program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1.2"}})]
{
    func main<ios18>(tensor<fp16, [1, 1, 64, 128]> cos, tensor<fp16, [1, 64, 2048]> hidden, state<tensor<fp16, [7, 8, 704, 128]>> k_cache, tensor<fp16, [1, 1, 64, 704]> mask, tensor<fp16, [64, 704]> onehot, tensor<fp16, [1, 1, 64, 128]> sin, state<tensor<fp16, [7, 8, 704, 128]>> v_cache) {
            tensor<int32, [1]> var_243_axes_0 = const()[name = string("op_243_axes_0"), val = tensor<int32, [1]>([0])];
            bool var_243_keep_dims_0 = const()[name = string("op_243_keep_dims_0"), val = bool(false)];
            tensor<fp16, [704]> var_243_cast_fp16 = reduce_sum(axes = var_243_axes_0, keep_dims = var_243_keep_dims_0, x = onehot)[name = string("op_243_cast_fp16")];
            fp16 const_0_to_fp16 = const()[name = string("const_0_to_fp16"), val = fp16(-inf)];
            fp16 var_245_promoted_to_fp16 = const()[name = string("op_245_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [704]> clip_0_cast_fp16 = clip(alpha = const_0_to_fp16, beta = var_245_promoted_to_fp16, x = var_243_cast_fp16)[name = string("clip_0_cast_fp16")];
            fp16 var_247_promoted_to_fp16 = const()[name = string("op_247_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [704]> keep_cast_fp16 = sub(x = var_247_promoted_to_fp16, y = clip_0_cast_fp16)[name = string("keep_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_250_cast_fp16 = abs(x = hidden)[name = string("op_250_cast_fp16")];
            bool var_253 = const()[name = string("op_253"), val = bool(true)];
            tensor<int32, [1]> var_254_axes_0 = const()[name = string("op_254_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_254_cast_fp16 = reduce_max(axes = var_254_axes_0, keep_dims = var_253, x = var_250_cast_fp16)[name = string("op_254_cast_fp16")];
            fp16 var_255_to_fp16 = const()[name = string("op_255_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_1_to_fp16 = const()[name = string("const_1_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_1_cast_fp16 = clip(alpha = var_255_to_fp16, beta = const_1_to_fp16, x = var_254_cast_fp16)[name = string("clip_1_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> y_1_cast_fp16 = real_div(x = hidden, y = clip_1_cast_fp16)[name = string("y_1_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_259_cast_fp16 = mul(x = y_1_cast_fp16, y = y_1_cast_fp16)[name = string("op_259_cast_fp16")];
            tensor<int32, [1]> var_264_axes_0 = const()[name = string("op_264_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_264_keep_dims_0 = const()[name = string("op_264_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_264_cast_fp16 = reduce_mean(axes = var_264_axes_0, keep_dims = var_264_keep_dims_0, x = var_259_cast_fp16)[name = string("op_264_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_265_cast_fp16 = mul(x = clip_1_cast_fp16, y = clip_1_cast_fp16)[name = string("op_265_cast_fp16")];
            fp32 var_266_epsilon_0 = const()[name = string("op_266_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_266_cast_fp16 = inverse(epsilon = var_266_epsilon_0, x = var_265_cast_fp16)[name = string("op_266_cast_fp16")];
            fp16 var_267_to_fp16 = const()[name = string("op_267_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_268_cast_fp16 = mul(x = var_266_cast_fp16, y = var_267_to_fp16)[name = string("op_268_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_270_cast_fp16 = add(x = var_264_cast_fp16, y = var_268_cast_fp16)[name = string("op_270_cast_fp16")];
            fp32 var_271_epsilon_0 = const()[name = string("op_271_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_271_cast_fp16 = rsqrt(epsilon = var_271_epsilon_0, x = var_270_cast_fp16)[name = string("op_271_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_272_cast_fp16 = mul(x = y_1_cast_fp16, y = var_271_cast_fp16)[name = string("op_272_cast_fp16")];
            tensor<fp16, [2048]> layers_0_input_norm_to_fp16 = const()[name = string("layers_0_input_norm_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            tensor<fp16, [1, 64, 2048]> input_1_cast_fp16 = mul(x = var_272_cast_fp16, y = layers_0_input_norm_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [2048, 2048]> layers_0_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4198592))))[name = string("layers_0_q_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(4202752)))];
            tensor<fp16, [1, 64, 2048]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_q_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> var_281 = const()[name = string("op_281"), val = tensor<int32, [4]>([1, 64, 16, 128])];
            tensor<fp16, [1, 64, 16, 128]> var_282_cast_fp16 = reshape(shape = var_281, x = linear_0_cast_fp16)[name = string("op_282_cast_fp16")];
            tensor<int32, [4]> x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 2048]> layers_0_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4206912))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6304128))))[name = string("layers_0_k_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(6306240)))];
            tensor<fp16, [1, 64, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_k_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_293 = const()[name = string("op_293"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<fp16, [1, 64, 8, 128]> var_294_cast_fp16 = reshape(shape = var_293, x = linear_1_cast_fp16)[name = string("op_294_cast_fp16")];
            tensor<int32, [4]> x_3_perm_0 = const()[name = string("x_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 2048]> layers_0_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6308352))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8405568))))[name = string("layers_0_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_2_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_v_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_305 = const()[name = string("op_305"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<fp16, [1, 64, 8, 128]> var_306_cast_fp16 = reshape(shape = var_305, x = linear_2_cast_fp16)[name = string("op_306_cast_fp16")];
            tensor<int32, [4]> v_1_perm_0 = const()[name = string("v_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 64, 128]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = var_282_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 16, 64, 128]> var_310_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_310_cast_fp16")];
            bool var_313 = const()[name = string("op_313"), val = bool(true)];
            tensor<int32, [1]> var_314_axes_0 = const()[name = string("op_314_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 64, 1]> var_314_cast_fp16 = reduce_max(axes = var_314_axes_0, keep_dims = var_313, x = var_310_cast_fp16)[name = string("op_314_cast_fp16")];
            fp16 var_315_to_fp16 = const()[name = string("op_315_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_2_to_fp16 = const()[name = string("const_2_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 16, 64, 1]> clip_2_cast_fp16 = clip(alpha = var_315_to_fp16, beta = const_2_to_fp16, x = var_314_cast_fp16)[name = string("clip_2_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> y_3_cast_fp16 = real_div(x = x_1_cast_fp16, y = clip_2_cast_fp16)[name = string("y_3_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_319_cast_fp16 = mul(x = y_3_cast_fp16, y = y_3_cast_fp16)[name = string("op_319_cast_fp16")];
            tensor<int32, [1]> var_324_axes_0 = const()[name = string("op_324_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_324_keep_dims_0 = const()[name = string("op_324_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 64, 1]> var_324_cast_fp16 = reduce_mean(axes = var_324_axes_0, keep_dims = var_324_keep_dims_0, x = var_319_cast_fp16)[name = string("op_324_cast_fp16")];
            tensor<fp16, [1, 16, 64, 1]> var_325_cast_fp16 = mul(x = clip_2_cast_fp16, y = clip_2_cast_fp16)[name = string("op_325_cast_fp16")];
            fp32 var_326_epsilon_0 = const()[name = string("op_326_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 64, 1]> var_326_cast_fp16 = inverse(epsilon = var_326_epsilon_0, x = var_325_cast_fp16)[name = string("op_326_cast_fp16")];
            fp16 var_327_to_fp16 = const()[name = string("op_327_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 64, 1]> var_328_cast_fp16 = mul(x = var_326_cast_fp16, y = var_327_to_fp16)[name = string("op_328_cast_fp16")];
            tensor<fp16, [1, 16, 64, 1]> var_330_cast_fp16 = add(x = var_324_cast_fp16, y = var_328_cast_fp16)[name = string("op_330_cast_fp16")];
            fp32 var_331_epsilon_0 = const()[name = string("op_331_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 64, 1]> var_331_cast_fp16 = rsqrt(epsilon = var_331_epsilon_0, x = var_330_cast_fp16)[name = string("op_331_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_332_cast_fp16 = mul(x = y_3_cast_fp16, y = var_331_cast_fp16)[name = string("op_332_cast_fp16")];
            tensor<fp16, [128]> layers_0_q_norm_to_fp16 = const()[name = string("layers_0_q_norm_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8407680)))];
            tensor<fp16, [1, 16, 64, 128]> q_1_cast_fp16 = mul(x = var_332_cast_fp16, y = layers_0_q_norm_to_fp16)[name = string("q_1_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> x_3_cast_fp16 = transpose(perm = x_3_perm_0, x = var_294_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 8, 64, 128]> var_334_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_334_cast_fp16")];
            bool var_337 = const()[name = string("op_337"), val = bool(true)];
            tensor<int32, [1]> var_338_axes_0 = const()[name = string("op_338_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 64, 1]> var_338_cast_fp16 = reduce_max(axes = var_338_axes_0, keep_dims = var_337, x = var_334_cast_fp16)[name = string("op_338_cast_fp16")];
            fp16 var_339_to_fp16 = const()[name = string("op_339_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_3_to_fp16 = const()[name = string("const_3_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 8, 64, 1]> clip_3_cast_fp16 = clip(alpha = var_339_to_fp16, beta = const_3_to_fp16, x = var_338_cast_fp16)[name = string("clip_3_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> y_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = clip_3_cast_fp16)[name = string("y_5_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_343_cast_fp16 = mul(x = y_5_cast_fp16, y = y_5_cast_fp16)[name = string("op_343_cast_fp16")];
            tensor<int32, [1]> var_348_axes_0 = const()[name = string("op_348_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_348_keep_dims_0 = const()[name = string("op_348_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 64, 1]> var_348_cast_fp16 = reduce_mean(axes = var_348_axes_0, keep_dims = var_348_keep_dims_0, x = var_343_cast_fp16)[name = string("op_348_cast_fp16")];
            tensor<fp16, [1, 8, 64, 1]> var_349_cast_fp16 = mul(x = clip_3_cast_fp16, y = clip_3_cast_fp16)[name = string("op_349_cast_fp16")];
            fp32 var_350_epsilon_0 = const()[name = string("op_350_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 64, 1]> var_350_cast_fp16 = inverse(epsilon = var_350_epsilon_0, x = var_349_cast_fp16)[name = string("op_350_cast_fp16")];
            fp16 var_351_to_fp16 = const()[name = string("op_351_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 64, 1]> var_352_cast_fp16 = mul(x = var_350_cast_fp16, y = var_351_to_fp16)[name = string("op_352_cast_fp16")];
            tensor<fp16, [1, 8, 64, 1]> var_354_cast_fp16 = add(x = var_348_cast_fp16, y = var_352_cast_fp16)[name = string("op_354_cast_fp16")];
            fp32 var_355_epsilon_0 = const()[name = string("op_355_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 64, 1]> var_355_cast_fp16 = rsqrt(epsilon = var_355_epsilon_0, x = var_354_cast_fp16)[name = string("op_355_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_356_cast_fp16 = mul(x = y_5_cast_fp16, y = var_355_cast_fp16)[name = string("op_356_cast_fp16")];
            tensor<fp16, [128]> layers_0_k_norm_to_fp16 = const()[name = string("layers_0_k_norm_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8408000)))];
            tensor<fp16, [1, 8, 64, 128]> k_1_cast_fp16 = mul(x = var_356_cast_fp16, y = layers_0_k_norm_to_fp16)[name = string("k_1_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_358_cast_fp16 = mul(x = q_1_cast_fp16, y = cos)[name = string("op_358_cast_fp16")];
            tensor<int32, [2]> var_359_split_sizes_0 = const()[name = string("op_359_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_359_axis_0 = const()[name = string("op_359_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 64, 64]> var_359_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_359_cast_fp16_1 = split(axis = var_359_axis_0, split_sizes = var_359_split_sizes_0, x = q_1_cast_fp16)[name = string("op_359_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_361_cast_fp16 = mul(x = var_359_cast_fp16_1, y = const_4_promoted_to_fp16)[name = string("op_361_cast_fp16")];
            int32 var_363 = const()[name = string("op_363"), val = int32(-1)];
            bool var_364_interleave_0 = const()[name = string("op_364_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> var_364_cast_fp16 = concat(axis = var_363, interleave = var_364_interleave_0, values = (var_361_cast_fp16, var_359_cast_fp16_0))[name = string("op_364_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_365_cast_fp16 = mul(x = var_364_cast_fp16, y = sin)[name = string("op_365_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> q_3_cast_fp16 = add(x = var_358_cast_fp16, y = var_365_cast_fp16)[name = string("q_3_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_368_cast_fp16 = mul(x = k_1_cast_fp16, y = cos)[name = string("op_368_cast_fp16")];
            tensor<int32, [2]> var_369_split_sizes_0 = const()[name = string("op_369_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_369_axis_0 = const()[name = string("op_369_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 64, 64]> var_369_cast_fp16_0, tensor<fp16, [1, 8, 64, 64]> var_369_cast_fp16_1 = split(axis = var_369_axis_0, split_sizes = var_369_split_sizes_0, x = k_1_cast_fp16)[name = string("op_369_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 64, 64]> var_371_cast_fp16 = mul(x = var_369_cast_fp16_1, y = const_5_promoted_to_fp16)[name = string("op_371_cast_fp16")];
            int32 var_373 = const()[name = string("op_373"), val = int32(-1)];
            bool var_374_interleave_0 = const()[name = string("op_374_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_374_cast_fp16 = concat(axis = var_373, interleave = var_374_interleave_0, values = (var_371_cast_fp16, var_369_cast_fp16_0))[name = string("op_374_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_375_cast_fp16 = mul(x = var_374_cast_fp16, y = sin)[name = string("op_375_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> k_3_cast_fp16 = add(x = var_368_cast_fp16, y = var_375_cast_fp16)[name = string("k_3_cast_fp16")];
            tensor<fp16, [7, 8, 704, 128]> read_state_0 = read_state(input = k_cache)[name = string("read_state_0")];
            tensor<int32, [4]> var_382_begin_0 = const()[name = string("op_382_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_382_end_0 = const()[name = string("op_382_end_0"), val = tensor<int32, [4]>([1, 8, 704, 128])];
            tensor<bool, [4]> var_382_end_mask_0 = const()[name = string("op_382_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_382_cast_fp16 = slice_by_index(begin = var_382_begin_0, end = var_382_end_0, end_mask = var_382_end_mask_0, x = read_state_0)[name = string("op_382_cast_fp16")];
            tensor<int32, [4]> var_392 = const()[name = string("op_392"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            tensor<fp16, [1, 1, 704, 1]> var_393_cast_fp16 = reshape(shape = var_392, x = keep_cast_fp16)[name = string("op_393_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_394_cast_fp16 = mul(x = var_382_cast_fp16, y = var_393_cast_fp16)[name = string("op_394_cast_fp16")];
            tensor<int32, [4]> var_398_begin_0 = const()[name = string("op_398_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_398_end_0 = const()[name = string("op_398_end_0"), val = tensor<int32, [4]>([1, 8, 64, 128])];
            tensor<bool, [4]> var_398_end_mask_0 = const()[name = string("op_398_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_398_squeeze_mask_0 = const()[name = string("op_398_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [8, 64, 128]> var_398_cast_fp16 = slice_by_index(begin = var_398_begin_0, end = var_398_end_0, end_mask = var_398_end_mask_0, squeeze_mask = var_398_squeeze_mask_0, x = k_3_cast_fp16)[name = string("op_398_cast_fp16")];
            bool var_399_transpose_x_1 = const()[name = string("op_399_transpose_x_1"), val = bool(true)];
            bool var_399_transpose_y_1 = const()[name = string("op_399_transpose_y_1"), val = bool(false)];
            tensor<fp16, [8, 704, 128]> var_399_cast_fp16 = matmul(transpose_x = var_399_transpose_x_1, transpose_y = var_399_transpose_y_1, x = onehot, y = var_398_cast_fp16)[name = string("op_399_cast_fp16")];
            tensor<int32, [1]> var_401_axes_0 = const()[name = string("op_401_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8, 704, 128]> var_401_cast_fp16 = expand_dims(axes = var_401_axes_0, x = var_399_cast_fp16)[name = string("op_401_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_403_cast_fp16 = add(x = var_394_cast_fp16, y = var_401_cast_fp16)[name = string("op_403_cast_fp16")];
            tensor<int32, [4]> concat_0 = const()[name = string("concat_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> concat_1 = const()[name = string("concat_1"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> k_cache_internal_tensor_assign_1_stride_0 = const()[name = string("k_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("k_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("k_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("k_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [7, 8, 704, 128]> k_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_0, begin_mask = k_cache_internal_tensor_assign_1_begin_mask_0, end = concat_1, end_mask = k_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = k_cache_internal_tensor_assign_1_squeeze_mask_0, stride = k_cache_internal_tensor_assign_1_stride_0, update = var_403_cast_fp16, x = read_state_0)[name = string("k_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = k_cache_internal_tensor_assign_1_cast_fp16, input = k_cache)[name = string("coreml_update_state_14_write_state")];
            tensor<fp16, [7, 8, 704, 128]> coreml_update_state_14 = read_state(input = k_cache)[name = string("coreml_update_state_14")];
            tensor<fp16, [7, 8, 704, 128]> read_state_1 = read_state(input = v_cache)[name = string("read_state_1")];
            tensor<int32, [4]> var_420_begin_0 = const()[name = string("op_420_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_420_end_0 = const()[name = string("op_420_end_0"), val = tensor<int32, [4]>([1, 8, 704, 128])];
            tensor<bool, [4]> var_420_end_mask_0 = const()[name = string("op_420_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_420_cast_fp16 = slice_by_index(begin = var_420_begin_0, end = var_420_end_0, end_mask = var_420_end_mask_0, x = read_state_1)[name = string("op_420_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_432_cast_fp16 = mul(x = var_420_cast_fp16, y = var_393_cast_fp16)[name = string("op_432_cast_fp16")];
            tensor<int32, [4]> var_436_begin_0 = const()[name = string("op_436_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_436_end_0 = const()[name = string("op_436_end_0"), val = tensor<int32, [4]>([1, 8, 64, 128])];
            tensor<bool, [4]> var_436_end_mask_0 = const()[name = string("op_436_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_436_squeeze_mask_0 = const()[name = string("op_436_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 8, 64, 128]> v_1_cast_fp16 = transpose(perm = v_1_perm_0, x = var_306_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [8, 64, 128]> var_436_cast_fp16 = slice_by_index(begin = var_436_begin_0, end = var_436_end_0, end_mask = var_436_end_mask_0, squeeze_mask = var_436_squeeze_mask_0, x = v_1_cast_fp16)[name = string("op_436_cast_fp16")];
            bool var_437_transpose_x_1 = const()[name = string("op_437_transpose_x_1"), val = bool(true)];
            bool var_437_transpose_y_1 = const()[name = string("op_437_transpose_y_1"), val = bool(false)];
            tensor<fp16, [8, 704, 128]> var_437_cast_fp16 = matmul(transpose_x = var_437_transpose_x_1, transpose_y = var_437_transpose_y_1, x = onehot, y = var_436_cast_fp16)[name = string("op_437_cast_fp16")];
            tensor<int32, [1]> var_439_axes_0 = const()[name = string("op_439_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8, 704, 128]> var_439_cast_fp16 = expand_dims(axes = var_439_axes_0, x = var_437_cast_fp16)[name = string("op_439_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_441_cast_fp16 = add(x = var_432_cast_fp16, y = var_439_cast_fp16)[name = string("op_441_cast_fp16")];
            tensor<int32, [4]> concat_2 = const()[name = string("concat_2"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> concat_3 = const()[name = string("concat_3"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> v_cache_internal_tensor_assign_1_stride_0 = const()[name = string("v_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = string("v_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_1_end_mask_0 = const()[name = string("v_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = string("v_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [7, 8, 704, 128]> v_cache_internal_tensor_assign_1_cast_fp16 = slice_update(begin = concat_2, begin_mask = v_cache_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = v_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = v_cache_internal_tensor_assign_1_squeeze_mask_0, stride = v_cache_internal_tensor_assign_1_stride_0, update = var_441_cast_fp16, x = read_state_1)[name = string("v_cache_internal_tensor_assign_1_cast_fp16")];
            write_state(data = v_cache_internal_tensor_assign_1_cast_fp16, input = v_cache)[name = string("coreml_update_state_15_write_state")];
            tensor<fp16, [7, 8, 704, 128]> coreml_update_state_15 = read_state(input = v_cache)[name = string("coreml_update_state_15")];
            tensor<int32, [4]> var_458_begin_0 = const()[name = string("op_458_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_458_end_0 = const()[name = string("op_458_end_0"), val = tensor<int32, [4]>([1, 8, 704, 128])];
            tensor<bool, [4]> var_458_end_mask_0 = const()[name = string("op_458_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_458_cast_fp16 = slice_by_index(begin = var_458_begin_0, end = var_458_end_0, end_mask = var_458_end_mask_0, x = coreml_update_state_14)[name = string("op_458_cast_fp16")];
            tensor<int32, [4]> transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_0_reps_0 = const()[name = string("tile_0_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [8, 1, 704, 128]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = var_458_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [16, 1, 704, 128]> tile_0_cast_fp16 = tile(reps = tile_0_reps_0, x = transpose_0_cast_fp16)[name = string("tile_0_cast_fp16")];
            tensor<int32, [5]> concat_4 = const()[name = string("concat_4"), val = tensor<int32, [5]>([2, 8, 1, 704, 128])];
            tensor<fp16, [2, 8, 1, 704, 128]> reshape_0_cast_fp16 = reshape(shape = concat_4, x = tile_0_cast_fp16)[name = string("reshape_0_cast_fp16")];
            tensor<int32, [5]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_5 = const()[name = string("concat_5"), val = tensor<int32, [4]>([-1, 1, 704, 128])];
            tensor<fp16, [8, 2, 1, 704, 128]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = reshape_0_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [16, 1, 704, 128]> reshape_1_cast_fp16 = reshape(shape = concat_5, x = transpose_1_cast_fp16)[name = string("reshape_1_cast_fp16")];
            tensor<int32, [4]> transpose_28_perm_0 = const()[name = string("transpose_28_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> var_472_begin_0 = const()[name = string("op_472_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_472_end_0 = const()[name = string("op_472_end_0"), val = tensor<int32, [4]>([1, 8, 704, 128])];
            tensor<bool, [4]> var_472_end_mask_0 = const()[name = string("op_472_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_472_cast_fp16 = slice_by_index(begin = var_472_begin_0, end = var_472_end_0, end_mask = var_472_end_mask_0, x = coreml_update_state_15)[name = string("op_472_cast_fp16")];
            tensor<int32, [4]> transpose_2_perm_0 = const()[name = string("transpose_2_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_1_reps_0 = const()[name = string("tile_1_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [8, 1, 704, 128]> transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = var_472_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [16, 1, 704, 128]> tile_1_cast_fp16 = tile(reps = tile_1_reps_0, x = transpose_2_cast_fp16)[name = string("tile_1_cast_fp16")];
            tensor<int32, [5]> concat_6 = const()[name = string("concat_6"), val = tensor<int32, [5]>([2, 8, 1, 704, 128])];
            tensor<fp16, [2, 8, 1, 704, 128]> reshape_2_cast_fp16 = reshape(shape = concat_6, x = tile_1_cast_fp16)[name = string("reshape_2_cast_fp16")];
            tensor<int32, [5]> transpose_3_perm_0 = const()[name = string("transpose_3_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_7 = const()[name = string("concat_7"), val = tensor<int32, [4]>([-1, 1, 704, 128])];
            tensor<fp16, [8, 2, 1, 704, 128]> transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = reshape_2_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [16, 1, 704, 128]> reshape_3_cast_fp16 = reshape(shape = concat_7, x = transpose_3_cast_fp16)[name = string("reshape_3_cast_fp16")];
            tensor<int32, [4]> values_1_perm_0 = const()[name = string("values_1_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_485_transpose_x_0 = const()[name = string("op_485_transpose_x_0"), val = bool(false)];
            bool var_485_transpose_y_0 = const()[name = string("op_485_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 704]> transpose_28_cast_fp16 = transpose(perm = transpose_28_perm_0, x = reshape_1_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 16, 64, 704]> var_485_cast_fp16 = matmul(transpose_x = var_485_transpose_x_0, transpose_y = var_485_transpose_y_0, x = q_3_cast_fp16, y = transpose_28_cast_fp16)[name = string("op_485_cast_fp16")];
            fp16 var_486_to_fp16 = const()[name = string("op_486_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 64, 704]> var_487_cast_fp16 = mul(x = var_485_cast_fp16, y = var_486_to_fp16)[name = string("op_487_cast_fp16")];
            tensor<fp16, [1, 16, 64, 704]> scores_1_cast_fp16 = add(x = var_487_cast_fp16, y = mask)[name = string("scores_1_cast_fp16")];
            int32 var_490 = const()[name = string("op_490"), val = int32(-1)];
            tensor<fp16, [1, 16, 64, 704]> var_492_cast_fp16 = softmax(axis = var_490, x = scores_1_cast_fp16)[name = string("op_492_cast_fp16")];
            bool attn_1_transpose_x_0 = const()[name = string("attn_1_transpose_x_0"), val = bool(false)];
            bool attn_1_transpose_y_0 = const()[name = string("attn_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 704, 128]> values_1_cast_fp16 = transpose(perm = values_1_perm_0, x = reshape_3_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 16, 64, 128]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_492_cast_fp16, y = values_1_cast_fp16)[name = string("attn_1_cast_fp16")];
            tensor<int32, [4]> var_496_perm_0 = const()[name = string("op_496_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_500 = const()[name = string("op_500"), val = tensor<int32, [3]>([1, 64, 2048])];
