fp16 = mul(x = var_1672_cast_fp16, y = var_1673_to_fp16)[name = string("attn_33_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_81_cast_fp16 = add(x = attn_33_cast_fp16, y = attention_mask)[name = string("input_81_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_35_cast_fp16 = softmax(axis = var_1548, x = input_81_cast_fp16)[name = string("attn_35_cast_fp16")];
            bool out_17_transpose_x_0 = const()[name = string("out_17_transpose_x_0"), val = bool(false)];
            bool out_17_transpose_y_0 = const()[name = string("out_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_17_cast_fp16 = matmul(transpose_x = out_17_transpose_x_0, transpose_y = out_17_transpose_y_0, x = attn_35_cast_fp16, y = v_exp_35_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [4]> var_1678_perm_0 = const()[name = string("op_1678_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_53_axis_0 = const()[name = string("concat_53_axis_0"), val = int32(0)];
            bool concat_53_interleave_0 = const()[name = string("concat_53_interleave_0"), val = bool(false)];
            int32 gather_49_cast_uint16_to_int32 = cast(dtype = gather_49_cast_uint16_to_int32_dtype_0, x = gather_49_cast_uint16)[name = string("cast_38")];
            tensor<int32, [3]> concat_53 = concat(axis = concat_53_axis_0, interleave = concat_53_interleave_0, values = (gather_48, gather_49_cast_uint16_to_int32, var_1548))[name = string("concat_53")];
            tensor<fp16, [1, ?, 16, 128]> var_1678_cast_fp16 = transpose(perm = var_1678_perm_0, x = out_17_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, ?, ?]> input_83_cast_fp16 = reshape(shape = concat_53, x = var_1678_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138211840))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140309056))))[name = string("layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_227_cast_fp16 = add(x = x_207_cast_fp16, y = linear_59_cast_fp16)[name = string("x_227_cast_fp16")];
            fp16 var_1547_promoted_3_to_fp16 = const()[name = string("op_1547_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_1686_cast_fp16 = pow(x = x_227_cast_fp16, y = var_1547_promoted_3_to_fp16)[name = string("op_1686_cast_fp16")];
            tensor<int32, [1]> var_1688_axes_0 = const()[name = string("op_1688_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1688_keep_dims_0 = const()[name = string("op_1688_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_1688_cast_fp16 = reduce_mean(axes = var_1688_axes_0, keep_dims = var_1688_keep_dims_0, x = var_1686_cast_fp16)[name = string("op_1688_cast_fp16")];
            fp16 var_1689_to_fp16 = const()[name = string("op_1689_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1690_cast_fp16 = add(x = var_1688_cast_fp16, y = var_1689_to_fp16)[name = string("op_1690_cast_fp16")];
            fp32 norm_71_epsilon_0 = const()[name = string("norm_71_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_71_cast_fp16 = rsqrt(epsilon = norm_71_epsilon_0, x = var_1690_cast_fp16)[name = string("norm_71_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_1692_cast_fp16 = mul(x = x_227_cast_fp16, y = norm_71_cast_fp16)[name = string("op_1692_cast_fp16")];
            tensor<fp16, [1024]> layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140440192)))];
            tensor<fp16, [1, ?, 1024]> var_1693_cast_fp16 = mul(x = var_1692_cast_fp16, y = layers_8_post_attention_layernorm_weight_to_fp16)[name = string("op_1693_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140442304))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143588096))))[name = string("layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = var_1693_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_1703_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_1703_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143784768))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146930560))))[name = string("layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_8_mlp_up_proj_weight_to_fp16_quantized, x = var_1693_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_89_cast_fp16 = mul(x = var_1703_cast_fp16, y = linear_61_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147127232))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150273024))))[name = string("layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_62_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_233_cast_fp16 = add(x = x_227_cast_fp16, y = linear_62_cast_fp16)[name = string("x_233_cast_fp16")];
            int32 var_1723 = const()[name = string("op_1723"), val = int32(-1)];
            fp16 var_1722_promoted_to_fp16 = const()[name = string("op_1722_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_1732_cast_fp16 = pow(x = x_233_cast_fp16, y = var_1722_promoted_to_fp16)[name = string("op_1732_cast_fp16")];
            tensor<int32, [1]> var_1734_axes_0 = const()[name = string("op_1734_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1734_keep_dims_0 = const()[name = string("op_1734_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_1734_cast_fp16 = reduce_mean(axes = var_1734_axes_0, keep_dims = var_1734_keep_dims_0, x = var_1732_cast_fp16)[name = string("op_1734_cast_fp16")];
            fp16 var_1735_to_fp16 = const()[name = string("op_1735_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1736_cast_fp16 = add(x = var_1734_cast_fp16, y = var_1735_to_fp16)[name = string("op_1736_cast_fp16")];
            fp32 norm_73_epsilon_0 = const()[name = string("norm_73_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_73_cast_fp16 = rsqrt(epsilon = norm_73_epsilon_0, x = var_1736_cast_fp16)[name = string("norm_73_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_1738_cast_fp16 = mul(x = x_233_cast_fp16, y = norm_73_cast_fp16)[name = string("op_1738_cast_fp16")];
            tensor<fp16, [1024]> layers_9_input_layernorm_weight_to_fp16 = const()[name = string("layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150469696)))];
            tensor<fp16, [1, ?, 1024]> var_1739_cast_fp16 = mul(x = var_1738_cast_fp16, y = layers_9_input_layernorm_weight_to_fp16)[name = string("op_1739_cast_fp16")];
            tensor<int32, [3]> var_1751_shape_cast_fp16 = shape(x = var_1739_cast_fp16)[name = string("op_1751_shape_cast_fp16")];
            int32 gather_54 = const()[name = string("gather_54"), val = int32(1)];
            int32 gather_55_axis_0 = const()[name = string("gather_55_axis_0"), val = int32(0)];
            int32 gather_55_batch_dims_0 = const()[name = string("gather_55_batch_dims_0"), val = int32(0)];
            bool gather_55_validate_indices_0 = const()[name = string("gather_55_validate_indices_0"), val = bool(false)];
            string var_1751_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1751_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_55_to_uint16 = const()[name = string("select_55_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_1751_shape_cast_fp16_to_uint16 = cast(dtype = var_1751_shape_cast_fp16_to_uint16_dtype_0, x = var_1751_shape_cast_fp16)[name = string("cast_37")];
            uint16 gather_55_cast_uint16 = gather(axis = gather_55_axis_0, batch_dims = gather_55_batch_dims_0, indices = select_55_to_uint16, validate_indices = gather_55_validate_indices_0, x = var_1751_shape_cast_fp16_to_uint16)[name = string("gather_55_cast_uint16")];
            string gather_55_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_55_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150471808))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152569024))))[name = string("layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_63_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = var_1739_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> concat_54x = const()[name = string("concat_54x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_1756_cast_fp16 = reshape(shape = concat_54x, x = linear_63_cast_fp16)[name = string("op_1756_cast_fp16")];
            tensor<int32, [4]> x_239_perm_0 = const()[name = string("x_239_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152700160))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153748800))))[name = string("layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = var_1739_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> concat_55x = const()[name = string("concat_55x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_1761_cast_fp16 = reshape(shape = concat_55x, x = linear_64_cast_fp16)[name = string("op_1761_cast_fp16")];
            tensor<int32, [4]> x_243_perm_0 = const()[name = string("x_243_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153814400))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154863040))))[name = string("layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = var_1739_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> concat_56x = const()[name = string("concat_56x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_1766_cast_fp16 = reshape(shape = concat_56x, x = linear_65_cast_fp16)[name = string("op_1766_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1722_promoted_1_to_fp16 = const()[name = string("op_1722_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_239_cast_fp16 = transpose(perm = x_239_perm_0, x = var_1756_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 16, ?, 128]> var_1770_cast_fp16 = pow(x = x_239_cast_fp16, y = var_1722_promoted_1_to_fp16)[name = string("op_1770_cast_fp16")];
            tensor<int32, [1]> var_1772_axes_0 = const()[name = string("op_1772_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1772_keep_dims_0 = const()[name = string("op_1772_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_1772_cast_fp16 = reduce_mean(axes = var_1772_axes_0, keep_dims = var_1772_keep_dims_0, x = var_1770_cast_fp16)[name = string("op_1772_cast_fp16")];
            fp16 var_1773_to_fp16 = const()[name = string("op_1773_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_1774_cast_fp16 = add(x = var_1772_cast_fp16, y = var_1773_to_fp16)[name = string("op_1774_cast_fp16")];
            fp32 norm_75_epsilon_0 = const()[name = string("norm_75_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_75_cast_fp16 = rsqrt(epsilon = norm_75_epsilon_0, x = var_1774_cast_fp16)[name = string("norm_75_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_1776_cast_fp16 = mul(x = x_239_cast_fp16, y = norm_75_cast_fp16)[name = string("op_1776_cast_fp16")];
            tensor<fp16, [128]> layers_9_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154928640)))];
            tensor<fp16, [1, 16, ?, 128]> var_1777_cast_fp16 = mul(x = var_1776_cast_fp16, y = layers_9_self_attn_q_norm_weight_to_fp16)[name = string("op_1777_cast_fp16")];
            fp16 var_1722_promoted_2_to_fp16 = const()[name = string("op_1722_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_243_cast_fp16 = transpose(perm = x_243_perm_0, x = var_1761_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 8, ?, 128]> var_1781_cast_fp16 = pow(x = x_243_cast_fp16, y = var_1722_promoted_2_to_fp16)[name = string("op_1781_cast_fp16")];
            tensor<int32, [1]> var_1783_axes_0 = const()[name = string("op_1783_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1783_keep_dims_0 = const()[name = string("op_1783_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_1783_cast_fp16 = reduce_mean(axes = var_1783_axes_0, keep_dims = var_1783_keep_dims_0, x = var_1781_cast_fp16)[name = string("op_1783_cast_fp16")];
            fp16 var_1784_to_fp16 = const()[name = string("op_1784_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_1785_cast_fp16 = add(x = var_1783_cast_fp16, y = var_1784_to_fp16)[name = string("op_1785_cast_fp16")];
            fp32 norm_77_epsilon_0 = const()[name = string("norm_77_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_77_cast_fp16 = rsqrt(epsilon = norm_77_epsilon_0, x = var_1785_cast_fp16)[name = string("norm_77_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_1787_cast_fp16 = mul(x = x_243_cast_fp16, y = norm_77_cast_fp16)[name = string("op_1787_cast_fp16")];
            tensor<fp16, [128]> layers_9_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154928960)))];
            tensor<fp16, [1, 8, ?, 128]> var_1788_cast_fp16 = mul(x = var_1787_cast_fp16, y = layers_9_self_attn_k_norm_weight_to_fp16)[name = string("op_1788_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = var_1777_cast_fp16)[name = string("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = var_1777_cast_fp16)[name = string("x2_37_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1804_cast_fp16 = mul(x = x1_37_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1804_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1805_cast_fp16 = mul(x = x2_37_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1805_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1806_cast_fp16 = sub(x = var_1804_cast_fp16, y = var_1805_cast_fp16)[name = string("op_1806_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1807_cast_fp16 = mul(x = x2_37_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1808_cast_fp16 = mul(x = x1_37_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1808_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1809_cast_fp16 = add(x = var_1807_cast_fp16, y = var_1808_cast_fp16)[name = string("op_1809_cast_fp16")];
            bool q_19_interleave_0 = const()[name = string("q_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_19_cast_fp16 = concat(axis = var_1723, interleave = q_19_interleave_0, values = (var_1806_cast_fp16, var_1809_cast_fp16))[name = string("q_19_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = var_1788_cast_fp16)[name = string("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = var_1788_cast_fp16)[name = string("x2_39_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1826_cast_fp16 = mul(x = x1_39_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1826_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1827_cast_fp16 = mul(x = x2_39_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1827_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1828_cast_fp16 = sub(x = var_1826_cast_fp16, y = var_1827_cast_fp16)[name = string("op_1828_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1829_cast_fp16 = mul(x = x2_39_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1829_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1830_cast_fp16 = mul(x = x1_39_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1830_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_1831_cast_fp16 = add(x = var_1829_cast_fp16, y = var_1830_cast_fp16)[name = string("op_1831_cast_fp16")];
            bool k_19_interleave_0 = const()[name = string("k_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_19_cast_fp16 = concat(axis = var_1723, interleave = k_19_interleave_0, values = (var_1828_cast_fp16, var_1831_cast_fp16))[name = string("k_19_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_18 = read_state(input = k_cache_9)[name = string("read_state_18")];
            int32 k_cache_39_axis_0 = const()[name = string("k_cache_39_axis_0"), val = int32(2)];
            string k_cache_39_mode_0 = const()[name = string("k_cache_39_mode_0"), val = string("update")];
            bool k_cache_39_validate_indices_0 = const()[name = string("k_cache_39_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_39_cast_fp16 = scatter_along_axis(axis = k_cache_39_axis_0, data = read_state_18, indices = scatter_indices, mode = k_cache_39_mode_0, updates = k_19_cast_fp16, validate_indices = k_cache_39_validate_indices_0)[name = string("k_cache_39_cast_fp16")];
            write_state(data = k_cache_39_cast_fp16, input = k_cache_9)[name = string("coreml_update_state_18_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_74 = read_state(input = k_cache_9)[name = string("coreml_update_state_18")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_19 = read_state(input = v_cache_9)[name = string("read_state_19")];
            int32 v_cache_39_axis_0 = const()[name = string("v_cache_39_axis_0"), val = int32(2)];
            string v_cache_39_mode_0 = const()[name = string("v_cache_39_mode_0"), val = string("update")];
            bool v_cache_39_validate_indices_0 = const()[name = string("v_cache_39_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = var_1766_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_39_cast_fp16 = scatter_along_axis(axis = v_cache_39_axis_0, data = read_state_19, indices = scatter_indices, mode = v_cache_39_mode_0, updates = v_19_cast_fp16, validate_indices = v_cache_39_validate_indices_0)[name = string("v_cache_39_cast_fp16")];
            write_state(data = v_cache_39_cast_fp16, input = v_cache_9)[name = string("coreml_update_state_19_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_75 = read_state(input = v_cache_9)[name = string("coreml_update_state_19")];
            tensor<int32, [1]> var_1836_axes_0 = const()[name = string("op_1836_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1836_cast_fp16 = expand_dims(axes = var_1836_axes_0, x = coreml_update_state_74)[name = string("op_1836_cast_fp16")];
            tensor<int32, [5]> k_exp_37_reps_0 = const()[name = string("k_exp_37_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_37_cast_fp16 = tile(reps = k_exp_37_reps_0, x = var_1836_cast_fp16)[name = string("k_exp_37_cast_fp16")];
            tensor<int32, [4]> var_1839 = const()[name = string("op_1839"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_39_cast_fp16 = reshape(shape = var_1839, x = k_exp_37_cast_fp16)[name = string("k_exp_39_cast_fp16")];
            tensor<int32, [1]> var_1841_axes_0 = const()[name = string("op_1841_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1841_cast_fp16 = expand_dims(axes = var_1841_axes_0, x = coreml_update_state_75)[name = string("op_1841_cast_fp16")];
            tensor<int32, [5]> v_exp_37_reps_0 = const()[name = string("v_exp_37_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_37_cast_fp16 = tile(reps = v_exp_37_reps_0, x = var_1841_cast_fp16)[name = string("v_exp_37_cast_fp16")];
            tensor<int32, [4]> var_1844 = const()[name = string("op_1844"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_39_cast_fp16 = reshape(shape = var_1844, x = v_exp_37_cast_fp16)[name = string("v_exp_39_cast_fp16")];
            bool var_1847_transpose_x_1 = const()[name = string("op_1847_transpose_x_1"), val = bool(false)];
            bool var_1847_transpose_y_1 = const()[name = string("op_1847_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_1847_cast_fp16 = matmul(transpose_x = var_1847_transpose_x_1, transpose_y = var_1847_transpose_y_1, x = q_19_cast_fp16, y = k_exp_39_cast_fp16)[name = string("op_1847_cast_fp16")];
            fp16 var_1848_to_fp16 = const()[name = string("op_1848_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_37_cast_fp16 = mul(x = var_1847_cast_fp16, y = var_1848_to_fp16)[name = string("attn_37_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_91_cast_fp16 = add(x = attn_37_cast_fp16, y = attention_mask)[name = string("input_91_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_39_cast_fp16 = softmax(axis = var_1723, x = input_91_cast_fp16)[name = string("attn_39_cast_fp16")];
            bool out_19_transpose_x_0 = const()[name = string("out_19_transpose_x_0"), val = bool(false)];
            bool out_19_transpose_y_0 = const()[name = string("out_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_19_cast_fp16 = matmul(transpose_x = out_19_transpose_x_0, transpose_y = out_19_transpose_y_0, x = attn_39_cast_fp16, y = v_exp_39_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [4]> var_1853_perm_0 = const()[name = string("op_1853_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_59_axis_0 = const()[name = string("concat_59_axis_0"), val = int32(0)];
            bool concat_59_interleave_0 = const()[name = string("concat_59_interleave_0"), val = bool(false)];
            int32 gather_55_cast_uint16_to_int32 = cast(dtype = gather_55_cast_uint16_to_int32_dtype_0, x = gather_55_cast_uint16)[name = string("cast_36")];
            tensor<int32, [3]> concat_59 = concat(axis = concat_59_axis_0, interleave = concat_59_interleave_0, values = (gather_54, gather_55_cast_uint16_to_int32, var_1723))[name = string("concat_59")];
            tensor<fp16, [1, ?, 16, 128]> var_1853_cast_fp16 = transpose(perm = var_1853_perm_0, x = out_19_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, ?, ?]> input_93_cast_fp16 = reshape(shape = concat_59, x = var_1853_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154929280))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157026496))))[name = string("layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_66_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_253_cast_fp16 = add(x = x_233_cast_fp16, y = linear_66_cast_fp16)[name = string("x_253_cast_fp16")];
            fp16 var_1722_promoted_3_to_fp16 = const()[name = string("op_1722_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_1861_cast_fp16 = pow(x = x_253_cast_fp16, y = var_1722_promoted_3_to_fp16)[name = string("op_1861_cast_fp16")];
            tensor<int32, [1]> var_1863_axes_0 = const()[name = string("op_1863_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1863_keep_dims_0 = const()[name = string("op_1863_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_1863_cast_fp16 = reduce_mean(axes = var_1863_axes_0, keep_dims = var_1863_keep_dims_0, x = var_1861_cast_fp16)[name = string("op_1863_cast_fp16")];
            fp16 var_1864_to_fp16 = const()[name = string("op_1864_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1865_cast_fp16 = add(x = var_1863_cast_fp16, y = var_1864_to_fp16)[name = string("op_1865_cast_fp16")];
            fp32 norm_79_epsilon_0 = const()[name = string("norm_79_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_79_cast_fp16 = rsqrt(epsilon = norm_79_epsilon_0, x = var_1865_cast_fp16)[name = string("norm_79_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_1867_cast_fp16 = mul(x = x_253_cast_fp16, y = norm_79_cast_fp16)[name = string("op_1867_cast_fp16")];
            tensor<fp16, [1024]> layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157157632)))];
            tensor<fp16, [1, ?, 1024]> var_1868_cast_fp16 = mul(x = var_1867_cast_fp16, y = layers_9_post_attention_layernorm_weight_to_fp16)[name = string("op_1868_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157159744))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160305536))))[name = string("layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = var_1868_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_1878_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_1878_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160502208))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163648000))))[name = string("layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_9_mlp_up_proj_weight_to_fp16_quantized, x = var_1868_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_99_cast_fp16 = mul(x = var_1878_cast_fp16, y = linear_68_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163844672))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166990464))))[name = string("layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_259_cast_fp16 = add(x = x_253_cast_fp16, y = linear_69_cast_fp16)[name = string("x_259_cast_fp16")];
            int32 var_1898 = const()[name = string("op_1898"), val = int32(-1)];
            fp16 var_1897_promoted_to_fp16 = const()[name = string("op_1897_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_1907_cast_fp16 = pow(x = x_259_cast_fp16, y = var_1897_promoted_to_fp16)[name = string("op_1907_cast_fp16")];
            tensor<int32, [1]> var_1909_axes_0 = const()[name = string("op_1909_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1909_keep_dims_0 = const()[name = string("op_1909_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_1909_cast_fp16 = reduce_mean(axes = var_1909_axes_0, keep_dims = var_1909_keep_dims_0, x = var_1907_cast_fp16)[name = string("op_1909_cast_fp16")];
            fp16 var_1910_to_fp16 = const()[name = string("op_1910_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_1911_cast_fp16 = add(x = var_1909_cast_fp16, y = var_1910_to_fp16)[name = string("op_1911_cast_fp16")];
            fp32 norm_81_epsilon_0 = const()[name = string("norm_81_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_81_cast_fp16 = rsqrt(epsilon = norm_81_epsilon_0, x = var_1911_cast_fp16)[name = string("norm_81_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_1913_cast_fp16 = mul(x = x_259_cast_fp16, y = norm_81_cast_fp16)[name = string("op_1913_cast_fp16")];
            tensor<fp16, [1024]> layers_10_input_layernorm_weight_to_fp16 = const()[name = string("layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167187136)))];
            tensor<fp16, [1, ?, 1024]> var_1914_cast_fp16 = mul(x = var_1913_cast_fp16, y = layers_10_input_layernorm_weight_to_fp16)[name = string("op_1914_cast_fp16")];
            tensor<int32, [3]> var_1926_shape_cast_fp16 = shape(x = var_1914_cast_fp16)[name = string("op_1926_shape_cast_fp16")];
            int32 gather_60 = const()[name = string("gather_60"), val = int32(1)];
            int32 gather_61_axis_0 = const()[name = string("gather_61_axis_0"), val = int32(0)];
            int32 gather_61_batch_dims_0 = const()[name = string("gather_61_batch_dims_0"), val = int32(0)];
            bool gather_61_validate_indices_0 = const()[name = string("gather_61_validate_indices_0"), val = bool(false)];
            string var_1926_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_1926_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_61_to_uint16 = const()[name = string("select_61_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_1926_shape_cast_fp16_to_uint16 = cast(dtype = var_1926_shape_cast_fp16_to_uint16_dtype_0, x = var_1926_shape_cast_fp16)[name = string("cast_35")];
            uint16 gather_61_cast_uint16 = gather(axis = gather_61_axis_0, batch_dims = gather_61_batch_dims_0, indices = select_61_to_uint16, validate_indices = gather_61_validate_indices_0, x = var_1926_shape_cast_fp16_to_uint16)[name = string("gather_61_cast_uint16")];
            string gather_61_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_61_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167189248))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169286464))))[name = string("layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = var_1914_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [4]> concat_60x = const()[name = string("concat_60x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_1931_cast_fp16 = reshape(shape = concat_60x, x = linear_70_cast_fp16)[name = string("op_1931_cast_fp16")];
            tensor<int32, [4]> x_265_perm_0 = const()[name = string("x_265_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169417600))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170466240))))[name = string("layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = var_1914_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_1936_cast_fp16 = reshape(shape = concat_61x, x = linear_71_cast_fp16)[name = string("op_1936_cast_fp16")];
            tensor<int32, [4]> x_269_perm_0 = const()[name = string("x_269_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170531840))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171580480))))[name = string("layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_72_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = var_1914_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> concat_62x = const()[name = string("concat_62x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_1941_cast_fp16 = reshape(shape = concat_62x, x = linear_72_cast_fp16)[name = string("op_1941_cast_fp16")];
            tensor<int32, [4]> v_21_perm_0 = const()[name = string("v_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1897_promoted_1_to_fp16 = const()[name = string("op_1897_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_265_cast_fp16 = transpose(perm = x_265_perm_0, x = var_1931_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 16, ?, 128]> var_1945_cast_fp16 = pow(x = x_265_cast_fp16, y = var_1897_promoted_1_to_fp16)[name = string("op_1945_cast_fp16")];
            tensor<int32, [1]> var_1947_axes_0 = const()[name = string("op_1947_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1947_keep_dims_0 = const()[name = string("op_1947_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_1947_cast_fp16 = reduce_mean(axes = var_1947_axes_0, keep_dims = var_1947_keep_dims_0, x = var_1945_cast_fp16)[name = string("op_1947_cast_fp16")];
            fp16 var_1948_to_fp16 = const()[name = string("op_1948_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_1949_cast_fp16 = add(x = var_1947_cast_fp16, y = var_1948_to_fp16)[name = string("op_1949_cast_fp16")];
            fp32 norm_83_epsilon_0 = const()[name = string("norm_83_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_83_cast_fp16 = rsqrt(epsilon = norm_83_epsilon_0, x = var_1949_cast_fp16)[name = string("norm_83_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_1951_cast_fp16 = mul(x = x_265_cast_fp16, y = norm_83_cast_fp16)[name = string("op_1951_cast_fp16")];
            tensor<fp16, [128]> layers_10_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_10_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171646080)))];
            tensor<fp16, [1, 16, ?, 128]> var_1952_cast_fp16 = mul(x = var_1951_cast_fp16, y = layers_10_self_attn_q_norm_weight_to_fp16)[name = string("op_1952_cast_fp16")];
            fp16 var_1897_promoted_2_to_fp16 = const()[name = string("op_1897_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_269_cast_fp16 = transpose(perm = x_269_perm_0, x = var_1936_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 8, ?, 128]> var_1956_cast_fp16 = pow(x = x_269_cast_fp16, y = var_1897_promoted_2_to_fp16)[name = string("op_1956_cast_fp16")];
            tensor<int32, [1]> var_1958_axes_0 = const()[name = string("op_1958_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1958_keep_dims_0 = const()[name = string("op_1958_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_1958_cast_fp16 = reduce_mean(axes = var_1958_axes_0, keep_dims = var_1958_keep_dims_0, x = var_1956_cast_fp16)[name = string("op_1958_cast_fp16")];
            fp16 var_1959_to_fp16 = const()[name = string("op_1959_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_1960_cast_fp16 = add(x = var_1958_cast_fp16, y = var_1959_to_fp16)[name = string("op_1960_cast_fp16")];
            fp32 norm_85_epsilon_0 = const()[name = string("norm_85_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_85_cast_fp16 = rsqrt(epsilon = norm_85_epsilon_0, x = var_1960_cast_fp16)[name = string("norm_85_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_1962_cast_fp16 = mul(x = x_269_cast_fp16, y = norm_85_cast_fp16)[name = string("op_1962_cast_fp16")];
            tensor<fp16, [128]> layers_10_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_10_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171646400)))];
            tensor<fp16, [1, 8, ?, 128]> var_1963_cast_fp16 = mul(x = var_1962_cast_fp16, y = layers_10_self_attn_k_norm_weight_to_fp16)[name = string("op_1963_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = var_1952_cast_fp16)[name = string("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = var_1952_cast_fp16)[name = string("x2_41_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1979_cast_fp16 = mul(x = x1_41_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1979_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1980_cast_fp16 = mul(x = x2_41_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1980_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1981_cast_fp16 = sub(x = var_1979_cast_fp16, y = var_1980_cast_fp16)[name = string("op_1981_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1982_cast_fp16 = mul(x = x2_41_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1982_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1983_cast_fp16 = mul(x = x1_41_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1983_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_1984_cast_fp16 = add(x = var_1982_cast_fp16, y = var_1983_cast_fp16)[name = string("op_1984_cast_fp16")];
            bool q_21_interleave_0 = const()[name = string("q_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_21_cast_fp16 = concat(axis = var_1898, interleave = q_21_interleave_0, values = (var_1981_cast_fp16, var_1984_cast_fp16))[name = string("q_21_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = var_1963_cast_fp16)[name = string("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = var_1963_cast_fp16)[name = string("x2_43_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2001_cast_fp16 = mul(x = x1_43_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2001_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2002_cast_fp16 = mul(x = x2_43_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2002_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2003_cast_fp16 = sub(x = var_2001_cast_fp16, y = var_2002_cast_fp16)[name = string("op_2003_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2004_cast_fp16 = mul(x = x2_43_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2004_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2005_cast_fp16 = mul(x = x1_43_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2005_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2006_cast_fp16 = add(x = var_2004_cast_fp16, y = var_2005_cast_fp16)[name = string("op_2006_cast_fp16")];
            bool k_21_interleave_0 = const()[name = string("k_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_21_cast_fp16 = concat(axis = var_1898, interleave = k_21_interleave_0, values = (var_2003_cast_fp16, var_2006_cast_fp16))[name = string("k_21_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_20 = read_state(input = k_cache_10)[name = string("read_state_20")];
            int32 k_cache_43_axis_0 = const()[name = string("k_cache_43_axis_0"), val = int32(2)];
            string k_cache_43_mode_0 = const()[name = string("k_cache_43_mode_0"), val = string("update")];
            bool k_cache_43_validate_indices_0 = const()[name = string("k_cache_43_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_43_cast_fp16 = scatter_along_axis(axis = k_cache_43_axis_0, data = read_state_20, indices = scatter_indices, mode = k_cache_43_mode_0, updates = k_21_cast_fp16, validate_indices = k_cache_43_validate_indices_0)[name = string("k_cache_43_cast_fp16")];
            write_state(data = k_cache_43_cast_fp16, input = k_cache_10)[name = string("coreml_update_state_20_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_76 = read_state(input = k_cache_10)[name = string("coreml_update_state_20")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_21 = read_state(input = v_cache_10)[name = string("read_state_21")];
            int32 v_cache_43_axis_0 = const()[name = string("v_cache_43_axis_0"), val = int32(2)];
            string v_cache_43_mode_0 = const()[name = string("v_cache_43_mode_0"), val = string("update")];
            bool v_cache_43_validate_indices_0 = const()[name = string("v_cache_43_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_21_cast_fp16 = transpose(perm = v_21_perm_0, x = var_1941_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_43_cast_fp16 = scatter_along_axis(axis = v_cache_43_axis_0, data = read_state_21, indices = scatter_indices, mode = v_cache_43_mode_0, updates = v_21_cast_fp16, validate_indices = v_cache_43_validate_indices_0)[name = string("v_cache_43_cast_fp16")];
            write_state(data = v_cache_43_cast_fp16, input = v_cache_10)[name = string("coreml_update_state_21_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_77 = read_state(input = v_cache_10)[name = string("coreml_update_state_21")];
            tensor<int32, [1]> var_2011_axes_0 = const()[name = string("op_2011_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2011_cast_fp16 = expand_dims(axes = var_2011_axes_0, x = coreml_update_state_76)[name = string("op_2011_cast_fp16")];
            tensor<int32, [5]> k_exp_41_reps_0 = const()[name = string("k_exp_41_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_41_cast_fp16 = tile(reps = k_exp_41_reps_0, x = var_2011_cast_fp16)[name = string("k_exp_41_cast_fp16")];
            tensor<int32, [4]> var_2014 = const()[name = string("op_2014"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_43_cast_fp16 = reshape(shape = var_2014, x = k_exp_41_cast_fp16)[name = string("k_exp_43_cast_fp16")];
            tensor<int32, [1]> var_2016_axes_0 = const()[name = string("op_2016_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2016_cast_fp16 = expand_dims(axes = var_2016_axes_0, x = coreml_update_state_77)[name = string("op_2016_cast_fp16")];
            tensor<int32, [5]> v_exp_41_reps_0 = const()[name = string("v_exp_41_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_41_cast_fp16 = tile(reps = v_exp_41_reps_0, x = var_2016_cast_fp16)[name = string("v_exp_41_cast_fp16")];
            tensor<int32, [4]> var_2019 = const()[name = string("op_2019"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_43_cast_fp16 = reshape(shape = var_2019, x = v_exp_41_cast_fp16)[name = string("v_exp_43_cast_fp16")];
            bool var_2022_transpose_x_1 = const()[name = string("op_2022_transpose_x_1"), val = bool(false)];
            bool var_2022_transpose_y_1 = const()[name = string("op_2022_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_2022_cast_fp16 = matmul(transpose_x = var_2022_transpose_x_1, transpose_y = var_2022_transpose_y_1, x = q_21_cast_fp16, y = k_exp_43_cast_fp16)[name = string("op_2022_cast_fp16")];
            fp16 var_2023_to_fp16 = const()[name = string("op_2023_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_41_cast_fp16 = mul(x = var_2022_cast_fp16, y = var_2023_to_fp16)[name = string("attn_41_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_101_cast_fp16 = add(x = attn_41_cast_fp16, y = attention_mask)[name = string("input_101_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_43_cast_fp16 = softmax(axis = var_1898, x = input_101_cast_fp16)[name = string("attn_43_cast_fp16")];
            bool out_21_transpose_x_0 = const()[name = string("out_21_transpose_x_0"), val = bool(false)];
            bool out_21_transpose_y_0 = const()[name = string("out_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_21_cast_fp16 = matmul(transpose_x = out_21_transpose_x_0, transpose_y = out_21_transpose_y_0, x = attn_43_cast_fp16, y = v_exp_43_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [4]> var_2028_perm_0 = const()[name = string("op_2028_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_65_axis_0 = const()[name = string("concat_65_axis_0"), val = int32(0)];
            bool concat_65_interleave_0 = const()[name = string("concat_65_interleave_0"), val = bool(false)];
            int32 gather_61_cast_uint16_to_int32 = cast(dtype = gather_61_cast_uint16_to_int32_dtype_0, x = gather_61_cast_uint16)[name = string("cast_34")];
            tensor<int32, [3]> concat_65 = concat(axis = concat_65_axis_0, interleave = concat_65_interleave_0, values = (gather_60, gather_61_cast_uint16_to_int32, var_1898))[name = string("concat_65")];
            tensor<fp16, [1, ?, 16, 128]> var_2028_cast_fp16 = transpose(perm = var_2028_perm_0, x = out_21_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, ?, ?]> input_103_cast_fp16 = reshape(shape = concat_65, x = var_2028_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171646720))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173743936))))[name = string("layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_279_cast_fp16 = add(x = x_259_cast_fp16, y = linear_73_cast_fp16)[name = string("x_279_cast_fp16")];
            fp16 var_1897_promoted_3_to_fp16 = const()[name = string("op_1897_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_2036_cast_fp16 = pow(x = x_279_cast_fp16, y = var_1897_promoted_3_to_fp16)[name = string("op_2036_cast_fp16")];
            tensor<int32, [1]> var_2038_axes_0 = const()[name = string("op_2038_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2038_keep_dims_0 = const()[name = string("op_2038_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_2038_cast_fp16 = reduce_mean(axes = var_2038_axes_0, keep_dims = var_2038_keep_dims_0, x = var_2036_cast_fp16)[name = string("op_2038_cast_fp16")];
            fp16 var_2039_to_fp16 = const()[name = string("op_2039_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2040_cast_fp16 = add(x = var_2038_cast_fp16, y = var_2039_to_fp16)[name = string("op_2040_cast_fp16")];
            fp32 norm_87_epsilon_0 = const()[name = string("norm_87_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_87_cast_fp16 = rsqrt(epsilon = norm_87_epsilon_0, x = var_2040_cast_fp16)[name = string("norm_87_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_2042_cast_fp16 = mul(x = x_279_cast_fp16, y = norm_87_cast_fp16)[name = string("op_2042_cast_fp16")];
            tensor<fp16, [1024]> layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173875072)))];
            tensor<fp16, [1, ?, 1024]> var_2043_cast_fp16 = mul(x = var_2042_cast_fp16, y = layers_10_post_attention_layernorm_weight_to_fp16)[name = string("op_2043_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173877184))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177022976))))[name = string("layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = var_2043_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_2053_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_2053_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177219648))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180365440))))[name = string("layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_10_mlp_up_proj_weight_to_fp16_quantized, x = var_2043_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_109_cast_fp16 = mul(x = var_2053_cast_fp16, y = linear_75_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180562112))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183707904))))[name = string("layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_76_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_285_cast_fp16 = add(x = x_279_cast_fp16, y = linear_76_cast_fp16)[name = string("x_285_cast_fp16")];
            int32 var_2073 = const()[name = string("op_2073"), val = int32(-1)];
            fp16 var_2072_promoted_to_fp16 = const()[name = string("op_2072_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_2082_cast_fp16 = pow(x = x_285_cast_fp16, y = var_2072_promoted_to_fp16)[name = string("op_2082_cast_fp16")];
            tensor<int32, [1]> var_2084_axes_0 = const()[name = string("op_2084_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2084_keep_dims_0 = const()[name = string("op_2084_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_2084_cast_fp16 = reduce_mean(axes = var_2084_axes_0, keep_dims = var_2084_keep_dims_0, x = var_2082_cast_fp16)[name = string("op_2084_cast_fp16")];
            fp16 var_2085_to_fp16 = const()[name = string("op_2085_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2086_cast_fp16 = add(x = var_2084_cast_fp16, y = var_2085_to_fp16)[name = string("op_2086_cast_fp16")];
            fp32 norm_89_epsilon_0 = const()[name = string("norm_89_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_89_cast_fp16 = rsqrt(epsilon = norm_89_epsilon_0, x = var_2086_cast_fp16)[name = string("norm_89_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_2088_cast_fp16 = mul(x = x_285_cast_fp16, y = norm_89_cast_fp16)[name = string("op_2088_cast_fp16")];
            tensor<fp16, [1024]> layers_11_input_layernorm_weight_to_fp16 = const()[name = string("layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183904576)))];
            tensor<fp16, [1, ?, 1024]> var_2089_cast_fp16 = mul(x = var_2088_cast_fp16, y = layers_11_input_layernorm_weight_to_fp16)[name = string("op_2089_cast_fp16")];
            tensor<int32, [3]> var_2101_shape_cast_fp16 = shape(x = var_2089_cast_fp16)[name = string("op_2101_shape_cast_fp16")];
            int32 gather_66 = const()[name = string("gather_66"), val = int32(1)];
            int32 gather_67_axis_0 = const()[name = string("gather_67_axis_0"), val = int32(0)];
            int32 gather_67_batch_dims_0 = const()[name = string("gather_67_batch_dims_0"), val = int32(0)];
            bool gather_67_validate_indices_0 = const()[name = string("gather_67_validate_indices_0"), val = bool(false)];
            string var_2101_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2101_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_67_to_uint16 = const()[name = string("select_67_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_2101_shape_cast_fp16_to_uint16 = cast(dtype = var_2101_shape_cast_fp16_to_uint16_dtype_0, x = var_2101_shape_cast_fp16)[name = string("cast_33")];
            uint16 gather_67_cast_uint16 = gather(axis = gather_67_axis_0, batch_dims = gather_67_batch_dims_0, indices = select_67_to_uint16, validate_indices = gather_67_validate_indices_0, x = var_2101_shape_cast_fp16_to_uint16)[name = string("gather_67_cast_uint16")];
            string gather_67_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_67_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183906688))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186003904))))[name = string("layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_77_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = var_2089_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> concat_66x = const()[name = string("concat_66x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_2106_cast_fp16 = reshape(shape = concat_66x, x = linear_77_cast_fp16)[name = string("op_2106_cast_fp16")];
            tensor<int32, [4]> x_291_perm_0 = const()[name = string("x_291_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186135040))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187183680))))[name = string("layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_78_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = var_2089_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> concat_67x = const()[name = string("concat_67x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_2111_cast_fp16 = reshape(shape = concat_67x, x = linear_78_cast_fp16)[name = string("op_2111_cast_fp16")];
            tensor<int32, [4]> x_295_perm_0 = const()[name = string("x_295_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187249280))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188297920))))[name = string("layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = var_2089_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> concat_68x = const()[name = string("concat_68x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_2116_cast_fp16 = reshape(shape = concat_68x, x = linear_79_cast_fp16)[name = string("op_2116_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2072_promoted_1_to_fp16 = const()[name = string("op_2072_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_291_cast_fp16 = transpose(perm = x_291_perm_0, x = var_2106_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 16, ?, 128]> var_2120_cast_fp16 = pow(x = x_291_cast_fp16, y = var_2072_promoted_1_to_fp16)[name = string("op_2120_cast_fp16")];
            tensor<int32, [1]> var_2122_axes_0 = const()[name = string("op_2122_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2122_keep_dims_0 = const()[name = string("op_2122_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_2122_cast_fp16 = reduce_mean(axes = var_2122_axes_0, keep_dims = var_2122_keep_dims_0, x = var_2120_cast_fp16)[name = string("op_2122_cast_fp16")];
            fp16 var_2123_to_fp16 = const()[name = string("op_2123_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_2124_cast_fp16 = add(x = var_2122_cast_fp16, y = var_2123_to_fp16)[name = string("op_2124_cast_fp16")];
            fp32 norm_91_epsilon_0 = const()[name = string("norm_91_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_91_cast_fp16 = rsqrt(epsilon = norm_91_epsilon_0, x = var_2124_cast_fp16)[name = string("norm_91_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_2126_cast_fp16 = mul(x = x_291_cast_fp16, y = norm_91_cast_fp16)[name = string("op_2126_cast_fp16")];
            tensor<fp16, [128]> layers_11_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_11_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188363520)))];
            tensor<fp16, [1, 16, ?, 128]> var_2127_cast_fp16 = mul(x = var_2126_cast_fp16, y = layers_11_self_attn_q_norm_weight_to_fp16)[name = string("op_2127_cast_fp16")];
            fp16 var_2072_promoted_2_to_fp16 = const()[name = string("op_2072_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_295_cast_fp16 = transpose(perm = x_295_perm_0, x = var_2111_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 8, ?, 128]> var_2131_cast_fp16 = pow(x = x_295_cast_fp16, y = var_2072_promoted_2_to_fp16)[name = string("op_2131_cast_fp16")];
            tensor<int32, [1]> var_2133_axes_0 = const()[name = string("op_2133_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2133_keep_dims_0 = const()[name = string("op_2133_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_2133_cast_fp16 = reduce_mean(axes = var_2133_axes_0, keep_dims = var_2133_keep_dims_0, x = var_2131_cast_fp16)[name = string("op_2133_cast_fp16")];
            fp16 var_2134_to_fp16 = const()[name = string("op_2134_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_2135_cast_fp16 = add(x = var_2133_cast_fp16, y = var_2134_to_fp16)[name = string("op_2135_cast_fp16")];
