program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3600.16.1"}, {"coremlc-version", "3600.25.1"}})]
{
    func n16<ios18>(tensor<int32, [1, 16]> ids, tensor<int32, [1]> last) {
            tensor<int32, [1, 16]> pos = const()[name = string("pos"), val = tensor<int32, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119700544)))];
            int32 x_1_batch_dims_0 = const()[name = string("x_1_batch_dims_0"), val = int32(0)];
            bool x_1_validate_indices_0 = const()[name = string("x_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [16000, 640]> emb_to_fp16 = const()[name = string("emb_to_fp16"), val = tensor<fp16, [16000, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            string ids_to_int16_dtype_0 = const()[name = string("ids_to_int16_dtype_0"), val = string("int16")];
            string cast_69_dtype_0 = const()[name = string("cast_69_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int16, [1, 16]> ids_to_int16 = cast(dtype = ids_to_int16_dtype_0, x = ids)[name = string("cast_74")];
            tensor<int32, [1, 16]> cast_69 = cast(dtype = cast_69_dtype_0, x = ids_to_int16)[name = string("cast_73")];
            tensor<bool, [1, 16]> greater_equal_0 = greater_equal(x = cast_69, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(16000)];
            tensor<int32, [1, 16]> add_0 = add(x = cast_69, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 16]> select_0 = select(a = cast_69, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 x_1_cast_fp16_cast_uint16_axis_0 = const()[name = string("x_1_cast_fp16_cast_uint16_axis_0"), val = int32(0)];
            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
            tensor<int16, [1, 16]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_72")];
            tensor<fp16, [1, 16, 640]> x_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = x_1_cast_fp16_cast_uint16_axis_0, batch_dims = x_1_batch_dims_0, indices = select_0_to_int16, validate_indices = x_1_validate_indices_0, x = emb_to_fp16)[name = string("x_1_cast_fp16_cast_uint16_cast_uint16")];
            fp16 var_253_to_fp16 = const()[name = string("op_253_to_fp16"), val = fp16(0x1p+5)];
            tensor<fp16, [1, 16, 640]> x_3_cast_fp16 = mul(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = var_253_to_fp16)[name = string("x_3_cast_fp16")];
            fp16 var_251_promoted_to_fp16 = const()[name = string("op_251_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> var_255_cast_fp16 = pow(x = x_3_cast_fp16, y = var_251_promoted_to_fp16)[name = string("op_255_cast_fp16")];
            tensor<int32, [1]> var_257_axes_0 = const()[name = string("op_257_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_257_keep_dims_0 = const()[name = string("op_257_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_257_cast_fp16 = reduce_mean(axes = var_257_axes_0, keep_dims = var_257_keep_dims_0, x = var_255_cast_fp16)[name = string("op_257_cast_fp16")];
            fp16 var_258_to_fp16 = const()[name = string("op_258_to_fp16"), val = fp16(0x1.0c8p-10)];
            tensor<fp16, [1, 16, 1]> var_259_cast_fp16 = add(x = var_257_cast_fp16, y = var_258_to_fp16)[name = string("op_259_cast_fp16")];
            fp32 var_260_epsilon_0 = const()[name = string("op_260_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_260_cast_fp16 = rsqrt(epsilon = var_260_epsilon_0, x = var_259_cast_fp16)[name = string("op_260_cast_fp16")];
            tensor<fp16, [1, 16, 640]> var_261_cast_fp16 = mul(x = x_3_cast_fp16, y = var_260_cast_fp16)[name = string("op_261_cast_fp16")];
            tensor<fp16, [640]> layers_0_n1_w1_to_fp16 = const()[name = string("layers_0_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20480128)))];
            tensor<fp16, [1, 16, 640]> input_1_cast_fp16 = mul(x = var_261_cast_fp16, y = layers_0_n1_w1_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_0_qkvzba_weight_to_fp16 = const()[name = string("layers_0_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20481472)))];
            tensor<fp16, [1544]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [1544]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22457856)))];
            tensor<fp16, [1, 16, 1544]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_qkvzba_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [3]> qkv_1_begin_0 = const()[name = string("qkv_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_1_end_0 = const()[name = string("qkv_1_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> qkv_1_end_mask_0 = const()[name = string("qkv_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 1024]> qkv_1_cast_fp16 = slice_by_index(begin = qkv_1_begin_0, end = qkv_1_end_0, end_mask = qkv_1_end_mask_0, x = linear_0_cast_fp16)[name = string("qkv_1_cast_fp16")];
            tensor<int32, [3]> z_1_begin_0 = const()[name = string("z_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_1_end_0 = const()[name = string("z_1_end_0"), val = tensor<int32, [3]>([1, 16, 1536])];
            tensor<bool, [3]> z_1_end_mask_0 = const()[name = string("z_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 512]> z_1_cast_fp16 = slice_by_index(begin = z_1_begin_0, end = z_1_end_0, end_mask = z_1_end_mask_0, x = linear_0_cast_fp16)[name = string("z_1_cast_fp16")];
            tensor<int32, [3]> b_1_begin_0 = const()[name = string("b_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_1_end_0 = const()[name = string("b_1_end_0"), val = tensor<int32, [3]>([1, 16, 1540])];
            tensor<bool, [3]> b_1_end_mask_0 = const()[name = string("b_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 4]> b_1_cast_fp16 = slice_by_index(begin = b_1_begin_0, end = b_1_end_0, end_mask = b_1_end_mask_0, x = linear_0_cast_fp16)[name = string("b_1_cast_fp16")];
            tensor<int32, [3]> a_1_begin_0 = const()[name = string("a_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_1_end_0 = const()[name = string("a_1_end_0"), val = tensor<int32, [3]>([1, 16, 1])];
            tensor<bool, [3]> a_1_end_mask_0 = const()[name = string("a_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 4]> a_1_cast_fp16 = slice_by_index(begin = a_1_begin_0, end = a_1_end_0, end_mask = a_1_end_mask_0, x = linear_0_cast_fp16)[name = string("a_1_cast_fp16")];
            int32 var_300 = const()[name = string("op_300"), val = int32(1)];
            bool xp_1_interleave_0 = const()[name = string("xp_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 1024]> zpad_to_fp16 = const()[name = string("zpad_to_fp16"), val = tensor<fp16, [1, 3, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119700672)))];
            tensor<fp16, [1, 19, 1024]> xp_1_cast_fp16 = concat(axis = var_300, interleave = xp_1_interleave_0, values = (zpad_to_fp16, qkv_1_cast_fp16))[name = string("xp_1_cast_fp16")];
            tensor<int32, [3]> var_311_begin_0 = const()[name = string("op_311_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_311_end_0 = const()[name = string("op_311_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_311_end_mask_0 = const()[name = string("op_311_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 1024]> var_311_cast_fp16 = slice_by_index(begin = var_311_begin_0, end = var_311_end_0, end_mask = var_311_end_mask_0, x = xp_1_cast_fp16)[name = string("op_311_cast_fp16")];
            tensor<fp16, [1024]> var_319_to_fp16 = const()[name = string("op_319_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22479552)))];
            tensor<fp16, [1, 16, 1024]> var_320_cast_fp16 = mul(x = var_311_cast_fp16, y = var_319_to_fp16)[name = string("op_320_cast_fp16")];
            tensor<int32, [3]> var_330_begin_0 = const()[name = string("op_330_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_330_end_0 = const()[name = string("op_330_end_0"), val = tensor<int32, [3]>([1, 18, 1024])];
            tensor<bool, [3]> var_330_end_mask_0 = const()[name = string("op_330_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_330_cast_fp16 = slice_by_index(begin = var_330_begin_0, end = var_330_end_0, end_mask = var_330_end_mask_0, x = xp_1_cast_fp16)[name = string("op_330_cast_fp16")];
            tensor<fp16, [1024]> var_338_to_fp16 = const()[name = string("op_338_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22481664)))];
            tensor<fp16, [1, 16, 1024]> var_339_cast_fp16 = mul(x = var_330_cast_fp16, y = var_338_to_fp16)[name = string("op_339_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_341_cast_fp16 = add(x = var_320_cast_fp16, y = var_339_cast_fp16)[name = string("op_341_cast_fp16")];
            tensor<int32, [3]> var_351_begin_0 = const()[name = string("op_351_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_351_end_0 = const()[name = string("op_351_end_0"), val = tensor<int32, [3]>([1, 17, 1024])];
            tensor<bool, [3]> var_351_end_mask_0 = const()[name = string("op_351_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_351_cast_fp16 = slice_by_index(begin = var_351_begin_0, end = var_351_end_0, end_mask = var_351_end_mask_0, x = xp_1_cast_fp16)[name = string("op_351_cast_fp16")];
            tensor<fp16, [1024]> var_359_to_fp16 = const()[name = string("op_359_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22483776)))];
            tensor<fp16, [1, 16, 1024]> var_360_cast_fp16 = mul(x = var_351_cast_fp16, y = var_359_to_fp16)[name = string("op_360_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_362_cast_fp16 = add(x = var_341_cast_fp16, y = var_360_cast_fp16)[name = string("op_362_cast_fp16")];
            tensor<int32, [3]> var_372_begin_0 = const()[name = string("op_372_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_372_end_0 = const()[name = string("op_372_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> var_372_end_mask_0 = const()[name = string("op_372_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_372_cast_fp16 = slice_by_index(begin = var_372_begin_0, end = var_372_end_0, end_mask = var_372_end_mask_0, x = xp_1_cast_fp16)[name = string("op_372_cast_fp16")];
            tensor<fp16, [1024]> var_380_to_fp16 = const()[name = string("op_380_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22485888)))];
            tensor<fp16, [1, 16, 1024]> var_381_cast_fp16 = mul(x = var_372_cast_fp16, y = var_380_to_fp16)[name = string("op_381_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> input_3_cast_fp16 = add(x = var_362_cast_fp16, y = var_381_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> c_1_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("c_1_cast_fp16")];
            tensor<int32, [3]> var_389_begin_0 = const()[name = string("op_389_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_389_end_0 = const()[name = string("op_389_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
            tensor<bool, [3]> var_389_end_mask_0 = const()[name = string("op_389_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 256]> var_389_cast_fp16 = slice_by_index(begin = var_389_begin_0, end = var_389_end_0, end_mask = var_389_end_mask_0, x = c_1_cast_fp16)[name = string("op_389_cast_fp16")];
            tensor<int32, [4]> var_394 = const()[name = string("op_394"), val = tensor<int32, [4]>([1, 16, 2, 128])];
            tensor<fp16, [1, 16, 2, 128]> q_1_cast_fp16 = reshape(shape = var_394, x = var_389_cast_fp16)[name = string("q_1_cast_fp16")];
            tensor<int32, [3]> var_400_begin_0 = const()[name = string("op_400_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_400_end_0 = const()[name = string("op_400_end_0"), val = tensor<int32, [3]>([1, 16, 512])];
            tensor<bool, [3]> var_400_end_mask_0 = const()[name = string("op_400_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 256]> var_400_cast_fp16 = slice_by_index(begin = var_400_begin_0, end = var_400_end_0, end_mask = var_400_end_mask_0, x = c_1_cast_fp16)[name = string("op_400_cast_fp16")];
            tensor<int32, [4]> var_405 = const()[name = string("op_405"), val = tensor<int32, [4]>([1, 16, 2, 128])];
            tensor<fp16, [1, 16, 2, 128]> k_1_cast_fp16 = reshape(shape = var_405, x = var_400_cast_fp16)[name = string("k_1_cast_fp16")];
            tensor<int32, [3]> var_411_begin_0 = const()[name = string("op_411_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_411_end_0 = const()[name = string("op_411_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> var_411_end_mask_0 = const()[name = string("op_411_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 512]> var_411_cast_fp16 = slice_by_index(begin = var_411_begin_0, end = var_411_end_0, end_mask = var_411_end_mask_0, x = c_1_cast_fp16)[name = string("op_411_cast_fp16")];
            tensor<int32, [4]> var_416 = const()[name = string("op_416"), val = tensor<int32, [4]>([1, 16, 4, 128])];
            tensor<fp16, [1, 16, 4, 128]> var_417_cast_fp16 = reshape(shape = var_416, x = var_411_cast_fp16)[name = string("op_417_cast_fp16")];
            tensor<int32, [4]> var_422 = const()[name = string("op_422"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_424_to_fp16 = const()[name = string("op_424_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 16, 2, 128]> q_3_cast_fp16 = mul(x = q_1_cast_fp16, y = var_424_to_fp16)[name = string("q_3_cast_fp16")];
            fp16 var_426_to_fp16 = const()[name = string("op_426_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 16, 2, 128]> k_3_cast_fp16 = mul(x = k_1_cast_fp16, y = var_426_to_fp16)[name = string("k_3_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_428_cast_fp16 = mul(x = q_3_cast_fp16, y = q_3_cast_fp16)[name = string("op_428_cast_fp16")];
            tensor<int32, [1]> var_433_axes_0 = const()[name = string("op_433_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_433_keep_dims_0 = const()[name = string("op_433_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 2, 1]> var_433_cast_fp16 = reduce_mean(axes = var_433_axes_0, keep_dims = var_433_keep_dims_0, x = var_428_cast_fp16)[name = string("op_433_cast_fp16")];
            fp16 var_434_promoted_to_fp16 = const()[name = string("op_434_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 16, 2, 1]> var_435_cast_fp16 = mul(x = var_433_cast_fp16, y = var_434_promoted_to_fp16)[name = string("op_435_cast_fp16")];
            fp16 var_437_to_fp16 = const()[name = string("op_437_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 16, 2, 1]> var_438_cast_fp16 = add(x = var_435_cast_fp16, y = var_437_to_fp16)[name = string("op_438_cast_fp16")];
            fp32 var_439_epsilon_0 = const()[name = string("op_439_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 2, 1]> var_439_cast_fp16 = rsqrt(epsilon = var_439_epsilon_0, x = var_438_cast_fp16)[name = string("op_439_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_440_cast_fp16 = mul(x = q_3_cast_fp16, y = var_439_cast_fp16)[name = string("op_440_cast_fp16")];
            fp16 var_441_to_fp16 = const()[name = string("op_441_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 16, 2, 128]> q_5_cast_fp16 = mul(x = var_440_cast_fp16, y = var_441_to_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_443_cast_fp16 = mul(x = k_3_cast_fp16, y = k_3_cast_fp16)[name = string("op_443_cast_fp16")];
            tensor<int32, [1]> var_448_axes_0 = const()[name = string("op_448_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_448_keep_dims_0 = const()[name = string("op_448_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 2, 1]> var_448_cast_fp16 = reduce_mean(axes = var_448_axes_0, keep_dims = var_448_keep_dims_0, x = var_443_cast_fp16)[name = string("op_448_cast_fp16")];
            fp16 var_449_promoted_to_fp16 = const()[name = string("op_449_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 16, 2, 1]> var_450_cast_fp16 = mul(x = var_448_cast_fp16, y = var_449_promoted_to_fp16)[name = string("op_450_cast_fp16")];
            fp16 var_452_to_fp16 = const()[name = string("op_452_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 16, 2, 1]> var_453_cast_fp16 = add(x = var_450_cast_fp16, y = var_452_to_fp16)[name = string("op_453_cast_fp16")];
            fp32 var_454_epsilon_0 = const()[name = string("op_454_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 2, 1]> var_454_cast_fp16 = rsqrt(epsilon = var_454_epsilon_0, x = var_453_cast_fp16)[name = string("op_454_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> k_5_cast_fp16 = mul(x = k_3_cast_fp16, y = var_454_cast_fp16)[name = string("k_5_cast_fp16")];
            tensor<int32, [4]> var_460 = const()[name = string("op_460"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_473_axes_0 = const()[name = string("op_473_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 16, 128]> var_461_cast_fp16 = transpose(perm = var_460, x = q_5_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 2, 1, 16, 128]> var_473_cast_fp16 = expand_dims(axes = var_473_axes_0, x = var_461_cast_fp16)[name = string("op_473_cast_fp16")];
            tensor<int32, [5]> var_481_reps_0 = const()[name = string("op_481_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 16, 128]> var_481_cast_fp16 = tile(reps = var_481_reps_0, x = var_473_cast_fp16)[name = string("op_481_cast_fp16")];
            tensor<int32, [4]> var_486 = const()[name = string("op_486"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<fp16, [1, 4, 16, 128]> q_7_cast_fp16 = reshape(shape = var_486, x = var_481_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<int32, [4]> var_492 = const()[name = string("op_492"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_505_axes_0 = const()[name = string("op_505_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 16, 128]> var_493_cast_fp16 = transpose(perm = var_492, x = k_5_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 2, 1, 16, 128]> var_505_cast_fp16 = expand_dims(axes = var_505_axes_0, x = var_493_cast_fp16)[name = string("op_505_cast_fp16")];
            tensor<int32, [5]> var_513_reps_0 = const()[name = string("op_513_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 16, 128]> var_513_cast_fp16 = tile(reps = var_513_reps_0, x = var_505_cast_fp16)[name = string("op_513_cast_fp16")];
            tensor<int32, [4]> var_518 = const()[name = string("op_518"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<fp16, [1, 4, 16, 128]> k_7_cast_fp16 = reshape(shape = var_518, x = var_513_cast_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_520_cast_fp16 = sigmoid(x = b_1_cast_fp16)[name = string("op_520_cast_fp16")];
            tensor<int32, [3]> var_524 = const()[name = string("op_524"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_0_dt_bias_to_fp16 = const()[name = string("layers_0_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 16, 4]> input_5_cast_fp16 = add(x = a_1_cast_fp16, y = layers_0_dt_bias_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_528_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("op_528_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_529_cast_fp16 = abs(x = input_5_cast_fp16)[name = string("op_529_cast_fp16")];
            fp16 const_0_promoted_to_fp16 = const()[name = string("const_0_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 4]> var_530_cast_fp16 = mul(x = var_529_cast_fp16, y = const_0_promoted_to_fp16)[name = string("op_530_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_531_cast_fp16 = exp(x = var_530_cast_fp16)[name = string("op_531_cast_fp16")];
            fp16 var_533_to_fp16 = const()[name = string("op_533_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 16, 4]> var_534_cast_fp16 = add(x = var_531_cast_fp16, y = var_533_to_fp16)[name = string("op_534_cast_fp16")];
            fp32 var_535_epsilon_0 = const()[name = string("op_535_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 16, 4]> var_535_cast_fp16 = log(epsilon = var_535_epsilon_0, x = var_534_cast_fp16)[name = string("op_535_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_537_cast_fp16 = add(x = var_528_cast_fp16, y = var_535_cast_fp16)[name = string("op_537_cast_fp16")];
            tensor<fp16, [4]> layers_0_negA_to_fp16 = const()[name = string("layers_0_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.b3p+3, -0x1.a7cp+0, -0x1.ac4p+3, -0x1.704p+2])];
            tensor<fp16, [1, 16, 4]> var_538_cast_fp16 = mul(x = layers_0_negA_to_fp16, y = var_537_cast_fp16)[name = string("op_538_cast_fp16")];
            tensor<int32, [3]> var_542 = const()[name = string("op_542"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_545_axes_0 = const()[name = string("op_545_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16]> beta_1_cast_fp16 = transpose(perm = var_524, x = var_520_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 4, 16, 1]> var_545_cast_fp16 = expand_dims(axes = var_545_axes_0, x = beta_1_cast_fp16)[name = string("op_545_cast_fp16")];
            fp16 var_546_to_fp16 = const()[name = string("op_546_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 16, 1]> var_547_cast_fp16 = mul(x = var_545_cast_fp16, y = var_546_to_fp16)[name = string("op_547_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> v_1_cast_fp16 = transpose(perm = var_422, x = var_417_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 4, 16, 128]> vb_1_cast_fp16 = mul(x = v_1_cast_fp16, y = var_547_cast_fp16)[name = string("vb_1_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> kb_1_cast_fp16 = mul(x = k_7_cast_fp16, y = var_545_cast_fp16)[name = string("kb_1_cast_fp16")];
            tensor<int32, [2]> var_554 = const()[name = string("op_554"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [1, 4, 16]> gdec_1_cast_fp16 = transpose(perm = var_542, x = var_538_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [4, 16]> var_555_cast_fp16 = reshape(shape = var_554, x = gdec_1_cast_fp16)[name = string("op_555_cast_fp16")];
            tensor<fp16, [256, 16]> transpose_0_to_fp16 = const()[name = string("transpose_0_to_fp16"), val = tensor<fp16, [256, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22488000)))];
            tensor<fp16, [256]> var_556_bias_0_to_fp16 = const()[name = string("op_556_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22496256)))];
            tensor<fp16, [4, 256]> var_556_cast_fp16 = linear(bias = var_556_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_555_cast_fp16)[name = string("op_556_cast_fp16")];
            tensor<int32, [4]> var_561 = const()[name = string("op_561"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_1_cast_fp16 = reshape(shape = var_561, x = var_556_cast_fp16)[name = string("P_1_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_563_cast_fp16 = exp(x = P_1_cast_fp16)[name = string("op_563_cast_fp16")];
            tensor<fp16, [16, 16]> tril_incl_to_fp16 = const()[name = string("tril_incl_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22496832)))];
            tensor<fp16, [1, 4, 16, 16]> D_1_cast_fp16 = mul(x = var_563_cast_fp16, y = tril_incl_to_fp16)[name = string("D_1_cast_fp16")];
            bool kk_1_transpose_x_1 = const()[name = string("kk_1_transpose_x_1"), val = bool(false)];
            bool kk_1_transpose_y_1 = const()[name = string("kk_1_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_1_cast_fp16 = matmul(transpose_x = kk_1_transpose_x_1, transpose_y = kk_1_transpose_y_1, x = kb_1_cast_fp16, y = k_7_cast_fp16)[name = string("kk_1_cast_fp16")];
            fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_569_cast_fp16 = mul(x = kk_1_cast_fp16, y = const_1_promoted_to_fp16)[name = string("op_569_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_570_cast_fp16 = mul(x = var_569_cast_fp16, y = D_1_cast_fp16)[name = string("op_570_cast_fp16")];
            tensor<fp16, [16, 16]> tril_strict_to_fp16 = const()[name = string("tril_strict_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22497408)))];
            tensor<fp16, [1, 4, 16, 16]> N_1_cast_fp16 = mul(x = var_570_cast_fp16, y = tril_strict_to_fp16)[name = string("N_1_cast_fp16")];
            tensor<fp16, [16, 16]> neg_tril_strict_to_fp16 = const()[name = string("neg_tril_strict_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22497984)))];
            tensor<fp16, [1, 4, 16, 16]> var_572_cast_fp16 = mul(x = D_1_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_572_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_1_cast_fp16 = mul(x = kk_1_cast_fp16, y = var_572_cast_fp16)[name = string("Na_1_cast_fp16")];
            tensor<fp16, [16, 16]> eye_to_fp16 = const()[name = string("eye_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22498560)))];
            tensor<fp16, [1, 4, 16, 16]> Minv_1_cast_fp16 = add(x = eye_to_fp16, y = N_1_cast_fp16)[name = string("Minv_1_cast_fp16")];
            bool Np_1_transpose_x_0 = const()[name = string("Np_1_transpose_x_0"), val = bool(false)];
            bool Np_1_transpose_y_0 = const()[name = string("Np_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_1_cast_fp16 = matmul(transpose_x = Np_1_transpose_x_0, transpose_y = Np_1_transpose_y_0, x = N_1_cast_fp16, y = Na_1_cast_fp16)[name = string("Np_1_cast_fp16")];
            bool Npa_1_transpose_x_0 = const()[name = string("Npa_1_transpose_x_0"), val = bool(false)];
            bool Npa_1_transpose_y_0 = const()[name = string("Npa_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_1_cast_fp16 = matmul(transpose_x = Npa_1_transpose_x_0, transpose_y = Npa_1_transpose_y_0, x = Na_1_cast_fp16, y = N_1_cast_fp16)[name = string("Npa_1_cast_fp16")];
            bool var_578_transpose_x_0 = const()[name = string("op_578_transpose_x_0"), val = bool(false)];
            bool var_578_transpose_y_0 = const()[name = string("op_578_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_578_cast_fp16 = matmul(transpose_x = var_578_transpose_x_0, transpose_y = var_578_transpose_y_0, x = Minv_1_cast_fp16, y = Np_1_cast_fp16)[name = string("op_578_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_3_cast_fp16 = add(x = Minv_1_cast_fp16, y = var_578_cast_fp16)[name = string("Minv_3_cast_fp16")];
            bool Np_3_transpose_x_0 = const()[name = string("Np_3_transpose_x_0"), val = bool(false)];
            bool Np_3_transpose_y_0 = const()[name = string("Np_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_3_cast_fp16 = matmul(transpose_x = Np_3_transpose_x_0, transpose_y = Np_3_transpose_y_0, x = Np_1_cast_fp16, y = Npa_1_cast_fp16)[name = string("Np_3_cast_fp16")];
            bool Npa_3_transpose_x_0 = const()[name = string("Npa_3_transpose_x_0"), val = bool(false)];
            bool Npa_3_transpose_y_0 = const()[name = string("Npa_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_3_cast_fp16 = matmul(transpose_x = Npa_3_transpose_x_0, transpose_y = Npa_3_transpose_y_0, x = Npa_1_cast_fp16, y = Np_1_cast_fp16)[name = string("Npa_3_cast_fp16")];
            bool var_583_transpose_x_0 = const()[name = string("op_583_transpose_x_0"), val = bool(false)];
            bool var_583_transpose_y_0 = const()[name = string("op_583_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_583_cast_fp16 = matmul(transpose_x = var_583_transpose_x_0, transpose_y = var_583_transpose_y_0, x = Minv_3_cast_fp16, y = Np_3_cast_fp16)[name = string("op_583_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_5_cast_fp16 = add(x = Minv_3_cast_fp16, y = var_583_cast_fp16)[name = string("Minv_5_cast_fp16")];
            bool Np_5_transpose_x_0 = const()[name = string("Np_5_transpose_x_0"), val = bool(false)];
            bool Np_5_transpose_y_0 = const()[name = string("Np_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_5_cast_fp16 = matmul(transpose_x = Np_5_transpose_x_0, transpose_y = Np_5_transpose_y_0, x = Np_3_cast_fp16, y = Npa_3_cast_fp16)[name = string("Np_5_cast_fp16")];
            bool var_587_transpose_x_0 = const()[name = string("op_587_transpose_x_0"), val = bool(false)];
            bool var_587_transpose_y_0 = const()[name = string("op_587_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_587_cast_fp16 = matmul(transpose_x = var_587_transpose_x_0, transpose_y = var_587_transpose_y_0, x = Minv_5_cast_fp16, y = Np_5_cast_fp16)[name = string("op_587_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_7_cast_fp16 = add(x = Minv_5_cast_fp16, y = var_587_cast_fp16)[name = string("Minv_7_cast_fp16")];
            bool u_1_transpose_x_0 = const()[name = string("u_1_transpose_x_0"), val = bool(false)];
            bool u_1_transpose_y_0 = const()[name = string("u_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_1_cast_fp16 = matmul(transpose_x = u_1_transpose_x_0, transpose_y = u_1_transpose_y_0, x = Minv_7_cast_fp16, y = vb_1_cast_fp16)[name = string("u_1_cast_fp16")];
            bool var_594_transpose_x_1 = const()[name = string("op_594_transpose_x_1"), val = bool(false)];
            bool var_594_transpose_y_1 = const()[name = string("op_594_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_594_cast_fp16 = matmul(transpose_x = var_594_transpose_x_1, transpose_y = var_594_transpose_y_1, x = q_7_cast_fp16, y = k_7_cast_fp16)[name = string("op_594_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_1_cast_fp16 = mul(x = var_594_cast_fp16, y = D_1_cast_fp16)[name = string("A_1_cast_fp16")];
            bool o_1_transpose_x_0 = const()[name = string("o_1_transpose_x_0"), val = bool(false)];
            bool o_1_transpose_y_0 = const()[name = string("o_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> o_1_cast_fp16 = matmul(transpose_x = o_1_transpose_x_0, transpose_y = o_1_transpose_y_0, x = A_1_cast_fp16, y = u_1_cast_fp16)[name = string("o_1_cast_fp16")];
            tensor<int32, [4]> var_601 = const()[name = string("op_601"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_607 = const()[name = string("op_607"), val = tensor<int32, [4]>([1, 16, 4, 128])];
            tensor<fp16, [1, 16, 4, 128]> input_7_cast_fp16 = reshape(shape = var_607, x = z_1_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> o_3_cast_fp16 = transpose(perm = var_601, x = o_1_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 16, 4, 128]> var_609_cast_fp16 = mul(x = o_3_cast_fp16, y = o_3_cast_fp16)[name = string("op_609_cast_fp16")];
            tensor<int32, [1]> var_614_axes_0 = const()[name = string("op_614_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_614_keep_dims_0 = const()[name = string("op_614_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 4, 1]> var_614_cast_fp16 = reduce_mean(axes = var_614_axes_0, keep_dims = var_614_keep_dims_0, x = var_609_cast_fp16)[name = string("op_614_cast_fp16")];
            fp16 var_616_to_fp16 = const()[name = string("op_616_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 16, 4, 1]> var_617_cast_fp16 = add(x = var_614_cast_fp16, y = var_616_to_fp16)[name = string("op_617_cast_fp16")];
            fp32 var_618_epsilon_0 = const()[name = string("op_618_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 4, 1]> var_618_cast_fp16 = rsqrt(epsilon = var_618_epsilon_0, x = var_617_cast_fp16)[name = string("op_618_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> var_619_cast_fp16 = mul(x = o_3_cast_fp16, y = var_618_cast_fp16)[name = string("op_619_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> var_621_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_621_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> o_5_cast_fp16 = mul(x = var_619_cast_fp16, y = var_621_cast_fp16)[name = string("o_5_cast_fp16")];
            tensor<int32, [3]> var_626 = const()[name = string("op_626"), val = tensor<int32, [3]>([1, 16, 512])];
            tensor<fp16, [1, 16, 512]> input_9_cast_fp16 = reshape(shape = var_626, x = o_5_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [640, 512]> layers_0_out_weight_to_fp16 = const()[name = string("layers_0_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22499136)))];
            tensor<fp16, [640]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23154560)))];
            tensor<fp16, [1, 16, 640]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_out_weight_to_fp16, x = input_9_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [1, 16, 640]> x_5_cast_fp16 = add(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = linear_1_cast_fp16)[name = string("x_5_cast_fp16")];
            fp16 var_639_to_fp16 = const()[name = string("op_639_to_fp16"), val = fp16(0x1p+5)];
            tensor<fp16, [1, 16, 640]> x_7_cast_fp16 = mul(x = x_5_cast_fp16, y = var_639_to_fp16)[name = string("x_7_cast_fp16")];
            fp16 var_637_promoted_to_fp16 = const()[name = string("op_637_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> var_641_cast_fp16 = pow(x = x_7_cast_fp16, y = var_637_promoted_to_fp16)[name = string("op_641_cast_fp16")];
            tensor<int32, [1]> var_643_axes_0 = const()[name = string("op_643_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_643_keep_dims_0 = const()[name = string("op_643_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_643_cast_fp16 = reduce_mean(axes = var_643_axes_0, keep_dims = var_643_keep_dims_0, x = var_641_cast_fp16)[name = string("op_643_cast_fp16")];
            fp16 var_644_to_fp16 = const()[name = string("op_644_to_fp16"), val = fp16(0x1.0c8p-10)];
            tensor<fp16, [1, 16, 1]> var_645_cast_fp16 = add(x = var_643_cast_fp16, y = var_644_to_fp16)[name = string("op_645_cast_fp16")];
            fp32 var_646_epsilon_0 = const()[name = string("op_646_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_646_cast_fp16 = rsqrt(epsilon = var_646_epsilon_0, x = var_645_cast_fp16)[name = string("op_646_cast_fp16")];
            tensor<fp16, [1, 16, 640]> var_647_cast_fp16 = mul(x = x_7_cast_fp16, y = var_646_cast_fp16)[name = string("op_647_cast_fp16")];
            tensor<fp16, [640]> layers_0_n2_w1_to_fp16 = const()[name = string("layers_0_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23155904)))];
            tensor<fp16, [1, 16, 640]> input_11_cast_fp16 = mul(x = var_647_cast_fp16, y = layers_0_n2_w1_to_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_0_gate_up_weight_to_fp16 = const()[name = string("layers_0_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23157248)))];
            tensor<fp16, [5120]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29710912)))];
            tensor<fp16, [1, 16, 5120]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_0_gate_up_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [3]> input_13_begin_0 = const()[name = string("input_13_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_13_end_0 = const()[name = string("input_13_end_0"), val = tensor<int32, [3]>([1, 16, 2560])];
            tensor<bool, [3]> input_13_end_mask_0 = const()[name = string("input_13_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 2560]> input_13_cast_fp16 = slice_by_index(begin = input_13_begin_0, end = input_13_end_0, end_mask = input_13_end_mask_0, x = linear_2_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 16, 2560]> var_657_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_657_cast_fp16")];
            tensor<int32, [3]> var_662_begin_0 = const()[name = string("op_662_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_662_end_0 = const()[name = string("op_662_end_0"), val = tensor<int32, [3]>([1, 16, 5120])];
            tensor<bool, [3]> var_662_end_mask_0 = const()[name = string("op_662_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 2560]> var_662_cast_fp16 = slice_by_index(begin = var_662_begin_0, end = var_662_end_0, end_mask = var_662_end_mask_0, x = linear_2_cast_fp16)[name = string("op_662_cast_fp16")];
            tensor<fp16, [1, 16, 2560]> input_15_cast_fp16 = mul(x = var_657_cast_fp16, y = var_662_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_0_down_weight_to_fp16 = const()[name = string("layers_0_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29721216)))];
            tensor<fp16, [1, 16, 640]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_down_weight_to_fp16, x = input_15_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 16, 640]> x_9_cast_fp16 = add(x = x_5_cast_fp16, y = linear_3_cast_fp16)[name = string("x_9_cast_fp16")];
            fp16 var_675_to_fp16 = const()[name = string("op_675_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> x_11_cast_fp16 = mul(x = x_9_cast_fp16, y = var_675_to_fp16)[name = string("x_11_cast_fp16")];
            fp16 var_673_promoted_to_fp16 = const()[name = string("op_673_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> var_677_cast_fp16 = pow(x = x_11_cast_fp16, y = var_673_promoted_to_fp16)[name = string("op_677_cast_fp16")];
            tensor<int32, [1]> var_679_axes_0 = const()[name = string("op_679_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_679_keep_dims_0 = const()[name = string("op_679_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_679_cast_fp16 = reduce_mean(axes = var_679_axes_0, keep_dims = var_679_keep_dims_0, x = var_677_cast_fp16)[name = string("op_679_cast_fp16")];
            fp16 var_680_to_fp16 = const()[name = string("op_680_to_fp16"), val = fp16(0x1.0cp-18)];
            tensor<fp16, [1, 16, 1]> var_681_cast_fp16 = add(x = var_679_cast_fp16, y = var_680_to_fp16)[name = string("op_681_cast_fp16")];
            fp32 var_682_epsilon_0 = const()[name = string("op_682_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_682_cast_fp16 = rsqrt(epsilon = var_682_epsilon_0, x = var_681_cast_fp16)[name = string("op_682_cast_fp16")];
            tensor<fp16, [1, 16, 640]> var_683_cast_fp16 = mul(x = x_11_cast_fp16, y = var_682_cast_fp16)[name = string("op_683_cast_fp16")];
            tensor<fp16, [640]> layers_1_n1_w1_to_fp16 = const()[name = string("layers_1_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32998080)))];
            tensor<fp16, [1, 16, 640]> input_17_cast_fp16 = mul(x = var_683_cast_fp16, y = layers_1_n1_w1_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_1_qkvzba_weight_to_fp16 = const()[name = string("layers_1_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32999424)))];
            tensor<fp16, [1, 16, 1544]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_qkvzba_weight_to_fp16, x = input_17_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<int32, [3]> qkv_3_begin_0 = const()[name = string("qkv_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_3_end_0 = const()[name = string("qkv_3_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> qkv_3_end_mask_0 = const()[name = string("qkv_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 1024]> qkv_3_cast_fp16 = slice_by_index(begin = qkv_3_begin_0, end = qkv_3_end_0, end_mask = qkv_3_end_mask_0, x = linear_4_cast_fp16)[name = string("qkv_3_cast_fp16")];
            tensor<int32, [3]> z_3_begin_0 = const()[name = string("z_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_3_end_0 = const()[name = string("z_3_end_0"), val = tensor<int32, [3]>([1, 16, 1536])];
            tensor<bool, [3]> z_3_end_mask_0 = const()[name = string("z_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 512]> z_3_cast_fp16 = slice_by_index(begin = z_3_begin_0, end = z_3_end_0, end_mask = z_3_end_mask_0, x = linear_4_cast_fp16)[name = string("z_3_cast_fp16")];
            tensor<int32, [3]> b_3_begin_0 = const()[name = string("b_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_3_end_0 = const()[name = string("b_3_end_0"), val = tensor<int32, [3]>([1, 16, 1540])];
            tensor<bool, [3]> b_3_end_mask_0 = const()[name = string("b_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 4]> b_3_cast_fp16 = slice_by_index(begin = b_3_begin_0, end = b_3_end_0, end_mask = b_3_end_mask_0, x = linear_4_cast_fp16)[name = string("b_3_cast_fp16")];
            tensor<int32, [3]> a_3_begin_0 = const()[name = string("a_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_3_end_0 = const()[name = string("a_3_end_0"), val = tensor<int32, [3]>([1, 16, 1])];
            tensor<bool, [3]> a_3_end_mask_0 = const()[name = string("a_3_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 4]> a_3_cast_fp16 = slice_by_index(begin = a_3_begin_0, end = a_3_end_0, end_mask = a_3_end_mask_0, x = linear_4_cast_fp16)[name = string("a_3_cast_fp16")];
            int32 var_722 = const()[name = string("op_722"), val = int32(1)];
            bool xp_3_interleave_0 = const()[name = string("xp_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 19, 1024]> xp_3_cast_fp16 = concat(axis = var_722, interleave = xp_3_interleave_0, values = (zpad_to_fp16, qkv_3_cast_fp16))[name = string("xp_3_cast_fp16")];
            tensor<int32, [3]> var_733_begin_0 = const()[name = string("op_733_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_733_end_0 = const()[name = string("op_733_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_733_end_mask_0 = const()[name = string("op_733_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 1024]> var_733_cast_fp16 = slice_by_index(begin = var_733_begin_0, end = var_733_end_0, end_mask = var_733_end_mask_0, x = xp_3_cast_fp16)[name = string("op_733_cast_fp16")];
            tensor<fp16, [1024]> var_741_to_fp16 = const()[name = string("op_741_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34975808)))];
            tensor<fp16, [1, 16, 1024]> var_742_cast_fp16 = mul(x = var_733_cast_fp16, y = var_741_to_fp16)[name = string("op_742_cast_fp16")];
            tensor<int32, [3]> var_752_begin_0 = const()[name = string("op_752_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_752_end_0 = const()[name = string("op_752_end_0"), val = tensor<int32, [3]>([1, 18, 1024])];
            tensor<bool, [3]> var_752_end_mask_0 = const()[name = string("op_752_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_752_cast_fp16 = slice_by_index(begin = var_752_begin_0, end = var_752_end_0, end_mask = var_752_end_mask_0, x = xp_3_cast_fp16)[name = string("op_752_cast_fp16")];
            tensor<fp16, [1024]> var_760_to_fp16 = const()[name = string("op_760_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34977920)))];
            tensor<fp16, [1, 16, 1024]> var_761_cast_fp16 = mul(x = var_752_cast_fp16, y = var_760_to_fp16)[name = string("op_761_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_763_cast_fp16 = add(x = var_742_cast_fp16, y = var_761_cast_fp16)[name = string("op_763_cast_fp16")];
            tensor<int32, [3]> var_773_begin_0 = const()[name = string("op_773_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_773_end_0 = const()[name = string("op_773_end_0"), val = tensor<int32, [3]>([1, 17, 1024])];
            tensor<bool, [3]> var_773_end_mask_0 = const()[name = string("op_773_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_773_cast_fp16 = slice_by_index(begin = var_773_begin_0, end = var_773_end_0, end_mask = var_773_end_mask_0, x = xp_3_cast_fp16)[name = string("op_773_cast_fp16")];
            tensor<fp16, [1024]> var_781_to_fp16 = const()[name = string("op_781_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34980032)))];
            tensor<fp16, [1, 16, 1024]> var_782_cast_fp16 = mul(x = var_773_cast_fp16, y = var_781_to_fp16)[name = string("op_782_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_784_cast_fp16 = add(x = var_763_cast_fp16, y = var_782_cast_fp16)[name = string("op_784_cast_fp16")];
            tensor<int32, [3]> var_794_begin_0 = const()[name = string("op_794_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_794_end_0 = const()[name = string("op_794_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> var_794_end_mask_0 = const()[name = string("op_794_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_794_cast_fp16 = slice_by_index(begin = var_794_begin_0, end = var_794_end_0, end_mask = var_794_end_mask_0, x = xp_3_cast_fp16)[name = string("op_794_cast_fp16")];
            tensor<fp16, [1024]> var_802_to_fp16 = const()[name = string("op_802_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34982144)))];
            tensor<fp16, [1, 16, 1024]> var_803_cast_fp16 = mul(x = var_794_cast_fp16, y = var_802_to_fp16)[name = string("op_803_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> input_19_cast_fp16 = add(x = var_784_cast_fp16, y = var_803_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> c_3_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("c_3_cast_fp16")];
            tensor<int32, [3]> var_811_begin_0 = const()[name = string("op_811_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_811_end_0 = const()[name = string("op_811_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
            tensor<bool, [3]> var_811_end_mask_0 = const()[name = string("op_811_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 256]> var_811_cast_fp16 = slice_by_index(begin = var_811_begin_0, end = var_811_end_0, end_mask = var_811_end_mask_0, x = c_3_cast_fp16)[name = string("op_811_cast_fp16")];
            tensor<int32, [4]> var_816 = const()[name = string("op_816"), val = tensor<int32, [4]>([1, 16, 2, 128])];
            tensor<fp16, [1, 16, 2, 128]> q_9_cast_fp16 = reshape(shape = var_816, x = var_811_cast_fp16)[name = string("q_9_cast_fp16")];
            tensor<int32, [3]> var_822_begin_0 = const()[name = string("op_822_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_822_end_0 = const()[name = string("op_822_end_0"), val = tensor<int32, [3]>([1, 16, 512])];
            tensor<bool, [3]> var_822_end_mask_0 = const()[name = string("op_822_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 256]> var_822_cast_fp16 = slice_by_index(begin = var_822_begin_0, end = var_822_end_0, end_mask = var_822_end_mask_0, x = c_3_cast_fp16)[name = string("op_822_cast_fp16")];
            tensor<int32, [4]> var_827 = const()[name = string("op_827"), val = tensor<int32, [4]>([1, 16, 2, 128])];
            tensor<fp16, [1, 16, 2, 128]> k_9_cast_fp16 = reshape(shape = var_827, x = var_822_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<int32, [3]> var_833_begin_0 = const()[name = string("op_833_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_833_end_0 = const()[name = string("op_833_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> var_833_end_mask_0 = const()[name = string("op_833_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 512]> var_833_cast_fp16 = slice_by_index(begin = var_833_begin_0, end = var_833_end_0, end_mask = var_833_end_mask_0, x = c_3_cast_fp16)[name = string("op_833_cast_fp16")];
            tensor<int32, [4]> var_838 = const()[name = string("op_838"), val = tensor<int32, [4]>([1, 16, 4, 128])];
            tensor<fp16, [1, 16, 4, 128]> var_839_cast_fp16 = reshape(shape = var_838, x = var_833_cast_fp16)[name = string("op_839_cast_fp16")];
            tensor<int32, [4]> var_844 = const()[name = string("op_844"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_846_to_fp16 = const()[name = string("op_846_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 16, 2, 128]> q_11_cast_fp16 = mul(x = q_9_cast_fp16, y = var_846_to_fp16)[name = string("q_11_cast_fp16")];
            fp16 var_848_to_fp16 = const()[name = string("op_848_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 16, 2, 128]> k_11_cast_fp16 = mul(x = k_9_cast_fp16, y = var_848_to_fp16)[name = string("k_11_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_850_cast_fp16 = mul(x = q_11_cast_fp16, y = q_11_cast_fp16)[name = string("op_850_cast_fp16")];
            tensor<int32, [1]> var_855_axes_0 = const()[name = string("op_855_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_855_keep_dims_0 = const()[name = string("op_855_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 2, 1]> var_855_cast_fp16 = reduce_mean(axes = var_855_axes_0, keep_dims = var_855_keep_dims_0, x = var_850_cast_fp16)[name = string("op_855_cast_fp16")];
            fp16 var_856_promoted_to_fp16 = const()[name = string("op_856_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 16, 2, 1]> var_857_cast_fp16 = mul(x = var_855_cast_fp16, y = var_856_promoted_to_fp16)[name = string("op_857_cast_fp16")];
            fp16 var_859_to_fp16 = const()[name = string("op_859_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 16, 2, 1]> var_860_cast_fp16 = add(x = var_857_cast_fp16, y = var_859_to_fp16)[name = string("op_860_cast_fp16")];
            fp32 var_861_epsilon_0 = const()[name = string("op_861_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 2, 1]> var_861_cast_fp16 = rsqrt(epsilon = var_861_epsilon_0, x = var_860_cast_fp16)[name = string("op_861_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_862_cast_fp16 = mul(x = q_11_cast_fp16, y = var_861_cast_fp16)[name = string("op_862_cast_fp16")];
            fp16 var_863_to_fp16 = const()[name = string("op_863_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 16, 2, 128]> q_13_cast_fp16 = mul(x = var_862_cast_fp16, y = var_863_to_fp16)[name = string("q_13_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_865_cast_fp16 = mul(x = k_11_cast_fp16, y = k_11_cast_fp16)[name = string("op_865_cast_fp16")];
            tensor<int32, [1]> var_870_axes_0 = const()[name = string("op_870_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_870_keep_dims_0 = const()[name = string("op_870_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 2, 1]> var_870_cast_fp16 = reduce_mean(axes = var_870_axes_0, keep_dims = var_870_keep_dims_0, x = var_865_cast_fp16)[name = string("op_870_cast_fp16")];
            fp16 var_871_promoted_to_fp16 = const()[name = string("op_871_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 16, 2, 1]> var_872_cast_fp16 = mul(x = var_870_cast_fp16, y = var_871_promoted_to_fp16)[name = string("op_872_cast_fp16")];
            fp16 var_874_to_fp16 = const()[name = string("op_874_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 16, 2, 1]> var_875_cast_fp16 = add(x = var_872_cast_fp16, y = var_874_to_fp16)[name = string("op_875_cast_fp16")];
            fp32 var_876_epsilon_0 = const()[name = string("op_876_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 2, 1]> var_876_cast_fp16 = rsqrt(epsilon = var_876_epsilon_0, x = var_875_cast_fp16)[name = string("op_876_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> k_13_cast_fp16 = mul(x = k_11_cast_fp16, y = var_876_cast_fp16)[name = string("k_13_cast_fp16")];
            tensor<int32, [4]> var_882 = const()[name = string("op_882"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_895_axes_0 = const()[name = string("op_895_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 16, 128]> var_883_cast_fp16 = transpose(perm = var_882, x = q_13_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 2, 1, 16, 128]> var_895_cast_fp16 = expand_dims(axes = var_895_axes_0, x = var_883_cast_fp16)[name = string("op_895_cast_fp16")];
            tensor<int32, [5]> var_903_reps_0 = const()[name = string("op_903_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 16, 128]> var_903_cast_fp16 = tile(reps = var_903_reps_0, x = var_895_cast_fp16)[name = string("op_903_cast_fp16")];
            tensor<int32, [4]> var_908 = const()[name = string("op_908"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<fp16, [1, 4, 16, 128]> q_15_cast_fp16 = reshape(shape = var_908, x = var_903_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<int32, [4]> var_914 = const()[name = string("op_914"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_927_axes_0 = const()[name = string("op_927_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 16, 128]> var_915_cast_fp16 = transpose(perm = var_914, x = k_13_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 2, 1, 16, 128]> var_927_cast_fp16 = expand_dims(axes = var_927_axes_0, x = var_915_cast_fp16)[name = string("op_927_cast_fp16")];
            tensor<int32, [5]> var_935_reps_0 = const()[name = string("op_935_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 16, 128]> var_935_cast_fp16 = tile(reps = var_935_reps_0, x = var_927_cast_fp16)[name = string("op_935_cast_fp16")];
            tensor<int32, [4]> var_940 = const()[name = string("op_940"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<fp16, [1, 4, 16, 128]> k_15_cast_fp16 = reshape(shape = var_940, x = var_935_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_942_cast_fp16 = sigmoid(x = b_3_cast_fp16)[name = string("op_942_cast_fp16")];
            tensor<int32, [3]> var_946 = const()[name = string("op_946"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_1_dt_bias_to_fp16 = const()[name = string("layers_1_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 16, 4]> input_21_cast_fp16 = add(x = a_3_cast_fp16, y = layers_1_dt_bias_to_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_950_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("op_950_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_951_cast_fp16 = abs(x = input_21_cast_fp16)[name = string("op_951_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 4]> var_952_cast_fp16 = mul(x = var_951_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_952_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_953_cast_fp16 = exp(x = var_952_cast_fp16)[name = string("op_953_cast_fp16")];
            fp16 var_955_to_fp16 = const()[name = string("op_955_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 16, 4]> var_956_cast_fp16 = add(x = var_953_cast_fp16, y = var_955_to_fp16)[name = string("op_956_cast_fp16")];
            fp32 var_957_epsilon_0 = const()[name = string("op_957_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 16, 4]> var_957_cast_fp16 = log(epsilon = var_957_epsilon_0, x = var_956_cast_fp16)[name = string("op_957_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_959_cast_fp16 = add(x = var_950_cast_fp16, y = var_957_cast_fp16)[name = string("op_959_cast_fp16")];
            tensor<fp16, [4]> layers_1_negA_to_fp16 = const()[name = string("layers_1_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.ff4p+2, -0x1.8cp+3, -0x1.c0cp+3, -0x1.104p+3])];
            tensor<fp16, [1, 16, 4]> var_960_cast_fp16 = mul(x = layers_1_negA_to_fp16, y = var_959_cast_fp16)[name = string("op_960_cast_fp16")];
            tensor<int32, [3]> var_964 = const()[name = string("op_964"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_967_axes_0 = const()[name = string("op_967_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16]> beta_3_cast_fp16 = transpose(perm = var_946, x = var_942_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 4, 16, 1]> var_967_cast_fp16 = expand_dims(axes = var_967_axes_0, x = beta_3_cast_fp16)[name = string("op_967_cast_fp16")];
            fp16 var_968_to_fp16 = const()[name = string("op_968_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 16, 1]> var_969_cast_fp16 = mul(x = var_967_cast_fp16, y = var_968_to_fp16)[name = string("op_969_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> v_3_cast_fp16 = transpose(perm = var_844, x = var_839_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 4, 16, 128]> vb_3_cast_fp16 = mul(x = v_3_cast_fp16, y = var_969_cast_fp16)[name = string("vb_3_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> kb_3_cast_fp16 = mul(x = k_15_cast_fp16, y = var_967_cast_fp16)[name = string("kb_3_cast_fp16")];
            tensor<int32, [2]> var_976 = const()[name = string("op_976"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [1, 4, 16]> gdec_3_cast_fp16 = transpose(perm = var_964, x = var_960_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [4, 16]> var_977_cast_fp16 = reshape(shape = var_976, x = gdec_3_cast_fp16)[name = string("op_977_cast_fp16")];
            tensor<fp16, [4, 256]> var_978_cast_fp16 = linear(bias = var_556_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_977_cast_fp16)[name = string("op_978_cast_fp16")];
            tensor<int32, [4]> var_983 = const()[name = string("op_983"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_3_cast_fp16 = reshape(shape = var_983, x = var_978_cast_fp16)[name = string("P_3_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_985_cast_fp16 = exp(x = P_3_cast_fp16)[name = string("op_985_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_3_cast_fp16 = mul(x = var_985_cast_fp16, y = tril_incl_to_fp16)[name = string("D_3_cast_fp16")];
            bool kk_3_transpose_x_1 = const()[name = string("kk_3_transpose_x_1"), val = bool(false)];
            bool kk_3_transpose_y_1 = const()[name = string("kk_3_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_3_cast_fp16 = matmul(transpose_x = kk_3_transpose_x_1, transpose_y = kk_3_transpose_y_1, x = kb_3_cast_fp16, y = k_15_cast_fp16)[name = string("kk_3_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_991_cast_fp16 = mul(x = kk_3_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_991_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_992_cast_fp16 = mul(x = var_991_cast_fp16, y = D_3_cast_fp16)[name = string("op_992_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_3_cast_fp16 = mul(x = var_992_cast_fp16, y = tril_strict_to_fp16)[name = string("N_3_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_994_cast_fp16 = mul(x = D_3_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_994_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_3_cast_fp16 = mul(x = kk_3_cast_fp16, y = var_994_cast_fp16)[name = string("Na_3_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_9_cast_fp16 = add(x = eye_to_fp16, y = N_3_cast_fp16)[name = string("Minv_9_cast_fp16")];
            bool Np_7_transpose_x_0 = const()[name = string("Np_7_transpose_x_0"), val = bool(false)];
            bool Np_7_transpose_y_0 = const()[name = string("Np_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_7_cast_fp16 = matmul(transpose_x = Np_7_transpose_x_0, transpose_y = Np_7_transpose_y_0, x = N_3_cast_fp16, y = Na_3_cast_fp16)[name = string("Np_7_cast_fp16")];
            bool Npa_5_transpose_x_0 = const()[name = string("Npa_5_transpose_x_0"), val = bool(false)];
            bool Npa_5_transpose_y_0 = const()[name = string("Npa_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_5_cast_fp16 = matmul(transpose_x = Npa_5_transpose_x_0, transpose_y = Npa_5_transpose_y_0, x = Na_3_cast_fp16, y = N_3_cast_fp16)[name = string("Npa_5_cast_fp16")];
            bool var_1000_transpose_x_0 = const()[name = string("op_1000_transpose_x_0"), val = bool(false)];
            bool var_1000_transpose_y_0 = const()[name = string("op_1000_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1000_cast_fp16 = matmul(transpose_x = var_1000_transpose_x_0, transpose_y = var_1000_transpose_y_0, x = Minv_9_cast_fp16, y = Np_7_cast_fp16)[name = string("op_1000_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_11_cast_fp16 = add(x = Minv_9_cast_fp16, y = var_1000_cast_fp16)[name = string("Minv_11_cast_fp16")];
            bool Np_9_transpose_x_0 = const()[name = string("Np_9_transpose_x_0"), val = bool(false)];
            bool Np_9_transpose_y_0 = const()[name = string("Np_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_9_cast_fp16 = matmul(transpose_x = Np_9_transpose_x_0, transpose_y = Np_9_transpose_y_0, x = Np_7_cast_fp16, y = Npa_5_cast_fp16)[name = string("Np_9_cast_fp16")];
            bool Npa_7_transpose_x_0 = const()[name = string("Npa_7_transpose_x_0"), val = bool(false)];
            bool Npa_7_transpose_y_0 = const()[name = string("Npa_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_7_cast_fp16 = matmul(transpose_x = Npa_7_transpose_x_0, transpose_y = Npa_7_transpose_y_0, x = Npa_5_cast_fp16, y = Np_7_cast_fp16)[name = string("Npa_7_cast_fp16")];
            bool var_1005_transpose_x_0 = const()[name = string("op_1005_transpose_x_0"), val = bool(false)];
            bool var_1005_transpose_y_0 = const()[name = string("op_1005_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1005_cast_fp16 = matmul(transpose_x = var_1005_transpose_x_0, transpose_y = var_1005_transpose_y_0, x = Minv_11_cast_fp16, y = Np_9_cast_fp16)[name = string("op_1005_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_13_cast_fp16 = add(x = Minv_11_cast_fp16, y = var_1005_cast_fp16)[name = string("Minv_13_cast_fp16")];
            bool Np_11_transpose_x_0 = const()[name = string("Np_11_transpose_x_0"), val = bool(false)];
            bool Np_11_transpose_y_0 = const()[name = string("Np_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_11_cast_fp16 = matmul(transpose_x = Np_11_transpose_x_0, transpose_y = Np_11_transpose_y_0, x = Np_9_cast_fp16, y = Npa_7_cast_fp16)[name = string("Np_11_cast_fp16")];
            bool var_1009_transpose_x_0 = const()[name = string("op_1009_transpose_x_0"), val = bool(false)];
            bool var_1009_transpose_y_0 = const()[name = string("op_1009_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1009_cast_fp16 = matmul(transpose_x = var_1009_transpose_x_0, transpose_y = var_1009_transpose_y_0, x = Minv_13_cast_fp16, y = Np_11_cast_fp16)[name = string("op_1009_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_15_cast_fp16 = add(x = Minv_13_cast_fp16, y = var_1009_cast_fp16)[name = string("Minv_15_cast_fp16")];
            bool u_3_transpose_x_0 = const()[name = string("u_3_transpose_x_0"), val = bool(false)];
            bool u_3_transpose_y_0 = const()[name = string("u_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_3_cast_fp16 = matmul(transpose_x = u_3_transpose_x_0, transpose_y = u_3_transpose_y_0, x = Minv_15_cast_fp16, y = vb_3_cast_fp16)[name = string("u_3_cast_fp16")];
            bool var_1016_transpose_x_1 = const()[name = string("op_1016_transpose_x_1"), val = bool(false)];
            bool var_1016_transpose_y_1 = const()[name = string("op_1016_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_1016_cast_fp16 = matmul(transpose_x = var_1016_transpose_x_1, transpose_y = var_1016_transpose_y_1, x = q_15_cast_fp16, y = k_15_cast_fp16)[name = string("op_1016_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_3_cast_fp16 = mul(x = var_1016_cast_fp16, y = D_3_cast_fp16)[name = string("A_3_cast_fp16")];
            bool o_7_transpose_x_0 = const()[name = string("o_7_transpose_x_0"), val = bool(false)];
            bool o_7_transpose_y_0 = const()[name = string("o_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> o_7_cast_fp16 = matmul(transpose_x = o_7_transpose_x_0, transpose_y = o_7_transpose_y_0, x = A_3_cast_fp16, y = u_3_cast_fp16)[name = string("o_7_cast_fp16")];
            tensor<int32, [4]> var_1023 = const()[name = string("op_1023"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1029 = const()[name = string("op_1029"), val = tensor<int32, [4]>([1, 16, 4, 128])];
            tensor<fp16, [1, 16, 4, 128]> input_23_cast_fp16 = reshape(shape = var_1029, x = z_3_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> o_9_cast_fp16 = transpose(perm = var_1023, x = o_7_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 16, 4, 128]> var_1031_cast_fp16 = mul(x = o_9_cast_fp16, y = o_9_cast_fp16)[name = string("op_1031_cast_fp16")];
            tensor<int32, [1]> var_1036_axes_0 = const()[name = string("op_1036_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1036_keep_dims_0 = const()[name = string("op_1036_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 4, 1]> var_1036_cast_fp16 = reduce_mean(axes = var_1036_axes_0, keep_dims = var_1036_keep_dims_0, x = var_1031_cast_fp16)[name = string("op_1036_cast_fp16")];
            fp16 var_1038_to_fp16 = const()[name = string("op_1038_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 16, 4, 1]> var_1039_cast_fp16 = add(x = var_1036_cast_fp16, y = var_1038_to_fp16)[name = string("op_1039_cast_fp16")];
            fp32 var_1040_epsilon_0 = const()[name = string("op_1040_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 4, 1]> var_1040_cast_fp16 = rsqrt(epsilon = var_1040_epsilon_0, x = var_1039_cast_fp16)[name = string("op_1040_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> var_1041_cast_fp16 = mul(x = o_9_cast_fp16, y = var_1040_cast_fp16)[name = string("op_1041_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> var_1043_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_1043_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> o_11_cast_fp16 = mul(x = var_1041_cast_fp16, y = var_1043_cast_fp16)[name = string("o_11_cast_fp16")];
            tensor<int32, [3]> var_1048 = const()[name = string("op_1048"), val = tensor<int32, [3]>([1, 16, 512])];
            tensor<fp16, [1, 16, 512]> input_25_cast_fp16 = reshape(shape = var_1048, x = o_11_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [640, 512]> layers_1_out_weight_to_fp16 = const()[name = string("layers_1_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34984256)))];
            tensor<fp16, [1, 16, 640]> linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_out_weight_to_fp16, x = input_25_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 16, 640]> x_13_cast_fp16 = add(x = x_9_cast_fp16, y = linear_5_cast_fp16)[name = string("x_13_cast_fp16")];
            fp16 var_1061_to_fp16 = const()[name = string("op_1061_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> x_15_cast_fp16 = mul(x = x_13_cast_fp16, y = var_1061_to_fp16)[name = string("x_15_cast_fp16")];
            fp16 var_1059_promoted_to_fp16 = const()[name = string("op_1059_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> var_1063_cast_fp16 = pow(x = x_15_cast_fp16, y = var_1059_promoted_to_fp16)[name = string("op_1063_cast_fp16")];
            tensor<int32, [1]> var_1065_axes_0 = const()[name = string("op_1065_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1065_keep_dims_0 = const()[name = string("op_1065_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1065_cast_fp16 = reduce_mean(axes = var_1065_axes_0, keep_dims = var_1065_keep_dims_0, x = var_1063_cast_fp16)[name = string("op_1065_cast_fp16")];
            fp16 var_1066_to_fp16 = const()[name = string("op_1066_to_fp16"), val = fp16(0x1.0cp-18)];
            tensor<fp16, [1, 16, 1]> var_1067_cast_fp16 = add(x = var_1065_cast_fp16, y = var_1066_to_fp16)[name = string("op_1067_cast_fp16")];
            fp32 var_1068_epsilon_0 = const()[name = string("op_1068_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1068_cast_fp16 = rsqrt(epsilon = var_1068_epsilon_0, x = var_1067_cast_fp16)[name = string("op_1068_cast_fp16")];
            tensor<fp16, [1, 16, 640]> var_1069_cast_fp16 = mul(x = x_15_cast_fp16, y = var_1068_cast_fp16)[name = string("op_1069_cast_fp16")];
            tensor<fp16, [640]> layers_1_n2_w1_to_fp16 = const()[name = string("layers_1_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35639680)))];
            tensor<fp16, [1, 16, 640]> input_27_cast_fp16 = mul(x = var_1069_cast_fp16, y = layers_1_n2_w1_to_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_1_gate_up_weight_to_fp16 = const()[name = string("layers_1_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35641024)))];
            tensor<fp16, [1, 16, 5120]> linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_1_gate_up_weight_to_fp16, x = input_27_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<int32, [3]> input_29_begin_0 = const()[name = string("input_29_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_29_end_0 = const()[name = string("input_29_end_0"), val = tensor<int32, [3]>([1, 16, 2560])];
            tensor<bool, [3]> input_29_end_mask_0 = const()[name = string("input_29_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 2560]> input_29_cast_fp16 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = linear_6_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [1, 16, 2560]> var_1079_cast_fp16 = silu(x = input_29_cast_fp16)[name = string("op_1079_cast_fp16")];
            tensor<int32, [3]> var_1084_begin_0 = const()[name = string("op_1084_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_1084_end_0 = const()[name = string("op_1084_end_0"), val = tensor<int32, [3]>([1, 16, 5120])];
            tensor<bool, [3]> var_1084_end_mask_0 = const()[name = string("op_1084_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 2560]> var_1084_cast_fp16 = slice_by_index(begin = var_1084_begin_0, end = var_1084_end_0, end_mask = var_1084_end_mask_0, x = linear_6_cast_fp16)[name = string("op_1084_cast_fp16")];
            tensor<fp16, [1, 16, 2560]> input_31_cast_fp16 = mul(x = var_1079_cast_fp16, y = var_1084_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_1_down_weight_to_fp16 = const()[name = string("layers_1_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42194688)))];
            tensor<fp16, [1, 16, 640]> linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_down_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 16, 640]> x_17_cast_fp16 = add(x = x_13_cast_fp16, y = linear_7_cast_fp16)[name = string("x_17_cast_fp16")];
            fp16 var_1095_promoted_to_fp16 = const()[name = string("op_1095_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> var_1099_cast_fp16 = pow(x = x_17_cast_fp16, y = var_1095_promoted_to_fp16)[name = string("op_1099_cast_fp16")];
            tensor<int32, [1]> var_1101_axes_0 = const()[name = string("op_1101_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1101_keep_dims_0 = const()[name = string("op_1101_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1101_cast_fp16 = reduce_mean(axes = var_1101_axes_0, keep_dims = var_1101_keep_dims_0, x = var_1099_cast_fp16)[name = string("op_1101_cast_fp16")];
            fp16 var_1102_to_fp16 = const()[name = string("op_1102_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_1103_cast_fp16 = add(x = var_1101_cast_fp16, y = var_1102_to_fp16)[name = string("op_1103_cast_fp16")];
            fp32 var_1104_epsilon_0 = const()[name = string("op_1104_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1104_cast_fp16 = rsqrt(epsilon = var_1104_epsilon_0, x = var_1103_cast_fp16)[name = string("op_1104_cast_fp16")];
            tensor<fp16, [1, 16, 640]> var_1105_cast_fp16 = mul(x = x_17_cast_fp16, y = var_1104_cast_fp16)[name = string("op_1105_cast_fp16")];
            tensor<fp16, [640]> layers_2_n1_w1_to_fp16 = const()[name = string("layers_2_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45471552)))];
            tensor<fp16, [1, 16, 640]> input_33_cast_fp16 = mul(x = var_1105_cast_fp16, y = layers_2_n1_w1_to_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_2_qkvzba_weight_to_fp16 = const()[name = string("layers_2_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45472896)))];
            tensor<fp16, [1, 16, 1544]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_qkvzba_weight_to_fp16, x = input_33_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [3]> qkv_5_begin_0 = const()[name = string("qkv_5_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_5_end_0 = const()[name = string("qkv_5_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> qkv_5_end_mask_0 = const()[name = string("qkv_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 1024]> qkv_5_cast_fp16 = slice_by_index(begin = qkv_5_begin_0, end = qkv_5_end_0, end_mask = qkv_5_end_mask_0, x = linear_8_cast_fp16)[name = string("qkv_5_cast_fp16")];
            tensor<int32, [3]> z_5_begin_0 = const()[name = string("z_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_5_end_0 = const()[name = string("z_5_end_0"), val = tensor<int32, [3]>([1, 16, 1536])];
            tensor<bool, [3]> z_5_end_mask_0 = const()[name = string("z_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 512]> z_5_cast_fp16 = slice_by_index(begin = z_5_begin_0, end = z_5_end_0, end_mask = z_5_end_mask_0, x = linear_8_cast_fp16)[name = string("z_5_cast_fp16")];
            tensor<int32, [3]> b_5_begin_0 = const()[name = string("b_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_5_end_0 = const()[name = string("b_5_end_0"), val = tensor<int32, [3]>([1, 16, 1540])];
            tensor<bool, [3]> b_5_end_mask_0 = const()[name = string("b_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 4]> b_5_cast_fp16 = slice_by_index(begin = b_5_begin_0, end = b_5_end_0, end_mask = b_5_end_mask_0, x = linear_8_cast_fp16)[name = string("b_5_cast_fp16")];
            tensor<int32, [3]> a_5_begin_0 = const()[name = string("a_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_5_end_0 = const()[name = string("a_5_end_0"), val = tensor<int32, [3]>([1, 16, 1])];
            tensor<bool, [3]> a_5_end_mask_0 = const()[name = string("a_5_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 4]> a_5_cast_fp16 = slice_by_index(begin = a_5_begin_0, end = a_5_end_0, end_mask = a_5_end_mask_0, x = linear_8_cast_fp16)[name = string("a_5_cast_fp16")];
            int32 var_1144 = const()[name = string("op_1144"), val = int32(1)];
            bool xp_5_interleave_0 = const()[name = string("xp_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 19, 1024]> xp_5_cast_fp16 = concat(axis = var_1144, interleave = xp_5_interleave_0, values = (zpad_to_fp16, qkv_5_cast_fp16))[name = string("xp_5_cast_fp16")];
            tensor<int32, [3]> var_1155_begin_0 = const()[name = string("op_1155_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_1155_end_0 = const()[name = string("op_1155_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_1155_end_mask_0 = const()[name = string("op_1155_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 1024]> var_1155_cast_fp16 = slice_by_index(begin = var_1155_begin_0, end = var_1155_end_0, end_mask = var_1155_end_mask_0, x = xp_5_cast_fp16)[name = string("op_1155_cast_fp16")];
            tensor<fp16, [1024]> var_1163_to_fp16 = const()[name = string("op_1163_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47449280)))];
            tensor<fp16, [1, 16, 1024]> var_1164_cast_fp16 = mul(x = var_1155_cast_fp16, y = var_1163_to_fp16)[name = string("op_1164_cast_fp16")];
            tensor<int32, [3]> var_1174_begin_0 = const()[name = string("op_1174_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_1174_end_0 = const()[name = string("op_1174_end_0"), val = tensor<int32, [3]>([1, 18, 1024])];
            tensor<bool, [3]> var_1174_end_mask_0 = const()[name = string("op_1174_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_1174_cast_fp16 = slice_by_index(begin = var_1174_begin_0, end = var_1174_end_0, end_mask = var_1174_end_mask_0, x = xp_5_cast_fp16)[name = string("op_1174_cast_fp16")];
            tensor<fp16, [1024]> var_1182_to_fp16 = const()[name = string("op_1182_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47451392)))];
            tensor<fp16, [1, 16, 1024]> var_1183_cast_fp16 = mul(x = var_1174_cast_fp16, y = var_1182_to_fp16)[name = string("op_1183_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1185_cast_fp16 = add(x = var_1164_cast_fp16, y = var_1183_cast_fp16)[name = string("op_1185_cast_fp16")];
            tensor<int32, [3]> var_1195_begin_0 = const()[name = string("op_1195_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1195_end_0 = const()[name = string("op_1195_end_0"), val = tensor<int32, [3]>([1, 17, 1024])];
            tensor<bool, [3]> var_1195_end_mask_0 = const()[name = string("op_1195_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_1195_cast_fp16 = slice_by_index(begin = var_1195_begin_0, end = var_1195_end_0, end_mask = var_1195_end_mask_0, x = xp_5_cast_fp16)[name = string("op_1195_cast_fp16")];
            tensor<fp16, [1024]> var_1203_to_fp16 = const()[name = string("op_1203_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47453504)))];
            tensor<fp16, [1, 16, 1024]> var_1204_cast_fp16 = mul(x = var_1195_cast_fp16, y = var_1203_to_fp16)[name = string("op_1204_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1206_cast_fp16 = add(x = var_1185_cast_fp16, y = var_1204_cast_fp16)[name = string("op_1206_cast_fp16")];
            tensor<int32, [3]> var_1216_begin_0 = const()[name = string("op_1216_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1216_end_0 = const()[name = string("op_1216_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> var_1216_end_mask_0 = const()[name = string("op_1216_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_1216_cast_fp16 = slice_by_index(begin = var_1216_begin_0, end = var_1216_end_0, end_mask = var_1216_end_mask_0, x = xp_5_cast_fp16)[name = string("op_1216_cast_fp16")];
            tensor<fp16, [1024]> var_1224_to_fp16 = const()[name = string("op_1224_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47455616)))];
            tensor<fp16, [1, 16, 1024]> var_1225_cast_fp16 = mul(x = var_1216_cast_fp16, y = var_1224_to_fp16)[name = string("op_1225_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> input_35_cast_fp16 = add(x = var_1206_cast_fp16, y = var_1225_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> c_5_cast_fp16 = silu(x = input_35_cast_fp16)[name = string("c_5_cast_fp16")];
            tensor<int32, [3]> var_1233_begin_0 = const()[name = string("op_1233_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1233_end_0 = const()[name = string("op_1233_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
            tensor<bool, [3]> var_1233_end_mask_0 = const()[name = string("op_1233_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 256]> var_1233_cast_fp16 = slice_by_index(begin = var_1233_begin_0, end = var_1233_end_0, end_mask = var_1233_end_mask_0, x = c_5_cast_fp16)[name = string("op_1233_cast_fp16")];
            tensor<int32, [4]> var_1238 = const()[name = string("op_1238"), val = tensor<int32, [4]>([1, 16, 2, 128])];
            tensor<fp16, [1, 16, 2, 128]> q_17_cast_fp16 = reshape(shape = var_1238, x = var_1233_cast_fp16)[name = string("q_17_cast_fp16")];
            tensor<int32, [3]> var_1244_begin_0 = const()[name = string("op_1244_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_1244_end_0 = const()[name = string("op_1244_end_0"), val = tensor<int32, [3]>([1, 16, 512])];
            tensor<bool, [3]> var_1244_end_mask_0 = const()[name = string("op_1244_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 256]> var_1244_cast_fp16 = slice_by_index(begin = var_1244_begin_0, end = var_1244_end_0, end_mask = var_1244_end_mask_0, x = c_5_cast_fp16)[name = string("op_1244_cast_fp16")];
            tensor<int32, [4]> var_1249 = const()[name = string("op_1249"), val = tensor<int32, [4]>([1, 16, 2, 128])];
            tensor<fp16, [1, 16, 2, 128]> k_17_cast_fp16 = reshape(shape = var_1249, x = var_1244_cast_fp16)[name = string("k_17_cast_fp16")];
            tensor<int32, [3]> var_1255_begin_0 = const()[name = string("op_1255_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_1255_end_0 = const()[name = string("op_1255_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> var_1255_end_mask_0 = const()[name = string("op_1255_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 512]> var_1255_cast_fp16 = slice_by_index(begin = var_1255_begin_0, end = var_1255_end_0, end_mask = var_1255_end_mask_0, x = c_5_cast_fp16)[name = string("op_1255_cast_fp16")];
            tensor<int32, [4]> var_1260 = const()[name = string("op_1260"), val = tensor<int32, [4]>([1, 16, 4, 128])];
            tensor<fp16, [1, 16, 4, 128]> var_1261_cast_fp16 = reshape(shape = var_1260, x = var_1255_cast_fp16)[name = string("op_1261_cast_fp16")];
            tensor<int32, [4]> var_1266 = const()[name = string("op_1266"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1268_to_fp16 = const()[name = string("op_1268_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 16, 2, 128]> q_19_cast_fp16 = mul(x = q_17_cast_fp16, y = var_1268_to_fp16)[name = string("q_19_cast_fp16")];
            fp16 var_1270_to_fp16 = const()[name = string("op_1270_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 16, 2, 128]> k_19_cast_fp16 = mul(x = k_17_cast_fp16, y = var_1270_to_fp16)[name = string("k_19_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_1272_cast_fp16 = mul(x = q_19_cast_fp16, y = q_19_cast_fp16)[name = string("op_1272_cast_fp16")];
            tensor<int32, [1]> var_1277_axes_0 = const()[name = string("op_1277_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1277_keep_dims_0 = const()[name = string("op_1277_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 2, 1]> var_1277_cast_fp16 = reduce_mean(axes = var_1277_axes_0, keep_dims = var_1277_keep_dims_0, x = var_1272_cast_fp16)[name = string("op_1277_cast_fp16")];
            fp16 var_1278_promoted_to_fp16 = const()[name = string("op_1278_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 16, 2, 1]> var_1279_cast_fp16 = mul(x = var_1277_cast_fp16, y = var_1278_promoted_to_fp16)[name = string("op_1279_cast_fp16")];
            fp16 var_1281_to_fp16 = const()[name = string("op_1281_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 16, 2, 1]> var_1282_cast_fp16 = add(x = var_1279_cast_fp16, y = var_1281_to_fp16)[name = string("op_1282_cast_fp16")];
            fp32 var_1283_epsilon_0 = const()[name = string("op_1283_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 2, 1]> var_1283_cast_fp16 = rsqrt(epsilon = var_1283_epsilon_0, x = var_1282_cast_fp16)[name = string("op_1283_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_1284_cast_fp16 = mul(x = q_19_cast_fp16, y = var_1283_cast_fp16)[name = string("op_1284_cast_fp16")];
            fp16 var_1285_to_fp16 = const()[name = string("op_1285_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 16, 2, 128]> q_21_cast_fp16 = mul(x = var_1284_cast_fp16, y = var_1285_to_fp16)[name = string("q_21_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_1287_cast_fp16 = mul(x = k_19_cast_fp16, y = k_19_cast_fp16)[name = string("op_1287_cast_fp16")];
            tensor<int32, [1]> var_1292_axes_0 = const()[name = string("op_1292_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1292_keep_dims_0 = const()[name = string("op_1292_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 2, 1]> var_1292_cast_fp16 = reduce_mean(axes = var_1292_axes_0, keep_dims = var_1292_keep_dims_0, x = var_1287_cast_fp16)[name = string("op_1292_cast_fp16")];
            fp16 var_1293_promoted_to_fp16 = const()[name = string("op_1293_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 16, 2, 1]> var_1294_cast_fp16 = mul(x = var_1292_cast_fp16, y = var_1293_promoted_to_fp16)[name = string("op_1294_cast_fp16")];
            fp16 var_1296_to_fp16 = const()[name = string("op_1296_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 16, 2, 1]> var_1297_cast_fp16 = add(x = var_1294_cast_fp16, y = var_1296_to_fp16)[name = string("op_1297_cast_fp16")];
            fp32 var_1298_epsilon_0 = const()[name = string("op_1298_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 2, 1]> var_1298_cast_fp16 = rsqrt(epsilon = var_1298_epsilon_0, x = var_1297_cast_fp16)[name = string("op_1298_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> k_21_cast_fp16 = mul(x = k_19_cast_fp16, y = var_1298_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<int32, [4]> var_1304 = const()[name = string("op_1304"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1317_axes_0 = const()[name = string("op_1317_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 16, 128]> var_1305_cast_fp16 = transpose(perm = var_1304, x = q_21_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 2, 1, 16, 128]> var_1317_cast_fp16 = expand_dims(axes = var_1317_axes_0, x = var_1305_cast_fp16)[name = string("op_1317_cast_fp16")];
            tensor<int32, [5]> var_1325_reps_0 = const()[name = string("op_1325_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 16, 128]> var_1325_cast_fp16 = tile(reps = var_1325_reps_0, x = var_1317_cast_fp16)[name = string("op_1325_cast_fp16")];
            tensor<int32, [4]> var_1330 = const()[name = string("op_1330"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<fp16, [1, 4, 16, 128]> q_23_cast_fp16 = reshape(shape = var_1330, x = var_1325_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<int32, [4]> var_1336 = const()[name = string("op_1336"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1349_axes_0 = const()[name = string("op_1349_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 16, 128]> var_1337_cast_fp16 = transpose(perm = var_1336, x = k_21_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 2, 1, 16, 128]> var_1349_cast_fp16 = expand_dims(axes = var_1349_axes_0, x = var_1337_cast_fp16)[name = string("op_1349_cast_fp16")];
            tensor<int32, [5]> var_1357_reps_0 = const()[name = string("op_1357_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 16, 128]> var_1357_cast_fp16 = tile(reps = var_1357_reps_0, x = var_1349_cast_fp16)[name = string("op_1357_cast_fp16")];
            tensor<int32, [4]> var_1362 = const()[name = string("op_1362"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<fp16, [1, 4, 16, 128]> k_23_cast_fp16 = reshape(shape = var_1362, x = var_1357_cast_fp16)[name = string("k_23_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_1364_cast_fp16 = sigmoid(x = b_5_cast_fp16)[name = string("op_1364_cast_fp16")];
            tensor<int32, [3]> var_1368 = const()[name = string("op_1368"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_2_dt_bias_to_fp16 = const()[name = string("layers_2_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 16, 4]> input_37_cast_fp16 = add(x = a_5_cast_fp16, y = layers_2_dt_bias_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_1372_cast_fp16 = relu(x = input_37_cast_fp16)[name = string("op_1372_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_1373_cast_fp16 = abs(x = input_37_cast_fp16)[name = string("op_1373_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 4]> var_1374_cast_fp16 = mul(x = var_1373_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_1374_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_1375_cast_fp16 = exp(x = var_1374_cast_fp16)[name = string("op_1375_cast_fp16")];
            fp16 var_1377_to_fp16 = const()[name = string("op_1377_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 16, 4]> var_1378_cast_fp16 = add(x = var_1375_cast_fp16, y = var_1377_to_fp16)[name = string("op_1378_cast_fp16")];
            fp32 var_1379_epsilon_0 = const()[name = string("op_1379_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 16, 4]> var_1379_cast_fp16 = log(epsilon = var_1379_epsilon_0, x = var_1378_cast_fp16)[name = string("op_1379_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_1381_cast_fp16 = add(x = var_1372_cast_fp16, y = var_1379_cast_fp16)[name = string("op_1381_cast_fp16")];
            tensor<fp16, [4]> layers_2_negA_to_fp16 = const()[name = string("layers_2_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.c0cp+3, -0x1.3e4p+3, -0x1.568p+1, -0x1.9acp+2])];
            tensor<fp16, [1, 16, 4]> var_1382_cast_fp16 = mul(x = layers_2_negA_to_fp16, y = var_1381_cast_fp16)[name = string("op_1382_cast_fp16")];
            tensor<int32, [3]> var_1386 = const()[name = string("op_1386"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_1389_axes_0 = const()[name = string("op_1389_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16]> beta_5_cast_fp16 = transpose(perm = var_1368, x = var_1364_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 4, 16, 1]> var_1389_cast_fp16 = expand_dims(axes = var_1389_axes_0, x = beta_5_cast_fp16)[name = string("op_1389_cast_fp16")];
            fp16 var_1390_to_fp16 = const()[name = string("op_1390_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 16, 1]> var_1391_cast_fp16 = mul(x = var_1389_cast_fp16, y = var_1390_to_fp16)[name = string("op_1391_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> v_5_cast_fp16 = transpose(perm = var_1266, x = var_1261_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 4, 16, 128]> vb_5_cast_fp16 = mul(x = v_5_cast_fp16, y = var_1391_cast_fp16)[name = string("vb_5_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> kb_5_cast_fp16 = mul(x = k_23_cast_fp16, y = var_1389_cast_fp16)[name = string("kb_5_cast_fp16")];
            tensor<int32, [2]> var_1398 = const()[name = string("op_1398"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [1, 4, 16]> gdec_5_cast_fp16 = transpose(perm = var_1386, x = var_1382_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [4, 16]> var_1399_cast_fp16 = reshape(shape = var_1398, x = gdec_5_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<fp16, [4, 256]> var_1400_cast_fp16 = linear(bias = var_556_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1399_cast_fp16)[name = string("op_1400_cast_fp16")];
            tensor<int32, [4]> var_1405 = const()[name = string("op_1405"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_5_cast_fp16 = reshape(shape = var_1405, x = var_1400_cast_fp16)[name = string("P_5_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1407_cast_fp16 = exp(x = P_5_cast_fp16)[name = string("op_1407_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_5_cast_fp16 = mul(x = var_1407_cast_fp16, y = tril_incl_to_fp16)[name = string("D_5_cast_fp16")];
            bool kk_5_transpose_x_1 = const()[name = string("kk_5_transpose_x_1"), val = bool(false)];
            bool kk_5_transpose_y_1 = const()[name = string("kk_5_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_5_cast_fp16 = matmul(transpose_x = kk_5_transpose_x_1, transpose_y = kk_5_transpose_y_1, x = kb_5_cast_fp16, y = k_23_cast_fp16)[name = string("kk_5_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_1413_cast_fp16 = mul(x = kk_5_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_1413_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1414_cast_fp16 = mul(x = var_1413_cast_fp16, y = D_5_cast_fp16)[name = string("op_1414_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_5_cast_fp16 = mul(x = var_1414_cast_fp16, y = tril_strict_to_fp16)[name = string("N_5_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1416_cast_fp16 = mul(x = D_5_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1416_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_5_cast_fp16 = mul(x = kk_5_cast_fp16, y = var_1416_cast_fp16)[name = string("Na_5_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_17_cast_fp16 = add(x = eye_to_fp16, y = N_5_cast_fp16)[name = string("Minv_17_cast_fp16")];
            bool Np_13_transpose_x_0 = const()[name = string("Np_13_transpose_x_0"), val = bool(false)];
            bool Np_13_transpose_y_0 = const()[name = string("Np_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_13_cast_fp16 = matmul(transpose_x = Np_13_transpose_x_0, transpose_y = Np_13_transpose_y_0, x = N_5_cast_fp16, y = Na_5_cast_fp16)[name = string("Np_13_cast_fp16")];
            bool Npa_9_transpose_x_0 = const()[name = string("Npa_9_transpose_x_0"), val = bool(false)];
            bool Npa_9_transpose_y_0 = const()[name = string("Npa_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_9_cast_fp16 = matmul(transpose_x = Npa_9_transpose_x_0, transpose_y = Npa_9_transpose_y_0, x = Na_5_cast_fp16, y = N_5_cast_fp16)[name = string("Npa_9_cast_fp16")];
            bool var_1422_transpose_x_0 = const()[name = string("op_1422_transpose_x_0"), val = bool(false)];
            bool var_1422_transpose_y_0 = const()[name = string("op_1422_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1422_cast_fp16 = matmul(transpose_x = var_1422_transpose_x_0, transpose_y = var_1422_transpose_y_0, x = Minv_17_cast_fp16, y = Np_13_cast_fp16)[name = string("op_1422_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_19_cast_fp16 = add(x = Minv_17_cast_fp16, y = var_1422_cast_fp16)[name = string("Minv_19_cast_fp16")];
            bool Np_15_transpose_x_0 = const()[name = string("Np_15_transpose_x_0"), val = bool(false)];
            bool Np_15_transpose_y_0 = const()[name = string("Np_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_15_cast_fp16 = matmul(transpose_x = Np_15_transpose_x_0, transpose_y = Np_15_transpose_y_0, x = Np_13_cast_fp16, y = Npa_9_cast_fp16)[name = string("Np_15_cast_fp16")];
            bool Npa_11_transpose_x_0 = const()[name = string("Npa_11_transpose_x_0"), val = bool(false)];
            bool Npa_11_transpose_y_0 = const()[name = string("Npa_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_11_cast_fp16 = matmul(transpose_x = Npa_11_transpose_x_0, transpose_y = Npa_11_transpose_y_0, x = Npa_9_cast_fp16, y = Np_13_cast_fp16)[name = string("Npa_11_cast_fp16")];
            bool var_1427_transpose_x_0 = const()[name = string("op_1427_transpose_x_0"), val = bool(false)];
            bool var_1427_transpose_y_0 = const()[name = string("op_1427_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1427_cast_fp16 = matmul(transpose_x = var_1427_transpose_x_0, transpose_y = var_1427_transpose_y_0, x = Minv_19_cast_fp16, y = Np_15_cast_fp16)[name = string("op_1427_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_21_cast_fp16 = add(x = Minv_19_cast_fp16, y = var_1427_cast_fp16)[name = string("Minv_21_cast_fp16")];
            bool Np_17_transpose_x_0 = const()[name = string("Np_17_transpose_x_0"), val = bool(false)];
            bool Np_17_transpose_y_0 = const()[name = string("Np_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_17_cast_fp16 = matmul(transpose_x = Np_17_transpose_x_0, transpose_y = Np_17_transpose_y_0, x = Np_15_cast_fp16, y = Npa_11_cast_fp16)[name = string("Np_17_cast_fp16")];
            bool var_1431_transpose_x_0 = const()[name = string("op_1431_transpose_x_0"), val = bool(false)];
            bool var_1431_transpose_y_0 = const()[name = string("op_1431_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1431_cast_fp16 = matmul(transpose_x = var_1431_transpose_x_0, transpose_y = var_1431_transpose_y_0, x = Minv_21_cast_fp16, y = Np_17_cast_fp16)[name = string("op_1431_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_23_cast_fp16 = add(x = Minv_21_cast_fp16, y = var_1431_cast_fp16)[name = string("Minv_23_cast_fp16")];
            bool u_5_transpose_x_0 = const()[name = string("u_5_transpose_x_0"), val = bool(false)];
            bool u_5_transpose_y_0 = const()[name = string("u_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_5_cast_fp16 = matmul(transpose_x = u_5_transpose_x_0, transpose_y = u_5_transpose_y_0, x = Minv_23_cast_fp16, y = vb_5_cast_fp16)[name = string("u_5_cast_fp16")];
            bool var_1438_transpose_x_1 = const()[name = string("op_1438_transpose_x_1"), val = bool(false)];
            bool var_1438_transpose_y_1 = const()[name = string("op_1438_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_1438_cast_fp16 = matmul(transpose_x = var_1438_transpose_x_1, transpose_y = var_1438_transpose_y_1, x = q_23_cast_fp16, y = k_23_cast_fp16)[name = string("op_1438_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_5_cast_fp16 = mul(x = var_1438_cast_fp16, y = D_5_cast_fp16)[name = string("A_5_cast_fp16")];
            bool o_13_transpose_x_0 = const()[name = string("o_13_transpose_x_0"), val = bool(false)];
            bool o_13_transpose_y_0 = const()[name = string("o_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> o_13_cast_fp16 = matmul(transpose_x = o_13_transpose_x_0, transpose_y = o_13_transpose_y_0, x = A_5_cast_fp16, y = u_5_cast_fp16)[name = string("o_13_cast_fp16")];
            tensor<int32, [4]> var_1445 = const()[name = string("op_1445"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1451 = const()[name = string("op_1451"), val = tensor<int32, [4]>([1, 16, 4, 128])];
            tensor<fp16, [1, 16, 4, 128]> input_39_cast_fp16 = reshape(shape = var_1451, x = z_5_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> o_15_cast_fp16 = transpose(perm = var_1445, x = o_13_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 16, 4, 128]> var_1453_cast_fp16 = mul(x = o_15_cast_fp16, y = o_15_cast_fp16)[name = string("op_1453_cast_fp16")];
            tensor<int32, [1]> var_1458_axes_0 = const()[name = string("op_1458_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1458_keep_dims_0 = const()[name = string("op_1458_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 4, 1]> var_1458_cast_fp16 = reduce_mean(axes = var_1458_axes_0, keep_dims = var_1458_keep_dims_0, x = var_1453_cast_fp16)[name = string("op_1458_cast_fp16")];
            fp16 var_1460_to_fp16 = const()[name = string("op_1460_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 16, 4, 1]> var_1461_cast_fp16 = add(x = var_1458_cast_fp16, y = var_1460_to_fp16)[name = string("op_1461_cast_fp16")];
            fp32 var_1462_epsilon_0 = const()[name = string("op_1462_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 4, 1]> var_1462_cast_fp16 = rsqrt(epsilon = var_1462_epsilon_0, x = var_1461_cast_fp16)[name = string("op_1462_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> var_1463_cast_fp16 = mul(x = o_15_cast_fp16, y = var_1462_cast_fp16)[name = string("op_1463_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> var_1465_cast_fp16 = silu(x = input_39_cast_fp16)[name = string("op_1465_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> o_17_cast_fp16 = mul(x = var_1463_cast_fp16, y = var_1465_cast_fp16)[name = string("o_17_cast_fp16")];
            tensor<int32, [3]> var_1470 = const()[name = string("op_1470"), val = tensor<int32, [3]>([1, 16, 512])];
            tensor<fp16, [1, 16, 512]> input_41_cast_fp16 = reshape(shape = var_1470, x = o_17_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [640, 512]> layers_2_out_weight_to_fp16 = const()[name = string("layers_2_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47457728)))];
            tensor<fp16, [1, 16, 640]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_out_weight_to_fp16, x = input_41_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [1, 16, 640]> x_21_cast_fp16 = add(x = x_17_cast_fp16, y = linear_9_cast_fp16)[name = string("x_21_cast_fp16")];
            fp16 var_1481_promoted_to_fp16 = const()[name = string("op_1481_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> var_1485_cast_fp16 = pow(x = x_21_cast_fp16, y = var_1481_promoted_to_fp16)[name = string("op_1485_cast_fp16")];
            tensor<int32, [1]> var_1487_axes_0 = const()[name = string("op_1487_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1487_keep_dims_0 = const()[name = string("op_1487_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1487_cast_fp16 = reduce_mean(axes = var_1487_axes_0, keep_dims = var_1487_keep_dims_0, x = var_1485_cast_fp16)[name = string("op_1487_cast_fp16")];
            fp16 var_1488_to_fp16 = const()[name = string("op_1488_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_1489_cast_fp16 = add(x = var_1487_cast_fp16, y = var_1488_to_fp16)[name = string("op_1489_cast_fp16")];
            fp32 var_1490_epsilon_0 = const()[name = string("op_1490_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1490_cast_fp16 = rsqrt(epsilon = var_1490_epsilon_0, x = var_1489_cast_fp16)[name = string("op_1490_cast_fp16")];
            tensor<fp16, [1, 16, 640]> var_1491_cast_fp16 = mul(x = x_21_cast_fp16, y = var_1490_cast_fp16)[name = string("op_1491_cast_fp16")];
            tensor<fp16, [640]> layers_2_n2_w1_to_fp16 = const()[name = string("layers_2_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48113152)))];
            tensor<fp16, [1, 16, 640]> input_43_cast_fp16 = mul(x = var_1491_cast_fp16, y = layers_2_n2_w1_to_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_2_gate_up_weight_to_fp16 = const()[name = string("layers_2_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48114496)))];
            tensor<fp16, [1, 16, 5120]> linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_2_gate_up_weight_to_fp16, x = input_43_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [3]> input_45_begin_0 = const()[name = string("input_45_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_45_end_0 = const()[name = string("input_45_end_0"), val = tensor<int32, [3]>([1, 16, 2560])];
            tensor<bool, [3]> input_45_end_mask_0 = const()[name = string("input_45_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 2560]> input_45_cast_fp16 = slice_by_index(begin = input_45_begin_0, end = input_45_end_0, end_mask = input_45_end_mask_0, x = linear_10_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [1, 16, 2560]> var_1501_cast_fp16 = silu(x = input_45_cast_fp16)[name = string("op_1501_cast_fp16")];
            tensor<int32, [3]> var_1506_begin_0 = const()[name = string("op_1506_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_1506_end_0 = const()[name = string("op_1506_end_0"), val = tensor<int32, [3]>([1, 16, 5120])];
            tensor<bool, [3]> var_1506_end_mask_0 = const()[name = string("op_1506_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 2560]> var_1506_cast_fp16 = slice_by_index(begin = var_1506_begin_0, end = var_1506_end_0, end_mask = var_1506_end_mask_0, x = linear_10_cast_fp16)[name = string("op_1506_cast_fp16")];
            tensor<fp16, [1, 16, 2560]> input_47_cast_fp16 = mul(x = var_1501_cast_fp16, y = var_1506_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_2_down_weight_to_fp16 = const()[name = string("layers_2_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54668160)))];
            tensor<fp16, [1, 16, 640]> linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_down_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 16, 640]> x_25_cast_fp16 = add(x = x_21_cast_fp16, y = linear_11_cast_fp16)[name = string("x_25_cast_fp16")];
            fp16 var_1519_to_fp16 = const()[name = string("op_1519_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 16, 640]> x_27_cast_fp16 = mul(x = x_25_cast_fp16, y = var_1519_to_fp16)[name = string("x_27_cast_fp16")];
            fp16 var_1517_promoted_to_fp16 = const()[name = string("op_1517_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> var_1521_cast_fp16 = pow(x = x_27_cast_fp16, y = var_1517_promoted_to_fp16)[name = string("op_1521_cast_fp16")];
            tensor<int32, [1]> var_1523_axes_0 = const()[name = string("op_1523_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1523_keep_dims_0 = const()[name = string("op_1523_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1523_cast_fp16 = reduce_mean(axes = var_1523_axes_0, keep_dims = var_1523_keep_dims_0, x = var_1521_cast_fp16)[name = string("op_1523_cast_fp16")];
            fp16 var_1524_to_fp16 = const()[name = string("op_1524_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 16, 1]> var_1525_cast_fp16 = add(x = var_1523_cast_fp16, y = var_1524_to_fp16)[name = string("op_1525_cast_fp16")];
            fp32 var_1526_epsilon_0 = const()[name = string("op_1526_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1526_cast_fp16 = rsqrt(epsilon = var_1526_epsilon_0, x = var_1525_cast_fp16)[name = string("op_1526_cast_fp16")];
            tensor<fp16, [1, 16, 640]> var_1527_cast_fp16 = mul(x = x_27_cast_fp16, y = var_1526_cast_fp16)[name = string("op_1527_cast_fp16")];
            tensor<fp16, [640]> layers_3_n1_w1_to_fp16 = const()[name = string("layers_3_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57945024)))];
            tensor<fp16, [1, 16, 640]> input_49_cast_fp16 = mul(x = var_1527_cast_fp16, y = layers_3_n1_w1_to_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1280, 640]> layers_3_qkv_weight_to_fp16 = const()[name = string("layers_3_qkv_weight_to_fp16"), val = tensor<fp16, [1280, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57946368)))];
            tensor<fp16, [1280]> linear_12_bias_0_to_fp16 = const()[name = string("linear_12_bias_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59584832)))];
            tensor<fp16, [1, 16, 1280]> linear_12_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = layers_3_qkv_weight_to_fp16, x = input_49_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [3]> var_1536_begin_0 = const()[name = string("op_1536_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1536_end_0 = const()[name = string("op_1536_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> var_1536_end_mask_0 = const()[name = string("op_1536_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 1024]> var_1536_cast_fp16 = slice_by_index(begin = var_1536_begin_0, end = var_1536_end_0, end_mask = var_1536_end_mask_0, x = linear_12_cast_fp16)[name = string("op_1536_cast_fp16")];
            tensor<int32, [4]> var_1541 = const()[name = string("op_1541"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> qg_1_cast_fp16 = reshape(shape = var_1541, x = var_1536_cast_fp16)[name = string("qg_1_cast_fp16")];
            tensor<int32, [4]> x_29_begin_0 = const()[name = string("x_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_29_end_0 = const()[name = string("x_29_end_0"), val = tensor<int32, [4]>([1, 16, 8, 64])];
            tensor<bool, [4]> x_29_end_mask_0 = const()[name = string("x_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 8, 64]> x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, x = qg_1_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [4]> var_1552_begin_0 = const()[name = string("op_1552_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> var_1552_end_0 = const()[name = string("op_1552_end_0"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<bool, [4]> var_1552_end_mask_0 = const()[name = string("op_1552_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 8, 64]> var_1552_cast_fp16 = slice_by_index(begin = var_1552_begin_0, end = var_1552_end_0, end_mask = var_1552_end_mask_0, x = qg_1_cast_fp16)[name = string("op_1552_cast_fp16")];
            tensor<int32, [3]> var_1556 = const()[name = string("op_1556"), val = tensor<int32, [3]>([1, 16, 512])];
            tensor<fp16, [1, 16, 512]> gate_1_cast_fp16 = reshape(shape = var_1556, x = var_1552_cast_fp16)[name = string("gate_1_cast_fp16")];
            tensor<int32, [3]> var_1562_begin_0 = const()[name = string("op_1562_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> var_1562_end_0 = const()[name = string("op_1562_end_0"), val = tensor<int32, [3]>([1, 16, 1152])];
            tensor<bool, [3]> var_1562_end_mask_0 = const()[name = string("op_1562_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 128]> var_1562_cast_fp16 = slice_by_index(begin = var_1562_begin_0, end = var_1562_end_0, end_mask = var_1562_end_mask_0, x = linear_12_cast_fp16)[name = string("op_1562_cast_fp16")];
            tensor<int32, [4]> var_1567 = const()[name = string("op_1567"), val = tensor<int32, [4]>([1, 16, 2, 64])];
            tensor<fp16, [1, 16, 2, 64]> x_35_cast_fp16 = reshape(shape = var_1567, x = var_1562_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<int32, [3]> var_1573_begin_0 = const()[name = string("op_1573_begin_0"), val = tensor<int32, [3]>([0, 0, 1152])];
            tensor<int32, [3]> var_1573_end_0 = const()[name = string("op_1573_end_0"), val = tensor<int32, [3]>([1, 16, 1280])];
            tensor<bool, [3]> var_1573_end_mask_0 = const()[name = string("op_1573_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 128]> var_1573_cast_fp16 = slice_by_index(begin = var_1573_begin_0, end = var_1573_end_0, end_mask = var_1573_end_mask_0, x = linear_12_cast_fp16)[name = string("op_1573_cast_fp16")];
            tensor<int32, [4]> var_1578 = const()[name = string("op_1578"), val = tensor<int32, [4]>([1, 16, 2, 64])];
            tensor<fp16, [1, 16, 2, 64]> var_1579_cast_fp16 = reshape(shape = var_1578, x = var_1573_cast_fp16)[name = string("op_1579_cast_fp16")];
            tensor<int32, [4]> var_1584 = const()[name = string("op_1584"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1590_promoted_to_fp16 = const()[name = string("op_1590_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 8, 64]> var_1594_cast_fp16 = pow(x = x_29_cast_fp16, y = var_1590_promoted_to_fp16)[name = string("op_1594_cast_fp16")];
            tensor<int32, [1]> var_1596_axes_0 = const()[name = string("op_1596_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1596_keep_dims_0 = const()[name = string("op_1596_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 8, 1]> var_1596_cast_fp16 = reduce_mean(axes = var_1596_axes_0, keep_dims = var_1596_keep_dims_0, x = var_1594_cast_fp16)[name = string("op_1596_cast_fp16")];
            fp16 var_1597_to_fp16 = const()[name = string("op_1597_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 8, 1]> var_1598_cast_fp16 = add(x = var_1596_cast_fp16, y = var_1597_to_fp16)[name = string("op_1598_cast_fp16")];
            fp32 var_1599_epsilon_0 = const()[name = string("op_1599_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 8, 1]> var_1599_cast_fp16 = rsqrt(epsilon = var_1599_epsilon_0, x = var_1598_cast_fp16)[name = string("op_1599_cast_fp16")];
            tensor<fp16, [1, 16, 8, 64]> var_1600_cast_fp16 = mul(x = x_29_cast_fp16, y = var_1599_cast_fp16)[name = string("op_1600_cast_fp16")];
            tensor<fp16, [64]> layers_3_qn_w1_to_fp16 = const()[name = string("layers_3_qn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59587456)))];
            tensor<fp16, [1, 16, 8, 64]> var_1601_cast_fp16 = mul(x = var_1600_cast_fp16, y = layers_3_qn_w1_to_fp16)[name = string("op_1601_cast_fp16")];
            tensor<int32, [4]> var_1606 = const()[name = string("op_1606"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_1_begin_0 = const()[name = string("r_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_1_end_0 = const()[name = string("r_1_end_0"), val = tensor<int32, [4]>([1, 8, 16, 16])];
            tensor<bool, [4]> r_1_end_mask_0 = const()[name = string("r_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> x_33_cast_fp16 = transpose(perm = var_1606, x = var_1601_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 8, 16, 16]> r_1_cast_fp16 = slice_by_index(begin = r_1_begin_0, end = r_1_end_0, end_mask = r_1_end_mask_0, x = x_33_cast_fp16)[name = string("r_1_cast_fp16")];
            tensor<int32, [4]> p_1_begin_0 = const()[name = string("p_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_1_end_0 = const()[name = string("p_1_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> p_1_end_mask_0 = const()[name = string("p_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 48]> p_1_cast_fp16 = slice_by_index(begin = p_1_begin_0, end = p_1_end_0, end_mask = p_1_end_mask_0, x = x_33_cast_fp16)[name = string("p_1_cast_fp16")];
            tensor<int32, [4]> var_1622_begin_0 = const()[name = string("op_1622_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_1622_end_0 = const()[name = string("op_1622_end_0"), val = tensor<int32, [4]>([1, 8, 16, 16])];
            tensor<bool, [4]> var_1622_end_mask_0 = const()[name = string("op_1622_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 8]> var_1622_cast_fp16 = slice_by_index(begin = var_1622_begin_0, end = var_1622_end_0, end_mask = var_1622_end_mask_0, x = r_1_cast_fp16)[name = string("op_1622_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 16, 8]> var_1623_cast_fp16 = mul(x = var_1622_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_1623_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, 16, 8])];
            tensor<bool, [4]> var_1628_end_mask_0 = const()[name = string("op_1628_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 8]> var_1628_cast_fp16 = slice_by_index(begin = var_1628_begin_0, end = var_1628_end_0, end_mask = var_1628_end_mask_0, x = r_1_cast_fp16)[name = string("op_1628_cast_fp16")];
            int32 var_1630 = const()[name = string("op_1630"), val = int32(-1)];
            bool rh_1_interleave_0 = const()[name = string("rh_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 16]> rh_1_cast_fp16 = concat(axis = var_1630, interleave = rh_1_interleave_0, values = (var_1623_cast_fp16, var_1628_cast_fp16))[name = string("rh_1_cast_fp16")];
            tensor<fp16, [16, 16]> cos_to_fp16 = const()[name = string("cos_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59587648)))];
            tensor<fp16, [1, 8, 16, 16]> var_1632_cast_fp16 = mul(x = r_1_cast_fp16, y = cos_to_fp16)[name = string("op_1632_cast_fp16")];
            tensor<fp16, [16, 16]> sin_to_fp16 = const()[name = string("sin_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59588224)))];
            tensor<fp16, [1, 8, 16, 16]> var_1633_cast_fp16 = mul(x = rh_1_cast_fp16, y = sin_to_fp16)[name = string("op_1633_cast_fp16")];
            tensor<fp16, [1, 8, 16, 16]> var_1635_cast_fp16 = add(x = var_1632_cast_fp16, y = var_1633_cast_fp16)[name = string("op_1635_cast_fp16")];
            int32 var_1637 = const()[name = string("op_1637"), val = int32(-1)];
            bool q_25_interleave_0 = const()[name = string("q_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 64]> q_25_cast_fp16 = concat(axis = var_1637, interleave = q_25_interleave_0, values = (var_1635_cast_fp16, p_1_cast_fp16))[name = string("q_25_cast_fp16")];
            fp16 var_1643_promoted_to_fp16 = const()[name = string("op_1643_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 2, 64]> var_1647_cast_fp16 = pow(x = x_35_cast_fp16, y = var_1643_promoted_to_fp16)[name = string("op_1647_cast_fp16")];
            tensor<int32, [1]> var_1649_axes_0 = const()[name = string("op_1649_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1649_keep_dims_0 = const()[name = string("op_1649_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 2, 1]> var_1649_cast_fp16 = reduce_mean(axes = var_1649_axes_0, keep_dims = var_1649_keep_dims_0, x = var_1647_cast_fp16)[name = string("op_1649_cast_fp16")];
            fp16 var_1650_to_fp16 = const()[name = string("op_1650_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 2, 1]> var_1651_cast_fp16 = add(x = var_1649_cast_fp16, y = var_1650_to_fp16)[name = string("op_1651_cast_fp16")];
            fp32 var_1652_epsilon_0 = const()[name = string("op_1652_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 2, 1]> var_1652_cast_fp16 = rsqrt(epsilon = var_1652_epsilon_0, x = var_1651_cast_fp16)[name = string("op_1652_cast_fp16")];
            tensor<fp16, [1, 16, 2, 64]> var_1653_cast_fp16 = mul(x = x_35_cast_fp16, y = var_1652_cast_fp16)[name = string("op_1653_cast_fp16")];
            tensor<fp16, [64]> layers_3_kn_w1_to_fp16 = const()[name = string("layers_3_kn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59588800)))];
            tensor<fp16, [1, 16, 2, 64]> var_1654_cast_fp16 = mul(x = var_1653_cast_fp16, y = layers_3_kn_w1_to_fp16)[name = string("op_1654_cast_fp16")];
            tensor<int32, [4]> var_1659 = const()[name = string("op_1659"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_3_begin_0 = const()[name = string("r_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_3_end_0 = const()[name = string("r_3_end_0"), val = tensor<int32, [4]>([1, 2, 16, 16])];
            tensor<bool, [4]> r_3_end_mask_0 = const()[name = string("r_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 16, 64]> x_39_cast_fp16 = transpose(perm = var_1659, x = var_1654_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 2, 16, 16]> r_3_cast_fp16 = slice_by_index(begin = r_3_begin_0, end = r_3_end_0, end_mask = r_3_end_mask_0, x = x_39_cast_fp16)[name = string("r_3_cast_fp16")];
            tensor<int32, [4]> p_3_begin_0 = const()[name = string("p_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_3_end_0 = const()[name = string("p_3_end_0"), val = tensor<int32, [4]>([1, 2, 16, 64])];
            tensor<bool, [4]> p_3_end_mask_0 = const()[name = string("p_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 16, 48]> p_3_cast_fp16 = slice_by_index(begin = p_3_begin_0, end = p_3_end_0, end_mask = p_3_end_mask_0, x = x_39_cast_fp16)[name = string("p_3_cast_fp16")];
            tensor<int32, [4]> var_1675_begin_0 = const()[name = string("op_1675_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_1675_end_0 = const()[name = string("op_1675_end_0"), val = tensor<int32, [4]>([1, 2, 16, 16])];
            tensor<bool, [4]> var_1675_end_mask_0 = const()[name = string("op_1675_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 16, 8]> var_1675_cast_fp16 = slice_by_index(begin = var_1675_begin_0, end = var_1675_end_0, end_mask = var_1675_end_mask_0, x = r_3_cast_fp16)[name = string("op_1675_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 16, 8]> var_1676_cast_fp16 = mul(x = var_1675_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_1676_cast_fp16")];
            tensor<int32, [4]> var_1681_begin_0 = const()[name = string("op_1681_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1681_end_0 = const()[name = string("op_1681_end_0"), val = tensor<int32, [4]>([1, 2, 16, 8])];
            tensor<bool, [4]> var_1681_end_mask_0 = const()[name = string("op_1681_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 16, 8]> var_1681_cast_fp16 = slice_by_index(begin = var_1681_begin_0, end = var_1681_end_0, end_mask = var_1681_end_mask_0, x = r_3_cast_fp16)[name = string("op_1681_cast_fp16")];
            int32 var_1683 = const()[name = string("op_1683"), val = int32(-1)];
            bool rh_3_interleave_0 = const()[name = string("rh_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 16, 16]> rh_3_cast_fp16 = concat(axis = var_1683, interleave = rh_3_interleave_0, values = (var_1676_cast_fp16, var_1681_cast_fp16))[name = string("rh_3_cast_fp16")];
            tensor<fp16, [1, 2, 16, 16]> var_1685_cast_fp16 = mul(x = r_3_cast_fp16, y = cos_to_fp16)[name = string("op_1685_cast_fp16")];
            tensor<fp16, [1, 2, 16, 16]> var_1686_cast_fp16 = mul(x = rh_3_cast_fp16, y = sin_to_fp16)[name = string("op_1686_cast_fp16")];
            tensor<fp16, [1, 2, 16, 16]> var_1688_cast_fp16 = add(x = var_1685_cast_fp16, y = var_1686_cast_fp16)[name = string("op_1688_cast_fp16")];
            int32 var_1690 = const()[name = string("op_1690"), val = int32(-1)];
            bool k_25_interleave_0 = const()[name = string("k_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 16, 64]> k_25_cast_fp16 = concat(axis = var_1690, interleave = k_25_interleave_0, values = (var_1688_cast_fp16, p_3_cast_fp16))[name = string("k_25_cast_fp16")];
            tensor<int32, [1]> var_1703_axes_0 = const()[name = string("op_1703_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 16, 64]> var_1703_cast_fp16 = expand_dims(axes = var_1703_axes_0, x = k_25_cast_fp16)[name = string("op_1703_cast_fp16")];
            tensor<int32, [5]> var_1711_reps_0 = const()[name = string("op_1711_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 2, 4, 16, 64]> var_1711_cast_fp16 = tile(reps = var_1711_reps_0, x = var_1703_cast_fp16)[name = string("op_1711_cast_fp16")];
            tensor<int32, [4]> var_1716 = const()[name = string("op_1716"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<fp16, [1, 8, 16, 64]> k_27_cast_fp16 = reshape(shape = var_1716, x = var_1711_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<int32, [1]> var_1729_axes_0 = const()[name = string("op_1729_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 16, 64]> v_7_cast_fp16 = transpose(perm = var_1584, x = var_1579_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 2, 1, 16, 64]> var_1729_cast_fp16 = expand_dims(axes = var_1729_axes_0, x = v_7_cast_fp16)[name = string("op_1729_cast_fp16")];
            tensor<int32, [5]> var_1737_reps_0 = const()[name = string("op_1737_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 2, 4, 16, 64]> var_1737_cast_fp16 = tile(reps = var_1737_reps_0, x = var_1729_cast_fp16)[name = string("op_1737_cast_fp16")];
            tensor<int32, [4]> var_1742 = const()[name = string("op_1742"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<fp16, [1, 8, 16, 64]> v_9_cast_fp16 = reshape(shape = var_1742, x = var_1737_cast_fp16)[name = string("v_9_cast_fp16")];
            bool var_1747_transpose_x_1 = const()[name = string("op_1747_transpose_x_1"), val = bool(false)];
            bool var_1747_transpose_y_1 = const()[name = string("op_1747_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 16]> var_1747_cast_fp16 = matmul(transpose_x = var_1747_transpose_x_1, transpose_y = var_1747_transpose_y_1, x = q_25_cast_fp16, y = k_27_cast_fp16)[name = string("op_1747_cast_fp16")];
            fp16 var_1748_to_fp16 = const()[name = string("op_1748_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 8, 16, 16]> var_1749_cast_fp16 = mul(x = var_1747_cast_fp16, y = var_1748_to_fp16)[name = string("op_1749_cast_fp16")];
            tensor<fp16, [16, 16]> causal_add_to_fp16 = const()[name = string("causal_add_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59588992)))];
            tensor<fp16, [1, 8, 16, 16]> s_1_cast_fp16 = add(x = var_1749_cast_fp16, y = causal_add_to_fp16)[name = string("s_1_cast_fp16")];
            int32 var_1752 = const()[name = string("op_1752"), val = int32(-1)];
            tensor<fp16, [1, 8, 16, 16]> var_1754_cast_fp16 = softmax(axis = var_1752, x = s_1_cast_fp16)[name = string("op_1754_cast_fp16")];
            bool o_19_transpose_x_0 = const()[name = string("o_19_transpose_x_0"), val = bool(false)];
            bool o_19_transpose_y_0 = const()[name = string("o_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 64]> o_19_cast_fp16 = matmul(transpose_x = o_19_transpose_x_0, transpose_y = o_19_transpose_y_0, x = var_1754_cast_fp16, y = v_9_cast_fp16)[name = string("o_19_cast_fp16")];
            tensor<int32, [4]> var_1760 = const()[name = string("op_1760"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1765 = const()[name = string("op_1765"), val = tensor<int32, [3]>([1, 16, 512])];
            tensor<fp16, [1, 16, 8, 64]> var_1761_cast_fp16 = transpose(perm = var_1760, x = o_19_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 16, 512]> var_1766_cast_fp16 = reshape(shape = var_1765, x = var_1761_cast_fp16)[name = string("op_1766_cast_fp16")];
            tensor<fp16, [1, 16, 512]> var_1767_cast_fp16 = sigmoid(x = gate_1_cast_fp16)[name = string("op_1767_cast_fp16")];
            tensor<fp16, [1, 16, 512]> input_51_cast_fp16 = mul(x = var_1766_cast_fp16, y = var_1767_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [640, 512]> layers_3_o_weight_to_fp16 = const()[name = string("layers_3_o_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59589568)))];
            tensor<fp16, [1, 16, 640]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_o_weight_to_fp16, x = input_51_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 16, 640]> x_41_cast_fp16 = add(x = x_25_cast_fp16, y = linear_13_cast_fp16)[name = string("x_41_cast_fp16")];
            fp16 var_1780_to_fp16 = const()[name = string("op_1780_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 16, 640]> x_43_cast_fp16 = mul(x = x_41_cast_fp16, y = var_1780_to_fp16)[name = string("x_43_cast_fp16")];
            fp16 var_1778_promoted_to_fp16 = const()[name = string("op_1778_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> var_1782_cast_fp16 = pow(x = x_43_cast_fp16, y = var_1778_promoted_to_fp16)[name = string("op_1782_cast_fp16")];
            tensor<int32, [1]> var_1784_axes_0 = const()[name = string("op_1784_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1784_keep_dims_0 = const()[name = string("op_1784_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1784_cast_fp16 = reduce_mean(axes = var_1784_axes_0, keep_dims = var_1784_keep_dims_0, x = var_1782_cast_fp16)[name = string("op_1784_cast_fp16")];
            fp16 var_1785_to_fp16 = const()[name = string("op_1785_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 16, 1]> var_1786_cast_fp16 = add(x = var_1784_cast_fp16, y = var_1785_to_fp16)[name = string("op_1786_cast_fp16")];
            fp32 var_1787_epsilon_0 = const()[name = string("op_1787_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1787_cast_fp16 = rsqrt(epsilon = var_1787_epsilon_0, x = var_1786_cast_fp16)[name = string("op_1787_cast_fp16")];
            tensor<fp16, [1, 16, 640]> var_1788_cast_fp16 = mul(x = x_43_cast_fp16, y = var_1787_cast_fp16)[name = string("op_1788_cast_fp16")];
            tensor<fp16, [640]> layers_3_n2_w1_to_fp16 = const()[name = string("layers_3_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60244992)))];
            tensor<fp16, [1, 16, 640]> input_53_cast_fp16 = mul(x = var_1788_cast_fp16, y = layers_3_n2_w1_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_3_gate_up_weight_to_fp16 = const()[name = string("layers_3_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60246336)))];
            tensor<fp16, [1, 16, 5120]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_3_gate_up_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [3]> input_55_begin_0 = const()[name = string("input_55_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_55_end_0 = const()[name = string("input_55_end_0"), val = tensor<int32, [3]>([1, 16, 2560])];
            tensor<bool, [3]> input_55_end_mask_0 = const()[name = string("input_55_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 2560]> input_55_cast_fp16 = slice_by_index(begin = input_55_begin_0, end = input_55_end_0, end_mask = input_55_end_mask_0, x = linear_14_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [1, 16, 2560]> var_1798_cast_fp16 = silu(x = input_55_cast_fp16)[name = string("op_1798_cast_fp16")];
            tensor<int32, [3]> var_1803_begin_0 = const()[name = string("op_1803_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_1803_end_0 = const()[name = string("op_1803_end_0"), val = tensor<int32, [3]>([1, 16, 5120])];
            tensor<bool, [3]> var_1803_end_mask_0 = const()[name = string("op_1803_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 2560]> var_1803_cast_fp16 = slice_by_index(begin = var_1803_begin_0, end = var_1803_end_0, end_mask = var_1803_end_mask_0, x = linear_14_cast_fp16)[name = string("op_1803_cast_fp16")];
            tensor<fp16, [1, 16, 2560]> input_57_cast_fp16 = mul(x = var_1798_cast_fp16, y = var_1803_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_3_down_weight_to_fp16 = const()[name = string("layers_3_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66800000)))];
            tensor<fp16, [1, 16, 640]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_down_weight_to_fp16, x = input_57_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, 16, 640]> x_45_cast_fp16 = add(x = x_41_cast_fp16, y = linear_15_cast_fp16)[name = string("x_45_cast_fp16")];
            fp16 var_1816_to_fp16 = const()[name = string("op_1816_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 16, 640]> x_47_cast_fp16 = mul(x = x_45_cast_fp16, y = var_1816_to_fp16)[name = string("x_47_cast_fp16")];
            fp16 var_1814_promoted_to_fp16 = const()[name = string("op_1814_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> var_1818_cast_fp16 = pow(x = x_47_cast_fp16, y = var_1814_promoted_to_fp16)[name = string("op_1818_cast_fp16")];
            tensor<int32, [1]> var_1820_axes_0 = const()[name = string("op_1820_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1820_keep_dims_0 = const()[name = string("op_1820_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1820_cast_fp16 = reduce_mean(axes = var_1820_axes_0, keep_dims = var_1820_keep_dims_0, x = var_1818_cast_fp16)[name = string("op_1820_cast_fp16")];
            fp16 var_1821_to_fp16 = const()[name = string("op_1821_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 16, 1]> var_1822_cast_fp16 = add(x = var_1820_cast_fp16, y = var_1821_to_fp16)[name = string("op_1822_cast_fp16")];
            fp32 var_1823_epsilon_0 = const()[name = string("op_1823_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1823_cast_fp16 = rsqrt(epsilon = var_1823_epsilon_0, x = var_1822_cast_fp16)[name = string("op_1823_cast_fp16")];
            tensor<fp16, [1, 16, 640]> var_1824_cast_fp16 = mul(x = x_47_cast_fp16, y = var_1823_cast_fp16)[name = string("op_1824_cast_fp16")];
            tensor<fp16, [640]> layers_4_n1_w1_to_fp16 = const()[name = string("layers_4_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70076864)))];
            tensor<fp16, [1, 16, 640]> input_59_cast_fp16 = mul(x = var_1824_cast_fp16, y = layers_4_n1_w1_to_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_4_qkvzba_weight_to_fp16 = const()[name = string("layers_4_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70078208)))];
            tensor<fp16, [1, 16, 1544]> linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_qkvzba_weight_to_fp16, x = input_59_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [3]> qkv_7_begin_0 = const()[name = string("qkv_7_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_7_end_0 = const()[name = string("qkv_7_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> qkv_7_end_mask_0 = const()[name = string("qkv_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 1024]> qkv_7_cast_fp16 = slice_by_index(begin = qkv_7_begin_0, end = qkv_7_end_0, end_mask = qkv_7_end_mask_0, x = linear_16_cast_fp16)[name = string("qkv_7_cast_fp16")];
            tensor<int32, [3]> z_7_begin_0 = const()[name = string("z_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_7_end_0 = const()[name = string("z_7_end_0"), val = tensor<int32, [3]>([1, 16, 1536])];
            tensor<bool, [3]> z_7_end_mask_0 = const()[name = string("z_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 512]> z_7_cast_fp16 = slice_by_index(begin = z_7_begin_0, end = z_7_end_0, end_mask = z_7_end_mask_0, x = linear_16_cast_fp16)[name = string("z_7_cast_fp16")];
            tensor<int32, [3]> b_7_begin_0 = const()[name = string("b_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_7_end_0 = const()[name = string("b_7_end_0"), val = tensor<int32, [3]>([1, 16, 1540])];
            tensor<bool, [3]> b_7_end_mask_0 = const()[name = string("b_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 4]> b_7_cast_fp16 = slice_by_index(begin = b_7_begin_0, end = b_7_end_0, end_mask = b_7_end_mask_0, x = linear_16_cast_fp16)[name = string("b_7_cast_fp16")];
            tensor<int32, [3]> a_7_begin_0 = const()[name = string("a_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_7_end_0 = const()[name = string("a_7_end_0"), val = tensor<int32, [3]>([1, 16, 1])];
            tensor<bool, [3]> a_7_end_mask_0 = const()[name = string("a_7_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 4]> a_7_cast_fp16 = slice_by_index(begin = a_7_begin_0, end = a_7_end_0, end_mask = a_7_end_mask_0, x = linear_16_cast_fp16)[name = string("a_7_cast_fp16")];
            int32 var_1863 = const()[name = string("op_1863"), val = int32(1)];
            bool xp_7_interleave_0 = const()[name = string("xp_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 19, 1024]> xp_7_cast_fp16 = concat(axis = var_1863, interleave = xp_7_interleave_0, values = (zpad_to_fp16, qkv_7_cast_fp16))[name = string("xp_7_cast_fp16")];
            tensor<int32, [3]> var_1874_begin_0 = const()[name = string("op_1874_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_1874_end_0 = const()[name = string("op_1874_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_1874_end_mask_0 = const()[name = string("op_1874_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 1024]> var_1874_cast_fp16 = slice_by_index(begin = var_1874_begin_0, end = var_1874_end_0, end_mask = var_1874_end_mask_0, x = xp_7_cast_fp16)[name = string("op_1874_cast_fp16")];
            tensor<fp16, [1024]> var_1882_to_fp16 = const()[name = string("op_1882_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72054592)))];
            tensor<fp16, [1, 16, 1024]> var_1883_cast_fp16 = mul(x = var_1874_cast_fp16, y = var_1882_to_fp16)[name = string("op_1883_cast_fp16")];
            tensor<int32, [3]> var_1893_begin_0 = const()[name = string("op_1893_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_1893_end_0 = const()[name = string("op_1893_end_0"), val = tensor<int32, [3]>([1, 18, 1024])];
            tensor<bool, [3]> var_1893_end_mask_0 = const()[name = string("op_1893_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_1893_cast_fp16 = slice_by_index(begin = var_1893_begin_0, end = var_1893_end_0, end_mask = var_1893_end_mask_0, x = xp_7_cast_fp16)[name = string("op_1893_cast_fp16")];
            tensor<fp16, [1024]> var_1901_to_fp16 = const()[name = string("op_1901_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72056704)))];
            tensor<fp16, [1, 16, 1024]> var_1902_cast_fp16 = mul(x = var_1893_cast_fp16, y = var_1901_to_fp16)[name = string("op_1902_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1904_cast_fp16 = add(x = var_1883_cast_fp16, y = var_1902_cast_fp16)[name = string("op_1904_cast_fp16")];
            tensor<int32, [3]> var_1914_begin_0 = const()[name = string("op_1914_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1914_end_0 = const()[name = string("op_1914_end_0"), val = tensor<int32, [3]>([1, 17, 1024])];
            tensor<bool, [3]> var_1914_end_mask_0 = const()[name = string("op_1914_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_1914_cast_fp16 = slice_by_index(begin = var_1914_begin_0, end = var_1914_end_0, end_mask = var_1914_end_mask_0, x = xp_7_cast_fp16)[name = string("op_1914_cast_fp16")];
            tensor<fp16, [1024]> var_1922_to_fp16 = const()[name = string("op_1922_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72058816)))];
            tensor<fp16, [1, 16, 1024]> var_1923_cast_fp16 = mul(x = var_1914_cast_fp16, y = var_1922_to_fp16)[name = string("op_1923_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1925_cast_fp16 = add(x = var_1904_cast_fp16, y = var_1923_cast_fp16)[name = string("op_1925_cast_fp16")];
            tensor<int32, [3]> var_1935_begin_0 = const()[name = string("op_1935_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1935_end_0 = const()[name = string("op_1935_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> var_1935_end_mask_0 = const()[name = string("op_1935_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_1935_cast_fp16 = slice_by_index(begin = var_1935_begin_0, end = var_1935_end_0, end_mask = var_1935_end_mask_0, x = xp_7_cast_fp16)[name = string("op_1935_cast_fp16")];
            tensor<fp16, [1024]> var_1943_to_fp16 = const()[name = string("op_1943_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72060928)))];
            tensor<fp16, [1, 16, 1024]> var_1944_cast_fp16 = mul(x = var_1935_cast_fp16, y = var_1943_to_fp16)[name = string("op_1944_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> input_61_cast_fp16 = add(x = var_1925_cast_fp16, y = var_1944_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> c_7_cast_fp16 = silu(x = input_61_cast_fp16)[name = string("c_7_cast_fp16")];
            tensor<int32, [3]> var_1952_begin_0 = const()[name = string("op_1952_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1952_end_0 = const()[name = string("op_1952_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
            tensor<bool, [3]> var_1952_end_mask_0 = const()[name = string("op_1952_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 256]> var_1952_cast_fp16 = slice_by_index(begin = var_1952_begin_0, end = var_1952_end_0, end_mask = var_1952_end_mask_0, x = c_7_cast_fp16)[name = string("op_1952_cast_fp16")];
            tensor<int32, [4]> var_1957 = const()[name = string("op_1957"), val = tensor<int32, [4]>([1, 16, 2, 128])];
            tensor<fp16, [1, 16, 2, 128]> q_27_cast_fp16 = reshape(shape = var_1957, x = var_1952_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<int32, [3]> var_1963_begin_0 = const()[name = string("op_1963_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_1963_end_0 = const()[name = string("op_1963_end_0"), val = tensor<int32, [3]>([1, 16, 512])];
            tensor<bool, [3]> var_1963_end_mask_0 = const()[name = string("op_1963_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 256]> var_1963_cast_fp16 = slice_by_index(begin = var_1963_begin_0, end = var_1963_end_0, end_mask = var_1963_end_mask_0, x = c_7_cast_fp16)[name = string("op_1963_cast_fp16")];
            tensor<int32, [4]> var_1968 = const()[name = string("op_1968"), val = tensor<int32, [4]>([1, 16, 2, 128])];
            tensor<fp16, [1, 16, 2, 128]> k_29_cast_fp16 = reshape(shape = var_1968, x = var_1963_cast_fp16)[name = string("k_29_cast_fp16")];
            tensor<int32, [3]> var_1974_begin_0 = const()[name = string("op_1974_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_1974_end_0 = const()[name = string("op_1974_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> var_1974_end_mask_0 = const()[name = string("op_1974_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 512]> var_1974_cast_fp16 = slice_by_index(begin = var_1974_begin_0, end = var_1974_end_0, end_mask = var_1974_end_mask_0, x = c_7_cast_fp16)[name = string("op_1974_cast_fp16")];
            tensor<int32, [4]> var_1979 = const()[name = string("op_1979"), val = tensor<int32, [4]>([1, 16, 4, 128])];
            tensor<fp16, [1, 16, 4, 128]> var_1980_cast_fp16 = reshape(shape = var_1979, x = var_1974_cast_fp16)[name = string("op_1980_cast_fp16")];
            tensor<int32, [4]> var_1985 = const()[name = string("op_1985"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1987_to_fp16 = const()[name = string("op_1987_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 16, 2, 128]> q_29_cast_fp16 = mul(x = q_27_cast_fp16, y = var_1987_to_fp16)[name = string("q_29_cast_fp16")];
            fp16 var_1989_to_fp16 = const()[name = string("op_1989_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 16, 2, 128]> k_31_cast_fp16 = mul(x = k_29_cast_fp16, y = var_1989_to_fp16)[name = string("k_31_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_1991_cast_fp16 = mul(x = q_29_cast_fp16, y = q_29_cast_fp16)[name = string("op_1991_cast_fp16")];
            tensor<int32, [1]> var_1996_axes_0 = const()[name = string("op_1996_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1996_keep_dims_0 = const()[name = string("op_1996_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 2, 1]> var_1996_cast_fp16 = reduce_mean(axes = var_1996_axes_0, keep_dims = var_1996_keep_dims_0, x = var_1991_cast_fp16)[name = string("op_1996_cast_fp16")];
            fp16 var_1997_promoted_to_fp16 = const()[name = string("op_1997_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 16, 2, 1]> var_1998_cast_fp16 = mul(x = var_1996_cast_fp16, y = var_1997_promoted_to_fp16)[name = string("op_1998_cast_fp16")];
            fp16 var_2000_to_fp16 = const()[name = string("op_2000_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 16, 2, 1]> var_2001_cast_fp16 = add(x = var_1998_cast_fp16, y = var_2000_to_fp16)[name = string("op_2001_cast_fp16")];
            fp32 var_2002_epsilon_0 = const()[name = string("op_2002_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 2, 1]> var_2002_cast_fp16 = rsqrt(epsilon = var_2002_epsilon_0, x = var_2001_cast_fp16)[name = string("op_2002_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_2003_cast_fp16 = mul(x = q_29_cast_fp16, y = var_2002_cast_fp16)[name = string("op_2003_cast_fp16")];
            fp16 var_2004_to_fp16 = const()[name = string("op_2004_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 16, 2, 128]> q_31_cast_fp16 = mul(x = var_2003_cast_fp16, y = var_2004_to_fp16)[name = string("q_31_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_2006_cast_fp16 = mul(x = k_31_cast_fp16, y = k_31_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, 16, 2, 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")];
            fp16 var_2012_promoted_to_fp16 = const()[name = string("op_2012_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 16, 2, 1]> var_2013_cast_fp16 = mul(x = var_2011_cast_fp16, y = var_2012_promoted_to_fp16)[name = string("op_2013_cast_fp16")];
            fp16 var_2015_to_fp16 = const()[name = string("op_2015_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 16, 2, 1]> var_2016_cast_fp16 = add(x = var_2013_cast_fp16, y = var_2015_to_fp16)[name = string("op_2016_cast_fp16")];
            fp32 var_2017_epsilon_0 = const()[name = string("op_2017_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 2, 1]> var_2017_cast_fp16 = rsqrt(epsilon = var_2017_epsilon_0, x = var_2016_cast_fp16)[name = string("op_2017_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> k_33_cast_fp16 = mul(x = k_31_cast_fp16, y = var_2017_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<int32, [4]> var_2023 = const()[name = string("op_2023"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2036_axes_0 = const()[name = string("op_2036_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 16, 128]> var_2024_cast_fp16 = transpose(perm = var_2023, x = q_31_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 2, 1, 16, 128]> var_2036_cast_fp16 = expand_dims(axes = var_2036_axes_0, x = var_2024_cast_fp16)[name = string("op_2036_cast_fp16")];
            tensor<int32, [5]> var_2044_reps_0 = const()[name = string("op_2044_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 16, 128]> var_2044_cast_fp16 = tile(reps = var_2044_reps_0, x = var_2036_cast_fp16)[name = string("op_2044_cast_fp16")];
            tensor<int32, [4]> var_2049 = const()[name = string("op_2049"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<fp16, [1, 4, 16, 128]> q_33_cast_fp16 = reshape(shape = var_2049, x = var_2044_cast_fp16)[name = string("q_33_cast_fp16")];
            tensor<int32, [4]> var_2055 = const()[name = string("op_2055"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2068_axes_0 = const()[name = string("op_2068_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 16, 128]> var_2056_cast_fp16 = transpose(perm = var_2055, x = k_33_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 2, 1, 16, 128]> var_2068_cast_fp16 = expand_dims(axes = var_2068_axes_0, x = var_2056_cast_fp16)[name = string("op_2068_cast_fp16")];
            tensor<int32, [5]> var_2076_reps_0 = const()[name = string("op_2076_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 16, 128]> var_2076_cast_fp16 = tile(reps = var_2076_reps_0, x = var_2068_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<int32, [4]> var_2081 = const()[name = string("op_2081"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<fp16, [1, 4, 16, 128]> k_35_cast_fp16 = reshape(shape = var_2081, x = var_2076_cast_fp16)[name = string("k_35_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_2083_cast_fp16 = sigmoid(x = b_7_cast_fp16)[name = string("op_2083_cast_fp16")];
            tensor<int32, [3]> var_2087 = const()[name = string("op_2087"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_4_dt_bias_to_fp16 = const()[name = string("layers_4_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 16, 4]> input_63_cast_fp16 = add(x = a_7_cast_fp16, y = layers_4_dt_bias_to_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_2091_cast_fp16 = relu(x = input_63_cast_fp16)[name = string("op_2091_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_2092_cast_fp16 = abs(x = input_63_cast_fp16)[name = string("op_2092_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 4]> var_2093_cast_fp16 = mul(x = var_2092_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_2093_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_2094_cast_fp16 = exp(x = var_2093_cast_fp16)[name = string("op_2094_cast_fp16")];
            fp16 var_2096_to_fp16 = const()[name = string("op_2096_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 16, 4]> var_2097_cast_fp16 = add(x = var_2094_cast_fp16, y = var_2096_to_fp16)[name = string("op_2097_cast_fp16")];
            fp32 var_2098_epsilon_0 = const()[name = string("op_2098_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 16, 4]> var_2098_cast_fp16 = log(epsilon = var_2098_epsilon_0, x = var_2097_cast_fp16)[name = string("op_2098_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_2100_cast_fp16 = add(x = var_2091_cast_fp16, y = var_2098_cast_fp16)[name = string("op_2100_cast_fp16")];
            tensor<fp16, [4]> layers_4_negA_to_fp16 = const()[name = string("layers_4_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.348p+3, -0x1.734p+2, -0x1.5a8p-1, -0x1.264p+3])];
            tensor<fp16, [1, 16, 4]> var_2101_cast_fp16 = mul(x = layers_4_negA_to_fp16, y = var_2100_cast_fp16)[name = string("op_2101_cast_fp16")];
            tensor<int32, [3]> var_2105 = const()[name = string("op_2105"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_2108_axes_0 = const()[name = string("op_2108_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16]> beta_7_cast_fp16 = transpose(perm = var_2087, x = var_2083_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 4, 16, 1]> var_2108_cast_fp16 = expand_dims(axes = var_2108_axes_0, x = beta_7_cast_fp16)[name = string("op_2108_cast_fp16")];
            fp16 var_2109_to_fp16 = const()[name = string("op_2109_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 16, 1]> var_2110_cast_fp16 = mul(x = var_2108_cast_fp16, y = var_2109_to_fp16)[name = string("op_2110_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> v_11_cast_fp16 = transpose(perm = var_1985, x = var_1980_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 4, 16, 128]> vb_7_cast_fp16 = mul(x = v_11_cast_fp16, y = var_2110_cast_fp16)[name = string("vb_7_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> kb_7_cast_fp16 = mul(x = k_35_cast_fp16, y = var_2108_cast_fp16)[name = string("kb_7_cast_fp16")];
            tensor<int32, [2]> var_2117 = const()[name = string("op_2117"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [1, 4, 16]> gdec_7_cast_fp16 = transpose(perm = var_2105, x = var_2101_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [4, 16]> var_2118_cast_fp16 = reshape(shape = var_2117, x = gdec_7_cast_fp16)[name = string("op_2118_cast_fp16")];
            tensor<fp16, [4, 256]> var_2119_cast_fp16 = linear(bias = var_556_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2118_cast_fp16)[name = string("op_2119_cast_fp16")];
            tensor<int32, [4]> var_2124 = const()[name = string("op_2124"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_7_cast_fp16 = reshape(shape = var_2124, x = var_2119_cast_fp16)[name = string("P_7_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2126_cast_fp16 = exp(x = P_7_cast_fp16)[name = string("op_2126_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_7_cast_fp16 = mul(x = var_2126_cast_fp16, y = tril_incl_to_fp16)[name = string("D_7_cast_fp16")];
            bool kk_7_transpose_x_1 = const()[name = string("kk_7_transpose_x_1"), val = bool(false)];
            bool kk_7_transpose_y_1 = const()[name = string("kk_7_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_7_cast_fp16 = matmul(transpose_x = kk_7_transpose_x_1, transpose_y = kk_7_transpose_y_1, x = kb_7_cast_fp16, y = k_35_cast_fp16)[name = string("kk_7_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_2132_cast_fp16 = mul(x = kk_7_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_2132_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2133_cast_fp16 = mul(x = var_2132_cast_fp16, y = D_7_cast_fp16)[name = string("op_2133_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_7_cast_fp16 = mul(x = var_2133_cast_fp16, y = tril_strict_to_fp16)[name = string("N_7_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2135_cast_fp16 = mul(x = D_7_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2135_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_7_cast_fp16 = mul(x = kk_7_cast_fp16, y = var_2135_cast_fp16)[name = string("Na_7_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_25_cast_fp16 = add(x = eye_to_fp16, y = N_7_cast_fp16)[name = string("Minv_25_cast_fp16")];
            bool Np_19_transpose_x_0 = const()[name = string("Np_19_transpose_x_0"), val = bool(false)];
            bool Np_19_transpose_y_0 = const()[name = string("Np_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_19_cast_fp16 = matmul(transpose_x = Np_19_transpose_x_0, transpose_y = Np_19_transpose_y_0, x = N_7_cast_fp16, y = Na_7_cast_fp16)[name = string("Np_19_cast_fp16")];
            bool Npa_13_transpose_x_0 = const()[name = string("Npa_13_transpose_x_0"), val = bool(false)];
            bool Npa_13_transpose_y_0 = const()[name = string("Npa_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_13_cast_fp16 = matmul(transpose_x = Npa_13_transpose_x_0, transpose_y = Npa_13_transpose_y_0, x = Na_7_cast_fp16, y = N_7_cast_fp16)[name = string("Npa_13_cast_fp16")];
            bool var_2141_transpose_x_0 = const()[name = string("op_2141_transpose_x_0"), val = bool(false)];
            bool var_2141_transpose_y_0 = const()[name = string("op_2141_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2141_cast_fp16 = matmul(transpose_x = var_2141_transpose_x_0, transpose_y = var_2141_transpose_y_0, x = Minv_25_cast_fp16, y = Np_19_cast_fp16)[name = string("op_2141_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_27_cast_fp16 = add(x = Minv_25_cast_fp16, y = var_2141_cast_fp16)[name = string("Minv_27_cast_fp16")];
            bool Np_21_transpose_x_0 = const()[name = string("Np_21_transpose_x_0"), val = bool(false)];
            bool Np_21_transpose_y_0 = const()[name = string("Np_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_21_cast_fp16 = matmul(transpose_x = Np_21_transpose_x_0, transpose_y = Np_21_transpose_y_0, x = Np_19_cast_fp16, y = Npa_13_cast_fp16)[name = string("Np_21_cast_fp16")];
            bool Npa_15_transpose_x_0 = const()[name = string("Npa_15_transpose_x_0"), val = bool(false)];
            bool Npa_15_transpose_y_0 = const()[name = string("Npa_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_15_cast_fp16 = matmul(transpose_x = Npa_15_transpose_x_0, transpose_y = Npa_15_transpose_y_0, x = Npa_13_cast_fp16, y = Np_19_cast_fp16)[name = string("Npa_15_cast_fp16")];
            bool var_2146_transpose_x_0 = const()[name = string("op_2146_transpose_x_0"), val = bool(false)];
            bool var_2146_transpose_y_0 = const()[name = string("op_2146_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2146_cast_fp16 = matmul(transpose_x = var_2146_transpose_x_0, transpose_y = var_2146_transpose_y_0, x = Minv_27_cast_fp16, y = Np_21_cast_fp16)[name = string("op_2146_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_29_cast_fp16 = add(x = Minv_27_cast_fp16, y = var_2146_cast_fp16)[name = string("Minv_29_cast_fp16")];
            bool Np_23_transpose_x_0 = const()[name = string("Np_23_transpose_x_0"), val = bool(false)];
            bool Np_23_transpose_y_0 = const()[name = string("Np_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_23_cast_fp16 = matmul(transpose_x = Np_23_transpose_x_0, transpose_y = Np_23_transpose_y_0, x = Np_21_cast_fp16, y = Npa_15_cast_fp16)[name = string("Np_23_cast_fp16")];
            bool var_2150_transpose_x_0 = const()[name = string("op_2150_transpose_x_0"), val = bool(false)];
            bool var_2150_transpose_y_0 = const()[name = string("op_2150_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2150_cast_fp16 = matmul(transpose_x = var_2150_transpose_x_0, transpose_y = var_2150_transpose_y_0, x = Minv_29_cast_fp16, y = Np_23_cast_fp16)[name = string("op_2150_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_31_cast_fp16 = add(x = Minv_29_cast_fp16, y = var_2150_cast_fp16)[name = string("Minv_31_cast_fp16")];
            bool u_7_transpose_x_0 = const()[name = string("u_7_transpose_x_0"), val = bool(false)];
            bool u_7_transpose_y_0 = const()[name = string("u_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_7_cast_fp16 = matmul(transpose_x = u_7_transpose_x_0, transpose_y = u_7_transpose_y_0, x = Minv_31_cast_fp16, y = vb_7_cast_fp16)[name = string("u_7_cast_fp16")];
            bool var_2157_transpose_x_1 = const()[name = string("op_2157_transpose_x_1"), val = bool(false)];
            bool var_2157_transpose_y_1 = const()[name = string("op_2157_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_2157_cast_fp16 = matmul(transpose_x = var_2157_transpose_x_1, transpose_y = var_2157_transpose_y_1, x = q_33_cast_fp16, y = k_35_cast_fp16)[name = string("op_2157_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_7_cast_fp16 = mul(x = var_2157_cast_fp16, y = D_7_cast_fp16)[name = string("A_7_cast_fp16")];
            bool o_21_transpose_x_0 = const()[name = string("o_21_transpose_x_0"), val = bool(false)];
            bool o_21_transpose_y_0 = const()[name = string("o_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> o_21_cast_fp16 = matmul(transpose_x = o_21_transpose_x_0, transpose_y = o_21_transpose_y_0, x = A_7_cast_fp16, y = u_7_cast_fp16)[name = string("o_21_cast_fp16")];
            tensor<int32, [4]> var_2164 = const()[name = string("op_2164"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2170 = const()[name = string("op_2170"), val = tensor<int32, [4]>([1, 16, 4, 128])];
            tensor<fp16, [1, 16, 4, 128]> input_65_cast_fp16 = reshape(shape = var_2170, x = z_7_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> o_23_cast_fp16 = transpose(perm = var_2164, x = o_21_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 16, 4, 128]> var_2172_cast_fp16 = mul(x = o_23_cast_fp16, y = o_23_cast_fp16)[name = string("op_2172_cast_fp16")];
            tensor<int32, [1]> var_2177_axes_0 = const()[name = string("op_2177_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2177_keep_dims_0 = const()[name = string("op_2177_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 4, 1]> var_2177_cast_fp16 = reduce_mean(axes = var_2177_axes_0, keep_dims = var_2177_keep_dims_0, x = var_2172_cast_fp16)[name = string("op_2177_cast_fp16")];
            fp16 var_2179_to_fp16 = const()[name = string("op_2179_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 16, 4, 1]> var_2180_cast_fp16 = add(x = var_2177_cast_fp16, y = var_2179_to_fp16)[name = string("op_2180_cast_fp16")];
            fp32 var_2181_epsilon_0 = const()[name = string("op_2181_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 4, 1]> var_2181_cast_fp16 = rsqrt(epsilon = var_2181_epsilon_0, x = var_2180_cast_fp16)[name = string("op_2181_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> var_2182_cast_fp16 = mul(x = o_23_cast_fp16, y = var_2181_cast_fp16)[name = string("op_2182_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> var_2184_cast_fp16 = silu(x = input_65_cast_fp16)[name = string("op_2184_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> o_25_cast_fp16 = mul(x = var_2182_cast_fp16, y = var_2184_cast_fp16)[name = string("o_25_cast_fp16")];
            tensor<int32, [3]> var_2189 = const()[name = string("op_2189"), val = tensor<int32, [3]>([1, 16, 512])];
            tensor<fp16, [1, 16, 512]> input_67_cast_fp16 = reshape(shape = var_2189, x = o_25_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [640, 512]> layers_4_out_weight_to_fp16 = const()[name = string("layers_4_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72063040)))];
            tensor<fp16, [1, 16, 640]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_out_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 16, 640]> x_49_cast_fp16 = add(x = x_45_cast_fp16, y = linear_17_cast_fp16)[name = string("x_49_cast_fp16")];
            fp16 var_2202_to_fp16 = const()[name = string("op_2202_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 16, 640]> x_51_cast_fp16 = mul(x = x_49_cast_fp16, y = var_2202_to_fp16)[name = string("x_51_cast_fp16")];
            fp16 var_2200_promoted_to_fp16 = const()[name = string("op_2200_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> var_2204_cast_fp16 = pow(x = x_51_cast_fp16, y = var_2200_promoted_to_fp16)[name = string("op_2204_cast_fp16")];
            tensor<int32, [1]> var_2206_axes_0 = const()[name = string("op_2206_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2206_keep_dims_0 = const()[name = string("op_2206_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2206_cast_fp16 = reduce_mean(axes = var_2206_axes_0, keep_dims = var_2206_keep_dims_0, x = var_2204_cast_fp16)[name = string("op_2206_cast_fp16")];
            fp16 var_2207_to_fp16 = const()[name = string("op_2207_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 16, 1]> var_2208_cast_fp16 = add(x = var_2206_cast_fp16, y = var_2207_to_fp16)[name = string("op_2208_cast_fp16")];
            fp32 var_2209_epsilon_0 = const()[name = string("op_2209_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2209_cast_fp16 = rsqrt(epsilon = var_2209_epsilon_0, x = var_2208_cast_fp16)[name = string("op_2209_cast_fp16")];
            tensor<fp16, [1, 16, 640]> var_2210_cast_fp16 = mul(x = x_51_cast_fp16, y = var_2209_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<fp16, [640]> layers_4_n2_w1_to_fp16 = const()[name = string("layers_4_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72718464)))];
            tensor<fp16, [1, 16, 640]> input_69_cast_fp16 = mul(x = var_2210_cast_fp16, y = layers_4_n2_w1_to_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_4_gate_up_weight_to_fp16 = const()[name = string("layers_4_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72719808)))];
            tensor<fp16, [1, 16, 5120]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_4_gate_up_weight_to_fp16, x = input_69_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<int32, [3]> input_71_begin_0 = const()[name = string("input_71_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_71_end_0 = const()[name = string("input_71_end_0"), val = tensor<int32, [3]>([1, 16, 2560])];
            tensor<bool, [3]> input_71_end_mask_0 = const()[name = string("input_71_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 2560]> input_71_cast_fp16 = slice_by_index(begin = input_71_begin_0, end = input_71_end_0, end_mask = input_71_end_mask_0, x = linear_18_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [1, 16, 2560]> var_2220_cast_fp16 = silu(x = input_71_cast_fp16)[name = string("op_2220_cast_fp16")];
            tensor<int32, [3]> var_2225_begin_0 = const()[name = string("op_2225_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_2225_end_0 = const()[name = string("op_2225_end_0"), val = tensor<int32, [3]>([1, 16, 5120])];
            tensor<bool, [3]> var_2225_end_mask_0 = const()[name = string("op_2225_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 2560]> var_2225_cast_fp16 = slice_by_index(begin = var_2225_begin_0, end = var_2225_end_0, end_mask = var_2225_end_mask_0, x = linear_18_cast_fp16)[name = string("op_2225_cast_fp16")];
            tensor<fp16, [1, 16, 2560]> input_73_cast_fp16 = mul(x = var_2220_cast_fp16, y = var_2225_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_4_down_weight_to_fp16 = const()[name = string("layers_4_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79273472)))];
            tensor<fp16, [1, 16, 640]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_down_weight_to_fp16, x = input_73_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 16, 640]> x_53_cast_fp16 = add(x = x_49_cast_fp16, y = linear_19_cast_fp16)[name = string("x_53_cast_fp16")];
            fp16 var_2238_to_fp16 = const()[name = string("op_2238_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 16, 640]> x_55_cast_fp16 = mul(x = x_53_cast_fp16, y = var_2238_to_fp16)[name = string("x_55_cast_fp16")];
            fp16 var_2236_promoted_to_fp16 = const()[name = string("op_2236_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> var_2240_cast_fp16 = pow(x = x_55_cast_fp16, y = var_2236_promoted_to_fp16)[name = string("op_2240_cast_fp16")];
            tensor<int32, [1]> var_2242_axes_0 = const()[name = string("op_2242_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2242_keep_dims_0 = const()[name = string("op_2242_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2242_cast_fp16 = reduce_mean(axes = var_2242_axes_0, keep_dims = var_2242_keep_dims_0, x = var_2240_cast_fp16)[name = string("op_2242_cast_fp16")];
            fp16 var_2243_to_fp16 = const()[name = string("op_2243_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 16, 1]> var_2244_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2243_to_fp16)[name = string("op_2244_cast_fp16")];
            fp32 var_2245_epsilon_0 = const()[name = string("op_2245_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2245_cast_fp16 = rsqrt(epsilon = var_2245_epsilon_0, x = var_2244_cast_fp16)[name = string("op_2245_cast_fp16")];
            tensor<fp16, [1, 16, 640]> var_2246_cast_fp16 = mul(x = x_55_cast_fp16, y = var_2245_cast_fp16)[name = string("op_2246_cast_fp16")];
            tensor<fp16, [640]> layers_5_n1_w1_to_fp16 = const()[name = string("layers_5_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82550336)))];
            tensor<fp16, [1, 16, 640]> input_75_cast_fp16 = mul(x = var_2246_cast_fp16, y = layers_5_n1_w1_to_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_5_qkvzba_weight_to_fp16 = const()[name = string("layers_5_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82551680)))];
            tensor<fp16, [1, 16, 1544]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_qkvzba_weight_to_fp16, x = input_75_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [3]> qkv_9_begin_0 = const()[name = string("qkv_9_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_9_end_0 = const()[name = string("qkv_9_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> qkv_9_end_mask_0 = const()[name = string("qkv_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 1024]> qkv_9_cast_fp16 = slice_by_index(begin = qkv_9_begin_0, end = qkv_9_end_0, end_mask = qkv_9_end_mask_0, x = linear_20_cast_fp16)[name = string("qkv_9_cast_fp16")];
            tensor<int32, [3]> z_9_begin_0 = const()[name = string("z_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_9_end_0 = const()[name = string("z_9_end_0"), val = tensor<int32, [3]>([1, 16, 1536])];
            tensor<bool, [3]> z_9_end_mask_0 = const()[name = string("z_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 512]> z_9_cast_fp16 = slice_by_index(begin = z_9_begin_0, end = z_9_end_0, end_mask = z_9_end_mask_0, x = linear_20_cast_fp16)[name = string("z_9_cast_fp16")];
            tensor<int32, [3]> b_9_begin_0 = const()[name = string("b_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_9_end_0 = const()[name = string("b_9_end_0"), val = tensor<int32, [3]>([1, 16, 1540])];
            tensor<bool, [3]> b_9_end_mask_0 = const()[name = string("b_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 4]> b_9_cast_fp16 = slice_by_index(begin = b_9_begin_0, end = b_9_end_0, end_mask = b_9_end_mask_0, x = linear_20_cast_fp16)[name = string("b_9_cast_fp16")];
            tensor<int32, [3]> a_9_begin_0 = const()[name = string("a_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_9_end_0 = const()[name = string("a_9_end_0"), val = tensor<int32, [3]>([1, 16, 1])];
            tensor<bool, [3]> a_9_end_mask_0 = const()[name = string("a_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 4]> a_9_cast_fp16 = slice_by_index(begin = a_9_begin_0, end = a_9_end_0, end_mask = a_9_end_mask_0, x = linear_20_cast_fp16)[name = string("a_9_cast_fp16")];
            int32 var_2285 = const()[name = string("op_2285"), val = int32(1)];
            bool xp_9_interleave_0 = const()[name = string("xp_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 19, 1024]> xp_9_cast_fp16 = concat(axis = var_2285, interleave = xp_9_interleave_0, values = (zpad_to_fp16, qkv_9_cast_fp16))[name = string("xp_9_cast_fp16")];
            tensor<int32, [3]> var_2296_begin_0 = const()[name = string("op_2296_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_2296_end_0 = const()[name = string("op_2296_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_2296_end_mask_0 = const()[name = string("op_2296_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 1024]> var_2296_cast_fp16 = slice_by_index(begin = var_2296_begin_0, end = var_2296_end_0, end_mask = var_2296_end_mask_0, x = xp_9_cast_fp16)[name = string("op_2296_cast_fp16")];
            tensor<fp16, [1024]> var_2304_to_fp16 = const()[name = string("op_2304_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84528064)))];
            tensor<fp16, [1, 16, 1024]> var_2305_cast_fp16 = mul(x = var_2296_cast_fp16, y = var_2304_to_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [3]> var_2315_begin_0 = const()[name = string("op_2315_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_2315_end_0 = const()[name = string("op_2315_end_0"), val = tensor<int32, [3]>([1, 18, 1024])];
            tensor<bool, [3]> var_2315_end_mask_0 = const()[name = string("op_2315_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_2315_cast_fp16 = slice_by_index(begin = var_2315_begin_0, end = var_2315_end_0, end_mask = var_2315_end_mask_0, x = xp_9_cast_fp16)[name = string("op_2315_cast_fp16")];
            tensor<fp16, [1024]> var_2323_to_fp16 = const()[name = string("op_2323_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84530176)))];
            tensor<fp16, [1, 16, 1024]> var_2324_cast_fp16 = mul(x = var_2315_cast_fp16, y = var_2323_to_fp16)[name = string("op_2324_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2326_cast_fp16 = add(x = var_2305_cast_fp16, y = var_2324_cast_fp16)[name = string("op_2326_cast_fp16")];
            tensor<int32, [3]> var_2336_begin_0 = const()[name = string("op_2336_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2336_end_0 = const()[name = string("op_2336_end_0"), val = tensor<int32, [3]>([1, 17, 1024])];
            tensor<bool, [3]> var_2336_end_mask_0 = const()[name = string("op_2336_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_2336_cast_fp16 = slice_by_index(begin = var_2336_begin_0, end = var_2336_end_0, end_mask = var_2336_end_mask_0, x = xp_9_cast_fp16)[name = string("op_2336_cast_fp16")];
            tensor<fp16, [1024]> var_2344_to_fp16 = const()[name = string("op_2344_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84532288)))];
            tensor<fp16, [1, 16, 1024]> var_2345_cast_fp16 = mul(x = var_2336_cast_fp16, y = var_2344_to_fp16)[name = string("op_2345_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2347_cast_fp16 = add(x = var_2326_cast_fp16, y = var_2345_cast_fp16)[name = string("op_2347_cast_fp16")];
            tensor<int32, [3]> var_2357_begin_0 = const()[name = string("op_2357_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2357_end_0 = const()[name = string("op_2357_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> var_2357_end_mask_0 = const()[name = string("op_2357_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_2357_cast_fp16 = slice_by_index(begin = var_2357_begin_0, end = var_2357_end_0, end_mask = var_2357_end_mask_0, x = xp_9_cast_fp16)[name = string("op_2357_cast_fp16")];
            tensor<fp16, [1024]> var_2365_to_fp16 = const()[name = string("op_2365_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84534400)))];
            tensor<fp16, [1, 16, 1024]> var_2366_cast_fp16 = mul(x = var_2357_cast_fp16, y = var_2365_to_fp16)[name = string("op_2366_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> input_77_cast_fp16 = add(x = var_2347_cast_fp16, y = var_2366_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> c_9_cast_fp16 = silu(x = input_77_cast_fp16)[name = string("c_9_cast_fp16")];
            tensor<int32, [3]> var_2374_begin_0 = const()[name = string("op_2374_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2374_end_0 = const()[name = string("op_2374_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
            tensor<bool, [3]> var_2374_end_mask_0 = const()[name = string("op_2374_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 256]> var_2374_cast_fp16 = slice_by_index(begin = var_2374_begin_0, end = var_2374_end_0, end_mask = var_2374_end_mask_0, x = c_9_cast_fp16)[name = string("op_2374_cast_fp16")];
            tensor<int32, [4]> var_2379 = const()[name = string("op_2379"), val = tensor<int32, [4]>([1, 16, 2, 128])];
            tensor<fp16, [1, 16, 2, 128]> q_35_cast_fp16 = reshape(shape = var_2379, x = var_2374_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<int32, [3]> var_2385_begin_0 = const()[name = string("op_2385_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_2385_end_0 = const()[name = string("op_2385_end_0"), val = tensor<int32, [3]>([1, 16, 512])];
            tensor<bool, [3]> var_2385_end_mask_0 = const()[name = string("op_2385_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 256]> var_2385_cast_fp16 = slice_by_index(begin = var_2385_begin_0, end = var_2385_end_0, end_mask = var_2385_end_mask_0, x = c_9_cast_fp16)[name = string("op_2385_cast_fp16")];
            tensor<int32, [4]> var_2390 = const()[name = string("op_2390"), val = tensor<int32, [4]>([1, 16, 2, 128])];
            tensor<fp16, [1, 16, 2, 128]> k_37_cast_fp16 = reshape(shape = var_2390, x = var_2385_cast_fp16)[name = string("k_37_cast_fp16")];
            tensor<int32, [3]> var_2396_begin_0 = const()[name = string("op_2396_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_2396_end_0 = const()[name = string("op_2396_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> var_2396_end_mask_0 = const()[name = string("op_2396_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 512]> var_2396_cast_fp16 = slice_by_index(begin = var_2396_begin_0, end = var_2396_end_0, end_mask = var_2396_end_mask_0, x = c_9_cast_fp16)[name = string("op_2396_cast_fp16")];
            tensor<int32, [4]> var_2401 = const()[name = string("op_2401"), val = tensor<int32, [4]>([1, 16, 4, 128])];
            tensor<fp16, [1, 16, 4, 128]> var_2402_cast_fp16 = reshape(shape = var_2401, x = var_2396_cast_fp16)[name = string("op_2402_cast_fp16")];
            tensor<int32, [4]> var_2407 = const()[name = string("op_2407"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2409_to_fp16 = const()[name = string("op_2409_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 16, 2, 128]> q_37_cast_fp16 = mul(x = q_35_cast_fp16, y = var_2409_to_fp16)[name = string("q_37_cast_fp16")];
            fp16 var_2411_to_fp16 = const()[name = string("op_2411_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 16, 2, 128]> k_39_cast_fp16 = mul(x = k_37_cast_fp16, y = var_2411_to_fp16)[name = string("k_39_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_2413_cast_fp16 = mul(x = q_37_cast_fp16, y = q_37_cast_fp16)[name = string("op_2413_cast_fp16")];
            tensor<int32, [1]> var_2418_axes_0 = const()[name = string("op_2418_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2418_keep_dims_0 = const()[name = string("op_2418_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 2, 1]> var_2418_cast_fp16 = reduce_mean(axes = var_2418_axes_0, keep_dims = var_2418_keep_dims_0, x = var_2413_cast_fp16)[name = string("op_2418_cast_fp16")];
            fp16 var_2419_promoted_to_fp16 = const()[name = string("op_2419_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 16, 2, 1]> var_2420_cast_fp16 = mul(x = var_2418_cast_fp16, y = var_2419_promoted_to_fp16)[name = string("op_2420_cast_fp16")];
            fp16 var_2422_to_fp16 = const()[name = string("op_2422_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 16, 2, 1]> var_2423_cast_fp16 = add(x = var_2420_cast_fp16, y = var_2422_to_fp16)[name = string("op_2423_cast_fp16")];
            fp32 var_2424_epsilon_0 = const()[name = string("op_2424_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 2, 1]> var_2424_cast_fp16 = rsqrt(epsilon = var_2424_epsilon_0, x = var_2423_cast_fp16)[name = string("op_2424_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_2425_cast_fp16 = mul(x = q_37_cast_fp16, y = var_2424_cast_fp16)[name = string("op_2425_cast_fp16")];
            fp16 var_2426_to_fp16 = const()[name = string("op_2426_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 16, 2, 128]> q_39_cast_fp16 = mul(x = var_2425_cast_fp16, y = var_2426_to_fp16)[name = string("q_39_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_2428_cast_fp16 = mul(x = k_39_cast_fp16, y = k_39_cast_fp16)[name = string("op_2428_cast_fp16")];
            tensor<int32, [1]> var_2433_axes_0 = const()[name = string("op_2433_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2433_keep_dims_0 = const()[name = string("op_2433_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 2, 1]> var_2433_cast_fp16 = reduce_mean(axes = var_2433_axes_0, keep_dims = var_2433_keep_dims_0, x = var_2428_cast_fp16)[name = string("op_2433_cast_fp16")];
            fp16 var_2434_promoted_to_fp16 = const()[name = string("op_2434_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 16, 2, 1]> var_2435_cast_fp16 = mul(x = var_2433_cast_fp16, y = var_2434_promoted_to_fp16)[name = string("op_2435_cast_fp16")];
            fp16 var_2437_to_fp16 = const()[name = string("op_2437_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 16, 2, 1]> var_2438_cast_fp16 = add(x = var_2435_cast_fp16, y = var_2437_to_fp16)[name = string("op_2438_cast_fp16")];
            fp32 var_2439_epsilon_0 = const()[name = string("op_2439_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 2, 1]> var_2439_cast_fp16 = rsqrt(epsilon = var_2439_epsilon_0, x = var_2438_cast_fp16)[name = string("op_2439_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> k_41_cast_fp16 = mul(x = k_39_cast_fp16, y = var_2439_cast_fp16)[name = string("k_41_cast_fp16")];
            tensor<int32, [4]> var_2445 = const()[name = string("op_2445"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2458_axes_0 = const()[name = string("op_2458_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 16, 128]> var_2446_cast_fp16 = transpose(perm = var_2445, x = q_39_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 2, 1, 16, 128]> var_2458_cast_fp16 = expand_dims(axes = var_2458_axes_0, x = var_2446_cast_fp16)[name = string("op_2458_cast_fp16")];
            tensor<int32, [5]> var_2466_reps_0 = const()[name = string("op_2466_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 16, 128]> var_2466_cast_fp16 = tile(reps = var_2466_reps_0, x = var_2458_cast_fp16)[name = string("op_2466_cast_fp16")];
            tensor<int32, [4]> var_2471 = const()[name = string("op_2471"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<fp16, [1, 4, 16, 128]> q_41_cast_fp16 = reshape(shape = var_2471, x = var_2466_cast_fp16)[name = string("q_41_cast_fp16")];
            tensor<int32, [4]> var_2477 = const()[name = string("op_2477"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2490_axes_0 = const()[name = string("op_2490_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 16, 128]> var_2478_cast_fp16 = transpose(perm = var_2477, x = k_41_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 2, 1, 16, 128]> var_2490_cast_fp16 = expand_dims(axes = var_2490_axes_0, x = var_2478_cast_fp16)[name = string("op_2490_cast_fp16")];
            tensor<int32, [5]> var_2498_reps_0 = const()[name = string("op_2498_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 16, 128]> var_2498_cast_fp16 = tile(reps = var_2498_reps_0, x = var_2490_cast_fp16)[name = string("op_2498_cast_fp16")];
            tensor<int32, [4]> var_2503 = const()[name = string("op_2503"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<fp16, [1, 4, 16, 128]> k_43_cast_fp16 = reshape(shape = var_2503, x = var_2498_cast_fp16)[name = string("k_43_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_2505_cast_fp16 = sigmoid(x = b_9_cast_fp16)[name = string("op_2505_cast_fp16")];
            tensor<int32, [3]> var_2509 = const()[name = string("op_2509"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_5_dt_bias_to_fp16 = const()[name = string("layers_5_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 16, 4]> input_79_cast_fp16 = add(x = a_9_cast_fp16, y = layers_5_dt_bias_to_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_2513_cast_fp16 = relu(x = input_79_cast_fp16)[name = string("op_2513_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_2514_cast_fp16 = abs(x = input_79_cast_fp16)[name = string("op_2514_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 4]> var_2515_cast_fp16 = mul(x = var_2514_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_2515_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_2516_cast_fp16 = exp(x = var_2515_cast_fp16)[name = string("op_2516_cast_fp16")];
            fp16 var_2518_to_fp16 = const()[name = string("op_2518_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 16, 4]> var_2519_cast_fp16 = add(x = var_2516_cast_fp16, y = var_2518_to_fp16)[name = string("op_2519_cast_fp16")];
            fp32 var_2520_epsilon_0 = const()[name = string("op_2520_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 16, 4]> var_2520_cast_fp16 = log(epsilon = var_2520_epsilon_0, x = var_2519_cast_fp16)[name = string("op_2520_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_2522_cast_fp16 = add(x = var_2513_cast_fp16, y = var_2520_cast_fp16)[name = string("op_2522_cast_fp16")];
            tensor<fp16, [4]> layers_5_negA_to_fp16 = const()[name = string("layers_5_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.2f8p+3, -0x1.ae8p+2, -0x1.548p+2, -0x1.63p+3])];
            tensor<fp16, [1, 16, 4]> var_2523_cast_fp16 = mul(x = layers_5_negA_to_fp16, y = var_2522_cast_fp16)[name = string("op_2523_cast_fp16")];
            tensor<int32, [3]> var_2527 = const()[name = string("op_2527"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_2530_axes_0 = const()[name = string("op_2530_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16]> beta_9_cast_fp16 = transpose(perm = var_2509, x = var_2505_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 4, 16, 1]> var_2530_cast_fp16 = expand_dims(axes = var_2530_axes_0, x = beta_9_cast_fp16)[name = string("op_2530_cast_fp16")];
            fp16 var_2531_to_fp16 = const()[name = string("op_2531_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 16, 1]> var_2532_cast_fp16 = mul(x = var_2530_cast_fp16, y = var_2531_to_fp16)[name = string("op_2532_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> v_13_cast_fp16 = transpose(perm = var_2407, x = var_2402_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 4, 16, 128]> vb_9_cast_fp16 = mul(x = v_13_cast_fp16, y = var_2532_cast_fp16)[name = string("vb_9_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> kb_9_cast_fp16 = mul(x = k_43_cast_fp16, y = var_2530_cast_fp16)[name = string("kb_9_cast_fp16")];
            tensor<int32, [2]> var_2539 = const()[name = string("op_2539"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [1, 4, 16]> gdec_9_cast_fp16 = transpose(perm = var_2527, x = var_2523_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [4, 16]> var_2540_cast_fp16 = reshape(shape = var_2539, x = gdec_9_cast_fp16)[name = string("op_2540_cast_fp16")];
            tensor<fp16, [4, 256]> var_2541_cast_fp16 = linear(bias = var_556_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2540_cast_fp16)[name = string("op_2541_cast_fp16")];
            tensor<int32, [4]> var_2546 = const()[name = string("op_2546"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_9_cast_fp16 = reshape(shape = var_2546, x = var_2541_cast_fp16)[name = string("P_9_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2548_cast_fp16 = exp(x = P_9_cast_fp16)[name = string("op_2548_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_9_cast_fp16 = mul(x = var_2548_cast_fp16, y = tril_incl_to_fp16)[name = string("D_9_cast_fp16")];
            bool kk_9_transpose_x_1 = const()[name = string("kk_9_transpose_x_1"), val = bool(false)];
            bool kk_9_transpose_y_1 = const()[name = string("kk_9_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_9_cast_fp16 = matmul(transpose_x = kk_9_transpose_x_1, transpose_y = kk_9_transpose_y_1, x = kb_9_cast_fp16, y = k_43_cast_fp16)[name = string("kk_9_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_2554_cast_fp16 = mul(x = kk_9_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_2554_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2555_cast_fp16 = mul(x = var_2554_cast_fp16, y = D_9_cast_fp16)[name = string("op_2555_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_9_cast_fp16 = mul(x = var_2555_cast_fp16, y = tril_strict_to_fp16)[name = string("N_9_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2557_cast_fp16 = mul(x = D_9_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2557_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_9_cast_fp16 = mul(x = kk_9_cast_fp16, y = var_2557_cast_fp16)[name = string("Na_9_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_33_cast_fp16 = add(x = eye_to_fp16, y = N_9_cast_fp16)[name = string("Minv_33_cast_fp16")];
            bool Np_25_transpose_x_0 = const()[name = string("Np_25_transpose_x_0"), val = bool(false)];
            bool Np_25_transpose_y_0 = const()[name = string("Np_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_25_cast_fp16 = matmul(transpose_x = Np_25_transpose_x_0, transpose_y = Np_25_transpose_y_0, x = N_9_cast_fp16, y = Na_9_cast_fp16)[name = string("Np_25_cast_fp16")];
            bool Npa_17_transpose_x_0 = const()[name = string("Npa_17_transpose_x_0"), val = bool(false)];
            bool Npa_17_transpose_y_0 = const()[name = string("Npa_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_17_cast_fp16 = matmul(transpose_x = Npa_17_transpose_x_0, transpose_y = Npa_17_transpose_y_0, x = Na_9_cast_fp16, y = N_9_cast_fp16)[name = string("Npa_17_cast_fp16")];
            bool var_2563_transpose_x_0 = const()[name = string("op_2563_transpose_x_0"), val = bool(false)];
            bool var_2563_transpose_y_0 = const()[name = string("op_2563_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2563_cast_fp16 = matmul(transpose_x = var_2563_transpose_x_0, transpose_y = var_2563_transpose_y_0, x = Minv_33_cast_fp16, y = Np_25_cast_fp16)[name = string("op_2563_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_35_cast_fp16 = add(x = Minv_33_cast_fp16, y = var_2563_cast_fp16)[name = string("Minv_35_cast_fp16")];
            bool Np_27_transpose_x_0 = const()[name = string("Np_27_transpose_x_0"), val = bool(false)];
            bool Np_27_transpose_y_0 = const()[name = string("Np_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_27_cast_fp16 = matmul(transpose_x = Np_27_transpose_x_0, transpose_y = Np_27_transpose_y_0, x = Np_25_cast_fp16, y = Npa_17_cast_fp16)[name = string("Np_27_cast_fp16")];
            bool Npa_19_transpose_x_0 = const()[name = string("Npa_19_transpose_x_0"), val = bool(false)];
            bool Npa_19_transpose_y_0 = const()[name = string("Npa_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_19_cast_fp16 = matmul(transpose_x = Npa_19_transpose_x_0, transpose_y = Npa_19_transpose_y_0, x = Npa_17_cast_fp16, y = Np_25_cast_fp16)[name = string("Npa_19_cast_fp16")];
            bool var_2568_transpose_x_0 = const()[name = string("op_2568_transpose_x_0"), val = bool(false)];
            bool var_2568_transpose_y_0 = const()[name = string("op_2568_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2568_cast_fp16 = matmul(transpose_x = var_2568_transpose_x_0, transpose_y = var_2568_transpose_y_0, x = Minv_35_cast_fp16, y = Np_27_cast_fp16)[name = string("op_2568_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_37_cast_fp16 = add(x = Minv_35_cast_fp16, y = var_2568_cast_fp16)[name = string("Minv_37_cast_fp16")];
            bool Np_29_transpose_x_0 = const()[name = string("Np_29_transpose_x_0"), val = bool(false)];
            bool Np_29_transpose_y_0 = const()[name = string("Np_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_29_cast_fp16 = matmul(transpose_x = Np_29_transpose_x_0, transpose_y = Np_29_transpose_y_0, x = Np_27_cast_fp16, y = Npa_19_cast_fp16)[name = string("Np_29_cast_fp16")];
            bool var_2572_transpose_x_0 = const()[name = string("op_2572_transpose_x_0"), val = bool(false)];
            bool var_2572_transpose_y_0 = const()[name = string("op_2572_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2572_cast_fp16 = matmul(transpose_x = var_2572_transpose_x_0, transpose_y = var_2572_transpose_y_0, x = Minv_37_cast_fp16, y = Np_29_cast_fp16)[name = string("op_2572_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_39_cast_fp16 = add(x = Minv_37_cast_fp16, y = var_2572_cast_fp16)[name = string("Minv_39_cast_fp16")];
            bool u_9_transpose_x_0 = const()[name = string("u_9_transpose_x_0"), val = bool(false)];
            bool u_9_transpose_y_0 = const()[name = string("u_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_9_cast_fp16 = matmul(transpose_x = u_9_transpose_x_0, transpose_y = u_9_transpose_y_0, x = Minv_39_cast_fp16, y = vb_9_cast_fp16)[name = string("u_9_cast_fp16")];
            bool var_2579_transpose_x_1 = const()[name = string("op_2579_transpose_x_1"), val = bool(false)];
            bool var_2579_transpose_y_1 = const()[name = string("op_2579_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_2579_cast_fp16 = matmul(transpose_x = var_2579_transpose_x_1, transpose_y = var_2579_transpose_y_1, x = q_41_cast_fp16, y = k_43_cast_fp16)[name = string("op_2579_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_9_cast_fp16 = mul(x = var_2579_cast_fp16, y = D_9_cast_fp16)[name = string("A_9_cast_fp16")];
            bool o_27_transpose_x_0 = const()[name = string("o_27_transpose_x_0"), val = bool(false)];
            bool o_27_transpose_y_0 = const()[name = string("o_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> o_27_cast_fp16 = matmul(transpose_x = o_27_transpose_x_0, transpose_y = o_27_transpose_y_0, x = A_9_cast_fp16, y = u_9_cast_fp16)[name = string("o_27_cast_fp16")];
            tensor<int32, [4]> var_2586 = const()[name = string("op_2586"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2592 = const()[name = string("op_2592"), val = tensor<int32, [4]>([1, 16, 4, 128])];
            tensor<fp16, [1, 16, 4, 128]> input_81_cast_fp16 = reshape(shape = var_2592, x = z_9_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> o_29_cast_fp16 = transpose(perm = var_2586, x = o_27_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 16, 4, 128]> var_2594_cast_fp16 = mul(x = o_29_cast_fp16, y = o_29_cast_fp16)[name = string("op_2594_cast_fp16")];
            tensor<int32, [1]> var_2599_axes_0 = const()[name = string("op_2599_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2599_keep_dims_0 = const()[name = string("op_2599_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 4, 1]> var_2599_cast_fp16 = reduce_mean(axes = var_2599_axes_0, keep_dims = var_2599_keep_dims_0, x = var_2594_cast_fp16)[name = string("op_2599_cast_fp16")];
            fp16 var_2601_to_fp16 = const()[name = string("op_2601_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 16, 4, 1]> var_2602_cast_fp16 = add(x = var_2599_cast_fp16, y = var_2601_to_fp16)[name = string("op_2602_cast_fp16")];
            fp32 var_2603_epsilon_0 = const()[name = string("op_2603_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 4, 1]> var_2603_cast_fp16 = rsqrt(epsilon = var_2603_epsilon_0, x = var_2602_cast_fp16)[name = string("op_2603_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> var_2604_cast_fp16 = mul(x = o_29_cast_fp16, y = var_2603_cast_fp16)[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> var_2606_cast_fp16 = silu(x = input_81_cast_fp16)[name = string("op_2606_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> o_31_cast_fp16 = mul(x = var_2604_cast_fp16, y = var_2606_cast_fp16)[name = string("o_31_cast_fp16")];
            tensor<int32, [3]> var_2611 = const()[name = string("op_2611"), val = tensor<int32, [3]>([1, 16, 512])];
            tensor<fp16, [1, 16, 512]> input_83_cast_fp16 = reshape(shape = var_2611, x = o_31_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [640, 512]> layers_5_out_weight_to_fp16 = const()[name = string("layers_5_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84536512)))];
            tensor<fp16, [1, 16, 640]> linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_out_weight_to_fp16, x = input_83_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [1, 16, 640]> x_57_cast_fp16 = add(x = x_53_cast_fp16, y = linear_21_cast_fp16)[name = string("x_57_cast_fp16")];
            fp16 var_2624_to_fp16 = const()[name = string("op_2624_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 16, 640]> x_59_cast_fp16 = mul(x = x_57_cast_fp16, y = var_2624_to_fp16)[name = string("x_59_cast_fp16")];
            fp16 var_2622_promoted_to_fp16 = const()[name = string("op_2622_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> var_2626_cast_fp16 = pow(x = x_59_cast_fp16, y = var_2622_promoted_to_fp16)[name = string("op_2626_cast_fp16")];
            tensor<int32, [1]> var_2628_axes_0 = const()[name = string("op_2628_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2628_keep_dims_0 = const()[name = string("op_2628_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2628_cast_fp16 = reduce_mean(axes = var_2628_axes_0, keep_dims = var_2628_keep_dims_0, x = var_2626_cast_fp16)[name = string("op_2628_cast_fp16")];
            fp16 var_2629_to_fp16 = const()[name = string("op_2629_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 16, 1]> var_2630_cast_fp16 = add(x = var_2628_cast_fp16, y = var_2629_to_fp16)[name = string("op_2630_cast_fp16")];
            fp32 var_2631_epsilon_0 = const()[name = string("op_2631_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2631_cast_fp16 = rsqrt(epsilon = var_2631_epsilon_0, x = var_2630_cast_fp16)[name = string("op_2631_cast_fp16")];
            tensor<fp16, [1, 16, 640]> var_2632_cast_fp16 = mul(x = x_59_cast_fp16, y = var_2631_cast_fp16)[name = string("op_2632_cast_fp16")];
            tensor<fp16, [640]> layers_5_n2_w1_to_fp16 = const()[name = string("layers_5_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85191936)))];
            tensor<fp16, [1, 16, 640]> input_85_cast_fp16 = mul(x = var_2632_cast_fp16, y = layers_5_n2_w1_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_5_gate_up_weight_to_fp16 = const()[name = string("layers_5_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85193280)))];
            tensor<fp16, [1, 16, 5120]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_5_gate_up_weight_to_fp16, x = input_85_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [3]> input_87_begin_0 = const()[name = string("input_87_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_87_end_0 = const()[name = string("input_87_end_0"), val = tensor<int32, [3]>([1, 16, 2560])];
            tensor<bool, [3]> input_87_end_mask_0 = const()[name = string("input_87_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 2560]> input_87_cast_fp16 = slice_by_index(begin = input_87_begin_0, end = input_87_end_0, end_mask = input_87_end_mask_0, x = linear_22_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [1, 16, 2560]> var_2642_cast_fp16 = silu(x = input_87_cast_fp16)[name = string("op_2642_cast_fp16")];
            tensor<int32, [3]> var_2647_begin_0 = const()[name = string("op_2647_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_2647_end_0 = const()[name = string("op_2647_end_0"), val = tensor<int32, [3]>([1, 16, 5120])];
            tensor<bool, [3]> var_2647_end_mask_0 = const()[name = string("op_2647_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 2560]> var_2647_cast_fp16 = slice_by_index(begin = var_2647_begin_0, end = var_2647_end_0, end_mask = var_2647_end_mask_0, x = linear_22_cast_fp16)[name = string("op_2647_cast_fp16")];
            tensor<fp16, [1, 16, 2560]> input_89_cast_fp16 = mul(x = var_2642_cast_fp16, y = var_2647_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_5_down_weight_to_fp16 = const()[name = string("layers_5_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91746944)))];
            tensor<fp16, [1, 16, 640]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_down_weight_to_fp16, x = input_89_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<fp16, [1, 16, 640]> x_61_cast_fp16 = add(x = x_57_cast_fp16, y = linear_23_cast_fp16)[name = string("x_61_cast_fp16")];
            fp16 var_2660_to_fp16 = const()[name = string("op_2660_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [1, 16, 640]> x_63_cast_fp16 = mul(x = x_61_cast_fp16, y = var_2660_to_fp16)[name = string("x_63_cast_fp16")];
            fp16 var_2658_promoted_to_fp16 = const()[name = string("op_2658_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> var_2662_cast_fp16 = pow(x = x_63_cast_fp16, y = var_2658_promoted_to_fp16)[name = string("op_2662_cast_fp16")];
            tensor<int32, [1]> var_2664_axes_0 = const()[name = string("op_2664_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2664_keep_dims_0 = const()[name = string("op_2664_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2664_cast_fp16 = reduce_mean(axes = var_2664_axes_0, keep_dims = var_2664_keep_dims_0, x = var_2662_cast_fp16)[name = string("op_2664_cast_fp16")];
            fp16 var_2665_to_fp16 = const()[name = string("op_2665_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 16, 1]> var_2666_cast_fp16 = add(x = var_2664_cast_fp16, y = var_2665_to_fp16)[name = string("op_2666_cast_fp16")];
            fp32 var_2667_epsilon_0 = const()[name = string("op_2667_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2667_cast_fp16 = rsqrt(epsilon = var_2667_epsilon_0, x = var_2666_cast_fp16)[name = string("op_2667_cast_fp16")];
            tensor<fp16, [1, 16, 640]> var_2668_cast_fp16 = mul(x = x_63_cast_fp16, y = var_2667_cast_fp16)[name = string("op_2668_cast_fp16")];
            tensor<fp16, [640]> layers_6_n1_w1_to_fp16 = const()[name = string("layers_6_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95023808)))];
            tensor<fp16, [1, 16, 640]> input_91_cast_fp16 = mul(x = var_2668_cast_fp16, y = layers_6_n1_w1_to_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_6_qkvzba_weight_to_fp16 = const()[name = string("layers_6_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95025152)))];
            tensor<fp16, [1, 16, 1544]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_qkvzba_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<int32, [3]> qkv_11_begin_0 = const()[name = string("qkv_11_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_11_end_0 = const()[name = string("qkv_11_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> qkv_11_end_mask_0 = const()[name = string("qkv_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 1024]> qkv_11_cast_fp16 = slice_by_index(begin = qkv_11_begin_0, end = qkv_11_end_0, end_mask = qkv_11_end_mask_0, x = linear_24_cast_fp16)[name = string("qkv_11_cast_fp16")];
            tensor<int32, [3]> z_begin_0 = const()[name = string("z_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_end_0 = const()[name = string("z_end_0"), val = tensor<int32, [3]>([1, 16, 1536])];
            tensor<bool, [3]> z_end_mask_0 = const()[name = string("z_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 512]> z_cast_fp16 = slice_by_index(begin = z_begin_0, end = z_end_0, end_mask = z_end_mask_0, x = linear_24_cast_fp16)[name = string("z_cast_fp16")];
            tensor<int32, [3]> b_begin_0 = const()[name = string("b_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_end_0 = const()[name = string("b_end_0"), val = tensor<int32, [3]>([1, 16, 1540])];
            tensor<bool, [3]> b_end_mask_0 = const()[name = string("b_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 4]> b_cast_fp16 = slice_by_index(begin = b_begin_0, end = b_end_0, end_mask = b_end_mask_0, x = linear_24_cast_fp16)[name = string("b_cast_fp16")];
            tensor<int32, [3]> a_begin_0 = const()[name = string("a_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_end_0 = const()[name = string("a_end_0"), val = tensor<int32, [3]>([1, 16, 1])];
            tensor<bool, [3]> a_end_mask_0 = const()[name = string("a_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 4]> a_cast_fp16 = slice_by_index(begin = a_begin_0, end = a_end_0, end_mask = a_end_mask_0, x = linear_24_cast_fp16)[name = string("a_cast_fp16")];
            int32 var_2707 = const()[name = string("op_2707"), val = int32(1)];
            bool xp_interleave_0 = const()[name = string("xp_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 19, 1024]> xp_cast_fp16 = concat(axis = var_2707, interleave = xp_interleave_0, values = (zpad_to_fp16, qkv_11_cast_fp16))[name = string("xp_cast_fp16")];
            tensor<int32, [3]> var_2718_begin_0 = const()[name = string("op_2718_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_2718_end_0 = const()[name = string("op_2718_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_2718_end_mask_0 = const()[name = string("op_2718_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 1024]> var_2718_cast_fp16 = slice_by_index(begin = var_2718_begin_0, end = var_2718_end_0, end_mask = var_2718_end_mask_0, x = xp_cast_fp16)[name = string("op_2718_cast_fp16")];
            tensor<fp16, [1024]> var_2726_to_fp16 = const()[name = string("op_2726_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97001536)))];
            tensor<fp16, [1, 16, 1024]> var_2727_cast_fp16 = mul(x = var_2718_cast_fp16, y = var_2726_to_fp16)[name = string("op_2727_cast_fp16")];
            tensor<int32, [3]> var_2737_begin_0 = const()[name = string("op_2737_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_2737_end_0 = const()[name = string("op_2737_end_0"), val = tensor<int32, [3]>([1, 18, 1024])];
            tensor<bool, [3]> var_2737_end_mask_0 = const()[name = string("op_2737_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_2737_cast_fp16 = slice_by_index(begin = var_2737_begin_0, end = var_2737_end_0, end_mask = var_2737_end_mask_0, x = xp_cast_fp16)[name = string("op_2737_cast_fp16")];
            tensor<fp16, [1024]> var_2745_to_fp16 = const()[name = string("op_2745_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97003648)))];
            tensor<fp16, [1, 16, 1024]> var_2746_cast_fp16 = mul(x = var_2737_cast_fp16, y = var_2745_to_fp16)[name = string("op_2746_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2748_cast_fp16 = add(x = var_2727_cast_fp16, y = var_2746_cast_fp16)[name = string("op_2748_cast_fp16")];
            tensor<int32, [3]> var_2758_begin_0 = const()[name = string("op_2758_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2758_end_0 = const()[name = string("op_2758_end_0"), val = tensor<int32, [3]>([1, 17, 1024])];
            tensor<bool, [3]> var_2758_end_mask_0 = const()[name = string("op_2758_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_2758_cast_fp16 = slice_by_index(begin = var_2758_begin_0, end = var_2758_end_0, end_mask = var_2758_end_mask_0, x = xp_cast_fp16)[name = string("op_2758_cast_fp16")];
            tensor<fp16, [1024]> var_2766_to_fp16 = const()[name = string("op_2766_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97005760)))];
            tensor<fp16, [1, 16, 1024]> var_2767_cast_fp16 = mul(x = var_2758_cast_fp16, y = var_2766_to_fp16)[name = string("op_2767_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2769_cast_fp16 = add(x = var_2748_cast_fp16, y = var_2767_cast_fp16)[name = string("op_2769_cast_fp16")];
            tensor<int32, [3]> var_2779_begin_0 = const()[name = string("op_2779_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2779_end_0 = const()[name = string("op_2779_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> var_2779_end_mask_0 = const()[name = string("op_2779_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 16, 1024]> var_2779_cast_fp16 = slice_by_index(begin = var_2779_begin_0, end = var_2779_end_0, end_mask = var_2779_end_mask_0, x = xp_cast_fp16)[name = string("op_2779_cast_fp16")];
            tensor<fp16, [1024]> var_2787_to_fp16 = const()[name = string("op_2787_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97007872)))];
            tensor<fp16, [1, 16, 1024]> var_2788_cast_fp16 = mul(x = var_2779_cast_fp16, y = var_2787_to_fp16)[name = string("op_2788_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> input_93_cast_fp16 = add(x = var_2769_cast_fp16, y = var_2788_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> c_cast_fp16 = silu(x = input_93_cast_fp16)[name = string("c_cast_fp16")];
            tensor<int32, [3]> var_2796_begin_0 = const()[name = string("op_2796_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2796_end_0 = const()[name = string("op_2796_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
            tensor<bool, [3]> var_2796_end_mask_0 = const()[name = string("op_2796_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 256]> var_2796_cast_fp16 = slice_by_index(begin = var_2796_begin_0, end = var_2796_end_0, end_mask = var_2796_end_mask_0, x = c_cast_fp16)[name = string("op_2796_cast_fp16")];
            tensor<int32, [4]> var_2801 = const()[name = string("op_2801"), val = tensor<int32, [4]>([1, 16, 2, 128])];
            tensor<fp16, [1, 16, 2, 128]> q_43_cast_fp16 = reshape(shape = var_2801, x = var_2796_cast_fp16)[name = string("q_43_cast_fp16")];
            tensor<int32, [3]> var_2807_begin_0 = const()[name = string("op_2807_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_2807_end_0 = const()[name = string("op_2807_end_0"), val = tensor<int32, [3]>([1, 16, 512])];
            tensor<bool, [3]> var_2807_end_mask_0 = const()[name = string("op_2807_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 256]> var_2807_cast_fp16 = slice_by_index(begin = var_2807_begin_0, end = var_2807_end_0, end_mask = var_2807_end_mask_0, x = c_cast_fp16)[name = string("op_2807_cast_fp16")];
            tensor<int32, [4]> var_2812 = const()[name = string("op_2812"), val = tensor<int32, [4]>([1, 16, 2, 128])];
            tensor<fp16, [1, 16, 2, 128]> k_45_cast_fp16 = reshape(shape = var_2812, x = var_2807_cast_fp16)[name = string("k_45_cast_fp16")];
            tensor<int32, [3]> var_2818_begin_0 = const()[name = string("op_2818_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_2818_end_0 = const()[name = string("op_2818_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> var_2818_end_mask_0 = const()[name = string("op_2818_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 512]> var_2818_cast_fp16 = slice_by_index(begin = var_2818_begin_0, end = var_2818_end_0, end_mask = var_2818_end_mask_0, x = c_cast_fp16)[name = string("op_2818_cast_fp16")];
            tensor<int32, [4]> var_2823 = const()[name = string("op_2823"), val = tensor<int32, [4]>([1, 16, 4, 128])];
            tensor<fp16, [1, 16, 4, 128]> var_2824_cast_fp16 = reshape(shape = var_2823, x = var_2818_cast_fp16)[name = string("op_2824_cast_fp16")];
            tensor<int32, [4]> var_2829 = const()[name = string("op_2829"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2831_to_fp16 = const()[name = string("op_2831_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 16, 2, 128]> q_45_cast_fp16 = mul(x = q_43_cast_fp16, y = var_2831_to_fp16)[name = string("q_45_cast_fp16")];
            fp16 var_2833_to_fp16 = const()[name = string("op_2833_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 16, 2, 128]> k_47_cast_fp16 = mul(x = k_45_cast_fp16, y = var_2833_to_fp16)[name = string("k_47_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_2835_cast_fp16 = mul(x = q_45_cast_fp16, y = q_45_cast_fp16)[name = string("op_2835_cast_fp16")];