            tensor<fp16, [1, 64, 16, 128]> var_496_cast_fp16 = transpose(perm = var_496_perm_0, x = attn_1_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 64, 2048]> input_3_cast_fp16 = reshape(shape = var_500, x = var_496_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [2048, 2048]> layers_0_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8408320))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12602688))))[name = string("layers_0_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_3_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_o_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> x_5_cast_fp16 = add(x = hidden, y = linear_3_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_507_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_507_cast_fp16")];
            bool var_510 = const()[name = string("op_510"), val = bool(true)];
            tensor<int32, [1]> var_511_axes_0 = const()[name = string("op_511_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_511_cast_fp16 = reduce_max(axes = var_511_axes_0, keep_dims = var_510, x = var_507_cast_fp16)[name = string("op_511_cast_fp16")];
            fp16 var_512_to_fp16 = const()[name = string("op_512_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_6_to_fp16 = const()[name = string("const_6_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_4_cast_fp16 = clip(alpha = var_512_to_fp16, beta = const_6_to_fp16, x = var_511_cast_fp16)[name = string("clip_4_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> y_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = clip_4_cast_fp16)[name = string("y_7_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_516_cast_fp16 = mul(x = y_7_cast_fp16, y = y_7_cast_fp16)[name = string("op_516_cast_fp16")];
            tensor<int32, [1]> var_521_axes_0 = const()[name = string("op_521_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_521_keep_dims_0 = const()[name = string("op_521_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_521_cast_fp16 = reduce_mean(axes = var_521_axes_0, keep_dims = var_521_keep_dims_0, x = var_516_cast_fp16)[name = string("op_521_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_522_cast_fp16 = mul(x = clip_4_cast_fp16, y = clip_4_cast_fp16)[name = string("op_522_cast_fp16")];
            fp32 var_523_epsilon_0 = const()[name = string("op_523_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_523_cast_fp16 = inverse(epsilon = var_523_epsilon_0, x = var_522_cast_fp16)[name = string("op_523_cast_fp16")];
            fp16 var_524_to_fp16 = const()[name = string("op_524_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_525_cast_fp16 = mul(x = var_523_cast_fp16, y = var_524_to_fp16)[name = string("op_525_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_527_cast_fp16 = add(x = var_521_cast_fp16, y = var_525_cast_fp16)[name = string("op_527_cast_fp16")];
            fp32 var_528_epsilon_0 = const()[name = string("op_528_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_528_cast_fp16 = rsqrt(epsilon = var_528_epsilon_0, x = var_527_cast_fp16)[name = string("op_528_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_529_cast_fp16 = mul(x = y_7_cast_fp16, y = var_528_cast_fp16)[name = string("op_529_cast_fp16")];
            tensor<fp16, [2048]> layers_0_post_norm_to_fp16 = const()[name = string("layers_0_post_norm_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12606848)))];
            tensor<fp16, [1, 64, 2048]> input_5_cast_fp16 = mul(x = var_529_cast_fp16, y = layers_0_post_norm_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [6144, 2048]> layers_0_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12611008))), scale = tensor<fp16, [6144, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25193984))))[name = string("layers_0_gate_weight_to_fp16_quantized")];
            tensor<fp16, [6144]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [6144]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25206336)))];
            tensor<fp16, [1, 64, 6144]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_0_gate_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 64, 6144]> var_534_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_534_cast_fp16")];
            tensor<fp16, [6144, 2048]> layers_0_up_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25218688))), scale = tensor<fp16, [6144, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37801664))))[name = string("layers_0_up_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 6144]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_0_up_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 64, 6144]> var_538_cast_fp16 = mul(x = var_534_cast_fp16, y = linear_5_cast_fp16)[name = string("op_538_cast_fp16")];
            fp16 var_539_to_fp16 = const()[name = string("op_539_to_fp16"), val = fp16(0x1p-8)];
            tensor<fp16, [1, 64, 6144]> input_9_cast_fp16 = mul(x = var_538_cast_fp16, y = var_539_to_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [2048, 6144]> layers_0_down_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37814016))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50396992))))[name = string("layers_0_down_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_down_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_6_cast_fp16")];
            fp16 var_544_to_fp16 = const()[name = string("op_544_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 64, 2048]> var_545_cast_fp16 = mul(x = linear_6_cast_fp16, y = var_544_to_fp16)[name = string("op_545_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> x_7_cast_fp16 = add(x = x_5_cast_fp16, y = var_545_cast_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_548_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_548_cast_fp16")];
            bool var_551 = const()[name = string("op_551"), val = bool(true)];
            tensor<int32, [1]> var_552_axes_0 = const()[name = string("op_552_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_552_cast_fp16 = reduce_max(axes = var_552_axes_0, keep_dims = var_551, x = var_548_cast_fp16)[name = string("op_552_cast_fp16")];
            fp16 var_553_to_fp16 = const()[name = string("op_553_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_7_to_fp16 = const()[name = string("const_7_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_5_cast_fp16 = clip(alpha = var_553_to_fp16, beta = const_7_to_fp16, x = var_552_cast_fp16)[name = string("clip_5_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> y_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = clip_5_cast_fp16)[name = string("y_9_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_557_cast_fp16 = mul(x = y_9_cast_fp16, y = y_9_cast_fp16)[name = string("op_557_cast_fp16")];
            tensor<int32, [1]> var_562_axes_0 = const()[name = string("op_562_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_562_keep_dims_0 = const()[name = string("op_562_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_562_cast_fp16 = reduce_mean(axes = var_562_axes_0, keep_dims = var_562_keep_dims_0, x = var_557_cast_fp16)[name = string("op_562_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_563_cast_fp16 = mul(x = clip_5_cast_fp16, y = clip_5_cast_fp16)[name = string("op_563_cast_fp16")];
            fp32 var_564_epsilon_0 = const()[name = string("op_564_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_564_cast_fp16 = inverse(epsilon = var_564_epsilon_0, x = var_563_cast_fp16)[name = string("op_564_cast_fp16")];
            fp16 var_565_to_fp16 = const()[name = string("op_565_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_566_cast_fp16 = mul(x = var_564_cast_fp16, y = var_565_to_fp16)[name = string("op_566_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_568_cast_fp16 = add(x = var_562_cast_fp16, y = var_566_cast_fp16)[name = string("op_568_cast_fp16")];
            fp32 var_569_epsilon_0 = const()[name = string("op_569_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_569_cast_fp16 = rsqrt(epsilon = var_569_epsilon_0, x = var_568_cast_fp16)[name = string("op_569_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_570_cast_fp16 = mul(x = y_9_cast_fp16, y = var_569_cast_fp16)[name = string("op_570_cast_fp16")];
            tensor<fp16, [2048]> layers_1_input_norm_to_fp16 = const()[name = string("layers_1_input_norm_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50401152)))];
            tensor<fp16, [1, 64, 2048]> input_11_cast_fp16 = mul(x = var_570_cast_fp16, y = layers_1_input_norm_to_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [2048, 2048]> layers_1_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50405312))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54599680))))[name = string("layers_1_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_7_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_q_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> var_579 = const()[name = string("op_579"), val = tensor<int32, [4]>([1, 64, 16, 128])];
            tensor<fp16, [1, 64, 16, 128]> var_580_cast_fp16 = reshape(shape = var_579, x = linear_7_cast_fp16)[name = string("op_580_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, 2048]> layers_1_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54603840))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56701056))))[name = string("layers_1_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_8_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_k_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_591 = const()[name = string("op_591"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<fp16, [1, 64, 8, 128]> var_592_cast_fp16 = reshape(shape = var_591, x = linear_8_cast_fp16)[name = string("op_592_cast_fp16")];
            tensor<int32, [4]> x_11_perm_0 = const()[name = string("x_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 2048]> layers_1_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56703168))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58800384))))[name = string("layers_1_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_v_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_603 = const()[name = string("op_603"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<fp16, [1, 64, 8, 128]> var_604_cast_fp16 = reshape(shape = var_603, x = linear_9_cast_fp16)[name = string("op_604_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 64, 128]> x_9_cast_fp16 = transpose(perm = x_9_perm_0, x = var_580_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 16, 64, 128]> var_608_cast_fp16 = abs(x = x_9_cast_fp16)[name = string("op_608_cast_fp16")];
            bool var_611 = const()[name = string("op_611"), val = bool(true)];
            tensor<int32, [1]> var_612_axes_0 = const()[name = string("op_612_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 64, 1]> var_612_cast_fp16 = reduce_max(axes = var_612_axes_0, keep_dims = var_611, x = var_608_cast_fp16)[name = string("op_612_cast_fp16")];
            fp16 var_613_to_fp16 = const()[name = string("op_613_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_8_to_fp16 = const()[name = string("const_8_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 16, 64, 1]> clip_6_cast_fp16 = clip(alpha = var_613_to_fp16, beta = const_8_to_fp16, x = var_612_cast_fp16)[name = string("clip_6_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> y_11_cast_fp16 = real_div(x = x_9_cast_fp16, y = clip_6_cast_fp16)[name = string("y_11_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_617_cast_fp16 = mul(x = y_11_cast_fp16, y = y_11_cast_fp16)[name = string("op_617_cast_fp16")];
            tensor<int32, [1]> var_622_axes_0 = const()[name = string("op_622_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_622_keep_dims_0 = const()[name = string("op_622_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 64, 1]> var_622_cast_fp16 = reduce_mean(axes = var_622_axes_0, keep_dims = var_622_keep_dims_0, x = var_617_cast_fp16)[name = string("op_622_cast_fp16")];
            tensor<fp16, [1, 16, 64, 1]> var_623_cast_fp16 = mul(x = clip_6_cast_fp16, y = clip_6_cast_fp16)[name = string("op_623_cast_fp16")];
            fp32 var_624_epsilon_0 = const()[name = string("op_624_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 64, 1]> var_624_cast_fp16 = inverse(epsilon = var_624_epsilon_0, x = var_623_cast_fp16)[name = string("op_624_cast_fp16")];
            fp16 var_625_to_fp16 = const()[name = string("op_625_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 64, 1]> var_626_cast_fp16 = mul(x = var_624_cast_fp16, y = var_625_to_fp16)[name = string("op_626_cast_fp16")];
            tensor<fp16, [1, 16, 64, 1]> var_628_cast_fp16 = add(x = var_622_cast_fp16, y = var_626_cast_fp16)[name = string("op_628_cast_fp16")];
            fp32 var_629_epsilon_0 = const()[name = string("op_629_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 64, 1]> var_629_cast_fp16 = rsqrt(epsilon = var_629_epsilon_0, x = var_628_cast_fp16)[name = string("op_629_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_630_cast_fp16 = mul(x = y_11_cast_fp16, y = var_629_cast_fp16)[name = string("op_630_cast_fp16")];
            tensor<fp16, [128]> layers_1_q_norm_to_fp16 = const()[name = string("layers_1_q_norm_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58802496)))];
            tensor<fp16, [1, 16, 64, 128]> q_5_cast_fp16 = mul(x = var_630_cast_fp16, y = layers_1_q_norm_to_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> x_11_cast_fp16 = transpose(perm = x_11_perm_0, x = var_592_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 8, 64, 128]> var_632_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_632_cast_fp16")];
            bool var_635 = const()[name = string("op_635"), val = bool(true)];
            tensor<int32, [1]> var_636_axes_0 = const()[name = string("op_636_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 64, 1]> var_636_cast_fp16 = reduce_max(axes = var_636_axes_0, keep_dims = var_635, x = var_632_cast_fp16)[name = string("op_636_cast_fp16")];
            fp16 var_637_to_fp16 = const()[name = string("op_637_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 8, 64, 1]> clip_7_cast_fp16 = clip(alpha = var_637_to_fp16, beta = const_9_to_fp16, x = var_636_cast_fp16)[name = string("clip_7_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> y_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = clip_7_cast_fp16)[name = string("y_13_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_641_cast_fp16 = mul(x = y_13_cast_fp16, y = y_13_cast_fp16)[name = string("op_641_cast_fp16")];
            tensor<int32, [1]> var_646_axes_0 = const()[name = string("op_646_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_646_keep_dims_0 = const()[name = string("op_646_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 64, 1]> var_646_cast_fp16 = reduce_mean(axes = var_646_axes_0, keep_dims = var_646_keep_dims_0, x = var_641_cast_fp16)[name = string("op_646_cast_fp16")];
            tensor<fp16, [1, 8, 64, 1]> var_647_cast_fp16 = mul(x = clip_7_cast_fp16, y = clip_7_cast_fp16)[name = string("op_647_cast_fp16")];
            fp32 var_648_epsilon_0 = const()[name = string("op_648_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 64, 1]> var_648_cast_fp16 = inverse(epsilon = var_648_epsilon_0, x = var_647_cast_fp16)[name = string("op_648_cast_fp16")];
            fp16 var_649_to_fp16 = const()[name = string("op_649_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 64, 1]> var_650_cast_fp16 = mul(x = var_648_cast_fp16, y = var_649_to_fp16)[name = string("op_650_cast_fp16")];
            tensor<fp16, [1, 8, 64, 1]> var_652_cast_fp16 = add(x = var_646_cast_fp16, y = var_650_cast_fp16)[name = string("op_652_cast_fp16")];
            fp32 var_653_epsilon_0 = const()[name = string("op_653_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 64, 1]> var_653_cast_fp16 = rsqrt(epsilon = var_653_epsilon_0, x = var_652_cast_fp16)[name = string("op_653_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_654_cast_fp16 = mul(x = y_13_cast_fp16, y = var_653_cast_fp16)[name = string("op_654_cast_fp16")];
            tensor<fp16, [128]> layers_1_k_norm_to_fp16 = const()[name = string("layers_1_k_norm_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58802816)))];
            tensor<fp16, [1, 8, 64, 128]> k_5_cast_fp16 = mul(x = var_654_cast_fp16, y = layers_1_k_norm_to_fp16)[name = string("k_5_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_656_cast_fp16 = mul(x = q_5_cast_fp16, y = cos)[name = string("op_656_cast_fp16")];
            tensor<int32, [2]> var_657_split_sizes_0 = const()[name = string("op_657_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_657_axis_0 = const()[name = string("op_657_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 64, 64]> var_657_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_657_cast_fp16_1 = split(axis = var_657_axis_0, split_sizes = var_657_split_sizes_0, x = q_5_cast_fp16)[name = string("op_657_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_659_cast_fp16 = mul(x = var_657_cast_fp16_1, y = const_10_promoted_to_fp16)[name = string("op_659_cast_fp16")];
            int32 var_661 = const()[name = string("op_661"), val = int32(-1)];
            bool var_662_interleave_0 = const()[name = string("op_662_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> var_662_cast_fp16 = concat(axis = var_661, interleave = var_662_interleave_0, values = (var_659_cast_fp16, var_657_cast_fp16_0))[name = string("op_662_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_663_cast_fp16 = mul(x = var_662_cast_fp16, y = sin)[name = string("op_663_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> q_7_cast_fp16 = add(x = var_656_cast_fp16, y = var_663_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_666_cast_fp16 = mul(x = k_5_cast_fp16, y = cos)[name = string("op_666_cast_fp16")];
            tensor<int32, [2]> var_667_split_sizes_0 = const()[name = string("op_667_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_667_axis_0 = const()[name = string("op_667_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 64, 64]> var_667_cast_fp16_0, tensor<fp16, [1, 8, 64, 64]> var_667_cast_fp16_1 = split(axis = var_667_axis_0, split_sizes = var_667_split_sizes_0, x = k_5_cast_fp16)[name = string("op_667_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 64, 64]> var_669_cast_fp16 = mul(x = var_667_cast_fp16_1, y = const_11_promoted_to_fp16)[name = string("op_669_cast_fp16")];
            int32 var_671 = const()[name = string("op_671"), val = int32(-1)];
            bool var_672_interleave_0 = const()[name = string("op_672_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_672_cast_fp16 = concat(axis = var_671, interleave = var_672_interleave_0, values = (var_669_cast_fp16, var_667_cast_fp16_0))[name = string("op_672_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_673_cast_fp16 = mul(x = var_672_cast_fp16, y = sin)[name = string("op_673_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> k_7_cast_fp16 = add(x = var_666_cast_fp16, y = var_673_cast_fp16)[name = string("k_7_cast_fp16")];
            tensor<int32, [4]> var_680_begin_0 = const()[name = string("op_680_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_680_end_0 = const()[name = string("op_680_end_0"), val = tensor<int32, [4]>([2, 8, 704, 128])];
            tensor<bool, [4]> var_680_end_mask_0 = const()[name = string("op_680_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_680_cast_fp16 = slice_by_index(begin = var_680_begin_0, end = var_680_end_0, end_mask = var_680_end_mask_0, x = coreml_update_state_14)[name = string("op_680_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_692_cast_fp16 = mul(x = var_680_cast_fp16, y = var_393_cast_fp16)[name = string("op_692_cast_fp16")];
            tensor<int32, [4]> var_696_begin_0 = const()[name = string("op_696_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_696_end_0 = const()[name = string("op_696_end_0"), val = tensor<int32, [4]>([1, 8, 64, 128])];
            tensor<bool, [4]> var_696_end_mask_0 = const()[name = string("op_696_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_696_squeeze_mask_0 = const()[name = string("op_696_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [8, 64, 128]> var_696_cast_fp16 = slice_by_index(begin = var_696_begin_0, end = var_696_end_0, end_mask = var_696_end_mask_0, squeeze_mask = var_696_squeeze_mask_0, x = k_7_cast_fp16)[name = string("op_696_cast_fp16")];
            bool var_697_transpose_x_1 = const()[name = string("op_697_transpose_x_1"), val = bool(true)];
            bool var_697_transpose_y_1 = const()[name = string("op_697_transpose_y_1"), val = bool(false)];
            tensor<fp16, [8, 704, 128]> var_697_cast_fp16 = matmul(transpose_x = var_697_transpose_x_1, transpose_y = var_697_transpose_y_1, x = onehot, y = var_696_cast_fp16)[name = string("op_697_cast_fp16")];
            tensor<int32, [1]> var_699_axes_0 = const()[name = string("op_699_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8, 704, 128]> var_699_cast_fp16 = expand_dims(axes = var_699_axes_0, x = var_697_cast_fp16)[name = string("op_699_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_701_cast_fp16 = add(x = var_692_cast_fp16, y = var_699_cast_fp16)[name = string("op_701_cast_fp16")];
            tensor<int32, [4]> concat_8 = const()[name = string("concat_8"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> concat_9 = const()[name = string("concat_9"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> k_cache_internal_tensor_assign_2_stride_0 = const()[name = string("k_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("k_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("k_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("k_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [7, 8, 704, 128]> k_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_8, begin_mask = k_cache_internal_tensor_assign_2_begin_mask_0, end = concat_9, end_mask = k_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = k_cache_internal_tensor_assign_2_squeeze_mask_0, stride = k_cache_internal_tensor_assign_2_stride_0, update = var_701_cast_fp16, x = coreml_update_state_14)[name = string("k_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = k_cache_internal_tensor_assign_2_cast_fp16, input = k_cache)[name = string("coreml_update_state_16_write_state")];
            tensor<fp16, [7, 8, 704, 128]> coreml_update_state_16 = read_state(input = k_cache)[name = string("coreml_update_state_16")];
            tensor<int32, [4]> var_718_begin_0 = const()[name = string("op_718_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_718_end_0 = const()[name = string("op_718_end_0"), val = tensor<int32, [4]>([2, 8, 704, 128])];
            tensor<bool, [4]> var_718_end_mask_0 = const()[name = string("op_718_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_718_cast_fp16 = slice_by_index(begin = var_718_begin_0, end = var_718_end_0, end_mask = var_718_end_mask_0, x = coreml_update_state_15)[name = string("op_718_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_730_cast_fp16 = mul(x = var_718_cast_fp16, y = var_393_cast_fp16)[name = string("op_730_cast_fp16")];
            tensor<int32, [4]> var_734_begin_0 = const()[name = string("op_734_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_734_end_0 = const()[name = string("op_734_end_0"), val = tensor<int32, [4]>([1, 8, 64, 128])];
            tensor<bool, [4]> var_734_end_mask_0 = const()[name = string("op_734_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_734_squeeze_mask_0 = const()[name = string("op_734_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 8, 64, 128]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_604_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [8, 64, 128]> var_734_cast_fp16 = slice_by_index(begin = var_734_begin_0, end = var_734_end_0, end_mask = var_734_end_mask_0, squeeze_mask = var_734_squeeze_mask_0, x = v_3_cast_fp16)[name = string("op_734_cast_fp16")];
            bool var_735_transpose_x_1 = const()[name = string("op_735_transpose_x_1"), val = bool(true)];
            bool var_735_transpose_y_1 = const()[name = string("op_735_transpose_y_1"), val = bool(false)];
            tensor<fp16, [8, 704, 128]> var_735_cast_fp16 = matmul(transpose_x = var_735_transpose_x_1, transpose_y = var_735_transpose_y_1, x = onehot, y = var_734_cast_fp16)[name = string("op_735_cast_fp16")];
            tensor<int32, [1]> var_737_axes_0 = const()[name = string("op_737_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8, 704, 128]> var_737_cast_fp16 = expand_dims(axes = var_737_axes_0, x = var_735_cast_fp16)[name = string("op_737_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_739_cast_fp16 = add(x = var_730_cast_fp16, y = var_737_cast_fp16)[name = string("op_739_cast_fp16")];
            tensor<int32, [4]> concat_10 = const()[name = string("concat_10"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> concat_11 = const()[name = string("concat_11"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> v_cache_internal_tensor_assign_2_stride_0 = const()[name = string("v_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = string("v_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_2_end_mask_0 = const()[name = string("v_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = string("v_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [7, 8, 704, 128]> v_cache_internal_tensor_assign_2_cast_fp16 = slice_update(begin = concat_10, begin_mask = v_cache_internal_tensor_assign_2_begin_mask_0, end = concat_11, end_mask = v_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = v_cache_internal_tensor_assign_2_squeeze_mask_0, stride = v_cache_internal_tensor_assign_2_stride_0, update = var_739_cast_fp16, x = coreml_update_state_15)[name = string("v_cache_internal_tensor_assign_2_cast_fp16")];
            write_state(data = v_cache_internal_tensor_assign_2_cast_fp16, input = v_cache)[name = string("coreml_update_state_17_write_state")];
            tensor<fp16, [7, 8, 704, 128]> coreml_update_state_17 = read_state(input = v_cache)[name = string("coreml_update_state_17")];
            tensor<int32, [4]> var_756_begin_0 = const()[name = string("op_756_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_756_end_0 = const()[name = string("op_756_end_0"), val = tensor<int32, [4]>([2, 8, 704, 128])];
            tensor<bool, [4]> var_756_end_mask_0 = const()[name = string("op_756_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_756_cast_fp16 = slice_by_index(begin = var_756_begin_0, end = var_756_end_0, end_mask = var_756_end_mask_0, x = coreml_update_state_16)[name = string("op_756_cast_fp16")];
            tensor<int32, [4]> transpose_4_perm_0 = const()[name = string("transpose_4_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_2_reps_0 = const()[name = string("tile_2_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [8, 1, 704, 128]> transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = var_756_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [16, 1, 704, 128]> tile_2_cast_fp16 = tile(reps = tile_2_reps_0, x = transpose_4_cast_fp16)[name = string("tile_2_cast_fp16")];
            tensor<int32, [5]> concat_12 = const()[name = string("concat_12"), val = tensor<int32, [5]>([2, 8, 1, 704, 128])];
            tensor<fp16, [2, 8, 1, 704, 128]> reshape_4_cast_fp16 = reshape(shape = concat_12, x = tile_2_cast_fp16)[name = string("reshape_4_cast_fp16")];
            tensor<int32, [5]> transpose_5_perm_0 = const()[name = string("transpose_5_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_13 = const()[name = string("concat_13"), val = tensor<int32, [4]>([-1, 1, 704, 128])];
            tensor<fp16, [8, 2, 1, 704, 128]> transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = reshape_4_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [16, 1, 704, 128]> reshape_5_cast_fp16 = reshape(shape = concat_13, x = transpose_5_cast_fp16)[name = string("reshape_5_cast_fp16")];
            tensor<int32, [4]> transpose_29_perm_0 = const()[name = string("transpose_29_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> var_770_begin_0 = const()[name = string("op_770_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_770_end_0 = const()[name = string("op_770_end_0"), val = tensor<int32, [4]>([2, 8, 704, 128])];
            tensor<bool, [4]> var_770_end_mask_0 = const()[name = string("op_770_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_770_cast_fp16 = slice_by_index(begin = var_770_begin_0, end = var_770_end_0, end_mask = var_770_end_mask_0, x = coreml_update_state_17)[name = string("op_770_cast_fp16")];
            tensor<int32, [4]> transpose_6_perm_0 = const()[name = string("transpose_6_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_3_reps_0 = const()[name = string("tile_3_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [8, 1, 704, 128]> transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = var_770_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [16, 1, 704, 128]> tile_3_cast_fp16 = tile(reps = tile_3_reps_0, x = transpose_6_cast_fp16)[name = string("tile_3_cast_fp16")];