            fp32 norm_93_epsilon_0 = const()[name = string("norm_93_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_93_cast_fp16 = rsqrt(epsilon = norm_93_epsilon_0, x = var_2135_cast_fp16)[name = string("norm_93_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_2137_cast_fp16 = mul(x = x_295_cast_fp16, y = norm_93_cast_fp16)[name = string("op_2137_cast_fp16")];
            tensor<fp16, [128]> layers_11_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_11_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188363840)))];
            tensor<fp16, [1, 8, ?, 128]> var_2138_cast_fp16 = mul(x = var_2137_cast_fp16, y = layers_11_self_attn_k_norm_weight_to_fp16)[name = string("op_2138_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = var_2127_cast_fp16)[name = string("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = var_2127_cast_fp16)[name = string("x2_45_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2154_cast_fp16 = mul(x = x1_45_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2154_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2155_cast_fp16 = mul(x = x2_45_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2155_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2156_cast_fp16 = sub(x = var_2154_cast_fp16, y = var_2155_cast_fp16)[name = string("op_2156_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2157_cast_fp16 = mul(x = x2_45_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2157_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2158_cast_fp16 = mul(x = x1_45_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2158_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2159_cast_fp16 = add(x = var_2157_cast_fp16, y = var_2158_cast_fp16)[name = string("op_2159_cast_fp16")];
            bool q_23_interleave_0 = const()[name = string("q_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_23_cast_fp16 = concat(axis = var_2073, interleave = q_23_interleave_0, values = (var_2156_cast_fp16, var_2159_cast_fp16))[name = string("q_23_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = var_2138_cast_fp16)[name = string("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = var_2138_cast_fp16)[name = string("x2_47_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2176_cast_fp16 = mul(x = x1_47_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2176_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2177_cast_fp16 = mul(x = x2_47_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2177_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2178_cast_fp16 = sub(x = var_2176_cast_fp16, y = var_2177_cast_fp16)[name = string("op_2178_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2179_cast_fp16 = mul(x = x2_47_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2179_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2180_cast_fp16 = mul(x = x1_47_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2180_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2181_cast_fp16 = add(x = var_2179_cast_fp16, y = var_2180_cast_fp16)[name = string("op_2181_cast_fp16")];
            bool k_23_interleave_0 = const()[name = string("k_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_23_cast_fp16 = concat(axis = var_2073, interleave = k_23_interleave_0, values = (var_2178_cast_fp16, var_2181_cast_fp16))[name = string("k_23_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_22 = read_state(input = k_cache_11)[name = string("read_state_22")];
            int32 k_cache_47_axis_0 = const()[name = string("k_cache_47_axis_0"), val = int32(2)];
            string k_cache_47_mode_0 = const()[name = string("k_cache_47_mode_0"), val = string("update")];
            bool k_cache_47_validate_indices_0 = const()[name = string("k_cache_47_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_47_cast_fp16 = scatter_along_axis(axis = k_cache_47_axis_0, data = read_state_22, indices = scatter_indices, mode = k_cache_47_mode_0, updates = k_23_cast_fp16, validate_indices = k_cache_47_validate_indices_0)[name = string("k_cache_47_cast_fp16")];
            write_state(data = k_cache_47_cast_fp16, input = k_cache_11)[name = string("coreml_update_state_22_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_78 = read_state(input = k_cache_11)[name = string("coreml_update_state_22")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_23 = read_state(input = v_cache_11)[name = string("read_state_23")];
            int32 v_cache_47_axis_0 = const()[name = string("v_cache_47_axis_0"), val = int32(2)];
            string v_cache_47_mode_0 = const()[name = string("v_cache_47_mode_0"), val = string("update")];
            bool v_cache_47_validate_indices_0 = const()[name = string("v_cache_47_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = var_2116_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_47_cast_fp16 = scatter_along_axis(axis = v_cache_47_axis_0, data = read_state_23, indices = scatter_indices, mode = v_cache_47_mode_0, updates = v_23_cast_fp16, validate_indices = v_cache_47_validate_indices_0)[name = string("v_cache_47_cast_fp16")];
            write_state(data = v_cache_47_cast_fp16, input = v_cache_11)[name = string("coreml_update_state_23_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_79 = read_state(input = v_cache_11)[name = string("coreml_update_state_23")];
            tensor<int32, [1]> var_2186_axes_0 = const()[name = string("op_2186_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2186_cast_fp16 = expand_dims(axes = var_2186_axes_0, x = coreml_update_state_78)[name = string("op_2186_cast_fp16")];
            tensor<int32, [5]> k_exp_45_reps_0 = const()[name = string("k_exp_45_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_45_cast_fp16 = tile(reps = k_exp_45_reps_0, x = var_2186_cast_fp16)[name = string("k_exp_45_cast_fp16")];
            tensor<int32, [4]> var_2189 = const()[name = string("op_2189"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_47_cast_fp16 = reshape(shape = var_2189, x = k_exp_45_cast_fp16)[name = string("k_exp_47_cast_fp16")];
            tensor<int32, [1]> var_2191_axes_0 = const()[name = string("op_2191_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2191_cast_fp16 = expand_dims(axes = var_2191_axes_0, x = coreml_update_state_79)[name = string("op_2191_cast_fp16")];
            tensor<int32, [5]> v_exp_45_reps_0 = const()[name = string("v_exp_45_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_45_cast_fp16 = tile(reps = v_exp_45_reps_0, x = var_2191_cast_fp16)[name = string("v_exp_45_cast_fp16")];
            tensor<int32, [4]> var_2194 = const()[name = string("op_2194"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_47_cast_fp16 = reshape(shape = var_2194, x = v_exp_45_cast_fp16)[name = string("v_exp_47_cast_fp16")];
            bool var_2197_transpose_x_1 = const()[name = string("op_2197_transpose_x_1"), val = bool(false)];
            bool var_2197_transpose_y_1 = const()[name = string("op_2197_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_2197_cast_fp16 = matmul(transpose_x = var_2197_transpose_x_1, transpose_y = var_2197_transpose_y_1, x = q_23_cast_fp16, y = k_exp_47_cast_fp16)[name = string("op_2197_cast_fp16")];
            fp16 var_2198_to_fp16 = const()[name = string("op_2198_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_45_cast_fp16 = mul(x = var_2197_cast_fp16, y = var_2198_to_fp16)[name = string("attn_45_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_111_cast_fp16 = add(x = attn_45_cast_fp16, y = attention_mask)[name = string("input_111_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_47_cast_fp16 = softmax(axis = var_2073, x = input_111_cast_fp16)[name = string("attn_47_cast_fp16")];
            bool out_23_transpose_x_0 = const()[name = string("out_23_transpose_x_0"), val = bool(false)];
            bool out_23_transpose_y_0 = const()[name = string("out_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_23_cast_fp16 = matmul(transpose_x = out_23_transpose_x_0, transpose_y = out_23_transpose_y_0, x = attn_47_cast_fp16, y = v_exp_47_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [4]> var_2203_perm_0 = const()[name = string("op_2203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_71_axis_0 = const()[name = string("concat_71_axis_0"), val = int32(0)];
            bool concat_71_interleave_0 = const()[name = string("concat_71_interleave_0"), val = bool(false)];
            int32 gather_67_cast_uint16_to_int32 = cast(dtype = gather_67_cast_uint16_to_int32_dtype_0, x = gather_67_cast_uint16)[name = string("cast_32")];
            tensor<int32, [3]> concat_71 = concat(axis = concat_71_axis_0, interleave = concat_71_interleave_0, values = (gather_66, gather_67_cast_uint16_to_int32, var_2073))[name = string("concat_71")];
            tensor<fp16, [1, ?, 16, 128]> var_2203_cast_fp16 = transpose(perm = var_2203_perm_0, x = out_23_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, ?, ?]> input_113_cast_fp16 = reshape(shape = concat_71, x = var_2203_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188364160))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190461376))))[name = string("layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_80_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_305_cast_fp16 = add(x = x_285_cast_fp16, y = linear_80_cast_fp16)[name = string("x_305_cast_fp16")];
            fp16 var_2072_promoted_3_to_fp16 = const()[name = string("op_2072_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_2211_cast_fp16 = pow(x = x_305_cast_fp16, y = var_2072_promoted_3_to_fp16)[name = string("op_2211_cast_fp16")];
            tensor<int32, [1]> var_2213_axes_0 = const()[name = string("op_2213_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2213_keep_dims_0 = const()[name = string("op_2213_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_2213_cast_fp16 = reduce_mean(axes = var_2213_axes_0, keep_dims = var_2213_keep_dims_0, x = var_2211_cast_fp16)[name = string("op_2213_cast_fp16")];
            fp16 var_2214_to_fp16 = const()[name = string("op_2214_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2215_cast_fp16 = add(x = var_2213_cast_fp16, y = var_2214_to_fp16)[name = string("op_2215_cast_fp16")];
            fp32 norm_95_epsilon_0 = const()[name = string("norm_95_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_95_cast_fp16 = rsqrt(epsilon = norm_95_epsilon_0, x = var_2215_cast_fp16)[name = string("norm_95_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_2217_cast_fp16 = mul(x = x_305_cast_fp16, y = norm_95_cast_fp16)[name = string("op_2217_cast_fp16")];
            tensor<fp16, [1024]> layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190592512)))];
            tensor<fp16, [1, ?, 1024]> var_2218_cast_fp16 = mul(x = var_2217_cast_fp16, y = layers_11_post_attention_layernorm_weight_to_fp16)[name = string("op_2218_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190594624))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193740416))))[name = string("layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = var_2218_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_2228_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_2228_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193937088))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197082880))))[name = string("layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_11_mlp_up_proj_weight_to_fp16_quantized, x = var_2218_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_119_cast_fp16 = mul(x = var_2228_cast_fp16, y = linear_82_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197279552))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200425344))))[name = string("layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_311_cast_fp16 = add(x = x_305_cast_fp16, y = linear_83_cast_fp16)[name = string("x_311_cast_fp16")];
            int32 var_2248 = const()[name = string("op_2248"), val = int32(-1)];
            fp16 var_2247_promoted_to_fp16 = const()[name = string("op_2247_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_2257_cast_fp16 = pow(x = x_311_cast_fp16, y = var_2247_promoted_to_fp16)[name = string("op_2257_cast_fp16")];
            tensor<int32, [1]> var_2259_axes_0 = const()[name = string("op_2259_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2259_keep_dims_0 = const()[name = string("op_2259_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_2259_cast_fp16 = reduce_mean(axes = var_2259_axes_0, keep_dims = var_2259_keep_dims_0, x = var_2257_cast_fp16)[name = string("op_2259_cast_fp16")];
            fp16 var_2260_to_fp16 = const()[name = string("op_2260_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2261_cast_fp16 = add(x = var_2259_cast_fp16, y = var_2260_to_fp16)[name = string("op_2261_cast_fp16")];
            fp32 norm_97_epsilon_0 = const()[name = string("norm_97_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_97_cast_fp16 = rsqrt(epsilon = norm_97_epsilon_0, x = var_2261_cast_fp16)[name = string("norm_97_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_2263_cast_fp16 = mul(x = x_311_cast_fp16, y = norm_97_cast_fp16)[name = string("op_2263_cast_fp16")];
            tensor<fp16, [1024]> layers_12_input_layernorm_weight_to_fp16 = const()[name = string("layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200622016)))];
            tensor<fp16, [1, ?, 1024]> var_2264_cast_fp16 = mul(x = var_2263_cast_fp16, y = layers_12_input_layernorm_weight_to_fp16)[name = string("op_2264_cast_fp16")];
            tensor<int32, [3]> var_2276_shape_cast_fp16 = shape(x = var_2264_cast_fp16)[name = string("op_2276_shape_cast_fp16")];
            int32 gather_72 = const()[name = string("gather_72"), val = int32(1)];
            int32 gather_73_axis_0 = const()[name = string("gather_73_axis_0"), val = int32(0)];
            int32 gather_73_batch_dims_0 = const()[name = string("gather_73_batch_dims_0"), val = int32(0)];
            bool gather_73_validate_indices_0 = const()[name = string("gather_73_validate_indices_0"), val = bool(false)];
            string var_2276_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2276_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_73_to_uint16 = const()[name = string("select_73_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_2276_shape_cast_fp16_to_uint16 = cast(dtype = var_2276_shape_cast_fp16_to_uint16_dtype_0, x = var_2276_shape_cast_fp16)[name = string("cast_31")];
            uint16 gather_73_cast_uint16 = gather(axis = gather_73_axis_0, batch_dims = gather_73_batch_dims_0, indices = select_73_to_uint16, validate_indices = gather_73_validate_indices_0, x = var_2276_shape_cast_fp16_to_uint16)[name = string("gather_73_cast_uint16")];
            string gather_73_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_73_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200624128))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202721344))))[name = string("layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = var_2264_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> concat_72x = const()[name = string("concat_72x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_2281_cast_fp16 = reshape(shape = concat_72x, x = linear_84_cast_fp16)[name = string("op_2281_cast_fp16")];
            tensor<int32, [4]> x_317_perm_0 = const()[name = string("x_317_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202852480))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203901120))))[name = string("layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = var_2264_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_2286_cast_fp16 = reshape(shape = concat_73x, x = linear_85_cast_fp16)[name = string("op_2286_cast_fp16")];
            tensor<int32, [4]> x_321_perm_0 = const()[name = string("x_321_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203966720))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205015360))))[name = string("layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_86_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = var_2264_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> concat_74x = const()[name = string("concat_74x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_2291_cast_fp16 = reshape(shape = concat_74x, x = linear_86_cast_fp16)[name = string("op_2291_cast_fp16")];
            tensor<int32, [4]> v_25_perm_0 = const()[name = string("v_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2247_promoted_1_to_fp16 = const()[name = string("op_2247_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_317_cast_fp16 = transpose(perm = x_317_perm_0, x = var_2281_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 16, ?, 128]> var_2295_cast_fp16 = pow(x = x_317_cast_fp16, y = var_2247_promoted_1_to_fp16)[name = string("op_2295_cast_fp16")];
            tensor<int32, [1]> var_2297_axes_0 = const()[name = string("op_2297_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2297_keep_dims_0 = const()[name = string("op_2297_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_2297_cast_fp16 = reduce_mean(axes = var_2297_axes_0, keep_dims = var_2297_keep_dims_0, x = var_2295_cast_fp16)[name = string("op_2297_cast_fp16")];
            fp16 var_2298_to_fp16 = const()[name = string("op_2298_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_2299_cast_fp16 = add(x = var_2297_cast_fp16, y = var_2298_to_fp16)[name = string("op_2299_cast_fp16")];
            fp32 norm_99_epsilon_0 = const()[name = string("norm_99_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_99_cast_fp16 = rsqrt(epsilon = norm_99_epsilon_0, x = var_2299_cast_fp16)[name = string("norm_99_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_2301_cast_fp16 = mul(x = x_317_cast_fp16, y = norm_99_cast_fp16)[name = string("op_2301_cast_fp16")];
            tensor<fp16, [128]> layers_12_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_12_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205080960)))];
            tensor<fp16, [1, 16, ?, 128]> var_2302_cast_fp16 = mul(x = var_2301_cast_fp16, y = layers_12_self_attn_q_norm_weight_to_fp16)[name = string("op_2302_cast_fp16")];
            fp16 var_2247_promoted_2_to_fp16 = const()[name = string("op_2247_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_321_cast_fp16 = transpose(perm = x_321_perm_0, x = var_2286_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 8, ?, 128]> var_2306_cast_fp16 = pow(x = x_321_cast_fp16, y = var_2247_promoted_2_to_fp16)[name = string("op_2306_cast_fp16")];
            tensor<int32, [1]> var_2308_axes_0 = const()[name = string("op_2308_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2308_keep_dims_0 = const()[name = string("op_2308_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_2308_cast_fp16 = reduce_mean(axes = var_2308_axes_0, keep_dims = var_2308_keep_dims_0, x = var_2306_cast_fp16)[name = string("op_2308_cast_fp16")];
            fp16 var_2309_to_fp16 = const()[name = string("op_2309_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_2310_cast_fp16 = add(x = var_2308_cast_fp16, y = var_2309_to_fp16)[name = string("op_2310_cast_fp16")];
            fp32 norm_101_epsilon_0 = const()[name = string("norm_101_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_101_cast_fp16 = rsqrt(epsilon = norm_101_epsilon_0, x = var_2310_cast_fp16)[name = string("norm_101_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_2312_cast_fp16 = mul(x = x_321_cast_fp16, y = norm_101_cast_fp16)[name = string("op_2312_cast_fp16")];
            tensor<fp16, [128]> layers_12_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_12_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205081280)))];
            tensor<fp16, [1, 8, ?, 128]> var_2313_cast_fp16 = mul(x = var_2312_cast_fp16, y = layers_12_self_attn_k_norm_weight_to_fp16)[name = string("op_2313_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = var_2302_cast_fp16)[name = string("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = var_2302_cast_fp16)[name = string("x2_49_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2329_cast_fp16 = mul(x = x1_49_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2329_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2330_cast_fp16 = mul(x = x2_49_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2330_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2331_cast_fp16 = sub(x = var_2329_cast_fp16, y = var_2330_cast_fp16)[name = string("op_2331_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2332_cast_fp16 = mul(x = x2_49_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2332_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2333_cast_fp16 = mul(x = x1_49_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2333_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2334_cast_fp16 = add(x = var_2332_cast_fp16, y = var_2333_cast_fp16)[name = string("op_2334_cast_fp16")];
            bool q_25_interleave_0 = const()[name = string("q_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_25_cast_fp16 = concat(axis = var_2248, interleave = q_25_interleave_0, values = (var_2331_cast_fp16, var_2334_cast_fp16))[name = string("q_25_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = var_2313_cast_fp16)[name = string("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = var_2313_cast_fp16)[name = string("x2_51_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2351_cast_fp16 = mul(x = x1_51_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2351_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2352_cast_fp16 = mul(x = x2_51_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2352_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2353_cast_fp16 = sub(x = var_2351_cast_fp16, y = var_2352_cast_fp16)[name = string("op_2353_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2354_cast_fp16 = mul(x = x2_51_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2354_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2355_cast_fp16 = mul(x = x1_51_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2355_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2356_cast_fp16 = add(x = var_2354_cast_fp16, y = var_2355_cast_fp16)[name = string("op_2356_cast_fp16")];
            bool k_25_interleave_0 = const()[name = string("k_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_25_cast_fp16 = concat(axis = var_2248, interleave = k_25_interleave_0, values = (var_2353_cast_fp16, var_2356_cast_fp16))[name = string("k_25_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_24 = read_state(input = k_cache_12)[name = string("read_state_24")];
            int32 k_cache_51_axis_0 = const()[name = string("k_cache_51_axis_0"), val = int32(2)];
            string k_cache_51_mode_0 = const()[name = string("k_cache_51_mode_0"), val = string("update")];
            bool k_cache_51_validate_indices_0 = const()[name = string("k_cache_51_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_51_cast_fp16 = scatter_along_axis(axis = k_cache_51_axis_0, data = read_state_24, indices = scatter_indices, mode = k_cache_51_mode_0, updates = k_25_cast_fp16, validate_indices = k_cache_51_validate_indices_0)[name = string("k_cache_51_cast_fp16")];
            write_state(data = k_cache_51_cast_fp16, input = k_cache_12)[name = string("coreml_update_state_24_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_80 = read_state(input = k_cache_12)[name = string("coreml_update_state_24")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_25 = read_state(input = v_cache_12)[name = string("read_state_25")];
            int32 v_cache_51_axis_0 = const()[name = string("v_cache_51_axis_0"), val = int32(2)];
            string v_cache_51_mode_0 = const()[name = string("v_cache_51_mode_0"), val = string("update")];
            bool v_cache_51_validate_indices_0 = const()[name = string("v_cache_51_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_25_cast_fp16 = transpose(perm = v_25_perm_0, x = var_2291_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_51_cast_fp16 = scatter_along_axis(axis = v_cache_51_axis_0, data = read_state_25, indices = scatter_indices, mode = v_cache_51_mode_0, updates = v_25_cast_fp16, validate_indices = v_cache_51_validate_indices_0)[name = string("v_cache_51_cast_fp16")];
            write_state(data = v_cache_51_cast_fp16, input = v_cache_12)[name = string("coreml_update_state_25_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_81 = read_state(input = v_cache_12)[name = string("coreml_update_state_25")];
            tensor<int32, [1]> var_2361_axes_0 = const()[name = string("op_2361_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2361_cast_fp16 = expand_dims(axes = var_2361_axes_0, x = coreml_update_state_80)[name = string("op_2361_cast_fp16")];
            tensor<int32, [5]> k_exp_49_reps_0 = const()[name = string("k_exp_49_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_49_cast_fp16 = tile(reps = k_exp_49_reps_0, x = var_2361_cast_fp16)[name = string("k_exp_49_cast_fp16")];
            tensor<int32, [4]> var_2364 = const()[name = string("op_2364"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_51_cast_fp16 = reshape(shape = var_2364, x = k_exp_49_cast_fp16)[name = string("k_exp_51_cast_fp16")];
            tensor<int32, [1]> var_2366_axes_0 = const()[name = string("op_2366_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2366_cast_fp16 = expand_dims(axes = var_2366_axes_0, x = coreml_update_state_81)[name = string("op_2366_cast_fp16")];
            tensor<int32, [5]> v_exp_49_reps_0 = const()[name = string("v_exp_49_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_49_cast_fp16 = tile(reps = v_exp_49_reps_0, x = var_2366_cast_fp16)[name = string("v_exp_49_cast_fp16")];
            tensor<int32, [4]> var_2369 = const()[name = string("op_2369"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_51_cast_fp16 = reshape(shape = var_2369, x = v_exp_49_cast_fp16)[name = string("v_exp_51_cast_fp16")];
            bool var_2372_transpose_x_1 = const()[name = string("op_2372_transpose_x_1"), val = bool(false)];
            bool var_2372_transpose_y_1 = const()[name = string("op_2372_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_2372_cast_fp16 = matmul(transpose_x = var_2372_transpose_x_1, transpose_y = var_2372_transpose_y_1, x = q_25_cast_fp16, y = k_exp_51_cast_fp16)[name = string("op_2372_cast_fp16")];
            fp16 var_2373_to_fp16 = const()[name = string("op_2373_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_49_cast_fp16 = mul(x = var_2372_cast_fp16, y = var_2373_to_fp16)[name = string("attn_49_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_121_cast_fp16 = add(x = attn_49_cast_fp16, y = attention_mask)[name = string("input_121_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_51_cast_fp16 = softmax(axis = var_2248, x = input_121_cast_fp16)[name = string("attn_51_cast_fp16")];
            bool out_25_transpose_x_0 = const()[name = string("out_25_transpose_x_0"), val = bool(false)];
            bool out_25_transpose_y_0 = const()[name = string("out_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_25_cast_fp16 = matmul(transpose_x = out_25_transpose_x_0, transpose_y = out_25_transpose_y_0, x = attn_51_cast_fp16, y = v_exp_51_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [4]> var_2378_perm_0 = const()[name = string("op_2378_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_77_axis_0 = const()[name = string("concat_77_axis_0"), val = int32(0)];
            bool concat_77_interleave_0 = const()[name = string("concat_77_interleave_0"), val = bool(false)];
            int32 gather_73_cast_uint16_to_int32 = cast(dtype = gather_73_cast_uint16_to_int32_dtype_0, x = gather_73_cast_uint16)[name = string("cast_30")];
            tensor<int32, [3]> concat_77 = concat(axis = concat_77_axis_0, interleave = concat_77_interleave_0, values = (gather_72, gather_73_cast_uint16_to_int32, var_2248))[name = string("concat_77")];
            tensor<fp16, [1, ?, 16, 128]> var_2378_cast_fp16 = transpose(perm = var_2378_perm_0, x = out_25_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, ?, ?]> input_123_cast_fp16 = reshape(shape = concat_77, x = var_2378_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205081600))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207178816))))[name = string("layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_331_cast_fp16 = add(x = x_311_cast_fp16, y = linear_87_cast_fp16)[name = string("x_331_cast_fp16")];
            fp16 var_2247_promoted_3_to_fp16 = const()[name = string("op_2247_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_2386_cast_fp16 = pow(x = x_331_cast_fp16, y = var_2247_promoted_3_to_fp16)[name = string("op_2386_cast_fp16")];
            tensor<int32, [1]> var_2388_axes_0 = const()[name = string("op_2388_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2388_keep_dims_0 = const()[name = string("op_2388_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_2388_cast_fp16 = reduce_mean(axes = var_2388_axes_0, keep_dims = var_2388_keep_dims_0, x = var_2386_cast_fp16)[name = string("op_2388_cast_fp16")];
            fp16 var_2389_to_fp16 = const()[name = string("op_2389_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2390_cast_fp16 = add(x = var_2388_cast_fp16, y = var_2389_to_fp16)[name = string("op_2390_cast_fp16")];
            fp32 norm_103_epsilon_0 = const()[name = string("norm_103_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_103_cast_fp16 = rsqrt(epsilon = norm_103_epsilon_0, x = var_2390_cast_fp16)[name = string("norm_103_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_2392_cast_fp16 = mul(x = x_331_cast_fp16, y = norm_103_cast_fp16)[name = string("op_2392_cast_fp16")];
            tensor<fp16, [1024]> layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207309952)))];
            tensor<fp16, [1, ?, 1024]> var_2393_cast_fp16 = mul(x = var_2392_cast_fp16, y = layers_12_post_attention_layernorm_weight_to_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207312064))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210457856))))[name = string("layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = var_2393_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_2403_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_2403_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210654528))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213800320))))[name = string("layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_12_mlp_up_proj_weight_to_fp16_quantized, x = var_2393_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_129_cast_fp16 = mul(x = var_2403_cast_fp16, y = linear_89_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213996992))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217142784))))[name = string("layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_90_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_129_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_337_cast_fp16 = add(x = x_331_cast_fp16, y = linear_90_cast_fp16)[name = string("x_337_cast_fp16")];
            int32 var_2423 = const()[name = string("op_2423"), val = int32(-1)];
            fp16 var_2422_promoted_to_fp16 = const()[name = string("op_2422_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_2432_cast_fp16 = pow(x = x_337_cast_fp16, y = var_2422_promoted_to_fp16)[name = string("op_2432_cast_fp16")];
            tensor<int32, [1]> var_2434_axes_0 = const()[name = string("op_2434_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2434_keep_dims_0 = const()[name = string("op_2434_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_2434_cast_fp16 = reduce_mean(axes = var_2434_axes_0, keep_dims = var_2434_keep_dims_0, x = var_2432_cast_fp16)[name = string("op_2434_cast_fp16")];
            fp16 var_2435_to_fp16 = const()[name = string("op_2435_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2436_cast_fp16 = add(x = var_2434_cast_fp16, y = var_2435_to_fp16)[name = string("op_2436_cast_fp16")];
            fp32 norm_105_epsilon_0 = const()[name = string("norm_105_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_105_cast_fp16 = rsqrt(epsilon = norm_105_epsilon_0, x = var_2436_cast_fp16)[name = string("norm_105_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_2438_cast_fp16 = mul(x = x_337_cast_fp16, y = norm_105_cast_fp16)[name = string("op_2438_cast_fp16")];
            tensor<fp16, [1024]> layers_13_input_layernorm_weight_to_fp16 = const()[name = string("layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217339456)))];
            tensor<fp16, [1, ?, 1024]> var_2439_cast_fp16 = mul(x = var_2438_cast_fp16, y = layers_13_input_layernorm_weight_to_fp16)[name = string("op_2439_cast_fp16")];
            tensor<int32, [3]> var_2451_shape_cast_fp16 = shape(x = var_2439_cast_fp16)[name = string("op_2451_shape_cast_fp16")];
            int32 gather_78 = const()[name = string("gather_78"), val = int32(1)];
            int32 gather_79_axis_0 = const()[name = string("gather_79_axis_0"), val = int32(0)];
            int32 gather_79_batch_dims_0 = const()[name = string("gather_79_batch_dims_0"), val = int32(0)];
            bool gather_79_validate_indices_0 = const()[name = string("gather_79_validate_indices_0"), val = bool(false)];
            string var_2451_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2451_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_79_to_uint16 = const()[name = string("select_79_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_2451_shape_cast_fp16_to_uint16 = cast(dtype = var_2451_shape_cast_fp16_to_uint16_dtype_0, x = var_2451_shape_cast_fp16)[name = string("cast_29")];
            uint16 gather_79_cast_uint16 = gather(axis = gather_79_axis_0, batch_dims = gather_79_batch_dims_0, indices = select_79_to_uint16, validate_indices = gather_79_validate_indices_0, x = var_2451_shape_cast_fp16_to_uint16)[name = string("gather_79_cast_uint16")];
            string gather_79_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_79_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217341568))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219438784))))[name = string("layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_91_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = var_2439_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> concat_78x = const()[name = string("concat_78x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_2456_cast_fp16 = reshape(shape = concat_78x, x = linear_91_cast_fp16)[name = string("op_2456_cast_fp16")];
            tensor<int32, [4]> x_343_perm_0 = const()[name = string("x_343_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219569920))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220618560))))[name = string("layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_92_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = var_2439_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> concat_79x = const()[name = string("concat_79x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_2461_cast_fp16 = reshape(shape = concat_79x, x = linear_92_cast_fp16)[name = string("op_2461_cast_fp16")];
            tensor<int32, [4]> x_347_perm_0 = const()[name = string("x_347_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220684160))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221732800))))[name = string("layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = var_2439_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> concat_80x = const()[name = string("concat_80x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_2466_cast_fp16 = reshape(shape = concat_80x, x = linear_93_cast_fp16)[name = string("op_2466_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2422_promoted_1_to_fp16 = const()[name = string("op_2422_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_343_cast_fp16 = transpose(perm = x_343_perm_0, x = var_2456_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 16, ?, 128]> var_2470_cast_fp16 = pow(x = x_343_cast_fp16, y = var_2422_promoted_1_to_fp16)[name = string("op_2470_cast_fp16")];
            tensor<int32, [1]> var_2472_axes_0 = const()[name = string("op_2472_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2472_keep_dims_0 = const()[name = string("op_2472_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_2472_cast_fp16 = reduce_mean(axes = var_2472_axes_0, keep_dims = var_2472_keep_dims_0, x = var_2470_cast_fp16)[name = string("op_2472_cast_fp16")];
            fp16 var_2473_to_fp16 = const()[name = string("op_2473_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_2474_cast_fp16 = add(x = var_2472_cast_fp16, y = var_2473_to_fp16)[name = string("op_2474_cast_fp16")];
            fp32 norm_107_epsilon_0 = const()[name = string("norm_107_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_107_cast_fp16 = rsqrt(epsilon = norm_107_epsilon_0, x = var_2474_cast_fp16)[name = string("norm_107_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_2476_cast_fp16 = mul(x = x_343_cast_fp16, y = norm_107_cast_fp16)[name = string("op_2476_cast_fp16")];
            tensor<fp16, [128]> layers_13_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_13_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221798400)))];
            tensor<fp16, [1, 16, ?, 128]> var_2477_cast_fp16 = mul(x = var_2476_cast_fp16, y = layers_13_self_attn_q_norm_weight_to_fp16)[name = string("op_2477_cast_fp16")];
            fp16 var_2422_promoted_2_to_fp16 = const()[name = string("op_2422_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_347_cast_fp16 = transpose(perm = x_347_perm_0, x = var_2461_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 8, ?, 128]> var_2481_cast_fp16 = pow(x = x_347_cast_fp16, y = var_2422_promoted_2_to_fp16)[name = string("op_2481_cast_fp16")];
            tensor<int32, [1]> var_2483_axes_0 = const()[name = string("op_2483_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2483_keep_dims_0 = const()[name = string("op_2483_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_2483_cast_fp16 = reduce_mean(axes = var_2483_axes_0, keep_dims = var_2483_keep_dims_0, x = var_2481_cast_fp16)[name = string("op_2483_cast_fp16")];
            fp16 var_2484_to_fp16 = const()[name = string("op_2484_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_2485_cast_fp16 = add(x = var_2483_cast_fp16, y = var_2484_to_fp16)[name = string("op_2485_cast_fp16")];
            fp32 norm_109_epsilon_0 = const()[name = string("norm_109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_109_cast_fp16 = rsqrt(epsilon = norm_109_epsilon_0, x = var_2485_cast_fp16)[name = string("norm_109_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_2487_cast_fp16 = mul(x = x_347_cast_fp16, y = norm_109_cast_fp16)[name = string("op_2487_cast_fp16")];
            tensor<fp16, [128]> layers_13_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_13_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221798720)))];
            tensor<fp16, [1, 8, ?, 128]> var_2488_cast_fp16 = mul(x = var_2487_cast_fp16, y = layers_13_self_attn_k_norm_weight_to_fp16)[name = string("op_2488_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = var_2477_cast_fp16)[name = string("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = var_2477_cast_fp16)[name = string("x2_53_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2504_cast_fp16 = mul(x = x1_53_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2504_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2505_cast_fp16 = mul(x = x2_53_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2505_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2506_cast_fp16 = sub(x = var_2504_cast_fp16, y = var_2505_cast_fp16)[name = string("op_2506_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2507_cast_fp16 = mul(x = x2_53_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2507_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2508_cast_fp16 = mul(x = x1_53_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2508_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2509_cast_fp16 = add(x = var_2507_cast_fp16, y = var_2508_cast_fp16)[name = string("op_2509_cast_fp16")];
            bool q_27_interleave_0 = const()[name = string("q_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_27_cast_fp16 = concat(axis = var_2423, interleave = q_27_interleave_0, values = (var_2506_cast_fp16, var_2509_cast_fp16))[name = string("q_27_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = var_2488_cast_fp16)[name = string("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = var_2488_cast_fp16)[name = string("x2_55_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2526_cast_fp16 = mul(x = x1_55_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2526_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2527_cast_fp16 = mul(x = x2_55_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2527_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2528_cast_fp16 = sub(x = var_2526_cast_fp16, y = var_2527_cast_fp16)[name = string("op_2528_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2529_cast_fp16 = mul(x = x2_55_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2530_cast_fp16 = mul(x = x1_55_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2530_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2531_cast_fp16 = add(x = var_2529_cast_fp16, y = var_2530_cast_fp16)[name = string("op_2531_cast_fp16")];
            bool k_27_interleave_0 = const()[name = string("k_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_27_cast_fp16 = concat(axis = var_2423, interleave = k_27_interleave_0, values = (var_2528_cast_fp16, var_2531_cast_fp16))[name = string("k_27_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_26 = read_state(input = k_cache_13)[name = string("read_state_26")];
            int32 k_cache_55_axis_0 = const()[name = string("k_cache_55_axis_0"), val = int32(2)];
            string k_cache_55_mode_0 = const()[name = string("k_cache_55_mode_0"), val = string("update")];
            bool k_cache_55_validate_indices_0 = const()[name = string("k_cache_55_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_55_cast_fp16 = scatter_along_axis(axis = k_cache_55_axis_0, data = read_state_26, indices = scatter_indices, mode = k_cache_55_mode_0, updates = k_27_cast_fp16, validate_indices = k_cache_55_validate_indices_0)[name = string("k_cache_55_cast_fp16")];
            write_state(data = k_cache_55_cast_fp16, input = k_cache_13)[name = string("coreml_update_state_26_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_82 = read_state(input = k_cache_13)[name = string("coreml_update_state_26")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_27 = read_state(input = v_cache_13)[name = string("read_state_27")];
            int32 v_cache_55_axis_0 = const()[name = string("v_cache_55_axis_0"), val = int32(2)];
            string v_cache_55_mode_0 = const()[name = string("v_cache_55_mode_0"), val = string("update")];
            bool v_cache_55_validate_indices_0 = const()[name = string("v_cache_55_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_2466_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_55_cast_fp16 = scatter_along_axis(axis = v_cache_55_axis_0, data = read_state_27, indices = scatter_indices, mode = v_cache_55_mode_0, updates = v_27_cast_fp16, validate_indices = v_cache_55_validate_indices_0)[name = string("v_cache_55_cast_fp16")];
            write_state(data = v_cache_55_cast_fp16, input = v_cache_13)[name = string("coreml_update_state_27_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_83 = read_state(input = v_cache_13)[name = string("coreml_update_state_27")];
            tensor<int32, [1]> var_2536_axes_0 = const()[name = string("op_2536_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2536_cast_fp16 = expand_dims(axes = var_2536_axes_0, x = coreml_update_state_82)[name = string("op_2536_cast_fp16")];
            tensor<int32, [5]> k_exp_53_reps_0 = const()[name = string("k_exp_53_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_53_cast_fp16 = tile(reps = k_exp_53_reps_0, x = var_2536_cast_fp16)[name = string("k_exp_53_cast_fp16")];
            tensor<int32, [4]> var_2539 = const()[name = string("op_2539"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_55_cast_fp16 = reshape(shape = var_2539, x = k_exp_53_cast_fp16)[name = string("k_exp_55_cast_fp16")];
            tensor<int32, [1]> var_2541_axes_0 = const()[name = string("op_2541_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2541_cast_fp16 = expand_dims(axes = var_2541_axes_0, x = coreml_update_state_83)[name = string("op_2541_cast_fp16")];
            tensor<int32, [5]> v_exp_53_reps_0 = const()[name = string("v_exp_53_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_53_cast_fp16 = tile(reps = v_exp_53_reps_0, x = var_2541_cast_fp16)[name = string("v_exp_53_cast_fp16")];
            tensor<int32, [4]> var_2544 = const()[name = string("op_2544"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_55_cast_fp16 = reshape(shape = var_2544, x = v_exp_53_cast_fp16)[name = string("v_exp_55_cast_fp16")];
            bool var_2547_transpose_x_1 = const()[name = string("op_2547_transpose_x_1"), val = bool(false)];
            bool var_2547_transpose_y_1 = const()[name = string("op_2547_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_2547_cast_fp16 = matmul(transpose_x = var_2547_transpose_x_1, transpose_y = var_2547_transpose_y_1, x = q_27_cast_fp16, y = k_exp_55_cast_fp16)[name = string("op_2547_cast_fp16")];
            fp16 var_2548_to_fp16 = const()[name = string("op_2548_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_53_cast_fp16 = mul(x = var_2547_cast_fp16, y = var_2548_to_fp16)[name = string("attn_53_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_131_cast_fp16 = add(x = attn_53_cast_fp16, y = attention_mask)[name = string("input_131_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_55_cast_fp16 = softmax(axis = var_2423, x = input_131_cast_fp16)[name = string("attn_55_cast_fp16")];
            bool out_27_transpose_x_0 = const()[name = string("out_27_transpose_x_0"), val = bool(false)];
            bool out_27_transpose_y_0 = const()[name = string("out_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_27_cast_fp16 = matmul(transpose_x = out_27_transpose_x_0, transpose_y = out_27_transpose_y_0, x = attn_55_cast_fp16, y = v_exp_55_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [4]> var_2553_perm_0 = const()[name = string("op_2553_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_83_axis_0 = const()[name = string("concat_83_axis_0"), val = int32(0)];
            bool concat_83_interleave_0 = const()[name = string("concat_83_interleave_0"), val = bool(false)];
            int32 gather_79_cast_uint16_to_int32 = cast(dtype = gather_79_cast_uint16_to_int32_dtype_0, x = gather_79_cast_uint16)[name = string("cast_28")];
            tensor<int32, [3]> concat_83 = concat(axis = concat_83_axis_0, interleave = concat_83_interleave_0, values = (gather_78, gather_79_cast_uint16_to_int32, var_2423))[name = string("concat_83")];
            tensor<fp16, [1, ?, 16, 128]> var_2553_cast_fp16 = transpose(perm = var_2553_perm_0, x = out_27_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, ?, ?]> input_133_cast_fp16 = reshape(shape = concat_83, x = var_2553_cast_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221799040))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223896256))))[name = string("layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_94_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_357_cast_fp16 = add(x = x_337_cast_fp16, y = linear_94_cast_fp16)[name = string("x_357_cast_fp16")];
            fp16 var_2422_promoted_3_to_fp16 = const()[name = string("op_2422_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_2561_cast_fp16 = pow(x = x_357_cast_fp16, y = var_2422_promoted_3_to_fp16)[name = string("op_2561_cast_fp16")];
            tensor<int32, [1]> var_2563_axes_0 = const()[name = string("op_2563_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2563_keep_dims_0 = const()[name = string("op_2563_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_2563_cast_fp16 = reduce_mean(axes = var_2563_axes_0, keep_dims = var_2563_keep_dims_0, x = var_2561_cast_fp16)[name = string("op_2563_cast_fp16")];
            fp16 var_2564_to_fp16 = const()[name = string("op_2564_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2565_cast_fp16 = add(x = var_2563_cast_fp16, y = var_2564_to_fp16)[name = string("op_2565_cast_fp16")];
            fp32 norm_111_epsilon_0 = const()[name = string("norm_111_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_111_cast_fp16 = rsqrt(epsilon = norm_111_epsilon_0, x = var_2565_cast_fp16)[name = string("norm_111_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_2567_cast_fp16 = mul(x = x_357_cast_fp16, y = norm_111_cast_fp16)[name = string("op_2567_cast_fp16")];
            tensor<fp16, [1024]> layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224027392)))];
            tensor<fp16, [1, ?, 1024]> var_2568_cast_fp16 = mul(x = var_2567_cast_fp16, y = layers_13_post_attention_layernorm_weight_to_fp16)[name = string("op_2568_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224029504))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227175296))))[name = string("layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = var_2568_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_2578_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_2578_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227371968))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230517760))))[name = string("layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_13_mlp_up_proj_weight_to_fp16_quantized, x = var_2568_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_139_cast_fp16 = mul(x = var_2578_cast_fp16, y = linear_96_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230714432))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233860224))))[name = string("layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_97_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_363_cast_fp16 = add(x = x_357_cast_fp16, y = linear_97_cast_fp16)[name = string("x_363_cast_fp16")];