            tensor<int32, [1]> var_2840_axes_0 = const()[name = string("op_2840_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2840_keep_dims_0 = const()[name = string("op_2840_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 2, 1]> var_2840_cast_fp16 = reduce_mean(axes = var_2840_axes_0, keep_dims = var_2840_keep_dims_0, x = var_2835_cast_fp16)[name = string("op_2840_cast_fp16")];
            fp16 var_2841_promoted_to_fp16 = const()[name = string("op_2841_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 16, 2, 1]> var_2842_cast_fp16 = mul(x = var_2840_cast_fp16, y = var_2841_promoted_to_fp16)[name = string("op_2842_cast_fp16")];
            fp16 var_2844_to_fp16 = const()[name = string("op_2844_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 16, 2, 1]> var_2845_cast_fp16 = add(x = var_2842_cast_fp16, y = var_2844_to_fp16)[name = string("op_2845_cast_fp16")];
            fp32 var_2846_epsilon_0 = const()[name = string("op_2846_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 2, 1]> var_2846_cast_fp16 = rsqrt(epsilon = var_2846_epsilon_0, x = var_2845_cast_fp16)[name = string("op_2846_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_2847_cast_fp16 = mul(x = q_45_cast_fp16, y = var_2846_cast_fp16)[name = string("op_2847_cast_fp16")];
            fp16 var_2848_to_fp16 = const()[name = string("op_2848_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 16, 2, 128]> q_47_cast_fp16 = mul(x = var_2847_cast_fp16, y = var_2848_to_fp16)[name = string("q_47_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> var_2850_cast_fp16 = mul(x = k_47_cast_fp16, y = k_47_cast_fp16)[name = string("op_2850_cast_fp16")];
            tensor<int32, [1]> var_2855_axes_0 = const()[name = string("op_2855_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2855_keep_dims_0 = const()[name = string("op_2855_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 2, 1]> var_2855_cast_fp16 = reduce_mean(axes = var_2855_axes_0, keep_dims = var_2855_keep_dims_0, x = var_2850_cast_fp16)[name = string("op_2855_cast_fp16")];
            fp16 var_2856_promoted_to_fp16 = const()[name = string("op_2856_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 16, 2, 1]> var_2857_cast_fp16 = mul(x = var_2855_cast_fp16, y = var_2856_promoted_to_fp16)[name = string("op_2857_cast_fp16")];
            fp16 var_2859_to_fp16 = const()[name = string("op_2859_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 16, 2, 1]> var_2860_cast_fp16 = add(x = var_2857_cast_fp16, y = var_2859_to_fp16)[name = string("op_2860_cast_fp16")];
            fp32 var_2861_epsilon_0 = const()[name = string("op_2861_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 2, 1]> var_2861_cast_fp16 = rsqrt(epsilon = var_2861_epsilon_0, x = var_2860_cast_fp16)[name = string("op_2861_cast_fp16")];
            tensor<fp16, [1, 16, 2, 128]> k_49_cast_fp16 = mul(x = k_47_cast_fp16, y = var_2861_cast_fp16)[name = string("k_49_cast_fp16")];
            tensor<int32, [4]> var_2867 = const()[name = string("op_2867"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2880_axes_0 = const()[name = string("op_2880_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 16, 128]> var_2868_cast_fp16 = transpose(perm = var_2867, x = q_47_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 2, 1, 16, 128]> var_2880_cast_fp16 = expand_dims(axes = var_2880_axes_0, x = var_2868_cast_fp16)[name = string("op_2880_cast_fp16")];
            tensor<int32, [5]> var_2888_reps_0 = const()[name = string("op_2888_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 16, 128]> var_2888_cast_fp16 = tile(reps = var_2888_reps_0, x = var_2880_cast_fp16)[name = string("op_2888_cast_fp16")];
            tensor<int32, [4]> var_2893 = const()[name = string("op_2893"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<fp16, [1, 4, 16, 128]> q_49_cast_fp16 = reshape(shape = var_2893, x = var_2888_cast_fp16)[name = string("q_49_cast_fp16")];
            tensor<int32, [4]> var_2899 = const()[name = string("op_2899"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2912_axes_0 = const()[name = string("op_2912_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 16, 128]> var_2900_cast_fp16 = transpose(perm = var_2899, x = k_49_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 2, 1, 16, 128]> var_2912_cast_fp16 = expand_dims(axes = var_2912_axes_0, x = var_2900_cast_fp16)[name = string("op_2912_cast_fp16")];
            tensor<int32, [5]> var_2920_reps_0 = const()[name = string("op_2920_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 16, 128]> var_2920_cast_fp16 = tile(reps = var_2920_reps_0, x = var_2912_cast_fp16)[name = string("op_2920_cast_fp16")];
            tensor<int32, [4]> var_2925 = const()[name = string("op_2925"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<fp16, [1, 4, 16, 128]> k_51_cast_fp16 = reshape(shape = var_2925, x = var_2920_cast_fp16)[name = string("k_51_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_2927_cast_fp16 = sigmoid(x = b_cast_fp16)[name = string("op_2927_cast_fp16")];
            tensor<int32, [3]> var_2931 = const()[name = string("op_2931"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_6_dt_bias_to_fp16 = const()[name = string("layers_6_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 16, 4]> input_95_cast_fp16 = add(x = a_cast_fp16, y = layers_6_dt_bias_to_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_2935_cast_fp16 = relu(x = input_95_cast_fp16)[name = string("op_2935_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_2936_cast_fp16 = abs(x = input_95_cast_fp16)[name = string("op_2936_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 4]> var_2937_cast_fp16 = mul(x = var_2936_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_2937_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_2938_cast_fp16 = exp(x = var_2937_cast_fp16)[name = string("op_2938_cast_fp16")];
            fp16 var_2940_to_fp16 = const()[name = string("op_2940_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 16, 4]> var_2941_cast_fp16 = add(x = var_2938_cast_fp16, y = var_2940_to_fp16)[name = string("op_2941_cast_fp16")];
            fp32 var_2942_epsilon_0 = const()[name = string("op_2942_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 16, 4]> var_2942_cast_fp16 = log(epsilon = var_2942_epsilon_0, x = var_2941_cast_fp16)[name = string("op_2942_cast_fp16")];
            tensor<fp16, [1, 16, 4]> var_2944_cast_fp16 = add(x = var_2935_cast_fp16, y = var_2942_cast_fp16)[name = string("op_2944_cast_fp16")];
            tensor<fp16, [4]> layers_6_negA_to_fp16 = const()[name = string("layers_6_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.07cp+1, -0x1.eccp+3, -0x1.63p+3, -0x1.728p+1])];
            tensor<fp16, [1, 16, 4]> var_2945_cast_fp16 = mul(x = layers_6_negA_to_fp16, y = var_2944_cast_fp16)[name = string("op_2945_cast_fp16")];
            tensor<int32, [3]> var_2949 = const()[name = string("op_2949"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_2952_axes_0 = const()[name = string("op_2952_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16]> beta_cast_fp16 = transpose(perm = var_2931, x = var_2927_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 4, 16, 1]> var_2952_cast_fp16 = expand_dims(axes = var_2952_axes_0, x = beta_cast_fp16)[name = string("op_2952_cast_fp16")];
            fp16 var_2953_to_fp16 = const()[name = string("op_2953_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 16, 1]> var_2954_cast_fp16 = mul(x = var_2952_cast_fp16, y = var_2953_to_fp16)[name = string("op_2954_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> v_15_cast_fp16 = transpose(perm = var_2829, x = var_2824_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 4, 16, 128]> vb_cast_fp16 = mul(x = v_15_cast_fp16, y = var_2954_cast_fp16)[name = string("vb_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> kb_cast_fp16 = mul(x = k_51_cast_fp16, y = var_2952_cast_fp16)[name = string("kb_cast_fp16")];
            tensor<int32, [2]> var_2961 = const()[name = string("op_2961"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [1, 4, 16]> gdec_cast_fp16 = transpose(perm = var_2949, x = var_2945_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [4, 16]> var_2962_cast_fp16 = reshape(shape = var_2961, x = gdec_cast_fp16)[name = string("op_2962_cast_fp16")];
            tensor<fp16, [4, 256]> var_2963_cast_fp16 = linear(bias = var_556_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2962_cast_fp16)[name = string("op_2963_cast_fp16")];
            tensor<int32, [4]> var_2968 = const()[name = string("op_2968"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_cast_fp16 = reshape(shape = var_2968, x = var_2963_cast_fp16)[name = string("P_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2970_cast_fp16 = exp(x = P_cast_fp16)[name = string("op_2970_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_cast_fp16 = mul(x = var_2970_cast_fp16, y = tril_incl_to_fp16)[name = string("D_cast_fp16")];
            bool kk_transpose_x_1 = const()[name = string("kk_transpose_x_1"), val = bool(false)];
            bool kk_transpose_y_1 = const()[name = string("kk_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_cast_fp16 = matmul(transpose_x = kk_transpose_x_1, transpose_y = kk_transpose_y_1, x = kb_cast_fp16, y = k_51_cast_fp16)[name = string("kk_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_2976_cast_fp16 = mul(x = kk_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_2976_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2977_cast_fp16 = mul(x = var_2976_cast_fp16, y = D_cast_fp16)[name = string("op_2977_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_cast_fp16 = mul(x = var_2977_cast_fp16, y = tril_strict_to_fp16)[name = string("N_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2979_cast_fp16 = mul(x = D_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2979_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_cast_fp16 = mul(x = kk_cast_fp16, y = var_2979_cast_fp16)[name = string("Na_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_41_cast_fp16 = add(x = eye_to_fp16, y = N_cast_fp16)[name = string("Minv_41_cast_fp16")];
            bool Np_31_transpose_x_0 = const()[name = string("Np_31_transpose_x_0"), val = bool(false)];
            bool Np_31_transpose_y_0 = const()[name = string("Np_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_31_cast_fp16 = matmul(transpose_x = Np_31_transpose_x_0, transpose_y = Np_31_transpose_y_0, x = N_cast_fp16, y = Na_cast_fp16)[name = string("Np_31_cast_fp16")];
            bool Npa_21_transpose_x_0 = const()[name = string("Npa_21_transpose_x_0"), val = bool(false)];
            bool Npa_21_transpose_y_0 = const()[name = string("Npa_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_21_cast_fp16 = matmul(transpose_x = Npa_21_transpose_x_0, transpose_y = Npa_21_transpose_y_0, x = Na_cast_fp16, y = N_cast_fp16)[name = string("Npa_21_cast_fp16")];
            bool var_2985_transpose_x_0 = const()[name = string("op_2985_transpose_x_0"), val = bool(false)];
            bool var_2985_transpose_y_0 = const()[name = string("op_2985_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2985_cast_fp16 = matmul(transpose_x = var_2985_transpose_x_0, transpose_y = var_2985_transpose_y_0, x = Minv_41_cast_fp16, y = Np_31_cast_fp16)[name = string("op_2985_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_43_cast_fp16 = add(x = Minv_41_cast_fp16, y = var_2985_cast_fp16)[name = string("Minv_43_cast_fp16")];
            bool Np_33_transpose_x_0 = const()[name = string("Np_33_transpose_x_0"), val = bool(false)];
            bool Np_33_transpose_y_0 = const()[name = string("Np_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_33_cast_fp16 = matmul(transpose_x = Np_33_transpose_x_0, transpose_y = Np_33_transpose_y_0, x = Np_31_cast_fp16, y = Npa_21_cast_fp16)[name = string("Np_33_cast_fp16")];
            bool Npa_transpose_x_0 = const()[name = string("Npa_transpose_x_0"), val = bool(false)];
            bool Npa_transpose_y_0 = const()[name = string("Npa_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_cast_fp16 = matmul(transpose_x = Npa_transpose_x_0, transpose_y = Npa_transpose_y_0, x = Npa_21_cast_fp16, y = Np_31_cast_fp16)[name = string("Npa_cast_fp16")];
            bool var_2990_transpose_x_0 = const()[name = string("op_2990_transpose_x_0"), val = bool(false)];
            bool var_2990_transpose_y_0 = const()[name = string("op_2990_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2990_cast_fp16 = matmul(transpose_x = var_2990_transpose_x_0, transpose_y = var_2990_transpose_y_0, x = Minv_43_cast_fp16, y = Np_33_cast_fp16)[name = string("op_2990_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_45_cast_fp16 = add(x = Minv_43_cast_fp16, y = var_2990_cast_fp16)[name = string("Minv_45_cast_fp16")];
            bool Np_transpose_x_0 = const()[name = string("Np_transpose_x_0"), val = bool(false)];
            bool Np_transpose_y_0 = const()[name = string("Np_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_cast_fp16 = matmul(transpose_x = Np_transpose_x_0, transpose_y = Np_transpose_y_0, x = Np_33_cast_fp16, y = Npa_cast_fp16)[name = string("Np_cast_fp16")];
            bool var_2994_transpose_x_0 = const()[name = string("op_2994_transpose_x_0"), val = bool(false)];
            bool var_2994_transpose_y_0 = const()[name = string("op_2994_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2994_cast_fp16 = matmul(transpose_x = var_2994_transpose_x_0, transpose_y = var_2994_transpose_y_0, x = Minv_45_cast_fp16, y = Np_cast_fp16)[name = string("op_2994_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_cast_fp16 = add(x = Minv_45_cast_fp16, y = var_2994_cast_fp16)[name = string("Minv_cast_fp16")];
            bool u_transpose_x_0 = const()[name = string("u_transpose_x_0"), val = bool(false)];
            bool u_transpose_y_0 = const()[name = string("u_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_cast_fp16 = matmul(transpose_x = u_transpose_x_0, transpose_y = u_transpose_y_0, x = Minv_cast_fp16, y = vb_cast_fp16)[name = string("u_cast_fp16")];
            bool var_3001_transpose_x_1 = const()[name = string("op_3001_transpose_x_1"), val = bool(false)];
            bool var_3001_transpose_y_1 = const()[name = string("op_3001_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_3001_cast_fp16 = matmul(transpose_x = var_3001_transpose_x_1, transpose_y = var_3001_transpose_y_1, x = q_49_cast_fp16, y = k_51_cast_fp16)[name = string("op_3001_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_cast_fp16 = mul(x = var_3001_cast_fp16, y = D_cast_fp16)[name = string("A_cast_fp16")];
            bool o_33_transpose_x_0 = const()[name = string("o_33_transpose_x_0"), val = bool(false)];
            bool o_33_transpose_y_0 = const()[name = string("o_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> o_33_cast_fp16 = matmul(transpose_x = o_33_transpose_x_0, transpose_y = o_33_transpose_y_0, x = A_cast_fp16, y = u_cast_fp16)[name = string("o_33_cast_fp16")];
            tensor<int32, [4]> var_3008 = const()[name = string("op_3008"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3014 = const()[name = string("op_3014"), val = tensor<int32, [4]>([1, 16, 4, 128])];
            tensor<fp16, [1, 16, 4, 128]> input_97_cast_fp16 = reshape(shape = var_3014, x = z_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> o_35_cast_fp16 = transpose(perm = var_3008, x = o_33_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 16, 4, 128]> var_3016_cast_fp16 = mul(x = o_35_cast_fp16, y = o_35_cast_fp16)[name = string("op_3016_cast_fp16")];
            tensor<int32, [1]> var_3021_axes_0 = const()[name = string("op_3021_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3021_keep_dims_0 = const()[name = string("op_3021_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 4, 1]> var_3021_cast_fp16 = reduce_mean(axes = var_3021_axes_0, keep_dims = var_3021_keep_dims_0, x = var_3016_cast_fp16)[name = string("op_3021_cast_fp16")];
            fp16 var_3023_to_fp16 = const()[name = string("op_3023_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 16, 4, 1]> var_3024_cast_fp16 = add(x = var_3021_cast_fp16, y = var_3023_to_fp16)[name = string("op_3024_cast_fp16")];
            fp32 var_3025_epsilon_0 = const()[name = string("op_3025_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 4, 1]> var_3025_cast_fp16 = rsqrt(epsilon = var_3025_epsilon_0, x = var_3024_cast_fp16)[name = string("op_3025_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> var_3026_cast_fp16 = mul(x = o_35_cast_fp16, y = var_3025_cast_fp16)[name = string("op_3026_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> var_3028_cast_fp16 = silu(x = input_97_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<fp16, [1, 16, 4, 128]> o_37_cast_fp16 = mul(x = var_3026_cast_fp16, y = var_3028_cast_fp16)[name = string("o_37_cast_fp16")];
            tensor<int32, [3]> var_3033 = const()[name = string("op_3033"), val = tensor<int32, [3]>([1, 16, 512])];
            tensor<fp16, [1, 16, 512]> input_99_cast_fp16 = reshape(shape = var_3033, x = o_37_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [640, 512]> layers_6_out_weight_to_fp16 = const()[name = string("layers_6_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97009984)))];
            tensor<fp16, [1, 16, 640]> linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_out_weight_to_fp16, x = input_99_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 16, 640]> x_65_cast_fp16 = add(x = x_61_cast_fp16, y = linear_25_cast_fp16)[name = string("x_65_cast_fp16")];
            fp16 var_3046_to_fp16 = const()[name = string("op_3046_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [1, 16, 640]> x_67_cast_fp16 = mul(x = x_65_cast_fp16, y = var_3046_to_fp16)[name = string("x_67_cast_fp16")];
            fp16 var_3044_promoted_to_fp16 = const()[name = string("op_3044_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> var_3048_cast_fp16 = pow(x = x_67_cast_fp16, y = var_3044_promoted_to_fp16)[name = string("op_3048_cast_fp16")];
            tensor<int32, [1]> var_3050_axes_0 = const()[name = string("op_3050_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3050_keep_dims_0 = const()[name = string("op_3050_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3050_cast_fp16 = reduce_mean(axes = var_3050_axes_0, keep_dims = var_3050_keep_dims_0, x = var_3048_cast_fp16)[name = string("op_3050_cast_fp16")];
            fp16 var_3051_to_fp16 = const()[name = string("op_3051_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 16, 1]> var_3052_cast_fp16 = add(x = var_3050_cast_fp16, y = var_3051_to_fp16)[name = string("op_3052_cast_fp16")];
            fp32 var_3053_epsilon_0 = const()[name = string("op_3053_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3053_cast_fp16 = rsqrt(epsilon = var_3053_epsilon_0, x = var_3052_cast_fp16)[name = string("op_3053_cast_fp16")];
            tensor<fp16, [1, 16, 640]> var_3054_cast_fp16 = mul(x = x_67_cast_fp16, y = var_3053_cast_fp16)[name = string("op_3054_cast_fp16")];
            tensor<fp16, [640]> layers_6_n2_w1_to_fp16 = const()[name = string("layers_6_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97665408)))];
            tensor<fp16, [1, 16, 640]> input_101_cast_fp16 = mul(x = var_3054_cast_fp16, y = layers_6_n2_w1_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_6_gate_up_weight_to_fp16 = const()[name = string("layers_6_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97666752)))];
            tensor<fp16, [1, 16, 5120]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_6_gate_up_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<int32, [3]> input_103_begin_0 = const()[name = string("input_103_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_103_end_0 = const()[name = string("input_103_end_0"), val = tensor<int32, [3]>([1, 16, 2560])];
            tensor<bool, [3]> input_103_end_mask_0 = const()[name = string("input_103_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 2560]> input_103_cast_fp16 = slice_by_index(begin = input_103_begin_0, end = input_103_end_0, end_mask = input_103_end_mask_0, x = linear_26_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [1, 16, 2560]> var_3064_cast_fp16 = silu(x = input_103_cast_fp16)[name = string("op_3064_cast_fp16")];
            tensor<int32, [3]> var_3069_begin_0 = const()[name = string("op_3069_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_3069_end_0 = const()[name = string("op_3069_end_0"), val = tensor<int32, [3]>([1, 16, 5120])];
            tensor<bool, [3]> var_3069_end_mask_0 = const()[name = string("op_3069_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 2560]> var_3069_cast_fp16 = slice_by_index(begin = var_3069_begin_0, end = var_3069_end_0, end_mask = var_3069_end_mask_0, x = linear_26_cast_fp16)[name = string("op_3069_cast_fp16")];
            tensor<fp16, [1, 16, 2560]> input_105_cast_fp16 = mul(x = var_3064_cast_fp16, y = var_3069_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_6_down_weight_to_fp16 = const()[name = string("layers_6_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104220416)))];
            tensor<fp16, [1, 16, 640]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_down_weight_to_fp16, x = input_105_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 16, 640]> x_69_cast_fp16 = add(x = x_65_cast_fp16, y = linear_27_cast_fp16)[name = string("x_69_cast_fp16")];
            fp16 var_3082_to_fp16 = const()[name = string("op_3082_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [1, 16, 640]> x_71_cast_fp16 = mul(x = x_69_cast_fp16, y = var_3082_to_fp16)[name = string("x_71_cast_fp16")];
            fp16 var_3080_promoted_to_fp16 = const()[name = string("op_3080_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> var_3084_cast_fp16 = pow(x = x_71_cast_fp16, y = var_3080_promoted_to_fp16)[name = string("op_3084_cast_fp16")];
            tensor<int32, [1]> var_3086_axes_0 = const()[name = string("op_3086_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3086_keep_dims_0 = const()[name = string("op_3086_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3086_cast_fp16 = reduce_mean(axes = var_3086_axes_0, keep_dims = var_3086_keep_dims_0, x = var_3084_cast_fp16)[name = string("op_3086_cast_fp16")];
            fp16 var_3087_to_fp16 = const()[name = string("op_3087_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 16, 1]> var_3088_cast_fp16 = add(x = var_3086_cast_fp16, y = var_3087_to_fp16)[name = string("op_3088_cast_fp16")];
            fp32 var_3089_epsilon_0 = const()[name = string("op_3089_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3089_cast_fp16 = rsqrt(epsilon = var_3089_epsilon_0, x = var_3088_cast_fp16)[name = string("op_3089_cast_fp16")];
            tensor<fp16, [1, 16, 640]> var_3090_cast_fp16 = mul(x = x_71_cast_fp16, y = var_3089_cast_fp16)[name = string("op_3090_cast_fp16")];
            tensor<fp16, [640]> layers_7_n1_w1_to_fp16 = const()[name = string("layers_7_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107497280)))];
            tensor<fp16, [1, 16, 640]> input_107_cast_fp16 = mul(x = var_3090_cast_fp16, y = layers_7_n1_w1_to_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [1280, 640]> layers_7_qkv_weight_to_fp16 = const()[name = string("layers_7_qkv_weight_to_fp16"), val = tensor<fp16, [1280, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107498624)))];
            tensor<fp16, [1, 16, 1280]> linear_28_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = layers_7_qkv_weight_to_fp16, x = input_107_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [3]> var_3099_begin_0 = const()[name = string("op_3099_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3099_end_0 = const()[name = string("op_3099_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> var_3099_end_mask_0 = const()[name = string("op_3099_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 1024]> var_3099_cast_fp16 = slice_by_index(begin = var_3099_begin_0, end = var_3099_end_0, end_mask = var_3099_end_mask_0, x = linear_28_cast_fp16)[name = string("op_3099_cast_fp16")];
            tensor<int32, [4]> var_3104 = const()[name = string("op_3104"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> qg_cast_fp16 = reshape(shape = var_3104, x = var_3099_cast_fp16)[name = string("qg_cast_fp16")];
            tensor<int32, [4]> x_73_begin_0 = const()[name = string("x_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_73_end_0 = const()[name = string("x_73_end_0"), val = tensor<int32, [4]>([1, 16, 8, 64])];
            tensor<bool, [4]> x_73_end_mask_0 = const()[name = string("x_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 8, 64]> x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, x = qg_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<int32, [4]> var_3115_begin_0 = const()[name = string("op_3115_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> var_3115_end_0 = const()[name = string("op_3115_end_0"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<bool, [4]> var_3115_end_mask_0 = const()[name = string("op_3115_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 8, 64]> var_3115_cast_fp16 = slice_by_index(begin = var_3115_begin_0, end = var_3115_end_0, end_mask = var_3115_end_mask_0, x = qg_cast_fp16)[name = string("op_3115_cast_fp16")];
            tensor<int32, [3]> var_3119 = const()[name = string("op_3119"), val = tensor<int32, [3]>([1, 16, 512])];
            tensor<fp16, [1, 16, 512]> gate_cast_fp16 = reshape(shape = var_3119, x = var_3115_cast_fp16)[name = string("gate_cast_fp16")];
            tensor<int32, [3]> var_3125_begin_0 = const()[name = string("op_3125_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> var_3125_end_0 = const()[name = string("op_3125_end_0"), val = tensor<int32, [3]>([1, 16, 1152])];
            tensor<bool, [3]> var_3125_end_mask_0 = const()[name = string("op_3125_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 128]> var_3125_cast_fp16 = slice_by_index(begin = var_3125_begin_0, end = var_3125_end_0, end_mask = var_3125_end_mask_0, x = linear_28_cast_fp16)[name = string("op_3125_cast_fp16")];
            tensor<int32, [4]> var_3130 = const()[name = string("op_3130"), val = tensor<int32, [4]>([1, 16, 2, 64])];
            tensor<fp16, [1, 16, 2, 64]> x_79_cast_fp16 = reshape(shape = var_3130, x = var_3125_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<int32, [3]> var_3136_begin_0 = const()[name = string("op_3136_begin_0"), val = tensor<int32, [3]>([0, 0, 1152])];
            tensor<int32, [3]> var_3136_end_0 = const()[name = string("op_3136_end_0"), val = tensor<int32, [3]>([1, 16, 1280])];
            tensor<bool, [3]> var_3136_end_mask_0 = const()[name = string("op_3136_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 128]> var_3136_cast_fp16 = slice_by_index(begin = var_3136_begin_0, end = var_3136_end_0, end_mask = var_3136_end_mask_0, x = linear_28_cast_fp16)[name = string("op_3136_cast_fp16")];
            tensor<int32, [4]> var_3141 = const()[name = string("op_3141"), val = tensor<int32, [4]>([1, 16, 2, 64])];
            tensor<fp16, [1, 16, 2, 64]> var_3142_cast_fp16 = reshape(shape = var_3141, x = var_3136_cast_fp16)[name = string("op_3142_cast_fp16")];
            tensor<int32, [4]> var_3147 = const()[name = string("op_3147"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3153_promoted_to_fp16 = const()[name = string("op_3153_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 8, 64]> var_3157_cast_fp16 = pow(x = x_73_cast_fp16, y = var_3153_promoted_to_fp16)[name = string("op_3157_cast_fp16")];
            tensor<int32, [1]> var_3159_axes_0 = const()[name = string("op_3159_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3159_keep_dims_0 = const()[name = string("op_3159_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 8, 1]> var_3159_cast_fp16 = reduce_mean(axes = var_3159_axes_0, keep_dims = var_3159_keep_dims_0, x = var_3157_cast_fp16)[name = string("op_3159_cast_fp16")];
            fp16 var_3160_to_fp16 = const()[name = string("op_3160_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 8, 1]> var_3161_cast_fp16 = add(x = var_3159_cast_fp16, y = var_3160_to_fp16)[name = string("op_3161_cast_fp16")];
            fp32 var_3162_epsilon_0 = const()[name = string("op_3162_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 8, 1]> var_3162_cast_fp16 = rsqrt(epsilon = var_3162_epsilon_0, x = var_3161_cast_fp16)[name = string("op_3162_cast_fp16")];
            tensor<fp16, [1, 16, 8, 64]> var_3163_cast_fp16 = mul(x = x_73_cast_fp16, y = var_3162_cast_fp16)[name = string("op_3163_cast_fp16")];
            tensor<fp16, [64]> layers_7_qn_w1_to_fp16 = const()[name = string("layers_7_qn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137088)))];
            tensor<fp16, [1, 16, 8, 64]> var_3164_cast_fp16 = mul(x = var_3163_cast_fp16, y = layers_7_qn_w1_to_fp16)[name = string("op_3164_cast_fp16")];
            tensor<int32, [4]> var_3169 = const()[name = string("op_3169"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_5_begin_0 = const()[name = string("r_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_5_end_0 = const()[name = string("r_5_end_0"), val = tensor<int32, [4]>([1, 8, 16, 16])];
            tensor<bool, [4]> r_5_end_mask_0 = const()[name = string("r_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> x_77_cast_fp16 = transpose(perm = var_3169, x = var_3164_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 8, 16, 16]> r_5_cast_fp16 = slice_by_index(begin = r_5_begin_0, end = r_5_end_0, end_mask = r_5_end_mask_0, x = x_77_cast_fp16)[name = string("r_5_cast_fp16")];
            tensor<int32, [4]> p_5_begin_0 = const()[name = string("p_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_5_end_0 = const()[name = string("p_5_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> p_5_end_mask_0 = const()[name = string("p_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 48]> p_5_cast_fp16 = slice_by_index(begin = p_5_begin_0, end = p_5_end_0, end_mask = p_5_end_mask_0, x = x_77_cast_fp16)[name = string("p_5_cast_fp16")];
            tensor<int32, [4]> var_3185_begin_0 = const()[name = string("op_3185_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_3185_end_0 = const()[name = string("op_3185_end_0"), val = tensor<int32, [4]>([1, 8, 16, 16])];
            tensor<bool, [4]> var_3185_end_mask_0 = const()[name = string("op_3185_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 8]> var_3185_cast_fp16 = slice_by_index(begin = var_3185_begin_0, end = var_3185_end_0, end_mask = var_3185_end_mask_0, x = r_5_cast_fp16)[name = string("op_3185_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 16, 8]> var_3186_cast_fp16 = mul(x = var_3185_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_3186_cast_fp16")];
            tensor<int32, [4]> var_3191_begin_0 = const()[name = string("op_3191_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3191_end_0 = const()[name = string("op_3191_end_0"), val = tensor<int32, [4]>([1, 8, 16, 8])];
            tensor<bool, [4]> var_3191_end_mask_0 = const()[name = string("op_3191_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 8]> var_3191_cast_fp16 = slice_by_index(begin = var_3191_begin_0, end = var_3191_end_0, end_mask = var_3191_end_mask_0, x = r_5_cast_fp16)[name = string("op_3191_cast_fp16")];
            int32 var_3193 = const()[name = string("op_3193"), val = int32(-1)];
            bool rh_5_interleave_0 = const()[name = string("rh_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 16]> rh_5_cast_fp16 = concat(axis = var_3193, interleave = rh_5_interleave_0, values = (var_3186_cast_fp16, var_3191_cast_fp16))[name = string("rh_5_cast_fp16")];
            tensor<fp16, [1, 8, 16, 16]> var_3195_cast_fp16 = mul(x = r_5_cast_fp16, y = cos_to_fp16)[name = string("op_3195_cast_fp16")];
            tensor<fp16, [1, 8, 16, 16]> var_3196_cast_fp16 = mul(x = rh_5_cast_fp16, y = sin_to_fp16)[name = string("op_3196_cast_fp16")];
            tensor<fp16, [1, 8, 16, 16]> var_3198_cast_fp16 = add(x = var_3195_cast_fp16, y = var_3196_cast_fp16)[name = string("op_3198_cast_fp16")];
            int32 var_3200 = const()[name = string("op_3200"), val = int32(-1)];
            bool q_interleave_0 = const()[name = string("q_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 64]> q_cast_fp16 = concat(axis = var_3200, interleave = q_interleave_0, values = (var_3198_cast_fp16, p_5_cast_fp16))[name = string("q_cast_fp16")];
            fp16 var_3206_promoted_to_fp16 = const()[name = string("op_3206_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 2, 64]> var_3210_cast_fp16 = pow(x = x_79_cast_fp16, y = var_3206_promoted_to_fp16)[name = string("op_3210_cast_fp16")];
            tensor<int32, [1]> var_3212_axes_0 = const()[name = string("op_3212_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3212_keep_dims_0 = const()[name = string("op_3212_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 2, 1]> var_3212_cast_fp16 = reduce_mean(axes = var_3212_axes_0, keep_dims = var_3212_keep_dims_0, x = var_3210_cast_fp16)[name = string("op_3212_cast_fp16")];
            fp16 var_3213_to_fp16 = const()[name = string("op_3213_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 2, 1]> var_3214_cast_fp16 = add(x = var_3212_cast_fp16, y = var_3213_to_fp16)[name = string("op_3214_cast_fp16")];
            fp32 var_3215_epsilon_0 = const()[name = string("op_3215_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 2, 1]> var_3215_cast_fp16 = rsqrt(epsilon = var_3215_epsilon_0, x = var_3214_cast_fp16)[name = string("op_3215_cast_fp16")];
            tensor<fp16, [1, 16, 2, 64]> var_3216_cast_fp16 = mul(x = x_79_cast_fp16, y = var_3215_cast_fp16)[name = string("op_3216_cast_fp16")];
            tensor<fp16, [64]> layers_7_kn_w1_to_fp16 = const()[name = string("layers_7_kn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137280)))];
            tensor<fp16, [1, 16, 2, 64]> var_3217_cast_fp16 = mul(x = var_3216_cast_fp16, y = layers_7_kn_w1_to_fp16)[name = string("op_3217_cast_fp16")];
            tensor<int32, [4]> var_3222 = const()[name = string("op_3222"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_begin_0 = const()[name = string("r_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_end_0 = const()[name = string("r_end_0"), val = tensor<int32, [4]>([1, 2, 16, 16])];
            tensor<bool, [4]> r_end_mask_0 = const()[name = string("r_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 16, 64]> x_83_cast_fp16 = transpose(perm = var_3222, x = var_3217_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 2, 16, 16]> r_cast_fp16 = slice_by_index(begin = r_begin_0, end = r_end_0, end_mask = r_end_mask_0, x = x_83_cast_fp16)[name = string("r_cast_fp16")];
            tensor<int32, [4]> p_begin_0 = const()[name = string("p_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_end_0 = const()[name = string("p_end_0"), val = tensor<int32, [4]>([1, 2, 16, 64])];
            tensor<bool, [4]> p_end_mask_0 = const()[name = string("p_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 16, 48]> p_cast_fp16 = slice_by_index(begin = p_begin_0, end = p_end_0, end_mask = p_end_mask_0, x = x_83_cast_fp16)[name = string("p_cast_fp16")];
            tensor<int32, [4]> var_3238_begin_0 = const()[name = string("op_3238_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_3238_end_0 = const()[name = string("op_3238_end_0"), val = tensor<int32, [4]>([1, 2, 16, 16])];
            tensor<bool, [4]> var_3238_end_mask_0 = const()[name = string("op_3238_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 16, 8]> var_3238_cast_fp16 = slice_by_index(begin = var_3238_begin_0, end = var_3238_end_0, end_mask = var_3238_end_mask_0, x = r_cast_fp16)[name = string("op_3238_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 16, 8]> var_3239_cast_fp16 = mul(x = var_3238_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_3239_cast_fp16")];
            tensor<int32, [4]> var_3244_begin_0 = const()[name = string("op_3244_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3244_end_0 = const()[name = string("op_3244_end_0"), val = tensor<int32, [4]>([1, 2, 16, 8])];
            tensor<bool, [4]> var_3244_end_mask_0 = const()[name = string("op_3244_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 16, 8]> var_3244_cast_fp16 = slice_by_index(begin = var_3244_begin_0, end = var_3244_end_0, end_mask = var_3244_end_mask_0, x = r_cast_fp16)[name = string("op_3244_cast_fp16")];
            int32 var_3246 = const()[name = string("op_3246"), val = int32(-1)];
            bool rh_interleave_0 = const()[name = string("rh_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 16, 16]> rh_cast_fp16 = concat(axis = var_3246, interleave = rh_interleave_0, values = (var_3239_cast_fp16, var_3244_cast_fp16))[name = string("rh_cast_fp16")];
            tensor<fp16, [1, 2, 16, 16]> var_3248_cast_fp16 = mul(x = r_cast_fp16, y = cos_to_fp16)[name = string("op_3248_cast_fp16")];
            tensor<fp16, [1, 2, 16, 16]> var_3249_cast_fp16 = mul(x = rh_cast_fp16, y = sin_to_fp16)[name = string("op_3249_cast_fp16")];
            tensor<fp16, [1, 2, 16, 16]> var_3251_cast_fp16 = add(x = var_3248_cast_fp16, y = var_3249_cast_fp16)[name = string("op_3251_cast_fp16")];
            int32 var_3253 = const()[name = string("op_3253"), val = int32(-1)];
            bool k_53_interleave_0 = const()[name = string("k_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 16, 64]> k_53_cast_fp16 = concat(axis = var_3253, interleave = k_53_interleave_0, values = (var_3251_cast_fp16, p_cast_fp16))[name = string("k_53_cast_fp16")];
            tensor<int32, [1]> var_3266_axes_0 = const()[name = string("op_3266_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 16, 64]> var_3266_cast_fp16 = expand_dims(axes = var_3266_axes_0, x = k_53_cast_fp16)[name = string("op_3266_cast_fp16")];
            tensor<int32, [5]> var_3274_reps_0 = const()[name = string("op_3274_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 2, 4, 16, 64]> var_3274_cast_fp16 = tile(reps = var_3274_reps_0, x = var_3266_cast_fp16)[name = string("op_3274_cast_fp16")];
            tensor<int32, [4]> var_3279 = const()[name = string("op_3279"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<fp16, [1, 8, 16, 64]> k_cast_fp16 = reshape(shape = var_3279, x = var_3274_cast_fp16)[name = string("k_cast_fp16")];
            tensor<int32, [1]> var_3292_axes_0 = const()[name = string("op_3292_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 16, 64]> v_17_cast_fp16 = transpose(perm = var_3147, x = var_3142_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 2, 1, 16, 64]> var_3292_cast_fp16 = expand_dims(axes = var_3292_axes_0, x = v_17_cast_fp16)[name = string("op_3292_cast_fp16")];
            tensor<int32, [5]> var_3300_reps_0 = const()[name = string("op_3300_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 2, 4, 16, 64]> var_3300_cast_fp16 = tile(reps = var_3300_reps_0, x = var_3292_cast_fp16)[name = string("op_3300_cast_fp16")];
            tensor<int32, [4]> var_3305 = const()[name = string("op_3305"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<fp16, [1, 8, 16, 64]> v_cast_fp16 = reshape(shape = var_3305, x = var_3300_cast_fp16)[name = string("v_cast_fp16")];
            bool var_3310_transpose_x_1 = const()[name = string("op_3310_transpose_x_1"), val = bool(false)];
            bool var_3310_transpose_y_1 = const()[name = string("op_3310_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 16]> var_3310_cast_fp16 = matmul(transpose_x = var_3310_transpose_x_1, transpose_y = var_3310_transpose_y_1, x = q_cast_fp16, y = k_cast_fp16)[name = string("op_3310_cast_fp16")];
            fp16 var_3311_to_fp16 = const()[name = string("op_3311_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 8, 16, 16]> var_3312_cast_fp16 = mul(x = var_3310_cast_fp16, y = var_3311_to_fp16)[name = string("op_3312_cast_fp16")];
            tensor<fp16, [1, 8, 16, 16]> s_cast_fp16 = add(x = var_3312_cast_fp16, y = causal_add_to_fp16)[name = string("s_cast_fp16")];
            int32 var_3315 = const()[name = string("op_3315"), val = int32(-1)];
            tensor<fp16, [1, 8, 16, 16]> var_3317_cast_fp16 = softmax(axis = var_3315, x = s_cast_fp16)[name = string("op_3317_cast_fp16")];
            bool o_39_transpose_x_0 = const()[name = string("o_39_transpose_x_0"), val = bool(false)];
            bool o_39_transpose_y_0 = const()[name = string("o_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 64]> o_39_cast_fp16 = matmul(transpose_x = o_39_transpose_x_0, transpose_y = o_39_transpose_y_0, x = var_3317_cast_fp16, y = v_cast_fp16)[name = string("o_39_cast_fp16")];
            tensor<int32, [4]> var_3323 = const()[name = string("op_3323"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3328 = const()[name = string("op_3328"), val = tensor<int32, [3]>([1, 16, 512])];
            tensor<fp16, [1, 16, 8, 64]> var_3324_cast_fp16 = transpose(perm = var_3323, x = o_39_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 16, 512]> var_3329_cast_fp16 = reshape(shape = var_3328, x = var_3324_cast_fp16)[name = string("op_3329_cast_fp16")];
            tensor<fp16, [1, 16, 512]> var_3330_cast_fp16 = sigmoid(x = gate_cast_fp16)[name = string("op_3330_cast_fp16")];
            tensor<fp16, [1, 16, 512]> input_109_cast_fp16 = mul(x = var_3329_cast_fp16, y = var_3330_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [640, 512]> layers_7_o_weight_to_fp16 = const()[name = string("layers_7_o_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137472)))];
            tensor<fp16, [1, 16, 640]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_o_weight_to_fp16, x = input_109_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<fp16, [1, 16, 640]> x_85_cast_fp16 = add(x = x_69_cast_fp16, y = linear_29_cast_fp16)[name = string("x_85_cast_fp16")];
            fp16 var_3343_to_fp16 = const()[name = string("op_3343_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [1, 16, 640]> x_87_cast_fp16 = mul(x = x_85_cast_fp16, y = var_3343_to_fp16)[name = string("x_87_cast_fp16")];
            fp16 var_3341_promoted_to_fp16 = const()[name = string("op_3341_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 640]> var_3345_cast_fp16 = pow(x = x_87_cast_fp16, y = var_3341_promoted_to_fp16)[name = string("op_3345_cast_fp16")];
            tensor<int32, [1]> var_3347_axes_0 = const()[name = string("op_3347_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3347_keep_dims_0 = const()[name = string("op_3347_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3347_cast_fp16 = reduce_mean(axes = var_3347_axes_0, keep_dims = var_3347_keep_dims_0, x = var_3345_cast_fp16)[name = string("op_3347_cast_fp16")];
            fp16 var_3348_to_fp16 = const()[name = string("op_3348_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 16, 1]> var_3349_cast_fp16 = add(x = var_3347_cast_fp16, y = var_3348_to_fp16)[name = string("op_3349_cast_fp16")];
            fp32 var_3350_epsilon_0 = const()[name = string("op_3350_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3350_cast_fp16 = rsqrt(epsilon = var_3350_epsilon_0, x = var_3349_cast_fp16)[name = string("op_3350_cast_fp16")];
            tensor<fp16, [1, 16, 640]> var_3351_cast_fp16 = mul(x = x_87_cast_fp16, y = var_3350_cast_fp16)[name = string("op_3351_cast_fp16")];
            tensor<fp16, [640]> layers_7_n2_w1_to_fp16 = const()[name = string("layers_7_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109792896)))];
            tensor<fp16, [1, 16, 640]> input_111_cast_fp16 = mul(x = var_3351_cast_fp16, y = layers_7_n2_w1_to_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_7_gate_up_weight_to_fp16 = const()[name = string("layers_7_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109794240)))];
            tensor<fp16, [1, 16, 5120]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_7_gate_up_weight_to_fp16, x = input_111_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [3]> input_113_begin_0 = const()[name = string("input_113_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_113_end_0 = const()[name = string("input_113_end_0"), val = tensor<int32, [3]>([1, 16, 2560])];
            tensor<bool, [3]> input_113_end_mask_0 = const()[name = string("input_113_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 16, 2560]> input_113_cast_fp16 = slice_by_index(begin = input_113_begin_0, end = input_113_end_0, end_mask = input_113_end_mask_0, x = linear_30_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [1, 16, 2560]> var_3361_cast_fp16 = silu(x = input_113_cast_fp16)[name = string("op_3361_cast_fp16")];
            tensor<int32, [3]> var_3366_begin_0 = const()[name = string("op_3366_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_3366_end_0 = const()[name = string("op_3366_end_0"), val = tensor<int32, [3]>([1, 16, 5120])];
            tensor<bool, [3]> var_3366_end_mask_0 = const()[name = string("op_3366_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 16, 2560]> var_3366_cast_fp16 = slice_by_index(begin = var_3366_begin_0, end = var_3366_end_0, end_mask = var_3366_end_mask_0, x = linear_30_cast_fp16)[name = string("op_3366_cast_fp16")];
            tensor<fp16, [1, 16, 2560]> input_cast_fp16 = mul(x = var_3361_cast_fp16, y = var_3366_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_7_down_weight_to_fp16 = const()[name = string("layers_7_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116347904)))];
            tensor<fp16, [1, 16, 640]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_down_weight_to_fp16, x = input_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 16, 640]> x_89_cast_fp16 = add(x = x_85_cast_fp16, y = linear_31_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<int32, [1]> var_3384_axes_0 = const()[name = string("op_3384_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_3384 = expand_dims(axes = var_3384_axes_0, x = last)[name = string("op_3384")];
            tensor<bool, [1, 16]> var_3385 = equal(x = pos, y = var_3384)[name = string("op_3385")];
            tensor<int32, [1]> var_3402_axes_0 = const()[name = string("op_3402_axes_0"), val = tensor<int32, [1]>([2])];
            string sel_to_fp16_dtype_0 = const()[name = string("sel_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 16]> var_3385_to_fp16 = cast(dtype = sel_to_fp16_dtype_0, x = var_3385)[name = string("cast_71")];
            tensor<fp16, [1, 16, 1]> var_3402_cast_fp16 = expand_dims(axes = var_3402_axes_0, x = var_3385_to_fp16)[name = string("op_3402_cast_fp16")];
            tensor<fp16, [1, 16, 640]> var_3403_cast_fp16 = mul(x = var_3402_cast_fp16, y = x_89_cast_fp16)[name = string("op_3403_cast_fp16")];
            tensor<int32, [1]> x_91_axes_0 = const()[name = string("x_91_axes_0"), val = tensor<int32, [1]>([1])];
            bool x_91_keep_dims_0 = const()[name = string("x_91_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 640]> x_91_cast_fp16 = reduce_sum(axes = x_91_axes_0, keep_dims = x_91_keep_dims_0, x = var_3403_cast_fp16)[name = string("x_91_cast_fp16")];
            fp16 var_3415_to_fp16 = const()[name = string("op_3415_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 640]> x_cast_fp16 = mul(x = x_91_cast_fp16, y = var_3415_to_fp16)[name = string("x_cast_fp16")];
            fp16 var_3413_promoted_to_fp16 = const()[name = string("op_3413_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640]> var_3417_cast_fp16 = pow(x = x_cast_fp16, y = var_3413_promoted_to_fp16)[name = string("op_3417_cast_fp16")];
            tensor<int32, [1]> var_3419_axes_0 = const()[name = string("op_3419_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3419_keep_dims_0 = const()[name = string("op_3419_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_3419_cast_fp16 = reduce_mean(axes = var_3419_axes_0, keep_dims = var_3419_keep_dims_0, x = var_3417_cast_fp16)[name = string("op_3419_cast_fp16")];
            fp16 var_3420_to_fp16 = const()[name = string("op_3420_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 1]> var_3421_cast_fp16 = add(x = var_3419_cast_fp16, y = var_3420_to_fp16)[name = string("op_3421_cast_fp16")];
            fp32 var_3422_epsilon_0 = const()[name = string("op_3422_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1]> var_3422_cast_fp16 = rsqrt(epsilon = var_3422_epsilon_0, x = var_3421_cast_fp16)[name = string("op_3422_cast_fp16")];
            tensor<fp16, [1, 640]> var_3423_cast_fp16 = mul(x = x_cast_fp16, y = var_3422_cast_fp16)[name = string("op_3423_cast_fp16")];
            tensor<fp16, [640]> norm_w1_to_fp16 = const()[name = string("norm_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119624768)))];
            tensor<fp16, [1, 640]> h_cast_fp16 = mul(x = var_3423_cast_fp16, y = norm_w1_to_fp16)[name = string("h_cast_fp16")];
            tensor<fp16, [16000]> logits_bias_0_to_fp16 = const()[name = string("logits_bias_0_to_fp16"), val = tensor<fp16, [16000]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119626112)))];
            tensor<fp16, [1, 16000]> logits_cast_fp16 = linear(bias = logits_bias_0_to_fp16, weight = emb_to_fp16, x = h_cast_fp16)[name = string("logits_cast_fp16")];
            tensor<int32, [1]> reduce_max_0_axes_0 = const()[name = string("reduce_max_0_axes_0"), val = tensor<int32, [1]>([-1])];
            bool reduce_max_0_keep_dims_0 = const()[name = string("reduce_max_0_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> reduce_max_0_cast_fp16 = reduce_max(axes = reduce_max_0_axes_0, keep_dims = reduce_max_0_keep_dims_0, x = logits_cast_fp16)[name = string("reduce_max_0_cast_fp16")];
            tensor<fp16, [1, 16000]> var_3432_cast_fp16 = sub(x = logits_cast_fp16, y = reduce_max_0_cast_fp16)[name = string("op_3432_cast_fp16")];
            tensor<fp16, [1, 16000]> var_3433_cast_fp16 = exp(x = var_3432_cast_fp16)[name = string("op_3433_cast_fp16")];
            tensor<int32, [1]> var_3438_axes_0 = const()[name = string("op_3438_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3438_keep_dims_0 = const()[name = string("op_3438_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_3438_cast_fp16 = reduce_sum(axes = var_3438_axes_0, keep_dims = var_3438_keep_dims_0, x = var_3433_cast_fp16)[name = string("op_3438_cast_fp16")];
            fp32 var_3439_epsilon_0 = const()[name = string("op_3439_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 1]> var_3439_cast_fp16 = log(epsilon = var_3439_epsilon_0, x = var_3438_cast_fp16)[name = string("op_3439_cast_fp16")];
            tensor<fp16, [1, 1]> var_3441_cast_fp16 = add(x = reduce_max_0_cast_fp16, y = var_3439_cast_fp16)[name = string("op_3441_cast_fp16")];
            tensor<fp16, [1, 16000]> var_3443_cast_fp16 = sub(x = logits_cast_fp16, y = var_3441_cast_fp16)[name = string("op_3443_cast_fp16")];
            string var_3443_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_3443_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 16000]> lp = cast(dtype = var_3443_cast_fp16_to_fp32_dtype_0, x = var_3443_cast_fp16)[name = string("cast_70")];
        } -> (lp);
    func n32<ios18>(tensor<int32, [1, 32]> ids, tensor<int32, [1]> last) {
            tensor<int32, [1, 32]> pos = const()[name = string("pos"), val = tensor<int32, [1, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119706880)))];
            int32 x_1_batch_dims_0 = const()[name = string("x_1_batch_dims_0"), val = int32(0)];
            bool x_1_validate_indices_0 = const()[name = string("x_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [16000, 640]> emb_to_fp16 = const()[name = string("emb_to_fp16"), val = tensor<fp16, [16000, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            string ids_to_int16_dtype_0 = const()[name = string("ids_to_int16_dtype_0"), val = string("int16")];
            string cast_81_dtype_0 = const()[name = string("cast_81_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int16, [1, 32]> ids_to_int16 = cast(dtype = ids_to_int16_dtype_0, x = ids)[name = string("cast_86")];
            tensor<int32, [1, 32]> cast_81 = cast(dtype = cast_81_dtype_0, x = ids_to_int16)[name = string("cast_85")];
            tensor<bool, [1, 32]> greater_equal_0 = greater_equal(x = cast_81, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(16000)];
            tensor<int32, [1, 32]> add_0 = add(x = cast_81, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 32]> select_0 = select(a = cast_81, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 x_1_cast_fp16_cast_uint16_axis_0 = const()[name = string("x_1_cast_fp16_cast_uint16_axis_0"), val = int32(0)];
            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
            tensor<int16, [1, 32]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_84")];
            tensor<fp16, [1, 32, 640]> x_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = x_1_cast_fp16_cast_uint16_axis_0, batch_dims = x_1_batch_dims_0, indices = select_0_to_int16, validate_indices = x_1_validate_indices_0, x = emb_to_fp16)[name = string("x_1_cast_fp16_cast_uint16_cast_uint16")];
            fp16 var_255_to_fp16 = const()[name = string("op_255_to_fp16"), val = fp16(0x1p+5)];
            tensor<fp16, [1, 32, 640]> x_3_cast_fp16 = mul(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = var_255_to_fp16)[name = string("x_3_cast_fp16")];
            fp16 var_253_promoted_to_fp16 = const()[name = string("op_253_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> var_257_cast_fp16 = pow(x = x_3_cast_fp16, y = var_253_promoted_to_fp16)[name = string("op_257_cast_fp16")];
            tensor<int32, [1]> var_259_axes_0 = const()[name = string("op_259_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_259_keep_dims_0 = const()[name = string("op_259_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_259_cast_fp16 = reduce_mean(axes = var_259_axes_0, keep_dims = var_259_keep_dims_0, x = var_257_cast_fp16)[name = string("op_259_cast_fp16")];
            fp16 var_260_to_fp16 = const()[name = string("op_260_to_fp16"), val = fp16(0x1.0c8p-10)];
            tensor<fp16, [1, 32, 1]> var_261_cast_fp16 = add(x = var_259_cast_fp16, y = var_260_to_fp16)[name = string("op_261_cast_fp16")];
            fp32 var_262_epsilon_0 = const()[name = string("op_262_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_262_cast_fp16 = rsqrt(epsilon = var_262_epsilon_0, x = var_261_cast_fp16)[name = string("op_262_cast_fp16")];
            tensor<fp16, [1, 32, 640]> var_263_cast_fp16 = mul(x = x_3_cast_fp16, y = var_262_cast_fp16)[name = string("op_263_cast_fp16")];
            tensor<fp16, [640]> layers_0_n1_w1_to_fp16 = const()[name = string("layers_0_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20480128)))];
            tensor<fp16, [1, 32, 640]> input_1_cast_fp16 = mul(x = var_263_cast_fp16, y = layers_0_n1_w1_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_0_qkvzba_weight_to_fp16 = const()[name = string("layers_0_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20481472)))];
            tensor<fp16, [1544]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [1544]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22457856)))];
            tensor<fp16, [1, 32, 1544]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_qkvzba_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [3]> qkv_1_begin_0 = const()[name = string("qkv_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_1_end_0 = const()[name = string("qkv_1_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> qkv_1_end_mask_0 = const()[name = string("qkv_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 1024]> qkv_1_cast_fp16 = slice_by_index(begin = qkv_1_begin_0, end = qkv_1_end_0, end_mask = qkv_1_end_mask_0, x = linear_0_cast_fp16)[name = string("qkv_1_cast_fp16")];
            tensor<int32, [3]> z_1_begin_0 = const()[name = string("z_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_1_end_0 = const()[name = string("z_1_end_0"), val = tensor<int32, [3]>([1, 32, 1536])];
            tensor<bool, [3]> z_1_end_mask_0 = const()[name = string("z_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 512]> z_1_cast_fp16 = slice_by_index(begin = z_1_begin_0, end = z_1_end_0, end_mask = z_1_end_mask_0, x = linear_0_cast_fp16)[name = string("z_1_cast_fp16")];
            tensor<int32, [3]> b_1_begin_0 = const()[name = string("b_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_1_end_0 = const()[name = string("b_1_end_0"), val = tensor<int32, [3]>([1, 32, 1540])];
            tensor<bool, [3]> b_1_end_mask_0 = const()[name = string("b_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 4]> b_1_cast_fp16 = slice_by_index(begin = b_1_begin_0, end = b_1_end_0, end_mask = b_1_end_mask_0, x = linear_0_cast_fp16)[name = string("b_1_cast_fp16")];
            tensor<int32, [3]> a_1_begin_0 = const()[name = string("a_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_1_end_0 = const()[name = string("a_1_end_0"), val = tensor<int32, [3]>([1, 32, 1])];
            tensor<bool, [3]> a_1_end_mask_0 = const()[name = string("a_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 4]> a_1_cast_fp16 = slice_by_index(begin = a_1_begin_0, end = a_1_end_0, end_mask = a_1_end_mask_0, x = linear_0_cast_fp16)[name = string("a_1_cast_fp16")];
            int32 var_302 = const()[name = string("op_302"), val = int32(1)];
            bool xp_1_interleave_0 = const()[name = string("xp_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 1024]> zpad_to_fp16 = const()[name = string("zpad_to_fp16"), val = tensor<fp16, [1, 3, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119700672)))];
            tensor<fp16, [1, 35, 1024]> xp_1_cast_fp16 = concat(axis = var_302, interleave = xp_1_interleave_0, values = (zpad_to_fp16, qkv_1_cast_fp16))[name = string("xp_1_cast_fp16")];
            tensor<int32, [3]> var_313_begin_0 = const()[name = string("op_313_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_313_end_0 = const()[name = string("op_313_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_313_end_mask_0 = const()[name = string("op_313_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 1024]> var_313_cast_fp16 = slice_by_index(begin = var_313_begin_0, end = var_313_end_0, end_mask = var_313_end_mask_0, x = xp_1_cast_fp16)[name = string("op_313_cast_fp16")];
            tensor<fp16, [1024]> var_321_to_fp16 = const()[name = string("op_321_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22479552)))];
            tensor<fp16, [1, 32, 1024]> var_322_cast_fp16 = mul(x = var_313_cast_fp16, y = var_321_to_fp16)[name = string("op_322_cast_fp16")];
            tensor<int32, [3]> var_332_begin_0 = const()[name = string("op_332_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_332_end_0 = const()[name = string("op_332_end_0"), val = tensor<int32, [3]>([1, 34, 1024])];
            tensor<bool, [3]> var_332_end_mask_0 = const()[name = string("op_332_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_332_cast_fp16 = slice_by_index(begin = var_332_begin_0, end = var_332_end_0, end_mask = var_332_end_mask_0, x = xp_1_cast_fp16)[name = string("op_332_cast_fp16")];
            tensor<fp16, [1024]> var_340_to_fp16 = const()[name = string("op_340_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22481664)))];
            tensor<fp16, [1, 32, 1024]> var_341_cast_fp16 = mul(x = var_332_cast_fp16, y = var_340_to_fp16)[name = string("op_341_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> var_343_cast_fp16 = add(x = var_322_cast_fp16, y = var_341_cast_fp16)[name = string("op_343_cast_fp16")];
            tensor<int32, [3]> var_353_begin_0 = const()[name = string("op_353_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_353_end_0 = const()[name = string("op_353_end_0"), val = tensor<int32, [3]>([1, 33, 1024])];
            tensor<bool, [3]> var_353_end_mask_0 = const()[name = string("op_353_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_353_cast_fp16 = slice_by_index(begin = var_353_begin_0, end = var_353_end_0, end_mask = var_353_end_mask_0, x = xp_1_cast_fp16)[name = string("op_353_cast_fp16")];
            tensor<fp16, [1024]> var_361_to_fp16 = const()[name = string("op_361_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22483776)))];
            tensor<fp16, [1, 32, 1024]> var_362_cast_fp16 = mul(x = var_353_cast_fp16, y = var_361_to_fp16)[name = string("op_362_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> var_364_cast_fp16 = add(x = var_343_cast_fp16, y = var_362_cast_fp16)[name = string("op_364_cast_fp16")];
            tensor<int32, [3]> var_374_begin_0 = const()[name = string("op_374_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_374_end_0 = const()[name = string("op_374_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> var_374_end_mask_0 = const()[name = string("op_374_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_374_cast_fp16 = slice_by_index(begin = var_374_begin_0, end = var_374_end_0, end_mask = var_374_end_mask_0, x = xp_1_cast_fp16)[name = string("op_374_cast_fp16")];
            tensor<fp16, [1024]> var_382_to_fp16 = const()[name = string("op_382_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22485888)))];
            tensor<fp16, [1, 32, 1024]> var_383_cast_fp16 = mul(x = var_374_cast_fp16, y = var_382_to_fp16)[name = string("op_383_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> input_3_cast_fp16 = add(x = var_364_cast_fp16, y = var_383_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> c_1_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("c_1_cast_fp16")];
            tensor<int32, [3]> var_391_begin_0 = const()[name = string("op_391_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_391_end_0 = const()[name = string("op_391_end_0"), val = tensor<int32, [3]>([1, 32, 256])];
            tensor<bool, [3]> var_391_end_mask_0 = const()[name = string("op_391_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 256]> var_391_cast_fp16 = slice_by_index(begin = var_391_begin_0, end = var_391_end_0, end_mask = var_391_end_mask_0, x = c_1_cast_fp16)[name = string("op_391_cast_fp16")];
            tensor<int32, [4]> var_396 = const()[name = string("op_396"), val = tensor<int32, [4]>([1, 32, 2, 128])];
            tensor<fp16, [1, 32, 2, 128]> q_1_cast_fp16 = reshape(shape = var_396, x = var_391_cast_fp16)[name = string("q_1_cast_fp16")];
            tensor<int32, [3]> var_402_begin_0 = const()[name = string("op_402_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_402_end_0 = const()[name = string("op_402_end_0"), val = tensor<int32, [3]>([1, 32, 512])];
            tensor<bool, [3]> var_402_end_mask_0 = const()[name = string("op_402_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 256]> var_402_cast_fp16 = slice_by_index(begin = var_402_begin_0, end = var_402_end_0, end_mask = var_402_end_mask_0, x = c_1_cast_fp16)[name = string("op_402_cast_fp16")];
            tensor<int32, [4]> var_407 = const()[name = string("op_407"), val = tensor<int32, [4]>([1, 32, 2, 128])];
            tensor<fp16, [1, 32, 2, 128]> k_1_cast_fp16 = reshape(shape = var_407, x = var_402_cast_fp16)[name = string("k_1_cast_fp16")];
            tensor<int32, [3]> var_413_begin_0 = const()[name = string("op_413_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_413_end_0 = const()[name = string("op_413_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> var_413_end_mask_0 = const()[name = string("op_413_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 512]> var_413_cast_fp16 = slice_by_index(begin = var_413_begin_0, end = var_413_end_0, end_mask = var_413_end_mask_0, x = c_1_cast_fp16)[name = string("op_413_cast_fp16")];
            tensor<int32, [4]> var_418 = const()[name = string("op_418"), val = tensor<int32, [4]>([1, 32, 4, 128])];
            tensor<fp16, [1, 32, 4, 128]> var_419_cast_fp16 = reshape(shape = var_418, x = var_413_cast_fp16)[name = string("op_419_cast_fp16")];
            tensor<int32, [4]> var_424 = const()[name = string("op_424"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_426_to_fp16 = const()[name = string("op_426_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 32, 2, 128]> q_3_cast_fp16 = mul(x = q_1_cast_fp16, y = var_426_to_fp16)[name = string("q_3_cast_fp16")];
            fp16 var_428_to_fp16 = const()[name = string("op_428_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 32, 2, 128]> k_3_cast_fp16 = mul(x = k_1_cast_fp16, y = var_428_to_fp16)[name = string("k_3_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_430_cast_fp16 = mul(x = q_3_cast_fp16, y = q_3_cast_fp16)[name = string("op_430_cast_fp16")];
            tensor<int32, [1]> var_435_axes_0 = const()[name = string("op_435_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_435_keep_dims_0 = const()[name = string("op_435_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 2, 1]> var_435_cast_fp16 = reduce_mean(axes = var_435_axes_0, keep_dims = var_435_keep_dims_0, x = var_430_cast_fp16)[name = string("op_435_cast_fp16")];
            fp16 var_436_promoted_to_fp16 = const()[name = string("op_436_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 32, 2, 1]> var_437_cast_fp16 = mul(x = var_435_cast_fp16, y = var_436_promoted_to_fp16)[name = string("op_437_cast_fp16")];
            fp16 var_439_to_fp16 = const()[name = string("op_439_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 32, 2, 1]> var_440_cast_fp16 = add(x = var_437_cast_fp16, y = var_439_to_fp16)[name = string("op_440_cast_fp16")];
            fp32 var_441_epsilon_0 = const()[name = string("op_441_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 2, 1]> var_441_cast_fp16 = rsqrt(epsilon = var_441_epsilon_0, x = var_440_cast_fp16)[name = string("op_441_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_442_cast_fp16 = mul(x = q_3_cast_fp16, y = var_441_cast_fp16)[name = string("op_442_cast_fp16")];
            fp16 var_443_to_fp16 = const()[name = string("op_443_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 32, 2, 128]> q_5_cast_fp16 = mul(x = var_442_cast_fp16, y = var_443_to_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_445_cast_fp16 = mul(x = k_3_cast_fp16, y = k_3_cast_fp16)[name = string("op_445_cast_fp16")];
            tensor<int32, [1]> var_450_axes_0 = const()[name = string("op_450_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_450_keep_dims_0 = const()[name = string("op_450_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 2, 1]> var_450_cast_fp16 = reduce_mean(axes = var_450_axes_0, keep_dims = var_450_keep_dims_0, x = var_445_cast_fp16)[name = string("op_450_cast_fp16")];
            fp16 var_451_promoted_to_fp16 = const()[name = string("op_451_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 32, 2, 1]> var_452_cast_fp16 = mul(x = var_450_cast_fp16, y = var_451_promoted_to_fp16)[name = string("op_452_cast_fp16")];
            fp16 var_454_to_fp16 = const()[name = string("op_454_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 32, 2, 1]> var_455_cast_fp16 = add(x = var_452_cast_fp16, y = var_454_to_fp16)[name = string("op_455_cast_fp16")];
            fp32 var_456_epsilon_0 = const()[name = string("op_456_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 2, 1]> var_456_cast_fp16 = rsqrt(epsilon = var_456_epsilon_0, x = var_455_cast_fp16)[name = string("op_456_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> k_5_cast_fp16 = mul(x = k_3_cast_fp16, y = var_456_cast_fp16)[name = string("k_5_cast_fp16")];
            tensor<int32, [4]> var_462 = const()[name = string("op_462"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_475_axes_0 = const()[name = string("op_475_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 128]> var_463_cast_fp16 = transpose(perm = var_462, x = q_5_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 2, 1, 32, 128]> var_475_cast_fp16 = expand_dims(axes = var_475_axes_0, x = var_463_cast_fp16)[name = string("op_475_cast_fp16")];
            tensor<int32, [5]> var_483_reps_0 = const()[name = string("op_483_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 32, 128]> var_483_cast_fp16 = tile(reps = var_483_reps_0, x = var_475_cast_fp16)[name = string("op_483_cast_fp16")];
            tensor<int32, [4]> var_488 = const()[name = string("op_488"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<fp16, [1, 4, 32, 128]> q_7_cast_fp16 = reshape(shape = var_488, x = var_483_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<int32, [4]> var_494 = const()[name = string("op_494"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_507_axes_0 = const()[name = string("op_507_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 128]> var_495_cast_fp16 = transpose(perm = var_494, x = k_5_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 2, 1, 32, 128]> var_507_cast_fp16 = expand_dims(axes = var_507_axes_0, x = var_495_cast_fp16)[name = string("op_507_cast_fp16")];
            tensor<int32, [5]> var_515_reps_0 = const()[name = string("op_515_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 32, 128]> var_515_cast_fp16 = tile(reps = var_515_reps_0, x = var_507_cast_fp16)[name = string("op_515_cast_fp16")];
            tensor<int32, [4]> var_520 = const()[name = string("op_520"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<fp16, [1, 4, 32, 128]> k_7_cast_fp16 = reshape(shape = var_520, x = var_515_cast_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_522_cast_fp16 = sigmoid(x = b_1_cast_fp16)[name = string("op_522_cast_fp16")];
            tensor<int32, [3]> var_526 = const()[name = string("op_526"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_0_dt_bias_to_fp16 = const()[name = string("layers_0_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 32, 4]> input_5_cast_fp16 = add(x = a_1_cast_fp16, y = layers_0_dt_bias_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_530_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("op_530_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_531_cast_fp16 = abs(x = input_5_cast_fp16)[name = string("op_531_cast_fp16")];
            fp16 const_0_promoted_to_fp16 = const()[name = string("const_0_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, 4]> var_532_cast_fp16 = mul(x = var_531_cast_fp16, y = const_0_promoted_to_fp16)[name = string("op_532_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_533_cast_fp16 = exp(x = var_532_cast_fp16)[name = string("op_533_cast_fp16")];
            fp16 var_535_to_fp16 = const()[name = string("op_535_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 32, 4]> var_536_cast_fp16 = add(x = var_533_cast_fp16, y = var_535_to_fp16)[name = string("op_536_cast_fp16")];
            fp32 var_537_epsilon_0 = const()[name = string("op_537_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 32, 4]> var_537_cast_fp16 = log(epsilon = var_537_epsilon_0, x = var_536_cast_fp16)[name = string("op_537_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_539_cast_fp16 = add(x = var_530_cast_fp16, y = var_537_cast_fp16)[name = string("op_539_cast_fp16")];
            tensor<fp16, [4]> layers_0_negA_to_fp16 = const()[name = string("layers_0_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.b3p+3, -0x1.a7cp+0, -0x1.ac4p+3, -0x1.704p+2])];
            tensor<fp16, [1, 32, 4]> var_540_cast_fp16 = mul(x = layers_0_negA_to_fp16, y = var_539_cast_fp16)[name = string("op_540_cast_fp16")];
            tensor<int32, [3]> var_544 = const()[name = string("op_544"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_547_axes_0 = const()[name = string("op_547_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 32]> beta_1_cast_fp16 = transpose(perm = var_526, x = var_522_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 4, 32, 1]> var_547_cast_fp16 = expand_dims(axes = var_547_axes_0, x = beta_1_cast_fp16)[name = string("op_547_cast_fp16")];
            fp16 var_548_to_fp16 = const()[name = string("op_548_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 32, 1]> var_549_cast_fp16 = mul(x = var_547_cast_fp16, y = var_548_to_fp16)[name = string("op_549_cast_fp16")];
            tensor<fp16, [1, 4, 32, 128]> v_1_cast_fp16 = transpose(perm = var_424, x = var_419_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 4, 32, 128]> vb_1_cast_fp16 = mul(x = v_1_cast_fp16, y = var_549_cast_fp16)[name = string("vb_1_cast_fp16")];
            tensor<int32, [4]> q_9_begin_0 = const()[name = string("q_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_9_end_0 = const()[name = string("q_9_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_9_end_mask_0 = const()[name = string("q_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_9_cast_fp16 = slice_by_index(begin = q_9_begin_0, end = q_9_end_0, end_mask = q_9_end_mask_0, x = q_7_cast_fp16)[name = string("q_9_cast_fp16")];
            tensor<int32, [4]> k_9_begin_0 = const()[name = string("k_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_9_end_0 = const()[name = string("k_9_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_9_end_mask_0 = const()[name = string("k_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, x = k_7_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<fp16, [1, 4, 32, 128]> var_583_cast_fp16 = mul(x = k_7_cast_fp16, y = var_547_cast_fp16)[name = string("op_583_cast_fp16")];
            tensor<int32, [4]> kb_1_begin_0 = const()[name = string("kb_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_1_end_0 = const()[name = string("kb_1_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_1_end_mask_0 = const()[name = string("kb_1_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_1_cast_fp16 = slice_by_index(begin = kb_1_begin_0, end = kb_1_end_0, end_mask = kb_1_end_mask_0, x = var_583_cast_fp16)[name = string("kb_1_cast_fp16")];
            tensor<int32, [4]> vb_3_begin_0 = const()[name = string("vb_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_3_end_0 = const()[name = string("vb_3_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_3_end_mask_0 = const()[name = string("vb_3_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_3_cast_fp16 = slice_by_index(begin = vb_3_begin_0, end = vb_3_end_0, end_mask = vb_3_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_3_cast_fp16")];
            tensor<int32, [3]> gdec_3_begin_0 = const()[name = string("gdec_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_3_end_0 = const()[name = string("gdec_3_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_3_end_mask_0 = const()[name = string("gdec_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 32]> gdec_1_cast_fp16 = transpose(perm = var_544, x = var_540_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 4, 16]> gdec_3_cast_fp16 = slice_by_index(begin = gdec_3_begin_0, end = gdec_3_end_0, end_mask = gdec_3_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_3_cast_fp16")];
            tensor<int32, [2]> var_631 = const()[name = string("op_631"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_632_cast_fp16 = reshape(shape = var_631, x = gdec_3_cast_fp16)[name = string("op_632_cast_fp16")];
            tensor<fp16, [256, 16]> transpose_0_to_fp16 = const()[name = string("transpose_0_to_fp16"), val = tensor<fp16, [256, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22488000)))];
            tensor<fp16, [256]> var_633_bias_0_to_fp16 = const()[name = string("op_633_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22496256)))];
            tensor<fp16, [4, 256]> var_633_cast_fp16 = linear(bias = var_633_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_632_cast_fp16)[name = string("op_633_cast_fp16")];
            tensor<int32, [4]> var_638 = const()[name = string("op_638"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_1_cast_fp16 = reshape(shape = var_638, x = var_633_cast_fp16)[name = string("P_1_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_640_cast_fp16 = exp(x = P_1_cast_fp16)[name = string("op_640_cast_fp16")];
            tensor<fp16, [16, 16]> tril_incl_to_fp16 = const()[name = string("tril_incl_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22496832)))];
            tensor<fp16, [1, 4, 16, 16]> D_1_cast_fp16 = mul(x = var_640_cast_fp16, y = tril_incl_to_fp16)[name = string("D_1_cast_fp16")];
            bool kk_1_transpose_x_1 = const()[name = string("kk_1_transpose_x_1"), val = bool(false)];
            bool kk_1_transpose_y_1 = const()[name = string("kk_1_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_1_cast_fp16 = matmul(transpose_x = kk_1_transpose_x_1, transpose_y = kk_1_transpose_y_1, x = kb_1_cast_fp16, y = k_9_cast_fp16)[name = string("kk_1_cast_fp16")];
            fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_646_cast_fp16 = mul(x = kk_1_cast_fp16, y = const_1_promoted_to_fp16)[name = string("op_646_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_647_cast_fp16 = mul(x = var_646_cast_fp16, y = D_1_cast_fp16)[name = string("op_647_cast_fp16")];
            tensor<fp16, [16, 16]> cum_tail_to_fp16 = const()[name = string("cum_tail_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22497408)))];
            tensor<fp16, [1, 4, 16, 16]> N_1_cast_fp16 = mul(x = var_647_cast_fp16, y = cum_tail_to_fp16)[name = string("N_1_cast_fp16")];
            tensor<fp16, [16, 16]> neg_tril_strict_to_fp16 = const()[name = string("neg_tril_strict_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22497984)))];
            tensor<fp16, [1, 4, 16, 16]> var_649_cast_fp16 = mul(x = D_1_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_649_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_1_cast_fp16 = mul(x = kk_1_cast_fp16, y = var_649_cast_fp16)[name = string("Na_1_cast_fp16")];
            tensor<fp16, [16, 16]> eye_to_fp16 = const()[name = string("eye_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22498560)))];
            tensor<fp16, [1, 4, 16, 16]> Minv_1_cast_fp16 = add(x = eye_to_fp16, y = N_1_cast_fp16)[name = string("Minv_1_cast_fp16")];
            bool Np_1_transpose_x_0 = const()[name = string("Np_1_transpose_x_0"), val = bool(false)];
            bool Np_1_transpose_y_0 = const()[name = string("Np_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_1_cast_fp16 = matmul(transpose_x = Np_1_transpose_x_0, transpose_y = Np_1_transpose_y_0, x = N_1_cast_fp16, y = Na_1_cast_fp16)[name = string("Np_1_cast_fp16")];
            bool Npa_1_transpose_x_0 = const()[name = string("Npa_1_transpose_x_0"), val = bool(false)];
            bool Npa_1_transpose_y_0 = const()[name = string("Npa_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_1_cast_fp16 = matmul(transpose_x = Npa_1_transpose_x_0, transpose_y = Npa_1_transpose_y_0, x = Na_1_cast_fp16, y = N_1_cast_fp16)[name = string("Npa_1_cast_fp16")];
            bool var_655_transpose_x_0 = const()[name = string("op_655_transpose_x_0"), val = bool(false)];
            bool var_655_transpose_y_0 = const()[name = string("op_655_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_655_cast_fp16 = matmul(transpose_x = var_655_transpose_x_0, transpose_y = var_655_transpose_y_0, x = Minv_1_cast_fp16, y = Np_1_cast_fp16)[name = string("op_655_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_3_cast_fp16 = add(x = Minv_1_cast_fp16, y = var_655_cast_fp16)[name = string("Minv_3_cast_fp16")];
            bool Np_3_transpose_x_0 = const()[name = string("Np_3_transpose_x_0"), val = bool(false)];
            bool Np_3_transpose_y_0 = const()[name = string("Np_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_3_cast_fp16 = matmul(transpose_x = Np_3_transpose_x_0, transpose_y = Np_3_transpose_y_0, x = Np_1_cast_fp16, y = Npa_1_cast_fp16)[name = string("Np_3_cast_fp16")];
            bool Npa_3_transpose_x_0 = const()[name = string("Npa_3_transpose_x_0"), val = bool(false)];
            bool Npa_3_transpose_y_0 = const()[name = string("Npa_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_3_cast_fp16 = matmul(transpose_x = Npa_3_transpose_x_0, transpose_y = Npa_3_transpose_y_0, x = Npa_1_cast_fp16, y = Np_1_cast_fp16)[name = string("Npa_3_cast_fp16")];
            bool var_660_transpose_x_0 = const()[name = string("op_660_transpose_x_0"), val = bool(false)];
            bool var_660_transpose_y_0 = const()[name = string("op_660_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_660_cast_fp16 = matmul(transpose_x = var_660_transpose_x_0, transpose_y = var_660_transpose_y_0, x = Minv_3_cast_fp16, y = Np_3_cast_fp16)[name = string("op_660_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_5_cast_fp16 = add(x = Minv_3_cast_fp16, y = var_660_cast_fp16)[name = string("Minv_5_cast_fp16")];
            bool Np_5_transpose_x_0 = const()[name = string("Np_5_transpose_x_0"), val = bool(false)];
            bool Np_5_transpose_y_0 = const()[name = string("Np_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_5_cast_fp16 = matmul(transpose_x = Np_5_transpose_x_0, transpose_y = Np_5_transpose_y_0, x = Np_3_cast_fp16, y = Npa_3_cast_fp16)[name = string("Np_5_cast_fp16")];
            bool var_664_transpose_x_0 = const()[name = string("op_664_transpose_x_0"), val = bool(false)];
            bool var_664_transpose_y_0 = const()[name = string("op_664_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_664_cast_fp16 = matmul(transpose_x = var_664_transpose_x_0, transpose_y = var_664_transpose_y_0, x = Minv_5_cast_fp16, y = Np_5_cast_fp16)[name = string("op_664_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_7_cast_fp16 = add(x = Minv_5_cast_fp16, y = var_664_cast_fp16)[name = string("Minv_7_cast_fp16")];
            bool u_1_transpose_x_0 = const()[name = string("u_1_transpose_x_0"), val = bool(false)];
            bool u_1_transpose_y_0 = const()[name = string("u_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_1_cast_fp16 = matmul(transpose_x = u_1_transpose_x_0, transpose_y = u_1_transpose_y_0, x = Minv_7_cast_fp16, y = vb_3_cast_fp16)[name = string("u_1_cast_fp16")];
            bool var_671_transpose_x_1 = const()[name = string("op_671_transpose_x_1"), val = bool(false)];
            bool var_671_transpose_y_1 = const()[name = string("op_671_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_671_cast_fp16 = matmul(transpose_x = var_671_transpose_x_1, transpose_y = var_671_transpose_y_1, x = q_9_cast_fp16, y = k_9_cast_fp16)[name = string("op_671_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_1_cast_fp16 = mul(x = var_671_cast_fp16, y = D_1_cast_fp16)[name = string("A_1_cast_fp16")];
            bool var_673_transpose_x_0 = const()[name = string("op_673_transpose_x_0"), val = bool(false)];
            bool var_673_transpose_y_0 = const()[name = string("op_673_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_673_cast_fp16 = matmul(transpose_x = var_673_transpose_x_0, transpose_y = var_673_transpose_y_0, x = A_1_cast_fp16, y = u_1_cast_fp16)[name = string("op_673_cast_fp16")];
            tensor<fp16, [16, 16]> transpose_1_to_fp16 = const()[name = string("transpose_1_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658176)))];
            tensor<fp16, [16]> var_674_bias_0_to_fp16 = const()[name = string("op_674_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_674_cast_fp16 = linear(bias = var_674_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_3_cast_fp16)[name = string("op_674_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_1_cast_fp16 = exp(x = var_674_cast_fp16)[name = string("tail_1_cast_fp16")];
            tensor<int32, [1]> var_677_axes_0 = const()[name = string("op_677_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_677_cast_fp16 = expand_dims(axes = var_677_axes_0, x = tail_1_cast_fp16)[name = string("op_677_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_678_cast_fp16 = mul(x = k_9_cast_fp16, y = var_677_cast_fp16)[name = string("op_678_cast_fp16")];
            bool S_1_transpose_x_1 = const()[name = string("S_1_transpose_x_1"), val = bool(true)];
            bool S_1_transpose_y_1 = const()[name = string("S_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_1_cast_fp16 = matmul(transpose_x = S_1_transpose_x_1, transpose_y = S_1_transpose_y_1, x = var_678_cast_fp16, y = u_1_cast_fp16)[name = string("S_1_cast_fp16")];
            tensor<int32, [4]> q_11_begin_0 = const()[name = string("q_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_11_end_0 = const()[name = string("q_11_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_11_end_mask_0 = const()[name = string("q_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_11_cast_fp16 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, x = q_7_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<int32, [4]> k_11_begin_0 = const()[name = string("k_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_11_end_0 = const()[name = string("k_11_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_11_end_mask_0 = const()[name = string("k_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_11_cast_fp16 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, x = k_7_cast_fp16)[name = string("k_11_cast_fp16")];
            tensor<int32, [4]> kb_3_begin_0 = const()[name = string("kb_3_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_3_end_0 = const()[name = string("kb_3_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_3_end_mask_0 = const()[name = string("kb_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_3_cast_fp16 = slice_by_index(begin = kb_3_begin_0, end = kb_3_end_0, end_mask = kb_3_end_mask_0, x = var_583_cast_fp16)[name = string("kb_3_cast_fp16")];
            tensor<int32, [4]> vb_5_begin_0 = const()[name = string("vb_5_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_5_end_0 = const()[name = string("vb_5_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_5_end_mask_0 = const()[name = string("vb_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_5_cast_fp16 = slice_by_index(begin = vb_5_begin_0, end = vb_5_end_0, end_mask = vb_5_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_5_cast_fp16")];
            tensor<int32, [3]> gdec_5_begin_0 = const()[name = string("gdec_5_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_5_end_0 = const()[name = string("gdec_5_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_5_end_mask_0 = const()[name = string("gdec_5_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_5_cast_fp16 = slice_by_index(begin = gdec_5_begin_0, end = gdec_5_end_0, end_mask = gdec_5_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_5_cast_fp16")];
            tensor<int32, [2]> var_763 = const()[name = string("op_763"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_764_cast_fp16 = reshape(shape = var_763, x = gdec_5_cast_fp16)[name = string("op_764_cast_fp16")];
            tensor<fp16, [4, 256]> var_765_cast_fp16 = linear(bias = var_633_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_764_cast_fp16)[name = string("op_765_cast_fp16")];
            tensor<int32, [4]> var_770 = const()[name = string("op_770"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_3_cast_fp16 = reshape(shape = var_770, x = var_765_cast_fp16)[name = string("P_3_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_772_cast_fp16 = exp(x = P_3_cast_fp16)[name = string("op_772_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_3_cast_fp16 = mul(x = var_772_cast_fp16, y = tril_incl_to_fp16)[name = string("D_3_cast_fp16")];
            bool kk_3_transpose_x_1 = const()[name = string("kk_3_transpose_x_1"), val = bool(false)];
            bool kk_3_transpose_y_1 = const()[name = string("kk_3_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_3_cast_fp16 = matmul(transpose_x = kk_3_transpose_x_1, transpose_y = kk_3_transpose_y_1, x = kb_3_cast_fp16, y = k_11_cast_fp16)[name = string("kk_3_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_778_cast_fp16 = mul(x = kk_3_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_778_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_779_cast_fp16 = mul(x = var_778_cast_fp16, y = D_3_cast_fp16)[name = string("op_779_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_3_cast_fp16 = mul(x = var_779_cast_fp16, y = cum_tail_to_fp16)[name = string("N_3_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_781_cast_fp16 = mul(x = D_3_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_781_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_3_cast_fp16 = mul(x = kk_3_cast_fp16, y = var_781_cast_fp16)[name = string("Na_3_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_9_cast_fp16 = add(x = eye_to_fp16, y = N_3_cast_fp16)[name = string("Minv_9_cast_fp16")];
            bool Np_7_transpose_x_0 = const()[name = string("Np_7_transpose_x_0"), val = bool(false)];
            bool Np_7_transpose_y_0 = const()[name = string("Np_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_7_cast_fp16 = matmul(transpose_x = Np_7_transpose_x_0, transpose_y = Np_7_transpose_y_0, x = N_3_cast_fp16, y = Na_3_cast_fp16)[name = string("Np_7_cast_fp16")];
            bool Npa_5_transpose_x_0 = const()[name = string("Npa_5_transpose_x_0"), val = bool(false)];
            bool Npa_5_transpose_y_0 = const()[name = string("Npa_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_5_cast_fp16 = matmul(transpose_x = Npa_5_transpose_x_0, transpose_y = Npa_5_transpose_y_0, x = Na_3_cast_fp16, y = N_3_cast_fp16)[name = string("Npa_5_cast_fp16")];
            bool var_787_transpose_x_0 = const()[name = string("op_787_transpose_x_0"), val = bool(false)];
            bool var_787_transpose_y_0 = const()[name = string("op_787_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_787_cast_fp16 = matmul(transpose_x = var_787_transpose_x_0, transpose_y = var_787_transpose_y_0, x = Minv_9_cast_fp16, y = Np_7_cast_fp16)[name = string("op_787_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_11_cast_fp16 = add(x = Minv_9_cast_fp16, y = var_787_cast_fp16)[name = string("Minv_11_cast_fp16")];
            bool Np_9_transpose_x_0 = const()[name = string("Np_9_transpose_x_0"), val = bool(false)];
            bool Np_9_transpose_y_0 = const()[name = string("Np_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_9_cast_fp16 = matmul(transpose_x = Np_9_transpose_x_0, transpose_y = Np_9_transpose_y_0, x = Np_7_cast_fp16, y = Npa_5_cast_fp16)[name = string("Np_9_cast_fp16")];
            bool Npa_7_transpose_x_0 = const()[name = string("Npa_7_transpose_x_0"), val = bool(false)];
            bool Npa_7_transpose_y_0 = const()[name = string("Npa_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_7_cast_fp16 = matmul(transpose_x = Npa_7_transpose_x_0, transpose_y = Npa_7_transpose_y_0, x = Npa_5_cast_fp16, y = Np_7_cast_fp16)[name = string("Npa_7_cast_fp16")];
            bool var_792_transpose_x_0 = const()[name = string("op_792_transpose_x_0"), val = bool(false)];
            bool var_792_transpose_y_0 = const()[name = string("op_792_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_792_cast_fp16 = matmul(transpose_x = var_792_transpose_x_0, transpose_y = var_792_transpose_y_0, x = Minv_11_cast_fp16, y = Np_9_cast_fp16)[name = string("op_792_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_13_cast_fp16 = add(x = Minv_11_cast_fp16, y = var_792_cast_fp16)[name = string("Minv_13_cast_fp16")];
            bool Np_11_transpose_x_0 = const()[name = string("Np_11_transpose_x_0"), val = bool(false)];
            bool Np_11_transpose_y_0 = const()[name = string("Np_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_11_cast_fp16 = matmul(transpose_x = Np_11_transpose_x_0, transpose_y = Np_11_transpose_y_0, x = Np_9_cast_fp16, y = Npa_7_cast_fp16)[name = string("Np_11_cast_fp16")];
            bool var_796_transpose_x_0 = const()[name = string("op_796_transpose_x_0"), val = bool(false)];
            bool var_796_transpose_y_0 = const()[name = string("op_796_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_796_cast_fp16 = matmul(transpose_x = var_796_transpose_x_0, transpose_y = var_796_transpose_y_0, x = Minv_13_cast_fp16, y = Np_11_cast_fp16)[name = string("op_796_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_15_cast_fp16 = add(x = Minv_13_cast_fp16, y = var_796_cast_fp16)[name = string("Minv_15_cast_fp16")];
            bool u_3_transpose_x_0 = const()[name = string("u_3_transpose_x_0"), val = bool(false)];
            bool u_3_transpose_y_0 = const()[name = string("u_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_3_cast_fp16 = matmul(transpose_x = u_3_transpose_x_0, transpose_y = u_3_transpose_y_0, x = Minv_15_cast_fp16, y = vb_5_cast_fp16)[name = string("u_3_cast_fp16")];
            bool var_803_transpose_x_1 = const()[name = string("op_803_transpose_x_1"), val = bool(false)];
            bool var_803_transpose_y_1 = const()[name = string("op_803_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_803_cast_fp16 = matmul(transpose_x = var_803_transpose_x_1, transpose_y = var_803_transpose_y_1, x = q_11_cast_fp16, y = k_11_cast_fp16)[name = string("op_803_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_3_cast_fp16 = mul(x = var_803_cast_fp16, y = D_3_cast_fp16)[name = string("A_3_cast_fp16")];
            tensor<fp16, [16]> var_805_bias_0_to_fp16 = const()[name = string("op_805_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_805_cast_fp16 = linear(bias = var_805_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_5_cast_fp16)[name = string("op_805_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_1_cast_fp16 = exp(x = var_805_cast_fp16)[name = string("cum_1_cast_fp16")];
            tensor<int32, [1]> var_808_axes_0 = const()[name = string("op_808_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_808_cast_fp16 = expand_dims(axes = var_808_axes_0, x = cum_1_cast_fp16)[name = string("op_808_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_809_cast_fp16 = mul(x = kb_3_cast_fp16, y = var_808_cast_fp16)[name = string("op_809_cast_fp16")];
            bool w_3_transpose_x_0 = const()[name = string("w_3_transpose_x_0"), val = bool(false)];
            bool w_3_transpose_y_0 = const()[name = string("w_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_3_cast_fp16 = matmul(transpose_x = w_3_transpose_x_0, transpose_y = w_3_transpose_y_0, x = Minv_15_cast_fp16, y = var_809_cast_fp16)[name = string("w_3_cast_fp16")];
            bool var_811_transpose_x_0 = const()[name = string("op_811_transpose_x_0"), val = bool(false)];
            bool var_811_transpose_y_0 = const()[name = string("op_811_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_811_cast_fp16 = matmul(transpose_x = var_811_transpose_x_0, transpose_y = var_811_transpose_y_0, x = w_3_cast_fp16, y = S_1_cast_fp16)[name = string("op_811_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_1_cast_fp16 = sub(x = u_3_cast_fp16, y = var_811_cast_fp16)[name = string("vnew_1_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_816_cast_fp16 = mul(x = q_11_cast_fp16, y = var_808_cast_fp16)[name = string("op_816_cast_fp16")];
            bool var_817_transpose_x_0 = const()[name = string("op_817_transpose_x_0"), val = bool(false)];
            bool var_817_transpose_y_0 = const()[name = string("op_817_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_817_cast_fp16 = matmul(transpose_x = var_817_transpose_x_0, transpose_y = var_817_transpose_y_0, x = var_816_cast_fp16, y = S_1_cast_fp16)[name = string("op_817_cast_fp16")];
            bool var_818_transpose_x_0 = const()[name = string("op_818_transpose_x_0"), val = bool(false)];
            bool var_818_transpose_y_0 = const()[name = string("op_818_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_818_cast_fp16 = matmul(transpose_x = var_818_transpose_x_0, transpose_y = var_818_transpose_y_0, x = A_3_cast_fp16, y = vnew_1_cast_fp16)[name = string("op_818_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_1_cast_fp16 = add(x = var_817_cast_fp16, y = var_818_cast_fp16)[name = string("oc_1_cast_fp16")];
            int32 var_822 = const()[name = string("op_822"), val = int32(2)];
            bool o_1_interleave_0 = const()[name = string("o_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 128]> o_1_cast_fp16 = concat(axis = var_822, interleave = o_1_interleave_0, values = (var_673_cast_fp16, oc_1_cast_fp16))[name = string("o_1_cast_fp16")];
            tensor<int32, [4]> var_828 = const()[name = string("op_828"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_834 = const()[name = string("op_834"), val = tensor<int32, [4]>([1, 32, 4, 128])];
            tensor<fp16, [1, 32, 4, 128]> input_7_cast_fp16 = reshape(shape = var_834, x = z_1_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> o_3_cast_fp16 = transpose(perm = var_828, x = o_1_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 32, 4, 128]> var_836_cast_fp16 = mul(x = o_3_cast_fp16, y = o_3_cast_fp16)[name = string("op_836_cast_fp16")];
            tensor<int32, [1]> var_841_axes_0 = const()[name = string("op_841_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_841_keep_dims_0 = const()[name = string("op_841_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 4, 1]> var_841_cast_fp16 = reduce_mean(axes = var_841_axes_0, keep_dims = var_841_keep_dims_0, x = var_836_cast_fp16)[name = string("op_841_cast_fp16")];
            fp16 var_843_to_fp16 = const()[name = string("op_843_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 32, 4, 1]> var_844_cast_fp16 = add(x = var_841_cast_fp16, y = var_843_to_fp16)[name = string("op_844_cast_fp16")];
            fp32 var_845_epsilon_0 = const()[name = string("op_845_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 4, 1]> var_845_cast_fp16 = rsqrt(epsilon = var_845_epsilon_0, x = var_844_cast_fp16)[name = string("op_845_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> var_846_cast_fp16 = mul(x = o_3_cast_fp16, y = var_845_cast_fp16)[name = string("op_846_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> var_848_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_848_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> o_5_cast_fp16 = mul(x = var_846_cast_fp16, y = var_848_cast_fp16)[name = string("o_5_cast_fp16")];
            tensor<int32, [3]> var_853 = const()[name = string("op_853"), val = tensor<int32, [3]>([1, 32, 512])];
            tensor<fp16, [1, 32, 512]> input_9_cast_fp16 = reshape(shape = var_853, x = o_5_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [640, 512]> layers_0_out_weight_to_fp16 = const()[name = string("layers_0_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22499136)))];
            tensor<fp16, [640]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23154560)))];
            tensor<fp16, [1, 32, 640]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_out_weight_to_fp16, x = input_9_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [1, 32, 640]> x_5_cast_fp16 = add(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = linear_1_cast_fp16)[name = string("x_5_cast_fp16")];
            fp16 var_866_to_fp16 = const()[name = string("op_866_to_fp16"), val = fp16(0x1p+5)];
            tensor<fp16, [1, 32, 640]> x_7_cast_fp16 = mul(x = x_5_cast_fp16, y = var_866_to_fp16)[name = string("x_7_cast_fp16")];
            fp16 var_864_promoted_to_fp16 = const()[name = string("op_864_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> var_868_cast_fp16 = pow(x = x_7_cast_fp16, y = var_864_promoted_to_fp16)[name = string("op_868_cast_fp16")];
            tensor<int32, [1]> var_870_axes_0 = const()[name = string("op_870_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_870_keep_dims_0 = const()[name = string("op_870_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_870_cast_fp16 = reduce_mean(axes = var_870_axes_0, keep_dims = var_870_keep_dims_0, x = var_868_cast_fp16)[name = string("op_870_cast_fp16")];
            fp16 var_871_to_fp16 = const()[name = string("op_871_to_fp16"), val = fp16(0x1.0c8p-10)];
            tensor<fp16, [1, 32, 1]> var_872_cast_fp16 = add(x = var_870_cast_fp16, y = var_871_to_fp16)[name = string("op_872_cast_fp16")];
            fp32 var_873_epsilon_0 = const()[name = string("op_873_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_873_cast_fp16 = rsqrt(epsilon = var_873_epsilon_0, x = var_872_cast_fp16)[name = string("op_873_cast_fp16")];
            tensor<fp16, [1, 32, 640]> var_874_cast_fp16 = mul(x = x_7_cast_fp16, y = var_873_cast_fp16)[name = string("op_874_cast_fp16")];
            tensor<fp16, [640]> layers_0_n2_w1_to_fp16 = const()[name = string("layers_0_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23155904)))];
            tensor<fp16, [1, 32, 640]> input_11_cast_fp16 = mul(x = var_874_cast_fp16, y = layers_0_n2_w1_to_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_0_gate_up_weight_to_fp16 = const()[name = string("layers_0_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23157248)))];
            tensor<fp16, [5120]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29710912)))];
            tensor<fp16, [1, 32, 5120]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_0_gate_up_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [3]> input_13_begin_0 = const()[name = string("input_13_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_13_end_0 = const()[name = string("input_13_end_0"), val = tensor<int32, [3]>([1, 32, 2560])];
            tensor<bool, [3]> input_13_end_mask_0 = const()[name = string("input_13_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 2560]> input_13_cast_fp16 = slice_by_index(begin = input_13_begin_0, end = input_13_end_0, end_mask = input_13_end_mask_0, x = linear_2_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 32, 2560]> var_884_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_884_cast_fp16")];
            tensor<int32, [3]> var_889_begin_0 = const()[name = string("op_889_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_889_end_0 = const()[name = string("op_889_end_0"), val = tensor<int32, [3]>([1, 32, 5120])];
            tensor<bool, [3]> var_889_end_mask_0 = const()[name = string("op_889_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 2560]> var_889_cast_fp16 = slice_by_index(begin = var_889_begin_0, end = var_889_end_0, end_mask = var_889_end_mask_0, x = linear_2_cast_fp16)[name = string("op_889_cast_fp16")];
            tensor<fp16, [1, 32, 2560]> input_15_cast_fp16 = mul(x = var_884_cast_fp16, y = var_889_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_0_down_weight_to_fp16 = const()[name = string("layers_0_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29721216)))];
            tensor<fp16, [1, 32, 640]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_down_weight_to_fp16, x = input_15_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 32, 640]> x_9_cast_fp16 = add(x = x_5_cast_fp16, y = linear_3_cast_fp16)[name = string("x_9_cast_fp16")];
            fp16 var_902_to_fp16 = const()[name = string("op_902_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> x_11_cast_fp16 = mul(x = x_9_cast_fp16, y = var_902_to_fp16)[name = string("x_11_cast_fp16")];
            fp16 var_900_promoted_to_fp16 = const()[name = string("op_900_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> var_904_cast_fp16 = pow(x = x_11_cast_fp16, y = var_900_promoted_to_fp16)[name = string("op_904_cast_fp16")];
            tensor<int32, [1]> var_906_axes_0 = const()[name = string("op_906_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_906_keep_dims_0 = const()[name = string("op_906_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_906_cast_fp16 = reduce_mean(axes = var_906_axes_0, keep_dims = var_906_keep_dims_0, x = var_904_cast_fp16)[name = string("op_906_cast_fp16")];
            fp16 var_907_to_fp16 = const()[name = string("op_907_to_fp16"), val = fp16(0x1.0cp-18)];
            tensor<fp16, [1, 32, 1]> var_908_cast_fp16 = add(x = var_906_cast_fp16, y = var_907_to_fp16)[name = string("op_908_cast_fp16")];
            fp32 var_909_epsilon_0 = const()[name = string("op_909_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_909_cast_fp16 = rsqrt(epsilon = var_909_epsilon_0, x = var_908_cast_fp16)[name = string("op_909_cast_fp16")];
            tensor<fp16, [1, 32, 640]> var_910_cast_fp16 = mul(x = x_11_cast_fp16, y = var_909_cast_fp16)[name = string("op_910_cast_fp16")];
            tensor<fp16, [640]> layers_1_n1_w1_to_fp16 = const()[name = string("layers_1_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32998080)))];
            tensor<fp16, [1, 32, 640]> input_17_cast_fp16 = mul(x = var_910_cast_fp16, y = layers_1_n1_w1_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_1_qkvzba_weight_to_fp16 = const()[name = string("layers_1_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32999424)))];
            tensor<fp16, [1, 32, 1544]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_qkvzba_weight_to_fp16, x = input_17_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<int32, [3]> qkv_3_begin_0 = const()[name = string("qkv_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_3_end_0 = const()[name = string("qkv_3_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> qkv_3_end_mask_0 = const()[name = string("qkv_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 1024]> qkv_3_cast_fp16 = slice_by_index(begin = qkv_3_begin_0, end = qkv_3_end_0, end_mask = qkv_3_end_mask_0, x = linear_4_cast_fp16)[name = string("qkv_3_cast_fp16")];
            tensor<int32, [3]> z_3_begin_0 = const()[name = string("z_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_3_end_0 = const()[name = string("z_3_end_0"), val = tensor<int32, [3]>([1, 32, 1536])];
            tensor<bool, [3]> z_3_end_mask_0 = const()[name = string("z_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 512]> z_3_cast_fp16 = slice_by_index(begin = z_3_begin_0, end = z_3_end_0, end_mask = z_3_end_mask_0, x = linear_4_cast_fp16)[name = string("z_3_cast_fp16")];
            tensor<int32, [3]> b_3_begin_0 = const()[name = string("b_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_3_end_0 = const()[name = string("b_3_end_0"), val = tensor<int32, [3]>([1, 32, 1540])];
            tensor<bool, [3]> b_3_end_mask_0 = const()[name = string("b_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 4]> b_3_cast_fp16 = slice_by_index(begin = b_3_begin_0, end = b_3_end_0, end_mask = b_3_end_mask_0, x = linear_4_cast_fp16)[name = string("b_3_cast_fp16")];
            tensor<int32, [3]> a_3_begin_0 = const()[name = string("a_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_3_end_0 = const()[name = string("a_3_end_0"), val = tensor<int32, [3]>([1, 32, 1])];
            tensor<bool, [3]> a_3_end_mask_0 = const()[name = string("a_3_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 4]> a_3_cast_fp16 = slice_by_index(begin = a_3_begin_0, end = a_3_end_0, end_mask = a_3_end_mask_0, x = linear_4_cast_fp16)[name = string("a_3_cast_fp16")];
            int32 var_949 = const()[name = string("op_949"), val = int32(1)];
            bool xp_3_interleave_0 = const()[name = string("xp_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 35, 1024]> xp_3_cast_fp16 = concat(axis = var_949, interleave = xp_3_interleave_0, values = (zpad_to_fp16, qkv_3_cast_fp16))[name = string("xp_3_cast_fp16")];
            tensor<int32, [3]> var_960_begin_0 = const()[name = string("op_960_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_960_end_0 = const()[name = string("op_960_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_960_end_mask_0 = const()[name = string("op_960_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 1024]> var_960_cast_fp16 = slice_by_index(begin = var_960_begin_0, end = var_960_end_0, end_mask = var_960_end_mask_0, x = xp_3_cast_fp16)[name = string("op_960_cast_fp16")];
            tensor<fp16, [1024]> var_968_to_fp16 = const()[name = string("op_968_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34975808)))];
            tensor<fp16, [1, 32, 1024]> var_969_cast_fp16 = mul(x = var_960_cast_fp16, y = var_968_to_fp16)[name = string("op_969_cast_fp16")];
            tensor<int32, [3]> var_979_begin_0 = const()[name = string("op_979_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_979_end_0 = const()[name = string("op_979_end_0"), val = tensor<int32, [3]>([1, 34, 1024])];
            tensor<bool, [3]> var_979_end_mask_0 = const()[name = string("op_979_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_979_cast_fp16 = slice_by_index(begin = var_979_begin_0, end = var_979_end_0, end_mask = var_979_end_mask_0, x = xp_3_cast_fp16)[name = string("op_979_cast_fp16")];
            tensor<fp16, [1024]> var_987_to_fp16 = const()[name = string("op_987_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34977920)))];
            tensor<fp16, [1, 32, 1024]> var_988_cast_fp16 = mul(x = var_979_cast_fp16, y = var_987_to_fp16)[name = string("op_988_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> var_990_cast_fp16 = add(x = var_969_cast_fp16, y = var_988_cast_fp16)[name = string("op_990_cast_fp16")];
            tensor<int32, [3]> var_1000_begin_0 = const()[name = string("op_1000_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1000_end_0 = const()[name = string("op_1000_end_0"), val = tensor<int32, [3]>([1, 33, 1024])];
            tensor<bool, [3]> var_1000_end_mask_0 = const()[name = string("op_1000_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_1000_cast_fp16 = slice_by_index(begin = var_1000_begin_0, end = var_1000_end_0, end_mask = var_1000_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1000_cast_fp16")];
            tensor<fp16, [1024]> var_1008_to_fp16 = const()[name = string("op_1008_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34980032)))];
            tensor<fp16, [1, 32, 1024]> var_1009_cast_fp16 = mul(x = var_1000_cast_fp16, y = var_1008_to_fp16)[name = string("op_1009_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> var_1011_cast_fp16 = add(x = var_990_cast_fp16, y = var_1009_cast_fp16)[name = string("op_1011_cast_fp16")];
            tensor<int32, [3]> var_1021_begin_0 = const()[name = string("op_1021_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1021_end_0 = const()[name = string("op_1021_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> var_1021_end_mask_0 = const()[name = string("op_1021_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_1021_cast_fp16 = slice_by_index(begin = var_1021_begin_0, end = var_1021_end_0, end_mask = var_1021_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1021_cast_fp16")];
            tensor<fp16, [1024]> var_1029_to_fp16 = const()[name = string("op_1029_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34982144)))];
            tensor<fp16, [1, 32, 1024]> var_1030_cast_fp16 = mul(x = var_1021_cast_fp16, y = var_1029_to_fp16)[name = string("op_1030_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> input_19_cast_fp16 = add(x = var_1011_cast_fp16, y = var_1030_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> c_3_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("c_3_cast_fp16")];
            tensor<int32, [3]> var_1038_begin_0 = const()[name = string("op_1038_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1038_end_0 = const()[name = string("op_1038_end_0"), val = tensor<int32, [3]>([1, 32, 256])];
            tensor<bool, [3]> var_1038_end_mask_0 = const()[name = string("op_1038_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 256]> var_1038_cast_fp16 = slice_by_index(begin = var_1038_begin_0, end = var_1038_end_0, end_mask = var_1038_end_mask_0, x = c_3_cast_fp16)[name = string("op_1038_cast_fp16")];
            tensor<int32, [4]> var_1043 = const()[name = string("op_1043"), val = tensor<int32, [4]>([1, 32, 2, 128])];
            tensor<fp16, [1, 32, 2, 128]> q_13_cast_fp16 = reshape(shape = var_1043, x = var_1038_cast_fp16)[name = string("q_13_cast_fp16")];
            tensor<int32, [3]> var_1049_begin_0 = const()[name = string("op_1049_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_1049_end_0 = const()[name = string("op_1049_end_0"), val = tensor<int32, [3]>([1, 32, 512])];
            tensor<bool, [3]> var_1049_end_mask_0 = const()[name = string("op_1049_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 256]> var_1049_cast_fp16 = slice_by_index(begin = var_1049_begin_0, end = var_1049_end_0, end_mask = var_1049_end_mask_0, x = c_3_cast_fp16)[name = string("op_1049_cast_fp16")];
            tensor<int32, [4]> var_1054 = const()[name = string("op_1054"), val = tensor<int32, [4]>([1, 32, 2, 128])];
            tensor<fp16, [1, 32, 2, 128]> k_13_cast_fp16 = reshape(shape = var_1054, x = var_1049_cast_fp16)[name = string("k_13_cast_fp16")];
            tensor<int32, [3]> var_1060_begin_0 = const()[name = string("op_1060_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_1060_end_0 = const()[name = string("op_1060_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> var_1060_end_mask_0 = const()[name = string("op_1060_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 512]> var_1060_cast_fp16 = slice_by_index(begin = var_1060_begin_0, end = var_1060_end_0, end_mask = var_1060_end_mask_0, x = c_3_cast_fp16)[name = string("op_1060_cast_fp16")];
            tensor<int32, [4]> var_1065 = const()[name = string("op_1065"), val = tensor<int32, [4]>([1, 32, 4, 128])];
            tensor<fp16, [1, 32, 4, 128]> var_1066_cast_fp16 = reshape(shape = var_1065, x = var_1060_cast_fp16)[name = string("op_1066_cast_fp16")];
            tensor<int32, [4]> var_1071 = const()[name = string("op_1071"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1073_to_fp16 = const()[name = string("op_1073_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 32, 2, 128]> q_15_cast_fp16 = mul(x = q_13_cast_fp16, y = var_1073_to_fp16)[name = string("q_15_cast_fp16")];
            fp16 var_1075_to_fp16 = const()[name = string("op_1075_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 32, 2, 128]> k_15_cast_fp16 = mul(x = k_13_cast_fp16, y = var_1075_to_fp16)[name = string("k_15_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_1077_cast_fp16 = mul(x = q_15_cast_fp16, y = q_15_cast_fp16)[name = string("op_1077_cast_fp16")];
            tensor<int32, [1]> var_1082_axes_0 = const()[name = string("op_1082_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1082_keep_dims_0 = const()[name = string("op_1082_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 2, 1]> var_1082_cast_fp16 = reduce_mean(axes = var_1082_axes_0, keep_dims = var_1082_keep_dims_0, x = var_1077_cast_fp16)[name = string("op_1082_cast_fp16")];
            fp16 var_1083_promoted_to_fp16 = const()[name = string("op_1083_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 32, 2, 1]> var_1084_cast_fp16 = mul(x = var_1082_cast_fp16, y = var_1083_promoted_to_fp16)[name = string("op_1084_cast_fp16")];
            fp16 var_1086_to_fp16 = const()[name = string("op_1086_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 32, 2, 1]> var_1087_cast_fp16 = add(x = var_1084_cast_fp16, y = var_1086_to_fp16)[name = string("op_1087_cast_fp16")];
            fp32 var_1088_epsilon_0 = const()[name = string("op_1088_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 2, 1]> var_1088_cast_fp16 = rsqrt(epsilon = var_1088_epsilon_0, x = var_1087_cast_fp16)[name = string("op_1088_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_1089_cast_fp16 = mul(x = q_15_cast_fp16, y = var_1088_cast_fp16)[name = string("op_1089_cast_fp16")];
            fp16 var_1090_to_fp16 = const()[name = string("op_1090_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 32, 2, 128]> q_17_cast_fp16 = mul(x = var_1089_cast_fp16, y = var_1090_to_fp16)[name = string("q_17_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_1092_cast_fp16 = mul(x = k_15_cast_fp16, y = k_15_cast_fp16)[name = string("op_1092_cast_fp16")];
            tensor<int32, [1]> var_1097_axes_0 = const()[name = string("op_1097_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1097_keep_dims_0 = const()[name = string("op_1097_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 2, 1]> var_1097_cast_fp16 = reduce_mean(axes = var_1097_axes_0, keep_dims = var_1097_keep_dims_0, x = var_1092_cast_fp16)[name = string("op_1097_cast_fp16")];
            fp16 var_1098_promoted_to_fp16 = const()[name = string("op_1098_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 32, 2, 1]> var_1099_cast_fp16 = mul(x = var_1097_cast_fp16, y = var_1098_promoted_to_fp16)[name = string("op_1099_cast_fp16")];
            fp16 var_1101_to_fp16 = const()[name = string("op_1101_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 32, 2, 1]> var_1102_cast_fp16 = add(x = var_1099_cast_fp16, y = var_1101_to_fp16)[name = string("op_1102_cast_fp16")];
            fp32 var_1103_epsilon_0 = const()[name = string("op_1103_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 2, 1]> var_1103_cast_fp16 = rsqrt(epsilon = var_1103_epsilon_0, x = var_1102_cast_fp16)[name = string("op_1103_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> k_17_cast_fp16 = mul(x = k_15_cast_fp16, y = var_1103_cast_fp16)[name = string("k_17_cast_fp16")];
            tensor<int32, [4]> var_1109 = const()[name = string("op_1109"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1122_axes_0 = const()[name = string("op_1122_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 128]> var_1110_cast_fp16 = transpose(perm = var_1109, x = q_17_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 2, 1, 32, 128]> var_1122_cast_fp16 = expand_dims(axes = var_1122_axes_0, x = var_1110_cast_fp16)[name = string("op_1122_cast_fp16")];
            tensor<int32, [5]> var_1130_reps_0 = const()[name = string("op_1130_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 32, 128]> var_1130_cast_fp16 = tile(reps = var_1130_reps_0, x = var_1122_cast_fp16)[name = string("op_1130_cast_fp16")];
            tensor<int32, [4]> var_1135 = const()[name = string("op_1135"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<fp16, [1, 4, 32, 128]> q_19_cast_fp16 = reshape(shape = var_1135, x = var_1130_cast_fp16)[name = string("q_19_cast_fp16")];
            tensor<int32, [4]> var_1141 = const()[name = string("op_1141"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1154_axes_0 = const()[name = string("op_1154_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 128]> var_1142_cast_fp16 = transpose(perm = var_1141, x = k_17_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 2, 1, 32, 128]> var_1154_cast_fp16 = expand_dims(axes = var_1154_axes_0, x = var_1142_cast_fp16)[name = string("op_1154_cast_fp16")];
            tensor<int32, [5]> var_1162_reps_0 = const()[name = string("op_1162_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 32, 128]> var_1162_cast_fp16 = tile(reps = var_1162_reps_0, x = var_1154_cast_fp16)[name = string("op_1162_cast_fp16")];
            tensor<int32, [4]> var_1167 = const()[name = string("op_1167"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<fp16, [1, 4, 32, 128]> k_19_cast_fp16 = reshape(shape = var_1167, x = var_1162_cast_fp16)[name = string("k_19_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_1169_cast_fp16 = sigmoid(x = b_3_cast_fp16)[name = string("op_1169_cast_fp16")];
            tensor<int32, [3]> var_1173 = const()[name = string("op_1173"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_1_dt_bias_to_fp16 = const()[name = string("layers_1_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 32, 4]> input_21_cast_fp16 = add(x = a_3_cast_fp16, y = layers_1_dt_bias_to_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_1177_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("op_1177_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_1178_cast_fp16 = abs(x = input_21_cast_fp16)[name = string("op_1178_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, 4]> var_1179_cast_fp16 = mul(x = var_1178_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_1179_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_1180_cast_fp16 = exp(x = var_1179_cast_fp16)[name = string("op_1180_cast_fp16")];
            fp16 var_1182_to_fp16 = const()[name = string("op_1182_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 32, 4]> var_1183_cast_fp16 = add(x = var_1180_cast_fp16, y = var_1182_to_fp16)[name = string("op_1183_cast_fp16")];
            fp32 var_1184_epsilon_0 = const()[name = string("op_1184_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 32, 4]> var_1184_cast_fp16 = log(epsilon = var_1184_epsilon_0, x = var_1183_cast_fp16)[name = string("op_1184_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_1186_cast_fp16 = add(x = var_1177_cast_fp16, y = var_1184_cast_fp16)[name = string("op_1186_cast_fp16")];
            tensor<fp16, [4]> layers_1_negA_to_fp16 = const()[name = string("layers_1_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.ff4p+2, -0x1.8cp+3, -0x1.c0cp+3, -0x1.104p+3])];
            tensor<fp16, [1, 32, 4]> var_1187_cast_fp16 = mul(x = layers_1_negA_to_fp16, y = var_1186_cast_fp16)[name = string("op_1187_cast_fp16")];
            tensor<int32, [3]> var_1191 = const()[name = string("op_1191"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_1194_axes_0 = const()[name = string("op_1194_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 32]> beta_3_cast_fp16 = transpose(perm = var_1173, x = var_1169_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 4, 32, 1]> var_1194_cast_fp16 = expand_dims(axes = var_1194_axes_0, x = beta_3_cast_fp16)[name = string("op_1194_cast_fp16")];
            fp16 var_1195_to_fp16 = const()[name = string("op_1195_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 32, 1]> var_1196_cast_fp16 = mul(x = var_1194_cast_fp16, y = var_1195_to_fp16)[name = string("op_1196_cast_fp16")];
            tensor<fp16, [1, 4, 32, 128]> v_3_cast_fp16 = transpose(perm = var_1071, x = var_1066_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 4, 32, 128]> vb_7_cast_fp16 = mul(x = v_3_cast_fp16, y = var_1196_cast_fp16)[name = string("vb_7_cast_fp16")];
            tensor<int32, [4]> q_21_begin_0 = const()[name = string("q_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_21_end_0 = const()[name = string("q_21_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_21_end_mask_0 = const()[name = string("q_21_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_21_cast_fp16 = slice_by_index(begin = q_21_begin_0, end = q_21_end_0, end_mask = q_21_end_mask_0, x = q_19_cast_fp16)[name = string("q_21_cast_fp16")];
            tensor<int32, [4]> k_21_begin_0 = const()[name = string("k_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_21_end_0 = const()[name = string("k_21_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_21_end_mask_0 = const()[name = string("k_21_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, x = k_19_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<fp16, [1, 4, 32, 128]> var_1230_cast_fp16 = mul(x = k_19_cast_fp16, y = var_1194_cast_fp16)[name = string("op_1230_cast_fp16")];
            tensor<int32, [4]> kb_5_begin_0 = const()[name = string("kb_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_5_end_0 = const()[name = string("kb_5_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_5_end_mask_0 = const()[name = string("kb_5_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_5_cast_fp16 = slice_by_index(begin = kb_5_begin_0, end = kb_5_end_0, end_mask = kb_5_end_mask_0, x = var_1230_cast_fp16)[name = string("kb_5_cast_fp16")];
            tensor<int32, [4]> vb_9_begin_0 = const()[name = string("vb_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_9_end_0 = const()[name = string("vb_9_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_9_end_mask_0 = const()[name = string("vb_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_9_cast_fp16 = slice_by_index(begin = vb_9_begin_0, end = vb_9_end_0, end_mask = vb_9_end_mask_0, x = vb_7_cast_fp16)[name = string("vb_9_cast_fp16")];
            tensor<int32, [3]> gdec_9_begin_0 = const()[name = string("gdec_9_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_9_end_0 = const()[name = string("gdec_9_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_9_end_mask_0 = const()[name = string("gdec_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 32]> gdec_7_cast_fp16 = transpose(perm = var_1191, x = var_1187_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 4, 16]> gdec_9_cast_fp16 = slice_by_index(begin = gdec_9_begin_0, end = gdec_9_end_0, end_mask = gdec_9_end_mask_0, x = gdec_7_cast_fp16)[name = string("gdec_9_cast_fp16")];
            tensor<int32, [2]> var_1278 = const()[name = string("op_1278"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_1279_cast_fp16 = reshape(shape = var_1278, x = gdec_9_cast_fp16)[name = string("op_1279_cast_fp16")];
            tensor<fp16, [4, 256]> var_1280_cast_fp16 = linear(bias = var_633_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1279_cast_fp16)[name = string("op_1280_cast_fp16")];
            tensor<int32, [4]> var_1285 = const()[name = string("op_1285"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_5_cast_fp16 = reshape(shape = var_1285, x = var_1280_cast_fp16)[name = string("P_5_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1287_cast_fp16 = exp(x = P_5_cast_fp16)[name = string("op_1287_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_5_cast_fp16 = mul(x = var_1287_cast_fp16, y = tril_incl_to_fp16)[name = string("D_5_cast_fp16")];
            bool kk_5_transpose_x_1 = const()[name = string("kk_5_transpose_x_1"), val = bool(false)];
            bool kk_5_transpose_y_1 = const()[name = string("kk_5_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_5_cast_fp16 = matmul(transpose_x = kk_5_transpose_x_1, transpose_y = kk_5_transpose_y_1, x = kb_5_cast_fp16, y = k_21_cast_fp16)[name = string("kk_5_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_1293_cast_fp16 = mul(x = kk_5_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_1293_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1294_cast_fp16 = mul(x = var_1293_cast_fp16, y = D_5_cast_fp16)[name = string("op_1294_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_5_cast_fp16 = mul(x = var_1294_cast_fp16, y = cum_tail_to_fp16)[name = string("N_5_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1296_cast_fp16 = mul(x = D_5_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1296_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_5_cast_fp16 = mul(x = kk_5_cast_fp16, y = var_1296_cast_fp16)[name = string("Na_5_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_17_cast_fp16 = add(x = eye_to_fp16, y = N_5_cast_fp16)[name = string("Minv_17_cast_fp16")];
            bool Np_13_transpose_x_0 = const()[name = string("Np_13_transpose_x_0"), val = bool(false)];
            bool Np_13_transpose_y_0 = const()[name = string("Np_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_13_cast_fp16 = matmul(transpose_x = Np_13_transpose_x_0, transpose_y = Np_13_transpose_y_0, x = N_5_cast_fp16, y = Na_5_cast_fp16)[name = string("Np_13_cast_fp16")];
            bool Npa_9_transpose_x_0 = const()[name = string("Npa_9_transpose_x_0"), val = bool(false)];
            bool Npa_9_transpose_y_0 = const()[name = string("Npa_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_9_cast_fp16 = matmul(transpose_x = Npa_9_transpose_x_0, transpose_y = Npa_9_transpose_y_0, x = Na_5_cast_fp16, y = N_5_cast_fp16)[name = string("Npa_9_cast_fp16")];
            bool var_1302_transpose_x_0 = const()[name = string("op_1302_transpose_x_0"), val = bool(false)];
            bool var_1302_transpose_y_0 = const()[name = string("op_1302_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1302_cast_fp16 = matmul(transpose_x = var_1302_transpose_x_0, transpose_y = var_1302_transpose_y_0, x = Minv_17_cast_fp16, y = Np_13_cast_fp16)[name = string("op_1302_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_19_cast_fp16 = add(x = Minv_17_cast_fp16, y = var_1302_cast_fp16)[name = string("Minv_19_cast_fp16")];
            bool Np_15_transpose_x_0 = const()[name = string("Np_15_transpose_x_0"), val = bool(false)];
            bool Np_15_transpose_y_0 = const()[name = string("Np_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_15_cast_fp16 = matmul(transpose_x = Np_15_transpose_x_0, transpose_y = Np_15_transpose_y_0, x = Np_13_cast_fp16, y = Npa_9_cast_fp16)[name = string("Np_15_cast_fp16")];
            bool Npa_11_transpose_x_0 = const()[name = string("Npa_11_transpose_x_0"), val = bool(false)];
            bool Npa_11_transpose_y_0 = const()[name = string("Npa_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_11_cast_fp16 = matmul(transpose_x = Npa_11_transpose_x_0, transpose_y = Npa_11_transpose_y_0, x = Npa_9_cast_fp16, y = Np_13_cast_fp16)[name = string("Npa_11_cast_fp16")];
            bool var_1307_transpose_x_0 = const()[name = string("op_1307_transpose_x_0"), val = bool(false)];
            bool var_1307_transpose_y_0 = const()[name = string("op_1307_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1307_cast_fp16 = matmul(transpose_x = var_1307_transpose_x_0, transpose_y = var_1307_transpose_y_0, x = Minv_19_cast_fp16, y = Np_15_cast_fp16)[name = string("op_1307_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_21_cast_fp16 = add(x = Minv_19_cast_fp16, y = var_1307_cast_fp16)[name = string("Minv_21_cast_fp16")];
            bool Np_17_transpose_x_0 = const()[name = string("Np_17_transpose_x_0"), val = bool(false)];
            bool Np_17_transpose_y_0 = const()[name = string("Np_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_17_cast_fp16 = matmul(transpose_x = Np_17_transpose_x_0, transpose_y = Np_17_transpose_y_0, x = Np_15_cast_fp16, y = Npa_11_cast_fp16)[name = string("Np_17_cast_fp16")];
            bool var_1311_transpose_x_0 = const()[name = string("op_1311_transpose_x_0"), val = bool(false)];
            bool var_1311_transpose_y_0 = const()[name = string("op_1311_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1311_cast_fp16 = matmul(transpose_x = var_1311_transpose_x_0, transpose_y = var_1311_transpose_y_0, x = Minv_21_cast_fp16, y = Np_17_cast_fp16)[name = string("op_1311_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_23_cast_fp16 = add(x = Minv_21_cast_fp16, y = var_1311_cast_fp16)[name = string("Minv_23_cast_fp16")];
            bool u_5_transpose_x_0 = const()[name = string("u_5_transpose_x_0"), val = bool(false)];
            bool u_5_transpose_y_0 = const()[name = string("u_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_5_cast_fp16 = matmul(transpose_x = u_5_transpose_x_0, transpose_y = u_5_transpose_y_0, x = Minv_23_cast_fp16, y = vb_9_cast_fp16)[name = string("u_5_cast_fp16")];
            bool var_1318_transpose_x_1 = const()[name = string("op_1318_transpose_x_1"), val = bool(false)];
            bool var_1318_transpose_y_1 = const()[name = string("op_1318_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_1318_cast_fp16 = matmul(transpose_x = var_1318_transpose_x_1, transpose_y = var_1318_transpose_y_1, x = q_21_cast_fp16, y = k_21_cast_fp16)[name = string("op_1318_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_5_cast_fp16 = mul(x = var_1318_cast_fp16, y = D_5_cast_fp16)[name = string("A_5_cast_fp16")];
            bool var_1320_transpose_x_0 = const()[name = string("op_1320_transpose_x_0"), val = bool(false)];
            bool var_1320_transpose_y_0 = const()[name = string("op_1320_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1320_cast_fp16 = matmul(transpose_x = var_1320_transpose_x_0, transpose_y = var_1320_transpose_y_0, x = A_5_cast_fp16, y = u_5_cast_fp16)[name = string("op_1320_cast_fp16")];
            tensor<fp16, [16]> var_1321_bias_0_to_fp16 = const()[name = string("op_1321_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_1321_cast_fp16 = linear(bias = var_1321_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_9_cast_fp16)[name = string("op_1321_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_3_cast_fp16 = exp(x = var_1321_cast_fp16)[name = string("tail_3_cast_fp16")];
            tensor<int32, [1]> var_1324_axes_0 = const()[name = string("op_1324_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_1324_cast_fp16 = expand_dims(axes = var_1324_axes_0, x = tail_3_cast_fp16)[name = string("op_1324_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1325_cast_fp16 = mul(x = k_21_cast_fp16, y = var_1324_cast_fp16)[name = string("op_1325_cast_fp16")];
            bool S_3_transpose_x_1 = const()[name = string("S_3_transpose_x_1"), val = bool(true)];
            bool S_3_transpose_y_1 = const()[name = string("S_3_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_3_cast_fp16 = matmul(transpose_x = S_3_transpose_x_1, transpose_y = S_3_transpose_y_1, x = var_1325_cast_fp16, y = u_5_cast_fp16)[name = string("S_3_cast_fp16")];
            tensor<int32, [4]> q_23_begin_0 = const()[name = string("q_23_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_23_end_0 = const()[name = string("q_23_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_23_end_mask_0 = const()[name = string("q_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_23_cast_fp16 = slice_by_index(begin = q_23_begin_0, end = q_23_end_0, end_mask = q_23_end_mask_0, x = q_19_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<int32, [4]> k_23_begin_0 = const()[name = string("k_23_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_23_end_0 = const()[name = string("k_23_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_23_end_mask_0 = const()[name = string("k_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_23_cast_fp16 = slice_by_index(begin = k_23_begin_0, end = k_23_end_0, end_mask = k_23_end_mask_0, x = k_19_cast_fp16)[name = string("k_23_cast_fp16")];
            tensor<int32, [4]> kb_7_begin_0 = const()[name = string("kb_7_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_7_end_0 = const()[name = string("kb_7_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_7_end_mask_0 = const()[name = string("kb_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_7_cast_fp16 = slice_by_index(begin = kb_7_begin_0, end = kb_7_end_0, end_mask = kb_7_end_mask_0, x = var_1230_cast_fp16)[name = string("kb_7_cast_fp16")];
            tensor<int32, [4]> vb_11_begin_0 = const()[name = string("vb_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_11_end_0 = const()[name = string("vb_11_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_11_end_mask_0 = const()[name = string("vb_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_11_cast_fp16 = slice_by_index(begin = vb_11_begin_0, end = vb_11_end_0, end_mask = vb_11_end_mask_0, x = vb_7_cast_fp16)[name = string("vb_11_cast_fp16")];
            tensor<int32, [3]> gdec_11_begin_0 = const()[name = string("gdec_11_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_11_end_0 = const()[name = string("gdec_11_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_11_end_mask_0 = const()[name = string("gdec_11_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_11_cast_fp16 = slice_by_index(begin = gdec_11_begin_0, end = gdec_11_end_0, end_mask = gdec_11_end_mask_0, x = gdec_7_cast_fp16)[name = string("gdec_11_cast_fp16")];
            tensor<int32, [2]> var_1410 = const()[name = string("op_1410"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_1411_cast_fp16 = reshape(shape = var_1410, x = gdec_11_cast_fp16)[name = string("op_1411_cast_fp16")];
            tensor<fp16, [4, 256]> var_1412_cast_fp16 = linear(bias = var_633_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1411_cast_fp16)[name = string("op_1412_cast_fp16")];
            tensor<int32, [4]> var_1417 = const()[name = string("op_1417"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_7_cast_fp16 = reshape(shape = var_1417, x = var_1412_cast_fp16)[name = string("P_7_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1419_cast_fp16 = exp(x = P_7_cast_fp16)[name = string("op_1419_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_7_cast_fp16 = mul(x = var_1419_cast_fp16, y = tril_incl_to_fp16)[name = string("D_7_cast_fp16")];
            bool kk_7_transpose_x_1 = const()[name = string("kk_7_transpose_x_1"), val = bool(false)];
            bool kk_7_transpose_y_1 = const()[name = string("kk_7_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_7_cast_fp16 = matmul(transpose_x = kk_7_transpose_x_1, transpose_y = kk_7_transpose_y_1, x = kb_7_cast_fp16, y = k_23_cast_fp16)[name = string("kk_7_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_1425_cast_fp16 = mul(x = kk_7_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_1425_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1426_cast_fp16 = mul(x = var_1425_cast_fp16, y = D_7_cast_fp16)[name = string("op_1426_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_7_cast_fp16 = mul(x = var_1426_cast_fp16, y = cum_tail_to_fp16)[name = string("N_7_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1428_cast_fp16 = mul(x = D_7_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1428_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_7_cast_fp16 = mul(x = kk_7_cast_fp16, y = var_1428_cast_fp16)[name = string("Na_7_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_25_cast_fp16 = add(x = eye_to_fp16, y = N_7_cast_fp16)[name = string("Minv_25_cast_fp16")];
            bool Np_19_transpose_x_0 = const()[name = string("Np_19_transpose_x_0"), val = bool(false)];
            bool Np_19_transpose_y_0 = const()[name = string("Np_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_19_cast_fp16 = matmul(transpose_x = Np_19_transpose_x_0, transpose_y = Np_19_transpose_y_0, x = N_7_cast_fp16, y = Na_7_cast_fp16)[name = string("Np_19_cast_fp16")];
            bool Npa_13_transpose_x_0 = const()[name = string("Npa_13_transpose_x_0"), val = bool(false)];
            bool Npa_13_transpose_y_0 = const()[name = string("Npa_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_13_cast_fp16 = matmul(transpose_x = Npa_13_transpose_x_0, transpose_y = Npa_13_transpose_y_0, x = Na_7_cast_fp16, y = N_7_cast_fp16)[name = string("Npa_13_cast_fp16")];
            bool var_1434_transpose_x_0 = const()[name = string("op_1434_transpose_x_0"), val = bool(false)];
            bool var_1434_transpose_y_0 = const()[name = string("op_1434_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1434_cast_fp16 = matmul(transpose_x = var_1434_transpose_x_0, transpose_y = var_1434_transpose_y_0, x = Minv_25_cast_fp16, y = Np_19_cast_fp16)[name = string("op_1434_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_27_cast_fp16 = add(x = Minv_25_cast_fp16, y = var_1434_cast_fp16)[name = string("Minv_27_cast_fp16")];
            bool Np_21_transpose_x_0 = const()[name = string("Np_21_transpose_x_0"), val = bool(false)];
            bool Np_21_transpose_y_0 = const()[name = string("Np_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_21_cast_fp16 = matmul(transpose_x = Np_21_transpose_x_0, transpose_y = Np_21_transpose_y_0, x = Np_19_cast_fp16, y = Npa_13_cast_fp16)[name = string("Np_21_cast_fp16")];
            bool Npa_15_transpose_x_0 = const()[name = string("Npa_15_transpose_x_0"), val = bool(false)];
            bool Npa_15_transpose_y_0 = const()[name = string("Npa_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_15_cast_fp16 = matmul(transpose_x = Npa_15_transpose_x_0, transpose_y = Npa_15_transpose_y_0, x = Npa_13_cast_fp16, y = Np_19_cast_fp16)[name = string("Npa_15_cast_fp16")];
            bool var_1439_transpose_x_0 = const()[name = string("op_1439_transpose_x_0"), val = bool(false)];
            bool var_1439_transpose_y_0 = const()[name = string("op_1439_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1439_cast_fp16 = matmul(transpose_x = var_1439_transpose_x_0, transpose_y = var_1439_transpose_y_0, x = Minv_27_cast_fp16, y = Np_21_cast_fp16)[name = string("op_1439_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_29_cast_fp16 = add(x = Minv_27_cast_fp16, y = var_1439_cast_fp16)[name = string("Minv_29_cast_fp16")];
            bool Np_23_transpose_x_0 = const()[name = string("Np_23_transpose_x_0"), val = bool(false)];
            bool Np_23_transpose_y_0 = const()[name = string("Np_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_23_cast_fp16 = matmul(transpose_x = Np_23_transpose_x_0, transpose_y = Np_23_transpose_y_0, x = Np_21_cast_fp16, y = Npa_15_cast_fp16)[name = string("Np_23_cast_fp16")];
            bool var_1443_transpose_x_0 = const()[name = string("op_1443_transpose_x_0"), val = bool(false)];
            bool var_1443_transpose_y_0 = const()[name = string("op_1443_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1443_cast_fp16 = matmul(transpose_x = var_1443_transpose_x_0, transpose_y = var_1443_transpose_y_0, x = Minv_29_cast_fp16, y = Np_23_cast_fp16)[name = string("op_1443_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_31_cast_fp16 = add(x = Minv_29_cast_fp16, y = var_1443_cast_fp16)[name = string("Minv_31_cast_fp16")];
            bool u_7_transpose_x_0 = const()[name = string("u_7_transpose_x_0"), val = bool(false)];
            bool u_7_transpose_y_0 = const()[name = string("u_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_7_cast_fp16 = matmul(transpose_x = u_7_transpose_x_0, transpose_y = u_7_transpose_y_0, x = Minv_31_cast_fp16, y = vb_11_cast_fp16)[name = string("u_7_cast_fp16")];
            bool var_1450_transpose_x_1 = const()[name = string("op_1450_transpose_x_1"), val = bool(false)];
            bool var_1450_transpose_y_1 = const()[name = string("op_1450_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_1450_cast_fp16 = matmul(transpose_x = var_1450_transpose_x_1, transpose_y = var_1450_transpose_y_1, x = q_23_cast_fp16, y = k_23_cast_fp16)[name = string("op_1450_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_7_cast_fp16 = mul(x = var_1450_cast_fp16, y = D_7_cast_fp16)[name = string("A_7_cast_fp16")];
            tensor<fp16, [16]> var_1452_bias_0_to_fp16 = const()[name = string("op_1452_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_1452_cast_fp16 = linear(bias = var_1452_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_11_cast_fp16)[name = string("op_1452_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_3_cast_fp16 = exp(x = var_1452_cast_fp16)[name = string("cum_3_cast_fp16")];
            tensor<int32, [1]> var_1455_axes_0 = const()[name = string("op_1455_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_1455_cast_fp16 = expand_dims(axes = var_1455_axes_0, x = cum_3_cast_fp16)[name = string("op_1455_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1456_cast_fp16 = mul(x = kb_7_cast_fp16, y = var_1455_cast_fp16)[name = string("op_1456_cast_fp16")];
            bool w_7_transpose_x_0 = const()[name = string("w_7_transpose_x_0"), val = bool(false)];
            bool w_7_transpose_y_0 = const()[name = string("w_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_7_cast_fp16 = matmul(transpose_x = w_7_transpose_x_0, transpose_y = w_7_transpose_y_0, x = Minv_31_cast_fp16, y = var_1456_cast_fp16)[name = string("w_7_cast_fp16")];
            bool var_1458_transpose_x_0 = const()[name = string("op_1458_transpose_x_0"), val = bool(false)];
            bool var_1458_transpose_y_0 = const()[name = string("op_1458_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1458_cast_fp16 = matmul(transpose_x = var_1458_transpose_x_0, transpose_y = var_1458_transpose_y_0, x = w_7_cast_fp16, y = S_3_cast_fp16)[name = string("op_1458_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_3_cast_fp16 = sub(x = u_7_cast_fp16, y = var_1458_cast_fp16)[name = string("vnew_3_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1463_cast_fp16 = mul(x = q_23_cast_fp16, y = var_1455_cast_fp16)[name = string("op_1463_cast_fp16")];
            bool var_1464_transpose_x_0 = const()[name = string("op_1464_transpose_x_0"), val = bool(false)];
            bool var_1464_transpose_y_0 = const()[name = string("op_1464_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1464_cast_fp16 = matmul(transpose_x = var_1464_transpose_x_0, transpose_y = var_1464_transpose_y_0, x = var_1463_cast_fp16, y = S_3_cast_fp16)[name = string("op_1464_cast_fp16")];
            bool var_1465_transpose_x_0 = const()[name = string("op_1465_transpose_x_0"), val = bool(false)];
            bool var_1465_transpose_y_0 = const()[name = string("op_1465_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1465_cast_fp16 = matmul(transpose_x = var_1465_transpose_x_0, transpose_y = var_1465_transpose_y_0, x = A_7_cast_fp16, y = vnew_3_cast_fp16)[name = string("op_1465_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_3_cast_fp16 = add(x = var_1464_cast_fp16, y = var_1465_cast_fp16)[name = string("oc_3_cast_fp16")];
            int32 var_1469 = const()[name = string("op_1469"), val = int32(2)];
            bool o_7_interleave_0 = const()[name = string("o_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 128]> o_7_cast_fp16 = concat(axis = var_1469, interleave = o_7_interleave_0, values = (var_1320_cast_fp16, oc_3_cast_fp16))[name = string("o_7_cast_fp16")];
            tensor<int32, [4]> var_1475 = const()[name = string("op_1475"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1481 = const()[name = string("op_1481"), val = tensor<int32, [4]>([1, 32, 4, 128])];
            tensor<fp16, [1, 32, 4, 128]> input_23_cast_fp16 = reshape(shape = var_1481, x = z_3_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> o_9_cast_fp16 = transpose(perm = var_1475, x = o_7_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 32, 4, 128]> var_1483_cast_fp16 = mul(x = o_9_cast_fp16, y = o_9_cast_fp16)[name = string("op_1483_cast_fp16")];
            tensor<int32, [1]> var_1488_axes_0 = const()[name = string("op_1488_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1488_keep_dims_0 = const()[name = string("op_1488_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 4, 1]> var_1488_cast_fp16 = reduce_mean(axes = var_1488_axes_0, keep_dims = var_1488_keep_dims_0, x = var_1483_cast_fp16)[name = string("op_1488_cast_fp16")];
            fp16 var_1490_to_fp16 = const()[name = string("op_1490_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 32, 4, 1]> var_1491_cast_fp16 = add(x = var_1488_cast_fp16, y = var_1490_to_fp16)[name = string("op_1491_cast_fp16")];
            fp32 var_1492_epsilon_0 = const()[name = string("op_1492_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 4, 1]> var_1492_cast_fp16 = rsqrt(epsilon = var_1492_epsilon_0, x = var_1491_cast_fp16)[name = string("op_1492_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> var_1493_cast_fp16 = mul(x = o_9_cast_fp16, y = var_1492_cast_fp16)[name = string("op_1493_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> var_1495_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_1495_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> o_11_cast_fp16 = mul(x = var_1493_cast_fp16, y = var_1495_cast_fp16)[name = string("o_11_cast_fp16")];
            tensor<int32, [3]> var_1500 = const()[name = string("op_1500"), val = tensor<int32, [3]>([1, 32, 512])];
            tensor<fp16, [1, 32, 512]> input_25_cast_fp16 = reshape(shape = var_1500, x = o_11_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [640, 512]> layers_1_out_weight_to_fp16 = const()[name = string("layers_1_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34984256)))];
            tensor<fp16, [1, 32, 640]> linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_out_weight_to_fp16, x = input_25_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 32, 640]> x_13_cast_fp16 = add(x = x_9_cast_fp16, y = linear_5_cast_fp16)[name = string("x_13_cast_fp16")];
            fp16 var_1513_to_fp16 = const()[name = string("op_1513_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> x_15_cast_fp16 = mul(x = x_13_cast_fp16, y = var_1513_to_fp16)[name = string("x_15_cast_fp16")];
            fp16 var_1511_promoted_to_fp16 = const()[name = string("op_1511_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> var_1515_cast_fp16 = pow(x = x_15_cast_fp16, y = var_1511_promoted_to_fp16)[name = string("op_1515_cast_fp16")];
            tensor<int32, [1]> var_1517_axes_0 = const()[name = string("op_1517_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1517_keep_dims_0 = const()[name = string("op_1517_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_1517_cast_fp16 = reduce_mean(axes = var_1517_axes_0, keep_dims = var_1517_keep_dims_0, x = var_1515_cast_fp16)[name = string("op_1517_cast_fp16")];
            fp16 var_1518_to_fp16 = const()[name = string("op_1518_to_fp16"), val = fp16(0x1.0cp-18)];
            tensor<fp16, [1, 32, 1]> var_1519_cast_fp16 = add(x = var_1517_cast_fp16, y = var_1518_to_fp16)[name = string("op_1519_cast_fp16")];
            fp32 var_1520_epsilon_0 = const()[name = string("op_1520_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_1520_cast_fp16 = rsqrt(epsilon = var_1520_epsilon_0, x = var_1519_cast_fp16)[name = string("op_1520_cast_fp16")];
            tensor<fp16, [1, 32, 640]> var_1521_cast_fp16 = mul(x = x_15_cast_fp16, y = var_1520_cast_fp16)[name = string("op_1521_cast_fp16")];
            tensor<fp16, [640]> layers_1_n2_w1_to_fp16 = const()[name = string("layers_1_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35639680)))];
            tensor<fp16, [1, 32, 640]> input_27_cast_fp16 = mul(x = var_1521_cast_fp16, y = layers_1_n2_w1_to_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_1_gate_up_weight_to_fp16 = const()[name = string("layers_1_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35641024)))];
            tensor<fp16, [1, 32, 5120]> linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_1_gate_up_weight_to_fp16, x = input_27_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<int32, [3]> input_29_begin_0 = const()[name = string("input_29_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_29_end_0 = const()[name = string("input_29_end_0"), val = tensor<int32, [3]>([1, 32, 2560])];
            tensor<bool, [3]> input_29_end_mask_0 = const()[name = string("input_29_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 2560]> input_29_cast_fp16 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = linear_6_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [1, 32, 2560]> var_1531_cast_fp16 = silu(x = input_29_cast_fp16)[name = string("op_1531_cast_fp16")];
            tensor<int32, [3]> var_1536_begin_0 = const()[name = string("op_1536_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_1536_end_0 = const()[name = string("op_1536_end_0"), val = tensor<int32, [3]>([1, 32, 5120])];
            tensor<bool, [3]> var_1536_end_mask_0 = const()[name = string("op_1536_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 2560]> var_1536_cast_fp16 = slice_by_index(begin = var_1536_begin_0, end = var_1536_end_0, end_mask = var_1536_end_mask_0, x = linear_6_cast_fp16)[name = string("op_1536_cast_fp16")];
            tensor<fp16, [1, 32, 2560]> input_31_cast_fp16 = mul(x = var_1531_cast_fp16, y = var_1536_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_1_down_weight_to_fp16 = const()[name = string("layers_1_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42194688)))];
            tensor<fp16, [1, 32, 640]> linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_down_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 32, 640]> x_17_cast_fp16 = add(x = x_13_cast_fp16, y = linear_7_cast_fp16)[name = string("x_17_cast_fp16")];
            fp16 var_1547_promoted_to_fp16 = const()[name = string("op_1547_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> var_1551_cast_fp16 = pow(x = x_17_cast_fp16, y = var_1547_promoted_to_fp16)[name = string("op_1551_cast_fp16")];
            tensor<int32, [1]> var_1553_axes_0 = const()[name = string("op_1553_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1553_keep_dims_0 = const()[name = string("op_1553_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_1553_cast_fp16 = reduce_mean(axes = var_1553_axes_0, keep_dims = var_1553_keep_dims_0, x = var_1551_cast_fp16)[name = string("op_1553_cast_fp16")];
            fp16 var_1554_to_fp16 = const()[name = string("op_1554_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_1555_cast_fp16 = add(x = var_1553_cast_fp16, y = var_1554_to_fp16)[name = string("op_1555_cast_fp16")];
            fp32 var_1556_epsilon_0 = const()[name = string("op_1556_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_1556_cast_fp16 = rsqrt(epsilon = var_1556_epsilon_0, x = var_1555_cast_fp16)[name = string("op_1556_cast_fp16")];
            tensor<fp16, [1, 32, 640]> var_1557_cast_fp16 = mul(x = x_17_cast_fp16, y = var_1556_cast_fp16)[name = string("op_1557_cast_fp16")];
            tensor<fp16, [640]> layers_2_n1_w1_to_fp16 = const()[name = string("layers_2_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45471552)))];
            tensor<fp16, [1, 32, 640]> input_33_cast_fp16 = mul(x = var_1557_cast_fp16, y = layers_2_n1_w1_to_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_2_qkvzba_weight_to_fp16 = const()[name = string("layers_2_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45472896)))];
            tensor<fp16, [1, 32, 1544]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_qkvzba_weight_to_fp16, x = input_33_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [3]> qkv_5_begin_0 = const()[name = string("qkv_5_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_5_end_0 = const()[name = string("qkv_5_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> qkv_5_end_mask_0 = const()[name = string("qkv_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 1024]> qkv_5_cast_fp16 = slice_by_index(begin = qkv_5_begin_0, end = qkv_5_end_0, end_mask = qkv_5_end_mask_0, x = linear_8_cast_fp16)[name = string("qkv_5_cast_fp16")];
            tensor<int32, [3]> z_5_begin_0 = const()[name = string("z_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_5_end_0 = const()[name = string("z_5_end_0"), val = tensor<int32, [3]>([1, 32, 1536])];
            tensor<bool, [3]> z_5_end_mask_0 = const()[name = string("z_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 512]> z_5_cast_fp16 = slice_by_index(begin = z_5_begin_0, end = z_5_end_0, end_mask = z_5_end_mask_0, x = linear_8_cast_fp16)[name = string("z_5_cast_fp16")];
            tensor<int32, [3]> b_5_begin_0 = const()[name = string("b_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_5_end_0 = const()[name = string("b_5_end_0"), val = tensor<int32, [3]>([1, 32, 1540])];
            tensor<bool, [3]> b_5_end_mask_0 = const()[name = string("b_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 4]> b_5_cast_fp16 = slice_by_index(begin = b_5_begin_0, end = b_5_end_0, end_mask = b_5_end_mask_0, x = linear_8_cast_fp16)[name = string("b_5_cast_fp16")];
            tensor<int32, [3]> a_5_begin_0 = const()[name = string("a_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_5_end_0 = const()[name = string("a_5_end_0"), val = tensor<int32, [3]>([1, 32, 1])];
            tensor<bool, [3]> a_5_end_mask_0 = const()[name = string("a_5_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 4]> a_5_cast_fp16 = slice_by_index(begin = a_5_begin_0, end = a_5_end_0, end_mask = a_5_end_mask_0, x = linear_8_cast_fp16)[name = string("a_5_cast_fp16")];
            int32 var_1596 = const()[name = string("op_1596"), val = int32(1)];
            bool xp_5_interleave_0 = const()[name = string("xp_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 35, 1024]> xp_5_cast_fp16 = concat(axis = var_1596, interleave = xp_5_interleave_0, values = (zpad_to_fp16, qkv_5_cast_fp16))[name = string("xp_5_cast_fp16")];
            tensor<int32, [3]> var_1607_begin_0 = const()[name = string("op_1607_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_1607_end_0 = const()[name = string("op_1607_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_1607_end_mask_0 = const()[name = string("op_1607_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 1024]> var_1607_cast_fp16 = slice_by_index(begin = var_1607_begin_0, end = var_1607_end_0, end_mask = var_1607_end_mask_0, x = xp_5_cast_fp16)[name = string("op_1607_cast_fp16")];
            tensor<fp16, [1024]> var_1615_to_fp16 = const()[name = string("op_1615_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47449280)))];
            tensor<fp16, [1, 32, 1024]> var_1616_cast_fp16 = mul(x = var_1607_cast_fp16, y = var_1615_to_fp16)[name = string("op_1616_cast_fp16")];
            tensor<int32, [3]> var_1626_begin_0 = const()[name = string("op_1626_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_1626_end_0 = const()[name = string("op_1626_end_0"), val = tensor<int32, [3]>([1, 34, 1024])];
            tensor<bool, [3]> var_1626_end_mask_0 = const()[name = string("op_1626_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_1626_cast_fp16 = slice_by_index(begin = var_1626_begin_0, end = var_1626_end_0, end_mask = var_1626_end_mask_0, x = xp_5_cast_fp16)[name = string("op_1626_cast_fp16")];
            tensor<fp16, [1024]> var_1634_to_fp16 = const()[name = string("op_1634_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47451392)))];
            tensor<fp16, [1, 32, 1024]> var_1635_cast_fp16 = mul(x = var_1626_cast_fp16, y = var_1634_to_fp16)[name = string("op_1635_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> var_1637_cast_fp16 = add(x = var_1616_cast_fp16, y = var_1635_cast_fp16)[name = string("op_1637_cast_fp16")];
            tensor<int32, [3]> var_1647_begin_0 = const()[name = string("op_1647_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1647_end_0 = const()[name = string("op_1647_end_0"), val = tensor<int32, [3]>([1, 33, 1024])];
            tensor<bool, [3]> var_1647_end_mask_0 = const()[name = string("op_1647_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_1647_cast_fp16 = slice_by_index(begin = var_1647_begin_0, end = var_1647_end_0, end_mask = var_1647_end_mask_0, x = xp_5_cast_fp16)[name = string("op_1647_cast_fp16")];
            tensor<fp16, [1024]> var_1655_to_fp16 = const()[name = string("op_1655_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47453504)))];
            tensor<fp16, [1, 32, 1024]> var_1656_cast_fp16 = mul(x = var_1647_cast_fp16, y = var_1655_to_fp16)[name = string("op_1656_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> var_1658_cast_fp16 = add(x = var_1637_cast_fp16, y = var_1656_cast_fp16)[name = string("op_1658_cast_fp16")];
            tensor<int32, [3]> var_1668_begin_0 = const()[name = string("op_1668_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1668_end_0 = const()[name = string("op_1668_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> var_1668_end_mask_0 = const()[name = string("op_1668_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_1668_cast_fp16 = slice_by_index(begin = var_1668_begin_0, end = var_1668_end_0, end_mask = var_1668_end_mask_0, x = xp_5_cast_fp16)[name = string("op_1668_cast_fp16")];
            tensor<fp16, [1024]> var_1676_to_fp16 = const()[name = string("op_1676_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47455616)))];
            tensor<fp16, [1, 32, 1024]> var_1677_cast_fp16 = mul(x = var_1668_cast_fp16, y = var_1676_to_fp16)[name = string("op_1677_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> input_35_cast_fp16 = add(x = var_1658_cast_fp16, y = var_1677_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> c_5_cast_fp16 = silu(x = input_35_cast_fp16)[name = string("c_5_cast_fp16")];
            tensor<int32, [3]> var_1685_begin_0 = const()[name = string("op_1685_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1685_end_0 = const()[name = string("op_1685_end_0"), val = tensor<int32, [3]>([1, 32, 256])];
            tensor<bool, [3]> var_1685_end_mask_0 = const()[name = string("op_1685_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 256]> var_1685_cast_fp16 = slice_by_index(begin = var_1685_begin_0, end = var_1685_end_0, end_mask = var_1685_end_mask_0, x = c_5_cast_fp16)[name = string("op_1685_cast_fp16")];
            tensor<int32, [4]> var_1690 = const()[name = string("op_1690"), val = tensor<int32, [4]>([1, 32, 2, 128])];
            tensor<fp16, [1, 32, 2, 128]> q_25_cast_fp16 = reshape(shape = var_1690, x = var_1685_cast_fp16)[name = string("q_25_cast_fp16")];
            tensor<int32, [3]> var_1696_begin_0 = const()[name = string("op_1696_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_1696_end_0 = const()[name = string("op_1696_end_0"), val = tensor<int32, [3]>([1, 32, 512])];
            tensor<bool, [3]> var_1696_end_mask_0 = const()[name = string("op_1696_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 256]> var_1696_cast_fp16 = slice_by_index(begin = var_1696_begin_0, end = var_1696_end_0, end_mask = var_1696_end_mask_0, x = c_5_cast_fp16)[name = string("op_1696_cast_fp16")];
            tensor<int32, [4]> var_1701 = const()[name = string("op_1701"), val = tensor<int32, [4]>([1, 32, 2, 128])];
            tensor<fp16, [1, 32, 2, 128]> k_25_cast_fp16 = reshape(shape = var_1701, x = var_1696_cast_fp16)[name = string("k_25_cast_fp16")];
            tensor<int32, [3]> var_1707_begin_0 = const()[name = string("op_1707_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_1707_end_0 = const()[name = string("op_1707_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> var_1707_end_mask_0 = const()[name = string("op_1707_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 512]> var_1707_cast_fp16 = slice_by_index(begin = var_1707_begin_0, end = var_1707_end_0, end_mask = var_1707_end_mask_0, x = c_5_cast_fp16)[name = string("op_1707_cast_fp16")];
            tensor<int32, [4]> var_1712 = const()[name = string("op_1712"), val = tensor<int32, [4]>([1, 32, 4, 128])];
            tensor<fp16, [1, 32, 4, 128]> var_1713_cast_fp16 = reshape(shape = var_1712, x = var_1707_cast_fp16)[name = string("op_1713_cast_fp16")];
            tensor<int32, [4]> var_1718 = const()[name = string("op_1718"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1720_to_fp16 = const()[name = string("op_1720_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 32, 2, 128]> q_27_cast_fp16 = mul(x = q_25_cast_fp16, y = var_1720_to_fp16)[name = string("q_27_cast_fp16")];
            fp16 var_1722_to_fp16 = const()[name = string("op_1722_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 32, 2, 128]> k_27_cast_fp16 = mul(x = k_25_cast_fp16, y = var_1722_to_fp16)[name = string("k_27_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_1724_cast_fp16 = mul(x = q_27_cast_fp16, y = q_27_cast_fp16)[name = string("op_1724_cast_fp16")];
            tensor<int32, [1]> var_1729_axes_0 = const()[name = string("op_1729_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1729_keep_dims_0 = const()[name = string("op_1729_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 2, 1]> var_1729_cast_fp16 = reduce_mean(axes = var_1729_axes_0, keep_dims = var_1729_keep_dims_0, x = var_1724_cast_fp16)[name = string("op_1729_cast_fp16")];
            fp16 var_1730_promoted_to_fp16 = const()[name = string("op_1730_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 32, 2, 1]> var_1731_cast_fp16 = mul(x = var_1729_cast_fp16, y = var_1730_promoted_to_fp16)[name = string("op_1731_cast_fp16")];
            fp16 var_1733_to_fp16 = const()[name = string("op_1733_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 32, 2, 1]> var_1734_cast_fp16 = add(x = var_1731_cast_fp16, y = var_1733_to_fp16)[name = string("op_1734_cast_fp16")];
            fp32 var_1735_epsilon_0 = const()[name = string("op_1735_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 2, 1]> var_1735_cast_fp16 = rsqrt(epsilon = var_1735_epsilon_0, x = var_1734_cast_fp16)[name = string("op_1735_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_1736_cast_fp16 = mul(x = q_27_cast_fp16, y = var_1735_cast_fp16)[name = string("op_1736_cast_fp16")];
            fp16 var_1737_to_fp16 = const()[name = string("op_1737_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 32, 2, 128]> q_29_cast_fp16 = mul(x = var_1736_cast_fp16, y = var_1737_to_fp16)[name = string("q_29_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_1739_cast_fp16 = mul(x = k_27_cast_fp16, y = k_27_cast_fp16)[name = string("op_1739_cast_fp16")];
            tensor<int32, [1]> var_1744_axes_0 = const()[name = string("op_1744_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1744_keep_dims_0 = const()[name = string("op_1744_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 2, 1]> var_1744_cast_fp16 = reduce_mean(axes = var_1744_axes_0, keep_dims = var_1744_keep_dims_0, x = var_1739_cast_fp16)[name = string("op_1744_cast_fp16")];
            fp16 var_1745_promoted_to_fp16 = const()[name = string("op_1745_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 32, 2, 1]> var_1746_cast_fp16 = mul(x = var_1744_cast_fp16, y = var_1745_promoted_to_fp16)[name = string("op_1746_cast_fp16")];
            fp16 var_1748_to_fp16 = const()[name = string("op_1748_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 32, 2, 1]> var_1749_cast_fp16 = add(x = var_1746_cast_fp16, y = var_1748_to_fp16)[name = string("op_1749_cast_fp16")];
            fp32 var_1750_epsilon_0 = const()[name = string("op_1750_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 2, 1]> var_1750_cast_fp16 = rsqrt(epsilon = var_1750_epsilon_0, x = var_1749_cast_fp16)[name = string("op_1750_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> k_29_cast_fp16 = mul(x = k_27_cast_fp16, y = var_1750_cast_fp16)[name = string("k_29_cast_fp16")];
            tensor<int32, [4]> var_1756 = const()[name = string("op_1756"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1769_axes_0 = const()[name = string("op_1769_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 128]> var_1757_cast_fp16 = transpose(perm = var_1756, x = q_29_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 2, 1, 32, 128]> var_1769_cast_fp16 = expand_dims(axes = var_1769_axes_0, x = var_1757_cast_fp16)[name = string("op_1769_cast_fp16")];
            tensor<int32, [5]> var_1777_reps_0 = const()[name = string("op_1777_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 32, 128]> var_1777_cast_fp16 = tile(reps = var_1777_reps_0, x = var_1769_cast_fp16)[name = string("op_1777_cast_fp16")];
            tensor<int32, [4]> var_1782 = const()[name = string("op_1782"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<fp16, [1, 4, 32, 128]> q_31_cast_fp16 = reshape(shape = var_1782, x = var_1777_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<int32, [4]> var_1788 = const()[name = string("op_1788"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1801_axes_0 = const()[name = string("op_1801_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 128]> var_1789_cast_fp16 = transpose(perm = var_1788, x = k_29_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 2, 1, 32, 128]> var_1801_cast_fp16 = expand_dims(axes = var_1801_axes_0, x = var_1789_cast_fp16)[name = string("op_1801_cast_fp16")];
            tensor<int32, [5]> var_1809_reps_0 = const()[name = string("op_1809_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 32, 128]> var_1809_cast_fp16 = tile(reps = var_1809_reps_0, x = var_1801_cast_fp16)[name = string("op_1809_cast_fp16")];
            tensor<int32, [4]> var_1814 = const()[name = string("op_1814"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<fp16, [1, 4, 32, 128]> k_31_cast_fp16 = reshape(shape = var_1814, x = var_1809_cast_fp16)[name = string("k_31_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_1816_cast_fp16 = sigmoid(x = b_5_cast_fp16)[name = string("op_1816_cast_fp16")];
            tensor<int32, [3]> var_1820 = const()[name = string("op_1820"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_2_dt_bias_to_fp16 = const()[name = string("layers_2_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 32, 4]> input_37_cast_fp16 = add(x = a_5_cast_fp16, y = layers_2_dt_bias_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_1824_cast_fp16 = relu(x = input_37_cast_fp16)[name = string("op_1824_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_1825_cast_fp16 = abs(x = input_37_cast_fp16)[name = string("op_1825_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, 4]> var_1826_cast_fp16 = mul(x = var_1825_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_1826_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_1827_cast_fp16 = exp(x = var_1826_cast_fp16)[name = string("op_1827_cast_fp16")];
            fp16 var_1829_to_fp16 = const()[name = string("op_1829_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 32, 4]> var_1830_cast_fp16 = add(x = var_1827_cast_fp16, y = var_1829_to_fp16)[name = string("op_1830_cast_fp16")];
            fp32 var_1831_epsilon_0 = const()[name = string("op_1831_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 32, 4]> var_1831_cast_fp16 = log(epsilon = var_1831_epsilon_0, x = var_1830_cast_fp16)[name = string("op_1831_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_1833_cast_fp16 = add(x = var_1824_cast_fp16, y = var_1831_cast_fp16)[name = string("op_1833_cast_fp16")];
            tensor<fp16, [4]> layers_2_negA_to_fp16 = const()[name = string("layers_2_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.c0cp+3, -0x1.3e4p+3, -0x1.568p+1, -0x1.9acp+2])];
            tensor<fp16, [1, 32, 4]> var_1834_cast_fp16 = mul(x = layers_2_negA_to_fp16, y = var_1833_cast_fp16)[name = string("op_1834_cast_fp16")];
            tensor<int32, [3]> var_1838 = const()[name = string("op_1838"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_1841_axes_0 = const()[name = string("op_1841_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 32]> beta_5_cast_fp16 = transpose(perm = var_1820, x = var_1816_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 4, 32, 1]> var_1841_cast_fp16 = expand_dims(axes = var_1841_axes_0, x = beta_5_cast_fp16)[name = string("op_1841_cast_fp16")];
            fp16 var_1842_to_fp16 = const()[name = string("op_1842_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 32, 1]> var_1843_cast_fp16 = mul(x = var_1841_cast_fp16, y = var_1842_to_fp16)[name = string("op_1843_cast_fp16")];
            tensor<fp16, [1, 4, 32, 128]> v_5_cast_fp16 = transpose(perm = var_1718, x = var_1713_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 4, 32, 128]> vb_13_cast_fp16 = mul(x = v_5_cast_fp16, y = var_1843_cast_fp16)[name = string("vb_13_cast_fp16")];
            tensor<int32, [4]> q_33_begin_0 = const()[name = string("q_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_33_end_0 = const()[name = string("q_33_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_33_end_mask_0 = const()[name = string("q_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_33_cast_fp16 = slice_by_index(begin = q_33_begin_0, end = q_33_end_0, end_mask = q_33_end_mask_0, x = q_31_cast_fp16)[name = string("q_33_cast_fp16")];
            tensor<int32, [4]> k_33_begin_0 = const()[name = string("k_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_33_end_0 = const()[name = string("k_33_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_33_end_mask_0 = const()[name = string("k_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, x = k_31_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<fp16, [1, 4, 32, 128]> var_1877_cast_fp16 = mul(x = k_31_cast_fp16, y = var_1841_cast_fp16)[name = string("op_1877_cast_fp16")];
            tensor<int32, [4]> kb_9_begin_0 = const()[name = string("kb_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_9_end_0 = const()[name = string("kb_9_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_9_end_mask_0 = const()[name = string("kb_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_9_cast_fp16 = slice_by_index(begin = kb_9_begin_0, end = kb_9_end_0, end_mask = kb_9_end_mask_0, x = var_1877_cast_fp16)[name = string("kb_9_cast_fp16")];
            tensor<int32, [4]> vb_15_begin_0 = const()[name = string("vb_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_15_end_0 = const()[name = string("vb_15_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_15_end_mask_0 = const()[name = string("vb_15_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_15_cast_fp16 = slice_by_index(begin = vb_15_begin_0, end = vb_15_end_0, end_mask = vb_15_end_mask_0, x = vb_13_cast_fp16)[name = string("vb_15_cast_fp16")];
            tensor<int32, [3]> gdec_15_begin_0 = const()[name = string("gdec_15_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_15_end_0 = const()[name = string("gdec_15_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_15_end_mask_0 = const()[name = string("gdec_15_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 32]> gdec_13_cast_fp16 = transpose(perm = var_1838, x = var_1834_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 4, 16]> gdec_15_cast_fp16 = slice_by_index(begin = gdec_15_begin_0, end = gdec_15_end_0, end_mask = gdec_15_end_mask_0, x = gdec_13_cast_fp16)[name = string("gdec_15_cast_fp16")];
            tensor<int32, [2]> var_1925 = const()[name = string("op_1925"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_1926_cast_fp16 = reshape(shape = var_1925, x = gdec_15_cast_fp16)[name = string("op_1926_cast_fp16")];
            tensor<fp16, [4, 256]> var_1927_cast_fp16 = linear(bias = var_633_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1926_cast_fp16)[name = string("op_1927_cast_fp16")];
            tensor<int32, [4]> var_1932 = const()[name = string("op_1932"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_9_cast_fp16 = reshape(shape = var_1932, x = var_1927_cast_fp16)[name = string("P_9_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1934_cast_fp16 = exp(x = P_9_cast_fp16)[name = string("op_1934_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_9_cast_fp16 = mul(x = var_1934_cast_fp16, y = tril_incl_to_fp16)[name = string("D_9_cast_fp16")];
            bool kk_9_transpose_x_1 = const()[name = string("kk_9_transpose_x_1"), val = bool(false)];
            bool kk_9_transpose_y_1 = const()[name = string("kk_9_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_9_cast_fp16 = matmul(transpose_x = kk_9_transpose_x_1, transpose_y = kk_9_transpose_y_1, x = kb_9_cast_fp16, y = k_33_cast_fp16)[name = string("kk_9_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_1940_cast_fp16 = mul(x = kk_9_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_1940_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1941_cast_fp16 = mul(x = var_1940_cast_fp16, y = D_9_cast_fp16)[name = string("op_1941_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_9_cast_fp16 = mul(x = var_1941_cast_fp16, y = cum_tail_to_fp16)[name = string("N_9_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1943_cast_fp16 = mul(x = D_9_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1943_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_9_cast_fp16 = mul(x = kk_9_cast_fp16, y = var_1943_cast_fp16)[name = string("Na_9_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_33_cast_fp16 = add(x = eye_to_fp16, y = N_9_cast_fp16)[name = string("Minv_33_cast_fp16")];
            bool Np_25_transpose_x_0 = const()[name = string("Np_25_transpose_x_0"), val = bool(false)];
            bool Np_25_transpose_y_0 = const()[name = string("Np_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_25_cast_fp16 = matmul(transpose_x = Np_25_transpose_x_0, transpose_y = Np_25_transpose_y_0, x = N_9_cast_fp16, y = Na_9_cast_fp16)[name = string("Np_25_cast_fp16")];
            bool Npa_17_transpose_x_0 = const()[name = string("Npa_17_transpose_x_0"), val = bool(false)];
            bool Npa_17_transpose_y_0 = const()[name = string("Npa_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_17_cast_fp16 = matmul(transpose_x = Npa_17_transpose_x_0, transpose_y = Npa_17_transpose_y_0, x = Na_9_cast_fp16, y = N_9_cast_fp16)[name = string("Npa_17_cast_fp16")];
            bool var_1949_transpose_x_0 = const()[name = string("op_1949_transpose_x_0"), val = bool(false)];
            bool var_1949_transpose_y_0 = const()[name = string("op_1949_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1949_cast_fp16 = matmul(transpose_x = var_1949_transpose_x_0, transpose_y = var_1949_transpose_y_0, x = Minv_33_cast_fp16, y = Np_25_cast_fp16)[name = string("op_1949_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_35_cast_fp16 = add(x = Minv_33_cast_fp16, y = var_1949_cast_fp16)[name = string("Minv_35_cast_fp16")];
            bool Np_27_transpose_x_0 = const()[name = string("Np_27_transpose_x_0"), val = bool(false)];
            bool Np_27_transpose_y_0 = const()[name = string("Np_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_27_cast_fp16 = matmul(transpose_x = Np_27_transpose_x_0, transpose_y = Np_27_transpose_y_0, x = Np_25_cast_fp16, y = Npa_17_cast_fp16)[name = string("Np_27_cast_fp16")];
            bool Npa_19_transpose_x_0 = const()[name = string("Npa_19_transpose_x_0"), val = bool(false)];
            bool Npa_19_transpose_y_0 = const()[name = string("Npa_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_19_cast_fp16 = matmul(transpose_x = Npa_19_transpose_x_0, transpose_y = Npa_19_transpose_y_0, x = Npa_17_cast_fp16, y = Np_25_cast_fp16)[name = string("Npa_19_cast_fp16")];
            bool var_1954_transpose_x_0 = const()[name = string("op_1954_transpose_x_0"), val = bool(false)];
            bool var_1954_transpose_y_0 = const()[name = string("op_1954_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1954_cast_fp16 = matmul(transpose_x = var_1954_transpose_x_0, transpose_y = var_1954_transpose_y_0, x = Minv_35_cast_fp16, y = Np_27_cast_fp16)[name = string("op_1954_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_37_cast_fp16 = add(x = Minv_35_cast_fp16, y = var_1954_cast_fp16)[name = string("Minv_37_cast_fp16")];
            bool Np_29_transpose_x_0 = const()[name = string("Np_29_transpose_x_0"), val = bool(false)];
            bool Np_29_transpose_y_0 = const()[name = string("Np_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_29_cast_fp16 = matmul(transpose_x = Np_29_transpose_x_0, transpose_y = Np_29_transpose_y_0, x = Np_27_cast_fp16, y = Npa_19_cast_fp16)[name = string("Np_29_cast_fp16")];
            bool var_1958_transpose_x_0 = const()[name = string("op_1958_transpose_x_0"), val = bool(false)];
            bool var_1958_transpose_y_0 = const()[name = string("op_1958_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1958_cast_fp16 = matmul(transpose_x = var_1958_transpose_x_0, transpose_y = var_1958_transpose_y_0, x = Minv_37_cast_fp16, y = Np_29_cast_fp16)[name = string("op_1958_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_39_cast_fp16 = add(x = Minv_37_cast_fp16, y = var_1958_cast_fp16)[name = string("Minv_39_cast_fp16")];
            bool u_9_transpose_x_0 = const()[name = string("u_9_transpose_x_0"), val = bool(false)];
            bool u_9_transpose_y_0 = const()[name = string("u_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_9_cast_fp16 = matmul(transpose_x = u_9_transpose_x_0, transpose_y = u_9_transpose_y_0, x = Minv_39_cast_fp16, y = vb_15_cast_fp16)[name = string("u_9_cast_fp16")];
            bool var_1965_transpose_x_1 = const()[name = string("op_1965_transpose_x_1"), val = bool(false)];
            bool var_1965_transpose_y_1 = const()[name = string("op_1965_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_1965_cast_fp16 = matmul(transpose_x = var_1965_transpose_x_1, transpose_y = var_1965_transpose_y_1, x = q_33_cast_fp16, y = k_33_cast_fp16)[name = string("op_1965_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_9_cast_fp16 = mul(x = var_1965_cast_fp16, y = D_9_cast_fp16)[name = string("A_9_cast_fp16")];
            bool var_1967_transpose_x_0 = const()[name = string("op_1967_transpose_x_0"), val = bool(false)];
            bool var_1967_transpose_y_0 = const()[name = string("op_1967_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1967_cast_fp16 = matmul(transpose_x = var_1967_transpose_x_0, transpose_y = var_1967_transpose_y_0, x = A_9_cast_fp16, y = u_9_cast_fp16)[name = string("op_1967_cast_fp16")];
            tensor<fp16, [16]> var_1968_bias_0_to_fp16 = const()[name = string("op_1968_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_1968_cast_fp16 = linear(bias = var_1968_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_15_cast_fp16)[name = string("op_1968_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_5_cast_fp16 = exp(x = var_1968_cast_fp16)[name = string("tail_5_cast_fp16")];
            tensor<int32, [1]> var_1971_axes_0 = const()[name = string("op_1971_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_1971_cast_fp16 = expand_dims(axes = var_1971_axes_0, x = tail_5_cast_fp16)[name = string("op_1971_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1972_cast_fp16 = mul(x = k_33_cast_fp16, y = var_1971_cast_fp16)[name = string("op_1972_cast_fp16")];
            bool S_5_transpose_x_1 = const()[name = string("S_5_transpose_x_1"), val = bool(true)];
            bool S_5_transpose_y_1 = const()[name = string("S_5_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_5_cast_fp16 = matmul(transpose_x = S_5_transpose_x_1, transpose_y = S_5_transpose_y_1, x = var_1972_cast_fp16, y = u_9_cast_fp16)[name = string("S_5_cast_fp16")];
            tensor<int32, [4]> q_35_begin_0 = const()[name = string("q_35_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_35_end_0 = const()[name = string("q_35_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_35_end_mask_0 = const()[name = string("q_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_35_cast_fp16 = slice_by_index(begin = q_35_begin_0, end = q_35_end_0, end_mask = q_35_end_mask_0, x = q_31_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<int32, [4]> k_35_begin_0 = const()[name = string("k_35_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_35_end_0 = const()[name = string("k_35_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_35_end_mask_0 = const()[name = string("k_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_35_cast_fp16 = slice_by_index(begin = k_35_begin_0, end = k_35_end_0, end_mask = k_35_end_mask_0, x = k_31_cast_fp16)[name = string("k_35_cast_fp16")];
            tensor<int32, [4]> kb_11_begin_0 = const()[name = string("kb_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_11_end_0 = const()[name = string("kb_11_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_11_end_mask_0 = const()[name = string("kb_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_11_cast_fp16 = slice_by_index(begin = kb_11_begin_0, end = kb_11_end_0, end_mask = kb_11_end_mask_0, x = var_1877_cast_fp16)[name = string("kb_11_cast_fp16")];
            tensor<int32, [4]> vb_17_begin_0 = const()[name = string("vb_17_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_17_end_0 = const()[name = string("vb_17_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_17_end_mask_0 = const()[name = string("vb_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_17_cast_fp16 = slice_by_index(begin = vb_17_begin_0, end = vb_17_end_0, end_mask = vb_17_end_mask_0, x = vb_13_cast_fp16)[name = string("vb_17_cast_fp16")];
            tensor<int32, [3]> gdec_17_begin_0 = const()[name = string("gdec_17_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_17_end_0 = const()[name = string("gdec_17_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_17_end_mask_0 = const()[name = string("gdec_17_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_17_cast_fp16 = slice_by_index(begin = gdec_17_begin_0, end = gdec_17_end_0, end_mask = gdec_17_end_mask_0, x = gdec_13_cast_fp16)[name = string("gdec_17_cast_fp16")];
            tensor<int32, [2]> var_2057 = const()[name = string("op_2057"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_2058_cast_fp16 = reshape(shape = var_2057, x = gdec_17_cast_fp16)[name = string("op_2058_cast_fp16")];
            tensor<fp16, [4, 256]> var_2059_cast_fp16 = linear(bias = var_633_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2058_cast_fp16)[name = string("op_2059_cast_fp16")];
            tensor<int32, [4]> var_2064 = const()[name = string("op_2064"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_11_cast_fp16 = reshape(shape = var_2064, x = var_2059_cast_fp16)[name = string("P_11_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2066_cast_fp16 = exp(x = P_11_cast_fp16)[name = string("op_2066_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_11_cast_fp16 = mul(x = var_2066_cast_fp16, y = tril_incl_to_fp16)[name = string("D_11_cast_fp16")];
            bool kk_11_transpose_x_1 = const()[name = string("kk_11_transpose_x_1"), val = bool(false)];
            bool kk_11_transpose_y_1 = const()[name = string("kk_11_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_11_cast_fp16 = matmul(transpose_x = kk_11_transpose_x_1, transpose_y = kk_11_transpose_y_1, x = kb_11_cast_fp16, y = k_35_cast_fp16)[name = string("kk_11_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_2072_cast_fp16 = mul(x = kk_11_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_2072_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2073_cast_fp16 = mul(x = var_2072_cast_fp16, y = D_11_cast_fp16)[name = string("op_2073_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_11_cast_fp16 = mul(x = var_2073_cast_fp16, y = cum_tail_to_fp16)[name = string("N_11_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2075_cast_fp16 = mul(x = D_11_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2075_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_11_cast_fp16 = mul(x = kk_11_cast_fp16, y = var_2075_cast_fp16)[name = string("Na_11_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_41_cast_fp16 = add(x = eye_to_fp16, y = N_11_cast_fp16)[name = string("Minv_41_cast_fp16")];
            bool Np_31_transpose_x_0 = const()[name = string("Np_31_transpose_x_0"), val = bool(false)];
            bool Np_31_transpose_y_0 = const()[name = string("Np_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_31_cast_fp16 = matmul(transpose_x = Np_31_transpose_x_0, transpose_y = Np_31_transpose_y_0, x = N_11_cast_fp16, y = Na_11_cast_fp16)[name = string("Np_31_cast_fp16")];
            bool Npa_21_transpose_x_0 = const()[name = string("Npa_21_transpose_x_0"), val = bool(false)];
            bool Npa_21_transpose_y_0 = const()[name = string("Npa_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_21_cast_fp16 = matmul(transpose_x = Npa_21_transpose_x_0, transpose_y = Npa_21_transpose_y_0, x = Na_11_cast_fp16, y = N_11_cast_fp16)[name = string("Npa_21_cast_fp16")];
            bool var_2081_transpose_x_0 = const()[name = string("op_2081_transpose_x_0"), val = bool(false)];
            bool var_2081_transpose_y_0 = const()[name = string("op_2081_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2081_cast_fp16 = matmul(transpose_x = var_2081_transpose_x_0, transpose_y = var_2081_transpose_y_0, x = Minv_41_cast_fp16, y = Np_31_cast_fp16)[name = string("op_2081_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_43_cast_fp16 = add(x = Minv_41_cast_fp16, y = var_2081_cast_fp16)[name = string("Minv_43_cast_fp16")];
            bool Np_33_transpose_x_0 = const()[name = string("Np_33_transpose_x_0"), val = bool(false)];
            bool Np_33_transpose_y_0 = const()[name = string("Np_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_33_cast_fp16 = matmul(transpose_x = Np_33_transpose_x_0, transpose_y = Np_33_transpose_y_0, x = Np_31_cast_fp16, y = Npa_21_cast_fp16)[name = string("Np_33_cast_fp16")];
            bool Npa_23_transpose_x_0 = const()[name = string("Npa_23_transpose_x_0"), val = bool(false)];
            bool Npa_23_transpose_y_0 = const()[name = string("Npa_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_23_cast_fp16 = matmul(transpose_x = Npa_23_transpose_x_0, transpose_y = Npa_23_transpose_y_0, x = Npa_21_cast_fp16, y = Np_31_cast_fp16)[name = string("Npa_23_cast_fp16")];
            bool var_2086_transpose_x_0 = const()[name = string("op_2086_transpose_x_0"), val = bool(false)];
            bool var_2086_transpose_y_0 = const()[name = string("op_2086_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2086_cast_fp16 = matmul(transpose_x = var_2086_transpose_x_0, transpose_y = var_2086_transpose_y_0, x = Minv_43_cast_fp16, y = Np_33_cast_fp16)[name = string("op_2086_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_45_cast_fp16 = add(x = Minv_43_cast_fp16, y = var_2086_cast_fp16)[name = string("Minv_45_cast_fp16")];
            bool Np_35_transpose_x_0 = const()[name = string("Np_35_transpose_x_0"), val = bool(false)];
            bool Np_35_transpose_y_0 = const()[name = string("Np_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_35_cast_fp16 = matmul(transpose_x = Np_35_transpose_x_0, transpose_y = Np_35_transpose_y_0, x = Np_33_cast_fp16, y = Npa_23_cast_fp16)[name = string("Np_35_cast_fp16")];
            bool var_2090_transpose_x_0 = const()[name = string("op_2090_transpose_x_0"), val = bool(false)];
            bool var_2090_transpose_y_0 = const()[name = string("op_2090_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2090_cast_fp16 = matmul(transpose_x = var_2090_transpose_x_0, transpose_y = var_2090_transpose_y_0, x = Minv_45_cast_fp16, y = Np_35_cast_fp16)[name = string("op_2090_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_47_cast_fp16 = add(x = Minv_45_cast_fp16, y = var_2090_cast_fp16)[name = string("Minv_47_cast_fp16")];
            bool u_11_transpose_x_0 = const()[name = string("u_11_transpose_x_0"), val = bool(false)];
            bool u_11_transpose_y_0 = const()[name = string("u_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_11_cast_fp16 = matmul(transpose_x = u_11_transpose_x_0, transpose_y = u_11_transpose_y_0, x = Minv_47_cast_fp16, y = vb_17_cast_fp16)[name = string("u_11_cast_fp16")];
            bool var_2097_transpose_x_1 = const()[name = string("op_2097_transpose_x_1"), val = bool(false)];
            bool var_2097_transpose_y_1 = const()[name = string("op_2097_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_2097_cast_fp16 = matmul(transpose_x = var_2097_transpose_x_1, transpose_y = var_2097_transpose_y_1, x = q_35_cast_fp16, y = k_35_cast_fp16)[name = string("op_2097_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_11_cast_fp16 = mul(x = var_2097_cast_fp16, y = D_11_cast_fp16)[name = string("A_11_cast_fp16")];
            tensor<fp16, [16]> var_2099_bias_0_to_fp16 = const()[name = string("op_2099_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_2099_cast_fp16 = linear(bias = var_2099_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_17_cast_fp16)[name = string("op_2099_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_5_cast_fp16 = exp(x = var_2099_cast_fp16)[name = string("cum_5_cast_fp16")];
            tensor<int32, [1]> var_2102_axes_0 = const()[name = string("op_2102_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_2102_cast_fp16 = expand_dims(axes = var_2102_axes_0, x = cum_5_cast_fp16)[name = string("op_2102_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2103_cast_fp16 = mul(x = kb_11_cast_fp16, y = var_2102_cast_fp16)[name = string("op_2103_cast_fp16")];
            bool w_11_transpose_x_0 = const()[name = string("w_11_transpose_x_0"), val = bool(false)];
            bool w_11_transpose_y_0 = const()[name = string("w_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_11_cast_fp16 = matmul(transpose_x = w_11_transpose_x_0, transpose_y = w_11_transpose_y_0, x = Minv_47_cast_fp16, y = var_2103_cast_fp16)[name = string("w_11_cast_fp16")];
            bool var_2105_transpose_x_0 = const()[name = string("op_2105_transpose_x_0"), val = bool(false)];
            bool var_2105_transpose_y_0 = const()[name = string("op_2105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2105_cast_fp16 = matmul(transpose_x = var_2105_transpose_x_0, transpose_y = var_2105_transpose_y_0, x = w_11_cast_fp16, y = S_5_cast_fp16)[name = string("op_2105_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_5_cast_fp16 = sub(x = u_11_cast_fp16, y = var_2105_cast_fp16)[name = string("vnew_5_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2110_cast_fp16 = mul(x = q_35_cast_fp16, y = var_2102_cast_fp16)[name = string("op_2110_cast_fp16")];
            bool var_2111_transpose_x_0 = const()[name = string("op_2111_transpose_x_0"), val = bool(false)];
            bool var_2111_transpose_y_0 = const()[name = string("op_2111_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2111_cast_fp16 = matmul(transpose_x = var_2111_transpose_x_0, transpose_y = var_2111_transpose_y_0, x = var_2110_cast_fp16, y = S_5_cast_fp16)[name = string("op_2111_cast_fp16")];
            bool var_2112_transpose_x_0 = const()[name = string("op_2112_transpose_x_0"), val = bool(false)];
            bool var_2112_transpose_y_0 = const()[name = string("op_2112_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2112_cast_fp16 = matmul(transpose_x = var_2112_transpose_x_0, transpose_y = var_2112_transpose_y_0, x = A_11_cast_fp16, y = vnew_5_cast_fp16)[name = string("op_2112_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_5_cast_fp16 = add(x = var_2111_cast_fp16, y = var_2112_cast_fp16)[name = string("oc_5_cast_fp16")];
            int32 var_2116 = const()[name = string("op_2116"), val = int32(2)];
            bool o_13_interleave_0 = const()[name = string("o_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 128]> o_13_cast_fp16 = concat(axis = var_2116, interleave = o_13_interleave_0, values = (var_1967_cast_fp16, oc_5_cast_fp16))[name = string("o_13_cast_fp16")];
            tensor<int32, [4]> var_2122 = const()[name = string("op_2122"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2128 = const()[name = string("op_2128"), val = tensor<int32, [4]>([1, 32, 4, 128])];
            tensor<fp16, [1, 32, 4, 128]> input_39_cast_fp16 = reshape(shape = var_2128, x = z_5_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> o_15_cast_fp16 = transpose(perm = var_2122, x = o_13_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 32, 4, 128]> var_2130_cast_fp16 = mul(x = o_15_cast_fp16, y = o_15_cast_fp16)[name = string("op_2130_cast_fp16")];
            tensor<int32, [1]> var_2135_axes_0 = const()[name = string("op_2135_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2135_keep_dims_0 = const()[name = string("op_2135_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 4, 1]> var_2135_cast_fp16 = reduce_mean(axes = var_2135_axes_0, keep_dims = var_2135_keep_dims_0, x = var_2130_cast_fp16)[name = string("op_2135_cast_fp16")];
            fp16 var_2137_to_fp16 = const()[name = string("op_2137_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 32, 4, 1]> var_2138_cast_fp16 = add(x = var_2135_cast_fp16, y = var_2137_to_fp16)[name = string("op_2138_cast_fp16")];
            fp32 var_2139_epsilon_0 = const()[name = string("op_2139_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 4, 1]> var_2139_cast_fp16 = rsqrt(epsilon = var_2139_epsilon_0, x = var_2138_cast_fp16)[name = string("op_2139_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> var_2140_cast_fp16 = mul(x = o_15_cast_fp16, y = var_2139_cast_fp16)[name = string("op_2140_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> var_2142_cast_fp16 = silu(x = input_39_cast_fp16)[name = string("op_2142_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> o_17_cast_fp16 = mul(x = var_2140_cast_fp16, y = var_2142_cast_fp16)[name = string("o_17_cast_fp16")];
            tensor<int32, [3]> var_2147 = const()[name = string("op_2147"), val = tensor<int32, [3]>([1, 32, 512])];
            tensor<fp16, [1, 32, 512]> input_41_cast_fp16 = reshape(shape = var_2147, x = o_17_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [640, 512]> layers_2_out_weight_to_fp16 = const()[name = string("layers_2_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47457728)))];
            tensor<fp16, [1, 32, 640]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_out_weight_to_fp16, x = input_41_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [1, 32, 640]> x_21_cast_fp16 = add(x = x_17_cast_fp16, y = linear_9_cast_fp16)[name = string("x_21_cast_fp16")];
            fp16 var_2158_promoted_to_fp16 = const()[name = string("op_2158_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> var_2162_cast_fp16 = pow(x = x_21_cast_fp16, y = var_2158_promoted_to_fp16)[name = string("op_2162_cast_fp16")];
            tensor<int32, [1]> var_2164_axes_0 = const()[name = string("op_2164_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2164_keep_dims_0 = const()[name = string("op_2164_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_2164_cast_fp16 = reduce_mean(axes = var_2164_axes_0, keep_dims = var_2164_keep_dims_0, x = var_2162_cast_fp16)[name = string("op_2164_cast_fp16")];
            fp16 var_2165_to_fp16 = const()[name = string("op_2165_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2166_cast_fp16 = add(x = var_2164_cast_fp16, y = var_2165_to_fp16)[name = string("op_2166_cast_fp16")];
            fp32 var_2167_epsilon_0 = const()[name = string("op_2167_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2167_cast_fp16 = rsqrt(epsilon = var_2167_epsilon_0, x = var_2166_cast_fp16)[name = string("op_2167_cast_fp16")];
            tensor<fp16, [1, 32, 640]> var_2168_cast_fp16 = mul(x = x_21_cast_fp16, y = var_2167_cast_fp16)[name = string("op_2168_cast_fp16")];
            tensor<fp16, [640]> layers_2_n2_w1_to_fp16 = const()[name = string("layers_2_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48113152)))];
            tensor<fp16, [1, 32, 640]> input_43_cast_fp16 = mul(x = var_2168_cast_fp16, y = layers_2_n2_w1_to_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_2_gate_up_weight_to_fp16 = const()[name = string("layers_2_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48114496)))];
            tensor<fp16, [1, 32, 5120]> linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_2_gate_up_weight_to_fp16, x = input_43_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [3]> input_45_begin_0 = const()[name = string("input_45_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_45_end_0 = const()[name = string("input_45_end_0"), val = tensor<int32, [3]>([1, 32, 2560])];
            tensor<bool, [3]> input_45_end_mask_0 = const()[name = string("input_45_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 2560]> input_45_cast_fp16 = slice_by_index(begin = input_45_begin_0, end = input_45_end_0, end_mask = input_45_end_mask_0, x = linear_10_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [1, 32, 2560]> var_2178_cast_fp16 = silu(x = input_45_cast_fp16)[name = string("op_2178_cast_fp16")];
            tensor<int32, [3]> var_2183_begin_0 = const()[name = string("op_2183_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_2183_end_0 = const()[name = string("op_2183_end_0"), val = tensor<int32, [3]>([1, 32, 5120])];
            tensor<bool, [3]> var_2183_end_mask_0 = const()[name = string("op_2183_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 2560]> var_2183_cast_fp16 = slice_by_index(begin = var_2183_begin_0, end = var_2183_end_0, end_mask = var_2183_end_mask_0, x = linear_10_cast_fp16)[name = string("op_2183_cast_fp16")];
            tensor<fp16, [1, 32, 2560]> input_47_cast_fp16 = mul(x = var_2178_cast_fp16, y = var_2183_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_2_down_weight_to_fp16 = const()[name = string("layers_2_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54668160)))];
            tensor<fp16, [1, 32, 640]> linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_down_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 32, 640]> x_25_cast_fp16 = add(x = x_21_cast_fp16, y = linear_11_cast_fp16)[name = string("x_25_cast_fp16")];
            fp16 var_2196_to_fp16 = const()[name = string("op_2196_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 32, 640]> x_27_cast_fp16 = mul(x = x_25_cast_fp16, y = var_2196_to_fp16)[name = string("x_27_cast_fp16")];
            fp16 var_2194_promoted_to_fp16 = const()[name = string("op_2194_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> var_2198_cast_fp16 = pow(x = x_27_cast_fp16, y = var_2194_promoted_to_fp16)[name = string("op_2198_cast_fp16")];
            tensor<int32, [1]> var_2200_axes_0 = const()[name = string("op_2200_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2200_keep_dims_0 = const()[name = string("op_2200_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_2200_cast_fp16 = reduce_mean(axes = var_2200_axes_0, keep_dims = var_2200_keep_dims_0, x = var_2198_cast_fp16)[name = string("op_2200_cast_fp16")];
            fp16 var_2201_to_fp16 = const()[name = string("op_2201_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 32, 1]> var_2202_cast_fp16 = add(x = var_2200_cast_fp16, y = var_2201_to_fp16)[name = string("op_2202_cast_fp16")];
            fp32 var_2203_epsilon_0 = const()[name = string("op_2203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2203_cast_fp16 = rsqrt(epsilon = var_2203_epsilon_0, x = var_2202_cast_fp16)[name = string("op_2203_cast_fp16")];
            tensor<fp16, [1, 32, 640]> var_2204_cast_fp16 = mul(x = x_27_cast_fp16, y = var_2203_cast_fp16)[name = string("op_2204_cast_fp16")];
            tensor<fp16, [640]> layers_3_n1_w1_to_fp16 = const()[name = string("layers_3_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57945024)))];
            tensor<fp16, [1, 32, 640]> input_49_cast_fp16 = mul(x = var_2204_cast_fp16, y = layers_3_n1_w1_to_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1280, 640]> layers_3_qkv_weight_to_fp16 = const()[name = string("layers_3_qkv_weight_to_fp16"), val = tensor<fp16, [1280, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57946368)))];
            tensor<fp16, [1280]> linear_12_bias_0_to_fp16 = const()[name = string("linear_12_bias_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59584832)))];
            tensor<fp16, [1, 32, 1280]> linear_12_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = layers_3_qkv_weight_to_fp16, x = input_49_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [3]> var_2213_begin_0 = const()[name = string("op_2213_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2213_end_0 = const()[name = string("op_2213_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> var_2213_end_mask_0 = const()[name = string("op_2213_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 1024]> var_2213_cast_fp16 = slice_by_index(begin = var_2213_begin_0, end = var_2213_end_0, end_mask = var_2213_end_mask_0, x = linear_12_cast_fp16)[name = string("op_2213_cast_fp16")];
            tensor<int32, [4]> var_2218 = const()[name = string("op_2218"), val = tensor<int32, [4]>([1, 32, 8, 128])];
            tensor<fp16, [1, 32, 8, 128]> qg_1_cast_fp16 = reshape(shape = var_2218, x = var_2213_cast_fp16)[name = string("qg_1_cast_fp16")];
            tensor<int32, [4]> x_29_begin_0 = const()[name = string("x_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_29_end_0 = const()[name = string("x_29_end_0"), val = tensor<int32, [4]>([1, 32, 8, 64])];