            tensor<int32, [5]> concat_14 = const()[name = string("concat_14"), val = tensor<int32, [5]>([2, 8, 1, 704, 128])];
            tensor<fp16, [2, 8, 1, 704, 128]> reshape_6_cast_fp16 = reshape(shape = concat_14, x = tile_3_cast_fp16)[name = string("reshape_6_cast_fp16")];
            tensor<int32, [5]> transpose_7_perm_0 = const()[name = string("transpose_7_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_15 = const()[name = string("concat_15"), val = tensor<int32, [4]>([-1, 1, 704, 128])];
            tensor<fp16, [8, 2, 1, 704, 128]> transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = reshape_6_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [16, 1, 704, 128]> reshape_7_cast_fp16 = reshape(shape = concat_15, x = transpose_7_cast_fp16)[name = string("reshape_7_cast_fp16")];
            tensor<int32, [4]> values_3_perm_0 = const()[name = string("values_3_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_783_transpose_x_0 = const()[name = string("op_783_transpose_x_0"), val = bool(false)];
            bool var_783_transpose_y_0 = const()[name = string("op_783_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 704]> transpose_29_cast_fp16 = transpose(perm = transpose_29_perm_0, x = reshape_5_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 16, 64, 704]> var_783_cast_fp16 = matmul(transpose_x = var_783_transpose_x_0, transpose_y = var_783_transpose_y_0, x = q_7_cast_fp16, y = transpose_29_cast_fp16)[name = string("op_783_cast_fp16")];
            fp16 var_784_to_fp16 = const()[name = string("op_784_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 64, 704]> var_785_cast_fp16 = mul(x = var_783_cast_fp16, y = var_784_to_fp16)[name = string("op_785_cast_fp16")];
            tensor<fp16, [1, 16, 64, 704]> scores_3_cast_fp16 = add(x = var_785_cast_fp16, y = mask)[name = string("scores_3_cast_fp16")];
            int32 var_788 = const()[name = string("op_788"), val = int32(-1)];
            tensor<fp16, [1, 16, 64, 704]> var_790_cast_fp16 = softmax(axis = var_788, x = scores_3_cast_fp16)[name = string("op_790_cast_fp16")];
            bool attn_3_transpose_x_0 = const()[name = string("attn_3_transpose_x_0"), val = bool(false)];
            bool attn_3_transpose_y_0 = const()[name = string("attn_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 704, 128]> values_3_cast_fp16 = transpose(perm = values_3_perm_0, x = reshape_7_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 16, 64, 128]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_790_cast_fp16, y = values_3_cast_fp16)[name = string("attn_3_cast_fp16")];
            tensor<int32, [4]> var_794_perm_0 = const()[name = string("op_794_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_798 = const()[name = string("op_798"), val = tensor<int32, [3]>([1, 64, 2048])];
            tensor<fp16, [1, 64, 16, 128]> var_794_cast_fp16 = transpose(perm = var_794_perm_0, x = attn_3_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 64, 2048]> input_13_cast_fp16 = reshape(shape = var_798, x = var_794_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [2048, 2048]> layers_1_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58803136))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62997504))))[name = string("layers_1_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_o_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> x_13_cast_fp16 = add(x = x_7_cast_fp16, y = linear_10_cast_fp16)[name = string("x_13_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_805_cast_fp16 = abs(x = x_13_cast_fp16)[name = string("op_805_cast_fp16")];
            bool var_808 = const()[name = string("op_808"), val = bool(true)];
            tensor<int32, [1]> var_809_axes_0 = const()[name = string("op_809_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_809_cast_fp16 = reduce_max(axes = var_809_axes_0, keep_dims = var_808, x = var_805_cast_fp16)[name = string("op_809_cast_fp16")];
            fp16 var_810_to_fp16 = const()[name = string("op_810_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_12_to_fp16 = const()[name = string("const_12_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_8_cast_fp16 = clip(alpha = var_810_to_fp16, beta = const_12_to_fp16, x = var_809_cast_fp16)[name = string("clip_8_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> y_15_cast_fp16 = real_div(x = x_13_cast_fp16, y = clip_8_cast_fp16)[name = string("y_15_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_814_cast_fp16 = mul(x = y_15_cast_fp16, y = y_15_cast_fp16)[name = string("op_814_cast_fp16")];
            tensor<int32, [1]> var_819_axes_0 = const()[name = string("op_819_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_819_keep_dims_0 = const()[name = string("op_819_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_819_cast_fp16 = reduce_mean(axes = var_819_axes_0, keep_dims = var_819_keep_dims_0, x = var_814_cast_fp16)[name = string("op_819_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_820_cast_fp16 = mul(x = clip_8_cast_fp16, y = clip_8_cast_fp16)[name = string("op_820_cast_fp16")];
            fp32 var_821_epsilon_0 = const()[name = string("op_821_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_821_cast_fp16 = inverse(epsilon = var_821_epsilon_0, x = var_820_cast_fp16)[name = string("op_821_cast_fp16")];
            fp16 var_822_to_fp16 = const()[name = string("op_822_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_823_cast_fp16 = mul(x = var_821_cast_fp16, y = var_822_to_fp16)[name = string("op_823_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_825_cast_fp16 = add(x = var_819_cast_fp16, y = var_823_cast_fp16)[name = string("op_825_cast_fp16")];
            fp32 var_826_epsilon_0 = const()[name = string("op_826_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_826_cast_fp16 = rsqrt(epsilon = var_826_epsilon_0, x = var_825_cast_fp16)[name = string("op_826_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_827_cast_fp16 = mul(x = y_15_cast_fp16, y = var_826_cast_fp16)[name = string("op_827_cast_fp16")];
            tensor<fp16, [2048]> layers_1_post_norm_to_fp16 = const()[name = string("layers_1_post_norm_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63001664)))];
            tensor<fp16, [1, 64, 2048]> input_15_cast_fp16 = mul(x = var_827_cast_fp16, y = layers_1_post_norm_to_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [6144, 2048]> layers_1_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63005824))), scale = tensor<fp16, [6144, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75588800))))[name = string("layers_1_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 6144]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_1_gate_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 64, 6144]> var_832_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_832_cast_fp16")];
            tensor<fp16, [6144, 2048]> layers_1_up_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75601152))), scale = tensor<fp16, [6144, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88184128))))[name = string("layers_1_up_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 6144]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_1_up_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, 64, 6144]> var_836_cast_fp16 = mul(x = var_832_cast_fp16, y = linear_12_cast_fp16)[name = string("op_836_cast_fp16")];
            fp16 var_837_to_fp16 = const()[name = string("op_837_to_fp16"), val = fp16(0x1p-8)];
            tensor<fp16, [1, 64, 6144]> input_19_cast_fp16 = mul(x = var_836_cast_fp16, y = var_837_to_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [2048, 6144]> layers_1_down_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88196480))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100779456))))[name = string("layers_1_down_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_13_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_down_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_13_cast_fp16")];
            fp16 var_842_to_fp16 = const()[name = string("op_842_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 64, 2048]> var_843_cast_fp16 = mul(x = linear_13_cast_fp16, y = var_842_to_fp16)[name = string("op_843_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> x_15_cast_fp16 = add(x = x_13_cast_fp16, y = var_843_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_846_cast_fp16 = abs(x = x_15_cast_fp16)[name = string("op_846_cast_fp16")];
            bool var_849 = const()[name = string("op_849"), val = bool(true)];
            tensor<int32, [1]> var_850_axes_0 = const()[name = string("op_850_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_850_cast_fp16 = reduce_max(axes = var_850_axes_0, keep_dims = var_849, x = var_846_cast_fp16)[name = string("op_850_cast_fp16")];
            fp16 var_851_to_fp16 = const()[name = string("op_851_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_13_to_fp16 = const()[name = string("const_13_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_9_cast_fp16 = clip(alpha = var_851_to_fp16, beta = const_13_to_fp16, x = var_850_cast_fp16)[name = string("clip_9_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> y_17_cast_fp16 = real_div(x = x_15_cast_fp16, y = clip_9_cast_fp16)[name = string("y_17_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_855_cast_fp16 = mul(x = y_17_cast_fp16, y = y_17_cast_fp16)[name = string("op_855_cast_fp16")];
            tensor<int32, [1]> var_860_axes_0 = const()[name = string("op_860_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_860_keep_dims_0 = const()[name = string("op_860_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_860_cast_fp16 = reduce_mean(axes = var_860_axes_0, keep_dims = var_860_keep_dims_0, x = var_855_cast_fp16)[name = string("op_860_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_861_cast_fp16 = mul(x = clip_9_cast_fp16, y = clip_9_cast_fp16)[name = string("op_861_cast_fp16")];
            fp32 var_862_epsilon_0 = const()[name = string("op_862_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_862_cast_fp16 = inverse(epsilon = var_862_epsilon_0, x = var_861_cast_fp16)[name = string("op_862_cast_fp16")];
            fp16 var_863_to_fp16 = const()[name = string("op_863_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_864_cast_fp16 = mul(x = var_862_cast_fp16, y = var_863_to_fp16)[name = string("op_864_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_866_cast_fp16 = add(x = var_860_cast_fp16, y = var_864_cast_fp16)[name = string("op_866_cast_fp16")];
            fp32 var_867_epsilon_0 = const()[name = string("op_867_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_867_cast_fp16 = rsqrt(epsilon = var_867_epsilon_0, x = var_866_cast_fp16)[name = string("op_867_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_868_cast_fp16 = mul(x = y_17_cast_fp16, y = var_867_cast_fp16)[name = string("op_868_cast_fp16")];
            tensor<fp16, [2048]> layers_2_input_norm_to_fp16 = const()[name = string("layers_2_input_norm_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100783616)))];
            tensor<fp16, [1, 64, 2048]> input_21_cast_fp16 = mul(x = var_868_cast_fp16, y = layers_2_input_norm_to_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [2048, 2048]> layers_2_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100787776))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104982144))))[name = string("layers_2_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_q_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> var_877 = const()[name = string("op_877"), val = tensor<int32, [4]>([1, 64, 16, 128])];
            tensor<fp16, [1, 64, 16, 128]> var_878_cast_fp16 = reshape(shape = var_877, x = linear_14_cast_fp16)[name = string("op_878_cast_fp16")];
            tensor<int32, [4]> x_17_perm_0 = const()[name = string("x_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 2048]> layers_2_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104986304))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107083520))))[name = string("layers_2_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_k_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_889 = const()[name = string("op_889"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<fp16, [1, 64, 8, 128]> var_890_cast_fp16 = reshape(shape = var_889, x = linear_15_cast_fp16)[name = string("op_890_cast_fp16")];
            tensor<int32, [4]> x_19_perm_0 = const()[name = string("x_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 2048]> layers_2_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107085632))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109182848))))[name = string("layers_2_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_16_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_v_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_901 = const()[name = string("op_901"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<fp16, [1, 64, 8, 128]> var_902_cast_fp16 = reshape(shape = var_901, x = linear_16_cast_fp16)[name = string("op_902_cast_fp16")];
            tensor<int32, [4]> v_5_perm_0 = const()[name = string("v_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 64, 128]> x_17_cast_fp16 = transpose(perm = x_17_perm_0, x = var_878_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 16, 64, 128]> var_906_cast_fp16 = abs(x = x_17_cast_fp16)[name = string("op_906_cast_fp16")];
            bool var_909 = const()[name = string("op_909"), val = bool(true)];
            tensor<int32, [1]> var_910_axes_0 = const()[name = string("op_910_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 64, 1]> var_910_cast_fp16 = reduce_max(axes = var_910_axes_0, keep_dims = var_909, x = var_906_cast_fp16)[name = string("op_910_cast_fp16")];
            fp16 var_911_to_fp16 = const()[name = string("op_911_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_14_to_fp16 = const()[name = string("const_14_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 16, 64, 1]> clip_10_cast_fp16 = clip(alpha = var_911_to_fp16, beta = const_14_to_fp16, x = var_910_cast_fp16)[name = string("clip_10_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> y_19_cast_fp16 = real_div(x = x_17_cast_fp16, y = clip_10_cast_fp16)[name = string("y_19_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_915_cast_fp16 = mul(x = y_19_cast_fp16, y = y_19_cast_fp16)[name = string("op_915_cast_fp16")];
            tensor<int32, [1]> var_920_axes_0 = const()[name = string("op_920_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_920_keep_dims_0 = const()[name = string("op_920_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 64, 1]> var_920_cast_fp16 = reduce_mean(axes = var_920_axes_0, keep_dims = var_920_keep_dims_0, x = var_915_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<fp16, [1, 16, 64, 1]> var_921_cast_fp16 = mul(x = clip_10_cast_fp16, y = clip_10_cast_fp16)[name = string("op_921_cast_fp16")];
            fp32 var_922_epsilon_0 = const()[name = string("op_922_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 64, 1]> var_922_cast_fp16 = inverse(epsilon = var_922_epsilon_0, x = var_921_cast_fp16)[name = string("op_922_cast_fp16")];
            fp16 var_923_to_fp16 = const()[name = string("op_923_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 64, 1]> var_924_cast_fp16 = mul(x = var_922_cast_fp16, y = var_923_to_fp16)[name = string("op_924_cast_fp16")];
            tensor<fp16, [1, 16, 64, 1]> var_926_cast_fp16 = add(x = var_920_cast_fp16, y = var_924_cast_fp16)[name = string("op_926_cast_fp16")];
            fp32 var_927_epsilon_0 = const()[name = string("op_927_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 64, 1]> var_927_cast_fp16 = rsqrt(epsilon = var_927_epsilon_0, x = var_926_cast_fp16)[name = string("op_927_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_928_cast_fp16 = mul(x = y_19_cast_fp16, y = var_927_cast_fp16)[name = string("op_928_cast_fp16")];
            tensor<fp16, [128]> layers_2_q_norm_to_fp16 = const()[name = string("layers_2_q_norm_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109184960)))];
            tensor<fp16, [1, 16, 64, 128]> q_9_cast_fp16 = mul(x = var_928_cast_fp16, y = layers_2_q_norm_to_fp16)[name = string("q_9_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> x_19_cast_fp16 = transpose(perm = x_19_perm_0, x = var_890_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 8, 64, 128]> var_930_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_930_cast_fp16")];
            bool var_933 = const()[name = string("op_933"), val = bool(true)];
            tensor<int32, [1]> var_934_axes_0 = const()[name = string("op_934_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 64, 1]> var_934_cast_fp16 = reduce_max(axes = var_934_axes_0, keep_dims = var_933, x = var_930_cast_fp16)[name = string("op_934_cast_fp16")];
            fp16 var_935_to_fp16 = const()[name = string("op_935_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_15_to_fp16 = const()[name = string("const_15_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 8, 64, 1]> clip_11_cast_fp16 = clip(alpha = var_935_to_fp16, beta = const_15_to_fp16, x = var_934_cast_fp16)[name = string("clip_11_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> y_21_cast_fp16 = real_div(x = x_19_cast_fp16, y = clip_11_cast_fp16)[name = string("y_21_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_939_cast_fp16 = mul(x = y_21_cast_fp16, y = y_21_cast_fp16)[name = string("op_939_cast_fp16")];
            tensor<int32, [1]> var_944_axes_0 = const()[name = string("op_944_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_944_keep_dims_0 = const()[name = string("op_944_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 64, 1]> var_944_cast_fp16 = reduce_mean(axes = var_944_axes_0, keep_dims = var_944_keep_dims_0, x = var_939_cast_fp16)[name = string("op_944_cast_fp16")];
            tensor<fp16, [1, 8, 64, 1]> var_945_cast_fp16 = mul(x = clip_11_cast_fp16, y = clip_11_cast_fp16)[name = string("op_945_cast_fp16")];
            fp32 var_946_epsilon_0 = const()[name = string("op_946_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 64, 1]> var_946_cast_fp16 = inverse(epsilon = var_946_epsilon_0, x = var_945_cast_fp16)[name = string("op_946_cast_fp16")];
            fp16 var_947_to_fp16 = const()[name = string("op_947_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 64, 1]> var_948_cast_fp16 = mul(x = var_946_cast_fp16, y = var_947_to_fp16)[name = string("op_948_cast_fp16")];
            tensor<fp16, [1, 8, 64, 1]> var_950_cast_fp16 = add(x = var_944_cast_fp16, y = var_948_cast_fp16)[name = string("op_950_cast_fp16")];
            fp32 var_951_epsilon_0 = const()[name = string("op_951_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 64, 1]> var_951_cast_fp16 = rsqrt(epsilon = var_951_epsilon_0, x = var_950_cast_fp16)[name = string("op_951_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_952_cast_fp16 = mul(x = y_21_cast_fp16, y = var_951_cast_fp16)[name = string("op_952_cast_fp16")];
            tensor<fp16, [128]> layers_2_k_norm_to_fp16 = const()[name = string("layers_2_k_norm_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109185280)))];
            tensor<fp16, [1, 8, 64, 128]> k_9_cast_fp16 = mul(x = var_952_cast_fp16, y = layers_2_k_norm_to_fp16)[name = string("k_9_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_954_cast_fp16 = mul(x = q_9_cast_fp16, y = cos)[name = string("op_954_cast_fp16")];
            tensor<int32, [2]> var_955_split_sizes_0 = const()[name = string("op_955_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_955_axis_0 = const()[name = string("op_955_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 64, 64]> var_955_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_955_cast_fp16_1 = split(axis = var_955_axis_0, split_sizes = var_955_split_sizes_0, x = q_9_cast_fp16)[name = string("op_955_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_957_cast_fp16 = mul(x = var_955_cast_fp16_1, y = const_16_promoted_to_fp16)[name = string("op_957_cast_fp16")];
            int32 var_959 = const()[name = string("op_959"), val = int32(-1)];
            bool var_960_interleave_0 = const()[name = string("op_960_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> var_960_cast_fp16 = concat(axis = var_959, interleave = var_960_interleave_0, values = (var_957_cast_fp16, var_955_cast_fp16_0))[name = string("op_960_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_961_cast_fp16 = mul(x = var_960_cast_fp16, y = sin)[name = string("op_961_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> q_11_cast_fp16 = add(x = var_954_cast_fp16, y = var_961_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_964_cast_fp16 = mul(x = k_9_cast_fp16, y = cos)[name = string("op_964_cast_fp16")];
            tensor<int32, [2]> var_965_split_sizes_0 = const()[name = string("op_965_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_965_axis_0 = const()[name = string("op_965_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 64, 64]> var_965_cast_fp16_0, tensor<fp16, [1, 8, 64, 64]> var_965_cast_fp16_1 = split(axis = var_965_axis_0, split_sizes = var_965_split_sizes_0, x = k_9_cast_fp16)[name = string("op_965_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 64, 64]> var_967_cast_fp16 = mul(x = var_965_cast_fp16_1, y = const_17_promoted_to_fp16)[name = string("op_967_cast_fp16")];
            int32 var_969 = const()[name = string("op_969"), val = int32(-1)];
            bool var_970_interleave_0 = const()[name = string("op_970_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_970_cast_fp16 = concat(axis = var_969, interleave = var_970_interleave_0, values = (var_967_cast_fp16, var_965_cast_fp16_0))[name = string("op_970_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_971_cast_fp16 = mul(x = var_970_cast_fp16, y = sin)[name = string("op_971_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> k_11_cast_fp16 = add(x = var_964_cast_fp16, y = var_971_cast_fp16)[name = string("k_11_cast_fp16")];
            tensor<int32, [4]> var_978_begin_0 = const()[name = string("op_978_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_978_end_0 = const()[name = string("op_978_end_0"), val = tensor<int32, [4]>([3, 8, 704, 128])];
            tensor<bool, [4]> var_978_end_mask_0 = const()[name = string("op_978_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_978_cast_fp16 = slice_by_index(begin = var_978_begin_0, end = var_978_end_0, end_mask = var_978_end_mask_0, x = coreml_update_state_16)[name = string("op_978_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_990_cast_fp16 = mul(x = var_978_cast_fp16, y = var_393_cast_fp16)[name = string("op_990_cast_fp16")];
            tensor<int32, [4]> var_994_begin_0 = const()[name = string("op_994_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_994_end_0 = const()[name = string("op_994_end_0"), val = tensor<int32, [4]>([1, 8, 64, 128])];
            tensor<bool, [4]> var_994_end_mask_0 = const()[name = string("op_994_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_994_squeeze_mask_0 = const()[name = string("op_994_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [8, 64, 128]> var_994_cast_fp16 = slice_by_index(begin = var_994_begin_0, end = var_994_end_0, end_mask = var_994_end_mask_0, squeeze_mask = var_994_squeeze_mask_0, x = k_11_cast_fp16)[name = string("op_994_cast_fp16")];
            bool var_995_transpose_x_1 = const()[name = string("op_995_transpose_x_1"), val = bool(true)];
            bool var_995_transpose_y_1 = const()[name = string("op_995_transpose_y_1"), val = bool(false)];
            tensor<fp16, [8, 704, 128]> var_995_cast_fp16 = matmul(transpose_x = var_995_transpose_x_1, transpose_y = var_995_transpose_y_1, x = onehot, y = var_994_cast_fp16)[name = string("op_995_cast_fp16")];
            tensor<int32, [1]> var_997_axes_0 = const()[name = string("op_997_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8, 704, 128]> var_997_cast_fp16 = expand_dims(axes = var_997_axes_0, x = var_995_cast_fp16)[name = string("op_997_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_999_cast_fp16 = add(x = var_990_cast_fp16, y = var_997_cast_fp16)[name = string("op_999_cast_fp16")];
            tensor<int32, [4]> concat_16 = const()[name = string("concat_16"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> concat_17 = const()[name = string("concat_17"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> k_cache_internal_tensor_assign_3_stride_0 = const()[name = string("k_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("k_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("k_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("k_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [7, 8, 704, 128]> k_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_16, begin_mask = k_cache_internal_tensor_assign_3_begin_mask_0, end = concat_17, end_mask = k_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = k_cache_internal_tensor_assign_3_squeeze_mask_0, stride = k_cache_internal_tensor_assign_3_stride_0, update = var_999_cast_fp16, x = coreml_update_state_16)[name = string("k_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = k_cache_internal_tensor_assign_3_cast_fp16, input = k_cache)[name = string("coreml_update_state_18_write_state")];
            tensor<fp16, [7, 8, 704, 128]> coreml_update_state_18 = read_state(input = k_cache)[name = string("coreml_update_state_18")];
            tensor<int32, [4]> var_1016_begin_0 = const()[name = string("op_1016_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1016_end_0 = const()[name = string("op_1016_end_0"), val = tensor<int32, [4]>([3, 8, 704, 128])];
            tensor<bool, [4]> var_1016_end_mask_0 = const()[name = string("op_1016_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_1016_cast_fp16 = slice_by_index(begin = var_1016_begin_0, end = var_1016_end_0, end_mask = var_1016_end_mask_0, x = coreml_update_state_17)[name = string("op_1016_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_1028_cast_fp16 = mul(x = var_1016_cast_fp16, y = var_393_cast_fp16)[name = string("op_1028_cast_fp16")];
            tensor<int32, [4]> var_1032_begin_0 = const()[name = string("op_1032_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1032_end_0 = const()[name = string("op_1032_end_0"), val = tensor<int32, [4]>([1, 8, 64, 128])];
            tensor<bool, [4]> var_1032_end_mask_0 = const()[name = string("op_1032_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_1032_squeeze_mask_0 = const()[name = string("op_1032_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 8, 64, 128]> v_5_cast_fp16 = transpose(perm = v_5_perm_0, x = var_902_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [8, 64, 128]> var_1032_cast_fp16 = slice_by_index(begin = var_1032_begin_0, end = var_1032_end_0, end_mask = var_1032_end_mask_0, squeeze_mask = var_1032_squeeze_mask_0, x = v_5_cast_fp16)[name = string("op_1032_cast_fp16")];
            bool var_1033_transpose_x_1 = const()[name = string("op_1033_transpose_x_1"), val = bool(true)];
            bool var_1033_transpose_y_1 = const()[name = string("op_1033_transpose_y_1"), val = bool(false)];
            tensor<fp16, [8, 704, 128]> var_1033_cast_fp16 = matmul(transpose_x = var_1033_transpose_x_1, transpose_y = var_1033_transpose_y_1, x = onehot, y = var_1032_cast_fp16)[name = string("op_1033_cast_fp16")];
            tensor<int32, [1]> var_1035_axes_0 = const()[name = string("op_1035_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8, 704, 128]> var_1035_cast_fp16 = expand_dims(axes = var_1035_axes_0, x = var_1033_cast_fp16)[name = string("op_1035_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_1037_cast_fp16 = add(x = var_1028_cast_fp16, y = var_1035_cast_fp16)[name = string("op_1037_cast_fp16")];
            tensor<int32, [4]> concat_18 = const()[name = string("concat_18"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> concat_19 = const()[name = string("concat_19"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> v_cache_internal_tensor_assign_3_stride_0 = const()[name = string("v_cache_internal_tensor_assign_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_3_begin_mask_0 = const()[name = string("v_cache_internal_tensor_assign_3_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_3_end_mask_0 = const()[name = string("v_cache_internal_tensor_assign_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_3_squeeze_mask_0 = const()[name = string("v_cache_internal_tensor_assign_3_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [7, 8, 704, 128]> v_cache_internal_tensor_assign_3_cast_fp16 = slice_update(begin = concat_18, begin_mask = v_cache_internal_tensor_assign_3_begin_mask_0, end = concat_19, end_mask = v_cache_internal_tensor_assign_3_end_mask_0, squeeze_mask = v_cache_internal_tensor_assign_3_squeeze_mask_0, stride = v_cache_internal_tensor_assign_3_stride_0, update = var_1037_cast_fp16, x = coreml_update_state_17)[name = string("v_cache_internal_tensor_assign_3_cast_fp16")];