            int32 var_2598 = const()[name = string("op_2598"), val = int32(-1)];
            fp16 var_2597_promoted_to_fp16 = const()[name = string("op_2597_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_2607_cast_fp16 = pow(x = x_363_cast_fp16, y = var_2597_promoted_to_fp16)[name = string("op_2607_cast_fp16")];
            tensor<int32, [1]> var_2609_axes_0 = const()[name = string("op_2609_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2609_keep_dims_0 = const()[name = string("op_2609_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_2609_cast_fp16 = reduce_mean(axes = var_2609_axes_0, keep_dims = var_2609_keep_dims_0, x = var_2607_cast_fp16)[name = string("op_2609_cast_fp16")];
            fp16 var_2610_to_fp16 = const()[name = string("op_2610_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2611_cast_fp16 = add(x = var_2609_cast_fp16, y = var_2610_to_fp16)[name = string("op_2611_cast_fp16")];
            fp32 norm_113_epsilon_0 = const()[name = string("norm_113_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_113_cast_fp16 = rsqrt(epsilon = norm_113_epsilon_0, x = var_2611_cast_fp16)[name = string("norm_113_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_2613_cast_fp16 = mul(x = x_363_cast_fp16, y = norm_113_cast_fp16)[name = string("op_2613_cast_fp16")];
            tensor<fp16, [1024]> layers_14_input_layernorm_weight_to_fp16 = const()[name = string("layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234056896)))];
            tensor<fp16, [1, ?, 1024]> var_2614_cast_fp16 = mul(x = var_2613_cast_fp16, y = layers_14_input_layernorm_weight_to_fp16)[name = string("op_2614_cast_fp16")];
            tensor<int32, [3]> var_2626_shape_cast_fp16 = shape(x = var_2614_cast_fp16)[name = string("op_2626_shape_cast_fp16")];
            int32 gather_84 = const()[name = string("gather_84"), val = int32(1)];
            int32 gather_85_axis_0 = const()[name = string("gather_85_axis_0"), val = int32(0)];
            int32 gather_85_batch_dims_0 = const()[name = string("gather_85_batch_dims_0"), val = int32(0)];
            bool gather_85_validate_indices_0 = const()[name = string("gather_85_validate_indices_0"), val = bool(false)];
            string var_2626_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2626_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_85_to_uint16 = const()[name = string("select_85_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_2626_shape_cast_fp16_to_uint16 = cast(dtype = var_2626_shape_cast_fp16_to_uint16_dtype_0, x = var_2626_shape_cast_fp16)[name = string("cast_27")];
            uint16 gather_85_cast_uint16 = gather(axis = gather_85_axis_0, batch_dims = gather_85_batch_dims_0, indices = select_85_to_uint16, validate_indices = gather_85_validate_indices_0, x = var_2626_shape_cast_fp16_to_uint16)[name = string("gather_85_cast_uint16")];
            string gather_85_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_85_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234059008))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236156224))))[name = string("layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_98_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = var_2614_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [4]> concat_84x = const()[name = string("concat_84x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_2631_cast_fp16 = reshape(shape = concat_84x, x = linear_98_cast_fp16)[name = string("op_2631_cast_fp16")];
            tensor<int32, [4]> x_369_perm_0 = const()[name = string("x_369_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236287360))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237336000))))[name = string("layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_99_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = var_2614_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> concat_85x = const()[name = string("concat_85x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_2636_cast_fp16 = reshape(shape = concat_85x, x = linear_99_cast_fp16)[name = string("op_2636_cast_fp16")];
            tensor<int32, [4]> x_373_perm_0 = const()[name = string("x_373_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237401600))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238450240))))[name = string("layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_100_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = var_2614_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> concat_86x = const()[name = string("concat_86x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_2641_cast_fp16 = reshape(shape = concat_86x, x = linear_100_cast_fp16)[name = string("op_2641_cast_fp16")];
            tensor<int32, [4]> v_29_perm_0 = const()[name = string("v_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2597_promoted_1_to_fp16 = const()[name = string("op_2597_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_369_cast_fp16 = transpose(perm = x_369_perm_0, x = var_2631_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 16, ?, 128]> var_2645_cast_fp16 = pow(x = x_369_cast_fp16, y = var_2597_promoted_1_to_fp16)[name = string("op_2645_cast_fp16")];
            tensor<int32, [1]> var_2647_axes_0 = const()[name = string("op_2647_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2647_keep_dims_0 = const()[name = string("op_2647_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_2647_cast_fp16 = reduce_mean(axes = var_2647_axes_0, keep_dims = var_2647_keep_dims_0, x = var_2645_cast_fp16)[name = string("op_2647_cast_fp16")];
            fp16 var_2648_to_fp16 = const()[name = string("op_2648_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_2649_cast_fp16 = add(x = var_2647_cast_fp16, y = var_2648_to_fp16)[name = string("op_2649_cast_fp16")];
            fp32 norm_115_epsilon_0 = const()[name = string("norm_115_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_115_cast_fp16 = rsqrt(epsilon = norm_115_epsilon_0, x = var_2649_cast_fp16)[name = string("norm_115_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_2651_cast_fp16 = mul(x = x_369_cast_fp16, y = norm_115_cast_fp16)[name = string("op_2651_cast_fp16")];
            tensor<fp16, [128]> layers_14_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_14_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238515840)))];
            tensor<fp16, [1, 16, ?, 128]> var_2652_cast_fp16 = mul(x = var_2651_cast_fp16, y = layers_14_self_attn_q_norm_weight_to_fp16)[name = string("op_2652_cast_fp16")];
            fp16 var_2597_promoted_2_to_fp16 = const()[name = string("op_2597_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_373_cast_fp16 = transpose(perm = x_373_perm_0, x = var_2636_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 8, ?, 128]> var_2656_cast_fp16 = pow(x = x_373_cast_fp16, y = var_2597_promoted_2_to_fp16)[name = string("op_2656_cast_fp16")];
            tensor<int32, [1]> var_2658_axes_0 = const()[name = string("op_2658_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2658_keep_dims_0 = const()[name = string("op_2658_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_2658_cast_fp16 = reduce_mean(axes = var_2658_axes_0, keep_dims = var_2658_keep_dims_0, x = var_2656_cast_fp16)[name = string("op_2658_cast_fp16")];
            fp16 var_2659_to_fp16 = const()[name = string("op_2659_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_2660_cast_fp16 = add(x = var_2658_cast_fp16, y = var_2659_to_fp16)[name = string("op_2660_cast_fp16")];
            fp32 norm_117_epsilon_0 = const()[name = string("norm_117_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_117_cast_fp16 = rsqrt(epsilon = norm_117_epsilon_0, x = var_2660_cast_fp16)[name = string("norm_117_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_2662_cast_fp16 = mul(x = x_373_cast_fp16, y = norm_117_cast_fp16)[name = string("op_2662_cast_fp16")];
            tensor<fp16, [128]> layers_14_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_14_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238516160)))];
            tensor<fp16, [1, 8, ?, 128]> var_2663_cast_fp16 = mul(x = var_2662_cast_fp16, y = layers_14_self_attn_k_norm_weight_to_fp16)[name = string("op_2663_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = var_2652_cast_fp16)[name = string("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = var_2652_cast_fp16)[name = string("x2_57_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2679_cast_fp16 = mul(x = x1_57_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2679_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2680_cast_fp16 = mul(x = x2_57_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2680_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2681_cast_fp16 = sub(x = var_2679_cast_fp16, y = var_2680_cast_fp16)[name = string("op_2681_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2682_cast_fp16 = mul(x = x2_57_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2682_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2683_cast_fp16 = mul(x = x1_57_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2684_cast_fp16 = add(x = var_2682_cast_fp16, y = var_2683_cast_fp16)[name = string("op_2684_cast_fp16")];
            bool q_29_interleave_0 = const()[name = string("q_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_29_cast_fp16 = concat(axis = var_2598, interleave = q_29_interleave_0, values = (var_2681_cast_fp16, var_2684_cast_fp16))[name = string("q_29_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = var_2663_cast_fp16)[name = string("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = var_2663_cast_fp16)[name = string("x2_59_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2701_cast_fp16 = mul(x = x1_59_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2701_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2702_cast_fp16 = mul(x = x2_59_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2702_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2703_cast_fp16 = sub(x = var_2701_cast_fp16, y = var_2702_cast_fp16)[name = string("op_2703_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2704_cast_fp16 = mul(x = x2_59_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2704_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2705_cast_fp16 = mul(x = x1_59_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2705_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2706_cast_fp16 = add(x = var_2704_cast_fp16, y = var_2705_cast_fp16)[name = string("op_2706_cast_fp16")];
            bool k_29_interleave_0 = const()[name = string("k_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_29_cast_fp16 = concat(axis = var_2598, interleave = k_29_interleave_0, values = (var_2703_cast_fp16, var_2706_cast_fp16))[name = string("k_29_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_28 = read_state(input = k_cache_14)[name = string("read_state_28")];
            int32 k_cache_59_axis_0 = const()[name = string("k_cache_59_axis_0"), val = int32(2)];
            string k_cache_59_mode_0 = const()[name = string("k_cache_59_mode_0"), val = string("update")];
            bool k_cache_59_validate_indices_0 = const()[name = string("k_cache_59_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_59_cast_fp16 = scatter_along_axis(axis = k_cache_59_axis_0, data = read_state_28, indices = scatter_indices, mode = k_cache_59_mode_0, updates = k_29_cast_fp16, validate_indices = k_cache_59_validate_indices_0)[name = string("k_cache_59_cast_fp16")];
            write_state(data = k_cache_59_cast_fp16, input = k_cache_14)[name = string("coreml_update_state_28_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_84 = read_state(input = k_cache_14)[name = string("coreml_update_state_28")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_29 = read_state(input = v_cache_14)[name = string("read_state_29")];
            int32 v_cache_59_axis_0 = const()[name = string("v_cache_59_axis_0"), val = int32(2)];
            string v_cache_59_mode_0 = const()[name = string("v_cache_59_mode_0"), val = string("update")];
            bool v_cache_59_validate_indices_0 = const()[name = string("v_cache_59_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_29_cast_fp16 = transpose(perm = v_29_perm_0, x = var_2641_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_59_cast_fp16 = scatter_along_axis(axis = v_cache_59_axis_0, data = read_state_29, indices = scatter_indices, mode = v_cache_59_mode_0, updates = v_29_cast_fp16, validate_indices = v_cache_59_validate_indices_0)[name = string("v_cache_59_cast_fp16")];
            write_state(data = v_cache_59_cast_fp16, input = v_cache_14)[name = string("coreml_update_state_29_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_85 = read_state(input = v_cache_14)[name = string("coreml_update_state_29")];
            tensor<int32, [1]> var_2711_axes_0 = const()[name = string("op_2711_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2711_cast_fp16 = expand_dims(axes = var_2711_axes_0, x = coreml_update_state_84)[name = string("op_2711_cast_fp16")];
            tensor<int32, [5]> k_exp_57_reps_0 = const()[name = string("k_exp_57_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_57_cast_fp16 = tile(reps = k_exp_57_reps_0, x = var_2711_cast_fp16)[name = string("k_exp_57_cast_fp16")];
            tensor<int32, [4]> var_2714 = const()[name = string("op_2714"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_59_cast_fp16 = reshape(shape = var_2714, x = k_exp_57_cast_fp16)[name = string("k_exp_59_cast_fp16")];
            tensor<int32, [1]> var_2716_axes_0 = const()[name = string("op_2716_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2716_cast_fp16 = expand_dims(axes = var_2716_axes_0, x = coreml_update_state_85)[name = string("op_2716_cast_fp16")];
            tensor<int32, [5]> v_exp_57_reps_0 = const()[name = string("v_exp_57_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_57_cast_fp16 = tile(reps = v_exp_57_reps_0, x = var_2716_cast_fp16)[name = string("v_exp_57_cast_fp16")];
            tensor<int32, [4]> var_2719 = const()[name = string("op_2719"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_59_cast_fp16 = reshape(shape = var_2719, x = v_exp_57_cast_fp16)[name = string("v_exp_59_cast_fp16")];
            bool var_2722_transpose_x_1 = const()[name = string("op_2722_transpose_x_1"), val = bool(false)];
            bool var_2722_transpose_y_1 = const()[name = string("op_2722_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_2722_cast_fp16 = matmul(transpose_x = var_2722_transpose_x_1, transpose_y = var_2722_transpose_y_1, x = q_29_cast_fp16, y = k_exp_59_cast_fp16)[name = string("op_2722_cast_fp16")];
            fp16 var_2723_to_fp16 = const()[name = string("op_2723_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_57_cast_fp16 = mul(x = var_2722_cast_fp16, y = var_2723_to_fp16)[name = string("attn_57_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_141_cast_fp16 = add(x = attn_57_cast_fp16, y = attention_mask)[name = string("input_141_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_59_cast_fp16 = softmax(axis = var_2598, x = input_141_cast_fp16)[name = string("attn_59_cast_fp16")];
            bool out_29_transpose_x_0 = const()[name = string("out_29_transpose_x_0"), val = bool(false)];
            bool out_29_transpose_y_0 = const()[name = string("out_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_29_cast_fp16 = matmul(transpose_x = out_29_transpose_x_0, transpose_y = out_29_transpose_y_0, x = attn_59_cast_fp16, y = v_exp_59_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [4]> var_2728_perm_0 = const()[name = string("op_2728_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_89_axis_0 = const()[name = string("concat_89_axis_0"), val = int32(0)];
            bool concat_89_interleave_0 = const()[name = string("concat_89_interleave_0"), val = bool(false)];
            int32 gather_85_cast_uint16_to_int32 = cast(dtype = gather_85_cast_uint16_to_int32_dtype_0, x = gather_85_cast_uint16)[name = string("cast_26")];
            tensor<int32, [3]> concat_89 = concat(axis = concat_89_axis_0, interleave = concat_89_interleave_0, values = (gather_84, gather_85_cast_uint16_to_int32, var_2598))[name = string("concat_89")];
            tensor<fp16, [1, ?, 16, 128]> var_2728_cast_fp16 = transpose(perm = var_2728_perm_0, x = out_29_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, ?, ?]> input_143_cast_fp16 = reshape(shape = concat_89, x = var_2728_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238516480))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240613696))))[name = string("layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_101_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_383_cast_fp16 = add(x = x_363_cast_fp16, y = linear_101_cast_fp16)[name = string("x_383_cast_fp16")];
            fp16 var_2597_promoted_3_to_fp16 = const()[name = string("op_2597_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_2736_cast_fp16 = pow(x = x_383_cast_fp16, y = var_2597_promoted_3_to_fp16)[name = string("op_2736_cast_fp16")];
            tensor<int32, [1]> var_2738_axes_0 = const()[name = string("op_2738_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2738_keep_dims_0 = const()[name = string("op_2738_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_2738_cast_fp16 = reduce_mean(axes = var_2738_axes_0, keep_dims = var_2738_keep_dims_0, x = var_2736_cast_fp16)[name = string("op_2738_cast_fp16")];
            fp16 var_2739_to_fp16 = const()[name = string("op_2739_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2740_cast_fp16 = add(x = var_2738_cast_fp16, y = var_2739_to_fp16)[name = string("op_2740_cast_fp16")];
            fp32 norm_119_epsilon_0 = const()[name = string("norm_119_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_119_cast_fp16 = rsqrt(epsilon = norm_119_epsilon_0, x = var_2740_cast_fp16)[name = string("norm_119_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_2742_cast_fp16 = mul(x = x_383_cast_fp16, y = norm_119_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<fp16, [1024]> layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240744832)))];
            tensor<fp16, [1, ?, 1024]> var_2743_cast_fp16 = mul(x = var_2742_cast_fp16, y = layers_14_post_attention_layernorm_weight_to_fp16)[name = string("op_2743_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240746944))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243892736))))[name = string("layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = var_2743_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_2753_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_2753_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244089408))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247235200))))[name = string("layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_14_mlp_up_proj_weight_to_fp16_quantized, x = var_2743_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_149_cast_fp16 = mul(x = var_2753_cast_fp16, y = linear_103_cast_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247431872))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250577664))))[name = string("layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_104_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_389_cast_fp16 = add(x = x_383_cast_fp16, y = linear_104_cast_fp16)[name = string("x_389_cast_fp16")];
            int32 var_2773 = const()[name = string("op_2773"), val = int32(-1)];
            fp16 var_2772_promoted_to_fp16 = const()[name = string("op_2772_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_2782_cast_fp16 = pow(x = x_389_cast_fp16, y = var_2772_promoted_to_fp16)[name = string("op_2782_cast_fp16")];
            tensor<int32, [1]> var_2784_axes_0 = const()[name = string("op_2784_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2784_keep_dims_0 = const()[name = string("op_2784_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_2784_cast_fp16 = reduce_mean(axes = var_2784_axes_0, keep_dims = var_2784_keep_dims_0, x = var_2782_cast_fp16)[name = string("op_2784_cast_fp16")];
            fp16 var_2785_to_fp16 = const()[name = string("op_2785_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2786_cast_fp16 = add(x = var_2784_cast_fp16, y = var_2785_to_fp16)[name = string("op_2786_cast_fp16")];
            fp32 norm_121_epsilon_0 = const()[name = string("norm_121_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_121_cast_fp16 = rsqrt(epsilon = norm_121_epsilon_0, x = var_2786_cast_fp16)[name = string("norm_121_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_2788_cast_fp16 = mul(x = x_389_cast_fp16, y = norm_121_cast_fp16)[name = string("op_2788_cast_fp16")];
            tensor<fp16, [1024]> layers_15_input_layernorm_weight_to_fp16 = const()[name = string("layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250774336)))];
            tensor<fp16, [1, ?, 1024]> var_2789_cast_fp16 = mul(x = var_2788_cast_fp16, y = layers_15_input_layernorm_weight_to_fp16)[name = string("op_2789_cast_fp16")];
            tensor<int32, [3]> var_2801_shape_cast_fp16 = shape(x = var_2789_cast_fp16)[name = string("op_2801_shape_cast_fp16")];
            int32 gather_90 = const()[name = string("gather_90"), val = int32(1)];
            int32 gather_91_axis_0 = const()[name = string("gather_91_axis_0"), val = int32(0)];
            int32 gather_91_batch_dims_0 = const()[name = string("gather_91_batch_dims_0"), val = int32(0)];
            bool gather_91_validate_indices_0 = const()[name = string("gather_91_validate_indices_0"), val = bool(false)];
            string var_2801_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2801_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_91_to_uint16 = const()[name = string("select_91_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_2801_shape_cast_fp16_to_uint16 = cast(dtype = var_2801_shape_cast_fp16_to_uint16_dtype_0, x = var_2801_shape_cast_fp16)[name = string("cast_25")];
            uint16 gather_91_cast_uint16 = gather(axis = gather_91_axis_0, batch_dims = gather_91_batch_dims_0, indices = select_91_to_uint16, validate_indices = gather_91_validate_indices_0, x = var_2801_shape_cast_fp16_to_uint16)[name = string("gather_91_cast_uint16")];
            string gather_91_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_91_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250776448))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252873664))))[name = string("layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_105_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = var_2789_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [4]> concat_90x = const()[name = string("concat_90x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_2806_cast_fp16 = reshape(shape = concat_90x, x = linear_105_cast_fp16)[name = string("op_2806_cast_fp16")];
            tensor<int32, [4]> x_395_perm_0 = const()[name = string("x_395_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253004800))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254053440))))[name = string("layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_106_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = var_2789_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> concat_91x = const()[name = string("concat_91x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_2811_cast_fp16 = reshape(shape = concat_91x, x = linear_106_cast_fp16)[name = string("op_2811_cast_fp16")];
            tensor<int32, [4]> x_399_perm_0 = const()[name = string("x_399_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254119040))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255167680))))[name = string("layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_107_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = var_2789_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> concat_92x = const()[name = string("concat_92x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_2816_cast_fp16 = reshape(shape = concat_92x, x = linear_107_cast_fp16)[name = string("op_2816_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2772_promoted_1_to_fp16 = const()[name = string("op_2772_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_395_cast_fp16 = transpose(perm = x_395_perm_0, x = var_2806_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 16, ?, 128]> var_2820_cast_fp16 = pow(x = x_395_cast_fp16, y = var_2772_promoted_1_to_fp16)[name = string("op_2820_cast_fp16")];
            tensor<int32, [1]> var_2822_axes_0 = const()[name = string("op_2822_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2822_keep_dims_0 = const()[name = string("op_2822_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_2822_cast_fp16 = reduce_mean(axes = var_2822_axes_0, keep_dims = var_2822_keep_dims_0, x = var_2820_cast_fp16)[name = string("op_2822_cast_fp16")];
            fp16 var_2823_to_fp16 = const()[name = string("op_2823_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_2824_cast_fp16 = add(x = var_2822_cast_fp16, y = var_2823_to_fp16)[name = string("op_2824_cast_fp16")];
            fp32 norm_123_epsilon_0 = const()[name = string("norm_123_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_123_cast_fp16 = rsqrt(epsilon = norm_123_epsilon_0, x = var_2824_cast_fp16)[name = string("norm_123_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_2826_cast_fp16 = mul(x = x_395_cast_fp16, y = norm_123_cast_fp16)[name = string("op_2826_cast_fp16")];
            tensor<fp16, [128]> layers_15_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_15_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255233280)))];
            tensor<fp16, [1, 16, ?, 128]> var_2827_cast_fp16 = mul(x = var_2826_cast_fp16, y = layers_15_self_attn_q_norm_weight_to_fp16)[name = string("op_2827_cast_fp16")];
            fp16 var_2772_promoted_2_to_fp16 = const()[name = string("op_2772_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_399_cast_fp16 = transpose(perm = x_399_perm_0, x = var_2811_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 8, ?, 128]> var_2831_cast_fp16 = pow(x = x_399_cast_fp16, y = var_2772_promoted_2_to_fp16)[name = string("op_2831_cast_fp16")];
            tensor<int32, [1]> var_2833_axes_0 = const()[name = string("op_2833_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2833_keep_dims_0 = const()[name = string("op_2833_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_2833_cast_fp16 = reduce_mean(axes = var_2833_axes_0, keep_dims = var_2833_keep_dims_0, x = var_2831_cast_fp16)[name = string("op_2833_cast_fp16")];
            fp16 var_2834_to_fp16 = const()[name = string("op_2834_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_2835_cast_fp16 = add(x = var_2833_cast_fp16, y = var_2834_to_fp16)[name = string("op_2835_cast_fp16")];
            fp32 norm_125_epsilon_0 = const()[name = string("norm_125_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_125_cast_fp16 = rsqrt(epsilon = norm_125_epsilon_0, x = var_2835_cast_fp16)[name = string("norm_125_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_2837_cast_fp16 = mul(x = x_399_cast_fp16, y = norm_125_cast_fp16)[name = string("op_2837_cast_fp16")];
            tensor<fp16, [128]> layers_15_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_15_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255233600)))];
            tensor<fp16, [1, 8, ?, 128]> var_2838_cast_fp16 = mul(x = var_2837_cast_fp16, y = layers_15_self_attn_k_norm_weight_to_fp16)[name = string("op_2838_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = var_2827_cast_fp16)[name = string("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = var_2827_cast_fp16)[name = string("x2_61_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2854_cast_fp16 = mul(x = x1_61_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2854_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2855_cast_fp16 = mul(x = x2_61_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2855_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2856_cast_fp16 = sub(x = var_2854_cast_fp16, y = var_2855_cast_fp16)[name = string("op_2856_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2857_cast_fp16 = mul(x = x2_61_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2857_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2858_cast_fp16 = mul(x = x1_61_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2858_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_2859_cast_fp16 = add(x = var_2857_cast_fp16, y = var_2858_cast_fp16)[name = string("op_2859_cast_fp16")];
            bool q_31_interleave_0 = const()[name = string("q_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_31_cast_fp16 = concat(axis = var_2773, interleave = q_31_interleave_0, values = (var_2856_cast_fp16, var_2859_cast_fp16))[name = string("q_31_cast_fp16")];
            tensor<int32, [4]> x1_63_begin_0 = const()[name = string("x1_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_63_end_0 = const()[name = string("x1_63_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_63_end_mask_0 = const()[name = string("x1_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_63_cast_fp16 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = var_2838_cast_fp16)[name = string("x1_63_cast_fp16")];
            tensor<int32, [4]> x2_63_begin_0 = const()[name = string("x2_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_63_end_0 = const()[name = string("x2_63_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_63_end_mask_0 = const()[name = string("x2_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_63_cast_fp16 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = var_2838_cast_fp16)[name = string("x2_63_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2876_cast_fp16 = mul(x = x1_63_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2876_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2877_cast_fp16 = mul(x = x2_63_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2877_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2878_cast_fp16 = sub(x = var_2876_cast_fp16, y = var_2877_cast_fp16)[name = string("op_2878_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2879_cast_fp16 = mul(x = x2_63_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2879_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2880_cast_fp16 = mul(x = x1_63_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2880_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_2881_cast_fp16 = add(x = var_2879_cast_fp16, y = var_2880_cast_fp16)[name = string("op_2881_cast_fp16")];
            bool k_31_interleave_0 = const()[name = string("k_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_31_cast_fp16 = concat(axis = var_2773, interleave = k_31_interleave_0, values = (var_2878_cast_fp16, var_2881_cast_fp16))[name = string("k_31_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_30 = read_state(input = k_cache_15)[name = string("read_state_30")];
            int32 k_cache_63_axis_0 = const()[name = string("k_cache_63_axis_0"), val = int32(2)];
            string k_cache_63_mode_0 = const()[name = string("k_cache_63_mode_0"), val = string("update")];
            bool k_cache_63_validate_indices_0 = const()[name = string("k_cache_63_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_63_cast_fp16 = scatter_along_axis(axis = k_cache_63_axis_0, data = read_state_30, indices = scatter_indices, mode = k_cache_63_mode_0, updates = k_31_cast_fp16, validate_indices = k_cache_63_validate_indices_0)[name = string("k_cache_63_cast_fp16")];
            write_state(data = k_cache_63_cast_fp16, input = k_cache_15)[name = string("coreml_update_state_30_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_86 = read_state(input = k_cache_15)[name = string("coreml_update_state_30")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_31 = read_state(input = v_cache_15)[name = string("read_state_31")];
            int32 v_cache_63_axis_0 = const()[name = string("v_cache_63_axis_0"), val = int32(2)];
            string v_cache_63_mode_0 = const()[name = string("v_cache_63_mode_0"), val = string("update")];
            bool v_cache_63_validate_indices_0 = const()[name = string("v_cache_63_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = var_2816_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_63_cast_fp16 = scatter_along_axis(axis = v_cache_63_axis_0, data = read_state_31, indices = scatter_indices, mode = v_cache_63_mode_0, updates = v_31_cast_fp16, validate_indices = v_cache_63_validate_indices_0)[name = string("v_cache_63_cast_fp16")];
            write_state(data = v_cache_63_cast_fp16, input = v_cache_15)[name = string("coreml_update_state_31_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_87 = read_state(input = v_cache_15)[name = string("coreml_update_state_31")];
            tensor<int32, [1]> var_2886_axes_0 = const()[name = string("op_2886_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2886_cast_fp16 = expand_dims(axes = var_2886_axes_0, x = coreml_update_state_86)[name = string("op_2886_cast_fp16")];
            tensor<int32, [5]> k_exp_61_reps_0 = const()[name = string("k_exp_61_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_61_cast_fp16 = tile(reps = k_exp_61_reps_0, x = var_2886_cast_fp16)[name = string("k_exp_61_cast_fp16")];
            tensor<int32, [4]> var_2889 = const()[name = string("op_2889"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_63_cast_fp16 = reshape(shape = var_2889, x = k_exp_61_cast_fp16)[name = string("k_exp_63_cast_fp16")];
            tensor<int32, [1]> var_2891_axes_0 = const()[name = string("op_2891_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2891_cast_fp16 = expand_dims(axes = var_2891_axes_0, x = coreml_update_state_87)[name = string("op_2891_cast_fp16")];
            tensor<int32, [5]> v_exp_61_reps_0 = const()[name = string("v_exp_61_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_61_cast_fp16 = tile(reps = v_exp_61_reps_0, x = var_2891_cast_fp16)[name = string("v_exp_61_cast_fp16")];
            tensor<int32, [4]> var_2894 = const()[name = string("op_2894"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_63_cast_fp16 = reshape(shape = var_2894, x = v_exp_61_cast_fp16)[name = string("v_exp_63_cast_fp16")];
            bool var_2897_transpose_x_1 = const()[name = string("op_2897_transpose_x_1"), val = bool(false)];
            bool var_2897_transpose_y_1 = const()[name = string("op_2897_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_2897_cast_fp16 = matmul(transpose_x = var_2897_transpose_x_1, transpose_y = var_2897_transpose_y_1, x = q_31_cast_fp16, y = k_exp_63_cast_fp16)[name = string("op_2897_cast_fp16")];
            fp16 var_2898_to_fp16 = const()[name = string("op_2898_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_61_cast_fp16 = mul(x = var_2897_cast_fp16, y = var_2898_to_fp16)[name = string("attn_61_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_151_cast_fp16 = add(x = attn_61_cast_fp16, y = attention_mask)[name = string("input_151_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_63_cast_fp16 = softmax(axis = var_2773, x = input_151_cast_fp16)[name = string("attn_63_cast_fp16")];
            bool out_31_transpose_x_0 = const()[name = string("out_31_transpose_x_0"), val = bool(false)];
            bool out_31_transpose_y_0 = const()[name = string("out_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_31_cast_fp16 = matmul(transpose_x = out_31_transpose_x_0, transpose_y = out_31_transpose_y_0, x = attn_63_cast_fp16, y = v_exp_63_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [4]> var_2903_perm_0 = const()[name = string("op_2903_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_95_axis_0 = const()[name = string("concat_95_axis_0"), val = int32(0)];
            bool concat_95_interleave_0 = const()[name = string("concat_95_interleave_0"), val = bool(false)];
            int32 gather_91_cast_uint16_to_int32 = cast(dtype = gather_91_cast_uint16_to_int32_dtype_0, x = gather_91_cast_uint16)[name = string("cast_24")];
            tensor<int32, [3]> concat_95 = concat(axis = concat_95_axis_0, interleave = concat_95_interleave_0, values = (gather_90, gather_91_cast_uint16_to_int32, var_2773))[name = string("concat_95")];
            tensor<fp16, [1, ?, 16, 128]> var_2903_cast_fp16 = transpose(perm = var_2903_perm_0, x = out_31_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, ?, ?]> input_153_cast_fp16 = reshape(shape = concat_95, x = var_2903_cast_fp16)[name = string("input_153_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255233920))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257331136))))[name = string("layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_108_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_409_cast_fp16 = add(x = x_389_cast_fp16, y = linear_108_cast_fp16)[name = string("x_409_cast_fp16")];
            fp16 var_2772_promoted_3_to_fp16 = const()[name = string("op_2772_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_2911_cast_fp16 = pow(x = x_409_cast_fp16, y = var_2772_promoted_3_to_fp16)[name = string("op_2911_cast_fp16")];
            tensor<int32, [1]> var_2913_axes_0 = const()[name = string("op_2913_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2913_keep_dims_0 = const()[name = string("op_2913_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_2913_cast_fp16 = reduce_mean(axes = var_2913_axes_0, keep_dims = var_2913_keep_dims_0, x = var_2911_cast_fp16)[name = string("op_2913_cast_fp16")];
            fp16 var_2914_to_fp16 = const()[name = string("op_2914_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2915_cast_fp16 = add(x = var_2913_cast_fp16, y = var_2914_to_fp16)[name = string("op_2915_cast_fp16")];
            fp32 norm_127_epsilon_0 = const()[name = string("norm_127_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_127_cast_fp16 = rsqrt(epsilon = norm_127_epsilon_0, x = var_2915_cast_fp16)[name = string("norm_127_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_2917_cast_fp16 = mul(x = x_409_cast_fp16, y = norm_127_cast_fp16)[name = string("op_2917_cast_fp16")];
            tensor<fp16, [1024]> layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257462272)))];
            tensor<fp16, [1, ?, 1024]> var_2918_cast_fp16 = mul(x = var_2917_cast_fp16, y = layers_15_post_attention_layernorm_weight_to_fp16)[name = string("op_2918_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257464384))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260610176))))[name = string("layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = var_2918_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_2928_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_2928_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260806848))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263952640))))[name = string("layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_15_mlp_up_proj_weight_to_fp16_quantized, x = var_2918_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_159_cast_fp16 = mul(x = var_2928_cast_fp16, y = linear_110_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264149312))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267295104))))[name = string("layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_111_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_415_cast_fp16 = add(x = x_409_cast_fp16, y = linear_111_cast_fp16)[name = string("x_415_cast_fp16")];
            int32 var_2948 = const()[name = string("op_2948"), val = int32(-1)];
            fp16 var_2947_promoted_to_fp16 = const()[name = string("op_2947_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_2957_cast_fp16 = pow(x = x_415_cast_fp16, y = var_2947_promoted_to_fp16)[name = string("op_2957_cast_fp16")];
            tensor<int32, [1]> var_2959_axes_0 = const()[name = string("op_2959_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2959_keep_dims_0 = const()[name = string("op_2959_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_2959_cast_fp16 = reduce_mean(axes = var_2959_axes_0, keep_dims = var_2959_keep_dims_0, x = var_2957_cast_fp16)[name = string("op_2959_cast_fp16")];
            fp16 var_2960_to_fp16 = const()[name = string("op_2960_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_2961_cast_fp16 = add(x = var_2959_cast_fp16, y = var_2960_to_fp16)[name = string("op_2961_cast_fp16")];
            fp32 norm_129_epsilon_0 = const()[name = string("norm_129_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_129_cast_fp16 = rsqrt(epsilon = norm_129_epsilon_0, x = var_2961_cast_fp16)[name = string("norm_129_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_2963_cast_fp16 = mul(x = x_415_cast_fp16, y = norm_129_cast_fp16)[name = string("op_2963_cast_fp16")];
            tensor<fp16, [1024]> layers_16_input_layernorm_weight_to_fp16 = const()[name = string("layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267491776)))];
            tensor<fp16, [1, ?, 1024]> var_2964_cast_fp16 = mul(x = var_2963_cast_fp16, y = layers_16_input_layernorm_weight_to_fp16)[name = string("op_2964_cast_fp16")];
            tensor<int32, [3]> var_2976_shape_cast_fp16 = shape(x = var_2964_cast_fp16)[name = string("op_2976_shape_cast_fp16")];
            int32 gather_96 = const()[name = string("gather_96"), val = int32(1)];
            int32 gather_97_axis_0 = const()[name = string("gather_97_axis_0"), val = int32(0)];
            int32 gather_97_batch_dims_0 = const()[name = string("gather_97_batch_dims_0"), val = int32(0)];
            bool gather_97_validate_indices_0 = const()[name = string("gather_97_validate_indices_0"), val = bool(false)];
            string var_2976_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_2976_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_97_to_uint16 = const()[name = string("select_97_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_2976_shape_cast_fp16_to_uint16 = cast(dtype = var_2976_shape_cast_fp16_to_uint16_dtype_0, x = var_2976_shape_cast_fp16)[name = string("cast_23")];
            uint16 gather_97_cast_uint16 = gather(axis = gather_97_axis_0, batch_dims = gather_97_batch_dims_0, indices = select_97_to_uint16, validate_indices = gather_97_validate_indices_0, x = var_2976_shape_cast_fp16_to_uint16)[name = string("gather_97_cast_uint16")];
            string gather_97_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_97_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267493888))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269591104))))[name = string("layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_112_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = var_2964_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [4]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_2981_cast_fp16 = reshape(shape = concat_96x, x = linear_112_cast_fp16)[name = string("op_2981_cast_fp16")];
            tensor<int32, [4]> x_421_perm_0 = const()[name = string("x_421_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269722240))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270770880))))[name = string("layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_113_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = var_2964_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_2986_cast_fp16 = reshape(shape = concat_97x, x = linear_113_cast_fp16)[name = string("op_2986_cast_fp16")];
            tensor<int32, [4]> x_425_perm_0 = const()[name = string("x_425_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270836480))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271885120))))[name = string("layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_114_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = var_2964_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [4]> concat_98x = const()[name = string("concat_98x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_2991_cast_fp16 = reshape(shape = concat_98x, x = linear_114_cast_fp16)[name = string("op_2991_cast_fp16")];
            tensor<int32, [4]> v_33_perm_0 = const()[name = string("v_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2947_promoted_1_to_fp16 = const()[name = string("op_2947_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_421_cast_fp16 = transpose(perm = x_421_perm_0, x = var_2981_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 16, ?, 128]> var_2995_cast_fp16 = pow(x = x_421_cast_fp16, y = var_2947_promoted_1_to_fp16)[name = string("op_2995_cast_fp16")];
            tensor<int32, [1]> var_2997_axes_0 = const()[name = string("op_2997_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2997_keep_dims_0 = const()[name = string("op_2997_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_2997_cast_fp16 = reduce_mean(axes = var_2997_axes_0, keep_dims = var_2997_keep_dims_0, x = var_2995_cast_fp16)[name = string("op_2997_cast_fp16")];
            fp16 var_2998_to_fp16 = const()[name = string("op_2998_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_2999_cast_fp16 = add(x = var_2997_cast_fp16, y = var_2998_to_fp16)[name = string("op_2999_cast_fp16")];
            fp32 norm_131_epsilon_0 = const()[name = string("norm_131_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_131_cast_fp16 = rsqrt(epsilon = norm_131_epsilon_0, x = var_2999_cast_fp16)[name = string("norm_131_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_3001_cast_fp16 = mul(x = x_421_cast_fp16, y = norm_131_cast_fp16)[name = string("op_3001_cast_fp16")];
            tensor<fp16, [128]> layers_16_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_16_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271950720)))];
            tensor<fp16, [1, 16, ?, 128]> var_3002_cast_fp16 = mul(x = var_3001_cast_fp16, y = layers_16_self_attn_q_norm_weight_to_fp16)[name = string("op_3002_cast_fp16")];
            fp16 var_2947_promoted_2_to_fp16 = const()[name = string("op_2947_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_425_cast_fp16 = transpose(perm = x_425_perm_0, x = var_2986_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 8, ?, 128]> var_3006_cast_fp16 = pow(x = x_425_cast_fp16, y = var_2947_promoted_2_to_fp16)[name = string("op_3006_cast_fp16")];
            tensor<int32, [1]> var_3008_axes_0 = const()[name = string("op_3008_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3008_keep_dims_0 = const()[name = string("op_3008_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_3008_cast_fp16 = reduce_mean(axes = var_3008_axes_0, keep_dims = var_3008_keep_dims_0, x = var_3006_cast_fp16)[name = string("op_3008_cast_fp16")];
            fp16 var_3009_to_fp16 = const()[name = string("op_3009_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_3010_cast_fp16 = add(x = var_3008_cast_fp16, y = var_3009_to_fp16)[name = string("op_3010_cast_fp16")];
            fp32 norm_133_epsilon_0 = const()[name = string("norm_133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_133_cast_fp16 = rsqrt(epsilon = norm_133_epsilon_0, x = var_3010_cast_fp16)[name = string("norm_133_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_3012_cast_fp16 = mul(x = x_425_cast_fp16, y = norm_133_cast_fp16)[name = string("op_3012_cast_fp16")];
            tensor<fp16, [128]> layers_16_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_16_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271951040)))];
            tensor<fp16, [1, 8, ?, 128]> var_3013_cast_fp16 = mul(x = var_3012_cast_fp16, y = layers_16_self_attn_k_norm_weight_to_fp16)[name = string("op_3013_cast_fp16")];
            tensor<int32, [4]> x1_65_begin_0 = const()[name = string("x1_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_65_end_0 = const()[name = string("x1_65_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_65_end_mask_0 = const()[name = string("x1_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_65_cast_fp16 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = var_3002_cast_fp16)[name = string("x1_65_cast_fp16")];
            tensor<int32, [4]> x2_65_begin_0 = const()[name = string("x2_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_65_end_0 = const()[name = string("x2_65_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_65_end_mask_0 = const()[name = string("x2_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_65_cast_fp16 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = var_3002_cast_fp16)[name = string("x2_65_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3029_cast_fp16 = mul(x = x1_65_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3029_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3030_cast_fp16 = mul(x = x2_65_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3030_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3031_cast_fp16 = sub(x = var_3029_cast_fp16, y = var_3030_cast_fp16)[name = string("op_3031_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3032_cast_fp16 = mul(x = x2_65_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3032_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3033_cast_fp16 = mul(x = x1_65_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3033_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3034_cast_fp16 = add(x = var_3032_cast_fp16, y = var_3033_cast_fp16)[name = string("op_3034_cast_fp16")];
            bool q_33_interleave_0 = const()[name = string("q_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_33_cast_fp16 = concat(axis = var_2948, interleave = q_33_interleave_0, values = (var_3031_cast_fp16, var_3034_cast_fp16))[name = string("q_33_cast_fp16")];
            tensor<int32, [4]> x1_67_begin_0 = const()[name = string("x1_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_67_end_0 = const()[name = string("x1_67_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_67_end_mask_0 = const()[name = string("x1_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_67_cast_fp16 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = var_3013_cast_fp16)[name = string("x1_67_cast_fp16")];