            tensor<bool, [4]> x_29_end_mask_0 = const()[name = string("x_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, 8, 64]> x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, x = qg_1_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [4]> var_2229_begin_0 = const()[name = string("op_2229_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> var_2229_end_0 = const()[name = string("op_2229_end_0"), val = tensor<int32, [4]>([1, 32, 8, 128])];
            tensor<bool, [4]> var_2229_end_mask_0 = const()[name = string("op_2229_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, 8, 64]> var_2229_cast_fp16 = slice_by_index(begin = var_2229_begin_0, end = var_2229_end_0, end_mask = var_2229_end_mask_0, x = qg_1_cast_fp16)[name = string("op_2229_cast_fp16")];
            tensor<int32, [3]> var_2233 = const()[name = string("op_2233"), val = tensor<int32, [3]>([1, 32, 512])];
            tensor<fp16, [1, 32, 512]> gate_1_cast_fp16 = reshape(shape = var_2233, x = var_2229_cast_fp16)[name = string("gate_1_cast_fp16")];
            tensor<int32, [3]> var_2239_begin_0 = const()[name = string("op_2239_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> var_2239_end_0 = const()[name = string("op_2239_end_0"), val = tensor<int32, [3]>([1, 32, 1152])];
            tensor<bool, [3]> var_2239_end_mask_0 = const()[name = string("op_2239_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 128]> var_2239_cast_fp16 = slice_by_index(begin = var_2239_begin_0, end = var_2239_end_0, end_mask = var_2239_end_mask_0, x = linear_12_cast_fp16)[name = string("op_2239_cast_fp16")];
            tensor<int32, [4]> var_2244 = const()[name = string("op_2244"), val = tensor<int32, [4]>([1, 32, 2, 64])];
            tensor<fp16, [1, 32, 2, 64]> x_35_cast_fp16 = reshape(shape = var_2244, x = var_2239_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<int32, [3]> var_2250_begin_0 = const()[name = string("op_2250_begin_0"), val = tensor<int32, [3]>([0, 0, 1152])];
            tensor<int32, [3]> var_2250_end_0 = const()[name = string("op_2250_end_0"), val = tensor<int32, [3]>([1, 32, 1280])];
            tensor<bool, [3]> var_2250_end_mask_0 = const()[name = string("op_2250_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 128]> var_2250_cast_fp16 = slice_by_index(begin = var_2250_begin_0, end = var_2250_end_0, end_mask = var_2250_end_mask_0, x = linear_12_cast_fp16)[name = string("op_2250_cast_fp16")];
            tensor<int32, [4]> var_2255 = const()[name = string("op_2255"), val = tensor<int32, [4]>([1, 32, 2, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_2256_cast_fp16 = reshape(shape = var_2255, x = var_2250_cast_fp16)[name = string("op_2256_cast_fp16")];
            tensor<int32, [4]> var_2261 = const()[name = string("op_2261"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2267_promoted_to_fp16 = const()[name = string("op_2267_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 8, 64]> var_2271_cast_fp16 = pow(x = x_29_cast_fp16, y = var_2267_promoted_to_fp16)[name = string("op_2271_cast_fp16")];
            tensor<int32, [1]> var_2273_axes_0 = const()[name = string("op_2273_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2273_keep_dims_0 = const()[name = string("op_2273_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 8, 1]> var_2273_cast_fp16 = reduce_mean(axes = var_2273_axes_0, keep_dims = var_2273_keep_dims_0, x = var_2271_cast_fp16)[name = string("op_2273_cast_fp16")];
            fp16 var_2274_to_fp16 = const()[name = string("op_2274_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 8, 1]> var_2275_cast_fp16 = add(x = var_2273_cast_fp16, y = var_2274_to_fp16)[name = string("op_2275_cast_fp16")];
            fp32 var_2276_epsilon_0 = const()[name = string("op_2276_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 8, 1]> var_2276_cast_fp16 = rsqrt(epsilon = var_2276_epsilon_0, x = var_2275_cast_fp16)[name = string("op_2276_cast_fp16")];
            tensor<fp16, [1, 32, 8, 64]> var_2277_cast_fp16 = mul(x = x_29_cast_fp16, y = var_2276_cast_fp16)[name = string("op_2277_cast_fp16")];
            tensor<fp16, [64]> layers_3_qn_w1_to_fp16 = const()[name = string("layers_3_qn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59587456)))];
            tensor<fp16, [1, 32, 8, 64]> var_2278_cast_fp16 = mul(x = var_2277_cast_fp16, y = layers_3_qn_w1_to_fp16)[name = string("op_2278_cast_fp16")];
            tensor<int32, [4]> var_2283 = const()[name = string("op_2283"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_1_begin_0 = const()[name = string("r_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_1_end_0 = const()[name = string("r_1_end_0"), val = tensor<int32, [4]>([1, 8, 32, 16])];
            tensor<bool, [4]> r_1_end_mask_0 = const()[name = string("r_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 64]> x_33_cast_fp16 = transpose(perm = var_2283, x = var_2278_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 8, 32, 16]> r_1_cast_fp16 = slice_by_index(begin = r_1_begin_0, end = r_1_end_0, end_mask = r_1_end_mask_0, x = x_33_cast_fp16)[name = string("r_1_cast_fp16")];
            tensor<int32, [4]> p_1_begin_0 = const()[name = string("p_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_1_end_0 = const()[name = string("p_1_end_0"), val = tensor<int32, [4]>([1, 8, 32, 64])];
            tensor<bool, [4]> p_1_end_mask_0 = const()[name = string("p_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 48]> p_1_cast_fp16 = slice_by_index(begin = p_1_begin_0, end = p_1_end_0, end_mask = p_1_end_mask_0, x = x_33_cast_fp16)[name = string("p_1_cast_fp16")];
            tensor<int32, [4]> var_2299_begin_0 = const()[name = string("op_2299_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_2299_end_0 = const()[name = string("op_2299_end_0"), val = tensor<int32, [4]>([1, 8, 32, 16])];
            tensor<bool, [4]> var_2299_end_mask_0 = const()[name = string("op_2299_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 8]> var_2299_cast_fp16 = slice_by_index(begin = var_2299_begin_0, end = var_2299_end_0, end_mask = var_2299_end_mask_0, x = r_1_cast_fp16)[name = string("op_2299_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 32, 8]> var_2300_cast_fp16 = mul(x = var_2299_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_2300_cast_fp16")];
            tensor<int32, [4]> var_2305_begin_0 = const()[name = string("op_2305_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2305_end_0 = const()[name = string("op_2305_end_0"), val = tensor<int32, [4]>([1, 8, 32, 8])];
            tensor<bool, [4]> var_2305_end_mask_0 = const()[name = string("op_2305_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 8]> var_2305_cast_fp16 = slice_by_index(begin = var_2305_begin_0, end = var_2305_end_0, end_mask = var_2305_end_mask_0, x = r_1_cast_fp16)[name = string("op_2305_cast_fp16")];
            int32 var_2307 = const()[name = string("op_2307"), val = int32(-1)];
            bool rh_1_interleave_0 = const()[name = string("rh_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 16]> rh_1_cast_fp16 = concat(axis = var_2307, interleave = rh_1_interleave_0, values = (var_2300_cast_fp16, var_2305_cast_fp16))[name = string("rh_1_cast_fp16")];
            tensor<fp16, [32, 16]> cos_to_fp16 = const()[name = string("cos_to_fp16"), val = tensor<fp16, [32, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658880)))];
            tensor<fp16, [1, 8, 32, 16]> var_2309_cast_fp16 = mul(x = r_1_cast_fp16, y = cos_to_fp16)[name = string("op_2309_cast_fp16")];
            tensor<fp16, [32, 16]> sin_to_fp16 = const()[name = string("sin_to_fp16"), val = tensor<fp16, [32, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119659968)))];
            tensor<fp16, [1, 8, 32, 16]> var_2310_cast_fp16 = mul(x = rh_1_cast_fp16, y = sin_to_fp16)[name = string("op_2310_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> var_2312_cast_fp16 = add(x = var_2309_cast_fp16, y = var_2310_cast_fp16)[name = string("op_2312_cast_fp16")];
            int32 var_2314 = const()[name = string("op_2314"), val = int32(-1)];
            bool q_37_interleave_0 = const()[name = string("q_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 64]> q_37_cast_fp16 = concat(axis = var_2314, interleave = q_37_interleave_0, values = (var_2312_cast_fp16, p_1_cast_fp16))[name = string("q_37_cast_fp16")];
            fp16 var_2320_promoted_to_fp16 = const()[name = string("op_2320_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 2, 64]> var_2324_cast_fp16 = pow(x = x_35_cast_fp16, y = var_2320_promoted_to_fp16)[name = string("op_2324_cast_fp16")];
            tensor<int32, [1]> var_2326_axes_0 = const()[name = string("op_2326_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2326_keep_dims_0 = const()[name = string("op_2326_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 2, 1]> var_2326_cast_fp16 = reduce_mean(axes = var_2326_axes_0, keep_dims = var_2326_keep_dims_0, x = var_2324_cast_fp16)[name = string("op_2326_cast_fp16")];
            fp16 var_2327_to_fp16 = const()[name = string("op_2327_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 2, 1]> var_2328_cast_fp16 = add(x = var_2326_cast_fp16, y = var_2327_to_fp16)[name = string("op_2328_cast_fp16")];
            fp32 var_2329_epsilon_0 = const()[name = string("op_2329_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 2, 1]> var_2329_cast_fp16 = rsqrt(epsilon = var_2329_epsilon_0, x = var_2328_cast_fp16)[name = string("op_2329_cast_fp16")];
            tensor<fp16, [1, 32, 2, 64]> var_2330_cast_fp16 = mul(x = x_35_cast_fp16, y = var_2329_cast_fp16)[name = string("op_2330_cast_fp16")];
            tensor<fp16, [64]> layers_3_kn_w1_to_fp16 = const()[name = string("layers_3_kn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59588800)))];
            tensor<fp16, [1, 32, 2, 64]> var_2331_cast_fp16 = mul(x = var_2330_cast_fp16, y = layers_3_kn_w1_to_fp16)[name = string("op_2331_cast_fp16")];
            tensor<int32, [4]> var_2336 = const()[name = string("op_2336"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_3_begin_0 = const()[name = string("r_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_3_end_0 = const()[name = string("r_3_end_0"), val = tensor<int32, [4]>([1, 2, 32, 16])];
            tensor<bool, [4]> r_3_end_mask_0 = const()[name = string("r_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 64]> x_39_cast_fp16 = transpose(perm = var_2336, x = var_2331_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 2, 32, 16]> r_3_cast_fp16 = slice_by_index(begin = r_3_begin_0, end = r_3_end_0, end_mask = r_3_end_mask_0, x = x_39_cast_fp16)[name = string("r_3_cast_fp16")];
            tensor<int32, [4]> p_3_begin_0 = const()[name = string("p_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_3_end_0 = const()[name = string("p_3_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> p_3_end_mask_0 = const()[name = string("p_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 48]> p_3_cast_fp16 = slice_by_index(begin = p_3_begin_0, end = p_3_end_0, end_mask = p_3_end_mask_0, x = x_39_cast_fp16)[name = string("p_3_cast_fp16")];
            tensor<int32, [4]> var_2352_begin_0 = const()[name = string("op_2352_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_2352_end_0 = const()[name = string("op_2352_end_0"), val = tensor<int32, [4]>([1, 2, 32, 16])];
            tensor<bool, [4]> var_2352_end_mask_0 = const()[name = string("op_2352_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 8]> var_2352_cast_fp16 = slice_by_index(begin = var_2352_begin_0, end = var_2352_end_0, end_mask = var_2352_end_mask_0, x = r_3_cast_fp16)[name = string("op_2352_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 8]> var_2353_cast_fp16 = mul(x = var_2352_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_2353_cast_fp16")];
            tensor<int32, [4]> var_2358_begin_0 = const()[name = string("op_2358_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2358_end_0 = const()[name = string("op_2358_end_0"), val = tensor<int32, [4]>([1, 2, 32, 8])];
            tensor<bool, [4]> var_2358_end_mask_0 = const()[name = string("op_2358_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 8]> var_2358_cast_fp16 = slice_by_index(begin = var_2358_begin_0, end = var_2358_end_0, end_mask = var_2358_end_mask_0, x = r_3_cast_fp16)[name = string("op_2358_cast_fp16")];
            int32 var_2360 = const()[name = string("op_2360"), val = int32(-1)];
            bool rh_3_interleave_0 = const()[name = string("rh_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 16]> rh_3_cast_fp16 = concat(axis = var_2360, interleave = rh_3_interleave_0, values = (var_2353_cast_fp16, var_2358_cast_fp16))[name = string("rh_3_cast_fp16")];
            tensor<fp16, [1, 2, 32, 16]> var_2362_cast_fp16 = mul(x = r_3_cast_fp16, y = cos_to_fp16)[name = string("op_2362_cast_fp16")];
            tensor<fp16, [1, 2, 32, 16]> var_2363_cast_fp16 = mul(x = rh_3_cast_fp16, y = sin_to_fp16)[name = string("op_2363_cast_fp16")];
            tensor<fp16, [1, 2, 32, 16]> var_2365_cast_fp16 = add(x = var_2362_cast_fp16, y = var_2363_cast_fp16)[name = string("op_2365_cast_fp16")];
            int32 var_2367 = const()[name = string("op_2367"), val = int32(-1)];
            bool k_37_interleave_0 = const()[name = string("k_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> k_37_cast_fp16 = concat(axis = var_2367, interleave = k_37_interleave_0, values = (var_2365_cast_fp16, p_3_cast_fp16))[name = string("k_37_cast_fp16")];
            tensor<int32, [1]> var_2380_axes_0 = const()[name = string("op_2380_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_2380_cast_fp16 = expand_dims(axes = var_2380_axes_0, x = k_37_cast_fp16)[name = string("op_2380_cast_fp16")];
            tensor<int32, [5]> var_2388_reps_0 = const()[name = string("op_2388_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 2, 4, 32, 64]> var_2388_cast_fp16 = tile(reps = var_2388_reps_0, x = var_2380_cast_fp16)[name = string("op_2388_cast_fp16")];
            tensor<int32, [4]> var_2393 = const()[name = string("op_2393"), val = tensor<int32, [4]>([1, 8, 32, 64])];
            tensor<fp16, [1, 8, 32, 64]> k_39_cast_fp16 = reshape(shape = var_2393, x = var_2388_cast_fp16)[name = string("k_39_cast_fp16")];
            tensor<int32, [1]> var_2406_axes_0 = const()[name = string("op_2406_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> v_7_cast_fp16 = transpose(perm = var_2261, x = var_2256_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_2406_cast_fp16 = expand_dims(axes = var_2406_axes_0, x = v_7_cast_fp16)[name = string("op_2406_cast_fp16")];
            tensor<int32, [5]> var_2414_reps_0 = const()[name = string("op_2414_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 2, 4, 32, 64]> var_2414_cast_fp16 = tile(reps = var_2414_reps_0, x = var_2406_cast_fp16)[name = string("op_2414_cast_fp16")];
            tensor<int32, [4]> var_2419 = const()[name = string("op_2419"), val = tensor<int32, [4]>([1, 8, 32, 64])];
            tensor<fp16, [1, 8, 32, 64]> v_9_cast_fp16 = reshape(shape = var_2419, x = var_2414_cast_fp16)[name = string("v_9_cast_fp16")];
            bool var_2424_transpose_x_1 = const()[name = string("op_2424_transpose_x_1"), val = bool(false)];
            bool var_2424_transpose_y_1 = const()[name = string("op_2424_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 32]> var_2424_cast_fp16 = matmul(transpose_x = var_2424_transpose_x_1, transpose_y = var_2424_transpose_y_1, x = q_37_cast_fp16, y = k_39_cast_fp16)[name = string("op_2424_cast_fp16")];
            fp16 var_2425_to_fp16 = const()[name = string("op_2425_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 8, 32, 32]> var_2426_cast_fp16 = mul(x = var_2424_cast_fp16, y = var_2425_to_fp16)[name = string("op_2426_cast_fp16")];
            tensor<fp16, [32, 32]> causal_add_to_fp16 = const()[name = string("causal_add_to_fp16"), val = tensor<fp16, [32, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119661056)))];
            tensor<fp16, [1, 8, 32, 32]> s_1_cast_fp16 = add(x = var_2426_cast_fp16, y = causal_add_to_fp16)[name = string("s_1_cast_fp16")];
            int32 var_2429 = const()[name = string("op_2429"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 32]> var_2431_cast_fp16 = softmax(axis = var_2429, x = s_1_cast_fp16)[name = string("op_2431_cast_fp16")];
            bool o_19_transpose_x_0 = const()[name = string("o_19_transpose_x_0"), val = bool(false)];
            bool o_19_transpose_y_0 = const()[name = string("o_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 64]> o_19_cast_fp16 = matmul(transpose_x = o_19_transpose_x_0, transpose_y = o_19_transpose_y_0, x = var_2431_cast_fp16, y = v_9_cast_fp16)[name = string("o_19_cast_fp16")];
            tensor<int32, [4]> var_2437 = const()[name = string("op_2437"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2442 = const()[name = string("op_2442"), val = tensor<int32, [3]>([1, 32, 512])];
            tensor<fp16, [1, 32, 8, 64]> var_2438_cast_fp16 = transpose(perm = var_2437, x = o_19_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 32, 512]> var_2443_cast_fp16 = reshape(shape = var_2442, x = var_2438_cast_fp16)[name = string("op_2443_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2444_cast_fp16 = sigmoid(x = gate_1_cast_fp16)[name = string("op_2444_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_51_cast_fp16 = mul(x = var_2443_cast_fp16, y = var_2444_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [640, 512]> layers_3_o_weight_to_fp16 = const()[name = string("layers_3_o_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59589568)))];
            tensor<fp16, [1, 32, 640]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_o_weight_to_fp16, x = input_51_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 32, 640]> x_41_cast_fp16 = add(x = x_25_cast_fp16, y = linear_13_cast_fp16)[name = string("x_41_cast_fp16")];
            fp16 var_2457_to_fp16 = const()[name = string("op_2457_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 32, 640]> x_43_cast_fp16 = mul(x = x_41_cast_fp16, y = var_2457_to_fp16)[name = string("x_43_cast_fp16")];
            fp16 var_2455_promoted_to_fp16 = const()[name = string("op_2455_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> var_2459_cast_fp16 = pow(x = x_43_cast_fp16, y = var_2455_promoted_to_fp16)[name = string("op_2459_cast_fp16")];
            tensor<int32, [1]> var_2461_axes_0 = const()[name = string("op_2461_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2461_keep_dims_0 = const()[name = string("op_2461_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_2461_cast_fp16 = reduce_mean(axes = var_2461_axes_0, keep_dims = var_2461_keep_dims_0, x = var_2459_cast_fp16)[name = string("op_2461_cast_fp16")];
            fp16 var_2462_to_fp16 = const()[name = string("op_2462_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 32, 1]> var_2463_cast_fp16 = add(x = var_2461_cast_fp16, y = var_2462_to_fp16)[name = string("op_2463_cast_fp16")];
            fp32 var_2464_epsilon_0 = const()[name = string("op_2464_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2464_cast_fp16 = rsqrt(epsilon = var_2464_epsilon_0, x = var_2463_cast_fp16)[name = string("op_2464_cast_fp16")];
            tensor<fp16, [1, 32, 640]> var_2465_cast_fp16 = mul(x = x_43_cast_fp16, y = var_2464_cast_fp16)[name = string("op_2465_cast_fp16")];
            tensor<fp16, [640]> layers_3_n2_w1_to_fp16 = const()[name = string("layers_3_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60244992)))];
            tensor<fp16, [1, 32, 640]> input_53_cast_fp16 = mul(x = var_2465_cast_fp16, y = layers_3_n2_w1_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_3_gate_up_weight_to_fp16 = const()[name = string("layers_3_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60246336)))];
            tensor<fp16, [1, 32, 5120]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_3_gate_up_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [3]> input_55_begin_0 = const()[name = string("input_55_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_55_end_0 = const()[name = string("input_55_end_0"), val = tensor<int32, [3]>([1, 32, 2560])];
            tensor<bool, [3]> input_55_end_mask_0 = const()[name = string("input_55_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 2560]> input_55_cast_fp16 = slice_by_index(begin = input_55_begin_0, end = input_55_end_0, end_mask = input_55_end_mask_0, x = linear_14_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [1, 32, 2560]> var_2475_cast_fp16 = silu(x = input_55_cast_fp16)[name = string("op_2475_cast_fp16")];
            tensor<int32, [3]> var_2480_begin_0 = const()[name = string("op_2480_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_2480_end_0 = const()[name = string("op_2480_end_0"), val = tensor<int32, [3]>([1, 32, 5120])];
            tensor<bool, [3]> var_2480_end_mask_0 = const()[name = string("op_2480_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 2560]> var_2480_cast_fp16 = slice_by_index(begin = var_2480_begin_0, end = var_2480_end_0, end_mask = var_2480_end_mask_0, x = linear_14_cast_fp16)[name = string("op_2480_cast_fp16")];
            tensor<fp16, [1, 32, 2560]> input_57_cast_fp16 = mul(x = var_2475_cast_fp16, y = var_2480_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_3_down_weight_to_fp16 = const()[name = string("layers_3_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66800000)))];
            tensor<fp16, [1, 32, 640]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_down_weight_to_fp16, x = input_57_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, 32, 640]> x_45_cast_fp16 = add(x = x_41_cast_fp16, y = linear_15_cast_fp16)[name = string("x_45_cast_fp16")];
            fp16 var_2493_to_fp16 = const()[name = string("op_2493_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 32, 640]> x_47_cast_fp16 = mul(x = x_45_cast_fp16, y = var_2493_to_fp16)[name = string("x_47_cast_fp16")];
            fp16 var_2491_promoted_to_fp16 = const()[name = string("op_2491_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> var_2495_cast_fp16 = pow(x = x_47_cast_fp16, y = var_2491_promoted_to_fp16)[name = string("op_2495_cast_fp16")];
            tensor<int32, [1]> var_2497_axes_0 = const()[name = string("op_2497_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2497_keep_dims_0 = const()[name = string("op_2497_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_2497_cast_fp16 = reduce_mean(axes = var_2497_axes_0, keep_dims = var_2497_keep_dims_0, x = var_2495_cast_fp16)[name = string("op_2497_cast_fp16")];
            fp16 var_2498_to_fp16 = const()[name = string("op_2498_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 32, 1]> var_2499_cast_fp16 = add(x = var_2497_cast_fp16, y = var_2498_to_fp16)[name = string("op_2499_cast_fp16")];
            fp32 var_2500_epsilon_0 = const()[name = string("op_2500_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2500_cast_fp16 = rsqrt(epsilon = var_2500_epsilon_0, x = var_2499_cast_fp16)[name = string("op_2500_cast_fp16")];
            tensor<fp16, [1, 32, 640]> var_2501_cast_fp16 = mul(x = x_47_cast_fp16, y = var_2500_cast_fp16)[name = string("op_2501_cast_fp16")];
            tensor<fp16, [640]> layers_4_n1_w1_to_fp16 = const()[name = string("layers_4_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70076864)))];
            tensor<fp16, [1, 32, 640]> input_59_cast_fp16 = mul(x = var_2501_cast_fp16, y = layers_4_n1_w1_to_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_4_qkvzba_weight_to_fp16 = const()[name = string("layers_4_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70078208)))];
            tensor<fp16, [1, 32, 1544]> linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_qkvzba_weight_to_fp16, x = input_59_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [3]> qkv_7_begin_0 = const()[name = string("qkv_7_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_7_end_0 = const()[name = string("qkv_7_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> qkv_7_end_mask_0 = const()[name = string("qkv_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 1024]> qkv_7_cast_fp16 = slice_by_index(begin = qkv_7_begin_0, end = qkv_7_end_0, end_mask = qkv_7_end_mask_0, x = linear_16_cast_fp16)[name = string("qkv_7_cast_fp16")];
            tensor<int32, [3]> z_7_begin_0 = const()[name = string("z_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_7_end_0 = const()[name = string("z_7_end_0"), val = tensor<int32, [3]>([1, 32, 1536])];
            tensor<bool, [3]> z_7_end_mask_0 = const()[name = string("z_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 512]> z_7_cast_fp16 = slice_by_index(begin = z_7_begin_0, end = z_7_end_0, end_mask = z_7_end_mask_0, x = linear_16_cast_fp16)[name = string("z_7_cast_fp16")];
            tensor<int32, [3]> b_7_begin_0 = const()[name = string("b_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_7_end_0 = const()[name = string("b_7_end_0"), val = tensor<int32, [3]>([1, 32, 1540])];
            tensor<bool, [3]> b_7_end_mask_0 = const()[name = string("b_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 4]> b_7_cast_fp16 = slice_by_index(begin = b_7_begin_0, end = b_7_end_0, end_mask = b_7_end_mask_0, x = linear_16_cast_fp16)[name = string("b_7_cast_fp16")];
            tensor<int32, [3]> a_7_begin_0 = const()[name = string("a_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_7_end_0 = const()[name = string("a_7_end_0"), val = tensor<int32, [3]>([1, 32, 1])];
            tensor<bool, [3]> a_7_end_mask_0 = const()[name = string("a_7_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 4]> a_7_cast_fp16 = slice_by_index(begin = a_7_begin_0, end = a_7_end_0, end_mask = a_7_end_mask_0, x = linear_16_cast_fp16)[name = string("a_7_cast_fp16")];
            int32 var_2540 = const()[name = string("op_2540"), val = int32(1)];
            bool xp_7_interleave_0 = const()[name = string("xp_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 35, 1024]> xp_7_cast_fp16 = concat(axis = var_2540, interleave = xp_7_interleave_0, values = (zpad_to_fp16, qkv_7_cast_fp16))[name = string("xp_7_cast_fp16")];
            tensor<int32, [3]> var_2551_begin_0 = const()[name = string("op_2551_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_2551_end_0 = const()[name = string("op_2551_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_2551_end_mask_0 = const()[name = string("op_2551_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 1024]> var_2551_cast_fp16 = slice_by_index(begin = var_2551_begin_0, end = var_2551_end_0, end_mask = var_2551_end_mask_0, x = xp_7_cast_fp16)[name = string("op_2551_cast_fp16")];
            tensor<fp16, [1024]> var_2559_to_fp16 = const()[name = string("op_2559_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72054592)))];
            tensor<fp16, [1, 32, 1024]> var_2560_cast_fp16 = mul(x = var_2551_cast_fp16, y = var_2559_to_fp16)[name = string("op_2560_cast_fp16")];
            tensor<int32, [3]> var_2570_begin_0 = const()[name = string("op_2570_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_2570_end_0 = const()[name = string("op_2570_end_0"), val = tensor<int32, [3]>([1, 34, 1024])];
            tensor<bool, [3]> var_2570_end_mask_0 = const()[name = string("op_2570_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_2570_cast_fp16 = slice_by_index(begin = var_2570_begin_0, end = var_2570_end_0, end_mask = var_2570_end_mask_0, x = xp_7_cast_fp16)[name = string("op_2570_cast_fp16")];
            tensor<fp16, [1024]> var_2578_to_fp16 = const()[name = string("op_2578_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72056704)))];
            tensor<fp16, [1, 32, 1024]> var_2579_cast_fp16 = mul(x = var_2570_cast_fp16, y = var_2578_to_fp16)[name = string("op_2579_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> var_2581_cast_fp16 = add(x = var_2560_cast_fp16, y = var_2579_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [3]> var_2591_begin_0 = const()[name = string("op_2591_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2591_end_0 = const()[name = string("op_2591_end_0"), val = tensor<int32, [3]>([1, 33, 1024])];
            tensor<bool, [3]> var_2591_end_mask_0 = const()[name = string("op_2591_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_2591_cast_fp16 = slice_by_index(begin = var_2591_begin_0, end = var_2591_end_0, end_mask = var_2591_end_mask_0, x = xp_7_cast_fp16)[name = string("op_2591_cast_fp16")];
            tensor<fp16, [1024]> var_2599_to_fp16 = const()[name = string("op_2599_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72058816)))];
            tensor<fp16, [1, 32, 1024]> var_2600_cast_fp16 = mul(x = var_2591_cast_fp16, y = var_2599_to_fp16)[name = string("op_2600_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> var_2602_cast_fp16 = add(x = var_2581_cast_fp16, y = var_2600_cast_fp16)[name = string("op_2602_cast_fp16")];
            tensor<int32, [3]> var_2612_begin_0 = const()[name = string("op_2612_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2612_end_0 = const()[name = string("op_2612_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> var_2612_end_mask_0 = const()[name = string("op_2612_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_2612_cast_fp16 = slice_by_index(begin = var_2612_begin_0, end = var_2612_end_0, end_mask = var_2612_end_mask_0, x = xp_7_cast_fp16)[name = string("op_2612_cast_fp16")];
            tensor<fp16, [1024]> var_2620_to_fp16 = const()[name = string("op_2620_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72060928)))];
            tensor<fp16, [1, 32, 1024]> var_2621_cast_fp16 = mul(x = var_2612_cast_fp16, y = var_2620_to_fp16)[name = string("op_2621_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> input_61_cast_fp16 = add(x = var_2602_cast_fp16, y = var_2621_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> c_7_cast_fp16 = silu(x = input_61_cast_fp16)[name = string("c_7_cast_fp16")];
            tensor<int32, [3]> var_2629_begin_0 = const()[name = string("op_2629_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2629_end_0 = const()[name = string("op_2629_end_0"), val = tensor<int32, [3]>([1, 32, 256])];
            tensor<bool, [3]> var_2629_end_mask_0 = const()[name = string("op_2629_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 256]> var_2629_cast_fp16 = slice_by_index(begin = var_2629_begin_0, end = var_2629_end_0, end_mask = var_2629_end_mask_0, x = c_7_cast_fp16)[name = string("op_2629_cast_fp16")];
            tensor<int32, [4]> var_2634 = const()[name = string("op_2634"), val = tensor<int32, [4]>([1, 32, 2, 128])];
            tensor<fp16, [1, 32, 2, 128]> q_39_cast_fp16 = reshape(shape = var_2634, x = var_2629_cast_fp16)[name = string("q_39_cast_fp16")];
            tensor<int32, [3]> var_2640_begin_0 = const()[name = string("op_2640_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_2640_end_0 = const()[name = string("op_2640_end_0"), val = tensor<int32, [3]>([1, 32, 512])];
            tensor<bool, [3]> var_2640_end_mask_0 = const()[name = string("op_2640_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 256]> var_2640_cast_fp16 = slice_by_index(begin = var_2640_begin_0, end = var_2640_end_0, end_mask = var_2640_end_mask_0, x = c_7_cast_fp16)[name = string("op_2640_cast_fp16")];
            tensor<int32, [4]> var_2645 = const()[name = string("op_2645"), val = tensor<int32, [4]>([1, 32, 2, 128])];
            tensor<fp16, [1, 32, 2, 128]> k_41_cast_fp16 = reshape(shape = var_2645, x = var_2640_cast_fp16)[name = string("k_41_cast_fp16")];
            tensor<int32, [3]> var_2651_begin_0 = const()[name = string("op_2651_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_2651_end_0 = const()[name = string("op_2651_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> var_2651_end_mask_0 = const()[name = string("op_2651_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 512]> var_2651_cast_fp16 = slice_by_index(begin = var_2651_begin_0, end = var_2651_end_0, end_mask = var_2651_end_mask_0, x = c_7_cast_fp16)[name = string("op_2651_cast_fp16")];
            tensor<int32, [4]> var_2656 = const()[name = string("op_2656"), val = tensor<int32, [4]>([1, 32, 4, 128])];
            tensor<fp16, [1, 32, 4, 128]> var_2657_cast_fp16 = reshape(shape = var_2656, x = var_2651_cast_fp16)[name = string("op_2657_cast_fp16")];
            tensor<int32, [4]> var_2662 = const()[name = string("op_2662"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2664_to_fp16 = const()[name = string("op_2664_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 32, 2, 128]> q_41_cast_fp16 = mul(x = q_39_cast_fp16, y = var_2664_to_fp16)[name = string("q_41_cast_fp16")];
            fp16 var_2666_to_fp16 = const()[name = string("op_2666_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 32, 2, 128]> k_43_cast_fp16 = mul(x = k_41_cast_fp16, y = var_2666_to_fp16)[name = string("k_43_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_2668_cast_fp16 = mul(x = q_41_cast_fp16, y = q_41_cast_fp16)[name = string("op_2668_cast_fp16")];
            tensor<int32, [1]> var_2673_axes_0 = const()[name = string("op_2673_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2673_keep_dims_0 = const()[name = string("op_2673_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 2, 1]> var_2673_cast_fp16 = reduce_mean(axes = var_2673_axes_0, keep_dims = var_2673_keep_dims_0, x = var_2668_cast_fp16)[name = string("op_2673_cast_fp16")];
            fp16 var_2674_promoted_to_fp16 = const()[name = string("op_2674_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 32, 2, 1]> var_2675_cast_fp16 = mul(x = var_2673_cast_fp16, y = var_2674_promoted_to_fp16)[name = string("op_2675_cast_fp16")];
            fp16 var_2677_to_fp16 = const()[name = string("op_2677_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 32, 2, 1]> var_2678_cast_fp16 = add(x = var_2675_cast_fp16, y = var_2677_to_fp16)[name = string("op_2678_cast_fp16")];
            fp32 var_2679_epsilon_0 = const()[name = string("op_2679_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 2, 1]> var_2679_cast_fp16 = rsqrt(epsilon = var_2679_epsilon_0, x = var_2678_cast_fp16)[name = string("op_2679_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_2680_cast_fp16 = mul(x = q_41_cast_fp16, y = var_2679_cast_fp16)[name = string("op_2680_cast_fp16")];
            fp16 var_2681_to_fp16 = const()[name = string("op_2681_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 32, 2, 128]> q_43_cast_fp16 = mul(x = var_2680_cast_fp16, y = var_2681_to_fp16)[name = string("q_43_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_2683_cast_fp16 = mul(x = k_43_cast_fp16, y = k_43_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<int32, [1]> var_2688_axes_0 = const()[name = string("op_2688_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2688_keep_dims_0 = const()[name = string("op_2688_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 2, 1]> var_2688_cast_fp16 = reduce_mean(axes = var_2688_axes_0, keep_dims = var_2688_keep_dims_0, x = var_2683_cast_fp16)[name = string("op_2688_cast_fp16")];
            fp16 var_2689_promoted_to_fp16 = const()[name = string("op_2689_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 32, 2, 1]> var_2690_cast_fp16 = mul(x = var_2688_cast_fp16, y = var_2689_promoted_to_fp16)[name = string("op_2690_cast_fp16")];
            fp16 var_2692_to_fp16 = const()[name = string("op_2692_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 32, 2, 1]> var_2693_cast_fp16 = add(x = var_2690_cast_fp16, y = var_2692_to_fp16)[name = string("op_2693_cast_fp16")];
            fp32 var_2694_epsilon_0 = const()[name = string("op_2694_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 2, 1]> var_2694_cast_fp16 = rsqrt(epsilon = var_2694_epsilon_0, x = var_2693_cast_fp16)[name = string("op_2694_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> k_45_cast_fp16 = mul(x = k_43_cast_fp16, y = var_2694_cast_fp16)[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> var_2700 = const()[name = string("op_2700"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2713_axes_0 = const()[name = string("op_2713_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 128]> var_2701_cast_fp16 = transpose(perm = var_2700, x = q_43_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 2, 1, 32, 128]> var_2713_cast_fp16 = expand_dims(axes = var_2713_axes_0, x = var_2701_cast_fp16)[name = string("op_2713_cast_fp16")];
            tensor<int32, [5]> var_2721_reps_0 = const()[name = string("op_2721_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 32, 128]> var_2721_cast_fp16 = tile(reps = var_2721_reps_0, x = var_2713_cast_fp16)[name = string("op_2721_cast_fp16")];
            tensor<int32, [4]> var_2726 = const()[name = string("op_2726"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<fp16, [1, 4, 32, 128]> q_45_cast_fp16 = reshape(shape = var_2726, x = var_2721_cast_fp16)[name = string("q_45_cast_fp16")];
            tensor<int32, [4]> var_2732 = const()[name = string("op_2732"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2745_axes_0 = const()[name = string("op_2745_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 128]> var_2733_cast_fp16 = transpose(perm = var_2732, x = k_45_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 2, 1, 32, 128]> var_2745_cast_fp16 = expand_dims(axes = var_2745_axes_0, x = var_2733_cast_fp16)[name = string("op_2745_cast_fp16")];
            tensor<int32, [5]> var_2753_reps_0 = const()[name = string("op_2753_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 32, 128]> var_2753_cast_fp16 = tile(reps = var_2753_reps_0, x = var_2745_cast_fp16)[name = string("op_2753_cast_fp16")];
            tensor<int32, [4]> var_2758 = const()[name = string("op_2758"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<fp16, [1, 4, 32, 128]> k_47_cast_fp16 = reshape(shape = var_2758, x = var_2753_cast_fp16)[name = string("k_47_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_2760_cast_fp16 = sigmoid(x = b_7_cast_fp16)[name = string("op_2760_cast_fp16")];
            tensor<int32, [3]> var_2764 = const()[name = string("op_2764"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_4_dt_bias_to_fp16 = const()[name = string("layers_4_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 32, 4]> input_63_cast_fp16 = add(x = a_7_cast_fp16, y = layers_4_dt_bias_to_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_2768_cast_fp16 = relu(x = input_63_cast_fp16)[name = string("op_2768_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_2769_cast_fp16 = abs(x = input_63_cast_fp16)[name = string("op_2769_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, 4]> var_2770_cast_fp16 = mul(x = var_2769_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_2770_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_2771_cast_fp16 = exp(x = var_2770_cast_fp16)[name = string("op_2771_cast_fp16")];
            fp16 var_2773_to_fp16 = const()[name = string("op_2773_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 32, 4]> var_2774_cast_fp16 = add(x = var_2771_cast_fp16, y = var_2773_to_fp16)[name = string("op_2774_cast_fp16")];
            fp32 var_2775_epsilon_0 = const()[name = string("op_2775_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 32, 4]> var_2775_cast_fp16 = log(epsilon = var_2775_epsilon_0, x = var_2774_cast_fp16)[name = string("op_2775_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_2777_cast_fp16 = add(x = var_2768_cast_fp16, y = var_2775_cast_fp16)[name = string("op_2777_cast_fp16")];
            tensor<fp16, [4]> layers_4_negA_to_fp16 = const()[name = string("layers_4_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.348p+3, -0x1.734p+2, -0x1.5a8p-1, -0x1.264p+3])];
            tensor<fp16, [1, 32, 4]> var_2778_cast_fp16 = mul(x = layers_4_negA_to_fp16, y = var_2777_cast_fp16)[name = string("op_2778_cast_fp16")];
            tensor<int32, [3]> var_2782 = const()[name = string("op_2782"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_2785_axes_0 = const()[name = string("op_2785_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 32]> beta_7_cast_fp16 = transpose(perm = var_2764, x = var_2760_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 4, 32, 1]> var_2785_cast_fp16 = expand_dims(axes = var_2785_axes_0, x = beta_7_cast_fp16)[name = string("op_2785_cast_fp16")];
            fp16 var_2786_to_fp16 = const()[name = string("op_2786_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 32, 1]> var_2787_cast_fp16 = mul(x = var_2785_cast_fp16, y = var_2786_to_fp16)[name = string("op_2787_cast_fp16")];
            tensor<fp16, [1, 4, 32, 128]> v_11_cast_fp16 = transpose(perm = var_2662, x = var_2657_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 4, 32, 128]> vb_19_cast_fp16 = mul(x = v_11_cast_fp16, y = var_2787_cast_fp16)[name = string("vb_19_cast_fp16")];
            tensor<int32, [4]> q_47_begin_0 = const()[name = string("q_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_47_end_0 = const()[name = string("q_47_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_47_end_mask_0 = const()[name = string("q_47_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_47_cast_fp16 = slice_by_index(begin = q_47_begin_0, end = q_47_end_0, end_mask = q_47_end_mask_0, x = q_45_cast_fp16)[name = string("q_47_cast_fp16")];
            tensor<int32, [4]> k_49_begin_0 = const()[name = string("k_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_49_end_0 = const()[name = string("k_49_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_49_end_mask_0 = const()[name = string("k_49_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, x = k_47_cast_fp16)[name = string("k_49_cast_fp16")];
            tensor<fp16, [1, 4, 32, 128]> var_2821_cast_fp16 = mul(x = k_47_cast_fp16, y = var_2785_cast_fp16)[name = string("op_2821_cast_fp16")];
            tensor<int32, [4]> kb_13_begin_0 = const()[name = string("kb_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_13_end_0 = const()[name = string("kb_13_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_13_end_mask_0 = const()[name = string("kb_13_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_13_cast_fp16 = slice_by_index(begin = kb_13_begin_0, end = kb_13_end_0, end_mask = kb_13_end_mask_0, x = var_2821_cast_fp16)[name = string("kb_13_cast_fp16")];
            tensor<int32, [4]> vb_21_begin_0 = const()[name = string("vb_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_21_end_0 = const()[name = string("vb_21_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_21_end_mask_0 = const()[name = string("vb_21_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_21_cast_fp16 = slice_by_index(begin = vb_21_begin_0, end = vb_21_end_0, end_mask = vb_21_end_mask_0, x = vb_19_cast_fp16)[name = string("vb_21_cast_fp16")];
            tensor<int32, [3]> gdec_21_begin_0 = const()[name = string("gdec_21_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_21_end_0 = const()[name = string("gdec_21_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_21_end_mask_0 = const()[name = string("gdec_21_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 32]> gdec_19_cast_fp16 = transpose(perm = var_2782, x = var_2778_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 4, 16]> gdec_21_cast_fp16 = slice_by_index(begin = gdec_21_begin_0, end = gdec_21_end_0, end_mask = gdec_21_end_mask_0, x = gdec_19_cast_fp16)[name = string("gdec_21_cast_fp16")];
            tensor<int32, [2]> var_2869 = const()[name = string("op_2869"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_2870_cast_fp16 = reshape(shape = var_2869, x = gdec_21_cast_fp16)[name = string("op_2870_cast_fp16")];
            tensor<fp16, [4, 256]> var_2871_cast_fp16 = linear(bias = var_633_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2870_cast_fp16)[name = string("op_2871_cast_fp16")];
            tensor<int32, [4]> var_2876 = const()[name = string("op_2876"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_13_cast_fp16 = reshape(shape = var_2876, x = var_2871_cast_fp16)[name = string("P_13_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2878_cast_fp16 = exp(x = P_13_cast_fp16)[name = string("op_2878_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_13_cast_fp16 = mul(x = var_2878_cast_fp16, y = tril_incl_to_fp16)[name = string("D_13_cast_fp16")];
            bool kk_13_transpose_x_1 = const()[name = string("kk_13_transpose_x_1"), val = bool(false)];
            bool kk_13_transpose_y_1 = const()[name = string("kk_13_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_13_cast_fp16 = matmul(transpose_x = kk_13_transpose_x_1, transpose_y = kk_13_transpose_y_1, x = kb_13_cast_fp16, y = k_49_cast_fp16)[name = string("kk_13_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_2884_cast_fp16 = mul(x = kk_13_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_2884_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2885_cast_fp16 = mul(x = var_2884_cast_fp16, y = D_13_cast_fp16)[name = string("op_2885_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_13_cast_fp16 = mul(x = var_2885_cast_fp16, y = cum_tail_to_fp16)[name = string("N_13_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2887_cast_fp16 = mul(x = D_13_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2887_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_13_cast_fp16 = mul(x = kk_13_cast_fp16, y = var_2887_cast_fp16)[name = string("Na_13_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_49_cast_fp16 = add(x = eye_to_fp16, y = N_13_cast_fp16)[name = string("Minv_49_cast_fp16")];
            bool Np_37_transpose_x_0 = const()[name = string("Np_37_transpose_x_0"), val = bool(false)];
            bool Np_37_transpose_y_0 = const()[name = string("Np_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_37_cast_fp16 = matmul(transpose_x = Np_37_transpose_x_0, transpose_y = Np_37_transpose_y_0, x = N_13_cast_fp16, y = Na_13_cast_fp16)[name = string("Np_37_cast_fp16")];
            bool Npa_25_transpose_x_0 = const()[name = string("Npa_25_transpose_x_0"), val = bool(false)];
            bool Npa_25_transpose_y_0 = const()[name = string("Npa_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_25_cast_fp16 = matmul(transpose_x = Npa_25_transpose_x_0, transpose_y = Npa_25_transpose_y_0, x = Na_13_cast_fp16, y = N_13_cast_fp16)[name = string("Npa_25_cast_fp16")];
            bool var_2893_transpose_x_0 = const()[name = string("op_2893_transpose_x_0"), val = bool(false)];
            bool var_2893_transpose_y_0 = const()[name = string("op_2893_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2893_cast_fp16 = matmul(transpose_x = var_2893_transpose_x_0, transpose_y = var_2893_transpose_y_0, x = Minv_49_cast_fp16, y = Np_37_cast_fp16)[name = string("op_2893_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_51_cast_fp16 = add(x = Minv_49_cast_fp16, y = var_2893_cast_fp16)[name = string("Minv_51_cast_fp16")];
            bool Np_39_transpose_x_0 = const()[name = string("Np_39_transpose_x_0"), val = bool(false)];
            bool Np_39_transpose_y_0 = const()[name = string("Np_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_39_cast_fp16 = matmul(transpose_x = Np_39_transpose_x_0, transpose_y = Np_39_transpose_y_0, x = Np_37_cast_fp16, y = Npa_25_cast_fp16)[name = string("Np_39_cast_fp16")];
            bool Npa_27_transpose_x_0 = const()[name = string("Npa_27_transpose_x_0"), val = bool(false)];
            bool Npa_27_transpose_y_0 = const()[name = string("Npa_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_27_cast_fp16 = matmul(transpose_x = Npa_27_transpose_x_0, transpose_y = Npa_27_transpose_y_0, x = Npa_25_cast_fp16, y = Np_37_cast_fp16)[name = string("Npa_27_cast_fp16")];
            bool var_2898_transpose_x_0 = const()[name = string("op_2898_transpose_x_0"), val = bool(false)];
            bool var_2898_transpose_y_0 = const()[name = string("op_2898_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2898_cast_fp16 = matmul(transpose_x = var_2898_transpose_x_0, transpose_y = var_2898_transpose_y_0, x = Minv_51_cast_fp16, y = Np_39_cast_fp16)[name = string("op_2898_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_53_cast_fp16 = add(x = Minv_51_cast_fp16, y = var_2898_cast_fp16)[name = string("Minv_53_cast_fp16")];
            bool Np_41_transpose_x_0 = const()[name = string("Np_41_transpose_x_0"), val = bool(false)];
            bool Np_41_transpose_y_0 = const()[name = string("Np_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_41_cast_fp16 = matmul(transpose_x = Np_41_transpose_x_0, transpose_y = Np_41_transpose_y_0, x = Np_39_cast_fp16, y = Npa_27_cast_fp16)[name = string("Np_41_cast_fp16")];
            bool var_2902_transpose_x_0 = const()[name = string("op_2902_transpose_x_0"), val = bool(false)];
            bool var_2902_transpose_y_0 = const()[name = string("op_2902_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2902_cast_fp16 = matmul(transpose_x = var_2902_transpose_x_0, transpose_y = var_2902_transpose_y_0, x = Minv_53_cast_fp16, y = Np_41_cast_fp16)[name = string("op_2902_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_55_cast_fp16 = add(x = Minv_53_cast_fp16, y = var_2902_cast_fp16)[name = string("Minv_55_cast_fp16")];
            bool u_13_transpose_x_0 = const()[name = string("u_13_transpose_x_0"), val = bool(false)];
            bool u_13_transpose_y_0 = const()[name = string("u_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_13_cast_fp16 = matmul(transpose_x = u_13_transpose_x_0, transpose_y = u_13_transpose_y_0, x = Minv_55_cast_fp16, y = vb_21_cast_fp16)[name = string("u_13_cast_fp16")];
            bool var_2909_transpose_x_1 = const()[name = string("op_2909_transpose_x_1"), val = bool(false)];
            bool var_2909_transpose_y_1 = const()[name = string("op_2909_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_2909_cast_fp16 = matmul(transpose_x = var_2909_transpose_x_1, transpose_y = var_2909_transpose_y_1, x = q_47_cast_fp16, y = k_49_cast_fp16)[name = string("op_2909_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_13_cast_fp16 = mul(x = var_2909_cast_fp16, y = D_13_cast_fp16)[name = string("A_13_cast_fp16")];
            bool var_2911_transpose_x_0 = const()[name = string("op_2911_transpose_x_0"), val = bool(false)];
            bool var_2911_transpose_y_0 = const()[name = string("op_2911_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2911_cast_fp16 = matmul(transpose_x = var_2911_transpose_x_0, transpose_y = var_2911_transpose_y_0, x = A_13_cast_fp16, y = u_13_cast_fp16)[name = string("op_2911_cast_fp16")];
            tensor<fp16, [16]> var_2912_bias_0_to_fp16 = const()[name = string("op_2912_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_2912_cast_fp16 = linear(bias = var_2912_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_21_cast_fp16)[name = string("op_2912_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_7_cast_fp16 = exp(x = var_2912_cast_fp16)[name = string("tail_7_cast_fp16")];
            tensor<int32, [1]> var_2915_axes_0 = const()[name = string("op_2915_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_2915_cast_fp16 = expand_dims(axes = var_2915_axes_0, x = tail_7_cast_fp16)[name = string("op_2915_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2916_cast_fp16 = mul(x = k_49_cast_fp16, y = var_2915_cast_fp16)[name = string("op_2916_cast_fp16")];
            bool S_7_transpose_x_1 = const()[name = string("S_7_transpose_x_1"), val = bool(true)];
            bool S_7_transpose_y_1 = const()[name = string("S_7_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_7_cast_fp16 = matmul(transpose_x = S_7_transpose_x_1, transpose_y = S_7_transpose_y_1, x = var_2916_cast_fp16, y = u_13_cast_fp16)[name = string("S_7_cast_fp16")];
            tensor<int32, [4]> q_49_begin_0 = const()[name = string("q_49_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_49_end_0 = const()[name = string("q_49_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_49_end_mask_0 = const()[name = string("q_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, x = q_45_cast_fp16)[name = string("q_49_cast_fp16")];
            tensor<int32, [4]> k_51_begin_0 = const()[name = string("k_51_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_51_end_0 = const()[name = string("k_51_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_51_end_mask_0 = const()[name = string("k_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_51_cast_fp16 = slice_by_index(begin = k_51_begin_0, end = k_51_end_0, end_mask = k_51_end_mask_0, x = k_47_cast_fp16)[name = string("k_51_cast_fp16")];
            tensor<int32, [4]> kb_15_begin_0 = const()[name = string("kb_15_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_15_end_0 = const()[name = string("kb_15_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_15_end_mask_0 = const()[name = string("kb_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_15_cast_fp16 = slice_by_index(begin = kb_15_begin_0, end = kb_15_end_0, end_mask = kb_15_end_mask_0, x = var_2821_cast_fp16)[name = string("kb_15_cast_fp16")];
            tensor<int32, [4]> vb_23_begin_0 = const()[name = string("vb_23_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_23_end_0 = const()[name = string("vb_23_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_23_end_mask_0 = const()[name = string("vb_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_23_cast_fp16 = slice_by_index(begin = vb_23_begin_0, end = vb_23_end_0, end_mask = vb_23_end_mask_0, x = vb_19_cast_fp16)[name = string("vb_23_cast_fp16")];
            tensor<int32, [3]> gdec_23_begin_0 = const()[name = string("gdec_23_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_23_end_0 = const()[name = string("gdec_23_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_23_end_mask_0 = const()[name = string("gdec_23_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_23_cast_fp16 = slice_by_index(begin = gdec_23_begin_0, end = gdec_23_end_0, end_mask = gdec_23_end_mask_0, x = gdec_19_cast_fp16)[name = string("gdec_23_cast_fp16")];
            tensor<int32, [2]> var_3001 = const()[name = string("op_3001"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_3002_cast_fp16 = reshape(shape = var_3001, x = gdec_23_cast_fp16)[name = string("op_3002_cast_fp16")];
            tensor<fp16, [4, 256]> var_3003_cast_fp16 = linear(bias = var_633_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3002_cast_fp16)[name = string("op_3003_cast_fp16")];
            tensor<int32, [4]> var_3008 = const()[name = string("op_3008"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_15_cast_fp16 = reshape(shape = var_3008, x = var_3003_cast_fp16)[name = string("P_15_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3010_cast_fp16 = exp(x = P_15_cast_fp16)[name = string("op_3010_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_15_cast_fp16 = mul(x = var_3010_cast_fp16, y = tril_incl_to_fp16)[name = string("D_15_cast_fp16")];
            bool kk_15_transpose_x_1 = const()[name = string("kk_15_transpose_x_1"), val = bool(false)];
            bool kk_15_transpose_y_1 = const()[name = string("kk_15_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_15_cast_fp16 = matmul(transpose_x = kk_15_transpose_x_1, transpose_y = kk_15_transpose_y_1, x = kb_15_cast_fp16, y = k_51_cast_fp16)[name = string("kk_15_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_3016_cast_fp16 = mul(x = kk_15_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_3016_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3017_cast_fp16 = mul(x = var_3016_cast_fp16, y = D_15_cast_fp16)[name = string("op_3017_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_15_cast_fp16 = mul(x = var_3017_cast_fp16, y = cum_tail_to_fp16)[name = string("N_15_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3019_cast_fp16 = mul(x = D_15_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3019_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_15_cast_fp16 = mul(x = kk_15_cast_fp16, y = var_3019_cast_fp16)[name = string("Na_15_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_57_cast_fp16 = add(x = eye_to_fp16, y = N_15_cast_fp16)[name = string("Minv_57_cast_fp16")];
            bool Np_43_transpose_x_0 = const()[name = string("Np_43_transpose_x_0"), val = bool(false)];
            bool Np_43_transpose_y_0 = const()[name = string("Np_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_43_cast_fp16 = matmul(transpose_x = Np_43_transpose_x_0, transpose_y = Np_43_transpose_y_0, x = N_15_cast_fp16, y = Na_15_cast_fp16)[name = string("Np_43_cast_fp16")];
            bool Npa_29_transpose_x_0 = const()[name = string("Npa_29_transpose_x_0"), val = bool(false)];
            bool Npa_29_transpose_y_0 = const()[name = string("Npa_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_29_cast_fp16 = matmul(transpose_x = Npa_29_transpose_x_0, transpose_y = Npa_29_transpose_y_0, x = Na_15_cast_fp16, y = N_15_cast_fp16)[name = string("Npa_29_cast_fp16")];
            bool var_3025_transpose_x_0 = const()[name = string("op_3025_transpose_x_0"), val = bool(false)];
            bool var_3025_transpose_y_0 = const()[name = string("op_3025_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3025_cast_fp16 = matmul(transpose_x = var_3025_transpose_x_0, transpose_y = var_3025_transpose_y_0, x = Minv_57_cast_fp16, y = Np_43_cast_fp16)[name = string("op_3025_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_59_cast_fp16 = add(x = Minv_57_cast_fp16, y = var_3025_cast_fp16)[name = string("Minv_59_cast_fp16")];
            bool Np_45_transpose_x_0 = const()[name = string("Np_45_transpose_x_0"), val = bool(false)];
            bool Np_45_transpose_y_0 = const()[name = string("Np_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_45_cast_fp16 = matmul(transpose_x = Np_45_transpose_x_0, transpose_y = Np_45_transpose_y_0, x = Np_43_cast_fp16, y = Npa_29_cast_fp16)[name = string("Np_45_cast_fp16")];
            bool Npa_31_transpose_x_0 = const()[name = string("Npa_31_transpose_x_0"), val = bool(false)];
            bool Npa_31_transpose_y_0 = const()[name = string("Npa_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_31_cast_fp16 = matmul(transpose_x = Npa_31_transpose_x_0, transpose_y = Npa_31_transpose_y_0, x = Npa_29_cast_fp16, y = Np_43_cast_fp16)[name = string("Npa_31_cast_fp16")];
            bool var_3030_transpose_x_0 = const()[name = string("op_3030_transpose_x_0"), val = bool(false)];
            bool var_3030_transpose_y_0 = const()[name = string("op_3030_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3030_cast_fp16 = matmul(transpose_x = var_3030_transpose_x_0, transpose_y = var_3030_transpose_y_0, x = Minv_59_cast_fp16, y = Np_45_cast_fp16)[name = string("op_3030_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_61_cast_fp16 = add(x = Minv_59_cast_fp16, y = var_3030_cast_fp16)[name = string("Minv_61_cast_fp16")];
            bool Np_47_transpose_x_0 = const()[name = string("Np_47_transpose_x_0"), val = bool(false)];
            bool Np_47_transpose_y_0 = const()[name = string("Np_47_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_47_cast_fp16 = matmul(transpose_x = Np_47_transpose_x_0, transpose_y = Np_47_transpose_y_0, x = Np_45_cast_fp16, y = Npa_31_cast_fp16)[name = string("Np_47_cast_fp16")];
            bool var_3034_transpose_x_0 = const()[name = string("op_3034_transpose_x_0"), val = bool(false)];
            bool var_3034_transpose_y_0 = const()[name = string("op_3034_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3034_cast_fp16 = matmul(transpose_x = var_3034_transpose_x_0, transpose_y = var_3034_transpose_y_0, x = Minv_61_cast_fp16, y = Np_47_cast_fp16)[name = string("op_3034_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_63_cast_fp16 = add(x = Minv_61_cast_fp16, y = var_3034_cast_fp16)[name = string("Minv_63_cast_fp16")];
            bool u_15_transpose_x_0 = const()[name = string("u_15_transpose_x_0"), val = bool(false)];
            bool u_15_transpose_y_0 = const()[name = string("u_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_15_cast_fp16 = matmul(transpose_x = u_15_transpose_x_0, transpose_y = u_15_transpose_y_0, x = Minv_63_cast_fp16, y = vb_23_cast_fp16)[name = string("u_15_cast_fp16")];
            bool var_3041_transpose_x_1 = const()[name = string("op_3041_transpose_x_1"), val = bool(false)];
            bool var_3041_transpose_y_1 = const()[name = string("op_3041_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_3041_cast_fp16 = matmul(transpose_x = var_3041_transpose_x_1, transpose_y = var_3041_transpose_y_1, x = q_49_cast_fp16, y = k_51_cast_fp16)[name = string("op_3041_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_15_cast_fp16 = mul(x = var_3041_cast_fp16, y = D_15_cast_fp16)[name = string("A_15_cast_fp16")];
            tensor<fp16, [16]> var_3043_bias_0_to_fp16 = const()[name = string("op_3043_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_3043_cast_fp16 = linear(bias = var_3043_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_23_cast_fp16)[name = string("op_3043_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_7_cast_fp16 = exp(x = var_3043_cast_fp16)[name = string("cum_7_cast_fp16")];
            tensor<int32, [1]> var_3046_axes_0 = const()[name = string("op_3046_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_3046_cast_fp16 = expand_dims(axes = var_3046_axes_0, x = cum_7_cast_fp16)[name = string("op_3046_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3047_cast_fp16 = mul(x = kb_15_cast_fp16, y = var_3046_cast_fp16)[name = string("op_3047_cast_fp16")];
            bool w_15_transpose_x_0 = const()[name = string("w_15_transpose_x_0"), val = bool(false)];
            bool w_15_transpose_y_0 = const()[name = string("w_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_15_cast_fp16 = matmul(transpose_x = w_15_transpose_x_0, transpose_y = w_15_transpose_y_0, x = Minv_63_cast_fp16, y = var_3047_cast_fp16)[name = string("w_15_cast_fp16")];
            bool var_3049_transpose_x_0 = const()[name = string("op_3049_transpose_x_0"), val = bool(false)];
            bool var_3049_transpose_y_0 = const()[name = string("op_3049_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3049_cast_fp16 = matmul(transpose_x = var_3049_transpose_x_0, transpose_y = var_3049_transpose_y_0, x = w_15_cast_fp16, y = S_7_cast_fp16)[name = string("op_3049_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_7_cast_fp16 = sub(x = u_15_cast_fp16, y = var_3049_cast_fp16)[name = string("vnew_7_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3054_cast_fp16 = mul(x = q_49_cast_fp16, y = var_3046_cast_fp16)[name = string("op_3054_cast_fp16")];
            bool var_3055_transpose_x_0 = const()[name = string("op_3055_transpose_x_0"), val = bool(false)];
            bool var_3055_transpose_y_0 = const()[name = string("op_3055_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3055_cast_fp16 = matmul(transpose_x = var_3055_transpose_x_0, transpose_y = var_3055_transpose_y_0, x = var_3054_cast_fp16, y = S_7_cast_fp16)[name = string("op_3055_cast_fp16")];
            bool var_3056_transpose_x_0 = const()[name = string("op_3056_transpose_x_0"), val = bool(false)];
            bool var_3056_transpose_y_0 = const()[name = string("op_3056_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3056_cast_fp16 = matmul(transpose_x = var_3056_transpose_x_0, transpose_y = var_3056_transpose_y_0, x = A_15_cast_fp16, y = vnew_7_cast_fp16)[name = string("op_3056_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_7_cast_fp16 = add(x = var_3055_cast_fp16, y = var_3056_cast_fp16)[name = string("oc_7_cast_fp16")];
            int32 var_3060 = const()[name = string("op_3060"), val = int32(2)];
            bool o_21_interleave_0 = const()[name = string("o_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 128]> o_21_cast_fp16 = concat(axis = var_3060, interleave = o_21_interleave_0, values = (var_2911_cast_fp16, oc_7_cast_fp16))[name = string("o_21_cast_fp16")];
            tensor<int32, [4]> var_3066 = const()[name = string("op_3066"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3072 = const()[name = string("op_3072"), val = tensor<int32, [4]>([1, 32, 4, 128])];
            tensor<fp16, [1, 32, 4, 128]> input_65_cast_fp16 = reshape(shape = var_3072, x = z_7_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> o_23_cast_fp16 = transpose(perm = var_3066, x = o_21_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 32, 4, 128]> var_3074_cast_fp16 = mul(x = o_23_cast_fp16, y = o_23_cast_fp16)[name = string("op_3074_cast_fp16")];
            tensor<int32, [1]> var_3079_axes_0 = const()[name = string("op_3079_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3079_keep_dims_0 = const()[name = string("op_3079_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 4, 1]> var_3079_cast_fp16 = reduce_mean(axes = var_3079_axes_0, keep_dims = var_3079_keep_dims_0, x = var_3074_cast_fp16)[name = string("op_3079_cast_fp16")];
            fp16 var_3081_to_fp16 = const()[name = string("op_3081_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 32, 4, 1]> var_3082_cast_fp16 = add(x = var_3079_cast_fp16, y = var_3081_to_fp16)[name = string("op_3082_cast_fp16")];
            fp32 var_3083_epsilon_0 = const()[name = string("op_3083_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 4, 1]> var_3083_cast_fp16 = rsqrt(epsilon = var_3083_epsilon_0, x = var_3082_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> var_3084_cast_fp16 = mul(x = o_23_cast_fp16, y = var_3083_cast_fp16)[name = string("op_3084_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> var_3086_cast_fp16 = silu(x = input_65_cast_fp16)[name = string("op_3086_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> o_25_cast_fp16 = mul(x = var_3084_cast_fp16, y = var_3086_cast_fp16)[name = string("o_25_cast_fp16")];
            tensor<int32, [3]> var_3091 = const()[name = string("op_3091"), val = tensor<int32, [3]>([1, 32, 512])];
            tensor<fp16, [1, 32, 512]> input_67_cast_fp16 = reshape(shape = var_3091, x = o_25_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [640, 512]> layers_4_out_weight_to_fp16 = const()[name = string("layers_4_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72063040)))];
            tensor<fp16, [1, 32, 640]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_out_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 32, 640]> x_49_cast_fp16 = add(x = x_45_cast_fp16, y = linear_17_cast_fp16)[name = string("x_49_cast_fp16")];
            fp16 var_3104_to_fp16 = const()[name = string("op_3104_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 32, 640]> x_51_cast_fp16 = mul(x = x_49_cast_fp16, y = var_3104_to_fp16)[name = string("x_51_cast_fp16")];
            fp16 var_3102_promoted_to_fp16 = const()[name = string("op_3102_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> var_3106_cast_fp16 = pow(x = x_51_cast_fp16, y = var_3102_promoted_to_fp16)[name = string("op_3106_cast_fp16")];
            tensor<int32, [1]> var_3108_axes_0 = const()[name = string("op_3108_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3108_keep_dims_0 = const()[name = string("op_3108_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3108_cast_fp16 = reduce_mean(axes = var_3108_axes_0, keep_dims = var_3108_keep_dims_0, x = var_3106_cast_fp16)[name = string("op_3108_cast_fp16")];
            fp16 var_3109_to_fp16 = const()[name = string("op_3109_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 32, 1]> var_3110_cast_fp16 = add(x = var_3108_cast_fp16, y = var_3109_to_fp16)[name = string("op_3110_cast_fp16")];
            fp32 var_3111_epsilon_0 = const()[name = string("op_3111_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3111_cast_fp16 = rsqrt(epsilon = var_3111_epsilon_0, x = var_3110_cast_fp16)[name = string("op_3111_cast_fp16")];
            tensor<fp16, [1, 32, 640]> var_3112_cast_fp16 = mul(x = x_51_cast_fp16, y = var_3111_cast_fp16)[name = string("op_3112_cast_fp16")];
            tensor<fp16, [640]> layers_4_n2_w1_to_fp16 = const()[name = string("layers_4_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72718464)))];
            tensor<fp16, [1, 32, 640]> input_69_cast_fp16 = mul(x = var_3112_cast_fp16, y = layers_4_n2_w1_to_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_4_gate_up_weight_to_fp16 = const()[name = string("layers_4_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72719808)))];
            tensor<fp16, [1, 32, 5120]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_4_gate_up_weight_to_fp16, x = input_69_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<int32, [3]> input_71_begin_0 = const()[name = string("input_71_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_71_end_0 = const()[name = string("input_71_end_0"), val = tensor<int32, [3]>([1, 32, 2560])];
            tensor<bool, [3]> input_71_end_mask_0 = const()[name = string("input_71_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 2560]> input_71_cast_fp16 = slice_by_index(begin = input_71_begin_0, end = input_71_end_0, end_mask = input_71_end_mask_0, x = linear_18_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [1, 32, 2560]> var_3122_cast_fp16 = silu(x = input_71_cast_fp16)[name = string("op_3122_cast_fp16")];
            tensor<int32, [3]> var_3127_begin_0 = const()[name = string("op_3127_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_3127_end_0 = const()[name = string("op_3127_end_0"), val = tensor<int32, [3]>([1, 32, 5120])];
            tensor<bool, [3]> var_3127_end_mask_0 = const()[name = string("op_3127_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 2560]> var_3127_cast_fp16 = slice_by_index(begin = var_3127_begin_0, end = var_3127_end_0, end_mask = var_3127_end_mask_0, x = linear_18_cast_fp16)[name = string("op_3127_cast_fp16")];
            tensor<fp16, [1, 32, 2560]> input_73_cast_fp16 = mul(x = var_3122_cast_fp16, y = var_3127_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_4_down_weight_to_fp16 = const()[name = string("layers_4_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79273472)))];
            tensor<fp16, [1, 32, 640]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_down_weight_to_fp16, x = input_73_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 32, 640]> x_53_cast_fp16 = add(x = x_49_cast_fp16, y = linear_19_cast_fp16)[name = string("x_53_cast_fp16")];
            fp16 var_3140_to_fp16 = const()[name = string("op_3140_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 32, 640]> x_55_cast_fp16 = mul(x = x_53_cast_fp16, y = var_3140_to_fp16)[name = string("x_55_cast_fp16")];
            fp16 var_3138_promoted_to_fp16 = const()[name = string("op_3138_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> var_3142_cast_fp16 = pow(x = x_55_cast_fp16, y = var_3138_promoted_to_fp16)[name = string("op_3142_cast_fp16")];
            tensor<int32, [1]> var_3144_axes_0 = const()[name = string("op_3144_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3144_keep_dims_0 = const()[name = string("op_3144_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3144_cast_fp16 = reduce_mean(axes = var_3144_axes_0, keep_dims = var_3144_keep_dims_0, x = var_3142_cast_fp16)[name = string("op_3144_cast_fp16")];
            fp16 var_3145_to_fp16 = const()[name = string("op_3145_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 32, 1]> var_3146_cast_fp16 = add(x = var_3144_cast_fp16, y = var_3145_to_fp16)[name = string("op_3146_cast_fp16")];
            fp32 var_3147_epsilon_0 = const()[name = string("op_3147_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3147_cast_fp16 = rsqrt(epsilon = var_3147_epsilon_0, x = var_3146_cast_fp16)[name = string("op_3147_cast_fp16")];
            tensor<fp16, [1, 32, 640]> var_3148_cast_fp16 = mul(x = x_55_cast_fp16, y = var_3147_cast_fp16)[name = string("op_3148_cast_fp16")];
            tensor<fp16, [640]> layers_5_n1_w1_to_fp16 = const()[name = string("layers_5_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82550336)))];
            tensor<fp16, [1, 32, 640]> input_75_cast_fp16 = mul(x = var_3148_cast_fp16, y = layers_5_n1_w1_to_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_5_qkvzba_weight_to_fp16 = const()[name = string("layers_5_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82551680)))];
            tensor<fp16, [1, 32, 1544]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_qkvzba_weight_to_fp16, x = input_75_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [3]> qkv_9_begin_0 = const()[name = string("qkv_9_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_9_end_0 = const()[name = string("qkv_9_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> qkv_9_end_mask_0 = const()[name = string("qkv_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 1024]> qkv_9_cast_fp16 = slice_by_index(begin = qkv_9_begin_0, end = qkv_9_end_0, end_mask = qkv_9_end_mask_0, x = linear_20_cast_fp16)[name = string("qkv_9_cast_fp16")];
            tensor<int32, [3]> z_9_begin_0 = const()[name = string("z_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_9_end_0 = const()[name = string("z_9_end_0"), val = tensor<int32, [3]>([1, 32, 1536])];
            tensor<bool, [3]> z_9_end_mask_0 = const()[name = string("z_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 512]> z_9_cast_fp16 = slice_by_index(begin = z_9_begin_0, end = z_9_end_0, end_mask = z_9_end_mask_0, x = linear_20_cast_fp16)[name = string("z_9_cast_fp16")];
            tensor<int32, [3]> b_9_begin_0 = const()[name = string("b_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_9_end_0 = const()[name = string("b_9_end_0"), val = tensor<int32, [3]>([1, 32, 1540])];
            tensor<bool, [3]> b_9_end_mask_0 = const()[name = string("b_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 4]> b_9_cast_fp16 = slice_by_index(begin = b_9_begin_0, end = b_9_end_0, end_mask = b_9_end_mask_0, x = linear_20_cast_fp16)[name = string("b_9_cast_fp16")];
            tensor<int32, [3]> a_9_begin_0 = const()[name = string("a_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_9_end_0 = const()[name = string("a_9_end_0"), val = tensor<int32, [3]>([1, 32, 1])];
            tensor<bool, [3]> a_9_end_mask_0 = const()[name = string("a_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 4]> a_9_cast_fp16 = slice_by_index(begin = a_9_begin_0, end = a_9_end_0, end_mask = a_9_end_mask_0, x = linear_20_cast_fp16)[name = string("a_9_cast_fp16")];
            int32 var_3187 = const()[name = string("op_3187"), val = int32(1)];
            bool xp_9_interleave_0 = const()[name = string("xp_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 35, 1024]> xp_9_cast_fp16 = concat(axis = var_3187, interleave = xp_9_interleave_0, values = (zpad_to_fp16, qkv_9_cast_fp16))[name = string("xp_9_cast_fp16")];
            tensor<int32, [3]> var_3198_begin_0 = const()[name = string("op_3198_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_3198_end_0 = const()[name = string("op_3198_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_3198_end_mask_0 = const()[name = string("op_3198_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 1024]> var_3198_cast_fp16 = slice_by_index(begin = var_3198_begin_0, end = var_3198_end_0, end_mask = var_3198_end_mask_0, x = xp_9_cast_fp16)[name = string("op_3198_cast_fp16")];
            tensor<fp16, [1024]> var_3206_to_fp16 = const()[name = string("op_3206_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84528064)))];
            tensor<fp16, [1, 32, 1024]> var_3207_cast_fp16 = mul(x = var_3198_cast_fp16, y = var_3206_to_fp16)[name = string("op_3207_cast_fp16")];
            tensor<int32, [3]> var_3217_begin_0 = const()[name = string("op_3217_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_3217_end_0 = const()[name = string("op_3217_end_0"), val = tensor<int32, [3]>([1, 34, 1024])];
            tensor<bool, [3]> var_3217_end_mask_0 = const()[name = string("op_3217_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_3217_cast_fp16 = slice_by_index(begin = var_3217_begin_0, end = var_3217_end_0, end_mask = var_3217_end_mask_0, x = xp_9_cast_fp16)[name = string("op_3217_cast_fp16")];
            tensor<fp16, [1024]> var_3225_to_fp16 = const()[name = string("op_3225_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84530176)))];
            tensor<fp16, [1, 32, 1024]> var_3226_cast_fp16 = mul(x = var_3217_cast_fp16, y = var_3225_to_fp16)[name = string("op_3226_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> var_3228_cast_fp16 = add(x = var_3207_cast_fp16, y = var_3226_cast_fp16)[name = string("op_3228_cast_fp16")];
            tensor<int32, [3]> var_3238_begin_0 = const()[name = string("op_3238_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_3238_end_0 = const()[name = string("op_3238_end_0"), val = tensor<int32, [3]>([1, 33, 1024])];
            tensor<bool, [3]> var_3238_end_mask_0 = const()[name = string("op_3238_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_3238_cast_fp16 = slice_by_index(begin = var_3238_begin_0, end = var_3238_end_0, end_mask = var_3238_end_mask_0, x = xp_9_cast_fp16)[name = string("op_3238_cast_fp16")];
            tensor<fp16, [1024]> var_3246_to_fp16 = const()[name = string("op_3246_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84532288)))];
            tensor<fp16, [1, 32, 1024]> var_3247_cast_fp16 = mul(x = var_3238_cast_fp16, y = var_3246_to_fp16)[name = string("op_3247_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> var_3249_cast_fp16 = add(x = var_3228_cast_fp16, y = var_3247_cast_fp16)[name = string("op_3249_cast_fp16")];
            tensor<int32, [3]> var_3259_begin_0 = const()[name = string("op_3259_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3259_end_0 = const()[name = string("op_3259_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> var_3259_end_mask_0 = const()[name = string("op_3259_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_3259_cast_fp16 = slice_by_index(begin = var_3259_begin_0, end = var_3259_end_0, end_mask = var_3259_end_mask_0, x = xp_9_cast_fp16)[name = string("op_3259_cast_fp16")];
            tensor<fp16, [1024]> var_3267_to_fp16 = const()[name = string("op_3267_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84534400)))];
            tensor<fp16, [1, 32, 1024]> var_3268_cast_fp16 = mul(x = var_3259_cast_fp16, y = var_3267_to_fp16)[name = string("op_3268_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> input_77_cast_fp16 = add(x = var_3249_cast_fp16, y = var_3268_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> c_9_cast_fp16 = silu(x = input_77_cast_fp16)[name = string("c_9_cast_fp16")];
            tensor<int32, [3]> var_3276_begin_0 = const()[name = string("op_3276_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3276_end_0 = const()[name = string("op_3276_end_0"), val = tensor<int32, [3]>([1, 32, 256])];
            tensor<bool, [3]> var_3276_end_mask_0 = const()[name = string("op_3276_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 256]> var_3276_cast_fp16 = slice_by_index(begin = var_3276_begin_0, end = var_3276_end_0, end_mask = var_3276_end_mask_0, x = c_9_cast_fp16)[name = string("op_3276_cast_fp16")];
            tensor<int32, [4]> var_3281 = const()[name = string("op_3281"), val = tensor<int32, [4]>([1, 32, 2, 128])];
            tensor<fp16, [1, 32, 2, 128]> q_51_cast_fp16 = reshape(shape = var_3281, x = var_3276_cast_fp16)[name = string("q_51_cast_fp16")];
            tensor<int32, [3]> var_3287_begin_0 = const()[name = string("op_3287_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_3287_end_0 = const()[name = string("op_3287_end_0"), val = tensor<int32, [3]>([1, 32, 512])];
            tensor<bool, [3]> var_3287_end_mask_0 = const()[name = string("op_3287_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 256]> var_3287_cast_fp16 = slice_by_index(begin = var_3287_begin_0, end = var_3287_end_0, end_mask = var_3287_end_mask_0, x = c_9_cast_fp16)[name = string("op_3287_cast_fp16")];
            tensor<int32, [4]> var_3292 = const()[name = string("op_3292"), val = tensor<int32, [4]>([1, 32, 2, 128])];
            tensor<fp16, [1, 32, 2, 128]> k_53_cast_fp16 = reshape(shape = var_3292, x = var_3287_cast_fp16)[name = string("k_53_cast_fp16")];
            tensor<int32, [3]> var_3298_begin_0 = const()[name = string("op_3298_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_3298_end_0 = const()[name = string("op_3298_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> var_3298_end_mask_0 = const()[name = string("op_3298_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 512]> var_3298_cast_fp16 = slice_by_index(begin = var_3298_begin_0, end = var_3298_end_0, end_mask = var_3298_end_mask_0, x = c_9_cast_fp16)[name = string("op_3298_cast_fp16")];
            tensor<int32, [4]> var_3303 = const()[name = string("op_3303"), val = tensor<int32, [4]>([1, 32, 4, 128])];
            tensor<fp16, [1, 32, 4, 128]> var_3304_cast_fp16 = reshape(shape = var_3303, x = var_3298_cast_fp16)[name = string("op_3304_cast_fp16")];
            tensor<int32, [4]> var_3309 = const()[name = string("op_3309"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3311_to_fp16 = const()[name = string("op_3311_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 32, 2, 128]> q_53_cast_fp16 = mul(x = q_51_cast_fp16, y = var_3311_to_fp16)[name = string("q_53_cast_fp16")];
            fp16 var_3313_to_fp16 = const()[name = string("op_3313_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 32, 2, 128]> k_55_cast_fp16 = mul(x = k_53_cast_fp16, y = var_3313_to_fp16)[name = string("k_55_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_3315_cast_fp16 = mul(x = q_53_cast_fp16, y = q_53_cast_fp16)[name = string("op_3315_cast_fp16")];
            tensor<int32, [1]> var_3320_axes_0 = const()[name = string("op_3320_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3320_keep_dims_0 = const()[name = string("op_3320_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 2, 1]> var_3320_cast_fp16 = reduce_mean(axes = var_3320_axes_0, keep_dims = var_3320_keep_dims_0, x = var_3315_cast_fp16)[name = string("op_3320_cast_fp16")];
            fp16 var_3321_promoted_to_fp16 = const()[name = string("op_3321_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 32, 2, 1]> var_3322_cast_fp16 = mul(x = var_3320_cast_fp16, y = var_3321_promoted_to_fp16)[name = string("op_3322_cast_fp16")];
            fp16 var_3324_to_fp16 = const()[name = string("op_3324_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 32, 2, 1]> var_3325_cast_fp16 = add(x = var_3322_cast_fp16, y = var_3324_to_fp16)[name = string("op_3325_cast_fp16")];
            fp32 var_3326_epsilon_0 = const()[name = string("op_3326_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 2, 1]> var_3326_cast_fp16 = rsqrt(epsilon = var_3326_epsilon_0, x = var_3325_cast_fp16)[name = string("op_3326_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_3327_cast_fp16 = mul(x = q_53_cast_fp16, y = var_3326_cast_fp16)[name = string("op_3327_cast_fp16")];
            fp16 var_3328_to_fp16 = const()[name = string("op_3328_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 32, 2, 128]> q_55_cast_fp16 = mul(x = var_3327_cast_fp16, y = var_3328_to_fp16)[name = string("q_55_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_3330_cast_fp16 = mul(x = k_55_cast_fp16, y = k_55_cast_fp16)[name = string("op_3330_cast_fp16")];
            tensor<int32, [1]> var_3335_axes_0 = const()[name = string("op_3335_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3335_keep_dims_0 = const()[name = string("op_3335_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 2, 1]> var_3335_cast_fp16 = reduce_mean(axes = var_3335_axes_0, keep_dims = var_3335_keep_dims_0, x = var_3330_cast_fp16)[name = string("op_3335_cast_fp16")];
            fp16 var_3336_promoted_to_fp16 = const()[name = string("op_3336_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 32, 2, 1]> var_3337_cast_fp16 = mul(x = var_3335_cast_fp16, y = var_3336_promoted_to_fp16)[name = string("op_3337_cast_fp16")];
            fp16 var_3339_to_fp16 = const()[name = string("op_3339_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 32, 2, 1]> var_3340_cast_fp16 = add(x = var_3337_cast_fp16, y = var_3339_to_fp16)[name = string("op_3340_cast_fp16")];
            fp32 var_3341_epsilon_0 = const()[name = string("op_3341_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 2, 1]> var_3341_cast_fp16 = rsqrt(epsilon = var_3341_epsilon_0, x = var_3340_cast_fp16)[name = string("op_3341_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> k_57_cast_fp16 = mul(x = k_55_cast_fp16, y = var_3341_cast_fp16)[name = string("k_57_cast_fp16")];
            tensor<int32, [4]> var_3347 = const()[name = string("op_3347"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_3360_axes_0 = const()[name = string("op_3360_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 128]> var_3348_cast_fp16 = transpose(perm = var_3347, x = q_55_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 2, 1, 32, 128]> var_3360_cast_fp16 = expand_dims(axes = var_3360_axes_0, x = var_3348_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<int32, [5]> var_3368_reps_0 = const()[name = string("op_3368_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 32, 128]> var_3368_cast_fp16 = tile(reps = var_3368_reps_0, x = var_3360_cast_fp16)[name = string("op_3368_cast_fp16")];
            tensor<int32, [4]> var_3373 = const()[name = string("op_3373"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<fp16, [1, 4, 32, 128]> q_57_cast_fp16 = reshape(shape = var_3373, x = var_3368_cast_fp16)[name = string("q_57_cast_fp16")];
            tensor<int32, [4]> var_3379 = const()[name = string("op_3379"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_3392_axes_0 = const()[name = string("op_3392_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 128]> var_3380_cast_fp16 = transpose(perm = var_3379, x = k_57_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 2, 1, 32, 128]> var_3392_cast_fp16 = expand_dims(axes = var_3392_axes_0, x = var_3380_cast_fp16)[name = string("op_3392_cast_fp16")];
            tensor<int32, [5]> var_3400_reps_0 = const()[name = string("op_3400_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 32, 128]> var_3400_cast_fp16 = tile(reps = var_3400_reps_0, x = var_3392_cast_fp16)[name = string("op_3400_cast_fp16")];
            tensor<int32, [4]> var_3405 = const()[name = string("op_3405"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<fp16, [1, 4, 32, 128]> k_59_cast_fp16 = reshape(shape = var_3405, x = var_3400_cast_fp16)[name = string("k_59_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_3407_cast_fp16 = sigmoid(x = b_9_cast_fp16)[name = string("op_3407_cast_fp16")];
            tensor<int32, [3]> var_3411 = const()[name = string("op_3411"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_5_dt_bias_to_fp16 = const()[name = string("layers_5_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 32, 4]> input_79_cast_fp16 = add(x = a_9_cast_fp16, y = layers_5_dt_bias_to_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_3415_cast_fp16 = relu(x = input_79_cast_fp16)[name = string("op_3415_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_3416_cast_fp16 = abs(x = input_79_cast_fp16)[name = string("op_3416_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, 4]> var_3417_cast_fp16 = mul(x = var_3416_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_3417_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_3418_cast_fp16 = exp(x = var_3417_cast_fp16)[name = string("op_3418_cast_fp16")];
            fp16 var_3420_to_fp16 = const()[name = string("op_3420_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 32, 4]> var_3421_cast_fp16 = add(x = var_3418_cast_fp16, y = var_3420_to_fp16)[name = string("op_3421_cast_fp16")];
            fp32 var_3422_epsilon_0 = const()[name = string("op_3422_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 32, 4]> var_3422_cast_fp16 = log(epsilon = var_3422_epsilon_0, x = var_3421_cast_fp16)[name = string("op_3422_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_3424_cast_fp16 = add(x = var_3415_cast_fp16, y = var_3422_cast_fp16)[name = string("op_3424_cast_fp16")];
            tensor<fp16, [4]> layers_5_negA_to_fp16 = const()[name = string("layers_5_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.2f8p+3, -0x1.ae8p+2, -0x1.548p+2, -0x1.63p+3])];
            tensor<fp16, [1, 32, 4]> var_3425_cast_fp16 = mul(x = layers_5_negA_to_fp16, y = var_3424_cast_fp16)[name = string("op_3425_cast_fp16")];
            tensor<int32, [3]> var_3429 = const()[name = string("op_3429"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_3432_axes_0 = const()[name = string("op_3432_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 32]> beta_9_cast_fp16 = transpose(perm = var_3411, x = var_3407_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 4, 32, 1]> var_3432_cast_fp16 = expand_dims(axes = var_3432_axes_0, x = beta_9_cast_fp16)[name = string("op_3432_cast_fp16")];
            fp16 var_3433_to_fp16 = const()[name = string("op_3433_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 32, 1]> var_3434_cast_fp16 = mul(x = var_3432_cast_fp16, y = var_3433_to_fp16)[name = string("op_3434_cast_fp16")];
            tensor<fp16, [1, 4, 32, 128]> v_13_cast_fp16 = transpose(perm = var_3309, x = var_3304_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 4, 32, 128]> vb_25_cast_fp16 = mul(x = v_13_cast_fp16, y = var_3434_cast_fp16)[name = string("vb_25_cast_fp16")];
            tensor<int32, [4]> q_59_begin_0 = const()[name = string("q_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_59_end_0 = const()[name = string("q_59_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_59_end_mask_0 = const()[name = string("q_59_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_59_cast_fp16 = slice_by_index(begin = q_59_begin_0, end = q_59_end_0, end_mask = q_59_end_mask_0, x = q_57_cast_fp16)[name = string("q_59_cast_fp16")];
            tensor<int32, [4]> k_61_begin_0 = const()[name = string("k_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_61_end_0 = const()[name = string("k_61_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_61_end_mask_0 = const()[name = string("k_61_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, x = k_59_cast_fp16)[name = string("k_61_cast_fp16")];
            tensor<fp16, [1, 4, 32, 128]> var_3468_cast_fp16 = mul(x = k_59_cast_fp16, y = var_3432_cast_fp16)[name = string("op_3468_cast_fp16")];
            tensor<int32, [4]> kb_17_begin_0 = const()[name = string("kb_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_17_end_0 = const()[name = string("kb_17_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_17_end_mask_0 = const()[name = string("kb_17_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_17_cast_fp16 = slice_by_index(begin = kb_17_begin_0, end = kb_17_end_0, end_mask = kb_17_end_mask_0, x = var_3468_cast_fp16)[name = string("kb_17_cast_fp16")];
            tensor<int32, [4]> vb_27_begin_0 = const()[name = string("vb_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_27_end_0 = const()[name = string("vb_27_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_27_end_mask_0 = const()[name = string("vb_27_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_27_cast_fp16 = slice_by_index(begin = vb_27_begin_0, end = vb_27_end_0, end_mask = vb_27_end_mask_0, x = vb_25_cast_fp16)[name = string("vb_27_cast_fp16")];
            tensor<int32, [3]> gdec_27_begin_0 = const()[name = string("gdec_27_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_27_end_0 = const()[name = string("gdec_27_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_27_end_mask_0 = const()[name = string("gdec_27_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 32]> gdec_25_cast_fp16 = transpose(perm = var_3429, x = var_3425_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 4, 16]> gdec_27_cast_fp16 = slice_by_index(begin = gdec_27_begin_0, end = gdec_27_end_0, end_mask = gdec_27_end_mask_0, x = gdec_25_cast_fp16)[name = string("gdec_27_cast_fp16")];
            tensor<int32, [2]> var_3516 = const()[name = string("op_3516"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_3517_cast_fp16 = reshape(shape = var_3516, x = gdec_27_cast_fp16)[name = string("op_3517_cast_fp16")];
            tensor<fp16, [4, 256]> var_3518_cast_fp16 = linear(bias = var_633_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3517_cast_fp16)[name = string("op_3518_cast_fp16")];
            tensor<int32, [4]> var_3523 = const()[name = string("op_3523"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_17_cast_fp16 = reshape(shape = var_3523, x = var_3518_cast_fp16)[name = string("P_17_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3525_cast_fp16 = exp(x = P_17_cast_fp16)[name = string("op_3525_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_17_cast_fp16 = mul(x = var_3525_cast_fp16, y = tril_incl_to_fp16)[name = string("D_17_cast_fp16")];
            bool kk_17_transpose_x_1 = const()[name = string("kk_17_transpose_x_1"), val = bool(false)];
            bool kk_17_transpose_y_1 = const()[name = string("kk_17_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_17_cast_fp16 = matmul(transpose_x = kk_17_transpose_x_1, transpose_y = kk_17_transpose_y_1, x = kb_17_cast_fp16, y = k_61_cast_fp16)[name = string("kk_17_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_3531_cast_fp16 = mul(x = kk_17_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_3531_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3532_cast_fp16 = mul(x = var_3531_cast_fp16, y = D_17_cast_fp16)[name = string("op_3532_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_17_cast_fp16 = mul(x = var_3532_cast_fp16, y = cum_tail_to_fp16)[name = string("N_17_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3534_cast_fp16 = mul(x = D_17_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3534_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_17_cast_fp16 = mul(x = kk_17_cast_fp16, y = var_3534_cast_fp16)[name = string("Na_17_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_65_cast_fp16 = add(x = eye_to_fp16, y = N_17_cast_fp16)[name = string("Minv_65_cast_fp16")];
            bool Np_49_transpose_x_0 = const()[name = string("Np_49_transpose_x_0"), val = bool(false)];
            bool Np_49_transpose_y_0 = const()[name = string("Np_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_49_cast_fp16 = matmul(transpose_x = Np_49_transpose_x_0, transpose_y = Np_49_transpose_y_0, x = N_17_cast_fp16, y = Na_17_cast_fp16)[name = string("Np_49_cast_fp16")];
            bool Npa_33_transpose_x_0 = const()[name = string("Npa_33_transpose_x_0"), val = bool(false)];
            bool Npa_33_transpose_y_0 = const()[name = string("Npa_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_33_cast_fp16 = matmul(transpose_x = Npa_33_transpose_x_0, transpose_y = Npa_33_transpose_y_0, x = Na_17_cast_fp16, y = N_17_cast_fp16)[name = string("Npa_33_cast_fp16")];
            bool var_3540_transpose_x_0 = const()[name = string("op_3540_transpose_x_0"), val = bool(false)];
            bool var_3540_transpose_y_0 = const()[name = string("op_3540_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3540_cast_fp16 = matmul(transpose_x = var_3540_transpose_x_0, transpose_y = var_3540_transpose_y_0, x = Minv_65_cast_fp16, y = Np_49_cast_fp16)[name = string("op_3540_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_67_cast_fp16 = add(x = Minv_65_cast_fp16, y = var_3540_cast_fp16)[name = string("Minv_67_cast_fp16")];
            bool Np_51_transpose_x_0 = const()[name = string("Np_51_transpose_x_0"), val = bool(false)];
            bool Np_51_transpose_y_0 = const()[name = string("Np_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_51_cast_fp16 = matmul(transpose_x = Np_51_transpose_x_0, transpose_y = Np_51_transpose_y_0, x = Np_49_cast_fp16, y = Npa_33_cast_fp16)[name = string("Np_51_cast_fp16")];
            bool Npa_35_transpose_x_0 = const()[name = string("Npa_35_transpose_x_0"), val = bool(false)];
            bool Npa_35_transpose_y_0 = const()[name = string("Npa_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_35_cast_fp16 = matmul(transpose_x = Npa_35_transpose_x_0, transpose_y = Npa_35_transpose_y_0, x = Npa_33_cast_fp16, y = Np_49_cast_fp16)[name = string("Npa_35_cast_fp16")];
            bool var_3545_transpose_x_0 = const()[name = string("op_3545_transpose_x_0"), val = bool(false)];
            bool var_3545_transpose_y_0 = const()[name = string("op_3545_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3545_cast_fp16 = matmul(transpose_x = var_3545_transpose_x_0, transpose_y = var_3545_transpose_y_0, x = Minv_67_cast_fp16, y = Np_51_cast_fp16)[name = string("op_3545_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_69_cast_fp16 = add(x = Minv_67_cast_fp16, y = var_3545_cast_fp16)[name = string("Minv_69_cast_fp16")];
            bool Np_53_transpose_x_0 = const()[name = string("Np_53_transpose_x_0"), val = bool(false)];
            bool Np_53_transpose_y_0 = const()[name = string("Np_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_53_cast_fp16 = matmul(transpose_x = Np_53_transpose_x_0, transpose_y = Np_53_transpose_y_0, x = Np_51_cast_fp16, y = Npa_35_cast_fp16)[name = string("Np_53_cast_fp16")];
            bool var_3549_transpose_x_0 = const()[name = string("op_3549_transpose_x_0"), val = bool(false)];
            bool var_3549_transpose_y_0 = const()[name = string("op_3549_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3549_cast_fp16 = matmul(transpose_x = var_3549_transpose_x_0, transpose_y = var_3549_transpose_y_0, x = Minv_69_cast_fp16, y = Np_53_cast_fp16)[name = string("op_3549_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_71_cast_fp16 = add(x = Minv_69_cast_fp16, y = var_3549_cast_fp16)[name = string("Minv_71_cast_fp16")];
            bool u_17_transpose_x_0 = const()[name = string("u_17_transpose_x_0"), val = bool(false)];
            bool u_17_transpose_y_0 = const()[name = string("u_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_17_cast_fp16 = matmul(transpose_x = u_17_transpose_x_0, transpose_y = u_17_transpose_y_0, x = Minv_71_cast_fp16, y = vb_27_cast_fp16)[name = string("u_17_cast_fp16")];
            bool var_3556_transpose_x_1 = const()[name = string("op_3556_transpose_x_1"), val = bool(false)];
            bool var_3556_transpose_y_1 = const()[name = string("op_3556_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_3556_cast_fp16 = matmul(transpose_x = var_3556_transpose_x_1, transpose_y = var_3556_transpose_y_1, x = q_59_cast_fp16, y = k_61_cast_fp16)[name = string("op_3556_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_17_cast_fp16 = mul(x = var_3556_cast_fp16, y = D_17_cast_fp16)[name = string("A_17_cast_fp16")];
            bool var_3558_transpose_x_0 = const()[name = string("op_3558_transpose_x_0"), val = bool(false)];
            bool var_3558_transpose_y_0 = const()[name = string("op_3558_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3558_cast_fp16 = matmul(transpose_x = var_3558_transpose_x_0, transpose_y = var_3558_transpose_y_0, x = A_17_cast_fp16, y = u_17_cast_fp16)[name = string("op_3558_cast_fp16")];
            tensor<fp16, [16]> var_3559_bias_0_to_fp16 = const()[name = string("op_3559_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_3559_cast_fp16 = linear(bias = var_3559_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_27_cast_fp16)[name = string("op_3559_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_9_cast_fp16 = exp(x = var_3559_cast_fp16)[name = string("tail_9_cast_fp16")];
            tensor<int32, [1]> var_3562_axes_0 = const()[name = string("op_3562_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_3562_cast_fp16 = expand_dims(axes = var_3562_axes_0, x = tail_9_cast_fp16)[name = string("op_3562_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3563_cast_fp16 = mul(x = k_61_cast_fp16, y = var_3562_cast_fp16)[name = string("op_3563_cast_fp16")];
            bool S_9_transpose_x_1 = const()[name = string("S_9_transpose_x_1"), val = bool(true)];
            bool S_9_transpose_y_1 = const()[name = string("S_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_9_cast_fp16 = matmul(transpose_x = S_9_transpose_x_1, transpose_y = S_9_transpose_y_1, x = var_3563_cast_fp16, y = u_17_cast_fp16)[name = string("S_9_cast_fp16")];
            tensor<int32, [4]> q_61_begin_0 = const()[name = string("q_61_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_61_end_0 = const()[name = string("q_61_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_61_end_mask_0 = const()[name = string("q_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, x = q_57_cast_fp16)[name = string("q_61_cast_fp16")];
            tensor<int32, [4]> k_63_begin_0 = const()[name = string("k_63_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_63_end_0 = const()[name = string("k_63_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_63_end_mask_0 = const()[name = string("k_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_63_cast_fp16 = slice_by_index(begin = k_63_begin_0, end = k_63_end_0, end_mask = k_63_end_mask_0, x = k_59_cast_fp16)[name = string("k_63_cast_fp16")];
            tensor<int32, [4]> kb_19_begin_0 = const()[name = string("kb_19_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_19_end_0 = const()[name = string("kb_19_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_19_end_mask_0 = const()[name = string("kb_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_19_cast_fp16 = slice_by_index(begin = kb_19_begin_0, end = kb_19_end_0, end_mask = kb_19_end_mask_0, x = var_3468_cast_fp16)[name = string("kb_19_cast_fp16")];
            tensor<int32, [4]> vb_29_begin_0 = const()[name = string("vb_29_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_29_end_0 = const()[name = string("vb_29_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_29_end_mask_0 = const()[name = string("vb_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_29_cast_fp16 = slice_by_index(begin = vb_29_begin_0, end = vb_29_end_0, end_mask = vb_29_end_mask_0, x = vb_25_cast_fp16)[name = string("vb_29_cast_fp16")];
            tensor<int32, [3]> gdec_29_begin_0 = const()[name = string("gdec_29_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_29_end_0 = const()[name = string("gdec_29_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_29_end_mask_0 = const()[name = string("gdec_29_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_29_cast_fp16 = slice_by_index(begin = gdec_29_begin_0, end = gdec_29_end_0, end_mask = gdec_29_end_mask_0, x = gdec_25_cast_fp16)[name = string("gdec_29_cast_fp16")];
            tensor<int32, [2]> var_3648 = const()[name = string("op_3648"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_3649_cast_fp16 = reshape(shape = var_3648, x = gdec_29_cast_fp16)[name = string("op_3649_cast_fp16")];
            tensor<fp16, [4, 256]> var_3650_cast_fp16 = linear(bias = var_633_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3649_cast_fp16)[name = string("op_3650_cast_fp16")];
            tensor<int32, [4]> var_3655 = const()[name = string("op_3655"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_19_cast_fp16 = reshape(shape = var_3655, x = var_3650_cast_fp16)[name = string("P_19_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3657_cast_fp16 = exp(x = P_19_cast_fp16)[name = string("op_3657_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_19_cast_fp16 = mul(x = var_3657_cast_fp16, y = tril_incl_to_fp16)[name = string("D_19_cast_fp16")];
            bool kk_19_transpose_x_1 = const()[name = string("kk_19_transpose_x_1"), val = bool(false)];
            bool kk_19_transpose_y_1 = const()[name = string("kk_19_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_19_cast_fp16 = matmul(transpose_x = kk_19_transpose_x_1, transpose_y = kk_19_transpose_y_1, x = kb_19_cast_fp16, y = k_63_cast_fp16)[name = string("kk_19_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_3663_cast_fp16 = mul(x = kk_19_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_3663_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3664_cast_fp16 = mul(x = var_3663_cast_fp16, y = D_19_cast_fp16)[name = string("op_3664_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_19_cast_fp16 = mul(x = var_3664_cast_fp16, y = cum_tail_to_fp16)[name = string("N_19_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3666_cast_fp16 = mul(x = D_19_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3666_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_19_cast_fp16 = mul(x = kk_19_cast_fp16, y = var_3666_cast_fp16)[name = string("Na_19_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_73_cast_fp16 = add(x = eye_to_fp16, y = N_19_cast_fp16)[name = string("Minv_73_cast_fp16")];
            bool Np_55_transpose_x_0 = const()[name = string("Np_55_transpose_x_0"), val = bool(false)];
            bool Np_55_transpose_y_0 = const()[name = string("Np_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_55_cast_fp16 = matmul(transpose_x = Np_55_transpose_x_0, transpose_y = Np_55_transpose_y_0, x = N_19_cast_fp16, y = Na_19_cast_fp16)[name = string("Np_55_cast_fp16")];
            bool Npa_37_transpose_x_0 = const()[name = string("Npa_37_transpose_x_0"), val = bool(false)];
            bool Npa_37_transpose_y_0 = const()[name = string("Npa_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_37_cast_fp16 = matmul(transpose_x = Npa_37_transpose_x_0, transpose_y = Npa_37_transpose_y_0, x = Na_19_cast_fp16, y = N_19_cast_fp16)[name = string("Npa_37_cast_fp16")];
            bool var_3672_transpose_x_0 = const()[name = string("op_3672_transpose_x_0"), val = bool(false)];
            bool var_3672_transpose_y_0 = const()[name = string("op_3672_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3672_cast_fp16 = matmul(transpose_x = var_3672_transpose_x_0, transpose_y = var_3672_transpose_y_0, x = Minv_73_cast_fp16, y = Np_55_cast_fp16)[name = string("op_3672_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_75_cast_fp16 = add(x = Minv_73_cast_fp16, y = var_3672_cast_fp16)[name = string("Minv_75_cast_fp16")];
            bool Np_57_transpose_x_0 = const()[name = string("Np_57_transpose_x_0"), val = bool(false)];
            bool Np_57_transpose_y_0 = const()[name = string("Np_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_57_cast_fp16 = matmul(transpose_x = Np_57_transpose_x_0, transpose_y = Np_57_transpose_y_0, x = Np_55_cast_fp16, y = Npa_37_cast_fp16)[name = string("Np_57_cast_fp16")];
            bool Npa_39_transpose_x_0 = const()[name = string("Npa_39_transpose_x_0"), val = bool(false)];
            bool Npa_39_transpose_y_0 = const()[name = string("Npa_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_39_cast_fp16 = matmul(transpose_x = Npa_39_transpose_x_0, transpose_y = Npa_39_transpose_y_0, x = Npa_37_cast_fp16, y = Np_55_cast_fp16)[name = string("Npa_39_cast_fp16")];
            bool var_3677_transpose_x_0 = const()[name = string("op_3677_transpose_x_0"), val = bool(false)];
            bool var_3677_transpose_y_0 = const()[name = string("op_3677_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3677_cast_fp16 = matmul(transpose_x = var_3677_transpose_x_0, transpose_y = var_3677_transpose_y_0, x = Minv_75_cast_fp16, y = Np_57_cast_fp16)[name = string("op_3677_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_77_cast_fp16 = add(x = Minv_75_cast_fp16, y = var_3677_cast_fp16)[name = string("Minv_77_cast_fp16")];
            bool Np_59_transpose_x_0 = const()[name = string("Np_59_transpose_x_0"), val = bool(false)];
            bool Np_59_transpose_y_0 = const()[name = string("Np_59_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_59_cast_fp16 = matmul(transpose_x = Np_59_transpose_x_0, transpose_y = Np_59_transpose_y_0, x = Np_57_cast_fp16, y = Npa_39_cast_fp16)[name = string("Np_59_cast_fp16")];
            bool var_3681_transpose_x_0 = const()[name = string("op_3681_transpose_x_0"), val = bool(false)];
            bool var_3681_transpose_y_0 = const()[name = string("op_3681_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3681_cast_fp16 = matmul(transpose_x = var_3681_transpose_x_0, transpose_y = var_3681_transpose_y_0, x = Minv_77_cast_fp16, y = Np_59_cast_fp16)[name = string("op_3681_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_79_cast_fp16 = add(x = Minv_77_cast_fp16, y = var_3681_cast_fp16)[name = string("Minv_79_cast_fp16")];
            bool u_19_transpose_x_0 = const()[name = string("u_19_transpose_x_0"), val = bool(false)];
            bool u_19_transpose_y_0 = const()[name = string("u_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_19_cast_fp16 = matmul(transpose_x = u_19_transpose_x_0, transpose_y = u_19_transpose_y_0, x = Minv_79_cast_fp16, y = vb_29_cast_fp16)[name = string("u_19_cast_fp16")];
            bool var_3688_transpose_x_1 = const()[name = string("op_3688_transpose_x_1"), val = bool(false)];
            bool var_3688_transpose_y_1 = const()[name = string("op_3688_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_3688_cast_fp16 = matmul(transpose_x = var_3688_transpose_x_1, transpose_y = var_3688_transpose_y_1, x = q_61_cast_fp16, y = k_63_cast_fp16)[name = string("op_3688_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_19_cast_fp16 = mul(x = var_3688_cast_fp16, y = D_19_cast_fp16)[name = string("A_19_cast_fp16")];
            tensor<fp16, [16]> var_3690_bias_0_to_fp16 = const()[name = string("op_3690_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_3690_cast_fp16 = linear(bias = var_3690_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_29_cast_fp16)[name = string("op_3690_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_9_cast_fp16 = exp(x = var_3690_cast_fp16)[name = string("cum_9_cast_fp16")];
            tensor<int32, [1]> var_3693_axes_0 = const()[name = string("op_3693_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_3693_cast_fp16 = expand_dims(axes = var_3693_axes_0, x = cum_9_cast_fp16)[name = string("op_3693_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3694_cast_fp16 = mul(x = kb_19_cast_fp16, y = var_3693_cast_fp16)[name = string("op_3694_cast_fp16")];
            bool w_19_transpose_x_0 = const()[name = string("w_19_transpose_x_0"), val = bool(false)];
            bool w_19_transpose_y_0 = const()[name = string("w_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_19_cast_fp16 = matmul(transpose_x = w_19_transpose_x_0, transpose_y = w_19_transpose_y_0, x = Minv_79_cast_fp16, y = var_3694_cast_fp16)[name = string("w_19_cast_fp16")];
            bool var_3696_transpose_x_0 = const()[name = string("op_3696_transpose_x_0"), val = bool(false)];
            bool var_3696_transpose_y_0 = const()[name = string("op_3696_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3696_cast_fp16 = matmul(transpose_x = var_3696_transpose_x_0, transpose_y = var_3696_transpose_y_0, x = w_19_cast_fp16, y = S_9_cast_fp16)[name = string("op_3696_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_9_cast_fp16 = sub(x = u_19_cast_fp16, y = var_3696_cast_fp16)[name = string("vnew_9_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3701_cast_fp16 = mul(x = q_61_cast_fp16, y = var_3693_cast_fp16)[name = string("op_3701_cast_fp16")];
            bool var_3702_transpose_x_0 = const()[name = string("op_3702_transpose_x_0"), val = bool(false)];
            bool var_3702_transpose_y_0 = const()[name = string("op_3702_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3702_cast_fp16 = matmul(transpose_x = var_3702_transpose_x_0, transpose_y = var_3702_transpose_y_0, x = var_3701_cast_fp16, y = S_9_cast_fp16)[name = string("op_3702_cast_fp16")];
            bool var_3703_transpose_x_0 = const()[name = string("op_3703_transpose_x_0"), val = bool(false)];
            bool var_3703_transpose_y_0 = const()[name = string("op_3703_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3703_cast_fp16 = matmul(transpose_x = var_3703_transpose_x_0, transpose_y = var_3703_transpose_y_0, x = A_19_cast_fp16, y = vnew_9_cast_fp16)[name = string("op_3703_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_9_cast_fp16 = add(x = var_3702_cast_fp16, y = var_3703_cast_fp16)[name = string("oc_9_cast_fp16")];
            int32 var_3707 = const()[name = string("op_3707"), val = int32(2)];
            bool o_27_interleave_0 = const()[name = string("o_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 128]> o_27_cast_fp16 = concat(axis = var_3707, interleave = o_27_interleave_0, values = (var_3558_cast_fp16, oc_9_cast_fp16))[name = string("o_27_cast_fp16")];
            tensor<int32, [4]> var_3713 = const()[name = string("op_3713"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3719 = const()[name = string("op_3719"), val = tensor<int32, [4]>([1, 32, 4, 128])];
            tensor<fp16, [1, 32, 4, 128]> input_81_cast_fp16 = reshape(shape = var_3719, x = z_9_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> o_29_cast_fp16 = transpose(perm = var_3713, x = o_27_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 32, 4, 128]> var_3721_cast_fp16 = mul(x = o_29_cast_fp16, y = o_29_cast_fp16)[name = string("op_3721_cast_fp16")];
            tensor<int32, [1]> var_3726_axes_0 = const()[name = string("op_3726_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3726_keep_dims_0 = const()[name = string("op_3726_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 4, 1]> var_3726_cast_fp16 = reduce_mean(axes = var_3726_axes_0, keep_dims = var_3726_keep_dims_0, x = var_3721_cast_fp16)[name = string("op_3726_cast_fp16")];
            fp16 var_3728_to_fp16 = const()[name = string("op_3728_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 32, 4, 1]> var_3729_cast_fp16 = add(x = var_3726_cast_fp16, y = var_3728_to_fp16)[name = string("op_3729_cast_fp16")];
            fp32 var_3730_epsilon_0 = const()[name = string("op_3730_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 4, 1]> var_3730_cast_fp16 = rsqrt(epsilon = var_3730_epsilon_0, x = var_3729_cast_fp16)[name = string("op_3730_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> var_3731_cast_fp16 = mul(x = o_29_cast_fp16, y = var_3730_cast_fp16)[name = string("op_3731_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> var_3733_cast_fp16 = silu(x = input_81_cast_fp16)[name = string("op_3733_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> o_31_cast_fp16 = mul(x = var_3731_cast_fp16, y = var_3733_cast_fp16)[name = string("o_31_cast_fp16")];
            tensor<int32, [3]> var_3738 = const()[name = string("op_3738"), val = tensor<int32, [3]>([1, 32, 512])];
            tensor<fp16, [1, 32, 512]> input_83_cast_fp16 = reshape(shape = var_3738, x = o_31_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [640, 512]> layers_5_out_weight_to_fp16 = const()[name = string("layers_5_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84536512)))];
            tensor<fp16, [1, 32, 640]> linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_out_weight_to_fp16, x = input_83_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [1, 32, 640]> x_57_cast_fp16 = add(x = x_53_cast_fp16, y = linear_21_cast_fp16)[name = string("x_57_cast_fp16")];
            fp16 var_3751_to_fp16 = const()[name = string("op_3751_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 32, 640]> x_59_cast_fp16 = mul(x = x_57_cast_fp16, y = var_3751_to_fp16)[name = string("x_59_cast_fp16")];
            fp16 var_3749_promoted_to_fp16 = const()[name = string("op_3749_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> var_3753_cast_fp16 = pow(x = x_59_cast_fp16, y = var_3749_promoted_to_fp16)[name = string("op_3753_cast_fp16")];
            tensor<int32, [1]> var_3755_axes_0 = const()[name = string("op_3755_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3755_keep_dims_0 = const()[name = string("op_3755_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3755_cast_fp16 = reduce_mean(axes = var_3755_axes_0, keep_dims = var_3755_keep_dims_0, x = var_3753_cast_fp16)[name = string("op_3755_cast_fp16")];
            fp16 var_3756_to_fp16 = const()[name = string("op_3756_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 32, 1]> var_3757_cast_fp16 = add(x = var_3755_cast_fp16, y = var_3756_to_fp16)[name = string("op_3757_cast_fp16")];
            fp32 var_3758_epsilon_0 = const()[name = string("op_3758_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3758_cast_fp16 = rsqrt(epsilon = var_3758_epsilon_0, x = var_3757_cast_fp16)[name = string("op_3758_cast_fp16")];
            tensor<fp16, [1, 32, 640]> var_3759_cast_fp16 = mul(x = x_59_cast_fp16, y = var_3758_cast_fp16)[name = string("op_3759_cast_fp16")];
            tensor<fp16, [640]> layers_5_n2_w1_to_fp16 = const()[name = string("layers_5_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85191936)))];
            tensor<fp16, [1, 32, 640]> input_85_cast_fp16 = mul(x = var_3759_cast_fp16, y = layers_5_n2_w1_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_5_gate_up_weight_to_fp16 = const()[name = string("layers_5_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85193280)))];
            tensor<fp16, [1, 32, 5120]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_5_gate_up_weight_to_fp16, x = input_85_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [3]> input_87_begin_0 = const()[name = string("input_87_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_87_end_0 = const()[name = string("input_87_end_0"), val = tensor<int32, [3]>([1, 32, 2560])];
            tensor<bool, [3]> input_87_end_mask_0 = const()[name = string("input_87_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 2560]> input_87_cast_fp16 = slice_by_index(begin = input_87_begin_0, end = input_87_end_0, end_mask = input_87_end_mask_0, x = linear_22_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [1, 32, 2560]> var_3769_cast_fp16 = silu(x = input_87_cast_fp16)[name = string("op_3769_cast_fp16")];
            tensor<int32, [3]> var_3774_begin_0 = const()[name = string("op_3774_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_3774_end_0 = const()[name = string("op_3774_end_0"), val = tensor<int32, [3]>([1, 32, 5120])];
            tensor<bool, [3]> var_3774_end_mask_0 = const()[name = string("op_3774_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 2560]> var_3774_cast_fp16 = slice_by_index(begin = var_3774_begin_0, end = var_3774_end_0, end_mask = var_3774_end_mask_0, x = linear_22_cast_fp16)[name = string("op_3774_cast_fp16")];
            tensor<fp16, [1, 32, 2560]> input_89_cast_fp16 = mul(x = var_3769_cast_fp16, y = var_3774_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_5_down_weight_to_fp16 = const()[name = string("layers_5_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91746944)))];
            tensor<fp16, [1, 32, 640]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_down_weight_to_fp16, x = input_89_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<fp16, [1, 32, 640]> x_61_cast_fp16 = add(x = x_57_cast_fp16, y = linear_23_cast_fp16)[name = string("x_61_cast_fp16")];
            fp16 var_3787_to_fp16 = const()[name = string("op_3787_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [1, 32, 640]> x_63_cast_fp16 = mul(x = x_61_cast_fp16, y = var_3787_to_fp16)[name = string("x_63_cast_fp16")];
            fp16 var_3785_promoted_to_fp16 = const()[name = string("op_3785_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> var_3789_cast_fp16 = pow(x = x_63_cast_fp16, y = var_3785_promoted_to_fp16)[name = string("op_3789_cast_fp16")];
            tensor<int32, [1]> var_3791_axes_0 = const()[name = string("op_3791_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3791_keep_dims_0 = const()[name = string("op_3791_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3791_cast_fp16 = reduce_mean(axes = var_3791_axes_0, keep_dims = var_3791_keep_dims_0, x = var_3789_cast_fp16)[name = string("op_3791_cast_fp16")];
            fp16 var_3792_to_fp16 = const()[name = string("op_3792_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 32, 1]> var_3793_cast_fp16 = add(x = var_3791_cast_fp16, y = var_3792_to_fp16)[name = string("op_3793_cast_fp16")];
            fp32 var_3794_epsilon_0 = const()[name = string("op_3794_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3794_cast_fp16 = rsqrt(epsilon = var_3794_epsilon_0, x = var_3793_cast_fp16)[name = string("op_3794_cast_fp16")];
            tensor<fp16, [1, 32, 640]> var_3795_cast_fp16 = mul(x = x_63_cast_fp16, y = var_3794_cast_fp16)[name = string("op_3795_cast_fp16")];
            tensor<fp16, [640]> layers_6_n1_w1_to_fp16 = const()[name = string("layers_6_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95023808)))];
            tensor<fp16, [1, 32, 640]> input_91_cast_fp16 = mul(x = var_3795_cast_fp16, y = layers_6_n1_w1_to_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_6_qkvzba_weight_to_fp16 = const()[name = string("layers_6_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95025152)))];
            tensor<fp16, [1, 32, 1544]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_qkvzba_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<int32, [3]> qkv_11_begin_0 = const()[name = string("qkv_11_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_11_end_0 = const()[name = string("qkv_11_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> qkv_11_end_mask_0 = const()[name = string("qkv_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 1024]> qkv_11_cast_fp16 = slice_by_index(begin = qkv_11_begin_0, end = qkv_11_end_0, end_mask = qkv_11_end_mask_0, x = linear_24_cast_fp16)[name = string("qkv_11_cast_fp16")];
            tensor<int32, [3]> z_begin_0 = const()[name = string("z_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_end_0 = const()[name = string("z_end_0"), val = tensor<int32, [3]>([1, 32, 1536])];
            tensor<bool, [3]> z_end_mask_0 = const()[name = string("z_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 512]> z_cast_fp16 = slice_by_index(begin = z_begin_0, end = z_end_0, end_mask = z_end_mask_0, x = linear_24_cast_fp16)[name = string("z_cast_fp16")];
            tensor<int32, [3]> b_begin_0 = const()[name = string("b_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_end_0 = const()[name = string("b_end_0"), val = tensor<int32, [3]>([1, 32, 1540])];
            tensor<bool, [3]> b_end_mask_0 = const()[name = string("b_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 4]> b_cast_fp16 = slice_by_index(begin = b_begin_0, end = b_end_0, end_mask = b_end_mask_0, x = linear_24_cast_fp16)[name = string("b_cast_fp16")];
            tensor<int32, [3]> a_begin_0 = const()[name = string("a_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_end_0 = const()[name = string("a_end_0"), val = tensor<int32, [3]>([1, 32, 1])];
            tensor<bool, [3]> a_end_mask_0 = const()[name = string("a_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 4]> a_cast_fp16 = slice_by_index(begin = a_begin_0, end = a_end_0, end_mask = a_end_mask_0, x = linear_24_cast_fp16)[name = string("a_cast_fp16")];
            int32 var_3834 = const()[name = string("op_3834"), val = int32(1)];
            bool xp_interleave_0 = const()[name = string("xp_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 35, 1024]> xp_cast_fp16 = concat(axis = var_3834, interleave = xp_interleave_0, values = (zpad_to_fp16, qkv_11_cast_fp16))[name = string("xp_cast_fp16")];
            tensor<int32, [3]> var_3845_begin_0 = const()[name = string("op_3845_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_3845_end_0 = const()[name = string("op_3845_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_3845_end_mask_0 = const()[name = string("op_3845_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 1024]> var_3845_cast_fp16 = slice_by_index(begin = var_3845_begin_0, end = var_3845_end_0, end_mask = var_3845_end_mask_0, x = xp_cast_fp16)[name = string("op_3845_cast_fp16")];
            tensor<fp16, [1024]> var_3853_to_fp16 = const()[name = string("op_3853_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97001536)))];
            tensor<fp16, [1, 32, 1024]> var_3854_cast_fp16 = mul(x = var_3845_cast_fp16, y = var_3853_to_fp16)[name = string("op_3854_cast_fp16")];
            tensor<int32, [3]> var_3864_begin_0 = const()[name = string("op_3864_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_3864_end_0 = const()[name = string("op_3864_end_0"), val = tensor<int32, [3]>([1, 34, 1024])];
            tensor<bool, [3]> var_3864_end_mask_0 = const()[name = string("op_3864_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_3864_cast_fp16 = slice_by_index(begin = var_3864_begin_0, end = var_3864_end_0, end_mask = var_3864_end_mask_0, x = xp_cast_fp16)[name = string("op_3864_cast_fp16")];
            tensor<fp16, [1024]> var_3872_to_fp16 = const()[name = string("op_3872_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97003648)))];
            tensor<fp16, [1, 32, 1024]> var_3873_cast_fp16 = mul(x = var_3864_cast_fp16, y = var_3872_to_fp16)[name = string("op_3873_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> var_3875_cast_fp16 = add(x = var_3854_cast_fp16, y = var_3873_cast_fp16)[name = string("op_3875_cast_fp16")];
            tensor<int32, [3]> var_3885_begin_0 = const()[name = string("op_3885_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_3885_end_0 = const()[name = string("op_3885_end_0"), val = tensor<int32, [3]>([1, 33, 1024])];
            tensor<bool, [3]> var_3885_end_mask_0 = const()[name = string("op_3885_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_3885_cast_fp16 = slice_by_index(begin = var_3885_begin_0, end = var_3885_end_0, end_mask = var_3885_end_mask_0, x = xp_cast_fp16)[name = string("op_3885_cast_fp16")];
            tensor<fp16, [1024]> var_3893_to_fp16 = const()[name = string("op_3893_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97005760)))];
            tensor<fp16, [1, 32, 1024]> var_3894_cast_fp16 = mul(x = var_3885_cast_fp16, y = var_3893_to_fp16)[name = string("op_3894_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> var_3896_cast_fp16 = add(x = var_3875_cast_fp16, y = var_3894_cast_fp16)[name = string("op_3896_cast_fp16")];
            tensor<int32, [3]> var_3906_begin_0 = const()[name = string("op_3906_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3906_end_0 = const()[name = string("op_3906_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> var_3906_end_mask_0 = const()[name = string("op_3906_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 32, 1024]> var_3906_cast_fp16 = slice_by_index(begin = var_3906_begin_0, end = var_3906_end_0, end_mask = var_3906_end_mask_0, x = xp_cast_fp16)[name = string("op_3906_cast_fp16")];
            tensor<fp16, [1024]> var_3914_to_fp16 = const()[name = string("op_3914_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97007872)))];
            tensor<fp16, [1, 32, 1024]> var_3915_cast_fp16 = mul(x = var_3906_cast_fp16, y = var_3914_to_fp16)[name = string("op_3915_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> input_93_cast_fp16 = add(x = var_3896_cast_fp16, y = var_3915_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [1, 32, 1024]> c_cast_fp16 = silu(x = input_93_cast_fp16)[name = string("c_cast_fp16")];
            tensor<int32, [3]> var_3923_begin_0 = const()[name = string("op_3923_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3923_end_0 = const()[name = string("op_3923_end_0"), val = tensor<int32, [3]>([1, 32, 256])];
            tensor<bool, [3]> var_3923_end_mask_0 = const()[name = string("op_3923_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 256]> var_3923_cast_fp16 = slice_by_index(begin = var_3923_begin_0, end = var_3923_end_0, end_mask = var_3923_end_mask_0, x = c_cast_fp16)[name = string("op_3923_cast_fp16")];
            tensor<int32, [4]> var_3928 = const()[name = string("op_3928"), val = tensor<int32, [4]>([1, 32, 2, 128])];
            tensor<fp16, [1, 32, 2, 128]> q_63_cast_fp16 = reshape(shape = var_3928, x = var_3923_cast_fp16)[name = string("q_63_cast_fp16")];
            tensor<int32, [3]> var_3934_begin_0 = const()[name = string("op_3934_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_3934_end_0 = const()[name = string("op_3934_end_0"), val = tensor<int32, [3]>([1, 32, 512])];
            tensor<bool, [3]> var_3934_end_mask_0 = const()[name = string("op_3934_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 256]> var_3934_cast_fp16 = slice_by_index(begin = var_3934_begin_0, end = var_3934_end_0, end_mask = var_3934_end_mask_0, x = c_cast_fp16)[name = string("op_3934_cast_fp16")];
            tensor<int32, [4]> var_3939 = const()[name = string("op_3939"), val = tensor<int32, [4]>([1, 32, 2, 128])];
            tensor<fp16, [1, 32, 2, 128]> k_65_cast_fp16 = reshape(shape = var_3939, x = var_3934_cast_fp16)[name = string("k_65_cast_fp16")];
            tensor<int32, [3]> var_3945_begin_0 = const()[name = string("op_3945_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_3945_end_0 = const()[name = string("op_3945_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> var_3945_end_mask_0 = const()[name = string("op_3945_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 512]> var_3945_cast_fp16 = slice_by_index(begin = var_3945_begin_0, end = var_3945_end_0, end_mask = var_3945_end_mask_0, x = c_cast_fp16)[name = string("op_3945_cast_fp16")];
            tensor<int32, [4]> var_3950 = const()[name = string("op_3950"), val = tensor<int32, [4]>([1, 32, 4, 128])];
            tensor<fp16, [1, 32, 4, 128]> var_3951_cast_fp16 = reshape(shape = var_3950, x = var_3945_cast_fp16)[name = string("op_3951_cast_fp16")];
            tensor<int32, [4]> var_3956 = const()[name = string("op_3956"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3958_to_fp16 = const()[name = string("op_3958_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 32, 2, 128]> q_65_cast_fp16 = mul(x = q_63_cast_fp16, y = var_3958_to_fp16)[name = string("q_65_cast_fp16")];
            fp16 var_3960_to_fp16 = const()[name = string("op_3960_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 32, 2, 128]> k_67_cast_fp16 = mul(x = k_65_cast_fp16, y = var_3960_to_fp16)[name = string("k_67_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_3962_cast_fp16 = mul(x = q_65_cast_fp16, y = q_65_cast_fp16)[name = string("op_3962_cast_fp16")];
            tensor<int32, [1]> var_3967_axes_0 = const()[name = string("op_3967_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3967_keep_dims_0 = const()[name = string("op_3967_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 2, 1]> var_3967_cast_fp16 = reduce_mean(axes = var_3967_axes_0, keep_dims = var_3967_keep_dims_0, x = var_3962_cast_fp16)[name = string("op_3967_cast_fp16")];
            fp16 var_3968_promoted_to_fp16 = const()[name = string("op_3968_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 32, 2, 1]> var_3969_cast_fp16 = mul(x = var_3967_cast_fp16, y = var_3968_promoted_to_fp16)[name = string("op_3969_cast_fp16")];
            fp16 var_3971_to_fp16 = const()[name = string("op_3971_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 32, 2, 1]> var_3972_cast_fp16 = add(x = var_3969_cast_fp16, y = var_3971_to_fp16)[name = string("op_3972_cast_fp16")];
            fp32 var_3973_epsilon_0 = const()[name = string("op_3973_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 2, 1]> var_3973_cast_fp16 = rsqrt(epsilon = var_3973_epsilon_0, x = var_3972_cast_fp16)[name = string("op_3973_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_3974_cast_fp16 = mul(x = q_65_cast_fp16, y = var_3973_cast_fp16)[name = string("op_3974_cast_fp16")];
            fp16 var_3975_to_fp16 = const()[name = string("op_3975_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 32, 2, 128]> q_67_cast_fp16 = mul(x = var_3974_cast_fp16, y = var_3975_to_fp16)[name = string("q_67_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> var_3977_cast_fp16 = mul(x = k_67_cast_fp16, y = k_67_cast_fp16)[name = string("op_3977_cast_fp16")];
            tensor<int32, [1]> var_3982_axes_0 = const()[name = string("op_3982_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3982_keep_dims_0 = const()[name = string("op_3982_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 2, 1]> var_3982_cast_fp16 = reduce_mean(axes = var_3982_axes_0, keep_dims = var_3982_keep_dims_0, x = var_3977_cast_fp16)[name = string("op_3982_cast_fp16")];
            fp16 var_3983_promoted_to_fp16 = const()[name = string("op_3983_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 32, 2, 1]> var_3984_cast_fp16 = mul(x = var_3982_cast_fp16, y = var_3983_promoted_to_fp16)[name = string("op_3984_cast_fp16")];
            fp16 var_3986_to_fp16 = const()[name = string("op_3986_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 32, 2, 1]> var_3987_cast_fp16 = add(x = var_3984_cast_fp16, y = var_3986_to_fp16)[name = string("op_3987_cast_fp16")];
            fp32 var_3988_epsilon_0 = const()[name = string("op_3988_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 2, 1]> var_3988_cast_fp16 = rsqrt(epsilon = var_3988_epsilon_0, x = var_3987_cast_fp16)[name = string("op_3988_cast_fp16")];
            tensor<fp16, [1, 32, 2, 128]> k_69_cast_fp16 = mul(x = k_67_cast_fp16, y = var_3988_cast_fp16)[name = string("k_69_cast_fp16")];
            tensor<int32, [4]> var_3994 = const()[name = string("op_3994"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_4007_axes_0 = const()[name = string("op_4007_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 128]> var_3995_cast_fp16 = transpose(perm = var_3994, x = q_67_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 2, 1, 32, 128]> var_4007_cast_fp16 = expand_dims(axes = var_4007_axes_0, x = var_3995_cast_fp16)[name = string("op_4007_cast_fp16")];
            tensor<int32, [5]> var_4015_reps_0 = const()[name = string("op_4015_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 32, 128]> var_4015_cast_fp16 = tile(reps = var_4015_reps_0, x = var_4007_cast_fp16)[name = string("op_4015_cast_fp16")];
            tensor<int32, [4]> var_4020 = const()[name = string("op_4020"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<fp16, [1, 4, 32, 128]> q_69_cast_fp16 = reshape(shape = var_4020, x = var_4015_cast_fp16)[name = string("q_69_cast_fp16")];
            tensor<int32, [4]> var_4026 = const()[name = string("op_4026"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_4039_axes_0 = const()[name = string("op_4039_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 128]> var_4027_cast_fp16 = transpose(perm = var_4026, x = k_69_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 2, 1, 32, 128]> var_4039_cast_fp16 = expand_dims(axes = var_4039_axes_0, x = var_4027_cast_fp16)[name = string("op_4039_cast_fp16")];
            tensor<int32, [5]> var_4047_reps_0 = const()[name = string("op_4047_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 32, 128]> var_4047_cast_fp16 = tile(reps = var_4047_reps_0, x = var_4039_cast_fp16)[name = string("op_4047_cast_fp16")];
            tensor<int32, [4]> var_4052 = const()[name = string("op_4052"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<fp16, [1, 4, 32, 128]> k_71_cast_fp16 = reshape(shape = var_4052, x = var_4047_cast_fp16)[name = string("k_71_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_4054_cast_fp16 = sigmoid(x = b_cast_fp16)[name = string("op_4054_cast_fp16")];
            tensor<int32, [3]> var_4058 = const()[name = string("op_4058"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_6_dt_bias_to_fp16 = const()[name = string("layers_6_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 32, 4]> input_95_cast_fp16 = add(x = a_cast_fp16, y = layers_6_dt_bias_to_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_4062_cast_fp16 = relu(x = input_95_cast_fp16)[name = string("op_4062_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_4063_cast_fp16 = abs(x = input_95_cast_fp16)[name = string("op_4063_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 32, 4]> var_4064_cast_fp16 = mul(x = var_4063_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_4064_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_4065_cast_fp16 = exp(x = var_4064_cast_fp16)[name = string("op_4065_cast_fp16")];
            fp16 var_4067_to_fp16 = const()[name = string("op_4067_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 32, 4]> var_4068_cast_fp16 = add(x = var_4065_cast_fp16, y = var_4067_to_fp16)[name = string("op_4068_cast_fp16")];
            fp32 var_4069_epsilon_0 = const()[name = string("op_4069_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 32, 4]> var_4069_cast_fp16 = log(epsilon = var_4069_epsilon_0, x = var_4068_cast_fp16)[name = string("op_4069_cast_fp16")];
            tensor<fp16, [1, 32, 4]> var_4071_cast_fp16 = add(x = var_4062_cast_fp16, y = var_4069_cast_fp16)[name = string("op_4071_cast_fp16")];
            tensor<fp16, [4]> layers_6_negA_to_fp16 = const()[name = string("layers_6_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.07cp+1, -0x1.eccp+3, -0x1.63p+3, -0x1.728p+1])];
            tensor<fp16, [1, 32, 4]> var_4072_cast_fp16 = mul(x = layers_6_negA_to_fp16, y = var_4071_cast_fp16)[name = string("op_4072_cast_fp16")];
            tensor<int32, [3]> var_4076 = const()[name = string("op_4076"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_4079_axes_0 = const()[name = string("op_4079_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 32]> beta_cast_fp16 = transpose(perm = var_4058, x = var_4054_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 4, 32, 1]> var_4079_cast_fp16 = expand_dims(axes = var_4079_axes_0, x = beta_cast_fp16)[name = string("op_4079_cast_fp16")];
            fp16 var_4080_to_fp16 = const()[name = string("op_4080_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 32, 1]> var_4081_cast_fp16 = mul(x = var_4079_cast_fp16, y = var_4080_to_fp16)[name = string("op_4081_cast_fp16")];
            tensor<fp16, [1, 4, 32, 128]> v_15_cast_fp16 = transpose(perm = var_3956, x = var_3951_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 4, 32, 128]> vb_31_cast_fp16 = mul(x = v_15_cast_fp16, y = var_4081_cast_fp16)[name = string("vb_31_cast_fp16")];
            tensor<int32, [4]> q_71_begin_0 = const()[name = string("q_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_71_end_0 = const()[name = string("q_71_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_71_end_mask_0 = const()[name = string("q_71_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_71_cast_fp16 = slice_by_index(begin = q_71_begin_0, end = q_71_end_0, end_mask = q_71_end_mask_0, x = q_69_cast_fp16)[name = string("q_71_cast_fp16")];
            tensor<int32, [4]> k_73_begin_0 = const()[name = string("k_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_73_end_0 = const()[name = string("k_73_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_73_end_mask_0 = const()[name = string("k_73_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, x = k_71_cast_fp16)[name = string("k_73_cast_fp16")];
            tensor<fp16, [1, 4, 32, 128]> var_4115_cast_fp16 = mul(x = k_71_cast_fp16, y = var_4079_cast_fp16)[name = string("op_4115_cast_fp16")];
            tensor<int32, [4]> kb_21_begin_0 = const()[name = string("kb_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_21_end_0 = const()[name = string("kb_21_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_21_end_mask_0 = const()[name = string("kb_21_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_21_cast_fp16 = slice_by_index(begin = kb_21_begin_0, end = kb_21_end_0, end_mask = kb_21_end_mask_0, x = var_4115_cast_fp16)[name = string("kb_21_cast_fp16")];
            tensor<int32, [4]> vb_33_begin_0 = const()[name = string("vb_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_33_end_0 = const()[name = string("vb_33_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_33_end_mask_0 = const()[name = string("vb_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_33_cast_fp16 = slice_by_index(begin = vb_33_begin_0, end = vb_33_end_0, end_mask = vb_33_end_mask_0, x = vb_31_cast_fp16)[name = string("vb_33_cast_fp16")];
            tensor<int32, [3]> gdec_33_begin_0 = const()[name = string("gdec_33_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_33_end_0 = const()[name = string("gdec_33_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_33_end_mask_0 = const()[name = string("gdec_33_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 32]> gdec_31_cast_fp16 = transpose(perm = var_4076, x = var_4072_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 4, 16]> gdec_33_cast_fp16 = slice_by_index(begin = gdec_33_begin_0, end = gdec_33_end_0, end_mask = gdec_33_end_mask_0, x = gdec_31_cast_fp16)[name = string("gdec_33_cast_fp16")];
            tensor<int32, [2]> var_4163 = const()[name = string("op_4163"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_4164_cast_fp16 = reshape(shape = var_4163, x = gdec_33_cast_fp16)[name = string("op_4164_cast_fp16")];
            tensor<fp16, [4, 256]> var_4165_cast_fp16 = linear(bias = var_633_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_4164_cast_fp16)[name = string("op_4165_cast_fp16")];
            tensor<int32, [4]> var_4170 = const()[name = string("op_4170"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_21_cast_fp16 = reshape(shape = var_4170, x = var_4165_cast_fp16)[name = string("P_21_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4172_cast_fp16 = exp(x = P_21_cast_fp16)[name = string("op_4172_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_21_cast_fp16 = mul(x = var_4172_cast_fp16, y = tril_incl_to_fp16)[name = string("D_21_cast_fp16")];
            bool kk_21_transpose_x_1 = const()[name = string("kk_21_transpose_x_1"), val = bool(false)];
            bool kk_21_transpose_y_1 = const()[name = string("kk_21_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_21_cast_fp16 = matmul(transpose_x = kk_21_transpose_x_1, transpose_y = kk_21_transpose_y_1, x = kb_21_cast_fp16, y = k_73_cast_fp16)[name = string("kk_21_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_4178_cast_fp16 = mul(x = kk_21_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_4178_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4179_cast_fp16 = mul(x = var_4178_cast_fp16, y = D_21_cast_fp16)[name = string("op_4179_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_21_cast_fp16 = mul(x = var_4179_cast_fp16, y = cum_tail_to_fp16)[name = string("N_21_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4181_cast_fp16 = mul(x = D_21_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_4181_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_21_cast_fp16 = mul(x = kk_21_cast_fp16, y = var_4181_cast_fp16)[name = string("Na_21_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_81_cast_fp16 = add(x = eye_to_fp16, y = N_21_cast_fp16)[name = string("Minv_81_cast_fp16")];
            bool Np_61_transpose_x_0 = const()[name = string("Np_61_transpose_x_0"), val = bool(false)];
            bool Np_61_transpose_y_0 = const()[name = string("Np_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_61_cast_fp16 = matmul(transpose_x = Np_61_transpose_x_0, transpose_y = Np_61_transpose_y_0, x = N_21_cast_fp16, y = Na_21_cast_fp16)[name = string("Np_61_cast_fp16")];
            bool Npa_41_transpose_x_0 = const()[name = string("Npa_41_transpose_x_0"), val = bool(false)];
            bool Npa_41_transpose_y_0 = const()[name = string("Npa_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_41_cast_fp16 = matmul(transpose_x = Npa_41_transpose_x_0, transpose_y = Npa_41_transpose_y_0, x = Na_21_cast_fp16, y = N_21_cast_fp16)[name = string("Npa_41_cast_fp16")];
            bool var_4187_transpose_x_0 = const()[name = string("op_4187_transpose_x_0"), val = bool(false)];
            bool var_4187_transpose_y_0 = const()[name = string("op_4187_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4187_cast_fp16 = matmul(transpose_x = var_4187_transpose_x_0, transpose_y = var_4187_transpose_y_0, x = Minv_81_cast_fp16, y = Np_61_cast_fp16)[name = string("op_4187_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_83_cast_fp16 = add(x = Minv_81_cast_fp16, y = var_4187_cast_fp16)[name = string("Minv_83_cast_fp16")];
            bool Np_63_transpose_x_0 = const()[name = string("Np_63_transpose_x_0"), val = bool(false)];
            bool Np_63_transpose_y_0 = const()[name = string("Np_63_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_63_cast_fp16 = matmul(transpose_x = Np_63_transpose_x_0, transpose_y = Np_63_transpose_y_0, x = Np_61_cast_fp16, y = Npa_41_cast_fp16)[name = string("Np_63_cast_fp16")];
            bool Npa_43_transpose_x_0 = const()[name = string("Npa_43_transpose_x_0"), val = bool(false)];
            bool Npa_43_transpose_y_0 = const()[name = string("Npa_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_43_cast_fp16 = matmul(transpose_x = Npa_43_transpose_x_0, transpose_y = Npa_43_transpose_y_0, x = Npa_41_cast_fp16, y = Np_61_cast_fp16)[name = string("Npa_43_cast_fp16")];
            bool var_4192_transpose_x_0 = const()[name = string("op_4192_transpose_x_0"), val = bool(false)];
            bool var_4192_transpose_y_0 = const()[name = string("op_4192_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4192_cast_fp16 = matmul(transpose_x = var_4192_transpose_x_0, transpose_y = var_4192_transpose_y_0, x = Minv_83_cast_fp16, y = Np_63_cast_fp16)[name = string("op_4192_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_85_cast_fp16 = add(x = Minv_83_cast_fp16, y = var_4192_cast_fp16)[name = string("Minv_85_cast_fp16")];
            bool Np_65_transpose_x_0 = const()[name = string("Np_65_transpose_x_0"), val = bool(false)];
            bool Np_65_transpose_y_0 = const()[name = string("Np_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_65_cast_fp16 = matmul(transpose_x = Np_65_transpose_x_0, transpose_y = Np_65_transpose_y_0, x = Np_63_cast_fp16, y = Npa_43_cast_fp16)[name = string("Np_65_cast_fp16")];
            bool var_4196_transpose_x_0 = const()[name = string("op_4196_transpose_x_0"), val = bool(false)];
            bool var_4196_transpose_y_0 = const()[name = string("op_4196_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4196_cast_fp16 = matmul(transpose_x = var_4196_transpose_x_0, transpose_y = var_4196_transpose_y_0, x = Minv_85_cast_fp16, y = Np_65_cast_fp16)[name = string("op_4196_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_87_cast_fp16 = add(x = Minv_85_cast_fp16, y = var_4196_cast_fp16)[name = string("Minv_87_cast_fp16")];
            bool u_21_transpose_x_0 = const()[name = string("u_21_transpose_x_0"), val = bool(false)];
            bool u_21_transpose_y_0 = const()[name = string("u_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_21_cast_fp16 = matmul(transpose_x = u_21_transpose_x_0, transpose_y = u_21_transpose_y_0, x = Minv_87_cast_fp16, y = vb_33_cast_fp16)[name = string("u_21_cast_fp16")];
            bool var_4203_transpose_x_1 = const()[name = string("op_4203_transpose_x_1"), val = bool(false)];
            bool var_4203_transpose_y_1 = const()[name = string("op_4203_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_4203_cast_fp16 = matmul(transpose_x = var_4203_transpose_x_1, transpose_y = var_4203_transpose_y_1, x = q_71_cast_fp16, y = k_73_cast_fp16)[name = string("op_4203_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_21_cast_fp16 = mul(x = var_4203_cast_fp16, y = D_21_cast_fp16)[name = string("A_21_cast_fp16")];
            bool var_4205_transpose_x_0 = const()[name = string("op_4205_transpose_x_0"), val = bool(false)];
            bool var_4205_transpose_y_0 = const()[name = string("op_4205_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4205_cast_fp16 = matmul(transpose_x = var_4205_transpose_x_0, transpose_y = var_4205_transpose_y_0, x = A_21_cast_fp16, y = u_21_cast_fp16)[name = string("op_4205_cast_fp16")];
            tensor<fp16, [16]> var_4206_bias_0_to_fp16 = const()[name = string("op_4206_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_4206_cast_fp16 = linear(bias = var_4206_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_33_cast_fp16)[name = string("op_4206_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_cast_fp16 = exp(x = var_4206_cast_fp16)[name = string("tail_cast_fp16")];
            tensor<int32, [1]> var_4209_axes_0 = const()[name = string("op_4209_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_4209_cast_fp16 = expand_dims(axes = var_4209_axes_0, x = tail_cast_fp16)[name = string("op_4209_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4210_cast_fp16 = mul(x = k_73_cast_fp16, y = var_4209_cast_fp16)[name = string("op_4210_cast_fp16")];
            bool S_transpose_x_1 = const()[name = string("S_transpose_x_1"), val = bool(true)];
            bool S_transpose_y_1 = const()[name = string("S_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_cast_fp16 = matmul(transpose_x = S_transpose_x_1, transpose_y = S_transpose_y_1, x = var_4210_cast_fp16, y = u_21_cast_fp16)[name = string("S_cast_fp16")];
            tensor<int32, [4]> q_73_begin_0 = const()[name = string("q_73_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_73_end_0 = const()[name = string("q_73_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_73_end_mask_0 = const()[name = string("q_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, x = q_69_cast_fp16)[name = string("q_73_cast_fp16")];
            tensor<int32, [4]> k_75_begin_0 = const()[name = string("k_75_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_75_end_0 = const()[name = string("k_75_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_75_end_mask_0 = const()[name = string("k_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_75_cast_fp16 = slice_by_index(begin = k_75_begin_0, end = k_75_end_0, end_mask = k_75_end_mask_0, x = k_71_cast_fp16)[name = string("k_75_cast_fp16")];
            tensor<int32, [4]> kb_begin_0 = const()[name = string("kb_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_end_0 = const()[name = string("kb_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_end_mask_0 = const()[name = string("kb_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_cast_fp16 = slice_by_index(begin = kb_begin_0, end = kb_end_0, end_mask = kb_end_mask_0, x = var_4115_cast_fp16)[name = string("kb_cast_fp16")];
            tensor<int32, [4]> vb_begin_0 = const()[name = string("vb_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_end_0 = const()[name = string("vb_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_end_mask_0 = const()[name = string("vb_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_cast_fp16 = slice_by_index(begin = vb_begin_0, end = vb_end_0, end_mask = vb_end_mask_0, x = vb_31_cast_fp16)[name = string("vb_cast_fp16")];
            tensor<int32, [3]> gdec_begin_0 = const()[name = string("gdec_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_end_0 = const()[name = string("gdec_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_end_mask_0 = const()[name = string("gdec_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_cast_fp16 = slice_by_index(begin = gdec_begin_0, end = gdec_end_0, end_mask = gdec_end_mask_0, x = gdec_31_cast_fp16)[name = string("gdec_cast_fp16")];
            tensor<int32, [2]> var_4295 = const()[name = string("op_4295"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_4296_cast_fp16 = reshape(shape = var_4295, x = gdec_cast_fp16)[name = string("op_4296_cast_fp16")];
            tensor<fp16, [4, 256]> var_4297_cast_fp16 = linear(bias = var_633_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_4296_cast_fp16)[name = string("op_4297_cast_fp16")];
            tensor<int32, [4]> var_4302 = const()[name = string("op_4302"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_cast_fp16 = reshape(shape = var_4302, x = var_4297_cast_fp16)[name = string("P_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4304_cast_fp16 = exp(x = P_cast_fp16)[name = string("op_4304_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_cast_fp16 = mul(x = var_4304_cast_fp16, y = tril_incl_to_fp16)[name = string("D_cast_fp16")];
            bool kk_transpose_x_1 = const()[name = string("kk_transpose_x_1"), val = bool(false)];
            bool kk_transpose_y_1 = const()[name = string("kk_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_cast_fp16 = matmul(transpose_x = kk_transpose_x_1, transpose_y = kk_transpose_y_1, x = kb_cast_fp16, y = k_75_cast_fp16)[name = string("kk_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_4310_cast_fp16 = mul(x = kk_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_4310_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4311_cast_fp16 = mul(x = var_4310_cast_fp16, y = D_cast_fp16)[name = string("op_4311_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_cast_fp16 = mul(x = var_4311_cast_fp16, y = cum_tail_to_fp16)[name = string("N_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4313_cast_fp16 = mul(x = D_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_4313_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_cast_fp16 = mul(x = kk_cast_fp16, y = var_4313_cast_fp16)[name = string("Na_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_89_cast_fp16 = add(x = eye_to_fp16, y = N_cast_fp16)[name = string("Minv_89_cast_fp16")];
            bool Np_67_transpose_x_0 = const()[name = string("Np_67_transpose_x_0"), val = bool(false)];
            bool Np_67_transpose_y_0 = const()[name = string("Np_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_67_cast_fp16 = matmul(transpose_x = Np_67_transpose_x_0, transpose_y = Np_67_transpose_y_0, x = N_cast_fp16, y = Na_cast_fp16)[name = string("Np_67_cast_fp16")];
            bool Npa_45_transpose_x_0 = const()[name = string("Npa_45_transpose_x_0"), val = bool(false)];
            bool Npa_45_transpose_y_0 = const()[name = string("Npa_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_45_cast_fp16 = matmul(transpose_x = Npa_45_transpose_x_0, transpose_y = Npa_45_transpose_y_0, x = Na_cast_fp16, y = N_cast_fp16)[name = string("Npa_45_cast_fp16")];
            bool var_4319_transpose_x_0 = const()[name = string("op_4319_transpose_x_0"), val = bool(false)];
            bool var_4319_transpose_y_0 = const()[name = string("op_4319_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4319_cast_fp16 = matmul(transpose_x = var_4319_transpose_x_0, transpose_y = var_4319_transpose_y_0, x = Minv_89_cast_fp16, y = Np_67_cast_fp16)[name = string("op_4319_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_91_cast_fp16 = add(x = Minv_89_cast_fp16, y = var_4319_cast_fp16)[name = string("Minv_91_cast_fp16")];
            bool Np_69_transpose_x_0 = const()[name = string("Np_69_transpose_x_0"), val = bool(false)];
            bool Np_69_transpose_y_0 = const()[name = string("Np_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_69_cast_fp16 = matmul(transpose_x = Np_69_transpose_x_0, transpose_y = Np_69_transpose_y_0, x = Np_67_cast_fp16, y = Npa_45_cast_fp16)[name = string("Np_69_cast_fp16")];
            bool Npa_transpose_x_0 = const()[name = string("Npa_transpose_x_0"), val = bool(false)];
            bool Npa_transpose_y_0 = const()[name = string("Npa_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_cast_fp16 = matmul(transpose_x = Npa_transpose_x_0, transpose_y = Npa_transpose_y_0, x = Npa_45_cast_fp16, y = Np_67_cast_fp16)[name = string("Npa_cast_fp16")];
            bool var_4324_transpose_x_0 = const()[name = string("op_4324_transpose_x_0"), val = bool(false)];
            bool var_4324_transpose_y_0 = const()[name = string("op_4324_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4324_cast_fp16 = matmul(transpose_x = var_4324_transpose_x_0, transpose_y = var_4324_transpose_y_0, x = Minv_91_cast_fp16, y = Np_69_cast_fp16)[name = string("op_4324_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_93_cast_fp16 = add(x = Minv_91_cast_fp16, y = var_4324_cast_fp16)[name = string("Minv_93_cast_fp16")];
            bool Np_transpose_x_0 = const()[name = string("Np_transpose_x_0"), val = bool(false)];
            bool Np_transpose_y_0 = const()[name = string("Np_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_cast_fp16 = matmul(transpose_x = Np_transpose_x_0, transpose_y = Np_transpose_y_0, x = Np_69_cast_fp16, y = Npa_cast_fp16)[name = string("Np_cast_fp16")];
            bool var_4328_transpose_x_0 = const()[name = string("op_4328_transpose_x_0"), val = bool(false)];
            bool var_4328_transpose_y_0 = const()[name = string("op_4328_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4328_cast_fp16 = matmul(transpose_x = var_4328_transpose_x_0, transpose_y = var_4328_transpose_y_0, x = Minv_93_cast_fp16, y = Np_cast_fp16)[name = string("op_4328_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_cast_fp16 = add(x = Minv_93_cast_fp16, y = var_4328_cast_fp16)[name = string("Minv_cast_fp16")];
            bool u_transpose_x_0 = const()[name = string("u_transpose_x_0"), val = bool(false)];
            bool u_transpose_y_0 = const()[name = string("u_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_cast_fp16 = matmul(transpose_x = u_transpose_x_0, transpose_y = u_transpose_y_0, x = Minv_cast_fp16, y = vb_cast_fp16)[name = string("u_cast_fp16")];
            bool var_4335_transpose_x_1 = const()[name = string("op_4335_transpose_x_1"), val = bool(false)];
            bool var_4335_transpose_y_1 = const()[name = string("op_4335_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_4335_cast_fp16 = matmul(transpose_x = var_4335_transpose_x_1, transpose_y = var_4335_transpose_y_1, x = q_73_cast_fp16, y = k_75_cast_fp16)[name = string("op_4335_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_cast_fp16 = mul(x = var_4335_cast_fp16, y = D_cast_fp16)[name = string("A_cast_fp16")];
            tensor<fp16, [16]> var_4337_bias_0_to_fp16 = const()[name = string("op_4337_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_4337_cast_fp16 = linear(bias = var_4337_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_cast_fp16)[name = string("op_4337_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_cast_fp16 = exp(x = var_4337_cast_fp16)[name = string("cum_cast_fp16")];
            tensor<int32, [1]> var_4340_axes_0 = const()[name = string("op_4340_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_4340_cast_fp16 = expand_dims(axes = var_4340_axes_0, x = cum_cast_fp16)[name = string("op_4340_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4341_cast_fp16 = mul(x = kb_cast_fp16, y = var_4340_cast_fp16)[name = string("op_4341_cast_fp16")];
            bool w_transpose_x_0 = const()[name = string("w_transpose_x_0"), val = bool(false)];
            bool w_transpose_y_0 = const()[name = string("w_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_cast_fp16 = matmul(transpose_x = w_transpose_x_0, transpose_y = w_transpose_y_0, x = Minv_cast_fp16, y = var_4341_cast_fp16)[name = string("w_cast_fp16")];
            bool var_4343_transpose_x_0 = const()[name = string("op_4343_transpose_x_0"), val = bool(false)];
            bool var_4343_transpose_y_0 = const()[name = string("op_4343_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4343_cast_fp16 = matmul(transpose_x = var_4343_transpose_x_0, transpose_y = var_4343_transpose_y_0, x = w_cast_fp16, y = S_cast_fp16)[name = string("op_4343_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_cast_fp16 = sub(x = u_cast_fp16, y = var_4343_cast_fp16)[name = string("vnew_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4348_cast_fp16 = mul(x = q_73_cast_fp16, y = var_4340_cast_fp16)[name = string("op_4348_cast_fp16")];
            bool var_4349_transpose_x_0 = const()[name = string("op_4349_transpose_x_0"), val = bool(false)];
            bool var_4349_transpose_y_0 = const()[name = string("op_4349_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4349_cast_fp16 = matmul(transpose_x = var_4349_transpose_x_0, transpose_y = var_4349_transpose_y_0, x = var_4348_cast_fp16, y = S_cast_fp16)[name = string("op_4349_cast_fp16")];
            bool var_4350_transpose_x_0 = const()[name = string("op_4350_transpose_x_0"), val = bool(false)];
            bool var_4350_transpose_y_0 = const()[name = string("op_4350_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4350_cast_fp16 = matmul(transpose_x = var_4350_transpose_x_0, transpose_y = var_4350_transpose_y_0, x = A_cast_fp16, y = vnew_cast_fp16)[name = string("op_4350_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_cast_fp16 = add(x = var_4349_cast_fp16, y = var_4350_cast_fp16)[name = string("oc_cast_fp16")];
            int32 var_4354 = const()[name = string("op_4354"), val = int32(2)];
            bool o_33_interleave_0 = const()[name = string("o_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 128]> o_33_cast_fp16 = concat(axis = var_4354, interleave = o_33_interleave_0, values = (var_4205_cast_fp16, oc_cast_fp16))[name = string("o_33_cast_fp16")];
            tensor<int32, [4]> var_4360 = const()[name = string("op_4360"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_4366 = const()[name = string("op_4366"), val = tensor<int32, [4]>([1, 32, 4, 128])];
            tensor<fp16, [1, 32, 4, 128]> input_97_cast_fp16 = reshape(shape = var_4366, x = z_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> o_35_cast_fp16 = transpose(perm = var_4360, x = o_33_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 32, 4, 128]> var_4368_cast_fp16 = mul(x = o_35_cast_fp16, y = o_35_cast_fp16)[name = string("op_4368_cast_fp16")];
            tensor<int32, [1]> var_4373_axes_0 = const()[name = string("op_4373_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4373_keep_dims_0 = const()[name = string("op_4373_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 4, 1]> var_4373_cast_fp16 = reduce_mean(axes = var_4373_axes_0, keep_dims = var_4373_keep_dims_0, x = var_4368_cast_fp16)[name = string("op_4373_cast_fp16")];
            fp16 var_4375_to_fp16 = const()[name = string("op_4375_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 32, 4, 1]> var_4376_cast_fp16 = add(x = var_4373_cast_fp16, y = var_4375_to_fp16)[name = string("op_4376_cast_fp16")];
            fp32 var_4377_epsilon_0 = const()[name = string("op_4377_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 4, 1]> var_4377_cast_fp16 = rsqrt(epsilon = var_4377_epsilon_0, x = var_4376_cast_fp16)[name = string("op_4377_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> var_4378_cast_fp16 = mul(x = o_35_cast_fp16, y = var_4377_cast_fp16)[name = string("op_4378_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> var_4380_cast_fp16 = silu(x = input_97_cast_fp16)[name = string("op_4380_cast_fp16")];
            tensor<fp16, [1, 32, 4, 128]> o_37_cast_fp16 = mul(x = var_4378_cast_fp16, y = var_4380_cast_fp16)[name = string("o_37_cast_fp16")];
            tensor<int32, [3]> var_4385 = const()[name = string("op_4385"), val = tensor<int32, [3]>([1, 32, 512])];
            tensor<fp16, [1, 32, 512]> input_99_cast_fp16 = reshape(shape = var_4385, x = o_37_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [640, 512]> layers_6_out_weight_to_fp16 = const()[name = string("layers_6_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97009984)))];
            tensor<fp16, [1, 32, 640]> linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_out_weight_to_fp16, x = input_99_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 32, 640]> x_65_cast_fp16 = add(x = x_61_cast_fp16, y = linear_25_cast_fp16)[name = string("x_65_cast_fp16")];
            fp16 var_4398_to_fp16 = const()[name = string("op_4398_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [1, 32, 640]> x_67_cast_fp16 = mul(x = x_65_cast_fp16, y = var_4398_to_fp16)[name = string("x_67_cast_fp16")];
            fp16 var_4396_promoted_to_fp16 = const()[name = string("op_4396_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> var_4400_cast_fp16 = pow(x = x_67_cast_fp16, y = var_4396_promoted_to_fp16)[name = string("op_4400_cast_fp16")];
            tensor<int32, [1]> var_4402_axes_0 = const()[name = string("op_4402_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4402_keep_dims_0 = const()[name = string("op_4402_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4402_cast_fp16 = reduce_mean(axes = var_4402_axes_0, keep_dims = var_4402_keep_dims_0, x = var_4400_cast_fp16)[name = string("op_4402_cast_fp16")];
            fp16 var_4403_to_fp16 = const()[name = string("op_4403_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 32, 1]> var_4404_cast_fp16 = add(x = var_4402_cast_fp16, y = var_4403_to_fp16)[name = string("op_4404_cast_fp16")];
            fp32 var_4405_epsilon_0 = const()[name = string("op_4405_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4405_cast_fp16 = rsqrt(epsilon = var_4405_epsilon_0, x = var_4404_cast_fp16)[name = string("op_4405_cast_fp16")];
            tensor<fp16, [1, 32, 640]> var_4406_cast_fp16 = mul(x = x_67_cast_fp16, y = var_4405_cast_fp16)[name = string("op_4406_cast_fp16")];
            tensor<fp16, [640]> layers_6_n2_w1_to_fp16 = const()[name = string("layers_6_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97665408)))];
            tensor<fp16, [1, 32, 640]> input_101_cast_fp16 = mul(x = var_4406_cast_fp16, y = layers_6_n2_w1_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_6_gate_up_weight_to_fp16 = const()[name = string("layers_6_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97666752)))];
            tensor<fp16, [1, 32, 5120]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_6_gate_up_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<int32, [3]> input_103_begin_0 = const()[name = string("input_103_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_103_end_0 = const()[name = string("input_103_end_0"), val = tensor<int32, [3]>([1, 32, 2560])];
            tensor<bool, [3]> input_103_end_mask_0 = const()[name = string("input_103_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 2560]> input_103_cast_fp16 = slice_by_index(begin = input_103_begin_0, end = input_103_end_0, end_mask = input_103_end_mask_0, x = linear_26_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [1, 32, 2560]> var_4416_cast_fp16 = silu(x = input_103_cast_fp16)[name = string("op_4416_cast_fp16")];
            tensor<int32, [3]> var_4421_begin_0 = const()[name = string("op_4421_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_4421_end_0 = const()[name = string("op_4421_end_0"), val = tensor<int32, [3]>([1, 32, 5120])];
            tensor<bool, [3]> var_4421_end_mask_0 = const()[name = string("op_4421_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 2560]> var_4421_cast_fp16 = slice_by_index(begin = var_4421_begin_0, end = var_4421_end_0, end_mask = var_4421_end_mask_0, x = linear_26_cast_fp16)[name = string("op_4421_cast_fp16")];
            tensor<fp16, [1, 32, 2560]> input_105_cast_fp16 = mul(x = var_4416_cast_fp16, y = var_4421_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_6_down_weight_to_fp16 = const()[name = string("layers_6_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104220416)))];
            tensor<fp16, [1, 32, 640]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_down_weight_to_fp16, x = input_105_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 32, 640]> x_69_cast_fp16 = add(x = x_65_cast_fp16, y = linear_27_cast_fp16)[name = string("x_69_cast_fp16")];
            fp16 var_4434_to_fp16 = const()[name = string("op_4434_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [1, 32, 640]> x_71_cast_fp16 = mul(x = x_69_cast_fp16, y = var_4434_to_fp16)[name = string("x_71_cast_fp16")];
            fp16 var_4432_promoted_to_fp16 = const()[name = string("op_4432_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> var_4436_cast_fp16 = pow(x = x_71_cast_fp16, y = var_4432_promoted_to_fp16)[name = string("op_4436_cast_fp16")];
            tensor<int32, [1]> var_4438_axes_0 = const()[name = string("op_4438_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4438_keep_dims_0 = const()[name = string("op_4438_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4438_cast_fp16 = reduce_mean(axes = var_4438_axes_0, keep_dims = var_4438_keep_dims_0, x = var_4436_cast_fp16)[name = string("op_4438_cast_fp16")];
            fp16 var_4439_to_fp16 = const()[name = string("op_4439_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 32, 1]> var_4440_cast_fp16 = add(x = var_4438_cast_fp16, y = var_4439_to_fp16)[name = string("op_4440_cast_fp16")];
            fp32 var_4441_epsilon_0 = const()[name = string("op_4441_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4441_cast_fp16 = rsqrt(epsilon = var_4441_epsilon_0, x = var_4440_cast_fp16)[name = string("op_4441_cast_fp16")];
            tensor<fp16, [1, 32, 640]> var_4442_cast_fp16 = mul(x = x_71_cast_fp16, y = var_4441_cast_fp16)[name = string("op_4442_cast_fp16")];
            tensor<fp16, [640]> layers_7_n1_w1_to_fp16 = const()[name = string("layers_7_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107497280)))];
            tensor<fp16, [1, 32, 640]> input_107_cast_fp16 = mul(x = var_4442_cast_fp16, y = layers_7_n1_w1_to_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [1280, 640]> layers_7_qkv_weight_to_fp16 = const()[name = string("layers_7_qkv_weight_to_fp16"), val = tensor<fp16, [1280, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107498624)))];
            tensor<fp16, [1, 32, 1280]> linear_28_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = layers_7_qkv_weight_to_fp16, x = input_107_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [3]> var_4451_begin_0 = const()[name = string("op_4451_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_4451_end_0 = const()[name = string("op_4451_end_0"), val = tensor<int32, [3]>([1, 32, 1024])];
            tensor<bool, [3]> var_4451_end_mask_0 = const()[name = string("op_4451_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 1024]> var_4451_cast_fp16 = slice_by_index(begin = var_4451_begin_0, end = var_4451_end_0, end_mask = var_4451_end_mask_0, x = linear_28_cast_fp16)[name = string("op_4451_cast_fp16")];
            tensor<int32, [4]> var_4456 = const()[name = string("op_4456"), val = tensor<int32, [4]>([1, 32, 8, 128])];
            tensor<fp16, [1, 32, 8, 128]> qg_cast_fp16 = reshape(shape = var_4456, x = var_4451_cast_fp16)[name = string("qg_cast_fp16")];
            tensor<int32, [4]> x_73_begin_0 = const()[name = string("x_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_73_end_0 = const()[name = string("x_73_end_0"), val = tensor<int32, [4]>([1, 32, 8, 64])];
            tensor<bool, [4]> x_73_end_mask_0 = const()[name = string("x_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 32, 8, 64]> x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, x = qg_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<int32, [4]> var_4467_begin_0 = const()[name = string("op_4467_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> var_4467_end_0 = const()[name = string("op_4467_end_0"), val = tensor<int32, [4]>([1, 32, 8, 128])];
            tensor<bool, [4]> var_4467_end_mask_0 = const()[name = string("op_4467_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 32, 8, 64]> var_4467_cast_fp16 = slice_by_index(begin = var_4467_begin_0, end = var_4467_end_0, end_mask = var_4467_end_mask_0, x = qg_cast_fp16)[name = string("op_4467_cast_fp16")];
            tensor<int32, [3]> var_4471 = const()[name = string("op_4471"), val = tensor<int32, [3]>([1, 32, 512])];
            tensor<fp16, [1, 32, 512]> gate_cast_fp16 = reshape(shape = var_4471, x = var_4467_cast_fp16)[name = string("gate_cast_fp16")];
            tensor<int32, [3]> var_4477_begin_0 = const()[name = string("op_4477_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> var_4477_end_0 = const()[name = string("op_4477_end_0"), val = tensor<int32, [3]>([1, 32, 1152])];
            tensor<bool, [3]> var_4477_end_mask_0 = const()[name = string("op_4477_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 128]> var_4477_cast_fp16 = slice_by_index(begin = var_4477_begin_0, end = var_4477_end_0, end_mask = var_4477_end_mask_0, x = linear_28_cast_fp16)[name = string("op_4477_cast_fp16")];
            tensor<int32, [4]> var_4482 = const()[name = string("op_4482"), val = tensor<int32, [4]>([1, 32, 2, 64])];
            tensor<fp16, [1, 32, 2, 64]> x_79_cast_fp16 = reshape(shape = var_4482, x = var_4477_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<int32, [3]> var_4488_begin_0 = const()[name = string("op_4488_begin_0"), val = tensor<int32, [3]>([0, 0, 1152])];
            tensor<int32, [3]> var_4488_end_0 = const()[name = string("op_4488_end_0"), val = tensor<int32, [3]>([1, 32, 1280])];
            tensor<bool, [3]> var_4488_end_mask_0 = const()[name = string("op_4488_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 128]> var_4488_cast_fp16 = slice_by_index(begin = var_4488_begin_0, end = var_4488_end_0, end_mask = var_4488_end_mask_0, x = linear_28_cast_fp16)[name = string("op_4488_cast_fp16")];
            tensor<int32, [4]> var_4493 = const()[name = string("op_4493"), val = tensor<int32, [4]>([1, 32, 2, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_4494_cast_fp16 = reshape(shape = var_4493, x = var_4488_cast_fp16)[name = string("op_4494_cast_fp16")];
            tensor<int32, [4]> var_4499 = const()[name = string("op_4499"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4505_promoted_to_fp16 = const()[name = string("op_4505_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 8, 64]> var_4509_cast_fp16 = pow(x = x_73_cast_fp16, y = var_4505_promoted_to_fp16)[name = string("op_4509_cast_fp16")];
            tensor<int32, [1]> var_4511_axes_0 = const()[name = string("op_4511_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4511_keep_dims_0 = const()[name = string("op_4511_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 8, 1]> var_4511_cast_fp16 = reduce_mean(axes = var_4511_axes_0, keep_dims = var_4511_keep_dims_0, x = var_4509_cast_fp16)[name = string("op_4511_cast_fp16")];
            fp16 var_4512_to_fp16 = const()[name = string("op_4512_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 8, 1]> var_4513_cast_fp16 = add(x = var_4511_cast_fp16, y = var_4512_to_fp16)[name = string("op_4513_cast_fp16")];
            fp32 var_4514_epsilon_0 = const()[name = string("op_4514_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 8, 1]> var_4514_cast_fp16 = rsqrt(epsilon = var_4514_epsilon_0, x = var_4513_cast_fp16)[name = string("op_4514_cast_fp16")];
            tensor<fp16, [1, 32, 8, 64]> var_4515_cast_fp16 = mul(x = x_73_cast_fp16, y = var_4514_cast_fp16)[name = string("op_4515_cast_fp16")];
            tensor<fp16, [64]> layers_7_qn_w1_to_fp16 = const()[name = string("layers_7_qn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137088)))];
            tensor<fp16, [1, 32, 8, 64]> var_4516_cast_fp16 = mul(x = var_4515_cast_fp16, y = layers_7_qn_w1_to_fp16)[name = string("op_4516_cast_fp16")];
            tensor<int32, [4]> var_4521 = const()[name = string("op_4521"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_5_begin_0 = const()[name = string("r_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_5_end_0 = const()[name = string("r_5_end_0"), val = tensor<int32, [4]>([1, 8, 32, 16])];
            tensor<bool, [4]> r_5_end_mask_0 = const()[name = string("r_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 64]> x_77_cast_fp16 = transpose(perm = var_4521, x = var_4516_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 8, 32, 16]> r_5_cast_fp16 = slice_by_index(begin = r_5_begin_0, end = r_5_end_0, end_mask = r_5_end_mask_0, x = x_77_cast_fp16)[name = string("r_5_cast_fp16")];
            tensor<int32, [4]> p_5_begin_0 = const()[name = string("p_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_5_end_0 = const()[name = string("p_5_end_0"), val = tensor<int32, [4]>([1, 8, 32, 64])];
            tensor<bool, [4]> p_5_end_mask_0 = const()[name = string("p_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 48]> p_5_cast_fp16 = slice_by_index(begin = p_5_begin_0, end = p_5_end_0, end_mask = p_5_end_mask_0, x = x_77_cast_fp16)[name = string("p_5_cast_fp16")];
            tensor<int32, [4]> var_4537_begin_0 = const()[name = string("op_4537_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_4537_end_0 = const()[name = string("op_4537_end_0"), val = tensor<int32, [4]>([1, 8, 32, 16])];
            tensor<bool, [4]> var_4537_end_mask_0 = const()[name = string("op_4537_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 8]> var_4537_cast_fp16 = slice_by_index(begin = var_4537_begin_0, end = var_4537_end_0, end_mask = var_4537_end_mask_0, x = r_5_cast_fp16)[name = string("op_4537_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 32, 8]> var_4538_cast_fp16 = mul(x = var_4537_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_4538_cast_fp16")];
            tensor<int32, [4]> var_4543_begin_0 = const()[name = string("op_4543_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4543_end_0 = const()[name = string("op_4543_end_0"), val = tensor<int32, [4]>([1, 8, 32, 8])];
            tensor<bool, [4]> var_4543_end_mask_0 = const()[name = string("op_4543_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 8]> var_4543_cast_fp16 = slice_by_index(begin = var_4543_begin_0, end = var_4543_end_0, end_mask = var_4543_end_mask_0, x = r_5_cast_fp16)[name = string("op_4543_cast_fp16")];
            int32 var_4545 = const()[name = string("op_4545"), val = int32(-1)];
            bool rh_5_interleave_0 = const()[name = string("rh_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 16]> rh_5_cast_fp16 = concat(axis = var_4545, interleave = rh_5_interleave_0, values = (var_4538_cast_fp16, var_4543_cast_fp16))[name = string("rh_5_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> var_4547_cast_fp16 = mul(x = r_5_cast_fp16, y = cos_to_fp16)[name = string("op_4547_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> var_4548_cast_fp16 = mul(x = rh_5_cast_fp16, y = sin_to_fp16)[name = string("op_4548_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> var_4550_cast_fp16 = add(x = var_4547_cast_fp16, y = var_4548_cast_fp16)[name = string("op_4550_cast_fp16")];
            int32 var_4552 = const()[name = string("op_4552"), val = int32(-1)];
            bool q_interleave_0 = const()[name = string("q_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 64]> q_cast_fp16 = concat(axis = var_4552, interleave = q_interleave_0, values = (var_4550_cast_fp16, p_5_cast_fp16))[name = string("q_cast_fp16")];
            fp16 var_4558_promoted_to_fp16 = const()[name = string("op_4558_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 2, 64]> var_4562_cast_fp16 = pow(x = x_79_cast_fp16, y = var_4558_promoted_to_fp16)[name = string("op_4562_cast_fp16")];
            tensor<int32, [1]> var_4564_axes_0 = const()[name = string("op_4564_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4564_keep_dims_0 = const()[name = string("op_4564_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 2, 1]> var_4564_cast_fp16 = reduce_mean(axes = var_4564_axes_0, keep_dims = var_4564_keep_dims_0, x = var_4562_cast_fp16)[name = string("op_4564_cast_fp16")];
            fp16 var_4565_to_fp16 = const()[name = string("op_4565_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 2, 1]> var_4566_cast_fp16 = add(x = var_4564_cast_fp16, y = var_4565_to_fp16)[name = string("op_4566_cast_fp16")];
            fp32 var_4567_epsilon_0 = const()[name = string("op_4567_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 2, 1]> var_4567_cast_fp16 = rsqrt(epsilon = var_4567_epsilon_0, x = var_4566_cast_fp16)[name = string("op_4567_cast_fp16")];
            tensor<fp16, [1, 32, 2, 64]> var_4568_cast_fp16 = mul(x = x_79_cast_fp16, y = var_4567_cast_fp16)[name = string("op_4568_cast_fp16")];
            tensor<fp16, [64]> layers_7_kn_w1_to_fp16 = const()[name = string("layers_7_kn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137280)))];
            tensor<fp16, [1, 32, 2, 64]> var_4569_cast_fp16 = mul(x = var_4568_cast_fp16, y = layers_7_kn_w1_to_fp16)[name = string("op_4569_cast_fp16")];
            tensor<int32, [4]> var_4574 = const()[name = string("op_4574"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_begin_0 = const()[name = string("r_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_end_0 = const()[name = string("r_end_0"), val = tensor<int32, [4]>([1, 2, 32, 16])];
            tensor<bool, [4]> r_end_mask_0 = const()[name = string("r_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 64]> x_83_cast_fp16 = transpose(perm = var_4574, x = var_4569_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 2, 32, 16]> r_cast_fp16 = slice_by_index(begin = r_begin_0, end = r_end_0, end_mask = r_end_mask_0, x = x_83_cast_fp16)[name = string("r_cast_fp16")];
            tensor<int32, [4]> p_begin_0 = const()[name = string("p_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_end_0 = const()[name = string("p_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> p_end_mask_0 = const()[name = string("p_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 48]> p_cast_fp16 = slice_by_index(begin = p_begin_0, end = p_end_0, end_mask = p_end_mask_0, x = x_83_cast_fp16)[name = string("p_cast_fp16")];
            tensor<int32, [4]> var_4590_begin_0 = const()[name = string("op_4590_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_4590_end_0 = const()[name = string("op_4590_end_0"), val = tensor<int32, [4]>([1, 2, 32, 16])];
            tensor<bool, [4]> var_4590_end_mask_0 = const()[name = string("op_4590_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 8]> var_4590_cast_fp16 = slice_by_index(begin = var_4590_begin_0, end = var_4590_end_0, end_mask = var_4590_end_mask_0, x = r_cast_fp16)[name = string("op_4590_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 8]> var_4591_cast_fp16 = mul(x = var_4590_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_4591_cast_fp16")];
            tensor<int32, [4]> var_4596_begin_0 = const()[name = string("op_4596_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4596_end_0 = const()[name = string("op_4596_end_0"), val = tensor<int32, [4]>([1, 2, 32, 8])];
            tensor<bool, [4]> var_4596_end_mask_0 = const()[name = string("op_4596_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 8]> var_4596_cast_fp16 = slice_by_index(begin = var_4596_begin_0, end = var_4596_end_0, end_mask = var_4596_end_mask_0, x = r_cast_fp16)[name = string("op_4596_cast_fp16")];
            int32 var_4598 = const()[name = string("op_4598"), val = int32(-1)];
            bool rh_interleave_0 = const()[name = string("rh_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 16]> rh_cast_fp16 = concat(axis = var_4598, interleave = rh_interleave_0, values = (var_4591_cast_fp16, var_4596_cast_fp16))[name = string("rh_cast_fp16")];
            tensor<fp16, [1, 2, 32, 16]> var_4600_cast_fp16 = mul(x = r_cast_fp16, y = cos_to_fp16)[name = string("op_4600_cast_fp16")];
            tensor<fp16, [1, 2, 32, 16]> var_4601_cast_fp16 = mul(x = rh_cast_fp16, y = sin_to_fp16)[name = string("op_4601_cast_fp16")];
            tensor<fp16, [1, 2, 32, 16]> var_4603_cast_fp16 = add(x = var_4600_cast_fp16, y = var_4601_cast_fp16)[name = string("op_4603_cast_fp16")];
            int32 var_4605 = const()[name = string("op_4605"), val = int32(-1)];
            bool k_77_interleave_0 = const()[name = string("k_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> k_77_cast_fp16 = concat(axis = var_4605, interleave = k_77_interleave_0, values = (var_4603_cast_fp16, p_cast_fp16))[name = string("k_77_cast_fp16")];
            tensor<int32, [1]> var_4618_axes_0 = const()[name = string("op_4618_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_4618_cast_fp16 = expand_dims(axes = var_4618_axes_0, x = k_77_cast_fp16)[name = string("op_4618_cast_fp16")];
            tensor<int32, [5]> var_4626_reps_0 = const()[name = string("op_4626_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 2, 4, 32, 64]> var_4626_cast_fp16 = tile(reps = var_4626_reps_0, x = var_4618_cast_fp16)[name = string("op_4626_cast_fp16")];
            tensor<int32, [4]> var_4631 = const()[name = string("op_4631"), val = tensor<int32, [4]>([1, 8, 32, 64])];
            tensor<fp16, [1, 8, 32, 64]> k_cast_fp16 = reshape(shape = var_4631, x = var_4626_cast_fp16)[name = string("k_cast_fp16")];
            tensor<int32, [1]> var_4644_axes_0 = const()[name = string("op_4644_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> v_17_cast_fp16 = transpose(perm = var_4499, x = var_4494_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_4644_cast_fp16 = expand_dims(axes = var_4644_axes_0, x = v_17_cast_fp16)[name = string("op_4644_cast_fp16")];
            tensor<int32, [5]> var_4652_reps_0 = const()[name = string("op_4652_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 2, 4, 32, 64]> var_4652_cast_fp16 = tile(reps = var_4652_reps_0, x = var_4644_cast_fp16)[name = string("op_4652_cast_fp16")];
            tensor<int32, [4]> var_4657 = const()[name = string("op_4657"), val = tensor<int32, [4]>([1, 8, 32, 64])];
            tensor<fp16, [1, 8, 32, 64]> v_cast_fp16 = reshape(shape = var_4657, x = var_4652_cast_fp16)[name = string("v_cast_fp16")];
            bool var_4662_transpose_x_1 = const()[name = string("op_4662_transpose_x_1"), val = bool(false)];
            bool var_4662_transpose_y_1 = const()[name = string("op_4662_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 32]> var_4662_cast_fp16 = matmul(transpose_x = var_4662_transpose_x_1, transpose_y = var_4662_transpose_y_1, x = q_cast_fp16, y = k_cast_fp16)[name = string("op_4662_cast_fp16")];
            fp16 var_4663_to_fp16 = const()[name = string("op_4663_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 8, 32, 32]> var_4664_cast_fp16 = mul(x = var_4662_cast_fp16, y = var_4663_to_fp16)[name = string("op_4664_cast_fp16")];
            tensor<fp16, [1, 8, 32, 32]> s_cast_fp16 = add(x = var_4664_cast_fp16, y = causal_add_to_fp16)[name = string("s_cast_fp16")];
            int32 var_4667 = const()[name = string("op_4667"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 32]> var_4669_cast_fp16 = softmax(axis = var_4667, x = s_cast_fp16)[name = string("op_4669_cast_fp16")];
            bool o_39_transpose_x_0 = const()[name = string("o_39_transpose_x_0"), val = bool(false)];
            bool o_39_transpose_y_0 = const()[name = string("o_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 64]> o_39_cast_fp16 = matmul(transpose_x = o_39_transpose_x_0, transpose_y = o_39_transpose_y_0, x = var_4669_cast_fp16, y = v_cast_fp16)[name = string("o_39_cast_fp16")];
            tensor<int32, [4]> var_4675 = const()[name = string("op_4675"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4680 = const()[name = string("op_4680"), val = tensor<int32, [3]>([1, 32, 512])];
            tensor<fp16, [1, 32, 8, 64]> var_4676_cast_fp16 = transpose(perm = var_4675, x = o_39_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 32, 512]> var_4681_cast_fp16 = reshape(shape = var_4680, x = var_4676_cast_fp16)[name = string("op_4681_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4682_cast_fp16 = sigmoid(x = gate_cast_fp16)[name = string("op_4682_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_109_cast_fp16 = mul(x = var_4681_cast_fp16, y = var_4682_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [640, 512]> layers_7_o_weight_to_fp16 = const()[name = string("layers_7_o_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137472)))];
            tensor<fp16, [1, 32, 640]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_o_weight_to_fp16, x = input_109_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<fp16, [1, 32, 640]> x_85_cast_fp16 = add(x = x_69_cast_fp16, y = linear_29_cast_fp16)[name = string("x_85_cast_fp16")];
            fp16 var_4695_to_fp16 = const()[name = string("op_4695_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [1, 32, 640]> x_87_cast_fp16 = mul(x = x_85_cast_fp16, y = var_4695_to_fp16)[name = string("x_87_cast_fp16")];
            fp16 var_4693_promoted_to_fp16 = const()[name = string("op_4693_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 640]> var_4697_cast_fp16 = pow(x = x_87_cast_fp16, y = var_4693_promoted_to_fp16)[name = string("op_4697_cast_fp16")];
            tensor<int32, [1]> var_4699_axes_0 = const()[name = string("op_4699_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4699_keep_dims_0 = const()[name = string("op_4699_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4699_cast_fp16 = reduce_mean(axes = var_4699_axes_0, keep_dims = var_4699_keep_dims_0, x = var_4697_cast_fp16)[name = string("op_4699_cast_fp16")];
            fp16 var_4700_to_fp16 = const()[name = string("op_4700_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 32, 1]> var_4701_cast_fp16 = add(x = var_4699_cast_fp16, y = var_4700_to_fp16)[name = string("op_4701_cast_fp16")];
            fp32 var_4702_epsilon_0 = const()[name = string("op_4702_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4702_cast_fp16 = rsqrt(epsilon = var_4702_epsilon_0, x = var_4701_cast_fp16)[name = string("op_4702_cast_fp16")];
            tensor<fp16, [1, 32, 640]> var_4703_cast_fp16 = mul(x = x_87_cast_fp16, y = var_4702_cast_fp16)[name = string("op_4703_cast_fp16")];
            tensor<fp16, [640]> layers_7_n2_w1_to_fp16 = const()[name = string("layers_7_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109792896)))];
            tensor<fp16, [1, 32, 640]> input_111_cast_fp16 = mul(x = var_4703_cast_fp16, y = layers_7_n2_w1_to_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_7_gate_up_weight_to_fp16 = const()[name = string("layers_7_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109794240)))];
            tensor<fp16, [1, 32, 5120]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_7_gate_up_weight_to_fp16, x = input_111_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [3]> input_113_begin_0 = const()[name = string("input_113_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_113_end_0 = const()[name = string("input_113_end_0"), val = tensor<int32, [3]>([1, 32, 2560])];
            tensor<bool, [3]> input_113_end_mask_0 = const()[name = string("input_113_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 32, 2560]> input_113_cast_fp16 = slice_by_index(begin = input_113_begin_0, end = input_113_end_0, end_mask = input_113_end_mask_0, x = linear_30_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [1, 32, 2560]> var_4713_cast_fp16 = silu(x = input_113_cast_fp16)[name = string("op_4713_cast_fp16")];
            tensor<int32, [3]> var_4718_begin_0 = const()[name = string("op_4718_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_4718_end_0 = const()[name = string("op_4718_end_0"), val = tensor<int32, [3]>([1, 32, 5120])];
            tensor<bool, [3]> var_4718_end_mask_0 = const()[name = string("op_4718_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 32, 2560]> var_4718_cast_fp16 = slice_by_index(begin = var_4718_begin_0, end = var_4718_end_0, end_mask = var_4718_end_mask_0, x = linear_30_cast_fp16)[name = string("op_4718_cast_fp16")];
            tensor<fp16, [1, 32, 2560]> input_cast_fp16 = mul(x = var_4713_cast_fp16, y = var_4718_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_7_down_weight_to_fp16 = const()[name = string("layers_7_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116347904)))];
            tensor<fp16, [1, 32, 640]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_down_weight_to_fp16, x = input_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 32, 640]> x_89_cast_fp16 = add(x = x_85_cast_fp16, y = linear_31_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<int32, [1]> var_4736_axes_0 = const()[name = string("op_4736_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_4736 = expand_dims(axes = var_4736_axes_0, x = last)[name = string("op_4736")];
            tensor<bool, [1, 32]> var_4737 = equal(x = pos, y = var_4736)[name = string("op_4737")];
            tensor<int32, [1]> var_4754_axes_0 = const()[name = string("op_4754_axes_0"), val = tensor<int32, [1]>([2])];
            string sel_to_fp16_dtype_0 = const()[name = string("sel_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 32]> var_4737_to_fp16 = cast(dtype = sel_to_fp16_dtype_0, x = var_4737)[name = string("cast_83")];
            tensor<fp16, [1, 32, 1]> var_4754_cast_fp16 = expand_dims(axes = var_4754_axes_0, x = var_4737_to_fp16)[name = string("op_4754_cast_fp16")];
            tensor<fp16, [1, 32, 640]> var_4755_cast_fp16 = mul(x = var_4754_cast_fp16, y = x_89_cast_fp16)[name = string("op_4755_cast_fp16")];
            tensor<int32, [1]> x_91_axes_0 = const()[name = string("x_91_axes_0"), val = tensor<int32, [1]>([1])];
            bool x_91_keep_dims_0 = const()[name = string("x_91_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 640]> x_91_cast_fp16 = reduce_sum(axes = x_91_axes_0, keep_dims = x_91_keep_dims_0, x = var_4755_cast_fp16)[name = string("x_91_cast_fp16")];
            fp16 var_4767_to_fp16 = const()[name = string("op_4767_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 640]> x_cast_fp16 = mul(x = x_91_cast_fp16, y = var_4767_to_fp16)[name = string("x_cast_fp16")];
            fp16 var_4765_promoted_to_fp16 = const()[name = string("op_4765_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640]> var_4769_cast_fp16 = pow(x = x_cast_fp16, y = var_4765_promoted_to_fp16)[name = string("op_4769_cast_fp16")];
            tensor<int32, [1]> var_4771_axes_0 = const()[name = string("op_4771_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4771_keep_dims_0 = const()[name = string("op_4771_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_4771_cast_fp16 = reduce_mean(axes = var_4771_axes_0, keep_dims = var_4771_keep_dims_0, x = var_4769_cast_fp16)[name = string("op_4771_cast_fp16")];
            fp16 var_4772_to_fp16 = const()[name = string("op_4772_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 1]> var_4773_cast_fp16 = add(x = var_4771_cast_fp16, y = var_4772_to_fp16)[name = string("op_4773_cast_fp16")];
            fp32 var_4774_epsilon_0 = const()[name = string("op_4774_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1]> var_4774_cast_fp16 = rsqrt(epsilon = var_4774_epsilon_0, x = var_4773_cast_fp16)[name = string("op_4774_cast_fp16")];
            tensor<fp16, [1, 640]> var_4775_cast_fp16 = mul(x = x_cast_fp16, y = var_4774_cast_fp16)[name = string("op_4775_cast_fp16")];
            tensor<fp16, [640]> norm_w1_to_fp16 = const()[name = string("norm_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119624768)))];
            tensor<fp16, [1, 640]> h_cast_fp16 = mul(x = var_4775_cast_fp16, y = norm_w1_to_fp16)[name = string("h_cast_fp16")];
            tensor<fp16, [16000]> logits_bias_0_to_fp16 = const()[name = string("logits_bias_0_to_fp16"), val = tensor<fp16, [16000]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119626112)))];
            tensor<fp16, [1, 16000]> logits_cast_fp16 = linear(bias = logits_bias_0_to_fp16, weight = emb_to_fp16, x = h_cast_fp16)[name = string("logits_cast_fp16")];
            tensor<int32, [1]> reduce_max_0_axes_0 = const()[name = string("reduce_max_0_axes_0"), val = tensor<int32, [1]>([-1])];
            bool reduce_max_0_keep_dims_0 = const()[name = string("reduce_max_0_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> reduce_max_0_cast_fp16 = reduce_max(axes = reduce_max_0_axes_0, keep_dims = reduce_max_0_keep_dims_0, x = logits_cast_fp16)[name = string("reduce_max_0_cast_fp16")];
            tensor<fp16, [1, 16000]> var_4784_cast_fp16 = sub(x = logits_cast_fp16, y = reduce_max_0_cast_fp16)[name = string("op_4784_cast_fp16")];
            tensor<fp16, [1, 16000]> var_4785_cast_fp16 = exp(x = var_4784_cast_fp16)[name = string("op_4785_cast_fp16")];
            tensor<int32, [1]> var_4790_axes_0 = const()[name = string("op_4790_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4790_keep_dims_0 = const()[name = string("op_4790_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_4790_cast_fp16 = reduce_sum(axes = var_4790_axes_0, keep_dims = var_4790_keep_dims_0, x = var_4785_cast_fp16)[name = string("op_4790_cast_fp16")];
            fp32 var_4791_epsilon_0 = const()[name = string("op_4791_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 1]> var_4791_cast_fp16 = log(epsilon = var_4791_epsilon_0, x = var_4790_cast_fp16)[name = string("op_4791_cast_fp16")];
            tensor<fp16, [1, 1]> var_4793_cast_fp16 = add(x = reduce_max_0_cast_fp16, y = var_4791_cast_fp16)[name = string("op_4793_cast_fp16")];
            tensor<fp16, [1, 16000]> var_4795_cast_fp16 = sub(x = logits_cast_fp16, y = var_4793_cast_fp16)[name = string("op_4795_cast_fp16")];
            string var_4795_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_4795_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 16000]> lp = cast(dtype = var_4795_cast_fp16_to_fp32_dtype_0, x = var_4795_cast_fp16)[name = string("cast_82")];
        } -> (lp);
    func n64<ios18>(tensor<int32, [1, 64]> ids, tensor<int32, [1]> last) {
            tensor<int32, [1, 64]> pos = const()[name = string("pos"), val = tensor<int32, [1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119707072)))];
            int32 x_1_batch_dims_0 = const()[name = string("x_1_batch_dims_0"), val = int32(0)];
            bool x_1_validate_indices_0 = const()[name = string("x_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [16000, 640]> emb_to_fp16 = const()[name = string("emb_to_fp16"), val = tensor<fp16, [16000, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            string ids_to_int16_dtype_0 = const()[name = string("ids_to_int16_dtype_0"), val = string("int16")];
            string cast_105_dtype_0 = const()[name = string("cast_105_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int16, [1, 64]> ids_to_int16 = cast(dtype = ids_to_int16_dtype_0, x = ids)[name = string("cast_110")];
            tensor<int32, [1, 64]> cast_105 = cast(dtype = cast_105_dtype_0, x = ids_to_int16)[name = string("cast_109")];
            tensor<bool, [1, 64]> greater_equal_0 = greater_equal(x = cast_105, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(16000)];
            tensor<int32, [1, 64]> add_0 = add(x = cast_105, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 64]> select_0 = select(a = cast_105, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 x_1_cast_fp16_cast_uint16_axis_0 = const()[name = string("x_1_cast_fp16_cast_uint16_axis_0"), val = int32(0)];
            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
            tensor<int16, [1, 64]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_108")];
            tensor<fp16, [1, 64, 640]> x_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = x_1_cast_fp16_cast_uint16_axis_0, batch_dims = x_1_batch_dims_0, indices = select_0_to_int16, validate_indices = x_1_validate_indices_0, x = emb_to_fp16)[name = string("x_1_cast_fp16_cast_uint16_cast_uint16")];
            fp16 var_256_to_fp16 = const()[name = string("op_256_to_fp16"), val = fp16(0x1p+5)];
            tensor<fp16, [1, 64, 640]> x_3_cast_fp16 = mul(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = var_256_to_fp16)[name = string("x_3_cast_fp16")];
            fp16 var_254_promoted_to_fp16 = const()[name = string("op_254_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> var_258_cast_fp16 = pow(x = x_3_cast_fp16, y = var_254_promoted_to_fp16)[name = string("op_258_cast_fp16")];
            tensor<int32, [1]> var_260_axes_0 = const()[name = string("op_260_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_260_keep_dims_0 = const()[name = string("op_260_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_260_cast_fp16 = reduce_mean(axes = var_260_axes_0, keep_dims = var_260_keep_dims_0, x = var_258_cast_fp16)[name = string("op_260_cast_fp16")];
            fp16 var_261_to_fp16 = const()[name = string("op_261_to_fp16"), val = fp16(0x1.0c8p-10)];
            tensor<fp16, [1, 64, 1]> var_262_cast_fp16 = add(x = var_260_cast_fp16, y = var_261_to_fp16)[name = string("op_262_cast_fp16")];
            fp32 var_263_epsilon_0 = const()[name = string("op_263_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_263_cast_fp16 = rsqrt(epsilon = var_263_epsilon_0, x = var_262_cast_fp16)[name = string("op_263_cast_fp16")];
            tensor<fp16, [1, 64, 640]> var_264_cast_fp16 = mul(x = x_3_cast_fp16, y = var_263_cast_fp16)[name = string("op_264_cast_fp16")];
            tensor<fp16, [640]> layers_0_n1_w1_to_fp16 = const()[name = string("layers_0_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20480128)))];
            tensor<fp16, [1, 64, 640]> input_1_cast_fp16 = mul(x = var_264_cast_fp16, y = layers_0_n1_w1_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_0_qkvzba_weight_to_fp16 = const()[name = string("layers_0_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20481472)))];
            tensor<fp16, [1544]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [1544]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22457856)))];
            tensor<fp16, [1, 64, 1544]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_qkvzba_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [3]> qkv_1_begin_0 = const()[name = string("qkv_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_1_end_0 = const()[name = string("qkv_1_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> qkv_1_end_mask_0 = const()[name = string("qkv_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 1024]> qkv_1_cast_fp16 = slice_by_index(begin = qkv_1_begin_0, end = qkv_1_end_0, end_mask = qkv_1_end_mask_0, x = linear_0_cast_fp16)[name = string("qkv_1_cast_fp16")];
            tensor<int32, [3]> z_1_begin_0 = const()[name = string("z_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_1_end_0 = const()[name = string("z_1_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
            tensor<bool, [3]> z_1_end_mask_0 = const()[name = string("z_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 512]> z_1_cast_fp16 = slice_by_index(begin = z_1_begin_0, end = z_1_end_0, end_mask = z_1_end_mask_0, x = linear_0_cast_fp16)[name = string("z_1_cast_fp16")];
            tensor<int32, [3]> b_1_begin_0 = const()[name = string("b_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_1_end_0 = const()[name = string("b_1_end_0"), val = tensor<int32, [3]>([1, 64, 1540])];