            write_state(data = v_cache_internal_tensor_assign_3_cast_fp16, input = v_cache)[name = string("coreml_update_state_19_write_state")];
            tensor<fp16, [7, 8, 704, 128]> coreml_update_state_19 = read_state(input = v_cache)[name = string("coreml_update_state_19")];
            tensor<int32, [4]> var_1054_begin_0 = const()[name = string("op_1054_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1054_end_0 = const()[name = string("op_1054_end_0"), val = tensor<int32, [4]>([3, 8, 704, 128])];
            tensor<bool, [4]> var_1054_end_mask_0 = const()[name = string("op_1054_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_1054_cast_fp16 = slice_by_index(begin = var_1054_begin_0, end = var_1054_end_0, end_mask = var_1054_end_mask_0, x = coreml_update_state_18)[name = string("op_1054_cast_fp16")];
            tensor<int32, [4]> transpose_8_perm_0 = const()[name = string("transpose_8_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_4_reps_0 = const()[name = string("tile_4_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [8, 1, 704, 128]> transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = var_1054_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [16, 1, 704, 128]> tile_4_cast_fp16 = tile(reps = tile_4_reps_0, x = transpose_8_cast_fp16)[name = string("tile_4_cast_fp16")];
            tensor<int32, [5]> concat_20 = const()[name = string("concat_20"), val = tensor<int32, [5]>([2, 8, 1, 704, 128])];
            tensor<fp16, [2, 8, 1, 704, 128]> reshape_8_cast_fp16 = reshape(shape = concat_20, x = tile_4_cast_fp16)[name = string("reshape_8_cast_fp16")];
            tensor<int32, [5]> transpose_9_perm_0 = const()[name = string("transpose_9_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_21 = const()[name = string("concat_21"), val = tensor<int32, [4]>([-1, 1, 704, 128])];
            tensor<fp16, [8, 2, 1, 704, 128]> transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = reshape_8_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [16, 1, 704, 128]> reshape_9_cast_fp16 = reshape(shape = concat_21, x = transpose_9_cast_fp16)[name = string("reshape_9_cast_fp16")];
            tensor<int32, [4]> transpose_30_perm_0 = const()[name = string("transpose_30_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> var_1068_begin_0 = const()[name = string("op_1068_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_1068_end_0 = const()[name = string("op_1068_end_0"), val = tensor<int32, [4]>([3, 8, 704, 128])];
            tensor<bool, [4]> var_1068_end_mask_0 = const()[name = string("op_1068_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_1068_cast_fp16 = slice_by_index(begin = var_1068_begin_0, end = var_1068_end_0, end_mask = var_1068_end_mask_0, x = coreml_update_state_19)[name = string("op_1068_cast_fp16")];
            tensor<int32, [4]> transpose_10_perm_0 = const()[name = string("transpose_10_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_5_reps_0 = const()[name = string("tile_5_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [8, 1, 704, 128]> transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = var_1068_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [16, 1, 704, 128]> tile_5_cast_fp16 = tile(reps = tile_5_reps_0, x = transpose_10_cast_fp16)[name = string("tile_5_cast_fp16")];
            tensor<int32, [5]> concat_22 = const()[name = string("concat_22"), val = tensor<int32, [5]>([2, 8, 1, 704, 128])];
            tensor<fp16, [2, 8, 1, 704, 128]> reshape_10_cast_fp16 = reshape(shape = concat_22, x = tile_5_cast_fp16)[name = string("reshape_10_cast_fp16")];
            tensor<int32, [5]> transpose_11_perm_0 = const()[name = string("transpose_11_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_23 = const()[name = string("concat_23"), val = tensor<int32, [4]>([-1, 1, 704, 128])];
            tensor<fp16, [8, 2, 1, 704, 128]> transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = reshape_10_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [16, 1, 704, 128]> reshape_11_cast_fp16 = reshape(shape = concat_23, x = transpose_11_cast_fp16)[name = string("reshape_11_cast_fp16")];
            tensor<int32, [4]> values_5_perm_0 = const()[name = string("values_5_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_1081_transpose_x_0 = const()[name = string("op_1081_transpose_x_0"), val = bool(false)];
            bool var_1081_transpose_y_0 = const()[name = string("op_1081_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 704]> transpose_30_cast_fp16 = transpose(perm = transpose_30_perm_0, x = reshape_9_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 16, 64, 704]> var_1081_cast_fp16 = matmul(transpose_x = var_1081_transpose_x_0, transpose_y = var_1081_transpose_y_0, x = q_11_cast_fp16, y = transpose_30_cast_fp16)[name = string("op_1081_cast_fp16")];
            fp16 var_1082_to_fp16 = const()[name = string("op_1082_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 64, 704]> var_1083_cast_fp16 = mul(x = var_1081_cast_fp16, y = var_1082_to_fp16)[name = string("op_1083_cast_fp16")];
            tensor<fp16, [1, 16, 64, 704]> scores_5_cast_fp16 = add(x = var_1083_cast_fp16, y = mask)[name = string("scores_5_cast_fp16")];
            int32 var_1086 = const()[name = string("op_1086"), val = int32(-1)];
            tensor<fp16, [1, 16, 64, 704]> var_1088_cast_fp16 = softmax(axis = var_1086, x = scores_5_cast_fp16)[name = string("op_1088_cast_fp16")];
            bool attn_5_transpose_x_0 = const()[name = string("attn_5_transpose_x_0"), val = bool(false)];
            bool attn_5_transpose_y_0 = const()[name = string("attn_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 704, 128]> values_5_cast_fp16 = transpose(perm = values_5_perm_0, x = reshape_11_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 16, 64, 128]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_1088_cast_fp16, y = values_5_cast_fp16)[name = string("attn_5_cast_fp16")];
            tensor<int32, [4]> var_1092_perm_0 = const()[name = string("op_1092_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1096 = const()[name = string("op_1096"), val = tensor<int32, [3]>([1, 64, 2048])];
            tensor<fp16, [1, 64, 16, 128]> var_1092_cast_fp16 = transpose(perm = var_1092_perm_0, x = attn_5_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 64, 2048]> input_23_cast_fp16 = reshape(shape = var_1096, x = var_1092_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [2048, 2048]> layers_2_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109185600))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113379968))))[name = string("layers_2_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_17_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_o_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> x_21_cast_fp16 = add(x = x_15_cast_fp16, y = linear_17_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1103_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_1103_cast_fp16")];
            bool var_1106 = const()[name = string("op_1106"), val = bool(true)];
            tensor<int32, [1]> var_1107_axes_0 = const()[name = string("op_1107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1107_cast_fp16 = reduce_max(axes = var_1107_axes_0, keep_dims = var_1106, x = var_1103_cast_fp16)[name = string("op_1107_cast_fp16")];
            fp16 var_1108_to_fp16 = const()[name = string("op_1108_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_18_to_fp16 = const()[name = string("const_18_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_12_cast_fp16 = clip(alpha = var_1108_to_fp16, beta = const_18_to_fp16, x = var_1107_cast_fp16)[name = string("clip_12_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> y_23_cast_fp16 = real_div(x = x_21_cast_fp16, y = clip_12_cast_fp16)[name = string("y_23_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1112_cast_fp16 = mul(x = y_23_cast_fp16, y = y_23_cast_fp16)[name = string("op_1112_cast_fp16")];
            tensor<int32, [1]> var_1117_axes_0 = const()[name = string("op_1117_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1117_keep_dims_0 = const()[name = string("op_1117_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_1117_cast_fp16 = reduce_mean(axes = var_1117_axes_0, keep_dims = var_1117_keep_dims_0, x = var_1112_cast_fp16)[name = string("op_1117_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1118_cast_fp16 = mul(x = clip_12_cast_fp16, y = clip_12_cast_fp16)[name = string("op_1118_cast_fp16")];
            fp32 var_1119_epsilon_0 = const()[name = string("op_1119_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1119_cast_fp16 = inverse(epsilon = var_1119_epsilon_0, x = var_1118_cast_fp16)[name = string("op_1119_cast_fp16")];
            fp16 var_1120_to_fp16 = const()[name = string("op_1120_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1121_cast_fp16 = mul(x = var_1119_cast_fp16, y = var_1120_to_fp16)[name = string("op_1121_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1123_cast_fp16 = add(x = var_1117_cast_fp16, y = var_1121_cast_fp16)[name = string("op_1123_cast_fp16")];
            fp32 var_1124_epsilon_0 = const()[name = string("op_1124_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1124_cast_fp16 = rsqrt(epsilon = var_1124_epsilon_0, x = var_1123_cast_fp16)[name = string("op_1124_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1125_cast_fp16 = mul(x = y_23_cast_fp16, y = var_1124_cast_fp16)[name = string("op_1125_cast_fp16")];
            tensor<fp16, [2048]> layers_2_post_norm_to_fp16 = const()[name = string("layers_2_post_norm_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113384128)))];
            tensor<fp16, [1, 64, 2048]> input_25_cast_fp16 = mul(x = var_1125_cast_fp16, y = layers_2_post_norm_to_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [6144, 2048]> layers_2_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113388288))), scale = tensor<fp16, [6144, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125971264))))[name = string("layers_2_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 6144]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_2_gate_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 64, 6144]> var_1130_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_1130_cast_fp16")];
            tensor<fp16, [6144, 2048]> layers_2_up_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125983616))), scale = tensor<fp16, [6144, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138566592))))[name = string("layers_2_up_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 6144]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_2_up_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 64, 6144]> var_1134_cast_fp16 = mul(x = var_1130_cast_fp16, y = linear_19_cast_fp16)[name = string("op_1134_cast_fp16")];
            fp16 var_1135_to_fp16 = const()[name = string("op_1135_to_fp16"), val = fp16(0x1p-8)];
            tensor<fp16, [1, 64, 6144]> input_29_cast_fp16 = mul(x = var_1134_cast_fp16, y = var_1135_to_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [2048, 6144]> layers_2_down_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138578944))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151161920))))[name = string("layers_2_down_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_down_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_20_cast_fp16")];
            fp16 var_1140_to_fp16 = const()[name = string("op_1140_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 64, 2048]> var_1141_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_1140_to_fp16)[name = string("op_1141_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> x_23_cast_fp16 = add(x = x_21_cast_fp16, y = var_1141_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1144_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_1144_cast_fp16")];
            bool var_1147 = const()[name = string("op_1147"), val = bool(true)];
            tensor<int32, [1]> var_1148_axes_0 = const()[name = string("op_1148_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1148_cast_fp16 = reduce_max(axes = var_1148_axes_0, keep_dims = var_1147, x = var_1144_cast_fp16)[name = string("op_1148_cast_fp16")];
            fp16 var_1149_to_fp16 = const()[name = string("op_1149_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_19_to_fp16 = const()[name = string("const_19_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_13_cast_fp16 = clip(alpha = var_1149_to_fp16, beta = const_19_to_fp16, x = var_1148_cast_fp16)[name = string("clip_13_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> y_25_cast_fp16 = real_div(x = x_23_cast_fp16, y = clip_13_cast_fp16)[name = string("y_25_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1153_cast_fp16 = mul(x = y_25_cast_fp16, y = y_25_cast_fp16)[name = string("op_1153_cast_fp16")];
            tensor<int32, [1]> var_1158_axes_0 = const()[name = string("op_1158_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1158_keep_dims_0 = const()[name = string("op_1158_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_1158_cast_fp16 = reduce_mean(axes = var_1158_axes_0, keep_dims = var_1158_keep_dims_0, x = var_1153_cast_fp16)[name = string("op_1158_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1159_cast_fp16 = mul(x = clip_13_cast_fp16, y = clip_13_cast_fp16)[name = string("op_1159_cast_fp16")];
            fp32 var_1160_epsilon_0 = const()[name = string("op_1160_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1160_cast_fp16 = inverse(epsilon = var_1160_epsilon_0, x = var_1159_cast_fp16)[name = string("op_1160_cast_fp16")];
            fp16 var_1161_to_fp16 = const()[name = string("op_1161_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1162_cast_fp16 = mul(x = var_1160_cast_fp16, y = var_1161_to_fp16)[name = string("op_1162_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1164_cast_fp16 = add(x = var_1158_cast_fp16, y = var_1162_cast_fp16)[name = string("op_1164_cast_fp16")];
            fp32 var_1165_epsilon_0 = const()[name = string("op_1165_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1165_cast_fp16 = rsqrt(epsilon = var_1165_epsilon_0, x = var_1164_cast_fp16)[name = string("op_1165_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1166_cast_fp16 = mul(x = y_25_cast_fp16, y = var_1165_cast_fp16)[name = string("op_1166_cast_fp16")];
            tensor<fp16, [2048]> layers_3_input_norm_to_fp16 = const()[name = string("layers_3_input_norm_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151166080)))];
            tensor<fp16, [1, 64, 2048]> input_31_cast_fp16 = mul(x = var_1166_cast_fp16, y = layers_3_input_norm_to_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [2048, 2048]> layers_3_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151170240))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(155364608))))[name = string("layers_3_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_21_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_3_q_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_1175 = const()[name = string("op_1175"), val = tensor<int32, [4]>([1, 64, 16, 128])];
            tensor<fp16, [1, 64, 16, 128]> var_1176_cast_fp16 = reshape(shape = var_1175, x = linear_21_cast_fp16)[name = string("op_1176_cast_fp16")];
            tensor<int32, [4]> x_25_perm_0 = const()[name = string("x_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 2048]> layers_3_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(155368768))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157465984))))[name = string("layers_3_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_22_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_k_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_1187 = const()[name = string("op_1187"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<fp16, [1, 64, 8, 128]> var_1188_cast_fp16 = reshape(shape = var_1187, x = linear_22_cast_fp16)[name = string("op_1188_cast_fp16")];
            tensor<int32, [4]> x_27_perm_0 = const()[name = string("x_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 2048]> layers_3_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157468096))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159565312))))[name = string("layers_3_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_v_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_1199 = const()[name = string("op_1199"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<fp16, [1, 64, 8, 128]> var_1200_cast_fp16 = reshape(shape = var_1199, x = linear_23_cast_fp16)[name = string("op_1200_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 64, 128]> x_25_cast_fp16 = transpose(perm = x_25_perm_0, x = var_1176_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 16, 64, 128]> var_1204_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_1204_cast_fp16")];
            bool var_1207 = const()[name = string("op_1207"), val = bool(true)];
            tensor<int32, [1]> var_1208_axes_0 = const()[name = string("op_1208_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 64, 1]> var_1208_cast_fp16 = reduce_max(axes = var_1208_axes_0, keep_dims = var_1207, x = var_1204_cast_fp16)[name = string("op_1208_cast_fp16")];
            fp16 var_1209_to_fp16 = const()[name = string("op_1209_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_20_to_fp16 = const()[name = string("const_20_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 16, 64, 1]> clip_14_cast_fp16 = clip(alpha = var_1209_to_fp16, beta = const_20_to_fp16, x = var_1208_cast_fp16)[name = string("clip_14_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> y_27_cast_fp16 = real_div(x = x_25_cast_fp16, y = clip_14_cast_fp16)[name = string("y_27_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_1213_cast_fp16 = mul(x = y_27_cast_fp16, y = y_27_cast_fp16)[name = string("op_1213_cast_fp16")];
            tensor<int32, [1]> var_1218_axes_0 = const()[name = string("op_1218_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1218_keep_dims_0 = const()[name = string("op_1218_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 64, 1]> var_1218_cast_fp16 = reduce_mean(axes = var_1218_axes_0, keep_dims = var_1218_keep_dims_0, x = var_1213_cast_fp16)[name = string("op_1218_cast_fp16")];
            tensor<fp16, [1, 16, 64, 1]> var_1219_cast_fp16 = mul(x = clip_14_cast_fp16, y = clip_14_cast_fp16)[name = string("op_1219_cast_fp16")];
            fp32 var_1220_epsilon_0 = const()[name = string("op_1220_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 64, 1]> var_1220_cast_fp16 = inverse(epsilon = var_1220_epsilon_0, x = var_1219_cast_fp16)[name = string("op_1220_cast_fp16")];
            fp16 var_1221_to_fp16 = const()[name = string("op_1221_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 64, 1]> var_1222_cast_fp16 = mul(x = var_1220_cast_fp16, y = var_1221_to_fp16)[name = string("op_1222_cast_fp16")];
            tensor<fp16, [1, 16, 64, 1]> var_1224_cast_fp16 = add(x = var_1218_cast_fp16, y = var_1222_cast_fp16)[name = string("op_1224_cast_fp16")];
            fp32 var_1225_epsilon_0 = const()[name = string("op_1225_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 64, 1]> var_1225_cast_fp16 = rsqrt(epsilon = var_1225_epsilon_0, x = var_1224_cast_fp16)[name = string("op_1225_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_1226_cast_fp16 = mul(x = y_27_cast_fp16, y = var_1225_cast_fp16)[name = string("op_1226_cast_fp16")];
            tensor<fp16, [128]> layers_3_q_norm_to_fp16 = const()[name = string("layers_3_q_norm_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159567424)))];
            tensor<fp16, [1, 16, 64, 128]> q_13_cast_fp16 = mul(x = var_1226_cast_fp16, y = layers_3_q_norm_to_fp16)[name = string("q_13_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> x_27_cast_fp16 = transpose(perm = x_27_perm_0, x = var_1188_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 8, 64, 128]> var_1228_cast_fp16 = abs(x = x_27_cast_fp16)[name = string("op_1228_cast_fp16")];
            bool var_1231 = const()[name = string("op_1231"), val = bool(true)];
            tensor<int32, [1]> var_1232_axes_0 = const()[name = string("op_1232_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 64, 1]> var_1232_cast_fp16 = reduce_max(axes = var_1232_axes_0, keep_dims = var_1231, x = var_1228_cast_fp16)[name = string("op_1232_cast_fp16")];
            fp16 var_1233_to_fp16 = const()[name = string("op_1233_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_21_to_fp16 = const()[name = string("const_21_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 8, 64, 1]> clip_15_cast_fp16 = clip(alpha = var_1233_to_fp16, beta = const_21_to_fp16, x = var_1232_cast_fp16)[name = string("clip_15_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> y_29_cast_fp16 = real_div(x = x_27_cast_fp16, y = clip_15_cast_fp16)[name = string("y_29_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_1237_cast_fp16 = mul(x = y_29_cast_fp16, y = y_29_cast_fp16)[name = string("op_1237_cast_fp16")];
            tensor<int32, [1]> var_1242_axes_0 = const()[name = string("op_1242_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1242_keep_dims_0 = const()[name = string("op_1242_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 64, 1]> var_1242_cast_fp16 = reduce_mean(axes = var_1242_axes_0, keep_dims = var_1242_keep_dims_0, x = var_1237_cast_fp16)[name = string("op_1242_cast_fp16")];
            tensor<fp16, [1, 8, 64, 1]> var_1243_cast_fp16 = mul(x = clip_15_cast_fp16, y = clip_15_cast_fp16)[name = string("op_1243_cast_fp16")];
            fp32 var_1244_epsilon_0 = const()[name = string("op_1244_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 64, 1]> var_1244_cast_fp16 = inverse(epsilon = var_1244_epsilon_0, x = var_1243_cast_fp16)[name = string("op_1244_cast_fp16")];
            fp16 var_1245_to_fp16 = const()[name = string("op_1245_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 64, 1]> var_1246_cast_fp16 = mul(x = var_1244_cast_fp16, y = var_1245_to_fp16)[name = string("op_1246_cast_fp16")];
            tensor<fp16, [1, 8, 64, 1]> var_1248_cast_fp16 = add(x = var_1242_cast_fp16, y = var_1246_cast_fp16)[name = string("op_1248_cast_fp16")];
            fp32 var_1249_epsilon_0 = const()[name = string("op_1249_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 64, 1]> var_1249_cast_fp16 = rsqrt(epsilon = var_1249_epsilon_0, x = var_1248_cast_fp16)[name = string("op_1249_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_1250_cast_fp16 = mul(x = y_29_cast_fp16, y = var_1249_cast_fp16)[name = string("op_1250_cast_fp16")];
            tensor<fp16, [128]> layers_3_k_norm_to_fp16 = const()[name = string("layers_3_k_norm_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159567744)))];
            tensor<fp16, [1, 8, 64, 128]> k_13_cast_fp16 = mul(x = var_1250_cast_fp16, y = layers_3_k_norm_to_fp16)[name = string("k_13_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_1252_cast_fp16 = mul(x = q_13_cast_fp16, y = cos)[name = string("op_1252_cast_fp16")];
            tensor<int32, [2]> var_1253_split_sizes_0 = const()[name = string("op_1253_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1253_axis_0 = const()[name = string("op_1253_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 64, 64]> var_1253_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_1253_cast_fp16_1 = split(axis = var_1253_axis_0, split_sizes = var_1253_split_sizes_0, x = q_13_cast_fp16)[name = string("op_1253_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_1255_cast_fp16 = mul(x = var_1253_cast_fp16_1, y = const_22_promoted_to_fp16)[name = string("op_1255_cast_fp16")];
            int32 var_1257 = const()[name = string("op_1257"), val = int32(-1)];
            bool var_1258_interleave_0 = const()[name = string("op_1258_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> var_1258_cast_fp16 = concat(axis = var_1257, interleave = var_1258_interleave_0, values = (var_1255_cast_fp16, var_1253_cast_fp16_0))[name = string("op_1258_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_1259_cast_fp16 = mul(x = var_1258_cast_fp16, y = sin)[name = string("op_1259_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> q_15_cast_fp16 = add(x = var_1252_cast_fp16, y = var_1259_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_1262_cast_fp16 = mul(x = k_13_cast_fp16, y = cos)[name = string("op_1262_cast_fp16")];
            tensor<int32, [2]> var_1263_split_sizes_0 = const()[name = string("op_1263_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1263_axis_0 = const()[name = string("op_1263_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 64, 64]> var_1263_cast_fp16_0, tensor<fp16, [1, 8, 64, 64]> var_1263_cast_fp16_1 = split(axis = var_1263_axis_0, split_sizes = var_1263_split_sizes_0, x = k_13_cast_fp16)[name = string("op_1263_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 64, 64]> var_1265_cast_fp16 = mul(x = var_1263_cast_fp16_1, y = const_23_promoted_to_fp16)[name = string("op_1265_cast_fp16")];
            int32 var_1267 = const()[name = string("op_1267"), val = int32(-1)];
            bool var_1268_interleave_0 = const()[name = string("op_1268_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_1268_cast_fp16 = concat(axis = var_1267, interleave = var_1268_interleave_0, values = (var_1265_cast_fp16, var_1263_cast_fp16_0))[name = string("op_1268_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_1269_cast_fp16 = mul(x = var_1268_cast_fp16, y = sin)[name = string("op_1269_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> k_15_cast_fp16 = add(x = var_1262_cast_fp16, y = var_1269_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> var_1276_begin_0 = const()[name = string("op_1276_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1276_end_0 = const()[name = string("op_1276_end_0"), val = tensor<int32, [4]>([4, 8, 704, 128])];
            tensor<bool, [4]> var_1276_end_mask_0 = const()[name = string("op_1276_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_1276_cast_fp16 = slice_by_index(begin = var_1276_begin_0, end = var_1276_end_0, end_mask = var_1276_end_mask_0, x = coreml_update_state_18)[name = string("op_1276_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_1288_cast_fp16 = mul(x = var_1276_cast_fp16, y = var_393_cast_fp16)[name = string("op_1288_cast_fp16")];
            tensor<int32, [4]> var_1292_begin_0 = const()[name = string("op_1292_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1292_end_0 = const()[name = string("op_1292_end_0"), val = tensor<int32, [4]>([1, 8, 64, 128])];
            tensor<bool, [4]> var_1292_end_mask_0 = const()[name = string("op_1292_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_1292_squeeze_mask_0 = const()[name = string("op_1292_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [8, 64, 128]> var_1292_cast_fp16 = slice_by_index(begin = var_1292_begin_0, end = var_1292_end_0, end_mask = var_1292_end_mask_0, squeeze_mask = var_1292_squeeze_mask_0, x = k_15_cast_fp16)[name = string("op_1292_cast_fp16")];
            bool var_1293_transpose_x_1 = const()[name = string("op_1293_transpose_x_1"), val = bool(true)];
            bool var_1293_transpose_y_1 = const()[name = string("op_1293_transpose_y_1"), val = bool(false)];
            tensor<fp16, [8, 704, 128]> var_1293_cast_fp16 = matmul(transpose_x = var_1293_transpose_x_1, transpose_y = var_1293_transpose_y_1, x = onehot, y = var_1292_cast_fp16)[name = string("op_1293_cast_fp16")];
            tensor<int32, [1]> var_1295_axes_0 = const()[name = string("op_1295_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8, 704, 128]> var_1295_cast_fp16 = expand_dims(axes = var_1295_axes_0, x = var_1293_cast_fp16)[name = string("op_1295_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_1297_cast_fp16 = add(x = var_1288_cast_fp16, y = var_1295_cast_fp16)[name = string("op_1297_cast_fp16")];