            tensor<int32, [4]> x2_67_begin_0 = const()[name = string("x2_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_67_end_0 = const()[name = string("x2_67_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_67_end_mask_0 = const()[name = string("x2_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_67_cast_fp16 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = var_3013_cast_fp16)[name = string("x2_67_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3051_cast_fp16 = mul(x = x1_67_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3051_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3052_cast_fp16 = mul(x = x2_67_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3052_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3053_cast_fp16 = sub(x = var_3051_cast_fp16, y = var_3052_cast_fp16)[name = string("op_3053_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3054_cast_fp16 = mul(x = x2_67_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3054_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3055_cast_fp16 = mul(x = x1_67_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3055_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3056_cast_fp16 = add(x = var_3054_cast_fp16, y = var_3055_cast_fp16)[name = string("op_3056_cast_fp16")];
            bool k_33_interleave_0 = const()[name = string("k_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_33_cast_fp16 = concat(axis = var_2948, interleave = k_33_interleave_0, values = (var_3053_cast_fp16, var_3056_cast_fp16))[name = string("k_33_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_32 = read_state(input = k_cache_16)[name = string("read_state_32")];
            int32 k_cache_67_axis_0 = const()[name = string("k_cache_67_axis_0"), val = int32(2)];
            string k_cache_67_mode_0 = const()[name = string("k_cache_67_mode_0"), val = string("update")];
            bool k_cache_67_validate_indices_0 = const()[name = string("k_cache_67_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_67_cast_fp16 = scatter_along_axis(axis = k_cache_67_axis_0, data = read_state_32, indices = scatter_indices, mode = k_cache_67_mode_0, updates = k_33_cast_fp16, validate_indices = k_cache_67_validate_indices_0)[name = string("k_cache_67_cast_fp16")];
            write_state(data = k_cache_67_cast_fp16, input = k_cache_16)[name = string("coreml_update_state_32_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_88 = read_state(input = k_cache_16)[name = string("coreml_update_state_32")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_33 = read_state(input = v_cache_16)[name = string("read_state_33")];
            int32 v_cache_67_axis_0 = const()[name = string("v_cache_67_axis_0"), val = int32(2)];
            string v_cache_67_mode_0 = const()[name = string("v_cache_67_mode_0"), val = string("update")];
            bool v_cache_67_validate_indices_0 = const()[name = string("v_cache_67_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_33_cast_fp16 = transpose(perm = v_33_perm_0, x = var_2991_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_67_cast_fp16 = scatter_along_axis(axis = v_cache_67_axis_0, data = read_state_33, indices = scatter_indices, mode = v_cache_67_mode_0, updates = v_33_cast_fp16, validate_indices = v_cache_67_validate_indices_0)[name = string("v_cache_67_cast_fp16")];
            write_state(data = v_cache_67_cast_fp16, input = v_cache_16)[name = string("coreml_update_state_33_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_89 = read_state(input = v_cache_16)[name = string("coreml_update_state_33")];
            tensor<int32, [1]> var_3061_axes_0 = const()[name = string("op_3061_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3061_cast_fp16 = expand_dims(axes = var_3061_axes_0, x = coreml_update_state_88)[name = string("op_3061_cast_fp16")];
            tensor<int32, [5]> k_exp_65_reps_0 = const()[name = string("k_exp_65_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_65_cast_fp16 = tile(reps = k_exp_65_reps_0, x = var_3061_cast_fp16)[name = string("k_exp_65_cast_fp16")];
            tensor<int32, [4]> var_3064 = const()[name = string("op_3064"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_67_cast_fp16 = reshape(shape = var_3064, x = k_exp_65_cast_fp16)[name = string("k_exp_67_cast_fp16")];
            tensor<int32, [1]> var_3066_axes_0 = const()[name = string("op_3066_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3066_cast_fp16 = expand_dims(axes = var_3066_axes_0, x = coreml_update_state_89)[name = string("op_3066_cast_fp16")];
            tensor<int32, [5]> v_exp_65_reps_0 = const()[name = string("v_exp_65_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_65_cast_fp16 = tile(reps = v_exp_65_reps_0, x = var_3066_cast_fp16)[name = string("v_exp_65_cast_fp16")];
            tensor<int32, [4]> var_3069 = const()[name = string("op_3069"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_67_cast_fp16 = reshape(shape = var_3069, x = v_exp_65_cast_fp16)[name = string("v_exp_67_cast_fp16")];
            bool var_3072_transpose_x_1 = const()[name = string("op_3072_transpose_x_1"), val = bool(false)];
            bool var_3072_transpose_y_1 = const()[name = string("op_3072_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_3072_cast_fp16 = matmul(transpose_x = var_3072_transpose_x_1, transpose_y = var_3072_transpose_y_1, x = q_33_cast_fp16, y = k_exp_67_cast_fp16)[name = string("op_3072_cast_fp16")];
            fp16 var_3073_to_fp16 = const()[name = string("op_3073_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_65_cast_fp16 = mul(x = var_3072_cast_fp16, y = var_3073_to_fp16)[name = string("attn_65_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_161_cast_fp16 = add(x = attn_65_cast_fp16, y = attention_mask)[name = string("input_161_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_67_cast_fp16 = softmax(axis = var_2948, x = input_161_cast_fp16)[name = string("attn_67_cast_fp16")];
            bool out_33_transpose_x_0 = const()[name = string("out_33_transpose_x_0"), val = bool(false)];
            bool out_33_transpose_y_0 = const()[name = string("out_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_33_cast_fp16 = matmul(transpose_x = out_33_transpose_x_0, transpose_y = out_33_transpose_y_0, x = attn_67_cast_fp16, y = v_exp_67_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [4]> var_3078_perm_0 = const()[name = string("op_3078_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_101_axis_0 = const()[name = string("concat_101_axis_0"), val = int32(0)];
            bool concat_101_interleave_0 = const()[name = string("concat_101_interleave_0"), val = bool(false)];
            int32 gather_97_cast_uint16_to_int32 = cast(dtype = gather_97_cast_uint16_to_int32_dtype_0, x = gather_97_cast_uint16)[name = string("cast_22")];
            tensor<int32, [3]> concat_101 = concat(axis = concat_101_axis_0, interleave = concat_101_interleave_0, values = (gather_96, gather_97_cast_uint16_to_int32, var_2948))[name = string("concat_101")];
            tensor<fp16, [1, ?, 16, 128]> var_3078_cast_fp16 = transpose(perm = var_3078_perm_0, x = out_33_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, ?, ?]> input_163_cast_fp16 = reshape(shape = concat_101, x = var_3078_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271951360))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274048576))))[name = string("layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_115_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_435_cast_fp16 = add(x = x_415_cast_fp16, y = linear_115_cast_fp16)[name = string("x_435_cast_fp16")];
            fp16 var_2947_promoted_3_to_fp16 = const()[name = string("op_2947_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_3086_cast_fp16 = pow(x = x_435_cast_fp16, y = var_2947_promoted_3_to_fp16)[name = string("op_3086_cast_fp16")];
            tensor<int32, [1]> var_3088_axes_0 = const()[name = string("op_3088_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3088_keep_dims_0 = const()[name = string("op_3088_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_3088_cast_fp16 = reduce_mean(axes = var_3088_axes_0, keep_dims = var_3088_keep_dims_0, x = var_3086_cast_fp16)[name = string("op_3088_cast_fp16")];
            fp16 var_3089_to_fp16 = const()[name = string("op_3089_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3090_cast_fp16 = add(x = var_3088_cast_fp16, y = var_3089_to_fp16)[name = string("op_3090_cast_fp16")];
            fp32 norm_135_epsilon_0 = const()[name = string("norm_135_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_135_cast_fp16 = rsqrt(epsilon = norm_135_epsilon_0, x = var_3090_cast_fp16)[name = string("norm_135_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_3092_cast_fp16 = mul(x = x_435_cast_fp16, y = norm_135_cast_fp16)[name = string("op_3092_cast_fp16")];
            tensor<fp16, [1024]> layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274179712)))];
            tensor<fp16, [1, ?, 1024]> var_3093_cast_fp16 = mul(x = var_3092_cast_fp16, y = layers_16_post_attention_layernorm_weight_to_fp16)[name = string("op_3093_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274181824))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277327616))))[name = string("layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = var_3093_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_3103_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_3103_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277524288))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280670080))))[name = string("layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_16_mlp_up_proj_weight_to_fp16_quantized, x = var_3093_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_169_cast_fp16 = mul(x = var_3103_cast_fp16, y = linear_117_cast_fp16)[name = string("input_169_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280866752))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284012544))))[name = string("layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_441_cast_fp16 = add(x = x_435_cast_fp16, y = linear_118_cast_fp16)[name = string("x_441_cast_fp16")];
            int32 var_3123 = const()[name = string("op_3123"), val = int32(-1)];
            fp16 var_3122_promoted_to_fp16 = const()[name = string("op_3122_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_3132_cast_fp16 = pow(x = x_441_cast_fp16, y = var_3122_promoted_to_fp16)[name = string("op_3132_cast_fp16")];
            tensor<int32, [1]> var_3134_axes_0 = const()[name = string("op_3134_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3134_keep_dims_0 = const()[name = string("op_3134_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_3134_cast_fp16 = reduce_mean(axes = var_3134_axes_0, keep_dims = var_3134_keep_dims_0, x = var_3132_cast_fp16)[name = string("op_3134_cast_fp16")];
            fp16 var_3135_to_fp16 = const()[name = string("op_3135_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3136_cast_fp16 = add(x = var_3134_cast_fp16, y = var_3135_to_fp16)[name = string("op_3136_cast_fp16")];
            fp32 norm_137_epsilon_0 = const()[name = string("norm_137_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_137_cast_fp16 = rsqrt(epsilon = norm_137_epsilon_0, x = var_3136_cast_fp16)[name = string("norm_137_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_3138_cast_fp16 = mul(x = x_441_cast_fp16, y = norm_137_cast_fp16)[name = string("op_3138_cast_fp16")];
            tensor<fp16, [1024]> layers_17_input_layernorm_weight_to_fp16 = const()[name = string("layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284209216)))];
            tensor<fp16, [1, ?, 1024]> var_3139_cast_fp16 = mul(x = var_3138_cast_fp16, y = layers_17_input_layernorm_weight_to_fp16)[name = string("op_3139_cast_fp16")];
            tensor<int32, [3]> var_3151_shape_cast_fp16 = shape(x = var_3139_cast_fp16)[name = string("op_3151_shape_cast_fp16")];
            int32 gather_102 = const()[name = string("gather_102"), val = int32(1)];
            int32 gather_103_axis_0 = const()[name = string("gather_103_axis_0"), val = int32(0)];
            int32 gather_103_batch_dims_0 = const()[name = string("gather_103_batch_dims_0"), val = int32(0)];
            bool gather_103_validate_indices_0 = const()[name = string("gather_103_validate_indices_0"), val = bool(false)];
            string var_3151_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3151_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_103_to_uint16 = const()[name = string("select_103_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_3151_shape_cast_fp16_to_uint16 = cast(dtype = var_3151_shape_cast_fp16_to_uint16_dtype_0, x = var_3151_shape_cast_fp16)[name = string("cast_21")];
            uint16 gather_103_cast_uint16 = gather(axis = gather_103_axis_0, batch_dims = gather_103_batch_dims_0, indices = select_103_to_uint16, validate_indices = gather_103_validate_indices_0, x = var_3151_shape_cast_fp16_to_uint16)[name = string("gather_103_cast_uint16")];
            string gather_103_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_103_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284211328))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286308544))))[name = string("layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_119_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = var_3139_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> concat_102x = const()[name = string("concat_102x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_3156_cast_fp16 = reshape(shape = concat_102x, x = linear_119_cast_fp16)[name = string("op_3156_cast_fp16")];
            tensor<int32, [4]> x_447_perm_0 = const()[name = string("x_447_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286439680))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287488320))))[name = string("layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_120_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = var_3139_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> concat_103x = const()[name = string("concat_103x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_3161_cast_fp16 = reshape(shape = concat_103x, x = linear_120_cast_fp16)[name = string("op_3161_cast_fp16")];
            tensor<int32, [4]> x_451_perm_0 = const()[name = string("x_451_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287553920))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288602560))))[name = string("layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_121_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = var_3139_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> concat_104x = const()[name = string("concat_104x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_3166_cast_fp16 = reshape(shape = concat_104x, x = linear_121_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3122_promoted_1_to_fp16 = const()[name = string("op_3122_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_447_cast_fp16 = transpose(perm = x_447_perm_0, x = var_3156_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 16, ?, 128]> var_3170_cast_fp16 = pow(x = x_447_cast_fp16, y = var_3122_promoted_1_to_fp16)[name = string("op_3170_cast_fp16")];
            tensor<int32, [1]> var_3172_axes_0 = const()[name = string("op_3172_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3172_keep_dims_0 = const()[name = string("op_3172_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_3172_cast_fp16 = reduce_mean(axes = var_3172_axes_0, keep_dims = var_3172_keep_dims_0, x = var_3170_cast_fp16)[name = string("op_3172_cast_fp16")];
            fp16 var_3173_to_fp16 = const()[name = string("op_3173_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_3174_cast_fp16 = add(x = var_3172_cast_fp16, y = var_3173_to_fp16)[name = string("op_3174_cast_fp16")];
            fp32 norm_139_epsilon_0 = const()[name = string("norm_139_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_139_cast_fp16 = rsqrt(epsilon = norm_139_epsilon_0, x = var_3174_cast_fp16)[name = string("norm_139_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_3176_cast_fp16 = mul(x = x_447_cast_fp16, y = norm_139_cast_fp16)[name = string("op_3176_cast_fp16")];
            tensor<fp16, [128]> layers_17_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_17_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288668160)))];
            tensor<fp16, [1, 16, ?, 128]> var_3177_cast_fp16 = mul(x = var_3176_cast_fp16, y = layers_17_self_attn_q_norm_weight_to_fp16)[name = string("op_3177_cast_fp16")];
            fp16 var_3122_promoted_2_to_fp16 = const()[name = string("op_3122_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_451_cast_fp16 = transpose(perm = x_451_perm_0, x = var_3161_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 8, ?, 128]> var_3181_cast_fp16 = pow(x = x_451_cast_fp16, y = var_3122_promoted_2_to_fp16)[name = string("op_3181_cast_fp16")];
            tensor<int32, [1]> var_3183_axes_0 = const()[name = string("op_3183_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3183_keep_dims_0 = const()[name = string("op_3183_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_3183_cast_fp16 = reduce_mean(axes = var_3183_axes_0, keep_dims = var_3183_keep_dims_0, x = var_3181_cast_fp16)[name = string("op_3183_cast_fp16")];
            fp16 var_3184_to_fp16 = const()[name = string("op_3184_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_3185_cast_fp16 = add(x = var_3183_cast_fp16, y = var_3184_to_fp16)[name = string("op_3185_cast_fp16")];
            fp32 norm_141_epsilon_0 = const()[name = string("norm_141_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_141_cast_fp16 = rsqrt(epsilon = norm_141_epsilon_0, x = var_3185_cast_fp16)[name = string("norm_141_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_3187_cast_fp16 = mul(x = x_451_cast_fp16, y = norm_141_cast_fp16)[name = string("op_3187_cast_fp16")];
            tensor<fp16, [128]> layers_17_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_17_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288668480)))];
            tensor<fp16, [1, 8, ?, 128]> var_3188_cast_fp16 = mul(x = var_3187_cast_fp16, y = layers_17_self_attn_k_norm_weight_to_fp16)[name = string("op_3188_cast_fp16")];
            tensor<int32, [4]> x1_69_begin_0 = const()[name = string("x1_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_69_end_0 = const()[name = string("x1_69_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_69_end_mask_0 = const()[name = string("x1_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_69_cast_fp16 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = var_3177_cast_fp16)[name = string("x1_69_cast_fp16")];
            tensor<int32, [4]> x2_69_begin_0 = const()[name = string("x2_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_69_end_0 = const()[name = string("x2_69_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_69_end_mask_0 = const()[name = string("x2_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_69_cast_fp16 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = var_3177_cast_fp16)[name = string("x2_69_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3204_cast_fp16 = mul(x = x1_69_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3204_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3205_cast_fp16 = mul(x = x2_69_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3205_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3206_cast_fp16 = sub(x = var_3204_cast_fp16, y = var_3205_cast_fp16)[name = string("op_3206_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3207_cast_fp16 = mul(x = x2_69_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3207_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3208_cast_fp16 = mul(x = x1_69_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3208_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3209_cast_fp16 = add(x = var_3207_cast_fp16, y = var_3208_cast_fp16)[name = string("op_3209_cast_fp16")];
            bool q_35_interleave_0 = const()[name = string("q_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_35_cast_fp16 = concat(axis = var_3123, interleave = q_35_interleave_0, values = (var_3206_cast_fp16, var_3209_cast_fp16))[name = string("q_35_cast_fp16")];
            tensor<int32, [4]> x1_71_begin_0 = const()[name = string("x1_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_71_end_0 = const()[name = string("x1_71_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_71_end_mask_0 = const()[name = string("x1_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_71_cast_fp16 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = var_3188_cast_fp16)[name = string("x1_71_cast_fp16")];
            tensor<int32, [4]> x2_71_begin_0 = const()[name = string("x2_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_71_end_0 = const()[name = string("x2_71_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_71_end_mask_0 = const()[name = string("x2_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_71_cast_fp16 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = var_3188_cast_fp16)[name = string("x2_71_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3226_cast_fp16 = mul(x = x1_71_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3226_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3227_cast_fp16 = mul(x = x2_71_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3227_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3228_cast_fp16 = sub(x = var_3226_cast_fp16, y = var_3227_cast_fp16)[name = string("op_3228_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3229_cast_fp16 = mul(x = x2_71_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3229_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3230_cast_fp16 = mul(x = x1_71_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3230_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3231_cast_fp16 = add(x = var_3229_cast_fp16, y = var_3230_cast_fp16)[name = string("op_3231_cast_fp16")];
            bool k_35_interleave_0 = const()[name = string("k_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_35_cast_fp16 = concat(axis = var_3123, interleave = k_35_interleave_0, values = (var_3228_cast_fp16, var_3231_cast_fp16))[name = string("k_35_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_34 = read_state(input = k_cache_17)[name = string("read_state_34")];
            int32 k_cache_71_axis_0 = const()[name = string("k_cache_71_axis_0"), val = int32(2)];
            string k_cache_71_mode_0 = const()[name = string("k_cache_71_mode_0"), val = string("update")];
            bool k_cache_71_validate_indices_0 = const()[name = string("k_cache_71_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_71_cast_fp16 = scatter_along_axis(axis = k_cache_71_axis_0, data = read_state_34, indices = scatter_indices, mode = k_cache_71_mode_0, updates = k_35_cast_fp16, validate_indices = k_cache_71_validate_indices_0)[name = string("k_cache_71_cast_fp16")];
            write_state(data = k_cache_71_cast_fp16, input = k_cache_17)[name = string("coreml_update_state_34_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_90 = read_state(input = k_cache_17)[name = string("coreml_update_state_34")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_35 = read_state(input = v_cache_17)[name = string("read_state_35")];
            int32 v_cache_71_axis_0 = const()[name = string("v_cache_71_axis_0"), val = int32(2)];
            string v_cache_71_mode_0 = const()[name = string("v_cache_71_mode_0"), val = string("update")];
            bool v_cache_71_validate_indices_0 = const()[name = string("v_cache_71_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = var_3166_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_71_cast_fp16 = scatter_along_axis(axis = v_cache_71_axis_0, data = read_state_35, indices = scatter_indices, mode = v_cache_71_mode_0, updates = v_35_cast_fp16, validate_indices = v_cache_71_validate_indices_0)[name = string("v_cache_71_cast_fp16")];
            write_state(data = v_cache_71_cast_fp16, input = v_cache_17)[name = string("coreml_update_state_35_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_91 = read_state(input = v_cache_17)[name = string("coreml_update_state_35")];
            tensor<int32, [1]> var_3236_axes_0 = const()[name = string("op_3236_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3236_cast_fp16 = expand_dims(axes = var_3236_axes_0, x = coreml_update_state_90)[name = string("op_3236_cast_fp16")];
            tensor<int32, [5]> k_exp_69_reps_0 = const()[name = string("k_exp_69_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_69_cast_fp16 = tile(reps = k_exp_69_reps_0, x = var_3236_cast_fp16)[name = string("k_exp_69_cast_fp16")];
            tensor<int32, [4]> var_3239 = const()[name = string("op_3239"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_71_cast_fp16 = reshape(shape = var_3239, x = k_exp_69_cast_fp16)[name = string("k_exp_71_cast_fp16")];
            tensor<int32, [1]> var_3241_axes_0 = const()[name = string("op_3241_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3241_cast_fp16 = expand_dims(axes = var_3241_axes_0, x = coreml_update_state_91)[name = string("op_3241_cast_fp16")];
            tensor<int32, [5]> v_exp_69_reps_0 = const()[name = string("v_exp_69_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_69_cast_fp16 = tile(reps = v_exp_69_reps_0, x = var_3241_cast_fp16)[name = string("v_exp_69_cast_fp16")];
            tensor<int32, [4]> var_3244 = const()[name = string("op_3244"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_71_cast_fp16 = reshape(shape = var_3244, x = v_exp_69_cast_fp16)[name = string("v_exp_71_cast_fp16")];
            bool var_3247_transpose_x_1 = const()[name = string("op_3247_transpose_x_1"), val = bool(false)];
            bool var_3247_transpose_y_1 = const()[name = string("op_3247_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_3247_cast_fp16 = matmul(transpose_x = var_3247_transpose_x_1, transpose_y = var_3247_transpose_y_1, x = q_35_cast_fp16, y = k_exp_71_cast_fp16)[name = string("op_3247_cast_fp16")];
            fp16 var_3248_to_fp16 = const()[name = string("op_3248_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_69_cast_fp16 = mul(x = var_3247_cast_fp16, y = var_3248_to_fp16)[name = string("attn_69_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_171_cast_fp16 = add(x = attn_69_cast_fp16, y = attention_mask)[name = string("input_171_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_71_cast_fp16 = softmax(axis = var_3123, x = input_171_cast_fp16)[name = string("attn_71_cast_fp16")];
            bool out_35_transpose_x_0 = const()[name = string("out_35_transpose_x_0"), val = bool(false)];
            bool out_35_transpose_y_0 = const()[name = string("out_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_35_cast_fp16 = matmul(transpose_x = out_35_transpose_x_0, transpose_y = out_35_transpose_y_0, x = attn_71_cast_fp16, y = v_exp_71_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [4]> var_3253_perm_0 = const()[name = string("op_3253_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_107_axis_0 = const()[name = string("concat_107_axis_0"), val = int32(0)];
            bool concat_107_interleave_0 = const()[name = string("concat_107_interleave_0"), val = bool(false)];
            int32 gather_103_cast_uint16_to_int32 = cast(dtype = gather_103_cast_uint16_to_int32_dtype_0, x = gather_103_cast_uint16)[name = string("cast_20")];
            tensor<int32, [3]> concat_107 = concat(axis = concat_107_axis_0, interleave = concat_107_interleave_0, values = (gather_102, gather_103_cast_uint16_to_int32, var_3123))[name = string("concat_107")];
            tensor<fp16, [1, ?, 16, 128]> var_3253_cast_fp16 = transpose(perm = var_3253_perm_0, x = out_35_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, ?, ?]> input_173_cast_fp16 = reshape(shape = concat_107, x = var_3253_cast_fp16)[name = string("input_173_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288668800))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290766016))))[name = string("layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_122_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_461_cast_fp16 = add(x = x_441_cast_fp16, y = linear_122_cast_fp16)[name = string("x_461_cast_fp16")];
            fp16 var_3122_promoted_3_to_fp16 = const()[name = string("op_3122_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_3261_cast_fp16 = pow(x = x_461_cast_fp16, y = var_3122_promoted_3_to_fp16)[name = string("op_3261_cast_fp16")];
            tensor<int32, [1]> var_3263_axes_0 = const()[name = string("op_3263_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3263_keep_dims_0 = const()[name = string("op_3263_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_3263_cast_fp16 = reduce_mean(axes = var_3263_axes_0, keep_dims = var_3263_keep_dims_0, x = var_3261_cast_fp16)[name = string("op_3263_cast_fp16")];
            fp16 var_3264_to_fp16 = const()[name = string("op_3264_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3265_cast_fp16 = add(x = var_3263_cast_fp16, y = var_3264_to_fp16)[name = string("op_3265_cast_fp16")];
            fp32 norm_143_epsilon_0 = const()[name = string("norm_143_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_143_cast_fp16 = rsqrt(epsilon = norm_143_epsilon_0, x = var_3265_cast_fp16)[name = string("norm_143_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_3267_cast_fp16 = mul(x = x_461_cast_fp16, y = norm_143_cast_fp16)[name = string("op_3267_cast_fp16")];
            tensor<fp16, [1024]> layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290897152)))];
            tensor<fp16, [1, ?, 1024]> var_3268_cast_fp16 = mul(x = var_3267_cast_fp16, y = layers_17_post_attention_layernorm_weight_to_fp16)[name = string("op_3268_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290899264))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294045056))))[name = string("layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = var_3268_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_3278_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_3278_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294241728))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297387520))))[name = string("layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_124_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_17_mlp_up_proj_weight_to_fp16_quantized, x = var_3268_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_179_cast_fp16 = mul(x = var_3278_cast_fp16, y = linear_124_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297584192))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300729984))))[name = string("layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_125_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_467_cast_fp16 = add(x = x_461_cast_fp16, y = linear_125_cast_fp16)[name = string("x_467_cast_fp16")];
            int32 var_3298 = const()[name = string("op_3298"), val = int32(-1)];
            fp16 var_3297_promoted_to_fp16 = const()[name = string("op_3297_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_3307_cast_fp16 = pow(x = x_467_cast_fp16, y = var_3297_promoted_to_fp16)[name = string("op_3307_cast_fp16")];
            tensor<int32, [1]> var_3309_axes_0 = const()[name = string("op_3309_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3309_keep_dims_0 = const()[name = string("op_3309_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_3309_cast_fp16 = reduce_mean(axes = var_3309_axes_0, keep_dims = var_3309_keep_dims_0, x = var_3307_cast_fp16)[name = string("op_3309_cast_fp16")];
            fp16 var_3310_to_fp16 = const()[name = string("op_3310_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3311_cast_fp16 = add(x = var_3309_cast_fp16, y = var_3310_to_fp16)[name = string("op_3311_cast_fp16")];
            fp32 norm_145_epsilon_0 = const()[name = string("norm_145_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_145_cast_fp16 = rsqrt(epsilon = norm_145_epsilon_0, x = var_3311_cast_fp16)[name = string("norm_145_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_3313_cast_fp16 = mul(x = x_467_cast_fp16, y = norm_145_cast_fp16)[name = string("op_3313_cast_fp16")];
            tensor<fp16, [1024]> layers_18_input_layernorm_weight_to_fp16 = const()[name = string("layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300926656)))];
            tensor<fp16, [1, ?, 1024]> var_3314_cast_fp16 = mul(x = var_3313_cast_fp16, y = layers_18_input_layernorm_weight_to_fp16)[name = string("op_3314_cast_fp16")];
            tensor<int32, [3]> var_3326_shape_cast_fp16 = shape(x = var_3314_cast_fp16)[name = string("op_3326_shape_cast_fp16")];
            int32 gather_108 = const()[name = string("gather_108"), val = int32(1)];
            int32 gather_109_axis_0 = const()[name = string("gather_109_axis_0"), val = int32(0)];
            int32 gather_109_batch_dims_0 = const()[name = string("gather_109_batch_dims_0"), val = int32(0)];
            bool gather_109_validate_indices_0 = const()[name = string("gather_109_validate_indices_0"), val = bool(false)];
            string var_3326_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3326_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_109_to_uint16 = const()[name = string("select_109_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_3326_shape_cast_fp16_to_uint16 = cast(dtype = var_3326_shape_cast_fp16_to_uint16_dtype_0, x = var_3326_shape_cast_fp16)[name = string("cast_19")];
            uint16 gather_109_cast_uint16 = gather(axis = gather_109_axis_0, batch_dims = gather_109_batch_dims_0, indices = select_109_to_uint16, validate_indices = gather_109_validate_indices_0, x = var_3326_shape_cast_fp16_to_uint16)[name = string("gather_109_cast_uint16")];
            string gather_109_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_109_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300928768))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303025984))))[name = string("layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_126_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = var_3314_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<int32, [4]> concat_108x = const()[name = string("concat_108x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_3331_cast_fp16 = reshape(shape = concat_108x, x = linear_126_cast_fp16)[name = string("op_3331_cast_fp16")];
            tensor<int32, [4]> x_473_perm_0 = const()[name = string("x_473_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303157120))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304205760))))[name = string("layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = var_3314_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_3336_cast_fp16 = reshape(shape = concat_109x, x = linear_127_cast_fp16)[name = string("op_3336_cast_fp16")];
            tensor<int32, [4]> x_477_perm_0 = const()[name = string("x_477_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304271360))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305320000))))[name = string("layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_128_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = var_3314_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> concat_110x = const()[name = string("concat_110x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_3341_cast_fp16 = reshape(shape = concat_110x, x = linear_128_cast_fp16)[name = string("op_3341_cast_fp16")];
            tensor<int32, [4]> v_37_perm_0 = const()[name = string("v_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3297_promoted_1_to_fp16 = const()[name = string("op_3297_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_473_cast_fp16 = transpose(perm = x_473_perm_0, x = var_3331_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 16, ?, 128]> var_3345_cast_fp16 = pow(x = x_473_cast_fp16, y = var_3297_promoted_1_to_fp16)[name = string("op_3345_cast_fp16")];
            tensor<int32, [1]> var_3347_axes_0 = const()[name = string("op_3347_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3347_keep_dims_0 = const()[name = string("op_3347_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_3347_cast_fp16 = reduce_mean(axes = var_3347_axes_0, keep_dims = var_3347_keep_dims_0, x = var_3345_cast_fp16)[name = string("op_3347_cast_fp16")];
            fp16 var_3348_to_fp16 = const()[name = string("op_3348_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_3349_cast_fp16 = add(x = var_3347_cast_fp16, y = var_3348_to_fp16)[name = string("op_3349_cast_fp16")];
            fp32 norm_147_epsilon_0 = const()[name = string("norm_147_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_147_cast_fp16 = rsqrt(epsilon = norm_147_epsilon_0, x = var_3349_cast_fp16)[name = string("norm_147_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_3351_cast_fp16 = mul(x = x_473_cast_fp16, y = norm_147_cast_fp16)[name = string("op_3351_cast_fp16")];
            tensor<fp16, [128]> layers_18_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_18_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305385600)))];
            tensor<fp16, [1, 16, ?, 128]> var_3352_cast_fp16 = mul(x = var_3351_cast_fp16, y = layers_18_self_attn_q_norm_weight_to_fp16)[name = string("op_3352_cast_fp16")];
            fp16 var_3297_promoted_2_to_fp16 = const()[name = string("op_3297_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_477_cast_fp16 = transpose(perm = x_477_perm_0, x = var_3336_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 8, ?, 128]> var_3356_cast_fp16 = pow(x = x_477_cast_fp16, y = var_3297_promoted_2_to_fp16)[name = string("op_3356_cast_fp16")];
            tensor<int32, [1]> var_3358_axes_0 = const()[name = string("op_3358_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3358_keep_dims_0 = const()[name = string("op_3358_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_3358_cast_fp16 = reduce_mean(axes = var_3358_axes_0, keep_dims = var_3358_keep_dims_0, x = var_3356_cast_fp16)[name = string("op_3358_cast_fp16")];
            fp16 var_3359_to_fp16 = const()[name = string("op_3359_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_3360_cast_fp16 = add(x = var_3358_cast_fp16, y = var_3359_to_fp16)[name = string("op_3360_cast_fp16")];
            fp32 norm_149_epsilon_0 = const()[name = string("norm_149_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_149_cast_fp16 = rsqrt(epsilon = norm_149_epsilon_0, x = var_3360_cast_fp16)[name = string("norm_149_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_3362_cast_fp16 = mul(x = x_477_cast_fp16, y = norm_149_cast_fp16)[name = string("op_3362_cast_fp16")];
            tensor<fp16, [128]> layers_18_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_18_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305385920)))];
            tensor<fp16, [1, 8, ?, 128]> var_3363_cast_fp16 = mul(x = var_3362_cast_fp16, y = layers_18_self_attn_k_norm_weight_to_fp16)[name = string("op_3363_cast_fp16")];
            tensor<int32, [4]> x1_73_begin_0 = const()[name = string("x1_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_73_end_0 = const()[name = string("x1_73_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_73_end_mask_0 = const()[name = string("x1_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_73_cast_fp16 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = var_3352_cast_fp16)[name = string("x1_73_cast_fp16")];
            tensor<int32, [4]> x2_73_begin_0 = const()[name = string("x2_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_73_end_0 = const()[name = string("x2_73_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_73_end_mask_0 = const()[name = string("x2_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_73_cast_fp16 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = var_3352_cast_fp16)[name = string("x2_73_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3379_cast_fp16 = mul(x = x1_73_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3379_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3380_cast_fp16 = mul(x = x2_73_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3380_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3381_cast_fp16 = sub(x = var_3379_cast_fp16, y = var_3380_cast_fp16)[name = string("op_3381_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3382_cast_fp16 = mul(x = x2_73_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3382_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3383_cast_fp16 = mul(x = x1_73_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3383_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3384_cast_fp16 = add(x = var_3382_cast_fp16, y = var_3383_cast_fp16)[name = string("op_3384_cast_fp16")];
            bool q_37_interleave_0 = const()[name = string("q_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_37_cast_fp16 = concat(axis = var_3298, interleave = q_37_interleave_0, values = (var_3381_cast_fp16, var_3384_cast_fp16))[name = string("q_37_cast_fp16")];
            tensor<int32, [4]> x1_75_begin_0 = const()[name = string("x1_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_75_end_0 = const()[name = string("x1_75_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_75_end_mask_0 = const()[name = string("x1_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_75_cast_fp16 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = var_3363_cast_fp16)[name = string("x1_75_cast_fp16")];
            tensor<int32, [4]> x2_75_begin_0 = const()[name = string("x2_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_75_end_0 = const()[name = string("x2_75_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_75_end_mask_0 = const()[name = string("x2_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_75_cast_fp16 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = var_3363_cast_fp16)[name = string("x2_75_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3401_cast_fp16 = mul(x = x1_75_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3401_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3402_cast_fp16 = mul(x = x2_75_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3402_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3403_cast_fp16 = sub(x = var_3401_cast_fp16, y = var_3402_cast_fp16)[name = string("op_3403_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3404_cast_fp16 = mul(x = x2_75_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3404_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3405_cast_fp16 = mul(x = x1_75_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3405_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3406_cast_fp16 = add(x = var_3404_cast_fp16, y = var_3405_cast_fp16)[name = string("op_3406_cast_fp16")];
            bool k_37_interleave_0 = const()[name = string("k_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_37_cast_fp16 = concat(axis = var_3298, interleave = k_37_interleave_0, values = (var_3403_cast_fp16, var_3406_cast_fp16))[name = string("k_37_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_36 = read_state(input = k_cache_18)[name = string("read_state_36")];
            int32 k_cache_75_axis_0 = const()[name = string("k_cache_75_axis_0"), val = int32(2)];
            string k_cache_75_mode_0 = const()[name = string("k_cache_75_mode_0"), val = string("update")];
            bool k_cache_75_validate_indices_0 = const()[name = string("k_cache_75_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_75_cast_fp16 = scatter_along_axis(axis = k_cache_75_axis_0, data = read_state_36, indices = scatter_indices, mode = k_cache_75_mode_0, updates = k_37_cast_fp16, validate_indices = k_cache_75_validate_indices_0)[name = string("k_cache_75_cast_fp16")];
            write_state(data = k_cache_75_cast_fp16, input = k_cache_18)[name = string("coreml_update_state_36_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_92 = read_state(input = k_cache_18)[name = string("coreml_update_state_36")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_37 = read_state(input = v_cache_18)[name = string("read_state_37")];
            int32 v_cache_75_axis_0 = const()[name = string("v_cache_75_axis_0"), val = int32(2)];
            string v_cache_75_mode_0 = const()[name = string("v_cache_75_mode_0"), val = string("update")];
            bool v_cache_75_validate_indices_0 = const()[name = string("v_cache_75_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_37_cast_fp16 = transpose(perm = v_37_perm_0, x = var_3341_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_75_cast_fp16 = scatter_along_axis(axis = v_cache_75_axis_0, data = read_state_37, indices = scatter_indices, mode = v_cache_75_mode_0, updates = v_37_cast_fp16, validate_indices = v_cache_75_validate_indices_0)[name = string("v_cache_75_cast_fp16")];
            write_state(data = v_cache_75_cast_fp16, input = v_cache_18)[name = string("coreml_update_state_37_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_93 = read_state(input = v_cache_18)[name = string("coreml_update_state_37")];
            tensor<int32, [1]> var_3411_axes_0 = const()[name = string("op_3411_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3411_cast_fp16 = expand_dims(axes = var_3411_axes_0, x = coreml_update_state_92)[name = string("op_3411_cast_fp16")];
            tensor<int32, [5]> k_exp_73_reps_0 = const()[name = string("k_exp_73_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_73_cast_fp16 = tile(reps = k_exp_73_reps_0, x = var_3411_cast_fp16)[name = string("k_exp_73_cast_fp16")];
            tensor<int32, [4]> var_3414 = const()[name = string("op_3414"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_75_cast_fp16 = reshape(shape = var_3414, x = k_exp_73_cast_fp16)[name = string("k_exp_75_cast_fp16")];
            tensor<int32, [1]> var_3416_axes_0 = const()[name = string("op_3416_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3416_cast_fp16 = expand_dims(axes = var_3416_axes_0, x = coreml_update_state_93)[name = string("op_3416_cast_fp16")];
            tensor<int32, [5]> v_exp_73_reps_0 = const()[name = string("v_exp_73_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_73_cast_fp16 = tile(reps = v_exp_73_reps_0, x = var_3416_cast_fp16)[name = string("v_exp_73_cast_fp16")];
            tensor<int32, [4]> var_3419 = const()[name = string("op_3419"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_75_cast_fp16 = reshape(shape = var_3419, x = v_exp_73_cast_fp16)[name = string("v_exp_75_cast_fp16")];
            bool var_3422_transpose_x_1 = const()[name = string("op_3422_transpose_x_1"), val = bool(false)];
            bool var_3422_transpose_y_1 = const()[name = string("op_3422_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_3422_cast_fp16 = matmul(transpose_x = var_3422_transpose_x_1, transpose_y = var_3422_transpose_y_1, x = q_37_cast_fp16, y = k_exp_75_cast_fp16)[name = string("op_3422_cast_fp16")];
            fp16 var_3423_to_fp16 = const()[name = string("op_3423_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_73_cast_fp16 = mul(x = var_3422_cast_fp16, y = var_3423_to_fp16)[name = string("attn_73_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_181_cast_fp16 = add(x = attn_73_cast_fp16, y = attention_mask)[name = string("input_181_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_75_cast_fp16 = softmax(axis = var_3298, x = input_181_cast_fp16)[name = string("attn_75_cast_fp16")];
            bool out_37_transpose_x_0 = const()[name = string("out_37_transpose_x_0"), val = bool(false)];
            bool out_37_transpose_y_0 = const()[name = string("out_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_37_cast_fp16 = matmul(transpose_x = out_37_transpose_x_0, transpose_y = out_37_transpose_y_0, x = attn_75_cast_fp16, y = v_exp_75_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [4]> var_3428_perm_0 = const()[name = string("op_3428_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_113_axis_0 = const()[name = string("concat_113_axis_0"), val = int32(0)];
            bool concat_113_interleave_0 = const()[name = string("concat_113_interleave_0"), val = bool(false)];
            int32 gather_109_cast_uint16_to_int32 = cast(dtype = gather_109_cast_uint16_to_int32_dtype_0, x = gather_109_cast_uint16)[name = string("cast_18")];
            tensor<int32, [3]> concat_113 = concat(axis = concat_113_axis_0, interleave = concat_113_interleave_0, values = (gather_108, gather_109_cast_uint16_to_int32, var_3298))[name = string("concat_113")];
            tensor<fp16, [1, ?, 16, 128]> var_3428_cast_fp16 = transpose(perm = var_3428_perm_0, x = out_37_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, ?, ?]> input_183_cast_fp16 = reshape(shape = concat_113, x = var_3428_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305386240))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307483456))))[name = string("layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_129_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_487_cast_fp16 = add(x = x_467_cast_fp16, y = linear_129_cast_fp16)[name = string("x_487_cast_fp16")];
            fp16 var_3297_promoted_3_to_fp16 = const()[name = string("op_3297_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_3436_cast_fp16 = pow(x = x_487_cast_fp16, y = var_3297_promoted_3_to_fp16)[name = string("op_3436_cast_fp16")];
            tensor<int32, [1]> var_3438_axes_0 = const()[name = string("op_3438_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3438_keep_dims_0 = const()[name = string("op_3438_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_3438_cast_fp16 = reduce_mean(axes = var_3438_axes_0, keep_dims = var_3438_keep_dims_0, x = var_3436_cast_fp16)[name = string("op_3438_cast_fp16")];