            tensor<bool, [3]> b_1_end_mask_0 = const()[name = string("b_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 4]> b_1_cast_fp16 = slice_by_index(begin = b_1_begin_0, end = b_1_end_0, end_mask = b_1_end_mask_0, x = linear_0_cast_fp16)[name = string("b_1_cast_fp16")];
            tensor<int32, [3]> a_1_begin_0 = const()[name = string("a_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_1_end_0 = const()[name = string("a_1_end_0"), val = tensor<int32, [3]>([1, 64, 1])];
            tensor<bool, [3]> a_1_end_mask_0 = const()[name = string("a_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 4]> a_1_cast_fp16 = slice_by_index(begin = a_1_begin_0, end = a_1_end_0, end_mask = a_1_end_mask_0, x = linear_0_cast_fp16)[name = string("a_1_cast_fp16")];
            int32 var_303 = const()[name = string("op_303"), val = int32(1)];
            bool xp_1_interleave_0 = const()[name = string("xp_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 1024]> zpad_to_fp16 = const()[name = string("zpad_to_fp16"), val = tensor<fp16, [1, 3, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119700672)))];
            tensor<fp16, [1, 67, 1024]> xp_1_cast_fp16 = concat(axis = var_303, interleave = xp_1_interleave_0, values = (zpad_to_fp16, qkv_1_cast_fp16))[name = string("xp_1_cast_fp16")];
            tensor<int32, [3]> var_314_begin_0 = const()[name = string("op_314_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_314_end_0 = const()[name = string("op_314_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_314_end_mask_0 = const()[name = string("op_314_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 1024]> var_314_cast_fp16 = slice_by_index(begin = var_314_begin_0, end = var_314_end_0, end_mask = var_314_end_mask_0, x = xp_1_cast_fp16)[name = string("op_314_cast_fp16")];
            tensor<fp16, [1024]> var_322_to_fp16 = const()[name = string("op_322_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22479552)))];
            tensor<fp16, [1, 64, 1024]> var_323_cast_fp16 = mul(x = var_314_cast_fp16, y = var_322_to_fp16)[name = string("op_323_cast_fp16")];
            tensor<int32, [3]> var_333_begin_0 = const()[name = string("op_333_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_333_end_0 = const()[name = string("op_333_end_0"), val = tensor<int32, [3]>([1, 66, 1024])];
            tensor<bool, [3]> var_333_end_mask_0 = const()[name = string("op_333_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_333_cast_fp16 = slice_by_index(begin = var_333_begin_0, end = var_333_end_0, end_mask = var_333_end_mask_0, x = xp_1_cast_fp16)[name = string("op_333_cast_fp16")];
            tensor<fp16, [1024]> var_341_to_fp16 = const()[name = string("op_341_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22481664)))];
            tensor<fp16, [1, 64, 1024]> var_342_cast_fp16 = mul(x = var_333_cast_fp16, y = var_341_to_fp16)[name = string("op_342_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> var_344_cast_fp16 = add(x = var_323_cast_fp16, y = var_342_cast_fp16)[name = string("op_344_cast_fp16")];
            tensor<int32, [3]> var_354_begin_0 = const()[name = string("op_354_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_354_end_0 = const()[name = string("op_354_end_0"), val = tensor<int32, [3]>([1, 65, 1024])];
            tensor<bool, [3]> var_354_end_mask_0 = const()[name = string("op_354_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_354_cast_fp16 = slice_by_index(begin = var_354_begin_0, end = var_354_end_0, end_mask = var_354_end_mask_0, x = xp_1_cast_fp16)[name = string("op_354_cast_fp16")];
            tensor<fp16, [1024]> var_362_to_fp16 = const()[name = string("op_362_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22483776)))];
            tensor<fp16, [1, 64, 1024]> var_363_cast_fp16 = mul(x = var_354_cast_fp16, y = var_362_to_fp16)[name = string("op_363_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> var_365_cast_fp16 = add(x = var_344_cast_fp16, y = var_363_cast_fp16)[name = string("op_365_cast_fp16")];
            tensor<int32, [3]> var_375_begin_0 = const()[name = string("op_375_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_375_end_0 = const()[name = string("op_375_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> var_375_end_mask_0 = const()[name = string("op_375_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_375_cast_fp16 = slice_by_index(begin = var_375_begin_0, end = var_375_end_0, end_mask = var_375_end_mask_0, x = xp_1_cast_fp16)[name = string("op_375_cast_fp16")];
            tensor<fp16, [1024]> var_383_to_fp16 = const()[name = string("op_383_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22485888)))];
            tensor<fp16, [1, 64, 1024]> var_384_cast_fp16 = mul(x = var_375_cast_fp16, y = var_383_to_fp16)[name = string("op_384_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> input_3_cast_fp16 = add(x = var_365_cast_fp16, y = var_384_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> c_1_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("c_1_cast_fp16")];
            tensor<int32, [3]> var_392_begin_0 = const()[name = string("op_392_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_392_end_0 = const()[name = string("op_392_end_0"), val = tensor<int32, [3]>([1, 64, 256])];
            tensor<bool, [3]> var_392_end_mask_0 = const()[name = string("op_392_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 256]> var_392_cast_fp16 = slice_by_index(begin = var_392_begin_0, end = var_392_end_0, end_mask = var_392_end_mask_0, x = c_1_cast_fp16)[name = string("op_392_cast_fp16")];
            tensor<int32, [4]> var_397 = const()[name = string("op_397"), val = tensor<int32, [4]>([1, 64, 2, 128])];
            tensor<fp16, [1, 64, 2, 128]> q_1_cast_fp16 = reshape(shape = var_397, x = var_392_cast_fp16)[name = string("q_1_cast_fp16")];
            tensor<int32, [3]> var_403_begin_0 = const()[name = string("op_403_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_403_end_0 = const()[name = string("op_403_end_0"), val = tensor<int32, [3]>([1, 64, 512])];
            tensor<bool, [3]> var_403_end_mask_0 = const()[name = string("op_403_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 256]> var_403_cast_fp16 = slice_by_index(begin = var_403_begin_0, end = var_403_end_0, end_mask = var_403_end_mask_0, x = c_1_cast_fp16)[name = string("op_403_cast_fp16")];
            tensor<int32, [4]> var_408 = const()[name = string("op_408"), val = tensor<int32, [4]>([1, 64, 2, 128])];
            tensor<fp16, [1, 64, 2, 128]> k_1_cast_fp16 = reshape(shape = var_408, x = var_403_cast_fp16)[name = string("k_1_cast_fp16")];
            tensor<int32, [3]> var_414_begin_0 = const()[name = string("op_414_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_414_end_0 = const()[name = string("op_414_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> var_414_end_mask_0 = const()[name = string("op_414_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 512]> var_414_cast_fp16 = slice_by_index(begin = var_414_begin_0, end = var_414_end_0, end_mask = var_414_end_mask_0, x = c_1_cast_fp16)[name = string("op_414_cast_fp16")];
            tensor<int32, [4]> var_419 = const()[name = string("op_419"), val = tensor<int32, [4]>([1, 64, 4, 128])];
            tensor<fp16, [1, 64, 4, 128]> var_420_cast_fp16 = reshape(shape = var_419, x = var_414_cast_fp16)[name = string("op_420_cast_fp16")];
            tensor<int32, [4]> var_425 = const()[name = string("op_425"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_427_to_fp16 = const()[name = string("op_427_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 64, 2, 128]> q_3_cast_fp16 = mul(x = q_1_cast_fp16, y = var_427_to_fp16)[name = string("q_3_cast_fp16")];
            fp16 var_429_to_fp16 = const()[name = string("op_429_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 64, 2, 128]> k_3_cast_fp16 = mul(x = k_1_cast_fp16, y = var_429_to_fp16)[name = string("k_3_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_431_cast_fp16 = mul(x = q_3_cast_fp16, y = q_3_cast_fp16)[name = string("op_431_cast_fp16")];
            tensor<int32, [1]> var_436_axes_0 = const()[name = string("op_436_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_436_keep_dims_0 = const()[name = string("op_436_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 2, 1]> var_436_cast_fp16 = reduce_mean(axes = var_436_axes_0, keep_dims = var_436_keep_dims_0, x = var_431_cast_fp16)[name = string("op_436_cast_fp16")];
            fp16 var_437_promoted_to_fp16 = const()[name = string("op_437_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 64, 2, 1]> var_438_cast_fp16 = mul(x = var_436_cast_fp16, y = var_437_promoted_to_fp16)[name = string("op_438_cast_fp16")];
            fp16 var_440_to_fp16 = const()[name = string("op_440_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 64, 2, 1]> var_441_cast_fp16 = add(x = var_438_cast_fp16, y = var_440_to_fp16)[name = string("op_441_cast_fp16")];
            fp32 var_442_epsilon_0 = const()[name = string("op_442_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 2, 1]> var_442_cast_fp16 = rsqrt(epsilon = var_442_epsilon_0, x = var_441_cast_fp16)[name = string("op_442_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_443_cast_fp16 = mul(x = q_3_cast_fp16, y = var_442_cast_fp16)[name = string("op_443_cast_fp16")];
            fp16 var_444_to_fp16 = const()[name = string("op_444_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 64, 2, 128]> q_5_cast_fp16 = mul(x = var_443_cast_fp16, y = var_444_to_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_446_cast_fp16 = mul(x = k_3_cast_fp16, y = k_3_cast_fp16)[name = string("op_446_cast_fp16")];
            tensor<int32, [1]> var_451_axes_0 = const()[name = string("op_451_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_451_keep_dims_0 = const()[name = string("op_451_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 2, 1]> var_451_cast_fp16 = reduce_mean(axes = var_451_axes_0, keep_dims = var_451_keep_dims_0, x = var_446_cast_fp16)[name = string("op_451_cast_fp16")];
            fp16 var_452_promoted_to_fp16 = const()[name = string("op_452_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 64, 2, 1]> var_453_cast_fp16 = mul(x = var_451_cast_fp16, y = var_452_promoted_to_fp16)[name = string("op_453_cast_fp16")];
            fp16 var_455_to_fp16 = const()[name = string("op_455_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 64, 2, 1]> var_456_cast_fp16 = add(x = var_453_cast_fp16, y = var_455_to_fp16)[name = string("op_456_cast_fp16")];
            fp32 var_457_epsilon_0 = const()[name = string("op_457_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 2, 1]> var_457_cast_fp16 = rsqrt(epsilon = var_457_epsilon_0, x = var_456_cast_fp16)[name = string("op_457_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> k_5_cast_fp16 = mul(x = k_3_cast_fp16, y = var_457_cast_fp16)[name = string("k_5_cast_fp16")];
            tensor<int32, [4]> var_463 = const()[name = string("op_463"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_476_axes_0 = const()[name = string("op_476_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 128]> var_464_cast_fp16 = transpose(perm = var_463, x = q_5_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [1, 2, 1, 64, 128]> var_476_cast_fp16 = expand_dims(axes = var_476_axes_0, x = var_464_cast_fp16)[name = string("op_476_cast_fp16")];
            tensor<int32, [5]> var_484_reps_0 = const()[name = string("op_484_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 64, 128]> var_484_cast_fp16 = tile(reps = var_484_reps_0, x = var_476_cast_fp16)[name = string("op_484_cast_fp16")];
            tensor<int32, [4]> var_489 = const()[name = string("op_489"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> q_7_cast_fp16 = reshape(shape = var_489, x = var_484_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<int32, [4]> var_495 = const()[name = string("op_495"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_508_axes_0 = const()[name = string("op_508_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 128]> var_496_cast_fp16 = transpose(perm = var_495, x = k_5_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [1, 2, 1, 64, 128]> var_508_cast_fp16 = expand_dims(axes = var_508_axes_0, x = var_496_cast_fp16)[name = string("op_508_cast_fp16")];
            tensor<int32, [5]> var_516_reps_0 = const()[name = string("op_516_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 64, 128]> var_516_cast_fp16 = tile(reps = var_516_reps_0, x = var_508_cast_fp16)[name = string("op_516_cast_fp16")];
            tensor<int32, [4]> var_521 = const()[name = string("op_521"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> k_7_cast_fp16 = reshape(shape = var_521, x = var_516_cast_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_523_cast_fp16 = sigmoid(x = b_1_cast_fp16)[name = string("op_523_cast_fp16")];
            tensor<int32, [3]> var_527 = const()[name = string("op_527"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_0_dt_bias_to_fp16 = const()[name = string("layers_0_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 64, 4]> input_5_cast_fp16 = add(x = a_1_cast_fp16, y = layers_0_dt_bias_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_531_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("op_531_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_532_cast_fp16 = abs(x = input_5_cast_fp16)[name = string("op_532_cast_fp16")];
            fp16 const_0_promoted_to_fp16 = const()[name = string("const_0_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 64, 4]> var_533_cast_fp16 = mul(x = var_532_cast_fp16, y = const_0_promoted_to_fp16)[name = string("op_533_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_534_cast_fp16 = exp(x = var_533_cast_fp16)[name = string("op_534_cast_fp16")];
            fp16 var_536_to_fp16 = const()[name = string("op_536_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 64, 4]> var_537_cast_fp16 = add(x = var_534_cast_fp16, y = var_536_to_fp16)[name = string("op_537_cast_fp16")];
            fp32 var_538_epsilon_0 = const()[name = string("op_538_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 64, 4]> var_538_cast_fp16 = log(epsilon = var_538_epsilon_0, x = var_537_cast_fp16)[name = string("op_538_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_540_cast_fp16 = add(x = var_531_cast_fp16, y = var_538_cast_fp16)[name = string("op_540_cast_fp16")];
            tensor<fp16, [4]> layers_0_negA_to_fp16 = const()[name = string("layers_0_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.b3p+3, -0x1.a7cp+0, -0x1.ac4p+3, -0x1.704p+2])];
            tensor<fp16, [1, 64, 4]> var_541_cast_fp16 = mul(x = layers_0_negA_to_fp16, y = var_540_cast_fp16)[name = string("op_541_cast_fp16")];
            tensor<int32, [3]> var_545 = const()[name = string("op_545"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_548_axes_0 = const()[name = string("op_548_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 64]> beta_1_cast_fp16 = transpose(perm = var_527, x = var_523_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [1, 4, 64, 1]> var_548_cast_fp16 = expand_dims(axes = var_548_axes_0, x = beta_1_cast_fp16)[name = string("op_548_cast_fp16")];
            fp16 var_549_to_fp16 = const()[name = string("op_549_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 64, 1]> var_550_cast_fp16 = mul(x = var_548_cast_fp16, y = var_549_to_fp16)[name = string("op_550_cast_fp16")];
            tensor<fp16, [1, 4, 64, 128]> v_1_cast_fp16 = transpose(perm = var_425, x = var_420_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [1, 4, 64, 128]> vb_1_cast_fp16 = mul(x = v_1_cast_fp16, y = var_550_cast_fp16)[name = string("vb_1_cast_fp16")];
            tensor<int32, [4]> q_9_begin_0 = const()[name = string("q_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_9_end_0 = const()[name = string("q_9_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_9_end_mask_0 = const()[name = string("q_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_9_cast_fp16 = slice_by_index(begin = q_9_begin_0, end = q_9_end_0, end_mask = q_9_end_mask_0, x = q_7_cast_fp16)[name = string("q_9_cast_fp16")];
            tensor<int32, [4]> k_9_begin_0 = const()[name = string("k_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_9_end_0 = const()[name = string("k_9_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_9_end_mask_0 = const()[name = string("k_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, x = k_7_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<fp16, [1, 4, 64, 128]> var_584_cast_fp16 = mul(x = k_7_cast_fp16, y = var_548_cast_fp16)[name = string("op_584_cast_fp16")];
            tensor<int32, [4]> kb_1_begin_0 = const()[name = string("kb_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_1_end_0 = const()[name = string("kb_1_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_1_end_mask_0 = const()[name = string("kb_1_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_1_cast_fp16 = slice_by_index(begin = kb_1_begin_0, end = kb_1_end_0, end_mask = kb_1_end_mask_0, x = var_584_cast_fp16)[name = string("kb_1_cast_fp16")];
            tensor<int32, [4]> vb_3_begin_0 = const()[name = string("vb_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_3_end_0 = const()[name = string("vb_3_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_3_end_mask_0 = const()[name = string("vb_3_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_3_cast_fp16 = slice_by_index(begin = vb_3_begin_0, end = vb_3_end_0, end_mask = vb_3_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_3_cast_fp16")];
            tensor<int32, [3]> gdec_3_begin_0 = const()[name = string("gdec_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_3_end_0 = const()[name = string("gdec_3_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_3_end_mask_0 = const()[name = string("gdec_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 64]> gdec_1_cast_fp16 = transpose(perm = var_545, x = var_541_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 4, 16]> gdec_3_cast_fp16 = slice_by_index(begin = gdec_3_begin_0, end = gdec_3_end_0, end_mask = gdec_3_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_3_cast_fp16")];
            tensor<int32, [2]> var_632 = const()[name = string("op_632"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_633_cast_fp16 = reshape(shape = var_632, x = gdec_3_cast_fp16)[name = string("op_633_cast_fp16")];
            tensor<fp16, [256, 16]> transpose_0_to_fp16 = const()[name = string("transpose_0_to_fp16"), val = tensor<fp16, [256, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22488000)))];
            tensor<fp16, [256]> var_634_bias_0_to_fp16 = const()[name = string("op_634_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22496256)))];
            tensor<fp16, [4, 256]> var_634_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_633_cast_fp16)[name = string("op_634_cast_fp16")];
            tensor<int32, [4]> var_639 = const()[name = string("op_639"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_1_cast_fp16 = reshape(shape = var_639, x = var_634_cast_fp16)[name = string("P_1_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_641_cast_fp16 = exp(x = P_1_cast_fp16)[name = string("op_641_cast_fp16")];
            tensor<fp16, [16, 16]> tril_incl_to_fp16 = const()[name = string("tril_incl_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22496832)))];
            tensor<fp16, [1, 4, 16, 16]> D_1_cast_fp16 = mul(x = var_641_cast_fp16, y = tril_incl_to_fp16)[name = string("D_1_cast_fp16")];
            bool kk_1_transpose_x_1 = const()[name = string("kk_1_transpose_x_1"), val = bool(false)];
            bool kk_1_transpose_y_1 = const()[name = string("kk_1_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_1_cast_fp16 = matmul(transpose_x = kk_1_transpose_x_1, transpose_y = kk_1_transpose_y_1, x = kb_1_cast_fp16, y = k_9_cast_fp16)[name = string("kk_1_cast_fp16")];
            fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_647_cast_fp16 = mul(x = kk_1_cast_fp16, y = const_1_promoted_to_fp16)[name = string("op_647_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_648_cast_fp16 = mul(x = var_647_cast_fp16, y = D_1_cast_fp16)[name = string("op_648_cast_fp16")];
            tensor<fp16, [16, 16]> cum_tail_to_fp16 = const()[name = string("cum_tail_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22497408)))];
            tensor<fp16, [1, 4, 16, 16]> N_1_cast_fp16 = mul(x = var_648_cast_fp16, y = cum_tail_to_fp16)[name = string("N_1_cast_fp16")];
            tensor<fp16, [16, 16]> neg_tril_strict_to_fp16 = const()[name = string("neg_tril_strict_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22497984)))];
            tensor<fp16, [1, 4, 16, 16]> var_650_cast_fp16 = mul(x = D_1_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_650_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_1_cast_fp16 = mul(x = kk_1_cast_fp16, y = var_650_cast_fp16)[name = string("Na_1_cast_fp16")];
            tensor<fp16, [16, 16]> eye_to_fp16 = const()[name = string("eye_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22498560)))];
            tensor<fp16, [1, 4, 16, 16]> Minv_1_cast_fp16 = add(x = eye_to_fp16, y = N_1_cast_fp16)[name = string("Minv_1_cast_fp16")];
            bool Np_1_transpose_x_0 = const()[name = string("Np_1_transpose_x_0"), val = bool(false)];
            bool Np_1_transpose_y_0 = const()[name = string("Np_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_1_cast_fp16 = matmul(transpose_x = Np_1_transpose_x_0, transpose_y = Np_1_transpose_y_0, x = N_1_cast_fp16, y = Na_1_cast_fp16)[name = string("Np_1_cast_fp16")];
            bool Npa_1_transpose_x_0 = const()[name = string("Npa_1_transpose_x_0"), val = bool(false)];
            bool Npa_1_transpose_y_0 = const()[name = string("Npa_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_1_cast_fp16 = matmul(transpose_x = Npa_1_transpose_x_0, transpose_y = Npa_1_transpose_y_0, x = Na_1_cast_fp16, y = N_1_cast_fp16)[name = string("Npa_1_cast_fp16")];
            bool var_656_transpose_x_0 = const()[name = string("op_656_transpose_x_0"), val = bool(false)];
            bool var_656_transpose_y_0 = const()[name = string("op_656_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_656_cast_fp16 = matmul(transpose_x = var_656_transpose_x_0, transpose_y = var_656_transpose_y_0, x = Minv_1_cast_fp16, y = Np_1_cast_fp16)[name = string("op_656_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_3_cast_fp16 = add(x = Minv_1_cast_fp16, y = var_656_cast_fp16)[name = string("Minv_3_cast_fp16")];
            bool Np_3_transpose_x_0 = const()[name = string("Np_3_transpose_x_0"), val = bool(false)];
            bool Np_3_transpose_y_0 = const()[name = string("Np_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_3_cast_fp16 = matmul(transpose_x = Np_3_transpose_x_0, transpose_y = Np_3_transpose_y_0, x = Np_1_cast_fp16, y = Npa_1_cast_fp16)[name = string("Np_3_cast_fp16")];
            bool Npa_3_transpose_x_0 = const()[name = string("Npa_3_transpose_x_0"), val = bool(false)];
            bool Npa_3_transpose_y_0 = const()[name = string("Npa_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_3_cast_fp16 = matmul(transpose_x = Npa_3_transpose_x_0, transpose_y = Npa_3_transpose_y_0, x = Npa_1_cast_fp16, y = Np_1_cast_fp16)[name = string("Npa_3_cast_fp16")];
            bool var_661_transpose_x_0 = const()[name = string("op_661_transpose_x_0"), val = bool(false)];
            bool var_661_transpose_y_0 = const()[name = string("op_661_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_661_cast_fp16 = matmul(transpose_x = var_661_transpose_x_0, transpose_y = var_661_transpose_y_0, x = Minv_3_cast_fp16, y = Np_3_cast_fp16)[name = string("op_661_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_5_cast_fp16 = add(x = Minv_3_cast_fp16, y = var_661_cast_fp16)[name = string("Minv_5_cast_fp16")];
            bool Np_5_transpose_x_0 = const()[name = string("Np_5_transpose_x_0"), val = bool(false)];
            bool Np_5_transpose_y_0 = const()[name = string("Np_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_5_cast_fp16 = matmul(transpose_x = Np_5_transpose_x_0, transpose_y = Np_5_transpose_y_0, x = Np_3_cast_fp16, y = Npa_3_cast_fp16)[name = string("Np_5_cast_fp16")];
            bool var_665_transpose_x_0 = const()[name = string("op_665_transpose_x_0"), val = bool(false)];
            bool var_665_transpose_y_0 = const()[name = string("op_665_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_665_cast_fp16 = matmul(transpose_x = var_665_transpose_x_0, transpose_y = var_665_transpose_y_0, x = Minv_5_cast_fp16, y = Np_5_cast_fp16)[name = string("op_665_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_7_cast_fp16 = add(x = Minv_5_cast_fp16, y = var_665_cast_fp16)[name = string("Minv_7_cast_fp16")];
            bool u_1_transpose_x_0 = const()[name = string("u_1_transpose_x_0"), val = bool(false)];
            bool u_1_transpose_y_0 = const()[name = string("u_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_1_cast_fp16 = matmul(transpose_x = u_1_transpose_x_0, transpose_y = u_1_transpose_y_0, x = Minv_7_cast_fp16, y = vb_3_cast_fp16)[name = string("u_1_cast_fp16")];
            bool var_672_transpose_x_1 = const()[name = string("op_672_transpose_x_1"), val = bool(false)];
            bool var_672_transpose_y_1 = const()[name = string("op_672_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_672_cast_fp16 = matmul(transpose_x = var_672_transpose_x_1, transpose_y = var_672_transpose_y_1, x = q_9_cast_fp16, y = k_9_cast_fp16)[name = string("op_672_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_1_cast_fp16 = mul(x = var_672_cast_fp16, y = D_1_cast_fp16)[name = string("A_1_cast_fp16")];
            bool var_674_transpose_x_0 = const()[name = string("op_674_transpose_x_0"), val = bool(false)];
            bool var_674_transpose_y_0 = const()[name = string("op_674_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_674_cast_fp16 = matmul(transpose_x = var_674_transpose_x_0, transpose_y = var_674_transpose_y_0, x = A_1_cast_fp16, y = u_1_cast_fp16)[name = string("op_674_cast_fp16")];
            tensor<fp16, [16, 16]> transpose_1_to_fp16 = const()[name = string("transpose_1_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658176)))];
            tensor<fp16, [16]> var_675_bias_0_to_fp16 = const()[name = string("op_675_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_675_cast_fp16 = linear(bias = var_675_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_3_cast_fp16)[name = string("op_675_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_1_cast_fp16 = exp(x = var_675_cast_fp16)[name = string("tail_1_cast_fp16")];
            tensor<int32, [1]> var_678_axes_0 = const()[name = string("op_678_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_678_cast_fp16 = expand_dims(axes = var_678_axes_0, x = tail_1_cast_fp16)[name = string("op_678_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_679_cast_fp16 = mul(x = k_9_cast_fp16, y = var_678_cast_fp16)[name = string("op_679_cast_fp16")];
            bool S_1_transpose_x_1 = const()[name = string("S_1_transpose_x_1"), val = bool(true)];
            bool S_1_transpose_y_1 = const()[name = string("S_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_1_cast_fp16 = matmul(transpose_x = S_1_transpose_x_1, transpose_y = S_1_transpose_y_1, x = var_679_cast_fp16, y = u_1_cast_fp16)[name = string("S_1_cast_fp16")];
            tensor<int32, [4]> q_11_begin_0 = const()[name = string("q_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_11_end_0 = const()[name = string("q_11_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> q_11_end_mask_0 = const()[name = string("q_11_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_11_cast_fp16 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, x = q_7_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<int32, [4]> k_11_begin_0 = const()[name = string("k_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_11_end_0 = const()[name = string("k_11_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> k_11_end_mask_0 = const()[name = string("k_11_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_11_cast_fp16 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, x = k_7_cast_fp16)[name = string("k_11_cast_fp16")];
            tensor<int32, [4]> kb_3_begin_0 = const()[name = string("kb_3_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_3_end_0 = const()[name = string("kb_3_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> kb_3_end_mask_0 = const()[name = string("kb_3_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_3_cast_fp16 = slice_by_index(begin = kb_3_begin_0, end = kb_3_end_0, end_mask = kb_3_end_mask_0, x = var_584_cast_fp16)[name = string("kb_3_cast_fp16")];
            tensor<int32, [4]> vb_5_begin_0 = const()[name = string("vb_5_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_5_end_0 = const()[name = string("vb_5_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> vb_5_end_mask_0 = const()[name = string("vb_5_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_5_cast_fp16 = slice_by_index(begin = vb_5_begin_0, end = vb_5_end_0, end_mask = vb_5_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_5_cast_fp16")];
            tensor<int32, [3]> gdec_5_begin_0 = const()[name = string("gdec_5_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_5_end_0 = const()[name = string("gdec_5_end_0"), val = tensor<int32, [3]>([1, 4, 32])];
            tensor<bool, [3]> gdec_5_end_mask_0 = const()[name = string("gdec_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_5_cast_fp16 = slice_by_index(begin = gdec_5_begin_0, end = gdec_5_end_0, end_mask = gdec_5_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_5_cast_fp16")];
            tensor<int32, [2]> var_764 = const()[name = string("op_764"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_765_cast_fp16 = reshape(shape = var_764, x = gdec_5_cast_fp16)[name = string("op_765_cast_fp16")];
            tensor<fp16, [4, 256]> var_766_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_765_cast_fp16)[name = string("op_766_cast_fp16")];
            tensor<int32, [4]> var_771 = const()[name = string("op_771"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_3_cast_fp16 = reshape(shape = var_771, x = var_766_cast_fp16)[name = string("P_3_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_773_cast_fp16 = exp(x = P_3_cast_fp16)[name = string("op_773_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_3_cast_fp16 = mul(x = var_773_cast_fp16, y = tril_incl_to_fp16)[name = string("D_3_cast_fp16")];
            bool kk_3_transpose_x_1 = const()[name = string("kk_3_transpose_x_1"), val = bool(false)];
            bool kk_3_transpose_y_1 = const()[name = string("kk_3_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_3_cast_fp16 = matmul(transpose_x = kk_3_transpose_x_1, transpose_y = kk_3_transpose_y_1, x = kb_3_cast_fp16, y = k_11_cast_fp16)[name = string("kk_3_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_779_cast_fp16 = mul(x = kk_3_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_779_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_780_cast_fp16 = mul(x = var_779_cast_fp16, y = D_3_cast_fp16)[name = string("op_780_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_3_cast_fp16 = mul(x = var_780_cast_fp16, y = cum_tail_to_fp16)[name = string("N_3_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_782_cast_fp16 = mul(x = D_3_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_782_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_3_cast_fp16 = mul(x = kk_3_cast_fp16, y = var_782_cast_fp16)[name = string("Na_3_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_9_cast_fp16 = add(x = eye_to_fp16, y = N_3_cast_fp16)[name = string("Minv_9_cast_fp16")];
            bool Np_7_transpose_x_0 = const()[name = string("Np_7_transpose_x_0"), val = bool(false)];
            bool Np_7_transpose_y_0 = const()[name = string("Np_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_7_cast_fp16 = matmul(transpose_x = Np_7_transpose_x_0, transpose_y = Np_7_transpose_y_0, x = N_3_cast_fp16, y = Na_3_cast_fp16)[name = string("Np_7_cast_fp16")];
            bool Npa_5_transpose_x_0 = const()[name = string("Npa_5_transpose_x_0"), val = bool(false)];
            bool Npa_5_transpose_y_0 = const()[name = string("Npa_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_5_cast_fp16 = matmul(transpose_x = Npa_5_transpose_x_0, transpose_y = Npa_5_transpose_y_0, x = Na_3_cast_fp16, y = N_3_cast_fp16)[name = string("Npa_5_cast_fp16")];
            bool var_788_transpose_x_0 = const()[name = string("op_788_transpose_x_0"), val = bool(false)];
            bool var_788_transpose_y_0 = const()[name = string("op_788_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_788_cast_fp16 = matmul(transpose_x = var_788_transpose_x_0, transpose_y = var_788_transpose_y_0, x = Minv_9_cast_fp16, y = Np_7_cast_fp16)[name = string("op_788_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_11_cast_fp16 = add(x = Minv_9_cast_fp16, y = var_788_cast_fp16)[name = string("Minv_11_cast_fp16")];
            bool Np_9_transpose_x_0 = const()[name = string("Np_9_transpose_x_0"), val = bool(false)];
            bool Np_9_transpose_y_0 = const()[name = string("Np_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_9_cast_fp16 = matmul(transpose_x = Np_9_transpose_x_0, transpose_y = Np_9_transpose_y_0, x = Np_7_cast_fp16, y = Npa_5_cast_fp16)[name = string("Np_9_cast_fp16")];
            bool Npa_7_transpose_x_0 = const()[name = string("Npa_7_transpose_x_0"), val = bool(false)];
            bool Npa_7_transpose_y_0 = const()[name = string("Npa_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_7_cast_fp16 = matmul(transpose_x = Npa_7_transpose_x_0, transpose_y = Npa_7_transpose_y_0, x = Npa_5_cast_fp16, y = Np_7_cast_fp16)[name = string("Npa_7_cast_fp16")];
            bool var_793_transpose_x_0 = const()[name = string("op_793_transpose_x_0"), val = bool(false)];
            bool var_793_transpose_y_0 = const()[name = string("op_793_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_793_cast_fp16 = matmul(transpose_x = var_793_transpose_x_0, transpose_y = var_793_transpose_y_0, x = Minv_11_cast_fp16, y = Np_9_cast_fp16)[name = string("op_793_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_13_cast_fp16 = add(x = Minv_11_cast_fp16, y = var_793_cast_fp16)[name = string("Minv_13_cast_fp16")];
            bool Np_11_transpose_x_0 = const()[name = string("Np_11_transpose_x_0"), val = bool(false)];
            bool Np_11_transpose_y_0 = const()[name = string("Np_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_11_cast_fp16 = matmul(transpose_x = Np_11_transpose_x_0, transpose_y = Np_11_transpose_y_0, x = Np_9_cast_fp16, y = Npa_7_cast_fp16)[name = string("Np_11_cast_fp16")];
            bool var_797_transpose_x_0 = const()[name = string("op_797_transpose_x_0"), val = bool(false)];
            bool var_797_transpose_y_0 = const()[name = string("op_797_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_797_cast_fp16 = matmul(transpose_x = var_797_transpose_x_0, transpose_y = var_797_transpose_y_0, x = Minv_13_cast_fp16, y = Np_11_cast_fp16)[name = string("op_797_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_15_cast_fp16 = add(x = Minv_13_cast_fp16, y = var_797_cast_fp16)[name = string("Minv_15_cast_fp16")];
            bool u_3_transpose_x_0 = const()[name = string("u_3_transpose_x_0"), val = bool(false)];
            bool u_3_transpose_y_0 = const()[name = string("u_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_3_cast_fp16 = matmul(transpose_x = u_3_transpose_x_0, transpose_y = u_3_transpose_y_0, x = Minv_15_cast_fp16, y = vb_5_cast_fp16)[name = string("u_3_cast_fp16")];
            bool var_804_transpose_x_1 = const()[name = string("op_804_transpose_x_1"), val = bool(false)];
            bool var_804_transpose_y_1 = const()[name = string("op_804_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_804_cast_fp16 = matmul(transpose_x = var_804_transpose_x_1, transpose_y = var_804_transpose_y_1, x = q_11_cast_fp16, y = k_11_cast_fp16)[name = string("op_804_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_3_cast_fp16 = mul(x = var_804_cast_fp16, y = D_3_cast_fp16)[name = string("A_3_cast_fp16")];
            tensor<fp16, [16]> var_806_bias_0_to_fp16 = const()[name = string("op_806_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_806_cast_fp16 = linear(bias = var_806_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_5_cast_fp16)[name = string("op_806_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_1_cast_fp16 = exp(x = var_806_cast_fp16)[name = string("cum_1_cast_fp16")];
            tensor<int32, [1]> var_809_axes_0 = const()[name = string("op_809_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_809_cast_fp16 = expand_dims(axes = var_809_axes_0, x = cum_1_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_810_cast_fp16 = mul(x = kb_3_cast_fp16, y = var_809_cast_fp16)[name = string("op_810_cast_fp16")];
            bool w_3_transpose_x_0 = const()[name = string("w_3_transpose_x_0"), val = bool(false)];
            bool w_3_transpose_y_0 = const()[name = string("w_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_3_cast_fp16 = matmul(transpose_x = w_3_transpose_x_0, transpose_y = w_3_transpose_y_0, x = Minv_15_cast_fp16, y = var_810_cast_fp16)[name = string("w_3_cast_fp16")];
            bool var_812_transpose_x_0 = const()[name = string("op_812_transpose_x_0"), val = bool(false)];
            bool var_812_transpose_y_0 = const()[name = string("op_812_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_812_cast_fp16 = matmul(transpose_x = var_812_transpose_x_0, transpose_y = var_812_transpose_y_0, x = w_3_cast_fp16, y = S_1_cast_fp16)[name = string("op_812_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_1_cast_fp16 = sub(x = u_3_cast_fp16, y = var_812_cast_fp16)[name = string("vnew_1_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_817_cast_fp16 = mul(x = q_11_cast_fp16, y = var_809_cast_fp16)[name = string("op_817_cast_fp16")];
            bool var_818_transpose_x_0 = const()[name = string("op_818_transpose_x_0"), val = bool(false)];
            bool var_818_transpose_y_0 = const()[name = string("op_818_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_818_cast_fp16 = matmul(transpose_x = var_818_transpose_x_0, transpose_y = var_818_transpose_y_0, x = var_817_cast_fp16, y = S_1_cast_fp16)[name = string("op_818_cast_fp16")];
            bool var_819_transpose_x_0 = const()[name = string("op_819_transpose_x_0"), val = bool(false)];
            bool var_819_transpose_y_0 = const()[name = string("op_819_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_819_cast_fp16 = matmul(transpose_x = var_819_transpose_x_0, transpose_y = var_819_transpose_y_0, x = A_3_cast_fp16, y = vnew_1_cast_fp16)[name = string("op_819_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_821_cast_fp16 = add(x = var_818_cast_fp16, y = var_819_cast_fp16)[name = string("op_821_cast_fp16")];
            tensor<fp16, [16]> var_822_bias_0_to_fp16 = const()[name = string("op_822_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_822_cast_fp16 = linear(bias = var_822_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_5_cast_fp16)[name = string("op_822_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_3_cast_fp16 = exp(x = var_822_cast_fp16)[name = string("tail_3_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_5_to_fp16 = const()[name = string("transpose_5_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_824_bias_0_to_fp16 = const()[name = string("op_824_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_824_cast_fp16 = linear(bias = var_824_bias_0_to_fp16, weight = transpose_5_to_fp16, x = gdec_5_cast_fp16)[name = string("op_824_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_825_cast_fp16 = exp(x = var_824_cast_fp16)[name = string("op_825_cast_fp16")];
            tensor<int32, [1]> last_3_axes_0 = const()[name = string("last_3_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_3_cast_fp16 = expand_dims(axes = last_3_axes_0, x = var_825_cast_fp16)[name = string("last_3_cast_fp16")];
            tensor<int32, [1]> var_829_axes_0 = const()[name = string("op_829_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_829_cast_fp16 = expand_dims(axes = var_829_axes_0, x = tail_3_cast_fp16)[name = string("op_829_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_830_cast_fp16 = mul(x = k_11_cast_fp16, y = var_829_cast_fp16)[name = string("op_830_cast_fp16")];
            bool upd_1_transpose_x_1 = const()[name = string("upd_1_transpose_x_1"), val = bool(true)];
            bool upd_1_transpose_y_1 = const()[name = string("upd_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_1_cast_fp16 = matmul(transpose_x = upd_1_transpose_x_1, transpose_y = upd_1_transpose_y_1, x = var_830_cast_fp16, y = vnew_1_cast_fp16)[name = string("upd_1_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_835_cast_fp16 = mul(x = S_1_cast_fp16, y = last_3_cast_fp16)[name = string("op_835_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_3_cast_fp16 = add(x = var_835_cast_fp16, y = upd_1_cast_fp16)[name = string("S_3_cast_fp16")];
            tensor<int32, [4]> q_13_begin_0 = const()[name = string("q_13_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_13_end_0 = const()[name = string("q_13_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> q_13_end_mask_0 = const()[name = string("q_13_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, x = q_7_cast_fp16)[name = string("q_13_cast_fp16")];
            tensor<int32, [4]> k_13_begin_0 = const()[name = string("k_13_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_13_end_0 = const()[name = string("k_13_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> k_13_end_mask_0 = const()[name = string("k_13_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, x = k_7_cast_fp16)[name = string("k_13_cast_fp16")];
            tensor<int32, [4]> kb_5_begin_0 = const()[name = string("kb_5_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_5_end_0 = const()[name = string("kb_5_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> kb_5_end_mask_0 = const()[name = string("kb_5_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_5_cast_fp16 = slice_by_index(begin = kb_5_begin_0, end = kb_5_end_0, end_mask = kb_5_end_mask_0, x = var_584_cast_fp16)[name = string("kb_5_cast_fp16")];
            tensor<int32, [4]> vb_7_begin_0 = const()[name = string("vb_7_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_7_end_0 = const()[name = string("vb_7_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> vb_7_end_mask_0 = const()[name = string("vb_7_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_7_cast_fp16 = slice_by_index(begin = vb_7_begin_0, end = vb_7_end_0, end_mask = vb_7_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_7_cast_fp16")];
            tensor<int32, [3]> gdec_7_begin_0 = const()[name = string("gdec_7_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_7_end_0 = const()[name = string("gdec_7_end_0"), val = tensor<int32, [3]>([1, 4, 48])];
            tensor<bool, [3]> gdec_7_end_mask_0 = const()[name = string("gdec_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_7_cast_fp16 = slice_by_index(begin = gdec_7_begin_0, end = gdec_7_end_0, end_mask = gdec_7_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_7_cast_fp16")];
            tensor<int32, [2]> var_918 = const()[name = string("op_918"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_919_cast_fp16 = reshape(shape = var_918, x = gdec_7_cast_fp16)[name = string("op_919_cast_fp16")];
            tensor<fp16, [4, 256]> var_920_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_919_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<int32, [4]> var_925 = const()[name = string("op_925"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_5_cast_fp16 = reshape(shape = var_925, x = var_920_cast_fp16)[name = string("P_5_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_927_cast_fp16 = exp(x = P_5_cast_fp16)[name = string("op_927_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_5_cast_fp16 = mul(x = var_927_cast_fp16, y = tril_incl_to_fp16)[name = string("D_5_cast_fp16")];
            bool kk_5_transpose_x_1 = const()[name = string("kk_5_transpose_x_1"), val = bool(false)];
            bool kk_5_transpose_y_1 = const()[name = string("kk_5_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_5_cast_fp16 = matmul(transpose_x = kk_5_transpose_x_1, transpose_y = kk_5_transpose_y_1, x = kb_5_cast_fp16, y = k_13_cast_fp16)[name = string("kk_5_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_933_cast_fp16 = mul(x = kk_5_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_933_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_934_cast_fp16 = mul(x = var_933_cast_fp16, y = D_5_cast_fp16)[name = string("op_934_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_5_cast_fp16 = mul(x = var_934_cast_fp16, y = cum_tail_to_fp16)[name = string("N_5_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_936_cast_fp16 = mul(x = D_5_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_936_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_5_cast_fp16 = mul(x = kk_5_cast_fp16, y = var_936_cast_fp16)[name = string("Na_5_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_17_cast_fp16 = add(x = eye_to_fp16, y = N_5_cast_fp16)[name = string("Minv_17_cast_fp16")];
            bool Np_13_transpose_x_0 = const()[name = string("Np_13_transpose_x_0"), val = bool(false)];
            bool Np_13_transpose_y_0 = const()[name = string("Np_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_13_cast_fp16 = matmul(transpose_x = Np_13_transpose_x_0, transpose_y = Np_13_transpose_y_0, x = N_5_cast_fp16, y = Na_5_cast_fp16)[name = string("Np_13_cast_fp16")];
            bool Npa_9_transpose_x_0 = const()[name = string("Npa_9_transpose_x_0"), val = bool(false)];
            bool Npa_9_transpose_y_0 = const()[name = string("Npa_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_9_cast_fp16 = matmul(transpose_x = Npa_9_transpose_x_0, transpose_y = Npa_9_transpose_y_0, x = Na_5_cast_fp16, y = N_5_cast_fp16)[name = string("Npa_9_cast_fp16")];
            bool var_942_transpose_x_0 = const()[name = string("op_942_transpose_x_0"), val = bool(false)];
            bool var_942_transpose_y_0 = const()[name = string("op_942_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_942_cast_fp16 = matmul(transpose_x = var_942_transpose_x_0, transpose_y = var_942_transpose_y_0, x = Minv_17_cast_fp16, y = Np_13_cast_fp16)[name = string("op_942_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_19_cast_fp16 = add(x = Minv_17_cast_fp16, y = var_942_cast_fp16)[name = string("Minv_19_cast_fp16")];
            bool Np_15_transpose_x_0 = const()[name = string("Np_15_transpose_x_0"), val = bool(false)];
            bool Np_15_transpose_y_0 = const()[name = string("Np_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_15_cast_fp16 = matmul(transpose_x = Np_15_transpose_x_0, transpose_y = Np_15_transpose_y_0, x = Np_13_cast_fp16, y = Npa_9_cast_fp16)[name = string("Np_15_cast_fp16")];
            bool Npa_11_transpose_x_0 = const()[name = string("Npa_11_transpose_x_0"), val = bool(false)];
            bool Npa_11_transpose_y_0 = const()[name = string("Npa_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_11_cast_fp16 = matmul(transpose_x = Npa_11_transpose_x_0, transpose_y = Npa_11_transpose_y_0, x = Npa_9_cast_fp16, y = Np_13_cast_fp16)[name = string("Npa_11_cast_fp16")];
            bool var_947_transpose_x_0 = const()[name = string("op_947_transpose_x_0"), val = bool(false)];
            bool var_947_transpose_y_0 = const()[name = string("op_947_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_947_cast_fp16 = matmul(transpose_x = var_947_transpose_x_0, transpose_y = var_947_transpose_y_0, x = Minv_19_cast_fp16, y = Np_15_cast_fp16)[name = string("op_947_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_21_cast_fp16 = add(x = Minv_19_cast_fp16, y = var_947_cast_fp16)[name = string("Minv_21_cast_fp16")];
            bool Np_17_transpose_x_0 = const()[name = string("Np_17_transpose_x_0"), val = bool(false)];
            bool Np_17_transpose_y_0 = const()[name = string("Np_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_17_cast_fp16 = matmul(transpose_x = Np_17_transpose_x_0, transpose_y = Np_17_transpose_y_0, x = Np_15_cast_fp16, y = Npa_11_cast_fp16)[name = string("Np_17_cast_fp16")];
            bool var_951_transpose_x_0 = const()[name = string("op_951_transpose_x_0"), val = bool(false)];
            bool var_951_transpose_y_0 = const()[name = string("op_951_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_951_cast_fp16 = matmul(transpose_x = var_951_transpose_x_0, transpose_y = var_951_transpose_y_0, x = Minv_21_cast_fp16, y = Np_17_cast_fp16)[name = string("op_951_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_23_cast_fp16 = add(x = Minv_21_cast_fp16, y = var_951_cast_fp16)[name = string("Minv_23_cast_fp16")];
            bool u_5_transpose_x_0 = const()[name = string("u_5_transpose_x_0"), val = bool(false)];
            bool u_5_transpose_y_0 = const()[name = string("u_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_5_cast_fp16 = matmul(transpose_x = u_5_transpose_x_0, transpose_y = u_5_transpose_y_0, x = Minv_23_cast_fp16, y = vb_7_cast_fp16)[name = string("u_5_cast_fp16")];
            bool var_958_transpose_x_1 = const()[name = string("op_958_transpose_x_1"), val = bool(false)];
            bool var_958_transpose_y_1 = const()[name = string("op_958_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_958_cast_fp16 = matmul(transpose_x = var_958_transpose_x_1, transpose_y = var_958_transpose_y_1, x = q_13_cast_fp16, y = k_13_cast_fp16)[name = string("op_958_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_5_cast_fp16 = mul(x = var_958_cast_fp16, y = D_5_cast_fp16)[name = string("A_5_cast_fp16")];
            tensor<fp16, [16]> var_960_bias_0_to_fp16 = const()[name = string("op_960_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_960_cast_fp16 = linear(bias = var_960_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_7_cast_fp16)[name = string("op_960_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_3_cast_fp16 = exp(x = var_960_cast_fp16)[name = string("cum_3_cast_fp16")];
            tensor<int32, [1]> var_963_axes_0 = const()[name = string("op_963_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_963_cast_fp16 = expand_dims(axes = var_963_axes_0, x = cum_3_cast_fp16)[name = string("op_963_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_964_cast_fp16 = mul(x = kb_5_cast_fp16, y = var_963_cast_fp16)[name = string("op_964_cast_fp16")];
            bool w_5_transpose_x_0 = const()[name = string("w_5_transpose_x_0"), val = bool(false)];
            bool w_5_transpose_y_0 = const()[name = string("w_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_5_cast_fp16 = matmul(transpose_x = w_5_transpose_x_0, transpose_y = w_5_transpose_y_0, x = Minv_23_cast_fp16, y = var_964_cast_fp16)[name = string("w_5_cast_fp16")];
            bool var_966_transpose_x_0 = const()[name = string("op_966_transpose_x_0"), val = bool(false)];
            bool var_966_transpose_y_0 = const()[name = string("op_966_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_966_cast_fp16 = matmul(transpose_x = var_966_transpose_x_0, transpose_y = var_966_transpose_y_0, x = w_5_cast_fp16, y = S_3_cast_fp16)[name = string("op_966_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_3_cast_fp16 = sub(x = u_5_cast_fp16, y = var_966_cast_fp16)[name = string("vnew_3_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_971_cast_fp16 = mul(x = q_13_cast_fp16, y = var_963_cast_fp16)[name = string("op_971_cast_fp16")];
            bool var_972_transpose_x_0 = const()[name = string("op_972_transpose_x_0"), val = bool(false)];
            bool var_972_transpose_y_0 = const()[name = string("op_972_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_972_cast_fp16 = matmul(transpose_x = var_972_transpose_x_0, transpose_y = var_972_transpose_y_0, x = var_971_cast_fp16, y = S_3_cast_fp16)[name = string("op_972_cast_fp16")];
            bool var_973_transpose_x_0 = const()[name = string("op_973_transpose_x_0"), val = bool(false)];
            bool var_973_transpose_y_0 = const()[name = string("op_973_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_973_cast_fp16 = matmul(transpose_x = var_973_transpose_x_0, transpose_y = var_973_transpose_y_0, x = A_5_cast_fp16, y = vnew_3_cast_fp16)[name = string("op_973_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_975_cast_fp16 = add(x = var_972_cast_fp16, y = var_973_cast_fp16)[name = string("op_975_cast_fp16")];
            tensor<fp16, [16]> var_976_bias_0_to_fp16 = const()[name = string("op_976_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_976_cast_fp16 = linear(bias = var_976_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_7_cast_fp16)[name = string("op_976_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_5_cast_fp16 = exp(x = var_976_cast_fp16)[name = string("tail_5_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_9_to_fp16 = const()[name = string("transpose_9_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_978_bias_0_to_fp16 = const()[name = string("op_978_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_978_cast_fp16 = linear(bias = var_978_bias_0_to_fp16, weight = transpose_9_to_fp16, x = gdec_7_cast_fp16)[name = string("op_978_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_979_cast_fp16 = exp(x = var_978_cast_fp16)[name = string("op_979_cast_fp16")];
            tensor<int32, [1]> last_5_axes_0 = const()[name = string("last_5_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_5_cast_fp16 = expand_dims(axes = last_5_axes_0, x = var_979_cast_fp16)[name = string("last_5_cast_fp16")];
            tensor<int32, [1]> var_983_axes_0 = const()[name = string("op_983_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_983_cast_fp16 = expand_dims(axes = var_983_axes_0, x = tail_5_cast_fp16)[name = string("op_983_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_984_cast_fp16 = mul(x = k_13_cast_fp16, y = var_983_cast_fp16)[name = string("op_984_cast_fp16")];
            bool upd_3_transpose_x_1 = const()[name = string("upd_3_transpose_x_1"), val = bool(true)];
            bool upd_3_transpose_y_1 = const()[name = string("upd_3_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_3_cast_fp16 = matmul(transpose_x = upd_3_transpose_x_1, transpose_y = upd_3_transpose_y_1, x = var_984_cast_fp16, y = vnew_3_cast_fp16)[name = string("upd_3_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_989_cast_fp16 = mul(x = S_3_cast_fp16, y = last_5_cast_fp16)[name = string("op_989_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_5_cast_fp16 = add(x = var_989_cast_fp16, y = upd_3_cast_fp16)[name = string("S_5_cast_fp16")];
            tensor<int32, [4]> q_15_begin_0 = const()[name = string("q_15_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_15_end_0 = const()[name = string("q_15_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_15_end_mask_0 = const()[name = string("q_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_15_cast_fp16 = slice_by_index(begin = q_15_begin_0, end = q_15_end_0, end_mask = q_15_end_mask_0, x = q_7_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<int32, [4]> k_15_begin_0 = const()[name = string("k_15_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_15_end_0 = const()[name = string("k_15_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_15_end_mask_0 = const()[name = string("k_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_15_cast_fp16 = slice_by_index(begin = k_15_begin_0, end = k_15_end_0, end_mask = k_15_end_mask_0, x = k_7_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> kb_7_begin_0 = const()[name = string("kb_7_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_7_end_0 = const()[name = string("kb_7_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_7_end_mask_0 = const()[name = string("kb_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_7_cast_fp16 = slice_by_index(begin = kb_7_begin_0, end = kb_7_end_0, end_mask = kb_7_end_mask_0, x = var_584_cast_fp16)[name = string("kb_7_cast_fp16")];
            tensor<int32, [4]> vb_9_begin_0 = const()[name = string("vb_9_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_9_end_0 = const()[name = string("vb_9_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_9_end_mask_0 = const()[name = string("vb_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_9_cast_fp16 = slice_by_index(begin = vb_9_begin_0, end = vb_9_end_0, end_mask = vb_9_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_9_cast_fp16")];
            tensor<int32, [3]> gdec_9_begin_0 = const()[name = string("gdec_9_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_9_end_0 = const()[name = string("gdec_9_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_9_end_mask_0 = const()[name = string("gdec_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_9_cast_fp16 = slice_by_index(begin = gdec_9_begin_0, end = gdec_9_end_0, end_mask = gdec_9_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_9_cast_fp16")];
            tensor<int32, [2]> var_1072 = const()[name = string("op_1072"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_1073_cast_fp16 = reshape(shape = var_1072, x = gdec_9_cast_fp16)[name = string("op_1073_cast_fp16")];
            tensor<fp16, [4, 256]> var_1074_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1073_cast_fp16)[name = string("op_1074_cast_fp16")];
            tensor<int32, [4]> var_1079 = const()[name = string("op_1079"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_7_cast_fp16 = reshape(shape = var_1079, x = var_1074_cast_fp16)[name = string("P_7_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1081_cast_fp16 = exp(x = P_7_cast_fp16)[name = string("op_1081_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_7_cast_fp16 = mul(x = var_1081_cast_fp16, y = tril_incl_to_fp16)[name = string("D_7_cast_fp16")];
            bool kk_7_transpose_x_1 = const()[name = string("kk_7_transpose_x_1"), val = bool(false)];
            bool kk_7_transpose_y_1 = const()[name = string("kk_7_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_7_cast_fp16 = matmul(transpose_x = kk_7_transpose_x_1, transpose_y = kk_7_transpose_y_1, x = kb_7_cast_fp16, y = k_15_cast_fp16)[name = string("kk_7_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_1087_cast_fp16 = mul(x = kk_7_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_1087_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1088_cast_fp16 = mul(x = var_1087_cast_fp16, y = D_7_cast_fp16)[name = string("op_1088_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_7_cast_fp16 = mul(x = var_1088_cast_fp16, y = cum_tail_to_fp16)[name = string("N_7_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1090_cast_fp16 = mul(x = D_7_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1090_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_7_cast_fp16 = mul(x = kk_7_cast_fp16, y = var_1090_cast_fp16)[name = string("Na_7_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_25_cast_fp16 = add(x = eye_to_fp16, y = N_7_cast_fp16)[name = string("Minv_25_cast_fp16")];
            bool Np_19_transpose_x_0 = const()[name = string("Np_19_transpose_x_0"), val = bool(false)];
            bool Np_19_transpose_y_0 = const()[name = string("Np_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_19_cast_fp16 = matmul(transpose_x = Np_19_transpose_x_0, transpose_y = Np_19_transpose_y_0, x = N_7_cast_fp16, y = Na_7_cast_fp16)[name = string("Np_19_cast_fp16")];
            bool Npa_13_transpose_x_0 = const()[name = string("Npa_13_transpose_x_0"), val = bool(false)];
            bool Npa_13_transpose_y_0 = const()[name = string("Npa_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_13_cast_fp16 = matmul(transpose_x = Npa_13_transpose_x_0, transpose_y = Npa_13_transpose_y_0, x = Na_7_cast_fp16, y = N_7_cast_fp16)[name = string("Npa_13_cast_fp16")];
            bool var_1096_transpose_x_0 = const()[name = string("op_1096_transpose_x_0"), val = bool(false)];
            bool var_1096_transpose_y_0 = const()[name = string("op_1096_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1096_cast_fp16 = matmul(transpose_x = var_1096_transpose_x_0, transpose_y = var_1096_transpose_y_0, x = Minv_25_cast_fp16, y = Np_19_cast_fp16)[name = string("op_1096_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_27_cast_fp16 = add(x = Minv_25_cast_fp16, y = var_1096_cast_fp16)[name = string("Minv_27_cast_fp16")];
            bool Np_21_transpose_x_0 = const()[name = string("Np_21_transpose_x_0"), val = bool(false)];
            bool Np_21_transpose_y_0 = const()[name = string("Np_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_21_cast_fp16 = matmul(transpose_x = Np_21_transpose_x_0, transpose_y = Np_21_transpose_y_0, x = Np_19_cast_fp16, y = Npa_13_cast_fp16)[name = string("Np_21_cast_fp16")];
            bool Npa_15_transpose_x_0 = const()[name = string("Npa_15_transpose_x_0"), val = bool(false)];
            bool Npa_15_transpose_y_0 = const()[name = string("Npa_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_15_cast_fp16 = matmul(transpose_x = Npa_15_transpose_x_0, transpose_y = Npa_15_transpose_y_0, x = Npa_13_cast_fp16, y = Np_19_cast_fp16)[name = string("Npa_15_cast_fp16")];
            bool var_1101_transpose_x_0 = const()[name = string("op_1101_transpose_x_0"), val = bool(false)];
            bool var_1101_transpose_y_0 = const()[name = string("op_1101_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1101_cast_fp16 = matmul(transpose_x = var_1101_transpose_x_0, transpose_y = var_1101_transpose_y_0, x = Minv_27_cast_fp16, y = Np_21_cast_fp16)[name = string("op_1101_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_29_cast_fp16 = add(x = Minv_27_cast_fp16, y = var_1101_cast_fp16)[name = string("Minv_29_cast_fp16")];
            bool Np_23_transpose_x_0 = const()[name = string("Np_23_transpose_x_0"), val = bool(false)];
            bool Np_23_transpose_y_0 = const()[name = string("Np_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_23_cast_fp16 = matmul(transpose_x = Np_23_transpose_x_0, transpose_y = Np_23_transpose_y_0, x = Np_21_cast_fp16, y = Npa_15_cast_fp16)[name = string("Np_23_cast_fp16")];
            bool var_1105_transpose_x_0 = const()[name = string("op_1105_transpose_x_0"), val = bool(false)];
            bool var_1105_transpose_y_0 = const()[name = string("op_1105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1105_cast_fp16 = matmul(transpose_x = var_1105_transpose_x_0, transpose_y = var_1105_transpose_y_0, x = Minv_29_cast_fp16, y = Np_23_cast_fp16)[name = string("op_1105_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_31_cast_fp16 = add(x = Minv_29_cast_fp16, y = var_1105_cast_fp16)[name = string("Minv_31_cast_fp16")];
            bool u_7_transpose_x_0 = const()[name = string("u_7_transpose_x_0"), val = bool(false)];
            bool u_7_transpose_y_0 = const()[name = string("u_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_7_cast_fp16 = matmul(transpose_x = u_7_transpose_x_0, transpose_y = u_7_transpose_y_0, x = Minv_31_cast_fp16, y = vb_9_cast_fp16)[name = string("u_7_cast_fp16")];
            bool var_1112_transpose_x_1 = const()[name = string("op_1112_transpose_x_1"), val = bool(false)];
            bool var_1112_transpose_y_1 = const()[name = string("op_1112_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_1112_cast_fp16 = matmul(transpose_x = var_1112_transpose_x_1, transpose_y = var_1112_transpose_y_1, x = q_15_cast_fp16, y = k_15_cast_fp16)[name = string("op_1112_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_7_cast_fp16 = mul(x = var_1112_cast_fp16, y = D_7_cast_fp16)[name = string("A_7_cast_fp16")];
            tensor<fp16, [16]> var_1114_bias_0_to_fp16 = const()[name = string("op_1114_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_1114_cast_fp16 = linear(bias = var_1114_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_9_cast_fp16)[name = string("op_1114_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_5_cast_fp16 = exp(x = var_1114_cast_fp16)[name = string("cum_5_cast_fp16")];
            tensor<int32, [1]> var_1117_axes_0 = const()[name = string("op_1117_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_1117_cast_fp16 = expand_dims(axes = var_1117_axes_0, x = cum_5_cast_fp16)[name = string("op_1117_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1118_cast_fp16 = mul(x = kb_7_cast_fp16, y = var_1117_cast_fp16)[name = string("op_1118_cast_fp16")];
            bool w_7_transpose_x_0 = const()[name = string("w_7_transpose_x_0"), val = bool(false)];
            bool w_7_transpose_y_0 = const()[name = string("w_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_7_cast_fp16 = matmul(transpose_x = w_7_transpose_x_0, transpose_y = w_7_transpose_y_0, x = Minv_31_cast_fp16, y = var_1118_cast_fp16)[name = string("w_7_cast_fp16")];
            bool var_1120_transpose_x_0 = const()[name = string("op_1120_transpose_x_0"), val = bool(false)];
            bool var_1120_transpose_y_0 = const()[name = string("op_1120_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1120_cast_fp16 = matmul(transpose_x = var_1120_transpose_x_0, transpose_y = var_1120_transpose_y_0, x = w_7_cast_fp16, y = S_5_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_5_cast_fp16 = sub(x = u_7_cast_fp16, y = var_1120_cast_fp16)[name = string("vnew_5_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1125_cast_fp16 = mul(x = q_15_cast_fp16, y = var_1117_cast_fp16)[name = string("op_1125_cast_fp16")];
            bool var_1126_transpose_x_0 = const()[name = string("op_1126_transpose_x_0"), val = bool(false)];
            bool var_1126_transpose_y_0 = const()[name = string("op_1126_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1126_cast_fp16 = matmul(transpose_x = var_1126_transpose_x_0, transpose_y = var_1126_transpose_y_0, x = var_1125_cast_fp16, y = S_5_cast_fp16)[name = string("op_1126_cast_fp16")];
            bool var_1127_transpose_x_0 = const()[name = string("op_1127_transpose_x_0"), val = bool(false)];
            bool var_1127_transpose_y_0 = const()[name = string("op_1127_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1127_cast_fp16 = matmul(transpose_x = var_1127_transpose_x_0, transpose_y = var_1127_transpose_y_0, x = A_7_cast_fp16, y = vnew_5_cast_fp16)[name = string("op_1127_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_1_cast_fp16 = add(x = var_1126_cast_fp16, y = var_1127_cast_fp16)[name = string("oc_1_cast_fp16")];
            int32 var_1131 = const()[name = string("op_1131"), val = int32(2)];
            bool o_1_interleave_0 = const()[name = string("o_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 128]> o_1_cast_fp16 = concat(axis = var_1131, interleave = o_1_interleave_0, values = (var_674_cast_fp16, var_821_cast_fp16, var_975_cast_fp16, oc_1_cast_fp16))[name = string("o_1_cast_fp16")];
            tensor<int32, [4]> var_1137 = const()[name = string("op_1137"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1143 = const()[name = string("op_1143"), val = tensor<int32, [4]>([1, 64, 4, 128])];
            tensor<fp16, [1, 64, 4, 128]> input_7_cast_fp16 = reshape(shape = var_1143, x = z_1_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> o_3_cast_fp16 = transpose(perm = var_1137, x = o_1_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 64, 4, 128]> var_1145_cast_fp16 = mul(x = o_3_cast_fp16, y = o_3_cast_fp16)[name = string("op_1145_cast_fp16")];
            tensor<int32, [1]> var_1150_axes_0 = const()[name = string("op_1150_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1150_keep_dims_0 = const()[name = string("op_1150_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 4, 1]> var_1150_cast_fp16 = reduce_mean(axes = var_1150_axes_0, keep_dims = var_1150_keep_dims_0, x = var_1145_cast_fp16)[name = string("op_1150_cast_fp16")];
            fp16 var_1152_to_fp16 = const()[name = string("op_1152_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 64, 4, 1]> var_1153_cast_fp16 = add(x = var_1150_cast_fp16, y = var_1152_to_fp16)[name = string("op_1153_cast_fp16")];
            fp32 var_1154_epsilon_0 = const()[name = string("op_1154_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 4, 1]> var_1154_cast_fp16 = rsqrt(epsilon = var_1154_epsilon_0, x = var_1153_cast_fp16)[name = string("op_1154_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> var_1155_cast_fp16 = mul(x = o_3_cast_fp16, y = var_1154_cast_fp16)[name = string("op_1155_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> var_1157_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1157_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> o_5_cast_fp16 = mul(x = var_1155_cast_fp16, y = var_1157_cast_fp16)[name = string("o_5_cast_fp16")];
            tensor<int32, [3]> var_1162 = const()[name = string("op_1162"), val = tensor<int32, [3]>([1, 64, 512])];
            tensor<fp16, [1, 64, 512]> input_9_cast_fp16 = reshape(shape = var_1162, x = o_5_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [640, 512]> layers_0_out_weight_to_fp16 = const()[name = string("layers_0_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22499136)))];
            tensor<fp16, [640]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23154560)))];
            tensor<fp16, [1, 64, 640]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_out_weight_to_fp16, x = input_9_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [1, 64, 640]> x_5_cast_fp16 = add(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = linear_1_cast_fp16)[name = string("x_5_cast_fp16")];
            fp16 var_1175_to_fp16 = const()[name = string("op_1175_to_fp16"), val = fp16(0x1p+5)];
            tensor<fp16, [1, 64, 640]> x_7_cast_fp16 = mul(x = x_5_cast_fp16, y = var_1175_to_fp16)[name = string("x_7_cast_fp16")];
            fp16 var_1173_promoted_to_fp16 = const()[name = string("op_1173_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> var_1177_cast_fp16 = pow(x = x_7_cast_fp16, y = var_1173_promoted_to_fp16)[name = string("op_1177_cast_fp16")];
            tensor<int32, [1]> var_1179_axes_0 = const()[name = string("op_1179_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1179_keep_dims_0 = const()[name = string("op_1179_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_1179_cast_fp16 = reduce_mean(axes = var_1179_axes_0, keep_dims = var_1179_keep_dims_0, x = var_1177_cast_fp16)[name = string("op_1179_cast_fp16")];
            fp16 var_1180_to_fp16 = const()[name = string("op_1180_to_fp16"), val = fp16(0x1.0c8p-10)];
            tensor<fp16, [1, 64, 1]> var_1181_cast_fp16 = add(x = var_1179_cast_fp16, y = var_1180_to_fp16)[name = string("op_1181_cast_fp16")];
            fp32 var_1182_epsilon_0 = const()[name = string("op_1182_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1182_cast_fp16 = rsqrt(epsilon = var_1182_epsilon_0, x = var_1181_cast_fp16)[name = string("op_1182_cast_fp16")];
            tensor<fp16, [1, 64, 640]> var_1183_cast_fp16 = mul(x = x_7_cast_fp16, y = var_1182_cast_fp16)[name = string("op_1183_cast_fp16")];
            tensor<fp16, [640]> layers_0_n2_w1_to_fp16 = const()[name = string("layers_0_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23155904)))];
            tensor<fp16, [1, 64, 640]> input_11_cast_fp16 = mul(x = var_1183_cast_fp16, y = layers_0_n2_w1_to_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_0_gate_up_weight_to_fp16 = const()[name = string("layers_0_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23157248)))];
            tensor<fp16, [5120]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29710912)))];
            tensor<fp16, [1, 64, 5120]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_0_gate_up_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [3]> input_13_begin_0 = const()[name = string("input_13_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_13_end_0 = const()[name = string("input_13_end_0"), val = tensor<int32, [3]>([1, 64, 2560])];
            tensor<bool, [3]> input_13_end_mask_0 = const()[name = string("input_13_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 2560]> input_13_cast_fp16 = slice_by_index(begin = input_13_begin_0, end = input_13_end_0, end_mask = input_13_end_mask_0, x = linear_2_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 64, 2560]> var_1193_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_1193_cast_fp16")];
            tensor<int32, [3]> var_1198_begin_0 = const()[name = string("op_1198_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_1198_end_0 = const()[name = string("op_1198_end_0"), val = tensor<int32, [3]>([1, 64, 5120])];
            tensor<bool, [3]> var_1198_end_mask_0 = const()[name = string("op_1198_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 2560]> var_1198_cast_fp16 = slice_by_index(begin = var_1198_begin_0, end = var_1198_end_0, end_mask = var_1198_end_mask_0, x = linear_2_cast_fp16)[name = string("op_1198_cast_fp16")];
            tensor<fp16, [1, 64, 2560]> input_15_cast_fp16 = mul(x = var_1193_cast_fp16, y = var_1198_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_0_down_weight_to_fp16 = const()[name = string("layers_0_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29721216)))];
            tensor<fp16, [1, 64, 640]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_down_weight_to_fp16, x = input_15_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 64, 640]> x_9_cast_fp16 = add(x = x_5_cast_fp16, y = linear_3_cast_fp16)[name = string("x_9_cast_fp16")];
            fp16 var_1211_to_fp16 = const()[name = string("op_1211_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> x_11_cast_fp16 = mul(x = x_9_cast_fp16, y = var_1211_to_fp16)[name = string("x_11_cast_fp16")];
            fp16 var_1209_promoted_to_fp16 = const()[name = string("op_1209_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> var_1213_cast_fp16 = pow(x = x_11_cast_fp16, y = var_1209_promoted_to_fp16)[name = string("op_1213_cast_fp16")];
            tensor<int32, [1]> var_1215_axes_0 = const()[name = string("op_1215_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1215_keep_dims_0 = const()[name = string("op_1215_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_1215_cast_fp16 = reduce_mean(axes = var_1215_axes_0, keep_dims = var_1215_keep_dims_0, x = var_1213_cast_fp16)[name = string("op_1215_cast_fp16")];
            fp16 var_1216_to_fp16 = const()[name = string("op_1216_to_fp16"), val = fp16(0x1.0cp-18)];
            tensor<fp16, [1, 64, 1]> var_1217_cast_fp16 = add(x = var_1215_cast_fp16, y = var_1216_to_fp16)[name = string("op_1217_cast_fp16")];
            fp32 var_1218_epsilon_0 = const()[name = string("op_1218_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1218_cast_fp16 = rsqrt(epsilon = var_1218_epsilon_0, x = var_1217_cast_fp16)[name = string("op_1218_cast_fp16")];
            tensor<fp16, [1, 64, 640]> var_1219_cast_fp16 = mul(x = x_11_cast_fp16, y = var_1218_cast_fp16)[name = string("op_1219_cast_fp16")];
            tensor<fp16, [640]> layers_1_n1_w1_to_fp16 = const()[name = string("layers_1_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32998080)))];
            tensor<fp16, [1, 64, 640]> input_17_cast_fp16 = mul(x = var_1219_cast_fp16, y = layers_1_n1_w1_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_1_qkvzba_weight_to_fp16 = const()[name = string("layers_1_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32999424)))];
            tensor<fp16, [1, 64, 1544]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_qkvzba_weight_to_fp16, x = input_17_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<int32, [3]> qkv_3_begin_0 = const()[name = string("qkv_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_3_end_0 = const()[name = string("qkv_3_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> qkv_3_end_mask_0 = const()[name = string("qkv_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 1024]> qkv_3_cast_fp16 = slice_by_index(begin = qkv_3_begin_0, end = qkv_3_end_0, end_mask = qkv_3_end_mask_0, x = linear_4_cast_fp16)[name = string("qkv_3_cast_fp16")];
            tensor<int32, [3]> z_3_begin_0 = const()[name = string("z_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_3_end_0 = const()[name = string("z_3_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
            tensor<bool, [3]> z_3_end_mask_0 = const()[name = string("z_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 512]> z_3_cast_fp16 = slice_by_index(begin = z_3_begin_0, end = z_3_end_0, end_mask = z_3_end_mask_0, x = linear_4_cast_fp16)[name = string("z_3_cast_fp16")];
            tensor<int32, [3]> b_3_begin_0 = const()[name = string("b_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_3_end_0 = const()[name = string("b_3_end_0"), val = tensor<int32, [3]>([1, 64, 1540])];
            tensor<bool, [3]> b_3_end_mask_0 = const()[name = string("b_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 4]> b_3_cast_fp16 = slice_by_index(begin = b_3_begin_0, end = b_3_end_0, end_mask = b_3_end_mask_0, x = linear_4_cast_fp16)[name = string("b_3_cast_fp16")];
            tensor<int32, [3]> a_3_begin_0 = const()[name = string("a_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_3_end_0 = const()[name = string("a_3_end_0"), val = tensor<int32, [3]>([1, 64, 1])];
            tensor<bool, [3]> a_3_end_mask_0 = const()[name = string("a_3_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 4]> a_3_cast_fp16 = slice_by_index(begin = a_3_begin_0, end = a_3_end_0, end_mask = a_3_end_mask_0, x = linear_4_cast_fp16)[name = string("a_3_cast_fp16")];
            int32 var_1258 = const()[name = string("op_1258"), val = int32(1)];
            bool xp_3_interleave_0 = const()[name = string("xp_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 67, 1024]> xp_3_cast_fp16 = concat(axis = var_1258, interleave = xp_3_interleave_0, values = (zpad_to_fp16, qkv_3_cast_fp16))[name = string("xp_3_cast_fp16")];
            tensor<int32, [3]> var_1269_begin_0 = const()[name = string("op_1269_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_1269_end_0 = const()[name = string("op_1269_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_1269_end_mask_0 = const()[name = string("op_1269_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 1024]> var_1269_cast_fp16 = slice_by_index(begin = var_1269_begin_0, end = var_1269_end_0, end_mask = var_1269_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1269_cast_fp16")];
            tensor<fp16, [1024]> var_1277_to_fp16 = const()[name = string("op_1277_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34975808)))];
            tensor<fp16, [1, 64, 1024]> var_1278_cast_fp16 = mul(x = var_1269_cast_fp16, y = var_1277_to_fp16)[name = string("op_1278_cast_fp16")];
            tensor<int32, [3]> var_1288_begin_0 = const()[name = string("op_1288_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_1288_end_0 = const()[name = string("op_1288_end_0"), val = tensor<int32, [3]>([1, 66, 1024])];
            tensor<bool, [3]> var_1288_end_mask_0 = const()[name = string("op_1288_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_1288_cast_fp16 = slice_by_index(begin = var_1288_begin_0, end = var_1288_end_0, end_mask = var_1288_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1288_cast_fp16")];
            tensor<fp16, [1024]> var_1296_to_fp16 = const()[name = string("op_1296_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34977920)))];
            tensor<fp16, [1, 64, 1024]> var_1297_cast_fp16 = mul(x = var_1288_cast_fp16, y = var_1296_to_fp16)[name = string("op_1297_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> var_1299_cast_fp16 = add(x = var_1278_cast_fp16, y = var_1297_cast_fp16)[name = string("op_1299_cast_fp16")];
            tensor<int32, [3]> var_1309_begin_0 = const()[name = string("op_1309_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1309_end_0 = const()[name = string("op_1309_end_0"), val = tensor<int32, [3]>([1, 65, 1024])];
            tensor<bool, [3]> var_1309_end_mask_0 = const()[name = string("op_1309_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_1309_cast_fp16 = slice_by_index(begin = var_1309_begin_0, end = var_1309_end_0, end_mask = var_1309_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1309_cast_fp16")];
            tensor<fp16, [1024]> var_1317_to_fp16 = const()[name = string("op_1317_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34980032)))];
            tensor<fp16, [1, 64, 1024]> var_1318_cast_fp16 = mul(x = var_1309_cast_fp16, y = var_1317_to_fp16)[name = string("op_1318_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> var_1320_cast_fp16 = add(x = var_1299_cast_fp16, y = var_1318_cast_fp16)[name = string("op_1320_cast_fp16")];
            tensor<int32, [3]> var_1330_begin_0 = const()[name = string("op_1330_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1330_end_0 = const()[name = string("op_1330_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> var_1330_end_mask_0 = const()[name = string("op_1330_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_1330_cast_fp16 = slice_by_index(begin = var_1330_begin_0, end = var_1330_end_0, end_mask = var_1330_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1330_cast_fp16")];
            tensor<fp16, [1024]> var_1338_to_fp16 = const()[name = string("op_1338_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34982144)))];
            tensor<fp16, [1, 64, 1024]> var_1339_cast_fp16 = mul(x = var_1330_cast_fp16, y = var_1338_to_fp16)[name = string("op_1339_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> input_19_cast_fp16 = add(x = var_1320_cast_fp16, y = var_1339_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> c_3_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("c_3_cast_fp16")];
            tensor<int32, [3]> var_1347_begin_0 = const()[name = string("op_1347_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1347_end_0 = const()[name = string("op_1347_end_0"), val = tensor<int32, [3]>([1, 64, 256])];
            tensor<bool, [3]> var_1347_end_mask_0 = const()[name = string("op_1347_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 256]> var_1347_cast_fp16 = slice_by_index(begin = var_1347_begin_0, end = var_1347_end_0, end_mask = var_1347_end_mask_0, x = c_3_cast_fp16)[name = string("op_1347_cast_fp16")];
            tensor<int32, [4]> var_1352 = const()[name = string("op_1352"), val = tensor<int32, [4]>([1, 64, 2, 128])];
            tensor<fp16, [1, 64, 2, 128]> q_17_cast_fp16 = reshape(shape = var_1352, x = var_1347_cast_fp16)[name = string("q_17_cast_fp16")];
            tensor<int32, [3]> var_1358_begin_0 = const()[name = string("op_1358_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_1358_end_0 = const()[name = string("op_1358_end_0"), val = tensor<int32, [3]>([1, 64, 512])];
            tensor<bool, [3]> var_1358_end_mask_0 = const()[name = string("op_1358_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 256]> var_1358_cast_fp16 = slice_by_index(begin = var_1358_begin_0, end = var_1358_end_0, end_mask = var_1358_end_mask_0, x = c_3_cast_fp16)[name = string("op_1358_cast_fp16")];
            tensor<int32, [4]> var_1363 = const()[name = string("op_1363"), val = tensor<int32, [4]>([1, 64, 2, 128])];
            tensor<fp16, [1, 64, 2, 128]> k_17_cast_fp16 = reshape(shape = var_1363, x = var_1358_cast_fp16)[name = string("k_17_cast_fp16")];
            tensor<int32, [3]> var_1369_begin_0 = const()[name = string("op_1369_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_1369_end_0 = const()[name = string("op_1369_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> var_1369_end_mask_0 = const()[name = string("op_1369_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 512]> var_1369_cast_fp16 = slice_by_index(begin = var_1369_begin_0, end = var_1369_end_0, end_mask = var_1369_end_mask_0, x = c_3_cast_fp16)[name = string("op_1369_cast_fp16")];
            tensor<int32, [4]> var_1374 = const()[name = string("op_1374"), val = tensor<int32, [4]>([1, 64, 4, 128])];
            tensor<fp16, [1, 64, 4, 128]> var_1375_cast_fp16 = reshape(shape = var_1374, x = var_1369_cast_fp16)[name = string("op_1375_cast_fp16")];
            tensor<int32, [4]> var_1380 = const()[name = string("op_1380"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1382_to_fp16 = const()[name = string("op_1382_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 64, 2, 128]> q_19_cast_fp16 = mul(x = q_17_cast_fp16, y = var_1382_to_fp16)[name = string("q_19_cast_fp16")];
            fp16 var_1384_to_fp16 = const()[name = string("op_1384_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 64, 2, 128]> k_19_cast_fp16 = mul(x = k_17_cast_fp16, y = var_1384_to_fp16)[name = string("k_19_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_1386_cast_fp16 = mul(x = q_19_cast_fp16, y = q_19_cast_fp16)[name = string("op_1386_cast_fp16")];
            tensor<int32, [1]> var_1391_axes_0 = const()[name = string("op_1391_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1391_keep_dims_0 = const()[name = string("op_1391_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 2, 1]> var_1391_cast_fp16 = reduce_mean(axes = var_1391_axes_0, keep_dims = var_1391_keep_dims_0, x = var_1386_cast_fp16)[name = string("op_1391_cast_fp16")];
            fp16 var_1392_promoted_to_fp16 = const()[name = string("op_1392_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 64, 2, 1]> var_1393_cast_fp16 = mul(x = var_1391_cast_fp16, y = var_1392_promoted_to_fp16)[name = string("op_1393_cast_fp16")];
            fp16 var_1395_to_fp16 = const()[name = string("op_1395_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 64, 2, 1]> var_1396_cast_fp16 = add(x = var_1393_cast_fp16, y = var_1395_to_fp16)[name = string("op_1396_cast_fp16")];
            fp32 var_1397_epsilon_0 = const()[name = string("op_1397_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 2, 1]> var_1397_cast_fp16 = rsqrt(epsilon = var_1397_epsilon_0, x = var_1396_cast_fp16)[name = string("op_1397_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_1398_cast_fp16 = mul(x = q_19_cast_fp16, y = var_1397_cast_fp16)[name = string("op_1398_cast_fp16")];
            fp16 var_1399_to_fp16 = const()[name = string("op_1399_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 64, 2, 128]> q_21_cast_fp16 = mul(x = var_1398_cast_fp16, y = var_1399_to_fp16)[name = string("q_21_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_1401_cast_fp16 = mul(x = k_19_cast_fp16, y = k_19_cast_fp16)[name = string("op_1401_cast_fp16")];
            tensor<int32, [1]> var_1406_axes_0 = const()[name = string("op_1406_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1406_keep_dims_0 = const()[name = string("op_1406_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 2, 1]> var_1406_cast_fp16 = reduce_mean(axes = var_1406_axes_0, keep_dims = var_1406_keep_dims_0, x = var_1401_cast_fp16)[name = string("op_1406_cast_fp16")];
            fp16 var_1407_promoted_to_fp16 = const()[name = string("op_1407_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 64, 2, 1]> var_1408_cast_fp16 = mul(x = var_1406_cast_fp16, y = var_1407_promoted_to_fp16)[name = string("op_1408_cast_fp16")];
            fp16 var_1410_to_fp16 = const()[name = string("op_1410_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 64, 2, 1]> var_1411_cast_fp16 = add(x = var_1408_cast_fp16, y = var_1410_to_fp16)[name = string("op_1411_cast_fp16")];
            fp32 var_1412_epsilon_0 = const()[name = string("op_1412_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 2, 1]> var_1412_cast_fp16 = rsqrt(epsilon = var_1412_epsilon_0, x = var_1411_cast_fp16)[name = string("op_1412_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> k_21_cast_fp16 = mul(x = k_19_cast_fp16, y = var_1412_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<int32, [4]> var_1418 = const()[name = string("op_1418"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1431_axes_0 = const()[name = string("op_1431_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 128]> var_1419_cast_fp16 = transpose(perm = var_1418, x = q_21_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 2, 1, 64, 128]> var_1431_cast_fp16 = expand_dims(axes = var_1431_axes_0, x = var_1419_cast_fp16)[name = string("op_1431_cast_fp16")];
            tensor<int32, [5]> var_1439_reps_0 = const()[name = string("op_1439_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 64, 128]> var_1439_cast_fp16 = tile(reps = var_1439_reps_0, x = var_1431_cast_fp16)[name = string("op_1439_cast_fp16")];
            tensor<int32, [4]> var_1444 = const()[name = string("op_1444"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> q_23_cast_fp16 = reshape(shape = var_1444, x = var_1439_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<int32, [4]> var_1450 = const()[name = string("op_1450"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1463_axes_0 = const()[name = string("op_1463_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 128]> var_1451_cast_fp16 = transpose(perm = var_1450, x = k_21_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 2, 1, 64, 128]> var_1463_cast_fp16 = expand_dims(axes = var_1463_axes_0, x = var_1451_cast_fp16)[name = string("op_1463_cast_fp16")];
            tensor<int32, [5]> var_1471_reps_0 = const()[name = string("op_1471_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 64, 128]> var_1471_cast_fp16 = tile(reps = var_1471_reps_0, x = var_1463_cast_fp16)[name = string("op_1471_cast_fp16")];
            tensor<int32, [4]> var_1476 = const()[name = string("op_1476"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> k_23_cast_fp16 = reshape(shape = var_1476, x = var_1471_cast_fp16)[name = string("k_23_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_1478_cast_fp16 = sigmoid(x = b_3_cast_fp16)[name = string("op_1478_cast_fp16")];
            tensor<int32, [3]> var_1482 = const()[name = string("op_1482"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_1_dt_bias_to_fp16 = const()[name = string("layers_1_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 64, 4]> input_21_cast_fp16 = add(x = a_3_cast_fp16, y = layers_1_dt_bias_to_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_1486_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("op_1486_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_1487_cast_fp16 = abs(x = input_21_cast_fp16)[name = string("op_1487_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 64, 4]> var_1488_cast_fp16 = mul(x = var_1487_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_1488_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_1489_cast_fp16 = exp(x = var_1488_cast_fp16)[name = string("op_1489_cast_fp16")];
            fp16 var_1491_to_fp16 = const()[name = string("op_1491_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 64, 4]> var_1492_cast_fp16 = add(x = var_1489_cast_fp16, y = var_1491_to_fp16)[name = string("op_1492_cast_fp16")];
            fp32 var_1493_epsilon_0 = const()[name = string("op_1493_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 64, 4]> var_1493_cast_fp16 = log(epsilon = var_1493_epsilon_0, x = var_1492_cast_fp16)[name = string("op_1493_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_1495_cast_fp16 = add(x = var_1486_cast_fp16, y = var_1493_cast_fp16)[name = string("op_1495_cast_fp16")];
            tensor<fp16, [4]> layers_1_negA_to_fp16 = const()[name = string("layers_1_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.ff4p+2, -0x1.8cp+3, -0x1.c0cp+3, -0x1.104p+3])];
            tensor<fp16, [1, 64, 4]> var_1496_cast_fp16 = mul(x = layers_1_negA_to_fp16, y = var_1495_cast_fp16)[name = string("op_1496_cast_fp16")];
            tensor<int32, [3]> var_1500 = const()[name = string("op_1500"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_1503_axes_0 = const()[name = string("op_1503_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 64]> beta_3_cast_fp16 = transpose(perm = var_1482, x = var_1478_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 4, 64, 1]> var_1503_cast_fp16 = expand_dims(axes = var_1503_axes_0, x = beta_3_cast_fp16)[name = string("op_1503_cast_fp16")];
            fp16 var_1504_to_fp16 = const()[name = string("op_1504_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 64, 1]> var_1505_cast_fp16 = mul(x = var_1503_cast_fp16, y = var_1504_to_fp16)[name = string("op_1505_cast_fp16")];
            tensor<fp16, [1, 4, 64, 128]> v_3_cast_fp16 = transpose(perm = var_1380, x = var_1375_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 4, 64, 128]> vb_11_cast_fp16 = mul(x = v_3_cast_fp16, y = var_1505_cast_fp16)[name = string("vb_11_cast_fp16")];
            tensor<int32, [4]> q_25_begin_0 = const()[name = string("q_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_25_end_0 = const()[name = string("q_25_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_25_end_mask_0 = const()[name = string("q_25_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, x = q_23_cast_fp16)[name = string("q_25_cast_fp16")];
            tensor<int32, [4]> k_25_begin_0 = const()[name = string("k_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_25_end_0 = const()[name = string("k_25_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_25_end_mask_0 = const()[name = string("k_25_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, x = k_23_cast_fp16)[name = string("k_25_cast_fp16")];
            tensor<fp16, [1, 4, 64, 128]> var_1539_cast_fp16 = mul(x = k_23_cast_fp16, y = var_1503_cast_fp16)[name = string("op_1539_cast_fp16")];
            tensor<int32, [4]> kb_9_begin_0 = const()[name = string("kb_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_9_end_0 = const()[name = string("kb_9_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_9_end_mask_0 = const()[name = string("kb_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_9_cast_fp16 = slice_by_index(begin = kb_9_begin_0, end = kb_9_end_0, end_mask = kb_9_end_mask_0, x = var_1539_cast_fp16)[name = string("kb_9_cast_fp16")];
            tensor<int32, [4]> vb_13_begin_0 = const()[name = string("vb_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_13_end_0 = const()[name = string("vb_13_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_13_end_mask_0 = const()[name = string("vb_13_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_13_cast_fp16 = slice_by_index(begin = vb_13_begin_0, end = vb_13_end_0, end_mask = vb_13_end_mask_0, x = vb_11_cast_fp16)[name = string("vb_13_cast_fp16")];
            tensor<int32, [3]> gdec_13_begin_0 = const()[name = string("gdec_13_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_13_end_0 = const()[name = string("gdec_13_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_13_end_mask_0 = const()[name = string("gdec_13_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 64]> gdec_11_cast_fp16 = transpose(perm = var_1500, x = var_1496_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [1, 4, 16]> gdec_13_cast_fp16 = slice_by_index(begin = gdec_13_begin_0, end = gdec_13_end_0, end_mask = gdec_13_end_mask_0, x = gdec_11_cast_fp16)[name = string("gdec_13_cast_fp16")];
            tensor<int32, [2]> var_1587 = const()[name = string("op_1587"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_1588_cast_fp16 = reshape(shape = var_1587, x = gdec_13_cast_fp16)[name = string("op_1588_cast_fp16")];
            tensor<fp16, [4, 256]> var_1589_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1588_cast_fp16)[name = string("op_1589_cast_fp16")];
            tensor<int32, [4]> var_1594 = const()[name = string("op_1594"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_9_cast_fp16 = reshape(shape = var_1594, x = var_1589_cast_fp16)[name = string("P_9_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1596_cast_fp16 = exp(x = P_9_cast_fp16)[name = string("op_1596_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_9_cast_fp16 = mul(x = var_1596_cast_fp16, y = tril_incl_to_fp16)[name = string("D_9_cast_fp16")];
            bool kk_9_transpose_x_1 = const()[name = string("kk_9_transpose_x_1"), val = bool(false)];
            bool kk_9_transpose_y_1 = const()[name = string("kk_9_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_9_cast_fp16 = matmul(transpose_x = kk_9_transpose_x_1, transpose_y = kk_9_transpose_y_1, x = kb_9_cast_fp16, y = k_25_cast_fp16)[name = string("kk_9_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_1602_cast_fp16 = mul(x = kk_9_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_1602_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1603_cast_fp16 = mul(x = var_1602_cast_fp16, y = D_9_cast_fp16)[name = string("op_1603_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_9_cast_fp16 = mul(x = var_1603_cast_fp16, y = cum_tail_to_fp16)[name = string("N_9_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1605_cast_fp16 = mul(x = D_9_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1605_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_9_cast_fp16 = mul(x = kk_9_cast_fp16, y = var_1605_cast_fp16)[name = string("Na_9_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_33_cast_fp16 = add(x = eye_to_fp16, y = N_9_cast_fp16)[name = string("Minv_33_cast_fp16")];
            bool Np_25_transpose_x_0 = const()[name = string("Np_25_transpose_x_0"), val = bool(false)];
            bool Np_25_transpose_y_0 = const()[name = string("Np_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_25_cast_fp16 = matmul(transpose_x = Np_25_transpose_x_0, transpose_y = Np_25_transpose_y_0, x = N_9_cast_fp16, y = Na_9_cast_fp16)[name = string("Np_25_cast_fp16")];
            bool Npa_17_transpose_x_0 = const()[name = string("Npa_17_transpose_x_0"), val = bool(false)];
            bool Npa_17_transpose_y_0 = const()[name = string("Npa_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_17_cast_fp16 = matmul(transpose_x = Npa_17_transpose_x_0, transpose_y = Npa_17_transpose_y_0, x = Na_9_cast_fp16, y = N_9_cast_fp16)[name = string("Npa_17_cast_fp16")];
            bool var_1611_transpose_x_0 = const()[name = string("op_1611_transpose_x_0"), val = bool(false)];
            bool var_1611_transpose_y_0 = const()[name = string("op_1611_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1611_cast_fp16 = matmul(transpose_x = var_1611_transpose_x_0, transpose_y = var_1611_transpose_y_0, x = Minv_33_cast_fp16, y = Np_25_cast_fp16)[name = string("op_1611_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_35_cast_fp16 = add(x = Minv_33_cast_fp16, y = var_1611_cast_fp16)[name = string("Minv_35_cast_fp16")];
            bool Np_27_transpose_x_0 = const()[name = string("Np_27_transpose_x_0"), val = bool(false)];
            bool Np_27_transpose_y_0 = const()[name = string("Np_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_27_cast_fp16 = matmul(transpose_x = Np_27_transpose_x_0, transpose_y = Np_27_transpose_y_0, x = Np_25_cast_fp16, y = Npa_17_cast_fp16)[name = string("Np_27_cast_fp16")];
            bool Npa_19_transpose_x_0 = const()[name = string("Npa_19_transpose_x_0"), val = bool(false)];
            bool Npa_19_transpose_y_0 = const()[name = string("Npa_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_19_cast_fp16 = matmul(transpose_x = Npa_19_transpose_x_0, transpose_y = Npa_19_transpose_y_0, x = Npa_17_cast_fp16, y = Np_25_cast_fp16)[name = string("Npa_19_cast_fp16")];
            bool var_1616_transpose_x_0 = const()[name = string("op_1616_transpose_x_0"), val = bool(false)];
            bool var_1616_transpose_y_0 = const()[name = string("op_1616_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1616_cast_fp16 = matmul(transpose_x = var_1616_transpose_x_0, transpose_y = var_1616_transpose_y_0, x = Minv_35_cast_fp16, y = Np_27_cast_fp16)[name = string("op_1616_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_37_cast_fp16 = add(x = Minv_35_cast_fp16, y = var_1616_cast_fp16)[name = string("Minv_37_cast_fp16")];
            bool Np_29_transpose_x_0 = const()[name = string("Np_29_transpose_x_0"), val = bool(false)];
            bool Np_29_transpose_y_0 = const()[name = string("Np_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_29_cast_fp16 = matmul(transpose_x = Np_29_transpose_x_0, transpose_y = Np_29_transpose_y_0, x = Np_27_cast_fp16, y = Npa_19_cast_fp16)[name = string("Np_29_cast_fp16")];
            bool var_1620_transpose_x_0 = const()[name = string("op_1620_transpose_x_0"), val = bool(false)];
            bool var_1620_transpose_y_0 = const()[name = string("op_1620_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1620_cast_fp16 = matmul(transpose_x = var_1620_transpose_x_0, transpose_y = var_1620_transpose_y_0, x = Minv_37_cast_fp16, y = Np_29_cast_fp16)[name = string("op_1620_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_39_cast_fp16 = add(x = Minv_37_cast_fp16, y = var_1620_cast_fp16)[name = string("Minv_39_cast_fp16")];
            bool u_9_transpose_x_0 = const()[name = string("u_9_transpose_x_0"), val = bool(false)];
            bool u_9_transpose_y_0 = const()[name = string("u_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_9_cast_fp16 = matmul(transpose_x = u_9_transpose_x_0, transpose_y = u_9_transpose_y_0, x = Minv_39_cast_fp16, y = vb_13_cast_fp16)[name = string("u_9_cast_fp16")];
            bool var_1627_transpose_x_1 = const()[name = string("op_1627_transpose_x_1"), val = bool(false)];
            bool var_1627_transpose_y_1 = const()[name = string("op_1627_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_1627_cast_fp16 = matmul(transpose_x = var_1627_transpose_x_1, transpose_y = var_1627_transpose_y_1, x = q_25_cast_fp16, y = k_25_cast_fp16)[name = string("op_1627_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_9_cast_fp16 = mul(x = var_1627_cast_fp16, y = D_9_cast_fp16)[name = string("A_9_cast_fp16")];
            bool var_1629_transpose_x_0 = const()[name = string("op_1629_transpose_x_0"), val = bool(false)];
            bool var_1629_transpose_y_0 = const()[name = string("op_1629_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1629_cast_fp16 = matmul(transpose_x = var_1629_transpose_x_0, transpose_y = var_1629_transpose_y_0, x = A_9_cast_fp16, y = u_9_cast_fp16)[name = string("op_1629_cast_fp16")];
            tensor<fp16, [16]> var_1630_bias_0_to_fp16 = const()[name = string("op_1630_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_1630_cast_fp16 = linear(bias = var_1630_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_13_cast_fp16)[name = string("op_1630_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_7_cast_fp16 = exp(x = var_1630_cast_fp16)[name = string("tail_7_cast_fp16")];
            tensor<int32, [1]> var_1633_axes_0 = const()[name = string("op_1633_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_1633_cast_fp16 = expand_dims(axes = var_1633_axes_0, x = tail_7_cast_fp16)[name = string("op_1633_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1634_cast_fp16 = mul(x = k_25_cast_fp16, y = var_1633_cast_fp16)[name = string("op_1634_cast_fp16")];
            bool S_7_transpose_x_1 = const()[name = string("S_7_transpose_x_1"), val = bool(true)];
            bool S_7_transpose_y_1 = const()[name = string("S_7_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_7_cast_fp16 = matmul(transpose_x = S_7_transpose_x_1, transpose_y = S_7_transpose_y_1, x = var_1634_cast_fp16, y = u_9_cast_fp16)[name = string("S_7_cast_fp16")];
            tensor<int32, [4]> q_27_begin_0 = const()[name = string("q_27_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_27_end_0 = const()[name = string("q_27_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> q_27_end_mask_0 = const()[name = string("q_27_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_27_cast_fp16 = slice_by_index(begin = q_27_begin_0, end = q_27_end_0, end_mask = q_27_end_mask_0, x = q_23_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<int32, [4]> k_27_begin_0 = const()[name = string("k_27_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_27_end_0 = const()[name = string("k_27_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> k_27_end_mask_0 = const()[name = string("k_27_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_27_cast_fp16 = slice_by_index(begin = k_27_begin_0, end = k_27_end_0, end_mask = k_27_end_mask_0, x = k_23_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<int32, [4]> kb_11_begin_0 = const()[name = string("kb_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_11_end_0 = const()[name = string("kb_11_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> kb_11_end_mask_0 = const()[name = string("kb_11_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_11_cast_fp16 = slice_by_index(begin = kb_11_begin_0, end = kb_11_end_0, end_mask = kb_11_end_mask_0, x = var_1539_cast_fp16)[name = string("kb_11_cast_fp16")];
            tensor<int32, [4]> vb_15_begin_0 = const()[name = string("vb_15_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_15_end_0 = const()[name = string("vb_15_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> vb_15_end_mask_0 = const()[name = string("vb_15_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_15_cast_fp16 = slice_by_index(begin = vb_15_begin_0, end = vb_15_end_0, end_mask = vb_15_end_mask_0, x = vb_11_cast_fp16)[name = string("vb_15_cast_fp16")];
            tensor<int32, [3]> gdec_15_begin_0 = const()[name = string("gdec_15_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_15_end_0 = const()[name = string("gdec_15_end_0"), val = tensor<int32, [3]>([1, 4, 32])];
            tensor<bool, [3]> gdec_15_end_mask_0 = const()[name = string("gdec_15_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_15_cast_fp16 = slice_by_index(begin = gdec_15_begin_0, end = gdec_15_end_0, end_mask = gdec_15_end_mask_0, x = gdec_11_cast_fp16)[name = string("gdec_15_cast_fp16")];
            tensor<int32, [2]> var_1719 = const()[name = string("op_1719"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_1720_cast_fp16 = reshape(shape = var_1719, x = gdec_15_cast_fp16)[name = string("op_1720_cast_fp16")];
            tensor<fp16, [4, 256]> var_1721_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1720_cast_fp16)[name = string("op_1721_cast_fp16")];
            tensor<int32, [4]> var_1726 = const()[name = string("op_1726"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_11_cast_fp16 = reshape(shape = var_1726, x = var_1721_cast_fp16)[name = string("P_11_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1728_cast_fp16 = exp(x = P_11_cast_fp16)[name = string("op_1728_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_11_cast_fp16 = mul(x = var_1728_cast_fp16, y = tril_incl_to_fp16)[name = string("D_11_cast_fp16")];
            bool kk_11_transpose_x_1 = const()[name = string("kk_11_transpose_x_1"), val = bool(false)];
            bool kk_11_transpose_y_1 = const()[name = string("kk_11_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_11_cast_fp16 = matmul(transpose_x = kk_11_transpose_x_1, transpose_y = kk_11_transpose_y_1, x = kb_11_cast_fp16, y = k_27_cast_fp16)[name = string("kk_11_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_1734_cast_fp16 = mul(x = kk_11_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_1734_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1735_cast_fp16 = mul(x = var_1734_cast_fp16, y = D_11_cast_fp16)[name = string("op_1735_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_11_cast_fp16 = mul(x = var_1735_cast_fp16, y = cum_tail_to_fp16)[name = string("N_11_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1737_cast_fp16 = mul(x = D_11_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1737_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_11_cast_fp16 = mul(x = kk_11_cast_fp16, y = var_1737_cast_fp16)[name = string("Na_11_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_41_cast_fp16 = add(x = eye_to_fp16, y = N_11_cast_fp16)[name = string("Minv_41_cast_fp16")];
            bool Np_31_transpose_x_0 = const()[name = string("Np_31_transpose_x_0"), val = bool(false)];
            bool Np_31_transpose_y_0 = const()[name = string("Np_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_31_cast_fp16 = matmul(transpose_x = Np_31_transpose_x_0, transpose_y = Np_31_transpose_y_0, x = N_11_cast_fp16, y = Na_11_cast_fp16)[name = string("Np_31_cast_fp16")];
            bool Npa_21_transpose_x_0 = const()[name = string("Npa_21_transpose_x_0"), val = bool(false)];
            bool Npa_21_transpose_y_0 = const()[name = string("Npa_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_21_cast_fp16 = matmul(transpose_x = Npa_21_transpose_x_0, transpose_y = Npa_21_transpose_y_0, x = Na_11_cast_fp16, y = N_11_cast_fp16)[name = string("Npa_21_cast_fp16")];
            bool var_1743_transpose_x_0 = const()[name = string("op_1743_transpose_x_0"), val = bool(false)];
            bool var_1743_transpose_y_0 = const()[name = string("op_1743_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1743_cast_fp16 = matmul(transpose_x = var_1743_transpose_x_0, transpose_y = var_1743_transpose_y_0, x = Minv_41_cast_fp16, y = Np_31_cast_fp16)[name = string("op_1743_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_43_cast_fp16 = add(x = Minv_41_cast_fp16, y = var_1743_cast_fp16)[name = string("Minv_43_cast_fp16")];
            bool Np_33_transpose_x_0 = const()[name = string("Np_33_transpose_x_0"), val = bool(false)];
            bool Np_33_transpose_y_0 = const()[name = string("Np_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_33_cast_fp16 = matmul(transpose_x = Np_33_transpose_x_0, transpose_y = Np_33_transpose_y_0, x = Np_31_cast_fp16, y = Npa_21_cast_fp16)[name = string("Np_33_cast_fp16")];
            bool Npa_23_transpose_x_0 = const()[name = string("Npa_23_transpose_x_0"), val = bool(false)];
            bool Npa_23_transpose_y_0 = const()[name = string("Npa_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_23_cast_fp16 = matmul(transpose_x = Npa_23_transpose_x_0, transpose_y = Npa_23_transpose_y_0, x = Npa_21_cast_fp16, y = Np_31_cast_fp16)[name = string("Npa_23_cast_fp16")];
            bool var_1748_transpose_x_0 = const()[name = string("op_1748_transpose_x_0"), val = bool(false)];
            bool var_1748_transpose_y_0 = const()[name = string("op_1748_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1748_cast_fp16 = matmul(transpose_x = var_1748_transpose_x_0, transpose_y = var_1748_transpose_y_0, x = Minv_43_cast_fp16, y = Np_33_cast_fp16)[name = string("op_1748_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_45_cast_fp16 = add(x = Minv_43_cast_fp16, y = var_1748_cast_fp16)[name = string("Minv_45_cast_fp16")];
            bool Np_35_transpose_x_0 = const()[name = string("Np_35_transpose_x_0"), val = bool(false)];
            bool Np_35_transpose_y_0 = const()[name = string("Np_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_35_cast_fp16 = matmul(transpose_x = Np_35_transpose_x_0, transpose_y = Np_35_transpose_y_0, x = Np_33_cast_fp16, y = Npa_23_cast_fp16)[name = string("Np_35_cast_fp16")];
            bool var_1752_transpose_x_0 = const()[name = string("op_1752_transpose_x_0"), val = bool(false)];
            bool var_1752_transpose_y_0 = const()[name = string("op_1752_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1752_cast_fp16 = matmul(transpose_x = var_1752_transpose_x_0, transpose_y = var_1752_transpose_y_0, x = Minv_45_cast_fp16, y = Np_35_cast_fp16)[name = string("op_1752_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_47_cast_fp16 = add(x = Minv_45_cast_fp16, y = var_1752_cast_fp16)[name = string("Minv_47_cast_fp16")];
            bool u_11_transpose_x_0 = const()[name = string("u_11_transpose_x_0"), val = bool(false)];
            bool u_11_transpose_y_0 = const()[name = string("u_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_11_cast_fp16 = matmul(transpose_x = u_11_transpose_x_0, transpose_y = u_11_transpose_y_0, x = Minv_47_cast_fp16, y = vb_15_cast_fp16)[name = string("u_11_cast_fp16")];
            bool var_1759_transpose_x_1 = const()[name = string("op_1759_transpose_x_1"), val = bool(false)];
            bool var_1759_transpose_y_1 = const()[name = string("op_1759_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_1759_cast_fp16 = matmul(transpose_x = var_1759_transpose_x_1, transpose_y = var_1759_transpose_y_1, x = q_27_cast_fp16, y = k_27_cast_fp16)[name = string("op_1759_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_11_cast_fp16 = mul(x = var_1759_cast_fp16, y = D_11_cast_fp16)[name = string("A_11_cast_fp16")];
            tensor<fp16, [16]> var_1761_bias_0_to_fp16 = const()[name = string("op_1761_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_1761_cast_fp16 = linear(bias = var_1761_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_15_cast_fp16)[name = string("op_1761_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_7_cast_fp16 = exp(x = var_1761_cast_fp16)[name = string("cum_7_cast_fp16")];
            tensor<int32, [1]> var_1764_axes_0 = const()[name = string("op_1764_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_1764_cast_fp16 = expand_dims(axes = var_1764_axes_0, x = cum_7_cast_fp16)[name = string("op_1764_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1765_cast_fp16 = mul(x = kb_11_cast_fp16, y = var_1764_cast_fp16)[name = string("op_1765_cast_fp16")];
            bool w_11_transpose_x_0 = const()[name = string("w_11_transpose_x_0"), val = bool(false)];
            bool w_11_transpose_y_0 = const()[name = string("w_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_11_cast_fp16 = matmul(transpose_x = w_11_transpose_x_0, transpose_y = w_11_transpose_y_0, x = Minv_47_cast_fp16, y = var_1765_cast_fp16)[name = string("w_11_cast_fp16")];
            bool var_1767_transpose_x_0 = const()[name = string("op_1767_transpose_x_0"), val = bool(false)];
            bool var_1767_transpose_y_0 = const()[name = string("op_1767_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1767_cast_fp16 = matmul(transpose_x = var_1767_transpose_x_0, transpose_y = var_1767_transpose_y_0, x = w_11_cast_fp16, y = S_7_cast_fp16)[name = string("op_1767_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_7_cast_fp16 = sub(x = u_11_cast_fp16, y = var_1767_cast_fp16)[name = string("vnew_7_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1772_cast_fp16 = mul(x = q_27_cast_fp16, y = var_1764_cast_fp16)[name = string("op_1772_cast_fp16")];
            bool var_1773_transpose_x_0 = const()[name = string("op_1773_transpose_x_0"), val = bool(false)];
            bool var_1773_transpose_y_0 = const()[name = string("op_1773_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1773_cast_fp16 = matmul(transpose_x = var_1773_transpose_x_0, transpose_y = var_1773_transpose_y_0, x = var_1772_cast_fp16, y = S_7_cast_fp16)[name = string("op_1773_cast_fp16")];
            bool var_1774_transpose_x_0 = const()[name = string("op_1774_transpose_x_0"), val = bool(false)];
            bool var_1774_transpose_y_0 = const()[name = string("op_1774_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1774_cast_fp16 = matmul(transpose_x = var_1774_transpose_x_0, transpose_y = var_1774_transpose_y_0, x = A_11_cast_fp16, y = vnew_7_cast_fp16)[name = string("op_1774_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1776_cast_fp16 = add(x = var_1773_cast_fp16, y = var_1774_cast_fp16)[name = string("op_1776_cast_fp16")];
            tensor<fp16, [16]> var_1777_bias_0_to_fp16 = const()[name = string("op_1777_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_1777_cast_fp16 = linear(bias = var_1777_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_15_cast_fp16)[name = string("op_1777_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_9_cast_fp16 = exp(x = var_1777_cast_fp16)[name = string("tail_9_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_17_to_fp16 = const()[name = string("transpose_17_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_1779_bias_0_to_fp16 = const()[name = string("op_1779_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_1779_cast_fp16 = linear(bias = var_1779_bias_0_to_fp16, weight = transpose_17_to_fp16, x = gdec_15_cast_fp16)[name = string("op_1779_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_1780_cast_fp16 = exp(x = var_1779_cast_fp16)[name = string("op_1780_cast_fp16")];
            tensor<int32, [1]> last_7_axes_0 = const()[name = string("last_7_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_7_cast_fp16 = expand_dims(axes = last_7_axes_0, x = var_1780_cast_fp16)[name = string("last_7_cast_fp16")];
            tensor<int32, [1]> var_1784_axes_0 = const()[name = string("op_1784_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_1784_cast_fp16 = expand_dims(axes = var_1784_axes_0, x = tail_9_cast_fp16)[name = string("op_1784_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1785_cast_fp16 = mul(x = k_27_cast_fp16, y = var_1784_cast_fp16)[name = string("op_1785_cast_fp16")];
            bool upd_5_transpose_x_1 = const()[name = string("upd_5_transpose_x_1"), val = bool(true)];
            bool upd_5_transpose_y_1 = const()[name = string("upd_5_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_5_cast_fp16 = matmul(transpose_x = upd_5_transpose_x_1, transpose_y = upd_5_transpose_y_1, x = var_1785_cast_fp16, y = vnew_7_cast_fp16)[name = string("upd_5_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_1790_cast_fp16 = mul(x = S_7_cast_fp16, y = last_7_cast_fp16)[name = string("op_1790_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_9_cast_fp16 = add(x = var_1790_cast_fp16, y = upd_5_cast_fp16)[name = string("S_9_cast_fp16")];
            tensor<int32, [4]> q_29_begin_0 = const()[name = string("q_29_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_29_end_0 = const()[name = string("q_29_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> q_29_end_mask_0 = const()[name = string("q_29_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_29_cast_fp16 = slice_by_index(begin = q_29_begin_0, end = q_29_end_0, end_mask = q_29_end_mask_0, x = q_23_cast_fp16)[name = string("q_29_cast_fp16")];
            tensor<int32, [4]> k_29_begin_0 = const()[name = string("k_29_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_29_end_0 = const()[name = string("k_29_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> k_29_end_mask_0 = const()[name = string("k_29_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, x = k_23_cast_fp16)[name = string("k_29_cast_fp16")];
            tensor<int32, [4]> kb_13_begin_0 = const()[name = string("kb_13_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_13_end_0 = const()[name = string("kb_13_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> kb_13_end_mask_0 = const()[name = string("kb_13_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_13_cast_fp16 = slice_by_index(begin = kb_13_begin_0, end = kb_13_end_0, end_mask = kb_13_end_mask_0, x = var_1539_cast_fp16)[name = string("kb_13_cast_fp16")];
            tensor<int32, [4]> vb_17_begin_0 = const()[name = string("vb_17_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_17_end_0 = const()[name = string("vb_17_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> vb_17_end_mask_0 = const()[name = string("vb_17_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_17_cast_fp16 = slice_by_index(begin = vb_17_begin_0, end = vb_17_end_0, end_mask = vb_17_end_mask_0, x = vb_11_cast_fp16)[name = string("vb_17_cast_fp16")];
            tensor<int32, [3]> gdec_17_begin_0 = const()[name = string("gdec_17_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_17_end_0 = const()[name = string("gdec_17_end_0"), val = tensor<int32, [3]>([1, 4, 48])];
            tensor<bool, [3]> gdec_17_end_mask_0 = const()[name = string("gdec_17_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_17_cast_fp16 = slice_by_index(begin = gdec_17_begin_0, end = gdec_17_end_0, end_mask = gdec_17_end_mask_0, x = gdec_11_cast_fp16)[name = string("gdec_17_cast_fp16")];
            tensor<int32, [2]> var_1873 = const()[name = string("op_1873"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_1874_cast_fp16 = reshape(shape = var_1873, x = gdec_17_cast_fp16)[name = string("op_1874_cast_fp16")];
            tensor<fp16, [4, 256]> var_1875_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1874_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [4]> var_1880 = const()[name = string("op_1880"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_13_cast_fp16 = reshape(shape = var_1880, x = var_1875_cast_fp16)[name = string("P_13_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1882_cast_fp16 = exp(x = P_13_cast_fp16)[name = string("op_1882_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_13_cast_fp16 = mul(x = var_1882_cast_fp16, y = tril_incl_to_fp16)[name = string("D_13_cast_fp16")];
            bool kk_13_transpose_x_1 = const()[name = string("kk_13_transpose_x_1"), val = bool(false)];
            bool kk_13_transpose_y_1 = const()[name = string("kk_13_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_13_cast_fp16 = matmul(transpose_x = kk_13_transpose_x_1, transpose_y = kk_13_transpose_y_1, x = kb_13_cast_fp16, y = k_29_cast_fp16)[name = string("kk_13_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_1888_cast_fp16 = mul(x = kk_13_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_1888_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1889_cast_fp16 = mul(x = var_1888_cast_fp16, y = D_13_cast_fp16)[name = string("op_1889_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_13_cast_fp16 = mul(x = var_1889_cast_fp16, y = cum_tail_to_fp16)[name = string("N_13_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1891_cast_fp16 = mul(x = D_13_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1891_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_13_cast_fp16 = mul(x = kk_13_cast_fp16, y = var_1891_cast_fp16)[name = string("Na_13_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_49_cast_fp16 = add(x = eye_to_fp16, y = N_13_cast_fp16)[name = string("Minv_49_cast_fp16")];
            bool Np_37_transpose_x_0 = const()[name = string("Np_37_transpose_x_0"), val = bool(false)];
            bool Np_37_transpose_y_0 = const()[name = string("Np_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_37_cast_fp16 = matmul(transpose_x = Np_37_transpose_x_0, transpose_y = Np_37_transpose_y_0, x = N_13_cast_fp16, y = Na_13_cast_fp16)[name = string("Np_37_cast_fp16")];
            bool Npa_25_transpose_x_0 = const()[name = string("Npa_25_transpose_x_0"), val = bool(false)];
            bool Npa_25_transpose_y_0 = const()[name = string("Npa_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_25_cast_fp16 = matmul(transpose_x = Npa_25_transpose_x_0, transpose_y = Npa_25_transpose_y_0, x = Na_13_cast_fp16, y = N_13_cast_fp16)[name = string("Npa_25_cast_fp16")];
            bool var_1897_transpose_x_0 = const()[name = string("op_1897_transpose_x_0"), val = bool(false)];
            bool var_1897_transpose_y_0 = const()[name = string("op_1897_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1897_cast_fp16 = matmul(transpose_x = var_1897_transpose_x_0, transpose_y = var_1897_transpose_y_0, x = Minv_49_cast_fp16, y = Np_37_cast_fp16)[name = string("op_1897_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_51_cast_fp16 = add(x = Minv_49_cast_fp16, y = var_1897_cast_fp16)[name = string("Minv_51_cast_fp16")];
            bool Np_39_transpose_x_0 = const()[name = string("Np_39_transpose_x_0"), val = bool(false)];
            bool Np_39_transpose_y_0 = const()[name = string("Np_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_39_cast_fp16 = matmul(transpose_x = Np_39_transpose_x_0, transpose_y = Np_39_transpose_y_0, x = Np_37_cast_fp16, y = Npa_25_cast_fp16)[name = string("Np_39_cast_fp16")];
            bool Npa_27_transpose_x_0 = const()[name = string("Npa_27_transpose_x_0"), val = bool(false)];
            bool Npa_27_transpose_y_0 = const()[name = string("Npa_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_27_cast_fp16 = matmul(transpose_x = Npa_27_transpose_x_0, transpose_y = Npa_27_transpose_y_0, x = Npa_25_cast_fp16, y = Np_37_cast_fp16)[name = string("Npa_27_cast_fp16")];
            bool var_1902_transpose_x_0 = const()[name = string("op_1902_transpose_x_0"), val = bool(false)];
            bool var_1902_transpose_y_0 = const()[name = string("op_1902_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1902_cast_fp16 = matmul(transpose_x = var_1902_transpose_x_0, transpose_y = var_1902_transpose_y_0, x = Minv_51_cast_fp16, y = Np_39_cast_fp16)[name = string("op_1902_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_53_cast_fp16 = add(x = Minv_51_cast_fp16, y = var_1902_cast_fp16)[name = string("Minv_53_cast_fp16")];
            bool Np_41_transpose_x_0 = const()[name = string("Np_41_transpose_x_0"), val = bool(false)];
            bool Np_41_transpose_y_0 = const()[name = string("Np_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_41_cast_fp16 = matmul(transpose_x = Np_41_transpose_x_0, transpose_y = Np_41_transpose_y_0, x = Np_39_cast_fp16, y = Npa_27_cast_fp16)[name = string("Np_41_cast_fp16")];
            bool var_1906_transpose_x_0 = const()[name = string("op_1906_transpose_x_0"), val = bool(false)];
            bool var_1906_transpose_y_0 = const()[name = string("op_1906_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1906_cast_fp16 = matmul(transpose_x = var_1906_transpose_x_0, transpose_y = var_1906_transpose_y_0, x = Minv_53_cast_fp16, y = Np_41_cast_fp16)[name = string("op_1906_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_55_cast_fp16 = add(x = Minv_53_cast_fp16, y = var_1906_cast_fp16)[name = string("Minv_55_cast_fp16")];
            bool u_13_transpose_x_0 = const()[name = string("u_13_transpose_x_0"), val = bool(false)];
            bool u_13_transpose_y_0 = const()[name = string("u_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_13_cast_fp16 = matmul(transpose_x = u_13_transpose_x_0, transpose_y = u_13_transpose_y_0, x = Minv_55_cast_fp16, y = vb_17_cast_fp16)[name = string("u_13_cast_fp16")];
            bool var_1913_transpose_x_1 = const()[name = string("op_1913_transpose_x_1"), val = bool(false)];
            bool var_1913_transpose_y_1 = const()[name = string("op_1913_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_1913_cast_fp16 = matmul(transpose_x = var_1913_transpose_x_1, transpose_y = var_1913_transpose_y_1, x = q_29_cast_fp16, y = k_29_cast_fp16)[name = string("op_1913_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_13_cast_fp16 = mul(x = var_1913_cast_fp16, y = D_13_cast_fp16)[name = string("A_13_cast_fp16")];
            tensor<fp16, [16]> var_1915_bias_0_to_fp16 = const()[name = string("op_1915_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_1915_cast_fp16 = linear(bias = var_1915_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_17_cast_fp16)[name = string("op_1915_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_9_cast_fp16 = exp(x = var_1915_cast_fp16)[name = string("cum_9_cast_fp16")];
            tensor<int32, [1]> var_1918_axes_0 = const()[name = string("op_1918_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_1918_cast_fp16 = expand_dims(axes = var_1918_axes_0, x = cum_9_cast_fp16)[name = string("op_1918_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1919_cast_fp16 = mul(x = kb_13_cast_fp16, y = var_1918_cast_fp16)[name = string("op_1919_cast_fp16")];
            bool w_13_transpose_x_0 = const()[name = string("w_13_transpose_x_0"), val = bool(false)];
            bool w_13_transpose_y_0 = const()[name = string("w_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_13_cast_fp16 = matmul(transpose_x = w_13_transpose_x_0, transpose_y = w_13_transpose_y_0, x = Minv_55_cast_fp16, y = var_1919_cast_fp16)[name = string("w_13_cast_fp16")];
            bool var_1921_transpose_x_0 = const()[name = string("op_1921_transpose_x_0"), val = bool(false)];
            bool var_1921_transpose_y_0 = const()[name = string("op_1921_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1921_cast_fp16 = matmul(transpose_x = var_1921_transpose_x_0, transpose_y = var_1921_transpose_y_0, x = w_13_cast_fp16, y = S_9_cast_fp16)[name = string("op_1921_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_9_cast_fp16 = sub(x = u_13_cast_fp16, y = var_1921_cast_fp16)[name = string("vnew_9_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1926_cast_fp16 = mul(x = q_29_cast_fp16, y = var_1918_cast_fp16)[name = string("op_1926_cast_fp16")];
            bool var_1927_transpose_x_0 = const()[name = string("op_1927_transpose_x_0"), val = bool(false)];
            bool var_1927_transpose_y_0 = const()[name = string("op_1927_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1927_cast_fp16 = matmul(transpose_x = var_1927_transpose_x_0, transpose_y = var_1927_transpose_y_0, x = var_1926_cast_fp16, y = S_9_cast_fp16)[name = string("op_1927_cast_fp16")];
            bool var_1928_transpose_x_0 = const()[name = string("op_1928_transpose_x_0"), val = bool(false)];
            bool var_1928_transpose_y_0 = const()[name = string("op_1928_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1928_cast_fp16 = matmul(transpose_x = var_1928_transpose_x_0, transpose_y = var_1928_transpose_y_0, x = A_13_cast_fp16, y = vnew_9_cast_fp16)[name = string("op_1928_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1930_cast_fp16 = add(x = var_1927_cast_fp16, y = var_1928_cast_fp16)[name = string("op_1930_cast_fp16")];
            tensor<fp16, [16]> var_1931_bias_0_to_fp16 = const()[name = string("op_1931_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_1931_cast_fp16 = linear(bias = var_1931_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_17_cast_fp16)[name = string("op_1931_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_11_cast_fp16 = exp(x = var_1931_cast_fp16)[name = string("tail_11_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_21_to_fp16 = const()[name = string("transpose_21_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_1933_bias_0_to_fp16 = const()[name = string("op_1933_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_1933_cast_fp16 = linear(bias = var_1933_bias_0_to_fp16, weight = transpose_21_to_fp16, x = gdec_17_cast_fp16)[name = string("op_1933_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_1934_cast_fp16 = exp(x = var_1933_cast_fp16)[name = string("op_1934_cast_fp16")];
            tensor<int32, [1]> last_9_axes_0 = const()[name = string("last_9_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_9_cast_fp16 = expand_dims(axes = last_9_axes_0, x = var_1934_cast_fp16)[name = string("last_9_cast_fp16")];
            tensor<int32, [1]> var_1938_axes_0 = const()[name = string("op_1938_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_1938_cast_fp16 = expand_dims(axes = var_1938_axes_0, x = tail_11_cast_fp16)[name = string("op_1938_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1939_cast_fp16 = mul(x = k_29_cast_fp16, y = var_1938_cast_fp16)[name = string("op_1939_cast_fp16")];
            bool upd_7_transpose_x_1 = const()[name = string("upd_7_transpose_x_1"), val = bool(true)];
            bool upd_7_transpose_y_1 = const()[name = string("upd_7_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_7_cast_fp16 = matmul(transpose_x = upd_7_transpose_x_1, transpose_y = upd_7_transpose_y_1, x = var_1939_cast_fp16, y = vnew_9_cast_fp16)[name = string("upd_7_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_1944_cast_fp16 = mul(x = S_9_cast_fp16, y = last_9_cast_fp16)[name = string("op_1944_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_11_cast_fp16 = add(x = var_1944_cast_fp16, y = upd_7_cast_fp16)[name = string("S_11_cast_fp16")];
            tensor<int32, [4]> q_31_begin_0 = const()[name = string("q_31_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_31_end_0 = const()[name = string("q_31_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_31_end_mask_0 = const()[name = string("q_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, x = q_23_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<int32, [4]> k_31_begin_0 = const()[name = string("k_31_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_31_end_0 = const()[name = string("k_31_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_31_end_mask_0 = const()[name = string("k_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_31_cast_fp16 = slice_by_index(begin = k_31_begin_0, end = k_31_end_0, end_mask = k_31_end_mask_0, x = k_23_cast_fp16)[name = string("k_31_cast_fp16")];
            tensor<int32, [4]> kb_15_begin_0 = const()[name = string("kb_15_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_15_end_0 = const()[name = string("kb_15_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_15_end_mask_0 = const()[name = string("kb_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_15_cast_fp16 = slice_by_index(begin = kb_15_begin_0, end = kb_15_end_0, end_mask = kb_15_end_mask_0, x = var_1539_cast_fp16)[name = string("kb_15_cast_fp16")];
            tensor<int32, [4]> vb_19_begin_0 = const()[name = string("vb_19_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_19_end_0 = const()[name = string("vb_19_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_19_end_mask_0 = const()[name = string("vb_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_19_cast_fp16 = slice_by_index(begin = vb_19_begin_0, end = vb_19_end_0, end_mask = vb_19_end_mask_0, x = vb_11_cast_fp16)[name = string("vb_19_cast_fp16")];
            tensor<int32, [3]> gdec_19_begin_0 = const()[name = string("gdec_19_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_19_end_0 = const()[name = string("gdec_19_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_19_end_mask_0 = const()[name = string("gdec_19_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_19_cast_fp16 = slice_by_index(begin = gdec_19_begin_0, end = gdec_19_end_0, end_mask = gdec_19_end_mask_0, x = gdec_11_cast_fp16)[name = string("gdec_19_cast_fp16")];
            tensor<int32, [2]> var_2027 = const()[name = string("op_2027"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_2028_cast_fp16 = reshape(shape = var_2027, x = gdec_19_cast_fp16)[name = string("op_2028_cast_fp16")];
            tensor<fp16, [4, 256]> var_2029_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2028_cast_fp16)[name = string("op_2029_cast_fp16")];
            tensor<int32, [4]> var_2034 = const()[name = string("op_2034"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_15_cast_fp16 = reshape(shape = var_2034, x = var_2029_cast_fp16)[name = string("P_15_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2036_cast_fp16 = exp(x = P_15_cast_fp16)[name = string("op_2036_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_15_cast_fp16 = mul(x = var_2036_cast_fp16, y = tril_incl_to_fp16)[name = string("D_15_cast_fp16")];
            bool kk_15_transpose_x_1 = const()[name = string("kk_15_transpose_x_1"), val = bool(false)];
            bool kk_15_transpose_y_1 = const()[name = string("kk_15_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_15_cast_fp16 = matmul(transpose_x = kk_15_transpose_x_1, transpose_y = kk_15_transpose_y_1, x = kb_15_cast_fp16, y = k_31_cast_fp16)[name = string("kk_15_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_2042_cast_fp16 = mul(x = kk_15_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_2042_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2043_cast_fp16 = mul(x = var_2042_cast_fp16, y = D_15_cast_fp16)[name = string("op_2043_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_15_cast_fp16 = mul(x = var_2043_cast_fp16, y = cum_tail_to_fp16)[name = string("N_15_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2045_cast_fp16 = mul(x = D_15_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2045_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_15_cast_fp16 = mul(x = kk_15_cast_fp16, y = var_2045_cast_fp16)[name = string("Na_15_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_57_cast_fp16 = add(x = eye_to_fp16, y = N_15_cast_fp16)[name = string("Minv_57_cast_fp16")];
            bool Np_43_transpose_x_0 = const()[name = string("Np_43_transpose_x_0"), val = bool(false)];
            bool Np_43_transpose_y_0 = const()[name = string("Np_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_43_cast_fp16 = matmul(transpose_x = Np_43_transpose_x_0, transpose_y = Np_43_transpose_y_0, x = N_15_cast_fp16, y = Na_15_cast_fp16)[name = string("Np_43_cast_fp16")];
            bool Npa_29_transpose_x_0 = const()[name = string("Npa_29_transpose_x_0"), val = bool(false)];
            bool Npa_29_transpose_y_0 = const()[name = string("Npa_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_29_cast_fp16 = matmul(transpose_x = Npa_29_transpose_x_0, transpose_y = Npa_29_transpose_y_0, x = Na_15_cast_fp16, y = N_15_cast_fp16)[name = string("Npa_29_cast_fp16")];
            bool var_2051_transpose_x_0 = const()[name = string("op_2051_transpose_x_0"), val = bool(false)];
            bool var_2051_transpose_y_0 = const()[name = string("op_2051_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2051_cast_fp16 = matmul(transpose_x = var_2051_transpose_x_0, transpose_y = var_2051_transpose_y_0, x = Minv_57_cast_fp16, y = Np_43_cast_fp16)[name = string("op_2051_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_59_cast_fp16 = add(x = Minv_57_cast_fp16, y = var_2051_cast_fp16)[name = string("Minv_59_cast_fp16")];
            bool Np_45_transpose_x_0 = const()[name = string("Np_45_transpose_x_0"), val = bool(false)];
            bool Np_45_transpose_y_0 = const()[name = string("Np_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_45_cast_fp16 = matmul(transpose_x = Np_45_transpose_x_0, transpose_y = Np_45_transpose_y_0, x = Np_43_cast_fp16, y = Npa_29_cast_fp16)[name = string("Np_45_cast_fp16")];
            bool Npa_31_transpose_x_0 = const()[name = string("Npa_31_transpose_x_0"), val = bool(false)];
            bool Npa_31_transpose_y_0 = const()[name = string("Npa_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_31_cast_fp16 = matmul(transpose_x = Npa_31_transpose_x_0, transpose_y = Npa_31_transpose_y_0, x = Npa_29_cast_fp16, y = Np_43_cast_fp16)[name = string("Npa_31_cast_fp16")];
            bool var_2056_transpose_x_0 = const()[name = string("op_2056_transpose_x_0"), val = bool(false)];
            bool var_2056_transpose_y_0 = const()[name = string("op_2056_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2056_cast_fp16 = matmul(transpose_x = var_2056_transpose_x_0, transpose_y = var_2056_transpose_y_0, x = Minv_59_cast_fp16, y = Np_45_cast_fp16)[name = string("op_2056_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_61_cast_fp16 = add(x = Minv_59_cast_fp16, y = var_2056_cast_fp16)[name = string("Minv_61_cast_fp16")];
            bool Np_47_transpose_x_0 = const()[name = string("Np_47_transpose_x_0"), val = bool(false)];
            bool Np_47_transpose_y_0 = const()[name = string("Np_47_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_47_cast_fp16 = matmul(transpose_x = Np_47_transpose_x_0, transpose_y = Np_47_transpose_y_0, x = Np_45_cast_fp16, y = Npa_31_cast_fp16)[name = string("Np_47_cast_fp16")];
            bool var_2060_transpose_x_0 = const()[name = string("op_2060_transpose_x_0"), val = bool(false)];
            bool var_2060_transpose_y_0 = const()[name = string("op_2060_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2060_cast_fp16 = matmul(transpose_x = var_2060_transpose_x_0, transpose_y = var_2060_transpose_y_0, x = Minv_61_cast_fp16, y = Np_47_cast_fp16)[name = string("op_2060_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_63_cast_fp16 = add(x = Minv_61_cast_fp16, y = var_2060_cast_fp16)[name = string("Minv_63_cast_fp16")];
            bool u_15_transpose_x_0 = const()[name = string("u_15_transpose_x_0"), val = bool(false)];
            bool u_15_transpose_y_0 = const()[name = string("u_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_15_cast_fp16 = matmul(transpose_x = u_15_transpose_x_0, transpose_y = u_15_transpose_y_0, x = Minv_63_cast_fp16, y = vb_19_cast_fp16)[name = string("u_15_cast_fp16")];
            bool var_2067_transpose_x_1 = const()[name = string("op_2067_transpose_x_1"), val = bool(false)];
            bool var_2067_transpose_y_1 = const()[name = string("op_2067_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_2067_cast_fp16 = matmul(transpose_x = var_2067_transpose_x_1, transpose_y = var_2067_transpose_y_1, x = q_31_cast_fp16, y = k_31_cast_fp16)[name = string("op_2067_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_15_cast_fp16 = mul(x = var_2067_cast_fp16, y = D_15_cast_fp16)[name = string("A_15_cast_fp16")];
            tensor<fp16, [16]> var_2069_bias_0_to_fp16 = const()[name = string("op_2069_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_2069_cast_fp16 = linear(bias = var_2069_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_19_cast_fp16)[name = string("op_2069_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_11_cast_fp16 = exp(x = var_2069_cast_fp16)[name = string("cum_11_cast_fp16")];
            tensor<int32, [1]> var_2072_axes_0 = const()[name = string("op_2072_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_2072_cast_fp16 = expand_dims(axes = var_2072_axes_0, x = cum_11_cast_fp16)[name = string("op_2072_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2073_cast_fp16 = mul(x = kb_15_cast_fp16, y = var_2072_cast_fp16)[name = string("op_2073_cast_fp16")];
            bool w_15_transpose_x_0 = const()[name = string("w_15_transpose_x_0"), val = bool(false)];
            bool w_15_transpose_y_0 = const()[name = string("w_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_15_cast_fp16 = matmul(transpose_x = w_15_transpose_x_0, transpose_y = w_15_transpose_y_0, x = Minv_63_cast_fp16, y = var_2073_cast_fp16)[name = string("w_15_cast_fp16")];
            bool var_2075_transpose_x_0 = const()[name = string("op_2075_transpose_x_0"), val = bool(false)];
            bool var_2075_transpose_y_0 = const()[name = string("op_2075_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2075_cast_fp16 = matmul(transpose_x = var_2075_transpose_x_0, transpose_y = var_2075_transpose_y_0, x = w_15_cast_fp16, y = S_11_cast_fp16)[name = string("op_2075_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_11_cast_fp16 = sub(x = u_15_cast_fp16, y = var_2075_cast_fp16)[name = string("vnew_11_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2080_cast_fp16 = mul(x = q_31_cast_fp16, y = var_2072_cast_fp16)[name = string("op_2080_cast_fp16")];
            bool var_2081_transpose_x_0 = const()[name = string("op_2081_transpose_x_0"), val = bool(false)];
            bool var_2081_transpose_y_0 = const()[name = string("op_2081_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2081_cast_fp16 = matmul(transpose_x = var_2081_transpose_x_0, transpose_y = var_2081_transpose_y_0, x = var_2080_cast_fp16, y = S_11_cast_fp16)[name = string("op_2081_cast_fp16")];
            bool var_2082_transpose_x_0 = const()[name = string("op_2082_transpose_x_0"), val = bool(false)];
            bool var_2082_transpose_y_0 = const()[name = string("op_2082_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2082_cast_fp16 = matmul(transpose_x = var_2082_transpose_x_0, transpose_y = var_2082_transpose_y_0, x = A_15_cast_fp16, y = vnew_11_cast_fp16)[name = string("op_2082_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_3_cast_fp16 = add(x = var_2081_cast_fp16, y = var_2082_cast_fp16)[name = string("oc_3_cast_fp16")];
            int32 var_2086 = const()[name = string("op_2086"), val = int32(2)];
            bool o_7_interleave_0 = const()[name = string("o_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 128]> o_7_cast_fp16 = concat(axis = var_2086, interleave = o_7_interleave_0, values = (var_1629_cast_fp16, var_1776_cast_fp16, var_1930_cast_fp16, oc_3_cast_fp16))[name = string("o_7_cast_fp16")];
            tensor<int32, [4]> var_2092 = const()[name = string("op_2092"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2098 = const()[name = string("op_2098"), val = tensor<int32, [4]>([1, 64, 4, 128])];
            tensor<fp16, [1, 64, 4, 128]> input_23_cast_fp16 = reshape(shape = var_2098, x = z_3_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> o_9_cast_fp16 = transpose(perm = var_2092, x = o_7_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 64, 4, 128]> var_2100_cast_fp16 = mul(x = o_9_cast_fp16, y = o_9_cast_fp16)[name = string("op_2100_cast_fp16")];
            tensor<int32, [1]> var_2105_axes_0 = const()[name = string("op_2105_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2105_keep_dims_0 = const()[name = string("op_2105_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 4, 1]> var_2105_cast_fp16 = reduce_mean(axes = var_2105_axes_0, keep_dims = var_2105_keep_dims_0, x = var_2100_cast_fp16)[name = string("op_2105_cast_fp16")];
            fp16 var_2107_to_fp16 = const()[name = string("op_2107_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 64, 4, 1]> var_2108_cast_fp16 = add(x = var_2105_cast_fp16, y = var_2107_to_fp16)[name = string("op_2108_cast_fp16")];
            fp32 var_2109_epsilon_0 = const()[name = string("op_2109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 4, 1]> var_2109_cast_fp16 = rsqrt(epsilon = var_2109_epsilon_0, x = var_2108_cast_fp16)[name = string("op_2109_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> var_2110_cast_fp16 = mul(x = o_9_cast_fp16, y = var_2109_cast_fp16)[name = string("op_2110_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> var_2112_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_2112_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> o_11_cast_fp16 = mul(x = var_2110_cast_fp16, y = var_2112_cast_fp16)[name = string("o_11_cast_fp16")];
            tensor<int32, [3]> var_2117 = const()[name = string("op_2117"), val = tensor<int32, [3]>([1, 64, 512])];
            tensor<fp16, [1, 64, 512]> input_25_cast_fp16 = reshape(shape = var_2117, x = o_11_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [640, 512]> layers_1_out_weight_to_fp16 = const()[name = string("layers_1_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34984256)))];
            tensor<fp16, [1, 64, 640]> linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_out_weight_to_fp16, x = input_25_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 64, 640]> x_13_cast_fp16 = add(x = x_9_cast_fp16, y = linear_5_cast_fp16)[name = string("x_13_cast_fp16")];
            fp16 var_2130_to_fp16 = const()[name = string("op_2130_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> x_15_cast_fp16 = mul(x = x_13_cast_fp16, y = var_2130_to_fp16)[name = string("x_15_cast_fp16")];
            fp16 var_2128_promoted_to_fp16 = const()[name = string("op_2128_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> var_2132_cast_fp16 = pow(x = x_15_cast_fp16, y = var_2128_promoted_to_fp16)[name = string("op_2132_cast_fp16")];
            tensor<int32, [1]> var_2134_axes_0 = const()[name = string("op_2134_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2134_keep_dims_0 = const()[name = string("op_2134_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_2134_cast_fp16 = reduce_mean(axes = var_2134_axes_0, keep_dims = var_2134_keep_dims_0, x = var_2132_cast_fp16)[name = string("op_2134_cast_fp16")];
            fp16 var_2135_to_fp16 = const()[name = string("op_2135_to_fp16"), val = fp16(0x1.0cp-18)];
            tensor<fp16, [1, 64, 1]> var_2136_cast_fp16 = add(x = var_2134_cast_fp16, y = var_2135_to_fp16)[name = string("op_2136_cast_fp16")];
            fp32 var_2137_epsilon_0 = const()[name = string("op_2137_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2137_cast_fp16 = rsqrt(epsilon = var_2137_epsilon_0, x = var_2136_cast_fp16)[name = string("op_2137_cast_fp16")];
            tensor<fp16, [1, 64, 640]> var_2138_cast_fp16 = mul(x = x_15_cast_fp16, y = var_2137_cast_fp16)[name = string("op_2138_cast_fp16")];
            tensor<fp16, [640]> layers_1_n2_w1_to_fp16 = const()[name = string("layers_1_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35639680)))];
            tensor<fp16, [1, 64, 640]> input_27_cast_fp16 = mul(x = var_2138_cast_fp16, y = layers_1_n2_w1_to_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_1_gate_up_weight_to_fp16 = const()[name = string("layers_1_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35641024)))];
            tensor<fp16, [1, 64, 5120]> linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_1_gate_up_weight_to_fp16, x = input_27_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<int32, [3]> input_29_begin_0 = const()[name = string("input_29_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_29_end_0 = const()[name = string("input_29_end_0"), val = tensor<int32, [3]>([1, 64, 2560])];
            tensor<bool, [3]> input_29_end_mask_0 = const()[name = string("input_29_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 2560]> input_29_cast_fp16 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = linear_6_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [1, 64, 2560]> var_2148_cast_fp16 = silu(x = input_29_cast_fp16)[name = string("op_2148_cast_fp16")];
            tensor<int32, [3]> var_2153_begin_0 = const()[name = string("op_2153_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_2153_end_0 = const()[name = string("op_2153_end_0"), val = tensor<int32, [3]>([1, 64, 5120])];
            tensor<bool, [3]> var_2153_end_mask_0 = const()[name = string("op_2153_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 2560]> var_2153_cast_fp16 = slice_by_index(begin = var_2153_begin_0, end = var_2153_end_0, end_mask = var_2153_end_mask_0, x = linear_6_cast_fp16)[name = string("op_2153_cast_fp16")];
            tensor<fp16, [1, 64, 2560]> input_31_cast_fp16 = mul(x = var_2148_cast_fp16, y = var_2153_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_1_down_weight_to_fp16 = const()[name = string("layers_1_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42194688)))];
            tensor<fp16, [1, 64, 640]> linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_down_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 64, 640]> x_17_cast_fp16 = add(x = x_13_cast_fp16, y = linear_7_cast_fp16)[name = string("x_17_cast_fp16")];
            fp16 var_2164_promoted_to_fp16 = const()[name = string("op_2164_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> var_2168_cast_fp16 = pow(x = x_17_cast_fp16, y = var_2164_promoted_to_fp16)[name = string("op_2168_cast_fp16")];
            tensor<int32, [1]> var_2170_axes_0 = const()[name = string("op_2170_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2170_keep_dims_0 = const()[name = string("op_2170_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_2170_cast_fp16 = reduce_mean(axes = var_2170_axes_0, keep_dims = var_2170_keep_dims_0, x = var_2168_cast_fp16)[name = string("op_2170_cast_fp16")];
            fp16 var_2171_to_fp16 = const()[name = string("op_2171_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2172_cast_fp16 = add(x = var_2170_cast_fp16, y = var_2171_to_fp16)[name = string("op_2172_cast_fp16")];
            fp32 var_2173_epsilon_0 = const()[name = string("op_2173_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2173_cast_fp16 = rsqrt(epsilon = var_2173_epsilon_0, x = var_2172_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<fp16, [1, 64, 640]> var_2174_cast_fp16 = mul(x = x_17_cast_fp16, y = var_2173_cast_fp16)[name = string("op_2174_cast_fp16")];
            tensor<fp16, [640]> layers_2_n1_w1_to_fp16 = const()[name = string("layers_2_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45471552)))];
            tensor<fp16, [1, 64, 640]> input_33_cast_fp16 = mul(x = var_2174_cast_fp16, y = layers_2_n1_w1_to_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_2_qkvzba_weight_to_fp16 = const()[name = string("layers_2_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45472896)))];
            tensor<fp16, [1, 64, 1544]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_qkvzba_weight_to_fp16, x = input_33_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [3]> qkv_5_begin_0 = const()[name = string("qkv_5_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_5_end_0 = const()[name = string("qkv_5_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> qkv_5_end_mask_0 = const()[name = string("qkv_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 1024]> qkv_5_cast_fp16 = slice_by_index(begin = qkv_5_begin_0, end = qkv_5_end_0, end_mask = qkv_5_end_mask_0, x = linear_8_cast_fp16)[name = string("qkv_5_cast_fp16")];
            tensor<int32, [3]> z_5_begin_0 = const()[name = string("z_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_5_end_0 = const()[name = string("z_5_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
            tensor<bool, [3]> z_5_end_mask_0 = const()[name = string("z_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 512]> z_5_cast_fp16 = slice_by_index(begin = z_5_begin_0, end = z_5_end_0, end_mask = z_5_end_mask_0, x = linear_8_cast_fp16)[name = string("z_5_cast_fp16")];
            tensor<int32, [3]> b_5_begin_0 = const()[name = string("b_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_5_end_0 = const()[name = string("b_5_end_0"), val = tensor<int32, [3]>([1, 64, 1540])];
            tensor<bool, [3]> b_5_end_mask_0 = const()[name = string("b_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 4]> b_5_cast_fp16 = slice_by_index(begin = b_5_begin_0, end = b_5_end_0, end_mask = b_5_end_mask_0, x = linear_8_cast_fp16)[name = string("b_5_cast_fp16")];
            tensor<int32, [3]> a_5_begin_0 = const()[name = string("a_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_5_end_0 = const()[name = string("a_5_end_0"), val = tensor<int32, [3]>([1, 64, 1])];
            tensor<bool, [3]> a_5_end_mask_0 = const()[name = string("a_5_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 4]> a_5_cast_fp16 = slice_by_index(begin = a_5_begin_0, end = a_5_end_0, end_mask = a_5_end_mask_0, x = linear_8_cast_fp16)[name = string("a_5_cast_fp16")];
            int32 var_2213 = const()[name = string("op_2213"), val = int32(1)];
            bool xp_5_interleave_0 = const()[name = string("xp_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 67, 1024]> xp_5_cast_fp16 = concat(axis = var_2213, interleave = xp_5_interleave_0, values = (zpad_to_fp16, qkv_5_cast_fp16))[name = string("xp_5_cast_fp16")];
            tensor<int32, [3]> var_2224_begin_0 = const()[name = string("op_2224_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_2224_end_0 = const()[name = string("op_2224_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_2224_end_mask_0 = const()[name = string("op_2224_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 1024]> var_2224_cast_fp16 = slice_by_index(begin = var_2224_begin_0, end = var_2224_end_0, end_mask = var_2224_end_mask_0, x = xp_5_cast_fp16)[name = string("op_2224_cast_fp16")];
            tensor<fp16, [1024]> var_2232_to_fp16 = const()[name = string("op_2232_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47449280)))];
            tensor<fp16, [1, 64, 1024]> var_2233_cast_fp16 = mul(x = var_2224_cast_fp16, y = var_2232_to_fp16)[name = string("op_2233_cast_fp16")];
            tensor<int32, [3]> var_2243_begin_0 = const()[name = string("op_2243_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_2243_end_0 = const()[name = string("op_2243_end_0"), val = tensor<int32, [3]>([1, 66, 1024])];
            tensor<bool, [3]> var_2243_end_mask_0 = const()[name = string("op_2243_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_2243_cast_fp16 = slice_by_index(begin = var_2243_begin_0, end = var_2243_end_0, end_mask = var_2243_end_mask_0, x = xp_5_cast_fp16)[name = string("op_2243_cast_fp16")];
            tensor<fp16, [1024]> var_2251_to_fp16 = const()[name = string("op_2251_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47451392)))];
            tensor<fp16, [1, 64, 1024]> var_2252_cast_fp16 = mul(x = var_2243_cast_fp16, y = var_2251_to_fp16)[name = string("op_2252_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> var_2254_cast_fp16 = add(x = var_2233_cast_fp16, y = var_2252_cast_fp16)[name = string("op_2254_cast_fp16")];
            tensor<int32, [3]> var_2264_begin_0 = const()[name = string("op_2264_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2264_end_0 = const()[name = string("op_2264_end_0"), val = tensor<int32, [3]>([1, 65, 1024])];
            tensor<bool, [3]> var_2264_end_mask_0 = const()[name = string("op_2264_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_2264_cast_fp16 = slice_by_index(begin = var_2264_begin_0, end = var_2264_end_0, end_mask = var_2264_end_mask_0, x = xp_5_cast_fp16)[name = string("op_2264_cast_fp16")];
            tensor<fp16, [1024]> var_2272_to_fp16 = const()[name = string("op_2272_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47453504)))];
            tensor<fp16, [1, 64, 1024]> var_2273_cast_fp16 = mul(x = var_2264_cast_fp16, y = var_2272_to_fp16)[name = string("op_2273_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> var_2275_cast_fp16 = add(x = var_2254_cast_fp16, y = var_2273_cast_fp16)[name = string("op_2275_cast_fp16")];
            tensor<int32, [3]> var_2285_begin_0 = const()[name = string("op_2285_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2285_end_0 = const()[name = string("op_2285_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> var_2285_end_mask_0 = const()[name = string("op_2285_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_2285_cast_fp16 = slice_by_index(begin = var_2285_begin_0, end = var_2285_end_0, end_mask = var_2285_end_mask_0, x = xp_5_cast_fp16)[name = string("op_2285_cast_fp16")];
            tensor<fp16, [1024]> var_2293_to_fp16 = const()[name = string("op_2293_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47455616)))];
            tensor<fp16, [1, 64, 1024]> var_2294_cast_fp16 = mul(x = var_2285_cast_fp16, y = var_2293_to_fp16)[name = string("op_2294_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> input_35_cast_fp16 = add(x = var_2275_cast_fp16, y = var_2294_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> c_5_cast_fp16 = silu(x = input_35_cast_fp16)[name = string("c_5_cast_fp16")];
            tensor<int32, [3]> var_2302_begin_0 = const()[name = string("op_2302_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2302_end_0 = const()[name = string("op_2302_end_0"), val = tensor<int32, [3]>([1, 64, 256])];
            tensor<bool, [3]> var_2302_end_mask_0 = const()[name = string("op_2302_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 256]> var_2302_cast_fp16 = slice_by_index(begin = var_2302_begin_0, end = var_2302_end_0, end_mask = var_2302_end_mask_0, x = c_5_cast_fp16)[name = string("op_2302_cast_fp16")];
            tensor<int32, [4]> var_2307 = const()[name = string("op_2307"), val = tensor<int32, [4]>([1, 64, 2, 128])];
            tensor<fp16, [1, 64, 2, 128]> q_33_cast_fp16 = reshape(shape = var_2307, x = var_2302_cast_fp16)[name = string("q_33_cast_fp16")];
            tensor<int32, [3]> var_2313_begin_0 = const()[name = string("op_2313_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_2313_end_0 = const()[name = string("op_2313_end_0"), val = tensor<int32, [3]>([1, 64, 512])];
            tensor<bool, [3]> var_2313_end_mask_0 = const()[name = string("op_2313_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 256]> var_2313_cast_fp16 = slice_by_index(begin = var_2313_begin_0, end = var_2313_end_0, end_mask = var_2313_end_mask_0, x = c_5_cast_fp16)[name = string("op_2313_cast_fp16")];
            tensor<int32, [4]> var_2318 = const()[name = string("op_2318"), val = tensor<int32, [4]>([1, 64, 2, 128])];
            tensor<fp16, [1, 64, 2, 128]> k_33_cast_fp16 = reshape(shape = var_2318, x = var_2313_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<int32, [3]> var_2324_begin_0 = const()[name = string("op_2324_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_2324_end_0 = const()[name = string("op_2324_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> var_2324_end_mask_0 = const()[name = string("op_2324_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 512]> var_2324_cast_fp16 = slice_by_index(begin = var_2324_begin_0, end = var_2324_end_0, end_mask = var_2324_end_mask_0, x = c_5_cast_fp16)[name = string("op_2324_cast_fp16")];
            tensor<int32, [4]> var_2329 = const()[name = string("op_2329"), val = tensor<int32, [4]>([1, 64, 4, 128])];
            tensor<fp16, [1, 64, 4, 128]> var_2330_cast_fp16 = reshape(shape = var_2329, x = var_2324_cast_fp16)[name = string("op_2330_cast_fp16")];
            tensor<int32, [4]> var_2335 = const()[name = string("op_2335"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2337_to_fp16 = const()[name = string("op_2337_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 64, 2, 128]> q_35_cast_fp16 = mul(x = q_33_cast_fp16, y = var_2337_to_fp16)[name = string("q_35_cast_fp16")];
            fp16 var_2339_to_fp16 = const()[name = string("op_2339_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 64, 2, 128]> k_35_cast_fp16 = mul(x = k_33_cast_fp16, y = var_2339_to_fp16)[name = string("k_35_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_2341_cast_fp16 = mul(x = q_35_cast_fp16, y = q_35_cast_fp16)[name = string("op_2341_cast_fp16")];
            tensor<int32, [1]> var_2346_axes_0 = const()[name = string("op_2346_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2346_keep_dims_0 = const()[name = string("op_2346_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 2, 1]> var_2346_cast_fp16 = reduce_mean(axes = var_2346_axes_0, keep_dims = var_2346_keep_dims_0, x = var_2341_cast_fp16)[name = string("op_2346_cast_fp16")];
            fp16 var_2347_promoted_to_fp16 = const()[name = string("op_2347_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 64, 2, 1]> var_2348_cast_fp16 = mul(x = var_2346_cast_fp16, y = var_2347_promoted_to_fp16)[name = string("op_2348_cast_fp16")];
            fp16 var_2350_to_fp16 = const()[name = string("op_2350_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 64, 2, 1]> var_2351_cast_fp16 = add(x = var_2348_cast_fp16, y = var_2350_to_fp16)[name = string("op_2351_cast_fp16")];
            fp32 var_2352_epsilon_0 = const()[name = string("op_2352_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 2, 1]> var_2352_cast_fp16 = rsqrt(epsilon = var_2352_epsilon_0, x = var_2351_cast_fp16)[name = string("op_2352_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_2353_cast_fp16 = mul(x = q_35_cast_fp16, y = var_2352_cast_fp16)[name = string("op_2353_cast_fp16")];
            fp16 var_2354_to_fp16 = const()[name = string("op_2354_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 64, 2, 128]> q_37_cast_fp16 = mul(x = var_2353_cast_fp16, y = var_2354_to_fp16)[name = string("q_37_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_2356_cast_fp16 = mul(x = k_35_cast_fp16, y = k_35_cast_fp16)[name = string("op_2356_cast_fp16")];
            tensor<int32, [1]> var_2361_axes_0 = const()[name = string("op_2361_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2361_keep_dims_0 = const()[name = string("op_2361_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 2, 1]> var_2361_cast_fp16 = reduce_mean(axes = var_2361_axes_0, keep_dims = var_2361_keep_dims_0, x = var_2356_cast_fp16)[name = string("op_2361_cast_fp16")];
            fp16 var_2362_promoted_to_fp16 = const()[name = string("op_2362_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 64, 2, 1]> var_2363_cast_fp16 = mul(x = var_2361_cast_fp16, y = var_2362_promoted_to_fp16)[name = string("op_2363_cast_fp16")];
            fp16 var_2365_to_fp16 = const()[name = string("op_2365_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 64, 2, 1]> var_2366_cast_fp16 = add(x = var_2363_cast_fp16, y = var_2365_to_fp16)[name = string("op_2366_cast_fp16")];
            fp32 var_2367_epsilon_0 = const()[name = string("op_2367_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 2, 1]> var_2367_cast_fp16 = rsqrt(epsilon = var_2367_epsilon_0, x = var_2366_cast_fp16)[name = string("op_2367_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> k_37_cast_fp16 = mul(x = k_35_cast_fp16, y = var_2367_cast_fp16)[name = string("k_37_cast_fp16")];
            tensor<int32, [4]> var_2373 = const()[name = string("op_2373"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2386_axes_0 = const()[name = string("op_2386_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 128]> var_2374_cast_fp16 = transpose(perm = var_2373, x = q_37_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 2, 1, 64, 128]> var_2386_cast_fp16 = expand_dims(axes = var_2386_axes_0, x = var_2374_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<int32, [5]> var_2394_reps_0 = const()[name = string("op_2394_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 64, 128]> var_2394_cast_fp16 = tile(reps = var_2394_reps_0, x = var_2386_cast_fp16)[name = string("op_2394_cast_fp16")];
            tensor<int32, [4]> var_2399 = const()[name = string("op_2399"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> q_39_cast_fp16 = reshape(shape = var_2399, x = var_2394_cast_fp16)[name = string("q_39_cast_fp16")];
            tensor<int32, [4]> var_2405 = const()[name = string("op_2405"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2418_axes_0 = const()[name = string("op_2418_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 128]> var_2406_cast_fp16 = transpose(perm = var_2405, x = k_37_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 2, 1, 64, 128]> var_2418_cast_fp16 = expand_dims(axes = var_2418_axes_0, x = var_2406_cast_fp16)[name = string("op_2418_cast_fp16")];
            tensor<int32, [5]> var_2426_reps_0 = const()[name = string("op_2426_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 64, 128]> var_2426_cast_fp16 = tile(reps = var_2426_reps_0, x = var_2418_cast_fp16)[name = string("op_2426_cast_fp16")];
            tensor<int32, [4]> var_2431 = const()[name = string("op_2431"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> k_39_cast_fp16 = reshape(shape = var_2431, x = var_2426_cast_fp16)[name = string("k_39_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_2433_cast_fp16 = sigmoid(x = b_5_cast_fp16)[name = string("op_2433_cast_fp16")];
            tensor<int32, [3]> var_2437 = const()[name = string("op_2437"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_2_dt_bias_to_fp16 = const()[name = string("layers_2_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 64, 4]> input_37_cast_fp16 = add(x = a_5_cast_fp16, y = layers_2_dt_bias_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_2441_cast_fp16 = relu(x = input_37_cast_fp16)[name = string("op_2441_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_2442_cast_fp16 = abs(x = input_37_cast_fp16)[name = string("op_2442_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 64, 4]> var_2443_cast_fp16 = mul(x = var_2442_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_2443_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_2444_cast_fp16 = exp(x = var_2443_cast_fp16)[name = string("op_2444_cast_fp16")];
            fp16 var_2446_to_fp16 = const()[name = string("op_2446_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 64, 4]> var_2447_cast_fp16 = add(x = var_2444_cast_fp16, y = var_2446_to_fp16)[name = string("op_2447_cast_fp16")];
            fp32 var_2448_epsilon_0 = const()[name = string("op_2448_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 64, 4]> var_2448_cast_fp16 = log(epsilon = var_2448_epsilon_0, x = var_2447_cast_fp16)[name = string("op_2448_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_2450_cast_fp16 = add(x = var_2441_cast_fp16, y = var_2448_cast_fp16)[name = string("op_2450_cast_fp16")];
            tensor<fp16, [4]> layers_2_negA_to_fp16 = const()[name = string("layers_2_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.c0cp+3, -0x1.3e4p+3, -0x1.568p+1, -0x1.9acp+2])];
            tensor<fp16, [1, 64, 4]> var_2451_cast_fp16 = mul(x = layers_2_negA_to_fp16, y = var_2450_cast_fp16)[name = string("op_2451_cast_fp16")];
            tensor<int32, [3]> var_2455 = const()[name = string("op_2455"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_2458_axes_0 = const()[name = string("op_2458_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 64]> beta_5_cast_fp16 = transpose(perm = var_2437, x = var_2433_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 4, 64, 1]> var_2458_cast_fp16 = expand_dims(axes = var_2458_axes_0, x = beta_5_cast_fp16)[name = string("op_2458_cast_fp16")];
            fp16 var_2459_to_fp16 = const()[name = string("op_2459_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 64, 1]> var_2460_cast_fp16 = mul(x = var_2458_cast_fp16, y = var_2459_to_fp16)[name = string("op_2460_cast_fp16")];
            tensor<fp16, [1, 4, 64, 128]> v_5_cast_fp16 = transpose(perm = var_2335, x = var_2330_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [1, 4, 64, 128]> vb_21_cast_fp16 = mul(x = v_5_cast_fp16, y = var_2460_cast_fp16)[name = string("vb_21_cast_fp16")];
            tensor<int32, [4]> q_41_begin_0 = const()[name = string("q_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_41_end_0 = const()[name = string("q_41_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_41_end_mask_0 = const()[name = string("q_41_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_41_cast_fp16 = slice_by_index(begin = q_41_begin_0, end = q_41_end_0, end_mask = q_41_end_mask_0, x = q_39_cast_fp16)[name = string("q_41_cast_fp16")];
            tensor<int32, [4]> k_41_begin_0 = const()[name = string("k_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_41_end_0 = const()[name = string("k_41_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_41_end_mask_0 = const()[name = string("k_41_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, x = k_39_cast_fp16)[name = string("k_41_cast_fp16")];
            tensor<fp16, [1, 4, 64, 128]> var_2494_cast_fp16 = mul(x = k_39_cast_fp16, y = var_2458_cast_fp16)[name = string("op_2494_cast_fp16")];
            tensor<int32, [4]> kb_17_begin_0 = const()[name = string("kb_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_17_end_0 = const()[name = string("kb_17_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_17_end_mask_0 = const()[name = string("kb_17_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_17_cast_fp16 = slice_by_index(begin = kb_17_begin_0, end = kb_17_end_0, end_mask = kb_17_end_mask_0, x = var_2494_cast_fp16)[name = string("kb_17_cast_fp16")];
            tensor<int32, [4]> vb_23_begin_0 = const()[name = string("vb_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_23_end_0 = const()[name = string("vb_23_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_23_end_mask_0 = const()[name = string("vb_23_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_23_cast_fp16 = slice_by_index(begin = vb_23_begin_0, end = vb_23_end_0, end_mask = vb_23_end_mask_0, x = vb_21_cast_fp16)[name = string("vb_23_cast_fp16")];
            tensor<int32, [3]> gdec_23_begin_0 = const()[name = string("gdec_23_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_23_end_0 = const()[name = string("gdec_23_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_23_end_mask_0 = const()[name = string("gdec_23_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 64]> gdec_21_cast_fp16 = transpose(perm = var_2455, x = var_2451_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 4, 16]> gdec_23_cast_fp16 = slice_by_index(begin = gdec_23_begin_0, end = gdec_23_end_0, end_mask = gdec_23_end_mask_0, x = gdec_21_cast_fp16)[name = string("gdec_23_cast_fp16")];
            tensor<int32, [2]> var_2542 = const()[name = string("op_2542"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_2543_cast_fp16 = reshape(shape = var_2542, x = gdec_23_cast_fp16)[name = string("op_2543_cast_fp16")];
            tensor<fp16, [4, 256]> var_2544_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2543_cast_fp16)[name = string("op_2544_cast_fp16")];
            tensor<int32, [4]> var_2549 = const()[name = string("op_2549"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_17_cast_fp16 = reshape(shape = var_2549, x = var_2544_cast_fp16)[name = string("P_17_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2551_cast_fp16 = exp(x = P_17_cast_fp16)[name = string("op_2551_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_17_cast_fp16 = mul(x = var_2551_cast_fp16, y = tril_incl_to_fp16)[name = string("D_17_cast_fp16")];
            bool kk_17_transpose_x_1 = const()[name = string("kk_17_transpose_x_1"), val = bool(false)];
            bool kk_17_transpose_y_1 = const()[name = string("kk_17_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_17_cast_fp16 = matmul(transpose_x = kk_17_transpose_x_1, transpose_y = kk_17_transpose_y_1, x = kb_17_cast_fp16, y = k_41_cast_fp16)[name = string("kk_17_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_2557_cast_fp16 = mul(x = kk_17_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_2557_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2558_cast_fp16 = mul(x = var_2557_cast_fp16, y = D_17_cast_fp16)[name = string("op_2558_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_17_cast_fp16 = mul(x = var_2558_cast_fp16, y = cum_tail_to_fp16)[name = string("N_17_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2560_cast_fp16 = mul(x = D_17_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2560_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_17_cast_fp16 = mul(x = kk_17_cast_fp16, y = var_2560_cast_fp16)[name = string("Na_17_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_65_cast_fp16 = add(x = eye_to_fp16, y = N_17_cast_fp16)[name = string("Minv_65_cast_fp16")];
            bool Np_49_transpose_x_0 = const()[name = string("Np_49_transpose_x_0"), val = bool(false)];
            bool Np_49_transpose_y_0 = const()[name = string("Np_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_49_cast_fp16 = matmul(transpose_x = Np_49_transpose_x_0, transpose_y = Np_49_transpose_y_0, x = N_17_cast_fp16, y = Na_17_cast_fp16)[name = string("Np_49_cast_fp16")];
            bool Npa_33_transpose_x_0 = const()[name = string("Npa_33_transpose_x_0"), val = bool(false)];
            bool Npa_33_transpose_y_0 = const()[name = string("Npa_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_33_cast_fp16 = matmul(transpose_x = Npa_33_transpose_x_0, transpose_y = Npa_33_transpose_y_0, x = Na_17_cast_fp16, y = N_17_cast_fp16)[name = string("Npa_33_cast_fp16")];
            bool var_2566_transpose_x_0 = const()[name = string("op_2566_transpose_x_0"), val = bool(false)];
            bool var_2566_transpose_y_0 = const()[name = string("op_2566_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2566_cast_fp16 = matmul(transpose_x = var_2566_transpose_x_0, transpose_y = var_2566_transpose_y_0, x = Minv_65_cast_fp16, y = Np_49_cast_fp16)[name = string("op_2566_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_67_cast_fp16 = add(x = Minv_65_cast_fp16, y = var_2566_cast_fp16)[name = string("Minv_67_cast_fp16")];
            bool Np_51_transpose_x_0 = const()[name = string("Np_51_transpose_x_0"), val = bool(false)];
            bool Np_51_transpose_y_0 = const()[name = string("Np_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_51_cast_fp16 = matmul(transpose_x = Np_51_transpose_x_0, transpose_y = Np_51_transpose_y_0, x = Np_49_cast_fp16, y = Npa_33_cast_fp16)[name = string("Np_51_cast_fp16")];
            bool Npa_35_transpose_x_0 = const()[name = string("Npa_35_transpose_x_0"), val = bool(false)];
            bool Npa_35_transpose_y_0 = const()[name = string("Npa_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_35_cast_fp16 = matmul(transpose_x = Npa_35_transpose_x_0, transpose_y = Npa_35_transpose_y_0, x = Npa_33_cast_fp16, y = Np_49_cast_fp16)[name = string("Npa_35_cast_fp16")];
            bool var_2571_transpose_x_0 = const()[name = string("op_2571_transpose_x_0"), val = bool(false)];
            bool var_2571_transpose_y_0 = const()[name = string("op_2571_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2571_cast_fp16 = matmul(transpose_x = var_2571_transpose_x_0, transpose_y = var_2571_transpose_y_0, x = Minv_67_cast_fp16, y = Np_51_cast_fp16)[name = string("op_2571_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_69_cast_fp16 = add(x = Minv_67_cast_fp16, y = var_2571_cast_fp16)[name = string("Minv_69_cast_fp16")];
            bool Np_53_transpose_x_0 = const()[name = string("Np_53_transpose_x_0"), val = bool(false)];
            bool Np_53_transpose_y_0 = const()[name = string("Np_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_53_cast_fp16 = matmul(transpose_x = Np_53_transpose_x_0, transpose_y = Np_53_transpose_y_0, x = Np_51_cast_fp16, y = Npa_35_cast_fp16)[name = string("Np_53_cast_fp16")];
            bool var_2575_transpose_x_0 = const()[name = string("op_2575_transpose_x_0"), val = bool(false)];
            bool var_2575_transpose_y_0 = const()[name = string("op_2575_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2575_cast_fp16 = matmul(transpose_x = var_2575_transpose_x_0, transpose_y = var_2575_transpose_y_0, x = Minv_69_cast_fp16, y = Np_53_cast_fp16)[name = string("op_2575_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_71_cast_fp16 = add(x = Minv_69_cast_fp16, y = var_2575_cast_fp16)[name = string("Minv_71_cast_fp16")];
            bool u_17_transpose_x_0 = const()[name = string("u_17_transpose_x_0"), val = bool(false)];
            bool u_17_transpose_y_0 = const()[name = string("u_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_17_cast_fp16 = matmul(transpose_x = u_17_transpose_x_0, transpose_y = u_17_transpose_y_0, x = Minv_71_cast_fp16, y = vb_23_cast_fp16)[name = string("u_17_cast_fp16")];
            bool var_2582_transpose_x_1 = const()[name = string("op_2582_transpose_x_1"), val = bool(false)];
            bool var_2582_transpose_y_1 = const()[name = string("op_2582_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_2582_cast_fp16 = matmul(transpose_x = var_2582_transpose_x_1, transpose_y = var_2582_transpose_y_1, x = q_41_cast_fp16, y = k_41_cast_fp16)[name = string("op_2582_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_17_cast_fp16 = mul(x = var_2582_cast_fp16, y = D_17_cast_fp16)[name = string("A_17_cast_fp16")];
            bool var_2584_transpose_x_0 = const()[name = string("op_2584_transpose_x_0"), val = bool(false)];
            bool var_2584_transpose_y_0 = const()[name = string("op_2584_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2584_cast_fp16 = matmul(transpose_x = var_2584_transpose_x_0, transpose_y = var_2584_transpose_y_0, x = A_17_cast_fp16, y = u_17_cast_fp16)[name = string("op_2584_cast_fp16")];
            tensor<fp16, [16]> var_2585_bias_0_to_fp16 = const()[name = string("op_2585_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_2585_cast_fp16 = linear(bias = var_2585_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_23_cast_fp16)[name = string("op_2585_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_13_cast_fp16 = exp(x = var_2585_cast_fp16)[name = string("tail_13_cast_fp16")];
            tensor<int32, [1]> var_2588_axes_0 = const()[name = string("op_2588_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_2588_cast_fp16 = expand_dims(axes = var_2588_axes_0, x = tail_13_cast_fp16)[name = string("op_2588_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2589_cast_fp16 = mul(x = k_41_cast_fp16, y = var_2588_cast_fp16)[name = string("op_2589_cast_fp16")];
            bool S_13_transpose_x_1 = const()[name = string("S_13_transpose_x_1"), val = bool(true)];
            bool S_13_transpose_y_1 = const()[name = string("S_13_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_13_cast_fp16 = matmul(transpose_x = S_13_transpose_x_1, transpose_y = S_13_transpose_y_1, x = var_2589_cast_fp16, y = u_17_cast_fp16)[name = string("S_13_cast_fp16")];
            tensor<int32, [4]> q_43_begin_0 = const()[name = string("q_43_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_43_end_0 = const()[name = string("q_43_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> q_43_end_mask_0 = const()[name = string("q_43_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, x = q_39_cast_fp16)[name = string("q_43_cast_fp16")];
            tensor<int32, [4]> k_43_begin_0 = const()[name = string("k_43_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_43_end_0 = const()[name = string("k_43_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> k_43_end_mask_0 = const()[name = string("k_43_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_43_cast_fp16 = slice_by_index(begin = k_43_begin_0, end = k_43_end_0, end_mask = k_43_end_mask_0, x = k_39_cast_fp16)[name = string("k_43_cast_fp16")];
            tensor<int32, [4]> kb_19_begin_0 = const()[name = string("kb_19_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_19_end_0 = const()[name = string("kb_19_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> kb_19_end_mask_0 = const()[name = string("kb_19_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_19_cast_fp16 = slice_by_index(begin = kb_19_begin_0, end = kb_19_end_0, end_mask = kb_19_end_mask_0, x = var_2494_cast_fp16)[name = string("kb_19_cast_fp16")];
            tensor<int32, [4]> vb_25_begin_0 = const()[name = string("vb_25_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_25_end_0 = const()[name = string("vb_25_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> vb_25_end_mask_0 = const()[name = string("vb_25_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_25_cast_fp16 = slice_by_index(begin = vb_25_begin_0, end = vb_25_end_0, end_mask = vb_25_end_mask_0, x = vb_21_cast_fp16)[name = string("vb_25_cast_fp16")];
            tensor<int32, [3]> gdec_25_begin_0 = const()[name = string("gdec_25_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_25_end_0 = const()[name = string("gdec_25_end_0"), val = tensor<int32, [3]>([1, 4, 32])];
            tensor<bool, [3]> gdec_25_end_mask_0 = const()[name = string("gdec_25_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_25_cast_fp16 = slice_by_index(begin = gdec_25_begin_0, end = gdec_25_end_0, end_mask = gdec_25_end_mask_0, x = gdec_21_cast_fp16)[name = string("gdec_25_cast_fp16")];
            tensor<int32, [2]> var_2674 = const()[name = string("op_2674"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_2675_cast_fp16 = reshape(shape = var_2674, x = gdec_25_cast_fp16)[name = string("op_2675_cast_fp16")];
            tensor<fp16, [4, 256]> var_2676_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2675_cast_fp16)[name = string("op_2676_cast_fp16")];
            tensor<int32, [4]> var_2681 = const()[name = string("op_2681"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_19_cast_fp16 = reshape(shape = var_2681, x = var_2676_cast_fp16)[name = string("P_19_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2683_cast_fp16 = exp(x = P_19_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_19_cast_fp16 = mul(x = var_2683_cast_fp16, y = tril_incl_to_fp16)[name = string("D_19_cast_fp16")];
            bool kk_19_transpose_x_1 = const()[name = string("kk_19_transpose_x_1"), val = bool(false)];
            bool kk_19_transpose_y_1 = const()[name = string("kk_19_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_19_cast_fp16 = matmul(transpose_x = kk_19_transpose_x_1, transpose_y = kk_19_transpose_y_1, x = kb_19_cast_fp16, y = k_43_cast_fp16)[name = string("kk_19_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_2689_cast_fp16 = mul(x = kk_19_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_2689_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2690_cast_fp16 = mul(x = var_2689_cast_fp16, y = D_19_cast_fp16)[name = string("op_2690_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_19_cast_fp16 = mul(x = var_2690_cast_fp16, y = cum_tail_to_fp16)[name = string("N_19_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2692_cast_fp16 = mul(x = D_19_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2692_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_19_cast_fp16 = mul(x = kk_19_cast_fp16, y = var_2692_cast_fp16)[name = string("Na_19_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_73_cast_fp16 = add(x = eye_to_fp16, y = N_19_cast_fp16)[name = string("Minv_73_cast_fp16")];
            bool Np_55_transpose_x_0 = const()[name = string("Np_55_transpose_x_0"), val = bool(false)];
            bool Np_55_transpose_y_0 = const()[name = string("Np_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_55_cast_fp16 = matmul(transpose_x = Np_55_transpose_x_0, transpose_y = Np_55_transpose_y_0, x = N_19_cast_fp16, y = Na_19_cast_fp16)[name = string("Np_55_cast_fp16")];
            bool Npa_37_transpose_x_0 = const()[name = string("Npa_37_transpose_x_0"), val = bool(false)];
            bool Npa_37_transpose_y_0 = const()[name = string("Npa_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_37_cast_fp16 = matmul(transpose_x = Npa_37_transpose_x_0, transpose_y = Npa_37_transpose_y_0, x = Na_19_cast_fp16, y = N_19_cast_fp16)[name = string("Npa_37_cast_fp16")];
            bool var_2698_transpose_x_0 = const()[name = string("op_2698_transpose_x_0"), val = bool(false)];
            bool var_2698_transpose_y_0 = const()[name = string("op_2698_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2698_cast_fp16 = matmul(transpose_x = var_2698_transpose_x_0, transpose_y = var_2698_transpose_y_0, x = Minv_73_cast_fp16, y = Np_55_cast_fp16)[name = string("op_2698_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_75_cast_fp16 = add(x = Minv_73_cast_fp16, y = var_2698_cast_fp16)[name = string("Minv_75_cast_fp16")];
            bool Np_57_transpose_x_0 = const()[name = string("Np_57_transpose_x_0"), val = bool(false)];
            bool Np_57_transpose_y_0 = const()[name = string("Np_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_57_cast_fp16 = matmul(transpose_x = Np_57_transpose_x_0, transpose_y = Np_57_transpose_y_0, x = Np_55_cast_fp16, y = Npa_37_cast_fp16)[name = string("Np_57_cast_fp16")];
            bool Npa_39_transpose_x_0 = const()[name = string("Npa_39_transpose_x_0"), val = bool(false)];
            bool Npa_39_transpose_y_0 = const()[name = string("Npa_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_39_cast_fp16 = matmul(transpose_x = Npa_39_transpose_x_0, transpose_y = Npa_39_transpose_y_0, x = Npa_37_cast_fp16, y = Np_55_cast_fp16)[name = string("Npa_39_cast_fp16")];
            bool var_2703_transpose_x_0 = const()[name = string("op_2703_transpose_x_0"), val = bool(false)];
            bool var_2703_transpose_y_0 = const()[name = string("op_2703_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2703_cast_fp16 = matmul(transpose_x = var_2703_transpose_x_0, transpose_y = var_2703_transpose_y_0, x = Minv_75_cast_fp16, y = Np_57_cast_fp16)[name = string("op_2703_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_77_cast_fp16 = add(x = Minv_75_cast_fp16, y = var_2703_cast_fp16)[name = string("Minv_77_cast_fp16")];
            bool Np_59_transpose_x_0 = const()[name = string("Np_59_transpose_x_0"), val = bool(false)];
            bool Np_59_transpose_y_0 = const()[name = string("Np_59_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_59_cast_fp16 = matmul(transpose_x = Np_59_transpose_x_0, transpose_y = Np_59_transpose_y_0, x = Np_57_cast_fp16, y = Npa_39_cast_fp16)[name = string("Np_59_cast_fp16")];
            bool var_2707_transpose_x_0 = const()[name = string("op_2707_transpose_x_0"), val = bool(false)];
            bool var_2707_transpose_y_0 = const()[name = string("op_2707_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2707_cast_fp16 = matmul(transpose_x = var_2707_transpose_x_0, transpose_y = var_2707_transpose_y_0, x = Minv_77_cast_fp16, y = Np_59_cast_fp16)[name = string("op_2707_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_79_cast_fp16 = add(x = Minv_77_cast_fp16, y = var_2707_cast_fp16)[name = string("Minv_79_cast_fp16")];
            bool u_19_transpose_x_0 = const()[name = string("u_19_transpose_x_0"), val = bool(false)];
            bool u_19_transpose_y_0 = const()[name = string("u_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_19_cast_fp16 = matmul(transpose_x = u_19_transpose_x_0, transpose_y = u_19_transpose_y_0, x = Minv_79_cast_fp16, y = vb_25_cast_fp16)[name = string("u_19_cast_fp16")];
            bool var_2714_transpose_x_1 = const()[name = string("op_2714_transpose_x_1"), val = bool(false)];
            bool var_2714_transpose_y_1 = const()[name = string("op_2714_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_2714_cast_fp16 = matmul(transpose_x = var_2714_transpose_x_1, transpose_y = var_2714_transpose_y_1, x = q_43_cast_fp16, y = k_43_cast_fp16)[name = string("op_2714_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_19_cast_fp16 = mul(x = var_2714_cast_fp16, y = D_19_cast_fp16)[name = string("A_19_cast_fp16")];
            tensor<fp16, [16]> var_2716_bias_0_to_fp16 = const()[name = string("op_2716_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_2716_cast_fp16 = linear(bias = var_2716_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_25_cast_fp16)[name = string("op_2716_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_13_cast_fp16 = exp(x = var_2716_cast_fp16)[name = string("cum_13_cast_fp16")];
            tensor<int32, [1]> var_2719_axes_0 = const()[name = string("op_2719_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_2719_cast_fp16 = expand_dims(axes = var_2719_axes_0, x = cum_13_cast_fp16)[name = string("op_2719_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2720_cast_fp16 = mul(x = kb_19_cast_fp16, y = var_2719_cast_fp16)[name = string("op_2720_cast_fp16")];
            bool w_19_transpose_x_0 = const()[name = string("w_19_transpose_x_0"), val = bool(false)];
            bool w_19_transpose_y_0 = const()[name = string("w_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_19_cast_fp16 = matmul(transpose_x = w_19_transpose_x_0, transpose_y = w_19_transpose_y_0, x = Minv_79_cast_fp16, y = var_2720_cast_fp16)[name = string("w_19_cast_fp16")];
            bool var_2722_transpose_x_0 = const()[name = string("op_2722_transpose_x_0"), val = bool(false)];
            bool var_2722_transpose_y_0 = const()[name = string("op_2722_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2722_cast_fp16 = matmul(transpose_x = var_2722_transpose_x_0, transpose_y = var_2722_transpose_y_0, x = w_19_cast_fp16, y = S_13_cast_fp16)[name = string("op_2722_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_13_cast_fp16 = sub(x = u_19_cast_fp16, y = var_2722_cast_fp16)[name = string("vnew_13_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2727_cast_fp16 = mul(x = q_43_cast_fp16, y = var_2719_cast_fp16)[name = string("op_2727_cast_fp16")];
            bool var_2728_transpose_x_0 = const()[name = string("op_2728_transpose_x_0"), val = bool(false)];
            bool var_2728_transpose_y_0 = const()[name = string("op_2728_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2728_cast_fp16 = matmul(transpose_x = var_2728_transpose_x_0, transpose_y = var_2728_transpose_y_0, x = var_2727_cast_fp16, y = S_13_cast_fp16)[name = string("op_2728_cast_fp16")];
            bool var_2729_transpose_x_0 = const()[name = string("op_2729_transpose_x_0"), val = bool(false)];
            bool var_2729_transpose_y_0 = const()[name = string("op_2729_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2729_cast_fp16 = matmul(transpose_x = var_2729_transpose_x_0, transpose_y = var_2729_transpose_y_0, x = A_19_cast_fp16, y = vnew_13_cast_fp16)[name = string("op_2729_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2731_cast_fp16 = add(x = var_2728_cast_fp16, y = var_2729_cast_fp16)[name = string("op_2731_cast_fp16")];
            tensor<fp16, [16]> var_2732_bias_0_to_fp16 = const()[name = string("op_2732_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_2732_cast_fp16 = linear(bias = var_2732_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_25_cast_fp16)[name = string("op_2732_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_15_cast_fp16 = exp(x = var_2732_cast_fp16)[name = string("tail_15_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_29_to_fp16 = const()[name = string("transpose_29_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_2734_bias_0_to_fp16 = const()[name = string("op_2734_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_2734_cast_fp16 = linear(bias = var_2734_bias_0_to_fp16, weight = transpose_29_to_fp16, x = gdec_25_cast_fp16)[name = string("op_2734_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_2735_cast_fp16 = exp(x = var_2734_cast_fp16)[name = string("op_2735_cast_fp16")];
            tensor<int32, [1]> last_11_axes_0 = const()[name = string("last_11_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_11_cast_fp16 = expand_dims(axes = last_11_axes_0, x = var_2735_cast_fp16)[name = string("last_11_cast_fp16")];
            tensor<int32, [1]> var_2739_axes_0 = const()[name = string("op_2739_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_2739_cast_fp16 = expand_dims(axes = var_2739_axes_0, x = tail_15_cast_fp16)[name = string("op_2739_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2740_cast_fp16 = mul(x = k_43_cast_fp16, y = var_2739_cast_fp16)[name = string("op_2740_cast_fp16")];
            bool upd_9_transpose_x_1 = const()[name = string("upd_9_transpose_x_1"), val = bool(true)];
            bool upd_9_transpose_y_1 = const()[name = string("upd_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_9_cast_fp16 = matmul(transpose_x = upd_9_transpose_x_1, transpose_y = upd_9_transpose_y_1, x = var_2740_cast_fp16, y = vnew_13_cast_fp16)[name = string("upd_9_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_2745_cast_fp16 = mul(x = S_13_cast_fp16, y = last_11_cast_fp16)[name = string("op_2745_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_15_cast_fp16 = add(x = var_2745_cast_fp16, y = upd_9_cast_fp16)[name = string("S_15_cast_fp16")];
            tensor<int32, [4]> q_45_begin_0 = const()[name = string("q_45_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_45_end_0 = const()[name = string("q_45_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> q_45_end_mask_0 = const()[name = string("q_45_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_45_cast_fp16 = slice_by_index(begin = q_45_begin_0, end = q_45_end_0, end_mask = q_45_end_mask_0, x = q_39_cast_fp16)[name = string("q_45_cast_fp16")];
            tensor<int32, [4]> k_45_begin_0 = const()[name = string("k_45_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_45_end_0 = const()[name = string("k_45_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> k_45_end_mask_0 = const()[name = string("k_45_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, x = k_39_cast_fp16)[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> kb_21_begin_0 = const()[name = string("kb_21_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_21_end_0 = const()[name = string("kb_21_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> kb_21_end_mask_0 = const()[name = string("kb_21_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_21_cast_fp16 = slice_by_index(begin = kb_21_begin_0, end = kb_21_end_0, end_mask = kb_21_end_mask_0, x = var_2494_cast_fp16)[name = string("kb_21_cast_fp16")];
            tensor<int32, [4]> vb_27_begin_0 = const()[name = string("vb_27_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_27_end_0 = const()[name = string("vb_27_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> vb_27_end_mask_0 = const()[name = string("vb_27_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_27_cast_fp16 = slice_by_index(begin = vb_27_begin_0, end = vb_27_end_0, end_mask = vb_27_end_mask_0, x = vb_21_cast_fp16)[name = string("vb_27_cast_fp16")];
            tensor<int32, [3]> gdec_27_begin_0 = const()[name = string("gdec_27_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_27_end_0 = const()[name = string("gdec_27_end_0"), val = tensor<int32, [3]>([1, 4, 48])];
            tensor<bool, [3]> gdec_27_end_mask_0 = const()[name = string("gdec_27_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_27_cast_fp16 = slice_by_index(begin = gdec_27_begin_0, end = gdec_27_end_0, end_mask = gdec_27_end_mask_0, x = gdec_21_cast_fp16)[name = string("gdec_27_cast_fp16")];
            tensor<int32, [2]> var_2828 = const()[name = string("op_2828"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_2829_cast_fp16 = reshape(shape = var_2828, x = gdec_27_cast_fp16)[name = string("op_2829_cast_fp16")];
            tensor<fp16, [4, 256]> var_2830_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2829_cast_fp16)[name = string("op_2830_cast_fp16")];
            tensor<int32, [4]> var_2835 = const()[name = string("op_2835"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_21_cast_fp16 = reshape(shape = var_2835, x = var_2830_cast_fp16)[name = string("P_21_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2837_cast_fp16 = exp(x = P_21_cast_fp16)[name = string("op_2837_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_21_cast_fp16 = mul(x = var_2837_cast_fp16, y = tril_incl_to_fp16)[name = string("D_21_cast_fp16")];
            bool kk_21_transpose_x_1 = const()[name = string("kk_21_transpose_x_1"), val = bool(false)];
            bool kk_21_transpose_y_1 = const()[name = string("kk_21_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_21_cast_fp16 = matmul(transpose_x = kk_21_transpose_x_1, transpose_y = kk_21_transpose_y_1, x = kb_21_cast_fp16, y = k_45_cast_fp16)[name = string("kk_21_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_2843_cast_fp16 = mul(x = kk_21_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_2843_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2844_cast_fp16 = mul(x = var_2843_cast_fp16, y = D_21_cast_fp16)[name = string("op_2844_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_21_cast_fp16 = mul(x = var_2844_cast_fp16, y = cum_tail_to_fp16)[name = string("N_21_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2846_cast_fp16 = mul(x = D_21_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2846_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_21_cast_fp16 = mul(x = kk_21_cast_fp16, y = var_2846_cast_fp16)[name = string("Na_21_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_81_cast_fp16 = add(x = eye_to_fp16, y = N_21_cast_fp16)[name = string("Minv_81_cast_fp16")];
            bool Np_61_transpose_x_0 = const()[name = string("Np_61_transpose_x_0"), val = bool(false)];
            bool Np_61_transpose_y_0 = const()[name = string("Np_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_61_cast_fp16 = matmul(transpose_x = Np_61_transpose_x_0, transpose_y = Np_61_transpose_y_0, x = N_21_cast_fp16, y = Na_21_cast_fp16)[name = string("Np_61_cast_fp16")];
            bool Npa_41_transpose_x_0 = const()[name = string("Npa_41_transpose_x_0"), val = bool(false)];
            bool Npa_41_transpose_y_0 = const()[name = string("Npa_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_41_cast_fp16 = matmul(transpose_x = Npa_41_transpose_x_0, transpose_y = Npa_41_transpose_y_0, x = Na_21_cast_fp16, y = N_21_cast_fp16)[name = string("Npa_41_cast_fp16")];
            bool var_2852_transpose_x_0 = const()[name = string("op_2852_transpose_x_0"), val = bool(false)];
            bool var_2852_transpose_y_0 = const()[name = string("op_2852_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2852_cast_fp16 = matmul(transpose_x = var_2852_transpose_x_0, transpose_y = var_2852_transpose_y_0, x = Minv_81_cast_fp16, y = Np_61_cast_fp16)[name = string("op_2852_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_83_cast_fp16 = add(x = Minv_81_cast_fp16, y = var_2852_cast_fp16)[name = string("Minv_83_cast_fp16")];
            bool Np_63_transpose_x_0 = const()[name = string("Np_63_transpose_x_0"), val = bool(false)];
            bool Np_63_transpose_y_0 = const()[name = string("Np_63_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_63_cast_fp16 = matmul(transpose_x = Np_63_transpose_x_0, transpose_y = Np_63_transpose_y_0, x = Np_61_cast_fp16, y = Npa_41_cast_fp16)[name = string("Np_63_cast_fp16")];
            bool Npa_43_transpose_x_0 = const()[name = string("Npa_43_transpose_x_0"), val = bool(false)];
            bool Npa_43_transpose_y_0 = const()[name = string("Npa_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_43_cast_fp16 = matmul(transpose_x = Npa_43_transpose_x_0, transpose_y = Npa_43_transpose_y_0, x = Npa_41_cast_fp16, y = Np_61_cast_fp16)[name = string("Npa_43_cast_fp16")];
            bool var_2857_transpose_x_0 = const()[name = string("op_2857_transpose_x_0"), val = bool(false)];
            bool var_2857_transpose_y_0 = const()[name = string("op_2857_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2857_cast_fp16 = matmul(transpose_x = var_2857_transpose_x_0, transpose_y = var_2857_transpose_y_0, x = Minv_83_cast_fp16, y = Np_63_cast_fp16)[name = string("op_2857_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_85_cast_fp16 = add(x = Minv_83_cast_fp16, y = var_2857_cast_fp16)[name = string("Minv_85_cast_fp16")];
            bool Np_65_transpose_x_0 = const()[name = string("Np_65_transpose_x_0"), val = bool(false)];
            bool Np_65_transpose_y_0 = const()[name = string("Np_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_65_cast_fp16 = matmul(transpose_x = Np_65_transpose_x_0, transpose_y = Np_65_transpose_y_0, x = Np_63_cast_fp16, y = Npa_43_cast_fp16)[name = string("Np_65_cast_fp16")];
            bool var_2861_transpose_x_0 = const()[name = string("op_2861_transpose_x_0"), val = bool(false)];
            bool var_2861_transpose_y_0 = const()[name = string("op_2861_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2861_cast_fp16 = matmul(transpose_x = var_2861_transpose_x_0, transpose_y = var_2861_transpose_y_0, x = Minv_85_cast_fp16, y = Np_65_cast_fp16)[name = string("op_2861_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_87_cast_fp16 = add(x = Minv_85_cast_fp16, y = var_2861_cast_fp16)[name = string("Minv_87_cast_fp16")];
            bool u_21_transpose_x_0 = const()[name = string("u_21_transpose_x_0"), val = bool(false)];