            tensor<int32, [4]> concat_24 = const()[name = string("concat_24"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> concat_25 = const()[name = string("concat_25"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> k_cache_internal_tensor_assign_4_stride_0 = const()[name = string("k_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("k_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("k_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("k_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [7, 8, 704, 128]> k_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_24, begin_mask = k_cache_internal_tensor_assign_4_begin_mask_0, end = concat_25, end_mask = k_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = k_cache_internal_tensor_assign_4_squeeze_mask_0, stride = k_cache_internal_tensor_assign_4_stride_0, update = var_1297_cast_fp16, x = coreml_update_state_18)[name = string("k_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = k_cache_internal_tensor_assign_4_cast_fp16, input = k_cache)[name = string("coreml_update_state_20_write_state")];
            tensor<fp16, [7, 8, 704, 128]> coreml_update_state_20 = read_state(input = k_cache)[name = string("coreml_update_state_20")];
            tensor<int32, [4]> var_1314_begin_0 = const()[name = string("op_1314_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1314_end_0 = const()[name = string("op_1314_end_0"), val = tensor<int32, [4]>([4, 8, 704, 128])];
            tensor<bool, [4]> var_1314_end_mask_0 = const()[name = string("op_1314_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_1314_cast_fp16 = slice_by_index(begin = var_1314_begin_0, end = var_1314_end_0, end_mask = var_1314_end_mask_0, x = coreml_update_state_19)[name = string("op_1314_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_1326_cast_fp16 = mul(x = var_1314_cast_fp16, y = var_393_cast_fp16)[name = string("op_1326_cast_fp16")];
            tensor<int32, [4]> var_1330_begin_0 = const()[name = string("op_1330_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1330_end_0 = const()[name = string("op_1330_end_0"), val = tensor<int32, [4]>([1, 8, 64, 128])];
            tensor<bool, [4]> var_1330_end_mask_0 = const()[name = string("op_1330_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_1330_squeeze_mask_0 = const()[name = string("op_1330_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 8, 64, 128]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_1200_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [8, 64, 128]> var_1330_cast_fp16 = slice_by_index(begin = var_1330_begin_0, end = var_1330_end_0, end_mask = var_1330_end_mask_0, squeeze_mask = var_1330_squeeze_mask_0, x = v_7_cast_fp16)[name = string("op_1330_cast_fp16")];
            bool var_1331_transpose_x_1 = const()[name = string("op_1331_transpose_x_1"), val = bool(true)];
            bool var_1331_transpose_y_1 = const()[name = string("op_1331_transpose_y_1"), val = bool(false)];
            tensor<fp16, [8, 704, 128]> var_1331_cast_fp16 = matmul(transpose_x = var_1331_transpose_x_1, transpose_y = var_1331_transpose_y_1, x = onehot, y = var_1330_cast_fp16)[name = string("op_1331_cast_fp16")];
            tensor<int32, [1]> var_1333_axes_0 = const()[name = string("op_1333_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8, 704, 128]> var_1333_cast_fp16 = expand_dims(axes = var_1333_axes_0, x = var_1331_cast_fp16)[name = string("op_1333_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_1335_cast_fp16 = add(x = var_1326_cast_fp16, y = var_1333_cast_fp16)[name = string("op_1335_cast_fp16")];
            tensor<int32, [4]> concat_26 = const()[name = string("concat_26"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> concat_27 = const()[name = string("concat_27"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> v_cache_internal_tensor_assign_4_stride_0 = const()[name = string("v_cache_internal_tensor_assign_4_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_4_begin_mask_0 = const()[name = string("v_cache_internal_tensor_assign_4_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_4_end_mask_0 = const()[name = string("v_cache_internal_tensor_assign_4_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_4_squeeze_mask_0 = const()[name = string("v_cache_internal_tensor_assign_4_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [7, 8, 704, 128]> v_cache_internal_tensor_assign_4_cast_fp16 = slice_update(begin = concat_26, begin_mask = v_cache_internal_tensor_assign_4_begin_mask_0, end = concat_27, end_mask = v_cache_internal_tensor_assign_4_end_mask_0, squeeze_mask = v_cache_internal_tensor_assign_4_squeeze_mask_0, stride = v_cache_internal_tensor_assign_4_stride_0, update = var_1335_cast_fp16, x = coreml_update_state_19)[name = string("v_cache_internal_tensor_assign_4_cast_fp16")];
            write_state(data = v_cache_internal_tensor_assign_4_cast_fp16, input = v_cache)[name = string("coreml_update_state_21_write_state")];
            tensor<fp16, [7, 8, 704, 128]> coreml_update_state_21 = read_state(input = v_cache)[name = string("coreml_update_state_21")];
            tensor<int32, [4]> var_1352_begin_0 = const()[name = string("op_1352_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1352_end_0 = const()[name = string("op_1352_end_0"), val = tensor<int32, [4]>([4, 8, 704, 128])];
            tensor<bool, [4]> var_1352_end_mask_0 = const()[name = string("op_1352_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_1352_cast_fp16 = slice_by_index(begin = var_1352_begin_0, end = var_1352_end_0, end_mask = var_1352_end_mask_0, x = coreml_update_state_20)[name = string("op_1352_cast_fp16")];
            tensor<int32, [4]> transpose_12_perm_0 = const()[name = string("transpose_12_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_6_reps_0 = const()[name = string("tile_6_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [8, 1, 704, 128]> transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = var_1352_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [16, 1, 704, 128]> tile_6_cast_fp16 = tile(reps = tile_6_reps_0, x = transpose_12_cast_fp16)[name = string("tile_6_cast_fp16")];
            tensor<int32, [5]> concat_28 = const()[name = string("concat_28"), val = tensor<int32, [5]>([2, 8, 1, 704, 128])];
            tensor<fp16, [2, 8, 1, 704, 128]> reshape_12_cast_fp16 = reshape(shape = concat_28, x = tile_6_cast_fp16)[name = string("reshape_12_cast_fp16")];
            tensor<int32, [5]> transpose_13_perm_0 = const()[name = string("transpose_13_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_29 = const()[name = string("concat_29"), val = tensor<int32, [4]>([-1, 1, 704, 128])];
            tensor<fp16, [8, 2, 1, 704, 128]> transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = reshape_12_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [16, 1, 704, 128]> reshape_13_cast_fp16 = reshape(shape = concat_29, x = transpose_13_cast_fp16)[name = string("reshape_13_cast_fp16")];
            tensor<int32, [4]> transpose_31_perm_0 = const()[name = string("transpose_31_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> var_1366_begin_0 = const()[name = string("op_1366_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_1366_end_0 = const()[name = string("op_1366_end_0"), val = tensor<int32, [4]>([4, 8, 704, 128])];
            tensor<bool, [4]> var_1366_end_mask_0 = const()[name = string("op_1366_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_1366_cast_fp16 = slice_by_index(begin = var_1366_begin_0, end = var_1366_end_0, end_mask = var_1366_end_mask_0, x = coreml_update_state_21)[name = string("op_1366_cast_fp16")];
            tensor<int32, [4]> transpose_14_perm_0 = const()[name = string("transpose_14_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_7_reps_0 = const()[name = string("tile_7_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [8, 1, 704, 128]> transpose_14_cast_fp16 = transpose(perm = transpose_14_perm_0, x = var_1366_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [16, 1, 704, 128]> tile_7_cast_fp16 = tile(reps = tile_7_reps_0, x = transpose_14_cast_fp16)[name = string("tile_7_cast_fp16")];
            tensor<int32, [5]> concat_30 = const()[name = string("concat_30"), val = tensor<int32, [5]>([2, 8, 1, 704, 128])];
            tensor<fp16, [2, 8, 1, 704, 128]> reshape_14_cast_fp16 = reshape(shape = concat_30, x = tile_7_cast_fp16)[name = string("reshape_14_cast_fp16")];
            tensor<int32, [5]> transpose_15_perm_0 = const()[name = string("transpose_15_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_31 = const()[name = string("concat_31"), val = tensor<int32, [4]>([-1, 1, 704, 128])];
            tensor<fp16, [8, 2, 1, 704, 128]> transpose_15_cast_fp16 = transpose(perm = transpose_15_perm_0, x = reshape_14_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [16, 1, 704, 128]> reshape_15_cast_fp16 = reshape(shape = concat_31, x = transpose_15_cast_fp16)[name = string("reshape_15_cast_fp16")];
            tensor<int32, [4]> values_7_perm_0 = const()[name = string("values_7_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_1379_transpose_x_0 = const()[name = string("op_1379_transpose_x_0"), val = bool(false)];
            bool var_1379_transpose_y_0 = const()[name = string("op_1379_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 704]> transpose_31_cast_fp16 = transpose(perm = transpose_31_perm_0, x = reshape_13_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 16, 64, 704]> var_1379_cast_fp16 = matmul(transpose_x = var_1379_transpose_x_0, transpose_y = var_1379_transpose_y_0, x = q_15_cast_fp16, y = transpose_31_cast_fp16)[name = string("op_1379_cast_fp16")];
            fp16 var_1380_to_fp16 = const()[name = string("op_1380_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 64, 704]> var_1381_cast_fp16 = mul(x = var_1379_cast_fp16, y = var_1380_to_fp16)[name = string("op_1381_cast_fp16")];
            tensor<fp16, [1, 16, 64, 704]> scores_7_cast_fp16 = add(x = var_1381_cast_fp16, y = mask)[name = string("scores_7_cast_fp16")];
            int32 var_1384 = const()[name = string("op_1384"), val = int32(-1)];
            tensor<fp16, [1, 16, 64, 704]> var_1386_cast_fp16 = softmax(axis = var_1384, x = scores_7_cast_fp16)[name = string("op_1386_cast_fp16")];
            bool attn_7_transpose_x_0 = const()[name = string("attn_7_transpose_x_0"), val = bool(false)];
            bool attn_7_transpose_y_0 = const()[name = string("attn_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 704, 128]> values_7_cast_fp16 = transpose(perm = values_7_perm_0, x = reshape_15_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 16, 64, 128]> attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_1386_cast_fp16, y = values_7_cast_fp16)[name = string("attn_7_cast_fp16")];
            tensor<int32, [4]> var_1390_perm_0 = const()[name = string("op_1390_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1394 = const()[name = string("op_1394"), val = tensor<int32, [3]>([1, 64, 2048])];
            tensor<fp16, [1, 64, 16, 128]> var_1390_cast_fp16 = transpose(perm = var_1390_perm_0, x = attn_7_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 64, 2048]> input_33_cast_fp16 = reshape(shape = var_1394, x = var_1390_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [2048, 2048]> layers_3_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159568064))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163762432))))[name = string("layers_3_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_3_o_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> x_29_cast_fp16 = add(x = x_23_cast_fp16, y = linear_24_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1401_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_1401_cast_fp16")];
            bool var_1404 = const()[name = string("op_1404"), val = bool(true)];
            tensor<int32, [1]> var_1405_axes_0 = const()[name = string("op_1405_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1405_cast_fp16 = reduce_max(axes = var_1405_axes_0, keep_dims = var_1404, x = var_1401_cast_fp16)[name = string("op_1405_cast_fp16")];
            fp16 var_1406_to_fp16 = const()[name = string("op_1406_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_24_to_fp16 = const()[name = string("const_24_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_16_cast_fp16 = clip(alpha = var_1406_to_fp16, beta = const_24_to_fp16, x = var_1405_cast_fp16)[name = string("clip_16_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> y_31_cast_fp16 = real_div(x = x_29_cast_fp16, y = clip_16_cast_fp16)[name = string("y_31_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1410_cast_fp16 = mul(x = y_31_cast_fp16, y = y_31_cast_fp16)[name = string("op_1410_cast_fp16")];
            tensor<int32, [1]> var_1415_axes_0 = const()[name = string("op_1415_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1415_keep_dims_0 = const()[name = string("op_1415_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_1415_cast_fp16 = reduce_mean(axes = var_1415_axes_0, keep_dims = var_1415_keep_dims_0, x = var_1410_cast_fp16)[name = string("op_1415_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1416_cast_fp16 = mul(x = clip_16_cast_fp16, y = clip_16_cast_fp16)[name = string("op_1416_cast_fp16")];
            fp32 var_1417_epsilon_0 = const()[name = string("op_1417_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1417_cast_fp16 = inverse(epsilon = var_1417_epsilon_0, x = var_1416_cast_fp16)[name = string("op_1417_cast_fp16")];
            fp16 var_1418_to_fp16 = const()[name = string("op_1418_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1419_cast_fp16 = mul(x = var_1417_cast_fp16, y = var_1418_to_fp16)[name = string("op_1419_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1421_cast_fp16 = add(x = var_1415_cast_fp16, y = var_1419_cast_fp16)[name = string("op_1421_cast_fp16")];
            fp32 var_1422_epsilon_0 = const()[name = string("op_1422_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1422_cast_fp16 = rsqrt(epsilon = var_1422_epsilon_0, x = var_1421_cast_fp16)[name = string("op_1422_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1423_cast_fp16 = mul(x = y_31_cast_fp16, y = var_1422_cast_fp16)[name = string("op_1423_cast_fp16")];
            tensor<fp16, [2048]> layers_3_post_norm_to_fp16 = const()[name = string("layers_3_post_norm_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163766592)))];
            tensor<fp16, [1, 64, 2048]> input_35_cast_fp16 = mul(x = var_1423_cast_fp16, y = layers_3_post_norm_to_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [6144, 2048]> layers_3_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163770752))), scale = tensor<fp16, [6144, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176353728))))[name = string("layers_3_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 6144]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_3_gate_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 64, 6144]> var_1428_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_1428_cast_fp16")];
            tensor<fp16, [6144, 2048]> layers_3_up_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176366080))), scale = tensor<fp16, [6144, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188949056))))[name = string("layers_3_up_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 6144]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_3_up_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, 64, 6144]> var_1432_cast_fp16 = mul(x = var_1428_cast_fp16, y = linear_26_cast_fp16)[name = string("op_1432_cast_fp16")];
            fp16 var_1433_to_fp16 = const()[name = string("op_1433_to_fp16"), val = fp16(0x1p-8)];
            tensor<fp16, [1, 64, 6144]> input_39_cast_fp16 = mul(x = var_1432_cast_fp16, y = var_1433_to_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [2048, 6144]> layers_3_down_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188961408))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201544384))))[name = string("layers_3_down_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_27_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_3_down_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_27_cast_fp16")];
            fp16 var_1438_to_fp16 = const()[name = string("op_1438_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 64, 2048]> var_1439_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1438_to_fp16)[name = string("op_1439_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> x_31_cast_fp16 = add(x = x_29_cast_fp16, y = var_1439_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1442_cast_fp16 = abs(x = x_31_cast_fp16)[name = string("op_1442_cast_fp16")];
            bool var_1445 = const()[name = string("op_1445"), val = bool(true)];
            tensor<int32, [1]> var_1446_axes_0 = const()[name = string("op_1446_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1446_cast_fp16 = reduce_max(axes = var_1446_axes_0, keep_dims = var_1445, x = var_1442_cast_fp16)[name = string("op_1446_cast_fp16")];
            fp16 var_1447_to_fp16 = const()[name = string("op_1447_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_25_to_fp16 = const()[name = string("const_25_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_17_cast_fp16 = clip(alpha = var_1447_to_fp16, beta = const_25_to_fp16, x = var_1446_cast_fp16)[name = string("clip_17_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> y_33_cast_fp16 = real_div(x = x_31_cast_fp16, y = clip_17_cast_fp16)[name = string("y_33_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1451_cast_fp16 = mul(x = y_33_cast_fp16, y = y_33_cast_fp16)[name = string("op_1451_cast_fp16")];
            tensor<int32, [1]> var_1456_axes_0 = const()[name = string("op_1456_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1456_keep_dims_0 = const()[name = string("op_1456_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_1456_cast_fp16 = reduce_mean(axes = var_1456_axes_0, keep_dims = var_1456_keep_dims_0, x = var_1451_cast_fp16)[name = string("op_1456_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1457_cast_fp16 = mul(x = clip_17_cast_fp16, y = clip_17_cast_fp16)[name = string("op_1457_cast_fp16")];
            fp32 var_1458_epsilon_0 = const()[name = string("op_1458_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1458_cast_fp16 = inverse(epsilon = var_1458_epsilon_0, x = var_1457_cast_fp16)[name = string("op_1458_cast_fp16")];
            fp16 var_1459_to_fp16 = const()[name = string("op_1459_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1460_cast_fp16 = mul(x = var_1458_cast_fp16, y = var_1459_to_fp16)[name = string("op_1460_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1462_cast_fp16 = add(x = var_1456_cast_fp16, y = var_1460_cast_fp16)[name = string("op_1462_cast_fp16")];
            fp32 var_1463_epsilon_0 = const()[name = string("op_1463_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1463_cast_fp16 = rsqrt(epsilon = var_1463_epsilon_0, x = var_1462_cast_fp16)[name = string("op_1463_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1464_cast_fp16 = mul(x = y_33_cast_fp16, y = var_1463_cast_fp16)[name = string("op_1464_cast_fp16")];
            tensor<fp16, [2048]> layers_4_input_norm_to_fp16 = const()[name = string("layers_4_input_norm_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201548544)))];
            tensor<fp16, [1, 64, 2048]> input_41_cast_fp16 = mul(x = var_1464_cast_fp16, y = layers_4_input_norm_to_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [2048, 2048]> layers_4_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201552704))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205747072))))[name = string("layers_4_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_q_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_1473 = const()[name = string("op_1473"), val = tensor<int32, [4]>([1, 64, 16, 128])];
            tensor<fp16, [1, 64, 16, 128]> var_1474_cast_fp16 = reshape(shape = var_1473, x = linear_28_cast_fp16)[name = string("op_1474_cast_fp16")];
            tensor<int32, [4]> x_33_perm_0 = const()[name = string("x_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 2048]> layers_4_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205751232))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207848448))))[name = string("layers_4_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_k_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_1485 = const()[name = string("op_1485"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<fp16, [1, 64, 8, 128]> var_1486_cast_fp16 = reshape(shape = var_1485, x = linear_29_cast_fp16)[name = string("op_1486_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, 2048]> layers_4_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207850560))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209947776))))[name = string("layers_4_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_30_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_v_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_1497 = const()[name = string("op_1497"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<fp16, [1, 64, 8, 128]> var_1498_cast_fp16 = reshape(shape = var_1497, x = linear_30_cast_fp16)[name = string("op_1498_cast_fp16")];
            tensor<int32, [4]> v_9_perm_0 = const()[name = string("v_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 64, 128]> x_33_cast_fp16 = transpose(perm = x_33_perm_0, x = var_1474_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 16, 64, 128]> var_1502_cast_fp16 = abs(x = x_33_cast_fp16)[name = string("op_1502_cast_fp16")];
            bool var_1505 = const()[name = string("op_1505"), val = bool(true)];
            tensor<int32, [1]> var_1506_axes_0 = const()[name = string("op_1506_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 64, 1]> var_1506_cast_fp16 = reduce_max(axes = var_1506_axes_0, keep_dims = var_1505, x = var_1502_cast_fp16)[name = string("op_1506_cast_fp16")];
            fp16 var_1507_to_fp16 = const()[name = string("op_1507_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_26_to_fp16 = const()[name = string("const_26_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 16, 64, 1]> clip_18_cast_fp16 = clip(alpha = var_1507_to_fp16, beta = const_26_to_fp16, x = var_1506_cast_fp16)[name = string("clip_18_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> y_35_cast_fp16 = real_div(x = x_33_cast_fp16, y = clip_18_cast_fp16)[name = string("y_35_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_1511_cast_fp16 = mul(x = y_35_cast_fp16, y = y_35_cast_fp16)[name = string("op_1511_cast_fp16")];
            tensor<int32, [1]> var_1516_axes_0 = const()[name = string("op_1516_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1516_keep_dims_0 = const()[name = string("op_1516_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 64, 1]> var_1516_cast_fp16 = reduce_mean(axes = var_1516_axes_0, keep_dims = var_1516_keep_dims_0, x = var_1511_cast_fp16)[name = string("op_1516_cast_fp16")];
            tensor<fp16, [1, 16, 64, 1]> var_1517_cast_fp16 = mul(x = clip_18_cast_fp16, y = clip_18_cast_fp16)[name = string("op_1517_cast_fp16")];
            fp32 var_1518_epsilon_0 = const()[name = string("op_1518_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 64, 1]> var_1518_cast_fp16 = inverse(epsilon = var_1518_epsilon_0, x = var_1517_cast_fp16)[name = string("op_1518_cast_fp16")];
            fp16 var_1519_to_fp16 = const()[name = string("op_1519_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 64, 1]> var_1520_cast_fp16 = mul(x = var_1518_cast_fp16, y = var_1519_to_fp16)[name = string("op_1520_cast_fp16")];
            tensor<fp16, [1, 16, 64, 1]> var_1522_cast_fp16 = add(x = var_1516_cast_fp16, y = var_1520_cast_fp16)[name = string("op_1522_cast_fp16")];
            fp32 var_1523_epsilon_0 = const()[name = string("op_1523_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 64, 1]> var_1523_cast_fp16 = rsqrt(epsilon = var_1523_epsilon_0, x = var_1522_cast_fp16)[name = string("op_1523_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_1524_cast_fp16 = mul(x = y_35_cast_fp16, y = var_1523_cast_fp16)[name = string("op_1524_cast_fp16")];
            tensor<fp16, [128]> layers_4_q_norm_to_fp16 = const()[name = string("layers_4_q_norm_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209949888)))];
            tensor<fp16, [1, 16, 64, 128]> q_17_cast_fp16 = mul(x = var_1524_cast_fp16, y = layers_4_q_norm_to_fp16)[name = string("q_17_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> x_35_cast_fp16 = transpose(perm = x_35_perm_0, x = var_1486_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 8, 64, 128]> var_1526_cast_fp16 = abs(x = x_35_cast_fp16)[name = string("op_1526_cast_fp16")];
            bool var_1529 = const()[name = string("op_1529"), val = bool(true)];
            tensor<int32, [1]> var_1530_axes_0 = const()[name = string("op_1530_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 64, 1]> var_1530_cast_fp16 = reduce_max(axes = var_1530_axes_0, keep_dims = var_1529, x = var_1526_cast_fp16)[name = string("op_1530_cast_fp16")];
            fp16 var_1531_to_fp16 = const()[name = string("op_1531_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_27_to_fp16 = const()[name = string("const_27_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 8, 64, 1]> clip_19_cast_fp16 = clip(alpha = var_1531_to_fp16, beta = const_27_to_fp16, x = var_1530_cast_fp16)[name = string("clip_19_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> y_37_cast_fp16 = real_div(x = x_35_cast_fp16, y = clip_19_cast_fp16)[name = string("y_37_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_1535_cast_fp16 = mul(x = y_37_cast_fp16, y = y_37_cast_fp16)[name = string("op_1535_cast_fp16")];
            tensor<int32, [1]> var_1540_axes_0 = const()[name = string("op_1540_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1540_keep_dims_0 = const()[name = string("op_1540_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 64, 1]> var_1540_cast_fp16 = reduce_mean(axes = var_1540_axes_0, keep_dims = var_1540_keep_dims_0, x = var_1535_cast_fp16)[name = string("op_1540_cast_fp16")];
            tensor<fp16, [1, 8, 64, 1]> var_1541_cast_fp16 = mul(x = clip_19_cast_fp16, y = clip_19_cast_fp16)[name = string("op_1541_cast_fp16")];
            fp32 var_1542_epsilon_0 = const()[name = string("op_1542_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 64, 1]> var_1542_cast_fp16 = inverse(epsilon = var_1542_epsilon_0, x = var_1541_cast_fp16)[name = string("op_1542_cast_fp16")];
            fp16 var_1543_to_fp16 = const()[name = string("op_1543_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 64, 1]> var_1544_cast_fp16 = mul(x = var_1542_cast_fp16, y = var_1543_to_fp16)[name = string("op_1544_cast_fp16")];
            tensor<fp16, [1, 8, 64, 1]> var_1546_cast_fp16 = add(x = var_1540_cast_fp16, y = var_1544_cast_fp16)[name = string("op_1546_cast_fp16")];
            fp32 var_1547_epsilon_0 = const()[name = string("op_1547_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 64, 1]> var_1547_cast_fp16 = rsqrt(epsilon = var_1547_epsilon_0, x = var_1546_cast_fp16)[name = string("op_1547_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_1548_cast_fp16 = mul(x = y_37_cast_fp16, y = var_1547_cast_fp16)[name = string("op_1548_cast_fp16")];
            tensor<fp16, [128]> layers_4_k_norm_to_fp16 = const()[name = string("layers_4_k_norm_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209950208)))];
            tensor<fp16, [1, 8, 64, 128]> k_17_cast_fp16 = mul(x = var_1548_cast_fp16, y = layers_4_k_norm_to_fp16)[name = string("k_17_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_1550_cast_fp16 = mul(x = q_17_cast_fp16, y = cos)[name = string("op_1550_cast_fp16")];
            tensor<int32, [2]> var_1551_split_sizes_0 = const()[name = string("op_1551_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1551_axis_0 = const()[name = string("op_1551_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 64, 64]> var_1551_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_1551_cast_fp16_1 = split(axis = var_1551_axis_0, split_sizes = var_1551_split_sizes_0, x = q_17_cast_fp16)[name = string("op_1551_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_1553_cast_fp16 = mul(x = var_1551_cast_fp16_1, y = const_28_promoted_to_fp16)[name = string("op_1553_cast_fp16")];
            int32 var_1555 = const()[name = string("op_1555"), val = int32(-1)];
            bool var_1556_interleave_0 = const()[name = string("op_1556_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> var_1556_cast_fp16 = concat(axis = var_1555, interleave = var_1556_interleave_0, values = (var_1553_cast_fp16, var_1551_cast_fp16_0))[name = string("op_1556_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_1557_cast_fp16 = mul(x = var_1556_cast_fp16, y = sin)[name = string("op_1557_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> q_19_cast_fp16 = add(x = var_1550_cast_fp16, y = var_1557_cast_fp16)[name = string("q_19_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_1560_cast_fp16 = mul(x = k_17_cast_fp16, y = cos)[name = string("op_1560_cast_fp16")];