            fp16 var_3439_to_fp16 = const()[name = string("op_3439_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3440_cast_fp16 = add(x = var_3438_cast_fp16, y = var_3439_to_fp16)[name = string("op_3440_cast_fp16")];
            fp32 norm_151_epsilon_0 = const()[name = string("norm_151_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_151_cast_fp16 = rsqrt(epsilon = norm_151_epsilon_0, x = var_3440_cast_fp16)[name = string("norm_151_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_3442_cast_fp16 = mul(x = x_487_cast_fp16, y = norm_151_cast_fp16)[name = string("op_3442_cast_fp16")];
            tensor<fp16, [1024]> layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307614592)))];
            tensor<fp16, [1, ?, 1024]> var_3443_cast_fp16 = mul(x = var_3442_cast_fp16, y = layers_18_post_attention_layernorm_weight_to_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307616704))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310762496))))[name = string("layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = var_3443_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_3453_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_3453_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310959168))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314104960))))[name = string("layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_131_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_18_mlp_up_proj_weight_to_fp16_quantized, x = var_3443_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_189_cast_fp16 = mul(x = var_3453_cast_fp16, y = linear_131_cast_fp16)[name = string("input_189_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314301632))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317447424))))[name = string("layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_132_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_189_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_493_cast_fp16 = add(x = x_487_cast_fp16, y = linear_132_cast_fp16)[name = string("x_493_cast_fp16")];
            int32 var_3473 = const()[name = string("op_3473"), val = int32(-1)];
            fp16 var_3472_promoted_to_fp16 = const()[name = string("op_3472_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_3482_cast_fp16 = pow(x = x_493_cast_fp16, y = var_3472_promoted_to_fp16)[name = string("op_3482_cast_fp16")];
            tensor<int32, [1]> var_3484_axes_0 = const()[name = string("op_3484_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3484_keep_dims_0 = const()[name = string("op_3484_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_3484_cast_fp16 = reduce_mean(axes = var_3484_axes_0, keep_dims = var_3484_keep_dims_0, x = var_3482_cast_fp16)[name = string("op_3484_cast_fp16")];
            fp16 var_3485_to_fp16 = const()[name = string("op_3485_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3486_cast_fp16 = add(x = var_3484_cast_fp16, y = var_3485_to_fp16)[name = string("op_3486_cast_fp16")];
            fp32 norm_153_epsilon_0 = const()[name = string("norm_153_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_153_cast_fp16 = rsqrt(epsilon = norm_153_epsilon_0, x = var_3486_cast_fp16)[name = string("norm_153_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_3488_cast_fp16 = mul(x = x_493_cast_fp16, y = norm_153_cast_fp16)[name = string("op_3488_cast_fp16")];
            tensor<fp16, [1024]> layers_19_input_layernorm_weight_to_fp16 = const()[name = string("layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317644096)))];
            tensor<fp16, [1, ?, 1024]> var_3489_cast_fp16 = mul(x = var_3488_cast_fp16, y = layers_19_input_layernorm_weight_to_fp16)[name = string("op_3489_cast_fp16")];
            tensor<int32, [3]> var_3501_shape_cast_fp16 = shape(x = var_3489_cast_fp16)[name = string("op_3501_shape_cast_fp16")];
            int32 gather_114 = const()[name = string("gather_114"), val = int32(1)];
            int32 gather_115_axis_0 = const()[name = string("gather_115_axis_0"), val = int32(0)];
            int32 gather_115_batch_dims_0 = const()[name = string("gather_115_batch_dims_0"), val = int32(0)];
            bool gather_115_validate_indices_0 = const()[name = string("gather_115_validate_indices_0"), val = bool(false)];
            string var_3501_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3501_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_115_to_uint16 = const()[name = string("select_115_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_3501_shape_cast_fp16_to_uint16 = cast(dtype = var_3501_shape_cast_fp16_to_uint16_dtype_0, x = var_3501_shape_cast_fp16)[name = string("cast_17")];
            uint16 gather_115_cast_uint16 = gather(axis = gather_115_axis_0, batch_dims = gather_115_batch_dims_0, indices = select_115_to_uint16, validate_indices = gather_115_validate_indices_0, x = var_3501_shape_cast_fp16_to_uint16)[name = string("gather_115_cast_uint16")];
            string gather_115_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_115_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317646208))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319743424))))[name = string("layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_133_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = var_3489_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<int32, [4]> concat_114x = const()[name = string("concat_114x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_3506_cast_fp16 = reshape(shape = concat_114x, x = linear_133_cast_fp16)[name = string("op_3506_cast_fp16")];
            tensor<int32, [4]> x_499_perm_0 = const()[name = string("x_499_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319874560))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320923200))))[name = string("layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_134_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = var_3489_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<int32, [4]> concat_115x = const()[name = string("concat_115x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_3511_cast_fp16 = reshape(shape = concat_115x, x = linear_134_cast_fp16)[name = string("op_3511_cast_fp16")];
            tensor<int32, [4]> x_503_perm_0 = const()[name = string("x_503_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320988800))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322037440))))[name = string("layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_135_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = var_3489_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<int32, [4]> concat_116x = const()[name = string("concat_116x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_3516_cast_fp16 = reshape(shape = concat_116x, x = linear_135_cast_fp16)[name = string("op_3516_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = string("v_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3472_promoted_1_to_fp16 = const()[name = string("op_3472_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_499_cast_fp16 = transpose(perm = x_499_perm_0, x = var_3506_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 16, ?, 128]> var_3520_cast_fp16 = pow(x = x_499_cast_fp16, y = var_3472_promoted_1_to_fp16)[name = string("op_3520_cast_fp16")];
            tensor<int32, [1]> var_3522_axes_0 = const()[name = string("op_3522_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3522_keep_dims_0 = const()[name = string("op_3522_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_3522_cast_fp16 = reduce_mean(axes = var_3522_axes_0, keep_dims = var_3522_keep_dims_0, x = var_3520_cast_fp16)[name = string("op_3522_cast_fp16")];
            fp16 var_3523_to_fp16 = const()[name = string("op_3523_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_3524_cast_fp16 = add(x = var_3522_cast_fp16, y = var_3523_to_fp16)[name = string("op_3524_cast_fp16")];
            fp32 norm_155_epsilon_0 = const()[name = string("norm_155_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_155_cast_fp16 = rsqrt(epsilon = norm_155_epsilon_0, x = var_3524_cast_fp16)[name = string("norm_155_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_3526_cast_fp16 = mul(x = x_499_cast_fp16, y = norm_155_cast_fp16)[name = string("op_3526_cast_fp16")];
            tensor<fp16, [128]> layers_19_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_19_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322103040)))];
            tensor<fp16, [1, 16, ?, 128]> var_3527_cast_fp16 = mul(x = var_3526_cast_fp16, y = layers_19_self_attn_q_norm_weight_to_fp16)[name = string("op_3527_cast_fp16")];
            fp16 var_3472_promoted_2_to_fp16 = const()[name = string("op_3472_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_503_cast_fp16 = transpose(perm = x_503_perm_0, x = var_3511_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 8, ?, 128]> var_3531_cast_fp16 = pow(x = x_503_cast_fp16, y = var_3472_promoted_2_to_fp16)[name = string("op_3531_cast_fp16")];
            tensor<int32, [1]> var_3533_axes_0 = const()[name = string("op_3533_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3533_keep_dims_0 = const()[name = string("op_3533_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_3533_cast_fp16 = reduce_mean(axes = var_3533_axes_0, keep_dims = var_3533_keep_dims_0, x = var_3531_cast_fp16)[name = string("op_3533_cast_fp16")];
            fp16 var_3534_to_fp16 = const()[name = string("op_3534_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_3535_cast_fp16 = add(x = var_3533_cast_fp16, y = var_3534_to_fp16)[name = string("op_3535_cast_fp16")];
            fp32 norm_157_epsilon_0 = const()[name = string("norm_157_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_157_cast_fp16 = rsqrt(epsilon = norm_157_epsilon_0, x = var_3535_cast_fp16)[name = string("norm_157_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_3537_cast_fp16 = mul(x = x_503_cast_fp16, y = norm_157_cast_fp16)[name = string("op_3537_cast_fp16")];
            tensor<fp16, [128]> layers_19_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_19_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322103360)))];
            tensor<fp16, [1, 8, ?, 128]> var_3538_cast_fp16 = mul(x = var_3537_cast_fp16, y = layers_19_self_attn_k_norm_weight_to_fp16)[name = string("op_3538_cast_fp16")];
            tensor<int32, [4]> x1_77_begin_0 = const()[name = string("x1_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_77_end_0 = const()[name = string("x1_77_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_77_end_mask_0 = const()[name = string("x1_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_77_cast_fp16 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = var_3527_cast_fp16)[name = string("x1_77_cast_fp16")];
            tensor<int32, [4]> x2_77_begin_0 = const()[name = string("x2_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_77_end_0 = const()[name = string("x2_77_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_77_end_mask_0 = const()[name = string("x2_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_77_cast_fp16 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = var_3527_cast_fp16)[name = string("x2_77_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3554_cast_fp16 = mul(x = x1_77_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3554_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3555_cast_fp16 = mul(x = x2_77_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3555_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3556_cast_fp16 = sub(x = var_3554_cast_fp16, y = var_3555_cast_fp16)[name = string("op_3556_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3557_cast_fp16 = mul(x = x2_77_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3557_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3558_cast_fp16 = mul(x = x1_77_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3558_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3559_cast_fp16 = add(x = var_3557_cast_fp16, y = var_3558_cast_fp16)[name = string("op_3559_cast_fp16")];
            bool q_39_interleave_0 = const()[name = string("q_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_39_cast_fp16 = concat(axis = var_3473, interleave = q_39_interleave_0, values = (var_3556_cast_fp16, var_3559_cast_fp16))[name = string("q_39_cast_fp16")];
            tensor<int32, [4]> x1_79_begin_0 = const()[name = string("x1_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_79_end_0 = const()[name = string("x1_79_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_79_end_mask_0 = const()[name = string("x1_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_79_cast_fp16 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = var_3538_cast_fp16)[name = string("x1_79_cast_fp16")];
            tensor<int32, [4]> x2_79_begin_0 = const()[name = string("x2_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_79_end_0 = const()[name = string("x2_79_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_79_end_mask_0 = const()[name = string("x2_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_79_cast_fp16 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = var_3538_cast_fp16)[name = string("x2_79_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3576_cast_fp16 = mul(x = x1_79_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3576_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3577_cast_fp16 = mul(x = x2_79_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3577_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3578_cast_fp16 = sub(x = var_3576_cast_fp16, y = var_3577_cast_fp16)[name = string("op_3578_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3579_cast_fp16 = mul(x = x2_79_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3579_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3580_cast_fp16 = mul(x = x1_79_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3580_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3581_cast_fp16 = add(x = var_3579_cast_fp16, y = var_3580_cast_fp16)[name = string("op_3581_cast_fp16")];
            bool k_39_interleave_0 = const()[name = string("k_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_39_cast_fp16 = concat(axis = var_3473, interleave = k_39_interleave_0, values = (var_3578_cast_fp16, var_3581_cast_fp16))[name = string("k_39_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_38 = read_state(input = k_cache_19)[name = string("read_state_38")];
            int32 k_cache_79_axis_0 = const()[name = string("k_cache_79_axis_0"), val = int32(2)];
            string k_cache_79_mode_0 = const()[name = string("k_cache_79_mode_0"), val = string("update")];
            bool k_cache_79_validate_indices_0 = const()[name = string("k_cache_79_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_79_cast_fp16 = scatter_along_axis(axis = k_cache_79_axis_0, data = read_state_38, indices = scatter_indices, mode = k_cache_79_mode_0, updates = k_39_cast_fp16, validate_indices = k_cache_79_validate_indices_0)[name = string("k_cache_79_cast_fp16")];
            write_state(data = k_cache_79_cast_fp16, input = k_cache_19)[name = string("coreml_update_state_38_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_94 = read_state(input = k_cache_19)[name = string("coreml_update_state_38")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_39 = read_state(input = v_cache_19)[name = string("read_state_39")];
            int32 v_cache_79_axis_0 = const()[name = string("v_cache_79_axis_0"), val = int32(2)];
            string v_cache_79_mode_0 = const()[name = string("v_cache_79_mode_0"), val = string("update")];
            bool v_cache_79_validate_indices_0 = const()[name = string("v_cache_79_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = var_3516_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_79_cast_fp16 = scatter_along_axis(axis = v_cache_79_axis_0, data = read_state_39, indices = scatter_indices, mode = v_cache_79_mode_0, updates = v_39_cast_fp16, validate_indices = v_cache_79_validate_indices_0)[name = string("v_cache_79_cast_fp16")];
            write_state(data = v_cache_79_cast_fp16, input = v_cache_19)[name = string("coreml_update_state_39_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_95 = read_state(input = v_cache_19)[name = string("coreml_update_state_39")];
            tensor<int32, [1]> var_3586_axes_0 = const()[name = string("op_3586_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3586_cast_fp16 = expand_dims(axes = var_3586_axes_0, x = coreml_update_state_94)[name = string("op_3586_cast_fp16")];
            tensor<int32, [5]> k_exp_77_reps_0 = const()[name = string("k_exp_77_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_77_cast_fp16 = tile(reps = k_exp_77_reps_0, x = var_3586_cast_fp16)[name = string("k_exp_77_cast_fp16")];
            tensor<int32, [4]> var_3589 = const()[name = string("op_3589"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_79_cast_fp16 = reshape(shape = var_3589, x = k_exp_77_cast_fp16)[name = string("k_exp_79_cast_fp16")];
            tensor<int32, [1]> var_3591_axes_0 = const()[name = string("op_3591_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3591_cast_fp16 = expand_dims(axes = var_3591_axes_0, x = coreml_update_state_95)[name = string("op_3591_cast_fp16")];
            tensor<int32, [5]> v_exp_77_reps_0 = const()[name = string("v_exp_77_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_77_cast_fp16 = tile(reps = v_exp_77_reps_0, x = var_3591_cast_fp16)[name = string("v_exp_77_cast_fp16")];
            tensor<int32, [4]> var_3594 = const()[name = string("op_3594"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_79_cast_fp16 = reshape(shape = var_3594, x = v_exp_77_cast_fp16)[name = string("v_exp_79_cast_fp16")];
            bool var_3597_transpose_x_1 = const()[name = string("op_3597_transpose_x_1"), val = bool(false)];
            bool var_3597_transpose_y_1 = const()[name = string("op_3597_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_3597_cast_fp16 = matmul(transpose_x = var_3597_transpose_x_1, transpose_y = var_3597_transpose_y_1, x = q_39_cast_fp16, y = k_exp_79_cast_fp16)[name = string("op_3597_cast_fp16")];
            fp16 var_3598_to_fp16 = const()[name = string("op_3598_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_77_cast_fp16 = mul(x = var_3597_cast_fp16, y = var_3598_to_fp16)[name = string("attn_77_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_191_cast_fp16 = add(x = attn_77_cast_fp16, y = attention_mask)[name = string("input_191_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_79_cast_fp16 = softmax(axis = var_3473, x = input_191_cast_fp16)[name = string("attn_79_cast_fp16")];
            bool out_39_transpose_x_0 = const()[name = string("out_39_transpose_x_0"), val = bool(false)];
            bool out_39_transpose_y_0 = const()[name = string("out_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_39_cast_fp16 = matmul(transpose_x = out_39_transpose_x_0, transpose_y = out_39_transpose_y_0, x = attn_79_cast_fp16, y = v_exp_79_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<int32, [4]> var_3603_perm_0 = const()[name = string("op_3603_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_119_axis_0 = const()[name = string("concat_119_axis_0"), val = int32(0)];
            bool concat_119_interleave_0 = const()[name = string("concat_119_interleave_0"), val = bool(false)];
            int32 gather_115_cast_uint16_to_int32 = cast(dtype = gather_115_cast_uint16_to_int32_dtype_0, x = gather_115_cast_uint16)[name = string("cast_16")];
            tensor<int32, [3]> concat_119 = concat(axis = concat_119_axis_0, interleave = concat_119_interleave_0, values = (gather_114, gather_115_cast_uint16_to_int32, var_3473))[name = string("concat_119")];
            tensor<fp16, [1, ?, 16, 128]> var_3603_cast_fp16 = transpose(perm = var_3603_perm_0, x = out_39_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, ?, ?]> input_193_cast_fp16 = reshape(shape = concat_119, x = var_3603_cast_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322103680))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324200896))))[name = string("layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_513_cast_fp16 = add(x = x_493_cast_fp16, y = linear_136_cast_fp16)[name = string("x_513_cast_fp16")];
            fp16 var_3472_promoted_3_to_fp16 = const()[name = string("op_3472_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_3611_cast_fp16 = pow(x = x_513_cast_fp16, y = var_3472_promoted_3_to_fp16)[name = string("op_3611_cast_fp16")];
            tensor<int32, [1]> var_3613_axes_0 = const()[name = string("op_3613_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3613_keep_dims_0 = const()[name = string("op_3613_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_3613_cast_fp16 = reduce_mean(axes = var_3613_axes_0, keep_dims = var_3613_keep_dims_0, x = var_3611_cast_fp16)[name = string("op_3613_cast_fp16")];
            fp16 var_3614_to_fp16 = const()[name = string("op_3614_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3615_cast_fp16 = add(x = var_3613_cast_fp16, y = var_3614_to_fp16)[name = string("op_3615_cast_fp16")];
            fp32 norm_159_epsilon_0 = const()[name = string("norm_159_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_159_cast_fp16 = rsqrt(epsilon = norm_159_epsilon_0, x = var_3615_cast_fp16)[name = string("norm_159_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_3617_cast_fp16 = mul(x = x_513_cast_fp16, y = norm_159_cast_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [1024]> layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324332032)))];
            tensor<fp16, [1, ?, 1024]> var_3618_cast_fp16 = mul(x = var_3617_cast_fp16, y = layers_19_post_attention_layernorm_weight_to_fp16)[name = string("op_3618_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324334144))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327479936))))[name = string("layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = var_3618_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_3628_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_3628_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327676608))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330822400))))[name = string("layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_19_mlp_up_proj_weight_to_fp16_quantized, x = var_3618_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_199_cast_fp16 = mul(x = var_3628_cast_fp16, y = linear_138_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331019072))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334164864))))[name = string("layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_139_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_519_cast_fp16 = add(x = x_513_cast_fp16, y = linear_139_cast_fp16)[name = string("x_519_cast_fp16")];
            int32 var_3648 = const()[name = string("op_3648"), val = int32(-1)];
            fp16 var_3647_promoted_to_fp16 = const()[name = string("op_3647_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_3657_cast_fp16 = pow(x = x_519_cast_fp16, y = var_3647_promoted_to_fp16)[name = string("op_3657_cast_fp16")];
            tensor<int32, [1]> var_3659_axes_0 = const()[name = string("op_3659_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3659_keep_dims_0 = const()[name = string("op_3659_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_3659_cast_fp16 = reduce_mean(axes = var_3659_axes_0, keep_dims = var_3659_keep_dims_0, x = var_3657_cast_fp16)[name = string("op_3659_cast_fp16")];
            fp16 var_3660_to_fp16 = const()[name = string("op_3660_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3661_cast_fp16 = add(x = var_3659_cast_fp16, y = var_3660_to_fp16)[name = string("op_3661_cast_fp16")];
            fp32 norm_161_epsilon_0 = const()[name = string("norm_161_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_161_cast_fp16 = rsqrt(epsilon = norm_161_epsilon_0, x = var_3661_cast_fp16)[name = string("norm_161_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_3663_cast_fp16 = mul(x = x_519_cast_fp16, y = norm_161_cast_fp16)[name = string("op_3663_cast_fp16")];
            tensor<fp16, [1024]> layers_20_input_layernorm_weight_to_fp16 = const()[name = string("layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334361536)))];
            tensor<fp16, [1, ?, 1024]> var_3664_cast_fp16 = mul(x = var_3663_cast_fp16, y = layers_20_input_layernorm_weight_to_fp16)[name = string("op_3664_cast_fp16")];
            tensor<int32, [3]> var_3676_shape_cast_fp16 = shape(x = var_3664_cast_fp16)[name = string("op_3676_shape_cast_fp16")];
            int32 gather_120 = const()[name = string("gather_120"), val = int32(1)];
            int32 gather_121_axis_0 = const()[name = string("gather_121_axis_0"), val = int32(0)];
            int32 gather_121_batch_dims_0 = const()[name = string("gather_121_batch_dims_0"), val = int32(0)];
            bool gather_121_validate_indices_0 = const()[name = string("gather_121_validate_indices_0"), val = bool(false)];
            string var_3676_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3676_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_121_to_uint16 = const()[name = string("select_121_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_3676_shape_cast_fp16_to_uint16 = cast(dtype = var_3676_shape_cast_fp16_to_uint16_dtype_0, x = var_3676_shape_cast_fp16)[name = string("cast_15")];
            uint16 gather_121_cast_uint16 = gather(axis = gather_121_axis_0, batch_dims = gather_121_batch_dims_0, indices = select_121_to_uint16, validate_indices = gather_121_validate_indices_0, x = var_3676_shape_cast_fp16_to_uint16)[name = string("gather_121_cast_uint16")];
            string gather_121_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_121_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334363648))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336460864))))[name = string("layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_140_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = var_3664_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> concat_120x = const()[name = string("concat_120x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_3681_cast_fp16 = reshape(shape = concat_120x, x = linear_140_cast_fp16)[name = string("op_3681_cast_fp16")];
            tensor<int32, [4]> x_525_perm_0 = const()[name = string("x_525_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336592000))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337640640))))[name = string("layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_141_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = var_3664_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<int32, [4]> concat_121x = const()[name = string("concat_121x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_3686_cast_fp16 = reshape(shape = concat_121x, x = linear_141_cast_fp16)[name = string("op_3686_cast_fp16")];
            tensor<int32, [4]> x_529_perm_0 = const()[name = string("x_529_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337706240))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338754880))))[name = string("layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_142_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = var_3664_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<int32, [4]> concat_122x = const()[name = string("concat_122x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_3691_cast_fp16 = reshape(shape = concat_122x, x = linear_142_cast_fp16)[name = string("op_3691_cast_fp16")];
            tensor<int32, [4]> v_41_perm_0 = const()[name = string("v_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3647_promoted_1_to_fp16 = const()[name = string("op_3647_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_525_cast_fp16 = transpose(perm = x_525_perm_0, x = var_3681_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 16, ?, 128]> var_3695_cast_fp16 = pow(x = x_525_cast_fp16, y = var_3647_promoted_1_to_fp16)[name = string("op_3695_cast_fp16")];
            tensor<int32, [1]> var_3697_axes_0 = const()[name = string("op_3697_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3697_keep_dims_0 = const()[name = string("op_3697_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_3697_cast_fp16 = reduce_mean(axes = var_3697_axes_0, keep_dims = var_3697_keep_dims_0, x = var_3695_cast_fp16)[name = string("op_3697_cast_fp16")];
            fp16 var_3698_to_fp16 = const()[name = string("op_3698_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_3699_cast_fp16 = add(x = var_3697_cast_fp16, y = var_3698_to_fp16)[name = string("op_3699_cast_fp16")];
            fp32 norm_163_epsilon_0 = const()[name = string("norm_163_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_163_cast_fp16 = rsqrt(epsilon = norm_163_epsilon_0, x = var_3699_cast_fp16)[name = string("norm_163_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_3701_cast_fp16 = mul(x = x_525_cast_fp16, y = norm_163_cast_fp16)[name = string("op_3701_cast_fp16")];
            tensor<fp16, [128]> layers_20_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_20_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338820480)))];
            tensor<fp16, [1, 16, ?, 128]> var_3702_cast_fp16 = mul(x = var_3701_cast_fp16, y = layers_20_self_attn_q_norm_weight_to_fp16)[name = string("op_3702_cast_fp16")];
            fp16 var_3647_promoted_2_to_fp16 = const()[name = string("op_3647_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_529_cast_fp16 = transpose(perm = x_529_perm_0, x = var_3686_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 8, ?, 128]> var_3706_cast_fp16 = pow(x = x_529_cast_fp16, y = var_3647_promoted_2_to_fp16)[name = string("op_3706_cast_fp16")];
            tensor<int32, [1]> var_3708_axes_0 = const()[name = string("op_3708_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3708_keep_dims_0 = const()[name = string("op_3708_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_3708_cast_fp16 = reduce_mean(axes = var_3708_axes_0, keep_dims = var_3708_keep_dims_0, x = var_3706_cast_fp16)[name = string("op_3708_cast_fp16")];
            fp16 var_3709_to_fp16 = const()[name = string("op_3709_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_3710_cast_fp16 = add(x = var_3708_cast_fp16, y = var_3709_to_fp16)[name = string("op_3710_cast_fp16")];
            fp32 norm_165_epsilon_0 = const()[name = string("norm_165_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_165_cast_fp16 = rsqrt(epsilon = norm_165_epsilon_0, x = var_3710_cast_fp16)[name = string("norm_165_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_3712_cast_fp16 = mul(x = x_529_cast_fp16, y = norm_165_cast_fp16)[name = string("op_3712_cast_fp16")];
            tensor<fp16, [128]> layers_20_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_20_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338820800)))];
            tensor<fp16, [1, 8, ?, 128]> var_3713_cast_fp16 = mul(x = var_3712_cast_fp16, y = layers_20_self_attn_k_norm_weight_to_fp16)[name = string("op_3713_cast_fp16")];
            tensor<int32, [4]> x1_81_begin_0 = const()[name = string("x1_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_81_end_0 = const()[name = string("x1_81_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_81_end_mask_0 = const()[name = string("x1_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_81_cast_fp16 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = var_3702_cast_fp16)[name = string("x1_81_cast_fp16")];
            tensor<int32, [4]> x2_81_begin_0 = const()[name = string("x2_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_81_end_0 = const()[name = string("x2_81_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_81_end_mask_0 = const()[name = string("x2_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_81_cast_fp16 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = var_3702_cast_fp16)[name = string("x2_81_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3729_cast_fp16 = mul(x = x1_81_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3729_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3730_cast_fp16 = mul(x = x2_81_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3730_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3731_cast_fp16 = sub(x = var_3729_cast_fp16, y = var_3730_cast_fp16)[name = string("op_3731_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3732_cast_fp16 = mul(x = x2_81_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3732_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3733_cast_fp16 = mul(x = x1_81_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3733_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3734_cast_fp16 = add(x = var_3732_cast_fp16, y = var_3733_cast_fp16)[name = string("op_3734_cast_fp16")];
            bool q_41_interleave_0 = const()[name = string("q_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_41_cast_fp16 = concat(axis = var_3648, interleave = q_41_interleave_0, values = (var_3731_cast_fp16, var_3734_cast_fp16))[name = string("q_41_cast_fp16")];
            tensor<int32, [4]> x1_83_begin_0 = const()[name = string("x1_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_83_end_0 = const()[name = string("x1_83_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_83_end_mask_0 = const()[name = string("x1_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_83_cast_fp16 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = var_3713_cast_fp16)[name = string("x1_83_cast_fp16")];
            tensor<int32, [4]> x2_83_begin_0 = const()[name = string("x2_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_83_end_0 = const()[name = string("x2_83_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_83_end_mask_0 = const()[name = string("x2_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_83_cast_fp16 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = var_3713_cast_fp16)[name = string("x2_83_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3751_cast_fp16 = mul(x = x1_83_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3751_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3752_cast_fp16 = mul(x = x2_83_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3752_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3753_cast_fp16 = sub(x = var_3751_cast_fp16, y = var_3752_cast_fp16)[name = string("op_3753_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3754_cast_fp16 = mul(x = x2_83_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3754_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3755_cast_fp16 = mul(x = x1_83_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3755_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3756_cast_fp16 = add(x = var_3754_cast_fp16, y = var_3755_cast_fp16)[name = string("op_3756_cast_fp16")];
            bool k_41_interleave_0 = const()[name = string("k_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_41_cast_fp16 = concat(axis = var_3648, interleave = k_41_interleave_0, values = (var_3753_cast_fp16, var_3756_cast_fp16))[name = string("k_41_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_40 = read_state(input = k_cache_20)[name = string("read_state_40")];
            int32 k_cache_83_axis_0 = const()[name = string("k_cache_83_axis_0"), val = int32(2)];
            string k_cache_83_mode_0 = const()[name = string("k_cache_83_mode_0"), val = string("update")];
            bool k_cache_83_validate_indices_0 = const()[name = string("k_cache_83_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_83_cast_fp16 = scatter_along_axis(axis = k_cache_83_axis_0, data = read_state_40, indices = scatter_indices, mode = k_cache_83_mode_0, updates = k_41_cast_fp16, validate_indices = k_cache_83_validate_indices_0)[name = string("k_cache_83_cast_fp16")];
            write_state(data = k_cache_83_cast_fp16, input = k_cache_20)[name = string("coreml_update_state_40_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_96 = read_state(input = k_cache_20)[name = string("coreml_update_state_40")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_41 = read_state(input = v_cache_20)[name = string("read_state_41")];
            int32 v_cache_83_axis_0 = const()[name = string("v_cache_83_axis_0"), val = int32(2)];
            string v_cache_83_mode_0 = const()[name = string("v_cache_83_mode_0"), val = string("update")];
            bool v_cache_83_validate_indices_0 = const()[name = string("v_cache_83_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_41_cast_fp16 = transpose(perm = v_41_perm_0, x = var_3691_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_83_cast_fp16 = scatter_along_axis(axis = v_cache_83_axis_0, data = read_state_41, indices = scatter_indices, mode = v_cache_83_mode_0, updates = v_41_cast_fp16, validate_indices = v_cache_83_validate_indices_0)[name = string("v_cache_83_cast_fp16")];
            write_state(data = v_cache_83_cast_fp16, input = v_cache_20)[name = string("coreml_update_state_41_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_97 = read_state(input = v_cache_20)[name = string("coreml_update_state_41")];
            tensor<int32, [1]> var_3761_axes_0 = const()[name = string("op_3761_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3761_cast_fp16 = expand_dims(axes = var_3761_axes_0, x = coreml_update_state_96)[name = string("op_3761_cast_fp16")];
            tensor<int32, [5]> k_exp_81_reps_0 = const()[name = string("k_exp_81_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_81_cast_fp16 = tile(reps = k_exp_81_reps_0, x = var_3761_cast_fp16)[name = string("k_exp_81_cast_fp16")];
            tensor<int32, [4]> var_3764 = const()[name = string("op_3764"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_83_cast_fp16 = reshape(shape = var_3764, x = k_exp_81_cast_fp16)[name = string("k_exp_83_cast_fp16")];
            tensor<int32, [1]> var_3766_axes_0 = const()[name = string("op_3766_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3766_cast_fp16 = expand_dims(axes = var_3766_axes_0, x = coreml_update_state_97)[name = string("op_3766_cast_fp16")];
            tensor<int32, [5]> v_exp_81_reps_0 = const()[name = string("v_exp_81_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_81_cast_fp16 = tile(reps = v_exp_81_reps_0, x = var_3766_cast_fp16)[name = string("v_exp_81_cast_fp16")];
            tensor<int32, [4]> var_3769 = const()[name = string("op_3769"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_83_cast_fp16 = reshape(shape = var_3769, x = v_exp_81_cast_fp16)[name = string("v_exp_83_cast_fp16")];
            bool var_3772_transpose_x_1 = const()[name = string("op_3772_transpose_x_1"), val = bool(false)];
            bool var_3772_transpose_y_1 = const()[name = string("op_3772_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_3772_cast_fp16 = matmul(transpose_x = var_3772_transpose_x_1, transpose_y = var_3772_transpose_y_1, x = q_41_cast_fp16, y = k_exp_83_cast_fp16)[name = string("op_3772_cast_fp16")];
            fp16 var_3773_to_fp16 = const()[name = string("op_3773_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_81_cast_fp16 = mul(x = var_3772_cast_fp16, y = var_3773_to_fp16)[name = string("attn_81_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_201_cast_fp16 = add(x = attn_81_cast_fp16, y = attention_mask)[name = string("input_201_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_83_cast_fp16 = softmax(axis = var_3648, x = input_201_cast_fp16)[name = string("attn_83_cast_fp16")];
            bool out_41_transpose_x_0 = const()[name = string("out_41_transpose_x_0"), val = bool(false)];
            bool out_41_transpose_y_0 = const()[name = string("out_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_41_cast_fp16 = matmul(transpose_x = out_41_transpose_x_0, transpose_y = out_41_transpose_y_0, x = attn_83_cast_fp16, y = v_exp_83_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<int32, [4]> var_3778_perm_0 = const()[name = string("op_3778_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_125_axis_0 = const()[name = string("concat_125_axis_0"), val = int32(0)];
            bool concat_125_interleave_0 = const()[name = string("concat_125_interleave_0"), val = bool(false)];
            int32 gather_121_cast_uint16_to_int32 = cast(dtype = gather_121_cast_uint16_to_int32_dtype_0, x = gather_121_cast_uint16)[name = string("cast_14")];
            tensor<int32, [3]> concat_125 = concat(axis = concat_125_axis_0, interleave = concat_125_interleave_0, values = (gather_120, gather_121_cast_uint16_to_int32, var_3648))[name = string("concat_125")];
            tensor<fp16, [1, ?, 16, 128]> var_3778_cast_fp16 = transpose(perm = var_3778_perm_0, x = out_41_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, ?, ?]> input_203_cast_fp16 = reshape(shape = concat_125, x = var_3778_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338821120))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340918336))))[name = string("layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_143_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_539_cast_fp16 = add(x = x_519_cast_fp16, y = linear_143_cast_fp16)[name = string("x_539_cast_fp16")];
            fp16 var_3647_promoted_3_to_fp16 = const()[name = string("op_3647_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_3786_cast_fp16 = pow(x = x_539_cast_fp16, y = var_3647_promoted_3_to_fp16)[name = string("op_3786_cast_fp16")];
            tensor<int32, [1]> var_3788_axes_0 = const()[name = string("op_3788_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3788_keep_dims_0 = const()[name = string("op_3788_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_3788_cast_fp16 = reduce_mean(axes = var_3788_axes_0, keep_dims = var_3788_keep_dims_0, x = var_3786_cast_fp16)[name = string("op_3788_cast_fp16")];
            fp16 var_3789_to_fp16 = const()[name = string("op_3789_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3790_cast_fp16 = add(x = var_3788_cast_fp16, y = var_3789_to_fp16)[name = string("op_3790_cast_fp16")];
            fp32 norm_167_epsilon_0 = const()[name = string("norm_167_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_167_cast_fp16 = rsqrt(epsilon = norm_167_epsilon_0, x = var_3790_cast_fp16)[name = string("norm_167_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_3792_cast_fp16 = mul(x = x_539_cast_fp16, y = norm_167_cast_fp16)[name = string("op_3792_cast_fp16")];
            tensor<fp16, [1024]> layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341049472)))];
            tensor<fp16, [1, ?, 1024]> var_3793_cast_fp16 = mul(x = var_3792_cast_fp16, y = layers_20_post_attention_layernorm_weight_to_fp16)[name = string("op_3793_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341051584))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344197376))))[name = string("layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_144_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = var_3793_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_3803_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_3803_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344394048))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347539840))))[name = string("layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_20_mlp_up_proj_weight_to_fp16_quantized, x = var_3793_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_209_cast_fp16 = mul(x = var_3803_cast_fp16, y = linear_145_cast_fp16)[name = string("input_209_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347736512))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350882304))))[name = string("layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_146_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_545_cast_fp16 = add(x = x_539_cast_fp16, y = linear_146_cast_fp16)[name = string("x_545_cast_fp16")];
            int32 var_3823 = const()[name = string("op_3823"), val = int32(-1)];
            fp16 var_3822_promoted_to_fp16 = const()[name = string("op_3822_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_3832_cast_fp16 = pow(x = x_545_cast_fp16, y = var_3822_promoted_to_fp16)[name = string("op_3832_cast_fp16")];
            tensor<int32, [1]> var_3834_axes_0 = const()[name = string("op_3834_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3834_keep_dims_0 = const()[name = string("op_3834_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_3834_cast_fp16 = reduce_mean(axes = var_3834_axes_0, keep_dims = var_3834_keep_dims_0, x = var_3832_cast_fp16)[name = string("op_3834_cast_fp16")];
            fp16 var_3835_to_fp16 = const()[name = string("op_3835_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3836_cast_fp16 = add(x = var_3834_cast_fp16, y = var_3835_to_fp16)[name = string("op_3836_cast_fp16")];
            fp32 norm_169_epsilon_0 = const()[name = string("norm_169_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_169_cast_fp16 = rsqrt(epsilon = norm_169_epsilon_0, x = var_3836_cast_fp16)[name = string("norm_169_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_3838_cast_fp16 = mul(x = x_545_cast_fp16, y = norm_169_cast_fp16)[name = string("op_3838_cast_fp16")];
            tensor<fp16, [1024]> layers_21_input_layernorm_weight_to_fp16 = const()[name = string("layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351078976)))];
            tensor<fp16, [1, ?, 1024]> var_3839_cast_fp16 = mul(x = var_3838_cast_fp16, y = layers_21_input_layernorm_weight_to_fp16)[name = string("op_3839_cast_fp16")];
            tensor<int32, [3]> var_3851_shape_cast_fp16 = shape(x = var_3839_cast_fp16)[name = string("op_3851_shape_cast_fp16")];
            int32 gather_126 = const()[name = string("gather_126"), val = int32(1)];
            int32 gather_127_axis_0 = const()[name = string("gather_127_axis_0"), val = int32(0)];
            int32 gather_127_batch_dims_0 = const()[name = string("gather_127_batch_dims_0"), val = int32(0)];
            bool gather_127_validate_indices_0 = const()[name = string("gather_127_validate_indices_0"), val = bool(false)];
            string var_3851_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_3851_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_127_to_uint16 = const()[name = string("select_127_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_3851_shape_cast_fp16_to_uint16 = cast(dtype = var_3851_shape_cast_fp16_to_uint16_dtype_0, x = var_3851_shape_cast_fp16)[name = string("cast_13")];
            uint16 gather_127_cast_uint16 = gather(axis = gather_127_axis_0, batch_dims = gather_127_batch_dims_0, indices = select_127_to_uint16, validate_indices = gather_127_validate_indices_0, x = var_3851_shape_cast_fp16_to_uint16)[name = string("gather_127_cast_uint16")];
            string gather_127_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_127_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351081088))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353178304))))[name = string("layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_147_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = var_3839_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> concat_126x = const()[name = string("concat_126x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_3856_cast_fp16 = reshape(shape = concat_126x, x = linear_147_cast_fp16)[name = string("op_3856_cast_fp16")];
            tensor<int32, [4]> x_551_perm_0 = const()[name = string("x_551_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353309440))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354358080))))[name = string("layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_148_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = var_3839_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> concat_127x = const()[name = string("concat_127x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_3861_cast_fp16 = reshape(shape = concat_127x, x = linear_148_cast_fp16)[name = string("op_3861_cast_fp16")];
            tensor<int32, [4]> x_555_perm_0 = const()[name = string("x_555_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354423680))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355472320))))[name = string("layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_149_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = var_3839_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> concat_128x = const()[name = string("concat_128x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_3866_cast_fp16 = reshape(shape = concat_128x, x = linear_149_cast_fp16)[name = string("op_3866_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = string("v_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3822_promoted_1_to_fp16 = const()[name = string("op_3822_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_551_cast_fp16 = transpose(perm = x_551_perm_0, x = var_3856_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 16, ?, 128]> var_3870_cast_fp16 = pow(x = x_551_cast_fp16, y = var_3822_promoted_1_to_fp16)[name = string("op_3870_cast_fp16")];
            tensor<int32, [1]> var_3872_axes_0 = const()[name = string("op_3872_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3872_keep_dims_0 = const()[name = string("op_3872_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_3872_cast_fp16 = reduce_mean(axes = var_3872_axes_0, keep_dims = var_3872_keep_dims_0, x = var_3870_cast_fp16)[name = string("op_3872_cast_fp16")];
            fp16 var_3873_to_fp16 = const()[name = string("op_3873_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_3874_cast_fp16 = add(x = var_3872_cast_fp16, y = var_3873_to_fp16)[name = string("op_3874_cast_fp16")];
            fp32 norm_171_epsilon_0 = const()[name = string("norm_171_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_171_cast_fp16 = rsqrt(epsilon = norm_171_epsilon_0, x = var_3874_cast_fp16)[name = string("norm_171_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_3876_cast_fp16 = mul(x = x_551_cast_fp16, y = norm_171_cast_fp16)[name = string("op_3876_cast_fp16")];