            bool u_21_transpose_y_0 = const()[name = string("u_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_21_cast_fp16 = matmul(transpose_x = u_21_transpose_x_0, transpose_y = u_21_transpose_y_0, x = Minv_87_cast_fp16, y = vb_27_cast_fp16)[name = string("u_21_cast_fp16")];
            bool var_2868_transpose_x_1 = const()[name = string("op_2868_transpose_x_1"), val = bool(false)];
            bool var_2868_transpose_y_1 = const()[name = string("op_2868_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_2868_cast_fp16 = matmul(transpose_x = var_2868_transpose_x_1, transpose_y = var_2868_transpose_y_1, x = q_45_cast_fp16, y = k_45_cast_fp16)[name = string("op_2868_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_21_cast_fp16 = mul(x = var_2868_cast_fp16, y = D_21_cast_fp16)[name = string("A_21_cast_fp16")];
            tensor<fp16, [16]> var_2870_bias_0_to_fp16 = const()[name = string("op_2870_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_2870_cast_fp16 = linear(bias = var_2870_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_27_cast_fp16)[name = string("op_2870_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_15_cast_fp16 = exp(x = var_2870_cast_fp16)[name = string("cum_15_cast_fp16")];
            tensor<int32, [1]> var_2873_axes_0 = const()[name = string("op_2873_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_2873_cast_fp16 = expand_dims(axes = var_2873_axes_0, x = cum_15_cast_fp16)[name = string("op_2873_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2874_cast_fp16 = mul(x = kb_21_cast_fp16, y = var_2873_cast_fp16)[name = string("op_2874_cast_fp16")];
            bool w_21_transpose_x_0 = const()[name = string("w_21_transpose_x_0"), val = bool(false)];
            bool w_21_transpose_y_0 = const()[name = string("w_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_21_cast_fp16 = matmul(transpose_x = w_21_transpose_x_0, transpose_y = w_21_transpose_y_0, x = Minv_87_cast_fp16, y = var_2874_cast_fp16)[name = string("w_21_cast_fp16")];
            bool var_2876_transpose_x_0 = const()[name = string("op_2876_transpose_x_0"), val = bool(false)];
            bool var_2876_transpose_y_0 = const()[name = string("op_2876_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2876_cast_fp16 = matmul(transpose_x = var_2876_transpose_x_0, transpose_y = var_2876_transpose_y_0, x = w_21_cast_fp16, y = S_15_cast_fp16)[name = string("op_2876_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_15_cast_fp16 = sub(x = u_21_cast_fp16, y = var_2876_cast_fp16)[name = string("vnew_15_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2881_cast_fp16 = mul(x = q_45_cast_fp16, y = var_2873_cast_fp16)[name = string("op_2881_cast_fp16")];
            bool var_2882_transpose_x_0 = const()[name = string("op_2882_transpose_x_0"), val = bool(false)];
            bool var_2882_transpose_y_0 = const()[name = string("op_2882_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2882_cast_fp16 = matmul(transpose_x = var_2882_transpose_x_0, transpose_y = var_2882_transpose_y_0, x = var_2881_cast_fp16, y = S_15_cast_fp16)[name = string("op_2882_cast_fp16")];
            bool var_2883_transpose_x_0 = const()[name = string("op_2883_transpose_x_0"), val = bool(false)];
            bool var_2883_transpose_y_0 = const()[name = string("op_2883_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2883_cast_fp16 = matmul(transpose_x = var_2883_transpose_x_0, transpose_y = var_2883_transpose_y_0, x = A_21_cast_fp16, y = vnew_15_cast_fp16)[name = string("op_2883_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2885_cast_fp16 = add(x = var_2882_cast_fp16, y = var_2883_cast_fp16)[name = string("op_2885_cast_fp16")];
            tensor<fp16, [16]> var_2886_bias_0_to_fp16 = const()[name = string("op_2886_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_2886_cast_fp16 = linear(bias = var_2886_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_27_cast_fp16)[name = string("op_2886_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_17_cast_fp16 = exp(x = var_2886_cast_fp16)[name = string("tail_17_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_33_to_fp16 = const()[name = string("transpose_33_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_2888_bias_0_to_fp16 = const()[name = string("op_2888_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_2888_cast_fp16 = linear(bias = var_2888_bias_0_to_fp16, weight = transpose_33_to_fp16, x = gdec_27_cast_fp16)[name = string("op_2888_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_2889_cast_fp16 = exp(x = var_2888_cast_fp16)[name = string("op_2889_cast_fp16")];
            tensor<int32, [1]> last_13_axes_0 = const()[name = string("last_13_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_13_cast_fp16 = expand_dims(axes = last_13_axes_0, x = var_2889_cast_fp16)[name = string("last_13_cast_fp16")];
            tensor<int32, [1]> var_2893_axes_0 = const()[name = string("op_2893_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_2893_cast_fp16 = expand_dims(axes = var_2893_axes_0, x = tail_17_cast_fp16)[name = string("op_2893_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2894_cast_fp16 = mul(x = k_45_cast_fp16, y = var_2893_cast_fp16)[name = string("op_2894_cast_fp16")];
            bool upd_11_transpose_x_1 = const()[name = string("upd_11_transpose_x_1"), val = bool(true)];
            bool upd_11_transpose_y_1 = const()[name = string("upd_11_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_11_cast_fp16 = matmul(transpose_x = upd_11_transpose_x_1, transpose_y = upd_11_transpose_y_1, x = var_2894_cast_fp16, y = vnew_15_cast_fp16)[name = string("upd_11_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_2899_cast_fp16 = mul(x = S_15_cast_fp16, y = last_13_cast_fp16)[name = string("op_2899_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_17_cast_fp16 = add(x = var_2899_cast_fp16, y = upd_11_cast_fp16)[name = string("S_17_cast_fp16")];
            tensor<int32, [4]> q_47_begin_0 = const()[name = string("q_47_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_47_end_0 = const()[name = string("q_47_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_47_end_mask_0 = const()[name = string("q_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_47_cast_fp16 = slice_by_index(begin = q_47_begin_0, end = q_47_end_0, end_mask = q_47_end_mask_0, x = q_39_cast_fp16)[name = string("q_47_cast_fp16")];
            tensor<int32, [4]> k_47_begin_0 = const()[name = string("k_47_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_47_end_0 = const()[name = string("k_47_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_47_end_mask_0 = const()[name = string("k_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_47_cast_fp16 = slice_by_index(begin = k_47_begin_0, end = k_47_end_0, end_mask = k_47_end_mask_0, x = k_39_cast_fp16)[name = string("k_47_cast_fp16")];
            tensor<int32, [4]> kb_23_begin_0 = const()[name = string("kb_23_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_23_end_0 = const()[name = string("kb_23_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_23_end_mask_0 = const()[name = string("kb_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_23_cast_fp16 = slice_by_index(begin = kb_23_begin_0, end = kb_23_end_0, end_mask = kb_23_end_mask_0, x = var_2494_cast_fp16)[name = string("kb_23_cast_fp16")];
            tensor<int32, [4]> vb_29_begin_0 = const()[name = string("vb_29_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_29_end_0 = const()[name = string("vb_29_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_29_end_mask_0 = const()[name = string("vb_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_29_cast_fp16 = slice_by_index(begin = vb_29_begin_0, end = vb_29_end_0, end_mask = vb_29_end_mask_0, x = vb_21_cast_fp16)[name = string("vb_29_cast_fp16")];
            tensor<int32, [3]> gdec_29_begin_0 = const()[name = string("gdec_29_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_29_end_0 = const()[name = string("gdec_29_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_29_end_mask_0 = const()[name = string("gdec_29_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_29_cast_fp16 = slice_by_index(begin = gdec_29_begin_0, end = gdec_29_end_0, end_mask = gdec_29_end_mask_0, x = gdec_21_cast_fp16)[name = string("gdec_29_cast_fp16")];
            tensor<int32, [2]> var_2982 = const()[name = string("op_2982"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_2983_cast_fp16 = reshape(shape = var_2982, x = gdec_29_cast_fp16)[name = string("op_2983_cast_fp16")];
            tensor<fp16, [4, 256]> var_2984_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2983_cast_fp16)[name = string("op_2984_cast_fp16")];
            tensor<int32, [4]> var_2989 = const()[name = string("op_2989"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_23_cast_fp16 = reshape(shape = var_2989, x = var_2984_cast_fp16)[name = string("P_23_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2991_cast_fp16 = exp(x = P_23_cast_fp16)[name = string("op_2991_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_23_cast_fp16 = mul(x = var_2991_cast_fp16, y = tril_incl_to_fp16)[name = string("D_23_cast_fp16")];
            bool kk_23_transpose_x_1 = const()[name = string("kk_23_transpose_x_1"), val = bool(false)];
            bool kk_23_transpose_y_1 = const()[name = string("kk_23_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_23_cast_fp16 = matmul(transpose_x = kk_23_transpose_x_1, transpose_y = kk_23_transpose_y_1, x = kb_23_cast_fp16, y = k_47_cast_fp16)[name = string("kk_23_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_2997_cast_fp16 = mul(x = kk_23_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_2997_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2998_cast_fp16 = mul(x = var_2997_cast_fp16, y = D_23_cast_fp16)[name = string("op_2998_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_23_cast_fp16 = mul(x = var_2998_cast_fp16, y = cum_tail_to_fp16)[name = string("N_23_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3000_cast_fp16 = mul(x = D_23_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3000_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_23_cast_fp16 = mul(x = kk_23_cast_fp16, y = var_3000_cast_fp16)[name = string("Na_23_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_89_cast_fp16 = add(x = eye_to_fp16, y = N_23_cast_fp16)[name = string("Minv_89_cast_fp16")];
            bool Np_67_transpose_x_0 = const()[name = string("Np_67_transpose_x_0"), val = bool(false)];
            bool Np_67_transpose_y_0 = const()[name = string("Np_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_67_cast_fp16 = matmul(transpose_x = Np_67_transpose_x_0, transpose_y = Np_67_transpose_y_0, x = N_23_cast_fp16, y = Na_23_cast_fp16)[name = string("Np_67_cast_fp16")];
            bool Npa_45_transpose_x_0 = const()[name = string("Npa_45_transpose_x_0"), val = bool(false)];
            bool Npa_45_transpose_y_0 = const()[name = string("Npa_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_45_cast_fp16 = matmul(transpose_x = Npa_45_transpose_x_0, transpose_y = Npa_45_transpose_y_0, x = Na_23_cast_fp16, y = N_23_cast_fp16)[name = string("Npa_45_cast_fp16")];
            bool var_3006_transpose_x_0 = const()[name = string("op_3006_transpose_x_0"), val = bool(false)];
            bool var_3006_transpose_y_0 = const()[name = string("op_3006_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3006_cast_fp16 = matmul(transpose_x = var_3006_transpose_x_0, transpose_y = var_3006_transpose_y_0, x = Minv_89_cast_fp16, y = Np_67_cast_fp16)[name = string("op_3006_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_91_cast_fp16 = add(x = Minv_89_cast_fp16, y = var_3006_cast_fp16)[name = string("Minv_91_cast_fp16")];
            bool Np_69_transpose_x_0 = const()[name = string("Np_69_transpose_x_0"), val = bool(false)];
            bool Np_69_transpose_y_0 = const()[name = string("Np_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_69_cast_fp16 = matmul(transpose_x = Np_69_transpose_x_0, transpose_y = Np_69_transpose_y_0, x = Np_67_cast_fp16, y = Npa_45_cast_fp16)[name = string("Np_69_cast_fp16")];
            bool Npa_47_transpose_x_0 = const()[name = string("Npa_47_transpose_x_0"), val = bool(false)];
            bool Npa_47_transpose_y_0 = const()[name = string("Npa_47_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_47_cast_fp16 = matmul(transpose_x = Npa_47_transpose_x_0, transpose_y = Npa_47_transpose_y_0, x = Npa_45_cast_fp16, y = Np_67_cast_fp16)[name = string("Npa_47_cast_fp16")];
            bool var_3011_transpose_x_0 = const()[name = string("op_3011_transpose_x_0"), val = bool(false)];
            bool var_3011_transpose_y_0 = const()[name = string("op_3011_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3011_cast_fp16 = matmul(transpose_x = var_3011_transpose_x_0, transpose_y = var_3011_transpose_y_0, x = Minv_91_cast_fp16, y = Np_69_cast_fp16)[name = string("op_3011_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_93_cast_fp16 = add(x = Minv_91_cast_fp16, y = var_3011_cast_fp16)[name = string("Minv_93_cast_fp16")];
            bool Np_71_transpose_x_0 = const()[name = string("Np_71_transpose_x_0"), val = bool(false)];
            bool Np_71_transpose_y_0 = const()[name = string("Np_71_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_71_cast_fp16 = matmul(transpose_x = Np_71_transpose_x_0, transpose_y = Np_71_transpose_y_0, x = Np_69_cast_fp16, y = Npa_47_cast_fp16)[name = string("Np_71_cast_fp16")];
            bool var_3015_transpose_x_0 = const()[name = string("op_3015_transpose_x_0"), val = bool(false)];
            bool var_3015_transpose_y_0 = const()[name = string("op_3015_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3015_cast_fp16 = matmul(transpose_x = var_3015_transpose_x_0, transpose_y = var_3015_transpose_y_0, x = Minv_93_cast_fp16, y = Np_71_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_95_cast_fp16 = add(x = Minv_93_cast_fp16, y = var_3015_cast_fp16)[name = string("Minv_95_cast_fp16")];
            bool u_23_transpose_x_0 = const()[name = string("u_23_transpose_x_0"), val = bool(false)];
            bool u_23_transpose_y_0 = const()[name = string("u_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_23_cast_fp16 = matmul(transpose_x = u_23_transpose_x_0, transpose_y = u_23_transpose_y_0, x = Minv_95_cast_fp16, y = vb_29_cast_fp16)[name = string("u_23_cast_fp16")];
            bool var_3022_transpose_x_1 = const()[name = string("op_3022_transpose_x_1"), val = bool(false)];
            bool var_3022_transpose_y_1 = const()[name = string("op_3022_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_3022_cast_fp16 = matmul(transpose_x = var_3022_transpose_x_1, transpose_y = var_3022_transpose_y_1, x = q_47_cast_fp16, y = k_47_cast_fp16)[name = string("op_3022_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_23_cast_fp16 = mul(x = var_3022_cast_fp16, y = D_23_cast_fp16)[name = string("A_23_cast_fp16")];
            tensor<fp16, [16]> var_3024_bias_0_to_fp16 = const()[name = string("op_3024_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_3024_cast_fp16 = linear(bias = var_3024_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_29_cast_fp16)[name = string("op_3024_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_17_cast_fp16 = exp(x = var_3024_cast_fp16)[name = string("cum_17_cast_fp16")];
            tensor<int32, [1]> var_3027_axes_0 = const()[name = string("op_3027_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_3027_cast_fp16 = expand_dims(axes = var_3027_axes_0, x = cum_17_cast_fp16)[name = string("op_3027_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3028_cast_fp16 = mul(x = kb_23_cast_fp16, y = var_3027_cast_fp16)[name = string("op_3028_cast_fp16")];
            bool w_23_transpose_x_0 = const()[name = string("w_23_transpose_x_0"), val = bool(false)];
            bool w_23_transpose_y_0 = const()[name = string("w_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_23_cast_fp16 = matmul(transpose_x = w_23_transpose_x_0, transpose_y = w_23_transpose_y_0, x = Minv_95_cast_fp16, y = var_3028_cast_fp16)[name = string("w_23_cast_fp16")];
            bool var_3030_transpose_x_0 = const()[name = string("op_3030_transpose_x_0"), val = bool(false)];
            bool var_3030_transpose_y_0 = const()[name = string("op_3030_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3030_cast_fp16 = matmul(transpose_x = var_3030_transpose_x_0, transpose_y = var_3030_transpose_y_0, x = w_23_cast_fp16, y = S_17_cast_fp16)[name = string("op_3030_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_17_cast_fp16 = sub(x = u_23_cast_fp16, y = var_3030_cast_fp16)[name = string("vnew_17_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3035_cast_fp16 = mul(x = q_47_cast_fp16, y = var_3027_cast_fp16)[name = string("op_3035_cast_fp16")];
            bool var_3036_transpose_x_0 = const()[name = string("op_3036_transpose_x_0"), val = bool(false)];
            bool var_3036_transpose_y_0 = const()[name = string("op_3036_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3036_cast_fp16 = matmul(transpose_x = var_3036_transpose_x_0, transpose_y = var_3036_transpose_y_0, x = var_3035_cast_fp16, y = S_17_cast_fp16)[name = string("op_3036_cast_fp16")];
            bool var_3037_transpose_x_0 = const()[name = string("op_3037_transpose_x_0"), val = bool(false)];
            bool var_3037_transpose_y_0 = const()[name = string("op_3037_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3037_cast_fp16 = matmul(transpose_x = var_3037_transpose_x_0, transpose_y = var_3037_transpose_y_0, x = A_23_cast_fp16, y = vnew_17_cast_fp16)[name = string("op_3037_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_5_cast_fp16 = add(x = var_3036_cast_fp16, y = var_3037_cast_fp16)[name = string("oc_5_cast_fp16")];
            int32 var_3041 = const()[name = string("op_3041"), val = int32(2)];
            bool o_13_interleave_0 = const()[name = string("o_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 128]> o_13_cast_fp16 = concat(axis = var_3041, interleave = o_13_interleave_0, values = (var_2584_cast_fp16, var_2731_cast_fp16, var_2885_cast_fp16, oc_5_cast_fp16))[name = string("o_13_cast_fp16")];
            tensor<int32, [4]> var_3047 = const()[name = string("op_3047"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3053 = const()[name = string("op_3053"), val = tensor<int32, [4]>([1, 64, 4, 128])];
            tensor<fp16, [1, 64, 4, 128]> input_39_cast_fp16 = reshape(shape = var_3053, x = z_5_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> o_15_cast_fp16 = transpose(perm = var_3047, x = o_13_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 64, 4, 128]> var_3055_cast_fp16 = mul(x = o_15_cast_fp16, y = o_15_cast_fp16)[name = string("op_3055_cast_fp16")];
            tensor<int32, [1]> var_3060_axes_0 = const()[name = string("op_3060_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3060_keep_dims_0 = const()[name = string("op_3060_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 4, 1]> var_3060_cast_fp16 = reduce_mean(axes = var_3060_axes_0, keep_dims = var_3060_keep_dims_0, x = var_3055_cast_fp16)[name = string("op_3060_cast_fp16")];
            fp16 var_3062_to_fp16 = const()[name = string("op_3062_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 64, 4, 1]> var_3063_cast_fp16 = add(x = var_3060_cast_fp16, y = var_3062_to_fp16)[name = string("op_3063_cast_fp16")];
            fp32 var_3064_epsilon_0 = const()[name = string("op_3064_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 4, 1]> var_3064_cast_fp16 = rsqrt(epsilon = var_3064_epsilon_0, x = var_3063_cast_fp16)[name = string("op_3064_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> var_3065_cast_fp16 = mul(x = o_15_cast_fp16, y = var_3064_cast_fp16)[name = string("op_3065_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> var_3067_cast_fp16 = silu(x = input_39_cast_fp16)[name = string("op_3067_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> o_17_cast_fp16 = mul(x = var_3065_cast_fp16, y = var_3067_cast_fp16)[name = string("o_17_cast_fp16")];
            tensor<int32, [3]> var_3072 = const()[name = string("op_3072"), val = tensor<int32, [3]>([1, 64, 512])];
            tensor<fp16, [1, 64, 512]> input_41_cast_fp16 = reshape(shape = var_3072, x = o_17_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [640, 512]> layers_2_out_weight_to_fp16 = const()[name = string("layers_2_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47457728)))];
            tensor<fp16, [1, 64, 640]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_out_weight_to_fp16, x = input_41_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [1, 64, 640]> x_21_cast_fp16 = add(x = x_17_cast_fp16, y = linear_9_cast_fp16)[name = string("x_21_cast_fp16")];
            fp16 var_3083_promoted_to_fp16 = const()[name = string("op_3083_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> var_3087_cast_fp16 = pow(x = x_21_cast_fp16, y = var_3083_promoted_to_fp16)[name = string("op_3087_cast_fp16")];
            tensor<int32, [1]> var_3089_axes_0 = const()[name = string("op_3089_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3089_keep_dims_0 = const()[name = string("op_3089_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3089_cast_fp16 = reduce_mean(axes = var_3089_axes_0, keep_dims = var_3089_keep_dims_0, x = var_3087_cast_fp16)[name = string("op_3089_cast_fp16")];
            fp16 var_3090_to_fp16 = const()[name = string("op_3090_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3091_cast_fp16 = add(x = var_3089_cast_fp16, y = var_3090_to_fp16)[name = string("op_3091_cast_fp16")];
            fp32 var_3092_epsilon_0 = const()[name = string("op_3092_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3092_cast_fp16 = rsqrt(epsilon = var_3092_epsilon_0, x = var_3091_cast_fp16)[name = string("op_3092_cast_fp16")];
            tensor<fp16, [1, 64, 640]> var_3093_cast_fp16 = mul(x = x_21_cast_fp16, y = var_3092_cast_fp16)[name = string("op_3093_cast_fp16")];
            tensor<fp16, [640]> layers_2_n2_w1_to_fp16 = const()[name = string("layers_2_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48113152)))];
            tensor<fp16, [1, 64, 640]> input_43_cast_fp16 = mul(x = var_3093_cast_fp16, y = layers_2_n2_w1_to_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_2_gate_up_weight_to_fp16 = const()[name = string("layers_2_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48114496)))];
            tensor<fp16, [1, 64, 5120]> linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_2_gate_up_weight_to_fp16, x = input_43_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [3]> input_45_begin_0 = const()[name = string("input_45_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_45_end_0 = const()[name = string("input_45_end_0"), val = tensor<int32, [3]>([1, 64, 2560])];
            tensor<bool, [3]> input_45_end_mask_0 = const()[name = string("input_45_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 2560]> input_45_cast_fp16 = slice_by_index(begin = input_45_begin_0, end = input_45_end_0, end_mask = input_45_end_mask_0, x = linear_10_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [1, 64, 2560]> var_3103_cast_fp16 = silu(x = input_45_cast_fp16)[name = string("op_3103_cast_fp16")];
            tensor<int32, [3]> var_3108_begin_0 = const()[name = string("op_3108_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_3108_end_0 = const()[name = string("op_3108_end_0"), val = tensor<int32, [3]>([1, 64, 5120])];
            tensor<bool, [3]> var_3108_end_mask_0 = const()[name = string("op_3108_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 2560]> var_3108_cast_fp16 = slice_by_index(begin = var_3108_begin_0, end = var_3108_end_0, end_mask = var_3108_end_mask_0, x = linear_10_cast_fp16)[name = string("op_3108_cast_fp16")];
            tensor<fp16, [1, 64, 2560]> input_47_cast_fp16 = mul(x = var_3103_cast_fp16, y = var_3108_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_2_down_weight_to_fp16 = const()[name = string("layers_2_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54668160)))];
            tensor<fp16, [1, 64, 640]> linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_down_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 64, 640]> x_25_cast_fp16 = add(x = x_21_cast_fp16, y = linear_11_cast_fp16)[name = string("x_25_cast_fp16")];
            fp16 var_3121_to_fp16 = const()[name = string("op_3121_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 64, 640]> x_27_cast_fp16 = mul(x = x_25_cast_fp16, y = var_3121_to_fp16)[name = string("x_27_cast_fp16")];
            fp16 var_3119_promoted_to_fp16 = const()[name = string("op_3119_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> var_3123_cast_fp16 = pow(x = x_27_cast_fp16, y = var_3119_promoted_to_fp16)[name = string("op_3123_cast_fp16")];
            tensor<int32, [1]> var_3125_axes_0 = const()[name = string("op_3125_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3125_keep_dims_0 = const()[name = string("op_3125_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3125_cast_fp16 = reduce_mean(axes = var_3125_axes_0, keep_dims = var_3125_keep_dims_0, x = var_3123_cast_fp16)[name = string("op_3125_cast_fp16")];
            fp16 var_3126_to_fp16 = const()[name = string("op_3126_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 64, 1]> var_3127_cast_fp16 = add(x = var_3125_cast_fp16, y = var_3126_to_fp16)[name = string("op_3127_cast_fp16")];
            fp32 var_3128_epsilon_0 = const()[name = string("op_3128_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3128_cast_fp16 = rsqrt(epsilon = var_3128_epsilon_0, x = var_3127_cast_fp16)[name = string("op_3128_cast_fp16")];
            tensor<fp16, [1, 64, 640]> var_3129_cast_fp16 = mul(x = x_27_cast_fp16, y = var_3128_cast_fp16)[name = string("op_3129_cast_fp16")];
            tensor<fp16, [640]> layers_3_n1_w1_to_fp16 = const()[name = string("layers_3_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57945024)))];
            tensor<fp16, [1, 64, 640]> input_49_cast_fp16 = mul(x = var_3129_cast_fp16, y = layers_3_n1_w1_to_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1280, 640]> layers_3_qkv_weight_to_fp16 = const()[name = string("layers_3_qkv_weight_to_fp16"), val = tensor<fp16, [1280, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57946368)))];
            tensor<fp16, [1280]> linear_12_bias_0_to_fp16 = const()[name = string("linear_12_bias_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59584832)))];
            tensor<fp16, [1, 64, 1280]> linear_12_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = layers_3_qkv_weight_to_fp16, x = input_49_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [3]> var_3138_begin_0 = const()[name = string("op_3138_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3138_end_0 = const()[name = string("op_3138_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> var_3138_end_mask_0 = const()[name = string("op_3138_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 1024]> var_3138_cast_fp16 = slice_by_index(begin = var_3138_begin_0, end = var_3138_end_0, end_mask = var_3138_end_mask_0, x = linear_12_cast_fp16)[name = string("op_3138_cast_fp16")];
            tensor<int32, [4]> var_3143 = const()[name = string("op_3143"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<fp16, [1, 64, 8, 128]> qg_1_cast_fp16 = reshape(shape = var_3143, x = var_3138_cast_fp16)[name = string("qg_1_cast_fp16")];
            tensor<int32, [4]> x_29_begin_0 = const()[name = string("x_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_29_end_0 = const()[name = string("x_29_end_0"), val = tensor<int32, [4]>([1, 64, 8, 64])];
            tensor<bool, [4]> x_29_end_mask_0 = const()[name = string("x_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 64, 8, 64]> x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, x = qg_1_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [4]> var_3154_begin_0 = const()[name = string("op_3154_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> var_3154_end_0 = const()[name = string("op_3154_end_0"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<bool, [4]> var_3154_end_mask_0 = const()[name = string("op_3154_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 64, 8, 64]> var_3154_cast_fp16 = slice_by_index(begin = var_3154_begin_0, end = var_3154_end_0, end_mask = var_3154_end_mask_0, x = qg_1_cast_fp16)[name = string("op_3154_cast_fp16")];
            tensor<int32, [3]> var_3158 = const()[name = string("op_3158"), val = tensor<int32, [3]>([1, 64, 512])];
            tensor<fp16, [1, 64, 512]> gate_1_cast_fp16 = reshape(shape = var_3158, x = var_3154_cast_fp16)[name = string("gate_1_cast_fp16")];
            tensor<int32, [3]> var_3164_begin_0 = const()[name = string("op_3164_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> var_3164_end_0 = const()[name = string("op_3164_end_0"), val = tensor<int32, [3]>([1, 64, 1152])];
            tensor<bool, [3]> var_3164_end_mask_0 = const()[name = string("op_3164_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 128]> var_3164_cast_fp16 = slice_by_index(begin = var_3164_begin_0, end = var_3164_end_0, end_mask = var_3164_end_mask_0, x = linear_12_cast_fp16)[name = string("op_3164_cast_fp16")];
            tensor<int32, [4]> var_3169 = const()[name = string("op_3169"), val = tensor<int32, [4]>([1, 64, 2, 64])];
            tensor<fp16, [1, 64, 2, 64]> x_35_cast_fp16 = reshape(shape = var_3169, x = var_3164_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<int32, [3]> var_3175_begin_0 = const()[name = string("op_3175_begin_0"), val = tensor<int32, [3]>([0, 0, 1152])];
            tensor<int32, [3]> var_3175_end_0 = const()[name = string("op_3175_end_0"), val = tensor<int32, [3]>([1, 64, 1280])];
            tensor<bool, [3]> var_3175_end_mask_0 = const()[name = string("op_3175_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 128]> var_3175_cast_fp16 = slice_by_index(begin = var_3175_begin_0, end = var_3175_end_0, end_mask = var_3175_end_mask_0, x = linear_12_cast_fp16)[name = string("op_3175_cast_fp16")];
            tensor<int32, [4]> var_3180 = const()[name = string("op_3180"), val = tensor<int32, [4]>([1, 64, 2, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_3181_cast_fp16 = reshape(shape = var_3180, x = var_3175_cast_fp16)[name = string("op_3181_cast_fp16")];
            tensor<int32, [4]> var_3186 = const()[name = string("op_3186"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3192_promoted_to_fp16 = const()[name = string("op_3192_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 8, 64]> var_3196_cast_fp16 = pow(x = x_29_cast_fp16, y = var_3192_promoted_to_fp16)[name = string("op_3196_cast_fp16")];
            tensor<int32, [1]> var_3198_axes_0 = const()[name = string("op_3198_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3198_keep_dims_0 = const()[name = string("op_3198_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 8, 1]> var_3198_cast_fp16 = reduce_mean(axes = var_3198_axes_0, keep_dims = var_3198_keep_dims_0, x = var_3196_cast_fp16)[name = string("op_3198_cast_fp16")];
            fp16 var_3199_to_fp16 = const()[name = string("op_3199_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 8, 1]> var_3200_cast_fp16 = add(x = var_3198_cast_fp16, y = var_3199_to_fp16)[name = string("op_3200_cast_fp16")];
            fp32 var_3201_epsilon_0 = const()[name = string("op_3201_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 8, 1]> var_3201_cast_fp16 = rsqrt(epsilon = var_3201_epsilon_0, x = var_3200_cast_fp16)[name = string("op_3201_cast_fp16")];
            tensor<fp16, [1, 64, 8, 64]> var_3202_cast_fp16 = mul(x = x_29_cast_fp16, y = var_3201_cast_fp16)[name = string("op_3202_cast_fp16")];
            tensor<fp16, [64]> layers_3_qn_w1_to_fp16 = const()[name = string("layers_3_qn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59587456)))];
            tensor<fp16, [1, 64, 8, 64]> var_3203_cast_fp16 = mul(x = var_3202_cast_fp16, y = layers_3_qn_w1_to_fp16)[name = string("op_3203_cast_fp16")];
            tensor<int32, [4]> var_3208 = const()[name = string("op_3208"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_1_begin_0 = const()[name = string("r_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_1_end_0 = const()[name = string("r_1_end_0"), val = tensor<int32, [4]>([1, 8, 64, 16])];
            tensor<bool, [4]> r_1_end_mask_0 = const()[name = string("r_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 64, 64]> x_33_cast_fp16 = transpose(perm = var_3208, x = var_3203_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 8, 64, 16]> r_1_cast_fp16 = slice_by_index(begin = r_1_begin_0, end = r_1_end_0, end_mask = r_1_end_mask_0, x = x_33_cast_fp16)[name = string("r_1_cast_fp16")];
            tensor<int32, [4]> p_1_begin_0 = const()[name = string("p_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_1_end_0 = const()[name = string("p_1_end_0"), val = tensor<int32, [4]>([1, 8, 64, 64])];
            tensor<bool, [4]> p_1_end_mask_0 = const()[name = string("p_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 64, 48]> p_1_cast_fp16 = slice_by_index(begin = p_1_begin_0, end = p_1_end_0, end_mask = p_1_end_mask_0, x = x_33_cast_fp16)[name = string("p_1_cast_fp16")];
            tensor<int32, [4]> var_3224_begin_0 = const()[name = string("op_3224_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_3224_end_0 = const()[name = string("op_3224_end_0"), val = tensor<int32, [4]>([1, 8, 64, 16])];
            tensor<bool, [4]> var_3224_end_mask_0 = const()[name = string("op_3224_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 64, 8]> var_3224_cast_fp16 = slice_by_index(begin = var_3224_begin_0, end = var_3224_end_0, end_mask = var_3224_end_mask_0, x = r_1_cast_fp16)[name = string("op_3224_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 64, 8]> var_3225_cast_fp16 = mul(x = var_3224_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_3225_cast_fp16")];
            tensor<int32, [4]> var_3230_begin_0 = const()[name = string("op_3230_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3230_end_0 = const()[name = string("op_3230_end_0"), val = tensor<int32, [4]>([1, 8, 64, 8])];
            tensor<bool, [4]> var_3230_end_mask_0 = const()[name = string("op_3230_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 64, 8]> var_3230_cast_fp16 = slice_by_index(begin = var_3230_begin_0, end = var_3230_end_0, end_mask = var_3230_end_mask_0, x = r_1_cast_fp16)[name = string("op_3230_cast_fp16")];
            int32 var_3232 = const()[name = string("op_3232"), val = int32(-1)];
            bool rh_1_interleave_0 = const()[name = string("rh_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 16]> rh_1_cast_fp16 = concat(axis = var_3232, interleave = rh_1_interleave_0, values = (var_3225_cast_fp16, var_3230_cast_fp16))[name = string("rh_1_cast_fp16")];
            tensor<fp16, [64, 16]> cos_to_fp16 = const()[name = string("cos_to_fp16"), val = tensor<fp16, [64, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663296)))];
            tensor<fp16, [1, 8, 64, 16]> var_3234_cast_fp16 = mul(x = r_1_cast_fp16, y = cos_to_fp16)[name = string("op_3234_cast_fp16")];
            tensor<fp16, [64, 16]> sin_to_fp16 = const()[name = string("sin_to_fp16"), val = tensor<fp16, [64, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119665408)))];
            tensor<fp16, [1, 8, 64, 16]> var_3235_cast_fp16 = mul(x = rh_1_cast_fp16, y = sin_to_fp16)[name = string("op_3235_cast_fp16")];
            tensor<fp16, [1, 8, 64, 16]> var_3237_cast_fp16 = add(x = var_3234_cast_fp16, y = var_3235_cast_fp16)[name = string("op_3237_cast_fp16")];
            int32 var_3239 = const()[name = string("op_3239"), val = int32(-1)];
            bool q_49_interleave_0 = const()[name = string("q_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 64]> q_49_cast_fp16 = concat(axis = var_3239, interleave = q_49_interleave_0, values = (var_3237_cast_fp16, p_1_cast_fp16))[name = string("q_49_cast_fp16")];
            fp16 var_3245_promoted_to_fp16 = const()[name = string("op_3245_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 2, 64]> var_3249_cast_fp16 = pow(x = x_35_cast_fp16, y = var_3245_promoted_to_fp16)[name = string("op_3249_cast_fp16")];
            tensor<int32, [1]> var_3251_axes_0 = const()[name = string("op_3251_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3251_keep_dims_0 = const()[name = string("op_3251_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 2, 1]> var_3251_cast_fp16 = reduce_mean(axes = var_3251_axes_0, keep_dims = var_3251_keep_dims_0, x = var_3249_cast_fp16)[name = string("op_3251_cast_fp16")];
            fp16 var_3252_to_fp16 = const()[name = string("op_3252_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 2, 1]> var_3253_cast_fp16 = add(x = var_3251_cast_fp16, y = var_3252_to_fp16)[name = string("op_3253_cast_fp16")];
            fp32 var_3254_epsilon_0 = const()[name = string("op_3254_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 2, 1]> var_3254_cast_fp16 = rsqrt(epsilon = var_3254_epsilon_0, x = var_3253_cast_fp16)[name = string("op_3254_cast_fp16")];
            tensor<fp16, [1, 64, 2, 64]> var_3255_cast_fp16 = mul(x = x_35_cast_fp16, y = var_3254_cast_fp16)[name = string("op_3255_cast_fp16")];
            tensor<fp16, [64]> layers_3_kn_w1_to_fp16 = const()[name = string("layers_3_kn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59588800)))];
            tensor<fp16, [1, 64, 2, 64]> var_3256_cast_fp16 = mul(x = var_3255_cast_fp16, y = layers_3_kn_w1_to_fp16)[name = string("op_3256_cast_fp16")];
            tensor<int32, [4]> var_3261 = const()[name = string("op_3261"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_3_begin_0 = const()[name = string("r_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_3_end_0 = const()[name = string("r_3_end_0"), val = tensor<int32, [4]>([1, 2, 64, 16])];
            tensor<bool, [4]> r_3_end_mask_0 = const()[name = string("r_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 64]> x_39_cast_fp16 = transpose(perm = var_3261, x = var_3256_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 2, 64, 16]> r_3_cast_fp16 = slice_by_index(begin = r_3_begin_0, end = r_3_end_0, end_mask = r_3_end_mask_0, x = x_39_cast_fp16)[name = string("r_3_cast_fp16")];
            tensor<int32, [4]> p_3_begin_0 = const()[name = string("p_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_3_end_0 = const()[name = string("p_3_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> p_3_end_mask_0 = const()[name = string("p_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 48]> p_3_cast_fp16 = slice_by_index(begin = p_3_begin_0, end = p_3_end_0, end_mask = p_3_end_mask_0, x = x_39_cast_fp16)[name = string("p_3_cast_fp16")];
            tensor<int32, [4]> var_3277_begin_0 = const()[name = string("op_3277_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_3277_end_0 = const()[name = string("op_3277_end_0"), val = tensor<int32, [4]>([1, 2, 64, 16])];
            tensor<bool, [4]> var_3277_end_mask_0 = const()[name = string("op_3277_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 8]> var_3277_cast_fp16 = slice_by_index(begin = var_3277_begin_0, end = var_3277_end_0, end_mask = var_3277_end_mask_0, x = r_3_cast_fp16)[name = string("op_3277_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_3278_cast_fp16 = mul(x = var_3277_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_3278_cast_fp16")];
            tensor<int32, [4]> var_3283_begin_0 = const()[name = string("op_3283_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3283_end_0 = const()[name = string("op_3283_end_0"), val = tensor<int32, [4]>([1, 2, 64, 8])];
            tensor<bool, [4]> var_3283_end_mask_0 = const()[name = string("op_3283_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 8]> var_3283_cast_fp16 = slice_by_index(begin = var_3283_begin_0, end = var_3283_end_0, end_mask = var_3283_end_mask_0, x = r_3_cast_fp16)[name = string("op_3283_cast_fp16")];
            int32 var_3285 = const()[name = string("op_3285"), val = int32(-1)];
            bool rh_3_interleave_0 = const()[name = string("rh_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 16]> rh_3_cast_fp16 = concat(axis = var_3285, interleave = rh_3_interleave_0, values = (var_3278_cast_fp16, var_3283_cast_fp16))[name = string("rh_3_cast_fp16")];
            tensor<fp16, [1, 2, 64, 16]> var_3287_cast_fp16 = mul(x = r_3_cast_fp16, y = cos_to_fp16)[name = string("op_3287_cast_fp16")];
            tensor<fp16, [1, 2, 64, 16]> var_3288_cast_fp16 = mul(x = rh_3_cast_fp16, y = sin_to_fp16)[name = string("op_3288_cast_fp16")];
            tensor<fp16, [1, 2, 64, 16]> var_3290_cast_fp16 = add(x = var_3287_cast_fp16, y = var_3288_cast_fp16)[name = string("op_3290_cast_fp16")];
            int32 var_3292 = const()[name = string("op_3292"), val = int32(-1)];
            bool k_49_interleave_0 = const()[name = string("k_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> k_49_cast_fp16 = concat(axis = var_3292, interleave = k_49_interleave_0, values = (var_3290_cast_fp16, p_3_cast_fp16))[name = string("k_49_cast_fp16")];
            tensor<int32, [1]> var_3305_axes_0 = const()[name = string("op_3305_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_3305_cast_fp16 = expand_dims(axes = var_3305_axes_0, x = k_49_cast_fp16)[name = string("op_3305_cast_fp16")];
            tensor<int32, [5]> var_3313_reps_0 = const()[name = string("op_3313_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 2, 4, 64, 64]> var_3313_cast_fp16 = tile(reps = var_3313_reps_0, x = var_3305_cast_fp16)[name = string("op_3313_cast_fp16")];
            tensor<int32, [4]> var_3318 = const()[name = string("op_3318"), val = tensor<int32, [4]>([1, 8, 64, 64])];
            tensor<fp16, [1, 8, 64, 64]> k_51_cast_fp16 = reshape(shape = var_3318, x = var_3313_cast_fp16)[name = string("k_51_cast_fp16")];
            tensor<int32, [1]> var_3331_axes_0 = const()[name = string("op_3331_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> v_7_cast_fp16 = transpose(perm = var_3186, x = var_3181_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_3331_cast_fp16 = expand_dims(axes = var_3331_axes_0, x = v_7_cast_fp16)[name = string("op_3331_cast_fp16")];
            tensor<int32, [5]> var_3339_reps_0 = const()[name = string("op_3339_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 2, 4, 64, 64]> var_3339_cast_fp16 = tile(reps = var_3339_reps_0, x = var_3331_cast_fp16)[name = string("op_3339_cast_fp16")];
            tensor<int32, [4]> var_3344 = const()[name = string("op_3344"), val = tensor<int32, [4]>([1, 8, 64, 64])];
            tensor<fp16, [1, 8, 64, 64]> v_9_cast_fp16 = reshape(shape = var_3344, x = var_3339_cast_fp16)[name = string("v_9_cast_fp16")];
            bool var_3349_transpose_x_1 = const()[name = string("op_3349_transpose_x_1"), val = bool(false)];
            bool var_3349_transpose_y_1 = const()[name = string("op_3349_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 64, 64]> var_3349_cast_fp16 = matmul(transpose_x = var_3349_transpose_x_1, transpose_y = var_3349_transpose_y_1, x = q_49_cast_fp16, y = k_51_cast_fp16)[name = string("op_3349_cast_fp16")];
            fp16 var_3350_to_fp16 = const()[name = string("op_3350_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 8, 64, 64]> var_3351_cast_fp16 = mul(x = var_3349_cast_fp16, y = var_3350_to_fp16)[name = string("op_3351_cast_fp16")];
            tensor<fp16, [64, 64]> causal_add_to_fp16 = const()[name = string("causal_add_to_fp16"), val = tensor<fp16, [64, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119667520)))];
            tensor<fp16, [1, 8, 64, 64]> s_1_cast_fp16 = add(x = var_3351_cast_fp16, y = causal_add_to_fp16)[name = string("s_1_cast_fp16")];
            int32 var_3354 = const()[name = string("op_3354"), val = int32(-1)];
            tensor<fp16, [1, 8, 64, 64]> var_3356_cast_fp16 = softmax(axis = var_3354, x = s_1_cast_fp16)[name = string("op_3356_cast_fp16")];
            bool o_19_transpose_x_0 = const()[name = string("o_19_transpose_x_0"), val = bool(false)];
            bool o_19_transpose_y_0 = const()[name = string("o_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 64]> o_19_cast_fp16 = matmul(transpose_x = o_19_transpose_x_0, transpose_y = o_19_transpose_y_0, x = var_3356_cast_fp16, y = v_9_cast_fp16)[name = string("o_19_cast_fp16")];
            tensor<int32, [4]> var_3362 = const()[name = string("op_3362"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3367 = const()[name = string("op_3367"), val = tensor<int32, [3]>([1, 64, 512])];
            tensor<fp16, [1, 64, 8, 64]> var_3363_cast_fp16 = transpose(perm = var_3362, x = o_19_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 64, 512]> var_3368_cast_fp16 = reshape(shape = var_3367, x = var_3363_cast_fp16)[name = string("op_3368_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3369_cast_fp16 = sigmoid(x = gate_1_cast_fp16)[name = string("op_3369_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_51_cast_fp16 = mul(x = var_3368_cast_fp16, y = var_3369_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [640, 512]> layers_3_o_weight_to_fp16 = const()[name = string("layers_3_o_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59589568)))];
            tensor<fp16, [1, 64, 640]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_o_weight_to_fp16, x = input_51_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 64, 640]> x_41_cast_fp16 = add(x = x_25_cast_fp16, y = linear_13_cast_fp16)[name = string("x_41_cast_fp16")];
            fp16 var_3382_to_fp16 = const()[name = string("op_3382_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 64, 640]> x_43_cast_fp16 = mul(x = x_41_cast_fp16, y = var_3382_to_fp16)[name = string("x_43_cast_fp16")];
            fp16 var_3380_promoted_to_fp16 = const()[name = string("op_3380_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> var_3384_cast_fp16 = pow(x = x_43_cast_fp16, y = var_3380_promoted_to_fp16)[name = string("op_3384_cast_fp16")];
            tensor<int32, [1]> var_3386_axes_0 = const()[name = string("op_3386_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3386_keep_dims_0 = const()[name = string("op_3386_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3386_cast_fp16 = reduce_mean(axes = var_3386_axes_0, keep_dims = var_3386_keep_dims_0, x = var_3384_cast_fp16)[name = string("op_3386_cast_fp16")];
            fp16 var_3387_to_fp16 = const()[name = string("op_3387_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 64, 1]> var_3388_cast_fp16 = add(x = var_3386_cast_fp16, y = var_3387_to_fp16)[name = string("op_3388_cast_fp16")];
            fp32 var_3389_epsilon_0 = const()[name = string("op_3389_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3389_cast_fp16 = rsqrt(epsilon = var_3389_epsilon_0, x = var_3388_cast_fp16)[name = string("op_3389_cast_fp16")];
            tensor<fp16, [1, 64, 640]> var_3390_cast_fp16 = mul(x = x_43_cast_fp16, y = var_3389_cast_fp16)[name = string("op_3390_cast_fp16")];
            tensor<fp16, [640]> layers_3_n2_w1_to_fp16 = const()[name = string("layers_3_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60244992)))];
            tensor<fp16, [1, 64, 640]> input_53_cast_fp16 = mul(x = var_3390_cast_fp16, y = layers_3_n2_w1_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_3_gate_up_weight_to_fp16 = const()[name = string("layers_3_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60246336)))];
            tensor<fp16, [1, 64, 5120]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_3_gate_up_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [3]> input_55_begin_0 = const()[name = string("input_55_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_55_end_0 = const()[name = string("input_55_end_0"), val = tensor<int32, [3]>([1, 64, 2560])];
            tensor<bool, [3]> input_55_end_mask_0 = const()[name = string("input_55_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 2560]> input_55_cast_fp16 = slice_by_index(begin = input_55_begin_0, end = input_55_end_0, end_mask = input_55_end_mask_0, x = linear_14_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [1, 64, 2560]> var_3400_cast_fp16 = silu(x = input_55_cast_fp16)[name = string("op_3400_cast_fp16")];
            tensor<int32, [3]> var_3405_begin_0 = const()[name = string("op_3405_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_3405_end_0 = const()[name = string("op_3405_end_0"), val = tensor<int32, [3]>([1, 64, 5120])];
            tensor<bool, [3]> var_3405_end_mask_0 = const()[name = string("op_3405_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 2560]> var_3405_cast_fp16 = slice_by_index(begin = var_3405_begin_0, end = var_3405_end_0, end_mask = var_3405_end_mask_0, x = linear_14_cast_fp16)[name = string("op_3405_cast_fp16")];
            tensor<fp16, [1, 64, 2560]> input_57_cast_fp16 = mul(x = var_3400_cast_fp16, y = var_3405_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_3_down_weight_to_fp16 = const()[name = string("layers_3_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66800000)))];
            tensor<fp16, [1, 64, 640]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_down_weight_to_fp16, x = input_57_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, 64, 640]> x_45_cast_fp16 = add(x = x_41_cast_fp16, y = linear_15_cast_fp16)[name = string("x_45_cast_fp16")];
            fp16 var_3418_to_fp16 = const()[name = string("op_3418_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 64, 640]> x_47_cast_fp16 = mul(x = x_45_cast_fp16, y = var_3418_to_fp16)[name = string("x_47_cast_fp16")];
            fp16 var_3416_promoted_to_fp16 = const()[name = string("op_3416_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> var_3420_cast_fp16 = pow(x = x_47_cast_fp16, y = var_3416_promoted_to_fp16)[name = string("op_3420_cast_fp16")];
            tensor<int32, [1]> var_3422_axes_0 = const()[name = string("op_3422_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3422_keep_dims_0 = const()[name = string("op_3422_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3422_cast_fp16 = reduce_mean(axes = var_3422_axes_0, keep_dims = var_3422_keep_dims_0, x = var_3420_cast_fp16)[name = string("op_3422_cast_fp16")];
            fp16 var_3423_to_fp16 = const()[name = string("op_3423_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 64, 1]> var_3424_cast_fp16 = add(x = var_3422_cast_fp16, y = var_3423_to_fp16)[name = string("op_3424_cast_fp16")];
            fp32 var_3425_epsilon_0 = const()[name = string("op_3425_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3425_cast_fp16 = rsqrt(epsilon = var_3425_epsilon_0, x = var_3424_cast_fp16)[name = string("op_3425_cast_fp16")];
            tensor<fp16, [1, 64, 640]> var_3426_cast_fp16 = mul(x = x_47_cast_fp16, y = var_3425_cast_fp16)[name = string("op_3426_cast_fp16")];
            tensor<fp16, [640]> layers_4_n1_w1_to_fp16 = const()[name = string("layers_4_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70076864)))];
            tensor<fp16, [1, 64, 640]> input_59_cast_fp16 = mul(x = var_3426_cast_fp16, y = layers_4_n1_w1_to_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_4_qkvzba_weight_to_fp16 = const()[name = string("layers_4_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70078208)))];
            tensor<fp16, [1, 64, 1544]> linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_qkvzba_weight_to_fp16, x = input_59_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [3]> qkv_7_begin_0 = const()[name = string("qkv_7_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_7_end_0 = const()[name = string("qkv_7_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> qkv_7_end_mask_0 = const()[name = string("qkv_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 1024]> qkv_7_cast_fp16 = slice_by_index(begin = qkv_7_begin_0, end = qkv_7_end_0, end_mask = qkv_7_end_mask_0, x = linear_16_cast_fp16)[name = string("qkv_7_cast_fp16")];
            tensor<int32, [3]> z_7_begin_0 = const()[name = string("z_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_7_end_0 = const()[name = string("z_7_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
            tensor<bool, [3]> z_7_end_mask_0 = const()[name = string("z_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 512]> z_7_cast_fp16 = slice_by_index(begin = z_7_begin_0, end = z_7_end_0, end_mask = z_7_end_mask_0, x = linear_16_cast_fp16)[name = string("z_7_cast_fp16")];
            tensor<int32, [3]> b_7_begin_0 = const()[name = string("b_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_7_end_0 = const()[name = string("b_7_end_0"), val = tensor<int32, [3]>([1, 64, 1540])];
            tensor<bool, [3]> b_7_end_mask_0 = const()[name = string("b_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 4]> b_7_cast_fp16 = slice_by_index(begin = b_7_begin_0, end = b_7_end_0, end_mask = b_7_end_mask_0, x = linear_16_cast_fp16)[name = string("b_7_cast_fp16")];
            tensor<int32, [3]> a_7_begin_0 = const()[name = string("a_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_7_end_0 = const()[name = string("a_7_end_0"), val = tensor<int32, [3]>([1, 64, 1])];
            tensor<bool, [3]> a_7_end_mask_0 = const()[name = string("a_7_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 4]> a_7_cast_fp16 = slice_by_index(begin = a_7_begin_0, end = a_7_end_0, end_mask = a_7_end_mask_0, x = linear_16_cast_fp16)[name = string("a_7_cast_fp16")];
            int32 var_3465 = const()[name = string("op_3465"), val = int32(1)];
            bool xp_7_interleave_0 = const()[name = string("xp_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 67, 1024]> xp_7_cast_fp16 = concat(axis = var_3465, interleave = xp_7_interleave_0, values = (zpad_to_fp16, qkv_7_cast_fp16))[name = string("xp_7_cast_fp16")];
            tensor<int32, [3]> var_3476_begin_0 = const()[name = string("op_3476_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_3476_end_0 = const()[name = string("op_3476_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_3476_end_mask_0 = const()[name = string("op_3476_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 1024]> var_3476_cast_fp16 = slice_by_index(begin = var_3476_begin_0, end = var_3476_end_0, end_mask = var_3476_end_mask_0, x = xp_7_cast_fp16)[name = string("op_3476_cast_fp16")];
            tensor<fp16, [1024]> var_3484_to_fp16 = const()[name = string("op_3484_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72054592)))];
            tensor<fp16, [1, 64, 1024]> var_3485_cast_fp16 = mul(x = var_3476_cast_fp16, y = var_3484_to_fp16)[name = string("op_3485_cast_fp16")];
            tensor<int32, [3]> var_3495_begin_0 = const()[name = string("op_3495_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_3495_end_0 = const()[name = string("op_3495_end_0"), val = tensor<int32, [3]>([1, 66, 1024])];
            tensor<bool, [3]> var_3495_end_mask_0 = const()[name = string("op_3495_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_3495_cast_fp16 = slice_by_index(begin = var_3495_begin_0, end = var_3495_end_0, end_mask = var_3495_end_mask_0, x = xp_7_cast_fp16)[name = string("op_3495_cast_fp16")];
            tensor<fp16, [1024]> var_3503_to_fp16 = const()[name = string("op_3503_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72056704)))];
            tensor<fp16, [1, 64, 1024]> var_3504_cast_fp16 = mul(x = var_3495_cast_fp16, y = var_3503_to_fp16)[name = string("op_3504_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> var_3506_cast_fp16 = add(x = var_3485_cast_fp16, y = var_3504_cast_fp16)[name = string("op_3506_cast_fp16")];
            tensor<int32, [3]> var_3516_begin_0 = const()[name = string("op_3516_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_3516_end_0 = const()[name = string("op_3516_end_0"), val = tensor<int32, [3]>([1, 65, 1024])];
            tensor<bool, [3]> var_3516_end_mask_0 = const()[name = string("op_3516_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_3516_cast_fp16 = slice_by_index(begin = var_3516_begin_0, end = var_3516_end_0, end_mask = var_3516_end_mask_0, x = xp_7_cast_fp16)[name = string("op_3516_cast_fp16")];
            tensor<fp16, [1024]> var_3524_to_fp16 = const()[name = string("op_3524_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72058816)))];
            tensor<fp16, [1, 64, 1024]> var_3525_cast_fp16 = mul(x = var_3516_cast_fp16, y = var_3524_to_fp16)[name = string("op_3525_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> var_3527_cast_fp16 = add(x = var_3506_cast_fp16, y = var_3525_cast_fp16)[name = string("op_3527_cast_fp16")];
            tensor<int32, [3]> var_3537_begin_0 = const()[name = string("op_3537_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3537_end_0 = const()[name = string("op_3537_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> var_3537_end_mask_0 = const()[name = string("op_3537_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_3537_cast_fp16 = slice_by_index(begin = var_3537_begin_0, end = var_3537_end_0, end_mask = var_3537_end_mask_0, x = xp_7_cast_fp16)[name = string("op_3537_cast_fp16")];
            tensor<fp16, [1024]> var_3545_to_fp16 = const()[name = string("op_3545_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72060928)))];
            tensor<fp16, [1, 64, 1024]> var_3546_cast_fp16 = mul(x = var_3537_cast_fp16, y = var_3545_to_fp16)[name = string("op_3546_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> input_61_cast_fp16 = add(x = var_3527_cast_fp16, y = var_3546_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> c_7_cast_fp16 = silu(x = input_61_cast_fp16)[name = string("c_7_cast_fp16")];
            tensor<int32, [3]> var_3554_begin_0 = const()[name = string("op_3554_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3554_end_0 = const()[name = string("op_3554_end_0"), val = tensor<int32, [3]>([1, 64, 256])];
            tensor<bool, [3]> var_3554_end_mask_0 = const()[name = string("op_3554_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 256]> var_3554_cast_fp16 = slice_by_index(begin = var_3554_begin_0, end = var_3554_end_0, end_mask = var_3554_end_mask_0, x = c_7_cast_fp16)[name = string("op_3554_cast_fp16")];
            tensor<int32, [4]> var_3559 = const()[name = string("op_3559"), val = tensor<int32, [4]>([1, 64, 2, 128])];
            tensor<fp16, [1, 64, 2, 128]> q_51_cast_fp16 = reshape(shape = var_3559, x = var_3554_cast_fp16)[name = string("q_51_cast_fp16")];
            tensor<int32, [3]> var_3565_begin_0 = const()[name = string("op_3565_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_3565_end_0 = const()[name = string("op_3565_end_0"), val = tensor<int32, [3]>([1, 64, 512])];
            tensor<bool, [3]> var_3565_end_mask_0 = const()[name = string("op_3565_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 256]> var_3565_cast_fp16 = slice_by_index(begin = var_3565_begin_0, end = var_3565_end_0, end_mask = var_3565_end_mask_0, x = c_7_cast_fp16)[name = string("op_3565_cast_fp16")];
            tensor<int32, [4]> var_3570 = const()[name = string("op_3570"), val = tensor<int32, [4]>([1, 64, 2, 128])];
            tensor<fp16, [1, 64, 2, 128]> k_53_cast_fp16 = reshape(shape = var_3570, x = var_3565_cast_fp16)[name = string("k_53_cast_fp16")];
            tensor<int32, [3]> var_3576_begin_0 = const()[name = string("op_3576_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_3576_end_0 = const()[name = string("op_3576_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> var_3576_end_mask_0 = const()[name = string("op_3576_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 512]> var_3576_cast_fp16 = slice_by_index(begin = var_3576_begin_0, end = var_3576_end_0, end_mask = var_3576_end_mask_0, x = c_7_cast_fp16)[name = string("op_3576_cast_fp16")];
            tensor<int32, [4]> var_3581 = const()[name = string("op_3581"), val = tensor<int32, [4]>([1, 64, 4, 128])];
            tensor<fp16, [1, 64, 4, 128]> var_3582_cast_fp16 = reshape(shape = var_3581, x = var_3576_cast_fp16)[name = string("op_3582_cast_fp16")];
            tensor<int32, [4]> var_3587 = const()[name = string("op_3587"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3589_to_fp16 = const()[name = string("op_3589_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 64, 2, 128]> q_53_cast_fp16 = mul(x = q_51_cast_fp16, y = var_3589_to_fp16)[name = string("q_53_cast_fp16")];
            fp16 var_3591_to_fp16 = const()[name = string("op_3591_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 64, 2, 128]> k_55_cast_fp16 = mul(x = k_53_cast_fp16, y = var_3591_to_fp16)[name = string("k_55_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_3593_cast_fp16 = mul(x = q_53_cast_fp16, y = q_53_cast_fp16)[name = string("op_3593_cast_fp16")];
            tensor<int32, [1]> var_3598_axes_0 = const()[name = string("op_3598_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3598_keep_dims_0 = const()[name = string("op_3598_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 2, 1]> var_3598_cast_fp16 = reduce_mean(axes = var_3598_axes_0, keep_dims = var_3598_keep_dims_0, x = var_3593_cast_fp16)[name = string("op_3598_cast_fp16")];
            fp16 var_3599_promoted_to_fp16 = const()[name = string("op_3599_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 64, 2, 1]> var_3600_cast_fp16 = mul(x = var_3598_cast_fp16, y = var_3599_promoted_to_fp16)[name = string("op_3600_cast_fp16")];
            fp16 var_3602_to_fp16 = const()[name = string("op_3602_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 64, 2, 1]> var_3603_cast_fp16 = add(x = var_3600_cast_fp16, y = var_3602_to_fp16)[name = string("op_3603_cast_fp16")];
            fp32 var_3604_epsilon_0 = const()[name = string("op_3604_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 2, 1]> var_3604_cast_fp16 = rsqrt(epsilon = var_3604_epsilon_0, x = var_3603_cast_fp16)[name = string("op_3604_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_3605_cast_fp16 = mul(x = q_53_cast_fp16, y = var_3604_cast_fp16)[name = string("op_3605_cast_fp16")];
            fp16 var_3606_to_fp16 = const()[name = string("op_3606_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 64, 2, 128]> q_55_cast_fp16 = mul(x = var_3605_cast_fp16, y = var_3606_to_fp16)[name = string("q_55_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_3608_cast_fp16 = mul(x = k_55_cast_fp16, y = k_55_cast_fp16)[name = string("op_3608_cast_fp16")];
            tensor<int32, [1]> var_3613_axes_0 = const()[name = string("op_3613_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3613_keep_dims_0 = const()[name = string("op_3613_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 2, 1]> var_3613_cast_fp16 = reduce_mean(axes = var_3613_axes_0, keep_dims = var_3613_keep_dims_0, x = var_3608_cast_fp16)[name = string("op_3613_cast_fp16")];
            fp16 var_3614_promoted_to_fp16 = const()[name = string("op_3614_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 64, 2, 1]> var_3615_cast_fp16 = mul(x = var_3613_cast_fp16, y = var_3614_promoted_to_fp16)[name = string("op_3615_cast_fp16")];
            fp16 var_3617_to_fp16 = const()[name = string("op_3617_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 64, 2, 1]> var_3618_cast_fp16 = add(x = var_3615_cast_fp16, y = var_3617_to_fp16)[name = string("op_3618_cast_fp16")];
            fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 2, 1]> var_3619_cast_fp16 = rsqrt(epsilon = var_3619_epsilon_0, x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> k_57_cast_fp16 = mul(x = k_55_cast_fp16, y = var_3619_cast_fp16)[name = string("k_57_cast_fp16")];
            tensor<int32, [4]> var_3625 = const()[name = string("op_3625"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_3638_axes_0 = const()[name = string("op_3638_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 128]> var_3626_cast_fp16 = transpose(perm = var_3625, x = q_55_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 2, 1, 64, 128]> var_3638_cast_fp16 = expand_dims(axes = var_3638_axes_0, x = var_3626_cast_fp16)[name = string("op_3638_cast_fp16")];
            tensor<int32, [5]> var_3646_reps_0 = const()[name = string("op_3646_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 64, 128]> var_3646_cast_fp16 = tile(reps = var_3646_reps_0, x = var_3638_cast_fp16)[name = string("op_3646_cast_fp16")];
            tensor<int32, [4]> var_3651 = const()[name = string("op_3651"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> q_57_cast_fp16 = reshape(shape = var_3651, x = var_3646_cast_fp16)[name = string("q_57_cast_fp16")];
            tensor<int32, [4]> var_3657 = const()[name = string("op_3657"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_3670_axes_0 = const()[name = string("op_3670_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 128]> var_3658_cast_fp16 = transpose(perm = var_3657, x = k_57_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 2, 1, 64, 128]> var_3670_cast_fp16 = expand_dims(axes = var_3670_axes_0, x = var_3658_cast_fp16)[name = string("op_3670_cast_fp16")];
            tensor<int32, [5]> var_3678_reps_0 = const()[name = string("op_3678_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 64, 128]> var_3678_cast_fp16 = tile(reps = var_3678_reps_0, x = var_3670_cast_fp16)[name = string("op_3678_cast_fp16")];
            tensor<int32, [4]> var_3683 = const()[name = string("op_3683"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> k_59_cast_fp16 = reshape(shape = var_3683, x = var_3678_cast_fp16)[name = string("k_59_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_3685_cast_fp16 = sigmoid(x = b_7_cast_fp16)[name = string("op_3685_cast_fp16")];
            tensor<int32, [3]> var_3689 = const()[name = string("op_3689"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_4_dt_bias_to_fp16 = const()[name = string("layers_4_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 64, 4]> input_63_cast_fp16 = add(x = a_7_cast_fp16, y = layers_4_dt_bias_to_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_3693_cast_fp16 = relu(x = input_63_cast_fp16)[name = string("op_3693_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_3694_cast_fp16 = abs(x = input_63_cast_fp16)[name = string("op_3694_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 64, 4]> var_3695_cast_fp16 = mul(x = var_3694_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_3695_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_3696_cast_fp16 = exp(x = var_3695_cast_fp16)[name = string("op_3696_cast_fp16")];
            fp16 var_3698_to_fp16 = const()[name = string("op_3698_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 64, 4]> var_3699_cast_fp16 = add(x = var_3696_cast_fp16, y = var_3698_to_fp16)[name = string("op_3699_cast_fp16")];
            fp32 var_3700_epsilon_0 = const()[name = string("op_3700_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 64, 4]> var_3700_cast_fp16 = log(epsilon = var_3700_epsilon_0, x = var_3699_cast_fp16)[name = string("op_3700_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_3702_cast_fp16 = add(x = var_3693_cast_fp16, y = var_3700_cast_fp16)[name = string("op_3702_cast_fp16")];
            tensor<fp16, [4]> layers_4_negA_to_fp16 = const()[name = string("layers_4_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.348p+3, -0x1.734p+2, -0x1.5a8p-1, -0x1.264p+3])];
            tensor<fp16, [1, 64, 4]> var_3703_cast_fp16 = mul(x = layers_4_negA_to_fp16, y = var_3702_cast_fp16)[name = string("op_3703_cast_fp16")];
            tensor<int32, [3]> var_3707 = const()[name = string("op_3707"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_3710_axes_0 = const()[name = string("op_3710_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 64]> beta_7_cast_fp16 = transpose(perm = var_3689, x = var_3685_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 4, 64, 1]> var_3710_cast_fp16 = expand_dims(axes = var_3710_axes_0, x = beta_7_cast_fp16)[name = string("op_3710_cast_fp16")];
            fp16 var_3711_to_fp16 = const()[name = string("op_3711_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 64, 1]> var_3712_cast_fp16 = mul(x = var_3710_cast_fp16, y = var_3711_to_fp16)[name = string("op_3712_cast_fp16")];
            tensor<fp16, [1, 4, 64, 128]> v_11_cast_fp16 = transpose(perm = var_3587, x = var_3582_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 4, 64, 128]> vb_31_cast_fp16 = mul(x = v_11_cast_fp16, y = var_3712_cast_fp16)[name = string("vb_31_cast_fp16")];
            tensor<int32, [4]> q_59_begin_0 = const()[name = string("q_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_59_end_0 = const()[name = string("q_59_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_59_end_mask_0 = const()[name = string("q_59_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_59_cast_fp16 = slice_by_index(begin = q_59_begin_0, end = q_59_end_0, end_mask = q_59_end_mask_0, x = q_57_cast_fp16)[name = string("q_59_cast_fp16")];
            tensor<int32, [4]> k_61_begin_0 = const()[name = string("k_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_61_end_0 = const()[name = string("k_61_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_61_end_mask_0 = const()[name = string("k_61_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, x = k_59_cast_fp16)[name = string("k_61_cast_fp16")];
            tensor<fp16, [1, 4, 64, 128]> var_3746_cast_fp16 = mul(x = k_59_cast_fp16, y = var_3710_cast_fp16)[name = string("op_3746_cast_fp16")];
            tensor<int32, [4]> kb_25_begin_0 = const()[name = string("kb_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_25_end_0 = const()[name = string("kb_25_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_25_end_mask_0 = const()[name = string("kb_25_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_25_cast_fp16 = slice_by_index(begin = kb_25_begin_0, end = kb_25_end_0, end_mask = kb_25_end_mask_0, x = var_3746_cast_fp16)[name = string("kb_25_cast_fp16")];
            tensor<int32, [4]> vb_33_begin_0 = const()[name = string("vb_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_33_end_0 = const()[name = string("vb_33_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_33_end_mask_0 = const()[name = string("vb_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_33_cast_fp16 = slice_by_index(begin = vb_33_begin_0, end = vb_33_end_0, end_mask = vb_33_end_mask_0, x = vb_31_cast_fp16)[name = string("vb_33_cast_fp16")];
            tensor<int32, [3]> gdec_33_begin_0 = const()[name = string("gdec_33_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_33_end_0 = const()[name = string("gdec_33_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_33_end_mask_0 = const()[name = string("gdec_33_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 64]> gdec_31_cast_fp16 = transpose(perm = var_3707, x = var_3703_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 4, 16]> gdec_33_cast_fp16 = slice_by_index(begin = gdec_33_begin_0, end = gdec_33_end_0, end_mask = gdec_33_end_mask_0, x = gdec_31_cast_fp16)[name = string("gdec_33_cast_fp16")];
            tensor<int32, [2]> var_3794 = const()[name = string("op_3794"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_3795_cast_fp16 = reshape(shape = var_3794, x = gdec_33_cast_fp16)[name = string("op_3795_cast_fp16")];
            tensor<fp16, [4, 256]> var_3796_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3795_cast_fp16)[name = string("op_3796_cast_fp16")];
            tensor<int32, [4]> var_3801 = const()[name = string("op_3801"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_25_cast_fp16 = reshape(shape = var_3801, x = var_3796_cast_fp16)[name = string("P_25_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3803_cast_fp16 = exp(x = P_25_cast_fp16)[name = string("op_3803_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_25_cast_fp16 = mul(x = var_3803_cast_fp16, y = tril_incl_to_fp16)[name = string("D_25_cast_fp16")];
            bool kk_25_transpose_x_1 = const()[name = string("kk_25_transpose_x_1"), val = bool(false)];
            bool kk_25_transpose_y_1 = const()[name = string("kk_25_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_25_cast_fp16 = matmul(transpose_x = kk_25_transpose_x_1, transpose_y = kk_25_transpose_y_1, x = kb_25_cast_fp16, y = k_61_cast_fp16)[name = string("kk_25_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_3809_cast_fp16 = mul(x = kk_25_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_3809_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3810_cast_fp16 = mul(x = var_3809_cast_fp16, y = D_25_cast_fp16)[name = string("op_3810_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_25_cast_fp16 = mul(x = var_3810_cast_fp16, y = cum_tail_to_fp16)[name = string("N_25_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3812_cast_fp16 = mul(x = D_25_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3812_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_25_cast_fp16 = mul(x = kk_25_cast_fp16, y = var_3812_cast_fp16)[name = string("Na_25_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_97_cast_fp16 = add(x = eye_to_fp16, y = N_25_cast_fp16)[name = string("Minv_97_cast_fp16")];
            bool Np_73_transpose_x_0 = const()[name = string("Np_73_transpose_x_0"), val = bool(false)];
            bool Np_73_transpose_y_0 = const()[name = string("Np_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_73_cast_fp16 = matmul(transpose_x = Np_73_transpose_x_0, transpose_y = Np_73_transpose_y_0, x = N_25_cast_fp16, y = Na_25_cast_fp16)[name = string("Np_73_cast_fp16")];
            bool Npa_49_transpose_x_0 = const()[name = string("Npa_49_transpose_x_0"), val = bool(false)];
            bool Npa_49_transpose_y_0 = const()[name = string("Npa_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_49_cast_fp16 = matmul(transpose_x = Npa_49_transpose_x_0, transpose_y = Npa_49_transpose_y_0, x = Na_25_cast_fp16, y = N_25_cast_fp16)[name = string("Npa_49_cast_fp16")];
            bool var_3818_transpose_x_0 = const()[name = string("op_3818_transpose_x_0"), val = bool(false)];
            bool var_3818_transpose_y_0 = const()[name = string("op_3818_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3818_cast_fp16 = matmul(transpose_x = var_3818_transpose_x_0, transpose_y = var_3818_transpose_y_0, x = Minv_97_cast_fp16, y = Np_73_cast_fp16)[name = string("op_3818_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_99_cast_fp16 = add(x = Minv_97_cast_fp16, y = var_3818_cast_fp16)[name = string("Minv_99_cast_fp16")];
            bool Np_75_transpose_x_0 = const()[name = string("Np_75_transpose_x_0"), val = bool(false)];
            bool Np_75_transpose_y_0 = const()[name = string("Np_75_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_75_cast_fp16 = matmul(transpose_x = Np_75_transpose_x_0, transpose_y = Np_75_transpose_y_0, x = Np_73_cast_fp16, y = Npa_49_cast_fp16)[name = string("Np_75_cast_fp16")];
            bool Npa_51_transpose_x_0 = const()[name = string("Npa_51_transpose_x_0"), val = bool(false)];
            bool Npa_51_transpose_y_0 = const()[name = string("Npa_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_51_cast_fp16 = matmul(transpose_x = Npa_51_transpose_x_0, transpose_y = Npa_51_transpose_y_0, x = Npa_49_cast_fp16, y = Np_73_cast_fp16)[name = string("Npa_51_cast_fp16")];
            bool var_3823_transpose_x_0 = const()[name = string("op_3823_transpose_x_0"), val = bool(false)];
            bool var_3823_transpose_y_0 = const()[name = string("op_3823_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3823_cast_fp16 = matmul(transpose_x = var_3823_transpose_x_0, transpose_y = var_3823_transpose_y_0, x = Minv_99_cast_fp16, y = Np_75_cast_fp16)[name = string("op_3823_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_101_cast_fp16 = add(x = Minv_99_cast_fp16, y = var_3823_cast_fp16)[name = string("Minv_101_cast_fp16")];
            bool Np_77_transpose_x_0 = const()[name = string("Np_77_transpose_x_0"), val = bool(false)];
            bool Np_77_transpose_y_0 = const()[name = string("Np_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_77_cast_fp16 = matmul(transpose_x = Np_77_transpose_x_0, transpose_y = Np_77_transpose_y_0, x = Np_75_cast_fp16, y = Npa_51_cast_fp16)[name = string("Np_77_cast_fp16")];
            bool var_3827_transpose_x_0 = const()[name = string("op_3827_transpose_x_0"), val = bool(false)];
            bool var_3827_transpose_y_0 = const()[name = string("op_3827_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3827_cast_fp16 = matmul(transpose_x = var_3827_transpose_x_0, transpose_y = var_3827_transpose_y_0, x = Minv_101_cast_fp16, y = Np_77_cast_fp16)[name = string("op_3827_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_103_cast_fp16 = add(x = Minv_101_cast_fp16, y = var_3827_cast_fp16)[name = string("Minv_103_cast_fp16")];
            bool u_25_transpose_x_0 = const()[name = string("u_25_transpose_x_0"), val = bool(false)];
            bool u_25_transpose_y_0 = const()[name = string("u_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_25_cast_fp16 = matmul(transpose_x = u_25_transpose_x_0, transpose_y = u_25_transpose_y_0, x = Minv_103_cast_fp16, y = vb_33_cast_fp16)[name = string("u_25_cast_fp16")];
            bool var_3834_transpose_x_1 = const()[name = string("op_3834_transpose_x_1"), val = bool(false)];
            bool var_3834_transpose_y_1 = const()[name = string("op_3834_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_3834_cast_fp16 = matmul(transpose_x = var_3834_transpose_x_1, transpose_y = var_3834_transpose_y_1, x = q_59_cast_fp16, y = k_61_cast_fp16)[name = string("op_3834_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_25_cast_fp16 = mul(x = var_3834_cast_fp16, y = D_25_cast_fp16)[name = string("A_25_cast_fp16")];
            bool var_3836_transpose_x_0 = const()[name = string("op_3836_transpose_x_0"), val = bool(false)];
            bool var_3836_transpose_y_0 = const()[name = string("op_3836_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3836_cast_fp16 = matmul(transpose_x = var_3836_transpose_x_0, transpose_y = var_3836_transpose_y_0, x = A_25_cast_fp16, y = u_25_cast_fp16)[name = string("op_3836_cast_fp16")];
            tensor<fp16, [16]> var_3837_bias_0_to_fp16 = const()[name = string("op_3837_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_3837_cast_fp16 = linear(bias = var_3837_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_33_cast_fp16)[name = string("op_3837_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_19_cast_fp16 = exp(x = var_3837_cast_fp16)[name = string("tail_19_cast_fp16")];
            tensor<int32, [1]> var_3840_axes_0 = const()[name = string("op_3840_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_3840_cast_fp16 = expand_dims(axes = var_3840_axes_0, x = tail_19_cast_fp16)[name = string("op_3840_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3841_cast_fp16 = mul(x = k_61_cast_fp16, y = var_3840_cast_fp16)[name = string("op_3841_cast_fp16")];
            bool S_19_transpose_x_1 = const()[name = string("S_19_transpose_x_1"), val = bool(true)];
            bool S_19_transpose_y_1 = const()[name = string("S_19_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_19_cast_fp16 = matmul(transpose_x = S_19_transpose_x_1, transpose_y = S_19_transpose_y_1, x = var_3841_cast_fp16, y = u_25_cast_fp16)[name = string("S_19_cast_fp16")];
            tensor<int32, [4]> q_61_begin_0 = const()[name = string("q_61_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_61_end_0 = const()[name = string("q_61_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> q_61_end_mask_0 = const()[name = string("q_61_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, x = q_57_cast_fp16)[name = string("q_61_cast_fp16")];
            tensor<int32, [4]> k_63_begin_0 = const()[name = string("k_63_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_63_end_0 = const()[name = string("k_63_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> k_63_end_mask_0 = const()[name = string("k_63_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_63_cast_fp16 = slice_by_index(begin = k_63_begin_0, end = k_63_end_0, end_mask = k_63_end_mask_0, x = k_59_cast_fp16)[name = string("k_63_cast_fp16")];
            tensor<int32, [4]> kb_27_begin_0 = const()[name = string("kb_27_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_27_end_0 = const()[name = string("kb_27_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> kb_27_end_mask_0 = const()[name = string("kb_27_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_27_cast_fp16 = slice_by_index(begin = kb_27_begin_0, end = kb_27_end_0, end_mask = kb_27_end_mask_0, x = var_3746_cast_fp16)[name = string("kb_27_cast_fp16")];
            tensor<int32, [4]> vb_35_begin_0 = const()[name = string("vb_35_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_35_end_0 = const()[name = string("vb_35_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> vb_35_end_mask_0 = const()[name = string("vb_35_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_35_cast_fp16 = slice_by_index(begin = vb_35_begin_0, end = vb_35_end_0, end_mask = vb_35_end_mask_0, x = vb_31_cast_fp16)[name = string("vb_35_cast_fp16")];
            tensor<int32, [3]> gdec_35_begin_0 = const()[name = string("gdec_35_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_35_end_0 = const()[name = string("gdec_35_end_0"), val = tensor<int32, [3]>([1, 4, 32])];
            tensor<bool, [3]> gdec_35_end_mask_0 = const()[name = string("gdec_35_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_35_cast_fp16 = slice_by_index(begin = gdec_35_begin_0, end = gdec_35_end_0, end_mask = gdec_35_end_mask_0, x = gdec_31_cast_fp16)[name = string("gdec_35_cast_fp16")];
            tensor<int32, [2]> var_3926 = const()[name = string("op_3926"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_3927_cast_fp16 = reshape(shape = var_3926, x = gdec_35_cast_fp16)[name = string("op_3927_cast_fp16")];
            tensor<fp16, [4, 256]> var_3928_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3927_cast_fp16)[name = string("op_3928_cast_fp16")];
            tensor<int32, [4]> var_3933 = const()[name = string("op_3933"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_27_cast_fp16 = reshape(shape = var_3933, x = var_3928_cast_fp16)[name = string("P_27_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3935_cast_fp16 = exp(x = P_27_cast_fp16)[name = string("op_3935_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_27_cast_fp16 = mul(x = var_3935_cast_fp16, y = tril_incl_to_fp16)[name = string("D_27_cast_fp16")];
            bool kk_27_transpose_x_1 = const()[name = string("kk_27_transpose_x_1"), val = bool(false)];
            bool kk_27_transpose_y_1 = const()[name = string("kk_27_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_27_cast_fp16 = matmul(transpose_x = kk_27_transpose_x_1, transpose_y = kk_27_transpose_y_1, x = kb_27_cast_fp16, y = k_63_cast_fp16)[name = string("kk_27_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_3941_cast_fp16 = mul(x = kk_27_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_3941_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3942_cast_fp16 = mul(x = var_3941_cast_fp16, y = D_27_cast_fp16)[name = string("op_3942_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_27_cast_fp16 = mul(x = var_3942_cast_fp16, y = cum_tail_to_fp16)[name = string("N_27_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3944_cast_fp16 = mul(x = D_27_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3944_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_27_cast_fp16 = mul(x = kk_27_cast_fp16, y = var_3944_cast_fp16)[name = string("Na_27_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_105_cast_fp16 = add(x = eye_to_fp16, y = N_27_cast_fp16)[name = string("Minv_105_cast_fp16")];
            bool Np_79_transpose_x_0 = const()[name = string("Np_79_transpose_x_0"), val = bool(false)];
            bool Np_79_transpose_y_0 = const()[name = string("Np_79_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_79_cast_fp16 = matmul(transpose_x = Np_79_transpose_x_0, transpose_y = Np_79_transpose_y_0, x = N_27_cast_fp16, y = Na_27_cast_fp16)[name = string("Np_79_cast_fp16")];
            bool Npa_53_transpose_x_0 = const()[name = string("Npa_53_transpose_x_0"), val = bool(false)];
            bool Npa_53_transpose_y_0 = const()[name = string("Npa_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_53_cast_fp16 = matmul(transpose_x = Npa_53_transpose_x_0, transpose_y = Npa_53_transpose_y_0, x = Na_27_cast_fp16, y = N_27_cast_fp16)[name = string("Npa_53_cast_fp16")];
            bool var_3950_transpose_x_0 = const()[name = string("op_3950_transpose_x_0"), val = bool(false)];
            bool var_3950_transpose_y_0 = const()[name = string("op_3950_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3950_cast_fp16 = matmul(transpose_x = var_3950_transpose_x_0, transpose_y = var_3950_transpose_y_0, x = Minv_105_cast_fp16, y = Np_79_cast_fp16)[name = string("op_3950_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_107_cast_fp16 = add(x = Minv_105_cast_fp16, y = var_3950_cast_fp16)[name = string("Minv_107_cast_fp16")];
            bool Np_81_transpose_x_0 = const()[name = string("Np_81_transpose_x_0"), val = bool(false)];
            bool Np_81_transpose_y_0 = const()[name = string("Np_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_81_cast_fp16 = matmul(transpose_x = Np_81_transpose_x_0, transpose_y = Np_81_transpose_y_0, x = Np_79_cast_fp16, y = Npa_53_cast_fp16)[name = string("Np_81_cast_fp16")];
            bool Npa_55_transpose_x_0 = const()[name = string("Npa_55_transpose_x_0"), val = bool(false)];
            bool Npa_55_transpose_y_0 = const()[name = string("Npa_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_55_cast_fp16 = matmul(transpose_x = Npa_55_transpose_x_0, transpose_y = Npa_55_transpose_y_0, x = Npa_53_cast_fp16, y = Np_79_cast_fp16)[name = string("Npa_55_cast_fp16")];
            bool var_3955_transpose_x_0 = const()[name = string("op_3955_transpose_x_0"), val = bool(false)];
            bool var_3955_transpose_y_0 = const()[name = string("op_3955_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3955_cast_fp16 = matmul(transpose_x = var_3955_transpose_x_0, transpose_y = var_3955_transpose_y_0, x = Minv_107_cast_fp16, y = Np_81_cast_fp16)[name = string("op_3955_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_109_cast_fp16 = add(x = Minv_107_cast_fp16, y = var_3955_cast_fp16)[name = string("Minv_109_cast_fp16")];
            bool Np_83_transpose_x_0 = const()[name = string("Np_83_transpose_x_0"), val = bool(false)];
            bool Np_83_transpose_y_0 = const()[name = string("Np_83_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_83_cast_fp16 = matmul(transpose_x = Np_83_transpose_x_0, transpose_y = Np_83_transpose_y_0, x = Np_81_cast_fp16, y = Npa_55_cast_fp16)[name = string("Np_83_cast_fp16")];
            bool var_3959_transpose_x_0 = const()[name = string("op_3959_transpose_x_0"), val = bool(false)];
            bool var_3959_transpose_y_0 = const()[name = string("op_3959_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3959_cast_fp16 = matmul(transpose_x = var_3959_transpose_x_0, transpose_y = var_3959_transpose_y_0, x = Minv_109_cast_fp16, y = Np_83_cast_fp16)[name = string("op_3959_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_111_cast_fp16 = add(x = Minv_109_cast_fp16, y = var_3959_cast_fp16)[name = string("Minv_111_cast_fp16")];
            bool u_27_transpose_x_0 = const()[name = string("u_27_transpose_x_0"), val = bool(false)];
            bool u_27_transpose_y_0 = const()[name = string("u_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_27_cast_fp16 = matmul(transpose_x = u_27_transpose_x_0, transpose_y = u_27_transpose_y_0, x = Minv_111_cast_fp16, y = vb_35_cast_fp16)[name = string("u_27_cast_fp16")];
            bool var_3966_transpose_x_1 = const()[name = string("op_3966_transpose_x_1"), val = bool(false)];
            bool var_3966_transpose_y_1 = const()[name = string("op_3966_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_3966_cast_fp16 = matmul(transpose_x = var_3966_transpose_x_1, transpose_y = var_3966_transpose_y_1, x = q_61_cast_fp16, y = k_63_cast_fp16)[name = string("op_3966_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_27_cast_fp16 = mul(x = var_3966_cast_fp16, y = D_27_cast_fp16)[name = string("A_27_cast_fp16")];
            tensor<fp16, [16]> var_3968_bias_0_to_fp16 = const()[name = string("op_3968_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_3968_cast_fp16 = linear(bias = var_3968_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_35_cast_fp16)[name = string("op_3968_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_19_cast_fp16 = exp(x = var_3968_cast_fp16)[name = string("cum_19_cast_fp16")];
            tensor<int32, [1]> var_3971_axes_0 = const()[name = string("op_3971_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_3971_cast_fp16 = expand_dims(axes = var_3971_axes_0, x = cum_19_cast_fp16)[name = string("op_3971_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3972_cast_fp16 = mul(x = kb_27_cast_fp16, y = var_3971_cast_fp16)[name = string("op_3972_cast_fp16")];
            bool w_27_transpose_x_0 = const()[name = string("w_27_transpose_x_0"), val = bool(false)];
            bool w_27_transpose_y_0 = const()[name = string("w_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_27_cast_fp16 = matmul(transpose_x = w_27_transpose_x_0, transpose_y = w_27_transpose_y_0, x = Minv_111_cast_fp16, y = var_3972_cast_fp16)[name = string("w_27_cast_fp16")];
            bool var_3974_transpose_x_0 = const()[name = string("op_3974_transpose_x_0"), val = bool(false)];
            bool var_3974_transpose_y_0 = const()[name = string("op_3974_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3974_cast_fp16 = matmul(transpose_x = var_3974_transpose_x_0, transpose_y = var_3974_transpose_y_0, x = w_27_cast_fp16, y = S_19_cast_fp16)[name = string("op_3974_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_19_cast_fp16 = sub(x = u_27_cast_fp16, y = var_3974_cast_fp16)[name = string("vnew_19_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3979_cast_fp16 = mul(x = q_61_cast_fp16, y = var_3971_cast_fp16)[name = string("op_3979_cast_fp16")];
            bool var_3980_transpose_x_0 = const()[name = string("op_3980_transpose_x_0"), val = bool(false)];
            bool var_3980_transpose_y_0 = const()[name = string("op_3980_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3980_cast_fp16 = matmul(transpose_x = var_3980_transpose_x_0, transpose_y = var_3980_transpose_y_0, x = var_3979_cast_fp16, y = S_19_cast_fp16)[name = string("op_3980_cast_fp16")];
            bool var_3981_transpose_x_0 = const()[name = string("op_3981_transpose_x_0"), val = bool(false)];
            bool var_3981_transpose_y_0 = const()[name = string("op_3981_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3981_cast_fp16 = matmul(transpose_x = var_3981_transpose_x_0, transpose_y = var_3981_transpose_y_0, x = A_27_cast_fp16, y = vnew_19_cast_fp16)[name = string("op_3981_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3983_cast_fp16 = add(x = var_3980_cast_fp16, y = var_3981_cast_fp16)[name = string("op_3983_cast_fp16")];
            tensor<fp16, [16]> var_3984_bias_0_to_fp16 = const()[name = string("op_3984_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_3984_cast_fp16 = linear(bias = var_3984_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_35_cast_fp16)[name = string("op_3984_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_21_cast_fp16 = exp(x = var_3984_cast_fp16)[name = string("tail_21_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_41_to_fp16 = const()[name = string("transpose_41_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_3986_bias_0_to_fp16 = const()[name = string("op_3986_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_3986_cast_fp16 = linear(bias = var_3986_bias_0_to_fp16, weight = transpose_41_to_fp16, x = gdec_35_cast_fp16)[name = string("op_3986_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_3987_cast_fp16 = exp(x = var_3986_cast_fp16)[name = string("op_3987_cast_fp16")];
            tensor<int32, [1]> last_15_axes_0 = const()[name = string("last_15_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_15_cast_fp16 = expand_dims(axes = last_15_axes_0, x = var_3987_cast_fp16)[name = string("last_15_cast_fp16")];
            tensor<int32, [1]> var_3991_axes_0 = const()[name = string("op_3991_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_3991_cast_fp16 = expand_dims(axes = var_3991_axes_0, x = tail_21_cast_fp16)[name = string("op_3991_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3992_cast_fp16 = mul(x = k_63_cast_fp16, y = var_3991_cast_fp16)[name = string("op_3992_cast_fp16")];
            bool upd_13_transpose_x_1 = const()[name = string("upd_13_transpose_x_1"), val = bool(true)];
            bool upd_13_transpose_y_1 = const()[name = string("upd_13_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_13_cast_fp16 = matmul(transpose_x = upd_13_transpose_x_1, transpose_y = upd_13_transpose_y_1, x = var_3992_cast_fp16, y = vnew_19_cast_fp16)[name = string("upd_13_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_3997_cast_fp16 = mul(x = S_19_cast_fp16, y = last_15_cast_fp16)[name = string("op_3997_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_21_cast_fp16 = add(x = var_3997_cast_fp16, y = upd_13_cast_fp16)[name = string("S_21_cast_fp16")];
            tensor<int32, [4]> q_63_begin_0 = const()[name = string("q_63_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_63_end_0 = const()[name = string("q_63_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> q_63_end_mask_0 = const()[name = string("q_63_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_63_cast_fp16 = slice_by_index(begin = q_63_begin_0, end = q_63_end_0, end_mask = q_63_end_mask_0, x = q_57_cast_fp16)[name = string("q_63_cast_fp16")];
            tensor<int32, [4]> k_65_begin_0 = const()[name = string("k_65_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_65_end_0 = const()[name = string("k_65_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> k_65_end_mask_0 = const()[name = string("k_65_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_65_cast_fp16 = slice_by_index(begin = k_65_begin_0, end = k_65_end_0, end_mask = k_65_end_mask_0, x = k_59_cast_fp16)[name = string("k_65_cast_fp16")];
            tensor<int32, [4]> kb_29_begin_0 = const()[name = string("kb_29_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_29_end_0 = const()[name = string("kb_29_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> kb_29_end_mask_0 = const()[name = string("kb_29_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_29_cast_fp16 = slice_by_index(begin = kb_29_begin_0, end = kb_29_end_0, end_mask = kb_29_end_mask_0, x = var_3746_cast_fp16)[name = string("kb_29_cast_fp16")];
            tensor<int32, [4]> vb_37_begin_0 = const()[name = string("vb_37_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_37_end_0 = const()[name = string("vb_37_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> vb_37_end_mask_0 = const()[name = string("vb_37_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_37_cast_fp16 = slice_by_index(begin = vb_37_begin_0, end = vb_37_end_0, end_mask = vb_37_end_mask_0, x = vb_31_cast_fp16)[name = string("vb_37_cast_fp16")];
            tensor<int32, [3]> gdec_37_begin_0 = const()[name = string("gdec_37_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_37_end_0 = const()[name = string("gdec_37_end_0"), val = tensor<int32, [3]>([1, 4, 48])];
            tensor<bool, [3]> gdec_37_end_mask_0 = const()[name = string("gdec_37_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_37_cast_fp16 = slice_by_index(begin = gdec_37_begin_0, end = gdec_37_end_0, end_mask = gdec_37_end_mask_0, x = gdec_31_cast_fp16)[name = string("gdec_37_cast_fp16")];
            tensor<int32, [2]> var_4080 = const()[name = string("op_4080"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_4081_cast_fp16 = reshape(shape = var_4080, x = gdec_37_cast_fp16)[name = string("op_4081_cast_fp16")];
            tensor<fp16, [4, 256]> var_4082_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_4081_cast_fp16)[name = string("op_4082_cast_fp16")];
            tensor<int32, [4]> var_4087 = const()[name = string("op_4087"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_29_cast_fp16 = reshape(shape = var_4087, x = var_4082_cast_fp16)[name = string("P_29_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4089_cast_fp16 = exp(x = P_29_cast_fp16)[name = string("op_4089_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_29_cast_fp16 = mul(x = var_4089_cast_fp16, y = tril_incl_to_fp16)[name = string("D_29_cast_fp16")];
            bool kk_29_transpose_x_1 = const()[name = string("kk_29_transpose_x_1"), val = bool(false)];
            bool kk_29_transpose_y_1 = const()[name = string("kk_29_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_29_cast_fp16 = matmul(transpose_x = kk_29_transpose_x_1, transpose_y = kk_29_transpose_y_1, x = kb_29_cast_fp16, y = k_65_cast_fp16)[name = string("kk_29_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_4095_cast_fp16 = mul(x = kk_29_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_4095_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4096_cast_fp16 = mul(x = var_4095_cast_fp16, y = D_29_cast_fp16)[name = string("op_4096_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_29_cast_fp16 = mul(x = var_4096_cast_fp16, y = cum_tail_to_fp16)[name = string("N_29_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4098_cast_fp16 = mul(x = D_29_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_4098_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_29_cast_fp16 = mul(x = kk_29_cast_fp16, y = var_4098_cast_fp16)[name = string("Na_29_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_113_cast_fp16 = add(x = eye_to_fp16, y = N_29_cast_fp16)[name = string("Minv_113_cast_fp16")];
            bool Np_85_transpose_x_0 = const()[name = string("Np_85_transpose_x_0"), val = bool(false)];
            bool Np_85_transpose_y_0 = const()[name = string("Np_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_85_cast_fp16 = matmul(transpose_x = Np_85_transpose_x_0, transpose_y = Np_85_transpose_y_0, x = N_29_cast_fp16, y = Na_29_cast_fp16)[name = string("Np_85_cast_fp16")];
            bool Npa_57_transpose_x_0 = const()[name = string("Npa_57_transpose_x_0"), val = bool(false)];
            bool Npa_57_transpose_y_0 = const()[name = string("Npa_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_57_cast_fp16 = matmul(transpose_x = Npa_57_transpose_x_0, transpose_y = Npa_57_transpose_y_0, x = Na_29_cast_fp16, y = N_29_cast_fp16)[name = string("Npa_57_cast_fp16")];
            bool var_4104_transpose_x_0 = const()[name = string("op_4104_transpose_x_0"), val = bool(false)];
            bool var_4104_transpose_y_0 = const()[name = string("op_4104_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4104_cast_fp16 = matmul(transpose_x = var_4104_transpose_x_0, transpose_y = var_4104_transpose_y_0, x = Minv_113_cast_fp16, y = Np_85_cast_fp16)[name = string("op_4104_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_115_cast_fp16 = add(x = Minv_113_cast_fp16, y = var_4104_cast_fp16)[name = string("Minv_115_cast_fp16")];
            bool Np_87_transpose_x_0 = const()[name = string("Np_87_transpose_x_0"), val = bool(false)];
            bool Np_87_transpose_y_0 = const()[name = string("Np_87_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_87_cast_fp16 = matmul(transpose_x = Np_87_transpose_x_0, transpose_y = Np_87_transpose_y_0, x = Np_85_cast_fp16, y = Npa_57_cast_fp16)[name = string("Np_87_cast_fp16")];
            bool Npa_59_transpose_x_0 = const()[name = string("Npa_59_transpose_x_0"), val = bool(false)];
            bool Npa_59_transpose_y_0 = const()[name = string("Npa_59_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_59_cast_fp16 = matmul(transpose_x = Npa_59_transpose_x_0, transpose_y = Npa_59_transpose_y_0, x = Npa_57_cast_fp16, y = Np_85_cast_fp16)[name = string("Npa_59_cast_fp16")];
            bool var_4109_transpose_x_0 = const()[name = string("op_4109_transpose_x_0"), val = bool(false)];
            bool var_4109_transpose_y_0 = const()[name = string("op_4109_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4109_cast_fp16 = matmul(transpose_x = var_4109_transpose_x_0, transpose_y = var_4109_transpose_y_0, x = Minv_115_cast_fp16, y = Np_87_cast_fp16)[name = string("op_4109_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_117_cast_fp16 = add(x = Minv_115_cast_fp16, y = var_4109_cast_fp16)[name = string("Minv_117_cast_fp16")];
            bool Np_89_transpose_x_0 = const()[name = string("Np_89_transpose_x_0"), val = bool(false)];
            bool Np_89_transpose_y_0 = const()[name = string("Np_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_89_cast_fp16 = matmul(transpose_x = Np_89_transpose_x_0, transpose_y = Np_89_transpose_y_0, x = Np_87_cast_fp16, y = Npa_59_cast_fp16)[name = string("Np_89_cast_fp16")];
            bool var_4113_transpose_x_0 = const()[name = string("op_4113_transpose_x_0"), val = bool(false)];
            bool var_4113_transpose_y_0 = const()[name = string("op_4113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4113_cast_fp16 = matmul(transpose_x = var_4113_transpose_x_0, transpose_y = var_4113_transpose_y_0, x = Minv_117_cast_fp16, y = Np_89_cast_fp16)[name = string("op_4113_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_119_cast_fp16 = add(x = Minv_117_cast_fp16, y = var_4113_cast_fp16)[name = string("Minv_119_cast_fp16")];
            bool u_29_transpose_x_0 = const()[name = string("u_29_transpose_x_0"), val = bool(false)];
            bool u_29_transpose_y_0 = const()[name = string("u_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_29_cast_fp16 = matmul(transpose_x = u_29_transpose_x_0, transpose_y = u_29_transpose_y_0, x = Minv_119_cast_fp16, y = vb_37_cast_fp16)[name = string("u_29_cast_fp16")];
            bool var_4120_transpose_x_1 = const()[name = string("op_4120_transpose_x_1"), val = bool(false)];
            bool var_4120_transpose_y_1 = const()[name = string("op_4120_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_4120_cast_fp16 = matmul(transpose_x = var_4120_transpose_x_1, transpose_y = var_4120_transpose_y_1, x = q_63_cast_fp16, y = k_65_cast_fp16)[name = string("op_4120_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_29_cast_fp16 = mul(x = var_4120_cast_fp16, y = D_29_cast_fp16)[name = string("A_29_cast_fp16")];
            tensor<fp16, [16]> var_4122_bias_0_to_fp16 = const()[name = string("op_4122_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_4122_cast_fp16 = linear(bias = var_4122_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_37_cast_fp16)[name = string("op_4122_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_21_cast_fp16 = exp(x = var_4122_cast_fp16)[name = string("cum_21_cast_fp16")];
            tensor<int32, [1]> var_4125_axes_0 = const()[name = string("op_4125_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_4125_cast_fp16 = expand_dims(axes = var_4125_axes_0, x = cum_21_cast_fp16)[name = string("op_4125_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4126_cast_fp16 = mul(x = kb_29_cast_fp16, y = var_4125_cast_fp16)[name = string("op_4126_cast_fp16")];
            bool w_29_transpose_x_0 = const()[name = string("w_29_transpose_x_0"), val = bool(false)];
            bool w_29_transpose_y_0 = const()[name = string("w_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_29_cast_fp16 = matmul(transpose_x = w_29_transpose_x_0, transpose_y = w_29_transpose_y_0, x = Minv_119_cast_fp16, y = var_4126_cast_fp16)[name = string("w_29_cast_fp16")];
            bool var_4128_transpose_x_0 = const()[name = string("op_4128_transpose_x_0"), val = bool(false)];
            bool var_4128_transpose_y_0 = const()[name = string("op_4128_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4128_cast_fp16 = matmul(transpose_x = var_4128_transpose_x_0, transpose_y = var_4128_transpose_y_0, x = w_29_cast_fp16, y = S_21_cast_fp16)[name = string("op_4128_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_21_cast_fp16 = sub(x = u_29_cast_fp16, y = var_4128_cast_fp16)[name = string("vnew_21_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4133_cast_fp16 = mul(x = q_63_cast_fp16, y = var_4125_cast_fp16)[name = string("op_4133_cast_fp16")];
            bool var_4134_transpose_x_0 = const()[name = string("op_4134_transpose_x_0"), val = bool(false)];
            bool var_4134_transpose_y_0 = const()[name = string("op_4134_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4134_cast_fp16 = matmul(transpose_x = var_4134_transpose_x_0, transpose_y = var_4134_transpose_y_0, x = var_4133_cast_fp16, y = S_21_cast_fp16)[name = string("op_4134_cast_fp16")];
            bool var_4135_transpose_x_0 = const()[name = string("op_4135_transpose_x_0"), val = bool(false)];
            bool var_4135_transpose_y_0 = const()[name = string("op_4135_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4135_cast_fp16 = matmul(transpose_x = var_4135_transpose_x_0, transpose_y = var_4135_transpose_y_0, x = A_29_cast_fp16, y = vnew_21_cast_fp16)[name = string("op_4135_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4137_cast_fp16 = add(x = var_4134_cast_fp16, y = var_4135_cast_fp16)[name = string("op_4137_cast_fp16")];
            tensor<fp16, [16]> var_4138_bias_0_to_fp16 = const()[name = string("op_4138_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_4138_cast_fp16 = linear(bias = var_4138_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_37_cast_fp16)[name = string("op_4138_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_23_cast_fp16 = exp(x = var_4138_cast_fp16)[name = string("tail_23_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_45_to_fp16 = const()[name = string("transpose_45_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_4140_bias_0_to_fp16 = const()[name = string("op_4140_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_4140_cast_fp16 = linear(bias = var_4140_bias_0_to_fp16, weight = transpose_45_to_fp16, x = gdec_37_cast_fp16)[name = string("op_4140_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_4141_cast_fp16 = exp(x = var_4140_cast_fp16)[name = string("op_4141_cast_fp16")];
            tensor<int32, [1]> last_17_axes_0 = const()[name = string("last_17_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_17_cast_fp16 = expand_dims(axes = last_17_axes_0, x = var_4141_cast_fp16)[name = string("last_17_cast_fp16")];
            tensor<int32, [1]> var_4145_axes_0 = const()[name = string("op_4145_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_4145_cast_fp16 = expand_dims(axes = var_4145_axes_0, x = tail_23_cast_fp16)[name = string("op_4145_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4146_cast_fp16 = mul(x = k_65_cast_fp16, y = var_4145_cast_fp16)[name = string("op_4146_cast_fp16")];
            bool upd_15_transpose_x_1 = const()[name = string("upd_15_transpose_x_1"), val = bool(true)];
            bool upd_15_transpose_y_1 = const()[name = string("upd_15_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_15_cast_fp16 = matmul(transpose_x = upd_15_transpose_x_1, transpose_y = upd_15_transpose_y_1, x = var_4146_cast_fp16, y = vnew_21_cast_fp16)[name = string("upd_15_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_4151_cast_fp16 = mul(x = S_21_cast_fp16, y = last_17_cast_fp16)[name = string("op_4151_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_23_cast_fp16 = add(x = var_4151_cast_fp16, y = upd_15_cast_fp16)[name = string("S_23_cast_fp16")];
            tensor<int32, [4]> q_65_begin_0 = const()[name = string("q_65_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_65_end_0 = const()[name = string("q_65_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_65_end_mask_0 = const()[name = string("q_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_65_cast_fp16 = slice_by_index(begin = q_65_begin_0, end = q_65_end_0, end_mask = q_65_end_mask_0, x = q_57_cast_fp16)[name = string("q_65_cast_fp16")];
            tensor<int32, [4]> k_67_begin_0 = const()[name = string("k_67_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_67_end_0 = const()[name = string("k_67_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_67_end_mask_0 = const()[name = string("k_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_67_cast_fp16 = slice_by_index(begin = k_67_begin_0, end = k_67_end_0, end_mask = k_67_end_mask_0, x = k_59_cast_fp16)[name = string("k_67_cast_fp16")];
            tensor<int32, [4]> kb_31_begin_0 = const()[name = string("kb_31_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_31_end_0 = const()[name = string("kb_31_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_31_end_mask_0 = const()[name = string("kb_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_31_cast_fp16 = slice_by_index(begin = kb_31_begin_0, end = kb_31_end_0, end_mask = kb_31_end_mask_0, x = var_3746_cast_fp16)[name = string("kb_31_cast_fp16")];
            tensor<int32, [4]> vb_39_begin_0 = const()[name = string("vb_39_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_39_end_0 = const()[name = string("vb_39_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_39_end_mask_0 = const()[name = string("vb_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_39_cast_fp16 = slice_by_index(begin = vb_39_begin_0, end = vb_39_end_0, end_mask = vb_39_end_mask_0, x = vb_31_cast_fp16)[name = string("vb_39_cast_fp16")];
            tensor<int32, [3]> gdec_39_begin_0 = const()[name = string("gdec_39_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_39_end_0 = const()[name = string("gdec_39_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_39_end_mask_0 = const()[name = string("gdec_39_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_39_cast_fp16 = slice_by_index(begin = gdec_39_begin_0, end = gdec_39_end_0, end_mask = gdec_39_end_mask_0, x = gdec_31_cast_fp16)[name = string("gdec_39_cast_fp16")];
            tensor<int32, [2]> var_4234 = const()[name = string("op_4234"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_4235_cast_fp16 = reshape(shape = var_4234, x = gdec_39_cast_fp16)[name = string("op_4235_cast_fp16")];
            tensor<fp16, [4, 256]> var_4236_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_4235_cast_fp16)[name = string("op_4236_cast_fp16")];
            tensor<int32, [4]> var_4241 = const()[name = string("op_4241"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_31_cast_fp16 = reshape(shape = var_4241, x = var_4236_cast_fp16)[name = string("P_31_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4243_cast_fp16 = exp(x = P_31_cast_fp16)[name = string("op_4243_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_31_cast_fp16 = mul(x = var_4243_cast_fp16, y = tril_incl_to_fp16)[name = string("D_31_cast_fp16")];
            bool kk_31_transpose_x_1 = const()[name = string("kk_31_transpose_x_1"), val = bool(false)];
            bool kk_31_transpose_y_1 = const()[name = string("kk_31_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_31_cast_fp16 = matmul(transpose_x = kk_31_transpose_x_1, transpose_y = kk_31_transpose_y_1, x = kb_31_cast_fp16, y = k_67_cast_fp16)[name = string("kk_31_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_4249_cast_fp16 = mul(x = kk_31_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_4249_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4250_cast_fp16 = mul(x = var_4249_cast_fp16, y = D_31_cast_fp16)[name = string("op_4250_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_31_cast_fp16 = mul(x = var_4250_cast_fp16, y = cum_tail_to_fp16)[name = string("N_31_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4252_cast_fp16 = mul(x = D_31_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_4252_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_31_cast_fp16 = mul(x = kk_31_cast_fp16, y = var_4252_cast_fp16)[name = string("Na_31_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_121_cast_fp16 = add(x = eye_to_fp16, y = N_31_cast_fp16)[name = string("Minv_121_cast_fp16")];
            bool Np_91_transpose_x_0 = const()[name = string("Np_91_transpose_x_0"), val = bool(false)];
            bool Np_91_transpose_y_0 = const()[name = string("Np_91_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_91_cast_fp16 = matmul(transpose_x = Np_91_transpose_x_0, transpose_y = Np_91_transpose_y_0, x = N_31_cast_fp16, y = Na_31_cast_fp16)[name = string("Np_91_cast_fp16")];
            bool Npa_61_transpose_x_0 = const()[name = string("Npa_61_transpose_x_0"), val = bool(false)];
            bool Npa_61_transpose_y_0 = const()[name = string("Npa_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_61_cast_fp16 = matmul(transpose_x = Npa_61_transpose_x_0, transpose_y = Npa_61_transpose_y_0, x = Na_31_cast_fp16, y = N_31_cast_fp16)[name = string("Npa_61_cast_fp16")];
            bool var_4258_transpose_x_0 = const()[name = string("op_4258_transpose_x_0"), val = bool(false)];
            bool var_4258_transpose_y_0 = const()[name = string("op_4258_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4258_cast_fp16 = matmul(transpose_x = var_4258_transpose_x_0, transpose_y = var_4258_transpose_y_0, x = Minv_121_cast_fp16, y = Np_91_cast_fp16)[name = string("op_4258_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_123_cast_fp16 = add(x = Minv_121_cast_fp16, y = var_4258_cast_fp16)[name = string("Minv_123_cast_fp16")];
            bool Np_93_transpose_x_0 = const()[name = string("Np_93_transpose_x_0"), val = bool(false)];
            bool Np_93_transpose_y_0 = const()[name = string("Np_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_93_cast_fp16 = matmul(transpose_x = Np_93_transpose_x_0, transpose_y = Np_93_transpose_y_0, x = Np_91_cast_fp16, y = Npa_61_cast_fp16)[name = string("Np_93_cast_fp16")];
            bool Npa_63_transpose_x_0 = const()[name = string("Npa_63_transpose_x_0"), val = bool(false)];
            bool Npa_63_transpose_y_0 = const()[name = string("Npa_63_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_63_cast_fp16 = matmul(transpose_x = Npa_63_transpose_x_0, transpose_y = Npa_63_transpose_y_0, x = Npa_61_cast_fp16, y = Np_91_cast_fp16)[name = string("Npa_63_cast_fp16")];
            bool var_4263_transpose_x_0 = const()[name = string("op_4263_transpose_x_0"), val = bool(false)];
            bool var_4263_transpose_y_0 = const()[name = string("op_4263_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4263_cast_fp16 = matmul(transpose_x = var_4263_transpose_x_0, transpose_y = var_4263_transpose_y_0, x = Minv_123_cast_fp16, y = Np_93_cast_fp16)[name = string("op_4263_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_125_cast_fp16 = add(x = Minv_123_cast_fp16, y = var_4263_cast_fp16)[name = string("Minv_125_cast_fp16")];
            bool Np_95_transpose_x_0 = const()[name = string("Np_95_transpose_x_0"), val = bool(false)];
            bool Np_95_transpose_y_0 = const()[name = string("Np_95_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_95_cast_fp16 = matmul(transpose_x = Np_95_transpose_x_0, transpose_y = Np_95_transpose_y_0, x = Np_93_cast_fp16, y = Npa_63_cast_fp16)[name = string("Np_95_cast_fp16")];
            bool var_4267_transpose_x_0 = const()[name = string("op_4267_transpose_x_0"), val = bool(false)];
            bool var_4267_transpose_y_0 = const()[name = string("op_4267_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4267_cast_fp16 = matmul(transpose_x = var_4267_transpose_x_0, transpose_y = var_4267_transpose_y_0, x = Minv_125_cast_fp16, y = Np_95_cast_fp16)[name = string("op_4267_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_127_cast_fp16 = add(x = Minv_125_cast_fp16, y = var_4267_cast_fp16)[name = string("Minv_127_cast_fp16")];
            bool u_31_transpose_x_0 = const()[name = string("u_31_transpose_x_0"), val = bool(false)];
            bool u_31_transpose_y_0 = const()[name = string("u_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_31_cast_fp16 = matmul(transpose_x = u_31_transpose_x_0, transpose_y = u_31_transpose_y_0, x = Minv_127_cast_fp16, y = vb_39_cast_fp16)[name = string("u_31_cast_fp16")];
            bool var_4274_transpose_x_1 = const()[name = string("op_4274_transpose_x_1"), val = bool(false)];
            bool var_4274_transpose_y_1 = const()[name = string("op_4274_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_4274_cast_fp16 = matmul(transpose_x = var_4274_transpose_x_1, transpose_y = var_4274_transpose_y_1, x = q_65_cast_fp16, y = k_67_cast_fp16)[name = string("op_4274_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_31_cast_fp16 = mul(x = var_4274_cast_fp16, y = D_31_cast_fp16)[name = string("A_31_cast_fp16")];
            tensor<fp16, [16]> var_4276_bias_0_to_fp16 = const()[name = string("op_4276_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_4276_cast_fp16 = linear(bias = var_4276_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_39_cast_fp16)[name = string("op_4276_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_23_cast_fp16 = exp(x = var_4276_cast_fp16)[name = string("cum_23_cast_fp16")];
            tensor<int32, [1]> var_4279_axes_0 = const()[name = string("op_4279_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_4279_cast_fp16 = expand_dims(axes = var_4279_axes_0, x = cum_23_cast_fp16)[name = string("op_4279_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4280_cast_fp16 = mul(x = kb_31_cast_fp16, y = var_4279_cast_fp16)[name = string("op_4280_cast_fp16")];
            bool w_31_transpose_x_0 = const()[name = string("w_31_transpose_x_0"), val = bool(false)];
            bool w_31_transpose_y_0 = const()[name = string("w_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_31_cast_fp16 = matmul(transpose_x = w_31_transpose_x_0, transpose_y = w_31_transpose_y_0, x = Minv_127_cast_fp16, y = var_4280_cast_fp16)[name = string("w_31_cast_fp16")];
            bool var_4282_transpose_x_0 = const()[name = string("op_4282_transpose_x_0"), val = bool(false)];
            bool var_4282_transpose_y_0 = const()[name = string("op_4282_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4282_cast_fp16 = matmul(transpose_x = var_4282_transpose_x_0, transpose_y = var_4282_transpose_y_0, x = w_31_cast_fp16, y = S_23_cast_fp16)[name = string("op_4282_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_23_cast_fp16 = sub(x = u_31_cast_fp16, y = var_4282_cast_fp16)[name = string("vnew_23_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4287_cast_fp16 = mul(x = q_65_cast_fp16, y = var_4279_cast_fp16)[name = string("op_4287_cast_fp16")];
            bool var_4288_transpose_x_0 = const()[name = string("op_4288_transpose_x_0"), val = bool(false)];
            bool var_4288_transpose_y_0 = const()[name = string("op_4288_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4288_cast_fp16 = matmul(transpose_x = var_4288_transpose_x_0, transpose_y = var_4288_transpose_y_0, x = var_4287_cast_fp16, y = S_23_cast_fp16)[name = string("op_4288_cast_fp16")];
            bool var_4289_transpose_x_0 = const()[name = string("op_4289_transpose_x_0"), val = bool(false)];
            bool var_4289_transpose_y_0 = const()[name = string("op_4289_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4289_cast_fp16 = matmul(transpose_x = var_4289_transpose_x_0, transpose_y = var_4289_transpose_y_0, x = A_31_cast_fp16, y = vnew_23_cast_fp16)[name = string("op_4289_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_7_cast_fp16 = add(x = var_4288_cast_fp16, y = var_4289_cast_fp16)[name = string("oc_7_cast_fp16")];
            int32 var_4293 = const()[name = string("op_4293"), val = int32(2)];
            bool o_21_interleave_0 = const()[name = string("o_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 128]> o_21_cast_fp16 = concat(axis = var_4293, interleave = o_21_interleave_0, values = (var_3836_cast_fp16, var_3983_cast_fp16, var_4137_cast_fp16, oc_7_cast_fp16))[name = string("o_21_cast_fp16")];
            tensor<int32, [4]> var_4299 = const()[name = string("op_4299"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_4305 = const()[name = string("op_4305"), val = tensor<int32, [4]>([1, 64, 4, 128])];
            tensor<fp16, [1, 64, 4, 128]> input_65_cast_fp16 = reshape(shape = var_4305, x = z_7_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> o_23_cast_fp16 = transpose(perm = var_4299, x = o_21_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 64, 4, 128]> var_4307_cast_fp16 = mul(x = o_23_cast_fp16, y = o_23_cast_fp16)[name = string("op_4307_cast_fp16")];
            tensor<int32, [1]> var_4312_axes_0 = const()[name = string("op_4312_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4312_keep_dims_0 = const()[name = string("op_4312_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 4, 1]> var_4312_cast_fp16 = reduce_mean(axes = var_4312_axes_0, keep_dims = var_4312_keep_dims_0, x = var_4307_cast_fp16)[name = string("op_4312_cast_fp16")];
            fp16 var_4314_to_fp16 = const()[name = string("op_4314_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 64, 4, 1]> var_4315_cast_fp16 = add(x = var_4312_cast_fp16, y = var_4314_to_fp16)[name = string("op_4315_cast_fp16")];
            fp32 var_4316_epsilon_0 = const()[name = string("op_4316_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 4, 1]> var_4316_cast_fp16 = rsqrt(epsilon = var_4316_epsilon_0, x = var_4315_cast_fp16)[name = string("op_4316_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> var_4317_cast_fp16 = mul(x = o_23_cast_fp16, y = var_4316_cast_fp16)[name = string("op_4317_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> var_4319_cast_fp16 = silu(x = input_65_cast_fp16)[name = string("op_4319_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> o_25_cast_fp16 = mul(x = var_4317_cast_fp16, y = var_4319_cast_fp16)[name = string("o_25_cast_fp16")];
            tensor<int32, [3]> var_4324 = const()[name = string("op_4324"), val = tensor<int32, [3]>([1, 64, 512])];
            tensor<fp16, [1, 64, 512]> input_67_cast_fp16 = reshape(shape = var_4324, x = o_25_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [640, 512]> layers_4_out_weight_to_fp16 = const()[name = string("layers_4_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72063040)))];
            tensor<fp16, [1, 64, 640]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_out_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 64, 640]> x_49_cast_fp16 = add(x = x_45_cast_fp16, y = linear_17_cast_fp16)[name = string("x_49_cast_fp16")];
            fp16 var_4337_to_fp16 = const()[name = string("op_4337_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 64, 640]> x_51_cast_fp16 = mul(x = x_49_cast_fp16, y = var_4337_to_fp16)[name = string("x_51_cast_fp16")];
            fp16 var_4335_promoted_to_fp16 = const()[name = string("op_4335_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> var_4339_cast_fp16 = pow(x = x_51_cast_fp16, y = var_4335_promoted_to_fp16)[name = string("op_4339_cast_fp16")];
            tensor<int32, [1]> var_4341_axes_0 = const()[name = string("op_4341_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4341_keep_dims_0 = const()[name = string("op_4341_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_4341_cast_fp16 = reduce_mean(axes = var_4341_axes_0, keep_dims = var_4341_keep_dims_0, x = var_4339_cast_fp16)[name = string("op_4341_cast_fp16")];
            fp16 var_4342_to_fp16 = const()[name = string("op_4342_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 64, 1]> var_4343_cast_fp16 = add(x = var_4341_cast_fp16, y = var_4342_to_fp16)[name = string("op_4343_cast_fp16")];
            fp32 var_4344_epsilon_0 = const()[name = string("op_4344_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_4344_cast_fp16 = rsqrt(epsilon = var_4344_epsilon_0, x = var_4343_cast_fp16)[name = string("op_4344_cast_fp16")];
            tensor<fp16, [1, 64, 640]> var_4345_cast_fp16 = mul(x = x_51_cast_fp16, y = var_4344_cast_fp16)[name = string("op_4345_cast_fp16")];
            tensor<fp16, [640]> layers_4_n2_w1_to_fp16 = const()[name = string("layers_4_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72718464)))];
            tensor<fp16, [1, 64, 640]> input_69_cast_fp16 = mul(x = var_4345_cast_fp16, y = layers_4_n2_w1_to_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_4_gate_up_weight_to_fp16 = const()[name = string("layers_4_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72719808)))];
            tensor<fp16, [1, 64, 5120]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_4_gate_up_weight_to_fp16, x = input_69_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<int32, [3]> input_71_begin_0 = const()[name = string("input_71_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_71_end_0 = const()[name = string("input_71_end_0"), val = tensor<int32, [3]>([1, 64, 2560])];
            tensor<bool, [3]> input_71_end_mask_0 = const()[name = string("input_71_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 2560]> input_71_cast_fp16 = slice_by_index(begin = input_71_begin_0, end = input_71_end_0, end_mask = input_71_end_mask_0, x = linear_18_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [1, 64, 2560]> var_4355_cast_fp16 = silu(x = input_71_cast_fp16)[name = string("op_4355_cast_fp16")];
            tensor<int32, [3]> var_4360_begin_0 = const()[name = string("op_4360_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_4360_end_0 = const()[name = string("op_4360_end_0"), val = tensor<int32, [3]>([1, 64, 5120])];
            tensor<bool, [3]> var_4360_end_mask_0 = const()[name = string("op_4360_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 2560]> var_4360_cast_fp16 = slice_by_index(begin = var_4360_begin_0, end = var_4360_end_0, end_mask = var_4360_end_mask_0, x = linear_18_cast_fp16)[name = string("op_4360_cast_fp16")];
            tensor<fp16, [1, 64, 2560]> input_73_cast_fp16 = mul(x = var_4355_cast_fp16, y = var_4360_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_4_down_weight_to_fp16 = const()[name = string("layers_4_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79273472)))];
            tensor<fp16, [1, 64, 640]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_down_weight_to_fp16, x = input_73_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 64, 640]> x_53_cast_fp16 = add(x = x_49_cast_fp16, y = linear_19_cast_fp16)[name = string("x_53_cast_fp16")];
            fp16 var_4373_to_fp16 = const()[name = string("op_4373_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 64, 640]> x_55_cast_fp16 = mul(x = x_53_cast_fp16, y = var_4373_to_fp16)[name = string("x_55_cast_fp16")];
            fp16 var_4371_promoted_to_fp16 = const()[name = string("op_4371_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> var_4375_cast_fp16 = pow(x = x_55_cast_fp16, y = var_4371_promoted_to_fp16)[name = string("op_4375_cast_fp16")];
            tensor<int32, [1]> var_4377_axes_0 = const()[name = string("op_4377_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4377_keep_dims_0 = const()[name = string("op_4377_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_4377_cast_fp16 = reduce_mean(axes = var_4377_axes_0, keep_dims = var_4377_keep_dims_0, x = var_4375_cast_fp16)[name = string("op_4377_cast_fp16")];
            fp16 var_4378_to_fp16 = const()[name = string("op_4378_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 64, 1]> var_4379_cast_fp16 = add(x = var_4377_cast_fp16, y = var_4378_to_fp16)[name = string("op_4379_cast_fp16")];
            fp32 var_4380_epsilon_0 = const()[name = string("op_4380_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_4380_cast_fp16 = rsqrt(epsilon = var_4380_epsilon_0, x = var_4379_cast_fp16)[name = string("op_4380_cast_fp16")];
            tensor<fp16, [1, 64, 640]> var_4381_cast_fp16 = mul(x = x_55_cast_fp16, y = var_4380_cast_fp16)[name = string("op_4381_cast_fp16")];
            tensor<fp16, [640]> layers_5_n1_w1_to_fp16 = const()[name = string("layers_5_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82550336)))];
            tensor<fp16, [1, 64, 640]> input_75_cast_fp16 = mul(x = var_4381_cast_fp16, y = layers_5_n1_w1_to_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_5_qkvzba_weight_to_fp16 = const()[name = string("layers_5_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82551680)))];
            tensor<fp16, [1, 64, 1544]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_qkvzba_weight_to_fp16, x = input_75_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [3]> qkv_9_begin_0 = const()[name = string("qkv_9_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_9_end_0 = const()[name = string("qkv_9_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> qkv_9_end_mask_0 = const()[name = string("qkv_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 1024]> qkv_9_cast_fp16 = slice_by_index(begin = qkv_9_begin_0, end = qkv_9_end_0, end_mask = qkv_9_end_mask_0, x = linear_20_cast_fp16)[name = string("qkv_9_cast_fp16")];
            tensor<int32, [3]> z_9_begin_0 = const()[name = string("z_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_9_end_0 = const()[name = string("z_9_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
            tensor<bool, [3]> z_9_end_mask_0 = const()[name = string("z_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 512]> z_9_cast_fp16 = slice_by_index(begin = z_9_begin_0, end = z_9_end_0, end_mask = z_9_end_mask_0, x = linear_20_cast_fp16)[name = string("z_9_cast_fp16")];
            tensor<int32, [3]> b_9_begin_0 = const()[name = string("b_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_9_end_0 = const()[name = string("b_9_end_0"), val = tensor<int32, [3]>([1, 64, 1540])];
            tensor<bool, [3]> b_9_end_mask_0 = const()[name = string("b_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 4]> b_9_cast_fp16 = slice_by_index(begin = b_9_begin_0, end = b_9_end_0, end_mask = b_9_end_mask_0, x = linear_20_cast_fp16)[name = string("b_9_cast_fp16")];
            tensor<int32, [3]> a_9_begin_0 = const()[name = string("a_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_9_end_0 = const()[name = string("a_9_end_0"), val = tensor<int32, [3]>([1, 64, 1])];
            tensor<bool, [3]> a_9_end_mask_0 = const()[name = string("a_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 4]> a_9_cast_fp16 = slice_by_index(begin = a_9_begin_0, end = a_9_end_0, end_mask = a_9_end_mask_0, x = linear_20_cast_fp16)[name = string("a_9_cast_fp16")];
            int32 var_4420 = const()[name = string("op_4420"), val = int32(1)];
            bool xp_9_interleave_0 = const()[name = string("xp_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 67, 1024]> xp_9_cast_fp16 = concat(axis = var_4420, interleave = xp_9_interleave_0, values = (zpad_to_fp16, qkv_9_cast_fp16))[name = string("xp_9_cast_fp16")];
            tensor<int32, [3]> var_4431_begin_0 = const()[name = string("op_4431_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_4431_end_0 = const()[name = string("op_4431_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_4431_end_mask_0 = const()[name = string("op_4431_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 1024]> var_4431_cast_fp16 = slice_by_index(begin = var_4431_begin_0, end = var_4431_end_0, end_mask = var_4431_end_mask_0, x = xp_9_cast_fp16)[name = string("op_4431_cast_fp16")];
            tensor<fp16, [1024]> var_4439_to_fp16 = const()[name = string("op_4439_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84528064)))];
            tensor<fp16, [1, 64, 1024]> var_4440_cast_fp16 = mul(x = var_4431_cast_fp16, y = var_4439_to_fp16)[name = string("op_4440_cast_fp16")];
            tensor<int32, [3]> var_4450_begin_0 = const()[name = string("op_4450_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_4450_end_0 = const()[name = string("op_4450_end_0"), val = tensor<int32, [3]>([1, 66, 1024])];
            tensor<bool, [3]> var_4450_end_mask_0 = const()[name = string("op_4450_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_4450_cast_fp16 = slice_by_index(begin = var_4450_begin_0, end = var_4450_end_0, end_mask = var_4450_end_mask_0, x = xp_9_cast_fp16)[name = string("op_4450_cast_fp16")];
            tensor<fp16, [1024]> var_4458_to_fp16 = const()[name = string("op_4458_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84530176)))];
            tensor<fp16, [1, 64, 1024]> var_4459_cast_fp16 = mul(x = var_4450_cast_fp16, y = var_4458_to_fp16)[name = string("op_4459_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> var_4461_cast_fp16 = add(x = var_4440_cast_fp16, y = var_4459_cast_fp16)[name = string("op_4461_cast_fp16")];
            tensor<int32, [3]> var_4471_begin_0 = const()[name = string("op_4471_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_4471_end_0 = const()[name = string("op_4471_end_0"), val = tensor<int32, [3]>([1, 65, 1024])];
            tensor<bool, [3]> var_4471_end_mask_0 = const()[name = string("op_4471_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_4471_cast_fp16 = slice_by_index(begin = var_4471_begin_0, end = var_4471_end_0, end_mask = var_4471_end_mask_0, x = xp_9_cast_fp16)[name = string("op_4471_cast_fp16")];
            tensor<fp16, [1024]> var_4479_to_fp16 = const()[name = string("op_4479_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84532288)))];
            tensor<fp16, [1, 64, 1024]> var_4480_cast_fp16 = mul(x = var_4471_cast_fp16, y = var_4479_to_fp16)[name = string("op_4480_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> var_4482_cast_fp16 = add(x = var_4461_cast_fp16, y = var_4480_cast_fp16)[name = string("op_4482_cast_fp16")];
            tensor<int32, [3]> var_4492_begin_0 = const()[name = string("op_4492_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_4492_end_0 = const()[name = string("op_4492_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> var_4492_end_mask_0 = const()[name = string("op_4492_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_4492_cast_fp16 = slice_by_index(begin = var_4492_begin_0, end = var_4492_end_0, end_mask = var_4492_end_mask_0, x = xp_9_cast_fp16)[name = string("op_4492_cast_fp16")];
            tensor<fp16, [1024]> var_4500_to_fp16 = const()[name = string("op_4500_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84534400)))];
            tensor<fp16, [1, 64, 1024]> var_4501_cast_fp16 = mul(x = var_4492_cast_fp16, y = var_4500_to_fp16)[name = string("op_4501_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> input_77_cast_fp16 = add(x = var_4482_cast_fp16, y = var_4501_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> c_9_cast_fp16 = silu(x = input_77_cast_fp16)[name = string("c_9_cast_fp16")];
            tensor<int32, [3]> var_4509_begin_0 = const()[name = string("op_4509_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_4509_end_0 = const()[name = string("op_4509_end_0"), val = tensor<int32, [3]>([1, 64, 256])];
            tensor<bool, [3]> var_4509_end_mask_0 = const()[name = string("op_4509_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 256]> var_4509_cast_fp16 = slice_by_index(begin = var_4509_begin_0, end = var_4509_end_0, end_mask = var_4509_end_mask_0, x = c_9_cast_fp16)[name = string("op_4509_cast_fp16")];
            tensor<int32, [4]> var_4514 = const()[name = string("op_4514"), val = tensor<int32, [4]>([1, 64, 2, 128])];
            tensor<fp16, [1, 64, 2, 128]> q_67_cast_fp16 = reshape(shape = var_4514, x = var_4509_cast_fp16)[name = string("q_67_cast_fp16")];
            tensor<int32, [3]> var_4520_begin_0 = const()[name = string("op_4520_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_4520_end_0 = const()[name = string("op_4520_end_0"), val = tensor<int32, [3]>([1, 64, 512])];
            tensor<bool, [3]> var_4520_end_mask_0 = const()[name = string("op_4520_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 256]> var_4520_cast_fp16 = slice_by_index(begin = var_4520_begin_0, end = var_4520_end_0, end_mask = var_4520_end_mask_0, x = c_9_cast_fp16)[name = string("op_4520_cast_fp16")];
            tensor<int32, [4]> var_4525 = const()[name = string("op_4525"), val = tensor<int32, [4]>([1, 64, 2, 128])];
            tensor<fp16, [1, 64, 2, 128]> k_69_cast_fp16 = reshape(shape = var_4525, x = var_4520_cast_fp16)[name = string("k_69_cast_fp16")];
            tensor<int32, [3]> var_4531_begin_0 = const()[name = string("op_4531_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_4531_end_0 = const()[name = string("op_4531_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> var_4531_end_mask_0 = const()[name = string("op_4531_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 512]> var_4531_cast_fp16 = slice_by_index(begin = var_4531_begin_0, end = var_4531_end_0, end_mask = var_4531_end_mask_0, x = c_9_cast_fp16)[name = string("op_4531_cast_fp16")];
            tensor<int32, [4]> var_4536 = const()[name = string("op_4536"), val = tensor<int32, [4]>([1, 64, 4, 128])];
            tensor<fp16, [1, 64, 4, 128]> var_4537_cast_fp16 = reshape(shape = var_4536, x = var_4531_cast_fp16)[name = string("op_4537_cast_fp16")];
            tensor<int32, [4]> var_4542 = const()[name = string("op_4542"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4544_to_fp16 = const()[name = string("op_4544_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 64, 2, 128]> q_69_cast_fp16 = mul(x = q_67_cast_fp16, y = var_4544_to_fp16)[name = string("q_69_cast_fp16")];
            fp16 var_4546_to_fp16 = const()[name = string("op_4546_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 64, 2, 128]> k_71_cast_fp16 = mul(x = k_69_cast_fp16, y = var_4546_to_fp16)[name = string("k_71_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_4548_cast_fp16 = mul(x = q_69_cast_fp16, y = q_69_cast_fp16)[name = string("op_4548_cast_fp16")];
            tensor<int32, [1]> var_4553_axes_0 = const()[name = string("op_4553_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4553_keep_dims_0 = const()[name = string("op_4553_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 2, 1]> var_4553_cast_fp16 = reduce_mean(axes = var_4553_axes_0, keep_dims = var_4553_keep_dims_0, x = var_4548_cast_fp16)[name = string("op_4553_cast_fp16")];
            fp16 var_4554_promoted_to_fp16 = const()[name = string("op_4554_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 64, 2, 1]> var_4555_cast_fp16 = mul(x = var_4553_cast_fp16, y = var_4554_promoted_to_fp16)[name = string("op_4555_cast_fp16")];
            fp16 var_4557_to_fp16 = const()[name = string("op_4557_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 64, 2, 1]> var_4558_cast_fp16 = add(x = var_4555_cast_fp16, y = var_4557_to_fp16)[name = string("op_4558_cast_fp16")];
            fp32 var_4559_epsilon_0 = const()[name = string("op_4559_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 2, 1]> var_4559_cast_fp16 = rsqrt(epsilon = var_4559_epsilon_0, x = var_4558_cast_fp16)[name = string("op_4559_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_4560_cast_fp16 = mul(x = q_69_cast_fp16, y = var_4559_cast_fp16)[name = string("op_4560_cast_fp16")];
            fp16 var_4561_to_fp16 = const()[name = string("op_4561_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 64, 2, 128]> q_71_cast_fp16 = mul(x = var_4560_cast_fp16, y = var_4561_to_fp16)[name = string("q_71_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_4563_cast_fp16 = mul(x = k_71_cast_fp16, y = k_71_cast_fp16)[name = string("op_4563_cast_fp16")];
            tensor<int32, [1]> var_4568_axes_0 = const()[name = string("op_4568_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4568_keep_dims_0 = const()[name = string("op_4568_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 2, 1]> var_4568_cast_fp16 = reduce_mean(axes = var_4568_axes_0, keep_dims = var_4568_keep_dims_0, x = var_4563_cast_fp16)[name = string("op_4568_cast_fp16")];
            fp16 var_4569_promoted_to_fp16 = const()[name = string("op_4569_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 64, 2, 1]> var_4570_cast_fp16 = mul(x = var_4568_cast_fp16, y = var_4569_promoted_to_fp16)[name = string("op_4570_cast_fp16")];
            fp16 var_4572_to_fp16 = const()[name = string("op_4572_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 64, 2, 1]> var_4573_cast_fp16 = add(x = var_4570_cast_fp16, y = var_4572_to_fp16)[name = string("op_4573_cast_fp16")];
            fp32 var_4574_epsilon_0 = const()[name = string("op_4574_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 2, 1]> var_4574_cast_fp16 = rsqrt(epsilon = var_4574_epsilon_0, x = var_4573_cast_fp16)[name = string("op_4574_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> k_73_cast_fp16 = mul(x = k_71_cast_fp16, y = var_4574_cast_fp16)[name = string("k_73_cast_fp16")];
            tensor<int32, [4]> var_4580 = const()[name = string("op_4580"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_4593_axes_0 = const()[name = string("op_4593_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 128]> var_4581_cast_fp16 = transpose(perm = var_4580, x = q_71_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 2, 1, 64, 128]> var_4593_cast_fp16 = expand_dims(axes = var_4593_axes_0, x = var_4581_cast_fp16)[name = string("op_4593_cast_fp16")];
            tensor<int32, [5]> var_4601_reps_0 = const()[name = string("op_4601_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 64, 128]> var_4601_cast_fp16 = tile(reps = var_4601_reps_0, x = var_4593_cast_fp16)[name = string("op_4601_cast_fp16")];
            tensor<int32, [4]> var_4606 = const()[name = string("op_4606"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> q_73_cast_fp16 = reshape(shape = var_4606, x = var_4601_cast_fp16)[name = string("q_73_cast_fp16")];
            tensor<int32, [4]> var_4612 = const()[name = string("op_4612"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_4625_axes_0 = const()[name = string("op_4625_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 128]> var_4613_cast_fp16 = transpose(perm = var_4612, x = k_73_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 2, 1, 64, 128]> var_4625_cast_fp16 = expand_dims(axes = var_4625_axes_0, x = var_4613_cast_fp16)[name = string("op_4625_cast_fp16")];
            tensor<int32, [5]> var_4633_reps_0 = const()[name = string("op_4633_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 64, 128]> var_4633_cast_fp16 = tile(reps = var_4633_reps_0, x = var_4625_cast_fp16)[name = string("op_4633_cast_fp16")];
            tensor<int32, [4]> var_4638 = const()[name = string("op_4638"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> k_75_cast_fp16 = reshape(shape = var_4638, x = var_4633_cast_fp16)[name = string("k_75_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_4640_cast_fp16 = sigmoid(x = b_9_cast_fp16)[name = string("op_4640_cast_fp16")];
            tensor<int32, [3]> var_4644 = const()[name = string("op_4644"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_5_dt_bias_to_fp16 = const()[name = string("layers_5_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 64, 4]> input_79_cast_fp16 = add(x = a_9_cast_fp16, y = layers_5_dt_bias_to_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_4648_cast_fp16 = relu(x = input_79_cast_fp16)[name = string("op_4648_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_4649_cast_fp16 = abs(x = input_79_cast_fp16)[name = string("op_4649_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 64, 4]> var_4650_cast_fp16 = mul(x = var_4649_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_4650_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_4651_cast_fp16 = exp(x = var_4650_cast_fp16)[name = string("op_4651_cast_fp16")];
            fp16 var_4653_to_fp16 = const()[name = string("op_4653_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 64, 4]> var_4654_cast_fp16 = add(x = var_4651_cast_fp16, y = var_4653_to_fp16)[name = string("op_4654_cast_fp16")];
            fp32 var_4655_epsilon_0 = const()[name = string("op_4655_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 64, 4]> var_4655_cast_fp16 = log(epsilon = var_4655_epsilon_0, x = var_4654_cast_fp16)[name = string("op_4655_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_4657_cast_fp16 = add(x = var_4648_cast_fp16, y = var_4655_cast_fp16)[name = string("op_4657_cast_fp16")];
            tensor<fp16, [4]> layers_5_negA_to_fp16 = const()[name = string("layers_5_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.2f8p+3, -0x1.ae8p+2, -0x1.548p+2, -0x1.63p+3])];
            tensor<fp16, [1, 64, 4]> var_4658_cast_fp16 = mul(x = layers_5_negA_to_fp16, y = var_4657_cast_fp16)[name = string("op_4658_cast_fp16")];
            tensor<int32, [3]> var_4662 = const()[name = string("op_4662"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_4665_axes_0 = const()[name = string("op_4665_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 64]> beta_9_cast_fp16 = transpose(perm = var_4644, x = var_4640_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 4, 64, 1]> var_4665_cast_fp16 = expand_dims(axes = var_4665_axes_0, x = beta_9_cast_fp16)[name = string("op_4665_cast_fp16")];
            fp16 var_4666_to_fp16 = const()[name = string("op_4666_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 64, 1]> var_4667_cast_fp16 = mul(x = var_4665_cast_fp16, y = var_4666_to_fp16)[name = string("op_4667_cast_fp16")];
            tensor<fp16, [1, 4, 64, 128]> v_13_cast_fp16 = transpose(perm = var_4542, x = var_4537_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 4, 64, 128]> vb_41_cast_fp16 = mul(x = v_13_cast_fp16, y = var_4667_cast_fp16)[name = string("vb_41_cast_fp16")];
            tensor<int32, [4]> q_75_begin_0 = const()[name = string("q_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_75_end_0 = const()[name = string("q_75_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_75_end_mask_0 = const()[name = string("q_75_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_75_cast_fp16 = slice_by_index(begin = q_75_begin_0, end = q_75_end_0, end_mask = q_75_end_mask_0, x = q_73_cast_fp16)[name = string("q_75_cast_fp16")];
            tensor<int32, [4]> k_77_begin_0 = const()[name = string("k_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_77_end_0 = const()[name = string("k_77_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_77_end_mask_0 = const()[name = string("k_77_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, x = k_75_cast_fp16)[name = string("k_77_cast_fp16")];
            tensor<fp16, [1, 4, 64, 128]> var_4701_cast_fp16 = mul(x = k_75_cast_fp16, y = var_4665_cast_fp16)[name = string("op_4701_cast_fp16")];
            tensor<int32, [4]> kb_33_begin_0 = const()[name = string("kb_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_33_end_0 = const()[name = string("kb_33_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_33_end_mask_0 = const()[name = string("kb_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_33_cast_fp16 = slice_by_index(begin = kb_33_begin_0, end = kb_33_end_0, end_mask = kb_33_end_mask_0, x = var_4701_cast_fp16)[name = string("kb_33_cast_fp16")];
            tensor<int32, [4]> vb_43_begin_0 = const()[name = string("vb_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_43_end_0 = const()[name = string("vb_43_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_43_end_mask_0 = const()[name = string("vb_43_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_43_cast_fp16 = slice_by_index(begin = vb_43_begin_0, end = vb_43_end_0, end_mask = vb_43_end_mask_0, x = vb_41_cast_fp16)[name = string("vb_43_cast_fp16")];
            tensor<int32, [3]> gdec_43_begin_0 = const()[name = string("gdec_43_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_43_end_0 = const()[name = string("gdec_43_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_43_end_mask_0 = const()[name = string("gdec_43_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 64]> gdec_41_cast_fp16 = transpose(perm = var_4662, x = var_4658_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 4, 16]> gdec_43_cast_fp16 = slice_by_index(begin = gdec_43_begin_0, end = gdec_43_end_0, end_mask = gdec_43_end_mask_0, x = gdec_41_cast_fp16)[name = string("gdec_43_cast_fp16")];
            tensor<int32, [2]> var_4749 = const()[name = string("op_4749"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_4750_cast_fp16 = reshape(shape = var_4749, x = gdec_43_cast_fp16)[name = string("op_4750_cast_fp16")];
            tensor<fp16, [4, 256]> var_4751_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_4750_cast_fp16)[name = string("op_4751_cast_fp16")];
            tensor<int32, [4]> var_4756 = const()[name = string("op_4756"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_33_cast_fp16 = reshape(shape = var_4756, x = var_4751_cast_fp16)[name = string("P_33_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4758_cast_fp16 = exp(x = P_33_cast_fp16)[name = string("op_4758_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_33_cast_fp16 = mul(x = var_4758_cast_fp16, y = tril_incl_to_fp16)[name = string("D_33_cast_fp16")];
            bool kk_33_transpose_x_1 = const()[name = string("kk_33_transpose_x_1"), val = bool(false)];
            bool kk_33_transpose_y_1 = const()[name = string("kk_33_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_33_cast_fp16 = matmul(transpose_x = kk_33_transpose_x_1, transpose_y = kk_33_transpose_y_1, x = kb_33_cast_fp16, y = k_77_cast_fp16)[name = string("kk_33_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_4764_cast_fp16 = mul(x = kk_33_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_4764_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4765_cast_fp16 = mul(x = var_4764_cast_fp16, y = D_33_cast_fp16)[name = string("op_4765_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_33_cast_fp16 = mul(x = var_4765_cast_fp16, y = cum_tail_to_fp16)[name = string("N_33_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4767_cast_fp16 = mul(x = D_33_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_4767_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_33_cast_fp16 = mul(x = kk_33_cast_fp16, y = var_4767_cast_fp16)[name = string("Na_33_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_129_cast_fp16 = add(x = eye_to_fp16, y = N_33_cast_fp16)[name = string("Minv_129_cast_fp16")];
            bool Np_97_transpose_x_0 = const()[name = string("Np_97_transpose_x_0"), val = bool(false)];
            bool Np_97_transpose_y_0 = const()[name = string("Np_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_97_cast_fp16 = matmul(transpose_x = Np_97_transpose_x_0, transpose_y = Np_97_transpose_y_0, x = N_33_cast_fp16, y = Na_33_cast_fp16)[name = string("Np_97_cast_fp16")];
            bool Npa_65_transpose_x_0 = const()[name = string("Npa_65_transpose_x_0"), val = bool(false)];
            bool Npa_65_transpose_y_0 = const()[name = string("Npa_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_65_cast_fp16 = matmul(transpose_x = Npa_65_transpose_x_0, transpose_y = Npa_65_transpose_y_0, x = Na_33_cast_fp16, y = N_33_cast_fp16)[name = string("Npa_65_cast_fp16")];
            bool var_4773_transpose_x_0 = const()[name = string("op_4773_transpose_x_0"), val = bool(false)];
            bool var_4773_transpose_y_0 = const()[name = string("op_4773_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4773_cast_fp16 = matmul(transpose_x = var_4773_transpose_x_0, transpose_y = var_4773_transpose_y_0, x = Minv_129_cast_fp16, y = Np_97_cast_fp16)[name = string("op_4773_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_131_cast_fp16 = add(x = Minv_129_cast_fp16, y = var_4773_cast_fp16)[name = string("Minv_131_cast_fp16")];
            bool Np_99_transpose_x_0 = const()[name = string("Np_99_transpose_x_0"), val = bool(false)];
            bool Np_99_transpose_y_0 = const()[name = string("Np_99_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_99_cast_fp16 = matmul(transpose_x = Np_99_transpose_x_0, transpose_y = Np_99_transpose_y_0, x = Np_97_cast_fp16, y = Npa_65_cast_fp16)[name = string("Np_99_cast_fp16")];
            bool Npa_67_transpose_x_0 = const()[name = string("Npa_67_transpose_x_0"), val = bool(false)];
            bool Npa_67_transpose_y_0 = const()[name = string("Npa_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_67_cast_fp16 = matmul(transpose_x = Npa_67_transpose_x_0, transpose_y = Npa_67_transpose_y_0, x = Npa_65_cast_fp16, y = Np_97_cast_fp16)[name = string("Npa_67_cast_fp16")];
            bool var_4778_transpose_x_0 = const()[name = string("op_4778_transpose_x_0"), val = bool(false)];
            bool var_4778_transpose_y_0 = const()[name = string("op_4778_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4778_cast_fp16 = matmul(transpose_x = var_4778_transpose_x_0, transpose_y = var_4778_transpose_y_0, x = Minv_131_cast_fp16, y = Np_99_cast_fp16)[name = string("op_4778_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_133_cast_fp16 = add(x = Minv_131_cast_fp16, y = var_4778_cast_fp16)[name = string("Minv_133_cast_fp16")];
            bool Np_101_transpose_x_0 = const()[name = string("Np_101_transpose_x_0"), val = bool(false)];
            bool Np_101_transpose_y_0 = const()[name = string("Np_101_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_101_cast_fp16 = matmul(transpose_x = Np_101_transpose_x_0, transpose_y = Np_101_transpose_y_0, x = Np_99_cast_fp16, y = Npa_67_cast_fp16)[name = string("Np_101_cast_fp16")];
            bool var_4782_transpose_x_0 = const()[name = string("op_4782_transpose_x_0"), val = bool(false)];
            bool var_4782_transpose_y_0 = const()[name = string("op_4782_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4782_cast_fp16 = matmul(transpose_x = var_4782_transpose_x_0, transpose_y = var_4782_transpose_y_0, x = Minv_133_cast_fp16, y = Np_101_cast_fp16)[name = string("op_4782_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_135_cast_fp16 = add(x = Minv_133_cast_fp16, y = var_4782_cast_fp16)[name = string("Minv_135_cast_fp16")];
            bool u_33_transpose_x_0 = const()[name = string("u_33_transpose_x_0"), val = bool(false)];
            bool u_33_transpose_y_0 = const()[name = string("u_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_33_cast_fp16 = matmul(transpose_x = u_33_transpose_x_0, transpose_y = u_33_transpose_y_0, x = Minv_135_cast_fp16, y = vb_43_cast_fp16)[name = string("u_33_cast_fp16")];
            bool var_4789_transpose_x_1 = const()[name = string("op_4789_transpose_x_1"), val = bool(false)];
            bool var_4789_transpose_y_1 = const()[name = string("op_4789_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_4789_cast_fp16 = matmul(transpose_x = var_4789_transpose_x_1, transpose_y = var_4789_transpose_y_1, x = q_75_cast_fp16, y = k_77_cast_fp16)[name = string("op_4789_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_33_cast_fp16 = mul(x = var_4789_cast_fp16, y = D_33_cast_fp16)[name = string("A_33_cast_fp16")];
            bool var_4791_transpose_x_0 = const()[name = string("op_4791_transpose_x_0"), val = bool(false)];
            bool var_4791_transpose_y_0 = const()[name = string("op_4791_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4791_cast_fp16 = matmul(transpose_x = var_4791_transpose_x_0, transpose_y = var_4791_transpose_y_0, x = A_33_cast_fp16, y = u_33_cast_fp16)[name = string("op_4791_cast_fp16")];
            tensor<fp16, [16]> var_4792_bias_0_to_fp16 = const()[name = string("op_4792_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_4792_cast_fp16 = linear(bias = var_4792_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_43_cast_fp16)[name = string("op_4792_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_25_cast_fp16 = exp(x = var_4792_cast_fp16)[name = string("tail_25_cast_fp16")];
            tensor<int32, [1]> var_4795_axes_0 = const()[name = string("op_4795_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_4795_cast_fp16 = expand_dims(axes = var_4795_axes_0, x = tail_25_cast_fp16)[name = string("op_4795_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4796_cast_fp16 = mul(x = k_77_cast_fp16, y = var_4795_cast_fp16)[name = string("op_4796_cast_fp16")];
            bool S_25_transpose_x_1 = const()[name = string("S_25_transpose_x_1"), val = bool(true)];
            bool S_25_transpose_y_1 = const()[name = string("S_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_25_cast_fp16 = matmul(transpose_x = S_25_transpose_x_1, transpose_y = S_25_transpose_y_1, x = var_4796_cast_fp16, y = u_33_cast_fp16)[name = string("S_25_cast_fp16")];
            tensor<int32, [4]> q_77_begin_0 = const()[name = string("q_77_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_77_end_0 = const()[name = string("q_77_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> q_77_end_mask_0 = const()[name = string("q_77_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_77_cast_fp16 = slice_by_index(begin = q_77_begin_0, end = q_77_end_0, end_mask = q_77_end_mask_0, x = q_73_cast_fp16)[name = string("q_77_cast_fp16")];
            tensor<int32, [4]> k_79_begin_0 = const()[name = string("k_79_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_79_end_0 = const()[name = string("k_79_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> k_79_end_mask_0 = const()[name = string("k_79_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_79_cast_fp16 = slice_by_index(begin = k_79_begin_0, end = k_79_end_0, end_mask = k_79_end_mask_0, x = k_75_cast_fp16)[name = string("k_79_cast_fp16")];
            tensor<int32, [4]> kb_35_begin_0 = const()[name = string("kb_35_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_35_end_0 = const()[name = string("kb_35_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> kb_35_end_mask_0 = const()[name = string("kb_35_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_35_cast_fp16 = slice_by_index(begin = kb_35_begin_0, end = kb_35_end_0, end_mask = kb_35_end_mask_0, x = var_4701_cast_fp16)[name = string("kb_35_cast_fp16")];
            tensor<int32, [4]> vb_45_begin_0 = const()[name = string("vb_45_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_45_end_0 = const()[name = string("vb_45_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> vb_45_end_mask_0 = const()[name = string("vb_45_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_45_cast_fp16 = slice_by_index(begin = vb_45_begin_0, end = vb_45_end_0, end_mask = vb_45_end_mask_0, x = vb_41_cast_fp16)[name = string("vb_45_cast_fp16")];
            tensor<int32, [3]> gdec_45_begin_0 = const()[name = string("gdec_45_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_45_end_0 = const()[name = string("gdec_45_end_0"), val = tensor<int32, [3]>([1, 4, 32])];
            tensor<bool, [3]> gdec_45_end_mask_0 = const()[name = string("gdec_45_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_45_cast_fp16 = slice_by_index(begin = gdec_45_begin_0, end = gdec_45_end_0, end_mask = gdec_45_end_mask_0, x = gdec_41_cast_fp16)[name = string("gdec_45_cast_fp16")];
            tensor<int32, [2]> var_4881 = const()[name = string("op_4881"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_4882_cast_fp16 = reshape(shape = var_4881, x = gdec_45_cast_fp16)[name = string("op_4882_cast_fp16")];
            tensor<fp16, [4, 256]> var_4883_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_4882_cast_fp16)[name = string("op_4883_cast_fp16")];
            tensor<int32, [4]> var_4888 = const()[name = string("op_4888"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_35_cast_fp16 = reshape(shape = var_4888, x = var_4883_cast_fp16)[name = string("P_35_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4890_cast_fp16 = exp(x = P_35_cast_fp16)[name = string("op_4890_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_35_cast_fp16 = mul(x = var_4890_cast_fp16, y = tril_incl_to_fp16)[name = string("D_35_cast_fp16")];
            bool kk_35_transpose_x_1 = const()[name = string("kk_35_transpose_x_1"), val = bool(false)];
            bool kk_35_transpose_y_1 = const()[name = string("kk_35_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_35_cast_fp16 = matmul(transpose_x = kk_35_transpose_x_1, transpose_y = kk_35_transpose_y_1, x = kb_35_cast_fp16, y = k_79_cast_fp16)[name = string("kk_35_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_4896_cast_fp16 = mul(x = kk_35_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_4896_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4897_cast_fp16 = mul(x = var_4896_cast_fp16, y = D_35_cast_fp16)[name = string("op_4897_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_35_cast_fp16 = mul(x = var_4897_cast_fp16, y = cum_tail_to_fp16)[name = string("N_35_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4899_cast_fp16 = mul(x = D_35_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_4899_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_35_cast_fp16 = mul(x = kk_35_cast_fp16, y = var_4899_cast_fp16)[name = string("Na_35_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_137_cast_fp16 = add(x = eye_to_fp16, y = N_35_cast_fp16)[name = string("Minv_137_cast_fp16")];
            bool Np_103_transpose_x_0 = const()[name = string("Np_103_transpose_x_0"), val = bool(false)];
            bool Np_103_transpose_y_0 = const()[name = string("Np_103_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_103_cast_fp16 = matmul(transpose_x = Np_103_transpose_x_0, transpose_y = Np_103_transpose_y_0, x = N_35_cast_fp16, y = Na_35_cast_fp16)[name = string("Np_103_cast_fp16")];
            bool Npa_69_transpose_x_0 = const()[name = string("Npa_69_transpose_x_0"), val = bool(false)];
            bool Npa_69_transpose_y_0 = const()[name = string("Npa_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_69_cast_fp16 = matmul(transpose_x = Npa_69_transpose_x_0, transpose_y = Npa_69_transpose_y_0, x = Na_35_cast_fp16, y = N_35_cast_fp16)[name = string("Npa_69_cast_fp16")];
            bool var_4905_transpose_x_0 = const()[name = string("op_4905_transpose_x_0"), val = bool(false)];
            bool var_4905_transpose_y_0 = const()[name = string("op_4905_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4905_cast_fp16 = matmul(transpose_x = var_4905_transpose_x_0, transpose_y = var_4905_transpose_y_0, x = Minv_137_cast_fp16, y = Np_103_cast_fp16)[name = string("op_4905_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_139_cast_fp16 = add(x = Minv_137_cast_fp16, y = var_4905_cast_fp16)[name = string("Minv_139_cast_fp16")];
            bool Np_105_transpose_x_0 = const()[name = string("Np_105_transpose_x_0"), val = bool(false)];
            bool Np_105_transpose_y_0 = const()[name = string("Np_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_105_cast_fp16 = matmul(transpose_x = Np_105_transpose_x_0, transpose_y = Np_105_transpose_y_0, x = Np_103_cast_fp16, y = Npa_69_cast_fp16)[name = string("Np_105_cast_fp16")];
            bool Npa_71_transpose_x_0 = const()[name = string("Npa_71_transpose_x_0"), val = bool(false)];
            bool Npa_71_transpose_y_0 = const()[name = string("Npa_71_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_71_cast_fp16 = matmul(transpose_x = Npa_71_transpose_x_0, transpose_y = Npa_71_transpose_y_0, x = Npa_69_cast_fp16, y = Np_103_cast_fp16)[name = string("Npa_71_cast_fp16")];
            bool var_4910_transpose_x_0 = const()[name = string("op_4910_transpose_x_0"), val = bool(false)];
            bool var_4910_transpose_y_0 = const()[name = string("op_4910_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4910_cast_fp16 = matmul(transpose_x = var_4910_transpose_x_0, transpose_y = var_4910_transpose_y_0, x = Minv_139_cast_fp16, y = Np_105_cast_fp16)[name = string("op_4910_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_141_cast_fp16 = add(x = Minv_139_cast_fp16, y = var_4910_cast_fp16)[name = string("Minv_141_cast_fp16")];
            bool Np_107_transpose_x_0 = const()[name = string("Np_107_transpose_x_0"), val = bool(false)];
            bool Np_107_transpose_y_0 = const()[name = string("Np_107_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_107_cast_fp16 = matmul(transpose_x = Np_107_transpose_x_0, transpose_y = Np_107_transpose_y_0, x = Np_105_cast_fp16, y = Npa_71_cast_fp16)[name = string("Np_107_cast_fp16")];
            bool var_4914_transpose_x_0 = const()[name = string("op_4914_transpose_x_0"), val = bool(false)];
            bool var_4914_transpose_y_0 = const()[name = string("op_4914_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4914_cast_fp16 = matmul(transpose_x = var_4914_transpose_x_0, transpose_y = var_4914_transpose_y_0, x = Minv_141_cast_fp16, y = Np_107_cast_fp16)[name = string("op_4914_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_143_cast_fp16 = add(x = Minv_141_cast_fp16, y = var_4914_cast_fp16)[name = string("Minv_143_cast_fp16")];
            bool u_35_transpose_x_0 = const()[name = string("u_35_transpose_x_0"), val = bool(false)];
            bool u_35_transpose_y_0 = const()[name = string("u_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_35_cast_fp16 = matmul(transpose_x = u_35_transpose_x_0, transpose_y = u_35_transpose_y_0, x = Minv_143_cast_fp16, y = vb_45_cast_fp16)[name = string("u_35_cast_fp16")];
            bool var_4921_transpose_x_1 = const()[name = string("op_4921_transpose_x_1"), val = bool(false)];
            bool var_4921_transpose_y_1 = const()[name = string("op_4921_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_4921_cast_fp16 = matmul(transpose_x = var_4921_transpose_x_1, transpose_y = var_4921_transpose_y_1, x = q_77_cast_fp16, y = k_79_cast_fp16)[name = string("op_4921_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_35_cast_fp16 = mul(x = var_4921_cast_fp16, y = D_35_cast_fp16)[name = string("A_35_cast_fp16")];
            tensor<fp16, [16]> var_4923_bias_0_to_fp16 = const()[name = string("op_4923_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_4923_cast_fp16 = linear(bias = var_4923_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_45_cast_fp16)[name = string("op_4923_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_25_cast_fp16 = exp(x = var_4923_cast_fp16)[name = string("cum_25_cast_fp16")];
            tensor<int32, [1]> var_4926_axes_0 = const()[name = string("op_4926_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_4926_cast_fp16 = expand_dims(axes = var_4926_axes_0, x = cum_25_cast_fp16)[name = string("op_4926_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4927_cast_fp16 = mul(x = kb_35_cast_fp16, y = var_4926_cast_fp16)[name = string("op_4927_cast_fp16")];
            bool w_35_transpose_x_0 = const()[name = string("w_35_transpose_x_0"), val = bool(false)];
            bool w_35_transpose_y_0 = const()[name = string("w_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_35_cast_fp16 = matmul(transpose_x = w_35_transpose_x_0, transpose_y = w_35_transpose_y_0, x = Minv_143_cast_fp16, y = var_4927_cast_fp16)[name = string("w_35_cast_fp16")];
            bool var_4929_transpose_x_0 = const()[name = string("op_4929_transpose_x_0"), val = bool(false)];
            bool var_4929_transpose_y_0 = const()[name = string("op_4929_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4929_cast_fp16 = matmul(transpose_x = var_4929_transpose_x_0, transpose_y = var_4929_transpose_y_0, x = w_35_cast_fp16, y = S_25_cast_fp16)[name = string("op_4929_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_25_cast_fp16 = sub(x = u_35_cast_fp16, y = var_4929_cast_fp16)[name = string("vnew_25_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4934_cast_fp16 = mul(x = q_77_cast_fp16, y = var_4926_cast_fp16)[name = string("op_4934_cast_fp16")];
            bool var_4935_transpose_x_0 = const()[name = string("op_4935_transpose_x_0"), val = bool(false)];
            bool var_4935_transpose_y_0 = const()[name = string("op_4935_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4935_cast_fp16 = matmul(transpose_x = var_4935_transpose_x_0, transpose_y = var_4935_transpose_y_0, x = var_4934_cast_fp16, y = S_25_cast_fp16)[name = string("op_4935_cast_fp16")];
            bool var_4936_transpose_x_0 = const()[name = string("op_4936_transpose_x_0"), val = bool(false)];
            bool var_4936_transpose_y_0 = const()[name = string("op_4936_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4936_cast_fp16 = matmul(transpose_x = var_4936_transpose_x_0, transpose_y = var_4936_transpose_y_0, x = A_35_cast_fp16, y = vnew_25_cast_fp16)[name = string("op_4936_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4938_cast_fp16 = add(x = var_4935_cast_fp16, y = var_4936_cast_fp16)[name = string("op_4938_cast_fp16")];
            tensor<fp16, [16]> var_4939_bias_0_to_fp16 = const()[name = string("op_4939_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_4939_cast_fp16 = linear(bias = var_4939_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_45_cast_fp16)[name = string("op_4939_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_27_cast_fp16 = exp(x = var_4939_cast_fp16)[name = string("tail_27_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_53_to_fp16 = const()[name = string("transpose_53_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_4941_bias_0_to_fp16 = const()[name = string("op_4941_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_4941_cast_fp16 = linear(bias = var_4941_bias_0_to_fp16, weight = transpose_53_to_fp16, x = gdec_45_cast_fp16)[name = string("op_4941_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_4942_cast_fp16 = exp(x = var_4941_cast_fp16)[name = string("op_4942_cast_fp16")];
            tensor<int32, [1]> last_19_axes_0 = const()[name = string("last_19_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_19_cast_fp16 = expand_dims(axes = last_19_axes_0, x = var_4942_cast_fp16)[name = string("last_19_cast_fp16")];
            tensor<int32, [1]> var_4946_axes_0 = const()[name = string("op_4946_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_4946_cast_fp16 = expand_dims(axes = var_4946_axes_0, x = tail_27_cast_fp16)[name = string("op_4946_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4947_cast_fp16 = mul(x = k_79_cast_fp16, y = var_4946_cast_fp16)[name = string("op_4947_cast_fp16")];
            bool upd_17_transpose_x_1 = const()[name = string("upd_17_transpose_x_1"), val = bool(true)];
            bool upd_17_transpose_y_1 = const()[name = string("upd_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_17_cast_fp16 = matmul(transpose_x = upd_17_transpose_x_1, transpose_y = upd_17_transpose_y_1, x = var_4947_cast_fp16, y = vnew_25_cast_fp16)[name = string("upd_17_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_4952_cast_fp16 = mul(x = S_25_cast_fp16, y = last_19_cast_fp16)[name = string("op_4952_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_27_cast_fp16 = add(x = var_4952_cast_fp16, y = upd_17_cast_fp16)[name = string("S_27_cast_fp16")];
            tensor<int32, [4]> q_79_begin_0 = const()[name = string("q_79_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_79_end_0 = const()[name = string("q_79_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> q_79_end_mask_0 = const()[name = string("q_79_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, x = q_73_cast_fp16)[name = string("q_79_cast_fp16")];
            tensor<int32, [4]> k_81_begin_0 = const()[name = string("k_81_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_81_end_0 = const()[name = string("k_81_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> k_81_end_mask_0 = const()[name = string("k_81_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, x = k_75_cast_fp16)[name = string("k_81_cast_fp16")];
            tensor<int32, [4]> kb_37_begin_0 = const()[name = string("kb_37_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_37_end_0 = const()[name = string("kb_37_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> kb_37_end_mask_0 = const()[name = string("kb_37_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_37_cast_fp16 = slice_by_index(begin = kb_37_begin_0, end = kb_37_end_0, end_mask = kb_37_end_mask_0, x = var_4701_cast_fp16)[name = string("kb_37_cast_fp16")];
            tensor<int32, [4]> vb_47_begin_0 = const()[name = string("vb_47_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_47_end_0 = const()[name = string("vb_47_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> vb_47_end_mask_0 = const()[name = string("vb_47_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_47_cast_fp16 = slice_by_index(begin = vb_47_begin_0, end = vb_47_end_0, end_mask = vb_47_end_mask_0, x = vb_41_cast_fp16)[name = string("vb_47_cast_fp16")];
            tensor<int32, [3]> gdec_47_begin_0 = const()[name = string("gdec_47_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_47_end_0 = const()[name = string("gdec_47_end_0"), val = tensor<int32, [3]>([1, 4, 48])];
            tensor<bool, [3]> gdec_47_end_mask_0 = const()[name = string("gdec_47_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_47_cast_fp16 = slice_by_index(begin = gdec_47_begin_0, end = gdec_47_end_0, end_mask = gdec_47_end_mask_0, x = gdec_41_cast_fp16)[name = string("gdec_47_cast_fp16")];
            tensor<int32, [2]> var_5035 = const()[name = string("op_5035"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_5036_cast_fp16 = reshape(shape = var_5035, x = gdec_47_cast_fp16)[name = string("op_5036_cast_fp16")];
            tensor<fp16, [4, 256]> var_5037_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5036_cast_fp16)[name = string("op_5037_cast_fp16")];
            tensor<int32, [4]> var_5042 = const()[name = string("op_5042"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_37_cast_fp16 = reshape(shape = var_5042, x = var_5037_cast_fp16)[name = string("P_37_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5044_cast_fp16 = exp(x = P_37_cast_fp16)[name = string("op_5044_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_37_cast_fp16 = mul(x = var_5044_cast_fp16, y = tril_incl_to_fp16)[name = string("D_37_cast_fp16")];
            bool kk_37_transpose_x_1 = const()[name = string("kk_37_transpose_x_1"), val = bool(false)];
            bool kk_37_transpose_y_1 = const()[name = string("kk_37_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_37_cast_fp16 = matmul(transpose_x = kk_37_transpose_x_1, transpose_y = kk_37_transpose_y_1, x = kb_37_cast_fp16, y = k_81_cast_fp16)[name = string("kk_37_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_5050_cast_fp16 = mul(x = kk_37_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_5050_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5051_cast_fp16 = mul(x = var_5050_cast_fp16, y = D_37_cast_fp16)[name = string("op_5051_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_37_cast_fp16 = mul(x = var_5051_cast_fp16, y = cum_tail_to_fp16)[name = string("N_37_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5053_cast_fp16 = mul(x = D_37_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5053_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_37_cast_fp16 = mul(x = kk_37_cast_fp16, y = var_5053_cast_fp16)[name = string("Na_37_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_145_cast_fp16 = add(x = eye_to_fp16, y = N_37_cast_fp16)[name = string("Minv_145_cast_fp16")];
            bool Np_109_transpose_x_0 = const()[name = string("Np_109_transpose_x_0"), val = bool(false)];
            bool Np_109_transpose_y_0 = const()[name = string("Np_109_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_109_cast_fp16 = matmul(transpose_x = Np_109_transpose_x_0, transpose_y = Np_109_transpose_y_0, x = N_37_cast_fp16, y = Na_37_cast_fp16)[name = string("Np_109_cast_fp16")];
            bool Npa_73_transpose_x_0 = const()[name = string("Npa_73_transpose_x_0"), val = bool(false)];
            bool Npa_73_transpose_y_0 = const()[name = string("Npa_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_73_cast_fp16 = matmul(transpose_x = Npa_73_transpose_x_0, transpose_y = Npa_73_transpose_y_0, x = Na_37_cast_fp16, y = N_37_cast_fp16)[name = string("Npa_73_cast_fp16")];
            bool var_5059_transpose_x_0 = const()[name = string("op_5059_transpose_x_0"), val = bool(false)];
            bool var_5059_transpose_y_0 = const()[name = string("op_5059_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5059_cast_fp16 = matmul(transpose_x = var_5059_transpose_x_0, transpose_y = var_5059_transpose_y_0, x = Minv_145_cast_fp16, y = Np_109_cast_fp16)[name = string("op_5059_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_147_cast_fp16 = add(x = Minv_145_cast_fp16, y = var_5059_cast_fp16)[name = string("Minv_147_cast_fp16")];
            bool Np_111_transpose_x_0 = const()[name = string("Np_111_transpose_x_0"), val = bool(false)];
            bool Np_111_transpose_y_0 = const()[name = string("Np_111_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_111_cast_fp16 = matmul(transpose_x = Np_111_transpose_x_0, transpose_y = Np_111_transpose_y_0, x = Np_109_cast_fp16, y = Npa_73_cast_fp16)[name = string("Np_111_cast_fp16")];
            bool Npa_75_transpose_x_0 = const()[name = string("Npa_75_transpose_x_0"), val = bool(false)];
            bool Npa_75_transpose_y_0 = const()[name = string("Npa_75_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_75_cast_fp16 = matmul(transpose_x = Npa_75_transpose_x_0, transpose_y = Npa_75_transpose_y_0, x = Npa_73_cast_fp16, y = Np_109_cast_fp16)[name = string("Npa_75_cast_fp16")];
            bool var_5064_transpose_x_0 = const()[name = string("op_5064_transpose_x_0"), val = bool(false)];
            bool var_5064_transpose_y_0 = const()[name = string("op_5064_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5064_cast_fp16 = matmul(transpose_x = var_5064_transpose_x_0, transpose_y = var_5064_transpose_y_0, x = Minv_147_cast_fp16, y = Np_111_cast_fp16)[name = string("op_5064_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_149_cast_fp16 = add(x = Minv_147_cast_fp16, y = var_5064_cast_fp16)[name = string("Minv_149_cast_fp16")];
            bool Np_113_transpose_x_0 = const()[name = string("Np_113_transpose_x_0"), val = bool(false)];
            bool Np_113_transpose_y_0 = const()[name = string("Np_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_113_cast_fp16 = matmul(transpose_x = Np_113_transpose_x_0, transpose_y = Np_113_transpose_y_0, x = Np_111_cast_fp16, y = Npa_75_cast_fp16)[name = string("Np_113_cast_fp16")];
            bool var_5068_transpose_x_0 = const()[name = string("op_5068_transpose_x_0"), val = bool(false)];
            bool var_5068_transpose_y_0 = const()[name = string("op_5068_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5068_cast_fp16 = matmul(transpose_x = var_5068_transpose_x_0, transpose_y = var_5068_transpose_y_0, x = Minv_149_cast_fp16, y = Np_113_cast_fp16)[name = string("op_5068_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_151_cast_fp16 = add(x = Minv_149_cast_fp16, y = var_5068_cast_fp16)[name = string("Minv_151_cast_fp16")];
            bool u_37_transpose_x_0 = const()[name = string("u_37_transpose_x_0"), val = bool(false)];
            bool u_37_transpose_y_0 = const()[name = string("u_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_37_cast_fp16 = matmul(transpose_x = u_37_transpose_x_0, transpose_y = u_37_transpose_y_0, x = Minv_151_cast_fp16, y = vb_47_cast_fp16)[name = string("u_37_cast_fp16")];
            bool var_5075_transpose_x_1 = const()[name = string("op_5075_transpose_x_1"), val = bool(false)];
            bool var_5075_transpose_y_1 = const()[name = string("op_5075_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_5075_cast_fp16 = matmul(transpose_x = var_5075_transpose_x_1, transpose_y = var_5075_transpose_y_1, x = q_79_cast_fp16, y = k_81_cast_fp16)[name = string("op_5075_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_37_cast_fp16 = mul(x = var_5075_cast_fp16, y = D_37_cast_fp16)[name = string("A_37_cast_fp16")];
            tensor<fp16, [16]> var_5077_bias_0_to_fp16 = const()[name = string("op_5077_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_5077_cast_fp16 = linear(bias = var_5077_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_47_cast_fp16)[name = string("op_5077_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_27_cast_fp16 = exp(x = var_5077_cast_fp16)[name = string("cum_27_cast_fp16")];
            tensor<int32, [1]> var_5080_axes_0 = const()[name = string("op_5080_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_5080_cast_fp16 = expand_dims(axes = var_5080_axes_0, x = cum_27_cast_fp16)[name = string("op_5080_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5081_cast_fp16 = mul(x = kb_37_cast_fp16, y = var_5080_cast_fp16)[name = string("op_5081_cast_fp16")];
            bool w_37_transpose_x_0 = const()[name = string("w_37_transpose_x_0"), val = bool(false)];
            bool w_37_transpose_y_0 = const()[name = string("w_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_37_cast_fp16 = matmul(transpose_x = w_37_transpose_x_0, transpose_y = w_37_transpose_y_0, x = Minv_151_cast_fp16, y = var_5081_cast_fp16)[name = string("w_37_cast_fp16")];
            bool var_5083_transpose_x_0 = const()[name = string("op_5083_transpose_x_0"), val = bool(false)];
            bool var_5083_transpose_y_0 = const()[name = string("op_5083_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5083_cast_fp16 = matmul(transpose_x = var_5083_transpose_x_0, transpose_y = var_5083_transpose_y_0, x = w_37_cast_fp16, y = S_27_cast_fp16)[name = string("op_5083_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_27_cast_fp16 = sub(x = u_37_cast_fp16, y = var_5083_cast_fp16)[name = string("vnew_27_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5088_cast_fp16 = mul(x = q_79_cast_fp16, y = var_5080_cast_fp16)[name = string("op_5088_cast_fp16")];
            bool var_5089_transpose_x_0 = const()[name = string("op_5089_transpose_x_0"), val = bool(false)];
            bool var_5089_transpose_y_0 = const()[name = string("op_5089_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5089_cast_fp16 = matmul(transpose_x = var_5089_transpose_x_0, transpose_y = var_5089_transpose_y_0, x = var_5088_cast_fp16, y = S_27_cast_fp16)[name = string("op_5089_cast_fp16")];
            bool var_5090_transpose_x_0 = const()[name = string("op_5090_transpose_x_0"), val = bool(false)];
            bool var_5090_transpose_y_0 = const()[name = string("op_5090_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5090_cast_fp16 = matmul(transpose_x = var_5090_transpose_x_0, transpose_y = var_5090_transpose_y_0, x = A_37_cast_fp16, y = vnew_27_cast_fp16)[name = string("op_5090_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5092_cast_fp16 = add(x = var_5089_cast_fp16, y = var_5090_cast_fp16)[name = string("op_5092_cast_fp16")];
            tensor<fp16, [16]> var_5093_bias_0_to_fp16 = const()[name = string("op_5093_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_5093_cast_fp16 = linear(bias = var_5093_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_47_cast_fp16)[name = string("op_5093_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_29_cast_fp16 = exp(x = var_5093_cast_fp16)[name = string("tail_29_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_57_to_fp16 = const()[name = string("transpose_57_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_5095_bias_0_to_fp16 = const()[name = string("op_5095_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_5095_cast_fp16 = linear(bias = var_5095_bias_0_to_fp16, weight = transpose_57_to_fp16, x = gdec_47_cast_fp16)[name = string("op_5095_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_5096_cast_fp16 = exp(x = var_5095_cast_fp16)[name = string("op_5096_cast_fp16")];
            tensor<int32, [1]> last_21_axes_0 = const()[name = string("last_21_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_21_cast_fp16 = expand_dims(axes = last_21_axes_0, x = var_5096_cast_fp16)[name = string("last_21_cast_fp16")];
            tensor<int32, [1]> var_5100_axes_0 = const()[name = string("op_5100_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_5100_cast_fp16 = expand_dims(axes = var_5100_axes_0, x = tail_29_cast_fp16)[name = string("op_5100_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5101_cast_fp16 = mul(x = k_81_cast_fp16, y = var_5100_cast_fp16)[name = string("op_5101_cast_fp16")];
            bool upd_19_transpose_x_1 = const()[name = string("upd_19_transpose_x_1"), val = bool(true)];
            bool upd_19_transpose_y_1 = const()[name = string("upd_19_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_19_cast_fp16 = matmul(transpose_x = upd_19_transpose_x_1, transpose_y = upd_19_transpose_y_1, x = var_5101_cast_fp16, y = vnew_27_cast_fp16)[name = string("upd_19_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_5106_cast_fp16 = mul(x = S_27_cast_fp16, y = last_21_cast_fp16)[name = string("op_5106_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_29_cast_fp16 = add(x = var_5106_cast_fp16, y = upd_19_cast_fp16)[name = string("S_29_cast_fp16")];
            tensor<int32, [4]> q_81_begin_0 = const()[name = string("q_81_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_81_end_0 = const()[name = string("q_81_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_81_end_mask_0 = const()[name = string("q_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_81_cast_fp16 = slice_by_index(begin = q_81_begin_0, end = q_81_end_0, end_mask = q_81_end_mask_0, x = q_73_cast_fp16)[name = string("q_81_cast_fp16")];
            tensor<int32, [4]> k_83_begin_0 = const()[name = string("k_83_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_83_end_0 = const()[name = string("k_83_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_83_end_mask_0 = const()[name = string("k_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_83_cast_fp16 = slice_by_index(begin = k_83_begin_0, end = k_83_end_0, end_mask = k_83_end_mask_0, x = k_75_cast_fp16)[name = string("k_83_cast_fp16")];
            tensor<int32, [4]> kb_39_begin_0 = const()[name = string("kb_39_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_39_end_0 = const()[name = string("kb_39_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_39_end_mask_0 = const()[name = string("kb_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_39_cast_fp16 = slice_by_index(begin = kb_39_begin_0, end = kb_39_end_0, end_mask = kb_39_end_mask_0, x = var_4701_cast_fp16)[name = string("kb_39_cast_fp16")];
            tensor<int32, [4]> vb_49_begin_0 = const()[name = string("vb_49_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_49_end_0 = const()[name = string("vb_49_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_49_end_mask_0 = const()[name = string("vb_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_49_cast_fp16 = slice_by_index(begin = vb_49_begin_0, end = vb_49_end_0, end_mask = vb_49_end_mask_0, x = vb_41_cast_fp16)[name = string("vb_49_cast_fp16")];
            tensor<int32, [3]> gdec_49_begin_0 = const()[name = string("gdec_49_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_49_end_0 = const()[name = string("gdec_49_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_49_end_mask_0 = const()[name = string("gdec_49_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_49_cast_fp16 = slice_by_index(begin = gdec_49_begin_0, end = gdec_49_end_0, end_mask = gdec_49_end_mask_0, x = gdec_41_cast_fp16)[name = string("gdec_49_cast_fp16")];
            tensor<int32, [2]> var_5189 = const()[name = string("op_5189"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_5190_cast_fp16 = reshape(shape = var_5189, x = gdec_49_cast_fp16)[name = string("op_5190_cast_fp16")];
            tensor<fp16, [4, 256]> var_5191_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5190_cast_fp16)[name = string("op_5191_cast_fp16")];
            tensor<int32, [4]> var_5196 = const()[name = string("op_5196"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_39_cast_fp16 = reshape(shape = var_5196, x = var_5191_cast_fp16)[name = string("P_39_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5198_cast_fp16 = exp(x = P_39_cast_fp16)[name = string("op_5198_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_39_cast_fp16 = mul(x = var_5198_cast_fp16, y = tril_incl_to_fp16)[name = string("D_39_cast_fp16")];
            bool kk_39_transpose_x_1 = const()[name = string("kk_39_transpose_x_1"), val = bool(false)];
            bool kk_39_transpose_y_1 = const()[name = string("kk_39_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_39_cast_fp16 = matmul(transpose_x = kk_39_transpose_x_1, transpose_y = kk_39_transpose_y_1, x = kb_39_cast_fp16, y = k_83_cast_fp16)[name = string("kk_39_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_5204_cast_fp16 = mul(x = kk_39_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_5204_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5205_cast_fp16 = mul(x = var_5204_cast_fp16, y = D_39_cast_fp16)[name = string("op_5205_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_39_cast_fp16 = mul(x = var_5205_cast_fp16, y = cum_tail_to_fp16)[name = string("N_39_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5207_cast_fp16 = mul(x = D_39_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5207_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_39_cast_fp16 = mul(x = kk_39_cast_fp16, y = var_5207_cast_fp16)[name = string("Na_39_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_153_cast_fp16 = add(x = eye_to_fp16, y = N_39_cast_fp16)[name = string("Minv_153_cast_fp16")];
            bool Np_115_transpose_x_0 = const()[name = string("Np_115_transpose_x_0"), val = bool(false)];
            bool Np_115_transpose_y_0 = const()[name = string("Np_115_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_115_cast_fp16 = matmul(transpose_x = Np_115_transpose_x_0, transpose_y = Np_115_transpose_y_0, x = N_39_cast_fp16, y = Na_39_cast_fp16)[name = string("Np_115_cast_fp16")];
            bool Npa_77_transpose_x_0 = const()[name = string("Npa_77_transpose_x_0"), val = bool(false)];
            bool Npa_77_transpose_y_0 = const()[name = string("Npa_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_77_cast_fp16 = matmul(transpose_x = Npa_77_transpose_x_0, transpose_y = Npa_77_transpose_y_0, x = Na_39_cast_fp16, y = N_39_cast_fp16)[name = string("Npa_77_cast_fp16")];
            bool var_5213_transpose_x_0 = const()[name = string("op_5213_transpose_x_0"), val = bool(false)];
            bool var_5213_transpose_y_0 = const()[name = string("op_5213_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5213_cast_fp16 = matmul(transpose_x = var_5213_transpose_x_0, transpose_y = var_5213_transpose_y_0, x = Minv_153_cast_fp16, y = Np_115_cast_fp16)[name = string("op_5213_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_155_cast_fp16 = add(x = Minv_153_cast_fp16, y = var_5213_cast_fp16)[name = string("Minv_155_cast_fp16")];
            bool Np_117_transpose_x_0 = const()[name = string("Np_117_transpose_x_0"), val = bool(false)];
            bool Np_117_transpose_y_0 = const()[name = string("Np_117_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_117_cast_fp16 = matmul(transpose_x = Np_117_transpose_x_0, transpose_y = Np_117_transpose_y_0, x = Np_115_cast_fp16, y = Npa_77_cast_fp16)[name = string("Np_117_cast_fp16")];
            bool Npa_79_transpose_x_0 = const()[name = string("Npa_79_transpose_x_0"), val = bool(false)];
            bool Npa_79_transpose_y_0 = const()[name = string("Npa_79_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_79_cast_fp16 = matmul(transpose_x = Npa_79_transpose_x_0, transpose_y = Npa_79_transpose_y_0, x = Npa_77_cast_fp16, y = Np_115_cast_fp16)[name = string("Npa_79_cast_fp16")];
            bool var_5218_transpose_x_0 = const()[name = string("op_5218_transpose_x_0"), val = bool(false)];
            bool var_5218_transpose_y_0 = const()[name = string("op_5218_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5218_cast_fp16 = matmul(transpose_x = var_5218_transpose_x_0, transpose_y = var_5218_transpose_y_0, x = Minv_155_cast_fp16, y = Np_117_cast_fp16)[name = string("op_5218_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_157_cast_fp16 = add(x = Minv_155_cast_fp16, y = var_5218_cast_fp16)[name = string("Minv_157_cast_fp16")];
            bool Np_119_transpose_x_0 = const()[name = string("Np_119_transpose_x_0"), val = bool(false)];
            bool Np_119_transpose_y_0 = const()[name = string("Np_119_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_119_cast_fp16 = matmul(transpose_x = Np_119_transpose_x_0, transpose_y = Np_119_transpose_y_0, x = Np_117_cast_fp16, y = Npa_79_cast_fp16)[name = string("Np_119_cast_fp16")];
            bool var_5222_transpose_x_0 = const()[name = string("op_5222_transpose_x_0"), val = bool(false)];
            bool var_5222_transpose_y_0 = const()[name = string("op_5222_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5222_cast_fp16 = matmul(transpose_x = var_5222_transpose_x_0, transpose_y = var_5222_transpose_y_0, x = Minv_157_cast_fp16, y = Np_119_cast_fp16)[name = string("op_5222_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_159_cast_fp16 = add(x = Minv_157_cast_fp16, y = var_5222_cast_fp16)[name = string("Minv_159_cast_fp16")];
            bool u_39_transpose_x_0 = const()[name = string("u_39_transpose_x_0"), val = bool(false)];
            bool u_39_transpose_y_0 = const()[name = string("u_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_39_cast_fp16 = matmul(transpose_x = u_39_transpose_x_0, transpose_y = u_39_transpose_y_0, x = Minv_159_cast_fp16, y = vb_49_cast_fp16)[name = string("u_39_cast_fp16")];
            bool var_5229_transpose_x_1 = const()[name = string("op_5229_transpose_x_1"), val = bool(false)];
            bool var_5229_transpose_y_1 = const()[name = string("op_5229_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_5229_cast_fp16 = matmul(transpose_x = var_5229_transpose_x_1, transpose_y = var_5229_transpose_y_1, x = q_81_cast_fp16, y = k_83_cast_fp16)[name = string("op_5229_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_39_cast_fp16 = mul(x = var_5229_cast_fp16, y = D_39_cast_fp16)[name = string("A_39_cast_fp16")];
            tensor<fp16, [16]> var_5231_bias_0_to_fp16 = const()[name = string("op_5231_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_5231_cast_fp16 = linear(bias = var_5231_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_49_cast_fp16)[name = string("op_5231_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_29_cast_fp16 = exp(x = var_5231_cast_fp16)[name = string("cum_29_cast_fp16")];
            tensor<int32, [1]> var_5234_axes_0 = const()[name = string("op_5234_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_5234_cast_fp16 = expand_dims(axes = var_5234_axes_0, x = cum_29_cast_fp16)[name = string("op_5234_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5235_cast_fp16 = mul(x = kb_39_cast_fp16, y = var_5234_cast_fp16)[name = string("op_5235_cast_fp16")];
            bool w_39_transpose_x_0 = const()[name = string("w_39_transpose_x_0"), val = bool(false)];
            bool w_39_transpose_y_0 = const()[name = string("w_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_39_cast_fp16 = matmul(transpose_x = w_39_transpose_x_0, transpose_y = w_39_transpose_y_0, x = Minv_159_cast_fp16, y = var_5235_cast_fp16)[name = string("w_39_cast_fp16")];
            bool var_5237_transpose_x_0 = const()[name = string("op_5237_transpose_x_0"), val = bool(false)];
            bool var_5237_transpose_y_0 = const()[name = string("op_5237_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5237_cast_fp16 = matmul(transpose_x = var_5237_transpose_x_0, transpose_y = var_5237_transpose_y_0, x = w_39_cast_fp16, y = S_29_cast_fp16)[name = string("op_5237_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_29_cast_fp16 = sub(x = u_39_cast_fp16, y = var_5237_cast_fp16)[name = string("vnew_29_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5242_cast_fp16 = mul(x = q_81_cast_fp16, y = var_5234_cast_fp16)[name = string("op_5242_cast_fp16")];
            bool var_5243_transpose_x_0 = const()[name = string("op_5243_transpose_x_0"), val = bool(false)];
            bool var_5243_transpose_y_0 = const()[name = string("op_5243_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5243_cast_fp16 = matmul(transpose_x = var_5243_transpose_x_0, transpose_y = var_5243_transpose_y_0, x = var_5242_cast_fp16, y = S_29_cast_fp16)[name = string("op_5243_cast_fp16")];
            bool var_5244_transpose_x_0 = const()[name = string("op_5244_transpose_x_0"), val = bool(false)];
            bool var_5244_transpose_y_0 = const()[name = string("op_5244_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5244_cast_fp16 = matmul(transpose_x = var_5244_transpose_x_0, transpose_y = var_5244_transpose_y_0, x = A_39_cast_fp16, y = vnew_29_cast_fp16)[name = string("op_5244_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_9_cast_fp16 = add(x = var_5243_cast_fp16, y = var_5244_cast_fp16)[name = string("oc_9_cast_fp16")];
            int32 var_5248 = const()[name = string("op_5248"), val = int32(2)];
            bool o_27_interleave_0 = const()[name = string("o_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 128]> o_27_cast_fp16 = concat(axis = var_5248, interleave = o_27_interleave_0, values = (var_4791_cast_fp16, var_4938_cast_fp16, var_5092_cast_fp16, oc_9_cast_fp16))[name = string("o_27_cast_fp16")];
            tensor<int32, [4]> var_5254 = const()[name = string("op_5254"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_5260 = const()[name = string("op_5260"), val = tensor<int32, [4]>([1, 64, 4, 128])];
            tensor<fp16, [1, 64, 4, 128]> input_81_cast_fp16 = reshape(shape = var_5260, x = z_9_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> o_29_cast_fp16 = transpose(perm = var_5254, x = o_27_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 64, 4, 128]> var_5262_cast_fp16 = mul(x = o_29_cast_fp16, y = o_29_cast_fp16)[name = string("op_5262_cast_fp16")];
            tensor<int32, [1]> var_5267_axes_0 = const()[name = string("op_5267_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5267_keep_dims_0 = const()[name = string("op_5267_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 4, 1]> var_5267_cast_fp16 = reduce_mean(axes = var_5267_axes_0, keep_dims = var_5267_keep_dims_0, x = var_5262_cast_fp16)[name = string("op_5267_cast_fp16")];
            fp16 var_5269_to_fp16 = const()[name = string("op_5269_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 64, 4, 1]> var_5270_cast_fp16 = add(x = var_5267_cast_fp16, y = var_5269_to_fp16)[name = string("op_5270_cast_fp16")];
            fp32 var_5271_epsilon_0 = const()[name = string("op_5271_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 4, 1]> var_5271_cast_fp16 = rsqrt(epsilon = var_5271_epsilon_0, x = var_5270_cast_fp16)[name = string("op_5271_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> var_5272_cast_fp16 = mul(x = o_29_cast_fp16, y = var_5271_cast_fp16)[name = string("op_5272_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> var_5274_cast_fp16 = silu(x = input_81_cast_fp16)[name = string("op_5274_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> o_31_cast_fp16 = mul(x = var_5272_cast_fp16, y = var_5274_cast_fp16)[name = string("o_31_cast_fp16")];
            tensor<int32, [3]> var_5279 = const()[name = string("op_5279"), val = tensor<int32, [3]>([1, 64, 512])];
            tensor<fp16, [1, 64, 512]> input_83_cast_fp16 = reshape(shape = var_5279, x = o_31_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [640, 512]> layers_5_out_weight_to_fp16 = const()[name = string("layers_5_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84536512)))];
            tensor<fp16, [1, 64, 640]> linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_out_weight_to_fp16, x = input_83_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [1, 64, 640]> x_57_cast_fp16 = add(x = x_53_cast_fp16, y = linear_21_cast_fp16)[name = string("x_57_cast_fp16")];
            fp16 var_5292_to_fp16 = const()[name = string("op_5292_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 64, 640]> x_59_cast_fp16 = mul(x = x_57_cast_fp16, y = var_5292_to_fp16)[name = string("x_59_cast_fp16")];
            fp16 var_5290_promoted_to_fp16 = const()[name = string("op_5290_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> var_5294_cast_fp16 = pow(x = x_59_cast_fp16, y = var_5290_promoted_to_fp16)[name = string("op_5294_cast_fp16")];
            tensor<int32, [1]> var_5296_axes_0 = const()[name = string("op_5296_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5296_keep_dims_0 = const()[name = string("op_5296_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5296_cast_fp16 = reduce_mean(axes = var_5296_axes_0, keep_dims = var_5296_keep_dims_0, x = var_5294_cast_fp16)[name = string("op_5296_cast_fp16")];
            fp16 var_5297_to_fp16 = const()[name = string("op_5297_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 64, 1]> var_5298_cast_fp16 = add(x = var_5296_cast_fp16, y = var_5297_to_fp16)[name = string("op_5298_cast_fp16")];
            fp32 var_5299_epsilon_0 = const()[name = string("op_5299_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_5299_cast_fp16 = rsqrt(epsilon = var_5299_epsilon_0, x = var_5298_cast_fp16)[name = string("op_5299_cast_fp16")];
            tensor<fp16, [1, 64, 640]> var_5300_cast_fp16 = mul(x = x_59_cast_fp16, y = var_5299_cast_fp16)[name = string("op_5300_cast_fp16")];
            tensor<fp16, [640]> layers_5_n2_w1_to_fp16 = const()[name = string("layers_5_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85191936)))];
            tensor<fp16, [1, 64, 640]> input_85_cast_fp16 = mul(x = var_5300_cast_fp16, y = layers_5_n2_w1_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_5_gate_up_weight_to_fp16 = const()[name = string("layers_5_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85193280)))];
            tensor<fp16, [1, 64, 5120]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_5_gate_up_weight_to_fp16, x = input_85_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [3]> input_87_begin_0 = const()[name = string("input_87_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_87_end_0 = const()[name = string("input_87_end_0"), val = tensor<int32, [3]>([1, 64, 2560])];
            tensor<bool, [3]> input_87_end_mask_0 = const()[name = string("input_87_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 2560]> input_87_cast_fp16 = slice_by_index(begin = input_87_begin_0, end = input_87_end_0, end_mask = input_87_end_mask_0, x = linear_22_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [1, 64, 2560]> var_5310_cast_fp16 = silu(x = input_87_cast_fp16)[name = string("op_5310_cast_fp16")];