            tensor<int32, [2]> var_1561_split_sizes_0 = const()[name = string("op_1561_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1561_axis_0 = const()[name = string("op_1561_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 64, 64]> var_1561_cast_fp16_0, tensor<fp16, [1, 8, 64, 64]> var_1561_cast_fp16_1 = split(axis = var_1561_axis_0, split_sizes = var_1561_split_sizes_0, x = k_17_cast_fp16)[name = string("op_1561_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 64, 64]> var_1563_cast_fp16 = mul(x = var_1561_cast_fp16_1, y = const_29_promoted_to_fp16)[name = string("op_1563_cast_fp16")];
            int32 var_1565 = const()[name = string("op_1565"), val = int32(-1)];
            bool var_1566_interleave_0 = const()[name = string("op_1566_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_1566_cast_fp16 = concat(axis = var_1565, interleave = var_1566_interleave_0, values = (var_1563_cast_fp16, var_1561_cast_fp16_0))[name = string("op_1566_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_1567_cast_fp16 = mul(x = var_1566_cast_fp16, y = sin)[name = string("op_1567_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> k_19_cast_fp16 = add(x = var_1560_cast_fp16, y = var_1567_cast_fp16)[name = string("k_19_cast_fp16")];
            tensor<int32, [4]> var_1574_begin_0 = const()[name = string("op_1574_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1574_end_0 = const()[name = string("op_1574_end_0"), val = tensor<int32, [4]>([5, 8, 704, 128])];
            tensor<bool, [4]> var_1574_end_mask_0 = const()[name = string("op_1574_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_1574_cast_fp16 = slice_by_index(begin = var_1574_begin_0, end = var_1574_end_0, end_mask = var_1574_end_mask_0, x = coreml_update_state_20)[name = string("op_1574_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_1586_cast_fp16 = mul(x = var_1574_cast_fp16, y = var_393_cast_fp16)[name = string("op_1586_cast_fp16")];
            tensor<int32, [4]> var_1590_begin_0 = const()[name = string("op_1590_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1590_end_0 = const()[name = string("op_1590_end_0"), val = tensor<int32, [4]>([1, 8, 64, 128])];
            tensor<bool, [4]> var_1590_end_mask_0 = const()[name = string("op_1590_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_1590_squeeze_mask_0 = const()[name = string("op_1590_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [8, 64, 128]> var_1590_cast_fp16 = slice_by_index(begin = var_1590_begin_0, end = var_1590_end_0, end_mask = var_1590_end_mask_0, squeeze_mask = var_1590_squeeze_mask_0, x = k_19_cast_fp16)[name = string("op_1590_cast_fp16")];
            bool var_1591_transpose_x_1 = const()[name = string("op_1591_transpose_x_1"), val = bool(true)];
            bool var_1591_transpose_y_1 = const()[name = string("op_1591_transpose_y_1"), val = bool(false)];
            tensor<fp16, [8, 704, 128]> var_1591_cast_fp16 = matmul(transpose_x = var_1591_transpose_x_1, transpose_y = var_1591_transpose_y_1, x = onehot, y = var_1590_cast_fp16)[name = string("op_1591_cast_fp16")];
            tensor<int32, [1]> var_1593_axes_0 = const()[name = string("op_1593_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8, 704, 128]> var_1593_cast_fp16 = expand_dims(axes = var_1593_axes_0, x = var_1591_cast_fp16)[name = string("op_1593_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_1595_cast_fp16 = add(x = var_1586_cast_fp16, y = var_1593_cast_fp16)[name = string("op_1595_cast_fp16")];
            tensor<int32, [4]> concat_32 = const()[name = string("concat_32"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> concat_33 = const()[name = string("concat_33"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> k_cache_internal_tensor_assign_5_stride_0 = const()[name = string("k_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("k_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("k_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("k_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [7, 8, 704, 128]> k_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_32, begin_mask = k_cache_internal_tensor_assign_5_begin_mask_0, end = concat_33, end_mask = k_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = k_cache_internal_tensor_assign_5_squeeze_mask_0, stride = k_cache_internal_tensor_assign_5_stride_0, update = var_1595_cast_fp16, x = coreml_update_state_20)[name = string("k_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = k_cache_internal_tensor_assign_5_cast_fp16, input = k_cache)[name = string("coreml_update_state_22_write_state")];
            tensor<fp16, [7, 8, 704, 128]> coreml_update_state_22 = read_state(input = k_cache)[name = string("coreml_update_state_22")];
            tensor<int32, [4]> var_1612_begin_0 = const()[name = string("op_1612_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1612_end_0 = const()[name = string("op_1612_end_0"), val = tensor<int32, [4]>([5, 8, 704, 128])];
            tensor<bool, [4]> var_1612_end_mask_0 = const()[name = string("op_1612_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_1612_cast_fp16 = slice_by_index(begin = var_1612_begin_0, end = var_1612_end_0, end_mask = var_1612_end_mask_0, x = coreml_update_state_21)[name = string("op_1612_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_1624_cast_fp16 = mul(x = var_1612_cast_fp16, y = var_393_cast_fp16)[name = string("op_1624_cast_fp16")];
            tensor<int32, [4]> var_1628_begin_0 = const()[name = string("op_1628_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1628_end_0 = const()[name = string("op_1628_end_0"), val = tensor<int32, [4]>([1, 8, 64, 128])];
            tensor<bool, [4]> var_1628_end_mask_0 = const()[name = string("op_1628_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_1628_squeeze_mask_0 = const()[name = string("op_1628_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 8, 64, 128]> v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = var_1498_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [8, 64, 128]> var_1628_cast_fp16 = slice_by_index(begin = var_1628_begin_0, end = var_1628_end_0, end_mask = var_1628_end_mask_0, squeeze_mask = var_1628_squeeze_mask_0, x = v_9_cast_fp16)[name = string("op_1628_cast_fp16")];
            bool var_1629_transpose_x_1 = const()[name = string("op_1629_transpose_x_1"), val = bool(true)];
            bool var_1629_transpose_y_1 = const()[name = string("op_1629_transpose_y_1"), val = bool(false)];
            tensor<fp16, [8, 704, 128]> var_1629_cast_fp16 = matmul(transpose_x = var_1629_transpose_x_1, transpose_y = var_1629_transpose_y_1, x = onehot, y = var_1628_cast_fp16)[name = string("op_1629_cast_fp16")];
            tensor<int32, [1]> var_1631_axes_0 = const()[name = string("op_1631_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8, 704, 128]> var_1631_cast_fp16 = expand_dims(axes = var_1631_axes_0, x = var_1629_cast_fp16)[name = string("op_1631_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_1633_cast_fp16 = add(x = var_1624_cast_fp16, y = var_1631_cast_fp16)[name = string("op_1633_cast_fp16")];
            tensor<int32, [4]> concat_34 = const()[name = string("concat_34"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> concat_35 = const()[name = string("concat_35"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> v_cache_internal_tensor_assign_5_stride_0 = const()[name = string("v_cache_internal_tensor_assign_5_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_5_begin_mask_0 = const()[name = string("v_cache_internal_tensor_assign_5_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_5_end_mask_0 = const()[name = string("v_cache_internal_tensor_assign_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_5_squeeze_mask_0 = const()[name = string("v_cache_internal_tensor_assign_5_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [7, 8, 704, 128]> v_cache_internal_tensor_assign_5_cast_fp16 = slice_update(begin = concat_34, begin_mask = v_cache_internal_tensor_assign_5_begin_mask_0, end = concat_35, end_mask = v_cache_internal_tensor_assign_5_end_mask_0, squeeze_mask = v_cache_internal_tensor_assign_5_squeeze_mask_0, stride = v_cache_internal_tensor_assign_5_stride_0, update = var_1633_cast_fp16, x = coreml_update_state_21)[name = string("v_cache_internal_tensor_assign_5_cast_fp16")];
            write_state(data = v_cache_internal_tensor_assign_5_cast_fp16, input = v_cache)[name = string("coreml_update_state_23_write_state")];
            tensor<fp16, [7, 8, 704, 128]> coreml_update_state_23 = read_state(input = v_cache)[name = string("coreml_update_state_23")];
            tensor<int32, [4]> var_1650_begin_0 = const()[name = string("op_1650_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1650_end_0 = const()[name = string("op_1650_end_0"), val = tensor<int32, [4]>([5, 8, 704, 128])];
            tensor<bool, [4]> var_1650_end_mask_0 = const()[name = string("op_1650_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_1650_cast_fp16 = slice_by_index(begin = var_1650_begin_0, end = var_1650_end_0, end_mask = var_1650_end_mask_0, x = coreml_update_state_22)[name = string("op_1650_cast_fp16")];
            tensor<int32, [4]> transpose_16_perm_0 = const()[name = string("transpose_16_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_8_reps_0 = const()[name = string("tile_8_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [8, 1, 704, 128]> transpose_16_cast_fp16 = transpose(perm = transpose_16_perm_0, x = var_1650_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [16, 1, 704, 128]> tile_8_cast_fp16 = tile(reps = tile_8_reps_0, x = transpose_16_cast_fp16)[name = string("tile_8_cast_fp16")];
            tensor<int32, [5]> concat_36 = const()[name = string("concat_36"), val = tensor<int32, [5]>([2, 8, 1, 704, 128])];
            tensor<fp16, [2, 8, 1, 704, 128]> reshape_16_cast_fp16 = reshape(shape = concat_36, x = tile_8_cast_fp16)[name = string("reshape_16_cast_fp16")];
            tensor<int32, [5]> transpose_17_perm_0 = const()[name = string("transpose_17_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_37 = const()[name = string("concat_37"), val = tensor<int32, [4]>([-1, 1, 704, 128])];
            tensor<fp16, [8, 2, 1, 704, 128]> transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = reshape_16_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [16, 1, 704, 128]> reshape_17_cast_fp16 = reshape(shape = concat_37, x = transpose_17_cast_fp16)[name = string("reshape_17_cast_fp16")];
            tensor<int32, [4]> transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> var_1664_begin_0 = const()[name = string("op_1664_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_1664_end_0 = const()[name = string("op_1664_end_0"), val = tensor<int32, [4]>([5, 8, 704, 128])];
            tensor<bool, [4]> var_1664_end_mask_0 = const()[name = string("op_1664_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_1664_cast_fp16 = slice_by_index(begin = var_1664_begin_0, end = var_1664_end_0, end_mask = var_1664_end_mask_0, x = coreml_update_state_23)[name = string("op_1664_cast_fp16")];
            tensor<int32, [4]> transpose_18_perm_0 = const()[name = string("transpose_18_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_9_reps_0 = const()[name = string("tile_9_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [8, 1, 704, 128]> transpose_18_cast_fp16 = transpose(perm = transpose_18_perm_0, x = var_1664_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [16, 1, 704, 128]> tile_9_cast_fp16 = tile(reps = tile_9_reps_0, x = transpose_18_cast_fp16)[name = string("tile_9_cast_fp16")];
            tensor<int32, [5]> concat_38 = const()[name = string("concat_38"), val = tensor<int32, [5]>([2, 8, 1, 704, 128])];
            tensor<fp16, [2, 8, 1, 704, 128]> reshape_18_cast_fp16 = reshape(shape = concat_38, x = tile_9_cast_fp16)[name = string("reshape_18_cast_fp16")];
            tensor<int32, [5]> transpose_19_perm_0 = const()[name = string("transpose_19_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_39 = const()[name = string("concat_39"), val = tensor<int32, [4]>([-1, 1, 704, 128])];
            tensor<fp16, [8, 2, 1, 704, 128]> transpose_19_cast_fp16 = transpose(perm = transpose_19_perm_0, x = reshape_18_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [16, 1, 704, 128]> reshape_19_cast_fp16 = reshape(shape = concat_39, x = transpose_19_cast_fp16)[name = string("reshape_19_cast_fp16")];
            tensor<int32, [4]> values_9_perm_0 = const()[name = string("values_9_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_1677_transpose_x_0 = const()[name = string("op_1677_transpose_x_0"), val = bool(false)];
            bool var_1677_transpose_y_0 = const()[name = string("op_1677_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 704]> transpose_32_cast_fp16 = transpose(perm = transpose_32_perm_0, x = reshape_17_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 16, 64, 704]> var_1677_cast_fp16 = matmul(transpose_x = var_1677_transpose_x_0, transpose_y = var_1677_transpose_y_0, x = q_19_cast_fp16, y = transpose_32_cast_fp16)[name = string("op_1677_cast_fp16")];
            fp16 var_1678_to_fp16 = const()[name = string("op_1678_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 64, 704]> var_1679_cast_fp16 = mul(x = var_1677_cast_fp16, y = var_1678_to_fp16)[name = string("op_1679_cast_fp16")];
            tensor<fp16, [1, 16, 64, 704]> scores_9_cast_fp16 = add(x = var_1679_cast_fp16, y = mask)[name = string("scores_9_cast_fp16")];
            int32 var_1682 = const()[name = string("op_1682"), val = int32(-1)];
            tensor<fp16, [1, 16, 64, 704]> var_1684_cast_fp16 = softmax(axis = var_1682, x = scores_9_cast_fp16)[name = string("op_1684_cast_fp16")];
            bool attn_9_transpose_x_0 = const()[name = string("attn_9_transpose_x_0"), val = bool(false)];
            bool attn_9_transpose_y_0 = const()[name = string("attn_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 704, 128]> values_9_cast_fp16 = transpose(perm = values_9_perm_0, x = reshape_19_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 16, 64, 128]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_1684_cast_fp16, y = values_9_cast_fp16)[name = string("attn_9_cast_fp16")];
            tensor<int32, [4]> var_1688_perm_0 = const()[name = string("op_1688_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1692 = const()[name = string("op_1692"), val = tensor<int32, [3]>([1, 64, 2048])];
            tensor<fp16, [1, 64, 16, 128]> var_1688_cast_fp16 = transpose(perm = var_1688_perm_0, x = attn_9_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 64, 2048]> input_43_cast_fp16 = reshape(shape = var_1692, x = var_1688_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [2048, 2048]> layers_4_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209950528))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214144896))))[name = string("layers_4_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_31_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_o_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> x_37_cast_fp16 = add(x = x_31_cast_fp16, y = linear_31_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1699_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_1699_cast_fp16")];
            bool var_1702 = const()[name = string("op_1702"), val = bool(true)];
            tensor<int32, [1]> var_1703_axes_0 = const()[name = string("op_1703_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1703_cast_fp16 = reduce_max(axes = var_1703_axes_0, keep_dims = var_1702, x = var_1699_cast_fp16)[name = string("op_1703_cast_fp16")];
            fp16 var_1704_to_fp16 = const()[name = string("op_1704_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_30_to_fp16 = const()[name = string("const_30_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_20_cast_fp16 = clip(alpha = var_1704_to_fp16, beta = const_30_to_fp16, x = var_1703_cast_fp16)[name = string("clip_20_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> y_39_cast_fp16 = real_div(x = x_37_cast_fp16, y = clip_20_cast_fp16)[name = string("y_39_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1708_cast_fp16 = mul(x = y_39_cast_fp16, y = y_39_cast_fp16)[name = string("op_1708_cast_fp16")];
            tensor<int32, [1]> var_1713_axes_0 = const()[name = string("op_1713_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1713_keep_dims_0 = const()[name = string("op_1713_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_1713_cast_fp16 = reduce_mean(axes = var_1713_axes_0, keep_dims = var_1713_keep_dims_0, x = var_1708_cast_fp16)[name = string("op_1713_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1714_cast_fp16 = mul(x = clip_20_cast_fp16, y = clip_20_cast_fp16)[name = string("op_1714_cast_fp16")];
            fp32 var_1715_epsilon_0 = const()[name = string("op_1715_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1715_cast_fp16 = inverse(epsilon = var_1715_epsilon_0, x = var_1714_cast_fp16)[name = string("op_1715_cast_fp16")];
            fp16 var_1716_to_fp16 = const()[name = string("op_1716_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1717_cast_fp16 = mul(x = var_1715_cast_fp16, y = var_1716_to_fp16)[name = string("op_1717_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1719_cast_fp16 = add(x = var_1713_cast_fp16, y = var_1717_cast_fp16)[name = string("op_1719_cast_fp16")];
            fp32 var_1720_epsilon_0 = const()[name = string("op_1720_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1720_cast_fp16 = rsqrt(epsilon = var_1720_epsilon_0, x = var_1719_cast_fp16)[name = string("op_1720_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1721_cast_fp16 = mul(x = y_39_cast_fp16, y = var_1720_cast_fp16)[name = string("op_1721_cast_fp16")];
            tensor<fp16, [2048]> layers_4_post_norm_to_fp16 = const()[name = string("layers_4_post_norm_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214149056)))];
            tensor<fp16, [1, 64, 2048]> input_45_cast_fp16 = mul(x = var_1721_cast_fp16, y = layers_4_post_norm_to_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [6144, 2048]> layers_4_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214153216))), scale = tensor<fp16, [6144, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226736192))))[name = string("layers_4_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 6144]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_4_gate_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, 64, 6144]> var_1726_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_1726_cast_fp16")];
            tensor<fp16, [6144, 2048]> layers_4_up_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226748544))), scale = tensor<fp16, [6144, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239331520))))[name = string("layers_4_up_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 6144]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_4_up_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 64, 6144]> var_1730_cast_fp16 = mul(x = var_1726_cast_fp16, y = linear_33_cast_fp16)[name = string("op_1730_cast_fp16")];
            fp16 var_1731_to_fp16 = const()[name = string("op_1731_to_fp16"), val = fp16(0x1p-8)];
            tensor<fp16, [1, 64, 6144]> input_49_cast_fp16 = mul(x = var_1730_cast_fp16, y = var_1731_to_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [2048, 6144]> layers_4_down_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239343872))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251926848))))[name = string("layers_4_down_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_down_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_34_cast_fp16")];
            fp16 var_1736_to_fp16 = const()[name = string("op_1736_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 64, 2048]> var_1737_cast_fp16 = mul(x = linear_34_cast_fp16, y = var_1736_to_fp16)[name = string("op_1737_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> x_39_cast_fp16 = add(x = x_37_cast_fp16, y = var_1737_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1740_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_1740_cast_fp16")];
            bool var_1743 = const()[name = string("op_1743"), val = bool(true)];
            tensor<int32, [1]> var_1744_axes_0 = const()[name = string("op_1744_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1744_cast_fp16 = reduce_max(axes = var_1744_axes_0, keep_dims = var_1743, x = var_1740_cast_fp16)[name = string("op_1744_cast_fp16")];
            fp16 var_1745_to_fp16 = const()[name = string("op_1745_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_31_to_fp16 = const()[name = string("const_31_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_21_cast_fp16 = clip(alpha = var_1745_to_fp16, beta = const_31_to_fp16, x = var_1744_cast_fp16)[name = string("clip_21_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> y_41_cast_fp16 = real_div(x = x_39_cast_fp16, y = clip_21_cast_fp16)[name = string("y_41_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1749_cast_fp16 = mul(x = y_41_cast_fp16, y = y_41_cast_fp16)[name = string("op_1749_cast_fp16")];
            tensor<int32, [1]> var_1754_axes_0 = const()[name = string("op_1754_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1754_keep_dims_0 = const()[name = string("op_1754_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_1754_cast_fp16 = reduce_mean(axes = var_1754_axes_0, keep_dims = var_1754_keep_dims_0, x = var_1749_cast_fp16)[name = string("op_1754_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1755_cast_fp16 = mul(x = clip_21_cast_fp16, y = clip_21_cast_fp16)[name = string("op_1755_cast_fp16")];
            fp32 var_1756_epsilon_0 = const()[name = string("op_1756_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1756_cast_fp16 = inverse(epsilon = var_1756_epsilon_0, x = var_1755_cast_fp16)[name = string("op_1756_cast_fp16")];
            fp16 var_1757_to_fp16 = const()[name = string("op_1757_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1758_cast_fp16 = mul(x = var_1756_cast_fp16, y = var_1757_to_fp16)[name = string("op_1758_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1760_cast_fp16 = add(x = var_1754_cast_fp16, y = var_1758_cast_fp16)[name = string("op_1760_cast_fp16")];
            fp32 var_1761_epsilon_0 = const()[name = string("op_1761_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1761_cast_fp16 = rsqrt(epsilon = var_1761_epsilon_0, x = var_1760_cast_fp16)[name = string("op_1761_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1762_cast_fp16 = mul(x = y_41_cast_fp16, y = var_1761_cast_fp16)[name = string("op_1762_cast_fp16")];
            tensor<fp16, [2048]> layers_5_input_norm_to_fp16 = const()[name = string("layers_5_input_norm_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251931008)))];
            tensor<fp16, [1, 64, 2048]> input_51_cast_fp16 = mul(x = var_1762_cast_fp16, y = layers_5_input_norm_to_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [2048, 2048]> layers_5_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251935168))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256129536))))[name = string("layers_5_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_35_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_q_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> var_1771 = const()[name = string("op_1771"), val = tensor<int32, [4]>([1, 64, 16, 128])];
            tensor<fp16, [1, 64, 16, 128]> var_1772_cast_fp16 = reshape(shape = var_1771, x = linear_35_cast_fp16)[name = string("op_1772_cast_fp16")];
            tensor<int32, [4]> x_41_perm_0 = const()[name = string("x_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 2048]> layers_5_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256133696))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258230912))))[name = string("layers_5_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_36_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_k_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_1783 = const()[name = string("op_1783"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<fp16, [1, 64, 8, 128]> var_1784_cast_fp16 = reshape(shape = var_1783, x = linear_36_cast_fp16)[name = string("op_1784_cast_fp16")];
            tensor<int32, [4]> x_43_perm_0 = const()[name = string("x_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 2048]> layers_5_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258233024))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260330240))))[name = string("layers_5_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_v_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_1795 = const()[name = string("op_1795"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<fp16, [1, 64, 8, 128]> var_1796_cast_fp16 = reshape(shape = var_1795, x = linear_37_cast_fp16)[name = string("op_1796_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 64, 128]> x_41_cast_fp16 = transpose(perm = x_41_perm_0, x = var_1772_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 16, 64, 128]> var_1800_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_1800_cast_fp16")];
            bool var_1803 = const()[name = string("op_1803"), val = bool(true)];
            tensor<int32, [1]> var_1804_axes_0 = const()[name = string("op_1804_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 64, 1]> var_1804_cast_fp16 = reduce_max(axes = var_1804_axes_0, keep_dims = var_1803, x = var_1800_cast_fp16)[name = string("op_1804_cast_fp16")];
            fp16 var_1805_to_fp16 = const()[name = string("op_1805_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_32_to_fp16 = const()[name = string("const_32_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 16, 64, 1]> clip_22_cast_fp16 = clip(alpha = var_1805_to_fp16, beta = const_32_to_fp16, x = var_1804_cast_fp16)[name = string("clip_22_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> y_43_cast_fp16 = real_div(x = x_41_cast_fp16, y = clip_22_cast_fp16)[name = string("y_43_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_1809_cast_fp16 = mul(x = y_43_cast_fp16, y = y_43_cast_fp16)[name = string("op_1809_cast_fp16")];
            tensor<int32, [1]> var_1814_axes_0 = const()[name = string("op_1814_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1814_keep_dims_0 = const()[name = string("op_1814_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 64, 1]> var_1814_cast_fp16 = reduce_mean(axes = var_1814_axes_0, keep_dims = var_1814_keep_dims_0, x = var_1809_cast_fp16)[name = string("op_1814_cast_fp16")];
            tensor<fp16, [1, 16, 64, 1]> var_1815_cast_fp16 = mul(x = clip_22_cast_fp16, y = clip_22_cast_fp16)[name = string("op_1815_cast_fp16")];
            fp32 var_1816_epsilon_0 = const()[name = string("op_1816_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 64, 1]> var_1816_cast_fp16 = inverse(epsilon = var_1816_epsilon_0, x = var_1815_cast_fp16)[name = string("op_1816_cast_fp16")];
            fp16 var_1817_to_fp16 = const()[name = string("op_1817_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 64, 1]> var_1818_cast_fp16 = mul(x = var_1816_cast_fp16, y = var_1817_to_fp16)[name = string("op_1818_cast_fp16")];
            tensor<fp16, [1, 16, 64, 1]> var_1820_cast_fp16 = add(x = var_1814_cast_fp16, y = var_1818_cast_fp16)[name = string("op_1820_cast_fp16")];
            fp32 var_1821_epsilon_0 = const()[name = string("op_1821_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 64, 1]> var_1821_cast_fp16 = rsqrt(epsilon = var_1821_epsilon_0, x = var_1820_cast_fp16)[name = string("op_1821_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_1822_cast_fp16 = mul(x = y_43_cast_fp16, y = var_1821_cast_fp16)[name = string("op_1822_cast_fp16")];
            tensor<fp16, [128]> layers_5_q_norm_to_fp16 = const()[name = string("layers_5_q_norm_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260332352)))];
            tensor<fp16, [1, 16, 64, 128]> q_21_cast_fp16 = mul(x = var_1822_cast_fp16, y = layers_5_q_norm_to_fp16)[name = string("q_21_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> x_43_cast_fp16 = transpose(perm = x_43_perm_0, x = var_1784_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 8, 64, 128]> var_1824_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_1824_cast_fp16")];
            bool var_1827 = const()[name = string("op_1827"), val = bool(true)];