            tensor<fp16, [128]> layers_21_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_21_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355537920)))];
            tensor<fp16, [1, 16, ?, 128]> var_3877_cast_fp16 = mul(x = var_3876_cast_fp16, y = layers_21_self_attn_q_norm_weight_to_fp16)[name = string("op_3877_cast_fp16")];
            fp16 var_3822_promoted_2_to_fp16 = const()[name = string("op_3822_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_555_cast_fp16 = transpose(perm = x_555_perm_0, x = var_3861_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 8, ?, 128]> var_3881_cast_fp16 = pow(x = x_555_cast_fp16, y = var_3822_promoted_2_to_fp16)[name = string("op_3881_cast_fp16")];
            tensor<int32, [1]> var_3883_axes_0 = const()[name = string("op_3883_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3883_keep_dims_0 = const()[name = string("op_3883_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_3883_cast_fp16 = reduce_mean(axes = var_3883_axes_0, keep_dims = var_3883_keep_dims_0, x = var_3881_cast_fp16)[name = string("op_3883_cast_fp16")];
            fp16 var_3884_to_fp16 = const()[name = string("op_3884_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_3885_cast_fp16 = add(x = var_3883_cast_fp16, y = var_3884_to_fp16)[name = string("op_3885_cast_fp16")];
            fp32 norm_173_epsilon_0 = const()[name = string("norm_173_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_173_cast_fp16 = rsqrt(epsilon = norm_173_epsilon_0, x = var_3885_cast_fp16)[name = string("norm_173_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_3887_cast_fp16 = mul(x = x_555_cast_fp16, y = norm_173_cast_fp16)[name = string("op_3887_cast_fp16")];
            tensor<fp16, [128]> layers_21_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_21_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355538240)))];
            tensor<fp16, [1, 8, ?, 128]> var_3888_cast_fp16 = mul(x = var_3887_cast_fp16, y = layers_21_self_attn_k_norm_weight_to_fp16)[name = string("op_3888_cast_fp16")];
            tensor<int32, [4]> x1_85_begin_0 = const()[name = string("x1_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_85_end_0 = const()[name = string("x1_85_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_85_end_mask_0 = const()[name = string("x1_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_85_cast_fp16 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = var_3877_cast_fp16)[name = string("x1_85_cast_fp16")];
            tensor<int32, [4]> x2_85_begin_0 = const()[name = string("x2_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_85_end_0 = const()[name = string("x2_85_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_85_end_mask_0 = const()[name = string("x2_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_85_cast_fp16 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = var_3877_cast_fp16)[name = string("x2_85_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3904_cast_fp16 = mul(x = x1_85_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3904_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3905_cast_fp16 = mul(x = x2_85_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3905_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3906_cast_fp16 = sub(x = var_3904_cast_fp16, y = var_3905_cast_fp16)[name = string("op_3906_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3907_cast_fp16 = mul(x = x2_85_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3907_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3908_cast_fp16 = mul(x = x1_85_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3908_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_3909_cast_fp16 = add(x = var_3907_cast_fp16, y = var_3908_cast_fp16)[name = string("op_3909_cast_fp16")];
            bool q_43_interleave_0 = const()[name = string("q_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_43_cast_fp16 = concat(axis = var_3823, interleave = q_43_interleave_0, values = (var_3906_cast_fp16, var_3909_cast_fp16))[name = string("q_43_cast_fp16")];
            tensor<int32, [4]> x1_87_begin_0 = const()[name = string("x1_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_87_end_0 = const()[name = string("x1_87_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_87_end_mask_0 = const()[name = string("x1_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_87_cast_fp16 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = var_3888_cast_fp16)[name = string("x1_87_cast_fp16")];
            tensor<int32, [4]> x2_87_begin_0 = const()[name = string("x2_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_87_end_0 = const()[name = string("x2_87_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_87_end_mask_0 = const()[name = string("x2_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_87_cast_fp16 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = var_3888_cast_fp16)[name = string("x2_87_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3926_cast_fp16 = mul(x = x1_87_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3926_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3927_cast_fp16 = mul(x = x2_87_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3927_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3928_cast_fp16 = sub(x = var_3926_cast_fp16, y = var_3927_cast_fp16)[name = string("op_3928_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3929_cast_fp16 = mul(x = x2_87_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_3929_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3930_cast_fp16 = mul(x = x1_87_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_3930_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_3931_cast_fp16 = add(x = var_3929_cast_fp16, y = var_3930_cast_fp16)[name = string("op_3931_cast_fp16")];
            bool k_43_interleave_0 = const()[name = string("k_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_43_cast_fp16 = concat(axis = var_3823, interleave = k_43_interleave_0, values = (var_3928_cast_fp16, var_3931_cast_fp16))[name = string("k_43_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_42 = read_state(input = k_cache_21)[name = string("read_state_42")];
            int32 k_cache_87_axis_0 = const()[name = string("k_cache_87_axis_0"), val = int32(2)];
            string k_cache_87_mode_0 = const()[name = string("k_cache_87_mode_0"), val = string("update")];
            bool k_cache_87_validate_indices_0 = const()[name = string("k_cache_87_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_87_cast_fp16 = scatter_along_axis(axis = k_cache_87_axis_0, data = read_state_42, indices = scatter_indices, mode = k_cache_87_mode_0, updates = k_43_cast_fp16, validate_indices = k_cache_87_validate_indices_0)[name = string("k_cache_87_cast_fp16")];
            write_state(data = k_cache_87_cast_fp16, input = k_cache_21)[name = string("coreml_update_state_42_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_98 = read_state(input = k_cache_21)[name = string("coreml_update_state_42")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_43 = read_state(input = v_cache_21)[name = string("read_state_43")];
            int32 v_cache_87_axis_0 = const()[name = string("v_cache_87_axis_0"), val = int32(2)];
            string v_cache_87_mode_0 = const()[name = string("v_cache_87_mode_0"), val = string("update")];
            bool v_cache_87_validate_indices_0 = const()[name = string("v_cache_87_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = var_3866_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_87_cast_fp16 = scatter_along_axis(axis = v_cache_87_axis_0, data = read_state_43, indices = scatter_indices, mode = v_cache_87_mode_0, updates = v_43_cast_fp16, validate_indices = v_cache_87_validate_indices_0)[name = string("v_cache_87_cast_fp16")];
            write_state(data = v_cache_87_cast_fp16, input = v_cache_21)[name = string("coreml_update_state_43_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_99 = read_state(input = v_cache_21)[name = string("coreml_update_state_43")];
            tensor<int32, [1]> var_3936_axes_0 = const()[name = string("op_3936_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3936_cast_fp16 = expand_dims(axes = var_3936_axes_0, x = coreml_update_state_98)[name = string("op_3936_cast_fp16")];
            tensor<int32, [5]> k_exp_85_reps_0 = const()[name = string("k_exp_85_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_85_cast_fp16 = tile(reps = k_exp_85_reps_0, x = var_3936_cast_fp16)[name = string("k_exp_85_cast_fp16")];
            tensor<int32, [4]> var_3939 = const()[name = string("op_3939"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_87_cast_fp16 = reshape(shape = var_3939, x = k_exp_85_cast_fp16)[name = string("k_exp_87_cast_fp16")];
            tensor<int32, [1]> var_3941_axes_0 = const()[name = string("op_3941_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_3941_cast_fp16 = expand_dims(axes = var_3941_axes_0, x = coreml_update_state_99)[name = string("op_3941_cast_fp16")];
            tensor<int32, [5]> v_exp_85_reps_0 = const()[name = string("v_exp_85_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_85_cast_fp16 = tile(reps = v_exp_85_reps_0, x = var_3941_cast_fp16)[name = string("v_exp_85_cast_fp16")];
            tensor<int32, [4]> var_3944 = const()[name = string("op_3944"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_87_cast_fp16 = reshape(shape = var_3944, x = v_exp_85_cast_fp16)[name = string("v_exp_87_cast_fp16")];
            bool var_3947_transpose_x_1 = const()[name = string("op_3947_transpose_x_1"), val = bool(false)];
            bool var_3947_transpose_y_1 = const()[name = string("op_3947_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_3947_cast_fp16 = matmul(transpose_x = var_3947_transpose_x_1, transpose_y = var_3947_transpose_y_1, x = q_43_cast_fp16, y = k_exp_87_cast_fp16)[name = string("op_3947_cast_fp16")];
            fp16 var_3948_to_fp16 = const()[name = string("op_3948_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_85_cast_fp16 = mul(x = var_3947_cast_fp16, y = var_3948_to_fp16)[name = string("attn_85_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_211_cast_fp16 = add(x = attn_85_cast_fp16, y = attention_mask)[name = string("input_211_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_87_cast_fp16 = softmax(axis = var_3823, x = input_211_cast_fp16)[name = string("attn_87_cast_fp16")];
            bool out_43_transpose_x_0 = const()[name = string("out_43_transpose_x_0"), val = bool(false)];
            bool out_43_transpose_y_0 = const()[name = string("out_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_43_cast_fp16 = matmul(transpose_x = out_43_transpose_x_0, transpose_y = out_43_transpose_y_0, x = attn_87_cast_fp16, y = v_exp_87_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<int32, [4]> var_3953_perm_0 = const()[name = string("op_3953_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_131_axis_0 = const()[name = string("concat_131_axis_0"), val = int32(0)];
            bool concat_131_interleave_0 = const()[name = string("concat_131_interleave_0"), val = bool(false)];
            int32 gather_127_cast_uint16_to_int32 = cast(dtype = gather_127_cast_uint16_to_int32_dtype_0, x = gather_127_cast_uint16)[name = string("cast_12")];
            tensor<int32, [3]> concat_131 = concat(axis = concat_131_axis_0, interleave = concat_131_interleave_0, values = (gather_126, gather_127_cast_uint16_to_int32, var_3823))[name = string("concat_131")];
            tensor<fp16, [1, ?, 16, 128]> var_3953_cast_fp16 = transpose(perm = var_3953_perm_0, x = out_43_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, ?, ?]> input_213_cast_fp16 = reshape(shape = concat_131, x = var_3953_cast_fp16)[name = string("input_213_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355538560))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357635776))))[name = string("layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_150_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = input_213_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_565_cast_fp16 = add(x = x_545_cast_fp16, y = linear_150_cast_fp16)[name = string("x_565_cast_fp16")];
            fp16 var_3822_promoted_3_to_fp16 = const()[name = string("op_3822_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_3961_cast_fp16 = pow(x = x_565_cast_fp16, y = var_3822_promoted_3_to_fp16)[name = string("op_3961_cast_fp16")];
            tensor<int32, [1]> var_3963_axes_0 = const()[name = string("op_3963_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3963_keep_dims_0 = const()[name = string("op_3963_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_3963_cast_fp16 = reduce_mean(axes = var_3963_axes_0, keep_dims = var_3963_keep_dims_0, x = var_3961_cast_fp16)[name = string("op_3963_cast_fp16")];
            fp16 var_3964_to_fp16 = const()[name = string("op_3964_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_3965_cast_fp16 = add(x = var_3963_cast_fp16, y = var_3964_to_fp16)[name = string("op_3965_cast_fp16")];
            fp32 norm_175_epsilon_0 = const()[name = string("norm_175_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_175_cast_fp16 = rsqrt(epsilon = norm_175_epsilon_0, x = var_3965_cast_fp16)[name = string("norm_175_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_3967_cast_fp16 = mul(x = x_565_cast_fp16, y = norm_175_cast_fp16)[name = string("op_3967_cast_fp16")];
            tensor<fp16, [1024]> layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357766912)))];
            tensor<fp16, [1, ?, 1024]> var_3968_cast_fp16 = mul(x = var_3967_cast_fp16, y = layers_21_post_attention_layernorm_weight_to_fp16)[name = string("op_3968_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357769024))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360914816))))[name = string("layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = var_3968_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_3978_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_3978_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361111488))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364257280))))[name = string("layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_21_mlp_up_proj_weight_to_fp16_quantized, x = var_3968_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_219_cast_fp16 = mul(x = var_3978_cast_fp16, y = linear_152_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364453952))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367599744))))[name = string("layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_153_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_219_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_571_cast_fp16 = add(x = x_565_cast_fp16, y = linear_153_cast_fp16)[name = string("x_571_cast_fp16")];
            int32 var_3998 = const()[name = string("op_3998"), val = int32(-1)];
            fp16 var_3997_promoted_to_fp16 = const()[name = string("op_3997_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_4007_cast_fp16 = pow(x = x_571_cast_fp16, y = var_3997_promoted_to_fp16)[name = string("op_4007_cast_fp16")];
            tensor<int32, [1]> var_4009_axes_0 = const()[name = string("op_4009_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4009_keep_dims_0 = const()[name = string("op_4009_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_4009_cast_fp16 = reduce_mean(axes = var_4009_axes_0, keep_dims = var_4009_keep_dims_0, x = var_4007_cast_fp16)[name = string("op_4009_cast_fp16")];
            fp16 var_4010_to_fp16 = const()[name = string("op_4010_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4011_cast_fp16 = add(x = var_4009_cast_fp16, y = var_4010_to_fp16)[name = string("op_4011_cast_fp16")];
            fp32 norm_177_epsilon_0 = const()[name = string("norm_177_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_177_cast_fp16 = rsqrt(epsilon = norm_177_epsilon_0, x = var_4011_cast_fp16)[name = string("norm_177_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_4013_cast_fp16 = mul(x = x_571_cast_fp16, y = norm_177_cast_fp16)[name = string("op_4013_cast_fp16")];
            tensor<fp16, [1024]> layers_22_input_layernorm_weight_to_fp16 = const()[name = string("layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367796416)))];
            tensor<fp16, [1, ?, 1024]> var_4014_cast_fp16 = mul(x = var_4013_cast_fp16, y = layers_22_input_layernorm_weight_to_fp16)[name = string("op_4014_cast_fp16")];
            tensor<int32, [3]> var_4026_shape_cast_fp16 = shape(x = var_4014_cast_fp16)[name = string("op_4026_shape_cast_fp16")];
            int32 gather_132 = const()[name = string("gather_132"), val = int32(1)];
            int32 gather_133_axis_0 = const()[name = string("gather_133_axis_0"), val = int32(0)];
            int32 gather_133_batch_dims_0 = const()[name = string("gather_133_batch_dims_0"), val = int32(0)];
            bool gather_133_validate_indices_0 = const()[name = string("gather_133_validate_indices_0"), val = bool(false)];
            string var_4026_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4026_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_133_to_uint16 = const()[name = string("select_133_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_4026_shape_cast_fp16_to_uint16 = cast(dtype = var_4026_shape_cast_fp16_to_uint16_dtype_0, x = var_4026_shape_cast_fp16)[name = string("cast_11")];
            uint16 gather_133_cast_uint16 = gather(axis = gather_133_axis_0, batch_dims = gather_133_batch_dims_0, indices = select_133_to_uint16, validate_indices = gather_133_validate_indices_0, x = var_4026_shape_cast_fp16_to_uint16)[name = string("gather_133_cast_uint16")];
            string gather_133_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_133_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367798528))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369895744))))[name = string("layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_154_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = var_4014_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_4031_cast_fp16 = reshape(shape = concat_132x, x = linear_154_cast_fp16)[name = string("op_4031_cast_fp16")];
            tensor<int32, [4]> x_577_perm_0 = const()[name = string("x_577_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370026880))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371075520))))[name = string("layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_155_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = var_4014_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_4036_cast_fp16 = reshape(shape = concat_133x, x = linear_155_cast_fp16)[name = string("op_4036_cast_fp16")];
            tensor<int32, [4]> x_581_perm_0 = const()[name = string("x_581_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371141120))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372189760))))[name = string("layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_156_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = var_4014_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_4041_cast_fp16 = reshape(shape = concat_134x, x = linear_156_cast_fp16)[name = string("op_4041_cast_fp16")];
            tensor<int32, [4]> v_45_perm_0 = const()[name = string("v_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3997_promoted_1_to_fp16 = const()[name = string("op_3997_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_577_cast_fp16 = transpose(perm = x_577_perm_0, x = var_4031_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 16, ?, 128]> var_4045_cast_fp16 = pow(x = x_577_cast_fp16, y = var_3997_promoted_1_to_fp16)[name = string("op_4045_cast_fp16")];
            tensor<int32, [1]> var_4047_axes_0 = const()[name = string("op_4047_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4047_keep_dims_0 = const()[name = string("op_4047_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_4047_cast_fp16 = reduce_mean(axes = var_4047_axes_0, keep_dims = var_4047_keep_dims_0, x = var_4045_cast_fp16)[name = string("op_4047_cast_fp16")];
            fp16 var_4048_to_fp16 = const()[name = string("op_4048_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_4049_cast_fp16 = add(x = var_4047_cast_fp16, y = var_4048_to_fp16)[name = string("op_4049_cast_fp16")];
            fp32 norm_179_epsilon_0 = const()[name = string("norm_179_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_179_cast_fp16 = rsqrt(epsilon = norm_179_epsilon_0, x = var_4049_cast_fp16)[name = string("norm_179_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_4051_cast_fp16 = mul(x = x_577_cast_fp16, y = norm_179_cast_fp16)[name = string("op_4051_cast_fp16")];
            tensor<fp16, [128]> layers_22_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_22_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372255360)))];
            tensor<fp16, [1, 16, ?, 128]> var_4052_cast_fp16 = mul(x = var_4051_cast_fp16, y = layers_22_self_attn_q_norm_weight_to_fp16)[name = string("op_4052_cast_fp16")];
            fp16 var_3997_promoted_2_to_fp16 = const()[name = string("op_3997_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_581_cast_fp16 = transpose(perm = x_581_perm_0, x = var_4036_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 8, ?, 128]> var_4056_cast_fp16 = pow(x = x_581_cast_fp16, y = var_3997_promoted_2_to_fp16)[name = string("op_4056_cast_fp16")];
            tensor<int32, [1]> var_4058_axes_0 = const()[name = string("op_4058_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4058_keep_dims_0 = const()[name = string("op_4058_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_4058_cast_fp16 = reduce_mean(axes = var_4058_axes_0, keep_dims = var_4058_keep_dims_0, x = var_4056_cast_fp16)[name = string("op_4058_cast_fp16")];
            fp16 var_4059_to_fp16 = const()[name = string("op_4059_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_4060_cast_fp16 = add(x = var_4058_cast_fp16, y = var_4059_to_fp16)[name = string("op_4060_cast_fp16")];
            fp32 norm_181_epsilon_0 = const()[name = string("norm_181_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_181_cast_fp16 = rsqrt(epsilon = norm_181_epsilon_0, x = var_4060_cast_fp16)[name = string("norm_181_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_4062_cast_fp16 = mul(x = x_581_cast_fp16, y = norm_181_cast_fp16)[name = string("op_4062_cast_fp16")];
            tensor<fp16, [128]> layers_22_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_22_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372255680)))];
            tensor<fp16, [1, 8, ?, 128]> var_4063_cast_fp16 = mul(x = var_4062_cast_fp16, y = layers_22_self_attn_k_norm_weight_to_fp16)[name = string("op_4063_cast_fp16")];
            tensor<int32, [4]> x1_89_begin_0 = const()[name = string("x1_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_89_end_0 = const()[name = string("x1_89_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_89_end_mask_0 = const()[name = string("x1_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_89_cast_fp16 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = var_4052_cast_fp16)[name = string("x1_89_cast_fp16")];
            tensor<int32, [4]> x2_89_begin_0 = const()[name = string("x2_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_89_end_0 = const()[name = string("x2_89_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_89_end_mask_0 = const()[name = string("x2_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_89_cast_fp16 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = var_4052_cast_fp16)[name = string("x2_89_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4079_cast_fp16 = mul(x = x1_89_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4079_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4080_cast_fp16 = mul(x = x2_89_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4080_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4081_cast_fp16 = sub(x = var_4079_cast_fp16, y = var_4080_cast_fp16)[name = string("op_4081_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4082_cast_fp16 = mul(x = x2_89_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4082_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4083_cast_fp16 = mul(x = x1_89_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4083_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4084_cast_fp16 = add(x = var_4082_cast_fp16, y = var_4083_cast_fp16)[name = string("op_4084_cast_fp16")];
            bool q_45_interleave_0 = const()[name = string("q_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_45_cast_fp16 = concat(axis = var_3998, interleave = q_45_interleave_0, values = (var_4081_cast_fp16, var_4084_cast_fp16))[name = string("q_45_cast_fp16")];
            tensor<int32, [4]> x1_91_begin_0 = const()[name = string("x1_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_91_end_0 = const()[name = string("x1_91_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_91_end_mask_0 = const()[name = string("x1_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_91_cast_fp16 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = var_4063_cast_fp16)[name = string("x1_91_cast_fp16")];
            tensor<int32, [4]> x2_91_begin_0 = const()[name = string("x2_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_91_end_0 = const()[name = string("x2_91_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_91_end_mask_0 = const()[name = string("x2_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_91_cast_fp16 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = var_4063_cast_fp16)[name = string("x2_91_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4101_cast_fp16 = mul(x = x1_91_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4101_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4102_cast_fp16 = mul(x = x2_91_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4102_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4103_cast_fp16 = sub(x = var_4101_cast_fp16, y = var_4102_cast_fp16)[name = string("op_4103_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4104_cast_fp16 = mul(x = x2_91_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4104_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4105_cast_fp16 = mul(x = x1_91_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4105_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4106_cast_fp16 = add(x = var_4104_cast_fp16, y = var_4105_cast_fp16)[name = string("op_4106_cast_fp16")];
            bool k_45_interleave_0 = const()[name = string("k_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_45_cast_fp16 = concat(axis = var_3998, interleave = k_45_interleave_0, values = (var_4103_cast_fp16, var_4106_cast_fp16))[name = string("k_45_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_44 = read_state(input = k_cache_22)[name = string("read_state_44")];
            int32 k_cache_91_axis_0 = const()[name = string("k_cache_91_axis_0"), val = int32(2)];
            string k_cache_91_mode_0 = const()[name = string("k_cache_91_mode_0"), val = string("update")];
            bool k_cache_91_validate_indices_0 = const()[name = string("k_cache_91_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_91_cast_fp16 = scatter_along_axis(axis = k_cache_91_axis_0, data = read_state_44, indices = scatter_indices, mode = k_cache_91_mode_0, updates = k_45_cast_fp16, validate_indices = k_cache_91_validate_indices_0)[name = string("k_cache_91_cast_fp16")];
            write_state(data = k_cache_91_cast_fp16, input = k_cache_22)[name = string("coreml_update_state_44_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_100 = read_state(input = k_cache_22)[name = string("coreml_update_state_44")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_45 = read_state(input = v_cache_22)[name = string("read_state_45")];
            int32 v_cache_91_axis_0 = const()[name = string("v_cache_91_axis_0"), val = int32(2)];
            string v_cache_91_mode_0 = const()[name = string("v_cache_91_mode_0"), val = string("update")];
            bool v_cache_91_validate_indices_0 = const()[name = string("v_cache_91_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_45_cast_fp16 = transpose(perm = v_45_perm_0, x = var_4041_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_91_cast_fp16 = scatter_along_axis(axis = v_cache_91_axis_0, data = read_state_45, indices = scatter_indices, mode = v_cache_91_mode_0, updates = v_45_cast_fp16, validate_indices = v_cache_91_validate_indices_0)[name = string("v_cache_91_cast_fp16")];
            write_state(data = v_cache_91_cast_fp16, input = v_cache_22)[name = string("coreml_update_state_45_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_101 = read_state(input = v_cache_22)[name = string("coreml_update_state_45")];
            tensor<int32, [1]> var_4111_axes_0 = const()[name = string("op_4111_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4111_cast_fp16 = expand_dims(axes = var_4111_axes_0, x = coreml_update_state_100)[name = string("op_4111_cast_fp16")];
            tensor<int32, [5]> k_exp_89_reps_0 = const()[name = string("k_exp_89_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_89_cast_fp16 = tile(reps = k_exp_89_reps_0, x = var_4111_cast_fp16)[name = string("k_exp_89_cast_fp16")];
            tensor<int32, [4]> var_4114 = const()[name = string("op_4114"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_91_cast_fp16 = reshape(shape = var_4114, x = k_exp_89_cast_fp16)[name = string("k_exp_91_cast_fp16")];
            tensor<int32, [1]> var_4116_axes_0 = const()[name = string("op_4116_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4116_cast_fp16 = expand_dims(axes = var_4116_axes_0, x = coreml_update_state_101)[name = string("op_4116_cast_fp16")];
            tensor<int32, [5]> v_exp_89_reps_0 = const()[name = string("v_exp_89_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_89_cast_fp16 = tile(reps = v_exp_89_reps_0, x = var_4116_cast_fp16)[name = string("v_exp_89_cast_fp16")];
            tensor<int32, [4]> var_4119 = const()[name = string("op_4119"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_91_cast_fp16 = reshape(shape = var_4119, x = v_exp_89_cast_fp16)[name = string("v_exp_91_cast_fp16")];
            bool var_4122_transpose_x_1 = const()[name = string("op_4122_transpose_x_1"), val = bool(false)];
            bool var_4122_transpose_y_1 = const()[name = string("op_4122_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_4122_cast_fp16 = matmul(transpose_x = var_4122_transpose_x_1, transpose_y = var_4122_transpose_y_1, x = q_45_cast_fp16, y = k_exp_91_cast_fp16)[name = string("op_4122_cast_fp16")];
            fp16 var_4123_to_fp16 = const()[name = string("op_4123_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_89_cast_fp16 = mul(x = var_4122_cast_fp16, y = var_4123_to_fp16)[name = string("attn_89_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_221_cast_fp16 = add(x = attn_89_cast_fp16, y = attention_mask)[name = string("input_221_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_91_cast_fp16 = softmax(axis = var_3998, x = input_221_cast_fp16)[name = string("attn_91_cast_fp16")];
            bool out_45_transpose_x_0 = const()[name = string("out_45_transpose_x_0"), val = bool(false)];
            bool out_45_transpose_y_0 = const()[name = string("out_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_45_cast_fp16 = matmul(transpose_x = out_45_transpose_x_0, transpose_y = out_45_transpose_y_0, x = attn_91_cast_fp16, y = v_exp_91_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<int32, [4]> var_4128_perm_0 = const()[name = string("op_4128_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_137_axis_0 = const()[name = string("concat_137_axis_0"), val = int32(0)];
            bool concat_137_interleave_0 = const()[name = string("concat_137_interleave_0"), val = bool(false)];
            int32 gather_133_cast_uint16_to_int32 = cast(dtype = gather_133_cast_uint16_to_int32_dtype_0, x = gather_133_cast_uint16)[name = string("cast_10")];
            tensor<int32, [3]> concat_137 = concat(axis = concat_137_axis_0, interleave = concat_137_interleave_0, values = (gather_132, gather_133_cast_uint16_to_int32, var_3998))[name = string("concat_137")];
            tensor<fp16, [1, ?, 16, 128]> var_4128_cast_fp16 = transpose(perm = var_4128_perm_0, x = out_45_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, ?, ?]> input_223_cast_fp16 = reshape(shape = concat_137, x = var_4128_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372256000))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374353216))))[name = string("layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_157_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = input_223_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_591_cast_fp16 = add(x = x_571_cast_fp16, y = linear_157_cast_fp16)[name = string("x_591_cast_fp16")];
            fp16 var_3997_promoted_3_to_fp16 = const()[name = string("op_3997_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_4136_cast_fp16 = pow(x = x_591_cast_fp16, y = var_3997_promoted_3_to_fp16)[name = string("op_4136_cast_fp16")];
            tensor<int32, [1]> var_4138_axes_0 = const()[name = string("op_4138_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4138_keep_dims_0 = const()[name = string("op_4138_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_4138_cast_fp16 = reduce_mean(axes = var_4138_axes_0, keep_dims = var_4138_keep_dims_0, x = var_4136_cast_fp16)[name = string("op_4138_cast_fp16")];
            fp16 var_4139_to_fp16 = const()[name = string("op_4139_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4140_cast_fp16 = add(x = var_4138_cast_fp16, y = var_4139_to_fp16)[name = string("op_4140_cast_fp16")];
            fp32 norm_183_epsilon_0 = const()[name = string("norm_183_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_183_cast_fp16 = rsqrt(epsilon = norm_183_epsilon_0, x = var_4140_cast_fp16)[name = string("norm_183_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_4142_cast_fp16 = mul(x = x_591_cast_fp16, y = norm_183_cast_fp16)[name = string("op_4142_cast_fp16")];
            tensor<fp16, [1024]> layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374484352)))];
            tensor<fp16, [1, ?, 1024]> var_4143_cast_fp16 = mul(x = var_4142_cast_fp16, y = layers_22_post_attention_layernorm_weight_to_fp16)[name = string("op_4143_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374486464))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377632256))))[name = string("layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = var_4143_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_4153_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_4153_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377828928))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380974720))))[name = string("layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_22_mlp_up_proj_weight_to_fp16_quantized, x = var_4143_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_229_cast_fp16 = mul(x = var_4153_cast_fp16, y = linear_159_cast_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381171392))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384317184))))[name = string("layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_160_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_597_cast_fp16 = add(x = x_591_cast_fp16, y = linear_160_cast_fp16)[name = string("x_597_cast_fp16")];
            int32 var_4173 = const()[name = string("op_4173"), val = int32(-1)];
            fp16 var_4172_promoted_to_fp16 = const()[name = string("op_4172_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_4182_cast_fp16 = pow(x = x_597_cast_fp16, y = var_4172_promoted_to_fp16)[name = string("op_4182_cast_fp16")];
            tensor<int32, [1]> var_4184_axes_0 = const()[name = string("op_4184_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4184_keep_dims_0 = const()[name = string("op_4184_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_4184_cast_fp16 = reduce_mean(axes = var_4184_axes_0, keep_dims = var_4184_keep_dims_0, x = var_4182_cast_fp16)[name = string("op_4184_cast_fp16")];
            fp16 var_4185_to_fp16 = const()[name = string("op_4185_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4186_cast_fp16 = add(x = var_4184_cast_fp16, y = var_4185_to_fp16)[name = string("op_4186_cast_fp16")];
            fp32 norm_185_epsilon_0 = const()[name = string("norm_185_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_185_cast_fp16 = rsqrt(epsilon = norm_185_epsilon_0, x = var_4186_cast_fp16)[name = string("norm_185_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_4188_cast_fp16 = mul(x = x_597_cast_fp16, y = norm_185_cast_fp16)[name = string("op_4188_cast_fp16")];
            tensor<fp16, [1024]> layers_23_input_layernorm_weight_to_fp16 = const()[name = string("layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384513856)))];
            tensor<fp16, [1, ?, 1024]> var_4189_cast_fp16 = mul(x = var_4188_cast_fp16, y = layers_23_input_layernorm_weight_to_fp16)[name = string("op_4189_cast_fp16")];
            tensor<int32, [3]> var_4201_shape_cast_fp16 = shape(x = var_4189_cast_fp16)[name = string("op_4201_shape_cast_fp16")];
            int32 gather_138 = const()[name = string("gather_138"), val = int32(1)];
            int32 gather_139_axis_0 = const()[name = string("gather_139_axis_0"), val = int32(0)];
            int32 gather_139_batch_dims_0 = const()[name = string("gather_139_batch_dims_0"), val = int32(0)];
            bool gather_139_validate_indices_0 = const()[name = string("gather_139_validate_indices_0"), val = bool(false)];
            string var_4201_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4201_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_139_to_uint16 = const()[name = string("select_139_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_4201_shape_cast_fp16_to_uint16 = cast(dtype = var_4201_shape_cast_fp16_to_uint16_dtype_0, x = var_4201_shape_cast_fp16)[name = string("cast_9")];
            uint16 gather_139_cast_uint16 = gather(axis = gather_139_axis_0, batch_dims = gather_139_batch_dims_0, indices = select_139_to_uint16, validate_indices = gather_139_validate_indices_0, x = var_4201_shape_cast_fp16_to_uint16)[name = string("gather_139_cast_uint16")];
            string gather_139_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_139_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384515968))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386613184))))[name = string("layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_161_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = var_4189_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<int32, [4]> concat_138x = const()[name = string("concat_138x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_4206_cast_fp16 = reshape(shape = concat_138x, x = linear_161_cast_fp16)[name = string("op_4206_cast_fp16")];
            tensor<int32, [4]> x_603_perm_0 = const()[name = string("x_603_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386744320))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387792960))))[name = string("layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_162_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = var_4189_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [4]> concat_139x = const()[name = string("concat_139x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_4211_cast_fp16 = reshape(shape = concat_139x, x = linear_162_cast_fp16)[name = string("op_4211_cast_fp16")];
            tensor<int32, [4]> x_607_perm_0 = const()[name = string("x_607_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387858560))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388907200))))[name = string("layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_163_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = var_4189_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> concat_140x = const()[name = string("concat_140x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_4216_cast_fp16 = reshape(shape = concat_140x, x = linear_163_cast_fp16)[name = string("op_4216_cast_fp16")];
            tensor<int32, [4]> v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4172_promoted_1_to_fp16 = const()[name = string("op_4172_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_603_cast_fp16 = transpose(perm = x_603_perm_0, x = var_4206_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 16, ?, 128]> var_4220_cast_fp16 = pow(x = x_603_cast_fp16, y = var_4172_promoted_1_to_fp16)[name = string("op_4220_cast_fp16")];
            tensor<int32, [1]> var_4222_axes_0 = const()[name = string("op_4222_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4222_keep_dims_0 = const()[name = string("op_4222_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_4222_cast_fp16 = reduce_mean(axes = var_4222_axes_0, keep_dims = var_4222_keep_dims_0, x = var_4220_cast_fp16)[name = string("op_4222_cast_fp16")];
            fp16 var_4223_to_fp16 = const()[name = string("op_4223_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_4224_cast_fp16 = add(x = var_4222_cast_fp16, y = var_4223_to_fp16)[name = string("op_4224_cast_fp16")];
            fp32 norm_187_epsilon_0 = const()[name = string("norm_187_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_187_cast_fp16 = rsqrt(epsilon = norm_187_epsilon_0, x = var_4224_cast_fp16)[name = string("norm_187_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_4226_cast_fp16 = mul(x = x_603_cast_fp16, y = norm_187_cast_fp16)[name = string("op_4226_cast_fp16")];
            tensor<fp16, [128]> layers_23_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_23_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388972800)))];
            tensor<fp16, [1, 16, ?, 128]> var_4227_cast_fp16 = mul(x = var_4226_cast_fp16, y = layers_23_self_attn_q_norm_weight_to_fp16)[name = string("op_4227_cast_fp16")];
            fp16 var_4172_promoted_2_to_fp16 = const()[name = string("op_4172_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_607_cast_fp16 = transpose(perm = x_607_perm_0, x = var_4211_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 8, ?, 128]> var_4231_cast_fp16 = pow(x = x_607_cast_fp16, y = var_4172_promoted_2_to_fp16)[name = string("op_4231_cast_fp16")];
            tensor<int32, [1]> var_4233_axes_0 = const()[name = string("op_4233_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4233_keep_dims_0 = const()[name = string("op_4233_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_4233_cast_fp16 = reduce_mean(axes = var_4233_axes_0, keep_dims = var_4233_keep_dims_0, x = var_4231_cast_fp16)[name = string("op_4233_cast_fp16")];
            fp16 var_4234_to_fp16 = const()[name = string("op_4234_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_4235_cast_fp16 = add(x = var_4233_cast_fp16, y = var_4234_to_fp16)[name = string("op_4235_cast_fp16")];
            fp32 norm_189_epsilon_0 = const()[name = string("norm_189_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_189_cast_fp16 = rsqrt(epsilon = norm_189_epsilon_0, x = var_4235_cast_fp16)[name = string("norm_189_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_4237_cast_fp16 = mul(x = x_607_cast_fp16, y = norm_189_cast_fp16)[name = string("op_4237_cast_fp16")];
            tensor<fp16, [128]> layers_23_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_23_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388973120)))];
            tensor<fp16, [1, 8, ?, 128]> var_4238_cast_fp16 = mul(x = var_4237_cast_fp16, y = layers_23_self_attn_k_norm_weight_to_fp16)[name = string("op_4238_cast_fp16")];
            tensor<int32, [4]> x1_93_begin_0 = const()[name = string("x1_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_93_end_0 = const()[name = string("x1_93_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_93_end_mask_0 = const()[name = string("x1_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_93_cast_fp16 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = var_4227_cast_fp16)[name = string("x1_93_cast_fp16")];
            tensor<int32, [4]> x2_93_begin_0 = const()[name = string("x2_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_93_end_0 = const()[name = string("x2_93_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_93_end_mask_0 = const()[name = string("x2_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_93_cast_fp16 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = var_4227_cast_fp16)[name = string("x2_93_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4254_cast_fp16 = mul(x = x1_93_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4254_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4255_cast_fp16 = mul(x = x2_93_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4255_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4256_cast_fp16 = sub(x = var_4254_cast_fp16, y = var_4255_cast_fp16)[name = string("op_4256_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4257_cast_fp16 = mul(x = x2_93_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4257_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4258_cast_fp16 = mul(x = x1_93_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4258_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4259_cast_fp16 = add(x = var_4257_cast_fp16, y = var_4258_cast_fp16)[name = string("op_4259_cast_fp16")];
            bool q_47_interleave_0 = const()[name = string("q_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_47_cast_fp16 = concat(axis = var_4173, interleave = q_47_interleave_0, values = (var_4256_cast_fp16, var_4259_cast_fp16))[name = string("q_47_cast_fp16")];
            tensor<int32, [4]> x1_95_begin_0 = const()[name = string("x1_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_95_end_0 = const()[name = string("x1_95_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_95_end_mask_0 = const()[name = string("x1_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_95_cast_fp16 = slice_by_index(begin = x1_95_begin_0, end = x1_95_end_0, end_mask = x1_95_end_mask_0, x = var_4238_cast_fp16)[name = string("x1_95_cast_fp16")];
            tensor<int32, [4]> x2_95_begin_0 = const()[name = string("x2_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_95_end_0 = const()[name = string("x2_95_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_95_end_mask_0 = const()[name = string("x2_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_95_cast_fp16 = slice_by_index(begin = x2_95_begin_0, end = x2_95_end_0, end_mask = x2_95_end_mask_0, x = var_4238_cast_fp16)[name = string("x2_95_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4276_cast_fp16 = mul(x = x1_95_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4276_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4277_cast_fp16 = mul(x = x2_95_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4277_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4278_cast_fp16 = sub(x = var_4276_cast_fp16, y = var_4277_cast_fp16)[name = string("op_4278_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4279_cast_fp16 = mul(x = x2_95_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4279_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4280_cast_fp16 = mul(x = x1_95_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4280_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4281_cast_fp16 = add(x = var_4279_cast_fp16, y = var_4280_cast_fp16)[name = string("op_4281_cast_fp16")];
            bool k_47_interleave_0 = const()[name = string("k_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_47_cast_fp16 = concat(axis = var_4173, interleave = k_47_interleave_0, values = (var_4278_cast_fp16, var_4281_cast_fp16))[name = string("k_47_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_46 = read_state(input = k_cache_23)[name = string("read_state_46")];
            int32 k_cache_95_axis_0 = const()[name = string("k_cache_95_axis_0"), val = int32(2)];
            string k_cache_95_mode_0 = const()[name = string("k_cache_95_mode_0"), val = string("update")];
            bool k_cache_95_validate_indices_0 = const()[name = string("k_cache_95_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_95_cast_fp16 = scatter_along_axis(axis = k_cache_95_axis_0, data = read_state_46, indices = scatter_indices, mode = k_cache_95_mode_0, updates = k_47_cast_fp16, validate_indices = k_cache_95_validate_indices_0)[name = string("k_cache_95_cast_fp16")];
            write_state(data = k_cache_95_cast_fp16, input = k_cache_23)[name = string("coreml_update_state_46_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_102 = read_state(input = k_cache_23)[name = string("coreml_update_state_46")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_47 = read_state(input = v_cache_23)[name = string("read_state_47")];
            int32 v_cache_95_axis_0 = const()[name = string("v_cache_95_axis_0"), val = int32(2)];
            string v_cache_95_mode_0 = const()[name = string("v_cache_95_mode_0"), val = string("update")];
            bool v_cache_95_validate_indices_0 = const()[name = string("v_cache_95_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = var_4216_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_95_cast_fp16 = scatter_along_axis(axis = v_cache_95_axis_0, data = read_state_47, indices = scatter_indices, mode = v_cache_95_mode_0, updates = v_47_cast_fp16, validate_indices = v_cache_95_validate_indices_0)[name = string("v_cache_95_cast_fp16")];
            write_state(data = v_cache_95_cast_fp16, input = v_cache_23)[name = string("coreml_update_state_47_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_103 = read_state(input = v_cache_23)[name = string("coreml_update_state_47")];
            tensor<int32, [1]> var_4286_axes_0 = const()[name = string("op_4286_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4286_cast_fp16 = expand_dims(axes = var_4286_axes_0, x = coreml_update_state_102)[name = string("op_4286_cast_fp16")];