            tensor<int32, [3]> var_5315_begin_0 = const()[name = string("op_5315_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_5315_end_0 = const()[name = string("op_5315_end_0"), val = tensor<int32, [3]>([1, 64, 5120])];
            tensor<bool, [3]> var_5315_end_mask_0 = const()[name = string("op_5315_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 2560]> var_5315_cast_fp16 = slice_by_index(begin = var_5315_begin_0, end = var_5315_end_0, end_mask = var_5315_end_mask_0, x = linear_22_cast_fp16)[name = string("op_5315_cast_fp16")];
            tensor<fp16, [1, 64, 2560]> input_89_cast_fp16 = mul(x = var_5310_cast_fp16, y = var_5315_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_5_down_weight_to_fp16 = const()[name = string("layers_5_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91746944)))];
            tensor<fp16, [1, 64, 640]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_down_weight_to_fp16, x = input_89_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<fp16, [1, 64, 640]> x_61_cast_fp16 = add(x = x_57_cast_fp16, y = linear_23_cast_fp16)[name = string("x_61_cast_fp16")];
            fp16 var_5328_to_fp16 = const()[name = string("op_5328_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [1, 64, 640]> x_63_cast_fp16 = mul(x = x_61_cast_fp16, y = var_5328_to_fp16)[name = string("x_63_cast_fp16")];
            fp16 var_5326_promoted_to_fp16 = const()[name = string("op_5326_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> var_5330_cast_fp16 = pow(x = x_63_cast_fp16, y = var_5326_promoted_to_fp16)[name = string("op_5330_cast_fp16")];
            tensor<int32, [1]> var_5332_axes_0 = const()[name = string("op_5332_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5332_keep_dims_0 = const()[name = string("op_5332_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5332_cast_fp16 = reduce_mean(axes = var_5332_axes_0, keep_dims = var_5332_keep_dims_0, x = var_5330_cast_fp16)[name = string("op_5332_cast_fp16")];
            fp16 var_5333_to_fp16 = const()[name = string("op_5333_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 64, 1]> var_5334_cast_fp16 = add(x = var_5332_cast_fp16, y = var_5333_to_fp16)[name = string("op_5334_cast_fp16")];
            fp32 var_5335_epsilon_0 = const()[name = string("op_5335_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_5335_cast_fp16 = rsqrt(epsilon = var_5335_epsilon_0, x = var_5334_cast_fp16)[name = string("op_5335_cast_fp16")];
            tensor<fp16, [1, 64, 640]> var_5336_cast_fp16 = mul(x = x_63_cast_fp16, y = var_5335_cast_fp16)[name = string("op_5336_cast_fp16")];
            tensor<fp16, [640]> layers_6_n1_w1_to_fp16 = const()[name = string("layers_6_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95023808)))];
            tensor<fp16, [1, 64, 640]> input_91_cast_fp16 = mul(x = var_5336_cast_fp16, y = layers_6_n1_w1_to_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_6_qkvzba_weight_to_fp16 = const()[name = string("layers_6_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95025152)))];
            tensor<fp16, [1, 64, 1544]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_qkvzba_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<int32, [3]> qkv_11_begin_0 = const()[name = string("qkv_11_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_11_end_0 = const()[name = string("qkv_11_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> qkv_11_end_mask_0 = const()[name = string("qkv_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 1024]> qkv_11_cast_fp16 = slice_by_index(begin = qkv_11_begin_0, end = qkv_11_end_0, end_mask = qkv_11_end_mask_0, x = linear_24_cast_fp16)[name = string("qkv_11_cast_fp16")];
            tensor<int32, [3]> z_begin_0 = const()[name = string("z_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_end_0 = const()[name = string("z_end_0"), val = tensor<int32, [3]>([1, 64, 1536])];
            tensor<bool, [3]> z_end_mask_0 = const()[name = string("z_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 512]> z_cast_fp16 = slice_by_index(begin = z_begin_0, end = z_end_0, end_mask = z_end_mask_0, x = linear_24_cast_fp16)[name = string("z_cast_fp16")];
            tensor<int32, [3]> b_begin_0 = const()[name = string("b_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_end_0 = const()[name = string("b_end_0"), val = tensor<int32, [3]>([1, 64, 1540])];
            tensor<bool, [3]> b_end_mask_0 = const()[name = string("b_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 4]> b_cast_fp16 = slice_by_index(begin = b_begin_0, end = b_end_0, end_mask = b_end_mask_0, x = linear_24_cast_fp16)[name = string("b_cast_fp16")];
            tensor<int32, [3]> a_begin_0 = const()[name = string("a_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_end_0 = const()[name = string("a_end_0"), val = tensor<int32, [3]>([1, 64, 1])];
            tensor<bool, [3]> a_end_mask_0 = const()[name = string("a_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 4]> a_cast_fp16 = slice_by_index(begin = a_begin_0, end = a_end_0, end_mask = a_end_mask_0, x = linear_24_cast_fp16)[name = string("a_cast_fp16")];
            int32 var_5375 = const()[name = string("op_5375"), val = int32(1)];
            bool xp_interleave_0 = const()[name = string("xp_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 67, 1024]> xp_cast_fp16 = concat(axis = var_5375, interleave = xp_interleave_0, values = (zpad_to_fp16, qkv_11_cast_fp16))[name = string("xp_cast_fp16")];
            tensor<int32, [3]> var_5386_begin_0 = const()[name = string("op_5386_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_5386_end_0 = const()[name = string("op_5386_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_5386_end_mask_0 = const()[name = string("op_5386_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 1024]> var_5386_cast_fp16 = slice_by_index(begin = var_5386_begin_0, end = var_5386_end_0, end_mask = var_5386_end_mask_0, x = xp_cast_fp16)[name = string("op_5386_cast_fp16")];
            tensor<fp16, [1024]> var_5394_to_fp16 = const()[name = string("op_5394_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97001536)))];
            tensor<fp16, [1, 64, 1024]> var_5395_cast_fp16 = mul(x = var_5386_cast_fp16, y = var_5394_to_fp16)[name = string("op_5395_cast_fp16")];
            tensor<int32, [3]> var_5405_begin_0 = const()[name = string("op_5405_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_5405_end_0 = const()[name = string("op_5405_end_0"), val = tensor<int32, [3]>([1, 66, 1024])];
            tensor<bool, [3]> var_5405_end_mask_0 = const()[name = string("op_5405_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_5405_cast_fp16 = slice_by_index(begin = var_5405_begin_0, end = var_5405_end_0, end_mask = var_5405_end_mask_0, x = xp_cast_fp16)[name = string("op_5405_cast_fp16")];
            tensor<fp16, [1024]> var_5413_to_fp16 = const()[name = string("op_5413_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97003648)))];
            tensor<fp16, [1, 64, 1024]> var_5414_cast_fp16 = mul(x = var_5405_cast_fp16, y = var_5413_to_fp16)[name = string("op_5414_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> var_5416_cast_fp16 = add(x = var_5395_cast_fp16, y = var_5414_cast_fp16)[name = string("op_5416_cast_fp16")];
            tensor<int32, [3]> var_5426_begin_0 = const()[name = string("op_5426_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_5426_end_0 = const()[name = string("op_5426_end_0"), val = tensor<int32, [3]>([1, 65, 1024])];
            tensor<bool, [3]> var_5426_end_mask_0 = const()[name = string("op_5426_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_5426_cast_fp16 = slice_by_index(begin = var_5426_begin_0, end = var_5426_end_0, end_mask = var_5426_end_mask_0, x = xp_cast_fp16)[name = string("op_5426_cast_fp16")];
            tensor<fp16, [1024]> var_5434_to_fp16 = const()[name = string("op_5434_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97005760)))];
            tensor<fp16, [1, 64, 1024]> var_5435_cast_fp16 = mul(x = var_5426_cast_fp16, y = var_5434_to_fp16)[name = string("op_5435_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> var_5437_cast_fp16 = add(x = var_5416_cast_fp16, y = var_5435_cast_fp16)[name = string("op_5437_cast_fp16")];
            tensor<int32, [3]> var_5447_begin_0 = const()[name = string("op_5447_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_5447_end_0 = const()[name = string("op_5447_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> var_5447_end_mask_0 = const()[name = string("op_5447_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 64, 1024]> var_5447_cast_fp16 = slice_by_index(begin = var_5447_begin_0, end = var_5447_end_0, end_mask = var_5447_end_mask_0, x = xp_cast_fp16)[name = string("op_5447_cast_fp16")];
            tensor<fp16, [1024]> var_5455_to_fp16 = const()[name = string("op_5455_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97007872)))];
            tensor<fp16, [1, 64, 1024]> var_5456_cast_fp16 = mul(x = var_5447_cast_fp16, y = var_5455_to_fp16)[name = string("op_5456_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> input_93_cast_fp16 = add(x = var_5437_cast_fp16, y = var_5456_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [1, 64, 1024]> c_cast_fp16 = silu(x = input_93_cast_fp16)[name = string("c_cast_fp16")];
            tensor<int32, [3]> var_5464_begin_0 = const()[name = string("op_5464_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_5464_end_0 = const()[name = string("op_5464_end_0"), val = tensor<int32, [3]>([1, 64, 256])];
            tensor<bool, [3]> var_5464_end_mask_0 = const()[name = string("op_5464_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 256]> var_5464_cast_fp16 = slice_by_index(begin = var_5464_begin_0, end = var_5464_end_0, end_mask = var_5464_end_mask_0, x = c_cast_fp16)[name = string("op_5464_cast_fp16")];
            tensor<int32, [4]> var_5469 = const()[name = string("op_5469"), val = tensor<int32, [4]>([1, 64, 2, 128])];
            tensor<fp16, [1, 64, 2, 128]> q_83_cast_fp16 = reshape(shape = var_5469, x = var_5464_cast_fp16)[name = string("q_83_cast_fp16")];
            tensor<int32, [3]> var_5475_begin_0 = const()[name = string("op_5475_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_5475_end_0 = const()[name = string("op_5475_end_0"), val = tensor<int32, [3]>([1, 64, 512])];
            tensor<bool, [3]> var_5475_end_mask_0 = const()[name = string("op_5475_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 256]> var_5475_cast_fp16 = slice_by_index(begin = var_5475_begin_0, end = var_5475_end_0, end_mask = var_5475_end_mask_0, x = c_cast_fp16)[name = string("op_5475_cast_fp16")];
            tensor<int32, [4]> var_5480 = const()[name = string("op_5480"), val = tensor<int32, [4]>([1, 64, 2, 128])];
            tensor<fp16, [1, 64, 2, 128]> k_85_cast_fp16 = reshape(shape = var_5480, x = var_5475_cast_fp16)[name = string("k_85_cast_fp16")];
            tensor<int32, [3]> var_5486_begin_0 = const()[name = string("op_5486_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_5486_end_0 = const()[name = string("op_5486_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> var_5486_end_mask_0 = const()[name = string("op_5486_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 512]> var_5486_cast_fp16 = slice_by_index(begin = var_5486_begin_0, end = var_5486_end_0, end_mask = var_5486_end_mask_0, x = c_cast_fp16)[name = string("op_5486_cast_fp16")];
            tensor<int32, [4]> var_5491 = const()[name = string("op_5491"), val = tensor<int32, [4]>([1, 64, 4, 128])];
            tensor<fp16, [1, 64, 4, 128]> var_5492_cast_fp16 = reshape(shape = var_5491, x = var_5486_cast_fp16)[name = string("op_5492_cast_fp16")];
            tensor<int32, [4]> var_5497 = const()[name = string("op_5497"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5499_to_fp16 = const()[name = string("op_5499_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 64, 2, 128]> q_85_cast_fp16 = mul(x = q_83_cast_fp16, y = var_5499_to_fp16)[name = string("q_85_cast_fp16")];
            fp16 var_5501_to_fp16 = const()[name = string("op_5501_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 64, 2, 128]> k_87_cast_fp16 = mul(x = k_85_cast_fp16, y = var_5501_to_fp16)[name = string("k_87_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_5503_cast_fp16 = mul(x = q_85_cast_fp16, y = q_85_cast_fp16)[name = string("op_5503_cast_fp16")];
            tensor<int32, [1]> var_5508_axes_0 = const()[name = string("op_5508_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5508_keep_dims_0 = const()[name = string("op_5508_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 2, 1]> var_5508_cast_fp16 = reduce_mean(axes = var_5508_axes_0, keep_dims = var_5508_keep_dims_0, x = var_5503_cast_fp16)[name = string("op_5508_cast_fp16")];
            fp16 var_5509_promoted_to_fp16 = const()[name = string("op_5509_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 64, 2, 1]> var_5510_cast_fp16 = mul(x = var_5508_cast_fp16, y = var_5509_promoted_to_fp16)[name = string("op_5510_cast_fp16")];
            fp16 var_5512_to_fp16 = const()[name = string("op_5512_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 64, 2, 1]> var_5513_cast_fp16 = add(x = var_5510_cast_fp16, y = var_5512_to_fp16)[name = string("op_5513_cast_fp16")];
            fp32 var_5514_epsilon_0 = const()[name = string("op_5514_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 2, 1]> var_5514_cast_fp16 = rsqrt(epsilon = var_5514_epsilon_0, x = var_5513_cast_fp16)[name = string("op_5514_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_5515_cast_fp16 = mul(x = q_85_cast_fp16, y = var_5514_cast_fp16)[name = string("op_5515_cast_fp16")];
            fp16 var_5516_to_fp16 = const()[name = string("op_5516_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 64, 2, 128]> q_87_cast_fp16 = mul(x = var_5515_cast_fp16, y = var_5516_to_fp16)[name = string("q_87_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> var_5518_cast_fp16 = mul(x = k_87_cast_fp16, y = k_87_cast_fp16)[name = string("op_5518_cast_fp16")];
            tensor<int32, [1]> var_5523_axes_0 = const()[name = string("op_5523_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5523_keep_dims_0 = const()[name = string("op_5523_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 2, 1]> var_5523_cast_fp16 = reduce_mean(axes = var_5523_axes_0, keep_dims = var_5523_keep_dims_0, x = var_5518_cast_fp16)[name = string("op_5523_cast_fp16")];
            fp16 var_5524_promoted_to_fp16 = const()[name = string("op_5524_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 64, 2, 1]> var_5525_cast_fp16 = mul(x = var_5523_cast_fp16, y = var_5524_promoted_to_fp16)[name = string("op_5525_cast_fp16")];
            fp16 var_5527_to_fp16 = const()[name = string("op_5527_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 64, 2, 1]> var_5528_cast_fp16 = add(x = var_5525_cast_fp16, y = var_5527_to_fp16)[name = string("op_5528_cast_fp16")];
            fp32 var_5529_epsilon_0 = const()[name = string("op_5529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 2, 1]> var_5529_cast_fp16 = rsqrt(epsilon = var_5529_epsilon_0, x = var_5528_cast_fp16)[name = string("op_5529_cast_fp16")];
            tensor<fp16, [1, 64, 2, 128]> k_89_cast_fp16 = mul(x = k_87_cast_fp16, y = var_5529_cast_fp16)[name = string("k_89_cast_fp16")];
            tensor<int32, [4]> var_5535 = const()[name = string("op_5535"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_5548_axes_0 = const()[name = string("op_5548_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 128]> var_5536_cast_fp16 = transpose(perm = var_5535, x = q_87_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 2, 1, 64, 128]> var_5548_cast_fp16 = expand_dims(axes = var_5548_axes_0, x = var_5536_cast_fp16)[name = string("op_5548_cast_fp16")];
            tensor<int32, [5]> var_5556_reps_0 = const()[name = string("op_5556_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 64, 128]> var_5556_cast_fp16 = tile(reps = var_5556_reps_0, x = var_5548_cast_fp16)[name = string("op_5556_cast_fp16")];
            tensor<int32, [4]> var_5561 = const()[name = string("op_5561"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> q_89_cast_fp16 = reshape(shape = var_5561, x = var_5556_cast_fp16)[name = string("q_89_cast_fp16")];
            tensor<int32, [4]> var_5567 = const()[name = string("op_5567"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_5580_axes_0 = const()[name = string("op_5580_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 128]> var_5568_cast_fp16 = transpose(perm = var_5567, x = k_89_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 2, 1, 64, 128]> var_5580_cast_fp16 = expand_dims(axes = var_5580_axes_0, x = var_5568_cast_fp16)[name = string("op_5580_cast_fp16")];
            tensor<int32, [5]> var_5588_reps_0 = const()[name = string("op_5588_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 64, 128]> var_5588_cast_fp16 = tile(reps = var_5588_reps_0, x = var_5580_cast_fp16)[name = string("op_5588_cast_fp16")];
            tensor<int32, [4]> var_5593 = const()[name = string("op_5593"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<fp16, [1, 4, 64, 128]> k_91_cast_fp16 = reshape(shape = var_5593, x = var_5588_cast_fp16)[name = string("k_91_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_5595_cast_fp16 = sigmoid(x = b_cast_fp16)[name = string("op_5595_cast_fp16")];
            tensor<int32, [3]> var_5599 = const()[name = string("op_5599"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_6_dt_bias_to_fp16 = const()[name = string("layers_6_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 64, 4]> input_95_cast_fp16 = add(x = a_cast_fp16, y = layers_6_dt_bias_to_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_5603_cast_fp16 = relu(x = input_95_cast_fp16)[name = string("op_5603_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_5604_cast_fp16 = abs(x = input_95_cast_fp16)[name = string("op_5604_cast_fp16")];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 64, 4]> var_5605_cast_fp16 = mul(x = var_5604_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_5605_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_5606_cast_fp16 = exp(x = var_5605_cast_fp16)[name = string("op_5606_cast_fp16")];
            fp16 var_5608_to_fp16 = const()[name = string("op_5608_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 64, 4]> var_5609_cast_fp16 = add(x = var_5606_cast_fp16, y = var_5608_to_fp16)[name = string("op_5609_cast_fp16")];
            fp32 var_5610_epsilon_0 = const()[name = string("op_5610_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 64, 4]> var_5610_cast_fp16 = log(epsilon = var_5610_epsilon_0, x = var_5609_cast_fp16)[name = string("op_5610_cast_fp16")];
            tensor<fp16, [1, 64, 4]> var_5612_cast_fp16 = add(x = var_5603_cast_fp16, y = var_5610_cast_fp16)[name = string("op_5612_cast_fp16")];
            tensor<fp16, [4]> layers_6_negA_to_fp16 = const()[name = string("layers_6_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.07cp+1, -0x1.eccp+3, -0x1.63p+3, -0x1.728p+1])];
            tensor<fp16, [1, 64, 4]> var_5613_cast_fp16 = mul(x = layers_6_negA_to_fp16, y = var_5612_cast_fp16)[name = string("op_5613_cast_fp16")];
            tensor<int32, [3]> var_5617 = const()[name = string("op_5617"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_5620_axes_0 = const()[name = string("op_5620_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 64]> beta_cast_fp16 = transpose(perm = var_5599, x = var_5595_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 4, 64, 1]> var_5620_cast_fp16 = expand_dims(axes = var_5620_axes_0, x = beta_cast_fp16)[name = string("op_5620_cast_fp16")];
            fp16 var_5621_to_fp16 = const()[name = string("op_5621_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 64, 1]> var_5622_cast_fp16 = mul(x = var_5620_cast_fp16, y = var_5621_to_fp16)[name = string("op_5622_cast_fp16")];
            tensor<fp16, [1, 4, 64, 128]> v_15_cast_fp16 = transpose(perm = var_5497, x = var_5492_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 4, 64, 128]> vb_51_cast_fp16 = mul(x = v_15_cast_fp16, y = var_5622_cast_fp16)[name = string("vb_51_cast_fp16")];
            tensor<int32, [4]> q_91_begin_0 = const()[name = string("q_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_91_end_0 = const()[name = string("q_91_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_91_end_mask_0 = const()[name = string("q_91_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, x = q_89_cast_fp16)[name = string("q_91_cast_fp16")];
            tensor<int32, [4]> k_93_begin_0 = const()[name = string("k_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_93_end_0 = const()[name = string("k_93_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_93_end_mask_0 = const()[name = string("k_93_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, x = k_91_cast_fp16)[name = string("k_93_cast_fp16")];
            tensor<fp16, [1, 4, 64, 128]> var_5656_cast_fp16 = mul(x = k_91_cast_fp16, y = var_5620_cast_fp16)[name = string("op_5656_cast_fp16")];
            tensor<int32, [4]> kb_41_begin_0 = const()[name = string("kb_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_41_end_0 = const()[name = string("kb_41_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_41_end_mask_0 = const()[name = string("kb_41_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_41_cast_fp16 = slice_by_index(begin = kb_41_begin_0, end = kb_41_end_0, end_mask = kb_41_end_mask_0, x = var_5656_cast_fp16)[name = string("kb_41_cast_fp16")];
            tensor<int32, [4]> vb_53_begin_0 = const()[name = string("vb_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_53_end_0 = const()[name = string("vb_53_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_53_end_mask_0 = const()[name = string("vb_53_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_53_cast_fp16 = slice_by_index(begin = vb_53_begin_0, end = vb_53_end_0, end_mask = vb_53_end_mask_0, x = vb_51_cast_fp16)[name = string("vb_53_cast_fp16")];
            tensor<int32, [3]> gdec_53_begin_0 = const()[name = string("gdec_53_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_53_end_0 = const()[name = string("gdec_53_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_53_end_mask_0 = const()[name = string("gdec_53_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 64]> gdec_51_cast_fp16 = transpose(perm = var_5617, x = var_5613_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 4, 16]> gdec_53_cast_fp16 = slice_by_index(begin = gdec_53_begin_0, end = gdec_53_end_0, end_mask = gdec_53_end_mask_0, x = gdec_51_cast_fp16)[name = string("gdec_53_cast_fp16")];
            tensor<int32, [2]> var_5704 = const()[name = string("op_5704"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_5705_cast_fp16 = reshape(shape = var_5704, x = gdec_53_cast_fp16)[name = string("op_5705_cast_fp16")];
            tensor<fp16, [4, 256]> var_5706_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5705_cast_fp16)[name = string("op_5706_cast_fp16")];
            tensor<int32, [4]> var_5711 = const()[name = string("op_5711"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_41_cast_fp16 = reshape(shape = var_5711, x = var_5706_cast_fp16)[name = string("P_41_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5713_cast_fp16 = exp(x = P_41_cast_fp16)[name = string("op_5713_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_41_cast_fp16 = mul(x = var_5713_cast_fp16, y = tril_incl_to_fp16)[name = string("D_41_cast_fp16")];
            bool kk_41_transpose_x_1 = const()[name = string("kk_41_transpose_x_1"), val = bool(false)];
            bool kk_41_transpose_y_1 = const()[name = string("kk_41_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_41_cast_fp16 = matmul(transpose_x = kk_41_transpose_x_1, transpose_y = kk_41_transpose_y_1, x = kb_41_cast_fp16, y = k_93_cast_fp16)[name = string("kk_41_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_5719_cast_fp16 = mul(x = kk_41_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_5719_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5720_cast_fp16 = mul(x = var_5719_cast_fp16, y = D_41_cast_fp16)[name = string("op_5720_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_41_cast_fp16 = mul(x = var_5720_cast_fp16, y = cum_tail_to_fp16)[name = string("N_41_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5722_cast_fp16 = mul(x = D_41_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5722_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_41_cast_fp16 = mul(x = kk_41_cast_fp16, y = var_5722_cast_fp16)[name = string("Na_41_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_161_cast_fp16 = add(x = eye_to_fp16, y = N_41_cast_fp16)[name = string("Minv_161_cast_fp16")];
            bool Np_121_transpose_x_0 = const()[name = string("Np_121_transpose_x_0"), val = bool(false)];
            bool Np_121_transpose_y_0 = const()[name = string("Np_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_121_cast_fp16 = matmul(transpose_x = Np_121_transpose_x_0, transpose_y = Np_121_transpose_y_0, x = N_41_cast_fp16, y = Na_41_cast_fp16)[name = string("Np_121_cast_fp16")];
            bool Npa_81_transpose_x_0 = const()[name = string("Npa_81_transpose_x_0"), val = bool(false)];
            bool Npa_81_transpose_y_0 = const()[name = string("Npa_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_81_cast_fp16 = matmul(transpose_x = Npa_81_transpose_x_0, transpose_y = Npa_81_transpose_y_0, x = Na_41_cast_fp16, y = N_41_cast_fp16)[name = string("Npa_81_cast_fp16")];
            bool var_5728_transpose_x_0 = const()[name = string("op_5728_transpose_x_0"), val = bool(false)];
            bool var_5728_transpose_y_0 = const()[name = string("op_5728_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5728_cast_fp16 = matmul(transpose_x = var_5728_transpose_x_0, transpose_y = var_5728_transpose_y_0, x = Minv_161_cast_fp16, y = Np_121_cast_fp16)[name = string("op_5728_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_163_cast_fp16 = add(x = Minv_161_cast_fp16, y = var_5728_cast_fp16)[name = string("Minv_163_cast_fp16")];
            bool Np_123_transpose_x_0 = const()[name = string("Np_123_transpose_x_0"), val = bool(false)];
            bool Np_123_transpose_y_0 = const()[name = string("Np_123_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_123_cast_fp16 = matmul(transpose_x = Np_123_transpose_x_0, transpose_y = Np_123_transpose_y_0, x = Np_121_cast_fp16, y = Npa_81_cast_fp16)[name = string("Np_123_cast_fp16")];
            bool Npa_83_transpose_x_0 = const()[name = string("Npa_83_transpose_x_0"), val = bool(false)];
            bool Npa_83_transpose_y_0 = const()[name = string("Npa_83_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_83_cast_fp16 = matmul(transpose_x = Npa_83_transpose_x_0, transpose_y = Npa_83_transpose_y_0, x = Npa_81_cast_fp16, y = Np_121_cast_fp16)[name = string("Npa_83_cast_fp16")];
            bool var_5733_transpose_x_0 = const()[name = string("op_5733_transpose_x_0"), val = bool(false)];
            bool var_5733_transpose_y_0 = const()[name = string("op_5733_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5733_cast_fp16 = matmul(transpose_x = var_5733_transpose_x_0, transpose_y = var_5733_transpose_y_0, x = Minv_163_cast_fp16, y = Np_123_cast_fp16)[name = string("op_5733_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_165_cast_fp16 = add(x = Minv_163_cast_fp16, y = var_5733_cast_fp16)[name = string("Minv_165_cast_fp16")];
            bool Np_125_transpose_x_0 = const()[name = string("Np_125_transpose_x_0"), val = bool(false)];
            bool Np_125_transpose_y_0 = const()[name = string("Np_125_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_125_cast_fp16 = matmul(transpose_x = Np_125_transpose_x_0, transpose_y = Np_125_transpose_y_0, x = Np_123_cast_fp16, y = Npa_83_cast_fp16)[name = string("Np_125_cast_fp16")];
            bool var_5737_transpose_x_0 = const()[name = string("op_5737_transpose_x_0"), val = bool(false)];
            bool var_5737_transpose_y_0 = const()[name = string("op_5737_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5737_cast_fp16 = matmul(transpose_x = var_5737_transpose_x_0, transpose_y = var_5737_transpose_y_0, x = Minv_165_cast_fp16, y = Np_125_cast_fp16)[name = string("op_5737_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_167_cast_fp16 = add(x = Minv_165_cast_fp16, y = var_5737_cast_fp16)[name = string("Minv_167_cast_fp16")];
            bool u_41_transpose_x_0 = const()[name = string("u_41_transpose_x_0"), val = bool(false)];
            bool u_41_transpose_y_0 = const()[name = string("u_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_41_cast_fp16 = matmul(transpose_x = u_41_transpose_x_0, transpose_y = u_41_transpose_y_0, x = Minv_167_cast_fp16, y = vb_53_cast_fp16)[name = string("u_41_cast_fp16")];
            bool var_5744_transpose_x_1 = const()[name = string("op_5744_transpose_x_1"), val = bool(false)];
            bool var_5744_transpose_y_1 = const()[name = string("op_5744_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_5744_cast_fp16 = matmul(transpose_x = var_5744_transpose_x_1, transpose_y = var_5744_transpose_y_1, x = q_91_cast_fp16, y = k_93_cast_fp16)[name = string("op_5744_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_41_cast_fp16 = mul(x = var_5744_cast_fp16, y = D_41_cast_fp16)[name = string("A_41_cast_fp16")];
            bool var_5746_transpose_x_0 = const()[name = string("op_5746_transpose_x_0"), val = bool(false)];
            bool var_5746_transpose_y_0 = const()[name = string("op_5746_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5746_cast_fp16 = matmul(transpose_x = var_5746_transpose_x_0, transpose_y = var_5746_transpose_y_0, x = A_41_cast_fp16, y = u_41_cast_fp16)[name = string("op_5746_cast_fp16")];
            tensor<fp16, [16]> var_5747_bias_0_to_fp16 = const()[name = string("op_5747_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_5747_cast_fp16 = linear(bias = var_5747_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_53_cast_fp16)[name = string("op_5747_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_31_cast_fp16 = exp(x = var_5747_cast_fp16)[name = string("tail_31_cast_fp16")];
            tensor<int32, [1]> var_5750_axes_0 = const()[name = string("op_5750_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_5750_cast_fp16 = expand_dims(axes = var_5750_axes_0, x = tail_31_cast_fp16)[name = string("op_5750_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5751_cast_fp16 = mul(x = k_93_cast_fp16, y = var_5750_cast_fp16)[name = string("op_5751_cast_fp16")];
            bool S_31_transpose_x_1 = const()[name = string("S_31_transpose_x_1"), val = bool(true)];
            bool S_31_transpose_y_1 = const()[name = string("S_31_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_31_cast_fp16 = matmul(transpose_x = S_31_transpose_x_1, transpose_y = S_31_transpose_y_1, x = var_5751_cast_fp16, y = u_41_cast_fp16)[name = string("S_31_cast_fp16")];
            tensor<int32, [4]> q_93_begin_0 = const()[name = string("q_93_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_93_end_0 = const()[name = string("q_93_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> q_93_end_mask_0 = const()[name = string("q_93_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_93_cast_fp16 = slice_by_index(begin = q_93_begin_0, end = q_93_end_0, end_mask = q_93_end_mask_0, x = q_89_cast_fp16)[name = string("q_93_cast_fp16")];
            tensor<int32, [4]> k_95_begin_0 = const()[name = string("k_95_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_95_end_0 = const()[name = string("k_95_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> k_95_end_mask_0 = const()[name = string("k_95_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_95_cast_fp16 = slice_by_index(begin = k_95_begin_0, end = k_95_end_0, end_mask = k_95_end_mask_0, x = k_91_cast_fp16)[name = string("k_95_cast_fp16")];
            tensor<int32, [4]> kb_43_begin_0 = const()[name = string("kb_43_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_43_end_0 = const()[name = string("kb_43_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> kb_43_end_mask_0 = const()[name = string("kb_43_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_43_cast_fp16 = slice_by_index(begin = kb_43_begin_0, end = kb_43_end_0, end_mask = kb_43_end_mask_0, x = var_5656_cast_fp16)[name = string("kb_43_cast_fp16")];
            tensor<int32, [4]> vb_55_begin_0 = const()[name = string("vb_55_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_55_end_0 = const()[name = string("vb_55_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> vb_55_end_mask_0 = const()[name = string("vb_55_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_55_cast_fp16 = slice_by_index(begin = vb_55_begin_0, end = vb_55_end_0, end_mask = vb_55_end_mask_0, x = vb_51_cast_fp16)[name = string("vb_55_cast_fp16")];
            tensor<int32, [3]> gdec_55_begin_0 = const()[name = string("gdec_55_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_55_end_0 = const()[name = string("gdec_55_end_0"), val = tensor<int32, [3]>([1, 4, 32])];
            tensor<bool, [3]> gdec_55_end_mask_0 = const()[name = string("gdec_55_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_55_cast_fp16 = slice_by_index(begin = gdec_55_begin_0, end = gdec_55_end_0, end_mask = gdec_55_end_mask_0, x = gdec_51_cast_fp16)[name = string("gdec_55_cast_fp16")];
            tensor<int32, [2]> var_5836 = const()[name = string("op_5836"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_5837_cast_fp16 = reshape(shape = var_5836, x = gdec_55_cast_fp16)[name = string("op_5837_cast_fp16")];
            tensor<fp16, [4, 256]> var_5838_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5837_cast_fp16)[name = string("op_5838_cast_fp16")];
            tensor<int32, [4]> var_5843 = const()[name = string("op_5843"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_43_cast_fp16 = reshape(shape = var_5843, x = var_5838_cast_fp16)[name = string("P_43_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5845_cast_fp16 = exp(x = P_43_cast_fp16)[name = string("op_5845_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_43_cast_fp16 = mul(x = var_5845_cast_fp16, y = tril_incl_to_fp16)[name = string("D_43_cast_fp16")];
            bool kk_43_transpose_x_1 = const()[name = string("kk_43_transpose_x_1"), val = bool(false)];
            bool kk_43_transpose_y_1 = const()[name = string("kk_43_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_43_cast_fp16 = matmul(transpose_x = kk_43_transpose_x_1, transpose_y = kk_43_transpose_y_1, x = kb_43_cast_fp16, y = k_95_cast_fp16)[name = string("kk_43_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_5851_cast_fp16 = mul(x = kk_43_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_5851_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5852_cast_fp16 = mul(x = var_5851_cast_fp16, y = D_43_cast_fp16)[name = string("op_5852_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_43_cast_fp16 = mul(x = var_5852_cast_fp16, y = cum_tail_to_fp16)[name = string("N_43_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5854_cast_fp16 = mul(x = D_43_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5854_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_43_cast_fp16 = mul(x = kk_43_cast_fp16, y = var_5854_cast_fp16)[name = string("Na_43_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_169_cast_fp16 = add(x = eye_to_fp16, y = N_43_cast_fp16)[name = string("Minv_169_cast_fp16")];
            bool Np_127_transpose_x_0 = const()[name = string("Np_127_transpose_x_0"), val = bool(false)];
            bool Np_127_transpose_y_0 = const()[name = string("Np_127_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_127_cast_fp16 = matmul(transpose_x = Np_127_transpose_x_0, transpose_y = Np_127_transpose_y_0, x = N_43_cast_fp16, y = Na_43_cast_fp16)[name = string("Np_127_cast_fp16")];
            bool Npa_85_transpose_x_0 = const()[name = string("Npa_85_transpose_x_0"), val = bool(false)];
            bool Npa_85_transpose_y_0 = const()[name = string("Npa_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_85_cast_fp16 = matmul(transpose_x = Npa_85_transpose_x_0, transpose_y = Npa_85_transpose_y_0, x = Na_43_cast_fp16, y = N_43_cast_fp16)[name = string("Npa_85_cast_fp16")];
            bool var_5860_transpose_x_0 = const()[name = string("op_5860_transpose_x_0"), val = bool(false)];
            bool var_5860_transpose_y_0 = const()[name = string("op_5860_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5860_cast_fp16 = matmul(transpose_x = var_5860_transpose_x_0, transpose_y = var_5860_transpose_y_0, x = Minv_169_cast_fp16, y = Np_127_cast_fp16)[name = string("op_5860_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_171_cast_fp16 = add(x = Minv_169_cast_fp16, y = var_5860_cast_fp16)[name = string("Minv_171_cast_fp16")];
            bool Np_129_transpose_x_0 = const()[name = string("Np_129_transpose_x_0"), val = bool(false)];
            bool Np_129_transpose_y_0 = const()[name = string("Np_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_129_cast_fp16 = matmul(transpose_x = Np_129_transpose_x_0, transpose_y = Np_129_transpose_y_0, x = Np_127_cast_fp16, y = Npa_85_cast_fp16)[name = string("Np_129_cast_fp16")];
            bool Npa_87_transpose_x_0 = const()[name = string("Npa_87_transpose_x_0"), val = bool(false)];
            bool Npa_87_transpose_y_0 = const()[name = string("Npa_87_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_87_cast_fp16 = matmul(transpose_x = Npa_87_transpose_x_0, transpose_y = Npa_87_transpose_y_0, x = Npa_85_cast_fp16, y = Np_127_cast_fp16)[name = string("Npa_87_cast_fp16")];
            bool var_5865_transpose_x_0 = const()[name = string("op_5865_transpose_x_0"), val = bool(false)];
            bool var_5865_transpose_y_0 = const()[name = string("op_5865_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5865_cast_fp16 = matmul(transpose_x = var_5865_transpose_x_0, transpose_y = var_5865_transpose_y_0, x = Minv_171_cast_fp16, y = Np_129_cast_fp16)[name = string("op_5865_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_173_cast_fp16 = add(x = Minv_171_cast_fp16, y = var_5865_cast_fp16)[name = string("Minv_173_cast_fp16")];
            bool Np_131_transpose_x_0 = const()[name = string("Np_131_transpose_x_0"), val = bool(false)];
            bool Np_131_transpose_y_0 = const()[name = string("Np_131_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_131_cast_fp16 = matmul(transpose_x = Np_131_transpose_x_0, transpose_y = Np_131_transpose_y_0, x = Np_129_cast_fp16, y = Npa_87_cast_fp16)[name = string("Np_131_cast_fp16")];
            bool var_5869_transpose_x_0 = const()[name = string("op_5869_transpose_x_0"), val = bool(false)];
            bool var_5869_transpose_y_0 = const()[name = string("op_5869_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5869_cast_fp16 = matmul(transpose_x = var_5869_transpose_x_0, transpose_y = var_5869_transpose_y_0, x = Minv_173_cast_fp16, y = Np_131_cast_fp16)[name = string("op_5869_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_175_cast_fp16 = add(x = Minv_173_cast_fp16, y = var_5869_cast_fp16)[name = string("Minv_175_cast_fp16")];
            bool u_43_transpose_x_0 = const()[name = string("u_43_transpose_x_0"), val = bool(false)];
            bool u_43_transpose_y_0 = const()[name = string("u_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_43_cast_fp16 = matmul(transpose_x = u_43_transpose_x_0, transpose_y = u_43_transpose_y_0, x = Minv_175_cast_fp16, y = vb_55_cast_fp16)[name = string("u_43_cast_fp16")];
            bool var_5876_transpose_x_1 = const()[name = string("op_5876_transpose_x_1"), val = bool(false)];
            bool var_5876_transpose_y_1 = const()[name = string("op_5876_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_5876_cast_fp16 = matmul(transpose_x = var_5876_transpose_x_1, transpose_y = var_5876_transpose_y_1, x = q_93_cast_fp16, y = k_95_cast_fp16)[name = string("op_5876_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_43_cast_fp16 = mul(x = var_5876_cast_fp16, y = D_43_cast_fp16)[name = string("A_43_cast_fp16")];
            tensor<fp16, [16]> var_5878_bias_0_to_fp16 = const()[name = string("op_5878_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_5878_cast_fp16 = linear(bias = var_5878_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_55_cast_fp16)[name = string("op_5878_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_31_cast_fp16 = exp(x = var_5878_cast_fp16)[name = string("cum_31_cast_fp16")];
            tensor<int32, [1]> var_5881_axes_0 = const()[name = string("op_5881_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_5881_cast_fp16 = expand_dims(axes = var_5881_axes_0, x = cum_31_cast_fp16)[name = string("op_5881_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5882_cast_fp16 = mul(x = kb_43_cast_fp16, y = var_5881_cast_fp16)[name = string("op_5882_cast_fp16")];
            bool w_43_transpose_x_0 = const()[name = string("w_43_transpose_x_0"), val = bool(false)];
            bool w_43_transpose_y_0 = const()[name = string("w_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_43_cast_fp16 = matmul(transpose_x = w_43_transpose_x_0, transpose_y = w_43_transpose_y_0, x = Minv_175_cast_fp16, y = var_5882_cast_fp16)[name = string("w_43_cast_fp16")];
            bool var_5884_transpose_x_0 = const()[name = string("op_5884_transpose_x_0"), val = bool(false)];
            bool var_5884_transpose_y_0 = const()[name = string("op_5884_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5884_cast_fp16 = matmul(transpose_x = var_5884_transpose_x_0, transpose_y = var_5884_transpose_y_0, x = w_43_cast_fp16, y = S_31_cast_fp16)[name = string("op_5884_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_31_cast_fp16 = sub(x = u_43_cast_fp16, y = var_5884_cast_fp16)[name = string("vnew_31_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5889_cast_fp16 = mul(x = q_93_cast_fp16, y = var_5881_cast_fp16)[name = string("op_5889_cast_fp16")];
            bool var_5890_transpose_x_0 = const()[name = string("op_5890_transpose_x_0"), val = bool(false)];
            bool var_5890_transpose_y_0 = const()[name = string("op_5890_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5890_cast_fp16 = matmul(transpose_x = var_5890_transpose_x_0, transpose_y = var_5890_transpose_y_0, x = var_5889_cast_fp16, y = S_31_cast_fp16)[name = string("op_5890_cast_fp16")];
            bool var_5891_transpose_x_0 = const()[name = string("op_5891_transpose_x_0"), val = bool(false)];
            bool var_5891_transpose_y_0 = const()[name = string("op_5891_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5891_cast_fp16 = matmul(transpose_x = var_5891_transpose_x_0, transpose_y = var_5891_transpose_y_0, x = A_43_cast_fp16, y = vnew_31_cast_fp16)[name = string("op_5891_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5893_cast_fp16 = add(x = var_5890_cast_fp16, y = var_5891_cast_fp16)[name = string("op_5893_cast_fp16")];
            tensor<fp16, [16]> var_5894_bias_0_to_fp16 = const()[name = string("op_5894_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_5894_cast_fp16 = linear(bias = var_5894_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_55_cast_fp16)[name = string("op_5894_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_33_cast_fp16 = exp(x = var_5894_cast_fp16)[name = string("tail_33_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_65_to_fp16 = const()[name = string("transpose_65_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_5896_bias_0_to_fp16 = const()[name = string("op_5896_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_5896_cast_fp16 = linear(bias = var_5896_bias_0_to_fp16, weight = transpose_65_to_fp16, x = gdec_55_cast_fp16)[name = string("op_5896_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_5897_cast_fp16 = exp(x = var_5896_cast_fp16)[name = string("op_5897_cast_fp16")];
            tensor<int32, [1]> last_23_axes_0 = const()[name = string("last_23_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_23_cast_fp16 = expand_dims(axes = last_23_axes_0, x = var_5897_cast_fp16)[name = string("last_23_cast_fp16")];
            tensor<int32, [1]> var_5901_axes_0 = const()[name = string("op_5901_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_5901_cast_fp16 = expand_dims(axes = var_5901_axes_0, x = tail_33_cast_fp16)[name = string("op_5901_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5902_cast_fp16 = mul(x = k_95_cast_fp16, y = var_5901_cast_fp16)[name = string("op_5902_cast_fp16")];
            bool upd_21_transpose_x_1 = const()[name = string("upd_21_transpose_x_1"), val = bool(true)];
            bool upd_21_transpose_y_1 = const()[name = string("upd_21_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_21_cast_fp16 = matmul(transpose_x = upd_21_transpose_x_1, transpose_y = upd_21_transpose_y_1, x = var_5902_cast_fp16, y = vnew_31_cast_fp16)[name = string("upd_21_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_5907_cast_fp16 = mul(x = S_31_cast_fp16, y = last_23_cast_fp16)[name = string("op_5907_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_33_cast_fp16 = add(x = var_5907_cast_fp16, y = upd_21_cast_fp16)[name = string("S_33_cast_fp16")];
            tensor<int32, [4]> q_95_begin_0 = const()[name = string("q_95_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_95_end_0 = const()[name = string("q_95_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> q_95_end_mask_0 = const()[name = string("q_95_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_95_cast_fp16 = slice_by_index(begin = q_95_begin_0, end = q_95_end_0, end_mask = q_95_end_mask_0, x = q_89_cast_fp16)[name = string("q_95_cast_fp16")];
            tensor<int32, [4]> k_97_begin_0 = const()[name = string("k_97_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_97_end_0 = const()[name = string("k_97_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> k_97_end_mask_0 = const()[name = string("k_97_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_97_cast_fp16 = slice_by_index(begin = k_97_begin_0, end = k_97_end_0, end_mask = k_97_end_mask_0, x = k_91_cast_fp16)[name = string("k_97_cast_fp16")];
            tensor<int32, [4]> kb_45_begin_0 = const()[name = string("kb_45_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_45_end_0 = const()[name = string("kb_45_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> kb_45_end_mask_0 = const()[name = string("kb_45_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_45_cast_fp16 = slice_by_index(begin = kb_45_begin_0, end = kb_45_end_0, end_mask = kb_45_end_mask_0, x = var_5656_cast_fp16)[name = string("kb_45_cast_fp16")];
            tensor<int32, [4]> vb_57_begin_0 = const()[name = string("vb_57_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_57_end_0 = const()[name = string("vb_57_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> vb_57_end_mask_0 = const()[name = string("vb_57_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_57_cast_fp16 = slice_by_index(begin = vb_57_begin_0, end = vb_57_end_0, end_mask = vb_57_end_mask_0, x = vb_51_cast_fp16)[name = string("vb_57_cast_fp16")];
            tensor<int32, [3]> gdec_57_begin_0 = const()[name = string("gdec_57_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_57_end_0 = const()[name = string("gdec_57_end_0"), val = tensor<int32, [3]>([1, 4, 48])];
            tensor<bool, [3]> gdec_57_end_mask_0 = const()[name = string("gdec_57_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_57_cast_fp16 = slice_by_index(begin = gdec_57_begin_0, end = gdec_57_end_0, end_mask = gdec_57_end_mask_0, x = gdec_51_cast_fp16)[name = string("gdec_57_cast_fp16")];
            tensor<int32, [2]> var_5990 = const()[name = string("op_5990"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_5991_cast_fp16 = reshape(shape = var_5990, x = gdec_57_cast_fp16)[name = string("op_5991_cast_fp16")];
            tensor<fp16, [4, 256]> var_5992_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5991_cast_fp16)[name = string("op_5992_cast_fp16")];
            tensor<int32, [4]> var_5997 = const()[name = string("op_5997"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_45_cast_fp16 = reshape(shape = var_5997, x = var_5992_cast_fp16)[name = string("P_45_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5999_cast_fp16 = exp(x = P_45_cast_fp16)[name = string("op_5999_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_45_cast_fp16 = mul(x = var_5999_cast_fp16, y = tril_incl_to_fp16)[name = string("D_45_cast_fp16")];
            bool kk_45_transpose_x_1 = const()[name = string("kk_45_transpose_x_1"), val = bool(false)];
            bool kk_45_transpose_y_1 = const()[name = string("kk_45_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_45_cast_fp16 = matmul(transpose_x = kk_45_transpose_x_1, transpose_y = kk_45_transpose_y_1, x = kb_45_cast_fp16, y = k_97_cast_fp16)[name = string("kk_45_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_6005_cast_fp16 = mul(x = kk_45_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_6005_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6006_cast_fp16 = mul(x = var_6005_cast_fp16, y = D_45_cast_fp16)[name = string("op_6006_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_45_cast_fp16 = mul(x = var_6006_cast_fp16, y = cum_tail_to_fp16)[name = string("N_45_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6008_cast_fp16 = mul(x = D_45_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_6008_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_45_cast_fp16 = mul(x = kk_45_cast_fp16, y = var_6008_cast_fp16)[name = string("Na_45_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_177_cast_fp16 = add(x = eye_to_fp16, y = N_45_cast_fp16)[name = string("Minv_177_cast_fp16")];
            bool Np_133_transpose_x_0 = const()[name = string("Np_133_transpose_x_0"), val = bool(false)];
            bool Np_133_transpose_y_0 = const()[name = string("Np_133_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_133_cast_fp16 = matmul(transpose_x = Np_133_transpose_x_0, transpose_y = Np_133_transpose_y_0, x = N_45_cast_fp16, y = Na_45_cast_fp16)[name = string("Np_133_cast_fp16")];
            bool Npa_89_transpose_x_0 = const()[name = string("Npa_89_transpose_x_0"), val = bool(false)];
            bool Npa_89_transpose_y_0 = const()[name = string("Npa_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_89_cast_fp16 = matmul(transpose_x = Npa_89_transpose_x_0, transpose_y = Npa_89_transpose_y_0, x = Na_45_cast_fp16, y = N_45_cast_fp16)[name = string("Npa_89_cast_fp16")];
            bool var_6014_transpose_x_0 = const()[name = string("op_6014_transpose_x_0"), val = bool(false)];
            bool var_6014_transpose_y_0 = const()[name = string("op_6014_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6014_cast_fp16 = matmul(transpose_x = var_6014_transpose_x_0, transpose_y = var_6014_transpose_y_0, x = Minv_177_cast_fp16, y = Np_133_cast_fp16)[name = string("op_6014_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_179_cast_fp16 = add(x = Minv_177_cast_fp16, y = var_6014_cast_fp16)[name = string("Minv_179_cast_fp16")];
            bool Np_135_transpose_x_0 = const()[name = string("Np_135_transpose_x_0"), val = bool(false)];
            bool Np_135_transpose_y_0 = const()[name = string("Np_135_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_135_cast_fp16 = matmul(transpose_x = Np_135_transpose_x_0, transpose_y = Np_135_transpose_y_0, x = Np_133_cast_fp16, y = Npa_89_cast_fp16)[name = string("Np_135_cast_fp16")];
            bool Npa_91_transpose_x_0 = const()[name = string("Npa_91_transpose_x_0"), val = bool(false)];
            bool Npa_91_transpose_y_0 = const()[name = string("Npa_91_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_91_cast_fp16 = matmul(transpose_x = Npa_91_transpose_x_0, transpose_y = Npa_91_transpose_y_0, x = Npa_89_cast_fp16, y = Np_133_cast_fp16)[name = string("Npa_91_cast_fp16")];
            bool var_6019_transpose_x_0 = const()[name = string("op_6019_transpose_x_0"), val = bool(false)];
            bool var_6019_transpose_y_0 = const()[name = string("op_6019_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6019_cast_fp16 = matmul(transpose_x = var_6019_transpose_x_0, transpose_y = var_6019_transpose_y_0, x = Minv_179_cast_fp16, y = Np_135_cast_fp16)[name = string("op_6019_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_181_cast_fp16 = add(x = Minv_179_cast_fp16, y = var_6019_cast_fp16)[name = string("Minv_181_cast_fp16")];
            bool Np_137_transpose_x_0 = const()[name = string("Np_137_transpose_x_0"), val = bool(false)];
            bool Np_137_transpose_y_0 = const()[name = string("Np_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_137_cast_fp16 = matmul(transpose_x = Np_137_transpose_x_0, transpose_y = Np_137_transpose_y_0, x = Np_135_cast_fp16, y = Npa_91_cast_fp16)[name = string("Np_137_cast_fp16")];
            bool var_6023_transpose_x_0 = const()[name = string("op_6023_transpose_x_0"), val = bool(false)];
            bool var_6023_transpose_y_0 = const()[name = string("op_6023_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6023_cast_fp16 = matmul(transpose_x = var_6023_transpose_x_0, transpose_y = var_6023_transpose_y_0, x = Minv_181_cast_fp16, y = Np_137_cast_fp16)[name = string("op_6023_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_183_cast_fp16 = add(x = Minv_181_cast_fp16, y = var_6023_cast_fp16)[name = string("Minv_183_cast_fp16")];
            bool u_45_transpose_x_0 = const()[name = string("u_45_transpose_x_0"), val = bool(false)];
            bool u_45_transpose_y_0 = const()[name = string("u_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_45_cast_fp16 = matmul(transpose_x = u_45_transpose_x_0, transpose_y = u_45_transpose_y_0, x = Minv_183_cast_fp16, y = vb_57_cast_fp16)[name = string("u_45_cast_fp16")];
            bool var_6030_transpose_x_1 = const()[name = string("op_6030_transpose_x_1"), val = bool(false)];
            bool var_6030_transpose_y_1 = const()[name = string("op_6030_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_6030_cast_fp16 = matmul(transpose_x = var_6030_transpose_x_1, transpose_y = var_6030_transpose_y_1, x = q_95_cast_fp16, y = k_97_cast_fp16)[name = string("op_6030_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_45_cast_fp16 = mul(x = var_6030_cast_fp16, y = D_45_cast_fp16)[name = string("A_45_cast_fp16")];
            tensor<fp16, [16]> var_6032_bias_0_to_fp16 = const()[name = string("op_6032_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_6032_cast_fp16 = linear(bias = var_6032_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_57_cast_fp16)[name = string("op_6032_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_33_cast_fp16 = exp(x = var_6032_cast_fp16)[name = string("cum_33_cast_fp16")];
            tensor<int32, [1]> var_6035_axes_0 = const()[name = string("op_6035_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_6035_cast_fp16 = expand_dims(axes = var_6035_axes_0, x = cum_33_cast_fp16)[name = string("op_6035_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6036_cast_fp16 = mul(x = kb_45_cast_fp16, y = var_6035_cast_fp16)[name = string("op_6036_cast_fp16")];
            bool w_45_transpose_x_0 = const()[name = string("w_45_transpose_x_0"), val = bool(false)];
            bool w_45_transpose_y_0 = const()[name = string("w_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_45_cast_fp16 = matmul(transpose_x = w_45_transpose_x_0, transpose_y = w_45_transpose_y_0, x = Minv_183_cast_fp16, y = var_6036_cast_fp16)[name = string("w_45_cast_fp16")];
            bool var_6038_transpose_x_0 = const()[name = string("op_6038_transpose_x_0"), val = bool(false)];
            bool var_6038_transpose_y_0 = const()[name = string("op_6038_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6038_cast_fp16 = matmul(transpose_x = var_6038_transpose_x_0, transpose_y = var_6038_transpose_y_0, x = w_45_cast_fp16, y = S_33_cast_fp16)[name = string("op_6038_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_33_cast_fp16 = sub(x = u_45_cast_fp16, y = var_6038_cast_fp16)[name = string("vnew_33_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6043_cast_fp16 = mul(x = q_95_cast_fp16, y = var_6035_cast_fp16)[name = string("op_6043_cast_fp16")];
            bool var_6044_transpose_x_0 = const()[name = string("op_6044_transpose_x_0"), val = bool(false)];
            bool var_6044_transpose_y_0 = const()[name = string("op_6044_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6044_cast_fp16 = matmul(transpose_x = var_6044_transpose_x_0, transpose_y = var_6044_transpose_y_0, x = var_6043_cast_fp16, y = S_33_cast_fp16)[name = string("op_6044_cast_fp16")];
            bool var_6045_transpose_x_0 = const()[name = string("op_6045_transpose_x_0"), val = bool(false)];
            bool var_6045_transpose_y_0 = const()[name = string("op_6045_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6045_cast_fp16 = matmul(transpose_x = var_6045_transpose_x_0, transpose_y = var_6045_transpose_y_0, x = A_45_cast_fp16, y = vnew_33_cast_fp16)[name = string("op_6045_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6047_cast_fp16 = add(x = var_6044_cast_fp16, y = var_6045_cast_fp16)[name = string("op_6047_cast_fp16")];
            tensor<fp16, [16]> var_6048_bias_0_to_fp16 = const()[name = string("op_6048_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_6048_cast_fp16 = linear(bias = var_6048_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_57_cast_fp16)[name = string("op_6048_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_cast_fp16 = exp(x = var_6048_cast_fp16)[name = string("tail_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_69_to_fp16 = const()[name = string("transpose_69_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_6050_bias_0_to_fp16 = const()[name = string("op_6050_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_6050_cast_fp16 = linear(bias = var_6050_bias_0_to_fp16, weight = transpose_69_to_fp16, x = gdec_57_cast_fp16)[name = string("op_6050_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_6051_cast_fp16 = exp(x = var_6050_cast_fp16)[name = string("op_6051_cast_fp16")];
            tensor<int32, [1]> last_25_axes_0 = const()[name = string("last_25_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_25_cast_fp16 = expand_dims(axes = last_25_axes_0, x = var_6051_cast_fp16)[name = string("last_25_cast_fp16")];
            tensor<int32, [1]> var_6055_axes_0 = const()[name = string("op_6055_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_6055_cast_fp16 = expand_dims(axes = var_6055_axes_0, x = tail_cast_fp16)[name = string("op_6055_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6056_cast_fp16 = mul(x = k_97_cast_fp16, y = var_6055_cast_fp16)[name = string("op_6056_cast_fp16")];
            bool upd_transpose_x_1 = const()[name = string("upd_transpose_x_1"), val = bool(true)];
            bool upd_transpose_y_1 = const()[name = string("upd_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_cast_fp16 = matmul(transpose_x = upd_transpose_x_1, transpose_y = upd_transpose_y_1, x = var_6056_cast_fp16, y = vnew_33_cast_fp16)[name = string("upd_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_6061_cast_fp16 = mul(x = S_33_cast_fp16, y = last_25_cast_fp16)[name = string("op_6061_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_cast_fp16 = add(x = var_6061_cast_fp16, y = upd_cast_fp16)[name = string("S_cast_fp16")];
            tensor<int32, [4]> q_97_begin_0 = const()[name = string("q_97_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_97_end_0 = const()[name = string("q_97_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_97_end_mask_0 = const()[name = string("q_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, x = q_89_cast_fp16)[name = string("q_97_cast_fp16")];
            tensor<int32, [4]> k_99_begin_0 = const()[name = string("k_99_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_99_end_0 = const()[name = string("k_99_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_99_end_mask_0 = const()[name = string("k_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_99_cast_fp16 = slice_by_index(begin = k_99_begin_0, end = k_99_end_0, end_mask = k_99_end_mask_0, x = k_91_cast_fp16)[name = string("k_99_cast_fp16")];
            tensor<int32, [4]> kb_begin_0 = const()[name = string("kb_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_end_0 = const()[name = string("kb_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_end_mask_0 = const()[name = string("kb_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_cast_fp16 = slice_by_index(begin = kb_begin_0, end = kb_end_0, end_mask = kb_end_mask_0, x = var_5656_cast_fp16)[name = string("kb_cast_fp16")];
            tensor<int32, [4]> vb_begin_0 = const()[name = string("vb_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_end_0 = const()[name = string("vb_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_end_mask_0 = const()[name = string("vb_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_cast_fp16 = slice_by_index(begin = vb_begin_0, end = vb_end_0, end_mask = vb_end_mask_0, x = vb_51_cast_fp16)[name = string("vb_cast_fp16")];
            tensor<int32, [3]> gdec_begin_0 = const()[name = string("gdec_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_end_0 = const()[name = string("gdec_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_end_mask_0 = const()[name = string("gdec_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_cast_fp16 = slice_by_index(begin = gdec_begin_0, end = gdec_end_0, end_mask = gdec_end_mask_0, x = gdec_51_cast_fp16)[name = string("gdec_cast_fp16")];
            tensor<int32, [2]> var_6144 = const()[name = string("op_6144"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_6145_cast_fp16 = reshape(shape = var_6144, x = gdec_cast_fp16)[name = string("op_6145_cast_fp16")];
            tensor<fp16, [4, 256]> var_6146_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_6145_cast_fp16)[name = string("op_6146_cast_fp16")];
            tensor<int32, [4]> var_6151 = const()[name = string("op_6151"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_cast_fp16 = reshape(shape = var_6151, x = var_6146_cast_fp16)[name = string("P_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6153_cast_fp16 = exp(x = P_cast_fp16)[name = string("op_6153_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_cast_fp16 = mul(x = var_6153_cast_fp16, y = tril_incl_to_fp16)[name = string("D_cast_fp16")];
            bool kk_transpose_x_1 = const()[name = string("kk_transpose_x_1"), val = bool(false)];
            bool kk_transpose_y_1 = const()[name = string("kk_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_cast_fp16 = matmul(transpose_x = kk_transpose_x_1, transpose_y = kk_transpose_y_1, x = kb_cast_fp16, y = k_99_cast_fp16)[name = string("kk_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_6159_cast_fp16 = mul(x = kk_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_6159_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6160_cast_fp16 = mul(x = var_6159_cast_fp16, y = D_cast_fp16)[name = string("op_6160_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_cast_fp16 = mul(x = var_6160_cast_fp16, y = cum_tail_to_fp16)[name = string("N_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6162_cast_fp16 = mul(x = D_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_6162_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_cast_fp16 = mul(x = kk_cast_fp16, y = var_6162_cast_fp16)[name = string("Na_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_185_cast_fp16 = add(x = eye_to_fp16, y = N_cast_fp16)[name = string("Minv_185_cast_fp16")];
            bool Np_139_transpose_x_0 = const()[name = string("Np_139_transpose_x_0"), val = bool(false)];
            bool Np_139_transpose_y_0 = const()[name = string("Np_139_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_139_cast_fp16 = matmul(transpose_x = Np_139_transpose_x_0, transpose_y = Np_139_transpose_y_0, x = N_cast_fp16, y = Na_cast_fp16)[name = string("Np_139_cast_fp16")];
            bool Npa_93_transpose_x_0 = const()[name = string("Npa_93_transpose_x_0"), val = bool(false)];
            bool Npa_93_transpose_y_0 = const()[name = string("Npa_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_93_cast_fp16 = matmul(transpose_x = Npa_93_transpose_x_0, transpose_y = Npa_93_transpose_y_0, x = Na_cast_fp16, y = N_cast_fp16)[name = string("Npa_93_cast_fp16")];
            bool var_6168_transpose_x_0 = const()[name = string("op_6168_transpose_x_0"), val = bool(false)];
            bool var_6168_transpose_y_0 = const()[name = string("op_6168_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6168_cast_fp16 = matmul(transpose_x = var_6168_transpose_x_0, transpose_y = var_6168_transpose_y_0, x = Minv_185_cast_fp16, y = Np_139_cast_fp16)[name = string("op_6168_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_187_cast_fp16 = add(x = Minv_185_cast_fp16, y = var_6168_cast_fp16)[name = string("Minv_187_cast_fp16")];
            bool Np_141_transpose_x_0 = const()[name = string("Np_141_transpose_x_0"), val = bool(false)];
            bool Np_141_transpose_y_0 = const()[name = string("Np_141_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_141_cast_fp16 = matmul(transpose_x = Np_141_transpose_x_0, transpose_y = Np_141_transpose_y_0, x = Np_139_cast_fp16, y = Npa_93_cast_fp16)[name = string("Np_141_cast_fp16")];
            bool Npa_transpose_x_0 = const()[name = string("Npa_transpose_x_0"), val = bool(false)];
            bool Npa_transpose_y_0 = const()[name = string("Npa_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_cast_fp16 = matmul(transpose_x = Npa_transpose_x_0, transpose_y = Npa_transpose_y_0, x = Npa_93_cast_fp16, y = Np_139_cast_fp16)[name = string("Npa_cast_fp16")];
            bool var_6173_transpose_x_0 = const()[name = string("op_6173_transpose_x_0"), val = bool(false)];
            bool var_6173_transpose_y_0 = const()[name = string("op_6173_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6173_cast_fp16 = matmul(transpose_x = var_6173_transpose_x_0, transpose_y = var_6173_transpose_y_0, x = Minv_187_cast_fp16, y = Np_141_cast_fp16)[name = string("op_6173_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_189_cast_fp16 = add(x = Minv_187_cast_fp16, y = var_6173_cast_fp16)[name = string("Minv_189_cast_fp16")];
            bool Np_transpose_x_0 = const()[name = string("Np_transpose_x_0"), val = bool(false)];
            bool Np_transpose_y_0 = const()[name = string("Np_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_cast_fp16 = matmul(transpose_x = Np_transpose_x_0, transpose_y = Np_transpose_y_0, x = Np_141_cast_fp16, y = Npa_cast_fp16)[name = string("Np_cast_fp16")];
            bool var_6177_transpose_x_0 = const()[name = string("op_6177_transpose_x_0"), val = bool(false)];
            bool var_6177_transpose_y_0 = const()[name = string("op_6177_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6177_cast_fp16 = matmul(transpose_x = var_6177_transpose_x_0, transpose_y = var_6177_transpose_y_0, x = Minv_189_cast_fp16, y = Np_cast_fp16)[name = string("op_6177_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_cast_fp16 = add(x = Minv_189_cast_fp16, y = var_6177_cast_fp16)[name = string("Minv_cast_fp16")];
            bool u_transpose_x_0 = const()[name = string("u_transpose_x_0"), val = bool(false)];
            bool u_transpose_y_0 = const()[name = string("u_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_cast_fp16 = matmul(transpose_x = u_transpose_x_0, transpose_y = u_transpose_y_0, x = Minv_cast_fp16, y = vb_cast_fp16)[name = string("u_cast_fp16")];
            bool var_6184_transpose_x_1 = const()[name = string("op_6184_transpose_x_1"), val = bool(false)];
            bool var_6184_transpose_y_1 = const()[name = string("op_6184_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_6184_cast_fp16 = matmul(transpose_x = var_6184_transpose_x_1, transpose_y = var_6184_transpose_y_1, x = q_97_cast_fp16, y = k_99_cast_fp16)[name = string("op_6184_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_cast_fp16 = mul(x = var_6184_cast_fp16, y = D_cast_fp16)[name = string("A_cast_fp16")];
            tensor<fp16, [16]> var_6186_bias_0_to_fp16 = const()[name = string("op_6186_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_6186_cast_fp16 = linear(bias = var_6186_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_cast_fp16)[name = string("op_6186_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_cast_fp16 = exp(x = var_6186_cast_fp16)[name = string("cum_cast_fp16")];
            tensor<int32, [1]> var_6189_axes_0 = const()[name = string("op_6189_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_6189_cast_fp16 = expand_dims(axes = var_6189_axes_0, x = cum_cast_fp16)[name = string("op_6189_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6190_cast_fp16 = mul(x = kb_cast_fp16, y = var_6189_cast_fp16)[name = string("op_6190_cast_fp16")];
            bool w_transpose_x_0 = const()[name = string("w_transpose_x_0"), val = bool(false)];
            bool w_transpose_y_0 = const()[name = string("w_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_cast_fp16 = matmul(transpose_x = w_transpose_x_0, transpose_y = w_transpose_y_0, x = Minv_cast_fp16, y = var_6190_cast_fp16)[name = string("w_cast_fp16")];
            bool var_6192_transpose_x_0 = const()[name = string("op_6192_transpose_x_0"), val = bool(false)];
            bool var_6192_transpose_y_0 = const()[name = string("op_6192_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6192_cast_fp16 = matmul(transpose_x = var_6192_transpose_x_0, transpose_y = var_6192_transpose_y_0, x = w_cast_fp16, y = S_cast_fp16)[name = string("op_6192_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_cast_fp16 = sub(x = u_cast_fp16, y = var_6192_cast_fp16)[name = string("vnew_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6197_cast_fp16 = mul(x = q_97_cast_fp16, y = var_6189_cast_fp16)[name = string("op_6197_cast_fp16")];
            bool var_6198_transpose_x_0 = const()[name = string("op_6198_transpose_x_0"), val = bool(false)];
            bool var_6198_transpose_y_0 = const()[name = string("op_6198_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6198_cast_fp16 = matmul(transpose_x = var_6198_transpose_x_0, transpose_y = var_6198_transpose_y_0, x = var_6197_cast_fp16, y = S_cast_fp16)[name = string("op_6198_cast_fp16")];
            bool var_6199_transpose_x_0 = const()[name = string("op_6199_transpose_x_0"), val = bool(false)];
            bool var_6199_transpose_y_0 = const()[name = string("op_6199_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6199_cast_fp16 = matmul(transpose_x = var_6199_transpose_x_0, transpose_y = var_6199_transpose_y_0, x = A_cast_fp16, y = vnew_cast_fp16)[name = string("op_6199_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_cast_fp16 = add(x = var_6198_cast_fp16, y = var_6199_cast_fp16)[name = string("oc_cast_fp16")];
            int32 var_6203 = const()[name = string("op_6203"), val = int32(2)];
            bool o_33_interleave_0 = const()[name = string("o_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 128]> o_33_cast_fp16 = concat(axis = var_6203, interleave = o_33_interleave_0, values = (var_5746_cast_fp16, var_5893_cast_fp16, var_6047_cast_fp16, oc_cast_fp16))[name = string("o_33_cast_fp16")];
            tensor<int32, [4]> var_6209 = const()[name = string("op_6209"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_6215 = const()[name = string("op_6215"), val = tensor<int32, [4]>([1, 64, 4, 128])];
            tensor<fp16, [1, 64, 4, 128]> input_97_cast_fp16 = reshape(shape = var_6215, x = z_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> o_35_cast_fp16 = transpose(perm = var_6209, x = o_33_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 64, 4, 128]> var_6217_cast_fp16 = mul(x = o_35_cast_fp16, y = o_35_cast_fp16)[name = string("op_6217_cast_fp16")];
            tensor<int32, [1]> var_6222_axes_0 = const()[name = string("op_6222_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6222_keep_dims_0 = const()[name = string("op_6222_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 4, 1]> var_6222_cast_fp16 = reduce_mean(axes = var_6222_axes_0, keep_dims = var_6222_keep_dims_0, x = var_6217_cast_fp16)[name = string("op_6222_cast_fp16")];
            fp16 var_6224_to_fp16 = const()[name = string("op_6224_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 64, 4, 1]> var_6225_cast_fp16 = add(x = var_6222_cast_fp16, y = var_6224_to_fp16)[name = string("op_6225_cast_fp16")];
            fp32 var_6226_epsilon_0 = const()[name = string("op_6226_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 4, 1]> var_6226_cast_fp16 = rsqrt(epsilon = var_6226_epsilon_0, x = var_6225_cast_fp16)[name = string("op_6226_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> var_6227_cast_fp16 = mul(x = o_35_cast_fp16, y = var_6226_cast_fp16)[name = string("op_6227_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> var_6229_cast_fp16 = silu(x = input_97_cast_fp16)[name = string("op_6229_cast_fp16")];
            tensor<fp16, [1, 64, 4, 128]> o_37_cast_fp16 = mul(x = var_6227_cast_fp16, y = var_6229_cast_fp16)[name = string("o_37_cast_fp16")];
            tensor<int32, [3]> var_6234 = const()[name = string("op_6234"), val = tensor<int32, [3]>([1, 64, 512])];
            tensor<fp16, [1, 64, 512]> input_99_cast_fp16 = reshape(shape = var_6234, x = o_37_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [640, 512]> layers_6_out_weight_to_fp16 = const()[name = string("layers_6_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97009984)))];
            tensor<fp16, [1, 64, 640]> linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_out_weight_to_fp16, x = input_99_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 64, 640]> x_65_cast_fp16 = add(x = x_61_cast_fp16, y = linear_25_cast_fp16)[name = string("x_65_cast_fp16")];
            fp16 var_6247_to_fp16 = const()[name = string("op_6247_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [1, 64, 640]> x_67_cast_fp16 = mul(x = x_65_cast_fp16, y = var_6247_to_fp16)[name = string("x_67_cast_fp16")];
            fp16 var_6245_promoted_to_fp16 = const()[name = string("op_6245_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> var_6249_cast_fp16 = pow(x = x_67_cast_fp16, y = var_6245_promoted_to_fp16)[name = string("op_6249_cast_fp16")];
            tensor<int32, [1]> var_6251_axes_0 = const()[name = string("op_6251_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6251_keep_dims_0 = const()[name = string("op_6251_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6251_cast_fp16 = reduce_mean(axes = var_6251_axes_0, keep_dims = var_6251_keep_dims_0, x = var_6249_cast_fp16)[name = string("op_6251_cast_fp16")];
            fp16 var_6252_to_fp16 = const()[name = string("op_6252_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 64, 1]> var_6253_cast_fp16 = add(x = var_6251_cast_fp16, y = var_6252_to_fp16)[name = string("op_6253_cast_fp16")];
            fp32 var_6254_epsilon_0 = const()[name = string("op_6254_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6254_cast_fp16 = rsqrt(epsilon = var_6254_epsilon_0, x = var_6253_cast_fp16)[name = string("op_6254_cast_fp16")];
            tensor<fp16, [1, 64, 640]> var_6255_cast_fp16 = mul(x = x_67_cast_fp16, y = var_6254_cast_fp16)[name = string("op_6255_cast_fp16")];
            tensor<fp16, [640]> layers_6_n2_w1_to_fp16 = const()[name = string("layers_6_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97665408)))];
            tensor<fp16, [1, 64, 640]> input_101_cast_fp16 = mul(x = var_6255_cast_fp16, y = layers_6_n2_w1_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_6_gate_up_weight_to_fp16 = const()[name = string("layers_6_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97666752)))];
            tensor<fp16, [1, 64, 5120]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_6_gate_up_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<int32, [3]> input_103_begin_0 = const()[name = string("input_103_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_103_end_0 = const()[name = string("input_103_end_0"), val = tensor<int32, [3]>([1, 64, 2560])];
            tensor<bool, [3]> input_103_end_mask_0 = const()[name = string("input_103_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 2560]> input_103_cast_fp16 = slice_by_index(begin = input_103_begin_0, end = input_103_end_0, end_mask = input_103_end_mask_0, x = linear_26_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [1, 64, 2560]> var_6265_cast_fp16 = silu(x = input_103_cast_fp16)[name = string("op_6265_cast_fp16")];
            tensor<int32, [3]> var_6270_begin_0 = const()[name = string("op_6270_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_6270_end_0 = const()[name = string("op_6270_end_0"), val = tensor<int32, [3]>([1, 64, 5120])];
            tensor<bool, [3]> var_6270_end_mask_0 = const()[name = string("op_6270_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 2560]> var_6270_cast_fp16 = slice_by_index(begin = var_6270_begin_0, end = var_6270_end_0, end_mask = var_6270_end_mask_0, x = linear_26_cast_fp16)[name = string("op_6270_cast_fp16")];
            tensor<fp16, [1, 64, 2560]> input_105_cast_fp16 = mul(x = var_6265_cast_fp16, y = var_6270_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_6_down_weight_to_fp16 = const()[name = string("layers_6_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104220416)))];
            tensor<fp16, [1, 64, 640]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_down_weight_to_fp16, x = input_105_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 64, 640]> x_69_cast_fp16 = add(x = x_65_cast_fp16, y = linear_27_cast_fp16)[name = string("x_69_cast_fp16")];
            fp16 var_6283_to_fp16 = const()[name = string("op_6283_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [1, 64, 640]> x_71_cast_fp16 = mul(x = x_69_cast_fp16, y = var_6283_to_fp16)[name = string("x_71_cast_fp16")];
            fp16 var_6281_promoted_to_fp16 = const()[name = string("op_6281_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> var_6285_cast_fp16 = pow(x = x_71_cast_fp16, y = var_6281_promoted_to_fp16)[name = string("op_6285_cast_fp16")];
            tensor<int32, [1]> var_6287_axes_0 = const()[name = string("op_6287_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6287_keep_dims_0 = const()[name = string("op_6287_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6287_cast_fp16 = reduce_mean(axes = var_6287_axes_0, keep_dims = var_6287_keep_dims_0, x = var_6285_cast_fp16)[name = string("op_6287_cast_fp16")];
            fp16 var_6288_to_fp16 = const()[name = string("op_6288_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 64, 1]> var_6289_cast_fp16 = add(x = var_6287_cast_fp16, y = var_6288_to_fp16)[name = string("op_6289_cast_fp16")];
            fp32 var_6290_epsilon_0 = const()[name = string("op_6290_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6290_cast_fp16 = rsqrt(epsilon = var_6290_epsilon_0, x = var_6289_cast_fp16)[name = string("op_6290_cast_fp16")];
            tensor<fp16, [1, 64, 640]> var_6291_cast_fp16 = mul(x = x_71_cast_fp16, y = var_6290_cast_fp16)[name = string("op_6291_cast_fp16")];
            tensor<fp16, [640]> layers_7_n1_w1_to_fp16 = const()[name = string("layers_7_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107497280)))];
            tensor<fp16, [1, 64, 640]> input_107_cast_fp16 = mul(x = var_6291_cast_fp16, y = layers_7_n1_w1_to_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [1280, 640]> layers_7_qkv_weight_to_fp16 = const()[name = string("layers_7_qkv_weight_to_fp16"), val = tensor<fp16, [1280, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107498624)))];
            tensor<fp16, [1, 64, 1280]> linear_28_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = layers_7_qkv_weight_to_fp16, x = input_107_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [3]> var_6300_begin_0 = const()[name = string("op_6300_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_6300_end_0 = const()[name = string("op_6300_end_0"), val = tensor<int32, [3]>([1, 64, 1024])];
            tensor<bool, [3]> var_6300_end_mask_0 = const()[name = string("op_6300_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 1024]> var_6300_cast_fp16 = slice_by_index(begin = var_6300_begin_0, end = var_6300_end_0, end_mask = var_6300_end_mask_0, x = linear_28_cast_fp16)[name = string("op_6300_cast_fp16")];
            tensor<int32, [4]> var_6305 = const()[name = string("op_6305"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<fp16, [1, 64, 8, 128]> qg_cast_fp16 = reshape(shape = var_6305, x = var_6300_cast_fp16)[name = string("qg_cast_fp16")];
            tensor<int32, [4]> x_73_begin_0 = const()[name = string("x_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_73_end_0 = const()[name = string("x_73_end_0"), val = tensor<int32, [4]>([1, 64, 8, 64])];
            tensor<bool, [4]> x_73_end_mask_0 = const()[name = string("x_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 64, 8, 64]> x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, x = qg_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<int32, [4]> var_6316_begin_0 = const()[name = string("op_6316_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> var_6316_end_0 = const()[name = string("op_6316_end_0"), val = tensor<int32, [4]>([1, 64, 8, 128])];
            tensor<bool, [4]> var_6316_end_mask_0 = const()[name = string("op_6316_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 64, 8, 64]> var_6316_cast_fp16 = slice_by_index(begin = var_6316_begin_0, end = var_6316_end_0, end_mask = var_6316_end_mask_0, x = qg_cast_fp16)[name = string("op_6316_cast_fp16")];
            tensor<int32, [3]> var_6320 = const()[name = string("op_6320"), val = tensor<int32, [3]>([1, 64, 512])];
            tensor<fp16, [1, 64, 512]> gate_cast_fp16 = reshape(shape = var_6320, x = var_6316_cast_fp16)[name = string("gate_cast_fp16")];
            tensor<int32, [3]> var_6326_begin_0 = const()[name = string("op_6326_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> var_6326_end_0 = const()[name = string("op_6326_end_0"), val = tensor<int32, [3]>([1, 64, 1152])];
            tensor<bool, [3]> var_6326_end_mask_0 = const()[name = string("op_6326_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 128]> var_6326_cast_fp16 = slice_by_index(begin = var_6326_begin_0, end = var_6326_end_0, end_mask = var_6326_end_mask_0, x = linear_28_cast_fp16)[name = string("op_6326_cast_fp16")];
            tensor<int32, [4]> var_6331 = const()[name = string("op_6331"), val = tensor<int32, [4]>([1, 64, 2, 64])];
            tensor<fp16, [1, 64, 2, 64]> x_79_cast_fp16 = reshape(shape = var_6331, x = var_6326_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<int32, [3]> var_6337_begin_0 = const()[name = string("op_6337_begin_0"), val = tensor<int32, [3]>([0, 0, 1152])];
            tensor<int32, [3]> var_6337_end_0 = const()[name = string("op_6337_end_0"), val = tensor<int32, [3]>([1, 64, 1280])];
            tensor<bool, [3]> var_6337_end_mask_0 = const()[name = string("op_6337_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 128]> var_6337_cast_fp16 = slice_by_index(begin = var_6337_begin_0, end = var_6337_end_0, end_mask = var_6337_end_mask_0, x = linear_28_cast_fp16)[name = string("op_6337_cast_fp16")];
            tensor<int32, [4]> var_6342 = const()[name = string("op_6342"), val = tensor<int32, [4]>([1, 64, 2, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_6343_cast_fp16 = reshape(shape = var_6342, x = var_6337_cast_fp16)[name = string("op_6343_cast_fp16")];
            tensor<int32, [4]> var_6348 = const()[name = string("op_6348"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6354_promoted_to_fp16 = const()[name = string("op_6354_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 8, 64]> var_6358_cast_fp16 = pow(x = x_73_cast_fp16, y = var_6354_promoted_to_fp16)[name = string("op_6358_cast_fp16")];
            tensor<int32, [1]> var_6360_axes_0 = const()[name = string("op_6360_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6360_keep_dims_0 = const()[name = string("op_6360_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 8, 1]> var_6360_cast_fp16 = reduce_mean(axes = var_6360_axes_0, keep_dims = var_6360_keep_dims_0, x = var_6358_cast_fp16)[name = string("op_6360_cast_fp16")];
            fp16 var_6361_to_fp16 = const()[name = string("op_6361_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 8, 1]> var_6362_cast_fp16 = add(x = var_6360_cast_fp16, y = var_6361_to_fp16)[name = string("op_6362_cast_fp16")];
            fp32 var_6363_epsilon_0 = const()[name = string("op_6363_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 8, 1]> var_6363_cast_fp16 = rsqrt(epsilon = var_6363_epsilon_0, x = var_6362_cast_fp16)[name = string("op_6363_cast_fp16")];
            tensor<fp16, [1, 64, 8, 64]> var_6364_cast_fp16 = mul(x = x_73_cast_fp16, y = var_6363_cast_fp16)[name = string("op_6364_cast_fp16")];
            tensor<fp16, [64]> layers_7_qn_w1_to_fp16 = const()[name = string("layers_7_qn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137088)))];
            tensor<fp16, [1, 64, 8, 64]> var_6365_cast_fp16 = mul(x = var_6364_cast_fp16, y = layers_7_qn_w1_to_fp16)[name = string("op_6365_cast_fp16")];
            tensor<int32, [4]> var_6370 = const()[name = string("op_6370"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_5_begin_0 = const()[name = string("r_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_5_end_0 = const()[name = string("r_5_end_0"), val = tensor<int32, [4]>([1, 8, 64, 16])];
            tensor<bool, [4]> r_5_end_mask_0 = const()[name = string("r_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 64, 64]> x_77_cast_fp16 = transpose(perm = var_6370, x = var_6365_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 8, 64, 16]> r_5_cast_fp16 = slice_by_index(begin = r_5_begin_0, end = r_5_end_0, end_mask = r_5_end_mask_0, x = x_77_cast_fp16)[name = string("r_5_cast_fp16")];
            tensor<int32, [4]> p_5_begin_0 = const()[name = string("p_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_5_end_0 = const()[name = string("p_5_end_0"), val = tensor<int32, [4]>([1, 8, 64, 64])];
            tensor<bool, [4]> p_5_end_mask_0 = const()[name = string("p_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 64, 48]> p_5_cast_fp16 = slice_by_index(begin = p_5_begin_0, end = p_5_end_0, end_mask = p_5_end_mask_0, x = x_77_cast_fp16)[name = string("p_5_cast_fp16")];
            tensor<int32, [4]> var_6386_begin_0 = const()[name = string("op_6386_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_6386_end_0 = const()[name = string("op_6386_end_0"), val = tensor<int32, [4]>([1, 8, 64, 16])];
            tensor<bool, [4]> var_6386_end_mask_0 = const()[name = string("op_6386_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 64, 8]> var_6386_cast_fp16 = slice_by_index(begin = var_6386_begin_0, end = var_6386_end_0, end_mask = var_6386_end_mask_0, x = r_5_cast_fp16)[name = string("op_6386_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 64, 8]> var_6387_cast_fp16 = mul(x = var_6386_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_6387_cast_fp16")];
            tensor<int32, [4]> var_6392_begin_0 = const()[name = string("op_6392_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6392_end_0 = const()[name = string("op_6392_end_0"), val = tensor<int32, [4]>([1, 8, 64, 8])];
            tensor<bool, [4]> var_6392_end_mask_0 = const()[name = string("op_6392_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 64, 8]> var_6392_cast_fp16 = slice_by_index(begin = var_6392_begin_0, end = var_6392_end_0, end_mask = var_6392_end_mask_0, x = r_5_cast_fp16)[name = string("op_6392_cast_fp16")];
            int32 var_6394 = const()[name = string("op_6394"), val = int32(-1)];
            bool rh_5_interleave_0 = const()[name = string("rh_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 16]> rh_5_cast_fp16 = concat(axis = var_6394, interleave = rh_5_interleave_0, values = (var_6387_cast_fp16, var_6392_cast_fp16))[name = string("rh_5_cast_fp16")];
            tensor<fp16, [1, 8, 64, 16]> var_6396_cast_fp16 = mul(x = r_5_cast_fp16, y = cos_to_fp16)[name = string("op_6396_cast_fp16")];
            tensor<fp16, [1, 8, 64, 16]> var_6397_cast_fp16 = mul(x = rh_5_cast_fp16, y = sin_to_fp16)[name = string("op_6397_cast_fp16")];
            tensor<fp16, [1, 8, 64, 16]> var_6399_cast_fp16 = add(x = var_6396_cast_fp16, y = var_6397_cast_fp16)[name = string("op_6399_cast_fp16")];
            int32 var_6401 = const()[name = string("op_6401"), val = int32(-1)];
            bool q_interleave_0 = const()[name = string("q_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 64]> q_cast_fp16 = concat(axis = var_6401, interleave = q_interleave_0, values = (var_6399_cast_fp16, p_5_cast_fp16))[name = string("q_cast_fp16")];
            fp16 var_6407_promoted_to_fp16 = const()[name = string("op_6407_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 2, 64]> var_6411_cast_fp16 = pow(x = x_79_cast_fp16, y = var_6407_promoted_to_fp16)[name = string("op_6411_cast_fp16")];
            tensor<int32, [1]> var_6413_axes_0 = const()[name = string("op_6413_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6413_keep_dims_0 = const()[name = string("op_6413_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 2, 1]> var_6413_cast_fp16 = reduce_mean(axes = var_6413_axes_0, keep_dims = var_6413_keep_dims_0, x = var_6411_cast_fp16)[name = string("op_6413_cast_fp16")];
            fp16 var_6414_to_fp16 = const()[name = string("op_6414_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 2, 1]> var_6415_cast_fp16 = add(x = var_6413_cast_fp16, y = var_6414_to_fp16)[name = string("op_6415_cast_fp16")];
            fp32 var_6416_epsilon_0 = const()[name = string("op_6416_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 2, 1]> var_6416_cast_fp16 = rsqrt(epsilon = var_6416_epsilon_0, x = var_6415_cast_fp16)[name = string("op_6416_cast_fp16")];
            tensor<fp16, [1, 64, 2, 64]> var_6417_cast_fp16 = mul(x = x_79_cast_fp16, y = var_6416_cast_fp16)[name = string("op_6417_cast_fp16")];
            tensor<fp16, [64]> layers_7_kn_w1_to_fp16 = const()[name = string("layers_7_kn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137280)))];
            tensor<fp16, [1, 64, 2, 64]> var_6418_cast_fp16 = mul(x = var_6417_cast_fp16, y = layers_7_kn_w1_to_fp16)[name = string("op_6418_cast_fp16")];
            tensor<int32, [4]> var_6423 = const()[name = string("op_6423"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_begin_0 = const()[name = string("r_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_end_0 = const()[name = string("r_end_0"), val = tensor<int32, [4]>([1, 2, 64, 16])];
            tensor<bool, [4]> r_end_mask_0 = const()[name = string("r_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 64]> x_83_cast_fp16 = transpose(perm = var_6423, x = var_6418_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 2, 64, 16]> r_cast_fp16 = slice_by_index(begin = r_begin_0, end = r_end_0, end_mask = r_end_mask_0, x = x_83_cast_fp16)[name = string("r_cast_fp16")];
            tensor<int32, [4]> p_begin_0 = const()[name = string("p_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_end_0 = const()[name = string("p_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> p_end_mask_0 = const()[name = string("p_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 48]> p_cast_fp16 = slice_by_index(begin = p_begin_0, end = p_end_0, end_mask = p_end_mask_0, x = x_83_cast_fp16)[name = string("p_cast_fp16")];
            tensor<int32, [4]> var_6439_begin_0 = const()[name = string("op_6439_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_6439_end_0 = const()[name = string("op_6439_end_0"), val = tensor<int32, [4]>([1, 2, 64, 16])];
            tensor<bool, [4]> var_6439_end_mask_0 = const()[name = string("op_6439_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 8]> var_6439_cast_fp16 = slice_by_index(begin = var_6439_begin_0, end = var_6439_end_0, end_mask = var_6439_end_mask_0, x = r_cast_fp16)[name = string("op_6439_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 8]> var_6440_cast_fp16 = mul(x = var_6439_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_6440_cast_fp16")];
            tensor<int32, [4]> var_6445_begin_0 = const()[name = string("op_6445_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6445_end_0 = const()[name = string("op_6445_end_0"), val = tensor<int32, [4]>([1, 2, 64, 8])];
            tensor<bool, [4]> var_6445_end_mask_0 = const()[name = string("op_6445_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 8]> var_6445_cast_fp16 = slice_by_index(begin = var_6445_begin_0, end = var_6445_end_0, end_mask = var_6445_end_mask_0, x = r_cast_fp16)[name = string("op_6445_cast_fp16")];
            int32 var_6447 = const()[name = string("op_6447"), val = int32(-1)];
            bool rh_interleave_0 = const()[name = string("rh_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 16]> rh_cast_fp16 = concat(axis = var_6447, interleave = rh_interleave_0, values = (var_6440_cast_fp16, var_6445_cast_fp16))[name = string("rh_cast_fp16")];
            tensor<fp16, [1, 2, 64, 16]> var_6449_cast_fp16 = mul(x = r_cast_fp16, y = cos_to_fp16)[name = string("op_6449_cast_fp16")];
            tensor<fp16, [1, 2, 64, 16]> var_6450_cast_fp16 = mul(x = rh_cast_fp16, y = sin_to_fp16)[name = string("op_6450_cast_fp16")];
            tensor<fp16, [1, 2, 64, 16]> var_6452_cast_fp16 = add(x = var_6449_cast_fp16, y = var_6450_cast_fp16)[name = string("op_6452_cast_fp16")];
            int32 var_6454 = const()[name = string("op_6454"), val = int32(-1)];
            bool k_101_interleave_0 = const()[name = string("k_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> k_101_cast_fp16 = concat(axis = var_6454, interleave = k_101_interleave_0, values = (var_6452_cast_fp16, p_cast_fp16))[name = string("k_101_cast_fp16")];
            tensor<int32, [1]> var_6467_axes_0 = const()[name = string("op_6467_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_6467_cast_fp16 = expand_dims(axes = var_6467_axes_0, x = k_101_cast_fp16)[name = string("op_6467_cast_fp16")];
            tensor<int32, [5]> var_6475_reps_0 = const()[name = string("op_6475_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 2, 4, 64, 64]> var_6475_cast_fp16 = tile(reps = var_6475_reps_0, x = var_6467_cast_fp16)[name = string("op_6475_cast_fp16")];
            tensor<int32, [4]> var_6480 = const()[name = string("op_6480"), val = tensor<int32, [4]>([1, 8, 64, 64])];
            tensor<fp16, [1, 8, 64, 64]> k_cast_fp16 = reshape(shape = var_6480, x = var_6475_cast_fp16)[name = string("k_cast_fp16")];
            tensor<int32, [1]> var_6493_axes_0 = const()[name = string("op_6493_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> v_17_cast_fp16 = transpose(perm = var_6348, x = var_6343_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_6493_cast_fp16 = expand_dims(axes = var_6493_axes_0, x = v_17_cast_fp16)[name = string("op_6493_cast_fp16")];
            tensor<int32, [5]> var_6501_reps_0 = const()[name = string("op_6501_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 2, 4, 64, 64]> var_6501_cast_fp16 = tile(reps = var_6501_reps_0, x = var_6493_cast_fp16)[name = string("op_6501_cast_fp16")];
            tensor<int32, [4]> var_6506 = const()[name = string("op_6506"), val = tensor<int32, [4]>([1, 8, 64, 64])];
            tensor<fp16, [1, 8, 64, 64]> v_cast_fp16 = reshape(shape = var_6506, x = var_6501_cast_fp16)[name = string("v_cast_fp16")];
            bool var_6511_transpose_x_1 = const()[name = string("op_6511_transpose_x_1"), val = bool(false)];
            bool var_6511_transpose_y_1 = const()[name = string("op_6511_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 64, 64]> var_6511_cast_fp16 = matmul(transpose_x = var_6511_transpose_x_1, transpose_y = var_6511_transpose_y_1, x = q_cast_fp16, y = k_cast_fp16)[name = string("op_6511_cast_fp16")];
            fp16 var_6512_to_fp16 = const()[name = string("op_6512_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 8, 64, 64]> var_6513_cast_fp16 = mul(x = var_6511_cast_fp16, y = var_6512_to_fp16)[name = string("op_6513_cast_fp16")];
            tensor<fp16, [1, 8, 64, 64]> s_cast_fp16 = add(x = var_6513_cast_fp16, y = causal_add_to_fp16)[name = string("s_cast_fp16")];
            int32 var_6516 = const()[name = string("op_6516"), val = int32(-1)];
            tensor<fp16, [1, 8, 64, 64]> var_6518_cast_fp16 = softmax(axis = var_6516, x = s_cast_fp16)[name = string("op_6518_cast_fp16")];
            bool o_39_transpose_x_0 = const()[name = string("o_39_transpose_x_0"), val = bool(false)];
            bool o_39_transpose_y_0 = const()[name = string("o_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 64, 64]> o_39_cast_fp16 = matmul(transpose_x = o_39_transpose_x_0, transpose_y = o_39_transpose_y_0, x = var_6518_cast_fp16, y = v_cast_fp16)[name = string("o_39_cast_fp16")];
            tensor<int32, [4]> var_6524 = const()[name = string("op_6524"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6529 = const()[name = string("op_6529"), val = tensor<int32, [3]>([1, 64, 512])];
            tensor<fp16, [1, 64, 8, 64]> var_6525_cast_fp16 = transpose(perm = var_6524, x = o_39_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 64, 512]> var_6530_cast_fp16 = reshape(shape = var_6529, x = var_6525_cast_fp16)[name = string("op_6530_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6531_cast_fp16 = sigmoid(x = gate_cast_fp16)[name = string("op_6531_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_109_cast_fp16 = mul(x = var_6530_cast_fp16, y = var_6531_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [640, 512]> layers_7_o_weight_to_fp16 = const()[name = string("layers_7_o_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137472)))];
            tensor<fp16, [1, 64, 640]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_o_weight_to_fp16, x = input_109_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<fp16, [1, 64, 640]> x_85_cast_fp16 = add(x = x_69_cast_fp16, y = linear_29_cast_fp16)[name = string("x_85_cast_fp16")];
            fp16 var_6544_to_fp16 = const()[name = string("op_6544_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [1, 64, 640]> x_87_cast_fp16 = mul(x = x_85_cast_fp16, y = var_6544_to_fp16)[name = string("x_87_cast_fp16")];
            fp16 var_6542_promoted_to_fp16 = const()[name = string("op_6542_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 640]> var_6546_cast_fp16 = pow(x = x_87_cast_fp16, y = var_6542_promoted_to_fp16)[name = string("op_6546_cast_fp16")];
            tensor<int32, [1]> var_6548_axes_0 = const()[name = string("op_6548_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6548_keep_dims_0 = const()[name = string("op_6548_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6548_cast_fp16 = reduce_mean(axes = var_6548_axes_0, keep_dims = var_6548_keep_dims_0, x = var_6546_cast_fp16)[name = string("op_6548_cast_fp16")];
            fp16 var_6549_to_fp16 = const()[name = string("op_6549_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 64, 1]> var_6550_cast_fp16 = add(x = var_6548_cast_fp16, y = var_6549_to_fp16)[name = string("op_6550_cast_fp16")];
            fp32 var_6551_epsilon_0 = const()[name = string("op_6551_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6551_cast_fp16 = rsqrt(epsilon = var_6551_epsilon_0, x = var_6550_cast_fp16)[name = string("op_6551_cast_fp16")];
            tensor<fp16, [1, 64, 640]> var_6552_cast_fp16 = mul(x = x_87_cast_fp16, y = var_6551_cast_fp16)[name = string("op_6552_cast_fp16")];
            tensor<fp16, [640]> layers_7_n2_w1_to_fp16 = const()[name = string("layers_7_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109792896)))];
            tensor<fp16, [1, 64, 640]> input_111_cast_fp16 = mul(x = var_6552_cast_fp16, y = layers_7_n2_w1_to_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_7_gate_up_weight_to_fp16 = const()[name = string("layers_7_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109794240)))];
            tensor<fp16, [1, 64, 5120]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_7_gate_up_weight_to_fp16, x = input_111_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [3]> input_113_begin_0 = const()[name = string("input_113_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_113_end_0 = const()[name = string("input_113_end_0"), val = tensor<int32, [3]>([1, 64, 2560])];
            tensor<bool, [3]> input_113_end_mask_0 = const()[name = string("input_113_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 64, 2560]> input_113_cast_fp16 = slice_by_index(begin = input_113_begin_0, end = input_113_end_0, end_mask = input_113_end_mask_0, x = linear_30_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [1, 64, 2560]> var_6562_cast_fp16 = silu(x = input_113_cast_fp16)[name = string("op_6562_cast_fp16")];
            tensor<int32, [3]> var_6567_begin_0 = const()[name = string("op_6567_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_6567_end_0 = const()[name = string("op_6567_end_0"), val = tensor<int32, [3]>([1, 64, 5120])];
            tensor<bool, [3]> var_6567_end_mask_0 = const()[name = string("op_6567_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 64, 2560]> var_6567_cast_fp16 = slice_by_index(begin = var_6567_begin_0, end = var_6567_end_0, end_mask = var_6567_end_mask_0, x = linear_30_cast_fp16)[name = string("op_6567_cast_fp16")];
            tensor<fp16, [1, 64, 2560]> input_cast_fp16 = mul(x = var_6562_cast_fp16, y = var_6567_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_7_down_weight_to_fp16 = const()[name = string("layers_7_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116347904)))];
            tensor<fp16, [1, 64, 640]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_down_weight_to_fp16, x = input_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 64, 640]> x_89_cast_fp16 = add(x = x_85_cast_fp16, y = linear_31_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<int32, [1]> var_6585_axes_0 = const()[name = string("op_6585_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_6585 = expand_dims(axes = var_6585_axes_0, x = last)[name = string("op_6585")];
            tensor<bool, [1, 64]> var_6586 = equal(x = pos, y = var_6585)[name = string("op_6586")];
            tensor<int32, [1]> var_6603_axes_0 = const()[name = string("op_6603_axes_0"), val = tensor<int32, [1]>([2])];
            string sel_to_fp16_dtype_0 = const()[name = string("sel_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 64]> var_6586_to_fp16 = cast(dtype = sel_to_fp16_dtype_0, x = var_6586)[name = string("cast_107")];
            tensor<fp16, [1, 64, 1]> var_6603_cast_fp16 = expand_dims(axes = var_6603_axes_0, x = var_6586_to_fp16)[name = string("op_6603_cast_fp16")];
            tensor<fp16, [1, 64, 640]> var_6604_cast_fp16 = mul(x = var_6603_cast_fp16, y = x_89_cast_fp16)[name = string("op_6604_cast_fp16")];
            tensor<int32, [1]> x_91_axes_0 = const()[name = string("x_91_axes_0"), val = tensor<int32, [1]>([1])];
            bool x_91_keep_dims_0 = const()[name = string("x_91_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 640]> x_91_cast_fp16 = reduce_sum(axes = x_91_axes_0, keep_dims = x_91_keep_dims_0, x = var_6604_cast_fp16)[name = string("x_91_cast_fp16")];
            fp16 var_6616_to_fp16 = const()[name = string("op_6616_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 640]> x_cast_fp16 = mul(x = x_91_cast_fp16, y = var_6616_to_fp16)[name = string("x_cast_fp16")];
            fp16 var_6614_promoted_to_fp16 = const()[name = string("op_6614_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640]> var_6618_cast_fp16 = pow(x = x_cast_fp16, y = var_6614_promoted_to_fp16)[name = string("op_6618_cast_fp16")];
            tensor<int32, [1]> var_6620_axes_0 = const()[name = string("op_6620_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6620_keep_dims_0 = const()[name = string("op_6620_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_6620_cast_fp16 = reduce_mean(axes = var_6620_axes_0, keep_dims = var_6620_keep_dims_0, x = var_6618_cast_fp16)[name = string("op_6620_cast_fp16")];
            fp16 var_6621_to_fp16 = const()[name = string("op_6621_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 1]> var_6622_cast_fp16 = add(x = var_6620_cast_fp16, y = var_6621_to_fp16)[name = string("op_6622_cast_fp16")];
            fp32 var_6623_epsilon_0 = const()[name = string("op_6623_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1]> var_6623_cast_fp16 = rsqrt(epsilon = var_6623_epsilon_0, x = var_6622_cast_fp16)[name = string("op_6623_cast_fp16")];
            tensor<fp16, [1, 640]> var_6624_cast_fp16 = mul(x = x_cast_fp16, y = var_6623_cast_fp16)[name = string("op_6624_cast_fp16")];
            tensor<fp16, [640]> norm_w1_to_fp16 = const()[name = string("norm_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119624768)))];
            tensor<fp16, [1, 640]> h_cast_fp16 = mul(x = var_6624_cast_fp16, y = norm_w1_to_fp16)[name = string("h_cast_fp16")];
            tensor<fp16, [16000]> logits_bias_0_to_fp16 = const()[name = string("logits_bias_0_to_fp16"), val = tensor<fp16, [16000]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119626112)))];
            tensor<fp16, [1, 16000]> logits_cast_fp16 = linear(bias = logits_bias_0_to_fp16, weight = emb_to_fp16, x = h_cast_fp16)[name = string("logits_cast_fp16")];
            tensor<int32, [1]> reduce_max_0_axes_0 = const()[name = string("reduce_max_0_axes_0"), val = tensor<int32, [1]>([-1])];
            bool reduce_max_0_keep_dims_0 = const()[name = string("reduce_max_0_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> reduce_max_0_cast_fp16 = reduce_max(axes = reduce_max_0_axes_0, keep_dims = reduce_max_0_keep_dims_0, x = logits_cast_fp16)[name = string("reduce_max_0_cast_fp16")];
            tensor<fp16, [1, 16000]> var_6633_cast_fp16 = sub(x = logits_cast_fp16, y = reduce_max_0_cast_fp16)[name = string("op_6633_cast_fp16")];
            tensor<fp16, [1, 16000]> var_6634_cast_fp16 = exp(x = var_6633_cast_fp16)[name = string("op_6634_cast_fp16")];
            tensor<int32, [1]> var_6639_axes_0 = const()[name = string("op_6639_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6639_keep_dims_0 = const()[name = string("op_6639_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_6639_cast_fp16 = reduce_sum(axes = var_6639_axes_0, keep_dims = var_6639_keep_dims_0, x = var_6634_cast_fp16)[name = string("op_6639_cast_fp16")];
            fp32 var_6640_epsilon_0 = const()[name = string("op_6640_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 1]> var_6640_cast_fp16 = log(epsilon = var_6640_epsilon_0, x = var_6639_cast_fp16)[name = string("op_6640_cast_fp16")];
            tensor<fp16, [1, 1]> var_6642_cast_fp16 = add(x = reduce_max_0_cast_fp16, y = var_6640_cast_fp16)[name = string("op_6642_cast_fp16")];
            tensor<fp16, [1, 16000]> var_6644_cast_fp16 = sub(x = logits_cast_fp16, y = var_6642_cast_fp16)[name = string("op_6644_cast_fp16")];
            string var_6644_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_6644_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 16000]> lp = cast(dtype = var_6644_cast_fp16_to_fp32_dtype_0, x = var_6644_cast_fp16)[name = string("cast_106")];
        } -> (lp);
    func n96<ios18>(tensor<int32, [1, 96]> ids, tensor<int32, [1]> last) {
            tensor<int32, [1, 96]> pos = const()[name = string("pos"), val = tensor<int32, [1, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119707392)))];
            int32 x_1_batch_dims_0 = const()[name = string("x_1_batch_dims_0"), val = int32(0)];
            bool x_1_validate_indices_0 = const()[name = string("x_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [16000, 640]> emb_to_fp16 = const()[name = string("emb_to_fp16"), val = tensor<fp16, [16000, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            string ids_to_int16_dtype_0 = const()[name = string("ids_to_int16_dtype_0"), val = string("int16")];
            string cast_129_dtype_0 = const()[name = string("cast_129_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int16, [1, 96]> ids_to_int16 = cast(dtype = ids_to_int16_dtype_0, x = ids)[name = string("cast_134")];
            tensor<int32, [1, 96]> cast_129 = cast(dtype = cast_129_dtype_0, x = ids_to_int16)[name = string("cast_133")];
            tensor<bool, [1, 96]> greater_equal_0 = greater_equal(x = cast_129, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(16000)];
            tensor<int32, [1, 96]> add_0 = add(x = cast_129, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 96]> select_0 = select(a = cast_129, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 x_1_cast_fp16_cast_uint16_axis_0 = const()[name = string("x_1_cast_fp16_cast_uint16_axis_0"), val = int32(0)];
            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
            tensor<int16, [1, 96]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_132")];
            tensor<fp16, [1, 96, 640]> x_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = x_1_cast_fp16_cast_uint16_axis_0, batch_dims = x_1_batch_dims_0, indices = select_0_to_int16, validate_indices = x_1_validate_indices_0, x = emb_to_fp16)[name = string("x_1_cast_fp16_cast_uint16_cast_uint16")];
            fp16 var_256_to_fp16 = const()[name = string("op_256_to_fp16"), val = fp16(0x1p+5)];
            tensor<fp16, [1, 96, 640]> x_3_cast_fp16 = mul(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = var_256_to_fp16)[name = string("x_3_cast_fp16")];
            fp16 var_254_promoted_to_fp16 = const()[name = string("op_254_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> var_258_cast_fp16 = pow(x = x_3_cast_fp16, y = var_254_promoted_to_fp16)[name = string("op_258_cast_fp16")];
            tensor<int32, [1]> var_260_axes_0 = const()[name = string("op_260_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_260_keep_dims_0 = const()[name = string("op_260_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 1]> var_260_cast_fp16 = reduce_mean(axes = var_260_axes_0, keep_dims = var_260_keep_dims_0, x = var_258_cast_fp16)[name = string("op_260_cast_fp16")];
            fp16 var_261_to_fp16 = const()[name = string("op_261_to_fp16"), val = fp16(0x1.0c8p-10)];
            tensor<fp16, [1, 96, 1]> var_262_cast_fp16 = add(x = var_260_cast_fp16, y = var_261_to_fp16)[name = string("op_262_cast_fp16")];
            fp32 var_263_epsilon_0 = const()[name = string("op_263_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 1]> var_263_cast_fp16 = rsqrt(epsilon = var_263_epsilon_0, x = var_262_cast_fp16)[name = string("op_263_cast_fp16")];
            tensor<fp16, [1, 96, 640]> var_264_cast_fp16 = mul(x = x_3_cast_fp16, y = var_263_cast_fp16)[name = string("op_264_cast_fp16")];
            tensor<fp16, [640]> layers_0_n1_w1_to_fp16 = const()[name = string("layers_0_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20480128)))];
            tensor<fp16, [1, 96, 640]> input_1_cast_fp16 = mul(x = var_264_cast_fp16, y = layers_0_n1_w1_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_0_qkvzba_weight_to_fp16 = const()[name = string("layers_0_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20481472)))];
            tensor<fp16, [1544]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [1544]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22457856)))];
            tensor<fp16, [1, 96, 1544]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_qkvzba_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [3]> qkv_1_begin_0 = const()[name = string("qkv_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_1_end_0 = const()[name = string("qkv_1_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> qkv_1_end_mask_0 = const()[name = string("qkv_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 1024]> qkv_1_cast_fp16 = slice_by_index(begin = qkv_1_begin_0, end = qkv_1_end_0, end_mask = qkv_1_end_mask_0, x = linear_0_cast_fp16)[name = string("qkv_1_cast_fp16")];
            tensor<int32, [3]> z_1_begin_0 = const()[name = string("z_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_1_end_0 = const()[name = string("z_1_end_0"), val = tensor<int32, [3]>([1, 96, 1536])];
            tensor<bool, [3]> z_1_end_mask_0 = const()[name = string("z_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 512]> z_1_cast_fp16 = slice_by_index(begin = z_1_begin_0, end = z_1_end_0, end_mask = z_1_end_mask_0, x = linear_0_cast_fp16)[name = string("z_1_cast_fp16")];
            tensor<int32, [3]> b_1_begin_0 = const()[name = string("b_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_1_end_0 = const()[name = string("b_1_end_0"), val = tensor<int32, [3]>([1, 96, 1540])];
            tensor<bool, [3]> b_1_end_mask_0 = const()[name = string("b_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 4]> b_1_cast_fp16 = slice_by_index(begin = b_1_begin_0, end = b_1_end_0, end_mask = b_1_end_mask_0, x = linear_0_cast_fp16)[name = string("b_1_cast_fp16")];
            tensor<int32, [3]> a_1_begin_0 = const()[name = string("a_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_1_end_0 = const()[name = string("a_1_end_0"), val = tensor<int32, [3]>([1, 96, 1])];
            tensor<bool, [3]> a_1_end_mask_0 = const()[name = string("a_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 4]> a_1_cast_fp16 = slice_by_index(begin = a_1_begin_0, end = a_1_end_0, end_mask = a_1_end_mask_0, x = linear_0_cast_fp16)[name = string("a_1_cast_fp16")];
            int32 var_303 = const()[name = string("op_303"), val = int32(1)];
            bool xp_1_interleave_0 = const()[name = string("xp_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 3, 1024]> zpad_to_fp16 = const()[name = string("zpad_to_fp16"), val = tensor<fp16, [1, 3, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119700672)))];
            tensor<fp16, [1, 99, 1024]> xp_1_cast_fp16 = concat(axis = var_303, interleave = xp_1_interleave_0, values = (zpad_to_fp16, qkv_1_cast_fp16))[name = string("xp_1_cast_fp16")];
            tensor<int32, [3]> var_314_begin_0 = const()[name = string("op_314_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_314_end_0 = const()[name = string("op_314_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_314_end_mask_0 = const()[name = string("op_314_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 1024]> var_314_cast_fp16 = slice_by_index(begin = var_314_begin_0, end = var_314_end_0, end_mask = var_314_end_mask_0, x = xp_1_cast_fp16)[name = string("op_314_cast_fp16")];
            tensor<fp16, [1024]> var_322_to_fp16 = const()[name = string("op_322_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22479552)))];
            tensor<fp16, [1, 96, 1024]> var_323_cast_fp16 = mul(x = var_314_cast_fp16, y = var_322_to_fp16)[name = string("op_323_cast_fp16")];
            tensor<int32, [3]> var_333_begin_0 = const()[name = string("op_333_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_333_end_0 = const()[name = string("op_333_end_0"), val = tensor<int32, [3]>([1, 98, 1024])];
            tensor<bool, [3]> var_333_end_mask_0 = const()[name = string("op_333_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_333_cast_fp16 = slice_by_index(begin = var_333_begin_0, end = var_333_end_0, end_mask = var_333_end_mask_0, x = xp_1_cast_fp16)[name = string("op_333_cast_fp16")];
            tensor<fp16, [1024]> var_341_to_fp16 = const()[name = string("op_341_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22481664)))];
            tensor<fp16, [1, 96, 1024]> var_342_cast_fp16 = mul(x = var_333_cast_fp16, y = var_341_to_fp16)[name = string("op_342_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> var_344_cast_fp16 = add(x = var_323_cast_fp16, y = var_342_cast_fp16)[name = string("op_344_cast_fp16")];
            tensor<int32, [3]> var_354_begin_0 = const()[name = string("op_354_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_354_end_0 = const()[name = string("op_354_end_0"), val = tensor<int32, [3]>([1, 97, 1024])];
            tensor<bool, [3]> var_354_end_mask_0 = const()[name = string("op_354_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_354_cast_fp16 = slice_by_index(begin = var_354_begin_0, end = var_354_end_0, end_mask = var_354_end_mask_0, x = xp_1_cast_fp16)[name = string("op_354_cast_fp16")];
            tensor<fp16, [1024]> var_362_to_fp16 = const()[name = string("op_362_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22483776)))];
            tensor<fp16, [1, 96, 1024]> var_363_cast_fp16 = mul(x = var_354_cast_fp16, y = var_362_to_fp16)[name = string("op_363_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> var_365_cast_fp16 = add(x = var_344_cast_fp16, y = var_363_cast_fp16)[name = string("op_365_cast_fp16")];
            tensor<int32, [3]> var_375_begin_0 = const()[name = string("op_375_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_375_end_0 = const()[name = string("op_375_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> var_375_end_mask_0 = const()[name = string("op_375_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_375_cast_fp16 = slice_by_index(begin = var_375_begin_0, end = var_375_end_0, end_mask = var_375_end_mask_0, x = xp_1_cast_fp16)[name = string("op_375_cast_fp16")];
            tensor<fp16, [1024]> var_383_to_fp16 = const()[name = string("op_383_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22485888)))];
            tensor<fp16, [1, 96, 1024]> var_384_cast_fp16 = mul(x = var_375_cast_fp16, y = var_383_to_fp16)[name = string("op_384_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> input_3_cast_fp16 = add(x = var_365_cast_fp16, y = var_384_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> c_1_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("c_1_cast_fp16")];
            tensor<int32, [3]> var_392_begin_0 = const()[name = string("op_392_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_392_end_0 = const()[name = string("op_392_end_0"), val = tensor<int32, [3]>([1, 96, 256])];
            tensor<bool, [3]> var_392_end_mask_0 = const()[name = string("op_392_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 256]> var_392_cast_fp16 = slice_by_index(begin = var_392_begin_0, end = var_392_end_0, end_mask = var_392_end_mask_0, x = c_1_cast_fp16)[name = string("op_392_cast_fp16")];
            tensor<int32, [4]> var_397 = const()[name = string("op_397"), val = tensor<int32, [4]>([1, 96, 2, 128])];
            tensor<fp16, [1, 96, 2, 128]> q_1_cast_fp16 = reshape(shape = var_397, x = var_392_cast_fp16)[name = string("q_1_cast_fp16")];
            tensor<int32, [3]> var_403_begin_0 = const()[name = string("op_403_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_403_end_0 = const()[name = string("op_403_end_0"), val = tensor<int32, [3]>([1, 96, 512])];
            tensor<bool, [3]> var_403_end_mask_0 = const()[name = string("op_403_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 256]> var_403_cast_fp16 = slice_by_index(begin = var_403_begin_0, end = var_403_end_0, end_mask = var_403_end_mask_0, x = c_1_cast_fp16)[name = string("op_403_cast_fp16")];
            tensor<int32, [4]> var_408 = const()[name = string("op_408"), val = tensor<int32, [4]>([1, 96, 2, 128])];
            tensor<fp16, [1, 96, 2, 128]> k_1_cast_fp16 = reshape(shape = var_408, x = var_403_cast_fp16)[name = string("k_1_cast_fp16")];
            tensor<int32, [3]> var_414_begin_0 = const()[name = string("op_414_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_414_end_0 = const()[name = string("op_414_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> var_414_end_mask_0 = const()[name = string("op_414_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 512]> var_414_cast_fp16 = slice_by_index(begin = var_414_begin_0, end = var_414_end_0, end_mask = var_414_end_mask_0, x = c_1_cast_fp16)[name = string("op_414_cast_fp16")];
            tensor<int32, [4]> var_419 = const()[name = string("op_419"), val = tensor<int32, [4]>([1, 96, 4, 128])];
            tensor<fp16, [1, 96, 4, 128]> var_420_cast_fp16 = reshape(shape = var_419, x = var_414_cast_fp16)[name = string("op_420_cast_fp16")];
            tensor<int32, [4]> var_425 = const()[name = string("op_425"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_427_to_fp16 = const()[name = string("op_427_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 96, 2, 128]> q_3_cast_fp16 = mul(x = q_1_cast_fp16, y = var_427_to_fp16)[name = string("q_3_cast_fp16")];
            fp16 var_429_to_fp16 = const()[name = string("op_429_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 96, 2, 128]> k_3_cast_fp16 = mul(x = k_1_cast_fp16, y = var_429_to_fp16)[name = string("k_3_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_431_cast_fp16 = mul(x = q_3_cast_fp16, y = q_3_cast_fp16)[name = string("op_431_cast_fp16")];
            tensor<int32, [1]> var_436_axes_0 = const()[name = string("op_436_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_436_keep_dims_0 = const()[name = string("op_436_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 2, 1]> var_436_cast_fp16 = reduce_mean(axes = var_436_axes_0, keep_dims = var_436_keep_dims_0, x = var_431_cast_fp16)[name = string("op_436_cast_fp16")];
            fp16 var_437_promoted_to_fp16 = const()[name = string("op_437_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 96, 2, 1]> var_438_cast_fp16 = mul(x = var_436_cast_fp16, y = var_437_promoted_to_fp16)[name = string("op_438_cast_fp16")];
            fp16 var_440_to_fp16 = const()[name = string("op_440_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 96, 2, 1]> var_441_cast_fp16 = add(x = var_438_cast_fp16, y = var_440_to_fp16)[name = string("op_441_cast_fp16")];
            fp32 var_442_epsilon_0 = const()[name = string("op_442_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 2, 1]> var_442_cast_fp16 = rsqrt(epsilon = var_442_epsilon_0, x = var_441_cast_fp16)[name = string("op_442_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_443_cast_fp16 = mul(x = q_3_cast_fp16, y = var_442_cast_fp16)[name = string("op_443_cast_fp16")];
            fp16 var_444_to_fp16 = const()[name = string("op_444_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 96, 2, 128]> q_5_cast_fp16 = mul(x = var_443_cast_fp16, y = var_444_to_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_446_cast_fp16 = mul(x = k_3_cast_fp16, y = k_3_cast_fp16)[name = string("op_446_cast_fp16")];
            tensor<int32, [1]> var_451_axes_0 = const()[name = string("op_451_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_451_keep_dims_0 = const()[name = string("op_451_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 2, 1]> var_451_cast_fp16 = reduce_mean(axes = var_451_axes_0, keep_dims = var_451_keep_dims_0, x = var_446_cast_fp16)[name = string("op_451_cast_fp16")];
            fp16 var_452_promoted_to_fp16 = const()[name = string("op_452_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 96, 2, 1]> var_453_cast_fp16 = mul(x = var_451_cast_fp16, y = var_452_promoted_to_fp16)[name = string("op_453_cast_fp16")];
            fp16 var_455_to_fp16 = const()[name = string("op_455_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 96, 2, 1]> var_456_cast_fp16 = add(x = var_453_cast_fp16, y = var_455_to_fp16)[name = string("op_456_cast_fp16")];
            fp32 var_457_epsilon_0 = const()[name = string("op_457_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 2, 1]> var_457_cast_fp16 = rsqrt(epsilon = var_457_epsilon_0, x = var_456_cast_fp16)[name = string("op_457_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> k_5_cast_fp16 = mul(x = k_3_cast_fp16, y = var_457_cast_fp16)[name = string("k_5_cast_fp16")];
            tensor<int32, [4]> var_463 = const()[name = string("op_463"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_476_axes_0 = const()[name = string("op_476_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 96, 128]> var_464_cast_fp16 = transpose(perm = var_463, x = q_5_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [1, 2, 1, 96, 128]> var_476_cast_fp16 = expand_dims(axes = var_476_axes_0, x = var_464_cast_fp16)[name = string("op_476_cast_fp16")];
            tensor<int32, [5]> var_484_reps_0 = const()[name = string("op_484_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 96, 128]> var_484_cast_fp16 = tile(reps = var_484_reps_0, x = var_476_cast_fp16)[name = string("op_484_cast_fp16")];
            tensor<int32, [4]> var_489 = const()[name = string("op_489"), val = tensor<int32, [4]>([1, 4, 96, 128])];
            tensor<fp16, [1, 4, 96, 128]> q_7_cast_fp16 = reshape(shape = var_489, x = var_484_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<int32, [4]> var_495 = const()[name = string("op_495"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_508_axes_0 = const()[name = string("op_508_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 96, 128]> var_496_cast_fp16 = transpose(perm = var_495, x = k_5_cast_fp16)[name = string("transpose_162")];
            tensor<fp16, [1, 2, 1, 96, 128]> var_508_cast_fp16 = expand_dims(axes = var_508_axes_0, x = var_496_cast_fp16)[name = string("op_508_cast_fp16")];
            tensor<int32, [5]> var_516_reps_0 = const()[name = string("op_516_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 96, 128]> var_516_cast_fp16 = tile(reps = var_516_reps_0, x = var_508_cast_fp16)[name = string("op_516_cast_fp16")];
            tensor<int32, [4]> var_521 = const()[name = string("op_521"), val = tensor<int32, [4]>([1, 4, 96, 128])];
            tensor<fp16, [1, 4, 96, 128]> k_7_cast_fp16 = reshape(shape = var_521, x = var_516_cast_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_523_cast_fp16 = sigmoid(x = b_1_cast_fp16)[name = string("op_523_cast_fp16")];
            tensor<int32, [3]> var_527 = const()[name = string("op_527"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_0_dt_bias_to_fp16 = const()[name = string("layers_0_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 96, 4]> input_5_cast_fp16 = add(x = a_1_cast_fp16, y = layers_0_dt_bias_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_531_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("op_531_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_532_cast_fp16 = abs(x = input_5_cast_fp16)[name = string("op_532_cast_fp16")];
            fp16 const_0_promoted_to_fp16 = const()[name = string("const_0_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 96, 4]> var_533_cast_fp16 = mul(x = var_532_cast_fp16, y = const_0_promoted_to_fp16)[name = string("op_533_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_534_cast_fp16 = exp(x = var_533_cast_fp16)[name = string("op_534_cast_fp16")];
            fp16 var_536_to_fp16 = const()[name = string("op_536_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 96, 4]> var_537_cast_fp16 = add(x = var_534_cast_fp16, y = var_536_to_fp16)[name = string("op_537_cast_fp16")];
            fp32 var_538_epsilon_0 = const()[name = string("op_538_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 96, 4]> var_538_cast_fp16 = log(epsilon = var_538_epsilon_0, x = var_537_cast_fp16)[name = string("op_538_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_540_cast_fp16 = add(x = var_531_cast_fp16, y = var_538_cast_fp16)[name = string("op_540_cast_fp16")];
            tensor<fp16, [4]> layers_0_negA_to_fp16 = const()[name = string("layers_0_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.b3p+3, -0x1.a7cp+0, -0x1.ac4p+3, -0x1.704p+2])];
            tensor<fp16, [1, 96, 4]> var_541_cast_fp16 = mul(x = layers_0_negA_to_fp16, y = var_540_cast_fp16)[name = string("op_541_cast_fp16")];
            tensor<int32, [3]> var_545 = const()[name = string("op_545"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_548_axes_0 = const()[name = string("op_548_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 96]> beta_1_cast_fp16 = transpose(perm = var_527, x = var_523_cast_fp16)[name = string("transpose_161")];
            tensor<fp16, [1, 4, 96, 1]> var_548_cast_fp16 = expand_dims(axes = var_548_axes_0, x = beta_1_cast_fp16)[name = string("op_548_cast_fp16")];
            fp16 var_549_to_fp16 = const()[name = string("op_549_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 96, 1]> var_550_cast_fp16 = mul(x = var_548_cast_fp16, y = var_549_to_fp16)[name = string("op_550_cast_fp16")];
            tensor<fp16, [1, 4, 96, 128]> v_1_cast_fp16 = transpose(perm = var_425, x = var_420_cast_fp16)[name = string("transpose_164")];
            tensor<fp16, [1, 4, 96, 128]> vb_1_cast_fp16 = mul(x = v_1_cast_fp16, y = var_550_cast_fp16)[name = string("vb_1_cast_fp16")];
            tensor<int32, [4]> q_9_begin_0 = const()[name = string("q_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_9_end_0 = const()[name = string("q_9_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_9_end_mask_0 = const()[name = string("q_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_9_cast_fp16 = slice_by_index(begin = q_9_begin_0, end = q_9_end_0, end_mask = q_9_end_mask_0, x = q_7_cast_fp16)[name = string("q_9_cast_fp16")];
            tensor<int32, [4]> k_9_begin_0 = const()[name = string("k_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_9_end_0 = const()[name = string("k_9_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_9_end_mask_0 = const()[name = string("k_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, x = k_7_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<fp16, [1, 4, 96, 128]> var_584_cast_fp16 = mul(x = k_7_cast_fp16, y = var_548_cast_fp16)[name = string("op_584_cast_fp16")];
            tensor<int32, [4]> kb_1_begin_0 = const()[name = string("kb_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_1_end_0 = const()[name = string("kb_1_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_1_end_mask_0 = const()[name = string("kb_1_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_1_cast_fp16 = slice_by_index(begin = kb_1_begin_0, end = kb_1_end_0, end_mask = kb_1_end_mask_0, x = var_584_cast_fp16)[name = string("kb_1_cast_fp16")];
            tensor<int32, [4]> vb_3_begin_0 = const()[name = string("vb_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_3_end_0 = const()[name = string("vb_3_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_3_end_mask_0 = const()[name = string("vb_3_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_3_cast_fp16 = slice_by_index(begin = vb_3_begin_0, end = vb_3_end_0, end_mask = vb_3_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_3_cast_fp16")];
            tensor<int32, [3]> gdec_3_begin_0 = const()[name = string("gdec_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_3_end_0 = const()[name = string("gdec_3_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_3_end_mask_0 = const()[name = string("gdec_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 96]> gdec_1_cast_fp16 = transpose(perm = var_545, x = var_541_cast_fp16)[name = string("transpose_160")];
            tensor<fp16, [1, 4, 16]> gdec_3_cast_fp16 = slice_by_index(begin = gdec_3_begin_0, end = gdec_3_end_0, end_mask = gdec_3_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_3_cast_fp16")];
            tensor<int32, [2]> var_632 = const()[name = string("op_632"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_633_cast_fp16 = reshape(shape = var_632, x = gdec_3_cast_fp16)[name = string("op_633_cast_fp16")];
            tensor<fp16, [256, 16]> transpose_0_to_fp16 = const()[name = string("transpose_0_to_fp16"), val = tensor<fp16, [256, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22488000)))];
            tensor<fp16, [256]> var_634_bias_0_to_fp16 = const()[name = string("op_634_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22496256)))];
            tensor<fp16, [4, 256]> var_634_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_633_cast_fp16)[name = string("op_634_cast_fp16")];
            tensor<int32, [4]> var_639 = const()[name = string("op_639"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_1_cast_fp16 = reshape(shape = var_639, x = var_634_cast_fp16)[name = string("P_1_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_641_cast_fp16 = exp(x = P_1_cast_fp16)[name = string("op_641_cast_fp16")];
            tensor<fp16, [16, 16]> tril_incl_to_fp16 = const()[name = string("tril_incl_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22496832)))];
            tensor<fp16, [1, 4, 16, 16]> D_1_cast_fp16 = mul(x = var_641_cast_fp16, y = tril_incl_to_fp16)[name = string("D_1_cast_fp16")];
            bool kk_1_transpose_x_1 = const()[name = string("kk_1_transpose_x_1"), val = bool(false)];
            bool kk_1_transpose_y_1 = const()[name = string("kk_1_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_1_cast_fp16 = matmul(transpose_x = kk_1_transpose_x_1, transpose_y = kk_1_transpose_y_1, x = kb_1_cast_fp16, y = k_9_cast_fp16)[name = string("kk_1_cast_fp16")];
            fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_647_cast_fp16 = mul(x = kk_1_cast_fp16, y = const_1_promoted_to_fp16)[name = string("op_647_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_648_cast_fp16 = mul(x = var_647_cast_fp16, y = D_1_cast_fp16)[name = string("op_648_cast_fp16")];
            tensor<fp16, [16, 16]> cum_tail_to_fp16 = const()[name = string("cum_tail_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22497408)))];
            tensor<fp16, [1, 4, 16, 16]> N_1_cast_fp16 = mul(x = var_648_cast_fp16, y = cum_tail_to_fp16)[name = string("N_1_cast_fp16")];
            tensor<fp16, [16, 16]> neg_tril_strict_to_fp16 = const()[name = string("neg_tril_strict_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22497984)))];
            tensor<fp16, [1, 4, 16, 16]> var_650_cast_fp16 = mul(x = D_1_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_650_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_1_cast_fp16 = mul(x = kk_1_cast_fp16, y = var_650_cast_fp16)[name = string("Na_1_cast_fp16")];
            tensor<fp16, [16, 16]> eye_to_fp16 = const()[name = string("eye_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22498560)))];
            tensor<fp16, [1, 4, 16, 16]> Minv_1_cast_fp16 = add(x = eye_to_fp16, y = N_1_cast_fp16)[name = string("Minv_1_cast_fp16")];
            bool Np_1_transpose_x_0 = const()[name = string("Np_1_transpose_x_0"), val = bool(false)];
            bool Np_1_transpose_y_0 = const()[name = string("Np_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_1_cast_fp16 = matmul(transpose_x = Np_1_transpose_x_0, transpose_y = Np_1_transpose_y_0, x = N_1_cast_fp16, y = Na_1_cast_fp16)[name = string("Np_1_cast_fp16")];
            bool Npa_1_transpose_x_0 = const()[name = string("Npa_1_transpose_x_0"), val = bool(false)];
            bool Npa_1_transpose_y_0 = const()[name = string("Npa_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_1_cast_fp16 = matmul(transpose_x = Npa_1_transpose_x_0, transpose_y = Npa_1_transpose_y_0, x = Na_1_cast_fp16, y = N_1_cast_fp16)[name = string("Npa_1_cast_fp16")];
            bool var_656_transpose_x_0 = const()[name = string("op_656_transpose_x_0"), val = bool(false)];
            bool var_656_transpose_y_0 = const()[name = string("op_656_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_656_cast_fp16 = matmul(transpose_x = var_656_transpose_x_0, transpose_y = var_656_transpose_y_0, x = Minv_1_cast_fp16, y = Np_1_cast_fp16)[name = string("op_656_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_3_cast_fp16 = add(x = Minv_1_cast_fp16, y = var_656_cast_fp16)[name = string("Minv_3_cast_fp16")];
            bool Np_3_transpose_x_0 = const()[name = string("Np_3_transpose_x_0"), val = bool(false)];
            bool Np_3_transpose_y_0 = const()[name = string("Np_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_3_cast_fp16 = matmul(transpose_x = Np_3_transpose_x_0, transpose_y = Np_3_transpose_y_0, x = Np_1_cast_fp16, y = Npa_1_cast_fp16)[name = string("Np_3_cast_fp16")];
            bool Npa_3_transpose_x_0 = const()[name = string("Npa_3_transpose_x_0"), val = bool(false)];
            bool Npa_3_transpose_y_0 = const()[name = string("Npa_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_3_cast_fp16 = matmul(transpose_x = Npa_3_transpose_x_0, transpose_y = Npa_3_transpose_y_0, x = Npa_1_cast_fp16, y = Np_1_cast_fp16)[name = string("Npa_3_cast_fp16")];
            bool var_661_transpose_x_0 = const()[name = string("op_661_transpose_x_0"), val = bool(false)];
            bool var_661_transpose_y_0 = const()[name = string("op_661_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_661_cast_fp16 = matmul(transpose_x = var_661_transpose_x_0, transpose_y = var_661_transpose_y_0, x = Minv_3_cast_fp16, y = Np_3_cast_fp16)[name = string("op_661_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_5_cast_fp16 = add(x = Minv_3_cast_fp16, y = var_661_cast_fp16)[name = string("Minv_5_cast_fp16")];
            bool Np_5_transpose_x_0 = const()[name = string("Np_5_transpose_x_0"), val = bool(false)];
            bool Np_5_transpose_y_0 = const()[name = string("Np_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_5_cast_fp16 = matmul(transpose_x = Np_5_transpose_x_0, transpose_y = Np_5_transpose_y_0, x = Np_3_cast_fp16, y = Npa_3_cast_fp16)[name = string("Np_5_cast_fp16")];
            bool var_665_transpose_x_0 = const()[name = string("op_665_transpose_x_0"), val = bool(false)];
            bool var_665_transpose_y_0 = const()[name = string("op_665_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_665_cast_fp16 = matmul(transpose_x = var_665_transpose_x_0, transpose_y = var_665_transpose_y_0, x = Minv_5_cast_fp16, y = Np_5_cast_fp16)[name = string("op_665_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_7_cast_fp16 = add(x = Minv_5_cast_fp16, y = var_665_cast_fp16)[name = string("Minv_7_cast_fp16")];
            bool u_1_transpose_x_0 = const()[name = string("u_1_transpose_x_0"), val = bool(false)];
            bool u_1_transpose_y_0 = const()[name = string("u_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_1_cast_fp16 = matmul(transpose_x = u_1_transpose_x_0, transpose_y = u_1_transpose_y_0, x = Minv_7_cast_fp16, y = vb_3_cast_fp16)[name = string("u_1_cast_fp16")];
            bool var_672_transpose_x_1 = const()[name = string("op_672_transpose_x_1"), val = bool(false)];
            bool var_672_transpose_y_1 = const()[name = string("op_672_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_672_cast_fp16 = matmul(transpose_x = var_672_transpose_x_1, transpose_y = var_672_transpose_y_1, x = q_9_cast_fp16, y = k_9_cast_fp16)[name = string("op_672_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_1_cast_fp16 = mul(x = var_672_cast_fp16, y = D_1_cast_fp16)[name = string("A_1_cast_fp16")];
            bool var_674_transpose_x_0 = const()[name = string("op_674_transpose_x_0"), val = bool(false)];
            bool var_674_transpose_y_0 = const()[name = string("op_674_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_674_cast_fp16 = matmul(transpose_x = var_674_transpose_x_0, transpose_y = var_674_transpose_y_0, x = A_1_cast_fp16, y = u_1_cast_fp16)[name = string("op_674_cast_fp16")];
            tensor<fp16, [16, 16]> transpose_1_to_fp16 = const()[name = string("transpose_1_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658176)))];
            tensor<fp16, [16]> var_675_bias_0_to_fp16 = const()[name = string("op_675_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_675_cast_fp16 = linear(bias = var_675_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_3_cast_fp16)[name = string("op_675_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_1_cast_fp16 = exp(x = var_675_cast_fp16)[name = string("tail_1_cast_fp16")];
            tensor<int32, [1]> var_678_axes_0 = const()[name = string("op_678_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_678_cast_fp16 = expand_dims(axes = var_678_axes_0, x = tail_1_cast_fp16)[name = string("op_678_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_679_cast_fp16 = mul(x = k_9_cast_fp16, y = var_678_cast_fp16)[name = string("op_679_cast_fp16")];
            bool S_1_transpose_x_1 = const()[name = string("S_1_transpose_x_1"), val = bool(true)];
            bool S_1_transpose_y_1 = const()[name = string("S_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_1_cast_fp16 = matmul(transpose_x = S_1_transpose_x_1, transpose_y = S_1_transpose_y_1, x = var_679_cast_fp16, y = u_1_cast_fp16)[name = string("S_1_cast_fp16")];
            tensor<int32, [4]> q_11_begin_0 = const()[name = string("q_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_11_end_0 = const()[name = string("q_11_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> q_11_end_mask_0 = const()[name = string("q_11_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_11_cast_fp16 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, x = q_7_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<int32, [4]> k_11_begin_0 = const()[name = string("k_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_11_end_0 = const()[name = string("k_11_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> k_11_end_mask_0 = const()[name = string("k_11_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_11_cast_fp16 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, x = k_7_cast_fp16)[name = string("k_11_cast_fp16")];
            tensor<int32, [4]> kb_3_begin_0 = const()[name = string("kb_3_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_3_end_0 = const()[name = string("kb_3_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> kb_3_end_mask_0 = const()[name = string("kb_3_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_3_cast_fp16 = slice_by_index(begin = kb_3_begin_0, end = kb_3_end_0, end_mask = kb_3_end_mask_0, x = var_584_cast_fp16)[name = string("kb_3_cast_fp16")];
            tensor<int32, [4]> vb_5_begin_0 = const()[name = string("vb_5_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_5_end_0 = const()[name = string("vb_5_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> vb_5_end_mask_0 = const()[name = string("vb_5_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_5_cast_fp16 = slice_by_index(begin = vb_5_begin_0, end = vb_5_end_0, end_mask = vb_5_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_5_cast_fp16")];
            tensor<int32, [3]> gdec_5_begin_0 = const()[name = string("gdec_5_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_5_end_0 = const()[name = string("gdec_5_end_0"), val = tensor<int32, [3]>([1, 4, 32])];
            tensor<bool, [3]> gdec_5_end_mask_0 = const()[name = string("gdec_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_5_cast_fp16 = slice_by_index(begin = gdec_5_begin_0, end = gdec_5_end_0, end_mask = gdec_5_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_5_cast_fp16")];
            tensor<int32, [2]> var_764 = const()[name = string("op_764"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_765_cast_fp16 = reshape(shape = var_764, x = gdec_5_cast_fp16)[name = string("op_765_cast_fp16")];
            tensor<fp16, [4, 256]> var_766_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_765_cast_fp16)[name = string("op_766_cast_fp16")];
            tensor<int32, [4]> var_771 = const()[name = string("op_771"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_3_cast_fp16 = reshape(shape = var_771, x = var_766_cast_fp16)[name = string("P_3_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_773_cast_fp16 = exp(x = P_3_cast_fp16)[name = string("op_773_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_3_cast_fp16 = mul(x = var_773_cast_fp16, y = tril_incl_to_fp16)[name = string("D_3_cast_fp16")];
            bool kk_3_transpose_x_1 = const()[name = string("kk_3_transpose_x_1"), val = bool(false)];
            bool kk_3_transpose_y_1 = const()[name = string("kk_3_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_3_cast_fp16 = matmul(transpose_x = kk_3_transpose_x_1, transpose_y = kk_3_transpose_y_1, x = kb_3_cast_fp16, y = k_11_cast_fp16)[name = string("kk_3_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_779_cast_fp16 = mul(x = kk_3_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_779_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_780_cast_fp16 = mul(x = var_779_cast_fp16, y = D_3_cast_fp16)[name = string("op_780_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_3_cast_fp16 = mul(x = var_780_cast_fp16, y = cum_tail_to_fp16)[name = string("N_3_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_782_cast_fp16 = mul(x = D_3_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_782_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_3_cast_fp16 = mul(x = kk_3_cast_fp16, y = var_782_cast_fp16)[name = string("Na_3_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_9_cast_fp16 = add(x = eye_to_fp16, y = N_3_cast_fp16)[name = string("Minv_9_cast_fp16")];
            bool Np_7_transpose_x_0 = const()[name = string("Np_7_transpose_x_0"), val = bool(false)];
            bool Np_7_transpose_y_0 = const()[name = string("Np_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_7_cast_fp16 = matmul(transpose_x = Np_7_transpose_x_0, transpose_y = Np_7_transpose_y_0, x = N_3_cast_fp16, y = Na_3_cast_fp16)[name = string("Np_7_cast_fp16")];
            bool Npa_5_transpose_x_0 = const()[name = string("Npa_5_transpose_x_0"), val = bool(false)];
            bool Npa_5_transpose_y_0 = const()[name = string("Npa_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_5_cast_fp16 = matmul(transpose_x = Npa_5_transpose_x_0, transpose_y = Npa_5_transpose_y_0, x = Na_3_cast_fp16, y = N_3_cast_fp16)[name = string("Npa_5_cast_fp16")];
            bool var_788_transpose_x_0 = const()[name = string("op_788_transpose_x_0"), val = bool(false)];
            bool var_788_transpose_y_0 = const()[name = string("op_788_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_788_cast_fp16 = matmul(transpose_x = var_788_transpose_x_0, transpose_y = var_788_transpose_y_0, x = Minv_9_cast_fp16, y = Np_7_cast_fp16)[name = string("op_788_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_11_cast_fp16 = add(x = Minv_9_cast_fp16, y = var_788_cast_fp16)[name = string("Minv_11_cast_fp16")];
            bool Np_9_transpose_x_0 = const()[name = string("Np_9_transpose_x_0"), val = bool(false)];
            bool Np_9_transpose_y_0 = const()[name = string("Np_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_9_cast_fp16 = matmul(transpose_x = Np_9_transpose_x_0, transpose_y = Np_9_transpose_y_0, x = Np_7_cast_fp16, y = Npa_5_cast_fp16)[name = string("Np_9_cast_fp16")];
            bool Npa_7_transpose_x_0 = const()[name = string("Npa_7_transpose_x_0"), val = bool(false)];
            bool Npa_7_transpose_y_0 = const()[name = string("Npa_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_7_cast_fp16 = matmul(transpose_x = Npa_7_transpose_x_0, transpose_y = Npa_7_transpose_y_0, x = Npa_5_cast_fp16, y = Np_7_cast_fp16)[name = string("Npa_7_cast_fp16")];
            bool var_793_transpose_x_0 = const()[name = string("op_793_transpose_x_0"), val = bool(false)];
            bool var_793_transpose_y_0 = const()[name = string("op_793_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_793_cast_fp16 = matmul(transpose_x = var_793_transpose_x_0, transpose_y = var_793_transpose_y_0, x = Minv_11_cast_fp16, y = Np_9_cast_fp16)[name = string("op_793_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_13_cast_fp16 = add(x = Minv_11_cast_fp16, y = var_793_cast_fp16)[name = string("Minv_13_cast_fp16")];
            bool Np_11_transpose_x_0 = const()[name = string("Np_11_transpose_x_0"), val = bool(false)];
            bool Np_11_transpose_y_0 = const()[name = string("Np_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_11_cast_fp16 = matmul(transpose_x = Np_11_transpose_x_0, transpose_y = Np_11_transpose_y_0, x = Np_9_cast_fp16, y = Npa_7_cast_fp16)[name = string("Np_11_cast_fp16")];
            bool var_797_transpose_x_0 = const()[name = string("op_797_transpose_x_0"), val = bool(false)];
            bool var_797_transpose_y_0 = const()[name = string("op_797_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_797_cast_fp16 = matmul(transpose_x = var_797_transpose_x_0, transpose_y = var_797_transpose_y_0, x = Minv_13_cast_fp16, y = Np_11_cast_fp16)[name = string("op_797_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_15_cast_fp16 = add(x = Minv_13_cast_fp16, y = var_797_cast_fp16)[name = string("Minv_15_cast_fp16")];
            bool u_3_transpose_x_0 = const()[name = string("u_3_transpose_x_0"), val = bool(false)];
            bool u_3_transpose_y_0 = const()[name = string("u_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_3_cast_fp16 = matmul(transpose_x = u_3_transpose_x_0, transpose_y = u_3_transpose_y_0, x = Minv_15_cast_fp16, y = vb_5_cast_fp16)[name = string("u_3_cast_fp16")];
            bool var_804_transpose_x_1 = const()[name = string("op_804_transpose_x_1"), val = bool(false)];
            bool var_804_transpose_y_1 = const()[name = string("op_804_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_804_cast_fp16 = matmul(transpose_x = var_804_transpose_x_1, transpose_y = var_804_transpose_y_1, x = q_11_cast_fp16, y = k_11_cast_fp16)[name = string("op_804_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_3_cast_fp16 = mul(x = var_804_cast_fp16, y = D_3_cast_fp16)[name = string("A_3_cast_fp16")];
            tensor<fp16, [16]> var_806_bias_0_to_fp16 = const()[name = string("op_806_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_806_cast_fp16 = linear(bias = var_806_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_5_cast_fp16)[name = string("op_806_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_1_cast_fp16 = exp(x = var_806_cast_fp16)[name = string("cum_1_cast_fp16")];
            tensor<int32, [1]> var_809_axes_0 = const()[name = string("op_809_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_809_cast_fp16 = expand_dims(axes = var_809_axes_0, x = cum_1_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_810_cast_fp16 = mul(x = kb_3_cast_fp16, y = var_809_cast_fp16)[name = string("op_810_cast_fp16")];
            bool w_3_transpose_x_0 = const()[name = string("w_3_transpose_x_0"), val = bool(false)];
            bool w_3_transpose_y_0 = const()[name = string("w_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_3_cast_fp16 = matmul(transpose_x = w_3_transpose_x_0, transpose_y = w_3_transpose_y_0, x = Minv_15_cast_fp16, y = var_810_cast_fp16)[name = string("w_3_cast_fp16")];
            bool var_812_transpose_x_0 = const()[name = string("op_812_transpose_x_0"), val = bool(false)];
            bool var_812_transpose_y_0 = const()[name = string("op_812_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_812_cast_fp16 = matmul(transpose_x = var_812_transpose_x_0, transpose_y = var_812_transpose_y_0, x = w_3_cast_fp16, y = S_1_cast_fp16)[name = string("op_812_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_1_cast_fp16 = sub(x = u_3_cast_fp16, y = var_812_cast_fp16)[name = string("vnew_1_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_817_cast_fp16 = mul(x = q_11_cast_fp16, y = var_809_cast_fp16)[name = string("op_817_cast_fp16")];
            bool var_818_transpose_x_0 = const()[name = string("op_818_transpose_x_0"), val = bool(false)];
            bool var_818_transpose_y_0 = const()[name = string("op_818_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_818_cast_fp16 = matmul(transpose_x = var_818_transpose_x_0, transpose_y = var_818_transpose_y_0, x = var_817_cast_fp16, y = S_1_cast_fp16)[name = string("op_818_cast_fp16")];
            bool var_819_transpose_x_0 = const()[name = string("op_819_transpose_x_0"), val = bool(false)];
            bool var_819_transpose_y_0 = const()[name = string("op_819_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_819_cast_fp16 = matmul(transpose_x = var_819_transpose_x_0, transpose_y = var_819_transpose_y_0, x = A_3_cast_fp16, y = vnew_1_cast_fp16)[name = string("op_819_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_821_cast_fp16 = add(x = var_818_cast_fp16, y = var_819_cast_fp16)[name = string("op_821_cast_fp16")];
            tensor<fp16, [16]> var_822_bias_0_to_fp16 = const()[name = string("op_822_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_822_cast_fp16 = linear(bias = var_822_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_5_cast_fp16)[name = string("op_822_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_3_cast_fp16 = exp(x = var_822_cast_fp16)[name = string("tail_3_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_5_to_fp16 = const()[name = string("transpose_5_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_824_bias_0_to_fp16 = const()[name = string("op_824_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_824_cast_fp16 = linear(bias = var_824_bias_0_to_fp16, weight = transpose_5_to_fp16, x = gdec_5_cast_fp16)[name = string("op_824_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_825_cast_fp16 = exp(x = var_824_cast_fp16)[name = string("op_825_cast_fp16")];
            tensor<int32, [1]> last_3_axes_0 = const()[name = string("last_3_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_3_cast_fp16 = expand_dims(axes = last_3_axes_0, x = var_825_cast_fp16)[name = string("last_3_cast_fp16")];
            tensor<int32, [1]> var_829_axes_0 = const()[name = string("op_829_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_829_cast_fp16 = expand_dims(axes = var_829_axes_0, x = tail_3_cast_fp16)[name = string("op_829_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_830_cast_fp16 = mul(x = k_11_cast_fp16, y = var_829_cast_fp16)[name = string("op_830_cast_fp16")];
            bool upd_1_transpose_x_1 = const()[name = string("upd_1_transpose_x_1"), val = bool(true)];
            bool upd_1_transpose_y_1 = const()[name = string("upd_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_1_cast_fp16 = matmul(transpose_x = upd_1_transpose_x_1, transpose_y = upd_1_transpose_y_1, x = var_830_cast_fp16, y = vnew_1_cast_fp16)[name = string("upd_1_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_835_cast_fp16 = mul(x = S_1_cast_fp16, y = last_3_cast_fp16)[name = string("op_835_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_3_cast_fp16 = add(x = var_835_cast_fp16, y = upd_1_cast_fp16)[name = string("S_3_cast_fp16")];
            tensor<int32, [4]> q_13_begin_0 = const()[name = string("q_13_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_13_end_0 = const()[name = string("q_13_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> q_13_end_mask_0 = const()[name = string("q_13_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, x = q_7_cast_fp16)[name = string("q_13_cast_fp16")];
            tensor<int32, [4]> k_13_begin_0 = const()[name = string("k_13_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_13_end_0 = const()[name = string("k_13_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> k_13_end_mask_0 = const()[name = string("k_13_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, x = k_7_cast_fp16)[name = string("k_13_cast_fp16")];
            tensor<int32, [4]> kb_5_begin_0 = const()[name = string("kb_5_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_5_end_0 = const()[name = string("kb_5_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> kb_5_end_mask_0 = const()[name = string("kb_5_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_5_cast_fp16 = slice_by_index(begin = kb_5_begin_0, end = kb_5_end_0, end_mask = kb_5_end_mask_0, x = var_584_cast_fp16)[name = string("kb_5_cast_fp16")];
            tensor<int32, [4]> vb_7_begin_0 = const()[name = string("vb_7_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_7_end_0 = const()[name = string("vb_7_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> vb_7_end_mask_0 = const()[name = string("vb_7_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_7_cast_fp16 = slice_by_index(begin = vb_7_begin_0, end = vb_7_end_0, end_mask = vb_7_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_7_cast_fp16")];
            tensor<int32, [3]> gdec_7_begin_0 = const()[name = string("gdec_7_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_7_end_0 = const()[name = string("gdec_7_end_0"), val = tensor<int32, [3]>([1, 4, 48])];
            tensor<bool, [3]> gdec_7_end_mask_0 = const()[name = string("gdec_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_7_cast_fp16 = slice_by_index(begin = gdec_7_begin_0, end = gdec_7_end_0, end_mask = gdec_7_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_7_cast_fp16")];
            tensor<int32, [2]> var_918 = const()[name = string("op_918"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_919_cast_fp16 = reshape(shape = var_918, x = gdec_7_cast_fp16)[name = string("op_919_cast_fp16")];
            tensor<fp16, [4, 256]> var_920_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_919_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<int32, [4]> var_925 = const()[name = string("op_925"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_5_cast_fp16 = reshape(shape = var_925, x = var_920_cast_fp16)[name = string("P_5_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_927_cast_fp16 = exp(x = P_5_cast_fp16)[name = string("op_927_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_5_cast_fp16 = mul(x = var_927_cast_fp16, y = tril_incl_to_fp16)[name = string("D_5_cast_fp16")];
            bool kk_5_transpose_x_1 = const()[name = string("kk_5_transpose_x_1"), val = bool(false)];
            bool kk_5_transpose_y_1 = const()[name = string("kk_5_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_5_cast_fp16 = matmul(transpose_x = kk_5_transpose_x_1, transpose_y = kk_5_transpose_y_1, x = kb_5_cast_fp16, y = k_13_cast_fp16)[name = string("kk_5_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_933_cast_fp16 = mul(x = kk_5_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_933_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_934_cast_fp16 = mul(x = var_933_cast_fp16, y = D_5_cast_fp16)[name = string("op_934_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_5_cast_fp16 = mul(x = var_934_cast_fp16, y = cum_tail_to_fp16)[name = string("N_5_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_936_cast_fp16 = mul(x = D_5_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_936_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_5_cast_fp16 = mul(x = kk_5_cast_fp16, y = var_936_cast_fp16)[name = string("Na_5_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_17_cast_fp16 = add(x = eye_to_fp16, y = N_5_cast_fp16)[name = string("Minv_17_cast_fp16")];
            bool Np_13_transpose_x_0 = const()[name = string("Np_13_transpose_x_0"), val = bool(false)];
            bool Np_13_transpose_y_0 = const()[name = string("Np_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_13_cast_fp16 = matmul(transpose_x = Np_13_transpose_x_0, transpose_y = Np_13_transpose_y_0, x = N_5_cast_fp16, y = Na_5_cast_fp16)[name = string("Np_13_cast_fp16")];
            bool Npa_9_transpose_x_0 = const()[name = string("Npa_9_transpose_x_0"), val = bool(false)];
            bool Npa_9_transpose_y_0 = const()[name = string("Npa_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_9_cast_fp16 = matmul(transpose_x = Npa_9_transpose_x_0, transpose_y = Npa_9_transpose_y_0, x = Na_5_cast_fp16, y = N_5_cast_fp16)[name = string("Npa_9_cast_fp16")];
            bool var_942_transpose_x_0 = const()[name = string("op_942_transpose_x_0"), val = bool(false)];
            bool var_942_transpose_y_0 = const()[name = string("op_942_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_942_cast_fp16 = matmul(transpose_x = var_942_transpose_x_0, transpose_y = var_942_transpose_y_0, x = Minv_17_cast_fp16, y = Np_13_cast_fp16)[name = string("op_942_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_19_cast_fp16 = add(x = Minv_17_cast_fp16, y = var_942_cast_fp16)[name = string("Minv_19_cast_fp16")];
            bool Np_15_transpose_x_0 = const()[name = string("Np_15_transpose_x_0"), val = bool(false)];
            bool Np_15_transpose_y_0 = const()[name = string("Np_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_15_cast_fp16 = matmul(transpose_x = Np_15_transpose_x_0, transpose_y = Np_15_transpose_y_0, x = Np_13_cast_fp16, y = Npa_9_cast_fp16)[name = string("Np_15_cast_fp16")];
            bool Npa_11_transpose_x_0 = const()[name = string("Npa_11_transpose_x_0"), val = bool(false)];
            bool Npa_11_transpose_y_0 = const()[name = string("Npa_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_11_cast_fp16 = matmul(transpose_x = Npa_11_transpose_x_0, transpose_y = Npa_11_transpose_y_0, x = Npa_9_cast_fp16, y = Np_13_cast_fp16)[name = string("Npa_11_cast_fp16")];
            bool var_947_transpose_x_0 = const()[name = string("op_947_transpose_x_0"), val = bool(false)];
            bool var_947_transpose_y_0 = const()[name = string("op_947_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_947_cast_fp16 = matmul(transpose_x = var_947_transpose_x_0, transpose_y = var_947_transpose_y_0, x = Minv_19_cast_fp16, y = Np_15_cast_fp16)[name = string("op_947_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_21_cast_fp16 = add(x = Minv_19_cast_fp16, y = var_947_cast_fp16)[name = string("Minv_21_cast_fp16")];
            bool Np_17_transpose_x_0 = const()[name = string("Np_17_transpose_x_0"), val = bool(false)];
            bool Np_17_transpose_y_0 = const()[name = string("Np_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_17_cast_fp16 = matmul(transpose_x = Np_17_transpose_x_0, transpose_y = Np_17_transpose_y_0, x = Np_15_cast_fp16, y = Npa_11_cast_fp16)[name = string("Np_17_cast_fp16")];
            bool var_951_transpose_x_0 = const()[name = string("op_951_transpose_x_0"), val = bool(false)];
            bool var_951_transpose_y_0 = const()[name = string("op_951_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_951_cast_fp16 = matmul(transpose_x = var_951_transpose_x_0, transpose_y = var_951_transpose_y_0, x = Minv_21_cast_fp16, y = Np_17_cast_fp16)[name = string("op_951_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_23_cast_fp16 = add(x = Minv_21_cast_fp16, y = var_951_cast_fp16)[name = string("Minv_23_cast_fp16")];
            bool u_5_transpose_x_0 = const()[name = string("u_5_transpose_x_0"), val = bool(false)];
            bool u_5_transpose_y_0 = const()[name = string("u_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_5_cast_fp16 = matmul(transpose_x = u_5_transpose_x_0, transpose_y = u_5_transpose_y_0, x = Minv_23_cast_fp16, y = vb_7_cast_fp16)[name = string("u_5_cast_fp16")];
            bool var_958_transpose_x_1 = const()[name = string("op_958_transpose_x_1"), val = bool(false)];
            bool var_958_transpose_y_1 = const()[name = string("op_958_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_958_cast_fp16 = matmul(transpose_x = var_958_transpose_x_1, transpose_y = var_958_transpose_y_1, x = q_13_cast_fp16, y = k_13_cast_fp16)[name = string("op_958_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_5_cast_fp16 = mul(x = var_958_cast_fp16, y = D_5_cast_fp16)[name = string("A_5_cast_fp16")];
            tensor<fp16, [16]> var_960_bias_0_to_fp16 = const()[name = string("op_960_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_960_cast_fp16 = linear(bias = var_960_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_7_cast_fp16)[name = string("op_960_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_3_cast_fp16 = exp(x = var_960_cast_fp16)[name = string("cum_3_cast_fp16")];
            tensor<int32, [1]> var_963_axes_0 = const()[name = string("op_963_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_963_cast_fp16 = expand_dims(axes = var_963_axes_0, x = cum_3_cast_fp16)[name = string("op_963_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_964_cast_fp16 = mul(x = kb_5_cast_fp16, y = var_963_cast_fp16)[name = string("op_964_cast_fp16")];
            bool w_5_transpose_x_0 = const()[name = string("w_5_transpose_x_0"), val = bool(false)];
            bool w_5_transpose_y_0 = const()[name = string("w_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_5_cast_fp16 = matmul(transpose_x = w_5_transpose_x_0, transpose_y = w_5_transpose_y_0, x = Minv_23_cast_fp16, y = var_964_cast_fp16)[name = string("w_5_cast_fp16")];
            bool var_966_transpose_x_0 = const()[name = string("op_966_transpose_x_0"), val = bool(false)];
            bool var_966_transpose_y_0 = const()[name = string("op_966_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_966_cast_fp16 = matmul(transpose_x = var_966_transpose_x_0, transpose_y = var_966_transpose_y_0, x = w_5_cast_fp16, y = S_3_cast_fp16)[name = string("op_966_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_3_cast_fp16 = sub(x = u_5_cast_fp16, y = var_966_cast_fp16)[name = string("vnew_3_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_971_cast_fp16 = mul(x = q_13_cast_fp16, y = var_963_cast_fp16)[name = string("op_971_cast_fp16")];
            bool var_972_transpose_x_0 = const()[name = string("op_972_transpose_x_0"), val = bool(false)];
            bool var_972_transpose_y_0 = const()[name = string("op_972_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_972_cast_fp16 = matmul(transpose_x = var_972_transpose_x_0, transpose_y = var_972_transpose_y_0, x = var_971_cast_fp16, y = S_3_cast_fp16)[name = string("op_972_cast_fp16")];
            bool var_973_transpose_x_0 = const()[name = string("op_973_transpose_x_0"), val = bool(false)];
            bool var_973_transpose_y_0 = const()[name = string("op_973_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_973_cast_fp16 = matmul(transpose_x = var_973_transpose_x_0, transpose_y = var_973_transpose_y_0, x = A_5_cast_fp16, y = vnew_3_cast_fp16)[name = string("op_973_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_975_cast_fp16 = add(x = var_972_cast_fp16, y = var_973_cast_fp16)[name = string("op_975_cast_fp16")];
            tensor<fp16, [16]> var_976_bias_0_to_fp16 = const()[name = string("op_976_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_976_cast_fp16 = linear(bias = var_976_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_7_cast_fp16)[name = string("op_976_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_5_cast_fp16 = exp(x = var_976_cast_fp16)[name = string("tail_5_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_9_to_fp16 = const()[name = string("transpose_9_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_978_bias_0_to_fp16 = const()[name = string("op_978_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_978_cast_fp16 = linear(bias = var_978_bias_0_to_fp16, weight = transpose_9_to_fp16, x = gdec_7_cast_fp16)[name = string("op_978_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_979_cast_fp16 = exp(x = var_978_cast_fp16)[name = string("op_979_cast_fp16")];
            tensor<int32, [1]> last_5_axes_0 = const()[name = string("last_5_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_5_cast_fp16 = expand_dims(axes = last_5_axes_0, x = var_979_cast_fp16)[name = string("last_5_cast_fp16")];
            tensor<int32, [1]> var_983_axes_0 = const()[name = string("op_983_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_983_cast_fp16 = expand_dims(axes = var_983_axes_0, x = tail_5_cast_fp16)[name = string("op_983_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_984_cast_fp16 = mul(x = k_13_cast_fp16, y = var_983_cast_fp16)[name = string("op_984_cast_fp16")];
            bool upd_3_transpose_x_1 = const()[name = string("upd_3_transpose_x_1"), val = bool(true)];
            bool upd_3_transpose_y_1 = const()[name = string("upd_3_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_3_cast_fp16 = matmul(transpose_x = upd_3_transpose_x_1, transpose_y = upd_3_transpose_y_1, x = var_984_cast_fp16, y = vnew_3_cast_fp16)[name = string("upd_3_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_989_cast_fp16 = mul(x = S_3_cast_fp16, y = last_5_cast_fp16)[name = string("op_989_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_5_cast_fp16 = add(x = var_989_cast_fp16, y = upd_3_cast_fp16)[name = string("S_5_cast_fp16")];
            tensor<int32, [4]> q_15_begin_0 = const()[name = string("q_15_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_15_end_0 = const()[name = string("q_15_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> q_15_end_mask_0 = const()[name = string("q_15_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_15_cast_fp16 = slice_by_index(begin = q_15_begin_0, end = q_15_end_0, end_mask = q_15_end_mask_0, x = q_7_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<int32, [4]> k_15_begin_0 = const()[name = string("k_15_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_15_end_0 = const()[name = string("k_15_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> k_15_end_mask_0 = const()[name = string("k_15_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_15_cast_fp16 = slice_by_index(begin = k_15_begin_0, end = k_15_end_0, end_mask = k_15_end_mask_0, x = k_7_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> kb_7_begin_0 = const()[name = string("kb_7_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_7_end_0 = const()[name = string("kb_7_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> kb_7_end_mask_0 = const()[name = string("kb_7_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_7_cast_fp16 = slice_by_index(begin = kb_7_begin_0, end = kb_7_end_0, end_mask = kb_7_end_mask_0, x = var_584_cast_fp16)[name = string("kb_7_cast_fp16")];
            tensor<int32, [4]> vb_9_begin_0 = const()[name = string("vb_9_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_9_end_0 = const()[name = string("vb_9_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> vb_9_end_mask_0 = const()[name = string("vb_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_9_cast_fp16 = slice_by_index(begin = vb_9_begin_0, end = vb_9_end_0, end_mask = vb_9_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_9_cast_fp16")];
            tensor<int32, [3]> gdec_9_begin_0 = const()[name = string("gdec_9_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_9_end_0 = const()[name = string("gdec_9_end_0"), val = tensor<int32, [3]>([1, 4, 64])];
            tensor<bool, [3]> gdec_9_end_mask_0 = const()[name = string("gdec_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_9_cast_fp16 = slice_by_index(begin = gdec_9_begin_0, end = gdec_9_end_0, end_mask = gdec_9_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_9_cast_fp16")];
            tensor<int32, [2]> var_1072 = const()[name = string("op_1072"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_1073_cast_fp16 = reshape(shape = var_1072, x = gdec_9_cast_fp16)[name = string("op_1073_cast_fp16")];
            tensor<fp16, [4, 256]> var_1074_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1073_cast_fp16)[name = string("op_1074_cast_fp16")];
            tensor<int32, [4]> var_1079 = const()[name = string("op_1079"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_7_cast_fp16 = reshape(shape = var_1079, x = var_1074_cast_fp16)[name = string("P_7_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1081_cast_fp16 = exp(x = P_7_cast_fp16)[name = string("op_1081_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_7_cast_fp16 = mul(x = var_1081_cast_fp16, y = tril_incl_to_fp16)[name = string("D_7_cast_fp16")];
            bool kk_7_transpose_x_1 = const()[name = string("kk_7_transpose_x_1"), val = bool(false)];
            bool kk_7_transpose_y_1 = const()[name = string("kk_7_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_7_cast_fp16 = matmul(transpose_x = kk_7_transpose_x_1, transpose_y = kk_7_transpose_y_1, x = kb_7_cast_fp16, y = k_15_cast_fp16)[name = string("kk_7_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_1087_cast_fp16 = mul(x = kk_7_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_1087_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1088_cast_fp16 = mul(x = var_1087_cast_fp16, y = D_7_cast_fp16)[name = string("op_1088_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_7_cast_fp16 = mul(x = var_1088_cast_fp16, y = cum_tail_to_fp16)[name = string("N_7_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1090_cast_fp16 = mul(x = D_7_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1090_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_7_cast_fp16 = mul(x = kk_7_cast_fp16, y = var_1090_cast_fp16)[name = string("Na_7_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_25_cast_fp16 = add(x = eye_to_fp16, y = N_7_cast_fp16)[name = string("Minv_25_cast_fp16")];
            bool Np_19_transpose_x_0 = const()[name = string("Np_19_transpose_x_0"), val = bool(false)];
            bool Np_19_transpose_y_0 = const()[name = string("Np_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_19_cast_fp16 = matmul(transpose_x = Np_19_transpose_x_0, transpose_y = Np_19_transpose_y_0, x = N_7_cast_fp16, y = Na_7_cast_fp16)[name = string("Np_19_cast_fp16")];
            bool Npa_13_transpose_x_0 = const()[name = string("Npa_13_transpose_x_0"), val = bool(false)];
            bool Npa_13_transpose_y_0 = const()[name = string("Npa_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_13_cast_fp16 = matmul(transpose_x = Npa_13_transpose_x_0, transpose_y = Npa_13_transpose_y_0, x = Na_7_cast_fp16, y = N_7_cast_fp16)[name = string("Npa_13_cast_fp16")];
            bool var_1096_transpose_x_0 = const()[name = string("op_1096_transpose_x_0"), val = bool(false)];
            bool var_1096_transpose_y_0 = const()[name = string("op_1096_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1096_cast_fp16 = matmul(transpose_x = var_1096_transpose_x_0, transpose_y = var_1096_transpose_y_0, x = Minv_25_cast_fp16, y = Np_19_cast_fp16)[name = string("op_1096_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_27_cast_fp16 = add(x = Minv_25_cast_fp16, y = var_1096_cast_fp16)[name = string("Minv_27_cast_fp16")];
            bool Np_21_transpose_x_0 = const()[name = string("Np_21_transpose_x_0"), val = bool(false)];
            bool Np_21_transpose_y_0 = const()[name = string("Np_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_21_cast_fp16 = matmul(transpose_x = Np_21_transpose_x_0, transpose_y = Np_21_transpose_y_0, x = Np_19_cast_fp16, y = Npa_13_cast_fp16)[name = string("Np_21_cast_fp16")];
            bool Npa_15_transpose_x_0 = const()[name = string("Npa_15_transpose_x_0"), val = bool(false)];
            bool Npa_15_transpose_y_0 = const()[name = string("Npa_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_15_cast_fp16 = matmul(transpose_x = Npa_15_transpose_x_0, transpose_y = Npa_15_transpose_y_0, x = Npa_13_cast_fp16, y = Np_19_cast_fp16)[name = string("Npa_15_cast_fp16")];
            bool var_1101_transpose_x_0 = const()[name = string("op_1101_transpose_x_0"), val = bool(false)];
            bool var_1101_transpose_y_0 = const()[name = string("op_1101_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1101_cast_fp16 = matmul(transpose_x = var_1101_transpose_x_0, transpose_y = var_1101_transpose_y_0, x = Minv_27_cast_fp16, y = Np_21_cast_fp16)[name = string("op_1101_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_29_cast_fp16 = add(x = Minv_27_cast_fp16, y = var_1101_cast_fp16)[name = string("Minv_29_cast_fp16")];
            bool Np_23_transpose_x_0 = const()[name = string("Np_23_transpose_x_0"), val = bool(false)];
            bool Np_23_transpose_y_0 = const()[name = string("Np_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_23_cast_fp16 = matmul(transpose_x = Np_23_transpose_x_0, transpose_y = Np_23_transpose_y_0, x = Np_21_cast_fp16, y = Npa_15_cast_fp16)[name = string("Np_23_cast_fp16")];
            bool var_1105_transpose_x_0 = const()[name = string("op_1105_transpose_x_0"), val = bool(false)];
            bool var_1105_transpose_y_0 = const()[name = string("op_1105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1105_cast_fp16 = matmul(transpose_x = var_1105_transpose_x_0, transpose_y = var_1105_transpose_y_0, x = Minv_29_cast_fp16, y = Np_23_cast_fp16)[name = string("op_1105_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_31_cast_fp16 = add(x = Minv_29_cast_fp16, y = var_1105_cast_fp16)[name = string("Minv_31_cast_fp16")];
            bool u_7_transpose_x_0 = const()[name = string("u_7_transpose_x_0"), val = bool(false)];
            bool u_7_transpose_y_0 = const()[name = string("u_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_7_cast_fp16 = matmul(transpose_x = u_7_transpose_x_0, transpose_y = u_7_transpose_y_0, x = Minv_31_cast_fp16, y = vb_9_cast_fp16)[name = string("u_7_cast_fp16")];
            bool var_1112_transpose_x_1 = const()[name = string("op_1112_transpose_x_1"), val = bool(false)];
            bool var_1112_transpose_y_1 = const()[name = string("op_1112_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_1112_cast_fp16 = matmul(transpose_x = var_1112_transpose_x_1, transpose_y = var_1112_transpose_y_1, x = q_15_cast_fp16, y = k_15_cast_fp16)[name = string("op_1112_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_7_cast_fp16 = mul(x = var_1112_cast_fp16, y = D_7_cast_fp16)[name = string("A_7_cast_fp16")];
            tensor<fp16, [16]> var_1114_bias_0_to_fp16 = const()[name = string("op_1114_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_1114_cast_fp16 = linear(bias = var_1114_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_9_cast_fp16)[name = string("op_1114_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_5_cast_fp16 = exp(x = var_1114_cast_fp16)[name = string("cum_5_cast_fp16")];
            tensor<int32, [1]> var_1117_axes_0 = const()[name = string("op_1117_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_1117_cast_fp16 = expand_dims(axes = var_1117_axes_0, x = cum_5_cast_fp16)[name = string("op_1117_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1118_cast_fp16 = mul(x = kb_7_cast_fp16, y = var_1117_cast_fp16)[name = string("op_1118_cast_fp16")];
            bool w_7_transpose_x_0 = const()[name = string("w_7_transpose_x_0"), val = bool(false)];
            bool w_7_transpose_y_0 = const()[name = string("w_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_7_cast_fp16 = matmul(transpose_x = w_7_transpose_x_0, transpose_y = w_7_transpose_y_0, x = Minv_31_cast_fp16, y = var_1118_cast_fp16)[name = string("w_7_cast_fp16")];
            bool var_1120_transpose_x_0 = const()[name = string("op_1120_transpose_x_0"), val = bool(false)];
            bool var_1120_transpose_y_0 = const()[name = string("op_1120_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1120_cast_fp16 = matmul(transpose_x = var_1120_transpose_x_0, transpose_y = var_1120_transpose_y_0, x = w_7_cast_fp16, y = S_5_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_5_cast_fp16 = sub(x = u_7_cast_fp16, y = var_1120_cast_fp16)[name = string("vnew_5_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1125_cast_fp16 = mul(x = q_15_cast_fp16, y = var_1117_cast_fp16)[name = string("op_1125_cast_fp16")];
            bool var_1126_transpose_x_0 = const()[name = string("op_1126_transpose_x_0"), val = bool(false)];
            bool var_1126_transpose_y_0 = const()[name = string("op_1126_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1126_cast_fp16 = matmul(transpose_x = var_1126_transpose_x_0, transpose_y = var_1126_transpose_y_0, x = var_1125_cast_fp16, y = S_5_cast_fp16)[name = string("op_1126_cast_fp16")];
            bool var_1127_transpose_x_0 = const()[name = string("op_1127_transpose_x_0"), val = bool(false)];
            bool var_1127_transpose_y_0 = const()[name = string("op_1127_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1127_cast_fp16 = matmul(transpose_x = var_1127_transpose_x_0, transpose_y = var_1127_transpose_y_0, x = A_7_cast_fp16, y = vnew_5_cast_fp16)[name = string("op_1127_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1129_cast_fp16 = add(x = var_1126_cast_fp16, y = var_1127_cast_fp16)[name = string("op_1129_cast_fp16")];
            tensor<fp16, [16]> var_1130_bias_0_to_fp16 = const()[name = string("op_1130_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_1130_cast_fp16 = linear(bias = var_1130_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_9_cast_fp16)[name = string("op_1130_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_7_cast_fp16 = exp(x = var_1130_cast_fp16)[name = string("tail_7_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_13_to_fp16 = const()[name = string("transpose_13_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_1132_bias_0_to_fp16 = const()[name = string("op_1132_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_1132_cast_fp16 = linear(bias = var_1132_bias_0_to_fp16, weight = transpose_13_to_fp16, x = gdec_9_cast_fp16)[name = string("op_1132_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_1133_cast_fp16 = exp(x = var_1132_cast_fp16)[name = string("op_1133_cast_fp16")];
            tensor<int32, [1]> last_7_axes_0 = const()[name = string("last_7_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_7_cast_fp16 = expand_dims(axes = last_7_axes_0, x = var_1133_cast_fp16)[name = string("last_7_cast_fp16")];
            tensor<int32, [1]> var_1137_axes_0 = const()[name = string("op_1137_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_1137_cast_fp16 = expand_dims(axes = var_1137_axes_0, x = tail_7_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1138_cast_fp16 = mul(x = k_15_cast_fp16, y = var_1137_cast_fp16)[name = string("op_1138_cast_fp16")];
            bool upd_5_transpose_x_1 = const()[name = string("upd_5_transpose_x_1"), val = bool(true)];
            bool upd_5_transpose_y_1 = const()[name = string("upd_5_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_5_cast_fp16 = matmul(transpose_x = upd_5_transpose_x_1, transpose_y = upd_5_transpose_y_1, x = var_1138_cast_fp16, y = vnew_5_cast_fp16)[name = string("upd_5_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_1143_cast_fp16 = mul(x = S_5_cast_fp16, y = last_7_cast_fp16)[name = string("op_1143_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_7_cast_fp16 = add(x = var_1143_cast_fp16, y = upd_5_cast_fp16)[name = string("S_7_cast_fp16")];
            tensor<int32, [4]> q_17_begin_0 = const()[name = string("q_17_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> q_17_end_0 = const()[name = string("q_17_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> q_17_end_mask_0 = const()[name = string("q_17_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_17_cast_fp16 = slice_by_index(begin = q_17_begin_0, end = q_17_end_0, end_mask = q_17_end_mask_0, x = q_7_cast_fp16)[name = string("q_17_cast_fp16")];
            tensor<int32, [4]> k_17_begin_0 = const()[name = string("k_17_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> k_17_end_0 = const()[name = string("k_17_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> k_17_end_mask_0 = const()[name = string("k_17_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, x = k_7_cast_fp16)[name = string("k_17_cast_fp16")];
            tensor<int32, [4]> kb_9_begin_0 = const()[name = string("kb_9_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> kb_9_end_0 = const()[name = string("kb_9_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> kb_9_end_mask_0 = const()[name = string("kb_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_9_cast_fp16 = slice_by_index(begin = kb_9_begin_0, end = kb_9_end_0, end_mask = kb_9_end_mask_0, x = var_584_cast_fp16)[name = string("kb_9_cast_fp16")];
            tensor<int32, [4]> vb_11_begin_0 = const()[name = string("vb_11_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> vb_11_end_0 = const()[name = string("vb_11_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> vb_11_end_mask_0 = const()[name = string("vb_11_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_11_cast_fp16 = slice_by_index(begin = vb_11_begin_0, end = vb_11_end_0, end_mask = vb_11_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_11_cast_fp16")];
            tensor<int32, [3]> gdec_11_begin_0 = const()[name = string("gdec_11_begin_0"), val = tensor<int32, [3]>([0, 0, 64])];
            tensor<int32, [3]> gdec_11_end_0 = const()[name = string("gdec_11_end_0"), val = tensor<int32, [3]>([1, 4, 80])];
            tensor<bool, [3]> gdec_11_end_mask_0 = const()[name = string("gdec_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_11_cast_fp16 = slice_by_index(begin = gdec_11_begin_0, end = gdec_11_end_0, end_mask = gdec_11_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_11_cast_fp16")];
            tensor<int32, [2]> var_1226 = const()[name = string("op_1226"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_1227_cast_fp16 = reshape(shape = var_1226, x = gdec_11_cast_fp16)[name = string("op_1227_cast_fp16")];
            tensor<fp16, [4, 256]> var_1228_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1227_cast_fp16)[name = string("op_1228_cast_fp16")];
            tensor<int32, [4]> var_1233 = const()[name = string("op_1233"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_9_cast_fp16 = reshape(shape = var_1233, x = var_1228_cast_fp16)[name = string("P_9_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1235_cast_fp16 = exp(x = P_9_cast_fp16)[name = string("op_1235_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_9_cast_fp16 = mul(x = var_1235_cast_fp16, y = tril_incl_to_fp16)[name = string("D_9_cast_fp16")];
            bool kk_9_transpose_x_1 = const()[name = string("kk_9_transpose_x_1"), val = bool(false)];
            bool kk_9_transpose_y_1 = const()[name = string("kk_9_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_9_cast_fp16 = matmul(transpose_x = kk_9_transpose_x_1, transpose_y = kk_9_transpose_y_1, x = kb_9_cast_fp16, y = k_17_cast_fp16)[name = string("kk_9_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_1241_cast_fp16 = mul(x = kk_9_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_1241_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1242_cast_fp16 = mul(x = var_1241_cast_fp16, y = D_9_cast_fp16)[name = string("op_1242_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_9_cast_fp16 = mul(x = var_1242_cast_fp16, y = cum_tail_to_fp16)[name = string("N_9_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1244_cast_fp16 = mul(x = D_9_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1244_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_9_cast_fp16 = mul(x = kk_9_cast_fp16, y = var_1244_cast_fp16)[name = string("Na_9_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_33_cast_fp16 = add(x = eye_to_fp16, y = N_9_cast_fp16)[name = string("Minv_33_cast_fp16")];
            bool Np_25_transpose_x_0 = const()[name = string("Np_25_transpose_x_0"), val = bool(false)];
            bool Np_25_transpose_y_0 = const()[name = string("Np_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_25_cast_fp16 = matmul(transpose_x = Np_25_transpose_x_0, transpose_y = Np_25_transpose_y_0, x = N_9_cast_fp16, y = Na_9_cast_fp16)[name = string("Np_25_cast_fp16")];
            bool Npa_17_transpose_x_0 = const()[name = string("Npa_17_transpose_x_0"), val = bool(false)];
            bool Npa_17_transpose_y_0 = const()[name = string("Npa_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_17_cast_fp16 = matmul(transpose_x = Npa_17_transpose_x_0, transpose_y = Npa_17_transpose_y_0, x = Na_9_cast_fp16, y = N_9_cast_fp16)[name = string("Npa_17_cast_fp16")];
            bool var_1250_transpose_x_0 = const()[name = string("op_1250_transpose_x_0"), val = bool(false)];
            bool var_1250_transpose_y_0 = const()[name = string("op_1250_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1250_cast_fp16 = matmul(transpose_x = var_1250_transpose_x_0, transpose_y = var_1250_transpose_y_0, x = Minv_33_cast_fp16, y = Np_25_cast_fp16)[name = string("op_1250_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_35_cast_fp16 = add(x = Minv_33_cast_fp16, y = var_1250_cast_fp16)[name = string("Minv_35_cast_fp16")];
            bool Np_27_transpose_x_0 = const()[name = string("Np_27_transpose_x_0"), val = bool(false)];
            bool Np_27_transpose_y_0 = const()[name = string("Np_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_27_cast_fp16 = matmul(transpose_x = Np_27_transpose_x_0, transpose_y = Np_27_transpose_y_0, x = Np_25_cast_fp16, y = Npa_17_cast_fp16)[name = string("Np_27_cast_fp16")];
            bool Npa_19_transpose_x_0 = const()[name = string("Npa_19_transpose_x_0"), val = bool(false)];
            bool Npa_19_transpose_y_0 = const()[name = string("Npa_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_19_cast_fp16 = matmul(transpose_x = Npa_19_transpose_x_0, transpose_y = Npa_19_transpose_y_0, x = Npa_17_cast_fp16, y = Np_25_cast_fp16)[name = string("Npa_19_cast_fp16")];
            bool var_1255_transpose_x_0 = const()[name = string("op_1255_transpose_x_0"), val = bool(false)];
            bool var_1255_transpose_y_0 = const()[name = string("op_1255_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1255_cast_fp16 = matmul(transpose_x = var_1255_transpose_x_0, transpose_y = var_1255_transpose_y_0, x = Minv_35_cast_fp16, y = Np_27_cast_fp16)[name = string("op_1255_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_37_cast_fp16 = add(x = Minv_35_cast_fp16, y = var_1255_cast_fp16)[name = string("Minv_37_cast_fp16")];
            bool Np_29_transpose_x_0 = const()[name = string("Np_29_transpose_x_0"), val = bool(false)];
            bool Np_29_transpose_y_0 = const()[name = string("Np_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_29_cast_fp16 = matmul(transpose_x = Np_29_transpose_x_0, transpose_y = Np_29_transpose_y_0, x = Np_27_cast_fp16, y = Npa_19_cast_fp16)[name = string("Np_29_cast_fp16")];
            bool var_1259_transpose_x_0 = const()[name = string("op_1259_transpose_x_0"), val = bool(false)];
            bool var_1259_transpose_y_0 = const()[name = string("op_1259_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1259_cast_fp16 = matmul(transpose_x = var_1259_transpose_x_0, transpose_y = var_1259_transpose_y_0, x = Minv_37_cast_fp16, y = Np_29_cast_fp16)[name = string("op_1259_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_39_cast_fp16 = add(x = Minv_37_cast_fp16, y = var_1259_cast_fp16)[name = string("Minv_39_cast_fp16")];
            bool u_9_transpose_x_0 = const()[name = string("u_9_transpose_x_0"), val = bool(false)];
            bool u_9_transpose_y_0 = const()[name = string("u_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_9_cast_fp16 = matmul(transpose_x = u_9_transpose_x_0, transpose_y = u_9_transpose_y_0, x = Minv_39_cast_fp16, y = vb_11_cast_fp16)[name = string("u_9_cast_fp16")];
            bool var_1266_transpose_x_1 = const()[name = string("op_1266_transpose_x_1"), val = bool(false)];
            bool var_1266_transpose_y_1 = const()[name = string("op_1266_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_1266_cast_fp16 = matmul(transpose_x = var_1266_transpose_x_1, transpose_y = var_1266_transpose_y_1, x = q_17_cast_fp16, y = k_17_cast_fp16)[name = string("op_1266_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_9_cast_fp16 = mul(x = var_1266_cast_fp16, y = D_9_cast_fp16)[name = string("A_9_cast_fp16")];
            tensor<fp16, [16]> var_1268_bias_0_to_fp16 = const()[name = string("op_1268_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_1268_cast_fp16 = linear(bias = var_1268_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_11_cast_fp16)[name = string("op_1268_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_7_cast_fp16 = exp(x = var_1268_cast_fp16)[name = string("cum_7_cast_fp16")];
            tensor<int32, [1]> var_1271_axes_0 = const()[name = string("op_1271_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_1271_cast_fp16 = expand_dims(axes = var_1271_axes_0, x = cum_7_cast_fp16)[name = string("op_1271_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1272_cast_fp16 = mul(x = kb_9_cast_fp16, y = var_1271_cast_fp16)[name = string("op_1272_cast_fp16")];
            bool w_9_transpose_x_0 = const()[name = string("w_9_transpose_x_0"), val = bool(false)];
            bool w_9_transpose_y_0 = const()[name = string("w_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_9_cast_fp16 = matmul(transpose_x = w_9_transpose_x_0, transpose_y = w_9_transpose_y_0, x = Minv_39_cast_fp16, y = var_1272_cast_fp16)[name = string("w_9_cast_fp16")];
            bool var_1274_transpose_x_0 = const()[name = string("op_1274_transpose_x_0"), val = bool(false)];
            bool var_1274_transpose_y_0 = const()[name = string("op_1274_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1274_cast_fp16 = matmul(transpose_x = var_1274_transpose_x_0, transpose_y = var_1274_transpose_y_0, x = w_9_cast_fp16, y = S_7_cast_fp16)[name = string("op_1274_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_7_cast_fp16 = sub(x = u_9_cast_fp16, y = var_1274_cast_fp16)[name = string("vnew_7_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1279_cast_fp16 = mul(x = q_17_cast_fp16, y = var_1271_cast_fp16)[name = string("op_1279_cast_fp16")];
            bool var_1280_transpose_x_0 = const()[name = string("op_1280_transpose_x_0"), val = bool(false)];
            bool var_1280_transpose_y_0 = const()[name = string("op_1280_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1280_cast_fp16 = matmul(transpose_x = var_1280_transpose_x_0, transpose_y = var_1280_transpose_y_0, x = var_1279_cast_fp16, y = S_7_cast_fp16)[name = string("op_1280_cast_fp16")];
            bool var_1281_transpose_x_0 = const()[name = string("op_1281_transpose_x_0"), val = bool(false)];
            bool var_1281_transpose_y_0 = const()[name = string("op_1281_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1281_cast_fp16 = matmul(transpose_x = var_1281_transpose_x_0, transpose_y = var_1281_transpose_y_0, x = A_9_cast_fp16, y = vnew_7_cast_fp16)[name = string("op_1281_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1283_cast_fp16 = add(x = var_1280_cast_fp16, y = var_1281_cast_fp16)[name = string("op_1283_cast_fp16")];
            tensor<fp16, [16]> var_1284_bias_0_to_fp16 = const()[name = string("op_1284_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_1284_cast_fp16 = linear(bias = var_1284_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_11_cast_fp16)[name = string("op_1284_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_9_cast_fp16 = exp(x = var_1284_cast_fp16)[name = string("tail_9_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_17_to_fp16 = const()[name = string("transpose_17_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_1286_bias_0_to_fp16 = const()[name = string("op_1286_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_1286_cast_fp16 = linear(bias = var_1286_bias_0_to_fp16, weight = transpose_17_to_fp16, x = gdec_11_cast_fp16)[name = string("op_1286_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_1287_cast_fp16 = exp(x = var_1286_cast_fp16)[name = string("op_1287_cast_fp16")];
            tensor<int32, [1]> last_9_axes_0 = const()[name = string("last_9_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_9_cast_fp16 = expand_dims(axes = last_9_axes_0, x = var_1287_cast_fp16)[name = string("last_9_cast_fp16")];
            tensor<int32, [1]> var_1291_axes_0 = const()[name = string("op_1291_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_1291_cast_fp16 = expand_dims(axes = var_1291_axes_0, x = tail_9_cast_fp16)[name = string("op_1291_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1292_cast_fp16 = mul(x = k_17_cast_fp16, y = var_1291_cast_fp16)[name = string("op_1292_cast_fp16")];
            bool upd_7_transpose_x_1 = const()[name = string("upd_7_transpose_x_1"), val = bool(true)];
            bool upd_7_transpose_y_1 = const()[name = string("upd_7_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_7_cast_fp16 = matmul(transpose_x = upd_7_transpose_x_1, transpose_y = upd_7_transpose_y_1, x = var_1292_cast_fp16, y = vnew_7_cast_fp16)[name = string("upd_7_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_1297_cast_fp16 = mul(x = S_7_cast_fp16, y = last_9_cast_fp16)[name = string("op_1297_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_9_cast_fp16 = add(x = var_1297_cast_fp16, y = upd_7_cast_fp16)[name = string("S_9_cast_fp16")];
            tensor<int32, [4]> q_19_begin_0 = const()[name = string("q_19_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> q_19_end_0 = const()[name = string("q_19_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_19_end_mask_0 = const()[name = string("q_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, x = q_7_cast_fp16)[name = string("q_19_cast_fp16")];
            tensor<int32, [4]> k_19_begin_0 = const()[name = string("k_19_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> k_19_end_0 = const()[name = string("k_19_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_19_end_mask_0 = const()[name = string("k_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_19_cast_fp16 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, x = k_7_cast_fp16)[name = string("k_19_cast_fp16")];
            tensor<int32, [4]> kb_11_begin_0 = const()[name = string("kb_11_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> kb_11_end_0 = const()[name = string("kb_11_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_11_end_mask_0 = const()[name = string("kb_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_11_cast_fp16 = slice_by_index(begin = kb_11_begin_0, end = kb_11_end_0, end_mask = kb_11_end_mask_0, x = var_584_cast_fp16)[name = string("kb_11_cast_fp16")];
            tensor<int32, [4]> vb_13_begin_0 = const()[name = string("vb_13_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> vb_13_end_0 = const()[name = string("vb_13_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_13_end_mask_0 = const()[name = string("vb_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_13_cast_fp16 = slice_by_index(begin = vb_13_begin_0, end = vb_13_end_0, end_mask = vb_13_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_13_cast_fp16")];
            tensor<int32, [3]> gdec_13_begin_0 = const()[name = string("gdec_13_begin_0"), val = tensor<int32, [3]>([0, 0, 80])];
            tensor<int32, [3]> gdec_13_end_0 = const()[name = string("gdec_13_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_13_end_mask_0 = const()[name = string("gdec_13_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_13_cast_fp16 = slice_by_index(begin = gdec_13_begin_0, end = gdec_13_end_0, end_mask = gdec_13_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_13_cast_fp16")];
            tensor<int32, [2]> var_1380 = const()[name = string("op_1380"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_1381_cast_fp16 = reshape(shape = var_1380, x = gdec_13_cast_fp16)[name = string("op_1381_cast_fp16")];
            tensor<fp16, [4, 256]> var_1382_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1381_cast_fp16)[name = string("op_1382_cast_fp16")];
            tensor<int32, [4]> var_1387 = const()[name = string("op_1387"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_11_cast_fp16 = reshape(shape = var_1387, x = var_1382_cast_fp16)[name = string("P_11_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1389_cast_fp16 = exp(x = P_11_cast_fp16)[name = string("op_1389_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_11_cast_fp16 = mul(x = var_1389_cast_fp16, y = tril_incl_to_fp16)[name = string("D_11_cast_fp16")];
            bool kk_11_transpose_x_1 = const()[name = string("kk_11_transpose_x_1"), val = bool(false)];
            bool kk_11_transpose_y_1 = const()[name = string("kk_11_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_11_cast_fp16 = matmul(transpose_x = kk_11_transpose_x_1, transpose_y = kk_11_transpose_y_1, x = kb_11_cast_fp16, y = k_19_cast_fp16)[name = string("kk_11_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_1395_cast_fp16 = mul(x = kk_11_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_1395_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1396_cast_fp16 = mul(x = var_1395_cast_fp16, y = D_11_cast_fp16)[name = string("op_1396_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_11_cast_fp16 = mul(x = var_1396_cast_fp16, y = cum_tail_to_fp16)[name = string("N_11_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1398_cast_fp16 = mul(x = D_11_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1398_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_11_cast_fp16 = mul(x = kk_11_cast_fp16, y = var_1398_cast_fp16)[name = string("Na_11_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_41_cast_fp16 = add(x = eye_to_fp16, y = N_11_cast_fp16)[name = string("Minv_41_cast_fp16")];
            bool Np_31_transpose_x_0 = const()[name = string("Np_31_transpose_x_0"), val = bool(false)];
            bool Np_31_transpose_y_0 = const()[name = string("Np_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_31_cast_fp16 = matmul(transpose_x = Np_31_transpose_x_0, transpose_y = Np_31_transpose_y_0, x = N_11_cast_fp16, y = Na_11_cast_fp16)[name = string("Np_31_cast_fp16")];
            bool Npa_21_transpose_x_0 = const()[name = string("Npa_21_transpose_x_0"), val = bool(false)];
            bool Npa_21_transpose_y_0 = const()[name = string("Npa_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_21_cast_fp16 = matmul(transpose_x = Npa_21_transpose_x_0, transpose_y = Npa_21_transpose_y_0, x = Na_11_cast_fp16, y = N_11_cast_fp16)[name = string("Npa_21_cast_fp16")];
            bool var_1404_transpose_x_0 = const()[name = string("op_1404_transpose_x_0"), val = bool(false)];
            bool var_1404_transpose_y_0 = const()[name = string("op_1404_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1404_cast_fp16 = matmul(transpose_x = var_1404_transpose_x_0, transpose_y = var_1404_transpose_y_0, x = Minv_41_cast_fp16, y = Np_31_cast_fp16)[name = string("op_1404_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_43_cast_fp16 = add(x = Minv_41_cast_fp16, y = var_1404_cast_fp16)[name = string("Minv_43_cast_fp16")];
            bool Np_33_transpose_x_0 = const()[name = string("Np_33_transpose_x_0"), val = bool(false)];