            tensor<int32, [1]> var_1828_axes_0 = const()[name = string("op_1828_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 64, 1]> var_1828_cast_fp16 = reduce_max(axes = var_1828_axes_0, keep_dims = var_1827, x = var_1824_cast_fp16)[name = string("op_1828_cast_fp16")];
            fp16 var_1829_to_fp16 = const()[name = string("op_1829_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_33_to_fp16 = const()[name = string("const_33_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 8, 64, 1]> clip_23_cast_fp16 = clip(alpha = var_1829_to_fp16, beta = const_33_to_fp16, x = var_1828_cast_fp16)[name = string("clip_23_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> y_45_cast_fp16 = real_div(x = x_43_cast_fp16, y = clip_23_cast_fp16)[name = string("y_45_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_1833_cast_fp16 = mul(x = y_45_cast_fp16, y = y_45_cast_fp16)[name = string("op_1833_cast_fp16")];
            tensor<int32, [1]> var_1838_axes_0 = const()[name = string("op_1838_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1838_keep_dims_0 = const()[name = string("op_1838_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 64, 1]> var_1838_cast_fp16 = reduce_mean(axes = var_1838_axes_0, keep_dims = var_1838_keep_dims_0, x = var_1833_cast_fp16)[name = string("op_1838_cast_fp16")];
            tensor<fp16, [1, 8, 64, 1]> var_1839_cast_fp16 = mul(x = clip_23_cast_fp16, y = clip_23_cast_fp16)[name = string("op_1839_cast_fp16")];
            fp32 var_1840_epsilon_0 = const()[name = string("op_1840_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 64, 1]> var_1840_cast_fp16 = inverse(epsilon = var_1840_epsilon_0, x = var_1839_cast_fp16)[name = string("op_1840_cast_fp16")];
            fp16 var_1841_to_fp16 = const()[name = string("op_1841_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 64, 1]> var_1842_cast_fp16 = mul(x = var_1840_cast_fp16, y = var_1841_to_fp16)[name = string("op_1842_cast_fp16")];
            tensor<fp16, [1, 8, 64, 1]> var_1844_cast_fp16 = add(x = var_1838_cast_fp16, y = var_1842_cast_fp16)[name = string("op_1844_cast_fp16")];
            fp32 var_1845_epsilon_0 = const()[name = string("op_1845_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 64, 1]> var_1845_cast_fp16 = rsqrt(epsilon = var_1845_epsilon_0, x = var_1844_cast_fp16)[name = string("op_1845_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_1846_cast_fp16 = mul(x = y_45_cast_fp16, y = var_1845_cast_fp16)[name = string("op_1846_cast_fp16")];
            tensor<fp16, [128]> layers_5_k_norm_to_fp16 = const()[name = string("layers_5_k_norm_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260332672)))];
            tensor<fp16, [1, 8, 64, 128]> k_21_cast_fp16 = mul(x = var_1846_cast_fp16, y = layers_5_k_norm_to_fp16)[name = string("k_21_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_1848_cast_fp16 = mul(x = q_21_cast_fp16, y = cos)[name = string("op_1848_cast_fp16")];
            tensor<int32, [2]> var_1849_split_sizes_0 = const()[name = string("op_1849_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1849_axis_0 = const()[name = string("op_1849_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 64, 64]> var_1849_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_1849_cast_fp16_1 = split(axis = var_1849_axis_0, split_sizes = var_1849_split_sizes_0, x = q_21_cast_fp16)[name = string("op_1849_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_1851_cast_fp16 = mul(x = var_1849_cast_fp16_1, y = const_34_promoted_to_fp16)[name = string("op_1851_cast_fp16")];
            int32 var_1853 = const()[name = string("op_1853"), val = int32(-1)];
            bool var_1854_interleave_0 = const()[name = string("op_1854_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> var_1854_cast_fp16 = concat(axis = var_1853, interleave = var_1854_interleave_0, values = (var_1851_cast_fp16, var_1849_cast_fp16_0))[name = string("op_1854_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_1855_cast_fp16 = mul(x = var_1854_cast_fp16, y = sin)[name = string("op_1855_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> q_23_cast_fp16 = add(x = var_1848_cast_fp16, y = var_1855_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_1858_cast_fp16 = mul(x = k_21_cast_fp16, y = cos)[name = string("op_1858_cast_fp16")];
            tensor<int32, [2]> var_1859_split_sizes_0 = const()[name = string("op_1859_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1859_axis_0 = const()[name = string("op_1859_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 64, 64]> var_1859_cast_fp16_0, tensor<fp16, [1, 8, 64, 64]> var_1859_cast_fp16_1 = split(axis = var_1859_axis_0, split_sizes = var_1859_split_sizes_0, x = k_21_cast_fp16)[name = string("op_1859_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 64, 64]> var_1861_cast_fp16 = mul(x = var_1859_cast_fp16_1, y = const_35_promoted_to_fp16)[name = string("op_1861_cast_fp16")];
            int32 var_1863 = const()[name = string("op_1863"), val = int32(-1)];
            bool var_1864_interleave_0 = const()[name = string("op_1864_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_1864_cast_fp16 = concat(axis = var_1863, interleave = var_1864_interleave_0, values = (var_1861_cast_fp16, var_1859_cast_fp16_0))[name = string("op_1864_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_1865_cast_fp16 = mul(x = var_1864_cast_fp16, y = sin)[name = string("op_1865_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> k_23_cast_fp16 = add(x = var_1858_cast_fp16, y = var_1865_cast_fp16)[name = string("k_23_cast_fp16")];
            tensor<int32, [4]> var_1872_begin_0 = const()[name = string("op_1872_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_1872_end_0 = const()[name = string("op_1872_end_0"), val = tensor<int32, [4]>([6, 8, 704, 128])];
            tensor<bool, [4]> var_1872_end_mask_0 = const()[name = string("op_1872_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_1872_cast_fp16 = slice_by_index(begin = var_1872_begin_0, end = var_1872_end_0, end_mask = var_1872_end_mask_0, x = coreml_update_state_22)[name = string("op_1872_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_1884_cast_fp16 = mul(x = var_1872_cast_fp16, y = var_393_cast_fp16)[name = string("op_1884_cast_fp16")];
            tensor<int32, [4]> var_1888_begin_0 = const()[name = string("op_1888_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1888_end_0 = const()[name = string("op_1888_end_0"), val = tensor<int32, [4]>([1, 8, 64, 128])];
            tensor<bool, [4]> var_1888_end_mask_0 = const()[name = string("op_1888_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_1888_squeeze_mask_0 = const()[name = string("op_1888_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [8, 64, 128]> var_1888_cast_fp16 = slice_by_index(begin = var_1888_begin_0, end = var_1888_end_0, end_mask = var_1888_end_mask_0, squeeze_mask = var_1888_squeeze_mask_0, x = k_23_cast_fp16)[name = string("op_1888_cast_fp16")];
            bool var_1889_transpose_x_1 = const()[name = string("op_1889_transpose_x_1"), val = bool(true)];
            bool var_1889_transpose_y_1 = const()[name = string("op_1889_transpose_y_1"), val = bool(false)];
            tensor<fp16, [8, 704, 128]> var_1889_cast_fp16 = matmul(transpose_x = var_1889_transpose_x_1, transpose_y = var_1889_transpose_y_1, x = onehot, y = var_1888_cast_fp16)[name = string("op_1889_cast_fp16")];
            tensor<int32, [1]> var_1891_axes_0 = const()[name = string("op_1891_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8, 704, 128]> var_1891_cast_fp16 = expand_dims(axes = var_1891_axes_0, x = var_1889_cast_fp16)[name = string("op_1891_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_1893_cast_fp16 = add(x = var_1884_cast_fp16, y = var_1891_cast_fp16)[name = string("op_1893_cast_fp16")];
            tensor<int32, [4]> concat_40 = const()[name = string("concat_40"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> concat_41 = const()[name = string("concat_41"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> k_cache_internal_tensor_assign_6_stride_0 = const()[name = string("k_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("k_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("k_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("k_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [7, 8, 704, 128]> k_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_40, begin_mask = k_cache_internal_tensor_assign_6_begin_mask_0, end = concat_41, end_mask = k_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = k_cache_internal_tensor_assign_6_squeeze_mask_0, stride = k_cache_internal_tensor_assign_6_stride_0, update = var_1893_cast_fp16, x = coreml_update_state_22)[name = string("k_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = k_cache_internal_tensor_assign_6_cast_fp16, input = k_cache)[name = string("coreml_update_state_24_write_state")];
            tensor<fp16, [7, 8, 704, 128]> coreml_update_state_24 = read_state(input = k_cache)[name = string("coreml_update_state_24")];
            tensor<int32, [4]> var_1910_begin_0 = const()[name = string("op_1910_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_1910_end_0 = const()[name = string("op_1910_end_0"), val = tensor<int32, [4]>([6, 8, 704, 128])];
            tensor<bool, [4]> var_1910_end_mask_0 = const()[name = string("op_1910_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_1910_cast_fp16 = slice_by_index(begin = var_1910_begin_0, end = var_1910_end_0, end_mask = var_1910_end_mask_0, x = coreml_update_state_23)[name = string("op_1910_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_1922_cast_fp16 = mul(x = var_1910_cast_fp16, y = var_393_cast_fp16)[name = string("op_1922_cast_fp16")];
            tensor<int32, [4]> var_1926_begin_0 = const()[name = string("op_1926_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1926_end_0 = const()[name = string("op_1926_end_0"), val = tensor<int32, [4]>([1, 8, 64, 128])];
            tensor<bool, [4]> var_1926_end_mask_0 = const()[name = string("op_1926_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_1926_squeeze_mask_0 = const()[name = string("op_1926_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 8, 64, 128]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = var_1796_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [8, 64, 128]> var_1926_cast_fp16 = slice_by_index(begin = var_1926_begin_0, end = var_1926_end_0, end_mask = var_1926_end_mask_0, squeeze_mask = var_1926_squeeze_mask_0, x = v_11_cast_fp16)[name = string("op_1926_cast_fp16")];
            bool var_1927_transpose_x_1 = const()[name = string("op_1927_transpose_x_1"), val = bool(true)];
            bool var_1927_transpose_y_1 = const()[name = string("op_1927_transpose_y_1"), val = bool(false)];
            tensor<fp16, [8, 704, 128]> var_1927_cast_fp16 = matmul(transpose_x = var_1927_transpose_x_1, transpose_y = var_1927_transpose_y_1, x = onehot, y = var_1926_cast_fp16)[name = string("op_1927_cast_fp16")];
            tensor<int32, [1]> var_1929_axes_0 = const()[name = string("op_1929_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8, 704, 128]> var_1929_cast_fp16 = expand_dims(axes = var_1929_axes_0, x = var_1927_cast_fp16)[name = string("op_1929_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_1931_cast_fp16 = add(x = var_1922_cast_fp16, y = var_1929_cast_fp16)[name = string("op_1931_cast_fp16")];
            tensor<int32, [4]> concat_42 = const()[name = string("concat_42"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> concat_43 = const()[name = string("concat_43"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> v_cache_internal_tensor_assign_6_stride_0 = const()[name = string("v_cache_internal_tensor_assign_6_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_6_begin_mask_0 = const()[name = string("v_cache_internal_tensor_assign_6_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_6_end_mask_0 = const()[name = string("v_cache_internal_tensor_assign_6_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_6_squeeze_mask_0 = const()[name = string("v_cache_internal_tensor_assign_6_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [7, 8, 704, 128]> v_cache_internal_tensor_assign_6_cast_fp16 = slice_update(begin = concat_42, begin_mask = v_cache_internal_tensor_assign_6_begin_mask_0, end = concat_43, end_mask = v_cache_internal_tensor_assign_6_end_mask_0, squeeze_mask = v_cache_internal_tensor_assign_6_squeeze_mask_0, stride = v_cache_internal_tensor_assign_6_stride_0, update = var_1931_cast_fp16, x = coreml_update_state_23)[name = string("v_cache_internal_tensor_assign_6_cast_fp16")];
            write_state(data = v_cache_internal_tensor_assign_6_cast_fp16, input = v_cache)[name = string("coreml_update_state_25_write_state")];
            tensor<fp16, [7, 8, 704, 128]> coreml_update_state_25 = read_state(input = v_cache)[name = string("coreml_update_state_25")];
            tensor<int32, [4]> var_1948_begin_0 = const()[name = string("op_1948_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_1948_end_0 = const()[name = string("op_1948_end_0"), val = tensor<int32, [4]>([6, 8, 704, 128])];
            tensor<bool, [4]> var_1948_end_mask_0 = const()[name = string("op_1948_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_1948_cast_fp16 = slice_by_index(begin = var_1948_begin_0, end = var_1948_end_0, end_mask = var_1948_end_mask_0, x = coreml_update_state_24)[name = string("op_1948_cast_fp16")];
            tensor<int32, [4]> transpose_20_perm_0 = const()[name = string("transpose_20_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_10_reps_0 = const()[name = string("tile_10_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [8, 1, 704, 128]> transpose_20_cast_fp16 = transpose(perm = transpose_20_perm_0, x = var_1948_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [16, 1, 704, 128]> tile_10_cast_fp16 = tile(reps = tile_10_reps_0, x = transpose_20_cast_fp16)[name = string("tile_10_cast_fp16")];
            tensor<int32, [5]> concat_44 = const()[name = string("concat_44"), val = tensor<int32, [5]>([2, 8, 1, 704, 128])];
            tensor<fp16, [2, 8, 1, 704, 128]> reshape_20_cast_fp16 = reshape(shape = concat_44, x = tile_10_cast_fp16)[name = string("reshape_20_cast_fp16")];
            tensor<int32, [5]> transpose_21_perm_0 = const()[name = string("transpose_21_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_45 = const()[name = string("concat_45"), val = tensor<int32, [4]>([-1, 1, 704, 128])];
            tensor<fp16, [8, 2, 1, 704, 128]> transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = reshape_20_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [16, 1, 704, 128]> reshape_21_cast_fp16 = reshape(shape = concat_45, x = transpose_21_cast_fp16)[name = string("reshape_21_cast_fp16")];
            tensor<int32, [4]> transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> var_1962_begin_0 = const()[name = string("op_1962_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_1962_end_0 = const()[name = string("op_1962_end_0"), val = tensor<int32, [4]>([6, 8, 704, 128])];
            tensor<bool, [4]> var_1962_end_mask_0 = const()[name = string("op_1962_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_1962_cast_fp16 = slice_by_index(begin = var_1962_begin_0, end = var_1962_end_0, end_mask = var_1962_end_mask_0, x = coreml_update_state_25)[name = string("op_1962_cast_fp16")];
            tensor<int32, [4]> transpose_22_perm_0 = const()[name = string("transpose_22_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_11_reps_0 = const()[name = string("tile_11_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [8, 1, 704, 128]> transpose_22_cast_fp16 = transpose(perm = transpose_22_perm_0, x = var_1962_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [16, 1, 704, 128]> tile_11_cast_fp16 = tile(reps = tile_11_reps_0, x = transpose_22_cast_fp16)[name = string("tile_11_cast_fp16")];
            tensor<int32, [5]> concat_46 = const()[name = string("concat_46"), val = tensor<int32, [5]>([2, 8, 1, 704, 128])];
            tensor<fp16, [2, 8, 1, 704, 128]> reshape_22_cast_fp16 = reshape(shape = concat_46, x = tile_11_cast_fp16)[name = string("reshape_22_cast_fp16")];
            tensor<int32, [5]> transpose_23_perm_0 = const()[name = string("transpose_23_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_47 = const()[name = string("concat_47"), val = tensor<int32, [4]>([-1, 1, 704, 128])];
            tensor<fp16, [8, 2, 1, 704, 128]> transpose_23_cast_fp16 = transpose(perm = transpose_23_perm_0, x = reshape_22_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [16, 1, 704, 128]> reshape_23_cast_fp16 = reshape(shape = concat_47, x = transpose_23_cast_fp16)[name = string("reshape_23_cast_fp16")];
            tensor<int32, [4]> values_11_perm_0 = const()[name = string("values_11_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_1975_transpose_x_0 = const()[name = string("op_1975_transpose_x_0"), val = bool(false)];
            bool var_1975_transpose_y_0 = const()[name = string("op_1975_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 704]> transpose_33_cast_fp16 = transpose(perm = transpose_33_perm_0, x = reshape_21_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 16, 64, 704]> var_1975_cast_fp16 = matmul(transpose_x = var_1975_transpose_x_0, transpose_y = var_1975_transpose_y_0, x = q_23_cast_fp16, y = transpose_33_cast_fp16)[name = string("op_1975_cast_fp16")];
            fp16 var_1976_to_fp16 = const()[name = string("op_1976_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 64, 704]> var_1977_cast_fp16 = mul(x = var_1975_cast_fp16, y = var_1976_to_fp16)[name = string("op_1977_cast_fp16")];
            tensor<fp16, [1, 16, 64, 704]> scores_11_cast_fp16 = add(x = var_1977_cast_fp16, y = mask)[name = string("scores_11_cast_fp16")];
            int32 var_1980 = const()[name = string("op_1980"), val = int32(-1)];
            tensor<fp16, [1, 16, 64, 704]> var_1982_cast_fp16 = softmax(axis = var_1980, x = scores_11_cast_fp16)[name = string("op_1982_cast_fp16")];
            bool attn_11_transpose_x_0 = const()[name = string("attn_11_transpose_x_0"), val = bool(false)];
            bool attn_11_transpose_y_0 = const()[name = string("attn_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 704, 128]> values_11_cast_fp16 = transpose(perm = values_11_perm_0, x = reshape_23_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 16, 64, 128]> attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_1982_cast_fp16, y = values_11_cast_fp16)[name = string("attn_11_cast_fp16")];
            tensor<int32, [4]> var_1986_perm_0 = const()[name = string("op_1986_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1990 = const()[name = string("op_1990"), val = tensor<int32, [3]>([1, 64, 2048])];
            tensor<fp16, [1, 64, 16, 128]> var_1986_cast_fp16 = transpose(perm = var_1986_perm_0, x = attn_11_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 64, 2048]> input_53_cast_fp16 = reshape(shape = var_1990, x = var_1986_cast_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [2048, 2048]> layers_5_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260332992))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264527360))))[name = string("layers_5_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_o_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> x_45_cast_fp16 = add(x = x_39_cast_fp16, y = linear_38_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_1997_cast_fp16 = abs(x = x_45_cast_fp16)[name = string("op_1997_cast_fp16")];
            bool var_2000 = const()[name = string("op_2000"), val = bool(true)];
            tensor<int32, [1]> var_2001_axes_0 = const()[name = string("op_2001_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2001_cast_fp16 = reduce_max(axes = var_2001_axes_0, keep_dims = var_2000, x = var_1997_cast_fp16)[name = string("op_2001_cast_fp16")];
            fp16 var_2002_to_fp16 = const()[name = string("op_2002_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_36_to_fp16 = const()[name = string("const_36_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_24_cast_fp16 = clip(alpha = var_2002_to_fp16, beta = const_36_to_fp16, x = var_2001_cast_fp16)[name = string("clip_24_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> y_47_cast_fp16 = real_div(x = x_45_cast_fp16, y = clip_24_cast_fp16)[name = string("y_47_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_2006_cast_fp16 = mul(x = y_47_cast_fp16, y = y_47_cast_fp16)[name = string("op_2006_cast_fp16")];
            tensor<int32, [1]> var_2011_axes_0 = const()[name = string("op_2011_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2011_keep_dims_0 = const()[name = string("op_2011_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_2011_cast_fp16 = reduce_mean(axes = var_2011_axes_0, keep_dims = var_2011_keep_dims_0, x = var_2006_cast_fp16)[name = string("op_2011_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2012_cast_fp16 = mul(x = clip_24_cast_fp16, y = clip_24_cast_fp16)[name = string("op_2012_cast_fp16")];
            fp32 var_2013_epsilon_0 = const()[name = string("op_2013_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2013_cast_fp16 = inverse(epsilon = var_2013_epsilon_0, x = var_2012_cast_fp16)[name = string("op_2013_cast_fp16")];
            fp16 var_2014_to_fp16 = const()[name = string("op_2014_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2015_cast_fp16 = mul(x = var_2013_cast_fp16, y = var_2014_to_fp16)[name = string("op_2015_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2017_cast_fp16 = add(x = var_2011_cast_fp16, y = var_2015_cast_fp16)[name = string("op_2017_cast_fp16")];
            fp32 var_2018_epsilon_0 = const()[name = string("op_2018_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2018_cast_fp16 = rsqrt(epsilon = var_2018_epsilon_0, x = var_2017_cast_fp16)[name = string("op_2018_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_2019_cast_fp16 = mul(x = y_47_cast_fp16, y = var_2018_cast_fp16)[name = string("op_2019_cast_fp16")];
            tensor<fp16, [2048]> layers_5_post_norm_to_fp16 = const()[name = string("layers_5_post_norm_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264531520)))];
            tensor<fp16, [1, 64, 2048]> input_55_cast_fp16 = mul(x = var_2019_cast_fp16, y = layers_5_post_norm_to_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [6144, 2048]> layers_5_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264535680))), scale = tensor<fp16, [6144, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277118656))))[name = string("layers_5_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 6144]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_5_gate_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 64, 6144]> var_2024_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_2024_cast_fp16")];
            tensor<fp16, [6144, 2048]> layers_5_up_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277131008))), scale = tensor<fp16, [6144, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289713984))))[name = string("layers_5_up_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 6144]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_5_up_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 64, 6144]> var_2028_cast_fp16 = mul(x = var_2024_cast_fp16, y = linear_40_cast_fp16)[name = string("op_2028_cast_fp16")];
            fp16 var_2029_to_fp16 = const()[name = string("op_2029_to_fp16"), val = fp16(0x1p-8)];
            tensor<fp16, [1, 64, 6144]> input_59_cast_fp16 = mul(x = var_2028_cast_fp16, y = var_2029_to_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [2048, 6144]> layers_5_down_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289726336))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302309312))))[name = string("layers_5_down_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_41_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_down_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_41_cast_fp16")];
            fp16 var_2034_to_fp16 = const()[name = string("op_2034_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 64, 2048]> var_2035_cast_fp16 = mul(x = linear_41_cast_fp16, y = var_2034_to_fp16)[name = string("op_2035_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> x_47_cast_fp16 = add(x = x_45_cast_fp16, y = var_2035_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_2038_cast_fp16 = abs(x = x_47_cast_fp16)[name = string("op_2038_cast_fp16")];
            bool var_2041 = const()[name = string("op_2041"), val = bool(true)];
            tensor<int32, [1]> var_2042_axes_0 = const()[name = string("op_2042_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2042_cast_fp16 = reduce_max(axes = var_2042_axes_0, keep_dims = var_2041, x = var_2038_cast_fp16)[name = string("op_2042_cast_fp16")];
            fp16 var_2043_to_fp16 = const()[name = string("op_2043_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_37_to_fp16 = const()[name = string("const_37_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_25_cast_fp16 = clip(alpha = var_2043_to_fp16, beta = const_37_to_fp16, x = var_2042_cast_fp16)[name = string("clip_25_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> y_49_cast_fp16 = real_div(x = x_47_cast_fp16, y = clip_25_cast_fp16)[name = string("y_49_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_2047_cast_fp16 = mul(x = y_49_cast_fp16, y = y_49_cast_fp16)[name = string("op_2047_cast_fp16")];
            tensor<int32, [1]> var_2052_axes_0 = const()[name = string("op_2052_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2052_keep_dims_0 = const()[name = string("op_2052_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_2052_cast_fp16 = reduce_mean(axes = var_2052_axes_0, keep_dims = var_2052_keep_dims_0, x = var_2047_cast_fp16)[name = string("op_2052_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2053_cast_fp16 = mul(x = clip_25_cast_fp16, y = clip_25_cast_fp16)[name = string("op_2053_cast_fp16")];
            fp32 var_2054_epsilon_0 = const()[name = string("op_2054_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2054_cast_fp16 = inverse(epsilon = var_2054_epsilon_0, x = var_2053_cast_fp16)[name = string("op_2054_cast_fp16")];
            fp16 var_2055_to_fp16 = const()[name = string("op_2055_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2056_cast_fp16 = mul(x = var_2054_cast_fp16, y = var_2055_to_fp16)[name = string("op_2056_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2058_cast_fp16 = add(x = var_2052_cast_fp16, y = var_2056_cast_fp16)[name = string("op_2058_cast_fp16")];
            fp32 var_2059_epsilon_0 = const()[name = string("op_2059_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2059_cast_fp16 = rsqrt(epsilon = var_2059_epsilon_0, x = var_2058_cast_fp16)[name = string("op_2059_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_2060_cast_fp16 = mul(x = y_49_cast_fp16, y = var_2059_cast_fp16)[name = string("op_2060_cast_fp16")];
            tensor<fp16, [2048]> layers_6_input_norm_to_fp16 = const()[name = string("layers_6_input_norm_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302313472)))];
            tensor<fp16, [1, 64, 2048]> input_61_cast_fp16 = mul(x = var_2060_cast_fp16, y = layers_6_input_norm_to_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [2048, 2048]> layers_6_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302317632))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306512000))))[name = string("layers_6_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_q_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> var_2069 = const()[name = string("op_2069"), val = tensor<int32, [4]>([1, 64, 16, 128])];
            tensor<fp16, [1, 64, 16, 128]> var_2070_cast_fp16 = reshape(shape = var_2069, x = linear_42_cast_fp16)[name = string("op_2070_cast_fp16")];
            tensor<int32, [4]> x_49_perm_0 = const()[name = string("x_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 2048]> layers_6_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306516160))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308613376))))[name = string("layers_6_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_k_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_2081 = const()[name = string("op_2081"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<fp16, [1, 64, 8, 128]> var_2082_cast_fp16 = reshape(shape = var_2081, x = linear_43_cast_fp16)[name = string("op_2082_cast_fp16")];