            tensor<int32, [5]> k_exp_93_reps_0 = const()[name = string("k_exp_93_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_93_cast_fp16 = tile(reps = k_exp_93_reps_0, x = var_4286_cast_fp16)[name = string("k_exp_93_cast_fp16")];
            tensor<int32, [4]> var_4289 = const()[name = string("op_4289"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_95_cast_fp16 = reshape(shape = var_4289, x = k_exp_93_cast_fp16)[name = string("k_exp_95_cast_fp16")];
            tensor<int32, [1]> var_4291_axes_0 = const()[name = string("op_4291_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4291_cast_fp16 = expand_dims(axes = var_4291_axes_0, x = coreml_update_state_103)[name = string("op_4291_cast_fp16")];
            tensor<int32, [5]> v_exp_93_reps_0 = const()[name = string("v_exp_93_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_93_cast_fp16 = tile(reps = v_exp_93_reps_0, x = var_4291_cast_fp16)[name = string("v_exp_93_cast_fp16")];
            tensor<int32, [4]> var_4294 = const()[name = string("op_4294"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_95_cast_fp16 = reshape(shape = var_4294, x = v_exp_93_cast_fp16)[name = string("v_exp_95_cast_fp16")];
            bool var_4297_transpose_x_1 = const()[name = string("op_4297_transpose_x_1"), val = bool(false)];
            bool var_4297_transpose_y_1 = const()[name = string("op_4297_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_4297_cast_fp16 = matmul(transpose_x = var_4297_transpose_x_1, transpose_y = var_4297_transpose_y_1, x = q_47_cast_fp16, y = k_exp_95_cast_fp16)[name = string("op_4297_cast_fp16")];
            fp16 var_4298_to_fp16 = const()[name = string("op_4298_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_93_cast_fp16 = mul(x = var_4297_cast_fp16, y = var_4298_to_fp16)[name = string("attn_93_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_231_cast_fp16 = add(x = attn_93_cast_fp16, y = attention_mask)[name = string("input_231_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_95_cast_fp16 = softmax(axis = var_4173, x = input_231_cast_fp16)[name = string("attn_95_cast_fp16")];
            bool out_47_transpose_x_0 = const()[name = string("out_47_transpose_x_0"), val = bool(false)];
            bool out_47_transpose_y_0 = const()[name = string("out_47_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_47_cast_fp16 = matmul(transpose_x = out_47_transpose_x_0, transpose_y = out_47_transpose_y_0, x = attn_95_cast_fp16, y = v_exp_95_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<int32, [4]> var_4303_perm_0 = const()[name = string("op_4303_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_143_axis_0 = const()[name = string("concat_143_axis_0"), val = int32(0)];
            bool concat_143_interleave_0 = const()[name = string("concat_143_interleave_0"), val = bool(false)];
            int32 gather_139_cast_uint16_to_int32 = cast(dtype = gather_139_cast_uint16_to_int32_dtype_0, x = gather_139_cast_uint16)[name = string("cast_8")];
            tensor<int32, [3]> concat_143 = concat(axis = concat_143_axis_0, interleave = concat_143_interleave_0, values = (gather_138, gather_139_cast_uint16_to_int32, var_4173))[name = string("concat_143")];
            tensor<fp16, [1, ?, 16, 128]> var_4303_cast_fp16 = transpose(perm = var_4303_perm_0, x = out_47_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, ?, ?]> input_233_cast_fp16 = reshape(shape = concat_143, x = var_4303_cast_fp16)[name = string("input_233_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388973440))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391070656))))[name = string("layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_164_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_617_cast_fp16 = add(x = x_597_cast_fp16, y = linear_164_cast_fp16)[name = string("x_617_cast_fp16")];
            fp16 var_4172_promoted_3_to_fp16 = const()[name = string("op_4172_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_4311_cast_fp16 = pow(x = x_617_cast_fp16, y = var_4172_promoted_3_to_fp16)[name = string("op_4311_cast_fp16")];
            tensor<int32, [1]> var_4313_axes_0 = const()[name = string("op_4313_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4313_keep_dims_0 = const()[name = string("op_4313_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_4313_cast_fp16 = reduce_mean(axes = var_4313_axes_0, keep_dims = var_4313_keep_dims_0, x = var_4311_cast_fp16)[name = string("op_4313_cast_fp16")];
            fp16 var_4314_to_fp16 = const()[name = string("op_4314_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4315_cast_fp16 = add(x = var_4313_cast_fp16, y = var_4314_to_fp16)[name = string("op_4315_cast_fp16")];
            fp32 norm_191_epsilon_0 = const()[name = string("norm_191_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_191_cast_fp16 = rsqrt(epsilon = norm_191_epsilon_0, x = var_4315_cast_fp16)[name = string("norm_191_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_4317_cast_fp16 = mul(x = x_617_cast_fp16, y = norm_191_cast_fp16)[name = string("op_4317_cast_fp16")];
            tensor<fp16, [1024]> layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391201792)))];
            tensor<fp16, [1, ?, 1024]> var_4318_cast_fp16 = mul(x = var_4317_cast_fp16, y = layers_23_post_attention_layernorm_weight_to_fp16)[name = string("op_4318_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391203904))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394349696))))[name = string("layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = var_4318_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_4328_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_4328_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394546368))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397692160))))[name = string("layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_166_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_23_mlp_up_proj_weight_to_fp16_quantized, x = var_4318_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_239_cast_fp16 = mul(x = var_4328_cast_fp16, y = linear_166_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397888832))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401034624))))[name = string("layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_167_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_239_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_623_cast_fp16 = add(x = x_617_cast_fp16, y = linear_167_cast_fp16)[name = string("x_623_cast_fp16")];
            int32 var_4348 = const()[name = string("op_4348"), val = int32(-1)];
            fp16 var_4347_promoted_to_fp16 = const()[name = string("op_4347_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_4357_cast_fp16 = pow(x = x_623_cast_fp16, y = var_4347_promoted_to_fp16)[name = string("op_4357_cast_fp16")];
            tensor<int32, [1]> var_4359_axes_0 = const()[name = string("op_4359_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4359_keep_dims_0 = const()[name = string("op_4359_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_4359_cast_fp16 = reduce_mean(axes = var_4359_axes_0, keep_dims = var_4359_keep_dims_0, x = var_4357_cast_fp16)[name = string("op_4359_cast_fp16")];
            fp16 var_4360_to_fp16 = const()[name = string("op_4360_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4361_cast_fp16 = add(x = var_4359_cast_fp16, y = var_4360_to_fp16)[name = string("op_4361_cast_fp16")];
            fp32 norm_193_epsilon_0 = const()[name = string("norm_193_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_193_cast_fp16 = rsqrt(epsilon = norm_193_epsilon_0, x = var_4361_cast_fp16)[name = string("norm_193_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_4363_cast_fp16 = mul(x = x_623_cast_fp16, y = norm_193_cast_fp16)[name = string("op_4363_cast_fp16")];
            tensor<fp16, [1024]> layers_24_input_layernorm_weight_to_fp16 = const()[name = string("layers_24_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401231296)))];
            tensor<fp16, [1, ?, 1024]> var_4364_cast_fp16 = mul(x = var_4363_cast_fp16, y = layers_24_input_layernorm_weight_to_fp16)[name = string("op_4364_cast_fp16")];
            tensor<int32, [3]> var_4376_shape_cast_fp16 = shape(x = var_4364_cast_fp16)[name = string("op_4376_shape_cast_fp16")];
            int32 gather_144 = const()[name = string("gather_144"), val = int32(1)];
            int32 gather_145_axis_0 = const()[name = string("gather_145_axis_0"), val = int32(0)];
            int32 gather_145_batch_dims_0 = const()[name = string("gather_145_batch_dims_0"), val = int32(0)];
            bool gather_145_validate_indices_0 = const()[name = string("gather_145_validate_indices_0"), val = bool(false)];
            string var_4376_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4376_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_145_to_uint16 = const()[name = string("select_145_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_4376_shape_cast_fp16_to_uint16 = cast(dtype = var_4376_shape_cast_fp16_to_uint16_dtype_0, x = var_4376_shape_cast_fp16)[name = string("cast_7")];
            uint16 gather_145_cast_uint16 = gather(axis = gather_145_axis_0, batch_dims = gather_145_batch_dims_0, indices = select_145_to_uint16, validate_indices = gather_145_validate_indices_0, x = var_4376_shape_cast_fp16_to_uint16)[name = string("gather_145_cast_uint16")];
            string gather_145_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_145_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_24_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401233408))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403330624))))[name = string("layers_24_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_168_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_24_self_attn_q_proj_weight_to_fp16_quantized, x = var_4364_cast_fp16)[name = string("linear_168_cast_fp16")];
            tensor<int32, [4]> concat_144x = const()[name = string("concat_144x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_4381_cast_fp16 = reshape(shape = concat_144x, x = linear_168_cast_fp16)[name = string("op_4381_cast_fp16")];
            tensor<int32, [4]> x_629_perm_0 = const()[name = string("x_629_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_24_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403461760))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404510400))))[name = string("layers_24_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_169_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_24_self_attn_k_proj_weight_to_fp16_quantized, x = var_4364_cast_fp16)[name = string("linear_169_cast_fp16")];
            tensor<int32, [4]> concat_145x = const()[name = string("concat_145x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_4386_cast_fp16 = reshape(shape = concat_145x, x = linear_169_cast_fp16)[name = string("op_4386_cast_fp16")];
            tensor<int32, [4]> x_633_perm_0 = const()[name = string("x_633_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_24_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404576000))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405624640))))[name = string("layers_24_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_170_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_24_self_attn_v_proj_weight_to_fp16_quantized, x = var_4364_cast_fp16)[name = string("linear_170_cast_fp16")];
            tensor<int32, [4]> concat_146x = const()[name = string("concat_146x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_4391_cast_fp16 = reshape(shape = concat_146x, x = linear_170_cast_fp16)[name = string("op_4391_cast_fp16")];
            tensor<int32, [4]> v_49_perm_0 = const()[name = string("v_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4347_promoted_1_to_fp16 = const()[name = string("op_4347_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_629_cast_fp16 = transpose(perm = x_629_perm_0, x = var_4381_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 16, ?, 128]> var_4395_cast_fp16 = pow(x = x_629_cast_fp16, y = var_4347_promoted_1_to_fp16)[name = string("op_4395_cast_fp16")];
            tensor<int32, [1]> var_4397_axes_0 = const()[name = string("op_4397_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4397_keep_dims_0 = const()[name = string("op_4397_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_4397_cast_fp16 = reduce_mean(axes = var_4397_axes_0, keep_dims = var_4397_keep_dims_0, x = var_4395_cast_fp16)[name = string("op_4397_cast_fp16")];
            fp16 var_4398_to_fp16 = const()[name = string("op_4398_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_4399_cast_fp16 = add(x = var_4397_cast_fp16, y = var_4398_to_fp16)[name = string("op_4399_cast_fp16")];
            fp32 norm_195_epsilon_0 = const()[name = string("norm_195_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_195_cast_fp16 = rsqrt(epsilon = norm_195_epsilon_0, x = var_4399_cast_fp16)[name = string("norm_195_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_4401_cast_fp16 = mul(x = x_629_cast_fp16, y = norm_195_cast_fp16)[name = string("op_4401_cast_fp16")];
            tensor<fp16, [128]> layers_24_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_24_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405690240)))];
            tensor<fp16, [1, 16, ?, 128]> var_4402_cast_fp16 = mul(x = var_4401_cast_fp16, y = layers_24_self_attn_q_norm_weight_to_fp16)[name = string("op_4402_cast_fp16")];
            fp16 var_4347_promoted_2_to_fp16 = const()[name = string("op_4347_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_633_cast_fp16 = transpose(perm = x_633_perm_0, x = var_4386_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 8, ?, 128]> var_4406_cast_fp16 = pow(x = x_633_cast_fp16, y = var_4347_promoted_2_to_fp16)[name = string("op_4406_cast_fp16")];
            tensor<int32, [1]> var_4408_axes_0 = const()[name = string("op_4408_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4408_keep_dims_0 = const()[name = string("op_4408_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_4408_cast_fp16 = reduce_mean(axes = var_4408_axes_0, keep_dims = var_4408_keep_dims_0, x = var_4406_cast_fp16)[name = string("op_4408_cast_fp16")];
            fp16 var_4409_to_fp16 = const()[name = string("op_4409_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_4410_cast_fp16 = add(x = var_4408_cast_fp16, y = var_4409_to_fp16)[name = string("op_4410_cast_fp16")];
            fp32 norm_197_epsilon_0 = const()[name = string("norm_197_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_197_cast_fp16 = rsqrt(epsilon = norm_197_epsilon_0, x = var_4410_cast_fp16)[name = string("norm_197_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_4412_cast_fp16 = mul(x = x_633_cast_fp16, y = norm_197_cast_fp16)[name = string("op_4412_cast_fp16")];
            tensor<fp16, [128]> layers_24_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_24_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405690560)))];
            tensor<fp16, [1, 8, ?, 128]> var_4413_cast_fp16 = mul(x = var_4412_cast_fp16, y = layers_24_self_attn_k_norm_weight_to_fp16)[name = string("op_4413_cast_fp16")];
            tensor<int32, [4]> x1_97_begin_0 = const()[name = string("x1_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_97_end_0 = const()[name = string("x1_97_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_97_end_mask_0 = const()[name = string("x1_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_97_cast_fp16 = slice_by_index(begin = x1_97_begin_0, end = x1_97_end_0, end_mask = x1_97_end_mask_0, x = var_4402_cast_fp16)[name = string("x1_97_cast_fp16")];
            tensor<int32, [4]> x2_97_begin_0 = const()[name = string("x2_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_97_end_0 = const()[name = string("x2_97_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_97_end_mask_0 = const()[name = string("x2_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_97_cast_fp16 = slice_by_index(begin = x2_97_begin_0, end = x2_97_end_0, end_mask = x2_97_end_mask_0, x = var_4402_cast_fp16)[name = string("x2_97_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4429_cast_fp16 = mul(x = x1_97_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4429_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4430_cast_fp16 = mul(x = x2_97_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4430_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4431_cast_fp16 = sub(x = var_4429_cast_fp16, y = var_4430_cast_fp16)[name = string("op_4431_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4432_cast_fp16 = mul(x = x2_97_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4432_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4433_cast_fp16 = mul(x = x1_97_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4433_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4434_cast_fp16 = add(x = var_4432_cast_fp16, y = var_4433_cast_fp16)[name = string("op_4434_cast_fp16")];
            bool q_49_interleave_0 = const()[name = string("q_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_49_cast_fp16 = concat(axis = var_4348, interleave = q_49_interleave_0, values = (var_4431_cast_fp16, var_4434_cast_fp16))[name = string("q_49_cast_fp16")];
            tensor<int32, [4]> x1_99_begin_0 = const()[name = string("x1_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_99_end_0 = const()[name = string("x1_99_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_99_end_mask_0 = const()[name = string("x1_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_99_cast_fp16 = slice_by_index(begin = x1_99_begin_0, end = x1_99_end_0, end_mask = x1_99_end_mask_0, x = var_4413_cast_fp16)[name = string("x1_99_cast_fp16")];
            tensor<int32, [4]> x2_99_begin_0 = const()[name = string("x2_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_99_end_0 = const()[name = string("x2_99_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_99_end_mask_0 = const()[name = string("x2_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_99_cast_fp16 = slice_by_index(begin = x2_99_begin_0, end = x2_99_end_0, end_mask = x2_99_end_mask_0, x = var_4413_cast_fp16)[name = string("x2_99_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4451_cast_fp16 = mul(x = x1_99_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4451_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4452_cast_fp16 = mul(x = x2_99_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4452_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4453_cast_fp16 = sub(x = var_4451_cast_fp16, y = var_4452_cast_fp16)[name = string("op_4453_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4454_cast_fp16 = mul(x = x2_99_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4454_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4455_cast_fp16 = mul(x = x1_99_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4455_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4456_cast_fp16 = add(x = var_4454_cast_fp16, y = var_4455_cast_fp16)[name = string("op_4456_cast_fp16")];
            bool k_49_interleave_0 = const()[name = string("k_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_49_cast_fp16 = concat(axis = var_4348, interleave = k_49_interleave_0, values = (var_4453_cast_fp16, var_4456_cast_fp16))[name = string("k_49_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_48 = read_state(input = k_cache_24)[name = string("read_state_48")];
            int32 k_cache_99_axis_0 = const()[name = string("k_cache_99_axis_0"), val = int32(2)];
            string k_cache_99_mode_0 = const()[name = string("k_cache_99_mode_0"), val = string("update")];
            bool k_cache_99_validate_indices_0 = const()[name = string("k_cache_99_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_99_cast_fp16 = scatter_along_axis(axis = k_cache_99_axis_0, data = read_state_48, indices = scatter_indices, mode = k_cache_99_mode_0, updates = k_49_cast_fp16, validate_indices = k_cache_99_validate_indices_0)[name = string("k_cache_99_cast_fp16")];
            write_state(data = k_cache_99_cast_fp16, input = k_cache_24)[name = string("coreml_update_state_48_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_104 = read_state(input = k_cache_24)[name = string("coreml_update_state_48")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_49 = read_state(input = v_cache_24)[name = string("read_state_49")];
            int32 v_cache_99_axis_0 = const()[name = string("v_cache_99_axis_0"), val = int32(2)];
            string v_cache_99_mode_0 = const()[name = string("v_cache_99_mode_0"), val = string("update")];
            bool v_cache_99_validate_indices_0 = const()[name = string("v_cache_99_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_49_cast_fp16 = transpose(perm = v_49_perm_0, x = var_4391_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_99_cast_fp16 = scatter_along_axis(axis = v_cache_99_axis_0, data = read_state_49, indices = scatter_indices, mode = v_cache_99_mode_0, updates = v_49_cast_fp16, validate_indices = v_cache_99_validate_indices_0)[name = string("v_cache_99_cast_fp16")];
            write_state(data = v_cache_99_cast_fp16, input = v_cache_24)[name = string("coreml_update_state_49_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_105 = read_state(input = v_cache_24)[name = string("coreml_update_state_49")];
            tensor<int32, [1]> var_4461_axes_0 = const()[name = string("op_4461_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4461_cast_fp16 = expand_dims(axes = var_4461_axes_0, x = coreml_update_state_104)[name = string("op_4461_cast_fp16")];
            tensor<int32, [5]> k_exp_97_reps_0 = const()[name = string("k_exp_97_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_97_cast_fp16 = tile(reps = k_exp_97_reps_0, x = var_4461_cast_fp16)[name = string("k_exp_97_cast_fp16")];
            tensor<int32, [4]> var_4464 = const()[name = string("op_4464"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_99_cast_fp16 = reshape(shape = var_4464, x = k_exp_97_cast_fp16)[name = string("k_exp_99_cast_fp16")];
            tensor<int32, [1]> var_4466_axes_0 = const()[name = string("op_4466_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4466_cast_fp16 = expand_dims(axes = var_4466_axes_0, x = coreml_update_state_105)[name = string("op_4466_cast_fp16")];
            tensor<int32, [5]> v_exp_97_reps_0 = const()[name = string("v_exp_97_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_97_cast_fp16 = tile(reps = v_exp_97_reps_0, x = var_4466_cast_fp16)[name = string("v_exp_97_cast_fp16")];
            tensor<int32, [4]> var_4469 = const()[name = string("op_4469"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_99_cast_fp16 = reshape(shape = var_4469, x = v_exp_97_cast_fp16)[name = string("v_exp_99_cast_fp16")];
            bool var_4472_transpose_x_1 = const()[name = string("op_4472_transpose_x_1"), val = bool(false)];
            bool var_4472_transpose_y_1 = const()[name = string("op_4472_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_4472_cast_fp16 = matmul(transpose_x = var_4472_transpose_x_1, transpose_y = var_4472_transpose_y_1, x = q_49_cast_fp16, y = k_exp_99_cast_fp16)[name = string("op_4472_cast_fp16")];
            fp16 var_4473_to_fp16 = const()[name = string("op_4473_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_97_cast_fp16 = mul(x = var_4472_cast_fp16, y = var_4473_to_fp16)[name = string("attn_97_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_241_cast_fp16 = add(x = attn_97_cast_fp16, y = attention_mask)[name = string("input_241_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_99_cast_fp16 = softmax(axis = var_4348, x = input_241_cast_fp16)[name = string("attn_99_cast_fp16")];
            bool out_49_transpose_x_0 = const()[name = string("out_49_transpose_x_0"), val = bool(false)];
            bool out_49_transpose_y_0 = const()[name = string("out_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_49_cast_fp16 = matmul(transpose_x = out_49_transpose_x_0, transpose_y = out_49_transpose_y_0, x = attn_99_cast_fp16, y = v_exp_99_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<int32, [4]> var_4478_perm_0 = const()[name = string("op_4478_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_149_axis_0 = const()[name = string("concat_149_axis_0"), val = int32(0)];
            bool concat_149_interleave_0 = const()[name = string("concat_149_interleave_0"), val = bool(false)];
            int32 gather_145_cast_uint16_to_int32 = cast(dtype = gather_145_cast_uint16_to_int32_dtype_0, x = gather_145_cast_uint16)[name = string("cast_6")];
            tensor<int32, [3]> concat_149 = concat(axis = concat_149_axis_0, interleave = concat_149_interleave_0, values = (gather_144, gather_145_cast_uint16_to_int32, var_4348))[name = string("concat_149")];
            tensor<fp16, [1, ?, 16, 128]> var_4478_cast_fp16 = transpose(perm = var_4478_perm_0, x = out_49_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, ?, ?]> input_243_cast_fp16 = reshape(shape = concat_149, x = var_4478_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_24_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405690880))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407788096))))[name = string("layers_24_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_171_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_24_self_attn_o_proj_weight_to_fp16_quantized, x = input_243_cast_fp16)[name = string("linear_171_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_643_cast_fp16 = add(x = x_623_cast_fp16, y = linear_171_cast_fp16)[name = string("x_643_cast_fp16")];
            fp16 var_4347_promoted_3_to_fp16 = const()[name = string("op_4347_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_4486_cast_fp16 = pow(x = x_643_cast_fp16, y = var_4347_promoted_3_to_fp16)[name = string("op_4486_cast_fp16")];
            tensor<int32, [1]> var_4488_axes_0 = const()[name = string("op_4488_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4488_keep_dims_0 = const()[name = string("op_4488_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_4488_cast_fp16 = reduce_mean(axes = var_4488_axes_0, keep_dims = var_4488_keep_dims_0, x = var_4486_cast_fp16)[name = string("op_4488_cast_fp16")];
            fp16 var_4489_to_fp16 = const()[name = string("op_4489_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4490_cast_fp16 = add(x = var_4488_cast_fp16, y = var_4489_to_fp16)[name = string("op_4490_cast_fp16")];
            fp32 norm_199_epsilon_0 = const()[name = string("norm_199_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_199_cast_fp16 = rsqrt(epsilon = norm_199_epsilon_0, x = var_4490_cast_fp16)[name = string("norm_199_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_4492_cast_fp16 = mul(x = x_643_cast_fp16, y = norm_199_cast_fp16)[name = string("op_4492_cast_fp16")];
            tensor<fp16, [1024]> layers_24_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_24_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407919232)))];
            tensor<fp16, [1, ?, 1024]> var_4493_cast_fp16 = mul(x = var_4492_cast_fp16, y = layers_24_post_attention_layernorm_weight_to_fp16)[name = string("op_4493_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_24_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407921344))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411067136))))[name = string("layers_24_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_172_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_24_mlp_gate_proj_weight_to_fp16_quantized, x = var_4493_cast_fp16)[name = string("linear_172_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_4503_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("op_4503_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_24_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411263808))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(414409600))))[name = string("layers_24_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_173_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_24_mlp_up_proj_weight_to_fp16_quantized, x = var_4493_cast_fp16)[name = string("linear_173_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_249_cast_fp16 = mul(x = var_4503_cast_fp16, y = linear_173_cast_fp16)[name = string("input_249_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_24_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(414606272))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417752064))))[name = string("layers_24_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_174_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_24_mlp_down_proj_weight_to_fp16_quantized, x = input_249_cast_fp16)[name = string("linear_174_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_649_cast_fp16 = add(x = x_643_cast_fp16, y = linear_174_cast_fp16)[name = string("x_649_cast_fp16")];
            int32 var_4523 = const()[name = string("op_4523"), val = int32(-1)];
            fp16 var_4522_promoted_to_fp16 = const()[name = string("op_4522_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_4532_cast_fp16 = pow(x = x_649_cast_fp16, y = var_4522_promoted_to_fp16)[name = string("op_4532_cast_fp16")];
            tensor<int32, [1]> var_4534_axes_0 = const()[name = string("op_4534_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4534_keep_dims_0 = const()[name = string("op_4534_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_4534_cast_fp16 = reduce_mean(axes = var_4534_axes_0, keep_dims = var_4534_keep_dims_0, x = var_4532_cast_fp16)[name = string("op_4534_cast_fp16")];
            fp16 var_4535_to_fp16 = const()[name = string("op_4535_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4536_cast_fp16 = add(x = var_4534_cast_fp16, y = var_4535_to_fp16)[name = string("op_4536_cast_fp16")];
            fp32 norm_201_epsilon_0 = const()[name = string("norm_201_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_201_cast_fp16 = rsqrt(epsilon = norm_201_epsilon_0, x = var_4536_cast_fp16)[name = string("norm_201_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_4538_cast_fp16 = mul(x = x_649_cast_fp16, y = norm_201_cast_fp16)[name = string("op_4538_cast_fp16")];
            tensor<fp16, [1024]> layers_25_input_layernorm_weight_to_fp16 = const()[name = string("layers_25_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417948736)))];
            tensor<fp16, [1, ?, 1024]> var_4539_cast_fp16 = mul(x = var_4538_cast_fp16, y = layers_25_input_layernorm_weight_to_fp16)[name = string("op_4539_cast_fp16")];
            tensor<int32, [3]> var_4551_shape_cast_fp16 = shape(x = var_4539_cast_fp16)[name = string("op_4551_shape_cast_fp16")];
            int32 gather_150 = const()[name = string("gather_150"), val = int32(1)];
            int32 gather_151_axis_0 = const()[name = string("gather_151_axis_0"), val = int32(0)];
            int32 gather_151_batch_dims_0 = const()[name = string("gather_151_batch_dims_0"), val = int32(0)];
            bool gather_151_validate_indices_0 = const()[name = string("gather_151_validate_indices_0"), val = bool(false)];
            string var_4551_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4551_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_151_to_uint16 = const()[name = string("select_151_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_4551_shape_cast_fp16_to_uint16 = cast(dtype = var_4551_shape_cast_fp16_to_uint16_dtype_0, x = var_4551_shape_cast_fp16)[name = string("cast_5")];
            uint16 gather_151_cast_uint16 = gather(axis = gather_151_axis_0, batch_dims = gather_151_batch_dims_0, indices = select_151_to_uint16, validate_indices = gather_151_validate_indices_0, x = var_4551_shape_cast_fp16_to_uint16)[name = string("gather_151_cast_uint16")];
            string gather_151_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_151_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_25_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417950848))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420048064))))[name = string("layers_25_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_175_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_25_self_attn_q_proj_weight_to_fp16_quantized, x = var_4539_cast_fp16)[name = string("linear_175_cast_fp16")];
            tensor<int32, [4]> concat_150x = const()[name = string("concat_150x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_4556_cast_fp16 = reshape(shape = concat_150x, x = linear_175_cast_fp16)[name = string("op_4556_cast_fp16")];
            tensor<int32, [4]> x_655_perm_0 = const()[name = string("x_655_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_25_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420179200))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421227840))))[name = string("layers_25_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_176_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_25_self_attn_k_proj_weight_to_fp16_quantized, x = var_4539_cast_fp16)[name = string("linear_176_cast_fp16")];
            tensor<int32, [4]> concat_151x = const()[name = string("concat_151x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_4561_cast_fp16 = reshape(shape = concat_151x, x = linear_176_cast_fp16)[name = string("op_4561_cast_fp16")];
            tensor<int32, [4]> x_659_perm_0 = const()[name = string("x_659_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_25_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421293440))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422342080))))[name = string("layers_25_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_177_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_25_self_attn_v_proj_weight_to_fp16_quantized, x = var_4539_cast_fp16)[name = string("linear_177_cast_fp16")];
            tensor<int32, [4]> concat_152x = const()[name = string("concat_152x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_4566_cast_fp16 = reshape(shape = concat_152x, x = linear_177_cast_fp16)[name = string("op_4566_cast_fp16")];
            tensor<int32, [4]> v_51_perm_0 = const()[name = string("v_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4522_promoted_1_to_fp16 = const()[name = string("op_4522_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_655_cast_fp16 = transpose(perm = x_655_perm_0, x = var_4556_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 16, ?, 128]> var_4570_cast_fp16 = pow(x = x_655_cast_fp16, y = var_4522_promoted_1_to_fp16)[name = string("op_4570_cast_fp16")];
            tensor<int32, [1]> var_4572_axes_0 = const()[name = string("op_4572_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4572_keep_dims_0 = const()[name = string("op_4572_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_4572_cast_fp16 = reduce_mean(axes = var_4572_axes_0, keep_dims = var_4572_keep_dims_0, x = var_4570_cast_fp16)[name = string("op_4572_cast_fp16")];
            fp16 var_4573_to_fp16 = const()[name = string("op_4573_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_4574_cast_fp16 = add(x = var_4572_cast_fp16, y = var_4573_to_fp16)[name = string("op_4574_cast_fp16")];
            fp32 norm_203_epsilon_0 = const()[name = string("norm_203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_203_cast_fp16 = rsqrt(epsilon = norm_203_epsilon_0, x = var_4574_cast_fp16)[name = string("norm_203_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_4576_cast_fp16 = mul(x = x_655_cast_fp16, y = norm_203_cast_fp16)[name = string("op_4576_cast_fp16")];
            tensor<fp16, [128]> layers_25_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_25_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422407680)))];
            tensor<fp16, [1, 16, ?, 128]> var_4577_cast_fp16 = mul(x = var_4576_cast_fp16, y = layers_25_self_attn_q_norm_weight_to_fp16)[name = string("op_4577_cast_fp16")];
            fp16 var_4522_promoted_2_to_fp16 = const()[name = string("op_4522_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_659_cast_fp16 = transpose(perm = x_659_perm_0, x = var_4561_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 8, ?, 128]> var_4581_cast_fp16 = pow(x = x_659_cast_fp16, y = var_4522_promoted_2_to_fp16)[name = string("op_4581_cast_fp16")];
            tensor<int32, [1]> var_4583_axes_0 = const()[name = string("op_4583_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4583_keep_dims_0 = const()[name = string("op_4583_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_4583_cast_fp16 = reduce_mean(axes = var_4583_axes_0, keep_dims = var_4583_keep_dims_0, x = var_4581_cast_fp16)[name = string("op_4583_cast_fp16")];
            fp16 var_4584_to_fp16 = const()[name = string("op_4584_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_4585_cast_fp16 = add(x = var_4583_cast_fp16, y = var_4584_to_fp16)[name = string("op_4585_cast_fp16")];
            fp32 norm_205_epsilon_0 = const()[name = string("norm_205_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_205_cast_fp16 = rsqrt(epsilon = norm_205_epsilon_0, x = var_4585_cast_fp16)[name = string("norm_205_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_4587_cast_fp16 = mul(x = x_659_cast_fp16, y = norm_205_cast_fp16)[name = string("op_4587_cast_fp16")];
            tensor<fp16, [128]> layers_25_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_25_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422408000)))];
            tensor<fp16, [1, 8, ?, 128]> var_4588_cast_fp16 = mul(x = var_4587_cast_fp16, y = layers_25_self_attn_k_norm_weight_to_fp16)[name = string("op_4588_cast_fp16")];
            tensor<int32, [4]> x1_101_begin_0 = const()[name = string("x1_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_101_end_0 = const()[name = string("x1_101_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_101_end_mask_0 = const()[name = string("x1_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_101_cast_fp16 = slice_by_index(begin = x1_101_begin_0, end = x1_101_end_0, end_mask = x1_101_end_mask_0, x = var_4577_cast_fp16)[name = string("x1_101_cast_fp16")];
            tensor<int32, [4]> x2_101_begin_0 = const()[name = string("x2_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_101_end_0 = const()[name = string("x2_101_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_101_end_mask_0 = const()[name = string("x2_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_101_cast_fp16 = slice_by_index(begin = x2_101_begin_0, end = x2_101_end_0, end_mask = x2_101_end_mask_0, x = var_4577_cast_fp16)[name = string("x2_101_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4604_cast_fp16 = mul(x = x1_101_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4604_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4605_cast_fp16 = mul(x = x2_101_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4605_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4606_cast_fp16 = sub(x = var_4604_cast_fp16, y = var_4605_cast_fp16)[name = string("op_4606_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4607_cast_fp16 = mul(x = x2_101_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4607_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4608_cast_fp16 = mul(x = x1_101_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4608_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4609_cast_fp16 = add(x = var_4607_cast_fp16, y = var_4608_cast_fp16)[name = string("op_4609_cast_fp16")];
            bool q_51_interleave_0 = const()[name = string("q_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_51_cast_fp16 = concat(axis = var_4523, interleave = q_51_interleave_0, values = (var_4606_cast_fp16, var_4609_cast_fp16))[name = string("q_51_cast_fp16")];
            tensor<int32, [4]> x1_103_begin_0 = const()[name = string("x1_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_103_end_0 = const()[name = string("x1_103_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_103_end_mask_0 = const()[name = string("x1_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_103_cast_fp16 = slice_by_index(begin = x1_103_begin_0, end = x1_103_end_0, end_mask = x1_103_end_mask_0, x = var_4588_cast_fp16)[name = string("x1_103_cast_fp16")];
            tensor<int32, [4]> x2_103_begin_0 = const()[name = string("x2_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_103_end_0 = const()[name = string("x2_103_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_103_end_mask_0 = const()[name = string("x2_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_103_cast_fp16 = slice_by_index(begin = x2_103_begin_0, end = x2_103_end_0, end_mask = x2_103_end_mask_0, x = var_4588_cast_fp16)[name = string("x2_103_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4626_cast_fp16 = mul(x = x1_103_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4626_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4627_cast_fp16 = mul(x = x2_103_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4627_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4628_cast_fp16 = sub(x = var_4626_cast_fp16, y = var_4627_cast_fp16)[name = string("op_4628_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4629_cast_fp16 = mul(x = x2_103_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4629_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4630_cast_fp16 = mul(x = x1_103_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4630_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4631_cast_fp16 = add(x = var_4629_cast_fp16, y = var_4630_cast_fp16)[name = string("op_4631_cast_fp16")];
            bool k_51_interleave_0 = const()[name = string("k_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_51_cast_fp16 = concat(axis = var_4523, interleave = k_51_interleave_0, values = (var_4628_cast_fp16, var_4631_cast_fp16))[name = string("k_51_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_50 = read_state(input = k_cache_25)[name = string("read_state_50")];
            int32 k_cache_103_axis_0 = const()[name = string("k_cache_103_axis_0"), val = int32(2)];
            string k_cache_103_mode_0 = const()[name = string("k_cache_103_mode_0"), val = string("update")];
            bool k_cache_103_validate_indices_0 = const()[name = string("k_cache_103_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_103_cast_fp16 = scatter_along_axis(axis = k_cache_103_axis_0, data = read_state_50, indices = scatter_indices, mode = k_cache_103_mode_0, updates = k_51_cast_fp16, validate_indices = k_cache_103_validate_indices_0)[name = string("k_cache_103_cast_fp16")];
            write_state(data = k_cache_103_cast_fp16, input = k_cache_25)[name = string("coreml_update_state_50_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_106 = read_state(input = k_cache_25)[name = string("coreml_update_state_50")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_51 = read_state(input = v_cache_25)[name = string("read_state_51")];
            int32 v_cache_103_axis_0 = const()[name = string("v_cache_103_axis_0"), val = int32(2)];
            string v_cache_103_mode_0 = const()[name = string("v_cache_103_mode_0"), val = string("update")];
            bool v_cache_103_validate_indices_0 = const()[name = string("v_cache_103_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = var_4566_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_103_cast_fp16 = scatter_along_axis(axis = v_cache_103_axis_0, data = read_state_51, indices = scatter_indices, mode = v_cache_103_mode_0, updates = v_51_cast_fp16, validate_indices = v_cache_103_validate_indices_0)[name = string("v_cache_103_cast_fp16")];
            write_state(data = v_cache_103_cast_fp16, input = v_cache_25)[name = string("coreml_update_state_51_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_107 = read_state(input = v_cache_25)[name = string("coreml_update_state_51")];
            tensor<int32, [1]> var_4636_axes_0 = const()[name = string("op_4636_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4636_cast_fp16 = expand_dims(axes = var_4636_axes_0, x = coreml_update_state_106)[name = string("op_4636_cast_fp16")];
            tensor<int32, [5]> k_exp_101_reps_0 = const()[name = string("k_exp_101_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_101_cast_fp16 = tile(reps = k_exp_101_reps_0, x = var_4636_cast_fp16)[name = string("k_exp_101_cast_fp16")];
            tensor<int32, [4]> var_4639 = const()[name = string("op_4639"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_103_cast_fp16 = reshape(shape = var_4639, x = k_exp_101_cast_fp16)[name = string("k_exp_103_cast_fp16")];
            tensor<int32, [1]> var_4641_axes_0 = const()[name = string("op_4641_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4641_cast_fp16 = expand_dims(axes = var_4641_axes_0, x = coreml_update_state_107)[name = string("op_4641_cast_fp16")];
            tensor<int32, [5]> v_exp_101_reps_0 = const()[name = string("v_exp_101_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_101_cast_fp16 = tile(reps = v_exp_101_reps_0, x = var_4641_cast_fp16)[name = string("v_exp_101_cast_fp16")];
            tensor<int32, [4]> var_4644 = const()[name = string("op_4644"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_103_cast_fp16 = reshape(shape = var_4644, x = v_exp_101_cast_fp16)[name = string("v_exp_103_cast_fp16")];
            bool var_4647_transpose_x_1 = const()[name = string("op_4647_transpose_x_1"), val = bool(false)];
            bool var_4647_transpose_y_1 = const()[name = string("op_4647_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_4647_cast_fp16 = matmul(transpose_x = var_4647_transpose_x_1, transpose_y = var_4647_transpose_y_1, x = q_51_cast_fp16, y = k_exp_103_cast_fp16)[name = string("op_4647_cast_fp16")];
            fp16 var_4648_to_fp16 = const()[name = string("op_4648_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_101_cast_fp16 = mul(x = var_4647_cast_fp16, y = var_4648_to_fp16)[name = string("attn_101_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_251_cast_fp16 = add(x = attn_101_cast_fp16, y = attention_mask)[name = string("input_251_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_103_cast_fp16 = softmax(axis = var_4523, x = input_251_cast_fp16)[name = string("attn_103_cast_fp16")];
            bool out_51_transpose_x_0 = const()[name = string("out_51_transpose_x_0"), val = bool(false)];
            bool out_51_transpose_y_0 = const()[name = string("out_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_51_cast_fp16 = matmul(transpose_x = out_51_transpose_x_0, transpose_y = out_51_transpose_y_0, x = attn_103_cast_fp16, y = v_exp_103_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<int32, [4]> var_4653_perm_0 = const()[name = string("op_4653_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_155_axis_0 = const()[name = string("concat_155_axis_0"), val = int32(0)];
            bool concat_155_interleave_0 = const()[name = string("concat_155_interleave_0"), val = bool(false)];
            int32 gather_151_cast_uint16_to_int32 = cast(dtype = gather_151_cast_uint16_to_int32_dtype_0, x = gather_151_cast_uint16)[name = string("cast_4")];
            tensor<int32, [3]> concat_155 = concat(axis = concat_155_axis_0, interleave = concat_155_interleave_0, values = (gather_150, gather_151_cast_uint16_to_int32, var_4523))[name = string("concat_155")];
            tensor<fp16, [1, ?, 16, 128]> var_4653_cast_fp16 = transpose(perm = var_4653_perm_0, x = out_51_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, ?, ?]> input_253_cast_fp16 = reshape(shape = concat_155, x = var_4653_cast_fp16)[name = string("input_253_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_25_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422408320))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424505536))))[name = string("layers_25_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_178_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_25_self_attn_o_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_178_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_669_cast_fp16 = add(x = x_649_cast_fp16, y = linear_178_cast_fp16)[name = string("x_669_cast_fp16")];
            fp16 var_4522_promoted_3_to_fp16 = const()[name = string("op_4522_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_4661_cast_fp16 = pow(x = x_669_cast_fp16, y = var_4522_promoted_3_to_fp16)[name = string("op_4661_cast_fp16")];
            tensor<int32, [1]> var_4663_axes_0 = const()[name = string("op_4663_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4663_keep_dims_0 = const()[name = string("op_4663_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_4663_cast_fp16 = reduce_mean(axes = var_4663_axes_0, keep_dims = var_4663_keep_dims_0, x = var_4661_cast_fp16)[name = string("op_4663_cast_fp16")];
            fp16 var_4664_to_fp16 = const()[name = string("op_4664_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4665_cast_fp16 = add(x = var_4663_cast_fp16, y = var_4664_to_fp16)[name = string("op_4665_cast_fp16")];
            fp32 norm_207_epsilon_0 = const()[name = string("norm_207_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_207_cast_fp16 = rsqrt(epsilon = norm_207_epsilon_0, x = var_4665_cast_fp16)[name = string("norm_207_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_4667_cast_fp16 = mul(x = x_669_cast_fp16, y = norm_207_cast_fp16)[name = string("op_4667_cast_fp16")];
            tensor<fp16, [1024]> layers_25_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_25_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424636672)))];