            bool Np_33_transpose_y_0 = const()[name = string("Np_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_33_cast_fp16 = matmul(transpose_x = Np_33_transpose_x_0, transpose_y = Np_33_transpose_y_0, x = Np_31_cast_fp16, y = Npa_21_cast_fp16)[name = string("Np_33_cast_fp16")];
            bool Npa_23_transpose_x_0 = const()[name = string("Npa_23_transpose_x_0"), val = bool(false)];
            bool Npa_23_transpose_y_0 = const()[name = string("Npa_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_23_cast_fp16 = matmul(transpose_x = Npa_23_transpose_x_0, transpose_y = Npa_23_transpose_y_0, x = Npa_21_cast_fp16, y = Np_31_cast_fp16)[name = string("Npa_23_cast_fp16")];
            bool var_1409_transpose_x_0 = const()[name = string("op_1409_transpose_x_0"), val = bool(false)];
            bool var_1409_transpose_y_0 = const()[name = string("op_1409_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1409_cast_fp16 = matmul(transpose_x = var_1409_transpose_x_0, transpose_y = var_1409_transpose_y_0, x = Minv_43_cast_fp16, y = Np_33_cast_fp16)[name = string("op_1409_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_45_cast_fp16 = add(x = Minv_43_cast_fp16, y = var_1409_cast_fp16)[name = string("Minv_45_cast_fp16")];
            bool Np_35_transpose_x_0 = const()[name = string("Np_35_transpose_x_0"), val = bool(false)];
            bool Np_35_transpose_y_0 = const()[name = string("Np_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_35_cast_fp16 = matmul(transpose_x = Np_35_transpose_x_0, transpose_y = Np_35_transpose_y_0, x = Np_33_cast_fp16, y = Npa_23_cast_fp16)[name = string("Np_35_cast_fp16")];
            bool var_1413_transpose_x_0 = const()[name = string("op_1413_transpose_x_0"), val = bool(false)];
            bool var_1413_transpose_y_0 = const()[name = string("op_1413_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1413_cast_fp16 = matmul(transpose_x = var_1413_transpose_x_0, transpose_y = var_1413_transpose_y_0, x = Minv_45_cast_fp16, y = Np_35_cast_fp16)[name = string("op_1413_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_47_cast_fp16 = add(x = Minv_45_cast_fp16, y = var_1413_cast_fp16)[name = string("Minv_47_cast_fp16")];
            bool u_11_transpose_x_0 = const()[name = string("u_11_transpose_x_0"), val = bool(false)];
            bool u_11_transpose_y_0 = const()[name = string("u_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_11_cast_fp16 = matmul(transpose_x = u_11_transpose_x_0, transpose_y = u_11_transpose_y_0, x = Minv_47_cast_fp16, y = vb_13_cast_fp16)[name = string("u_11_cast_fp16")];
            bool var_1420_transpose_x_1 = const()[name = string("op_1420_transpose_x_1"), val = bool(false)];
            bool var_1420_transpose_y_1 = const()[name = string("op_1420_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_1420_cast_fp16 = matmul(transpose_x = var_1420_transpose_x_1, transpose_y = var_1420_transpose_y_1, x = q_19_cast_fp16, y = k_19_cast_fp16)[name = string("op_1420_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_11_cast_fp16 = mul(x = var_1420_cast_fp16, y = D_11_cast_fp16)[name = string("A_11_cast_fp16")];
            tensor<fp16, [16]> var_1422_bias_0_to_fp16 = const()[name = string("op_1422_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_1422_cast_fp16 = linear(bias = var_1422_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_13_cast_fp16)[name = string("op_1422_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_9_cast_fp16 = exp(x = var_1422_cast_fp16)[name = string("cum_9_cast_fp16")];
            tensor<int32, [1]> var_1425_axes_0 = const()[name = string("op_1425_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_1425_cast_fp16 = expand_dims(axes = var_1425_axes_0, x = cum_9_cast_fp16)[name = string("op_1425_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1426_cast_fp16 = mul(x = kb_11_cast_fp16, y = var_1425_cast_fp16)[name = string("op_1426_cast_fp16")];
            bool w_11_transpose_x_0 = const()[name = string("w_11_transpose_x_0"), val = bool(false)];
            bool w_11_transpose_y_0 = const()[name = string("w_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_11_cast_fp16 = matmul(transpose_x = w_11_transpose_x_0, transpose_y = w_11_transpose_y_0, x = Minv_47_cast_fp16, y = var_1426_cast_fp16)[name = string("w_11_cast_fp16")];
            bool var_1428_transpose_x_0 = const()[name = string("op_1428_transpose_x_0"), val = bool(false)];
            bool var_1428_transpose_y_0 = const()[name = string("op_1428_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1428_cast_fp16 = matmul(transpose_x = var_1428_transpose_x_0, transpose_y = var_1428_transpose_y_0, x = w_11_cast_fp16, y = S_9_cast_fp16)[name = string("op_1428_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_9_cast_fp16 = sub(x = u_11_cast_fp16, y = var_1428_cast_fp16)[name = string("vnew_9_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1433_cast_fp16 = mul(x = q_19_cast_fp16, y = var_1425_cast_fp16)[name = string("op_1433_cast_fp16")];
            bool var_1434_transpose_x_0 = const()[name = string("op_1434_transpose_x_0"), val = bool(false)];
            bool var_1434_transpose_y_0 = const()[name = string("op_1434_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1434_cast_fp16 = matmul(transpose_x = var_1434_transpose_x_0, transpose_y = var_1434_transpose_y_0, x = var_1433_cast_fp16, y = S_9_cast_fp16)[name = string("op_1434_cast_fp16")];
            bool var_1435_transpose_x_0 = const()[name = string("op_1435_transpose_x_0"), val = bool(false)];
            bool var_1435_transpose_y_0 = const()[name = string("op_1435_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1435_cast_fp16 = matmul(transpose_x = var_1435_transpose_x_0, transpose_y = var_1435_transpose_y_0, x = A_11_cast_fp16, y = vnew_9_cast_fp16)[name = string("op_1435_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_1_cast_fp16 = add(x = var_1434_cast_fp16, y = var_1435_cast_fp16)[name = string("oc_1_cast_fp16")];
            int32 var_1439 = const()[name = string("op_1439"), val = int32(2)];
            bool o_1_interleave_0 = const()[name = string("o_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 96, 128]> o_1_cast_fp16 = concat(axis = var_1439, interleave = o_1_interleave_0, values = (var_674_cast_fp16, var_821_cast_fp16, var_975_cast_fp16, var_1129_cast_fp16, var_1283_cast_fp16, oc_1_cast_fp16))[name = string("o_1_cast_fp16")];
            tensor<int32, [4]> var_1445 = const()[name = string("op_1445"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1451 = const()[name = string("op_1451"), val = tensor<int32, [4]>([1, 96, 4, 128])];
            tensor<fp16, [1, 96, 4, 128]> input_7_cast_fp16 = reshape(shape = var_1451, x = z_1_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> o_3_cast_fp16 = transpose(perm = var_1445, x = o_1_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 96, 4, 128]> var_1453_cast_fp16 = mul(x = o_3_cast_fp16, y = o_3_cast_fp16)[name = string("op_1453_cast_fp16")];
            tensor<int32, [1]> var_1458_axes_0 = const()[name = string("op_1458_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1458_keep_dims_0 = const()[name = string("op_1458_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 4, 1]> var_1458_cast_fp16 = reduce_mean(axes = var_1458_axes_0, keep_dims = var_1458_keep_dims_0, x = var_1453_cast_fp16)[name = string("op_1458_cast_fp16")];
            fp16 var_1460_to_fp16 = const()[name = string("op_1460_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 96, 4, 1]> var_1461_cast_fp16 = add(x = var_1458_cast_fp16, y = var_1460_to_fp16)[name = string("op_1461_cast_fp16")];
            fp32 var_1462_epsilon_0 = const()[name = string("op_1462_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 4, 1]> var_1462_cast_fp16 = rsqrt(epsilon = var_1462_epsilon_0, x = var_1461_cast_fp16)[name = string("op_1462_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> var_1463_cast_fp16 = mul(x = o_3_cast_fp16, y = var_1462_cast_fp16)[name = string("op_1463_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> var_1465_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1465_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> o_5_cast_fp16 = mul(x = var_1463_cast_fp16, y = var_1465_cast_fp16)[name = string("o_5_cast_fp16")];
            tensor<int32, [3]> var_1470 = const()[name = string("op_1470"), val = tensor<int32, [3]>([1, 96, 512])];
            tensor<fp16, [1, 96, 512]> input_9_cast_fp16 = reshape(shape = var_1470, x = o_5_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [640, 512]> layers_0_out_weight_to_fp16 = const()[name = string("layers_0_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22499136)))];
            tensor<fp16, [640]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23154560)))];
            tensor<fp16, [1, 96, 640]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_out_weight_to_fp16, x = input_9_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [1, 96, 640]> x_5_cast_fp16 = add(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = linear_1_cast_fp16)[name = string("x_5_cast_fp16")];
            fp16 var_1483_to_fp16 = const()[name = string("op_1483_to_fp16"), val = fp16(0x1p+5)];
            tensor<fp16, [1, 96, 640]> x_7_cast_fp16 = mul(x = x_5_cast_fp16, y = var_1483_to_fp16)[name = string("x_7_cast_fp16")];
            fp16 var_1481_promoted_to_fp16 = const()[name = string("op_1481_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> var_1485_cast_fp16 = pow(x = x_7_cast_fp16, y = var_1481_promoted_to_fp16)[name = string("op_1485_cast_fp16")];
            tensor<int32, [1]> var_1487_axes_0 = const()[name = string("op_1487_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1487_keep_dims_0 = const()[name = string("op_1487_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 1]> var_1487_cast_fp16 = reduce_mean(axes = var_1487_axes_0, keep_dims = var_1487_keep_dims_0, x = var_1485_cast_fp16)[name = string("op_1487_cast_fp16")];
            fp16 var_1488_to_fp16 = const()[name = string("op_1488_to_fp16"), val = fp16(0x1.0c8p-10)];
            tensor<fp16, [1, 96, 1]> var_1489_cast_fp16 = add(x = var_1487_cast_fp16, y = var_1488_to_fp16)[name = string("op_1489_cast_fp16")];
            fp32 var_1490_epsilon_0 = const()[name = string("op_1490_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 1]> var_1490_cast_fp16 = rsqrt(epsilon = var_1490_epsilon_0, x = var_1489_cast_fp16)[name = string("op_1490_cast_fp16")];
            tensor<fp16, [1, 96, 640]> var_1491_cast_fp16 = mul(x = x_7_cast_fp16, y = var_1490_cast_fp16)[name = string("op_1491_cast_fp16")];
            tensor<fp16, [640]> layers_0_n2_w1_to_fp16 = const()[name = string("layers_0_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23155904)))];
            tensor<fp16, [1, 96, 640]> input_11_cast_fp16 = mul(x = var_1491_cast_fp16, y = layers_0_n2_w1_to_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_0_gate_up_weight_to_fp16 = const()[name = string("layers_0_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23157248)))];
            tensor<fp16, [5120]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29710912)))];
            tensor<fp16, [1, 96, 5120]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_0_gate_up_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [3]> input_13_begin_0 = const()[name = string("input_13_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_13_end_0 = const()[name = string("input_13_end_0"), val = tensor<int32, [3]>([1, 96, 2560])];
            tensor<bool, [3]> input_13_end_mask_0 = const()[name = string("input_13_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 2560]> input_13_cast_fp16 = slice_by_index(begin = input_13_begin_0, end = input_13_end_0, end_mask = input_13_end_mask_0, x = linear_2_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1, 96, 2560]> var_1501_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_1501_cast_fp16")];
            tensor<int32, [3]> var_1506_begin_0 = const()[name = string("op_1506_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_1506_end_0 = const()[name = string("op_1506_end_0"), val = tensor<int32, [3]>([1, 96, 5120])];
            tensor<bool, [3]> var_1506_end_mask_0 = const()[name = string("op_1506_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 2560]> var_1506_cast_fp16 = slice_by_index(begin = var_1506_begin_0, end = var_1506_end_0, end_mask = var_1506_end_mask_0, x = linear_2_cast_fp16)[name = string("op_1506_cast_fp16")];
            tensor<fp16, [1, 96, 2560]> input_15_cast_fp16 = mul(x = var_1501_cast_fp16, y = var_1506_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_0_down_weight_to_fp16 = const()[name = string("layers_0_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29721216)))];
            tensor<fp16, [1, 96, 640]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_down_weight_to_fp16, x = input_15_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 96, 640]> x_9_cast_fp16 = add(x = x_5_cast_fp16, y = linear_3_cast_fp16)[name = string("x_9_cast_fp16")];
            fp16 var_1519_to_fp16 = const()[name = string("op_1519_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> x_11_cast_fp16 = mul(x = x_9_cast_fp16, y = var_1519_to_fp16)[name = string("x_11_cast_fp16")];
            fp16 var_1517_promoted_to_fp16 = const()[name = string("op_1517_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> var_1521_cast_fp16 = pow(x = x_11_cast_fp16, y = var_1517_promoted_to_fp16)[name = string("op_1521_cast_fp16")];
            tensor<int32, [1]> var_1523_axes_0 = const()[name = string("op_1523_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1523_keep_dims_0 = const()[name = string("op_1523_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 1]> var_1523_cast_fp16 = reduce_mean(axes = var_1523_axes_0, keep_dims = var_1523_keep_dims_0, x = var_1521_cast_fp16)[name = string("op_1523_cast_fp16")];
            fp16 var_1524_to_fp16 = const()[name = string("op_1524_to_fp16"), val = fp16(0x1.0cp-18)];
            tensor<fp16, [1, 96, 1]> var_1525_cast_fp16 = add(x = var_1523_cast_fp16, y = var_1524_to_fp16)[name = string("op_1525_cast_fp16")];
            fp32 var_1526_epsilon_0 = const()[name = string("op_1526_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 1]> var_1526_cast_fp16 = rsqrt(epsilon = var_1526_epsilon_0, x = var_1525_cast_fp16)[name = string("op_1526_cast_fp16")];
            tensor<fp16, [1, 96, 640]> var_1527_cast_fp16 = mul(x = x_11_cast_fp16, y = var_1526_cast_fp16)[name = string("op_1527_cast_fp16")];
            tensor<fp16, [640]> layers_1_n1_w1_to_fp16 = const()[name = string("layers_1_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32998080)))];
            tensor<fp16, [1, 96, 640]> input_17_cast_fp16 = mul(x = var_1527_cast_fp16, y = layers_1_n1_w1_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_1_qkvzba_weight_to_fp16 = const()[name = string("layers_1_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32999424)))];
            tensor<fp16, [1, 96, 1544]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_qkvzba_weight_to_fp16, x = input_17_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<int32, [3]> qkv_3_begin_0 = const()[name = string("qkv_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_3_end_0 = const()[name = string("qkv_3_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> qkv_3_end_mask_0 = const()[name = string("qkv_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 1024]> qkv_3_cast_fp16 = slice_by_index(begin = qkv_3_begin_0, end = qkv_3_end_0, end_mask = qkv_3_end_mask_0, x = linear_4_cast_fp16)[name = string("qkv_3_cast_fp16")];
            tensor<int32, [3]> z_3_begin_0 = const()[name = string("z_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_3_end_0 = const()[name = string("z_3_end_0"), val = tensor<int32, [3]>([1, 96, 1536])];
            tensor<bool, [3]> z_3_end_mask_0 = const()[name = string("z_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 512]> z_3_cast_fp16 = slice_by_index(begin = z_3_begin_0, end = z_3_end_0, end_mask = z_3_end_mask_0, x = linear_4_cast_fp16)[name = string("z_3_cast_fp16")];
            tensor<int32, [3]> b_3_begin_0 = const()[name = string("b_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_3_end_0 = const()[name = string("b_3_end_0"), val = tensor<int32, [3]>([1, 96, 1540])];
            tensor<bool, [3]> b_3_end_mask_0 = const()[name = string("b_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 4]> b_3_cast_fp16 = slice_by_index(begin = b_3_begin_0, end = b_3_end_0, end_mask = b_3_end_mask_0, x = linear_4_cast_fp16)[name = string("b_3_cast_fp16")];
            tensor<int32, [3]> a_3_begin_0 = const()[name = string("a_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_3_end_0 = const()[name = string("a_3_end_0"), val = tensor<int32, [3]>([1, 96, 1])];
            tensor<bool, [3]> a_3_end_mask_0 = const()[name = string("a_3_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 4]> a_3_cast_fp16 = slice_by_index(begin = a_3_begin_0, end = a_3_end_0, end_mask = a_3_end_mask_0, x = linear_4_cast_fp16)[name = string("a_3_cast_fp16")];
            int32 var_1566 = const()[name = string("op_1566"), val = int32(1)];
            bool xp_3_interleave_0 = const()[name = string("xp_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 99, 1024]> xp_3_cast_fp16 = concat(axis = var_1566, interleave = xp_3_interleave_0, values = (zpad_to_fp16, qkv_3_cast_fp16))[name = string("xp_3_cast_fp16")];
            tensor<int32, [3]> var_1577_begin_0 = const()[name = string("op_1577_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_1577_end_0 = const()[name = string("op_1577_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_1577_end_mask_0 = const()[name = string("op_1577_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 1024]> var_1577_cast_fp16 = slice_by_index(begin = var_1577_begin_0, end = var_1577_end_0, end_mask = var_1577_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1577_cast_fp16")];
            tensor<fp16, [1024]> var_1585_to_fp16 = const()[name = string("op_1585_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34975808)))];
            tensor<fp16, [1, 96, 1024]> var_1586_cast_fp16 = mul(x = var_1577_cast_fp16, y = var_1585_to_fp16)[name = string("op_1586_cast_fp16")];
            tensor<int32, [3]> var_1596_begin_0 = const()[name = string("op_1596_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_1596_end_0 = const()[name = string("op_1596_end_0"), val = tensor<int32, [3]>([1, 98, 1024])];
            tensor<bool, [3]> var_1596_end_mask_0 = const()[name = string("op_1596_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_1596_cast_fp16 = slice_by_index(begin = var_1596_begin_0, end = var_1596_end_0, end_mask = var_1596_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1596_cast_fp16")];
            tensor<fp16, [1024]> var_1604_to_fp16 = const()[name = string("op_1604_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34977920)))];
            tensor<fp16, [1, 96, 1024]> var_1605_cast_fp16 = mul(x = var_1596_cast_fp16, y = var_1604_to_fp16)[name = string("op_1605_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> var_1607_cast_fp16 = add(x = var_1586_cast_fp16, y = var_1605_cast_fp16)[name = string("op_1607_cast_fp16")];
            tensor<int32, [3]> var_1617_begin_0 = const()[name = string("op_1617_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1617_end_0 = const()[name = string("op_1617_end_0"), val = tensor<int32, [3]>([1, 97, 1024])];
            tensor<bool, [3]> var_1617_end_mask_0 = const()[name = string("op_1617_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_1617_cast_fp16 = slice_by_index(begin = var_1617_begin_0, end = var_1617_end_0, end_mask = var_1617_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1617_cast_fp16")];
            tensor<fp16, [1024]> var_1625_to_fp16 = const()[name = string("op_1625_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34980032)))];
            tensor<fp16, [1, 96, 1024]> var_1626_cast_fp16 = mul(x = var_1617_cast_fp16, y = var_1625_to_fp16)[name = string("op_1626_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> var_1628_cast_fp16 = add(x = var_1607_cast_fp16, y = var_1626_cast_fp16)[name = string("op_1628_cast_fp16")];
            tensor<int32, [3]> var_1638_begin_0 = const()[name = string("op_1638_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1638_end_0 = const()[name = string("op_1638_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> var_1638_end_mask_0 = const()[name = string("op_1638_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_1638_cast_fp16 = slice_by_index(begin = var_1638_begin_0, end = var_1638_end_0, end_mask = var_1638_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1638_cast_fp16")];
            tensor<fp16, [1024]> var_1646_to_fp16 = const()[name = string("op_1646_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34982144)))];
            tensor<fp16, [1, 96, 1024]> var_1647_cast_fp16 = mul(x = var_1638_cast_fp16, y = var_1646_to_fp16)[name = string("op_1647_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> input_19_cast_fp16 = add(x = var_1628_cast_fp16, y = var_1647_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> c_3_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("c_3_cast_fp16")];
            tensor<int32, [3]> var_1655_begin_0 = const()[name = string("op_1655_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1655_end_0 = const()[name = string("op_1655_end_0"), val = tensor<int32, [3]>([1, 96, 256])];
            tensor<bool, [3]> var_1655_end_mask_0 = const()[name = string("op_1655_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 256]> var_1655_cast_fp16 = slice_by_index(begin = var_1655_begin_0, end = var_1655_end_0, end_mask = var_1655_end_mask_0, x = c_3_cast_fp16)[name = string("op_1655_cast_fp16")];
            tensor<int32, [4]> var_1660 = const()[name = string("op_1660"), val = tensor<int32, [4]>([1, 96, 2, 128])];
            tensor<fp16, [1, 96, 2, 128]> q_21_cast_fp16 = reshape(shape = var_1660, x = var_1655_cast_fp16)[name = string("q_21_cast_fp16")];
            tensor<int32, [3]> var_1666_begin_0 = const()[name = string("op_1666_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_1666_end_0 = const()[name = string("op_1666_end_0"), val = tensor<int32, [3]>([1, 96, 512])];
            tensor<bool, [3]> var_1666_end_mask_0 = const()[name = string("op_1666_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 256]> var_1666_cast_fp16 = slice_by_index(begin = var_1666_begin_0, end = var_1666_end_0, end_mask = var_1666_end_mask_0, x = c_3_cast_fp16)[name = string("op_1666_cast_fp16")];
            tensor<int32, [4]> var_1671 = const()[name = string("op_1671"), val = tensor<int32, [4]>([1, 96, 2, 128])];
            tensor<fp16, [1, 96, 2, 128]> k_21_cast_fp16 = reshape(shape = var_1671, x = var_1666_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<int32, [3]> var_1677_begin_0 = const()[name = string("op_1677_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_1677_end_0 = const()[name = string("op_1677_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> var_1677_end_mask_0 = const()[name = string("op_1677_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 512]> var_1677_cast_fp16 = slice_by_index(begin = var_1677_begin_0, end = var_1677_end_0, end_mask = var_1677_end_mask_0, x = c_3_cast_fp16)[name = string("op_1677_cast_fp16")];
            tensor<int32, [4]> var_1682 = const()[name = string("op_1682"), val = tensor<int32, [4]>([1, 96, 4, 128])];
            tensor<fp16, [1, 96, 4, 128]> var_1683_cast_fp16 = reshape(shape = var_1682, x = var_1677_cast_fp16)[name = string("op_1683_cast_fp16")];
            tensor<int32, [4]> var_1688 = const()[name = string("op_1688"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1690_to_fp16 = const()[name = string("op_1690_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 96, 2, 128]> q_23_cast_fp16 = mul(x = q_21_cast_fp16, y = var_1690_to_fp16)[name = string("q_23_cast_fp16")];
            fp16 var_1692_to_fp16 = const()[name = string("op_1692_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 96, 2, 128]> k_23_cast_fp16 = mul(x = k_21_cast_fp16, y = var_1692_to_fp16)[name = string("k_23_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_1694_cast_fp16 = mul(x = q_23_cast_fp16, y = q_23_cast_fp16)[name = string("op_1694_cast_fp16")];
            tensor<int32, [1]> var_1699_axes_0 = const()[name = string("op_1699_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1699_keep_dims_0 = const()[name = string("op_1699_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 2, 1]> var_1699_cast_fp16 = reduce_mean(axes = var_1699_axes_0, keep_dims = var_1699_keep_dims_0, x = var_1694_cast_fp16)[name = string("op_1699_cast_fp16")];
            fp16 var_1700_promoted_to_fp16 = const()[name = string("op_1700_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 96, 2, 1]> var_1701_cast_fp16 = mul(x = var_1699_cast_fp16, y = var_1700_promoted_to_fp16)[name = string("op_1701_cast_fp16")];
            fp16 var_1703_to_fp16 = const()[name = string("op_1703_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 96, 2, 1]> var_1704_cast_fp16 = add(x = var_1701_cast_fp16, y = var_1703_to_fp16)[name = string("op_1704_cast_fp16")];
            fp32 var_1705_epsilon_0 = const()[name = string("op_1705_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 2, 1]> var_1705_cast_fp16 = rsqrt(epsilon = var_1705_epsilon_0, x = var_1704_cast_fp16)[name = string("op_1705_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_1706_cast_fp16 = mul(x = q_23_cast_fp16, y = var_1705_cast_fp16)[name = string("op_1706_cast_fp16")];
            fp16 var_1707_to_fp16 = const()[name = string("op_1707_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 96, 2, 128]> q_25_cast_fp16 = mul(x = var_1706_cast_fp16, y = var_1707_to_fp16)[name = string("q_25_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_1709_cast_fp16 = mul(x = k_23_cast_fp16, y = k_23_cast_fp16)[name = string("op_1709_cast_fp16")];
            tensor<int32, [1]> var_1714_axes_0 = const()[name = string("op_1714_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1714_keep_dims_0 = const()[name = string("op_1714_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 2, 1]> var_1714_cast_fp16 = reduce_mean(axes = var_1714_axes_0, keep_dims = var_1714_keep_dims_0, x = var_1709_cast_fp16)[name = string("op_1714_cast_fp16")];
            fp16 var_1715_promoted_to_fp16 = const()[name = string("op_1715_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 96, 2, 1]> var_1716_cast_fp16 = mul(x = var_1714_cast_fp16, y = var_1715_promoted_to_fp16)[name = string("op_1716_cast_fp16")];
            fp16 var_1718_to_fp16 = const()[name = string("op_1718_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 96, 2, 1]> var_1719_cast_fp16 = add(x = var_1716_cast_fp16, y = var_1718_to_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, 96, 2, 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, 96, 2, 128]> k_25_cast_fp16 = mul(x = k_23_cast_fp16, y = var_1720_cast_fp16)[name = string("k_25_cast_fp16")];
            tensor<int32, [4]> var_1726 = const()[name = string("op_1726"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1739_axes_0 = const()[name = string("op_1739_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 96, 128]> var_1727_cast_fp16 = transpose(perm = var_1726, x = q_25_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [1, 2, 1, 96, 128]> var_1739_cast_fp16 = expand_dims(axes = var_1739_axes_0, x = var_1727_cast_fp16)[name = string("op_1739_cast_fp16")];
            tensor<int32, [5]> var_1747_reps_0 = const()[name = string("op_1747_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 96, 128]> var_1747_cast_fp16 = tile(reps = var_1747_reps_0, x = var_1739_cast_fp16)[name = string("op_1747_cast_fp16")];
            tensor<int32, [4]> var_1752 = const()[name = string("op_1752"), val = tensor<int32, [4]>([1, 4, 96, 128])];
            tensor<fp16, [1, 4, 96, 128]> q_27_cast_fp16 = reshape(shape = var_1752, x = var_1747_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<int32, [4]> var_1758 = const()[name = string("op_1758"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1771_axes_0 = const()[name = string("op_1771_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 96, 128]> var_1759_cast_fp16 = transpose(perm = var_1758, x = k_25_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [1, 2, 1, 96, 128]> var_1771_cast_fp16 = expand_dims(axes = var_1771_axes_0, x = var_1759_cast_fp16)[name = string("op_1771_cast_fp16")];
            tensor<int32, [5]> var_1779_reps_0 = const()[name = string("op_1779_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 96, 128]> var_1779_cast_fp16 = tile(reps = var_1779_reps_0, x = var_1771_cast_fp16)[name = string("op_1779_cast_fp16")];
            tensor<int32, [4]> var_1784 = const()[name = string("op_1784"), val = tensor<int32, [4]>([1, 4, 96, 128])];
            tensor<fp16, [1, 4, 96, 128]> k_27_cast_fp16 = reshape(shape = var_1784, x = var_1779_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_1786_cast_fp16 = sigmoid(x = b_3_cast_fp16)[name = string("op_1786_cast_fp16")];
            tensor<int32, [3]> var_1790 = const()[name = string("op_1790"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_1_dt_bias_to_fp16 = const()[name = string("layers_1_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 96, 4]> input_21_cast_fp16 = add(x = a_3_cast_fp16, y = layers_1_dt_bias_to_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_1794_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("op_1794_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_1795_cast_fp16 = abs(x = input_21_cast_fp16)[name = string("op_1795_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 96, 4]> var_1796_cast_fp16 = mul(x = var_1795_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_1796_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_1797_cast_fp16 = exp(x = var_1796_cast_fp16)[name = string("op_1797_cast_fp16")];
            fp16 var_1799_to_fp16 = const()[name = string("op_1799_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 96, 4]> var_1800_cast_fp16 = add(x = var_1797_cast_fp16, y = var_1799_to_fp16)[name = string("op_1800_cast_fp16")];
            fp32 var_1801_epsilon_0 = const()[name = string("op_1801_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 96, 4]> var_1801_cast_fp16 = log(epsilon = var_1801_epsilon_0, x = var_1800_cast_fp16)[name = string("op_1801_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_1803_cast_fp16 = add(x = var_1794_cast_fp16, y = var_1801_cast_fp16)[name = string("op_1803_cast_fp16")];
            tensor<fp16, [4]> layers_1_negA_to_fp16 = const()[name = string("layers_1_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.ff4p+2, -0x1.8cp+3, -0x1.c0cp+3, -0x1.104p+3])];
            tensor<fp16, [1, 96, 4]> var_1804_cast_fp16 = mul(x = layers_1_negA_to_fp16, y = var_1803_cast_fp16)[name = string("op_1804_cast_fp16")];
            tensor<int32, [3]> var_1808 = const()[name = string("op_1808"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_1811_axes_0 = const()[name = string("op_1811_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 96]> beta_3_cast_fp16 = transpose(perm = var_1790, x = var_1786_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [1, 4, 96, 1]> var_1811_cast_fp16 = expand_dims(axes = var_1811_axes_0, x = beta_3_cast_fp16)[name = string("op_1811_cast_fp16")];
            fp16 var_1812_to_fp16 = const()[name = string("op_1812_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 96, 1]> var_1813_cast_fp16 = mul(x = var_1811_cast_fp16, y = var_1812_to_fp16)[name = string("op_1813_cast_fp16")];
            tensor<fp16, [1, 4, 96, 128]> v_3_cast_fp16 = transpose(perm = var_1688, x = var_1683_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [1, 4, 96, 128]> vb_15_cast_fp16 = mul(x = v_3_cast_fp16, y = var_1813_cast_fp16)[name = string("vb_15_cast_fp16")];
            tensor<int32, [4]> q_29_begin_0 = const()[name = string("q_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_29_end_0 = const()[name = string("q_29_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_29_end_mask_0 = const()[name = string("q_29_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_29_cast_fp16 = slice_by_index(begin = q_29_begin_0, end = q_29_end_0, end_mask = q_29_end_mask_0, x = q_27_cast_fp16)[name = string("q_29_cast_fp16")];
            tensor<int32, [4]> k_29_begin_0 = const()[name = string("k_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_29_end_0 = const()[name = string("k_29_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_29_end_mask_0 = const()[name = string("k_29_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, x = k_27_cast_fp16)[name = string("k_29_cast_fp16")];
            tensor<fp16, [1, 4, 96, 128]> var_1847_cast_fp16 = mul(x = k_27_cast_fp16, y = var_1811_cast_fp16)[name = string("op_1847_cast_fp16")];
            tensor<int32, [4]> kb_13_begin_0 = const()[name = string("kb_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_13_end_0 = const()[name = string("kb_13_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_13_end_mask_0 = const()[name = string("kb_13_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_13_cast_fp16 = slice_by_index(begin = kb_13_begin_0, end = kb_13_end_0, end_mask = kb_13_end_mask_0, x = var_1847_cast_fp16)[name = string("kb_13_cast_fp16")];
            tensor<int32, [4]> vb_17_begin_0 = const()[name = string("vb_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_17_end_0 = const()[name = string("vb_17_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_17_end_mask_0 = const()[name = string("vb_17_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_17_cast_fp16 = slice_by_index(begin = vb_17_begin_0, end = vb_17_end_0, end_mask = vb_17_end_mask_0, x = vb_15_cast_fp16)[name = string("vb_17_cast_fp16")];
            tensor<int32, [3]> gdec_17_begin_0 = const()[name = string("gdec_17_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_17_end_0 = const()[name = string("gdec_17_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_17_end_mask_0 = const()[name = string("gdec_17_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 96]> gdec_15_cast_fp16 = transpose(perm = var_1808, x = var_1804_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [1, 4, 16]> gdec_17_cast_fp16 = slice_by_index(begin = gdec_17_begin_0, end = gdec_17_end_0, end_mask = gdec_17_end_mask_0, x = gdec_15_cast_fp16)[name = string("gdec_17_cast_fp16")];
            tensor<int32, [2]> var_1895 = const()[name = string("op_1895"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_1896_cast_fp16 = reshape(shape = var_1895, x = gdec_17_cast_fp16)[name = string("op_1896_cast_fp16")];
            tensor<fp16, [4, 256]> var_1897_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1896_cast_fp16)[name = string("op_1897_cast_fp16")];
            tensor<int32, [4]> var_1902 = const()[name = string("op_1902"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_13_cast_fp16 = reshape(shape = var_1902, x = var_1897_cast_fp16)[name = string("P_13_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1904_cast_fp16 = exp(x = P_13_cast_fp16)[name = string("op_1904_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_13_cast_fp16 = mul(x = var_1904_cast_fp16, y = tril_incl_to_fp16)[name = string("D_13_cast_fp16")];
            bool kk_13_transpose_x_1 = const()[name = string("kk_13_transpose_x_1"), val = bool(false)];
            bool kk_13_transpose_y_1 = const()[name = string("kk_13_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_13_cast_fp16 = matmul(transpose_x = kk_13_transpose_x_1, transpose_y = kk_13_transpose_y_1, x = kb_13_cast_fp16, y = k_29_cast_fp16)[name = string("kk_13_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_1910_cast_fp16 = mul(x = kk_13_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_1910_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1911_cast_fp16 = mul(x = var_1910_cast_fp16, y = D_13_cast_fp16)[name = string("op_1911_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_13_cast_fp16 = mul(x = var_1911_cast_fp16, y = cum_tail_to_fp16)[name = string("N_13_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_1913_cast_fp16 = mul(x = D_13_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1913_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_13_cast_fp16 = mul(x = kk_13_cast_fp16, y = var_1913_cast_fp16)[name = string("Na_13_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_49_cast_fp16 = add(x = eye_to_fp16, y = N_13_cast_fp16)[name = string("Minv_49_cast_fp16")];
            bool Np_37_transpose_x_0 = const()[name = string("Np_37_transpose_x_0"), val = bool(false)];
            bool Np_37_transpose_y_0 = const()[name = string("Np_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_37_cast_fp16 = matmul(transpose_x = Np_37_transpose_x_0, transpose_y = Np_37_transpose_y_0, x = N_13_cast_fp16, y = Na_13_cast_fp16)[name = string("Np_37_cast_fp16")];
            bool Npa_25_transpose_x_0 = const()[name = string("Npa_25_transpose_x_0"), val = bool(false)];
            bool Npa_25_transpose_y_0 = const()[name = string("Npa_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_25_cast_fp16 = matmul(transpose_x = Npa_25_transpose_x_0, transpose_y = Npa_25_transpose_y_0, x = Na_13_cast_fp16, y = N_13_cast_fp16)[name = string("Npa_25_cast_fp16")];
            bool var_1919_transpose_x_0 = const()[name = string("op_1919_transpose_x_0"), val = bool(false)];
            bool var_1919_transpose_y_0 = const()[name = string("op_1919_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1919_cast_fp16 = matmul(transpose_x = var_1919_transpose_x_0, transpose_y = var_1919_transpose_y_0, x = Minv_49_cast_fp16, y = Np_37_cast_fp16)[name = string("op_1919_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_51_cast_fp16 = add(x = Minv_49_cast_fp16, y = var_1919_cast_fp16)[name = string("Minv_51_cast_fp16")];
            bool Np_39_transpose_x_0 = const()[name = string("Np_39_transpose_x_0"), val = bool(false)];
            bool Np_39_transpose_y_0 = const()[name = string("Np_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_39_cast_fp16 = matmul(transpose_x = Np_39_transpose_x_0, transpose_y = Np_39_transpose_y_0, x = Np_37_cast_fp16, y = Npa_25_cast_fp16)[name = string("Np_39_cast_fp16")];
            bool Npa_27_transpose_x_0 = const()[name = string("Npa_27_transpose_x_0"), val = bool(false)];
            bool Npa_27_transpose_y_0 = const()[name = string("Npa_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_27_cast_fp16 = matmul(transpose_x = Npa_27_transpose_x_0, transpose_y = Npa_27_transpose_y_0, x = Npa_25_cast_fp16, y = Np_37_cast_fp16)[name = string("Npa_27_cast_fp16")];
            bool var_1924_transpose_x_0 = const()[name = string("op_1924_transpose_x_0"), val = bool(false)];
            bool var_1924_transpose_y_0 = const()[name = string("op_1924_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1924_cast_fp16 = matmul(transpose_x = var_1924_transpose_x_0, transpose_y = var_1924_transpose_y_0, x = Minv_51_cast_fp16, y = Np_39_cast_fp16)[name = string("op_1924_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_53_cast_fp16 = add(x = Minv_51_cast_fp16, y = var_1924_cast_fp16)[name = string("Minv_53_cast_fp16")];
            bool Np_41_transpose_x_0 = const()[name = string("Np_41_transpose_x_0"), val = bool(false)];
            bool Np_41_transpose_y_0 = const()[name = string("Np_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_41_cast_fp16 = matmul(transpose_x = Np_41_transpose_x_0, transpose_y = Np_41_transpose_y_0, x = Np_39_cast_fp16, y = Npa_27_cast_fp16)[name = string("Np_41_cast_fp16")];
            bool var_1928_transpose_x_0 = const()[name = string("op_1928_transpose_x_0"), val = bool(false)];
            bool var_1928_transpose_y_0 = const()[name = string("op_1928_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_1928_cast_fp16 = matmul(transpose_x = var_1928_transpose_x_0, transpose_y = var_1928_transpose_y_0, x = Minv_53_cast_fp16, y = Np_41_cast_fp16)[name = string("op_1928_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_55_cast_fp16 = add(x = Minv_53_cast_fp16, y = var_1928_cast_fp16)[name = string("Minv_55_cast_fp16")];
            bool u_13_transpose_x_0 = const()[name = string("u_13_transpose_x_0"), val = bool(false)];
            bool u_13_transpose_y_0 = const()[name = string("u_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_13_cast_fp16 = matmul(transpose_x = u_13_transpose_x_0, transpose_y = u_13_transpose_y_0, x = Minv_55_cast_fp16, y = vb_17_cast_fp16)[name = string("u_13_cast_fp16")];
            bool var_1935_transpose_x_1 = const()[name = string("op_1935_transpose_x_1"), val = bool(false)];
            bool var_1935_transpose_y_1 = const()[name = string("op_1935_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_1935_cast_fp16 = matmul(transpose_x = var_1935_transpose_x_1, transpose_y = var_1935_transpose_y_1, x = q_29_cast_fp16, y = k_29_cast_fp16)[name = string("op_1935_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_13_cast_fp16 = mul(x = var_1935_cast_fp16, y = D_13_cast_fp16)[name = string("A_13_cast_fp16")];
            bool var_1937_transpose_x_0 = const()[name = string("op_1937_transpose_x_0"), val = bool(false)];
            bool var_1937_transpose_y_0 = const()[name = string("op_1937_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_1937_cast_fp16 = matmul(transpose_x = var_1937_transpose_x_0, transpose_y = var_1937_transpose_y_0, x = A_13_cast_fp16, y = u_13_cast_fp16)[name = string("op_1937_cast_fp16")];
            tensor<fp16, [16]> var_1938_bias_0_to_fp16 = const()[name = string("op_1938_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_1938_cast_fp16 = linear(bias = var_1938_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_17_cast_fp16)[name = string("op_1938_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_11_cast_fp16 = exp(x = var_1938_cast_fp16)[name = string("tail_11_cast_fp16")];
            tensor<int32, [1]> var_1941_axes_0 = const()[name = string("op_1941_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_1941_cast_fp16 = expand_dims(axes = var_1941_axes_0, x = tail_11_cast_fp16)[name = string("op_1941_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_1942_cast_fp16 = mul(x = k_29_cast_fp16, y = var_1941_cast_fp16)[name = string("op_1942_cast_fp16")];
            bool S_11_transpose_x_1 = const()[name = string("S_11_transpose_x_1"), val = bool(true)];
            bool S_11_transpose_y_1 = const()[name = string("S_11_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_11_cast_fp16 = matmul(transpose_x = S_11_transpose_x_1, transpose_y = S_11_transpose_y_1, x = var_1942_cast_fp16, y = u_13_cast_fp16)[name = string("S_11_cast_fp16")];
            tensor<int32, [4]> q_31_begin_0 = const()[name = string("q_31_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_31_end_0 = const()[name = string("q_31_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> q_31_end_mask_0 = const()[name = string("q_31_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, x = q_27_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<int32, [4]> k_31_begin_0 = const()[name = string("k_31_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_31_end_0 = const()[name = string("k_31_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> k_31_end_mask_0 = const()[name = string("k_31_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_31_cast_fp16 = slice_by_index(begin = k_31_begin_0, end = k_31_end_0, end_mask = k_31_end_mask_0, x = k_27_cast_fp16)[name = string("k_31_cast_fp16")];
            tensor<int32, [4]> kb_15_begin_0 = const()[name = string("kb_15_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_15_end_0 = const()[name = string("kb_15_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> kb_15_end_mask_0 = const()[name = string("kb_15_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_15_cast_fp16 = slice_by_index(begin = kb_15_begin_0, end = kb_15_end_0, end_mask = kb_15_end_mask_0, x = var_1847_cast_fp16)[name = string("kb_15_cast_fp16")];
            tensor<int32, [4]> vb_19_begin_0 = const()[name = string("vb_19_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_19_end_0 = const()[name = string("vb_19_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> vb_19_end_mask_0 = const()[name = string("vb_19_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_19_cast_fp16 = slice_by_index(begin = vb_19_begin_0, end = vb_19_end_0, end_mask = vb_19_end_mask_0, x = vb_15_cast_fp16)[name = string("vb_19_cast_fp16")];
            tensor<int32, [3]> gdec_19_begin_0 = const()[name = string("gdec_19_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_19_end_0 = const()[name = string("gdec_19_end_0"), val = tensor<int32, [3]>([1, 4, 32])];
            tensor<bool, [3]> gdec_19_end_mask_0 = const()[name = string("gdec_19_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_19_cast_fp16 = slice_by_index(begin = gdec_19_begin_0, end = gdec_19_end_0, end_mask = gdec_19_end_mask_0, x = gdec_15_cast_fp16)[name = string("gdec_19_cast_fp16")];
            tensor<int32, [2]> var_2027 = const()[name = string("op_2027"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_2028_cast_fp16 = reshape(shape = var_2027, x = gdec_19_cast_fp16)[name = string("op_2028_cast_fp16")];
            tensor<fp16, [4, 256]> var_2029_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2028_cast_fp16)[name = string("op_2029_cast_fp16")];
            tensor<int32, [4]> var_2034 = const()[name = string("op_2034"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_15_cast_fp16 = reshape(shape = var_2034, x = var_2029_cast_fp16)[name = string("P_15_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2036_cast_fp16 = exp(x = P_15_cast_fp16)[name = string("op_2036_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_15_cast_fp16 = mul(x = var_2036_cast_fp16, y = tril_incl_to_fp16)[name = string("D_15_cast_fp16")];
            bool kk_15_transpose_x_1 = const()[name = string("kk_15_transpose_x_1"), val = bool(false)];
            bool kk_15_transpose_y_1 = const()[name = string("kk_15_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_15_cast_fp16 = matmul(transpose_x = kk_15_transpose_x_1, transpose_y = kk_15_transpose_y_1, x = kb_15_cast_fp16, y = k_31_cast_fp16)[name = string("kk_15_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_2042_cast_fp16 = mul(x = kk_15_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_2042_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2043_cast_fp16 = mul(x = var_2042_cast_fp16, y = D_15_cast_fp16)[name = string("op_2043_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_15_cast_fp16 = mul(x = var_2043_cast_fp16, y = cum_tail_to_fp16)[name = string("N_15_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2045_cast_fp16 = mul(x = D_15_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2045_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_15_cast_fp16 = mul(x = kk_15_cast_fp16, y = var_2045_cast_fp16)[name = string("Na_15_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_57_cast_fp16 = add(x = eye_to_fp16, y = N_15_cast_fp16)[name = string("Minv_57_cast_fp16")];
            bool Np_43_transpose_x_0 = const()[name = string("Np_43_transpose_x_0"), val = bool(false)];
            bool Np_43_transpose_y_0 = const()[name = string("Np_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_43_cast_fp16 = matmul(transpose_x = Np_43_transpose_x_0, transpose_y = Np_43_transpose_y_0, x = N_15_cast_fp16, y = Na_15_cast_fp16)[name = string("Np_43_cast_fp16")];
            bool Npa_29_transpose_x_0 = const()[name = string("Npa_29_transpose_x_0"), val = bool(false)];
            bool Npa_29_transpose_y_0 = const()[name = string("Npa_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_29_cast_fp16 = matmul(transpose_x = Npa_29_transpose_x_0, transpose_y = Npa_29_transpose_y_0, x = Na_15_cast_fp16, y = N_15_cast_fp16)[name = string("Npa_29_cast_fp16")];
            bool var_2051_transpose_x_0 = const()[name = string("op_2051_transpose_x_0"), val = bool(false)];
            bool var_2051_transpose_y_0 = const()[name = string("op_2051_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2051_cast_fp16 = matmul(transpose_x = var_2051_transpose_x_0, transpose_y = var_2051_transpose_y_0, x = Minv_57_cast_fp16, y = Np_43_cast_fp16)[name = string("op_2051_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_59_cast_fp16 = add(x = Minv_57_cast_fp16, y = var_2051_cast_fp16)[name = string("Minv_59_cast_fp16")];
            bool Np_45_transpose_x_0 = const()[name = string("Np_45_transpose_x_0"), val = bool(false)];
            bool Np_45_transpose_y_0 = const()[name = string("Np_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_45_cast_fp16 = matmul(transpose_x = Np_45_transpose_x_0, transpose_y = Np_45_transpose_y_0, x = Np_43_cast_fp16, y = Npa_29_cast_fp16)[name = string("Np_45_cast_fp16")];
            bool Npa_31_transpose_x_0 = const()[name = string("Npa_31_transpose_x_0"), val = bool(false)];
            bool Npa_31_transpose_y_0 = const()[name = string("Npa_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_31_cast_fp16 = matmul(transpose_x = Npa_31_transpose_x_0, transpose_y = Npa_31_transpose_y_0, x = Npa_29_cast_fp16, y = Np_43_cast_fp16)[name = string("Npa_31_cast_fp16")];
            bool var_2056_transpose_x_0 = const()[name = string("op_2056_transpose_x_0"), val = bool(false)];
            bool var_2056_transpose_y_0 = const()[name = string("op_2056_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2056_cast_fp16 = matmul(transpose_x = var_2056_transpose_x_0, transpose_y = var_2056_transpose_y_0, x = Minv_59_cast_fp16, y = Np_45_cast_fp16)[name = string("op_2056_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_61_cast_fp16 = add(x = Minv_59_cast_fp16, y = var_2056_cast_fp16)[name = string("Minv_61_cast_fp16")];
            bool Np_47_transpose_x_0 = const()[name = string("Np_47_transpose_x_0"), val = bool(false)];
            bool Np_47_transpose_y_0 = const()[name = string("Np_47_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_47_cast_fp16 = matmul(transpose_x = Np_47_transpose_x_0, transpose_y = Np_47_transpose_y_0, x = Np_45_cast_fp16, y = Npa_31_cast_fp16)[name = string("Np_47_cast_fp16")];
            bool var_2060_transpose_x_0 = const()[name = string("op_2060_transpose_x_0"), val = bool(false)];
            bool var_2060_transpose_y_0 = const()[name = string("op_2060_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2060_cast_fp16 = matmul(transpose_x = var_2060_transpose_x_0, transpose_y = var_2060_transpose_y_0, x = Minv_61_cast_fp16, y = Np_47_cast_fp16)[name = string("op_2060_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_63_cast_fp16 = add(x = Minv_61_cast_fp16, y = var_2060_cast_fp16)[name = string("Minv_63_cast_fp16")];
            bool u_15_transpose_x_0 = const()[name = string("u_15_transpose_x_0"), val = bool(false)];
            bool u_15_transpose_y_0 = const()[name = string("u_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_15_cast_fp16 = matmul(transpose_x = u_15_transpose_x_0, transpose_y = u_15_transpose_y_0, x = Minv_63_cast_fp16, y = vb_19_cast_fp16)[name = string("u_15_cast_fp16")];
            bool var_2067_transpose_x_1 = const()[name = string("op_2067_transpose_x_1"), val = bool(false)];
            bool var_2067_transpose_y_1 = const()[name = string("op_2067_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_2067_cast_fp16 = matmul(transpose_x = var_2067_transpose_x_1, transpose_y = var_2067_transpose_y_1, x = q_31_cast_fp16, y = k_31_cast_fp16)[name = string("op_2067_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_15_cast_fp16 = mul(x = var_2067_cast_fp16, y = D_15_cast_fp16)[name = string("A_15_cast_fp16")];
            tensor<fp16, [16]> var_2069_bias_0_to_fp16 = const()[name = string("op_2069_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_2069_cast_fp16 = linear(bias = var_2069_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_19_cast_fp16)[name = string("op_2069_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_11_cast_fp16 = exp(x = var_2069_cast_fp16)[name = string("cum_11_cast_fp16")];
            tensor<int32, [1]> var_2072_axes_0 = const()[name = string("op_2072_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_2072_cast_fp16 = expand_dims(axes = var_2072_axes_0, x = cum_11_cast_fp16)[name = string("op_2072_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2073_cast_fp16 = mul(x = kb_15_cast_fp16, y = var_2072_cast_fp16)[name = string("op_2073_cast_fp16")];
            bool w_15_transpose_x_0 = const()[name = string("w_15_transpose_x_0"), val = bool(false)];
            bool w_15_transpose_y_0 = const()[name = string("w_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_15_cast_fp16 = matmul(transpose_x = w_15_transpose_x_0, transpose_y = w_15_transpose_y_0, x = Minv_63_cast_fp16, y = var_2073_cast_fp16)[name = string("w_15_cast_fp16")];
            bool var_2075_transpose_x_0 = const()[name = string("op_2075_transpose_x_0"), val = bool(false)];
            bool var_2075_transpose_y_0 = const()[name = string("op_2075_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2075_cast_fp16 = matmul(transpose_x = var_2075_transpose_x_0, transpose_y = var_2075_transpose_y_0, x = w_15_cast_fp16, y = S_11_cast_fp16)[name = string("op_2075_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_11_cast_fp16 = sub(x = u_15_cast_fp16, y = var_2075_cast_fp16)[name = string("vnew_11_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2080_cast_fp16 = mul(x = q_31_cast_fp16, y = var_2072_cast_fp16)[name = string("op_2080_cast_fp16")];
            bool var_2081_transpose_x_0 = const()[name = string("op_2081_transpose_x_0"), val = bool(false)];
            bool var_2081_transpose_y_0 = const()[name = string("op_2081_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2081_cast_fp16 = matmul(transpose_x = var_2081_transpose_x_0, transpose_y = var_2081_transpose_y_0, x = var_2080_cast_fp16, y = S_11_cast_fp16)[name = string("op_2081_cast_fp16")];
            bool var_2082_transpose_x_0 = const()[name = string("op_2082_transpose_x_0"), val = bool(false)];
            bool var_2082_transpose_y_0 = const()[name = string("op_2082_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2082_cast_fp16 = matmul(transpose_x = var_2082_transpose_x_0, transpose_y = var_2082_transpose_y_0, x = A_15_cast_fp16, y = vnew_11_cast_fp16)[name = string("op_2082_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2084_cast_fp16 = add(x = var_2081_cast_fp16, y = var_2082_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<fp16, [16]> var_2085_bias_0_to_fp16 = const()[name = string("op_2085_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_2085_cast_fp16 = linear(bias = var_2085_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_19_cast_fp16)[name = string("op_2085_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_13_cast_fp16 = exp(x = var_2085_cast_fp16)[name = string("tail_13_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_25_to_fp16 = const()[name = string("transpose_25_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_2087_bias_0_to_fp16 = const()[name = string("op_2087_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_2087_cast_fp16 = linear(bias = var_2087_bias_0_to_fp16, weight = transpose_25_to_fp16, x = gdec_19_cast_fp16)[name = string("op_2087_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_2088_cast_fp16 = exp(x = var_2087_cast_fp16)[name = string("op_2088_cast_fp16")];
            tensor<int32, [1]> last_11_axes_0 = const()[name = string("last_11_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_11_cast_fp16 = expand_dims(axes = last_11_axes_0, x = var_2088_cast_fp16)[name = string("last_11_cast_fp16")];
            tensor<int32, [1]> var_2092_axes_0 = const()[name = string("op_2092_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_2092_cast_fp16 = expand_dims(axes = var_2092_axes_0, x = tail_13_cast_fp16)[name = string("op_2092_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2093_cast_fp16 = mul(x = k_31_cast_fp16, y = var_2092_cast_fp16)[name = string("op_2093_cast_fp16")];
            bool upd_9_transpose_x_1 = const()[name = string("upd_9_transpose_x_1"), val = bool(true)];
            bool upd_9_transpose_y_1 = const()[name = string("upd_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_9_cast_fp16 = matmul(transpose_x = upd_9_transpose_x_1, transpose_y = upd_9_transpose_y_1, x = var_2093_cast_fp16, y = vnew_11_cast_fp16)[name = string("upd_9_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_2098_cast_fp16 = mul(x = S_11_cast_fp16, y = last_11_cast_fp16)[name = string("op_2098_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_13_cast_fp16 = add(x = var_2098_cast_fp16, y = upd_9_cast_fp16)[name = string("S_13_cast_fp16")];
            tensor<int32, [4]> q_33_begin_0 = const()[name = string("q_33_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_33_end_0 = const()[name = string("q_33_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> q_33_end_mask_0 = const()[name = string("q_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_33_cast_fp16 = slice_by_index(begin = q_33_begin_0, end = q_33_end_0, end_mask = q_33_end_mask_0, x = q_27_cast_fp16)[name = string("q_33_cast_fp16")];
            tensor<int32, [4]> k_33_begin_0 = const()[name = string("k_33_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_33_end_0 = const()[name = string("k_33_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> k_33_end_mask_0 = const()[name = string("k_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, x = k_27_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<int32, [4]> kb_17_begin_0 = const()[name = string("kb_17_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_17_end_0 = const()[name = string("kb_17_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> kb_17_end_mask_0 = const()[name = string("kb_17_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_17_cast_fp16 = slice_by_index(begin = kb_17_begin_0, end = kb_17_end_0, end_mask = kb_17_end_mask_0, x = var_1847_cast_fp16)[name = string("kb_17_cast_fp16")];
            tensor<int32, [4]> vb_21_begin_0 = const()[name = string("vb_21_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_21_end_0 = const()[name = string("vb_21_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> vb_21_end_mask_0 = const()[name = string("vb_21_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_21_cast_fp16 = slice_by_index(begin = vb_21_begin_0, end = vb_21_end_0, end_mask = vb_21_end_mask_0, x = vb_15_cast_fp16)[name = string("vb_21_cast_fp16")];
            tensor<int32, [3]> gdec_21_begin_0 = const()[name = string("gdec_21_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_21_end_0 = const()[name = string("gdec_21_end_0"), val = tensor<int32, [3]>([1, 4, 48])];
            tensor<bool, [3]> gdec_21_end_mask_0 = const()[name = string("gdec_21_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_21_cast_fp16 = slice_by_index(begin = gdec_21_begin_0, end = gdec_21_end_0, end_mask = gdec_21_end_mask_0, x = gdec_15_cast_fp16)[name = string("gdec_21_cast_fp16")];
            tensor<int32, [2]> var_2181 = const()[name = string("op_2181"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_2182_cast_fp16 = reshape(shape = var_2181, x = gdec_21_cast_fp16)[name = string("op_2182_cast_fp16")];
            tensor<fp16, [4, 256]> var_2183_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2182_cast_fp16)[name = string("op_2183_cast_fp16")];
            tensor<int32, [4]> var_2188 = const()[name = string("op_2188"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_17_cast_fp16 = reshape(shape = var_2188, x = var_2183_cast_fp16)[name = string("P_17_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2190_cast_fp16 = exp(x = P_17_cast_fp16)[name = string("op_2190_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_17_cast_fp16 = mul(x = var_2190_cast_fp16, y = tril_incl_to_fp16)[name = string("D_17_cast_fp16")];
            bool kk_17_transpose_x_1 = const()[name = string("kk_17_transpose_x_1"), val = bool(false)];
            bool kk_17_transpose_y_1 = const()[name = string("kk_17_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_17_cast_fp16 = matmul(transpose_x = kk_17_transpose_x_1, transpose_y = kk_17_transpose_y_1, x = kb_17_cast_fp16, y = k_33_cast_fp16)[name = string("kk_17_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_2196_cast_fp16 = mul(x = kk_17_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_2196_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2197_cast_fp16 = mul(x = var_2196_cast_fp16, y = D_17_cast_fp16)[name = string("op_2197_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_17_cast_fp16 = mul(x = var_2197_cast_fp16, y = cum_tail_to_fp16)[name = string("N_17_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2199_cast_fp16 = mul(x = D_17_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2199_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_17_cast_fp16 = mul(x = kk_17_cast_fp16, y = var_2199_cast_fp16)[name = string("Na_17_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_65_cast_fp16 = add(x = eye_to_fp16, y = N_17_cast_fp16)[name = string("Minv_65_cast_fp16")];
            bool Np_49_transpose_x_0 = const()[name = string("Np_49_transpose_x_0"), val = bool(false)];
            bool Np_49_transpose_y_0 = const()[name = string("Np_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_49_cast_fp16 = matmul(transpose_x = Np_49_transpose_x_0, transpose_y = Np_49_transpose_y_0, x = N_17_cast_fp16, y = Na_17_cast_fp16)[name = string("Np_49_cast_fp16")];
            bool Npa_33_transpose_x_0 = const()[name = string("Npa_33_transpose_x_0"), val = bool(false)];
            bool Npa_33_transpose_y_0 = const()[name = string("Npa_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_33_cast_fp16 = matmul(transpose_x = Npa_33_transpose_x_0, transpose_y = Npa_33_transpose_y_0, x = Na_17_cast_fp16, y = N_17_cast_fp16)[name = string("Npa_33_cast_fp16")];
            bool var_2205_transpose_x_0 = const()[name = string("op_2205_transpose_x_0"), val = bool(false)];
            bool var_2205_transpose_y_0 = const()[name = string("op_2205_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2205_cast_fp16 = matmul(transpose_x = var_2205_transpose_x_0, transpose_y = var_2205_transpose_y_0, x = Minv_65_cast_fp16, y = Np_49_cast_fp16)[name = string("op_2205_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_67_cast_fp16 = add(x = Minv_65_cast_fp16, y = var_2205_cast_fp16)[name = string("Minv_67_cast_fp16")];
            bool Np_51_transpose_x_0 = const()[name = string("Np_51_transpose_x_0"), val = bool(false)];
            bool Np_51_transpose_y_0 = const()[name = string("Np_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_51_cast_fp16 = matmul(transpose_x = Np_51_transpose_x_0, transpose_y = Np_51_transpose_y_0, x = Np_49_cast_fp16, y = Npa_33_cast_fp16)[name = string("Np_51_cast_fp16")];
            bool Npa_35_transpose_x_0 = const()[name = string("Npa_35_transpose_x_0"), val = bool(false)];
            bool Npa_35_transpose_y_0 = const()[name = string("Npa_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_35_cast_fp16 = matmul(transpose_x = Npa_35_transpose_x_0, transpose_y = Npa_35_transpose_y_0, x = Npa_33_cast_fp16, y = Np_49_cast_fp16)[name = string("Npa_35_cast_fp16")];
            bool var_2210_transpose_x_0 = const()[name = string("op_2210_transpose_x_0"), val = bool(false)];
            bool var_2210_transpose_y_0 = const()[name = string("op_2210_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2210_cast_fp16 = matmul(transpose_x = var_2210_transpose_x_0, transpose_y = var_2210_transpose_y_0, x = Minv_67_cast_fp16, y = Np_51_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_69_cast_fp16 = add(x = Minv_67_cast_fp16, y = var_2210_cast_fp16)[name = string("Minv_69_cast_fp16")];
            bool Np_53_transpose_x_0 = const()[name = string("Np_53_transpose_x_0"), val = bool(false)];
            bool Np_53_transpose_y_0 = const()[name = string("Np_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_53_cast_fp16 = matmul(transpose_x = Np_53_transpose_x_0, transpose_y = Np_53_transpose_y_0, x = Np_51_cast_fp16, y = Npa_35_cast_fp16)[name = string("Np_53_cast_fp16")];
            bool var_2214_transpose_x_0 = const()[name = string("op_2214_transpose_x_0"), val = bool(false)];
            bool var_2214_transpose_y_0 = const()[name = string("op_2214_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2214_cast_fp16 = matmul(transpose_x = var_2214_transpose_x_0, transpose_y = var_2214_transpose_y_0, x = Minv_69_cast_fp16, y = Np_53_cast_fp16)[name = string("op_2214_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_71_cast_fp16 = add(x = Minv_69_cast_fp16, y = var_2214_cast_fp16)[name = string("Minv_71_cast_fp16")];
            bool u_17_transpose_x_0 = const()[name = string("u_17_transpose_x_0"), val = bool(false)];
            bool u_17_transpose_y_0 = const()[name = string("u_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_17_cast_fp16 = matmul(transpose_x = u_17_transpose_x_0, transpose_y = u_17_transpose_y_0, x = Minv_71_cast_fp16, y = vb_21_cast_fp16)[name = string("u_17_cast_fp16")];
            bool var_2221_transpose_x_1 = const()[name = string("op_2221_transpose_x_1"), val = bool(false)];
            bool var_2221_transpose_y_1 = const()[name = string("op_2221_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_2221_cast_fp16 = matmul(transpose_x = var_2221_transpose_x_1, transpose_y = var_2221_transpose_y_1, x = q_33_cast_fp16, y = k_33_cast_fp16)[name = string("op_2221_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_17_cast_fp16 = mul(x = var_2221_cast_fp16, y = D_17_cast_fp16)[name = string("A_17_cast_fp16")];
            tensor<fp16, [16]> var_2223_bias_0_to_fp16 = const()[name = string("op_2223_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_2223_cast_fp16 = linear(bias = var_2223_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_21_cast_fp16)[name = string("op_2223_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_13_cast_fp16 = exp(x = var_2223_cast_fp16)[name = string("cum_13_cast_fp16")];
            tensor<int32, [1]> var_2226_axes_0 = const()[name = string("op_2226_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_2226_cast_fp16 = expand_dims(axes = var_2226_axes_0, x = cum_13_cast_fp16)[name = string("op_2226_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2227_cast_fp16 = mul(x = kb_17_cast_fp16, y = var_2226_cast_fp16)[name = string("op_2227_cast_fp16")];
            bool w_17_transpose_x_0 = const()[name = string("w_17_transpose_x_0"), val = bool(false)];
            bool w_17_transpose_y_0 = const()[name = string("w_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_17_cast_fp16 = matmul(transpose_x = w_17_transpose_x_0, transpose_y = w_17_transpose_y_0, x = Minv_71_cast_fp16, y = var_2227_cast_fp16)[name = string("w_17_cast_fp16")];
            bool var_2229_transpose_x_0 = const()[name = string("op_2229_transpose_x_0"), val = bool(false)];
            bool var_2229_transpose_y_0 = const()[name = string("op_2229_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2229_cast_fp16 = matmul(transpose_x = var_2229_transpose_x_0, transpose_y = var_2229_transpose_y_0, x = w_17_cast_fp16, y = S_13_cast_fp16)[name = string("op_2229_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_13_cast_fp16 = sub(x = u_17_cast_fp16, y = var_2229_cast_fp16)[name = string("vnew_13_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2234_cast_fp16 = mul(x = q_33_cast_fp16, y = var_2226_cast_fp16)[name = string("op_2234_cast_fp16")];
            bool var_2235_transpose_x_0 = const()[name = string("op_2235_transpose_x_0"), val = bool(false)];
            bool var_2235_transpose_y_0 = const()[name = string("op_2235_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2235_cast_fp16 = matmul(transpose_x = var_2235_transpose_x_0, transpose_y = var_2235_transpose_y_0, x = var_2234_cast_fp16, y = S_13_cast_fp16)[name = string("op_2235_cast_fp16")];
            bool var_2236_transpose_x_0 = const()[name = string("op_2236_transpose_x_0"), val = bool(false)];
            bool var_2236_transpose_y_0 = const()[name = string("op_2236_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2236_cast_fp16 = matmul(transpose_x = var_2236_transpose_x_0, transpose_y = var_2236_transpose_y_0, x = A_17_cast_fp16, y = vnew_13_cast_fp16)[name = string("op_2236_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2238_cast_fp16 = add(x = var_2235_cast_fp16, y = var_2236_cast_fp16)[name = string("op_2238_cast_fp16")];
            tensor<fp16, [16]> var_2239_bias_0_to_fp16 = const()[name = string("op_2239_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_2239_cast_fp16 = linear(bias = var_2239_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_21_cast_fp16)[name = string("op_2239_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_15_cast_fp16 = exp(x = var_2239_cast_fp16)[name = string("tail_15_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_29_to_fp16 = const()[name = string("transpose_29_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_2241_bias_0_to_fp16 = const()[name = string("op_2241_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_2241_cast_fp16 = linear(bias = var_2241_bias_0_to_fp16, weight = transpose_29_to_fp16, x = gdec_21_cast_fp16)[name = string("op_2241_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_2242_cast_fp16 = exp(x = var_2241_cast_fp16)[name = string("op_2242_cast_fp16")];
            tensor<int32, [1]> last_13_axes_0 = const()[name = string("last_13_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_13_cast_fp16 = expand_dims(axes = last_13_axes_0, x = var_2242_cast_fp16)[name = string("last_13_cast_fp16")];
            tensor<int32, [1]> var_2246_axes_0 = const()[name = string("op_2246_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_2246_cast_fp16 = expand_dims(axes = var_2246_axes_0, x = tail_15_cast_fp16)[name = string("op_2246_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2247_cast_fp16 = mul(x = k_33_cast_fp16, y = var_2246_cast_fp16)[name = string("op_2247_cast_fp16")];
            bool upd_11_transpose_x_1 = const()[name = string("upd_11_transpose_x_1"), val = bool(true)];
            bool upd_11_transpose_y_1 = const()[name = string("upd_11_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_11_cast_fp16 = matmul(transpose_x = upd_11_transpose_x_1, transpose_y = upd_11_transpose_y_1, x = var_2247_cast_fp16, y = vnew_13_cast_fp16)[name = string("upd_11_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_2252_cast_fp16 = mul(x = S_13_cast_fp16, y = last_13_cast_fp16)[name = string("op_2252_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_15_cast_fp16 = add(x = var_2252_cast_fp16, y = upd_11_cast_fp16)[name = string("S_15_cast_fp16")];
            tensor<int32, [4]> q_35_begin_0 = const()[name = string("q_35_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_35_end_0 = const()[name = string("q_35_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> q_35_end_mask_0 = const()[name = string("q_35_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_35_cast_fp16 = slice_by_index(begin = q_35_begin_0, end = q_35_end_0, end_mask = q_35_end_mask_0, x = q_27_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<int32, [4]> k_35_begin_0 = const()[name = string("k_35_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_35_end_0 = const()[name = string("k_35_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> k_35_end_mask_0 = const()[name = string("k_35_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_35_cast_fp16 = slice_by_index(begin = k_35_begin_0, end = k_35_end_0, end_mask = k_35_end_mask_0, x = k_27_cast_fp16)[name = string("k_35_cast_fp16")];
            tensor<int32, [4]> kb_19_begin_0 = const()[name = string("kb_19_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_19_end_0 = const()[name = string("kb_19_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> kb_19_end_mask_0 = const()[name = string("kb_19_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_19_cast_fp16 = slice_by_index(begin = kb_19_begin_0, end = kb_19_end_0, end_mask = kb_19_end_mask_0, x = var_1847_cast_fp16)[name = string("kb_19_cast_fp16")];
            tensor<int32, [4]> vb_23_begin_0 = const()[name = string("vb_23_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_23_end_0 = const()[name = string("vb_23_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> vb_23_end_mask_0 = const()[name = string("vb_23_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_23_cast_fp16 = slice_by_index(begin = vb_23_begin_0, end = vb_23_end_0, end_mask = vb_23_end_mask_0, x = vb_15_cast_fp16)[name = string("vb_23_cast_fp16")];
            tensor<int32, [3]> gdec_23_begin_0 = const()[name = string("gdec_23_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_23_end_0 = const()[name = string("gdec_23_end_0"), val = tensor<int32, [3]>([1, 4, 64])];
            tensor<bool, [3]> gdec_23_end_mask_0 = const()[name = string("gdec_23_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_23_cast_fp16 = slice_by_index(begin = gdec_23_begin_0, end = gdec_23_end_0, end_mask = gdec_23_end_mask_0, x = gdec_15_cast_fp16)[name = string("gdec_23_cast_fp16")];
            tensor<int32, [2]> var_2335 = const()[name = string("op_2335"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_2336_cast_fp16 = reshape(shape = var_2335, x = gdec_23_cast_fp16)[name = string("op_2336_cast_fp16")];
            tensor<fp16, [4, 256]> var_2337_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2336_cast_fp16)[name = string("op_2337_cast_fp16")];
            tensor<int32, [4]> var_2342 = const()[name = string("op_2342"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_19_cast_fp16 = reshape(shape = var_2342, x = var_2337_cast_fp16)[name = string("P_19_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2344_cast_fp16 = exp(x = P_19_cast_fp16)[name = string("op_2344_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_19_cast_fp16 = mul(x = var_2344_cast_fp16, y = tril_incl_to_fp16)[name = string("D_19_cast_fp16")];
            bool kk_19_transpose_x_1 = const()[name = string("kk_19_transpose_x_1"), val = bool(false)];
            bool kk_19_transpose_y_1 = const()[name = string("kk_19_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_19_cast_fp16 = matmul(transpose_x = kk_19_transpose_x_1, transpose_y = kk_19_transpose_y_1, x = kb_19_cast_fp16, y = k_35_cast_fp16)[name = string("kk_19_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_2350_cast_fp16 = mul(x = kk_19_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_2350_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2351_cast_fp16 = mul(x = var_2350_cast_fp16, y = D_19_cast_fp16)[name = string("op_2351_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_19_cast_fp16 = mul(x = var_2351_cast_fp16, y = cum_tail_to_fp16)[name = string("N_19_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2353_cast_fp16 = mul(x = D_19_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2353_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_19_cast_fp16 = mul(x = kk_19_cast_fp16, y = var_2353_cast_fp16)[name = string("Na_19_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_73_cast_fp16 = add(x = eye_to_fp16, y = N_19_cast_fp16)[name = string("Minv_73_cast_fp16")];
            bool Np_55_transpose_x_0 = const()[name = string("Np_55_transpose_x_0"), val = bool(false)];
            bool Np_55_transpose_y_0 = const()[name = string("Np_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_55_cast_fp16 = matmul(transpose_x = Np_55_transpose_x_0, transpose_y = Np_55_transpose_y_0, x = N_19_cast_fp16, y = Na_19_cast_fp16)[name = string("Np_55_cast_fp16")];
            bool Npa_37_transpose_x_0 = const()[name = string("Npa_37_transpose_x_0"), val = bool(false)];
            bool Npa_37_transpose_y_0 = const()[name = string("Npa_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_37_cast_fp16 = matmul(transpose_x = Npa_37_transpose_x_0, transpose_y = Npa_37_transpose_y_0, x = Na_19_cast_fp16, y = N_19_cast_fp16)[name = string("Npa_37_cast_fp16")];
            bool var_2359_transpose_x_0 = const()[name = string("op_2359_transpose_x_0"), val = bool(false)];
            bool var_2359_transpose_y_0 = const()[name = string("op_2359_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2359_cast_fp16 = matmul(transpose_x = var_2359_transpose_x_0, transpose_y = var_2359_transpose_y_0, x = Minv_73_cast_fp16, y = Np_55_cast_fp16)[name = string("op_2359_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_75_cast_fp16 = add(x = Minv_73_cast_fp16, y = var_2359_cast_fp16)[name = string("Minv_75_cast_fp16")];
            bool Np_57_transpose_x_0 = const()[name = string("Np_57_transpose_x_0"), val = bool(false)];
            bool Np_57_transpose_y_0 = const()[name = string("Np_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_57_cast_fp16 = matmul(transpose_x = Np_57_transpose_x_0, transpose_y = Np_57_transpose_y_0, x = Np_55_cast_fp16, y = Npa_37_cast_fp16)[name = string("Np_57_cast_fp16")];
            bool Npa_39_transpose_x_0 = const()[name = string("Npa_39_transpose_x_0"), val = bool(false)];
            bool Npa_39_transpose_y_0 = const()[name = string("Npa_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_39_cast_fp16 = matmul(transpose_x = Npa_39_transpose_x_0, transpose_y = Npa_39_transpose_y_0, x = Npa_37_cast_fp16, y = Np_55_cast_fp16)[name = string("Npa_39_cast_fp16")];
            bool var_2364_transpose_x_0 = const()[name = string("op_2364_transpose_x_0"), val = bool(false)];
            bool var_2364_transpose_y_0 = const()[name = string("op_2364_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2364_cast_fp16 = matmul(transpose_x = var_2364_transpose_x_0, transpose_y = var_2364_transpose_y_0, x = Minv_75_cast_fp16, y = Np_57_cast_fp16)[name = string("op_2364_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_77_cast_fp16 = add(x = Minv_75_cast_fp16, y = var_2364_cast_fp16)[name = string("Minv_77_cast_fp16")];
            bool Np_59_transpose_x_0 = const()[name = string("Np_59_transpose_x_0"), val = bool(false)];
            bool Np_59_transpose_y_0 = const()[name = string("Np_59_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_59_cast_fp16 = matmul(transpose_x = Np_59_transpose_x_0, transpose_y = Np_59_transpose_y_0, x = Np_57_cast_fp16, y = Npa_39_cast_fp16)[name = string("Np_59_cast_fp16")];
            bool var_2368_transpose_x_0 = const()[name = string("op_2368_transpose_x_0"), val = bool(false)];
            bool var_2368_transpose_y_0 = const()[name = string("op_2368_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2368_cast_fp16 = matmul(transpose_x = var_2368_transpose_x_0, transpose_y = var_2368_transpose_y_0, x = Minv_77_cast_fp16, y = Np_59_cast_fp16)[name = string("op_2368_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_79_cast_fp16 = add(x = Minv_77_cast_fp16, y = var_2368_cast_fp16)[name = string("Minv_79_cast_fp16")];
            bool u_19_transpose_x_0 = const()[name = string("u_19_transpose_x_0"), val = bool(false)];
            bool u_19_transpose_y_0 = const()[name = string("u_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_19_cast_fp16 = matmul(transpose_x = u_19_transpose_x_0, transpose_y = u_19_transpose_y_0, x = Minv_79_cast_fp16, y = vb_23_cast_fp16)[name = string("u_19_cast_fp16")];
            bool var_2375_transpose_x_1 = const()[name = string("op_2375_transpose_x_1"), val = bool(false)];
            bool var_2375_transpose_y_1 = const()[name = string("op_2375_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_2375_cast_fp16 = matmul(transpose_x = var_2375_transpose_x_1, transpose_y = var_2375_transpose_y_1, x = q_35_cast_fp16, y = k_35_cast_fp16)[name = string("op_2375_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_19_cast_fp16 = mul(x = var_2375_cast_fp16, y = D_19_cast_fp16)[name = string("A_19_cast_fp16")];
            tensor<fp16, [16]> var_2377_bias_0_to_fp16 = const()[name = string("op_2377_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_2377_cast_fp16 = linear(bias = var_2377_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_23_cast_fp16)[name = string("op_2377_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_15_cast_fp16 = exp(x = var_2377_cast_fp16)[name = string("cum_15_cast_fp16")];
            tensor<int32, [1]> var_2380_axes_0 = const()[name = string("op_2380_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_2380_cast_fp16 = expand_dims(axes = var_2380_axes_0, x = cum_15_cast_fp16)[name = string("op_2380_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2381_cast_fp16 = mul(x = kb_19_cast_fp16, y = var_2380_cast_fp16)[name = string("op_2381_cast_fp16")];
            bool w_19_transpose_x_0 = const()[name = string("w_19_transpose_x_0"), val = bool(false)];
            bool w_19_transpose_y_0 = const()[name = string("w_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_19_cast_fp16 = matmul(transpose_x = w_19_transpose_x_0, transpose_y = w_19_transpose_y_0, x = Minv_79_cast_fp16, y = var_2381_cast_fp16)[name = string("w_19_cast_fp16")];
            bool var_2383_transpose_x_0 = const()[name = string("op_2383_transpose_x_0"), val = bool(false)];
            bool var_2383_transpose_y_0 = const()[name = string("op_2383_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2383_cast_fp16 = matmul(transpose_x = var_2383_transpose_x_0, transpose_y = var_2383_transpose_y_0, x = w_19_cast_fp16, y = S_15_cast_fp16)[name = string("op_2383_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_15_cast_fp16 = sub(x = u_19_cast_fp16, y = var_2383_cast_fp16)[name = string("vnew_15_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2388_cast_fp16 = mul(x = q_35_cast_fp16, y = var_2380_cast_fp16)[name = string("op_2388_cast_fp16")];
            bool var_2389_transpose_x_0 = const()[name = string("op_2389_transpose_x_0"), val = bool(false)];
            bool var_2389_transpose_y_0 = const()[name = string("op_2389_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2389_cast_fp16 = matmul(transpose_x = var_2389_transpose_x_0, transpose_y = var_2389_transpose_y_0, x = var_2388_cast_fp16, y = S_15_cast_fp16)[name = string("op_2389_cast_fp16")];
            bool var_2390_transpose_x_0 = const()[name = string("op_2390_transpose_x_0"), val = bool(false)];
            bool var_2390_transpose_y_0 = const()[name = string("op_2390_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2390_cast_fp16 = matmul(transpose_x = var_2390_transpose_x_0, transpose_y = var_2390_transpose_y_0, x = A_19_cast_fp16, y = vnew_15_cast_fp16)[name = string("op_2390_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2392_cast_fp16 = add(x = var_2389_cast_fp16, y = var_2390_cast_fp16)[name = string("op_2392_cast_fp16")];
            tensor<fp16, [16]> var_2393_bias_0_to_fp16 = const()[name = string("op_2393_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_2393_cast_fp16 = linear(bias = var_2393_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_23_cast_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_17_cast_fp16 = exp(x = var_2393_cast_fp16)[name = string("tail_17_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_33_to_fp16 = const()[name = string("transpose_33_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_2395_bias_0_to_fp16 = const()[name = string("op_2395_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_2395_cast_fp16 = linear(bias = var_2395_bias_0_to_fp16, weight = transpose_33_to_fp16, x = gdec_23_cast_fp16)[name = string("op_2395_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_2396_cast_fp16 = exp(x = var_2395_cast_fp16)[name = string("op_2396_cast_fp16")];
            tensor<int32, [1]> last_15_axes_0 = const()[name = string("last_15_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_15_cast_fp16 = expand_dims(axes = last_15_axes_0, x = var_2396_cast_fp16)[name = string("last_15_cast_fp16")];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_2400_cast_fp16 = expand_dims(axes = var_2400_axes_0, x = tail_17_cast_fp16)[name = string("op_2400_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2401_cast_fp16 = mul(x = k_35_cast_fp16, y = var_2400_cast_fp16)[name = string("op_2401_cast_fp16")];
            bool upd_13_transpose_x_1 = const()[name = string("upd_13_transpose_x_1"), val = bool(true)];
            bool upd_13_transpose_y_1 = const()[name = string("upd_13_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_13_cast_fp16 = matmul(transpose_x = upd_13_transpose_x_1, transpose_y = upd_13_transpose_y_1, x = var_2401_cast_fp16, y = vnew_15_cast_fp16)[name = string("upd_13_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_2406_cast_fp16 = mul(x = S_15_cast_fp16, y = last_15_cast_fp16)[name = string("op_2406_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_17_cast_fp16 = add(x = var_2406_cast_fp16, y = upd_13_cast_fp16)[name = string("S_17_cast_fp16")];
            tensor<int32, [4]> q_37_begin_0 = const()[name = string("q_37_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> q_37_end_0 = const()[name = string("q_37_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> q_37_end_mask_0 = const()[name = string("q_37_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, x = q_27_cast_fp16)[name = string("q_37_cast_fp16")];
            tensor<int32, [4]> k_37_begin_0 = const()[name = string("k_37_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> k_37_end_0 = const()[name = string("k_37_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> k_37_end_mask_0 = const()[name = string("k_37_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, x = k_27_cast_fp16)[name = string("k_37_cast_fp16")];
            tensor<int32, [4]> kb_21_begin_0 = const()[name = string("kb_21_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> kb_21_end_0 = const()[name = string("kb_21_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> kb_21_end_mask_0 = const()[name = string("kb_21_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_21_cast_fp16 = slice_by_index(begin = kb_21_begin_0, end = kb_21_end_0, end_mask = kb_21_end_mask_0, x = var_1847_cast_fp16)[name = string("kb_21_cast_fp16")];
            tensor<int32, [4]> vb_25_begin_0 = const()[name = string("vb_25_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> vb_25_end_0 = const()[name = string("vb_25_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> vb_25_end_mask_0 = const()[name = string("vb_25_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_25_cast_fp16 = slice_by_index(begin = vb_25_begin_0, end = vb_25_end_0, end_mask = vb_25_end_mask_0, x = vb_15_cast_fp16)[name = string("vb_25_cast_fp16")];
            tensor<int32, [3]> gdec_25_begin_0 = const()[name = string("gdec_25_begin_0"), val = tensor<int32, [3]>([0, 0, 64])];
            tensor<int32, [3]> gdec_25_end_0 = const()[name = string("gdec_25_end_0"), val = tensor<int32, [3]>([1, 4, 80])];
            tensor<bool, [3]> gdec_25_end_mask_0 = const()[name = string("gdec_25_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_25_cast_fp16 = slice_by_index(begin = gdec_25_begin_0, end = gdec_25_end_0, end_mask = gdec_25_end_mask_0, x = gdec_15_cast_fp16)[name = string("gdec_25_cast_fp16")];
            tensor<int32, [2]> var_2489 = const()[name = string("op_2489"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_2490_cast_fp16 = reshape(shape = var_2489, x = gdec_25_cast_fp16)[name = string("op_2490_cast_fp16")];
            tensor<fp16, [4, 256]> var_2491_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2490_cast_fp16)[name = string("op_2491_cast_fp16")];
            tensor<int32, [4]> var_2496 = const()[name = string("op_2496"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_21_cast_fp16 = reshape(shape = var_2496, x = var_2491_cast_fp16)[name = string("P_21_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2498_cast_fp16 = exp(x = P_21_cast_fp16)[name = string("op_2498_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_21_cast_fp16 = mul(x = var_2498_cast_fp16, y = tril_incl_to_fp16)[name = string("D_21_cast_fp16")];
            bool kk_21_transpose_x_1 = const()[name = string("kk_21_transpose_x_1"), val = bool(false)];
            bool kk_21_transpose_y_1 = const()[name = string("kk_21_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_21_cast_fp16 = matmul(transpose_x = kk_21_transpose_x_1, transpose_y = kk_21_transpose_y_1, x = kb_21_cast_fp16, y = k_37_cast_fp16)[name = string("kk_21_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_2504_cast_fp16 = mul(x = kk_21_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_2504_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2505_cast_fp16 = mul(x = var_2504_cast_fp16, y = D_21_cast_fp16)[name = string("op_2505_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_21_cast_fp16 = mul(x = var_2505_cast_fp16, y = cum_tail_to_fp16)[name = string("N_21_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2507_cast_fp16 = mul(x = D_21_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2507_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_21_cast_fp16 = mul(x = kk_21_cast_fp16, y = var_2507_cast_fp16)[name = string("Na_21_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_81_cast_fp16 = add(x = eye_to_fp16, y = N_21_cast_fp16)[name = string("Minv_81_cast_fp16")];
            bool Np_61_transpose_x_0 = const()[name = string("Np_61_transpose_x_0"), val = bool(false)];
            bool Np_61_transpose_y_0 = const()[name = string("Np_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_61_cast_fp16 = matmul(transpose_x = Np_61_transpose_x_0, transpose_y = Np_61_transpose_y_0, x = N_21_cast_fp16, y = Na_21_cast_fp16)[name = string("Np_61_cast_fp16")];
            bool Npa_41_transpose_x_0 = const()[name = string("Npa_41_transpose_x_0"), val = bool(false)];
            bool Npa_41_transpose_y_0 = const()[name = string("Npa_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_41_cast_fp16 = matmul(transpose_x = Npa_41_transpose_x_0, transpose_y = Npa_41_transpose_y_0, x = Na_21_cast_fp16, y = N_21_cast_fp16)[name = string("Npa_41_cast_fp16")];
            bool var_2513_transpose_x_0 = const()[name = string("op_2513_transpose_x_0"), val = bool(false)];
            bool var_2513_transpose_y_0 = const()[name = string("op_2513_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2513_cast_fp16 = matmul(transpose_x = var_2513_transpose_x_0, transpose_y = var_2513_transpose_y_0, x = Minv_81_cast_fp16, y = Np_61_cast_fp16)[name = string("op_2513_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_83_cast_fp16 = add(x = Minv_81_cast_fp16, y = var_2513_cast_fp16)[name = string("Minv_83_cast_fp16")];
            bool Np_63_transpose_x_0 = const()[name = string("Np_63_transpose_x_0"), val = bool(false)];
            bool Np_63_transpose_y_0 = const()[name = string("Np_63_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_63_cast_fp16 = matmul(transpose_x = Np_63_transpose_x_0, transpose_y = Np_63_transpose_y_0, x = Np_61_cast_fp16, y = Npa_41_cast_fp16)[name = string("Np_63_cast_fp16")];
            bool Npa_43_transpose_x_0 = const()[name = string("Npa_43_transpose_x_0"), val = bool(false)];
            bool Npa_43_transpose_y_0 = const()[name = string("Npa_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_43_cast_fp16 = matmul(transpose_x = Npa_43_transpose_x_0, transpose_y = Npa_43_transpose_y_0, x = Npa_41_cast_fp16, y = Np_61_cast_fp16)[name = string("Npa_43_cast_fp16")];
            bool var_2518_transpose_x_0 = const()[name = string("op_2518_transpose_x_0"), val = bool(false)];
            bool var_2518_transpose_y_0 = const()[name = string("op_2518_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2518_cast_fp16 = matmul(transpose_x = var_2518_transpose_x_0, transpose_y = var_2518_transpose_y_0, x = Minv_83_cast_fp16, y = Np_63_cast_fp16)[name = string("op_2518_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_85_cast_fp16 = add(x = Minv_83_cast_fp16, y = var_2518_cast_fp16)[name = string("Minv_85_cast_fp16")];
            bool Np_65_transpose_x_0 = const()[name = string("Np_65_transpose_x_0"), val = bool(false)];
            bool Np_65_transpose_y_0 = const()[name = string("Np_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_65_cast_fp16 = matmul(transpose_x = Np_65_transpose_x_0, transpose_y = Np_65_transpose_y_0, x = Np_63_cast_fp16, y = Npa_43_cast_fp16)[name = string("Np_65_cast_fp16")];
            bool var_2522_transpose_x_0 = const()[name = string("op_2522_transpose_x_0"), val = bool(false)];
            bool var_2522_transpose_y_0 = const()[name = string("op_2522_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2522_cast_fp16 = matmul(transpose_x = var_2522_transpose_x_0, transpose_y = var_2522_transpose_y_0, x = Minv_85_cast_fp16, y = Np_65_cast_fp16)[name = string("op_2522_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_87_cast_fp16 = add(x = Minv_85_cast_fp16, y = var_2522_cast_fp16)[name = string("Minv_87_cast_fp16")];
            bool u_21_transpose_x_0 = const()[name = string("u_21_transpose_x_0"), val = bool(false)];
            bool u_21_transpose_y_0 = const()[name = string("u_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_21_cast_fp16 = matmul(transpose_x = u_21_transpose_x_0, transpose_y = u_21_transpose_y_0, x = Minv_87_cast_fp16, y = vb_25_cast_fp16)[name = string("u_21_cast_fp16")];
            bool var_2529_transpose_x_1 = const()[name = string("op_2529_transpose_x_1"), val = bool(false)];
            bool var_2529_transpose_y_1 = const()[name = string("op_2529_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_2529_cast_fp16 = matmul(transpose_x = var_2529_transpose_x_1, transpose_y = var_2529_transpose_y_1, x = q_37_cast_fp16, y = k_37_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_21_cast_fp16 = mul(x = var_2529_cast_fp16, y = D_21_cast_fp16)[name = string("A_21_cast_fp16")];
            tensor<fp16, [16]> var_2531_bias_0_to_fp16 = const()[name = string("op_2531_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_2531_cast_fp16 = linear(bias = var_2531_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_25_cast_fp16)[name = string("op_2531_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_17_cast_fp16 = exp(x = var_2531_cast_fp16)[name = string("cum_17_cast_fp16")];
            tensor<int32, [1]> var_2534_axes_0 = const()[name = string("op_2534_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_2534_cast_fp16 = expand_dims(axes = var_2534_axes_0, x = cum_17_cast_fp16)[name = string("op_2534_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2535_cast_fp16 = mul(x = kb_21_cast_fp16, y = var_2534_cast_fp16)[name = string("op_2535_cast_fp16")];
            bool w_21_transpose_x_0 = const()[name = string("w_21_transpose_x_0"), val = bool(false)];
            bool w_21_transpose_y_0 = const()[name = string("w_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_21_cast_fp16 = matmul(transpose_x = w_21_transpose_x_0, transpose_y = w_21_transpose_y_0, x = Minv_87_cast_fp16, y = var_2535_cast_fp16)[name = string("w_21_cast_fp16")];
            bool var_2537_transpose_x_0 = const()[name = string("op_2537_transpose_x_0"), val = bool(false)];
            bool var_2537_transpose_y_0 = const()[name = string("op_2537_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2537_cast_fp16 = matmul(transpose_x = var_2537_transpose_x_0, transpose_y = var_2537_transpose_y_0, x = w_21_cast_fp16, y = S_17_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_17_cast_fp16 = sub(x = u_21_cast_fp16, y = var_2537_cast_fp16)[name = string("vnew_17_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2542_cast_fp16 = mul(x = q_37_cast_fp16, y = var_2534_cast_fp16)[name = string("op_2542_cast_fp16")];
            bool var_2543_transpose_x_0 = const()[name = string("op_2543_transpose_x_0"), val = bool(false)];
            bool var_2543_transpose_y_0 = const()[name = string("op_2543_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2543_cast_fp16 = matmul(transpose_x = var_2543_transpose_x_0, transpose_y = var_2543_transpose_y_0, x = var_2542_cast_fp16, y = S_17_cast_fp16)[name = string("op_2543_cast_fp16")];
            bool var_2544_transpose_x_0 = const()[name = string("op_2544_transpose_x_0"), val = bool(false)];
            bool var_2544_transpose_y_0 = const()[name = string("op_2544_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2544_cast_fp16 = matmul(transpose_x = var_2544_transpose_x_0, transpose_y = var_2544_transpose_y_0, x = A_21_cast_fp16, y = vnew_17_cast_fp16)[name = string("op_2544_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2546_cast_fp16 = add(x = var_2543_cast_fp16, y = var_2544_cast_fp16)[name = string("op_2546_cast_fp16")];
            tensor<fp16, [16]> var_2547_bias_0_to_fp16 = const()[name = string("op_2547_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_2547_cast_fp16 = linear(bias = var_2547_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_25_cast_fp16)[name = string("op_2547_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_19_cast_fp16 = exp(x = var_2547_cast_fp16)[name = string("tail_19_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_37_to_fp16 = const()[name = string("transpose_37_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_2549_bias_0_to_fp16 = const()[name = string("op_2549_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_2549_cast_fp16 = linear(bias = var_2549_bias_0_to_fp16, weight = transpose_37_to_fp16, x = gdec_25_cast_fp16)[name = string("op_2549_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_2550_cast_fp16 = exp(x = var_2549_cast_fp16)[name = string("op_2550_cast_fp16")];
            tensor<int32, [1]> last_17_axes_0 = const()[name = string("last_17_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_17_cast_fp16 = expand_dims(axes = last_17_axes_0, x = var_2550_cast_fp16)[name = string("last_17_cast_fp16")];
            tensor<int32, [1]> var_2554_axes_0 = const()[name = string("op_2554_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_2554_cast_fp16 = expand_dims(axes = var_2554_axes_0, x = tail_19_cast_fp16)[name = string("op_2554_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2555_cast_fp16 = mul(x = k_37_cast_fp16, y = var_2554_cast_fp16)[name = string("op_2555_cast_fp16")];
            bool upd_15_transpose_x_1 = const()[name = string("upd_15_transpose_x_1"), val = bool(true)];
            bool upd_15_transpose_y_1 = const()[name = string("upd_15_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_15_cast_fp16 = matmul(transpose_x = upd_15_transpose_x_1, transpose_y = upd_15_transpose_y_1, x = var_2555_cast_fp16, y = vnew_17_cast_fp16)[name = string("upd_15_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_2560_cast_fp16 = mul(x = S_17_cast_fp16, y = last_17_cast_fp16)[name = string("op_2560_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_19_cast_fp16 = add(x = var_2560_cast_fp16, y = upd_15_cast_fp16)[name = string("S_19_cast_fp16")];
            tensor<int32, [4]> q_39_begin_0 = const()[name = string("q_39_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> q_39_end_0 = const()[name = string("q_39_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_39_end_mask_0 = const()[name = string("q_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_39_cast_fp16 = slice_by_index(begin = q_39_begin_0, end = q_39_end_0, end_mask = q_39_end_mask_0, x = q_27_cast_fp16)[name = string("q_39_cast_fp16")];
            tensor<int32, [4]> k_39_begin_0 = const()[name = string("k_39_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> k_39_end_0 = const()[name = string("k_39_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_39_end_mask_0 = const()[name = string("k_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_39_cast_fp16 = slice_by_index(begin = k_39_begin_0, end = k_39_end_0, end_mask = k_39_end_mask_0, x = k_27_cast_fp16)[name = string("k_39_cast_fp16")];
            tensor<int32, [4]> kb_23_begin_0 = const()[name = string("kb_23_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> kb_23_end_0 = const()[name = string("kb_23_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_23_end_mask_0 = const()[name = string("kb_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_23_cast_fp16 = slice_by_index(begin = kb_23_begin_0, end = kb_23_end_0, end_mask = kb_23_end_mask_0, x = var_1847_cast_fp16)[name = string("kb_23_cast_fp16")];
            tensor<int32, [4]> vb_27_begin_0 = const()[name = string("vb_27_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> vb_27_end_0 = const()[name = string("vb_27_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_27_end_mask_0 = const()[name = string("vb_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_27_cast_fp16 = slice_by_index(begin = vb_27_begin_0, end = vb_27_end_0, end_mask = vb_27_end_mask_0, x = vb_15_cast_fp16)[name = string("vb_27_cast_fp16")];
            tensor<int32, [3]> gdec_27_begin_0 = const()[name = string("gdec_27_begin_0"), val = tensor<int32, [3]>([0, 0, 80])];
            tensor<int32, [3]> gdec_27_end_0 = const()[name = string("gdec_27_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_27_end_mask_0 = const()[name = string("gdec_27_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_27_cast_fp16 = slice_by_index(begin = gdec_27_begin_0, end = gdec_27_end_0, end_mask = gdec_27_end_mask_0, x = gdec_15_cast_fp16)[name = string("gdec_27_cast_fp16")];
            tensor<int32, [2]> var_2643 = const()[name = string("op_2643"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_2644_cast_fp16 = reshape(shape = var_2643, x = gdec_27_cast_fp16)[name = string("op_2644_cast_fp16")];
            tensor<fp16, [4, 256]> var_2645_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2644_cast_fp16)[name = string("op_2645_cast_fp16")];
            tensor<int32, [4]> var_2650 = const()[name = string("op_2650"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_23_cast_fp16 = reshape(shape = var_2650, x = var_2645_cast_fp16)[name = string("P_23_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2652_cast_fp16 = exp(x = P_23_cast_fp16)[name = string("op_2652_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_23_cast_fp16 = mul(x = var_2652_cast_fp16, y = tril_incl_to_fp16)[name = string("D_23_cast_fp16")];
            bool kk_23_transpose_x_1 = const()[name = string("kk_23_transpose_x_1"), val = bool(false)];
            bool kk_23_transpose_y_1 = const()[name = string("kk_23_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_23_cast_fp16 = matmul(transpose_x = kk_23_transpose_x_1, transpose_y = kk_23_transpose_y_1, x = kb_23_cast_fp16, y = k_39_cast_fp16)[name = string("kk_23_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_2658_cast_fp16 = mul(x = kk_23_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_2658_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2659_cast_fp16 = mul(x = var_2658_cast_fp16, y = D_23_cast_fp16)[name = string("op_2659_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_23_cast_fp16 = mul(x = var_2659_cast_fp16, y = cum_tail_to_fp16)[name = string("N_23_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_2661_cast_fp16 = mul(x = D_23_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2661_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_23_cast_fp16 = mul(x = kk_23_cast_fp16, y = var_2661_cast_fp16)[name = string("Na_23_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_89_cast_fp16 = add(x = eye_to_fp16, y = N_23_cast_fp16)[name = string("Minv_89_cast_fp16")];
            bool Np_67_transpose_x_0 = const()[name = string("Np_67_transpose_x_0"), val = bool(false)];
            bool Np_67_transpose_y_0 = const()[name = string("Np_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_67_cast_fp16 = matmul(transpose_x = Np_67_transpose_x_0, transpose_y = Np_67_transpose_y_0, x = N_23_cast_fp16, y = Na_23_cast_fp16)[name = string("Np_67_cast_fp16")];
            bool Npa_45_transpose_x_0 = const()[name = string("Npa_45_transpose_x_0"), val = bool(false)];
            bool Npa_45_transpose_y_0 = const()[name = string("Npa_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_45_cast_fp16 = matmul(transpose_x = Npa_45_transpose_x_0, transpose_y = Npa_45_transpose_y_0, x = Na_23_cast_fp16, y = N_23_cast_fp16)[name = string("Npa_45_cast_fp16")];
            bool var_2667_transpose_x_0 = const()[name = string("op_2667_transpose_x_0"), val = bool(false)];
            bool var_2667_transpose_y_0 = const()[name = string("op_2667_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2667_cast_fp16 = matmul(transpose_x = var_2667_transpose_x_0, transpose_y = var_2667_transpose_y_0, x = Minv_89_cast_fp16, y = Np_67_cast_fp16)[name = string("op_2667_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_91_cast_fp16 = add(x = Minv_89_cast_fp16, y = var_2667_cast_fp16)[name = string("Minv_91_cast_fp16")];
            bool Np_69_transpose_x_0 = const()[name = string("Np_69_transpose_x_0"), val = bool(false)];
            bool Np_69_transpose_y_0 = const()[name = string("Np_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_69_cast_fp16 = matmul(transpose_x = Np_69_transpose_x_0, transpose_y = Np_69_transpose_y_0, x = Np_67_cast_fp16, y = Npa_45_cast_fp16)[name = string("Np_69_cast_fp16")];
            bool Npa_47_transpose_x_0 = const()[name = string("Npa_47_transpose_x_0"), val = bool(false)];
            bool Npa_47_transpose_y_0 = const()[name = string("Npa_47_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_47_cast_fp16 = matmul(transpose_x = Npa_47_transpose_x_0, transpose_y = Npa_47_transpose_y_0, x = Npa_45_cast_fp16, y = Np_67_cast_fp16)[name = string("Npa_47_cast_fp16")];
            bool var_2672_transpose_x_0 = const()[name = string("op_2672_transpose_x_0"), val = bool(false)];
            bool var_2672_transpose_y_0 = const()[name = string("op_2672_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2672_cast_fp16 = matmul(transpose_x = var_2672_transpose_x_0, transpose_y = var_2672_transpose_y_0, x = Minv_91_cast_fp16, y = Np_69_cast_fp16)[name = string("op_2672_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_93_cast_fp16 = add(x = Minv_91_cast_fp16, y = var_2672_cast_fp16)[name = string("Minv_93_cast_fp16")];
            bool Np_71_transpose_x_0 = const()[name = string("Np_71_transpose_x_0"), val = bool(false)];
            bool Np_71_transpose_y_0 = const()[name = string("Np_71_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_71_cast_fp16 = matmul(transpose_x = Np_71_transpose_x_0, transpose_y = Np_71_transpose_y_0, x = Np_69_cast_fp16, y = Npa_47_cast_fp16)[name = string("Np_71_cast_fp16")];
            bool var_2676_transpose_x_0 = const()[name = string("op_2676_transpose_x_0"), val = bool(false)];
            bool var_2676_transpose_y_0 = const()[name = string("op_2676_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_2676_cast_fp16 = matmul(transpose_x = var_2676_transpose_x_0, transpose_y = var_2676_transpose_y_0, x = Minv_93_cast_fp16, y = Np_71_cast_fp16)[name = string("op_2676_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_95_cast_fp16 = add(x = Minv_93_cast_fp16, y = var_2676_cast_fp16)[name = string("Minv_95_cast_fp16")];
            bool u_23_transpose_x_0 = const()[name = string("u_23_transpose_x_0"), val = bool(false)];
            bool u_23_transpose_y_0 = const()[name = string("u_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_23_cast_fp16 = matmul(transpose_x = u_23_transpose_x_0, transpose_y = u_23_transpose_y_0, x = Minv_95_cast_fp16, y = vb_27_cast_fp16)[name = string("u_23_cast_fp16")];
            bool var_2683_transpose_x_1 = const()[name = string("op_2683_transpose_x_1"), val = bool(false)];
            bool var_2683_transpose_y_1 = const()[name = string("op_2683_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_2683_cast_fp16 = matmul(transpose_x = var_2683_transpose_x_1, transpose_y = var_2683_transpose_y_1, x = q_39_cast_fp16, y = k_39_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_23_cast_fp16 = mul(x = var_2683_cast_fp16, y = D_23_cast_fp16)[name = string("A_23_cast_fp16")];
            tensor<fp16, [16]> var_2685_bias_0_to_fp16 = const()[name = string("op_2685_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_2685_cast_fp16 = linear(bias = var_2685_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_27_cast_fp16)[name = string("op_2685_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_19_cast_fp16 = exp(x = var_2685_cast_fp16)[name = string("cum_19_cast_fp16")];
            tensor<int32, [1]> var_2688_axes_0 = const()[name = string("op_2688_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_2688_cast_fp16 = expand_dims(axes = var_2688_axes_0, x = cum_19_cast_fp16)[name = string("op_2688_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2689_cast_fp16 = mul(x = kb_23_cast_fp16, y = var_2688_cast_fp16)[name = string("op_2689_cast_fp16")];
            bool w_23_transpose_x_0 = const()[name = string("w_23_transpose_x_0"), val = bool(false)];
            bool w_23_transpose_y_0 = const()[name = string("w_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_23_cast_fp16 = matmul(transpose_x = w_23_transpose_x_0, transpose_y = w_23_transpose_y_0, x = Minv_95_cast_fp16, y = var_2689_cast_fp16)[name = string("w_23_cast_fp16")];
            bool var_2691_transpose_x_0 = const()[name = string("op_2691_transpose_x_0"), val = bool(false)];
            bool var_2691_transpose_y_0 = const()[name = string("op_2691_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2691_cast_fp16 = matmul(transpose_x = var_2691_transpose_x_0, transpose_y = var_2691_transpose_y_0, x = w_23_cast_fp16, y = S_19_cast_fp16)[name = string("op_2691_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_19_cast_fp16 = sub(x = u_23_cast_fp16, y = var_2691_cast_fp16)[name = string("vnew_19_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_2696_cast_fp16 = mul(x = q_39_cast_fp16, y = var_2688_cast_fp16)[name = string("op_2696_cast_fp16")];
            bool var_2697_transpose_x_0 = const()[name = string("op_2697_transpose_x_0"), val = bool(false)];
            bool var_2697_transpose_y_0 = const()[name = string("op_2697_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2697_cast_fp16 = matmul(transpose_x = var_2697_transpose_x_0, transpose_y = var_2697_transpose_y_0, x = var_2696_cast_fp16, y = S_19_cast_fp16)[name = string("op_2697_cast_fp16")];
            bool var_2698_transpose_x_0 = const()[name = string("op_2698_transpose_x_0"), val = bool(false)];
            bool var_2698_transpose_y_0 = const()[name = string("op_2698_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_2698_cast_fp16 = matmul(transpose_x = var_2698_transpose_x_0, transpose_y = var_2698_transpose_y_0, x = A_23_cast_fp16, y = vnew_19_cast_fp16)[name = string("op_2698_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_3_cast_fp16 = add(x = var_2697_cast_fp16, y = var_2698_cast_fp16)[name = string("oc_3_cast_fp16")];
            int32 var_2702 = const()[name = string("op_2702"), val = int32(2)];
            bool o_7_interleave_0 = const()[name = string("o_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 96, 128]> o_7_cast_fp16 = concat(axis = var_2702, interleave = o_7_interleave_0, values = (var_1937_cast_fp16, var_2084_cast_fp16, var_2238_cast_fp16, var_2392_cast_fp16, var_2546_cast_fp16, oc_3_cast_fp16))[name = string("o_7_cast_fp16")];
            tensor<int32, [4]> var_2708 = const()[name = string("op_2708"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2714 = const()[name = string("op_2714"), val = tensor<int32, [4]>([1, 96, 4, 128])];
            tensor<fp16, [1, 96, 4, 128]> input_23_cast_fp16 = reshape(shape = var_2714, x = z_3_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> o_9_cast_fp16 = transpose(perm = var_2708, x = o_7_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [1, 96, 4, 128]> var_2716_cast_fp16 = mul(x = o_9_cast_fp16, y = o_9_cast_fp16)[name = string("op_2716_cast_fp16")];
            tensor<int32, [1]> var_2721_axes_0 = const()[name = string("op_2721_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2721_keep_dims_0 = const()[name = string("op_2721_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 4, 1]> var_2721_cast_fp16 = reduce_mean(axes = var_2721_axes_0, keep_dims = var_2721_keep_dims_0, x = var_2716_cast_fp16)[name = string("op_2721_cast_fp16")];
            fp16 var_2723_to_fp16 = const()[name = string("op_2723_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 96, 4, 1]> var_2724_cast_fp16 = add(x = var_2721_cast_fp16, y = var_2723_to_fp16)[name = string("op_2724_cast_fp16")];
            fp32 var_2725_epsilon_0 = const()[name = string("op_2725_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 4, 1]> var_2725_cast_fp16 = rsqrt(epsilon = var_2725_epsilon_0, x = var_2724_cast_fp16)[name = string("op_2725_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> var_2726_cast_fp16 = mul(x = o_9_cast_fp16, y = var_2725_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> var_2728_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_2728_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> o_11_cast_fp16 = mul(x = var_2726_cast_fp16, y = var_2728_cast_fp16)[name = string("o_11_cast_fp16")];
            tensor<int32, [3]> var_2733 = const()[name = string("op_2733"), val = tensor<int32, [3]>([1, 96, 512])];
            tensor<fp16, [1, 96, 512]> input_25_cast_fp16 = reshape(shape = var_2733, x = o_11_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [640, 512]> layers_1_out_weight_to_fp16 = const()[name = string("layers_1_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34984256)))];
            tensor<fp16, [1, 96, 640]> linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_out_weight_to_fp16, x = input_25_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 96, 640]> x_13_cast_fp16 = add(x = x_9_cast_fp16, y = linear_5_cast_fp16)[name = string("x_13_cast_fp16")];
            fp16 var_2746_to_fp16 = const()[name = string("op_2746_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> x_15_cast_fp16 = mul(x = x_13_cast_fp16, y = var_2746_to_fp16)[name = string("x_15_cast_fp16")];
            fp16 var_2744_promoted_to_fp16 = const()[name = string("op_2744_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> var_2748_cast_fp16 = pow(x = x_15_cast_fp16, y = var_2744_promoted_to_fp16)[name = string("op_2748_cast_fp16")];
            tensor<int32, [1]> var_2750_axes_0 = const()[name = string("op_2750_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2750_keep_dims_0 = const()[name = string("op_2750_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 1]> var_2750_cast_fp16 = reduce_mean(axes = var_2750_axes_0, keep_dims = var_2750_keep_dims_0, x = var_2748_cast_fp16)[name = string("op_2750_cast_fp16")];
            fp16 var_2751_to_fp16 = const()[name = string("op_2751_to_fp16"), val = fp16(0x1.0cp-18)];
            tensor<fp16, [1, 96, 1]> var_2752_cast_fp16 = add(x = var_2750_cast_fp16, y = var_2751_to_fp16)[name = string("op_2752_cast_fp16")];
            fp32 var_2753_epsilon_0 = const()[name = string("op_2753_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 1]> var_2753_cast_fp16 = rsqrt(epsilon = var_2753_epsilon_0, x = var_2752_cast_fp16)[name = string("op_2753_cast_fp16")];
            tensor<fp16, [1, 96, 640]> var_2754_cast_fp16 = mul(x = x_15_cast_fp16, y = var_2753_cast_fp16)[name = string("op_2754_cast_fp16")];
            tensor<fp16, [640]> layers_1_n2_w1_to_fp16 = const()[name = string("layers_1_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35639680)))];
            tensor<fp16, [1, 96, 640]> input_27_cast_fp16 = mul(x = var_2754_cast_fp16, y = layers_1_n2_w1_to_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_1_gate_up_weight_to_fp16 = const()[name = string("layers_1_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35641024)))];
            tensor<fp16, [1, 96, 5120]> linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_1_gate_up_weight_to_fp16, x = input_27_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<int32, [3]> input_29_begin_0 = const()[name = string("input_29_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_29_end_0 = const()[name = string("input_29_end_0"), val = tensor<int32, [3]>([1, 96, 2560])];
            tensor<bool, [3]> input_29_end_mask_0 = const()[name = string("input_29_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 2560]> input_29_cast_fp16 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = linear_6_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [1, 96, 2560]> var_2764_cast_fp16 = silu(x = input_29_cast_fp16)[name = string("op_2764_cast_fp16")];
            tensor<int32, [3]> var_2769_begin_0 = const()[name = string("op_2769_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_2769_end_0 = const()[name = string("op_2769_end_0"), val = tensor<int32, [3]>([1, 96, 5120])];
            tensor<bool, [3]> var_2769_end_mask_0 = const()[name = string("op_2769_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 2560]> var_2769_cast_fp16 = slice_by_index(begin = var_2769_begin_0, end = var_2769_end_0, end_mask = var_2769_end_mask_0, x = linear_6_cast_fp16)[name = string("op_2769_cast_fp16")];
            tensor<fp16, [1, 96, 2560]> input_31_cast_fp16 = mul(x = var_2764_cast_fp16, y = var_2769_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_1_down_weight_to_fp16 = const()[name = string("layers_1_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42194688)))];
            tensor<fp16, [1, 96, 640]> linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_down_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 96, 640]> x_17_cast_fp16 = add(x = x_13_cast_fp16, y = linear_7_cast_fp16)[name = string("x_17_cast_fp16")];
            fp16 var_2780_promoted_to_fp16 = const()[name = string("op_2780_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> var_2784_cast_fp16 = pow(x = x_17_cast_fp16, y = var_2780_promoted_to_fp16)[name = string("op_2784_cast_fp16")];
            tensor<int32, [1]> var_2786_axes_0 = const()[name = string("op_2786_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2786_keep_dims_0 = const()[name = string("op_2786_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 1]> var_2786_cast_fp16 = reduce_mean(axes = var_2786_axes_0, keep_dims = var_2786_keep_dims_0, x = var_2784_cast_fp16)[name = string("op_2786_cast_fp16")];
            fp16 var_2787_to_fp16 = const()[name = string("op_2787_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 96, 1]> var_2788_cast_fp16 = add(x = var_2786_cast_fp16, y = var_2787_to_fp16)[name = string("op_2788_cast_fp16")];
            fp32 var_2789_epsilon_0 = const()[name = string("op_2789_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 1]> var_2789_cast_fp16 = rsqrt(epsilon = var_2789_epsilon_0, x = var_2788_cast_fp16)[name = string("op_2789_cast_fp16")];
            tensor<fp16, [1, 96, 640]> var_2790_cast_fp16 = mul(x = x_17_cast_fp16, y = var_2789_cast_fp16)[name = string("op_2790_cast_fp16")];
            tensor<fp16, [640]> layers_2_n1_w1_to_fp16 = const()[name = string("layers_2_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45471552)))];
            tensor<fp16, [1, 96, 640]> input_33_cast_fp16 = mul(x = var_2790_cast_fp16, y = layers_2_n1_w1_to_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_2_qkvzba_weight_to_fp16 = const()[name = string("layers_2_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45472896)))];
            tensor<fp16, [1, 96, 1544]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_qkvzba_weight_to_fp16, x = input_33_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [3]> qkv_5_begin_0 = const()[name = string("qkv_5_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_5_end_0 = const()[name = string("qkv_5_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> qkv_5_end_mask_0 = const()[name = string("qkv_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 1024]> qkv_5_cast_fp16 = slice_by_index(begin = qkv_5_begin_0, end = qkv_5_end_0, end_mask = qkv_5_end_mask_0, x = linear_8_cast_fp16)[name = string("qkv_5_cast_fp16")];
            tensor<int32, [3]> z_5_begin_0 = const()[name = string("z_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_5_end_0 = const()[name = string("z_5_end_0"), val = tensor<int32, [3]>([1, 96, 1536])];
            tensor<bool, [3]> z_5_end_mask_0 = const()[name = string("z_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 512]> z_5_cast_fp16 = slice_by_index(begin = z_5_begin_0, end = z_5_end_0, end_mask = z_5_end_mask_0, x = linear_8_cast_fp16)[name = string("z_5_cast_fp16")];
            tensor<int32, [3]> b_5_begin_0 = const()[name = string("b_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_5_end_0 = const()[name = string("b_5_end_0"), val = tensor<int32, [3]>([1, 96, 1540])];
            tensor<bool, [3]> b_5_end_mask_0 = const()[name = string("b_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 4]> b_5_cast_fp16 = slice_by_index(begin = b_5_begin_0, end = b_5_end_0, end_mask = b_5_end_mask_0, x = linear_8_cast_fp16)[name = string("b_5_cast_fp16")];
            tensor<int32, [3]> a_5_begin_0 = const()[name = string("a_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_5_end_0 = const()[name = string("a_5_end_0"), val = tensor<int32, [3]>([1, 96, 1])];
            tensor<bool, [3]> a_5_end_mask_0 = const()[name = string("a_5_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 4]> a_5_cast_fp16 = slice_by_index(begin = a_5_begin_0, end = a_5_end_0, end_mask = a_5_end_mask_0, x = linear_8_cast_fp16)[name = string("a_5_cast_fp16")];
            int32 var_2829 = const()[name = string("op_2829"), val = int32(1)];
            bool xp_5_interleave_0 = const()[name = string("xp_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 99, 1024]> xp_5_cast_fp16 = concat(axis = var_2829, interleave = xp_5_interleave_0, values = (zpad_to_fp16, qkv_5_cast_fp16))[name = string("xp_5_cast_fp16")];
            tensor<int32, [3]> var_2840_begin_0 = const()[name = string("op_2840_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_2840_end_0 = const()[name = string("op_2840_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_2840_end_mask_0 = const()[name = string("op_2840_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 1024]> var_2840_cast_fp16 = slice_by_index(begin = var_2840_begin_0, end = var_2840_end_0, end_mask = var_2840_end_mask_0, x = xp_5_cast_fp16)[name = string("op_2840_cast_fp16")];
            tensor<fp16, [1024]> var_2848_to_fp16 = const()[name = string("op_2848_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47449280)))];
            tensor<fp16, [1, 96, 1024]> var_2849_cast_fp16 = mul(x = var_2840_cast_fp16, y = var_2848_to_fp16)[name = string("op_2849_cast_fp16")];
            tensor<int32, [3]> var_2859_begin_0 = const()[name = string("op_2859_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_2859_end_0 = const()[name = string("op_2859_end_0"), val = tensor<int32, [3]>([1, 98, 1024])];
            tensor<bool, [3]> var_2859_end_mask_0 = const()[name = string("op_2859_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_2859_cast_fp16 = slice_by_index(begin = var_2859_begin_0, end = var_2859_end_0, end_mask = var_2859_end_mask_0, x = xp_5_cast_fp16)[name = string("op_2859_cast_fp16")];
            tensor<fp16, [1024]> var_2867_to_fp16 = const()[name = string("op_2867_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47451392)))];
            tensor<fp16, [1, 96, 1024]> var_2868_cast_fp16 = mul(x = var_2859_cast_fp16, y = var_2867_to_fp16)[name = string("op_2868_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> var_2870_cast_fp16 = add(x = var_2849_cast_fp16, y = var_2868_cast_fp16)[name = string("op_2870_cast_fp16")];
            tensor<int32, [3]> var_2880_begin_0 = const()[name = string("op_2880_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2880_end_0 = const()[name = string("op_2880_end_0"), val = tensor<int32, [3]>([1, 97, 1024])];
            tensor<bool, [3]> var_2880_end_mask_0 = const()[name = string("op_2880_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_2880_cast_fp16 = slice_by_index(begin = var_2880_begin_0, end = var_2880_end_0, end_mask = var_2880_end_mask_0, x = xp_5_cast_fp16)[name = string("op_2880_cast_fp16")];
            tensor<fp16, [1024]> var_2888_to_fp16 = const()[name = string("op_2888_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47453504)))];
            tensor<fp16, [1, 96, 1024]> var_2889_cast_fp16 = mul(x = var_2880_cast_fp16, y = var_2888_to_fp16)[name = string("op_2889_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> var_2891_cast_fp16 = add(x = var_2870_cast_fp16, y = var_2889_cast_fp16)[name = string("op_2891_cast_fp16")];
            tensor<int32, [3]> var_2901_begin_0 = const()[name = string("op_2901_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2901_end_0 = const()[name = string("op_2901_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> var_2901_end_mask_0 = const()[name = string("op_2901_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_2901_cast_fp16 = slice_by_index(begin = var_2901_begin_0, end = var_2901_end_0, end_mask = var_2901_end_mask_0, x = xp_5_cast_fp16)[name = string("op_2901_cast_fp16")];
            tensor<fp16, [1024]> var_2909_to_fp16 = const()[name = string("op_2909_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47455616)))];
            tensor<fp16, [1, 96, 1024]> var_2910_cast_fp16 = mul(x = var_2901_cast_fp16, y = var_2909_to_fp16)[name = string("op_2910_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> input_35_cast_fp16 = add(x = var_2891_cast_fp16, y = var_2910_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> c_5_cast_fp16 = silu(x = input_35_cast_fp16)[name = string("c_5_cast_fp16")];
            tensor<int32, [3]> var_2918_begin_0 = const()[name = string("op_2918_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2918_end_0 = const()[name = string("op_2918_end_0"), val = tensor<int32, [3]>([1, 96, 256])];
            tensor<bool, [3]> var_2918_end_mask_0 = const()[name = string("op_2918_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 256]> var_2918_cast_fp16 = slice_by_index(begin = var_2918_begin_0, end = var_2918_end_0, end_mask = var_2918_end_mask_0, x = c_5_cast_fp16)[name = string("op_2918_cast_fp16")];
            tensor<int32, [4]> var_2923 = const()[name = string("op_2923"), val = tensor<int32, [4]>([1, 96, 2, 128])];
            tensor<fp16, [1, 96, 2, 128]> q_41_cast_fp16 = reshape(shape = var_2923, x = var_2918_cast_fp16)[name = string("q_41_cast_fp16")];
            tensor<int32, [3]> var_2929_begin_0 = const()[name = string("op_2929_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_2929_end_0 = const()[name = string("op_2929_end_0"), val = tensor<int32, [3]>([1, 96, 512])];
            tensor<bool, [3]> var_2929_end_mask_0 = const()[name = string("op_2929_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 256]> var_2929_cast_fp16 = slice_by_index(begin = var_2929_begin_0, end = var_2929_end_0, end_mask = var_2929_end_mask_0, x = c_5_cast_fp16)[name = string("op_2929_cast_fp16")];
            tensor<int32, [4]> var_2934 = const()[name = string("op_2934"), val = tensor<int32, [4]>([1, 96, 2, 128])];
            tensor<fp16, [1, 96, 2, 128]> k_41_cast_fp16 = reshape(shape = var_2934, x = var_2929_cast_fp16)[name = string("k_41_cast_fp16")];
            tensor<int32, [3]> var_2940_begin_0 = const()[name = string("op_2940_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_2940_end_0 = const()[name = string("op_2940_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> var_2940_end_mask_0 = const()[name = string("op_2940_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 512]> var_2940_cast_fp16 = slice_by_index(begin = var_2940_begin_0, end = var_2940_end_0, end_mask = var_2940_end_mask_0, x = c_5_cast_fp16)[name = string("op_2940_cast_fp16")];
            tensor<int32, [4]> var_2945 = const()[name = string("op_2945"), val = tensor<int32, [4]>([1, 96, 4, 128])];
            tensor<fp16, [1, 96, 4, 128]> var_2946_cast_fp16 = reshape(shape = var_2945, x = var_2940_cast_fp16)[name = string("op_2946_cast_fp16")];
            tensor<int32, [4]> var_2951 = const()[name = string("op_2951"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2953_to_fp16 = const()[name = string("op_2953_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 96, 2, 128]> q_43_cast_fp16 = mul(x = q_41_cast_fp16, y = var_2953_to_fp16)[name = string("q_43_cast_fp16")];
            fp16 var_2955_to_fp16 = const()[name = string("op_2955_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 96, 2, 128]> k_43_cast_fp16 = mul(x = k_41_cast_fp16, y = var_2955_to_fp16)[name = string("k_43_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_2957_cast_fp16 = mul(x = q_43_cast_fp16, y = q_43_cast_fp16)[name = string("op_2957_cast_fp16")];
            tensor<int32, [1]> var_2962_axes_0 = const()[name = string("op_2962_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2962_keep_dims_0 = const()[name = string("op_2962_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 2, 1]> var_2962_cast_fp16 = reduce_mean(axes = var_2962_axes_0, keep_dims = var_2962_keep_dims_0, x = var_2957_cast_fp16)[name = string("op_2962_cast_fp16")];
            fp16 var_2963_promoted_to_fp16 = const()[name = string("op_2963_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 96, 2, 1]> var_2964_cast_fp16 = mul(x = var_2962_cast_fp16, y = var_2963_promoted_to_fp16)[name = string("op_2964_cast_fp16")];
            fp16 var_2966_to_fp16 = const()[name = string("op_2966_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 96, 2, 1]> var_2967_cast_fp16 = add(x = var_2964_cast_fp16, y = var_2966_to_fp16)[name = string("op_2967_cast_fp16")];
            fp32 var_2968_epsilon_0 = const()[name = string("op_2968_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 2, 1]> var_2968_cast_fp16 = rsqrt(epsilon = var_2968_epsilon_0, x = var_2967_cast_fp16)[name = string("op_2968_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_2969_cast_fp16 = mul(x = q_43_cast_fp16, y = var_2968_cast_fp16)[name = string("op_2969_cast_fp16")];
            fp16 var_2970_to_fp16 = const()[name = string("op_2970_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 96, 2, 128]> q_45_cast_fp16 = mul(x = var_2969_cast_fp16, y = var_2970_to_fp16)[name = string("q_45_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_2972_cast_fp16 = mul(x = k_43_cast_fp16, y = k_43_cast_fp16)[name = string("op_2972_cast_fp16")];
            tensor<int32, [1]> var_2977_axes_0 = const()[name = string("op_2977_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2977_keep_dims_0 = const()[name = string("op_2977_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 2, 1]> var_2977_cast_fp16 = reduce_mean(axes = var_2977_axes_0, keep_dims = var_2977_keep_dims_0, x = var_2972_cast_fp16)[name = string("op_2977_cast_fp16")];
            fp16 var_2978_promoted_to_fp16 = const()[name = string("op_2978_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 96, 2, 1]> var_2979_cast_fp16 = mul(x = var_2977_cast_fp16, y = var_2978_promoted_to_fp16)[name = string("op_2979_cast_fp16")];
            fp16 var_2981_to_fp16 = const()[name = string("op_2981_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 96, 2, 1]> var_2982_cast_fp16 = add(x = var_2979_cast_fp16, y = var_2981_to_fp16)[name = string("op_2982_cast_fp16")];
            fp32 var_2983_epsilon_0 = const()[name = string("op_2983_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 2, 1]> var_2983_cast_fp16 = rsqrt(epsilon = var_2983_epsilon_0, x = var_2982_cast_fp16)[name = string("op_2983_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> k_45_cast_fp16 = mul(x = k_43_cast_fp16, y = var_2983_cast_fp16)[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> var_2989 = const()[name = string("op_2989"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_3002_axes_0 = const()[name = string("op_3002_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 96, 128]> var_2990_cast_fp16 = transpose(perm = var_2989, x = q_45_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [1, 2, 1, 96, 128]> var_3002_cast_fp16 = expand_dims(axes = var_3002_axes_0, x = var_2990_cast_fp16)[name = string("op_3002_cast_fp16")];
            tensor<int32, [5]> var_3010_reps_0 = const()[name = string("op_3010_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 96, 128]> var_3010_cast_fp16 = tile(reps = var_3010_reps_0, x = var_3002_cast_fp16)[name = string("op_3010_cast_fp16")];
            tensor<int32, [4]> var_3015 = const()[name = string("op_3015"), val = tensor<int32, [4]>([1, 4, 96, 128])];
            tensor<fp16, [1, 4, 96, 128]> q_47_cast_fp16 = reshape(shape = var_3015, x = var_3010_cast_fp16)[name = string("q_47_cast_fp16")];
            tensor<int32, [4]> var_3021 = const()[name = string("op_3021"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_3034_axes_0 = const()[name = string("op_3034_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 96, 128]> var_3022_cast_fp16 = transpose(perm = var_3021, x = k_45_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 2, 1, 96, 128]> var_3034_cast_fp16 = expand_dims(axes = var_3034_axes_0, x = var_3022_cast_fp16)[name = string("op_3034_cast_fp16")];
            tensor<int32, [5]> var_3042_reps_0 = const()[name = string("op_3042_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 96, 128]> var_3042_cast_fp16 = tile(reps = var_3042_reps_0, x = var_3034_cast_fp16)[name = string("op_3042_cast_fp16")];
            tensor<int32, [4]> var_3047 = const()[name = string("op_3047"), val = tensor<int32, [4]>([1, 4, 96, 128])];
            tensor<fp16, [1, 4, 96, 128]> k_47_cast_fp16 = reshape(shape = var_3047, x = var_3042_cast_fp16)[name = string("k_47_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_3049_cast_fp16 = sigmoid(x = b_5_cast_fp16)[name = string("op_3049_cast_fp16")];
            tensor<int32, [3]> var_3053 = const()[name = string("op_3053"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_2_dt_bias_to_fp16 = const()[name = string("layers_2_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 96, 4]> input_37_cast_fp16 = add(x = a_5_cast_fp16, y = layers_2_dt_bias_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_3057_cast_fp16 = relu(x = input_37_cast_fp16)[name = string("op_3057_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_3058_cast_fp16 = abs(x = input_37_cast_fp16)[name = string("op_3058_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 96, 4]> var_3059_cast_fp16 = mul(x = var_3058_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_3059_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_3060_cast_fp16 = exp(x = var_3059_cast_fp16)[name = string("op_3060_cast_fp16")];
            fp16 var_3062_to_fp16 = const()[name = string("op_3062_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 96, 4]> var_3063_cast_fp16 = add(x = var_3060_cast_fp16, y = var_3062_to_fp16)[name = string("op_3063_cast_fp16")];
            fp32 var_3064_epsilon_0 = const()[name = string("op_3064_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 96, 4]> var_3064_cast_fp16 = log(epsilon = var_3064_epsilon_0, x = var_3063_cast_fp16)[name = string("op_3064_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_3066_cast_fp16 = add(x = var_3057_cast_fp16, y = var_3064_cast_fp16)[name = string("op_3066_cast_fp16")];
            tensor<fp16, [4]> layers_2_negA_to_fp16 = const()[name = string("layers_2_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.c0cp+3, -0x1.3e4p+3, -0x1.568p+1, -0x1.9acp+2])];
            tensor<fp16, [1, 96, 4]> var_3067_cast_fp16 = mul(x = layers_2_negA_to_fp16, y = var_3066_cast_fp16)[name = string("op_3067_cast_fp16")];
            tensor<int32, [3]> var_3071 = const()[name = string("op_3071"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_3074_axes_0 = const()[name = string("op_3074_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 96]> beta_5_cast_fp16 = transpose(perm = var_3053, x = var_3049_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [1, 4, 96, 1]> var_3074_cast_fp16 = expand_dims(axes = var_3074_axes_0, x = beta_5_cast_fp16)[name = string("op_3074_cast_fp16")];
            fp16 var_3075_to_fp16 = const()[name = string("op_3075_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 96, 1]> var_3076_cast_fp16 = mul(x = var_3074_cast_fp16, y = var_3075_to_fp16)[name = string("op_3076_cast_fp16")];
            tensor<fp16, [1, 4, 96, 128]> v_5_cast_fp16 = transpose(perm = var_2951, x = var_2946_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [1, 4, 96, 128]> vb_29_cast_fp16 = mul(x = v_5_cast_fp16, y = var_3076_cast_fp16)[name = string("vb_29_cast_fp16")];
            tensor<int32, [4]> q_49_begin_0 = const()[name = string("q_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_49_end_0 = const()[name = string("q_49_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_49_end_mask_0 = const()[name = string("q_49_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, x = q_47_cast_fp16)[name = string("q_49_cast_fp16")];
            tensor<int32, [4]> k_49_begin_0 = const()[name = string("k_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_49_end_0 = const()[name = string("k_49_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_49_end_mask_0 = const()[name = string("k_49_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, x = k_47_cast_fp16)[name = string("k_49_cast_fp16")];
            tensor<fp16, [1, 4, 96, 128]> var_3110_cast_fp16 = mul(x = k_47_cast_fp16, y = var_3074_cast_fp16)[name = string("op_3110_cast_fp16")];
            tensor<int32, [4]> kb_25_begin_0 = const()[name = string("kb_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_25_end_0 = const()[name = string("kb_25_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_25_end_mask_0 = const()[name = string("kb_25_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_25_cast_fp16 = slice_by_index(begin = kb_25_begin_0, end = kb_25_end_0, end_mask = kb_25_end_mask_0, x = var_3110_cast_fp16)[name = string("kb_25_cast_fp16")];
            tensor<int32, [4]> vb_31_begin_0 = const()[name = string("vb_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_31_end_0 = const()[name = string("vb_31_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_31_end_mask_0 = const()[name = string("vb_31_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_31_cast_fp16 = slice_by_index(begin = vb_31_begin_0, end = vb_31_end_0, end_mask = vb_31_end_mask_0, x = vb_29_cast_fp16)[name = string("vb_31_cast_fp16")];
            tensor<int32, [3]> gdec_31_begin_0 = const()[name = string("gdec_31_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_31_end_0 = const()[name = string("gdec_31_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_31_end_mask_0 = const()[name = string("gdec_31_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 96]> gdec_29_cast_fp16 = transpose(perm = var_3071, x = var_3067_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [1, 4, 16]> gdec_31_cast_fp16 = slice_by_index(begin = gdec_31_begin_0, end = gdec_31_end_0, end_mask = gdec_31_end_mask_0, x = gdec_29_cast_fp16)[name = string("gdec_31_cast_fp16")];
            tensor<int32, [2]> var_3158 = const()[name = string("op_3158"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_3159_cast_fp16 = reshape(shape = var_3158, x = gdec_31_cast_fp16)[name = string("op_3159_cast_fp16")];
            tensor<fp16, [4, 256]> var_3160_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3159_cast_fp16)[name = string("op_3160_cast_fp16")];
            tensor<int32, [4]> var_3165 = const()[name = string("op_3165"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_25_cast_fp16 = reshape(shape = var_3165, x = var_3160_cast_fp16)[name = string("P_25_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3167_cast_fp16 = exp(x = P_25_cast_fp16)[name = string("op_3167_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_25_cast_fp16 = mul(x = var_3167_cast_fp16, y = tril_incl_to_fp16)[name = string("D_25_cast_fp16")];
            bool kk_25_transpose_x_1 = const()[name = string("kk_25_transpose_x_1"), val = bool(false)];
            bool kk_25_transpose_y_1 = const()[name = string("kk_25_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_25_cast_fp16 = matmul(transpose_x = kk_25_transpose_x_1, transpose_y = kk_25_transpose_y_1, x = kb_25_cast_fp16, y = k_49_cast_fp16)[name = string("kk_25_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_3173_cast_fp16 = mul(x = kk_25_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_3173_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3174_cast_fp16 = mul(x = var_3173_cast_fp16, y = D_25_cast_fp16)[name = string("op_3174_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_25_cast_fp16 = mul(x = var_3174_cast_fp16, y = cum_tail_to_fp16)[name = string("N_25_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3176_cast_fp16 = mul(x = D_25_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3176_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_25_cast_fp16 = mul(x = kk_25_cast_fp16, y = var_3176_cast_fp16)[name = string("Na_25_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_97_cast_fp16 = add(x = eye_to_fp16, y = N_25_cast_fp16)[name = string("Minv_97_cast_fp16")];
            bool Np_73_transpose_x_0 = const()[name = string("Np_73_transpose_x_0"), val = bool(false)];
            bool Np_73_transpose_y_0 = const()[name = string("Np_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_73_cast_fp16 = matmul(transpose_x = Np_73_transpose_x_0, transpose_y = Np_73_transpose_y_0, x = N_25_cast_fp16, y = Na_25_cast_fp16)[name = string("Np_73_cast_fp16")];
            bool Npa_49_transpose_x_0 = const()[name = string("Npa_49_transpose_x_0"), val = bool(false)];
            bool Npa_49_transpose_y_0 = const()[name = string("Npa_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_49_cast_fp16 = matmul(transpose_x = Npa_49_transpose_x_0, transpose_y = Npa_49_transpose_y_0, x = Na_25_cast_fp16, y = N_25_cast_fp16)[name = string("Npa_49_cast_fp16")];
            bool var_3182_transpose_x_0 = const()[name = string("op_3182_transpose_x_0"), val = bool(false)];
            bool var_3182_transpose_y_0 = const()[name = string("op_3182_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3182_cast_fp16 = matmul(transpose_x = var_3182_transpose_x_0, transpose_y = var_3182_transpose_y_0, x = Minv_97_cast_fp16, y = Np_73_cast_fp16)[name = string("op_3182_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_99_cast_fp16 = add(x = Minv_97_cast_fp16, y = var_3182_cast_fp16)[name = string("Minv_99_cast_fp16")];
            bool Np_75_transpose_x_0 = const()[name = string("Np_75_transpose_x_0"), val = bool(false)];
            bool Np_75_transpose_y_0 = const()[name = string("Np_75_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_75_cast_fp16 = matmul(transpose_x = Np_75_transpose_x_0, transpose_y = Np_75_transpose_y_0, x = Np_73_cast_fp16, y = Npa_49_cast_fp16)[name = string("Np_75_cast_fp16")];
            bool Npa_51_transpose_x_0 = const()[name = string("Npa_51_transpose_x_0"), val = bool(false)];
            bool Npa_51_transpose_y_0 = const()[name = string("Npa_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_51_cast_fp16 = matmul(transpose_x = Npa_51_transpose_x_0, transpose_y = Npa_51_transpose_y_0, x = Npa_49_cast_fp16, y = Np_73_cast_fp16)[name = string("Npa_51_cast_fp16")];
            bool var_3187_transpose_x_0 = const()[name = string("op_3187_transpose_x_0"), val = bool(false)];
            bool var_3187_transpose_y_0 = const()[name = string("op_3187_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3187_cast_fp16 = matmul(transpose_x = var_3187_transpose_x_0, transpose_y = var_3187_transpose_y_0, x = Minv_99_cast_fp16, y = Np_75_cast_fp16)[name = string("op_3187_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_101_cast_fp16 = add(x = Minv_99_cast_fp16, y = var_3187_cast_fp16)[name = string("Minv_101_cast_fp16")];
            bool Np_77_transpose_x_0 = const()[name = string("Np_77_transpose_x_0"), val = bool(false)];
            bool Np_77_transpose_y_0 = const()[name = string("Np_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_77_cast_fp16 = matmul(transpose_x = Np_77_transpose_x_0, transpose_y = Np_77_transpose_y_0, x = Np_75_cast_fp16, y = Npa_51_cast_fp16)[name = string("Np_77_cast_fp16")];
            bool var_3191_transpose_x_0 = const()[name = string("op_3191_transpose_x_0"), val = bool(false)];
            bool var_3191_transpose_y_0 = const()[name = string("op_3191_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3191_cast_fp16 = matmul(transpose_x = var_3191_transpose_x_0, transpose_y = var_3191_transpose_y_0, x = Minv_101_cast_fp16, y = Np_77_cast_fp16)[name = string("op_3191_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_103_cast_fp16 = add(x = Minv_101_cast_fp16, y = var_3191_cast_fp16)[name = string("Minv_103_cast_fp16")];
            bool u_25_transpose_x_0 = const()[name = string("u_25_transpose_x_0"), val = bool(false)];
            bool u_25_transpose_y_0 = const()[name = string("u_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_25_cast_fp16 = matmul(transpose_x = u_25_transpose_x_0, transpose_y = u_25_transpose_y_0, x = Minv_103_cast_fp16, y = vb_31_cast_fp16)[name = string("u_25_cast_fp16")];
            bool var_3198_transpose_x_1 = const()[name = string("op_3198_transpose_x_1"), val = bool(false)];
            bool var_3198_transpose_y_1 = const()[name = string("op_3198_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_3198_cast_fp16 = matmul(transpose_x = var_3198_transpose_x_1, transpose_y = var_3198_transpose_y_1, x = q_49_cast_fp16, y = k_49_cast_fp16)[name = string("op_3198_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_25_cast_fp16 = mul(x = var_3198_cast_fp16, y = D_25_cast_fp16)[name = string("A_25_cast_fp16")];
            bool var_3200_transpose_x_0 = const()[name = string("op_3200_transpose_x_0"), val = bool(false)];
            bool var_3200_transpose_y_0 = const()[name = string("op_3200_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3200_cast_fp16 = matmul(transpose_x = var_3200_transpose_x_0, transpose_y = var_3200_transpose_y_0, x = A_25_cast_fp16, y = u_25_cast_fp16)[name = string("op_3200_cast_fp16")];
            tensor<fp16, [16]> var_3201_bias_0_to_fp16 = const()[name = string("op_3201_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_3201_cast_fp16 = linear(bias = var_3201_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_31_cast_fp16)[name = string("op_3201_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_21_cast_fp16 = exp(x = var_3201_cast_fp16)[name = string("tail_21_cast_fp16")];
            tensor<int32, [1]> var_3204_axes_0 = const()[name = string("op_3204_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_3204_cast_fp16 = expand_dims(axes = var_3204_axes_0, x = tail_21_cast_fp16)[name = string("op_3204_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3205_cast_fp16 = mul(x = k_49_cast_fp16, y = var_3204_cast_fp16)[name = string("op_3205_cast_fp16")];
            bool S_21_transpose_x_1 = const()[name = string("S_21_transpose_x_1"), val = bool(true)];
            bool S_21_transpose_y_1 = const()[name = string("S_21_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_21_cast_fp16 = matmul(transpose_x = S_21_transpose_x_1, transpose_y = S_21_transpose_y_1, x = var_3205_cast_fp16, y = u_25_cast_fp16)[name = string("S_21_cast_fp16")];
            tensor<int32, [4]> q_51_begin_0 = const()[name = string("q_51_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_51_end_0 = const()[name = string("q_51_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> q_51_end_mask_0 = const()[name = string("q_51_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_51_cast_fp16 = slice_by_index(begin = q_51_begin_0, end = q_51_end_0, end_mask = q_51_end_mask_0, x = q_47_cast_fp16)[name = string("q_51_cast_fp16")];
            tensor<int32, [4]> k_51_begin_0 = const()[name = string("k_51_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_51_end_0 = const()[name = string("k_51_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> k_51_end_mask_0 = const()[name = string("k_51_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_51_cast_fp16 = slice_by_index(begin = k_51_begin_0, end = k_51_end_0, end_mask = k_51_end_mask_0, x = k_47_cast_fp16)[name = string("k_51_cast_fp16")];
            tensor<int32, [4]> kb_27_begin_0 = const()[name = string("kb_27_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_27_end_0 = const()[name = string("kb_27_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> kb_27_end_mask_0 = const()[name = string("kb_27_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_27_cast_fp16 = slice_by_index(begin = kb_27_begin_0, end = kb_27_end_0, end_mask = kb_27_end_mask_0, x = var_3110_cast_fp16)[name = string("kb_27_cast_fp16")];
            tensor<int32, [4]> vb_33_begin_0 = const()[name = string("vb_33_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_33_end_0 = const()[name = string("vb_33_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> vb_33_end_mask_0 = const()[name = string("vb_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_33_cast_fp16 = slice_by_index(begin = vb_33_begin_0, end = vb_33_end_0, end_mask = vb_33_end_mask_0, x = vb_29_cast_fp16)[name = string("vb_33_cast_fp16")];
            tensor<int32, [3]> gdec_33_begin_0 = const()[name = string("gdec_33_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_33_end_0 = const()[name = string("gdec_33_end_0"), val = tensor<int32, [3]>([1, 4, 32])];
            tensor<bool, [3]> gdec_33_end_mask_0 = const()[name = string("gdec_33_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_33_cast_fp16 = slice_by_index(begin = gdec_33_begin_0, end = gdec_33_end_0, end_mask = gdec_33_end_mask_0, x = gdec_29_cast_fp16)[name = string("gdec_33_cast_fp16")];
            tensor<int32, [2]> var_3290 = const()[name = string("op_3290"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_3291_cast_fp16 = reshape(shape = var_3290, x = gdec_33_cast_fp16)[name = string("op_3291_cast_fp16")];
            tensor<fp16, [4, 256]> var_3292_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3291_cast_fp16)[name = string("op_3292_cast_fp16")];
            tensor<int32, [4]> var_3297 = const()[name = string("op_3297"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_27_cast_fp16 = reshape(shape = var_3297, x = var_3292_cast_fp16)[name = string("P_27_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3299_cast_fp16 = exp(x = P_27_cast_fp16)[name = string("op_3299_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_27_cast_fp16 = mul(x = var_3299_cast_fp16, y = tril_incl_to_fp16)[name = string("D_27_cast_fp16")];
            bool kk_27_transpose_x_1 = const()[name = string("kk_27_transpose_x_1"), val = bool(false)];
            bool kk_27_transpose_y_1 = const()[name = string("kk_27_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_27_cast_fp16 = matmul(transpose_x = kk_27_transpose_x_1, transpose_y = kk_27_transpose_y_1, x = kb_27_cast_fp16, y = k_51_cast_fp16)[name = string("kk_27_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_3305_cast_fp16 = mul(x = kk_27_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_3305_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3306_cast_fp16 = mul(x = var_3305_cast_fp16, y = D_27_cast_fp16)[name = string("op_3306_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_27_cast_fp16 = mul(x = var_3306_cast_fp16, y = cum_tail_to_fp16)[name = string("N_27_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3308_cast_fp16 = mul(x = D_27_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3308_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_27_cast_fp16 = mul(x = kk_27_cast_fp16, y = var_3308_cast_fp16)[name = string("Na_27_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_105_cast_fp16 = add(x = eye_to_fp16, y = N_27_cast_fp16)[name = string("Minv_105_cast_fp16")];
            bool Np_79_transpose_x_0 = const()[name = string("Np_79_transpose_x_0"), val = bool(false)];
            bool Np_79_transpose_y_0 = const()[name = string("Np_79_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_79_cast_fp16 = matmul(transpose_x = Np_79_transpose_x_0, transpose_y = Np_79_transpose_y_0, x = N_27_cast_fp16, y = Na_27_cast_fp16)[name = string("Np_79_cast_fp16")];
            bool Npa_53_transpose_x_0 = const()[name = string("Npa_53_transpose_x_0"), val = bool(false)];
            bool Npa_53_transpose_y_0 = const()[name = string("Npa_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_53_cast_fp16 = matmul(transpose_x = Npa_53_transpose_x_0, transpose_y = Npa_53_transpose_y_0, x = Na_27_cast_fp16, y = N_27_cast_fp16)[name = string("Npa_53_cast_fp16")];
            bool var_3314_transpose_x_0 = const()[name = string("op_3314_transpose_x_0"), val = bool(false)];
            bool var_3314_transpose_y_0 = const()[name = string("op_3314_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3314_cast_fp16 = matmul(transpose_x = var_3314_transpose_x_0, transpose_y = var_3314_transpose_y_0, x = Minv_105_cast_fp16, y = Np_79_cast_fp16)[name = string("op_3314_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_107_cast_fp16 = add(x = Minv_105_cast_fp16, y = var_3314_cast_fp16)[name = string("Minv_107_cast_fp16")];
            bool Np_81_transpose_x_0 = const()[name = string("Np_81_transpose_x_0"), val = bool(false)];
            bool Np_81_transpose_y_0 = const()[name = string("Np_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_81_cast_fp16 = matmul(transpose_x = Np_81_transpose_x_0, transpose_y = Np_81_transpose_y_0, x = Np_79_cast_fp16, y = Npa_53_cast_fp16)[name = string("Np_81_cast_fp16")];
            bool Npa_55_transpose_x_0 = const()[name = string("Npa_55_transpose_x_0"), val = bool(false)];
            bool Npa_55_transpose_y_0 = const()[name = string("Npa_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_55_cast_fp16 = matmul(transpose_x = Npa_55_transpose_x_0, transpose_y = Npa_55_transpose_y_0, x = Npa_53_cast_fp16, y = Np_79_cast_fp16)[name = string("Npa_55_cast_fp16")];
            bool var_3319_transpose_x_0 = const()[name = string("op_3319_transpose_x_0"), val = bool(false)];
            bool var_3319_transpose_y_0 = const()[name = string("op_3319_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3319_cast_fp16 = matmul(transpose_x = var_3319_transpose_x_0, transpose_y = var_3319_transpose_y_0, x = Minv_107_cast_fp16, y = Np_81_cast_fp16)[name = string("op_3319_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_109_cast_fp16 = add(x = Minv_107_cast_fp16, y = var_3319_cast_fp16)[name = string("Minv_109_cast_fp16")];
            bool Np_83_transpose_x_0 = const()[name = string("Np_83_transpose_x_0"), val = bool(false)];
            bool Np_83_transpose_y_0 = const()[name = string("Np_83_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_83_cast_fp16 = matmul(transpose_x = Np_83_transpose_x_0, transpose_y = Np_83_transpose_y_0, x = Np_81_cast_fp16, y = Npa_55_cast_fp16)[name = string("Np_83_cast_fp16")];
            bool var_3323_transpose_x_0 = const()[name = string("op_3323_transpose_x_0"), val = bool(false)];
            bool var_3323_transpose_y_0 = const()[name = string("op_3323_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3323_cast_fp16 = matmul(transpose_x = var_3323_transpose_x_0, transpose_y = var_3323_transpose_y_0, x = Minv_109_cast_fp16, y = Np_83_cast_fp16)[name = string("op_3323_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_111_cast_fp16 = add(x = Minv_109_cast_fp16, y = var_3323_cast_fp16)[name = string("Minv_111_cast_fp16")];
            bool u_27_transpose_x_0 = const()[name = string("u_27_transpose_x_0"), val = bool(false)];
            bool u_27_transpose_y_0 = const()[name = string("u_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_27_cast_fp16 = matmul(transpose_x = u_27_transpose_x_0, transpose_y = u_27_transpose_y_0, x = Minv_111_cast_fp16, y = vb_33_cast_fp16)[name = string("u_27_cast_fp16")];
            bool var_3330_transpose_x_1 = const()[name = string("op_3330_transpose_x_1"), val = bool(false)];
            bool var_3330_transpose_y_1 = const()[name = string("op_3330_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_3330_cast_fp16 = matmul(transpose_x = var_3330_transpose_x_1, transpose_y = var_3330_transpose_y_1, x = q_51_cast_fp16, y = k_51_cast_fp16)[name = string("op_3330_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_27_cast_fp16 = mul(x = var_3330_cast_fp16, y = D_27_cast_fp16)[name = string("A_27_cast_fp16")];
            tensor<fp16, [16]> var_3332_bias_0_to_fp16 = const()[name = string("op_3332_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_3332_cast_fp16 = linear(bias = var_3332_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_33_cast_fp16)[name = string("op_3332_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_21_cast_fp16 = exp(x = var_3332_cast_fp16)[name = string("cum_21_cast_fp16")];
            tensor<int32, [1]> var_3335_axes_0 = const()[name = string("op_3335_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_3335_cast_fp16 = expand_dims(axes = var_3335_axes_0, x = cum_21_cast_fp16)[name = string("op_3335_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3336_cast_fp16 = mul(x = kb_27_cast_fp16, y = var_3335_cast_fp16)[name = string("op_3336_cast_fp16")];
            bool w_27_transpose_x_0 = const()[name = string("w_27_transpose_x_0"), val = bool(false)];
            bool w_27_transpose_y_0 = const()[name = string("w_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_27_cast_fp16 = matmul(transpose_x = w_27_transpose_x_0, transpose_y = w_27_transpose_y_0, x = Minv_111_cast_fp16, y = var_3336_cast_fp16)[name = string("w_27_cast_fp16")];
            bool var_3338_transpose_x_0 = const()[name = string("op_3338_transpose_x_0"), val = bool(false)];
            bool var_3338_transpose_y_0 = const()[name = string("op_3338_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3338_cast_fp16 = matmul(transpose_x = var_3338_transpose_x_0, transpose_y = var_3338_transpose_y_0, x = w_27_cast_fp16, y = S_21_cast_fp16)[name = string("op_3338_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_21_cast_fp16 = sub(x = u_27_cast_fp16, y = var_3338_cast_fp16)[name = string("vnew_21_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3343_cast_fp16 = mul(x = q_51_cast_fp16, y = var_3335_cast_fp16)[name = string("op_3343_cast_fp16")];
            bool var_3344_transpose_x_0 = const()[name = string("op_3344_transpose_x_0"), val = bool(false)];
            bool var_3344_transpose_y_0 = const()[name = string("op_3344_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3344_cast_fp16 = matmul(transpose_x = var_3344_transpose_x_0, transpose_y = var_3344_transpose_y_0, x = var_3343_cast_fp16, y = S_21_cast_fp16)[name = string("op_3344_cast_fp16")];
            bool var_3345_transpose_x_0 = const()[name = string("op_3345_transpose_x_0"), val = bool(false)];
            bool var_3345_transpose_y_0 = const()[name = string("op_3345_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3345_cast_fp16 = matmul(transpose_x = var_3345_transpose_x_0, transpose_y = var_3345_transpose_y_0, x = A_27_cast_fp16, y = vnew_21_cast_fp16)[name = string("op_3345_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3347_cast_fp16 = add(x = var_3344_cast_fp16, y = var_3345_cast_fp16)[name = string("op_3347_cast_fp16")];
            tensor<fp16, [16]> var_3348_bias_0_to_fp16 = const()[name = string("op_3348_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_3348_cast_fp16 = linear(bias = var_3348_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_33_cast_fp16)[name = string("op_3348_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_23_cast_fp16 = exp(x = var_3348_cast_fp16)[name = string("tail_23_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_45_to_fp16 = const()[name = string("transpose_45_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_3350_bias_0_to_fp16 = const()[name = string("op_3350_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_3350_cast_fp16 = linear(bias = var_3350_bias_0_to_fp16, weight = transpose_45_to_fp16, x = gdec_33_cast_fp16)[name = string("op_3350_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_3351_cast_fp16 = exp(x = var_3350_cast_fp16)[name = string("op_3351_cast_fp16")];
            tensor<int32, [1]> last_19_axes_0 = const()[name = string("last_19_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_19_cast_fp16 = expand_dims(axes = last_19_axes_0, x = var_3351_cast_fp16)[name = string("last_19_cast_fp16")];
            tensor<int32, [1]> var_3355_axes_0 = const()[name = string("op_3355_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_3355_cast_fp16 = expand_dims(axes = var_3355_axes_0, x = tail_23_cast_fp16)[name = string("op_3355_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3356_cast_fp16 = mul(x = k_51_cast_fp16, y = var_3355_cast_fp16)[name = string("op_3356_cast_fp16")];
            bool upd_17_transpose_x_1 = const()[name = string("upd_17_transpose_x_1"), val = bool(true)];
            bool upd_17_transpose_y_1 = const()[name = string("upd_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_17_cast_fp16 = matmul(transpose_x = upd_17_transpose_x_1, transpose_y = upd_17_transpose_y_1, x = var_3356_cast_fp16, y = vnew_21_cast_fp16)[name = string("upd_17_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_3361_cast_fp16 = mul(x = S_21_cast_fp16, y = last_19_cast_fp16)[name = string("op_3361_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_23_cast_fp16 = add(x = var_3361_cast_fp16, y = upd_17_cast_fp16)[name = string("S_23_cast_fp16")];
            tensor<int32, [4]> q_53_begin_0 = const()[name = string("q_53_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_53_end_0 = const()[name = string("q_53_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> q_53_end_mask_0 = const()[name = string("q_53_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_53_cast_fp16 = slice_by_index(begin = q_53_begin_0, end = q_53_end_0, end_mask = q_53_end_mask_0, x = q_47_cast_fp16)[name = string("q_53_cast_fp16")];
            tensor<int32, [4]> k_53_begin_0 = const()[name = string("k_53_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_53_end_0 = const()[name = string("k_53_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> k_53_end_mask_0 = const()[name = string("k_53_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_53_cast_fp16 = slice_by_index(begin = k_53_begin_0, end = k_53_end_0, end_mask = k_53_end_mask_0, x = k_47_cast_fp16)[name = string("k_53_cast_fp16")];
            tensor<int32, [4]> kb_29_begin_0 = const()[name = string("kb_29_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_29_end_0 = const()[name = string("kb_29_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> kb_29_end_mask_0 = const()[name = string("kb_29_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_29_cast_fp16 = slice_by_index(begin = kb_29_begin_0, end = kb_29_end_0, end_mask = kb_29_end_mask_0, x = var_3110_cast_fp16)[name = string("kb_29_cast_fp16")];
            tensor<int32, [4]> vb_35_begin_0 = const()[name = string("vb_35_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_35_end_0 = const()[name = string("vb_35_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> vb_35_end_mask_0 = const()[name = string("vb_35_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_35_cast_fp16 = slice_by_index(begin = vb_35_begin_0, end = vb_35_end_0, end_mask = vb_35_end_mask_0, x = vb_29_cast_fp16)[name = string("vb_35_cast_fp16")];
            tensor<int32, [3]> gdec_35_begin_0 = const()[name = string("gdec_35_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_35_end_0 = const()[name = string("gdec_35_end_0"), val = tensor<int32, [3]>([1, 4, 48])];
            tensor<bool, [3]> gdec_35_end_mask_0 = const()[name = string("gdec_35_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_35_cast_fp16 = slice_by_index(begin = gdec_35_begin_0, end = gdec_35_end_0, end_mask = gdec_35_end_mask_0, x = gdec_29_cast_fp16)[name = string("gdec_35_cast_fp16")];
            tensor<int32, [2]> var_3444 = const()[name = string("op_3444"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_3445_cast_fp16 = reshape(shape = var_3444, x = gdec_35_cast_fp16)[name = string("op_3445_cast_fp16")];
            tensor<fp16, [4, 256]> var_3446_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3445_cast_fp16)[name = string("op_3446_cast_fp16")];
            tensor<int32, [4]> var_3451 = const()[name = string("op_3451"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_29_cast_fp16 = reshape(shape = var_3451, x = var_3446_cast_fp16)[name = string("P_29_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3453_cast_fp16 = exp(x = P_29_cast_fp16)[name = string("op_3453_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_29_cast_fp16 = mul(x = var_3453_cast_fp16, y = tril_incl_to_fp16)[name = string("D_29_cast_fp16")];
            bool kk_29_transpose_x_1 = const()[name = string("kk_29_transpose_x_1"), val = bool(false)];
            bool kk_29_transpose_y_1 = const()[name = string("kk_29_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_29_cast_fp16 = matmul(transpose_x = kk_29_transpose_x_1, transpose_y = kk_29_transpose_y_1, x = kb_29_cast_fp16, y = k_53_cast_fp16)[name = string("kk_29_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_3459_cast_fp16 = mul(x = kk_29_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_3459_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3460_cast_fp16 = mul(x = var_3459_cast_fp16, y = D_29_cast_fp16)[name = string("op_3460_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_29_cast_fp16 = mul(x = var_3460_cast_fp16, y = cum_tail_to_fp16)[name = string("N_29_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3462_cast_fp16 = mul(x = D_29_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3462_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_29_cast_fp16 = mul(x = kk_29_cast_fp16, y = var_3462_cast_fp16)[name = string("Na_29_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_113_cast_fp16 = add(x = eye_to_fp16, y = N_29_cast_fp16)[name = string("Minv_113_cast_fp16")];
            bool Np_85_transpose_x_0 = const()[name = string("Np_85_transpose_x_0"), val = bool(false)];
            bool Np_85_transpose_y_0 = const()[name = string("Np_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_85_cast_fp16 = matmul(transpose_x = Np_85_transpose_x_0, transpose_y = Np_85_transpose_y_0, x = N_29_cast_fp16, y = Na_29_cast_fp16)[name = string("Np_85_cast_fp16")];
            bool Npa_57_transpose_x_0 = const()[name = string("Npa_57_transpose_x_0"), val = bool(false)];
            bool Npa_57_transpose_y_0 = const()[name = string("Npa_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_57_cast_fp16 = matmul(transpose_x = Npa_57_transpose_x_0, transpose_y = Npa_57_transpose_y_0, x = Na_29_cast_fp16, y = N_29_cast_fp16)[name = string("Npa_57_cast_fp16")];
            bool var_3468_transpose_x_0 = const()[name = string("op_3468_transpose_x_0"), val = bool(false)];
            bool var_3468_transpose_y_0 = const()[name = string("op_3468_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3468_cast_fp16 = matmul(transpose_x = var_3468_transpose_x_0, transpose_y = var_3468_transpose_y_0, x = Minv_113_cast_fp16, y = Np_85_cast_fp16)[name = string("op_3468_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_115_cast_fp16 = add(x = Minv_113_cast_fp16, y = var_3468_cast_fp16)[name = string("Minv_115_cast_fp16")];
            bool Np_87_transpose_x_0 = const()[name = string("Np_87_transpose_x_0"), val = bool(false)];
            bool Np_87_transpose_y_0 = const()[name = string("Np_87_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_87_cast_fp16 = matmul(transpose_x = Np_87_transpose_x_0, transpose_y = Np_87_transpose_y_0, x = Np_85_cast_fp16, y = Npa_57_cast_fp16)[name = string("Np_87_cast_fp16")];
            bool Npa_59_transpose_x_0 = const()[name = string("Npa_59_transpose_x_0"), val = bool(false)];
            bool Npa_59_transpose_y_0 = const()[name = string("Npa_59_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_59_cast_fp16 = matmul(transpose_x = Npa_59_transpose_x_0, transpose_y = Npa_59_transpose_y_0, x = Npa_57_cast_fp16, y = Np_85_cast_fp16)[name = string("Npa_59_cast_fp16")];
            bool var_3473_transpose_x_0 = const()[name = string("op_3473_transpose_x_0"), val = bool(false)];
            bool var_3473_transpose_y_0 = const()[name = string("op_3473_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3473_cast_fp16 = matmul(transpose_x = var_3473_transpose_x_0, transpose_y = var_3473_transpose_y_0, x = Minv_115_cast_fp16, y = Np_87_cast_fp16)[name = string("op_3473_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_117_cast_fp16 = add(x = Minv_115_cast_fp16, y = var_3473_cast_fp16)[name = string("Minv_117_cast_fp16")];
            bool Np_89_transpose_x_0 = const()[name = string("Np_89_transpose_x_0"), val = bool(false)];
            bool Np_89_transpose_y_0 = const()[name = string("Np_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_89_cast_fp16 = matmul(transpose_x = Np_89_transpose_x_0, transpose_y = Np_89_transpose_y_0, x = Np_87_cast_fp16, y = Npa_59_cast_fp16)[name = string("Np_89_cast_fp16")];
            bool var_3477_transpose_x_0 = const()[name = string("op_3477_transpose_x_0"), val = bool(false)];
            bool var_3477_transpose_y_0 = const()[name = string("op_3477_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3477_cast_fp16 = matmul(transpose_x = var_3477_transpose_x_0, transpose_y = var_3477_transpose_y_0, x = Minv_117_cast_fp16, y = Np_89_cast_fp16)[name = string("op_3477_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_119_cast_fp16 = add(x = Minv_117_cast_fp16, y = var_3477_cast_fp16)[name = string("Minv_119_cast_fp16")];
            bool u_29_transpose_x_0 = const()[name = string("u_29_transpose_x_0"), val = bool(false)];
            bool u_29_transpose_y_0 = const()[name = string("u_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_29_cast_fp16 = matmul(transpose_x = u_29_transpose_x_0, transpose_y = u_29_transpose_y_0, x = Minv_119_cast_fp16, y = vb_35_cast_fp16)[name = string("u_29_cast_fp16")];
            bool var_3484_transpose_x_1 = const()[name = string("op_3484_transpose_x_1"), val = bool(false)];
            bool var_3484_transpose_y_1 = const()[name = string("op_3484_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_3484_cast_fp16 = matmul(transpose_x = var_3484_transpose_x_1, transpose_y = var_3484_transpose_y_1, x = q_53_cast_fp16, y = k_53_cast_fp16)[name = string("op_3484_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_29_cast_fp16 = mul(x = var_3484_cast_fp16, y = D_29_cast_fp16)[name = string("A_29_cast_fp16")];
            tensor<fp16, [16]> var_3486_bias_0_to_fp16 = const()[name = string("op_3486_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_3486_cast_fp16 = linear(bias = var_3486_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_35_cast_fp16)[name = string("op_3486_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_23_cast_fp16 = exp(x = var_3486_cast_fp16)[name = string("cum_23_cast_fp16")];
            tensor<int32, [1]> var_3489_axes_0 = const()[name = string("op_3489_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_3489_cast_fp16 = expand_dims(axes = var_3489_axes_0, x = cum_23_cast_fp16)[name = string("op_3489_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3490_cast_fp16 = mul(x = kb_29_cast_fp16, y = var_3489_cast_fp16)[name = string("op_3490_cast_fp16")];
            bool w_29_transpose_x_0 = const()[name = string("w_29_transpose_x_0"), val = bool(false)];
            bool w_29_transpose_y_0 = const()[name = string("w_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_29_cast_fp16 = matmul(transpose_x = w_29_transpose_x_0, transpose_y = w_29_transpose_y_0, x = Minv_119_cast_fp16, y = var_3490_cast_fp16)[name = string("w_29_cast_fp16")];
            bool var_3492_transpose_x_0 = const()[name = string("op_3492_transpose_x_0"), val = bool(false)];
            bool var_3492_transpose_y_0 = const()[name = string("op_3492_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3492_cast_fp16 = matmul(transpose_x = var_3492_transpose_x_0, transpose_y = var_3492_transpose_y_0, x = w_29_cast_fp16, y = S_23_cast_fp16)[name = string("op_3492_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_23_cast_fp16 = sub(x = u_29_cast_fp16, y = var_3492_cast_fp16)[name = string("vnew_23_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3497_cast_fp16 = mul(x = q_53_cast_fp16, y = var_3489_cast_fp16)[name = string("op_3497_cast_fp16")];
            bool var_3498_transpose_x_0 = const()[name = string("op_3498_transpose_x_0"), val = bool(false)];
            bool var_3498_transpose_y_0 = const()[name = string("op_3498_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3498_cast_fp16 = matmul(transpose_x = var_3498_transpose_x_0, transpose_y = var_3498_transpose_y_0, x = var_3497_cast_fp16, y = S_23_cast_fp16)[name = string("op_3498_cast_fp16")];
            bool var_3499_transpose_x_0 = const()[name = string("op_3499_transpose_x_0"), val = bool(false)];
            bool var_3499_transpose_y_0 = const()[name = string("op_3499_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3499_cast_fp16 = matmul(transpose_x = var_3499_transpose_x_0, transpose_y = var_3499_transpose_y_0, x = A_29_cast_fp16, y = vnew_23_cast_fp16)[name = string("op_3499_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3501_cast_fp16 = add(x = var_3498_cast_fp16, y = var_3499_cast_fp16)[name = string("op_3501_cast_fp16")];
            tensor<fp16, [16]> var_3502_bias_0_to_fp16 = const()[name = string("op_3502_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_3502_cast_fp16 = linear(bias = var_3502_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_35_cast_fp16)[name = string("op_3502_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_25_cast_fp16 = exp(x = var_3502_cast_fp16)[name = string("tail_25_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_49_to_fp16 = const()[name = string("transpose_49_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_3504_bias_0_to_fp16 = const()[name = string("op_3504_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_3504_cast_fp16 = linear(bias = var_3504_bias_0_to_fp16, weight = transpose_49_to_fp16, x = gdec_35_cast_fp16)[name = string("op_3504_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_3505_cast_fp16 = exp(x = var_3504_cast_fp16)[name = string("op_3505_cast_fp16")];
            tensor<int32, [1]> last_21_axes_0 = const()[name = string("last_21_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_21_cast_fp16 = expand_dims(axes = last_21_axes_0, x = var_3505_cast_fp16)[name = string("last_21_cast_fp16")];
            tensor<int32, [1]> var_3509_axes_0 = const()[name = string("op_3509_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_3509_cast_fp16 = expand_dims(axes = var_3509_axes_0, x = tail_25_cast_fp16)[name = string("op_3509_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3510_cast_fp16 = mul(x = k_53_cast_fp16, y = var_3509_cast_fp16)[name = string("op_3510_cast_fp16")];
            bool upd_19_transpose_x_1 = const()[name = string("upd_19_transpose_x_1"), val = bool(true)];
            bool upd_19_transpose_y_1 = const()[name = string("upd_19_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_19_cast_fp16 = matmul(transpose_x = upd_19_transpose_x_1, transpose_y = upd_19_transpose_y_1, x = var_3510_cast_fp16, y = vnew_23_cast_fp16)[name = string("upd_19_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_3515_cast_fp16 = mul(x = S_23_cast_fp16, y = last_21_cast_fp16)[name = string("op_3515_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_25_cast_fp16 = add(x = var_3515_cast_fp16, y = upd_19_cast_fp16)[name = string("S_25_cast_fp16")];
            tensor<int32, [4]> q_55_begin_0 = const()[name = string("q_55_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_55_end_0 = const()[name = string("q_55_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> q_55_end_mask_0 = const()[name = string("q_55_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, x = q_47_cast_fp16)[name = string("q_55_cast_fp16")];
            tensor<int32, [4]> k_55_begin_0 = const()[name = string("k_55_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_55_end_0 = const()[name = string("k_55_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> k_55_end_mask_0 = const()[name = string("k_55_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_55_cast_fp16 = slice_by_index(begin = k_55_begin_0, end = k_55_end_0, end_mask = k_55_end_mask_0, x = k_47_cast_fp16)[name = string("k_55_cast_fp16")];
            tensor<int32, [4]> kb_31_begin_0 = const()[name = string("kb_31_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_31_end_0 = const()[name = string("kb_31_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> kb_31_end_mask_0 = const()[name = string("kb_31_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_31_cast_fp16 = slice_by_index(begin = kb_31_begin_0, end = kb_31_end_0, end_mask = kb_31_end_mask_0, x = var_3110_cast_fp16)[name = string("kb_31_cast_fp16")];
            tensor<int32, [4]> vb_37_begin_0 = const()[name = string("vb_37_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_37_end_0 = const()[name = string("vb_37_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> vb_37_end_mask_0 = const()[name = string("vb_37_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_37_cast_fp16 = slice_by_index(begin = vb_37_begin_0, end = vb_37_end_0, end_mask = vb_37_end_mask_0, x = vb_29_cast_fp16)[name = string("vb_37_cast_fp16")];
            tensor<int32, [3]> gdec_37_begin_0 = const()[name = string("gdec_37_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_37_end_0 = const()[name = string("gdec_37_end_0"), val = tensor<int32, [3]>([1, 4, 64])];
            tensor<bool, [3]> gdec_37_end_mask_0 = const()[name = string("gdec_37_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_37_cast_fp16 = slice_by_index(begin = gdec_37_begin_0, end = gdec_37_end_0, end_mask = gdec_37_end_mask_0, x = gdec_29_cast_fp16)[name = string("gdec_37_cast_fp16")];
            tensor<int32, [2]> var_3598 = const()[name = string("op_3598"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_3599_cast_fp16 = reshape(shape = var_3598, x = gdec_37_cast_fp16)[name = string("op_3599_cast_fp16")];
            tensor<fp16, [4, 256]> var_3600_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3599_cast_fp16)[name = string("op_3600_cast_fp16")];
            tensor<int32, [4]> var_3605 = const()[name = string("op_3605"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_31_cast_fp16 = reshape(shape = var_3605, x = var_3600_cast_fp16)[name = string("P_31_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3607_cast_fp16 = exp(x = P_31_cast_fp16)[name = string("op_3607_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_31_cast_fp16 = mul(x = var_3607_cast_fp16, y = tril_incl_to_fp16)[name = string("D_31_cast_fp16")];
            bool kk_31_transpose_x_1 = const()[name = string("kk_31_transpose_x_1"), val = bool(false)];
            bool kk_31_transpose_y_1 = const()[name = string("kk_31_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_31_cast_fp16 = matmul(transpose_x = kk_31_transpose_x_1, transpose_y = kk_31_transpose_y_1, x = kb_31_cast_fp16, y = k_55_cast_fp16)[name = string("kk_31_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_3613_cast_fp16 = mul(x = kk_31_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_3613_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3614_cast_fp16 = mul(x = var_3613_cast_fp16, y = D_31_cast_fp16)[name = string("op_3614_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_31_cast_fp16 = mul(x = var_3614_cast_fp16, y = cum_tail_to_fp16)[name = string("N_31_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3616_cast_fp16 = mul(x = D_31_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3616_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_31_cast_fp16 = mul(x = kk_31_cast_fp16, y = var_3616_cast_fp16)[name = string("Na_31_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_121_cast_fp16 = add(x = eye_to_fp16, y = N_31_cast_fp16)[name = string("Minv_121_cast_fp16")];
            bool Np_91_transpose_x_0 = const()[name = string("Np_91_transpose_x_0"), val = bool(false)];
            bool Np_91_transpose_y_0 = const()[name = string("Np_91_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_91_cast_fp16 = matmul(transpose_x = Np_91_transpose_x_0, transpose_y = Np_91_transpose_y_0, x = N_31_cast_fp16, y = Na_31_cast_fp16)[name = string("Np_91_cast_fp16")];
            bool Npa_61_transpose_x_0 = const()[name = string("Npa_61_transpose_x_0"), val = bool(false)];
            bool Npa_61_transpose_y_0 = const()[name = string("Npa_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_61_cast_fp16 = matmul(transpose_x = Npa_61_transpose_x_0, transpose_y = Npa_61_transpose_y_0, x = Na_31_cast_fp16, y = N_31_cast_fp16)[name = string("Npa_61_cast_fp16")];
            bool var_3622_transpose_x_0 = const()[name = string("op_3622_transpose_x_0"), val = bool(false)];
            bool var_3622_transpose_y_0 = const()[name = string("op_3622_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3622_cast_fp16 = matmul(transpose_x = var_3622_transpose_x_0, transpose_y = var_3622_transpose_y_0, x = Minv_121_cast_fp16, y = Np_91_cast_fp16)[name = string("op_3622_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_123_cast_fp16 = add(x = Minv_121_cast_fp16, y = var_3622_cast_fp16)[name = string("Minv_123_cast_fp16")];
            bool Np_93_transpose_x_0 = const()[name = string("Np_93_transpose_x_0"), val = bool(false)];
            bool Np_93_transpose_y_0 = const()[name = string("Np_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_93_cast_fp16 = matmul(transpose_x = Np_93_transpose_x_0, transpose_y = Np_93_transpose_y_0, x = Np_91_cast_fp16, y = Npa_61_cast_fp16)[name = string("Np_93_cast_fp16")];
            bool Npa_63_transpose_x_0 = const()[name = string("Npa_63_transpose_x_0"), val = bool(false)];
            bool Npa_63_transpose_y_0 = const()[name = string("Npa_63_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_63_cast_fp16 = matmul(transpose_x = Npa_63_transpose_x_0, transpose_y = Npa_63_transpose_y_0, x = Npa_61_cast_fp16, y = Np_91_cast_fp16)[name = string("Npa_63_cast_fp16")];
            bool var_3627_transpose_x_0 = const()[name = string("op_3627_transpose_x_0"), val = bool(false)];
            bool var_3627_transpose_y_0 = const()[name = string("op_3627_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3627_cast_fp16 = matmul(transpose_x = var_3627_transpose_x_0, transpose_y = var_3627_transpose_y_0, x = Minv_123_cast_fp16, y = Np_93_cast_fp16)[name = string("op_3627_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_125_cast_fp16 = add(x = Minv_123_cast_fp16, y = var_3627_cast_fp16)[name = string("Minv_125_cast_fp16")];
            bool Np_95_transpose_x_0 = const()[name = string("Np_95_transpose_x_0"), val = bool(false)];
            bool Np_95_transpose_y_0 = const()[name = string("Np_95_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_95_cast_fp16 = matmul(transpose_x = Np_95_transpose_x_0, transpose_y = Np_95_transpose_y_0, x = Np_93_cast_fp16, y = Npa_63_cast_fp16)[name = string("Np_95_cast_fp16")];
            bool var_3631_transpose_x_0 = const()[name = string("op_3631_transpose_x_0"), val = bool(false)];
            bool var_3631_transpose_y_0 = const()[name = string("op_3631_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3631_cast_fp16 = matmul(transpose_x = var_3631_transpose_x_0, transpose_y = var_3631_transpose_y_0, x = Minv_125_cast_fp16, y = Np_95_cast_fp16)[name = string("op_3631_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_127_cast_fp16 = add(x = Minv_125_cast_fp16, y = var_3631_cast_fp16)[name = string("Minv_127_cast_fp16")];
            bool u_31_transpose_x_0 = const()[name = string("u_31_transpose_x_0"), val = bool(false)];
            bool u_31_transpose_y_0 = const()[name = string("u_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_31_cast_fp16 = matmul(transpose_x = u_31_transpose_x_0, transpose_y = u_31_transpose_y_0, x = Minv_127_cast_fp16, y = vb_37_cast_fp16)[name = string("u_31_cast_fp16")];
            bool var_3638_transpose_x_1 = const()[name = string("op_3638_transpose_x_1"), val = bool(false)];
            bool var_3638_transpose_y_1 = const()[name = string("op_3638_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_3638_cast_fp16 = matmul(transpose_x = var_3638_transpose_x_1, transpose_y = var_3638_transpose_y_1, x = q_55_cast_fp16, y = k_55_cast_fp16)[name = string("op_3638_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_31_cast_fp16 = mul(x = var_3638_cast_fp16, y = D_31_cast_fp16)[name = string("A_31_cast_fp16")];
            tensor<fp16, [16]> var_3640_bias_0_to_fp16 = const()[name = string("op_3640_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_3640_cast_fp16 = linear(bias = var_3640_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_37_cast_fp16)[name = string("op_3640_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_25_cast_fp16 = exp(x = var_3640_cast_fp16)[name = string("cum_25_cast_fp16")];
            tensor<int32, [1]> var_3643_axes_0 = const()[name = string("op_3643_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_3643_cast_fp16 = expand_dims(axes = var_3643_axes_0, x = cum_25_cast_fp16)[name = string("op_3643_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3644_cast_fp16 = mul(x = kb_31_cast_fp16, y = var_3643_cast_fp16)[name = string("op_3644_cast_fp16")];
            bool w_31_transpose_x_0 = const()[name = string("w_31_transpose_x_0"), val = bool(false)];
            bool w_31_transpose_y_0 = const()[name = string("w_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_31_cast_fp16 = matmul(transpose_x = w_31_transpose_x_0, transpose_y = w_31_transpose_y_0, x = Minv_127_cast_fp16, y = var_3644_cast_fp16)[name = string("w_31_cast_fp16")];
            bool var_3646_transpose_x_0 = const()[name = string("op_3646_transpose_x_0"), val = bool(false)];
            bool var_3646_transpose_y_0 = const()[name = string("op_3646_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3646_cast_fp16 = matmul(transpose_x = var_3646_transpose_x_0, transpose_y = var_3646_transpose_y_0, x = w_31_cast_fp16, y = S_25_cast_fp16)[name = string("op_3646_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_25_cast_fp16 = sub(x = u_31_cast_fp16, y = var_3646_cast_fp16)[name = string("vnew_25_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3651_cast_fp16 = mul(x = q_55_cast_fp16, y = var_3643_cast_fp16)[name = string("op_3651_cast_fp16")];
            bool var_3652_transpose_x_0 = const()[name = string("op_3652_transpose_x_0"), val = bool(false)];
            bool var_3652_transpose_y_0 = const()[name = string("op_3652_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3652_cast_fp16 = matmul(transpose_x = var_3652_transpose_x_0, transpose_y = var_3652_transpose_y_0, x = var_3651_cast_fp16, y = S_25_cast_fp16)[name = string("op_3652_cast_fp16")];
            bool var_3653_transpose_x_0 = const()[name = string("op_3653_transpose_x_0"), val = bool(false)];
            bool var_3653_transpose_y_0 = const()[name = string("op_3653_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3653_cast_fp16 = matmul(transpose_x = var_3653_transpose_x_0, transpose_y = var_3653_transpose_y_0, x = A_31_cast_fp16, y = vnew_25_cast_fp16)[name = string("op_3653_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3655_cast_fp16 = add(x = var_3652_cast_fp16, y = var_3653_cast_fp16)[name = string("op_3655_cast_fp16")];
            tensor<fp16, [16]> var_3656_bias_0_to_fp16 = const()[name = string("op_3656_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_3656_cast_fp16 = linear(bias = var_3656_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_37_cast_fp16)[name = string("op_3656_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_27_cast_fp16 = exp(x = var_3656_cast_fp16)[name = string("tail_27_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_53_to_fp16 = const()[name = string("transpose_53_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_3658_bias_0_to_fp16 = const()[name = string("op_3658_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_3658_cast_fp16 = linear(bias = var_3658_bias_0_to_fp16, weight = transpose_53_to_fp16, x = gdec_37_cast_fp16)[name = string("op_3658_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_3659_cast_fp16 = exp(x = var_3658_cast_fp16)[name = string("op_3659_cast_fp16")];
            tensor<int32, [1]> last_23_axes_0 = const()[name = string("last_23_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_23_cast_fp16 = expand_dims(axes = last_23_axes_0, x = var_3659_cast_fp16)[name = string("last_23_cast_fp16")];
            tensor<int32, [1]> var_3663_axes_0 = const()[name = string("op_3663_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_3663_cast_fp16 = expand_dims(axes = var_3663_axes_0, x = tail_27_cast_fp16)[name = string("op_3663_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3664_cast_fp16 = mul(x = k_55_cast_fp16, y = var_3663_cast_fp16)[name = string("op_3664_cast_fp16")];
            bool upd_21_transpose_x_1 = const()[name = string("upd_21_transpose_x_1"), val = bool(true)];
            bool upd_21_transpose_y_1 = const()[name = string("upd_21_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_21_cast_fp16 = matmul(transpose_x = upd_21_transpose_x_1, transpose_y = upd_21_transpose_y_1, x = var_3664_cast_fp16, y = vnew_25_cast_fp16)[name = string("upd_21_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_3669_cast_fp16 = mul(x = S_25_cast_fp16, y = last_23_cast_fp16)[name = string("op_3669_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_27_cast_fp16 = add(x = var_3669_cast_fp16, y = upd_21_cast_fp16)[name = string("S_27_cast_fp16")];
            tensor<int32, [4]> q_57_begin_0 = const()[name = string("q_57_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> q_57_end_0 = const()[name = string("q_57_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> q_57_end_mask_0 = const()[name = string("q_57_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_57_cast_fp16 = slice_by_index(begin = q_57_begin_0, end = q_57_end_0, end_mask = q_57_end_mask_0, x = q_47_cast_fp16)[name = string("q_57_cast_fp16")];
            tensor<int32, [4]> k_57_begin_0 = const()[name = string("k_57_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> k_57_end_0 = const()[name = string("k_57_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> k_57_end_mask_0 = const()[name = string("k_57_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_57_cast_fp16 = slice_by_index(begin = k_57_begin_0, end = k_57_end_0, end_mask = k_57_end_mask_0, x = k_47_cast_fp16)[name = string("k_57_cast_fp16")];
            tensor<int32, [4]> kb_33_begin_0 = const()[name = string("kb_33_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> kb_33_end_0 = const()[name = string("kb_33_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> kb_33_end_mask_0 = const()[name = string("kb_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_33_cast_fp16 = slice_by_index(begin = kb_33_begin_0, end = kb_33_end_0, end_mask = kb_33_end_mask_0, x = var_3110_cast_fp16)[name = string("kb_33_cast_fp16")];
            tensor<int32, [4]> vb_39_begin_0 = const()[name = string("vb_39_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> vb_39_end_0 = const()[name = string("vb_39_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> vb_39_end_mask_0 = const()[name = string("vb_39_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_39_cast_fp16 = slice_by_index(begin = vb_39_begin_0, end = vb_39_end_0, end_mask = vb_39_end_mask_0, x = vb_29_cast_fp16)[name = string("vb_39_cast_fp16")];
            tensor<int32, [3]> gdec_39_begin_0 = const()[name = string("gdec_39_begin_0"), val = tensor<int32, [3]>([0, 0, 64])];
            tensor<int32, [3]> gdec_39_end_0 = const()[name = string("gdec_39_end_0"), val = tensor<int32, [3]>([1, 4, 80])];
            tensor<bool, [3]> gdec_39_end_mask_0 = const()[name = string("gdec_39_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_39_cast_fp16 = slice_by_index(begin = gdec_39_begin_0, end = gdec_39_end_0, end_mask = gdec_39_end_mask_0, x = gdec_29_cast_fp16)[name = string("gdec_39_cast_fp16")];
            tensor<int32, [2]> var_3752 = const()[name = string("op_3752"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_3753_cast_fp16 = reshape(shape = var_3752, x = gdec_39_cast_fp16)[name = string("op_3753_cast_fp16")];
            tensor<fp16, [4, 256]> var_3754_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3753_cast_fp16)[name = string("op_3754_cast_fp16")];
            tensor<int32, [4]> var_3759 = const()[name = string("op_3759"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_33_cast_fp16 = reshape(shape = var_3759, x = var_3754_cast_fp16)[name = string("P_33_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3761_cast_fp16 = exp(x = P_33_cast_fp16)[name = string("op_3761_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_33_cast_fp16 = mul(x = var_3761_cast_fp16, y = tril_incl_to_fp16)[name = string("D_33_cast_fp16")];
            bool kk_33_transpose_x_1 = const()[name = string("kk_33_transpose_x_1"), val = bool(false)];
            bool kk_33_transpose_y_1 = const()[name = string("kk_33_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_33_cast_fp16 = matmul(transpose_x = kk_33_transpose_x_1, transpose_y = kk_33_transpose_y_1, x = kb_33_cast_fp16, y = k_57_cast_fp16)[name = string("kk_33_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_3767_cast_fp16 = mul(x = kk_33_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_3767_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3768_cast_fp16 = mul(x = var_3767_cast_fp16, y = D_33_cast_fp16)[name = string("op_3768_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_33_cast_fp16 = mul(x = var_3768_cast_fp16, y = cum_tail_to_fp16)[name = string("N_33_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3770_cast_fp16 = mul(x = D_33_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3770_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_33_cast_fp16 = mul(x = kk_33_cast_fp16, y = var_3770_cast_fp16)[name = string("Na_33_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_129_cast_fp16 = add(x = eye_to_fp16, y = N_33_cast_fp16)[name = string("Minv_129_cast_fp16")];
            bool Np_97_transpose_x_0 = const()[name = string("Np_97_transpose_x_0"), val = bool(false)];
            bool Np_97_transpose_y_0 = const()[name = string("Np_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_97_cast_fp16 = matmul(transpose_x = Np_97_transpose_x_0, transpose_y = Np_97_transpose_y_0, x = N_33_cast_fp16, y = Na_33_cast_fp16)[name = string("Np_97_cast_fp16")];
            bool Npa_65_transpose_x_0 = const()[name = string("Npa_65_transpose_x_0"), val = bool(false)];
            bool Npa_65_transpose_y_0 = const()[name = string("Npa_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_65_cast_fp16 = matmul(transpose_x = Npa_65_transpose_x_0, transpose_y = Npa_65_transpose_y_0, x = Na_33_cast_fp16, y = N_33_cast_fp16)[name = string("Npa_65_cast_fp16")];
            bool var_3776_transpose_x_0 = const()[name = string("op_3776_transpose_x_0"), val = bool(false)];
            bool var_3776_transpose_y_0 = const()[name = string("op_3776_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3776_cast_fp16 = matmul(transpose_x = var_3776_transpose_x_0, transpose_y = var_3776_transpose_y_0, x = Minv_129_cast_fp16, y = Np_97_cast_fp16)[name = string("op_3776_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_131_cast_fp16 = add(x = Minv_129_cast_fp16, y = var_3776_cast_fp16)[name = string("Minv_131_cast_fp16")];
            bool Np_99_transpose_x_0 = const()[name = string("Np_99_transpose_x_0"), val = bool(false)];
            bool Np_99_transpose_y_0 = const()[name = string("Np_99_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_99_cast_fp16 = matmul(transpose_x = Np_99_transpose_x_0, transpose_y = Np_99_transpose_y_0, x = Np_97_cast_fp16, y = Npa_65_cast_fp16)[name = string("Np_99_cast_fp16")];
            bool Npa_67_transpose_x_0 = const()[name = string("Npa_67_transpose_x_0"), val = bool(false)];
            bool Npa_67_transpose_y_0 = const()[name = string("Npa_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_67_cast_fp16 = matmul(transpose_x = Npa_67_transpose_x_0, transpose_y = Npa_67_transpose_y_0, x = Npa_65_cast_fp16, y = Np_97_cast_fp16)[name = string("Npa_67_cast_fp16")];
            bool var_3781_transpose_x_0 = const()[name = string("op_3781_transpose_x_0"), val = bool(false)];
            bool var_3781_transpose_y_0 = const()[name = string("op_3781_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3781_cast_fp16 = matmul(transpose_x = var_3781_transpose_x_0, transpose_y = var_3781_transpose_y_0, x = Minv_131_cast_fp16, y = Np_99_cast_fp16)[name = string("op_3781_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_133_cast_fp16 = add(x = Minv_131_cast_fp16, y = var_3781_cast_fp16)[name = string("Minv_133_cast_fp16")];
            bool Np_101_transpose_x_0 = const()[name = string("Np_101_transpose_x_0"), val = bool(false)];
            bool Np_101_transpose_y_0 = const()[name = string("Np_101_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_101_cast_fp16 = matmul(transpose_x = Np_101_transpose_x_0, transpose_y = Np_101_transpose_y_0, x = Np_99_cast_fp16, y = Npa_67_cast_fp16)[name = string("Np_101_cast_fp16")];
            bool var_3785_transpose_x_0 = const()[name = string("op_3785_transpose_x_0"), val = bool(false)];
            bool var_3785_transpose_y_0 = const()[name = string("op_3785_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3785_cast_fp16 = matmul(transpose_x = var_3785_transpose_x_0, transpose_y = var_3785_transpose_y_0, x = Minv_133_cast_fp16, y = Np_101_cast_fp16)[name = string("op_3785_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_135_cast_fp16 = add(x = Minv_133_cast_fp16, y = var_3785_cast_fp16)[name = string("Minv_135_cast_fp16")];
            bool u_33_transpose_x_0 = const()[name = string("u_33_transpose_x_0"), val = bool(false)];
            bool u_33_transpose_y_0 = const()[name = string("u_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_33_cast_fp16 = matmul(transpose_x = u_33_transpose_x_0, transpose_y = u_33_transpose_y_0, x = Minv_135_cast_fp16, y = vb_39_cast_fp16)[name = string("u_33_cast_fp16")];
            bool var_3792_transpose_x_1 = const()[name = string("op_3792_transpose_x_1"), val = bool(false)];
            bool var_3792_transpose_y_1 = const()[name = string("op_3792_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_3792_cast_fp16 = matmul(transpose_x = var_3792_transpose_x_1, transpose_y = var_3792_transpose_y_1, x = q_57_cast_fp16, y = k_57_cast_fp16)[name = string("op_3792_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_33_cast_fp16 = mul(x = var_3792_cast_fp16, y = D_33_cast_fp16)[name = string("A_33_cast_fp16")];
            tensor<fp16, [16]> var_3794_bias_0_to_fp16 = const()[name = string("op_3794_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_3794_cast_fp16 = linear(bias = var_3794_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_39_cast_fp16)[name = string("op_3794_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_27_cast_fp16 = exp(x = var_3794_cast_fp16)[name = string("cum_27_cast_fp16")];
            tensor<int32, [1]> var_3797_axes_0 = const()[name = string("op_3797_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_3797_cast_fp16 = expand_dims(axes = var_3797_axes_0, x = cum_27_cast_fp16)[name = string("op_3797_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3798_cast_fp16 = mul(x = kb_33_cast_fp16, y = var_3797_cast_fp16)[name = string("op_3798_cast_fp16")];
            bool w_33_transpose_x_0 = const()[name = string("w_33_transpose_x_0"), val = bool(false)];
            bool w_33_transpose_y_0 = const()[name = string("w_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_33_cast_fp16 = matmul(transpose_x = w_33_transpose_x_0, transpose_y = w_33_transpose_y_0, x = Minv_135_cast_fp16, y = var_3798_cast_fp16)[name = string("w_33_cast_fp16")];
            bool var_3800_transpose_x_0 = const()[name = string("op_3800_transpose_x_0"), val = bool(false)];
            bool var_3800_transpose_y_0 = const()[name = string("op_3800_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3800_cast_fp16 = matmul(transpose_x = var_3800_transpose_x_0, transpose_y = var_3800_transpose_y_0, x = w_33_cast_fp16, y = S_27_cast_fp16)[name = string("op_3800_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_27_cast_fp16 = sub(x = u_33_cast_fp16, y = var_3800_cast_fp16)[name = string("vnew_27_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3805_cast_fp16 = mul(x = q_57_cast_fp16, y = var_3797_cast_fp16)[name = string("op_3805_cast_fp16")];
            bool var_3806_transpose_x_0 = const()[name = string("op_3806_transpose_x_0"), val = bool(false)];
            bool var_3806_transpose_y_0 = const()[name = string("op_3806_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3806_cast_fp16 = matmul(transpose_x = var_3806_transpose_x_0, transpose_y = var_3806_transpose_y_0, x = var_3805_cast_fp16, y = S_27_cast_fp16)[name = string("op_3806_cast_fp16")];
            bool var_3807_transpose_x_0 = const()[name = string("op_3807_transpose_x_0"), val = bool(false)];
            bool var_3807_transpose_y_0 = const()[name = string("op_3807_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3807_cast_fp16 = matmul(transpose_x = var_3807_transpose_x_0, transpose_y = var_3807_transpose_y_0, x = A_33_cast_fp16, y = vnew_27_cast_fp16)[name = string("op_3807_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3809_cast_fp16 = add(x = var_3806_cast_fp16, y = var_3807_cast_fp16)[name = string("op_3809_cast_fp16")];