            tensor<int32, [4]> x_51_perm_0 = const()[name = string("x_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 2048]> layers_6_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308615488))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310712704))))[name = string("layers_6_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_v_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_2093 = const()[name = string("op_2093"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<fp16, [1, 64, 8, 128]> var_2094_cast_fp16 = reshape(shape = var_2093, x = linear_44_cast_fp16)[name = string("op_2094_cast_fp16")];
            tensor<int32, [4]> v_13_perm_0 = const()[name = string("v_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 16, 64, 128]> x_49_cast_fp16 = transpose(perm = x_49_perm_0, x = var_2070_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 16, 64, 128]> var_2098_cast_fp16 = abs(x = x_49_cast_fp16)[name = string("op_2098_cast_fp16")];
            bool var_2101 = const()[name = string("op_2101"), val = bool(true)];
            tensor<int32, [1]> var_2102_axes_0 = const()[name = string("op_2102_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 64, 1]> var_2102_cast_fp16 = reduce_max(axes = var_2102_axes_0, keep_dims = var_2101, x = var_2098_cast_fp16)[name = string("op_2102_cast_fp16")];
            fp16 var_2103_to_fp16 = const()[name = string("op_2103_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_38_to_fp16 = const()[name = string("const_38_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 16, 64, 1]> clip_26_cast_fp16 = clip(alpha = var_2103_to_fp16, beta = const_38_to_fp16, x = var_2102_cast_fp16)[name = string("clip_26_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> y_51_cast_fp16 = real_div(x = x_49_cast_fp16, y = clip_26_cast_fp16)[name = string("y_51_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_2107_cast_fp16 = mul(x = y_51_cast_fp16, y = y_51_cast_fp16)[name = string("op_2107_cast_fp16")];
            tensor<int32, [1]> var_2112_axes_0 = const()[name = string("op_2112_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2112_keep_dims_0 = const()[name = string("op_2112_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 64, 1]> var_2112_cast_fp16 = reduce_mean(axes = var_2112_axes_0, keep_dims = var_2112_keep_dims_0, x = var_2107_cast_fp16)[name = string("op_2112_cast_fp16")];
            tensor<fp16, [1, 16, 64, 1]> var_2113_cast_fp16 = mul(x = clip_26_cast_fp16, y = clip_26_cast_fp16)[name = string("op_2113_cast_fp16")];
            fp32 var_2114_epsilon_0 = const()[name = string("op_2114_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 64, 1]> var_2114_cast_fp16 = inverse(epsilon = var_2114_epsilon_0, x = var_2113_cast_fp16)[name = string("op_2114_cast_fp16")];
            fp16 var_2115_to_fp16 = const()[name = string("op_2115_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 64, 1]> var_2116_cast_fp16 = mul(x = var_2114_cast_fp16, y = var_2115_to_fp16)[name = string("op_2116_cast_fp16")];
            tensor<fp16, [1, 16, 64, 1]> var_2118_cast_fp16 = add(x = var_2112_cast_fp16, y = var_2116_cast_fp16)[name = string("op_2118_cast_fp16")];
            fp32 var_2119_epsilon_0 = const()[name = string("op_2119_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 64, 1]> var_2119_cast_fp16 = rsqrt(epsilon = var_2119_epsilon_0, x = var_2118_cast_fp16)[name = string("op_2119_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_2120_cast_fp16 = mul(x = y_51_cast_fp16, y = var_2119_cast_fp16)[name = string("op_2120_cast_fp16")];
            tensor<fp16, [128]> layers_6_q_norm_to_fp16 = const()[name = string("layers_6_q_norm_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310714816)))];
            tensor<fp16, [1, 16, 64, 128]> q_25_cast_fp16 = mul(x = var_2120_cast_fp16, y = layers_6_q_norm_to_fp16)[name = string("q_25_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> x_51_cast_fp16 = transpose(perm = x_51_perm_0, x = var_2082_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 8, 64, 128]> var_2122_cast_fp16 = abs(x = x_51_cast_fp16)[name = string("op_2122_cast_fp16")];
            bool var_2125 = const()[name = string("op_2125"), val = bool(true)];
            tensor<int32, [1]> var_2126_axes_0 = const()[name = string("op_2126_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 64, 1]> var_2126_cast_fp16 = reduce_max(axes = var_2126_axes_0, keep_dims = var_2125, x = var_2122_cast_fp16)[name = string("op_2126_cast_fp16")];
            fp16 var_2127_to_fp16 = const()[name = string("op_2127_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_39_to_fp16 = const()[name = string("const_39_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 8, 64, 1]> clip_27_cast_fp16 = clip(alpha = var_2127_to_fp16, beta = const_39_to_fp16, x = var_2126_cast_fp16)[name = string("clip_27_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> y_53_cast_fp16 = real_div(x = x_51_cast_fp16, y = clip_27_cast_fp16)[name = string("y_53_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_2131_cast_fp16 = mul(x = y_53_cast_fp16, y = y_53_cast_fp16)[name = string("op_2131_cast_fp16")];
            tensor<int32, [1]> var_2136_axes_0 = const()[name = string("op_2136_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2136_keep_dims_0 = const()[name = string("op_2136_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 64, 1]> var_2136_cast_fp16 = reduce_mean(axes = var_2136_axes_0, keep_dims = var_2136_keep_dims_0, x = var_2131_cast_fp16)[name = string("op_2136_cast_fp16")];
            tensor<fp16, [1, 8, 64, 1]> var_2137_cast_fp16 = mul(x = clip_27_cast_fp16, y = clip_27_cast_fp16)[name = string("op_2137_cast_fp16")];
            fp32 var_2138_epsilon_0 = const()[name = string("op_2138_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 64, 1]> var_2138_cast_fp16 = inverse(epsilon = var_2138_epsilon_0, x = var_2137_cast_fp16)[name = string("op_2138_cast_fp16")];
            fp16 var_2139_to_fp16 = const()[name = string("op_2139_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 64, 1]> var_2140_cast_fp16 = mul(x = var_2138_cast_fp16, y = var_2139_to_fp16)[name = string("op_2140_cast_fp16")];
            tensor<fp16, [1, 8, 64, 1]> var_2142_cast_fp16 = add(x = var_2136_cast_fp16, y = var_2140_cast_fp16)[name = string("op_2142_cast_fp16")];
            fp32 var_2143_epsilon_0 = const()[name = string("op_2143_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 64, 1]> var_2143_cast_fp16 = rsqrt(epsilon = var_2143_epsilon_0, x = var_2142_cast_fp16)[name = string("op_2143_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_2144_cast_fp16 = mul(x = y_53_cast_fp16, y = var_2143_cast_fp16)[name = string("op_2144_cast_fp16")];
            tensor<fp16, [128]> layers_6_k_norm_to_fp16 = const()[name = string("layers_6_k_norm_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310715136)))];
            tensor<fp16, [1, 8, 64, 128]> k_25_cast_fp16 = mul(x = var_2144_cast_fp16, y = layers_6_k_norm_to_fp16)[name = string("k_25_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_2146_cast_fp16 = mul(x = q_25_cast_fp16, y = cos)[name = string("op_2146_cast_fp16")];
            tensor<int32, [2]> var_2147_split_sizes_0 = const()[name = string("op_2147_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2147_axis_0 = const()[name = string("op_2147_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 64, 64]> var_2147_cast_fp16_0, tensor<fp16, [1, 16, 64, 64]> var_2147_cast_fp16_1 = split(axis = var_2147_axis_0, split_sizes = var_2147_split_sizes_0, x = q_25_cast_fp16)[name = string("op_2147_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 64, 64]> var_2149_cast_fp16 = mul(x = var_2147_cast_fp16_1, y = const_40_promoted_to_fp16)[name = string("op_2149_cast_fp16")];
            int32 var_2151 = const()[name = string("op_2151"), val = int32(-1)];
            bool var_2152_interleave_0 = const()[name = string("op_2152_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 64, 128]> var_2152_cast_fp16 = concat(axis = var_2151, interleave = var_2152_interleave_0, values = (var_2149_cast_fp16, var_2147_cast_fp16_0))[name = string("op_2152_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> var_2153_cast_fp16 = mul(x = var_2152_cast_fp16, y = sin)[name = string("op_2153_cast_fp16")];
            tensor<fp16, [1, 16, 64, 128]> q_27_cast_fp16 = add(x = var_2146_cast_fp16, y = var_2153_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_2156_cast_fp16 = mul(x = k_25_cast_fp16, y = cos)[name = string("op_2156_cast_fp16")];
            tensor<int32, [2]> var_2157_split_sizes_0 = const()[name = string("op_2157_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2157_axis_0 = const()[name = string("op_2157_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 64, 64]> var_2157_cast_fp16_0, tensor<fp16, [1, 8, 64, 64]> var_2157_cast_fp16_1 = split(axis = var_2157_axis_0, split_sizes = var_2157_split_sizes_0, x = k_25_cast_fp16)[name = string("op_2157_cast_fp16")];
            fp16 const_41_promoted_to_fp16 = const()[name = string("const_41_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 64, 64]> var_2159_cast_fp16 = mul(x = var_2157_cast_fp16_1, y = const_41_promoted_to_fp16)[name = string("op_2159_cast_fp16")];
            int32 var_2161 = const()[name = string("op_2161"), val = int32(-1)];
            bool var_2162_interleave_0 = const()[name = string("op_2162_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 128]> var_2162_cast_fp16 = concat(axis = var_2161, interleave = var_2162_interleave_0, values = (var_2159_cast_fp16, var_2157_cast_fp16_0))[name = string("op_2162_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> var_2163_cast_fp16 = mul(x = var_2162_cast_fp16, y = sin)[name = string("op_2163_cast_fp16")];
            tensor<fp16, [1, 8, 64, 128]> k_27_cast_fp16 = add(x = var_2156_cast_fp16, y = var_2163_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<int32, [4]> var_2170_begin_0 = const()[name = string("op_2170_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2170_end_0 = const()[name = string("op_2170_end_0"), val = tensor<int32, [4]>([1, 8, 704, 128])];
            tensor<bool, [4]> var_2170_end_mask_0 = const()[name = string("op_2170_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_2170_cast_fp16 = slice_by_index(begin = var_2170_begin_0, end = var_2170_end_0, end_mask = var_2170_end_mask_0, x = coreml_update_state_24)[name = string("op_2170_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_2182_cast_fp16 = mul(x = var_2170_cast_fp16, y = var_393_cast_fp16)[name = string("op_2182_cast_fp16")];
            tensor<int32, [4]> var_2186_begin_0 = const()[name = string("op_2186_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2186_end_0 = const()[name = string("op_2186_end_0"), val = tensor<int32, [4]>([1, 8, 64, 128])];
            tensor<bool, [4]> var_2186_end_mask_0 = const()[name = string("op_2186_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_2186_squeeze_mask_0 = const()[name = string("op_2186_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [8, 64, 128]> var_2186_cast_fp16 = slice_by_index(begin = var_2186_begin_0, end = var_2186_end_0, end_mask = var_2186_end_mask_0, squeeze_mask = var_2186_squeeze_mask_0, x = k_27_cast_fp16)[name = string("op_2186_cast_fp16")];
            bool var_2187_transpose_x_1 = const()[name = string("op_2187_transpose_x_1"), val = bool(true)];
            bool var_2187_transpose_y_1 = const()[name = string("op_2187_transpose_y_1"), val = bool(false)];
            tensor<fp16, [8, 704, 128]> var_2187_cast_fp16 = matmul(transpose_x = var_2187_transpose_x_1, transpose_y = var_2187_transpose_y_1, x = onehot, y = var_2186_cast_fp16)[name = string("op_2187_cast_fp16")];
            tensor<int32, [1]> var_2189_axes_0 = const()[name = string("op_2189_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8, 704, 128]> var_2189_cast_fp16 = expand_dims(axes = var_2189_axes_0, x = var_2187_cast_fp16)[name = string("op_2189_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_2191_cast_fp16 = add(x = var_2182_cast_fp16, y = var_2189_cast_fp16)[name = string("op_2191_cast_fp16")];
            tensor<int32, [4]> concat_48 = const()[name = string("concat_48"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> concat_49 = const()[name = string("concat_49"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> k_cache_internal_tensor_assign_7_stride_0 = const()[name = string("k_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("k_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("k_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> k_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("k_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [7, 8, 704, 128]> k_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_48, begin_mask = k_cache_internal_tensor_assign_7_begin_mask_0, end = concat_49, end_mask = k_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = k_cache_internal_tensor_assign_7_squeeze_mask_0, stride = k_cache_internal_tensor_assign_7_stride_0, update = var_2191_cast_fp16, x = coreml_update_state_24)[name = string("k_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = k_cache_internal_tensor_assign_7_cast_fp16, input = k_cache)[name = string("coreml_update_state_26_write_state")];
            tensor<fp16, [7, 8, 704, 128]> coreml_update_state_26 = read_state(input = k_cache)[name = string("coreml_update_state_26")];
            tensor<int32, [4]> var_2208_begin_0 = const()[name = string("op_2208_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2208_end_0 = const()[name = string("op_2208_end_0"), val = tensor<int32, [4]>([1, 8, 704, 128])];
            tensor<bool, [4]> var_2208_end_mask_0 = const()[name = string("op_2208_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_2208_cast_fp16 = slice_by_index(begin = var_2208_begin_0, end = var_2208_end_0, end_mask = var_2208_end_mask_0, x = coreml_update_state_25)[name = string("op_2208_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_2220_cast_fp16 = mul(x = var_2208_cast_fp16, y = var_393_cast_fp16)[name = string("op_2220_cast_fp16")];
            tensor<int32, [4]> var_2224_begin_0 = const()[name = string("op_2224_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2224_end_0 = const()[name = string("op_2224_end_0"), val = tensor<int32, [4]>([1, 8, 64, 128])];
            tensor<bool, [4]> var_2224_end_mask_0 = const()[name = string("op_2224_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_2224_squeeze_mask_0 = const()[name = string("op_2224_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [1, 8, 64, 128]> v_13_cast_fp16 = transpose(perm = v_13_perm_0, x = var_2094_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [8, 64, 128]> var_2224_cast_fp16 = slice_by_index(begin = var_2224_begin_0, end = var_2224_end_0, end_mask = var_2224_end_mask_0, squeeze_mask = var_2224_squeeze_mask_0, x = v_13_cast_fp16)[name = string("op_2224_cast_fp16")];
            bool var_2225_transpose_x_1 = const()[name = string("op_2225_transpose_x_1"), val = bool(true)];
            bool var_2225_transpose_y_1 = const()[name = string("op_2225_transpose_y_1"), val = bool(false)];
            tensor<fp16, [8, 704, 128]> var_2225_cast_fp16 = matmul(transpose_x = var_2225_transpose_x_1, transpose_y = var_2225_transpose_y_1, x = onehot, y = var_2224_cast_fp16)[name = string("op_2225_cast_fp16")];
            tensor<int32, [1]> var_2227_axes_0 = const()[name = string("op_2227_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 8, 704, 128]> var_2227_cast_fp16 = expand_dims(axes = var_2227_axes_0, x = var_2225_cast_fp16)[name = string("op_2227_cast_fp16")];
            tensor<fp16, [1, 8, 704, 128]> var_2229_cast_fp16 = add(x = var_2220_cast_fp16, y = var_2227_cast_fp16)[name = string("op_2229_cast_fp16")];
            tensor<int32, [4]> concat_50 = const()[name = string("concat_50"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> concat_51 = const()[name = string("concat_51"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> v_cache_internal_tensor_assign_7_stride_0 = const()[name = string("v_cache_internal_tensor_assign_7_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_7_begin_mask_0 = const()[name = string("v_cache_internal_tensor_assign_7_begin_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_7_end_mask_0 = const()[name = string("v_cache_internal_tensor_assign_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> v_cache_internal_tensor_assign_7_squeeze_mask_0 = const()[name = string("v_cache_internal_tensor_assign_7_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, false, false])];
            tensor<fp16, [7, 8, 704, 128]> v_cache_internal_tensor_assign_7_cast_fp16 = slice_update(begin = concat_50, begin_mask = v_cache_internal_tensor_assign_7_begin_mask_0, end = concat_51, end_mask = v_cache_internal_tensor_assign_7_end_mask_0, squeeze_mask = v_cache_internal_tensor_assign_7_squeeze_mask_0, stride = v_cache_internal_tensor_assign_7_stride_0, update = var_2229_cast_fp16, x = coreml_update_state_25)[name = string("v_cache_internal_tensor_assign_7_cast_fp16")];
            write_state(data = v_cache_internal_tensor_assign_7_cast_fp16, input = v_cache)[name = string("coreml_update_state_27_write_state")];
            tensor<fp16, [7, 8, 704, 128]> coreml_update_state_27 = read_state(input = v_cache)[name = string("coreml_update_state_27")];
            tensor<int32, [4]> var_2246_begin_0 = const()[name = string("op_2246_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2246_end_0 = const()[name = string("op_2246_end_0"), val = tensor<int32, [4]>([1, 8, 704, 128])];
            tensor<bool, [4]> var_2246_end_mask_0 = const()[name = string("op_2246_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_2246_cast_fp16 = slice_by_index(begin = var_2246_begin_0, end = var_2246_end_0, end_mask = var_2246_end_mask_0, x = coreml_update_state_26)[name = string("op_2246_cast_fp16")];
            tensor<int32, [4]> transpose_24_perm_0 = const()[name = string("transpose_24_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_12_reps_0 = const()[name = string("tile_12_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [8, 1, 704, 128]> transpose_24_cast_fp16 = transpose(perm = transpose_24_perm_0, x = var_2246_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [16, 1, 704, 128]> tile_12_cast_fp16 = tile(reps = tile_12_reps_0, x = transpose_24_cast_fp16)[name = string("tile_12_cast_fp16")];
            tensor<int32, [5]> concat_52 = const()[name = string("concat_52"), val = tensor<int32, [5]>([2, 8, 1, 704, 128])];
            tensor<fp16, [2, 8, 1, 704, 128]> reshape_24_cast_fp16 = reshape(shape = concat_52, x = tile_12_cast_fp16)[name = string("reshape_24_cast_fp16")];
            tensor<int32, [5]> transpose_25_perm_0 = const()[name = string("transpose_25_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_53 = const()[name = string("concat_53"), val = tensor<int32, [4]>([-1, 1, 704, 128])];
            tensor<fp16, [8, 2, 1, 704, 128]> transpose_25_cast_fp16 = transpose(perm = transpose_25_perm_0, x = reshape_24_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [16, 1, 704, 128]> reshape_25_cast_fp16 = reshape(shape = concat_53, x = transpose_25_cast_fp16)[name = string("reshape_25_cast_fp16")];
            tensor<int32, [4]> transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> var_2260_begin_0 = const()[name = string("op_2260_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_2260_end_0 = const()[name = string("op_2260_end_0"), val = tensor<int32, [4]>([1, 8, 704, 128])];
            tensor<bool, [4]> var_2260_end_mask_0 = const()[name = string("op_2260_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 704, 128]> var_2260_cast_fp16 = slice_by_index(begin = var_2260_begin_0, end = var_2260_end_0, end_mask = var_2260_end_mask_0, x = coreml_update_state_27)[name = string("op_2260_cast_fp16")];
            tensor<int32, [4]> transpose_26_perm_0 = const()[name = string("transpose_26_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_13_reps_0 = const()[name = string("tile_13_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [8, 1, 704, 128]> transpose_26_cast_fp16 = transpose(perm = transpose_26_perm_0, x = var_2260_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [16, 1, 704, 128]> tile_13_cast_fp16 = tile(reps = tile_13_reps_0, x = transpose_26_cast_fp16)[name = string("tile_13_cast_fp16")];
            tensor<int32, [5]> concat_54 = const()[name = string("concat_54"), val = tensor<int32, [5]>([2, 8, 1, 704, 128])];
            tensor<fp16, [2, 8, 1, 704, 128]> reshape_26_cast_fp16 = reshape(shape = concat_54, x = tile_13_cast_fp16)[name = string("reshape_26_cast_fp16")];
            tensor<int32, [5]> transpose_27_perm_0 = const()[name = string("transpose_27_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_55 = const()[name = string("concat_55"), val = tensor<int32, [4]>([-1, 1, 704, 128])];
            tensor<fp16, [8, 2, 1, 704, 128]> transpose_27_cast_fp16 = transpose(perm = transpose_27_perm_0, x = reshape_26_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [16, 1, 704, 128]> reshape_27_cast_fp16 = reshape(shape = concat_55, x = transpose_27_cast_fp16)[name = string("reshape_27_cast_fp16")];
            tensor<int32, [4]> values_perm_0 = const()[name = string("values_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_2273_transpose_x_0 = const()[name = string("op_2273_transpose_x_0"), val = bool(false)];
            bool var_2273_transpose_y_0 = const()[name = string("op_2273_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 704]> transpose_34_cast_fp16 = transpose(perm = transpose_34_perm_0, x = reshape_25_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 16, 64, 704]> var_2273_cast_fp16 = matmul(transpose_x = var_2273_transpose_x_0, transpose_y = var_2273_transpose_y_0, x = q_27_cast_fp16, y = transpose_34_cast_fp16)[name = string("op_2273_cast_fp16")];
            fp16 var_2274_to_fp16 = const()[name = string("op_2274_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 64, 704]> var_2275_cast_fp16 = mul(x = var_2273_cast_fp16, y = var_2274_to_fp16)[name = string("op_2275_cast_fp16")];
            tensor<fp16, [1, 16, 64, 704]> scores_cast_fp16 = add(x = var_2275_cast_fp16, y = mask)[name = string("scores_cast_fp16")];
            int32 var_2278 = const()[name = string("op_2278"), val = int32(-1)];
            tensor<fp16, [1, 16, 64, 704]> var_2280_cast_fp16 = softmax(axis = var_2278, x = scores_cast_fp16)[name = string("op_2280_cast_fp16")];
            bool attn_transpose_x_0 = const()[name = string("attn_transpose_x_0"), val = bool(false)];
            bool attn_transpose_y_0 = const()[name = string("attn_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 704, 128]> values_cast_fp16 = transpose(perm = values_perm_0, x = reshape_27_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 16, 64, 128]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_2280_cast_fp16, y = values_cast_fp16)[name = string("attn_cast_fp16")];
            tensor<int32, [4]> var_2284_perm_0 = const()[name = string("op_2284_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2288 = const()[name = string("op_2288"), val = tensor<int32, [3]>([1, 64, 2048])];
            tensor<fp16, [1, 64, 16, 128]> var_2284_cast_fp16 = transpose(perm = var_2284_perm_0, x = attn_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 64, 2048]> input_63_cast_fp16 = reshape(shape = var_2288, x = var_2284_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [2048, 2048]> layers_6_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310715456))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314909824))))[name = string("layers_6_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_45_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_o_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> x_cast_fp16 = add(x = x_47_cast_fp16, y = linear_45_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_2295_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_2295_cast_fp16")];
            bool var_2298 = const()[name = string("op_2298"), val = bool(true)];
            tensor<int32, [1]> var_2299_axes_0 = const()[name = string("op_2299_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2299_cast_fp16 = reduce_max(axes = var_2299_axes_0, keep_dims = var_2298, x = var_2295_cast_fp16)[name = string("op_2299_cast_fp16")];
            fp16 var_2300_to_fp16 = const()[name = string("op_2300_to_fp16"), val = fp16(0x1.47cp-7)];
            fp16 const_42_to_fp16 = const()[name = string("const_42_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_28_cast_fp16 = clip(alpha = var_2300_to_fp16, beta = const_42_to_fp16, x = var_2299_cast_fp16)[name = string("clip_28_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> y_cast_fp16 = real_div(x = x_cast_fp16, y = clip_28_cast_fp16)[name = string("y_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_2304_cast_fp16 = mul(x = y_cast_fp16, y = y_cast_fp16)[name = string("op_2304_cast_fp16")];
            tensor<int32, [1]> var_2309_axes_0 = const()[name = string("op_2309_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2309_keep_dims_0 = const()[name = string("op_2309_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_2309_cast_fp16 = reduce_mean(axes = var_2309_axes_0, keep_dims = var_2309_keep_dims_0, x = var_2304_cast_fp16)[name = string("op_2309_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2310_cast_fp16 = mul(x = clip_28_cast_fp16, y = clip_28_cast_fp16)[name = string("op_2310_cast_fp16")];
            fp32 var_2311_epsilon_0 = const()[name = string("op_2311_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2311_cast_fp16 = inverse(epsilon = var_2311_epsilon_0, x = var_2310_cast_fp16)[name = string("op_2311_cast_fp16")];
            fp16 var_2312_to_fp16 = const()[name = string("op_2312_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2313_cast_fp16 = mul(x = var_2311_cast_fp16, y = var_2312_to_fp16)[name = string("op_2313_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2315_cast_fp16 = add(x = var_2309_cast_fp16, y = var_2313_cast_fp16)[name = string("op_2315_cast_fp16")];
            fp32 var_2316_epsilon_0 = const()[name = string("op_2316_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2316_cast_fp16 = rsqrt(epsilon = var_2316_epsilon_0, x = var_2315_cast_fp16)[name = string("op_2316_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> var_2317_cast_fp16 = mul(x = y_cast_fp16, y = var_2316_cast_fp16)[name = string("op_2317_cast_fp16")];
            tensor<fp16, [2048]> layers_6_post_norm_to_fp16 = const()[name = string("layers_6_post_norm_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314913984)))];
            tensor<fp16, [1, 64, 2048]> input_65_cast_fp16 = mul(x = var_2317_cast_fp16, y = layers_6_post_norm_to_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [6144, 2048]> layers_6_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314918144))), scale = tensor<fp16, [6144, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327501120))))[name = string("layers_6_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 6144]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_6_gate_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 64, 6144]> var_2322_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_2322_cast_fp16")];
            tensor<fp16, [6144, 2048]> layers_6_up_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [6144, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327513472))), scale = tensor<fp16, [6144, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340096448))))[name = string("layers_6_up_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 6144]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_6_up_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 64, 6144]> var_2326_cast_fp16 = mul(x = var_2322_cast_fp16, y = linear_47_cast_fp16)[name = string("op_2326_cast_fp16")];
            fp16 var_2327_to_fp16 = const()[name = string("op_2327_to_fp16"), val = fp16(0x1p-8)];
            tensor<fp16, [1, 64, 6144]> input_cast_fp16 = mul(x = var_2326_cast_fp16, y = var_2327_to_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [2048, 6144]> layers_6_down_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 6144]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340108800))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352691776))))[name = string("layers_6_down_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_down_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_48_cast_fp16")];
            fp16 var_2332_to_fp16 = const()[name = string("op_2332_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 64, 2048]> var_2333_cast_fp16 = mul(x = linear_48_cast_fp16, y = var_2332_to_fp16)[name = string("op_2333_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> hidden_out = add(x = x_cast_fp16, y = var_2333_cast_fp16)[name = string("op_2335_cast_fp16")];
        } -> (hidden_out);
}