            tensor<fp16, [1, ?, 1024]> var_4668_cast_fp16 = mul(x = var_4667_cast_fp16, y = layers_25_post_attention_layernorm_weight_to_fp16)[name = string("op_4668_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_25_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424638784))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427784576))))[name = string("layers_25_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_179_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_25_mlp_gate_proj_weight_to_fp16_quantized, x = var_4668_cast_fp16)[name = string("linear_179_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_4678_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("op_4678_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_25_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427981248))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431127040))))[name = string("layers_25_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_180_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_25_mlp_up_proj_weight_to_fp16_quantized, x = var_4668_cast_fp16)[name = string("linear_180_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_259_cast_fp16 = mul(x = var_4678_cast_fp16, y = linear_180_cast_fp16)[name = string("input_259_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_25_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431323712))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434469504))))[name = string("layers_25_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_181_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_25_mlp_down_proj_weight_to_fp16_quantized, x = input_259_cast_fp16)[name = string("linear_181_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_675_cast_fp16 = add(x = x_669_cast_fp16, y = linear_181_cast_fp16)[name = string("x_675_cast_fp16")];
            int32 var_4698 = const()[name = string("op_4698"), val = int32(-1)];
            fp16 var_4697_promoted_to_fp16 = const()[name = string("op_4697_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_4707_cast_fp16 = pow(x = x_675_cast_fp16, y = var_4697_promoted_to_fp16)[name = string("op_4707_cast_fp16")];
            tensor<int32, [1]> var_4709_axes_0 = const()[name = string("op_4709_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4709_keep_dims_0 = const()[name = string("op_4709_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_4709_cast_fp16 = reduce_mean(axes = var_4709_axes_0, keep_dims = var_4709_keep_dims_0, x = var_4707_cast_fp16)[name = string("op_4709_cast_fp16")];
            fp16 var_4710_to_fp16 = const()[name = string("op_4710_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4711_cast_fp16 = add(x = var_4709_cast_fp16, y = var_4710_to_fp16)[name = string("op_4711_cast_fp16")];
            fp32 norm_209_epsilon_0 = const()[name = string("norm_209_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_209_cast_fp16 = rsqrt(epsilon = norm_209_epsilon_0, x = var_4711_cast_fp16)[name = string("norm_209_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_4713_cast_fp16 = mul(x = x_675_cast_fp16, y = norm_209_cast_fp16)[name = string("op_4713_cast_fp16")];
            tensor<fp16, [1024]> layers_26_input_layernorm_weight_to_fp16 = const()[name = string("layers_26_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434666176)))];
            tensor<fp16, [1, ?, 1024]> var_4714_cast_fp16 = mul(x = var_4713_cast_fp16, y = layers_26_input_layernorm_weight_to_fp16)[name = string("op_4714_cast_fp16")];
            tensor<int32, [3]> var_4726_shape_cast_fp16 = shape(x = var_4714_cast_fp16)[name = string("op_4726_shape_cast_fp16")];
            int32 gather_156 = const()[name = string("gather_156"), val = int32(1)];
            int32 gather_157_axis_0 = const()[name = string("gather_157_axis_0"), val = int32(0)];
            int32 gather_157_batch_dims_0 = const()[name = string("gather_157_batch_dims_0"), val = int32(0)];
            bool gather_157_validate_indices_0 = const()[name = string("gather_157_validate_indices_0"), val = bool(false)];
            string var_4726_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4726_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_157_to_uint16 = const()[name = string("select_157_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_4726_shape_cast_fp16_to_uint16 = cast(dtype = var_4726_shape_cast_fp16_to_uint16_dtype_0, x = var_4726_shape_cast_fp16)[name = string("cast_3")];
            uint16 gather_157_cast_uint16 = gather(axis = gather_157_axis_0, batch_dims = gather_157_batch_dims_0, indices = select_157_to_uint16, validate_indices = gather_157_validate_indices_0, x = var_4726_shape_cast_fp16_to_uint16)[name = string("gather_157_cast_uint16")];
            string gather_157_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_157_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_26_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434668288))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436765504))))[name = string("layers_26_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_182_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_26_self_attn_q_proj_weight_to_fp16_quantized, x = var_4714_cast_fp16)[name = string("linear_182_cast_fp16")];
            tensor<int32, [4]> concat_156x = const()[name = string("concat_156x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_4731_cast_fp16 = reshape(shape = concat_156x, x = linear_182_cast_fp16)[name = string("op_4731_cast_fp16")];
            tensor<int32, [4]> x_681_perm_0 = const()[name = string("x_681_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_26_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436896640))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437945280))))[name = string("layers_26_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_183_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_26_self_attn_k_proj_weight_to_fp16_quantized, x = var_4714_cast_fp16)[name = string("linear_183_cast_fp16")];
            tensor<int32, [4]> concat_157x = const()[name = string("concat_157x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_4736_cast_fp16 = reshape(shape = concat_157x, x = linear_183_cast_fp16)[name = string("op_4736_cast_fp16")];
            tensor<int32, [4]> x_685_perm_0 = const()[name = string("x_685_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_26_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(438010880))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439059520))))[name = string("layers_26_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_184_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_26_self_attn_v_proj_weight_to_fp16_quantized, x = var_4714_cast_fp16)[name = string("linear_184_cast_fp16")];
            tensor<int32, [4]> concat_158x = const()[name = string("concat_158x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_4741_cast_fp16 = reshape(shape = concat_158x, x = linear_184_cast_fp16)[name = string("op_4741_cast_fp16")];
            tensor<int32, [4]> v_53_perm_0 = const()[name = string("v_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4697_promoted_1_to_fp16 = const()[name = string("op_4697_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_681_cast_fp16 = transpose(perm = x_681_perm_0, x = var_4731_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 16, ?, 128]> var_4745_cast_fp16 = pow(x = x_681_cast_fp16, y = var_4697_promoted_1_to_fp16)[name = string("op_4745_cast_fp16")];
            tensor<int32, [1]> var_4747_axes_0 = const()[name = string("op_4747_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4747_keep_dims_0 = const()[name = string("op_4747_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_4747_cast_fp16 = reduce_mean(axes = var_4747_axes_0, keep_dims = var_4747_keep_dims_0, x = var_4745_cast_fp16)[name = string("op_4747_cast_fp16")];
            fp16 var_4748_to_fp16 = const()[name = string("op_4748_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_4749_cast_fp16 = add(x = var_4747_cast_fp16, y = var_4748_to_fp16)[name = string("op_4749_cast_fp16")];
            fp32 norm_211_epsilon_0 = const()[name = string("norm_211_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_211_cast_fp16 = rsqrt(epsilon = norm_211_epsilon_0, x = var_4749_cast_fp16)[name = string("norm_211_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_4751_cast_fp16 = mul(x = x_681_cast_fp16, y = norm_211_cast_fp16)[name = string("op_4751_cast_fp16")];
            tensor<fp16, [128]> layers_26_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_26_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439125120)))];
            tensor<fp16, [1, 16, ?, 128]> var_4752_cast_fp16 = mul(x = var_4751_cast_fp16, y = layers_26_self_attn_q_norm_weight_to_fp16)[name = string("op_4752_cast_fp16")];
            fp16 var_4697_promoted_2_to_fp16 = const()[name = string("op_4697_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_685_cast_fp16 = transpose(perm = x_685_perm_0, x = var_4736_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 8, ?, 128]> var_4756_cast_fp16 = pow(x = x_685_cast_fp16, y = var_4697_promoted_2_to_fp16)[name = string("op_4756_cast_fp16")];
            tensor<int32, [1]> var_4758_axes_0 = const()[name = string("op_4758_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4758_keep_dims_0 = const()[name = string("op_4758_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_4758_cast_fp16 = reduce_mean(axes = var_4758_axes_0, keep_dims = var_4758_keep_dims_0, x = var_4756_cast_fp16)[name = string("op_4758_cast_fp16")];
            fp16 var_4759_to_fp16 = const()[name = string("op_4759_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_4760_cast_fp16 = add(x = var_4758_cast_fp16, y = var_4759_to_fp16)[name = string("op_4760_cast_fp16")];
            fp32 norm_213_epsilon_0 = const()[name = string("norm_213_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_213_cast_fp16 = rsqrt(epsilon = norm_213_epsilon_0, x = var_4760_cast_fp16)[name = string("norm_213_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_4762_cast_fp16 = mul(x = x_685_cast_fp16, y = norm_213_cast_fp16)[name = string("op_4762_cast_fp16")];
            tensor<fp16, [128]> layers_26_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_26_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439125440)))];
            tensor<fp16, [1, 8, ?, 128]> var_4763_cast_fp16 = mul(x = var_4762_cast_fp16, y = layers_26_self_attn_k_norm_weight_to_fp16)[name = string("op_4763_cast_fp16")];
            tensor<int32, [4]> x1_105_begin_0 = const()[name = string("x1_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_105_end_0 = const()[name = string("x1_105_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_105_end_mask_0 = const()[name = string("x1_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_105_cast_fp16 = slice_by_index(begin = x1_105_begin_0, end = x1_105_end_0, end_mask = x1_105_end_mask_0, x = var_4752_cast_fp16)[name = string("x1_105_cast_fp16")];
            tensor<int32, [4]> x2_105_begin_0 = const()[name = string("x2_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_105_end_0 = const()[name = string("x2_105_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_105_end_mask_0 = const()[name = string("x2_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_105_cast_fp16 = slice_by_index(begin = x2_105_begin_0, end = x2_105_end_0, end_mask = x2_105_end_mask_0, x = var_4752_cast_fp16)[name = string("x2_105_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4779_cast_fp16 = mul(x = x1_105_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4779_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4780_cast_fp16 = mul(x = x2_105_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4780_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4781_cast_fp16 = sub(x = var_4779_cast_fp16, y = var_4780_cast_fp16)[name = string("op_4781_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4782_cast_fp16 = mul(x = x2_105_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4782_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4783_cast_fp16 = mul(x = x1_105_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4784_cast_fp16 = add(x = var_4782_cast_fp16, y = var_4783_cast_fp16)[name = string("op_4784_cast_fp16")];
            bool q_53_interleave_0 = const()[name = string("q_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_53_cast_fp16 = concat(axis = var_4698, interleave = q_53_interleave_0, values = (var_4781_cast_fp16, var_4784_cast_fp16))[name = string("q_53_cast_fp16")];
            tensor<int32, [4]> x1_107_begin_0 = const()[name = string("x1_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_107_end_0 = const()[name = string("x1_107_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_107_end_mask_0 = const()[name = string("x1_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_107_cast_fp16 = slice_by_index(begin = x1_107_begin_0, end = x1_107_end_0, end_mask = x1_107_end_mask_0, x = var_4763_cast_fp16)[name = string("x1_107_cast_fp16")];
            tensor<int32, [4]> x2_107_begin_0 = const()[name = string("x2_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_107_end_0 = const()[name = string("x2_107_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_107_end_mask_0 = const()[name = string("x2_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_107_cast_fp16 = slice_by_index(begin = x2_107_begin_0, end = x2_107_end_0, end_mask = x2_107_end_mask_0, x = var_4763_cast_fp16)[name = string("x2_107_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4801_cast_fp16 = mul(x = x1_107_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4801_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4802_cast_fp16 = mul(x = x2_107_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4802_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4803_cast_fp16 = sub(x = var_4801_cast_fp16, y = var_4802_cast_fp16)[name = string("op_4803_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4804_cast_fp16 = mul(x = x2_107_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4804_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4805_cast_fp16 = mul(x = x1_107_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4805_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4806_cast_fp16 = add(x = var_4804_cast_fp16, y = var_4805_cast_fp16)[name = string("op_4806_cast_fp16")];
            bool k_53_interleave_0 = const()[name = string("k_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_53_cast_fp16 = concat(axis = var_4698, interleave = k_53_interleave_0, values = (var_4803_cast_fp16, var_4806_cast_fp16))[name = string("k_53_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_52 = read_state(input = k_cache_26)[name = string("read_state_52")];
            int32 k_cache_107_axis_0 = const()[name = string("k_cache_107_axis_0"), val = int32(2)];
            string k_cache_107_mode_0 = const()[name = string("k_cache_107_mode_0"), val = string("update")];
            bool k_cache_107_validate_indices_0 = const()[name = string("k_cache_107_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_107_cast_fp16 = scatter_along_axis(axis = k_cache_107_axis_0, data = read_state_52, indices = scatter_indices, mode = k_cache_107_mode_0, updates = k_53_cast_fp16, validate_indices = k_cache_107_validate_indices_0)[name = string("k_cache_107_cast_fp16")];
            write_state(data = k_cache_107_cast_fp16, input = k_cache_26)[name = string("coreml_update_state_52_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_108 = read_state(input = k_cache_26)[name = string("coreml_update_state_52")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_53 = read_state(input = v_cache_26)[name = string("read_state_53")];
            int32 v_cache_107_axis_0 = const()[name = string("v_cache_107_axis_0"), val = int32(2)];
            string v_cache_107_mode_0 = const()[name = string("v_cache_107_mode_0"), val = string("update")];
            bool v_cache_107_validate_indices_0 = const()[name = string("v_cache_107_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_53_cast_fp16 = transpose(perm = v_53_perm_0, x = var_4741_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_107_cast_fp16 = scatter_along_axis(axis = v_cache_107_axis_0, data = read_state_53, indices = scatter_indices, mode = v_cache_107_mode_0, updates = v_53_cast_fp16, validate_indices = v_cache_107_validate_indices_0)[name = string("v_cache_107_cast_fp16")];
            write_state(data = v_cache_107_cast_fp16, input = v_cache_26)[name = string("coreml_update_state_53_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_109 = read_state(input = v_cache_26)[name = string("coreml_update_state_53")];
            tensor<int32, [1]> var_4811_axes_0 = const()[name = string("op_4811_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4811_cast_fp16 = expand_dims(axes = var_4811_axes_0, x = coreml_update_state_108)[name = string("op_4811_cast_fp16")];
            tensor<int32, [5]> k_exp_105_reps_0 = const()[name = string("k_exp_105_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_105_cast_fp16 = tile(reps = k_exp_105_reps_0, x = var_4811_cast_fp16)[name = string("k_exp_105_cast_fp16")];
            tensor<int32, [4]> var_4814 = const()[name = string("op_4814"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_107_cast_fp16 = reshape(shape = var_4814, x = k_exp_105_cast_fp16)[name = string("k_exp_107_cast_fp16")];
            tensor<int32, [1]> var_4816_axes_0 = const()[name = string("op_4816_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4816_cast_fp16 = expand_dims(axes = var_4816_axes_0, x = coreml_update_state_109)[name = string("op_4816_cast_fp16")];
            tensor<int32, [5]> v_exp_105_reps_0 = const()[name = string("v_exp_105_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_105_cast_fp16 = tile(reps = v_exp_105_reps_0, x = var_4816_cast_fp16)[name = string("v_exp_105_cast_fp16")];
            tensor<int32, [4]> var_4819 = const()[name = string("op_4819"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_107_cast_fp16 = reshape(shape = var_4819, x = v_exp_105_cast_fp16)[name = string("v_exp_107_cast_fp16")];
            bool var_4822_transpose_x_1 = const()[name = string("op_4822_transpose_x_1"), val = bool(false)];
            bool var_4822_transpose_y_1 = const()[name = string("op_4822_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_4822_cast_fp16 = matmul(transpose_x = var_4822_transpose_x_1, transpose_y = var_4822_transpose_y_1, x = q_53_cast_fp16, y = k_exp_107_cast_fp16)[name = string("op_4822_cast_fp16")];
            fp16 var_4823_to_fp16 = const()[name = string("op_4823_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_105_cast_fp16 = mul(x = var_4822_cast_fp16, y = var_4823_to_fp16)[name = string("attn_105_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_261_cast_fp16 = add(x = attn_105_cast_fp16, y = attention_mask)[name = string("input_261_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_107_cast_fp16 = softmax(axis = var_4698, x = input_261_cast_fp16)[name = string("attn_107_cast_fp16")];
            bool out_53_transpose_x_0 = const()[name = string("out_53_transpose_x_0"), val = bool(false)];
            bool out_53_transpose_y_0 = const()[name = string("out_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_53_cast_fp16 = matmul(transpose_x = out_53_transpose_x_0, transpose_y = out_53_transpose_y_0, x = attn_107_cast_fp16, y = v_exp_107_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<int32, [4]> var_4828_perm_0 = const()[name = string("op_4828_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_161_axis_0 = const()[name = string("concat_161_axis_0"), val = int32(0)];
            bool concat_161_interleave_0 = const()[name = string("concat_161_interleave_0"), val = bool(false)];
            int32 gather_157_cast_uint16_to_int32 = cast(dtype = gather_157_cast_uint16_to_int32_dtype_0, x = gather_157_cast_uint16)[name = string("cast_2")];
            tensor<int32, [3]> concat_161 = concat(axis = concat_161_axis_0, interleave = concat_161_interleave_0, values = (gather_156, gather_157_cast_uint16_to_int32, var_4698))[name = string("concat_161")];
            tensor<fp16, [1, ?, 16, 128]> var_4828_cast_fp16 = transpose(perm = var_4828_perm_0, x = out_53_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, ?, ?]> input_263_cast_fp16 = reshape(shape = concat_161, x = var_4828_cast_fp16)[name = string("input_263_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_26_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439125760))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441222976))))[name = string("layers_26_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_185_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_26_self_attn_o_proj_weight_to_fp16_quantized, x = input_263_cast_fp16)[name = string("linear_185_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_695_cast_fp16 = add(x = x_675_cast_fp16, y = linear_185_cast_fp16)[name = string("x_695_cast_fp16")];
            fp16 var_4697_promoted_3_to_fp16 = const()[name = string("op_4697_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_4836_cast_fp16 = pow(x = x_695_cast_fp16, y = var_4697_promoted_3_to_fp16)[name = string("op_4836_cast_fp16")];
            tensor<int32, [1]> var_4838_axes_0 = const()[name = string("op_4838_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4838_keep_dims_0 = const()[name = string("op_4838_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_4838_cast_fp16 = reduce_mean(axes = var_4838_axes_0, keep_dims = var_4838_keep_dims_0, x = var_4836_cast_fp16)[name = string("op_4838_cast_fp16")];
            fp16 var_4839_to_fp16 = const()[name = string("op_4839_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4840_cast_fp16 = add(x = var_4838_cast_fp16, y = var_4839_to_fp16)[name = string("op_4840_cast_fp16")];
            fp32 norm_215_epsilon_0 = const()[name = string("norm_215_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_215_cast_fp16 = rsqrt(epsilon = norm_215_epsilon_0, x = var_4840_cast_fp16)[name = string("norm_215_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_4842_cast_fp16 = mul(x = x_695_cast_fp16, y = norm_215_cast_fp16)[name = string("op_4842_cast_fp16")];
            tensor<fp16, [1024]> layers_26_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_26_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441354112)))];
            tensor<fp16, [1, ?, 1024]> var_4843_cast_fp16 = mul(x = var_4842_cast_fp16, y = layers_26_post_attention_layernorm_weight_to_fp16)[name = string("op_4843_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_26_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441356224))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444502016))))[name = string("layers_26_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_186_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_26_mlp_gate_proj_weight_to_fp16_quantized, x = var_4843_cast_fp16)[name = string("linear_186_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_4853_cast_fp16 = silu(x = linear_186_cast_fp16)[name = string("op_4853_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_26_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444698688))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447844480))))[name = string("layers_26_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_187_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_26_mlp_up_proj_weight_to_fp16_quantized, x = var_4843_cast_fp16)[name = string("linear_187_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_269_cast_fp16 = mul(x = var_4853_cast_fp16, y = linear_187_cast_fp16)[name = string("input_269_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_26_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448041152))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451186944))))[name = string("layers_26_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_188_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_26_mlp_down_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_188_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_701_cast_fp16 = add(x = x_695_cast_fp16, y = linear_188_cast_fp16)[name = string("x_701_cast_fp16")];
            int32 var_4873 = const()[name = string("op_4873"), val = int32(-1)];
            fp16 var_4872_promoted_to_fp16 = const()[name = string("op_4872_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_4882_cast_fp16 = pow(x = x_701_cast_fp16, y = var_4872_promoted_to_fp16)[name = string("op_4882_cast_fp16")];
            tensor<int32, [1]> var_4884_axes_0 = const()[name = string("op_4884_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4884_keep_dims_0 = const()[name = string("op_4884_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_4884_cast_fp16 = reduce_mean(axes = var_4884_axes_0, keep_dims = var_4884_keep_dims_0, x = var_4882_cast_fp16)[name = string("op_4884_cast_fp16")];
            fp16 var_4885_to_fp16 = const()[name = string("op_4885_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_4886_cast_fp16 = add(x = var_4884_cast_fp16, y = var_4885_to_fp16)[name = string("op_4886_cast_fp16")];
            fp32 norm_217_epsilon_0 = const()[name = string("norm_217_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_217_cast_fp16 = rsqrt(epsilon = norm_217_epsilon_0, x = var_4886_cast_fp16)[name = string("norm_217_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_4888_cast_fp16 = mul(x = x_701_cast_fp16, y = norm_217_cast_fp16)[name = string("op_4888_cast_fp16")];
            tensor<fp16, [1024]> layers_27_input_layernorm_weight_to_fp16 = const()[name = string("layers_27_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451383616)))];
            tensor<fp16, [1, ?, 1024]> var_4889_cast_fp16 = mul(x = var_4888_cast_fp16, y = layers_27_input_layernorm_weight_to_fp16)[name = string("op_4889_cast_fp16")];
            tensor<int32, [3]> var_4901_shape_cast_fp16 = shape(x = var_4889_cast_fp16)[name = string("op_4901_shape_cast_fp16")];
            int32 gather_162 = const()[name = string("gather_162"), val = int32(1)];
            int32 gather_163_axis_0 = const()[name = string("gather_163_axis_0"), val = int32(0)];
            int32 gather_163_batch_dims_0 = const()[name = string("gather_163_batch_dims_0"), val = int32(0)];
            bool gather_163_validate_indices_0 = const()[name = string("gather_163_validate_indices_0"), val = bool(false)];
            string var_4901_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_4901_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
            uint16 select_163_to_uint16 = const()[name = string("select_163_to_uint16"), val = uint16(1)];
            tensor<uint16, [3]> var_4901_shape_cast_fp16_to_uint16 = cast(dtype = var_4901_shape_cast_fp16_to_uint16_dtype_0, x = var_4901_shape_cast_fp16)[name = string("cast_1")];
            uint16 gather_163_cast_uint16 = gather(axis = gather_163_axis_0, batch_dims = gather_163_batch_dims_0, indices = select_163_to_uint16, validate_indices = gather_163_validate_indices_0, x = var_4901_shape_cast_fp16_to_uint16)[name = string("gather_163_cast_uint16")];
            string gather_163_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_163_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<fp16, [2048, 1024]> layers_27_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451385728))), scale = tensor<fp16, [2048, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453482944))))[name = string("layers_27_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 2048]> linear_189_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_27_self_attn_q_proj_weight_to_fp16_quantized, x = var_4889_cast_fp16)[name = string("linear_189_cast_fp16")];
            tensor<int32, [4]> concat_162x = const()[name = string("concat_162x"), val = tensor<int32, [4]>([1, -1, 16, 128])];
            tensor<fp16, [1, ?, 16, 128]> var_4906_cast_fp16 = reshape(shape = concat_162x, x = linear_189_cast_fp16)[name = string("op_4906_cast_fp16")];
            tensor<int32, [4]> x_707_perm_0 = const()[name = string("x_707_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_27_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453614080))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454662720))))[name = string("layers_27_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_190_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_27_self_attn_k_proj_weight_to_fp16_quantized, x = var_4889_cast_fp16)[name = string("linear_190_cast_fp16")];
            tensor<int32, [4]> concat_163x = const()[name = string("concat_163x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_4911_cast_fp16 = reshape(shape = concat_163x, x = linear_190_cast_fp16)[name = string("op_4911_cast_fp16")];
            tensor<int32, [4]> x_711_perm_0 = const()[name = string("x_711_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_27_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454728320))), scale = tensor<fp16, [1024, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455776960))))[name = string("layers_27_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_191_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_27_self_attn_v_proj_weight_to_fp16_quantized, x = var_4889_cast_fp16)[name = string("linear_191_cast_fp16")];
            tensor<int32, [4]> concat_164x = const()[name = string("concat_164x"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, ?, 8, 128]> var_4916_cast_fp16 = reshape(shape = concat_164x, x = linear_191_cast_fp16)[name = string("op_4916_cast_fp16")];
            tensor<int32, [4]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4872_promoted_1_to_fp16 = const()[name = string("op_4872_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, ?, 128]> x_707_cast_fp16 = transpose(perm = x_707_perm_0, x = var_4906_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 16, ?, 128]> var_4920_cast_fp16 = pow(x = x_707_cast_fp16, y = var_4872_promoted_1_to_fp16)[name = string("op_4920_cast_fp16")];
            tensor<int32, [1]> var_4922_axes_0 = const()[name = string("op_4922_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4922_keep_dims_0 = const()[name = string("op_4922_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1]> var_4922_cast_fp16 = reduce_mean(axes = var_4922_axes_0, keep_dims = var_4922_keep_dims_0, x = var_4920_cast_fp16)[name = string("op_4922_cast_fp16")];
            fp16 var_4923_to_fp16 = const()[name = string("op_4923_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, ?, 1]> var_4924_cast_fp16 = add(x = var_4922_cast_fp16, y = var_4923_to_fp16)[name = string("op_4924_cast_fp16")];
            fp32 norm_219_epsilon_0 = const()[name = string("norm_219_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, ?, 1]> norm_219_cast_fp16 = rsqrt(epsilon = norm_219_epsilon_0, x = var_4924_cast_fp16)[name = string("norm_219_cast_fp16")];
            tensor<fp16, [1, 16, ?, 128]> var_4926_cast_fp16 = mul(x = x_707_cast_fp16, y = norm_219_cast_fp16)[name = string("op_4926_cast_fp16")];
            tensor<fp16, [128]> layers_27_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_27_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455842560)))];
            tensor<fp16, [1, 16, ?, 128]> var_4927_cast_fp16 = mul(x = var_4926_cast_fp16, y = layers_27_self_attn_q_norm_weight_to_fp16)[name = string("op_4927_cast_fp16")];
            fp16 var_4872_promoted_2_to_fp16 = const()[name = string("op_4872_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, ?, 128]> x_711_cast_fp16 = transpose(perm = x_711_perm_0, x = var_4911_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 8, ?, 128]> var_4931_cast_fp16 = pow(x = x_711_cast_fp16, y = var_4872_promoted_2_to_fp16)[name = string("op_4931_cast_fp16")];
            tensor<int32, [1]> var_4933_axes_0 = const()[name = string("op_4933_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4933_keep_dims_0 = const()[name = string("op_4933_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, ?, 1]> var_4933_cast_fp16 = reduce_mean(axes = var_4933_axes_0, keep_dims = var_4933_keep_dims_0, x = var_4931_cast_fp16)[name = string("op_4933_cast_fp16")];
            fp16 var_4934_to_fp16 = const()[name = string("op_4934_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, ?, 1]> var_4935_cast_fp16 = add(x = var_4933_cast_fp16, y = var_4934_to_fp16)[name = string("op_4935_cast_fp16")];
            fp32 norm_221_epsilon_0 = const()[name = string("norm_221_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, ?, 1]> norm_221_cast_fp16 = rsqrt(epsilon = norm_221_epsilon_0, x = var_4935_cast_fp16)[name = string("norm_221_cast_fp16")];
            tensor<fp16, [1, 8, ?, 128]> var_4937_cast_fp16 = mul(x = x_711_cast_fp16, y = norm_221_cast_fp16)[name = string("op_4937_cast_fp16")];
            tensor<fp16, [128]> layers_27_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_27_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455842880)))];
            tensor<fp16, [1, 8, ?, 128]> var_4938_cast_fp16 = mul(x = var_4937_cast_fp16, y = layers_27_self_attn_k_norm_weight_to_fp16)[name = string("op_4938_cast_fp16")];
            tensor<int32, [4]> x1_109_begin_0 = const()[name = string("x1_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_109_end_0 = const()[name = string("x1_109_end_0"), val = tensor<int32, [4]>([1, 16, 0, 64])];
            tensor<bool, [4]> x1_109_end_mask_0 = const()[name = string("x1_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, ?, 64]> x1_109_cast_fp16 = slice_by_index(begin = x1_109_begin_0, end = x1_109_end_0, end_mask = x1_109_end_mask_0, x = var_4927_cast_fp16)[name = string("x1_109_cast_fp16")];
            tensor<int32, [4]> x2_109_begin_0 = const()[name = string("x2_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_109_end_0 = const()[name = string("x2_109_end_0"), val = tensor<int32, [4]>([1, 16, 0, 128])];
            tensor<bool, [4]> x2_109_end_mask_0 = const()[name = string("x2_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, ?, 64]> x2_109_cast_fp16 = slice_by_index(begin = x2_109_begin_0, end = x2_109_end_0, end_mask = x2_109_end_mask_0, x = var_4927_cast_fp16)[name = string("x2_109_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4954_cast_fp16 = mul(x = x1_109_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4954_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4955_cast_fp16 = mul(x = x2_109_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4955_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4956_cast_fp16 = sub(x = var_4954_cast_fp16, y = var_4955_cast_fp16)[name = string("op_4956_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4957_cast_fp16 = mul(x = x2_109_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4957_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4958_cast_fp16 = mul(x = x1_109_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4958_cast_fp16")];
            tensor<fp16, [1, 16, ?, 64]> var_4959_cast_fp16 = add(x = var_4957_cast_fp16, y = var_4958_cast_fp16)[name = string("op_4959_cast_fp16")];
            bool q_interleave_0 = const()[name = string("q_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> q_cast_fp16 = concat(axis = var_4873, interleave = q_interleave_0, values = (var_4956_cast_fp16, var_4959_cast_fp16))[name = string("q_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([1, 8, 0, 64])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, ?, 64]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = var_4938_cast_fp16)[name = string("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([1, 8, 0, 128])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, ?, 64]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = var_4938_cast_fp16)[name = string("x2_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4976_cast_fp16 = mul(x = x1_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4976_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4977_cast_fp16 = mul(x = x2_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4977_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4978_cast_fp16 = sub(x = var_4976_cast_fp16, y = var_4977_cast_fp16)[name = string("op_4978_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4979_cast_fp16 = mul(x = x2_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_4979_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4980_cast_fp16 = mul(x = x1_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_4980_cast_fp16")];
            tensor<fp16, [1, 8, ?, 64]> var_4981_cast_fp16 = add(x = var_4979_cast_fp16, y = var_4980_cast_fp16)[name = string("op_4981_cast_fp16")];
            bool k_interleave_0 = const()[name = string("k_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> k_cast_fp16 = concat(axis = var_4873, interleave = k_interleave_0, values = (var_4978_cast_fp16, var_4981_cast_fp16))[name = string("k_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_54 = read_state(input = k_cache_27)[name = string("read_state_54")];
            int32 k_cache_axis_0 = const()[name = string("k_cache_axis_0"), val = int32(2)];
            string k_cache_mode_0 = const()[name = string("k_cache_mode_0"), val = string("update")];
            bool k_cache_validate_indices_0 = const()[name = string("k_cache_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_cast_fp16 = scatter_along_axis(axis = k_cache_axis_0, data = read_state_54, indices = scatter_indices, mode = k_cache_mode_0, updates = k_cast_fp16, validate_indices = k_cache_validate_indices_0)[name = string("k_cache_cast_fp16")];
            write_state(data = k_cache_cast_fp16, input = k_cache_27)[name = string("coreml_update_state_54_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_110 = read_state(input = k_cache_27)[name = string("coreml_update_state_54")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_55 = read_state(input = v_cache_27)[name = string("read_state_55")];
            int32 v_cache_axis_0 = const()[name = string("v_cache_axis_0"), val = int32(2)];
            string v_cache_mode_0 = const()[name = string("v_cache_mode_0"), val = string("update")];
            bool v_cache_validate_indices_0 = const()[name = string("v_cache_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1, 8, ?, 128]> v_cast_fp16 = transpose(perm = v_perm_0, x = var_4916_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_cast_fp16 = scatter_along_axis(axis = v_cache_axis_0, data = read_state_55, indices = scatter_indices, mode = v_cache_mode_0, updates = v_cast_fp16, validate_indices = v_cache_validate_indices_0)[name = string("v_cache_cast_fp16")];
            write_state(data = v_cache_cast_fp16, input = v_cache_27)[name = string("coreml_update_state_55_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_111 = read_state(input = v_cache_27)[name = string("coreml_update_state_55")];
            tensor<int32, [1]> var_4986_axes_0 = const()[name = string("op_4986_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4986_cast_fp16 = expand_dims(axes = var_4986_axes_0, x = coreml_update_state_110)[name = string("op_4986_cast_fp16")];
            tensor<int32, [5]> k_exp_109_reps_0 = const()[name = string("k_exp_109_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_109_cast_fp16 = tile(reps = k_exp_109_reps_0, x = var_4986_cast_fp16)[name = string("k_exp_109_cast_fp16")];
            tensor<int32, [4]> var_4989 = const()[name = string("op_4989"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_cast_fp16 = reshape(shape = var_4989, x = k_exp_109_cast_fp16)[name = string("k_exp_cast_fp16")];
            tensor<int32, [1]> var_4991_axes_0 = const()[name = string("op_4991_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_4991_cast_fp16 = expand_dims(axes = var_4991_axes_0, x = coreml_update_state_111)[name = string("op_4991_cast_fp16")];
            tensor<int32, [5]> v_exp_109_reps_0 = const()[name = string("v_exp_109_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_109_cast_fp16 = tile(reps = v_exp_109_reps_0, x = var_4991_cast_fp16)[name = string("v_exp_109_cast_fp16")];
            tensor<int32, [4]> var_4994 = const()[name = string("op_4994"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_cast_fp16 = reshape(shape = var_4994, x = v_exp_109_cast_fp16)[name = string("v_exp_cast_fp16")];
            bool var_4997_transpose_x_1 = const()[name = string("op_4997_transpose_x_1"), val = bool(false)];
            bool var_4997_transpose_y_1 = const()[name = string("op_4997_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, ?, 1024]> var_4997_cast_fp16 = matmul(transpose_x = var_4997_transpose_x_1, transpose_y = var_4997_transpose_y_1, x = q_cast_fp16, y = k_exp_cast_fp16)[name = string("op_4997_cast_fp16")];
            fp16 var_4998_to_fp16 = const()[name = string("op_4998_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, ?, 1024]> attn_109_cast_fp16 = mul(x = var_4997_cast_fp16, y = var_4998_to_fp16)[name = string("attn_109_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> input_271_cast_fp16 = add(x = attn_109_cast_fp16, y = attention_mask)[name = string("input_271_cast_fp16")];
            tensor<fp16, [1, 16, ?, 1024]> attn_cast_fp16 = softmax(axis = var_4873, x = input_271_cast_fp16)[name = string("attn_cast_fp16")];
            bool out_transpose_x_0 = const()[name = string("out_transpose_x_0"), val = bool(false)];
            bool out_transpose_y_0 = const()[name = string("out_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, ?, 128]> out_cast_fp16 = matmul(transpose_x = out_transpose_x_0, transpose_y = out_transpose_y_0, x = attn_cast_fp16, y = v_exp_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [4]> var_5003_perm_0 = const()[name = string("op_5003_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 concat_167_axis_0 = const()[name = string("concat_167_axis_0"), val = int32(0)];
            bool concat_167_interleave_0 = const()[name = string("concat_167_interleave_0"), val = bool(false)];
            int32 gather_163_cast_uint16_to_int32 = cast(dtype = gather_163_cast_uint16_to_int32_dtype_0, x = gather_163_cast_uint16)[name = string("cast_0")];
            tensor<int32, [3]> concat_167 = concat(axis = concat_167_axis_0, interleave = concat_167_interleave_0, values = (gather_162, gather_163_cast_uint16_to_int32, var_4873))[name = string("concat_167")];
            tensor<fp16, [1, ?, 16, 128]> var_5003_cast_fp16 = transpose(perm = var_5003_perm_0, x = out_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, ?, ?]> input_273_cast_fp16 = reshape(shape = concat_167, x = var_5003_cast_fp16)[name = string("input_273_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_27_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455843200))), scale = tensor<fp16, [1024, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457940416))))[name = string("layers_27_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_192_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_27_self_attn_o_proj_weight_to_fp16_quantized, x = input_273_cast_fp16)[name = string("linear_192_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_721_cast_fp16 = add(x = x_701_cast_fp16, y = linear_192_cast_fp16)[name = string("x_721_cast_fp16")];
            fp16 var_4872_promoted_3_to_fp16 = const()[name = string("op_4872_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, ?, 1024]> var_5011_cast_fp16 = pow(x = x_721_cast_fp16, y = var_4872_promoted_3_to_fp16)[name = string("op_5011_cast_fp16")];
            tensor<int32, [1]> var_5013_axes_0 = const()[name = string("op_5013_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5013_keep_dims_0 = const()[name = string("op_5013_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, ?, 1]> var_5013_cast_fp16 = reduce_mean(axes = var_5013_axes_0, keep_dims = var_5013_keep_dims_0, x = var_5011_cast_fp16)[name = string("op_5013_cast_fp16")];
            fp16 var_5014_to_fp16 = const()[name = string("op_5014_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, ?, 1]> var_5015_cast_fp16 = add(x = var_5013_cast_fp16, y = var_5014_to_fp16)[name = string("op_5015_cast_fp16")];
            fp32 norm_223_epsilon_0 = const()[name = string("norm_223_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, ?, 1]> norm_223_cast_fp16 = rsqrt(epsilon = norm_223_epsilon_0, x = var_5015_cast_fp16)[name = string("norm_223_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> var_5017_cast_fp16 = mul(x = x_721_cast_fp16, y = norm_223_cast_fp16)[name = string("op_5017_cast_fp16")];
            tensor<fp16, [1024]> layers_27_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_27_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458071552)))];
            tensor<fp16, [1, ?, 1024]> var_5018_cast_fp16 = mul(x = var_5017_cast_fp16, y = layers_27_post_attention_layernorm_weight_to_fp16)[name = string("op_5018_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_27_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458073664))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461219456))))[name = string("layers_27_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_193_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_27_mlp_gate_proj_weight_to_fp16_quantized, x = var_5018_cast_fp16)[name = string("linear_193_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> var_5028_cast_fp16 = silu(x = linear_193_cast_fp16)[name = string("op_5028_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_27_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461416128))), scale = tensor<fp16, [3072, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464561920))))[name = string("layers_27_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 3072]> linear_194_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_27_mlp_up_proj_weight_to_fp16_quantized, x = var_5018_cast_fp16)[name = string("linear_194_cast_fp16")];
            tensor<fp16, [1, ?, 3072]> input_279_cast_fp16 = mul(x = var_5028_cast_fp16, y = linear_194_cast_fp16)[name = string("input_279_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_27_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464758592))), scale = tensor<fp16, [1024, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467904384))))[name = string("layers_27_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, ?, 1024]> linear_195_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_27_mlp_down_proj_weight_to_fp16_quantized, x = input_279_cast_fp16)[name = string("linear_195_cast_fp16")];
            tensor<fp16, [1, ?, 1024]> x_727_cast_fp16 = add(x = x_721_cast_fp16, y = linear_195_cast_fp16)[name = string("x_727_cast_fp16")];
            tensor<int32, [3]> var_5044_begin_0 = const()[name = string("op_5044_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
            tensor<int32, [3]> var_5044_end_0 = const()[name = string("op_5044_end_0"), val = tensor<int32, [3]>([1, 0, 1024])];
            tensor<bool, [3]> var_5044_end_mask_0 = const()[name = string("op_5044_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 1, 1024]> var_5044_cast_fp16 = slice_by_index(begin = var_5044_begin_0, end = var_5044_end_0, end_mask = var_5044_end_mask_0, x = x_727_cast_fp16)[name = string("op_5044_cast_fp16")];
            fp16 var_5053_promoted_to_fp16 = const()[name = string("op_5053_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 1024]> var_5059_cast_fp16 = pow(x = var_5044_cast_fp16, y = var_5053_promoted_to_fp16)[name = string("op_5059_cast_fp16")];
            tensor<int32, [1]> var_5061_axes_0 = const()[name = string("op_5061_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5061_keep_dims_0 = const()[name = string("op_5061_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1]> var_5061_cast_fp16 = reduce_mean(axes = var_5061_axes_0, keep_dims = var_5061_keep_dims_0, x = var_5059_cast_fp16)[name = string("op_5061_cast_fp16")];
            fp16 var_5062_to_fp16 = const()[name = string("op_5062_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_5063_cast_fp16 = add(x = var_5061_cast_fp16, y = var_5062_to_fp16)[name = string("op_5063_cast_fp16")];
            fp32 norm_225_epsilon_0 = const()[name = string("norm_225_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1]> norm_225_cast_fp16 = rsqrt(epsilon = norm_225_epsilon_0, x = var_5063_cast_fp16)[name = string("norm_225_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> var_5065_cast_fp16 = mul(x = var_5044_cast_fp16, y = norm_225_cast_fp16)[name = string("op_5065_cast_fp16")];
            tensor<fp16, [1024]> norm_weight_to_fp16 = const()[name = string("norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468101056)))];
            tensor<fp16, [1, 1, 1024]> var_5066_cast_fp16 = mul(x = var_5065_cast_fp16, y = norm_weight_to_fp16)[name = string("op_5066_cast_fp16")];
            tensor<fp16, [151936, 1024]> lm_head_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [151936, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468103168))), scale = tensor<fp16, [151936, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(623685696))))[name = string("lm_head_weight_to_fp16_quantized")];
            tensor<fp16, [151936]> linear_196_bias_0_to_fp16 = const()[name = string("linear_196_bias_0_to_fp16"), val = tensor<fp16, [151936]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633409664)))];
            tensor<fp16, [1, 1, 151936]> logits = linear(bias = linear_196_bias_0_to_fp16, weight = lm_head_weight_to_fp16_quantized, x = var_5066_cast_fp16)[name = string("linear_196_cast_fp16")];
        } -> (logits);
    func embedding<ios18>(tensor<int32, [1, 1]> token_id) {
            int32 var_6_axis_0 = const()[name = string("op_6_axis_0"), val = int32(0)];
            int32 var_6_batch_dims_0 = const()[name = string("op_6_batch_dims_0"), val = int32(0)];
            bool var_6_validate_indices_0 = const()[name = string("op_6_validate_indices_0"), val = bool(false)];
            tensor<fp16, [151936, 1024]> embedding_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [151936, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468103168))), scale = tensor<fp16, [151936, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(623685696))))[name = string("embedding_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1, 1024]> embedding = gather(axis = var_6_axis_0, batch_dims = var_6_batch_dims_0, indices = token_id, validate_indices = var_6_validate_indices_0, x = embedding_weight_to_fp16_quantized)[name = string("op_6_cast_fp16")];
        } -> (embedding);
}