            tensor<fp16, [16]> var_3810_bias_0_to_fp16 = const()[name = string("op_3810_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_3810_cast_fp16 = linear(bias = var_3810_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_39_cast_fp16)[name = string("op_3810_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_29_cast_fp16 = exp(x = var_3810_cast_fp16)[name = string("tail_29_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_57_to_fp16 = const()[name = string("transpose_57_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_3812_bias_0_to_fp16 = const()[name = string("op_3812_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_3812_cast_fp16 = linear(bias = var_3812_bias_0_to_fp16, weight = transpose_57_to_fp16, x = gdec_39_cast_fp16)[name = string("op_3812_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_3813_cast_fp16 = exp(x = var_3812_cast_fp16)[name = string("op_3813_cast_fp16")];
            tensor<int32, [1]> last_25_axes_0 = const()[name = string("last_25_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_25_cast_fp16 = expand_dims(axes = last_25_axes_0, x = var_3813_cast_fp16)[name = string("last_25_cast_fp16")];
            tensor<int32, [1]> var_3817_axes_0 = const()[name = string("op_3817_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_3817_cast_fp16 = expand_dims(axes = var_3817_axes_0, x = tail_29_cast_fp16)[name = string("op_3817_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3818_cast_fp16 = mul(x = k_57_cast_fp16, y = var_3817_cast_fp16)[name = string("op_3818_cast_fp16")];
            bool upd_23_transpose_x_1 = const()[name = string("upd_23_transpose_x_1"), val = bool(true)];
            bool upd_23_transpose_y_1 = const()[name = string("upd_23_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_23_cast_fp16 = matmul(transpose_x = upd_23_transpose_x_1, transpose_y = upd_23_transpose_y_1, x = var_3818_cast_fp16, y = vnew_27_cast_fp16)[name = string("upd_23_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_3823_cast_fp16 = mul(x = S_27_cast_fp16, y = last_25_cast_fp16)[name = string("op_3823_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_29_cast_fp16 = add(x = var_3823_cast_fp16, y = upd_23_cast_fp16)[name = string("S_29_cast_fp16")];
            tensor<int32, [4]> q_59_begin_0 = const()[name = string("q_59_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> q_59_end_0 = const()[name = string("q_59_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_59_end_mask_0 = const()[name = string("q_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_59_cast_fp16 = slice_by_index(begin = q_59_begin_0, end = q_59_end_0, end_mask = q_59_end_mask_0, x = q_47_cast_fp16)[name = string("q_59_cast_fp16")];
            tensor<int32, [4]> k_59_begin_0 = const()[name = string("k_59_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> k_59_end_0 = const()[name = string("k_59_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_59_end_mask_0 = const()[name = string("k_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_59_cast_fp16 = slice_by_index(begin = k_59_begin_0, end = k_59_end_0, end_mask = k_59_end_mask_0, x = k_47_cast_fp16)[name = string("k_59_cast_fp16")];
            tensor<int32, [4]> kb_35_begin_0 = const()[name = string("kb_35_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> kb_35_end_0 = const()[name = string("kb_35_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_35_end_mask_0 = const()[name = string("kb_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_35_cast_fp16 = slice_by_index(begin = kb_35_begin_0, end = kb_35_end_0, end_mask = kb_35_end_mask_0, x = var_3110_cast_fp16)[name = string("kb_35_cast_fp16")];
            tensor<int32, [4]> vb_41_begin_0 = const()[name = string("vb_41_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> vb_41_end_0 = const()[name = string("vb_41_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_41_end_mask_0 = const()[name = string("vb_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_41_cast_fp16 = slice_by_index(begin = vb_41_begin_0, end = vb_41_end_0, end_mask = vb_41_end_mask_0, x = vb_29_cast_fp16)[name = string("vb_41_cast_fp16")];
            tensor<int32, [3]> gdec_41_begin_0 = const()[name = string("gdec_41_begin_0"), val = tensor<int32, [3]>([0, 0, 80])];
            tensor<int32, [3]> gdec_41_end_0 = const()[name = string("gdec_41_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_41_end_mask_0 = const()[name = string("gdec_41_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_41_cast_fp16 = slice_by_index(begin = gdec_41_begin_0, end = gdec_41_end_0, end_mask = gdec_41_end_mask_0, x = gdec_29_cast_fp16)[name = string("gdec_41_cast_fp16")];
            tensor<int32, [2]> var_3906 = const()[name = string("op_3906"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_3907_cast_fp16 = reshape(shape = var_3906, x = gdec_41_cast_fp16)[name = string("op_3907_cast_fp16")];
            tensor<fp16, [4, 256]> var_3908_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3907_cast_fp16)[name = string("op_3908_cast_fp16")];
            tensor<int32, [4]> var_3913 = const()[name = string("op_3913"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_35_cast_fp16 = reshape(shape = var_3913, x = var_3908_cast_fp16)[name = string("P_35_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3915_cast_fp16 = exp(x = P_35_cast_fp16)[name = string("op_3915_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_35_cast_fp16 = mul(x = var_3915_cast_fp16, y = tril_incl_to_fp16)[name = string("D_35_cast_fp16")];
            bool kk_35_transpose_x_1 = const()[name = string("kk_35_transpose_x_1"), val = bool(false)];
            bool kk_35_transpose_y_1 = const()[name = string("kk_35_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_35_cast_fp16 = matmul(transpose_x = kk_35_transpose_x_1, transpose_y = kk_35_transpose_y_1, x = kb_35_cast_fp16, y = k_59_cast_fp16)[name = string("kk_35_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_3921_cast_fp16 = mul(x = kk_35_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_3921_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3922_cast_fp16 = mul(x = var_3921_cast_fp16, y = D_35_cast_fp16)[name = string("op_3922_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_35_cast_fp16 = mul(x = var_3922_cast_fp16, y = cum_tail_to_fp16)[name = string("N_35_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_3924_cast_fp16 = mul(x = D_35_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3924_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_35_cast_fp16 = mul(x = kk_35_cast_fp16, y = var_3924_cast_fp16)[name = string("Na_35_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_137_cast_fp16 = add(x = eye_to_fp16, y = N_35_cast_fp16)[name = string("Minv_137_cast_fp16")];
            bool Np_103_transpose_x_0 = const()[name = string("Np_103_transpose_x_0"), val = bool(false)];
            bool Np_103_transpose_y_0 = const()[name = string("Np_103_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_103_cast_fp16 = matmul(transpose_x = Np_103_transpose_x_0, transpose_y = Np_103_transpose_y_0, x = N_35_cast_fp16, y = Na_35_cast_fp16)[name = string("Np_103_cast_fp16")];
            bool Npa_69_transpose_x_0 = const()[name = string("Npa_69_transpose_x_0"), val = bool(false)];
            bool Npa_69_transpose_y_0 = const()[name = string("Npa_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_69_cast_fp16 = matmul(transpose_x = Npa_69_transpose_x_0, transpose_y = Npa_69_transpose_y_0, x = Na_35_cast_fp16, y = N_35_cast_fp16)[name = string("Npa_69_cast_fp16")];
            bool var_3930_transpose_x_0 = const()[name = string("op_3930_transpose_x_0"), val = bool(false)];
            bool var_3930_transpose_y_0 = const()[name = string("op_3930_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3930_cast_fp16 = matmul(transpose_x = var_3930_transpose_x_0, transpose_y = var_3930_transpose_y_0, x = Minv_137_cast_fp16, y = Np_103_cast_fp16)[name = string("op_3930_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_139_cast_fp16 = add(x = Minv_137_cast_fp16, y = var_3930_cast_fp16)[name = string("Minv_139_cast_fp16")];
            bool Np_105_transpose_x_0 = const()[name = string("Np_105_transpose_x_0"), val = bool(false)];
            bool Np_105_transpose_y_0 = const()[name = string("Np_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_105_cast_fp16 = matmul(transpose_x = Np_105_transpose_x_0, transpose_y = Np_105_transpose_y_0, x = Np_103_cast_fp16, y = Npa_69_cast_fp16)[name = string("Np_105_cast_fp16")];
            bool Npa_71_transpose_x_0 = const()[name = string("Npa_71_transpose_x_0"), val = bool(false)];
            bool Npa_71_transpose_y_0 = const()[name = string("Npa_71_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_71_cast_fp16 = matmul(transpose_x = Npa_71_transpose_x_0, transpose_y = Npa_71_transpose_y_0, x = Npa_69_cast_fp16, y = Np_103_cast_fp16)[name = string("Npa_71_cast_fp16")];
            bool var_3935_transpose_x_0 = const()[name = string("op_3935_transpose_x_0"), val = bool(false)];
            bool var_3935_transpose_y_0 = const()[name = string("op_3935_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3935_cast_fp16 = matmul(transpose_x = var_3935_transpose_x_0, transpose_y = var_3935_transpose_y_0, x = Minv_139_cast_fp16, y = Np_105_cast_fp16)[name = string("op_3935_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_141_cast_fp16 = add(x = Minv_139_cast_fp16, y = var_3935_cast_fp16)[name = string("Minv_141_cast_fp16")];
            bool Np_107_transpose_x_0 = const()[name = string("Np_107_transpose_x_0"), val = bool(false)];
            bool Np_107_transpose_y_0 = const()[name = string("Np_107_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_107_cast_fp16 = matmul(transpose_x = Np_107_transpose_x_0, transpose_y = Np_107_transpose_y_0, x = Np_105_cast_fp16, y = Npa_71_cast_fp16)[name = string("Np_107_cast_fp16")];
            bool var_3939_transpose_x_0 = const()[name = string("op_3939_transpose_x_0"), val = bool(false)];
            bool var_3939_transpose_y_0 = const()[name = string("op_3939_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_3939_cast_fp16 = matmul(transpose_x = var_3939_transpose_x_0, transpose_y = var_3939_transpose_y_0, x = Minv_141_cast_fp16, y = Np_107_cast_fp16)[name = string("op_3939_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_143_cast_fp16 = add(x = Minv_141_cast_fp16, y = var_3939_cast_fp16)[name = string("Minv_143_cast_fp16")];
            bool u_35_transpose_x_0 = const()[name = string("u_35_transpose_x_0"), val = bool(false)];
            bool u_35_transpose_y_0 = const()[name = string("u_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_35_cast_fp16 = matmul(transpose_x = u_35_transpose_x_0, transpose_y = u_35_transpose_y_0, x = Minv_143_cast_fp16, y = vb_41_cast_fp16)[name = string("u_35_cast_fp16")];
            bool var_3946_transpose_x_1 = const()[name = string("op_3946_transpose_x_1"), val = bool(false)];
            bool var_3946_transpose_y_1 = const()[name = string("op_3946_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_3946_cast_fp16 = matmul(transpose_x = var_3946_transpose_x_1, transpose_y = var_3946_transpose_y_1, x = q_59_cast_fp16, y = k_59_cast_fp16)[name = string("op_3946_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_35_cast_fp16 = mul(x = var_3946_cast_fp16, y = D_35_cast_fp16)[name = string("A_35_cast_fp16")];
            tensor<fp16, [16]> var_3948_bias_0_to_fp16 = const()[name = string("op_3948_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_3948_cast_fp16 = linear(bias = var_3948_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_41_cast_fp16)[name = string("op_3948_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_29_cast_fp16 = exp(x = var_3948_cast_fp16)[name = string("cum_29_cast_fp16")];
            tensor<int32, [1]> var_3951_axes_0 = const()[name = string("op_3951_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_3951_cast_fp16 = expand_dims(axes = var_3951_axes_0, x = cum_29_cast_fp16)[name = string("op_3951_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3952_cast_fp16 = mul(x = kb_35_cast_fp16, y = var_3951_cast_fp16)[name = string("op_3952_cast_fp16")];
            bool w_35_transpose_x_0 = const()[name = string("w_35_transpose_x_0"), val = bool(false)];
            bool w_35_transpose_y_0 = const()[name = string("w_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_35_cast_fp16 = matmul(transpose_x = w_35_transpose_x_0, transpose_y = w_35_transpose_y_0, x = Minv_143_cast_fp16, y = var_3952_cast_fp16)[name = string("w_35_cast_fp16")];
            bool var_3954_transpose_x_0 = const()[name = string("op_3954_transpose_x_0"), val = bool(false)];
            bool var_3954_transpose_y_0 = const()[name = string("op_3954_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3954_cast_fp16 = matmul(transpose_x = var_3954_transpose_x_0, transpose_y = var_3954_transpose_y_0, x = w_35_cast_fp16, y = S_29_cast_fp16)[name = string("op_3954_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_29_cast_fp16 = sub(x = u_35_cast_fp16, y = var_3954_cast_fp16)[name = string("vnew_29_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_3959_cast_fp16 = mul(x = q_59_cast_fp16, y = var_3951_cast_fp16)[name = string("op_3959_cast_fp16")];
            bool var_3960_transpose_x_0 = const()[name = string("op_3960_transpose_x_0"), val = bool(false)];
            bool var_3960_transpose_y_0 = const()[name = string("op_3960_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3960_cast_fp16 = matmul(transpose_x = var_3960_transpose_x_0, transpose_y = var_3960_transpose_y_0, x = var_3959_cast_fp16, y = S_29_cast_fp16)[name = string("op_3960_cast_fp16")];
            bool var_3961_transpose_x_0 = const()[name = string("op_3961_transpose_x_0"), val = bool(false)];
            bool var_3961_transpose_y_0 = const()[name = string("op_3961_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_3961_cast_fp16 = matmul(transpose_x = var_3961_transpose_x_0, transpose_y = var_3961_transpose_y_0, x = A_35_cast_fp16, y = vnew_29_cast_fp16)[name = string("op_3961_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_5_cast_fp16 = add(x = var_3960_cast_fp16, y = var_3961_cast_fp16)[name = string("oc_5_cast_fp16")];
            int32 var_3965 = const()[name = string("op_3965"), val = int32(2)];
            bool o_13_interleave_0 = const()[name = string("o_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 96, 128]> o_13_cast_fp16 = concat(axis = var_3965, interleave = o_13_interleave_0, values = (var_3200_cast_fp16, var_3347_cast_fp16, var_3501_cast_fp16, var_3655_cast_fp16, var_3809_cast_fp16, oc_5_cast_fp16))[name = string("o_13_cast_fp16")];
            tensor<int32, [4]> var_3971 = const()[name = string("op_3971"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3977 = const()[name = string("op_3977"), val = tensor<int32, [4]>([1, 96, 4, 128])];
            tensor<fp16, [1, 96, 4, 128]> input_39_cast_fp16 = reshape(shape = var_3977, x = z_5_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> o_15_cast_fp16 = transpose(perm = var_3971, x = o_13_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [1, 96, 4, 128]> var_3979_cast_fp16 = mul(x = o_15_cast_fp16, y = o_15_cast_fp16)[name = string("op_3979_cast_fp16")];
            tensor<int32, [1]> var_3984_axes_0 = const()[name = string("op_3984_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3984_keep_dims_0 = const()[name = string("op_3984_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 4, 1]> var_3984_cast_fp16 = reduce_mean(axes = var_3984_axes_0, keep_dims = var_3984_keep_dims_0, x = var_3979_cast_fp16)[name = string("op_3984_cast_fp16")];
            fp16 var_3986_to_fp16 = const()[name = string("op_3986_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 96, 4, 1]> var_3987_cast_fp16 = add(x = var_3984_cast_fp16, y = var_3986_to_fp16)[name = string("op_3987_cast_fp16")];
            fp32 var_3988_epsilon_0 = const()[name = string("op_3988_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 4, 1]> var_3988_cast_fp16 = rsqrt(epsilon = var_3988_epsilon_0, x = var_3987_cast_fp16)[name = string("op_3988_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> var_3989_cast_fp16 = mul(x = o_15_cast_fp16, y = var_3988_cast_fp16)[name = string("op_3989_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> var_3991_cast_fp16 = silu(x = input_39_cast_fp16)[name = string("op_3991_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> o_17_cast_fp16 = mul(x = var_3989_cast_fp16, y = var_3991_cast_fp16)[name = string("o_17_cast_fp16")];
            tensor<int32, [3]> var_3996 = const()[name = string("op_3996"), val = tensor<int32, [3]>([1, 96, 512])];
            tensor<fp16, [1, 96, 512]> input_41_cast_fp16 = reshape(shape = var_3996, x = o_17_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [640, 512]> layers_2_out_weight_to_fp16 = const()[name = string("layers_2_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47457728)))];
            tensor<fp16, [1, 96, 640]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_out_weight_to_fp16, x = input_41_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [1, 96, 640]> x_21_cast_fp16 = add(x = x_17_cast_fp16, y = linear_9_cast_fp16)[name = string("x_21_cast_fp16")];
            fp16 var_4007_promoted_to_fp16 = const()[name = string("op_4007_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> var_4011_cast_fp16 = pow(x = x_21_cast_fp16, y = var_4007_promoted_to_fp16)[name = string("op_4011_cast_fp16")];
            tensor<int32, [1]> var_4013_axes_0 = const()[name = string("op_4013_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4013_keep_dims_0 = const()[name = string("op_4013_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 1]> var_4013_cast_fp16 = reduce_mean(axes = var_4013_axes_0, keep_dims = var_4013_keep_dims_0, x = var_4011_cast_fp16)[name = string("op_4013_cast_fp16")];
            fp16 var_4014_to_fp16 = const()[name = string("op_4014_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 96, 1]> var_4015_cast_fp16 = add(x = var_4013_cast_fp16, y = var_4014_to_fp16)[name = string("op_4015_cast_fp16")];
            fp32 var_4016_epsilon_0 = const()[name = string("op_4016_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 1]> var_4016_cast_fp16 = rsqrt(epsilon = var_4016_epsilon_0, x = var_4015_cast_fp16)[name = string("op_4016_cast_fp16")];
            tensor<fp16, [1, 96, 640]> var_4017_cast_fp16 = mul(x = x_21_cast_fp16, y = var_4016_cast_fp16)[name = string("op_4017_cast_fp16")];
            tensor<fp16, [640]> layers_2_n2_w1_to_fp16 = const()[name = string("layers_2_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48113152)))];
            tensor<fp16, [1, 96, 640]> input_43_cast_fp16 = mul(x = var_4017_cast_fp16, y = layers_2_n2_w1_to_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_2_gate_up_weight_to_fp16 = const()[name = string("layers_2_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48114496)))];
            tensor<fp16, [1, 96, 5120]> linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_2_gate_up_weight_to_fp16, x = input_43_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [3]> input_45_begin_0 = const()[name = string("input_45_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_45_end_0 = const()[name = string("input_45_end_0"), val = tensor<int32, [3]>([1, 96, 2560])];
            tensor<bool, [3]> input_45_end_mask_0 = const()[name = string("input_45_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 2560]> input_45_cast_fp16 = slice_by_index(begin = input_45_begin_0, end = input_45_end_0, end_mask = input_45_end_mask_0, x = linear_10_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [1, 96, 2560]> var_4027_cast_fp16 = silu(x = input_45_cast_fp16)[name = string("op_4027_cast_fp16")];
            tensor<int32, [3]> var_4032_begin_0 = const()[name = string("op_4032_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_4032_end_0 = const()[name = string("op_4032_end_0"), val = tensor<int32, [3]>([1, 96, 5120])];
            tensor<bool, [3]> var_4032_end_mask_0 = const()[name = string("op_4032_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 2560]> var_4032_cast_fp16 = slice_by_index(begin = var_4032_begin_0, end = var_4032_end_0, end_mask = var_4032_end_mask_0, x = linear_10_cast_fp16)[name = string("op_4032_cast_fp16")];
            tensor<fp16, [1, 96, 2560]> input_47_cast_fp16 = mul(x = var_4027_cast_fp16, y = var_4032_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_2_down_weight_to_fp16 = const()[name = string("layers_2_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54668160)))];
            tensor<fp16, [1, 96, 640]> linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_down_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 96, 640]> x_25_cast_fp16 = add(x = x_21_cast_fp16, y = linear_11_cast_fp16)[name = string("x_25_cast_fp16")];
            fp16 var_4045_to_fp16 = const()[name = string("op_4045_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 96, 640]> x_27_cast_fp16 = mul(x = x_25_cast_fp16, y = var_4045_to_fp16)[name = string("x_27_cast_fp16")];
            fp16 var_4043_promoted_to_fp16 = const()[name = string("op_4043_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> var_4047_cast_fp16 = pow(x = x_27_cast_fp16, y = var_4043_promoted_to_fp16)[name = string("op_4047_cast_fp16")];
            tensor<int32, [1]> var_4049_axes_0 = const()[name = string("op_4049_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4049_keep_dims_0 = const()[name = string("op_4049_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 1]> var_4049_cast_fp16 = reduce_mean(axes = var_4049_axes_0, keep_dims = var_4049_keep_dims_0, x = var_4047_cast_fp16)[name = string("op_4049_cast_fp16")];
            fp16 var_4050_to_fp16 = const()[name = string("op_4050_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 96, 1]> var_4051_cast_fp16 = add(x = var_4049_cast_fp16, y = var_4050_to_fp16)[name = string("op_4051_cast_fp16")];
            fp32 var_4052_epsilon_0 = const()[name = string("op_4052_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 1]> var_4052_cast_fp16 = rsqrt(epsilon = var_4052_epsilon_0, x = var_4051_cast_fp16)[name = string("op_4052_cast_fp16")];
            tensor<fp16, [1, 96, 640]> var_4053_cast_fp16 = mul(x = x_27_cast_fp16, y = var_4052_cast_fp16)[name = string("op_4053_cast_fp16")];
            tensor<fp16, [640]> layers_3_n1_w1_to_fp16 = const()[name = string("layers_3_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57945024)))];
            tensor<fp16, [1, 96, 640]> input_49_cast_fp16 = mul(x = var_4053_cast_fp16, y = layers_3_n1_w1_to_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1280, 640]> layers_3_qkv_weight_to_fp16 = const()[name = string("layers_3_qkv_weight_to_fp16"), val = tensor<fp16, [1280, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57946368)))];
            tensor<fp16, [1280]> linear_12_bias_0_to_fp16 = const()[name = string("linear_12_bias_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59584832)))];
            tensor<fp16, [1, 96, 1280]> linear_12_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = layers_3_qkv_weight_to_fp16, x = input_49_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [3]> var_4062_begin_0 = const()[name = string("op_4062_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_4062_end_0 = const()[name = string("op_4062_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> var_4062_end_mask_0 = const()[name = string("op_4062_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 1024]> var_4062_cast_fp16 = slice_by_index(begin = var_4062_begin_0, end = var_4062_end_0, end_mask = var_4062_end_mask_0, x = linear_12_cast_fp16)[name = string("op_4062_cast_fp16")];
            tensor<int32, [4]> var_4067 = const()[name = string("op_4067"), val = tensor<int32, [4]>([1, 96, 8, 128])];
            tensor<fp16, [1, 96, 8, 128]> qg_1_cast_fp16 = reshape(shape = var_4067, x = var_4062_cast_fp16)[name = string("qg_1_cast_fp16")];
            tensor<int32, [4]> x_29_begin_0 = const()[name = string("x_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_29_end_0 = const()[name = string("x_29_end_0"), val = tensor<int32, [4]>([1, 96, 8, 64])];
            tensor<bool, [4]> x_29_end_mask_0 = const()[name = string("x_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 96, 8, 64]> x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, x = qg_1_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [4]> var_4078_begin_0 = const()[name = string("op_4078_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> var_4078_end_0 = const()[name = string("op_4078_end_0"), val = tensor<int32, [4]>([1, 96, 8, 128])];
            tensor<bool, [4]> var_4078_end_mask_0 = const()[name = string("op_4078_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 96, 8, 64]> var_4078_cast_fp16 = slice_by_index(begin = var_4078_begin_0, end = var_4078_end_0, end_mask = var_4078_end_mask_0, x = qg_1_cast_fp16)[name = string("op_4078_cast_fp16")];
            tensor<int32, [3]> var_4082 = const()[name = string("op_4082"), val = tensor<int32, [3]>([1, 96, 512])];
            tensor<fp16, [1, 96, 512]> gate_1_cast_fp16 = reshape(shape = var_4082, x = var_4078_cast_fp16)[name = string("gate_1_cast_fp16")];
            tensor<int32, [3]> var_4088_begin_0 = const()[name = string("op_4088_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> var_4088_end_0 = const()[name = string("op_4088_end_0"), val = tensor<int32, [3]>([1, 96, 1152])];
            tensor<bool, [3]> var_4088_end_mask_0 = const()[name = string("op_4088_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 128]> var_4088_cast_fp16 = slice_by_index(begin = var_4088_begin_0, end = var_4088_end_0, end_mask = var_4088_end_mask_0, x = linear_12_cast_fp16)[name = string("op_4088_cast_fp16")];
            tensor<int32, [4]> var_4093 = const()[name = string("op_4093"), val = tensor<int32, [4]>([1, 96, 2, 64])];
            tensor<fp16, [1, 96, 2, 64]> x_35_cast_fp16 = reshape(shape = var_4093, x = var_4088_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<int32, [3]> var_4099_begin_0 = const()[name = string("op_4099_begin_0"), val = tensor<int32, [3]>([0, 0, 1152])];
            tensor<int32, [3]> var_4099_end_0 = const()[name = string("op_4099_end_0"), val = tensor<int32, [3]>([1, 96, 1280])];
            tensor<bool, [3]> var_4099_end_mask_0 = const()[name = string("op_4099_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 128]> var_4099_cast_fp16 = slice_by_index(begin = var_4099_begin_0, end = var_4099_end_0, end_mask = var_4099_end_mask_0, x = linear_12_cast_fp16)[name = string("op_4099_cast_fp16")];
            tensor<int32, [4]> var_4104 = const()[name = string("op_4104"), val = tensor<int32, [4]>([1, 96, 2, 64])];
            tensor<fp16, [1, 96, 2, 64]> var_4105_cast_fp16 = reshape(shape = var_4104, x = var_4099_cast_fp16)[name = string("op_4105_cast_fp16")];
            tensor<int32, [4]> var_4110 = const()[name = string("op_4110"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4116_promoted_to_fp16 = const()[name = string("op_4116_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 8, 64]> var_4120_cast_fp16 = pow(x = x_29_cast_fp16, y = var_4116_promoted_to_fp16)[name = string("op_4120_cast_fp16")];
            tensor<int32, [1]> var_4122_axes_0 = const()[name = string("op_4122_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4122_keep_dims_0 = const()[name = string("op_4122_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 8, 1]> var_4122_cast_fp16 = reduce_mean(axes = var_4122_axes_0, keep_dims = var_4122_keep_dims_0, x = var_4120_cast_fp16)[name = string("op_4122_cast_fp16")];
            fp16 var_4123_to_fp16 = const()[name = string("op_4123_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 96, 8, 1]> var_4124_cast_fp16 = add(x = var_4122_cast_fp16, y = var_4123_to_fp16)[name = string("op_4124_cast_fp16")];
            fp32 var_4125_epsilon_0 = const()[name = string("op_4125_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 8, 1]> var_4125_cast_fp16 = rsqrt(epsilon = var_4125_epsilon_0, x = var_4124_cast_fp16)[name = string("op_4125_cast_fp16")];
            tensor<fp16, [1, 96, 8, 64]> var_4126_cast_fp16 = mul(x = x_29_cast_fp16, y = var_4125_cast_fp16)[name = string("op_4126_cast_fp16")];
            tensor<fp16, [64]> layers_3_qn_w1_to_fp16 = const()[name = string("layers_3_qn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59587456)))];
            tensor<fp16, [1, 96, 8, 64]> var_4127_cast_fp16 = mul(x = var_4126_cast_fp16, y = layers_3_qn_w1_to_fp16)[name = string("op_4127_cast_fp16")];
            tensor<int32, [4]> var_4132 = const()[name = string("op_4132"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_1_begin_0 = const()[name = string("r_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_1_end_0 = const()[name = string("r_1_end_0"), val = tensor<int32, [4]>([1, 8, 96, 16])];
            tensor<bool, [4]> r_1_end_mask_0 = const()[name = string("r_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 96, 64]> x_33_cast_fp16 = transpose(perm = var_4132, x = var_4127_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [1, 8, 96, 16]> r_1_cast_fp16 = slice_by_index(begin = r_1_begin_0, end = r_1_end_0, end_mask = r_1_end_mask_0, x = x_33_cast_fp16)[name = string("r_1_cast_fp16")];
            tensor<int32, [4]> p_1_begin_0 = const()[name = string("p_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_1_end_0 = const()[name = string("p_1_end_0"), val = tensor<int32, [4]>([1, 8, 96, 64])];
            tensor<bool, [4]> p_1_end_mask_0 = const()[name = string("p_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 96, 48]> p_1_cast_fp16 = slice_by_index(begin = p_1_begin_0, end = p_1_end_0, end_mask = p_1_end_mask_0, x = x_33_cast_fp16)[name = string("p_1_cast_fp16")];
            tensor<int32, [4]> var_4148_begin_0 = const()[name = string("op_4148_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_4148_end_0 = const()[name = string("op_4148_end_0"), val = tensor<int32, [4]>([1, 8, 96, 16])];
            tensor<bool, [4]> var_4148_end_mask_0 = const()[name = string("op_4148_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 96, 8]> var_4148_cast_fp16 = slice_by_index(begin = var_4148_begin_0, end = var_4148_end_0, end_mask = var_4148_end_mask_0, x = r_1_cast_fp16)[name = string("op_4148_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 96, 8]> var_4149_cast_fp16 = mul(x = var_4148_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_4149_cast_fp16")];
            tensor<int32, [4]> var_4154_begin_0 = const()[name = string("op_4154_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4154_end_0 = const()[name = string("op_4154_end_0"), val = tensor<int32, [4]>([1, 8, 96, 8])];
            tensor<bool, [4]> var_4154_end_mask_0 = const()[name = string("op_4154_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 96, 8]> var_4154_cast_fp16 = slice_by_index(begin = var_4154_begin_0, end = var_4154_end_0, end_mask = var_4154_end_mask_0, x = r_1_cast_fp16)[name = string("op_4154_cast_fp16")];
            int32 var_4156 = const()[name = string("op_4156"), val = int32(-1)];
            bool rh_1_interleave_0 = const()[name = string("rh_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 96, 16]> rh_1_cast_fp16 = concat(axis = var_4156, interleave = rh_1_interleave_0, values = (var_4149_cast_fp16, var_4154_cast_fp16))[name = string("rh_1_cast_fp16")];
            tensor<fp16, [96, 16]> cos_to_fp16 = const()[name = string("cos_to_fp16"), val = tensor<fp16, [96, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119675776)))];
            tensor<fp16, [1, 8, 96, 16]> var_4158_cast_fp16 = mul(x = r_1_cast_fp16, y = cos_to_fp16)[name = string("op_4158_cast_fp16")];
            tensor<fp16, [96, 16]> sin_to_fp16 = const()[name = string("sin_to_fp16"), val = tensor<fp16, [96, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119678912)))];
            tensor<fp16, [1, 8, 96, 16]> var_4159_cast_fp16 = mul(x = rh_1_cast_fp16, y = sin_to_fp16)[name = string("op_4159_cast_fp16")];
            tensor<fp16, [1, 8, 96, 16]> var_4161_cast_fp16 = add(x = var_4158_cast_fp16, y = var_4159_cast_fp16)[name = string("op_4161_cast_fp16")];
            int32 var_4163 = const()[name = string("op_4163"), val = int32(-1)];
            bool q_61_interleave_0 = const()[name = string("q_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 96, 64]> q_61_cast_fp16 = concat(axis = var_4163, interleave = q_61_interleave_0, values = (var_4161_cast_fp16, p_1_cast_fp16))[name = string("q_61_cast_fp16")];
            fp16 var_4169_promoted_to_fp16 = const()[name = string("op_4169_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 2, 64]> var_4173_cast_fp16 = pow(x = x_35_cast_fp16, y = var_4169_promoted_to_fp16)[name = string("op_4173_cast_fp16")];
            tensor<int32, [1]> var_4175_axes_0 = const()[name = string("op_4175_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4175_keep_dims_0 = const()[name = string("op_4175_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 2, 1]> var_4175_cast_fp16 = reduce_mean(axes = var_4175_axes_0, keep_dims = var_4175_keep_dims_0, x = var_4173_cast_fp16)[name = string("op_4175_cast_fp16")];
            fp16 var_4176_to_fp16 = const()[name = string("op_4176_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 96, 2, 1]> var_4177_cast_fp16 = add(x = var_4175_cast_fp16, y = var_4176_to_fp16)[name = string("op_4177_cast_fp16")];
            fp32 var_4178_epsilon_0 = const()[name = string("op_4178_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 2, 1]> var_4178_cast_fp16 = rsqrt(epsilon = var_4178_epsilon_0, x = var_4177_cast_fp16)[name = string("op_4178_cast_fp16")];
            tensor<fp16, [1, 96, 2, 64]> var_4179_cast_fp16 = mul(x = x_35_cast_fp16, y = var_4178_cast_fp16)[name = string("op_4179_cast_fp16")];
            tensor<fp16, [64]> layers_3_kn_w1_to_fp16 = const()[name = string("layers_3_kn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59588800)))];
            tensor<fp16, [1, 96, 2, 64]> var_4180_cast_fp16 = mul(x = var_4179_cast_fp16, y = layers_3_kn_w1_to_fp16)[name = string("op_4180_cast_fp16")];
            tensor<int32, [4]> var_4185 = const()[name = string("op_4185"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_3_begin_0 = const()[name = string("r_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_3_end_0 = const()[name = string("r_3_end_0"), val = tensor<int32, [4]>([1, 2, 96, 16])];
            tensor<bool, [4]> r_3_end_mask_0 = const()[name = string("r_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 96, 64]> x_39_cast_fp16 = transpose(perm = var_4185, x = var_4180_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [1, 2, 96, 16]> r_3_cast_fp16 = slice_by_index(begin = r_3_begin_0, end = r_3_end_0, end_mask = r_3_end_mask_0, x = x_39_cast_fp16)[name = string("r_3_cast_fp16")];
            tensor<int32, [4]> p_3_begin_0 = const()[name = string("p_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_3_end_0 = const()[name = string("p_3_end_0"), val = tensor<int32, [4]>([1, 2, 96, 64])];
            tensor<bool, [4]> p_3_end_mask_0 = const()[name = string("p_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 96, 48]> p_3_cast_fp16 = slice_by_index(begin = p_3_begin_0, end = p_3_end_0, end_mask = p_3_end_mask_0, x = x_39_cast_fp16)[name = string("p_3_cast_fp16")];
            tensor<int32, [4]> var_4201_begin_0 = const()[name = string("op_4201_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_4201_end_0 = const()[name = string("op_4201_end_0"), val = tensor<int32, [4]>([1, 2, 96, 16])];
            tensor<bool, [4]> var_4201_end_mask_0 = const()[name = string("op_4201_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 96, 8]> var_4201_cast_fp16 = slice_by_index(begin = var_4201_begin_0, end = var_4201_end_0, end_mask = var_4201_end_mask_0, x = r_3_cast_fp16)[name = string("op_4201_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 96, 8]> var_4202_cast_fp16 = mul(x = var_4201_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_4202_cast_fp16")];
            tensor<int32, [4]> var_4207_begin_0 = const()[name = string("op_4207_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4207_end_0 = const()[name = string("op_4207_end_0"), val = tensor<int32, [4]>([1, 2, 96, 8])];
            tensor<bool, [4]> var_4207_end_mask_0 = const()[name = string("op_4207_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 96, 8]> var_4207_cast_fp16 = slice_by_index(begin = var_4207_begin_0, end = var_4207_end_0, end_mask = var_4207_end_mask_0, x = r_3_cast_fp16)[name = string("op_4207_cast_fp16")];
            int32 var_4209 = const()[name = string("op_4209"), val = int32(-1)];
            bool rh_3_interleave_0 = const()[name = string("rh_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 96, 16]> rh_3_cast_fp16 = concat(axis = var_4209, interleave = rh_3_interleave_0, values = (var_4202_cast_fp16, var_4207_cast_fp16))[name = string("rh_3_cast_fp16")];
            tensor<fp16, [1, 2, 96, 16]> var_4211_cast_fp16 = mul(x = r_3_cast_fp16, y = cos_to_fp16)[name = string("op_4211_cast_fp16")];
            tensor<fp16, [1, 2, 96, 16]> var_4212_cast_fp16 = mul(x = rh_3_cast_fp16, y = sin_to_fp16)[name = string("op_4212_cast_fp16")];
            tensor<fp16, [1, 2, 96, 16]> var_4214_cast_fp16 = add(x = var_4211_cast_fp16, y = var_4212_cast_fp16)[name = string("op_4214_cast_fp16")];
            int32 var_4216 = const()[name = string("op_4216"), val = int32(-1)];
            bool k_61_interleave_0 = const()[name = string("k_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 96, 64]> k_61_cast_fp16 = concat(axis = var_4216, interleave = k_61_interleave_0, values = (var_4214_cast_fp16, p_3_cast_fp16))[name = string("k_61_cast_fp16")];
            tensor<int32, [1]> var_4229_axes_0 = const()[name = string("op_4229_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 96, 64]> var_4229_cast_fp16 = expand_dims(axes = var_4229_axes_0, x = k_61_cast_fp16)[name = string("op_4229_cast_fp16")];
            tensor<int32, [5]> var_4237_reps_0 = const()[name = string("op_4237_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 2, 4, 96, 64]> var_4237_cast_fp16 = tile(reps = var_4237_reps_0, x = var_4229_cast_fp16)[name = string("op_4237_cast_fp16")];
            tensor<int32, [4]> var_4242 = const()[name = string("op_4242"), val = tensor<int32, [4]>([1, 8, 96, 64])];
            tensor<fp16, [1, 8, 96, 64]> k_63_cast_fp16 = reshape(shape = var_4242, x = var_4237_cast_fp16)[name = string("k_63_cast_fp16")];
            tensor<int32, [1]> var_4255_axes_0 = const()[name = string("op_4255_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 96, 64]> v_7_cast_fp16 = transpose(perm = var_4110, x = var_4105_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [1, 2, 1, 96, 64]> var_4255_cast_fp16 = expand_dims(axes = var_4255_axes_0, x = v_7_cast_fp16)[name = string("op_4255_cast_fp16")];
            tensor<int32, [5]> var_4263_reps_0 = const()[name = string("op_4263_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 2, 4, 96, 64]> var_4263_cast_fp16 = tile(reps = var_4263_reps_0, x = var_4255_cast_fp16)[name = string("op_4263_cast_fp16")];
            tensor<int32, [4]> var_4268 = const()[name = string("op_4268"), val = tensor<int32, [4]>([1, 8, 96, 64])];
            tensor<fp16, [1, 8, 96, 64]> v_9_cast_fp16 = reshape(shape = var_4268, x = var_4263_cast_fp16)[name = string("v_9_cast_fp16")];
            bool var_4273_transpose_x_1 = const()[name = string("op_4273_transpose_x_1"), val = bool(false)];
            bool var_4273_transpose_y_1 = const()[name = string("op_4273_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 96, 96]> var_4273_cast_fp16 = matmul(transpose_x = var_4273_transpose_x_1, transpose_y = var_4273_transpose_y_1, x = q_61_cast_fp16, y = k_63_cast_fp16)[name = string("op_4273_cast_fp16")];
            fp16 var_4274_to_fp16 = const()[name = string("op_4274_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 8, 96, 96]> var_4275_cast_fp16 = mul(x = var_4273_cast_fp16, y = var_4274_to_fp16)[name = string("op_4275_cast_fp16")];
            tensor<fp16, [96, 96]> causal_add_to_fp16 = const()[name = string("causal_add_to_fp16"), val = tensor<fp16, [96, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119682048)))];
            tensor<fp16, [1, 8, 96, 96]> s_1_cast_fp16 = add(x = var_4275_cast_fp16, y = causal_add_to_fp16)[name = string("s_1_cast_fp16")];
            int32 var_4278 = const()[name = string("op_4278"), val = int32(-1)];
            tensor<fp16, [1, 8, 96, 96]> var_4280_cast_fp16 = softmax(axis = var_4278, x = s_1_cast_fp16)[name = string("op_4280_cast_fp16")];
            bool o_19_transpose_x_0 = const()[name = string("o_19_transpose_x_0"), val = bool(false)];
            bool o_19_transpose_y_0 = const()[name = string("o_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 96, 64]> o_19_cast_fp16 = matmul(transpose_x = o_19_transpose_x_0, transpose_y = o_19_transpose_y_0, x = var_4280_cast_fp16, y = v_9_cast_fp16)[name = string("o_19_cast_fp16")];
            tensor<int32, [4]> var_4286 = const()[name = string("op_4286"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4291 = const()[name = string("op_4291"), val = tensor<int32, [3]>([1, 96, 512])];
            tensor<fp16, [1, 96, 8, 64]> var_4287_cast_fp16 = transpose(perm = var_4286, x = o_19_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [1, 96, 512]> var_4292_cast_fp16 = reshape(shape = var_4291, x = var_4287_cast_fp16)[name = string("op_4292_cast_fp16")];
            tensor<fp16, [1, 96, 512]> var_4293_cast_fp16 = sigmoid(x = gate_1_cast_fp16)[name = string("op_4293_cast_fp16")];
            tensor<fp16, [1, 96, 512]> input_51_cast_fp16 = mul(x = var_4292_cast_fp16, y = var_4293_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [640, 512]> layers_3_o_weight_to_fp16 = const()[name = string("layers_3_o_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59589568)))];
            tensor<fp16, [1, 96, 640]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_o_weight_to_fp16, x = input_51_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 96, 640]> x_41_cast_fp16 = add(x = x_25_cast_fp16, y = linear_13_cast_fp16)[name = string("x_41_cast_fp16")];
            fp16 var_4306_to_fp16 = const()[name = string("op_4306_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 96, 640]> x_43_cast_fp16 = mul(x = x_41_cast_fp16, y = var_4306_to_fp16)[name = string("x_43_cast_fp16")];
            fp16 var_4304_promoted_to_fp16 = const()[name = string("op_4304_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> var_4308_cast_fp16 = pow(x = x_43_cast_fp16, y = var_4304_promoted_to_fp16)[name = string("op_4308_cast_fp16")];
            tensor<int32, [1]> var_4310_axes_0 = const()[name = string("op_4310_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4310_keep_dims_0 = const()[name = string("op_4310_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 1]> var_4310_cast_fp16 = reduce_mean(axes = var_4310_axes_0, keep_dims = var_4310_keep_dims_0, x = var_4308_cast_fp16)[name = string("op_4310_cast_fp16")];
            fp16 var_4311_to_fp16 = const()[name = string("op_4311_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 96, 1]> var_4312_cast_fp16 = add(x = var_4310_cast_fp16, y = var_4311_to_fp16)[name = string("op_4312_cast_fp16")];
            fp32 var_4313_epsilon_0 = const()[name = string("op_4313_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 1]> var_4313_cast_fp16 = rsqrt(epsilon = var_4313_epsilon_0, x = var_4312_cast_fp16)[name = string("op_4313_cast_fp16")];
            tensor<fp16, [1, 96, 640]> var_4314_cast_fp16 = mul(x = x_43_cast_fp16, y = var_4313_cast_fp16)[name = string("op_4314_cast_fp16")];
            tensor<fp16, [640]> layers_3_n2_w1_to_fp16 = const()[name = string("layers_3_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60244992)))];
            tensor<fp16, [1, 96, 640]> input_53_cast_fp16 = mul(x = var_4314_cast_fp16, y = layers_3_n2_w1_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_3_gate_up_weight_to_fp16 = const()[name = string("layers_3_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60246336)))];
            tensor<fp16, [1, 96, 5120]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_3_gate_up_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [3]> input_55_begin_0 = const()[name = string("input_55_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_55_end_0 = const()[name = string("input_55_end_0"), val = tensor<int32, [3]>([1, 96, 2560])];
            tensor<bool, [3]> input_55_end_mask_0 = const()[name = string("input_55_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 2560]> input_55_cast_fp16 = slice_by_index(begin = input_55_begin_0, end = input_55_end_0, end_mask = input_55_end_mask_0, x = linear_14_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [1, 96, 2560]> var_4324_cast_fp16 = silu(x = input_55_cast_fp16)[name = string("op_4324_cast_fp16")];
            tensor<int32, [3]> var_4329_begin_0 = const()[name = string("op_4329_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_4329_end_0 = const()[name = string("op_4329_end_0"), val = tensor<int32, [3]>([1, 96, 5120])];
            tensor<bool, [3]> var_4329_end_mask_0 = const()[name = string("op_4329_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 2560]> var_4329_cast_fp16 = slice_by_index(begin = var_4329_begin_0, end = var_4329_end_0, end_mask = var_4329_end_mask_0, x = linear_14_cast_fp16)[name = string("op_4329_cast_fp16")];
            tensor<fp16, [1, 96, 2560]> input_57_cast_fp16 = mul(x = var_4324_cast_fp16, y = var_4329_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_3_down_weight_to_fp16 = const()[name = string("layers_3_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66800000)))];
            tensor<fp16, [1, 96, 640]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_down_weight_to_fp16, x = input_57_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, 96, 640]> x_45_cast_fp16 = add(x = x_41_cast_fp16, y = linear_15_cast_fp16)[name = string("x_45_cast_fp16")];
            fp16 var_4342_to_fp16 = const()[name = string("op_4342_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 96, 640]> x_47_cast_fp16 = mul(x = x_45_cast_fp16, y = var_4342_to_fp16)[name = string("x_47_cast_fp16")];
            fp16 var_4340_promoted_to_fp16 = const()[name = string("op_4340_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> var_4344_cast_fp16 = pow(x = x_47_cast_fp16, y = var_4340_promoted_to_fp16)[name = string("op_4344_cast_fp16")];
            tensor<int32, [1]> var_4346_axes_0 = const()[name = string("op_4346_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4346_keep_dims_0 = const()[name = string("op_4346_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 1]> var_4346_cast_fp16 = reduce_mean(axes = var_4346_axes_0, keep_dims = var_4346_keep_dims_0, x = var_4344_cast_fp16)[name = string("op_4346_cast_fp16")];
            fp16 var_4347_to_fp16 = const()[name = string("op_4347_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 96, 1]> var_4348_cast_fp16 = add(x = var_4346_cast_fp16, y = var_4347_to_fp16)[name = string("op_4348_cast_fp16")];
            fp32 var_4349_epsilon_0 = const()[name = string("op_4349_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 1]> var_4349_cast_fp16 = rsqrt(epsilon = var_4349_epsilon_0, x = var_4348_cast_fp16)[name = string("op_4349_cast_fp16")];
            tensor<fp16, [1, 96, 640]> var_4350_cast_fp16 = mul(x = x_47_cast_fp16, y = var_4349_cast_fp16)[name = string("op_4350_cast_fp16")];
            tensor<fp16, [640]> layers_4_n1_w1_to_fp16 = const()[name = string("layers_4_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70076864)))];
            tensor<fp16, [1, 96, 640]> input_59_cast_fp16 = mul(x = var_4350_cast_fp16, y = layers_4_n1_w1_to_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_4_qkvzba_weight_to_fp16 = const()[name = string("layers_4_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70078208)))];
            tensor<fp16, [1, 96, 1544]> linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_qkvzba_weight_to_fp16, x = input_59_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [3]> qkv_7_begin_0 = const()[name = string("qkv_7_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_7_end_0 = const()[name = string("qkv_7_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> qkv_7_end_mask_0 = const()[name = string("qkv_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 1024]> qkv_7_cast_fp16 = slice_by_index(begin = qkv_7_begin_0, end = qkv_7_end_0, end_mask = qkv_7_end_mask_0, x = linear_16_cast_fp16)[name = string("qkv_7_cast_fp16")];
            tensor<int32, [3]> z_7_begin_0 = const()[name = string("z_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_7_end_0 = const()[name = string("z_7_end_0"), val = tensor<int32, [3]>([1, 96, 1536])];
            tensor<bool, [3]> z_7_end_mask_0 = const()[name = string("z_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 512]> z_7_cast_fp16 = slice_by_index(begin = z_7_begin_0, end = z_7_end_0, end_mask = z_7_end_mask_0, x = linear_16_cast_fp16)[name = string("z_7_cast_fp16")];
            tensor<int32, [3]> b_7_begin_0 = const()[name = string("b_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_7_end_0 = const()[name = string("b_7_end_0"), val = tensor<int32, [3]>([1, 96, 1540])];
            tensor<bool, [3]> b_7_end_mask_0 = const()[name = string("b_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 4]> b_7_cast_fp16 = slice_by_index(begin = b_7_begin_0, end = b_7_end_0, end_mask = b_7_end_mask_0, x = linear_16_cast_fp16)[name = string("b_7_cast_fp16")];
            tensor<int32, [3]> a_7_begin_0 = const()[name = string("a_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_7_end_0 = const()[name = string("a_7_end_0"), val = tensor<int32, [3]>([1, 96, 1])];
            tensor<bool, [3]> a_7_end_mask_0 = const()[name = string("a_7_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 4]> a_7_cast_fp16 = slice_by_index(begin = a_7_begin_0, end = a_7_end_0, end_mask = a_7_end_mask_0, x = linear_16_cast_fp16)[name = string("a_7_cast_fp16")];
            int32 var_4389 = const()[name = string("op_4389"), val = int32(1)];
            bool xp_7_interleave_0 = const()[name = string("xp_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 99, 1024]> xp_7_cast_fp16 = concat(axis = var_4389, interleave = xp_7_interleave_0, values = (zpad_to_fp16, qkv_7_cast_fp16))[name = string("xp_7_cast_fp16")];
            tensor<int32, [3]> var_4400_begin_0 = const()[name = string("op_4400_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_4400_end_0 = const()[name = string("op_4400_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_4400_end_mask_0 = const()[name = string("op_4400_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 1024]> var_4400_cast_fp16 = slice_by_index(begin = var_4400_begin_0, end = var_4400_end_0, end_mask = var_4400_end_mask_0, x = xp_7_cast_fp16)[name = string("op_4400_cast_fp16")];
            tensor<fp16, [1024]> var_4408_to_fp16 = const()[name = string("op_4408_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72054592)))];
            tensor<fp16, [1, 96, 1024]> var_4409_cast_fp16 = mul(x = var_4400_cast_fp16, y = var_4408_to_fp16)[name = string("op_4409_cast_fp16")];
            tensor<int32, [3]> var_4419_begin_0 = const()[name = string("op_4419_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_4419_end_0 = const()[name = string("op_4419_end_0"), val = tensor<int32, [3]>([1, 98, 1024])];
            tensor<bool, [3]> var_4419_end_mask_0 = const()[name = string("op_4419_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_4419_cast_fp16 = slice_by_index(begin = var_4419_begin_0, end = var_4419_end_0, end_mask = var_4419_end_mask_0, x = xp_7_cast_fp16)[name = string("op_4419_cast_fp16")];
            tensor<fp16, [1024]> var_4427_to_fp16 = const()[name = string("op_4427_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72056704)))];
            tensor<fp16, [1, 96, 1024]> var_4428_cast_fp16 = mul(x = var_4419_cast_fp16, y = var_4427_to_fp16)[name = string("op_4428_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> var_4430_cast_fp16 = add(x = var_4409_cast_fp16, y = var_4428_cast_fp16)[name = string("op_4430_cast_fp16")];
            tensor<int32, [3]> var_4440_begin_0 = const()[name = string("op_4440_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_4440_end_0 = const()[name = string("op_4440_end_0"), val = tensor<int32, [3]>([1, 97, 1024])];
            tensor<bool, [3]> var_4440_end_mask_0 = const()[name = string("op_4440_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_4440_cast_fp16 = slice_by_index(begin = var_4440_begin_0, end = var_4440_end_0, end_mask = var_4440_end_mask_0, x = xp_7_cast_fp16)[name = string("op_4440_cast_fp16")];
            tensor<fp16, [1024]> var_4448_to_fp16 = const()[name = string("op_4448_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72058816)))];
            tensor<fp16, [1, 96, 1024]> var_4449_cast_fp16 = mul(x = var_4440_cast_fp16, y = var_4448_to_fp16)[name = string("op_4449_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> var_4451_cast_fp16 = add(x = var_4430_cast_fp16, y = var_4449_cast_fp16)[name = string("op_4451_cast_fp16")];
            tensor<int32, [3]> var_4461_begin_0 = const()[name = string("op_4461_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_4461_end_0 = const()[name = string("op_4461_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> var_4461_end_mask_0 = const()[name = string("op_4461_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_4461_cast_fp16 = slice_by_index(begin = var_4461_begin_0, end = var_4461_end_0, end_mask = var_4461_end_mask_0, x = xp_7_cast_fp16)[name = string("op_4461_cast_fp16")];
            tensor<fp16, [1024]> var_4469_to_fp16 = const()[name = string("op_4469_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72060928)))];
            tensor<fp16, [1, 96, 1024]> var_4470_cast_fp16 = mul(x = var_4461_cast_fp16, y = var_4469_to_fp16)[name = string("op_4470_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> input_61_cast_fp16 = add(x = var_4451_cast_fp16, y = var_4470_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> c_7_cast_fp16 = silu(x = input_61_cast_fp16)[name = string("c_7_cast_fp16")];
            tensor<int32, [3]> var_4478_begin_0 = const()[name = string("op_4478_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_4478_end_0 = const()[name = string("op_4478_end_0"), val = tensor<int32, [3]>([1, 96, 256])];
            tensor<bool, [3]> var_4478_end_mask_0 = const()[name = string("op_4478_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 256]> var_4478_cast_fp16 = slice_by_index(begin = var_4478_begin_0, end = var_4478_end_0, end_mask = var_4478_end_mask_0, x = c_7_cast_fp16)[name = string("op_4478_cast_fp16")];
            tensor<int32, [4]> var_4483 = const()[name = string("op_4483"), val = tensor<int32, [4]>([1, 96, 2, 128])];
            tensor<fp16, [1, 96, 2, 128]> q_63_cast_fp16 = reshape(shape = var_4483, x = var_4478_cast_fp16)[name = string("q_63_cast_fp16")];
            tensor<int32, [3]> var_4489_begin_0 = const()[name = string("op_4489_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_4489_end_0 = const()[name = string("op_4489_end_0"), val = tensor<int32, [3]>([1, 96, 512])];
            tensor<bool, [3]> var_4489_end_mask_0 = const()[name = string("op_4489_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 256]> var_4489_cast_fp16 = slice_by_index(begin = var_4489_begin_0, end = var_4489_end_0, end_mask = var_4489_end_mask_0, x = c_7_cast_fp16)[name = string("op_4489_cast_fp16")];
            tensor<int32, [4]> var_4494 = const()[name = string("op_4494"), val = tensor<int32, [4]>([1, 96, 2, 128])];
            tensor<fp16, [1, 96, 2, 128]> k_65_cast_fp16 = reshape(shape = var_4494, x = var_4489_cast_fp16)[name = string("k_65_cast_fp16")];
            tensor<int32, [3]> var_4500_begin_0 = const()[name = string("op_4500_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_4500_end_0 = const()[name = string("op_4500_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> var_4500_end_mask_0 = const()[name = string("op_4500_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 512]> var_4500_cast_fp16 = slice_by_index(begin = var_4500_begin_0, end = var_4500_end_0, end_mask = var_4500_end_mask_0, x = c_7_cast_fp16)[name = string("op_4500_cast_fp16")];
            tensor<int32, [4]> var_4505 = const()[name = string("op_4505"), val = tensor<int32, [4]>([1, 96, 4, 128])];
            tensor<fp16, [1, 96, 4, 128]> var_4506_cast_fp16 = reshape(shape = var_4505, x = var_4500_cast_fp16)[name = string("op_4506_cast_fp16")];
            tensor<int32, [4]> var_4511 = const()[name = string("op_4511"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4513_to_fp16 = const()[name = string("op_4513_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 96, 2, 128]> q_65_cast_fp16 = mul(x = q_63_cast_fp16, y = var_4513_to_fp16)[name = string("q_65_cast_fp16")];
            fp16 var_4515_to_fp16 = const()[name = string("op_4515_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 96, 2, 128]> k_67_cast_fp16 = mul(x = k_65_cast_fp16, y = var_4515_to_fp16)[name = string("k_67_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_4517_cast_fp16 = mul(x = q_65_cast_fp16, y = q_65_cast_fp16)[name = string("op_4517_cast_fp16")];
            tensor<int32, [1]> var_4522_axes_0 = const()[name = string("op_4522_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4522_keep_dims_0 = const()[name = string("op_4522_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 2, 1]> var_4522_cast_fp16 = reduce_mean(axes = var_4522_axes_0, keep_dims = var_4522_keep_dims_0, x = var_4517_cast_fp16)[name = string("op_4522_cast_fp16")];
            fp16 var_4523_promoted_to_fp16 = const()[name = string("op_4523_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 96, 2, 1]> var_4524_cast_fp16 = mul(x = var_4522_cast_fp16, y = var_4523_promoted_to_fp16)[name = string("op_4524_cast_fp16")];
            fp16 var_4526_to_fp16 = const()[name = string("op_4526_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 96, 2, 1]> var_4527_cast_fp16 = add(x = var_4524_cast_fp16, y = var_4526_to_fp16)[name = string("op_4527_cast_fp16")];
            fp32 var_4528_epsilon_0 = const()[name = string("op_4528_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 2, 1]> var_4528_cast_fp16 = rsqrt(epsilon = var_4528_epsilon_0, x = var_4527_cast_fp16)[name = string("op_4528_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_4529_cast_fp16 = mul(x = q_65_cast_fp16, y = var_4528_cast_fp16)[name = string("op_4529_cast_fp16")];
            fp16 var_4530_to_fp16 = const()[name = string("op_4530_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 96, 2, 128]> q_67_cast_fp16 = mul(x = var_4529_cast_fp16, y = var_4530_to_fp16)[name = string("q_67_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_4532_cast_fp16 = mul(x = k_67_cast_fp16, y = k_67_cast_fp16)[name = string("op_4532_cast_fp16")];
            tensor<int32, [1]> var_4537_axes_0 = const()[name = string("op_4537_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4537_keep_dims_0 = const()[name = string("op_4537_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 2, 1]> var_4537_cast_fp16 = reduce_mean(axes = var_4537_axes_0, keep_dims = var_4537_keep_dims_0, x = var_4532_cast_fp16)[name = string("op_4537_cast_fp16")];
            fp16 var_4538_promoted_to_fp16 = const()[name = string("op_4538_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 96, 2, 1]> var_4539_cast_fp16 = mul(x = var_4537_cast_fp16, y = var_4538_promoted_to_fp16)[name = string("op_4539_cast_fp16")];
            fp16 var_4541_to_fp16 = const()[name = string("op_4541_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 96, 2, 1]> var_4542_cast_fp16 = add(x = var_4539_cast_fp16, y = var_4541_to_fp16)[name = string("op_4542_cast_fp16")];
            fp32 var_4543_epsilon_0 = const()[name = string("op_4543_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 2, 1]> var_4543_cast_fp16 = rsqrt(epsilon = var_4543_epsilon_0, x = var_4542_cast_fp16)[name = string("op_4543_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> k_69_cast_fp16 = mul(x = k_67_cast_fp16, y = var_4543_cast_fp16)[name = string("k_69_cast_fp16")];
            tensor<int32, [4]> var_4549 = const()[name = string("op_4549"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_4562_axes_0 = const()[name = string("op_4562_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 96, 128]> var_4550_cast_fp16 = transpose(perm = var_4549, x = q_67_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [1, 2, 1, 96, 128]> var_4562_cast_fp16 = expand_dims(axes = var_4562_axes_0, x = var_4550_cast_fp16)[name = string("op_4562_cast_fp16")];
            tensor<int32, [5]> var_4570_reps_0 = const()[name = string("op_4570_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 96, 128]> var_4570_cast_fp16 = tile(reps = var_4570_reps_0, x = var_4562_cast_fp16)[name = string("op_4570_cast_fp16")];
            tensor<int32, [4]> var_4575 = const()[name = string("op_4575"), val = tensor<int32, [4]>([1, 4, 96, 128])];
            tensor<fp16, [1, 4, 96, 128]> q_69_cast_fp16 = reshape(shape = var_4575, x = var_4570_cast_fp16)[name = string("q_69_cast_fp16")];
            tensor<int32, [4]> var_4581 = const()[name = string("op_4581"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_4594_axes_0 = const()[name = string("op_4594_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 96, 128]> var_4582_cast_fp16 = transpose(perm = var_4581, x = k_69_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [1, 2, 1, 96, 128]> var_4594_cast_fp16 = expand_dims(axes = var_4594_axes_0, x = var_4582_cast_fp16)[name = string("op_4594_cast_fp16")];
            tensor<int32, [5]> var_4602_reps_0 = const()[name = string("op_4602_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 96, 128]> var_4602_cast_fp16 = tile(reps = var_4602_reps_0, x = var_4594_cast_fp16)[name = string("op_4602_cast_fp16")];
            tensor<int32, [4]> var_4607 = const()[name = string("op_4607"), val = tensor<int32, [4]>([1, 4, 96, 128])];
            tensor<fp16, [1, 4, 96, 128]> k_71_cast_fp16 = reshape(shape = var_4607, x = var_4602_cast_fp16)[name = string("k_71_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_4609_cast_fp16 = sigmoid(x = b_7_cast_fp16)[name = string("op_4609_cast_fp16")];
            tensor<int32, [3]> var_4613 = const()[name = string("op_4613"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_4_dt_bias_to_fp16 = const()[name = string("layers_4_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 96, 4]> input_63_cast_fp16 = add(x = a_7_cast_fp16, y = layers_4_dt_bias_to_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_4617_cast_fp16 = relu(x = input_63_cast_fp16)[name = string("op_4617_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_4618_cast_fp16 = abs(x = input_63_cast_fp16)[name = string("op_4618_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 96, 4]> var_4619_cast_fp16 = mul(x = var_4618_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_4619_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_4620_cast_fp16 = exp(x = var_4619_cast_fp16)[name = string("op_4620_cast_fp16")];
            fp16 var_4622_to_fp16 = const()[name = string("op_4622_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 96, 4]> var_4623_cast_fp16 = add(x = var_4620_cast_fp16, y = var_4622_to_fp16)[name = string("op_4623_cast_fp16")];
            fp32 var_4624_epsilon_0 = const()[name = string("op_4624_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 96, 4]> var_4624_cast_fp16 = log(epsilon = var_4624_epsilon_0, x = var_4623_cast_fp16)[name = string("op_4624_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_4626_cast_fp16 = add(x = var_4617_cast_fp16, y = var_4624_cast_fp16)[name = string("op_4626_cast_fp16")];
            tensor<fp16, [4]> layers_4_negA_to_fp16 = const()[name = string("layers_4_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.348p+3, -0x1.734p+2, -0x1.5a8p-1, -0x1.264p+3])];
            tensor<fp16, [1, 96, 4]> var_4627_cast_fp16 = mul(x = layers_4_negA_to_fp16, y = var_4626_cast_fp16)[name = string("op_4627_cast_fp16")];
            tensor<int32, [3]> var_4631 = const()[name = string("op_4631"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_4634_axes_0 = const()[name = string("op_4634_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 96]> beta_7_cast_fp16 = transpose(perm = var_4613, x = var_4609_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 4, 96, 1]> var_4634_cast_fp16 = expand_dims(axes = var_4634_axes_0, x = beta_7_cast_fp16)[name = string("op_4634_cast_fp16")];
            fp16 var_4635_to_fp16 = const()[name = string("op_4635_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 96, 1]> var_4636_cast_fp16 = mul(x = var_4634_cast_fp16, y = var_4635_to_fp16)[name = string("op_4636_cast_fp16")];
            tensor<fp16, [1, 4, 96, 128]> v_11_cast_fp16 = transpose(perm = var_4511, x = var_4506_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [1, 4, 96, 128]> vb_43_cast_fp16 = mul(x = v_11_cast_fp16, y = var_4636_cast_fp16)[name = string("vb_43_cast_fp16")];
            tensor<int32, [4]> q_71_begin_0 = const()[name = string("q_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_71_end_0 = const()[name = string("q_71_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_71_end_mask_0 = const()[name = string("q_71_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_71_cast_fp16 = slice_by_index(begin = q_71_begin_0, end = q_71_end_0, end_mask = q_71_end_mask_0, x = q_69_cast_fp16)[name = string("q_71_cast_fp16")];
            tensor<int32, [4]> k_73_begin_0 = const()[name = string("k_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_73_end_0 = const()[name = string("k_73_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_73_end_mask_0 = const()[name = string("k_73_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, x = k_71_cast_fp16)[name = string("k_73_cast_fp16")];
            tensor<fp16, [1, 4, 96, 128]> var_4670_cast_fp16 = mul(x = k_71_cast_fp16, y = var_4634_cast_fp16)[name = string("op_4670_cast_fp16")];
            tensor<int32, [4]> kb_37_begin_0 = const()[name = string("kb_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_37_end_0 = const()[name = string("kb_37_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_37_end_mask_0 = const()[name = string("kb_37_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_37_cast_fp16 = slice_by_index(begin = kb_37_begin_0, end = kb_37_end_0, end_mask = kb_37_end_mask_0, x = var_4670_cast_fp16)[name = string("kb_37_cast_fp16")];
            tensor<int32, [4]> vb_45_begin_0 = const()[name = string("vb_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_45_end_0 = const()[name = string("vb_45_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_45_end_mask_0 = const()[name = string("vb_45_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_45_cast_fp16 = slice_by_index(begin = vb_45_begin_0, end = vb_45_end_0, end_mask = vb_45_end_mask_0, x = vb_43_cast_fp16)[name = string("vb_45_cast_fp16")];
            tensor<int32, [3]> gdec_45_begin_0 = const()[name = string("gdec_45_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_45_end_0 = const()[name = string("gdec_45_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_45_end_mask_0 = const()[name = string("gdec_45_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 96]> gdec_43_cast_fp16 = transpose(perm = var_4631, x = var_4627_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 4, 16]> gdec_45_cast_fp16 = slice_by_index(begin = gdec_45_begin_0, end = gdec_45_end_0, end_mask = gdec_45_end_mask_0, x = gdec_43_cast_fp16)[name = string("gdec_45_cast_fp16")];
            tensor<int32, [2]> var_4718 = const()[name = string("op_4718"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_4719_cast_fp16 = reshape(shape = var_4718, x = gdec_45_cast_fp16)[name = string("op_4719_cast_fp16")];
            tensor<fp16, [4, 256]> var_4720_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_4719_cast_fp16)[name = string("op_4720_cast_fp16")];
            tensor<int32, [4]> var_4725 = const()[name = string("op_4725"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_37_cast_fp16 = reshape(shape = var_4725, x = var_4720_cast_fp16)[name = string("P_37_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4727_cast_fp16 = exp(x = P_37_cast_fp16)[name = string("op_4727_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_37_cast_fp16 = mul(x = var_4727_cast_fp16, y = tril_incl_to_fp16)[name = string("D_37_cast_fp16")];
            bool kk_37_transpose_x_1 = const()[name = string("kk_37_transpose_x_1"), val = bool(false)];
            bool kk_37_transpose_y_1 = const()[name = string("kk_37_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_37_cast_fp16 = matmul(transpose_x = kk_37_transpose_x_1, transpose_y = kk_37_transpose_y_1, x = kb_37_cast_fp16, y = k_73_cast_fp16)[name = string("kk_37_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_4733_cast_fp16 = mul(x = kk_37_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_4733_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4734_cast_fp16 = mul(x = var_4733_cast_fp16, y = D_37_cast_fp16)[name = string("op_4734_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_37_cast_fp16 = mul(x = var_4734_cast_fp16, y = cum_tail_to_fp16)[name = string("N_37_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4736_cast_fp16 = mul(x = D_37_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_4736_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_37_cast_fp16 = mul(x = kk_37_cast_fp16, y = var_4736_cast_fp16)[name = string("Na_37_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_145_cast_fp16 = add(x = eye_to_fp16, y = N_37_cast_fp16)[name = string("Minv_145_cast_fp16")];
            bool Np_109_transpose_x_0 = const()[name = string("Np_109_transpose_x_0"), val = bool(false)];
            bool Np_109_transpose_y_0 = const()[name = string("Np_109_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_109_cast_fp16 = matmul(transpose_x = Np_109_transpose_x_0, transpose_y = Np_109_transpose_y_0, x = N_37_cast_fp16, y = Na_37_cast_fp16)[name = string("Np_109_cast_fp16")];
            bool Npa_73_transpose_x_0 = const()[name = string("Npa_73_transpose_x_0"), val = bool(false)];
            bool Npa_73_transpose_y_0 = const()[name = string("Npa_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_73_cast_fp16 = matmul(transpose_x = Npa_73_transpose_x_0, transpose_y = Npa_73_transpose_y_0, x = Na_37_cast_fp16, y = N_37_cast_fp16)[name = string("Npa_73_cast_fp16")];
            bool var_4742_transpose_x_0 = const()[name = string("op_4742_transpose_x_0"), val = bool(false)];
            bool var_4742_transpose_y_0 = const()[name = string("op_4742_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4742_cast_fp16 = matmul(transpose_x = var_4742_transpose_x_0, transpose_y = var_4742_transpose_y_0, x = Minv_145_cast_fp16, y = Np_109_cast_fp16)[name = string("op_4742_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_147_cast_fp16 = add(x = Minv_145_cast_fp16, y = var_4742_cast_fp16)[name = string("Minv_147_cast_fp16")];
            bool Np_111_transpose_x_0 = const()[name = string("Np_111_transpose_x_0"), val = bool(false)];
            bool Np_111_transpose_y_0 = const()[name = string("Np_111_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_111_cast_fp16 = matmul(transpose_x = Np_111_transpose_x_0, transpose_y = Np_111_transpose_y_0, x = Np_109_cast_fp16, y = Npa_73_cast_fp16)[name = string("Np_111_cast_fp16")];
            bool Npa_75_transpose_x_0 = const()[name = string("Npa_75_transpose_x_0"), val = bool(false)];
            bool Npa_75_transpose_y_0 = const()[name = string("Npa_75_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_75_cast_fp16 = matmul(transpose_x = Npa_75_transpose_x_0, transpose_y = Npa_75_transpose_y_0, x = Npa_73_cast_fp16, y = Np_109_cast_fp16)[name = string("Npa_75_cast_fp16")];
            bool var_4747_transpose_x_0 = const()[name = string("op_4747_transpose_x_0"), val = bool(false)];
            bool var_4747_transpose_y_0 = const()[name = string("op_4747_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4747_cast_fp16 = matmul(transpose_x = var_4747_transpose_x_0, transpose_y = var_4747_transpose_y_0, x = Minv_147_cast_fp16, y = Np_111_cast_fp16)[name = string("op_4747_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_149_cast_fp16 = add(x = Minv_147_cast_fp16, y = var_4747_cast_fp16)[name = string("Minv_149_cast_fp16")];
            bool Np_113_transpose_x_0 = const()[name = string("Np_113_transpose_x_0"), val = bool(false)];
            bool Np_113_transpose_y_0 = const()[name = string("Np_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_113_cast_fp16 = matmul(transpose_x = Np_113_transpose_x_0, transpose_y = Np_113_transpose_y_0, x = Np_111_cast_fp16, y = Npa_75_cast_fp16)[name = string("Np_113_cast_fp16")];
            bool var_4751_transpose_x_0 = const()[name = string("op_4751_transpose_x_0"), val = bool(false)];
            bool var_4751_transpose_y_0 = const()[name = string("op_4751_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4751_cast_fp16 = matmul(transpose_x = var_4751_transpose_x_0, transpose_y = var_4751_transpose_y_0, x = Minv_149_cast_fp16, y = Np_113_cast_fp16)[name = string("op_4751_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_151_cast_fp16 = add(x = Minv_149_cast_fp16, y = var_4751_cast_fp16)[name = string("Minv_151_cast_fp16")];
            bool u_37_transpose_x_0 = const()[name = string("u_37_transpose_x_0"), val = bool(false)];
            bool u_37_transpose_y_0 = const()[name = string("u_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_37_cast_fp16 = matmul(transpose_x = u_37_transpose_x_0, transpose_y = u_37_transpose_y_0, x = Minv_151_cast_fp16, y = vb_45_cast_fp16)[name = string("u_37_cast_fp16")];
            bool var_4758_transpose_x_1 = const()[name = string("op_4758_transpose_x_1"), val = bool(false)];
            bool var_4758_transpose_y_1 = const()[name = string("op_4758_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_4758_cast_fp16 = matmul(transpose_x = var_4758_transpose_x_1, transpose_y = var_4758_transpose_y_1, x = q_71_cast_fp16, y = k_73_cast_fp16)[name = string("op_4758_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_37_cast_fp16 = mul(x = var_4758_cast_fp16, y = D_37_cast_fp16)[name = string("A_37_cast_fp16")];
            bool var_4760_transpose_x_0 = const()[name = string("op_4760_transpose_x_0"), val = bool(false)];
            bool var_4760_transpose_y_0 = const()[name = string("op_4760_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4760_cast_fp16 = matmul(transpose_x = var_4760_transpose_x_0, transpose_y = var_4760_transpose_y_0, x = A_37_cast_fp16, y = u_37_cast_fp16)[name = string("op_4760_cast_fp16")];
            tensor<fp16, [16]> var_4761_bias_0_to_fp16 = const()[name = string("op_4761_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_4761_cast_fp16 = linear(bias = var_4761_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_45_cast_fp16)[name = string("op_4761_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_31_cast_fp16 = exp(x = var_4761_cast_fp16)[name = string("tail_31_cast_fp16")];
            tensor<int32, [1]> var_4764_axes_0 = const()[name = string("op_4764_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_4764_cast_fp16 = expand_dims(axes = var_4764_axes_0, x = tail_31_cast_fp16)[name = string("op_4764_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4765_cast_fp16 = mul(x = k_73_cast_fp16, y = var_4764_cast_fp16)[name = string("op_4765_cast_fp16")];
            bool S_31_transpose_x_1 = const()[name = string("S_31_transpose_x_1"), val = bool(true)];
            bool S_31_transpose_y_1 = const()[name = string("S_31_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_31_cast_fp16 = matmul(transpose_x = S_31_transpose_x_1, transpose_y = S_31_transpose_y_1, x = var_4765_cast_fp16, y = u_37_cast_fp16)[name = string("S_31_cast_fp16")];
            tensor<int32, [4]> q_73_begin_0 = const()[name = string("q_73_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_73_end_0 = const()[name = string("q_73_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> q_73_end_mask_0 = const()[name = string("q_73_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, x = q_69_cast_fp16)[name = string("q_73_cast_fp16")];
            tensor<int32, [4]> k_75_begin_0 = const()[name = string("k_75_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_75_end_0 = const()[name = string("k_75_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> k_75_end_mask_0 = const()[name = string("k_75_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_75_cast_fp16 = slice_by_index(begin = k_75_begin_0, end = k_75_end_0, end_mask = k_75_end_mask_0, x = k_71_cast_fp16)[name = string("k_75_cast_fp16")];
            tensor<int32, [4]> kb_39_begin_0 = const()[name = string("kb_39_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_39_end_0 = const()[name = string("kb_39_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> kb_39_end_mask_0 = const()[name = string("kb_39_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_39_cast_fp16 = slice_by_index(begin = kb_39_begin_0, end = kb_39_end_0, end_mask = kb_39_end_mask_0, x = var_4670_cast_fp16)[name = string("kb_39_cast_fp16")];
            tensor<int32, [4]> vb_47_begin_0 = const()[name = string("vb_47_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_47_end_0 = const()[name = string("vb_47_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> vb_47_end_mask_0 = const()[name = string("vb_47_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_47_cast_fp16 = slice_by_index(begin = vb_47_begin_0, end = vb_47_end_0, end_mask = vb_47_end_mask_0, x = vb_43_cast_fp16)[name = string("vb_47_cast_fp16")];
            tensor<int32, [3]> gdec_47_begin_0 = const()[name = string("gdec_47_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_47_end_0 = const()[name = string("gdec_47_end_0"), val = tensor<int32, [3]>([1, 4, 32])];
            tensor<bool, [3]> gdec_47_end_mask_0 = const()[name = string("gdec_47_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_47_cast_fp16 = slice_by_index(begin = gdec_47_begin_0, end = gdec_47_end_0, end_mask = gdec_47_end_mask_0, x = gdec_43_cast_fp16)[name = string("gdec_47_cast_fp16")];
            tensor<int32, [2]> var_4850 = const()[name = string("op_4850"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_4851_cast_fp16 = reshape(shape = var_4850, x = gdec_47_cast_fp16)[name = string("op_4851_cast_fp16")];
            tensor<fp16, [4, 256]> var_4852_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_4851_cast_fp16)[name = string("op_4852_cast_fp16")];
            tensor<int32, [4]> var_4857 = const()[name = string("op_4857"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_39_cast_fp16 = reshape(shape = var_4857, x = var_4852_cast_fp16)[name = string("P_39_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4859_cast_fp16 = exp(x = P_39_cast_fp16)[name = string("op_4859_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_39_cast_fp16 = mul(x = var_4859_cast_fp16, y = tril_incl_to_fp16)[name = string("D_39_cast_fp16")];
            bool kk_39_transpose_x_1 = const()[name = string("kk_39_transpose_x_1"), val = bool(false)];
            bool kk_39_transpose_y_1 = const()[name = string("kk_39_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_39_cast_fp16 = matmul(transpose_x = kk_39_transpose_x_1, transpose_y = kk_39_transpose_y_1, x = kb_39_cast_fp16, y = k_75_cast_fp16)[name = string("kk_39_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_4865_cast_fp16 = mul(x = kk_39_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_4865_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4866_cast_fp16 = mul(x = var_4865_cast_fp16, y = D_39_cast_fp16)[name = string("op_4866_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_39_cast_fp16 = mul(x = var_4866_cast_fp16, y = cum_tail_to_fp16)[name = string("N_39_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_4868_cast_fp16 = mul(x = D_39_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_4868_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_39_cast_fp16 = mul(x = kk_39_cast_fp16, y = var_4868_cast_fp16)[name = string("Na_39_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_153_cast_fp16 = add(x = eye_to_fp16, y = N_39_cast_fp16)[name = string("Minv_153_cast_fp16")];
            bool Np_115_transpose_x_0 = const()[name = string("Np_115_transpose_x_0"), val = bool(false)];
            bool Np_115_transpose_y_0 = const()[name = string("Np_115_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_115_cast_fp16 = matmul(transpose_x = Np_115_transpose_x_0, transpose_y = Np_115_transpose_y_0, x = N_39_cast_fp16, y = Na_39_cast_fp16)[name = string("Np_115_cast_fp16")];
            bool Npa_77_transpose_x_0 = const()[name = string("Npa_77_transpose_x_0"), val = bool(false)];
            bool Npa_77_transpose_y_0 = const()[name = string("Npa_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_77_cast_fp16 = matmul(transpose_x = Npa_77_transpose_x_0, transpose_y = Npa_77_transpose_y_0, x = Na_39_cast_fp16, y = N_39_cast_fp16)[name = string("Npa_77_cast_fp16")];
            bool var_4874_transpose_x_0 = const()[name = string("op_4874_transpose_x_0"), val = bool(false)];
            bool var_4874_transpose_y_0 = const()[name = string("op_4874_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4874_cast_fp16 = matmul(transpose_x = var_4874_transpose_x_0, transpose_y = var_4874_transpose_y_0, x = Minv_153_cast_fp16, y = Np_115_cast_fp16)[name = string("op_4874_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_155_cast_fp16 = add(x = Minv_153_cast_fp16, y = var_4874_cast_fp16)[name = string("Minv_155_cast_fp16")];
            bool Np_117_transpose_x_0 = const()[name = string("Np_117_transpose_x_0"), val = bool(false)];
            bool Np_117_transpose_y_0 = const()[name = string("Np_117_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_117_cast_fp16 = matmul(transpose_x = Np_117_transpose_x_0, transpose_y = Np_117_transpose_y_0, x = Np_115_cast_fp16, y = Npa_77_cast_fp16)[name = string("Np_117_cast_fp16")];
            bool Npa_79_transpose_x_0 = const()[name = string("Npa_79_transpose_x_0"), val = bool(false)];
            bool Npa_79_transpose_y_0 = const()[name = string("Npa_79_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_79_cast_fp16 = matmul(transpose_x = Npa_79_transpose_x_0, transpose_y = Npa_79_transpose_y_0, x = Npa_77_cast_fp16, y = Np_115_cast_fp16)[name = string("Npa_79_cast_fp16")];
            bool var_4879_transpose_x_0 = const()[name = string("op_4879_transpose_x_0"), val = bool(false)];
            bool var_4879_transpose_y_0 = const()[name = string("op_4879_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4879_cast_fp16 = matmul(transpose_x = var_4879_transpose_x_0, transpose_y = var_4879_transpose_y_0, x = Minv_155_cast_fp16, y = Np_117_cast_fp16)[name = string("op_4879_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_157_cast_fp16 = add(x = Minv_155_cast_fp16, y = var_4879_cast_fp16)[name = string("Minv_157_cast_fp16")];
            bool Np_119_transpose_x_0 = const()[name = string("Np_119_transpose_x_0"), val = bool(false)];
            bool Np_119_transpose_y_0 = const()[name = string("Np_119_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_119_cast_fp16 = matmul(transpose_x = Np_119_transpose_x_0, transpose_y = Np_119_transpose_y_0, x = Np_117_cast_fp16, y = Npa_79_cast_fp16)[name = string("Np_119_cast_fp16")];
            bool var_4883_transpose_x_0 = const()[name = string("op_4883_transpose_x_0"), val = bool(false)];
            bool var_4883_transpose_y_0 = const()[name = string("op_4883_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_4883_cast_fp16 = matmul(transpose_x = var_4883_transpose_x_0, transpose_y = var_4883_transpose_y_0, x = Minv_157_cast_fp16, y = Np_119_cast_fp16)[name = string("op_4883_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_159_cast_fp16 = add(x = Minv_157_cast_fp16, y = var_4883_cast_fp16)[name = string("Minv_159_cast_fp16")];
            bool u_39_transpose_x_0 = const()[name = string("u_39_transpose_x_0"), val = bool(false)];
            bool u_39_transpose_y_0 = const()[name = string("u_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_39_cast_fp16 = matmul(transpose_x = u_39_transpose_x_0, transpose_y = u_39_transpose_y_0, x = Minv_159_cast_fp16, y = vb_47_cast_fp16)[name = string("u_39_cast_fp16")];
            bool var_4890_transpose_x_1 = const()[name = string("op_4890_transpose_x_1"), val = bool(false)];
            bool var_4890_transpose_y_1 = const()[name = string("op_4890_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_4890_cast_fp16 = matmul(transpose_x = var_4890_transpose_x_1, transpose_y = var_4890_transpose_y_1, x = q_73_cast_fp16, y = k_75_cast_fp16)[name = string("op_4890_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_39_cast_fp16 = mul(x = var_4890_cast_fp16, y = D_39_cast_fp16)[name = string("A_39_cast_fp16")];
            tensor<fp16, [16]> var_4892_bias_0_to_fp16 = const()[name = string("op_4892_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_4892_cast_fp16 = linear(bias = var_4892_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_47_cast_fp16)[name = string("op_4892_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_31_cast_fp16 = exp(x = var_4892_cast_fp16)[name = string("cum_31_cast_fp16")];
            tensor<int32, [1]> var_4895_axes_0 = const()[name = string("op_4895_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_4895_cast_fp16 = expand_dims(axes = var_4895_axes_0, x = cum_31_cast_fp16)[name = string("op_4895_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4896_cast_fp16 = mul(x = kb_39_cast_fp16, y = var_4895_cast_fp16)[name = string("op_4896_cast_fp16")];
            bool w_39_transpose_x_0 = const()[name = string("w_39_transpose_x_0"), val = bool(false)];
            bool w_39_transpose_y_0 = const()[name = string("w_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_39_cast_fp16 = matmul(transpose_x = w_39_transpose_x_0, transpose_y = w_39_transpose_y_0, x = Minv_159_cast_fp16, y = var_4896_cast_fp16)[name = string("w_39_cast_fp16")];
            bool var_4898_transpose_x_0 = const()[name = string("op_4898_transpose_x_0"), val = bool(false)];
            bool var_4898_transpose_y_0 = const()[name = string("op_4898_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4898_cast_fp16 = matmul(transpose_x = var_4898_transpose_x_0, transpose_y = var_4898_transpose_y_0, x = w_39_cast_fp16, y = S_31_cast_fp16)[name = string("op_4898_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_31_cast_fp16 = sub(x = u_39_cast_fp16, y = var_4898_cast_fp16)[name = string("vnew_31_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4903_cast_fp16 = mul(x = q_73_cast_fp16, y = var_4895_cast_fp16)[name = string("op_4903_cast_fp16")];
            bool var_4904_transpose_x_0 = const()[name = string("op_4904_transpose_x_0"), val = bool(false)];
            bool var_4904_transpose_y_0 = const()[name = string("op_4904_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4904_cast_fp16 = matmul(transpose_x = var_4904_transpose_x_0, transpose_y = var_4904_transpose_y_0, x = var_4903_cast_fp16, y = S_31_cast_fp16)[name = string("op_4904_cast_fp16")];
            bool var_4905_transpose_x_0 = const()[name = string("op_4905_transpose_x_0"), val = bool(false)];
            bool var_4905_transpose_y_0 = const()[name = string("op_4905_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_4905_cast_fp16 = matmul(transpose_x = var_4905_transpose_x_0, transpose_y = var_4905_transpose_y_0, x = A_39_cast_fp16, y = vnew_31_cast_fp16)[name = string("op_4905_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4907_cast_fp16 = add(x = var_4904_cast_fp16, y = var_4905_cast_fp16)[name = string("op_4907_cast_fp16")];
            tensor<fp16, [16]> var_4908_bias_0_to_fp16 = const()[name = string("op_4908_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_4908_cast_fp16 = linear(bias = var_4908_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_47_cast_fp16)[name = string("op_4908_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_33_cast_fp16 = exp(x = var_4908_cast_fp16)[name = string("tail_33_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_65_to_fp16 = const()[name = string("transpose_65_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_4910_bias_0_to_fp16 = const()[name = string("op_4910_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_4910_cast_fp16 = linear(bias = var_4910_bias_0_to_fp16, weight = transpose_65_to_fp16, x = gdec_47_cast_fp16)[name = string("op_4910_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_4911_cast_fp16 = exp(x = var_4910_cast_fp16)[name = string("op_4911_cast_fp16")];
            tensor<int32, [1]> last_27_axes_0 = const()[name = string("last_27_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_27_cast_fp16 = expand_dims(axes = last_27_axes_0, x = var_4911_cast_fp16)[name = string("last_27_cast_fp16")];
            tensor<int32, [1]> var_4915_axes_0 = const()[name = string("op_4915_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_4915_cast_fp16 = expand_dims(axes = var_4915_axes_0, x = tail_33_cast_fp16)[name = string("op_4915_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_4916_cast_fp16 = mul(x = k_75_cast_fp16, y = var_4915_cast_fp16)[name = string("op_4916_cast_fp16")];
            bool upd_25_transpose_x_1 = const()[name = string("upd_25_transpose_x_1"), val = bool(true)];
            bool upd_25_transpose_y_1 = const()[name = string("upd_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_25_cast_fp16 = matmul(transpose_x = upd_25_transpose_x_1, transpose_y = upd_25_transpose_y_1, x = var_4916_cast_fp16, y = vnew_31_cast_fp16)[name = string("upd_25_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_4921_cast_fp16 = mul(x = S_31_cast_fp16, y = last_27_cast_fp16)[name = string("op_4921_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_33_cast_fp16 = add(x = var_4921_cast_fp16, y = upd_25_cast_fp16)[name = string("S_33_cast_fp16")];
            tensor<int32, [4]> q_75_begin_0 = const()[name = string("q_75_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_75_end_0 = const()[name = string("q_75_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> q_75_end_mask_0 = const()[name = string("q_75_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_75_cast_fp16 = slice_by_index(begin = q_75_begin_0, end = q_75_end_0, end_mask = q_75_end_mask_0, x = q_69_cast_fp16)[name = string("q_75_cast_fp16")];
            tensor<int32, [4]> k_77_begin_0 = const()[name = string("k_77_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_77_end_0 = const()[name = string("k_77_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> k_77_end_mask_0 = const()[name = string("k_77_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, x = k_71_cast_fp16)[name = string("k_77_cast_fp16")];
            tensor<int32, [4]> kb_41_begin_0 = const()[name = string("kb_41_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_41_end_0 = const()[name = string("kb_41_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> kb_41_end_mask_0 = const()[name = string("kb_41_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_41_cast_fp16 = slice_by_index(begin = kb_41_begin_0, end = kb_41_end_0, end_mask = kb_41_end_mask_0, x = var_4670_cast_fp16)[name = string("kb_41_cast_fp16")];
            tensor<int32, [4]> vb_49_begin_0 = const()[name = string("vb_49_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_49_end_0 = const()[name = string("vb_49_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> vb_49_end_mask_0 = const()[name = string("vb_49_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_49_cast_fp16 = slice_by_index(begin = vb_49_begin_0, end = vb_49_end_0, end_mask = vb_49_end_mask_0, x = vb_43_cast_fp16)[name = string("vb_49_cast_fp16")];
            tensor<int32, [3]> gdec_49_begin_0 = const()[name = string("gdec_49_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_49_end_0 = const()[name = string("gdec_49_end_0"), val = tensor<int32, [3]>([1, 4, 48])];
            tensor<bool, [3]> gdec_49_end_mask_0 = const()[name = string("gdec_49_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_49_cast_fp16 = slice_by_index(begin = gdec_49_begin_0, end = gdec_49_end_0, end_mask = gdec_49_end_mask_0, x = gdec_43_cast_fp16)[name = string("gdec_49_cast_fp16")];
            tensor<int32, [2]> var_5004 = const()[name = string("op_5004"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_5005_cast_fp16 = reshape(shape = var_5004, x = gdec_49_cast_fp16)[name = string("op_5005_cast_fp16")];
            tensor<fp16, [4, 256]> var_5006_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5005_cast_fp16)[name = string("op_5006_cast_fp16")];
            tensor<int32, [4]> var_5011 = const()[name = string("op_5011"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_41_cast_fp16 = reshape(shape = var_5011, x = var_5006_cast_fp16)[name = string("P_41_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5013_cast_fp16 = exp(x = P_41_cast_fp16)[name = string("op_5013_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_41_cast_fp16 = mul(x = var_5013_cast_fp16, y = tril_incl_to_fp16)[name = string("D_41_cast_fp16")];
            bool kk_41_transpose_x_1 = const()[name = string("kk_41_transpose_x_1"), val = bool(false)];
            bool kk_41_transpose_y_1 = const()[name = string("kk_41_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_41_cast_fp16 = matmul(transpose_x = kk_41_transpose_x_1, transpose_y = kk_41_transpose_y_1, x = kb_41_cast_fp16, y = k_77_cast_fp16)[name = string("kk_41_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_5019_cast_fp16 = mul(x = kk_41_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_5019_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5020_cast_fp16 = mul(x = var_5019_cast_fp16, y = D_41_cast_fp16)[name = string("op_5020_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_41_cast_fp16 = mul(x = var_5020_cast_fp16, y = cum_tail_to_fp16)[name = string("N_41_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5022_cast_fp16 = mul(x = D_41_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5022_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_41_cast_fp16 = mul(x = kk_41_cast_fp16, y = var_5022_cast_fp16)[name = string("Na_41_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_161_cast_fp16 = add(x = eye_to_fp16, y = N_41_cast_fp16)[name = string("Minv_161_cast_fp16")];
            bool Np_121_transpose_x_0 = const()[name = string("Np_121_transpose_x_0"), val = bool(false)];
            bool Np_121_transpose_y_0 = const()[name = string("Np_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_121_cast_fp16 = matmul(transpose_x = Np_121_transpose_x_0, transpose_y = Np_121_transpose_y_0, x = N_41_cast_fp16, y = Na_41_cast_fp16)[name = string("Np_121_cast_fp16")];
            bool Npa_81_transpose_x_0 = const()[name = string("Npa_81_transpose_x_0"), val = bool(false)];
            bool Npa_81_transpose_y_0 = const()[name = string("Npa_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_81_cast_fp16 = matmul(transpose_x = Npa_81_transpose_x_0, transpose_y = Npa_81_transpose_y_0, x = Na_41_cast_fp16, y = N_41_cast_fp16)[name = string("Npa_81_cast_fp16")];
            bool var_5028_transpose_x_0 = const()[name = string("op_5028_transpose_x_0"), val = bool(false)];
            bool var_5028_transpose_y_0 = const()[name = string("op_5028_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5028_cast_fp16 = matmul(transpose_x = var_5028_transpose_x_0, transpose_y = var_5028_transpose_y_0, x = Minv_161_cast_fp16, y = Np_121_cast_fp16)[name = string("op_5028_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_163_cast_fp16 = add(x = Minv_161_cast_fp16, y = var_5028_cast_fp16)[name = string("Minv_163_cast_fp16")];
            bool Np_123_transpose_x_0 = const()[name = string("Np_123_transpose_x_0"), val = bool(false)];
            bool Np_123_transpose_y_0 = const()[name = string("Np_123_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_123_cast_fp16 = matmul(transpose_x = Np_123_transpose_x_0, transpose_y = Np_123_transpose_y_0, x = Np_121_cast_fp16, y = Npa_81_cast_fp16)[name = string("Np_123_cast_fp16")];
            bool Npa_83_transpose_x_0 = const()[name = string("Npa_83_transpose_x_0"), val = bool(false)];
            bool Npa_83_transpose_y_0 = const()[name = string("Npa_83_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_83_cast_fp16 = matmul(transpose_x = Npa_83_transpose_x_0, transpose_y = Npa_83_transpose_y_0, x = Npa_81_cast_fp16, y = Np_121_cast_fp16)[name = string("Npa_83_cast_fp16")];
            bool var_5033_transpose_x_0 = const()[name = string("op_5033_transpose_x_0"), val = bool(false)];
            bool var_5033_transpose_y_0 = const()[name = string("op_5033_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5033_cast_fp16 = matmul(transpose_x = var_5033_transpose_x_0, transpose_y = var_5033_transpose_y_0, x = Minv_163_cast_fp16, y = Np_123_cast_fp16)[name = string("op_5033_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_165_cast_fp16 = add(x = Minv_163_cast_fp16, y = var_5033_cast_fp16)[name = string("Minv_165_cast_fp16")];
            bool Np_125_transpose_x_0 = const()[name = string("Np_125_transpose_x_0"), val = bool(false)];
            bool Np_125_transpose_y_0 = const()[name = string("Np_125_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_125_cast_fp16 = matmul(transpose_x = Np_125_transpose_x_0, transpose_y = Np_125_transpose_y_0, x = Np_123_cast_fp16, y = Npa_83_cast_fp16)[name = string("Np_125_cast_fp16")];
            bool var_5037_transpose_x_0 = const()[name = string("op_5037_transpose_x_0"), val = bool(false)];
            bool var_5037_transpose_y_0 = const()[name = string("op_5037_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5037_cast_fp16 = matmul(transpose_x = var_5037_transpose_x_0, transpose_y = var_5037_transpose_y_0, x = Minv_165_cast_fp16, y = Np_125_cast_fp16)[name = string("op_5037_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_167_cast_fp16 = add(x = Minv_165_cast_fp16, y = var_5037_cast_fp16)[name = string("Minv_167_cast_fp16")];
            bool u_41_transpose_x_0 = const()[name = string("u_41_transpose_x_0"), val = bool(false)];
            bool u_41_transpose_y_0 = const()[name = string("u_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_41_cast_fp16 = matmul(transpose_x = u_41_transpose_x_0, transpose_y = u_41_transpose_y_0, x = Minv_167_cast_fp16, y = vb_49_cast_fp16)[name = string("u_41_cast_fp16")];
            bool var_5044_transpose_x_1 = const()[name = string("op_5044_transpose_x_1"), val = bool(false)];
            bool var_5044_transpose_y_1 = const()[name = string("op_5044_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_5044_cast_fp16 = matmul(transpose_x = var_5044_transpose_x_1, transpose_y = var_5044_transpose_y_1, x = q_75_cast_fp16, y = k_77_cast_fp16)[name = string("op_5044_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_41_cast_fp16 = mul(x = var_5044_cast_fp16, y = D_41_cast_fp16)[name = string("A_41_cast_fp16")];
            tensor<fp16, [16]> var_5046_bias_0_to_fp16 = const()[name = string("op_5046_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_5046_cast_fp16 = linear(bias = var_5046_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_49_cast_fp16)[name = string("op_5046_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_33_cast_fp16 = exp(x = var_5046_cast_fp16)[name = string("cum_33_cast_fp16")];
            tensor<int32, [1]> var_5049_axes_0 = const()[name = string("op_5049_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_5049_cast_fp16 = expand_dims(axes = var_5049_axes_0, x = cum_33_cast_fp16)[name = string("op_5049_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5050_cast_fp16 = mul(x = kb_41_cast_fp16, y = var_5049_cast_fp16)[name = string("op_5050_cast_fp16")];
            bool w_41_transpose_x_0 = const()[name = string("w_41_transpose_x_0"), val = bool(false)];
            bool w_41_transpose_y_0 = const()[name = string("w_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_41_cast_fp16 = matmul(transpose_x = w_41_transpose_x_0, transpose_y = w_41_transpose_y_0, x = Minv_167_cast_fp16, y = var_5050_cast_fp16)[name = string("w_41_cast_fp16")];
            bool var_5052_transpose_x_0 = const()[name = string("op_5052_transpose_x_0"), val = bool(false)];
            bool var_5052_transpose_y_0 = const()[name = string("op_5052_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5052_cast_fp16 = matmul(transpose_x = var_5052_transpose_x_0, transpose_y = var_5052_transpose_y_0, x = w_41_cast_fp16, y = S_33_cast_fp16)[name = string("op_5052_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_33_cast_fp16 = sub(x = u_41_cast_fp16, y = var_5052_cast_fp16)[name = string("vnew_33_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5057_cast_fp16 = mul(x = q_75_cast_fp16, y = var_5049_cast_fp16)[name = string("op_5057_cast_fp16")];
            bool var_5058_transpose_x_0 = const()[name = string("op_5058_transpose_x_0"), val = bool(false)];
            bool var_5058_transpose_y_0 = const()[name = string("op_5058_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5058_cast_fp16 = matmul(transpose_x = var_5058_transpose_x_0, transpose_y = var_5058_transpose_y_0, x = var_5057_cast_fp16, y = S_33_cast_fp16)[name = string("op_5058_cast_fp16")];
            bool var_5059_transpose_x_0 = const()[name = string("op_5059_transpose_x_0"), val = bool(false)];
            bool var_5059_transpose_y_0 = const()[name = string("op_5059_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5059_cast_fp16 = matmul(transpose_x = var_5059_transpose_x_0, transpose_y = var_5059_transpose_y_0, x = A_41_cast_fp16, y = vnew_33_cast_fp16)[name = string("op_5059_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5061_cast_fp16 = add(x = var_5058_cast_fp16, y = var_5059_cast_fp16)[name = string("op_5061_cast_fp16")];
            tensor<fp16, [16]> var_5062_bias_0_to_fp16 = const()[name = string("op_5062_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_5062_cast_fp16 = linear(bias = var_5062_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_49_cast_fp16)[name = string("op_5062_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_35_cast_fp16 = exp(x = var_5062_cast_fp16)[name = string("tail_35_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_69_to_fp16 = const()[name = string("transpose_69_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_5064_bias_0_to_fp16 = const()[name = string("op_5064_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_5064_cast_fp16 = linear(bias = var_5064_bias_0_to_fp16, weight = transpose_69_to_fp16, x = gdec_49_cast_fp16)[name = string("op_5064_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_5065_cast_fp16 = exp(x = var_5064_cast_fp16)[name = string("op_5065_cast_fp16")];
            tensor<int32, [1]> last_29_axes_0 = const()[name = string("last_29_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_29_cast_fp16 = expand_dims(axes = last_29_axes_0, x = var_5065_cast_fp16)[name = string("last_29_cast_fp16")];
            tensor<int32, [1]> var_5069_axes_0 = const()[name = string("op_5069_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_5069_cast_fp16 = expand_dims(axes = var_5069_axes_0, x = tail_35_cast_fp16)[name = string("op_5069_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5070_cast_fp16 = mul(x = k_77_cast_fp16, y = var_5069_cast_fp16)[name = string("op_5070_cast_fp16")];
            bool upd_27_transpose_x_1 = const()[name = string("upd_27_transpose_x_1"), val = bool(true)];
            bool upd_27_transpose_y_1 = const()[name = string("upd_27_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_27_cast_fp16 = matmul(transpose_x = upd_27_transpose_x_1, transpose_y = upd_27_transpose_y_1, x = var_5070_cast_fp16, y = vnew_33_cast_fp16)[name = string("upd_27_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_5075_cast_fp16 = mul(x = S_33_cast_fp16, y = last_29_cast_fp16)[name = string("op_5075_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_35_cast_fp16 = add(x = var_5075_cast_fp16, y = upd_27_cast_fp16)[name = string("S_35_cast_fp16")];
            tensor<int32, [4]> q_77_begin_0 = const()[name = string("q_77_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_77_end_0 = const()[name = string("q_77_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> q_77_end_mask_0 = const()[name = string("q_77_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_77_cast_fp16 = slice_by_index(begin = q_77_begin_0, end = q_77_end_0, end_mask = q_77_end_mask_0, x = q_69_cast_fp16)[name = string("q_77_cast_fp16")];
            tensor<int32, [4]> k_79_begin_0 = const()[name = string("k_79_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_79_end_0 = const()[name = string("k_79_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> k_79_end_mask_0 = const()[name = string("k_79_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_79_cast_fp16 = slice_by_index(begin = k_79_begin_0, end = k_79_end_0, end_mask = k_79_end_mask_0, x = k_71_cast_fp16)[name = string("k_79_cast_fp16")];
            tensor<int32, [4]> kb_43_begin_0 = const()[name = string("kb_43_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_43_end_0 = const()[name = string("kb_43_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> kb_43_end_mask_0 = const()[name = string("kb_43_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_43_cast_fp16 = slice_by_index(begin = kb_43_begin_0, end = kb_43_end_0, end_mask = kb_43_end_mask_0, x = var_4670_cast_fp16)[name = string("kb_43_cast_fp16")];
            tensor<int32, [4]> vb_51_begin_0 = const()[name = string("vb_51_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_51_end_0 = const()[name = string("vb_51_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> vb_51_end_mask_0 = const()[name = string("vb_51_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_51_cast_fp16 = slice_by_index(begin = vb_51_begin_0, end = vb_51_end_0, end_mask = vb_51_end_mask_0, x = vb_43_cast_fp16)[name = string("vb_51_cast_fp16")];
            tensor<int32, [3]> gdec_51_begin_0 = const()[name = string("gdec_51_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_51_end_0 = const()[name = string("gdec_51_end_0"), val = tensor<int32, [3]>([1, 4, 64])];
            tensor<bool, [3]> gdec_51_end_mask_0 = const()[name = string("gdec_51_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_51_cast_fp16 = slice_by_index(begin = gdec_51_begin_0, end = gdec_51_end_0, end_mask = gdec_51_end_mask_0, x = gdec_43_cast_fp16)[name = string("gdec_51_cast_fp16")];
            tensor<int32, [2]> var_5158 = const()[name = string("op_5158"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_5159_cast_fp16 = reshape(shape = var_5158, x = gdec_51_cast_fp16)[name = string("op_5159_cast_fp16")];
            tensor<fp16, [4, 256]> var_5160_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5159_cast_fp16)[name = string("op_5160_cast_fp16")];
            tensor<int32, [4]> var_5165 = const()[name = string("op_5165"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_43_cast_fp16 = reshape(shape = var_5165, x = var_5160_cast_fp16)[name = string("P_43_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5167_cast_fp16 = exp(x = P_43_cast_fp16)[name = string("op_5167_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_43_cast_fp16 = mul(x = var_5167_cast_fp16, y = tril_incl_to_fp16)[name = string("D_43_cast_fp16")];
            bool kk_43_transpose_x_1 = const()[name = string("kk_43_transpose_x_1"), val = bool(false)];
            bool kk_43_transpose_y_1 = const()[name = string("kk_43_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_43_cast_fp16 = matmul(transpose_x = kk_43_transpose_x_1, transpose_y = kk_43_transpose_y_1, x = kb_43_cast_fp16, y = k_79_cast_fp16)[name = string("kk_43_cast_fp16")];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_5173_cast_fp16 = mul(x = kk_43_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_5173_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5174_cast_fp16 = mul(x = var_5173_cast_fp16, y = D_43_cast_fp16)[name = string("op_5174_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_43_cast_fp16 = mul(x = var_5174_cast_fp16, y = cum_tail_to_fp16)[name = string("N_43_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5176_cast_fp16 = mul(x = D_43_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5176_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_43_cast_fp16 = mul(x = kk_43_cast_fp16, y = var_5176_cast_fp16)[name = string("Na_43_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_169_cast_fp16 = add(x = eye_to_fp16, y = N_43_cast_fp16)[name = string("Minv_169_cast_fp16")];
            bool Np_127_transpose_x_0 = const()[name = string("Np_127_transpose_x_0"), val = bool(false)];
            bool Np_127_transpose_y_0 = const()[name = string("Np_127_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_127_cast_fp16 = matmul(transpose_x = Np_127_transpose_x_0, transpose_y = Np_127_transpose_y_0, x = N_43_cast_fp16, y = Na_43_cast_fp16)[name = string("Np_127_cast_fp16")];
            bool Npa_85_transpose_x_0 = const()[name = string("Npa_85_transpose_x_0"), val = bool(false)];
            bool Npa_85_transpose_y_0 = const()[name = string("Npa_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_85_cast_fp16 = matmul(transpose_x = Npa_85_transpose_x_0, transpose_y = Npa_85_transpose_y_0, x = Na_43_cast_fp16, y = N_43_cast_fp16)[name = string("Npa_85_cast_fp16")];
            bool var_5182_transpose_x_0 = const()[name = string("op_5182_transpose_x_0"), val = bool(false)];
            bool var_5182_transpose_y_0 = const()[name = string("op_5182_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5182_cast_fp16 = matmul(transpose_x = var_5182_transpose_x_0, transpose_y = var_5182_transpose_y_0, x = Minv_169_cast_fp16, y = Np_127_cast_fp16)[name = string("op_5182_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_171_cast_fp16 = add(x = Minv_169_cast_fp16, y = var_5182_cast_fp16)[name = string("Minv_171_cast_fp16")];
            bool Np_129_transpose_x_0 = const()[name = string("Np_129_transpose_x_0"), val = bool(false)];
            bool Np_129_transpose_y_0 = const()[name = string("Np_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_129_cast_fp16 = matmul(transpose_x = Np_129_transpose_x_0, transpose_y = Np_129_transpose_y_0, x = Np_127_cast_fp16, y = Npa_85_cast_fp16)[name = string("Np_129_cast_fp16")];
            bool Npa_87_transpose_x_0 = const()[name = string("Npa_87_transpose_x_0"), val = bool(false)];
            bool Npa_87_transpose_y_0 = const()[name = string("Npa_87_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_87_cast_fp16 = matmul(transpose_x = Npa_87_transpose_x_0, transpose_y = Npa_87_transpose_y_0, x = Npa_85_cast_fp16, y = Np_127_cast_fp16)[name = string("Npa_87_cast_fp16")];
            bool var_5187_transpose_x_0 = const()[name = string("op_5187_transpose_x_0"), val = bool(false)];
            bool var_5187_transpose_y_0 = const()[name = string("op_5187_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5187_cast_fp16 = matmul(transpose_x = var_5187_transpose_x_0, transpose_y = var_5187_transpose_y_0, x = Minv_171_cast_fp16, y = Np_129_cast_fp16)[name = string("op_5187_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_173_cast_fp16 = add(x = Minv_171_cast_fp16, y = var_5187_cast_fp16)[name = string("Minv_173_cast_fp16")];
            bool Np_131_transpose_x_0 = const()[name = string("Np_131_transpose_x_0"), val = bool(false)];
            bool Np_131_transpose_y_0 = const()[name = string("Np_131_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_131_cast_fp16 = matmul(transpose_x = Np_131_transpose_x_0, transpose_y = Np_131_transpose_y_0, x = Np_129_cast_fp16, y = Npa_87_cast_fp16)[name = string("Np_131_cast_fp16")];
            bool var_5191_transpose_x_0 = const()[name = string("op_5191_transpose_x_0"), val = bool(false)];
            bool var_5191_transpose_y_0 = const()[name = string("op_5191_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5191_cast_fp16 = matmul(transpose_x = var_5191_transpose_x_0, transpose_y = var_5191_transpose_y_0, x = Minv_173_cast_fp16, y = Np_131_cast_fp16)[name = string("op_5191_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_175_cast_fp16 = add(x = Minv_173_cast_fp16, y = var_5191_cast_fp16)[name = string("Minv_175_cast_fp16")];
            bool u_43_transpose_x_0 = const()[name = string("u_43_transpose_x_0"), val = bool(false)];
            bool u_43_transpose_y_0 = const()[name = string("u_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_43_cast_fp16 = matmul(transpose_x = u_43_transpose_x_0, transpose_y = u_43_transpose_y_0, x = Minv_175_cast_fp16, y = vb_51_cast_fp16)[name = string("u_43_cast_fp16")];
            bool var_5198_transpose_x_1 = const()[name = string("op_5198_transpose_x_1"), val = bool(false)];
            bool var_5198_transpose_y_1 = const()[name = string("op_5198_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_5198_cast_fp16 = matmul(transpose_x = var_5198_transpose_x_1, transpose_y = var_5198_transpose_y_1, x = q_77_cast_fp16, y = k_79_cast_fp16)[name = string("op_5198_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_43_cast_fp16 = mul(x = var_5198_cast_fp16, y = D_43_cast_fp16)[name = string("A_43_cast_fp16")];
            tensor<fp16, [16]> var_5200_bias_0_to_fp16 = const()[name = string("op_5200_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_5200_cast_fp16 = linear(bias = var_5200_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_51_cast_fp16)[name = string("op_5200_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_35_cast_fp16 = exp(x = var_5200_cast_fp16)[name = string("cum_35_cast_fp16")];
            tensor<int32, [1]> var_5203_axes_0 = const()[name = string("op_5203_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_5203_cast_fp16 = expand_dims(axes = var_5203_axes_0, x = cum_35_cast_fp16)[name = string("op_5203_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5204_cast_fp16 = mul(x = kb_43_cast_fp16, y = var_5203_cast_fp16)[name = string("op_5204_cast_fp16")];
            bool w_43_transpose_x_0 = const()[name = string("w_43_transpose_x_0"), val = bool(false)];
            bool w_43_transpose_y_0 = const()[name = string("w_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_43_cast_fp16 = matmul(transpose_x = w_43_transpose_x_0, transpose_y = w_43_transpose_y_0, x = Minv_175_cast_fp16, y = var_5204_cast_fp16)[name = string("w_43_cast_fp16")];
            bool var_5206_transpose_x_0 = const()[name = string("op_5206_transpose_x_0"), val = bool(false)];
            bool var_5206_transpose_y_0 = const()[name = string("op_5206_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5206_cast_fp16 = matmul(transpose_x = var_5206_transpose_x_0, transpose_y = var_5206_transpose_y_0, x = w_43_cast_fp16, y = S_35_cast_fp16)[name = string("op_5206_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_35_cast_fp16 = sub(x = u_43_cast_fp16, y = var_5206_cast_fp16)[name = string("vnew_35_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5211_cast_fp16 = mul(x = q_77_cast_fp16, y = var_5203_cast_fp16)[name = string("op_5211_cast_fp16")];
            bool var_5212_transpose_x_0 = const()[name = string("op_5212_transpose_x_0"), val = bool(false)];
            bool var_5212_transpose_y_0 = const()[name = string("op_5212_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5212_cast_fp16 = matmul(transpose_x = var_5212_transpose_x_0, transpose_y = var_5212_transpose_y_0, x = var_5211_cast_fp16, y = S_35_cast_fp16)[name = string("op_5212_cast_fp16")];
            bool var_5213_transpose_x_0 = const()[name = string("op_5213_transpose_x_0"), val = bool(false)];
            bool var_5213_transpose_y_0 = const()[name = string("op_5213_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5213_cast_fp16 = matmul(transpose_x = var_5213_transpose_x_0, transpose_y = var_5213_transpose_y_0, x = A_43_cast_fp16, y = vnew_35_cast_fp16)[name = string("op_5213_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5215_cast_fp16 = add(x = var_5212_cast_fp16, y = var_5213_cast_fp16)[name = string("op_5215_cast_fp16")];
            tensor<fp16, [16]> var_5216_bias_0_to_fp16 = const()[name = string("op_5216_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_5216_cast_fp16 = linear(bias = var_5216_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_51_cast_fp16)[name = string("op_5216_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_37_cast_fp16 = exp(x = var_5216_cast_fp16)[name = string("tail_37_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_73_to_fp16 = const()[name = string("transpose_73_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_5218_bias_0_to_fp16 = const()[name = string("op_5218_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_5218_cast_fp16 = linear(bias = var_5218_bias_0_to_fp16, weight = transpose_73_to_fp16, x = gdec_51_cast_fp16)[name = string("op_5218_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_5219_cast_fp16 = exp(x = var_5218_cast_fp16)[name = string("op_5219_cast_fp16")];
            tensor<int32, [1]> last_31_axes_0 = const()[name = string("last_31_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_31_cast_fp16 = expand_dims(axes = last_31_axes_0, x = var_5219_cast_fp16)[name = string("last_31_cast_fp16")];
            tensor<int32, [1]> var_5223_axes_0 = const()[name = string("op_5223_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_5223_cast_fp16 = expand_dims(axes = var_5223_axes_0, x = tail_37_cast_fp16)[name = string("op_5223_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5224_cast_fp16 = mul(x = k_79_cast_fp16, y = var_5223_cast_fp16)[name = string("op_5224_cast_fp16")];
            bool upd_29_transpose_x_1 = const()[name = string("upd_29_transpose_x_1"), val = bool(true)];
            bool upd_29_transpose_y_1 = const()[name = string("upd_29_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_29_cast_fp16 = matmul(transpose_x = upd_29_transpose_x_1, transpose_y = upd_29_transpose_y_1, x = var_5224_cast_fp16, y = vnew_35_cast_fp16)[name = string("upd_29_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_5229_cast_fp16 = mul(x = S_35_cast_fp16, y = last_31_cast_fp16)[name = string("op_5229_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_37_cast_fp16 = add(x = var_5229_cast_fp16, y = upd_29_cast_fp16)[name = string("S_37_cast_fp16")];
            tensor<int32, [4]> q_79_begin_0 = const()[name = string("q_79_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> q_79_end_0 = const()[name = string("q_79_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> q_79_end_mask_0 = const()[name = string("q_79_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, x = q_69_cast_fp16)[name = string("q_79_cast_fp16")];
            tensor<int32, [4]> k_81_begin_0 = const()[name = string("k_81_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> k_81_end_0 = const()[name = string("k_81_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> k_81_end_mask_0 = const()[name = string("k_81_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, x = k_71_cast_fp16)[name = string("k_81_cast_fp16")];
            tensor<int32, [4]> kb_45_begin_0 = const()[name = string("kb_45_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> kb_45_end_0 = const()[name = string("kb_45_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> kb_45_end_mask_0 = const()[name = string("kb_45_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_45_cast_fp16 = slice_by_index(begin = kb_45_begin_0, end = kb_45_end_0, end_mask = kb_45_end_mask_0, x = var_4670_cast_fp16)[name = string("kb_45_cast_fp16")];
            tensor<int32, [4]> vb_53_begin_0 = const()[name = string("vb_53_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> vb_53_end_0 = const()[name = string("vb_53_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> vb_53_end_mask_0 = const()[name = string("vb_53_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_53_cast_fp16 = slice_by_index(begin = vb_53_begin_0, end = vb_53_end_0, end_mask = vb_53_end_mask_0, x = vb_43_cast_fp16)[name = string("vb_53_cast_fp16")];
            tensor<int32, [3]> gdec_53_begin_0 = const()[name = string("gdec_53_begin_0"), val = tensor<int32, [3]>([0, 0, 64])];
            tensor<int32, [3]> gdec_53_end_0 = const()[name = string("gdec_53_end_0"), val = tensor<int32, [3]>([1, 4, 80])];
            tensor<bool, [3]> gdec_53_end_mask_0 = const()[name = string("gdec_53_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_53_cast_fp16 = slice_by_index(begin = gdec_53_begin_0, end = gdec_53_end_0, end_mask = gdec_53_end_mask_0, x = gdec_43_cast_fp16)[name = string("gdec_53_cast_fp16")];
            tensor<int32, [2]> var_5312 = const()[name = string("op_5312"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_5313_cast_fp16 = reshape(shape = var_5312, x = gdec_53_cast_fp16)[name = string("op_5313_cast_fp16")];
            tensor<fp16, [4, 256]> var_5314_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5313_cast_fp16)[name = string("op_5314_cast_fp16")];
            tensor<int32, [4]> var_5319 = const()[name = string("op_5319"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_45_cast_fp16 = reshape(shape = var_5319, x = var_5314_cast_fp16)[name = string("P_45_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5321_cast_fp16 = exp(x = P_45_cast_fp16)[name = string("op_5321_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_45_cast_fp16 = mul(x = var_5321_cast_fp16, y = tril_incl_to_fp16)[name = string("D_45_cast_fp16")];
            bool kk_45_transpose_x_1 = const()[name = string("kk_45_transpose_x_1"), val = bool(false)];
            bool kk_45_transpose_y_1 = const()[name = string("kk_45_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_45_cast_fp16 = matmul(transpose_x = kk_45_transpose_x_1, transpose_y = kk_45_transpose_y_1, x = kb_45_cast_fp16, y = k_81_cast_fp16)[name = string("kk_45_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_5327_cast_fp16 = mul(x = kk_45_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_5327_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5328_cast_fp16 = mul(x = var_5327_cast_fp16, y = D_45_cast_fp16)[name = string("op_5328_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_45_cast_fp16 = mul(x = var_5328_cast_fp16, y = cum_tail_to_fp16)[name = string("N_45_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5330_cast_fp16 = mul(x = D_45_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5330_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_45_cast_fp16 = mul(x = kk_45_cast_fp16, y = var_5330_cast_fp16)[name = string("Na_45_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_177_cast_fp16 = add(x = eye_to_fp16, y = N_45_cast_fp16)[name = string("Minv_177_cast_fp16")];
            bool Np_133_transpose_x_0 = const()[name = string("Np_133_transpose_x_0"), val = bool(false)];
            bool Np_133_transpose_y_0 = const()[name = string("Np_133_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_133_cast_fp16 = matmul(transpose_x = Np_133_transpose_x_0, transpose_y = Np_133_transpose_y_0, x = N_45_cast_fp16, y = Na_45_cast_fp16)[name = string("Np_133_cast_fp16")];
            bool Npa_89_transpose_x_0 = const()[name = string("Npa_89_transpose_x_0"), val = bool(false)];
            bool Npa_89_transpose_y_0 = const()[name = string("Npa_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_89_cast_fp16 = matmul(transpose_x = Npa_89_transpose_x_0, transpose_y = Npa_89_transpose_y_0, x = Na_45_cast_fp16, y = N_45_cast_fp16)[name = string("Npa_89_cast_fp16")];
            bool var_5336_transpose_x_0 = const()[name = string("op_5336_transpose_x_0"), val = bool(false)];
            bool var_5336_transpose_y_0 = const()[name = string("op_5336_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5336_cast_fp16 = matmul(transpose_x = var_5336_transpose_x_0, transpose_y = var_5336_transpose_y_0, x = Minv_177_cast_fp16, y = Np_133_cast_fp16)[name = string("op_5336_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_179_cast_fp16 = add(x = Minv_177_cast_fp16, y = var_5336_cast_fp16)[name = string("Minv_179_cast_fp16")];
            bool Np_135_transpose_x_0 = const()[name = string("Np_135_transpose_x_0"), val = bool(false)];
            bool Np_135_transpose_y_0 = const()[name = string("Np_135_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_135_cast_fp16 = matmul(transpose_x = Np_135_transpose_x_0, transpose_y = Np_135_transpose_y_0, x = Np_133_cast_fp16, y = Npa_89_cast_fp16)[name = string("Np_135_cast_fp16")];
            bool Npa_91_transpose_x_0 = const()[name = string("Npa_91_transpose_x_0"), val = bool(false)];
            bool Npa_91_transpose_y_0 = const()[name = string("Npa_91_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_91_cast_fp16 = matmul(transpose_x = Npa_91_transpose_x_0, transpose_y = Npa_91_transpose_y_0, x = Npa_89_cast_fp16, y = Np_133_cast_fp16)[name = string("Npa_91_cast_fp16")];
            bool var_5341_transpose_x_0 = const()[name = string("op_5341_transpose_x_0"), val = bool(false)];
            bool var_5341_transpose_y_0 = const()[name = string("op_5341_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5341_cast_fp16 = matmul(transpose_x = var_5341_transpose_x_0, transpose_y = var_5341_transpose_y_0, x = Minv_179_cast_fp16, y = Np_135_cast_fp16)[name = string("op_5341_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_181_cast_fp16 = add(x = Minv_179_cast_fp16, y = var_5341_cast_fp16)[name = string("Minv_181_cast_fp16")];
            bool Np_137_transpose_x_0 = const()[name = string("Np_137_transpose_x_0"), val = bool(false)];
            bool Np_137_transpose_y_0 = const()[name = string("Np_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_137_cast_fp16 = matmul(transpose_x = Np_137_transpose_x_0, transpose_y = Np_137_transpose_y_0, x = Np_135_cast_fp16, y = Npa_91_cast_fp16)[name = string("Np_137_cast_fp16")];
            bool var_5345_transpose_x_0 = const()[name = string("op_5345_transpose_x_0"), val = bool(false)];
            bool var_5345_transpose_y_0 = const()[name = string("op_5345_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5345_cast_fp16 = matmul(transpose_x = var_5345_transpose_x_0, transpose_y = var_5345_transpose_y_0, x = Minv_181_cast_fp16, y = Np_137_cast_fp16)[name = string("op_5345_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_183_cast_fp16 = add(x = Minv_181_cast_fp16, y = var_5345_cast_fp16)[name = string("Minv_183_cast_fp16")];
            bool u_45_transpose_x_0 = const()[name = string("u_45_transpose_x_0"), val = bool(false)];
            bool u_45_transpose_y_0 = const()[name = string("u_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_45_cast_fp16 = matmul(transpose_x = u_45_transpose_x_0, transpose_y = u_45_transpose_y_0, x = Minv_183_cast_fp16, y = vb_53_cast_fp16)[name = string("u_45_cast_fp16")];
            bool var_5352_transpose_x_1 = const()[name = string("op_5352_transpose_x_1"), val = bool(false)];
            bool var_5352_transpose_y_1 = const()[name = string("op_5352_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_5352_cast_fp16 = matmul(transpose_x = var_5352_transpose_x_1, transpose_y = var_5352_transpose_y_1, x = q_79_cast_fp16, y = k_81_cast_fp16)[name = string("op_5352_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_45_cast_fp16 = mul(x = var_5352_cast_fp16, y = D_45_cast_fp16)[name = string("A_45_cast_fp16")];
            tensor<fp16, [16]> var_5354_bias_0_to_fp16 = const()[name = string("op_5354_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_5354_cast_fp16 = linear(bias = var_5354_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_53_cast_fp16)[name = string("op_5354_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_37_cast_fp16 = exp(x = var_5354_cast_fp16)[name = string("cum_37_cast_fp16")];
            tensor<int32, [1]> var_5357_axes_0 = const()[name = string("op_5357_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_5357_cast_fp16 = expand_dims(axes = var_5357_axes_0, x = cum_37_cast_fp16)[name = string("op_5357_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5358_cast_fp16 = mul(x = kb_45_cast_fp16, y = var_5357_cast_fp16)[name = string("op_5358_cast_fp16")];
            bool w_45_transpose_x_0 = const()[name = string("w_45_transpose_x_0"), val = bool(false)];
            bool w_45_transpose_y_0 = const()[name = string("w_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_45_cast_fp16 = matmul(transpose_x = w_45_transpose_x_0, transpose_y = w_45_transpose_y_0, x = Minv_183_cast_fp16, y = var_5358_cast_fp16)[name = string("w_45_cast_fp16")];
            bool var_5360_transpose_x_0 = const()[name = string("op_5360_transpose_x_0"), val = bool(false)];
            bool var_5360_transpose_y_0 = const()[name = string("op_5360_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5360_cast_fp16 = matmul(transpose_x = var_5360_transpose_x_0, transpose_y = var_5360_transpose_y_0, x = w_45_cast_fp16, y = S_37_cast_fp16)[name = string("op_5360_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_37_cast_fp16 = sub(x = u_45_cast_fp16, y = var_5360_cast_fp16)[name = string("vnew_37_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5365_cast_fp16 = mul(x = q_79_cast_fp16, y = var_5357_cast_fp16)[name = string("op_5365_cast_fp16")];
            bool var_5366_transpose_x_0 = const()[name = string("op_5366_transpose_x_0"), val = bool(false)];
            bool var_5366_transpose_y_0 = const()[name = string("op_5366_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5366_cast_fp16 = matmul(transpose_x = var_5366_transpose_x_0, transpose_y = var_5366_transpose_y_0, x = var_5365_cast_fp16, y = S_37_cast_fp16)[name = string("op_5366_cast_fp16")];
            bool var_5367_transpose_x_0 = const()[name = string("op_5367_transpose_x_0"), val = bool(false)];
            bool var_5367_transpose_y_0 = const()[name = string("op_5367_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5367_cast_fp16 = matmul(transpose_x = var_5367_transpose_x_0, transpose_y = var_5367_transpose_y_0, x = A_45_cast_fp16, y = vnew_37_cast_fp16)[name = string("op_5367_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5369_cast_fp16 = add(x = var_5366_cast_fp16, y = var_5367_cast_fp16)[name = string("op_5369_cast_fp16")];
            tensor<fp16, [16]> var_5370_bias_0_to_fp16 = const()[name = string("op_5370_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_5370_cast_fp16 = linear(bias = var_5370_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_53_cast_fp16)[name = string("op_5370_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_39_cast_fp16 = exp(x = var_5370_cast_fp16)[name = string("tail_39_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_77_to_fp16 = const()[name = string("transpose_77_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_5372_bias_0_to_fp16 = const()[name = string("op_5372_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_5372_cast_fp16 = linear(bias = var_5372_bias_0_to_fp16, weight = transpose_77_to_fp16, x = gdec_53_cast_fp16)[name = string("op_5372_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_5373_cast_fp16 = exp(x = var_5372_cast_fp16)[name = string("op_5373_cast_fp16")];
            tensor<int32, [1]> last_33_axes_0 = const()[name = string("last_33_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_33_cast_fp16 = expand_dims(axes = last_33_axes_0, x = var_5373_cast_fp16)[name = string("last_33_cast_fp16")];
            tensor<int32, [1]> var_5377_axes_0 = const()[name = string("op_5377_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_5377_cast_fp16 = expand_dims(axes = var_5377_axes_0, x = tail_39_cast_fp16)[name = string("op_5377_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5378_cast_fp16 = mul(x = k_81_cast_fp16, y = var_5377_cast_fp16)[name = string("op_5378_cast_fp16")];
            bool upd_31_transpose_x_1 = const()[name = string("upd_31_transpose_x_1"), val = bool(true)];
            bool upd_31_transpose_y_1 = const()[name = string("upd_31_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_31_cast_fp16 = matmul(transpose_x = upd_31_transpose_x_1, transpose_y = upd_31_transpose_y_1, x = var_5378_cast_fp16, y = vnew_37_cast_fp16)[name = string("upd_31_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_5383_cast_fp16 = mul(x = S_37_cast_fp16, y = last_33_cast_fp16)[name = string("op_5383_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_39_cast_fp16 = add(x = var_5383_cast_fp16, y = upd_31_cast_fp16)[name = string("S_39_cast_fp16")];
            tensor<int32, [4]> q_81_begin_0 = const()[name = string("q_81_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> q_81_end_0 = const()[name = string("q_81_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_81_end_mask_0 = const()[name = string("q_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_81_cast_fp16 = slice_by_index(begin = q_81_begin_0, end = q_81_end_0, end_mask = q_81_end_mask_0, x = q_69_cast_fp16)[name = string("q_81_cast_fp16")];
            tensor<int32, [4]> k_83_begin_0 = const()[name = string("k_83_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> k_83_end_0 = const()[name = string("k_83_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_83_end_mask_0 = const()[name = string("k_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_83_cast_fp16 = slice_by_index(begin = k_83_begin_0, end = k_83_end_0, end_mask = k_83_end_mask_0, x = k_71_cast_fp16)[name = string("k_83_cast_fp16")];
            tensor<int32, [4]> kb_47_begin_0 = const()[name = string("kb_47_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> kb_47_end_0 = const()[name = string("kb_47_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_47_end_mask_0 = const()[name = string("kb_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_47_cast_fp16 = slice_by_index(begin = kb_47_begin_0, end = kb_47_end_0, end_mask = kb_47_end_mask_0, x = var_4670_cast_fp16)[name = string("kb_47_cast_fp16")];
            tensor<int32, [4]> vb_55_begin_0 = const()[name = string("vb_55_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> vb_55_end_0 = const()[name = string("vb_55_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_55_end_mask_0 = const()[name = string("vb_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_55_cast_fp16 = slice_by_index(begin = vb_55_begin_0, end = vb_55_end_0, end_mask = vb_55_end_mask_0, x = vb_43_cast_fp16)[name = string("vb_55_cast_fp16")];
            tensor<int32, [3]> gdec_55_begin_0 = const()[name = string("gdec_55_begin_0"), val = tensor<int32, [3]>([0, 0, 80])];
            tensor<int32, [3]> gdec_55_end_0 = const()[name = string("gdec_55_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_55_end_mask_0 = const()[name = string("gdec_55_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_55_cast_fp16 = slice_by_index(begin = gdec_55_begin_0, end = gdec_55_end_0, end_mask = gdec_55_end_mask_0, x = gdec_43_cast_fp16)[name = string("gdec_55_cast_fp16")];
            tensor<int32, [2]> var_5466 = const()[name = string("op_5466"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_5467_cast_fp16 = reshape(shape = var_5466, x = gdec_55_cast_fp16)[name = string("op_5467_cast_fp16")];
            tensor<fp16, [4, 256]> var_5468_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5467_cast_fp16)[name = string("op_5468_cast_fp16")];
            tensor<int32, [4]> var_5473 = const()[name = string("op_5473"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_47_cast_fp16 = reshape(shape = var_5473, x = var_5468_cast_fp16)[name = string("P_47_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5475_cast_fp16 = exp(x = P_47_cast_fp16)[name = string("op_5475_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_47_cast_fp16 = mul(x = var_5475_cast_fp16, y = tril_incl_to_fp16)[name = string("D_47_cast_fp16")];
            bool kk_47_transpose_x_1 = const()[name = string("kk_47_transpose_x_1"), val = bool(false)];
            bool kk_47_transpose_y_1 = const()[name = string("kk_47_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_47_cast_fp16 = matmul(transpose_x = kk_47_transpose_x_1, transpose_y = kk_47_transpose_y_1, x = kb_47_cast_fp16, y = k_83_cast_fp16)[name = string("kk_47_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_5481_cast_fp16 = mul(x = kk_47_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_5481_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5482_cast_fp16 = mul(x = var_5481_cast_fp16, y = D_47_cast_fp16)[name = string("op_5482_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_47_cast_fp16 = mul(x = var_5482_cast_fp16, y = cum_tail_to_fp16)[name = string("N_47_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5484_cast_fp16 = mul(x = D_47_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5484_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_47_cast_fp16 = mul(x = kk_47_cast_fp16, y = var_5484_cast_fp16)[name = string("Na_47_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_185_cast_fp16 = add(x = eye_to_fp16, y = N_47_cast_fp16)[name = string("Minv_185_cast_fp16")];
            bool Np_139_transpose_x_0 = const()[name = string("Np_139_transpose_x_0"), val = bool(false)];
            bool Np_139_transpose_y_0 = const()[name = string("Np_139_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_139_cast_fp16 = matmul(transpose_x = Np_139_transpose_x_0, transpose_y = Np_139_transpose_y_0, x = N_47_cast_fp16, y = Na_47_cast_fp16)[name = string("Np_139_cast_fp16")];
            bool Npa_93_transpose_x_0 = const()[name = string("Npa_93_transpose_x_0"), val = bool(false)];
            bool Npa_93_transpose_y_0 = const()[name = string("Npa_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_93_cast_fp16 = matmul(transpose_x = Npa_93_transpose_x_0, transpose_y = Npa_93_transpose_y_0, x = Na_47_cast_fp16, y = N_47_cast_fp16)[name = string("Npa_93_cast_fp16")];
            bool var_5490_transpose_x_0 = const()[name = string("op_5490_transpose_x_0"), val = bool(false)];
            bool var_5490_transpose_y_0 = const()[name = string("op_5490_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5490_cast_fp16 = matmul(transpose_x = var_5490_transpose_x_0, transpose_y = var_5490_transpose_y_0, x = Minv_185_cast_fp16, y = Np_139_cast_fp16)[name = string("op_5490_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_187_cast_fp16 = add(x = Minv_185_cast_fp16, y = var_5490_cast_fp16)[name = string("Minv_187_cast_fp16")];
            bool Np_141_transpose_x_0 = const()[name = string("Np_141_transpose_x_0"), val = bool(false)];
            bool Np_141_transpose_y_0 = const()[name = string("Np_141_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_141_cast_fp16 = matmul(transpose_x = Np_141_transpose_x_0, transpose_y = Np_141_transpose_y_0, x = Np_139_cast_fp16, y = Npa_93_cast_fp16)[name = string("Np_141_cast_fp16")];
            bool Npa_95_transpose_x_0 = const()[name = string("Npa_95_transpose_x_0"), val = bool(false)];
            bool Npa_95_transpose_y_0 = const()[name = string("Npa_95_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_95_cast_fp16 = matmul(transpose_x = Npa_95_transpose_x_0, transpose_y = Npa_95_transpose_y_0, x = Npa_93_cast_fp16, y = Np_139_cast_fp16)[name = string("Npa_95_cast_fp16")];
            bool var_5495_transpose_x_0 = const()[name = string("op_5495_transpose_x_0"), val = bool(false)];
            bool var_5495_transpose_y_0 = const()[name = string("op_5495_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5495_cast_fp16 = matmul(transpose_x = var_5495_transpose_x_0, transpose_y = var_5495_transpose_y_0, x = Minv_187_cast_fp16, y = Np_141_cast_fp16)[name = string("op_5495_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_189_cast_fp16 = add(x = Minv_187_cast_fp16, y = var_5495_cast_fp16)[name = string("Minv_189_cast_fp16")];
            bool Np_143_transpose_x_0 = const()[name = string("Np_143_transpose_x_0"), val = bool(false)];
            bool Np_143_transpose_y_0 = const()[name = string("Np_143_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_143_cast_fp16 = matmul(transpose_x = Np_143_transpose_x_0, transpose_y = Np_143_transpose_y_0, x = Np_141_cast_fp16, y = Npa_95_cast_fp16)[name = string("Np_143_cast_fp16")];
            bool var_5499_transpose_x_0 = const()[name = string("op_5499_transpose_x_0"), val = bool(false)];
            bool var_5499_transpose_y_0 = const()[name = string("op_5499_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_5499_cast_fp16 = matmul(transpose_x = var_5499_transpose_x_0, transpose_y = var_5499_transpose_y_0, x = Minv_189_cast_fp16, y = Np_143_cast_fp16)[name = string("op_5499_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_191_cast_fp16 = add(x = Minv_189_cast_fp16, y = var_5499_cast_fp16)[name = string("Minv_191_cast_fp16")];
            bool u_47_transpose_x_0 = const()[name = string("u_47_transpose_x_0"), val = bool(false)];
            bool u_47_transpose_y_0 = const()[name = string("u_47_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_47_cast_fp16 = matmul(transpose_x = u_47_transpose_x_0, transpose_y = u_47_transpose_y_0, x = Minv_191_cast_fp16, y = vb_55_cast_fp16)[name = string("u_47_cast_fp16")];
            bool var_5506_transpose_x_1 = const()[name = string("op_5506_transpose_x_1"), val = bool(false)];
            bool var_5506_transpose_y_1 = const()[name = string("op_5506_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_5506_cast_fp16 = matmul(transpose_x = var_5506_transpose_x_1, transpose_y = var_5506_transpose_y_1, x = q_81_cast_fp16, y = k_83_cast_fp16)[name = string("op_5506_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_47_cast_fp16 = mul(x = var_5506_cast_fp16, y = D_47_cast_fp16)[name = string("A_47_cast_fp16")];
            tensor<fp16, [16]> var_5508_bias_0_to_fp16 = const()[name = string("op_5508_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_5508_cast_fp16 = linear(bias = var_5508_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_55_cast_fp16)[name = string("op_5508_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_39_cast_fp16 = exp(x = var_5508_cast_fp16)[name = string("cum_39_cast_fp16")];
            tensor<int32, [1]> var_5511_axes_0 = const()[name = string("op_5511_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_5511_cast_fp16 = expand_dims(axes = var_5511_axes_0, x = cum_39_cast_fp16)[name = string("op_5511_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5512_cast_fp16 = mul(x = kb_47_cast_fp16, y = var_5511_cast_fp16)[name = string("op_5512_cast_fp16")];
            bool w_47_transpose_x_0 = const()[name = string("w_47_transpose_x_0"), val = bool(false)];
            bool w_47_transpose_y_0 = const()[name = string("w_47_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_47_cast_fp16 = matmul(transpose_x = w_47_transpose_x_0, transpose_y = w_47_transpose_y_0, x = Minv_191_cast_fp16, y = var_5512_cast_fp16)[name = string("w_47_cast_fp16")];
            bool var_5514_transpose_x_0 = const()[name = string("op_5514_transpose_x_0"), val = bool(false)];
            bool var_5514_transpose_y_0 = const()[name = string("op_5514_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5514_cast_fp16 = matmul(transpose_x = var_5514_transpose_x_0, transpose_y = var_5514_transpose_y_0, x = w_47_cast_fp16, y = S_39_cast_fp16)[name = string("op_5514_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_39_cast_fp16 = sub(x = u_47_cast_fp16, y = var_5514_cast_fp16)[name = string("vnew_39_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_5519_cast_fp16 = mul(x = q_81_cast_fp16, y = var_5511_cast_fp16)[name = string("op_5519_cast_fp16")];
            bool var_5520_transpose_x_0 = const()[name = string("op_5520_transpose_x_0"), val = bool(false)];
            bool var_5520_transpose_y_0 = const()[name = string("op_5520_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5520_cast_fp16 = matmul(transpose_x = var_5520_transpose_x_0, transpose_y = var_5520_transpose_y_0, x = var_5519_cast_fp16, y = S_39_cast_fp16)[name = string("op_5520_cast_fp16")];
            bool var_5521_transpose_x_0 = const()[name = string("op_5521_transpose_x_0"), val = bool(false)];
            bool var_5521_transpose_y_0 = const()[name = string("op_5521_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_5521_cast_fp16 = matmul(transpose_x = var_5521_transpose_x_0, transpose_y = var_5521_transpose_y_0, x = A_47_cast_fp16, y = vnew_39_cast_fp16)[name = string("op_5521_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_7_cast_fp16 = add(x = var_5520_cast_fp16, y = var_5521_cast_fp16)[name = string("oc_7_cast_fp16")];
            int32 var_5525 = const()[name = string("op_5525"), val = int32(2)];
            bool o_21_interleave_0 = const()[name = string("o_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 96, 128]> o_21_cast_fp16 = concat(axis = var_5525, interleave = o_21_interleave_0, values = (var_4760_cast_fp16, var_4907_cast_fp16, var_5061_cast_fp16, var_5215_cast_fp16, var_5369_cast_fp16, oc_7_cast_fp16))[name = string("o_21_cast_fp16")];
            tensor<int32, [4]> var_5531 = const()[name = string("op_5531"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_5537 = const()[name = string("op_5537"), val = tensor<int32, [4]>([1, 96, 4, 128])];
            tensor<fp16, [1, 96, 4, 128]> input_65_cast_fp16 = reshape(shape = var_5537, x = z_7_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> o_23_cast_fp16 = transpose(perm = var_5531, x = o_21_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 96, 4, 128]> var_5539_cast_fp16 = mul(x = o_23_cast_fp16, y = o_23_cast_fp16)[name = string("op_5539_cast_fp16")];
            tensor<int32, [1]> var_5544_axes_0 = const()[name = string("op_5544_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5544_keep_dims_0 = const()[name = string("op_5544_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 4, 1]> var_5544_cast_fp16 = reduce_mean(axes = var_5544_axes_0, keep_dims = var_5544_keep_dims_0, x = var_5539_cast_fp16)[name = string("op_5544_cast_fp16")];
            fp16 var_5546_to_fp16 = const()[name = string("op_5546_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 96, 4, 1]> var_5547_cast_fp16 = add(x = var_5544_cast_fp16, y = var_5546_to_fp16)[name = string("op_5547_cast_fp16")];
            fp32 var_5548_epsilon_0 = const()[name = string("op_5548_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 4, 1]> var_5548_cast_fp16 = rsqrt(epsilon = var_5548_epsilon_0, x = var_5547_cast_fp16)[name = string("op_5548_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> var_5549_cast_fp16 = mul(x = o_23_cast_fp16, y = var_5548_cast_fp16)[name = string("op_5549_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> var_5551_cast_fp16 = silu(x = input_65_cast_fp16)[name = string("op_5551_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> o_25_cast_fp16 = mul(x = var_5549_cast_fp16, y = var_5551_cast_fp16)[name = string("o_25_cast_fp16")];
            tensor<int32, [3]> var_5556 = const()[name = string("op_5556"), val = tensor<int32, [3]>([1, 96, 512])];
            tensor<fp16, [1, 96, 512]> input_67_cast_fp16 = reshape(shape = var_5556, x = o_25_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [640, 512]> layers_4_out_weight_to_fp16 = const()[name = string("layers_4_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72063040)))];
            tensor<fp16, [1, 96, 640]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_out_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 96, 640]> x_49_cast_fp16 = add(x = x_45_cast_fp16, y = linear_17_cast_fp16)[name = string("x_49_cast_fp16")];
            fp16 var_5569_to_fp16 = const()[name = string("op_5569_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 96, 640]> x_51_cast_fp16 = mul(x = x_49_cast_fp16, y = var_5569_to_fp16)[name = string("x_51_cast_fp16")];
            fp16 var_5567_promoted_to_fp16 = const()[name = string("op_5567_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> var_5571_cast_fp16 = pow(x = x_51_cast_fp16, y = var_5567_promoted_to_fp16)[name = string("op_5571_cast_fp16")];
            tensor<int32, [1]> var_5573_axes_0 = const()[name = string("op_5573_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5573_keep_dims_0 = const()[name = string("op_5573_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 1]> var_5573_cast_fp16 = reduce_mean(axes = var_5573_axes_0, keep_dims = var_5573_keep_dims_0, x = var_5571_cast_fp16)[name = string("op_5573_cast_fp16")];
            fp16 var_5574_to_fp16 = const()[name = string("op_5574_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 96, 1]> var_5575_cast_fp16 = add(x = var_5573_cast_fp16, y = var_5574_to_fp16)[name = string("op_5575_cast_fp16")];
            fp32 var_5576_epsilon_0 = const()[name = string("op_5576_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 1]> var_5576_cast_fp16 = rsqrt(epsilon = var_5576_epsilon_0, x = var_5575_cast_fp16)[name = string("op_5576_cast_fp16")];
            tensor<fp16, [1, 96, 640]> var_5577_cast_fp16 = mul(x = x_51_cast_fp16, y = var_5576_cast_fp16)[name = string("op_5577_cast_fp16")];
            tensor<fp16, [640]> layers_4_n2_w1_to_fp16 = const()[name = string("layers_4_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72718464)))];
            tensor<fp16, [1, 96, 640]> input_69_cast_fp16 = mul(x = var_5577_cast_fp16, y = layers_4_n2_w1_to_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_4_gate_up_weight_to_fp16 = const()[name = string("layers_4_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72719808)))];
            tensor<fp16, [1, 96, 5120]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_4_gate_up_weight_to_fp16, x = input_69_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<int32, [3]> input_71_begin_0 = const()[name = string("input_71_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_71_end_0 = const()[name = string("input_71_end_0"), val = tensor<int32, [3]>([1, 96, 2560])];
            tensor<bool, [3]> input_71_end_mask_0 = const()[name = string("input_71_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 2560]> input_71_cast_fp16 = slice_by_index(begin = input_71_begin_0, end = input_71_end_0, end_mask = input_71_end_mask_0, x = linear_18_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [1, 96, 2560]> var_5587_cast_fp16 = silu(x = input_71_cast_fp16)[name = string("op_5587_cast_fp16")];
            tensor<int32, [3]> var_5592_begin_0 = const()[name = string("op_5592_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_5592_end_0 = const()[name = string("op_5592_end_0"), val = tensor<int32, [3]>([1, 96, 5120])];
            tensor<bool, [3]> var_5592_end_mask_0 = const()[name = string("op_5592_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 2560]> var_5592_cast_fp16 = slice_by_index(begin = var_5592_begin_0, end = var_5592_end_0, end_mask = var_5592_end_mask_0, x = linear_18_cast_fp16)[name = string("op_5592_cast_fp16")];
            tensor<fp16, [1, 96, 2560]> input_73_cast_fp16 = mul(x = var_5587_cast_fp16, y = var_5592_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_4_down_weight_to_fp16 = const()[name = string("layers_4_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79273472)))];
            tensor<fp16, [1, 96, 640]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_down_weight_to_fp16, x = input_73_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 96, 640]> x_53_cast_fp16 = add(x = x_49_cast_fp16, y = linear_19_cast_fp16)[name = string("x_53_cast_fp16")];
            fp16 var_5605_to_fp16 = const()[name = string("op_5605_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 96, 640]> x_55_cast_fp16 = mul(x = x_53_cast_fp16, y = var_5605_to_fp16)[name = string("x_55_cast_fp16")];
            fp16 var_5603_promoted_to_fp16 = const()[name = string("op_5603_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> var_5607_cast_fp16 = pow(x = x_55_cast_fp16, y = var_5603_promoted_to_fp16)[name = string("op_5607_cast_fp16")];
            tensor<int32, [1]> var_5609_axes_0 = const()[name = string("op_5609_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5609_keep_dims_0 = const()[name = string("op_5609_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 1]> var_5609_cast_fp16 = reduce_mean(axes = var_5609_axes_0, keep_dims = var_5609_keep_dims_0, x = var_5607_cast_fp16)[name = string("op_5609_cast_fp16")];
            fp16 var_5610_to_fp16 = const()[name = string("op_5610_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 96, 1]> var_5611_cast_fp16 = add(x = var_5609_cast_fp16, y = var_5610_to_fp16)[name = string("op_5611_cast_fp16")];
            fp32 var_5612_epsilon_0 = const()[name = string("op_5612_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 1]> var_5612_cast_fp16 = rsqrt(epsilon = var_5612_epsilon_0, x = var_5611_cast_fp16)[name = string("op_5612_cast_fp16")];
            tensor<fp16, [1, 96, 640]> var_5613_cast_fp16 = mul(x = x_55_cast_fp16, y = var_5612_cast_fp16)[name = string("op_5613_cast_fp16")];
            tensor<fp16, [640]> layers_5_n1_w1_to_fp16 = const()[name = string("layers_5_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82550336)))];
            tensor<fp16, [1, 96, 640]> input_75_cast_fp16 = mul(x = var_5613_cast_fp16, y = layers_5_n1_w1_to_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_5_qkvzba_weight_to_fp16 = const()[name = string("layers_5_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82551680)))];
            tensor<fp16, [1, 96, 1544]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_qkvzba_weight_to_fp16, x = input_75_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [3]> qkv_9_begin_0 = const()[name = string("qkv_9_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_9_end_0 = const()[name = string("qkv_9_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> qkv_9_end_mask_0 = const()[name = string("qkv_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 1024]> qkv_9_cast_fp16 = slice_by_index(begin = qkv_9_begin_0, end = qkv_9_end_0, end_mask = qkv_9_end_mask_0, x = linear_20_cast_fp16)[name = string("qkv_9_cast_fp16")];
            tensor<int32, [3]> z_9_begin_0 = const()[name = string("z_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_9_end_0 = const()[name = string("z_9_end_0"), val = tensor<int32, [3]>([1, 96, 1536])];
            tensor<bool, [3]> z_9_end_mask_0 = const()[name = string("z_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 512]> z_9_cast_fp16 = slice_by_index(begin = z_9_begin_0, end = z_9_end_0, end_mask = z_9_end_mask_0, x = linear_20_cast_fp16)[name = string("z_9_cast_fp16")];
            tensor<int32, [3]> b_9_begin_0 = const()[name = string("b_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_9_end_0 = const()[name = string("b_9_end_0"), val = tensor<int32, [3]>([1, 96, 1540])];
            tensor<bool, [3]> b_9_end_mask_0 = const()[name = string("b_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 4]> b_9_cast_fp16 = slice_by_index(begin = b_9_begin_0, end = b_9_end_0, end_mask = b_9_end_mask_0, x = linear_20_cast_fp16)[name = string("b_9_cast_fp16")];
            tensor<int32, [3]> a_9_begin_0 = const()[name = string("a_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_9_end_0 = const()[name = string("a_9_end_0"), val = tensor<int32, [3]>([1, 96, 1])];
            tensor<bool, [3]> a_9_end_mask_0 = const()[name = string("a_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 4]> a_9_cast_fp16 = slice_by_index(begin = a_9_begin_0, end = a_9_end_0, end_mask = a_9_end_mask_0, x = linear_20_cast_fp16)[name = string("a_9_cast_fp16")];
            int32 var_5652 = const()[name = string("op_5652"), val = int32(1)];
            bool xp_9_interleave_0 = const()[name = string("xp_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 99, 1024]> xp_9_cast_fp16 = concat(axis = var_5652, interleave = xp_9_interleave_0, values = (zpad_to_fp16, qkv_9_cast_fp16))[name = string("xp_9_cast_fp16")];
            tensor<int32, [3]> var_5663_begin_0 = const()[name = string("op_5663_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_5663_end_0 = const()[name = string("op_5663_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_5663_end_mask_0 = const()[name = string("op_5663_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 1024]> var_5663_cast_fp16 = slice_by_index(begin = var_5663_begin_0, end = var_5663_end_0, end_mask = var_5663_end_mask_0, x = xp_9_cast_fp16)[name = string("op_5663_cast_fp16")];
            tensor<fp16, [1024]> var_5671_to_fp16 = const()[name = string("op_5671_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84528064)))];
            tensor<fp16, [1, 96, 1024]> var_5672_cast_fp16 = mul(x = var_5663_cast_fp16, y = var_5671_to_fp16)[name = string("op_5672_cast_fp16")];
            tensor<int32, [3]> var_5682_begin_0 = const()[name = string("op_5682_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_5682_end_0 = const()[name = string("op_5682_end_0"), val = tensor<int32, [3]>([1, 98, 1024])];
            tensor<bool, [3]> var_5682_end_mask_0 = const()[name = string("op_5682_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_5682_cast_fp16 = slice_by_index(begin = var_5682_begin_0, end = var_5682_end_0, end_mask = var_5682_end_mask_0, x = xp_9_cast_fp16)[name = string("op_5682_cast_fp16")];
            tensor<fp16, [1024]> var_5690_to_fp16 = const()[name = string("op_5690_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84530176)))];
            tensor<fp16, [1, 96, 1024]> var_5691_cast_fp16 = mul(x = var_5682_cast_fp16, y = var_5690_to_fp16)[name = string("op_5691_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> var_5693_cast_fp16 = add(x = var_5672_cast_fp16, y = var_5691_cast_fp16)[name = string("op_5693_cast_fp16")];
            tensor<int32, [3]> var_5703_begin_0 = const()[name = string("op_5703_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_5703_end_0 = const()[name = string("op_5703_end_0"), val = tensor<int32, [3]>([1, 97, 1024])];
            tensor<bool, [3]> var_5703_end_mask_0 = const()[name = string("op_5703_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_5703_cast_fp16 = slice_by_index(begin = var_5703_begin_0, end = var_5703_end_0, end_mask = var_5703_end_mask_0, x = xp_9_cast_fp16)[name = string("op_5703_cast_fp16")];
            tensor<fp16, [1024]> var_5711_to_fp16 = const()[name = string("op_5711_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84532288)))];
            tensor<fp16, [1, 96, 1024]> var_5712_cast_fp16 = mul(x = var_5703_cast_fp16, y = var_5711_to_fp16)[name = string("op_5712_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> var_5714_cast_fp16 = add(x = var_5693_cast_fp16, y = var_5712_cast_fp16)[name = string("op_5714_cast_fp16")];
            tensor<int32, [3]> var_5724_begin_0 = const()[name = string("op_5724_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_5724_end_0 = const()[name = string("op_5724_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> var_5724_end_mask_0 = const()[name = string("op_5724_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_5724_cast_fp16 = slice_by_index(begin = var_5724_begin_0, end = var_5724_end_0, end_mask = var_5724_end_mask_0, x = xp_9_cast_fp16)[name = string("op_5724_cast_fp16")];
            tensor<fp16, [1024]> var_5732_to_fp16 = const()[name = string("op_5732_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84534400)))];
            tensor<fp16, [1, 96, 1024]> var_5733_cast_fp16 = mul(x = var_5724_cast_fp16, y = var_5732_to_fp16)[name = string("op_5733_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> input_77_cast_fp16 = add(x = var_5714_cast_fp16, y = var_5733_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> c_9_cast_fp16 = silu(x = input_77_cast_fp16)[name = string("c_9_cast_fp16")];
            tensor<int32, [3]> var_5741_begin_0 = const()[name = string("op_5741_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_5741_end_0 = const()[name = string("op_5741_end_0"), val = tensor<int32, [3]>([1, 96, 256])];
            tensor<bool, [3]> var_5741_end_mask_0 = const()[name = string("op_5741_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 256]> var_5741_cast_fp16 = slice_by_index(begin = var_5741_begin_0, end = var_5741_end_0, end_mask = var_5741_end_mask_0, x = c_9_cast_fp16)[name = string("op_5741_cast_fp16")];
            tensor<int32, [4]> var_5746 = const()[name = string("op_5746"), val = tensor<int32, [4]>([1, 96, 2, 128])];
            tensor<fp16, [1, 96, 2, 128]> q_83_cast_fp16 = reshape(shape = var_5746, x = var_5741_cast_fp16)[name = string("q_83_cast_fp16")];
            tensor<int32, [3]> var_5752_begin_0 = const()[name = string("op_5752_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_5752_end_0 = const()[name = string("op_5752_end_0"), val = tensor<int32, [3]>([1, 96, 512])];
            tensor<bool, [3]> var_5752_end_mask_0 = const()[name = string("op_5752_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 256]> var_5752_cast_fp16 = slice_by_index(begin = var_5752_begin_0, end = var_5752_end_0, end_mask = var_5752_end_mask_0, x = c_9_cast_fp16)[name = string("op_5752_cast_fp16")];
            tensor<int32, [4]> var_5757 = const()[name = string("op_5757"), val = tensor<int32, [4]>([1, 96, 2, 128])];
            tensor<fp16, [1, 96, 2, 128]> k_85_cast_fp16 = reshape(shape = var_5757, x = var_5752_cast_fp16)[name = string("k_85_cast_fp16")];
            tensor<int32, [3]> var_5763_begin_0 = const()[name = string("op_5763_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_5763_end_0 = const()[name = string("op_5763_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> var_5763_end_mask_0 = const()[name = string("op_5763_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 512]> var_5763_cast_fp16 = slice_by_index(begin = var_5763_begin_0, end = var_5763_end_0, end_mask = var_5763_end_mask_0, x = c_9_cast_fp16)[name = string("op_5763_cast_fp16")];
            tensor<int32, [4]> var_5768 = const()[name = string("op_5768"), val = tensor<int32, [4]>([1, 96, 4, 128])];
            tensor<fp16, [1, 96, 4, 128]> var_5769_cast_fp16 = reshape(shape = var_5768, x = var_5763_cast_fp16)[name = string("op_5769_cast_fp16")];
            tensor<int32, [4]> var_5774 = const()[name = string("op_5774"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5776_to_fp16 = const()[name = string("op_5776_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 96, 2, 128]> q_85_cast_fp16 = mul(x = q_83_cast_fp16, y = var_5776_to_fp16)[name = string("q_85_cast_fp16")];
            fp16 var_5778_to_fp16 = const()[name = string("op_5778_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 96, 2, 128]> k_87_cast_fp16 = mul(x = k_85_cast_fp16, y = var_5778_to_fp16)[name = string("k_87_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_5780_cast_fp16 = mul(x = q_85_cast_fp16, y = q_85_cast_fp16)[name = string("op_5780_cast_fp16")];
            tensor<int32, [1]> var_5785_axes_0 = const()[name = string("op_5785_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5785_keep_dims_0 = const()[name = string("op_5785_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 2, 1]> var_5785_cast_fp16 = reduce_mean(axes = var_5785_axes_0, keep_dims = var_5785_keep_dims_0, x = var_5780_cast_fp16)[name = string("op_5785_cast_fp16")];
            fp16 var_5786_promoted_to_fp16 = const()[name = string("op_5786_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 96, 2, 1]> var_5787_cast_fp16 = mul(x = var_5785_cast_fp16, y = var_5786_promoted_to_fp16)[name = string("op_5787_cast_fp16")];
            fp16 var_5789_to_fp16 = const()[name = string("op_5789_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 96, 2, 1]> var_5790_cast_fp16 = add(x = var_5787_cast_fp16, y = var_5789_to_fp16)[name = string("op_5790_cast_fp16")];
            fp32 var_5791_epsilon_0 = const()[name = string("op_5791_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 2, 1]> var_5791_cast_fp16 = rsqrt(epsilon = var_5791_epsilon_0, x = var_5790_cast_fp16)[name = string("op_5791_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_5792_cast_fp16 = mul(x = q_85_cast_fp16, y = var_5791_cast_fp16)[name = string("op_5792_cast_fp16")];
            fp16 var_5793_to_fp16 = const()[name = string("op_5793_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 96, 2, 128]> q_87_cast_fp16 = mul(x = var_5792_cast_fp16, y = var_5793_to_fp16)[name = string("q_87_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_5795_cast_fp16 = mul(x = k_87_cast_fp16, y = k_87_cast_fp16)[name = string("op_5795_cast_fp16")];
            tensor<int32, [1]> var_5800_axes_0 = const()[name = string("op_5800_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5800_keep_dims_0 = const()[name = string("op_5800_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 2, 1]> var_5800_cast_fp16 = reduce_mean(axes = var_5800_axes_0, keep_dims = var_5800_keep_dims_0, x = var_5795_cast_fp16)[name = string("op_5800_cast_fp16")];
            fp16 var_5801_promoted_to_fp16 = const()[name = string("op_5801_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 96, 2, 1]> var_5802_cast_fp16 = mul(x = var_5800_cast_fp16, y = var_5801_promoted_to_fp16)[name = string("op_5802_cast_fp16")];
            fp16 var_5804_to_fp16 = const()[name = string("op_5804_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 96, 2, 1]> var_5805_cast_fp16 = add(x = var_5802_cast_fp16, y = var_5804_to_fp16)[name = string("op_5805_cast_fp16")];
            fp32 var_5806_epsilon_0 = const()[name = string("op_5806_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 2, 1]> var_5806_cast_fp16 = rsqrt(epsilon = var_5806_epsilon_0, x = var_5805_cast_fp16)[name = string("op_5806_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> k_89_cast_fp16 = mul(x = k_87_cast_fp16, y = var_5806_cast_fp16)[name = string("k_89_cast_fp16")];
            tensor<int32, [4]> var_5812 = const()[name = string("op_5812"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_5825_axes_0 = const()[name = string("op_5825_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 96, 128]> var_5813_cast_fp16 = transpose(perm = var_5812, x = q_87_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [1, 2, 1, 96, 128]> var_5825_cast_fp16 = expand_dims(axes = var_5825_axes_0, x = var_5813_cast_fp16)[name = string("op_5825_cast_fp16")];
            tensor<int32, [5]> var_5833_reps_0 = const()[name = string("op_5833_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 96, 128]> var_5833_cast_fp16 = tile(reps = var_5833_reps_0, x = var_5825_cast_fp16)[name = string("op_5833_cast_fp16")];
            tensor<int32, [4]> var_5838 = const()[name = string("op_5838"), val = tensor<int32, [4]>([1, 4, 96, 128])];
            tensor<fp16, [1, 4, 96, 128]> q_89_cast_fp16 = reshape(shape = var_5838, x = var_5833_cast_fp16)[name = string("q_89_cast_fp16")];
            tensor<int32, [4]> var_5844 = const()[name = string("op_5844"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_5857_axes_0 = const()[name = string("op_5857_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 96, 128]> var_5845_cast_fp16 = transpose(perm = var_5844, x = k_89_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [1, 2, 1, 96, 128]> var_5857_cast_fp16 = expand_dims(axes = var_5857_axes_0, x = var_5845_cast_fp16)[name = string("op_5857_cast_fp16")];
            tensor<int32, [5]> var_5865_reps_0 = const()[name = string("op_5865_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 96, 128]> var_5865_cast_fp16 = tile(reps = var_5865_reps_0, x = var_5857_cast_fp16)[name = string("op_5865_cast_fp16")];
            tensor<int32, [4]> var_5870 = const()[name = string("op_5870"), val = tensor<int32, [4]>([1, 4, 96, 128])];
            tensor<fp16, [1, 4, 96, 128]> k_91_cast_fp16 = reshape(shape = var_5870, x = var_5865_cast_fp16)[name = string("k_91_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_5872_cast_fp16 = sigmoid(x = b_9_cast_fp16)[name = string("op_5872_cast_fp16")];
            tensor<int32, [3]> var_5876 = const()[name = string("op_5876"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_5_dt_bias_to_fp16 = const()[name = string("layers_5_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 96, 4]> input_79_cast_fp16 = add(x = a_9_cast_fp16, y = layers_5_dt_bias_to_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_5880_cast_fp16 = relu(x = input_79_cast_fp16)[name = string("op_5880_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_5881_cast_fp16 = abs(x = input_79_cast_fp16)[name = string("op_5881_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 96, 4]> var_5882_cast_fp16 = mul(x = var_5881_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_5882_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_5883_cast_fp16 = exp(x = var_5882_cast_fp16)[name = string("op_5883_cast_fp16")];
            fp16 var_5885_to_fp16 = const()[name = string("op_5885_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 96, 4]> var_5886_cast_fp16 = add(x = var_5883_cast_fp16, y = var_5885_to_fp16)[name = string("op_5886_cast_fp16")];
            fp32 var_5887_epsilon_0 = const()[name = string("op_5887_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 96, 4]> var_5887_cast_fp16 = log(epsilon = var_5887_epsilon_0, x = var_5886_cast_fp16)[name = string("op_5887_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_5889_cast_fp16 = add(x = var_5880_cast_fp16, y = var_5887_cast_fp16)[name = string("op_5889_cast_fp16")];
            tensor<fp16, [4]> layers_5_negA_to_fp16 = const()[name = string("layers_5_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.2f8p+3, -0x1.ae8p+2, -0x1.548p+2, -0x1.63p+3])];
            tensor<fp16, [1, 96, 4]> var_5890_cast_fp16 = mul(x = layers_5_negA_to_fp16, y = var_5889_cast_fp16)[name = string("op_5890_cast_fp16")];
            tensor<int32, [3]> var_5894 = const()[name = string("op_5894"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_5897_axes_0 = const()[name = string("op_5897_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 96]> beta_9_cast_fp16 = transpose(perm = var_5876, x = var_5872_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [1, 4, 96, 1]> var_5897_cast_fp16 = expand_dims(axes = var_5897_axes_0, x = beta_9_cast_fp16)[name = string("op_5897_cast_fp16")];
            fp16 var_5898_to_fp16 = const()[name = string("op_5898_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 96, 1]> var_5899_cast_fp16 = mul(x = var_5897_cast_fp16, y = var_5898_to_fp16)[name = string("op_5899_cast_fp16")];
            tensor<fp16, [1, 4, 96, 128]> v_13_cast_fp16 = transpose(perm = var_5774, x = var_5769_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [1, 4, 96, 128]> vb_57_cast_fp16 = mul(x = v_13_cast_fp16, y = var_5899_cast_fp16)[name = string("vb_57_cast_fp16")];
            tensor<int32, [4]> q_91_begin_0 = const()[name = string("q_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_91_end_0 = const()[name = string("q_91_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_91_end_mask_0 = const()[name = string("q_91_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, x = q_89_cast_fp16)[name = string("q_91_cast_fp16")];
            tensor<int32, [4]> k_93_begin_0 = const()[name = string("k_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_93_end_0 = const()[name = string("k_93_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_93_end_mask_0 = const()[name = string("k_93_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, x = k_91_cast_fp16)[name = string("k_93_cast_fp16")];
            tensor<fp16, [1, 4, 96, 128]> var_5933_cast_fp16 = mul(x = k_91_cast_fp16, y = var_5897_cast_fp16)[name = string("op_5933_cast_fp16")];
            tensor<int32, [4]> kb_49_begin_0 = const()[name = string("kb_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_49_end_0 = const()[name = string("kb_49_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_49_end_mask_0 = const()[name = string("kb_49_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_49_cast_fp16 = slice_by_index(begin = kb_49_begin_0, end = kb_49_end_0, end_mask = kb_49_end_mask_0, x = var_5933_cast_fp16)[name = string("kb_49_cast_fp16")];
            tensor<int32, [4]> vb_59_begin_0 = const()[name = string("vb_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_59_end_0 = const()[name = string("vb_59_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_59_end_mask_0 = const()[name = string("vb_59_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_59_cast_fp16 = slice_by_index(begin = vb_59_begin_0, end = vb_59_end_0, end_mask = vb_59_end_mask_0, x = vb_57_cast_fp16)[name = string("vb_59_cast_fp16")];
            tensor<int32, [3]> gdec_59_begin_0 = const()[name = string("gdec_59_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_59_end_0 = const()[name = string("gdec_59_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_59_end_mask_0 = const()[name = string("gdec_59_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 96]> gdec_57_cast_fp16 = transpose(perm = var_5894, x = var_5890_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 4, 16]> gdec_59_cast_fp16 = slice_by_index(begin = gdec_59_begin_0, end = gdec_59_end_0, end_mask = gdec_59_end_mask_0, x = gdec_57_cast_fp16)[name = string("gdec_59_cast_fp16")];
            tensor<int32, [2]> var_5981 = const()[name = string("op_5981"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_5982_cast_fp16 = reshape(shape = var_5981, x = gdec_59_cast_fp16)[name = string("op_5982_cast_fp16")];
            tensor<fp16, [4, 256]> var_5983_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5982_cast_fp16)[name = string("op_5983_cast_fp16")];
            tensor<int32, [4]> var_5988 = const()[name = string("op_5988"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_49_cast_fp16 = reshape(shape = var_5988, x = var_5983_cast_fp16)[name = string("P_49_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5990_cast_fp16 = exp(x = P_49_cast_fp16)[name = string("op_5990_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_49_cast_fp16 = mul(x = var_5990_cast_fp16, y = tril_incl_to_fp16)[name = string("D_49_cast_fp16")];
            bool kk_49_transpose_x_1 = const()[name = string("kk_49_transpose_x_1"), val = bool(false)];
            bool kk_49_transpose_y_1 = const()[name = string("kk_49_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_49_cast_fp16 = matmul(transpose_x = kk_49_transpose_x_1, transpose_y = kk_49_transpose_y_1, x = kb_49_cast_fp16, y = k_93_cast_fp16)[name = string("kk_49_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_5996_cast_fp16 = mul(x = kk_49_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_5996_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5997_cast_fp16 = mul(x = var_5996_cast_fp16, y = D_49_cast_fp16)[name = string("op_5997_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_49_cast_fp16 = mul(x = var_5997_cast_fp16, y = cum_tail_to_fp16)[name = string("N_49_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_5999_cast_fp16 = mul(x = D_49_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5999_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_49_cast_fp16 = mul(x = kk_49_cast_fp16, y = var_5999_cast_fp16)[name = string("Na_49_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_193_cast_fp16 = add(x = eye_to_fp16, y = N_49_cast_fp16)[name = string("Minv_193_cast_fp16")];
            bool Np_145_transpose_x_0 = const()[name = string("Np_145_transpose_x_0"), val = bool(false)];
            bool Np_145_transpose_y_0 = const()[name = string("Np_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_145_cast_fp16 = matmul(transpose_x = Np_145_transpose_x_0, transpose_y = Np_145_transpose_y_0, x = N_49_cast_fp16, y = Na_49_cast_fp16)[name = string("Np_145_cast_fp16")];
            bool Npa_97_transpose_x_0 = const()[name = string("Npa_97_transpose_x_0"), val = bool(false)];
            bool Npa_97_transpose_y_0 = const()[name = string("Npa_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_97_cast_fp16 = matmul(transpose_x = Npa_97_transpose_x_0, transpose_y = Npa_97_transpose_y_0, x = Na_49_cast_fp16, y = N_49_cast_fp16)[name = string("Npa_97_cast_fp16")];
            bool var_6005_transpose_x_0 = const()[name = string("op_6005_transpose_x_0"), val = bool(false)];
            bool var_6005_transpose_y_0 = const()[name = string("op_6005_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6005_cast_fp16 = matmul(transpose_x = var_6005_transpose_x_0, transpose_y = var_6005_transpose_y_0, x = Minv_193_cast_fp16, y = Np_145_cast_fp16)[name = string("op_6005_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_195_cast_fp16 = add(x = Minv_193_cast_fp16, y = var_6005_cast_fp16)[name = string("Minv_195_cast_fp16")];
            bool Np_147_transpose_x_0 = const()[name = string("Np_147_transpose_x_0"), val = bool(false)];
            bool Np_147_transpose_y_0 = const()[name = string("Np_147_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_147_cast_fp16 = matmul(transpose_x = Np_147_transpose_x_0, transpose_y = Np_147_transpose_y_0, x = Np_145_cast_fp16, y = Npa_97_cast_fp16)[name = string("Np_147_cast_fp16")];
            bool Npa_99_transpose_x_0 = const()[name = string("Npa_99_transpose_x_0"), val = bool(false)];
            bool Npa_99_transpose_y_0 = const()[name = string("Npa_99_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_99_cast_fp16 = matmul(transpose_x = Npa_99_transpose_x_0, transpose_y = Npa_99_transpose_y_0, x = Npa_97_cast_fp16, y = Np_145_cast_fp16)[name = string("Npa_99_cast_fp16")];
            bool var_6010_transpose_x_0 = const()[name = string("op_6010_transpose_x_0"), val = bool(false)];
            bool var_6010_transpose_y_0 = const()[name = string("op_6010_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6010_cast_fp16 = matmul(transpose_x = var_6010_transpose_x_0, transpose_y = var_6010_transpose_y_0, x = Minv_195_cast_fp16, y = Np_147_cast_fp16)[name = string("op_6010_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_197_cast_fp16 = add(x = Minv_195_cast_fp16, y = var_6010_cast_fp16)[name = string("Minv_197_cast_fp16")];
            bool Np_149_transpose_x_0 = const()[name = string("Np_149_transpose_x_0"), val = bool(false)];
            bool Np_149_transpose_y_0 = const()[name = string("Np_149_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_149_cast_fp16 = matmul(transpose_x = Np_149_transpose_x_0, transpose_y = Np_149_transpose_y_0, x = Np_147_cast_fp16, y = Npa_99_cast_fp16)[name = string("Np_149_cast_fp16")];
            bool var_6014_transpose_x_0 = const()[name = string("op_6014_transpose_x_0"), val = bool(false)];
            bool var_6014_transpose_y_0 = const()[name = string("op_6014_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6014_cast_fp16 = matmul(transpose_x = var_6014_transpose_x_0, transpose_y = var_6014_transpose_y_0, x = Minv_197_cast_fp16, y = Np_149_cast_fp16)[name = string("op_6014_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_199_cast_fp16 = add(x = Minv_197_cast_fp16, y = var_6014_cast_fp16)[name = string("Minv_199_cast_fp16")];
            bool u_49_transpose_x_0 = const()[name = string("u_49_transpose_x_0"), val = bool(false)];
            bool u_49_transpose_y_0 = const()[name = string("u_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_49_cast_fp16 = matmul(transpose_x = u_49_transpose_x_0, transpose_y = u_49_transpose_y_0, x = Minv_199_cast_fp16, y = vb_59_cast_fp16)[name = string("u_49_cast_fp16")];
            bool var_6021_transpose_x_1 = const()[name = string("op_6021_transpose_x_1"), val = bool(false)];
            bool var_6021_transpose_y_1 = const()[name = string("op_6021_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_6021_cast_fp16 = matmul(transpose_x = var_6021_transpose_x_1, transpose_y = var_6021_transpose_y_1, x = q_91_cast_fp16, y = k_93_cast_fp16)[name = string("op_6021_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_49_cast_fp16 = mul(x = var_6021_cast_fp16, y = D_49_cast_fp16)[name = string("A_49_cast_fp16")];
            bool var_6023_transpose_x_0 = const()[name = string("op_6023_transpose_x_0"), val = bool(false)];
            bool var_6023_transpose_y_0 = const()[name = string("op_6023_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6023_cast_fp16 = matmul(transpose_x = var_6023_transpose_x_0, transpose_y = var_6023_transpose_y_0, x = A_49_cast_fp16, y = u_49_cast_fp16)[name = string("op_6023_cast_fp16")];
            tensor<fp16, [16]> var_6024_bias_0_to_fp16 = const()[name = string("op_6024_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_6024_cast_fp16 = linear(bias = var_6024_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_59_cast_fp16)[name = string("op_6024_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_41_cast_fp16 = exp(x = var_6024_cast_fp16)[name = string("tail_41_cast_fp16")];
            tensor<int32, [1]> var_6027_axes_0 = const()[name = string("op_6027_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_6027_cast_fp16 = expand_dims(axes = var_6027_axes_0, x = tail_41_cast_fp16)[name = string("op_6027_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6028_cast_fp16 = mul(x = k_93_cast_fp16, y = var_6027_cast_fp16)[name = string("op_6028_cast_fp16")];
            bool S_41_transpose_x_1 = const()[name = string("S_41_transpose_x_1"), val = bool(true)];
            bool S_41_transpose_y_1 = const()[name = string("S_41_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_41_cast_fp16 = matmul(transpose_x = S_41_transpose_x_1, transpose_y = S_41_transpose_y_1, x = var_6028_cast_fp16, y = u_49_cast_fp16)[name = string("S_41_cast_fp16")];
            tensor<int32, [4]> q_93_begin_0 = const()[name = string("q_93_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_93_end_0 = const()[name = string("q_93_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> q_93_end_mask_0 = const()[name = string("q_93_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_93_cast_fp16 = slice_by_index(begin = q_93_begin_0, end = q_93_end_0, end_mask = q_93_end_mask_0, x = q_89_cast_fp16)[name = string("q_93_cast_fp16")];
            tensor<int32, [4]> k_95_begin_0 = const()[name = string("k_95_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_95_end_0 = const()[name = string("k_95_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> k_95_end_mask_0 = const()[name = string("k_95_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_95_cast_fp16 = slice_by_index(begin = k_95_begin_0, end = k_95_end_0, end_mask = k_95_end_mask_0, x = k_91_cast_fp16)[name = string("k_95_cast_fp16")];
            tensor<int32, [4]> kb_51_begin_0 = const()[name = string("kb_51_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_51_end_0 = const()[name = string("kb_51_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> kb_51_end_mask_0 = const()[name = string("kb_51_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_51_cast_fp16 = slice_by_index(begin = kb_51_begin_0, end = kb_51_end_0, end_mask = kb_51_end_mask_0, x = var_5933_cast_fp16)[name = string("kb_51_cast_fp16")];
            tensor<int32, [4]> vb_61_begin_0 = const()[name = string("vb_61_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_61_end_0 = const()[name = string("vb_61_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> vb_61_end_mask_0 = const()[name = string("vb_61_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_61_cast_fp16 = slice_by_index(begin = vb_61_begin_0, end = vb_61_end_0, end_mask = vb_61_end_mask_0, x = vb_57_cast_fp16)[name = string("vb_61_cast_fp16")];
            tensor<int32, [3]> gdec_61_begin_0 = const()[name = string("gdec_61_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_61_end_0 = const()[name = string("gdec_61_end_0"), val = tensor<int32, [3]>([1, 4, 32])];
            tensor<bool, [3]> gdec_61_end_mask_0 = const()[name = string("gdec_61_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_61_cast_fp16 = slice_by_index(begin = gdec_61_begin_0, end = gdec_61_end_0, end_mask = gdec_61_end_mask_0, x = gdec_57_cast_fp16)[name = string("gdec_61_cast_fp16")];
            tensor<int32, [2]> var_6113 = const()[name = string("op_6113"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_6114_cast_fp16 = reshape(shape = var_6113, x = gdec_61_cast_fp16)[name = string("op_6114_cast_fp16")];
            tensor<fp16, [4, 256]> var_6115_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_6114_cast_fp16)[name = string("op_6115_cast_fp16")];
            tensor<int32, [4]> var_6120 = const()[name = string("op_6120"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_51_cast_fp16 = reshape(shape = var_6120, x = var_6115_cast_fp16)[name = string("P_51_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6122_cast_fp16 = exp(x = P_51_cast_fp16)[name = string("op_6122_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_51_cast_fp16 = mul(x = var_6122_cast_fp16, y = tril_incl_to_fp16)[name = string("D_51_cast_fp16")];
            bool kk_51_transpose_x_1 = const()[name = string("kk_51_transpose_x_1"), val = bool(false)];
            bool kk_51_transpose_y_1 = const()[name = string("kk_51_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_51_cast_fp16 = matmul(transpose_x = kk_51_transpose_x_1, transpose_y = kk_51_transpose_y_1, x = kb_51_cast_fp16, y = k_95_cast_fp16)[name = string("kk_51_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_6128_cast_fp16 = mul(x = kk_51_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_6128_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6129_cast_fp16 = mul(x = var_6128_cast_fp16, y = D_51_cast_fp16)[name = string("op_6129_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_51_cast_fp16 = mul(x = var_6129_cast_fp16, y = cum_tail_to_fp16)[name = string("N_51_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6131_cast_fp16 = mul(x = D_51_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_6131_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_51_cast_fp16 = mul(x = kk_51_cast_fp16, y = var_6131_cast_fp16)[name = string("Na_51_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_201_cast_fp16 = add(x = eye_to_fp16, y = N_51_cast_fp16)[name = string("Minv_201_cast_fp16")];
            bool Np_151_transpose_x_0 = const()[name = string("Np_151_transpose_x_0"), val = bool(false)];
            bool Np_151_transpose_y_0 = const()[name = string("Np_151_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_151_cast_fp16 = matmul(transpose_x = Np_151_transpose_x_0, transpose_y = Np_151_transpose_y_0, x = N_51_cast_fp16, y = Na_51_cast_fp16)[name = string("Np_151_cast_fp16")];
            bool Npa_101_transpose_x_0 = const()[name = string("Npa_101_transpose_x_0"), val = bool(false)];
            bool Npa_101_transpose_y_0 = const()[name = string("Npa_101_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_101_cast_fp16 = matmul(transpose_x = Npa_101_transpose_x_0, transpose_y = Npa_101_transpose_y_0, x = Na_51_cast_fp16, y = N_51_cast_fp16)[name = string("Npa_101_cast_fp16")];
            bool var_6137_transpose_x_0 = const()[name = string("op_6137_transpose_x_0"), val = bool(false)];
            bool var_6137_transpose_y_0 = const()[name = string("op_6137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6137_cast_fp16 = matmul(transpose_x = var_6137_transpose_x_0, transpose_y = var_6137_transpose_y_0, x = Minv_201_cast_fp16, y = Np_151_cast_fp16)[name = string("op_6137_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_203_cast_fp16 = add(x = Minv_201_cast_fp16, y = var_6137_cast_fp16)[name = string("Minv_203_cast_fp16")];
            bool Np_153_transpose_x_0 = const()[name = string("Np_153_transpose_x_0"), val = bool(false)];
            bool Np_153_transpose_y_0 = const()[name = string("Np_153_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_153_cast_fp16 = matmul(transpose_x = Np_153_transpose_x_0, transpose_y = Np_153_transpose_y_0, x = Np_151_cast_fp16, y = Npa_101_cast_fp16)[name = string("Np_153_cast_fp16")];
            bool Npa_103_transpose_x_0 = const()[name = string("Npa_103_transpose_x_0"), val = bool(false)];
            bool Npa_103_transpose_y_0 = const()[name = string("Npa_103_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_103_cast_fp16 = matmul(transpose_x = Npa_103_transpose_x_0, transpose_y = Npa_103_transpose_y_0, x = Npa_101_cast_fp16, y = Np_151_cast_fp16)[name = string("Npa_103_cast_fp16")];
            bool var_6142_transpose_x_0 = const()[name = string("op_6142_transpose_x_0"), val = bool(false)];
            bool var_6142_transpose_y_0 = const()[name = string("op_6142_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6142_cast_fp16 = matmul(transpose_x = var_6142_transpose_x_0, transpose_y = var_6142_transpose_y_0, x = Minv_203_cast_fp16, y = Np_153_cast_fp16)[name = string("op_6142_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_205_cast_fp16 = add(x = Minv_203_cast_fp16, y = var_6142_cast_fp16)[name = string("Minv_205_cast_fp16")];
            bool Np_155_transpose_x_0 = const()[name = string("Np_155_transpose_x_0"), val = bool(false)];
            bool Np_155_transpose_y_0 = const()[name = string("Np_155_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_155_cast_fp16 = matmul(transpose_x = Np_155_transpose_x_0, transpose_y = Np_155_transpose_y_0, x = Np_153_cast_fp16, y = Npa_103_cast_fp16)[name = string("Np_155_cast_fp16")];
            bool var_6146_transpose_x_0 = const()[name = string("op_6146_transpose_x_0"), val = bool(false)];
            bool var_6146_transpose_y_0 = const()[name = string("op_6146_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6146_cast_fp16 = matmul(transpose_x = var_6146_transpose_x_0, transpose_y = var_6146_transpose_y_0, x = Minv_205_cast_fp16, y = Np_155_cast_fp16)[name = string("op_6146_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_207_cast_fp16 = add(x = Minv_205_cast_fp16, y = var_6146_cast_fp16)[name = string("Minv_207_cast_fp16")];
            bool u_51_transpose_x_0 = const()[name = string("u_51_transpose_x_0"), val = bool(false)];
            bool u_51_transpose_y_0 = const()[name = string("u_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_51_cast_fp16 = matmul(transpose_x = u_51_transpose_x_0, transpose_y = u_51_transpose_y_0, x = Minv_207_cast_fp16, y = vb_61_cast_fp16)[name = string("u_51_cast_fp16")];
            bool var_6153_transpose_x_1 = const()[name = string("op_6153_transpose_x_1"), val = bool(false)];
            bool var_6153_transpose_y_1 = const()[name = string("op_6153_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_6153_cast_fp16 = matmul(transpose_x = var_6153_transpose_x_1, transpose_y = var_6153_transpose_y_1, x = q_93_cast_fp16, y = k_95_cast_fp16)[name = string("op_6153_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_51_cast_fp16 = mul(x = var_6153_cast_fp16, y = D_51_cast_fp16)[name = string("A_51_cast_fp16")];
            tensor<fp16, [16]> var_6155_bias_0_to_fp16 = const()[name = string("op_6155_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_6155_cast_fp16 = linear(bias = var_6155_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_61_cast_fp16)[name = string("op_6155_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_41_cast_fp16 = exp(x = var_6155_cast_fp16)[name = string("cum_41_cast_fp16")];
            tensor<int32, [1]> var_6158_axes_0 = const()[name = string("op_6158_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_6158_cast_fp16 = expand_dims(axes = var_6158_axes_0, x = cum_41_cast_fp16)[name = string("op_6158_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6159_cast_fp16 = mul(x = kb_51_cast_fp16, y = var_6158_cast_fp16)[name = string("op_6159_cast_fp16")];
            bool w_51_transpose_x_0 = const()[name = string("w_51_transpose_x_0"), val = bool(false)];
            bool w_51_transpose_y_0 = const()[name = string("w_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_51_cast_fp16 = matmul(transpose_x = w_51_transpose_x_0, transpose_y = w_51_transpose_y_0, x = Minv_207_cast_fp16, y = var_6159_cast_fp16)[name = string("w_51_cast_fp16")];
            bool var_6161_transpose_x_0 = const()[name = string("op_6161_transpose_x_0"), val = bool(false)];
            bool var_6161_transpose_y_0 = const()[name = string("op_6161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6161_cast_fp16 = matmul(transpose_x = var_6161_transpose_x_0, transpose_y = var_6161_transpose_y_0, x = w_51_cast_fp16, y = S_41_cast_fp16)[name = string("op_6161_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_41_cast_fp16 = sub(x = u_51_cast_fp16, y = var_6161_cast_fp16)[name = string("vnew_41_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6166_cast_fp16 = mul(x = q_93_cast_fp16, y = var_6158_cast_fp16)[name = string("op_6166_cast_fp16")];
            bool var_6167_transpose_x_0 = const()[name = string("op_6167_transpose_x_0"), val = bool(false)];
            bool var_6167_transpose_y_0 = const()[name = string("op_6167_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6167_cast_fp16 = matmul(transpose_x = var_6167_transpose_x_0, transpose_y = var_6167_transpose_y_0, x = var_6166_cast_fp16, y = S_41_cast_fp16)[name = string("op_6167_cast_fp16")];
            bool var_6168_transpose_x_0 = const()[name = string("op_6168_transpose_x_0"), val = bool(false)];
            bool var_6168_transpose_y_0 = const()[name = string("op_6168_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6168_cast_fp16 = matmul(transpose_x = var_6168_transpose_x_0, transpose_y = var_6168_transpose_y_0, x = A_51_cast_fp16, y = vnew_41_cast_fp16)[name = string("op_6168_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6170_cast_fp16 = add(x = var_6167_cast_fp16, y = var_6168_cast_fp16)[name = string("op_6170_cast_fp16")];
            tensor<fp16, [16]> var_6171_bias_0_to_fp16 = const()[name = string("op_6171_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_6171_cast_fp16 = linear(bias = var_6171_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_61_cast_fp16)[name = string("op_6171_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_43_cast_fp16 = exp(x = var_6171_cast_fp16)[name = string("tail_43_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_85_to_fp16 = const()[name = string("transpose_85_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_6173_bias_0_to_fp16 = const()[name = string("op_6173_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_6173_cast_fp16 = linear(bias = var_6173_bias_0_to_fp16, weight = transpose_85_to_fp16, x = gdec_61_cast_fp16)[name = string("op_6173_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_6174_cast_fp16 = exp(x = var_6173_cast_fp16)[name = string("op_6174_cast_fp16")];
            tensor<int32, [1]> last_35_axes_0 = const()[name = string("last_35_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_35_cast_fp16 = expand_dims(axes = last_35_axes_0, x = var_6174_cast_fp16)[name = string("last_35_cast_fp16")];
            tensor<int32, [1]> var_6178_axes_0 = const()[name = string("op_6178_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_6178_cast_fp16 = expand_dims(axes = var_6178_axes_0, x = tail_43_cast_fp16)[name = string("op_6178_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6179_cast_fp16 = mul(x = k_95_cast_fp16, y = var_6178_cast_fp16)[name = string("op_6179_cast_fp16")];
            bool upd_33_transpose_x_1 = const()[name = string("upd_33_transpose_x_1"), val = bool(true)];
            bool upd_33_transpose_y_1 = const()[name = string("upd_33_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_33_cast_fp16 = matmul(transpose_x = upd_33_transpose_x_1, transpose_y = upd_33_transpose_y_1, x = var_6179_cast_fp16, y = vnew_41_cast_fp16)[name = string("upd_33_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_6184_cast_fp16 = mul(x = S_41_cast_fp16, y = last_35_cast_fp16)[name = string("op_6184_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_43_cast_fp16 = add(x = var_6184_cast_fp16, y = upd_33_cast_fp16)[name = string("S_43_cast_fp16")];
            tensor<int32, [4]> q_95_begin_0 = const()[name = string("q_95_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_95_end_0 = const()[name = string("q_95_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> q_95_end_mask_0 = const()[name = string("q_95_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_95_cast_fp16 = slice_by_index(begin = q_95_begin_0, end = q_95_end_0, end_mask = q_95_end_mask_0, x = q_89_cast_fp16)[name = string("q_95_cast_fp16")];
            tensor<int32, [4]> k_97_begin_0 = const()[name = string("k_97_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_97_end_0 = const()[name = string("k_97_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> k_97_end_mask_0 = const()[name = string("k_97_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_97_cast_fp16 = slice_by_index(begin = k_97_begin_0, end = k_97_end_0, end_mask = k_97_end_mask_0, x = k_91_cast_fp16)[name = string("k_97_cast_fp16")];
            tensor<int32, [4]> kb_53_begin_0 = const()[name = string("kb_53_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_53_end_0 = const()[name = string("kb_53_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> kb_53_end_mask_0 = const()[name = string("kb_53_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_53_cast_fp16 = slice_by_index(begin = kb_53_begin_0, end = kb_53_end_0, end_mask = kb_53_end_mask_0, x = var_5933_cast_fp16)[name = string("kb_53_cast_fp16")];
            tensor<int32, [4]> vb_63_begin_0 = const()[name = string("vb_63_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_63_end_0 = const()[name = string("vb_63_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> vb_63_end_mask_0 = const()[name = string("vb_63_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_63_cast_fp16 = slice_by_index(begin = vb_63_begin_0, end = vb_63_end_0, end_mask = vb_63_end_mask_0, x = vb_57_cast_fp16)[name = string("vb_63_cast_fp16")];
            tensor<int32, [3]> gdec_63_begin_0 = const()[name = string("gdec_63_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_63_end_0 = const()[name = string("gdec_63_end_0"), val = tensor<int32, [3]>([1, 4, 48])];
            tensor<bool, [3]> gdec_63_end_mask_0 = const()[name = string("gdec_63_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_63_cast_fp16 = slice_by_index(begin = gdec_63_begin_0, end = gdec_63_end_0, end_mask = gdec_63_end_mask_0, x = gdec_57_cast_fp16)[name = string("gdec_63_cast_fp16")];
            tensor<int32, [2]> var_6267 = const()[name = string("op_6267"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_6268_cast_fp16 = reshape(shape = var_6267, x = gdec_63_cast_fp16)[name = string("op_6268_cast_fp16")];
            tensor<fp16, [4, 256]> var_6269_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_6268_cast_fp16)[name = string("op_6269_cast_fp16")];
            tensor<int32, [4]> var_6274 = const()[name = string("op_6274"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_53_cast_fp16 = reshape(shape = var_6274, x = var_6269_cast_fp16)[name = string("P_53_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6276_cast_fp16 = exp(x = P_53_cast_fp16)[name = string("op_6276_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_53_cast_fp16 = mul(x = var_6276_cast_fp16, y = tril_incl_to_fp16)[name = string("D_53_cast_fp16")];
            bool kk_53_transpose_x_1 = const()[name = string("kk_53_transpose_x_1"), val = bool(false)];
            bool kk_53_transpose_y_1 = const()[name = string("kk_53_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_53_cast_fp16 = matmul(transpose_x = kk_53_transpose_x_1, transpose_y = kk_53_transpose_y_1, x = kb_53_cast_fp16, y = k_97_cast_fp16)[name = string("kk_53_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_6282_cast_fp16 = mul(x = kk_53_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_6282_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6283_cast_fp16 = mul(x = var_6282_cast_fp16, y = D_53_cast_fp16)[name = string("op_6283_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_53_cast_fp16 = mul(x = var_6283_cast_fp16, y = cum_tail_to_fp16)[name = string("N_53_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6285_cast_fp16 = mul(x = D_53_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_6285_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_53_cast_fp16 = mul(x = kk_53_cast_fp16, y = var_6285_cast_fp16)[name = string("Na_53_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_209_cast_fp16 = add(x = eye_to_fp16, y = N_53_cast_fp16)[name = string("Minv_209_cast_fp16")];
            bool Np_157_transpose_x_0 = const()[name = string("Np_157_transpose_x_0"), val = bool(false)];
            bool Np_157_transpose_y_0 = const()[name = string("Np_157_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_157_cast_fp16 = matmul(transpose_x = Np_157_transpose_x_0, transpose_y = Np_157_transpose_y_0, x = N_53_cast_fp16, y = Na_53_cast_fp16)[name = string("Np_157_cast_fp16")];
            bool Npa_105_transpose_x_0 = const()[name = string("Npa_105_transpose_x_0"), val = bool(false)];
            bool Npa_105_transpose_y_0 = const()[name = string("Npa_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_105_cast_fp16 = matmul(transpose_x = Npa_105_transpose_x_0, transpose_y = Npa_105_transpose_y_0, x = Na_53_cast_fp16, y = N_53_cast_fp16)[name = string("Npa_105_cast_fp16")];
            bool var_6291_transpose_x_0 = const()[name = string("op_6291_transpose_x_0"), val = bool(false)];
            bool var_6291_transpose_y_0 = const()[name = string("op_6291_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6291_cast_fp16 = matmul(transpose_x = var_6291_transpose_x_0, transpose_y = var_6291_transpose_y_0, x = Minv_209_cast_fp16, y = Np_157_cast_fp16)[name = string("op_6291_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_211_cast_fp16 = add(x = Minv_209_cast_fp16, y = var_6291_cast_fp16)[name = string("Minv_211_cast_fp16")];
            bool Np_159_transpose_x_0 = const()[name = string("Np_159_transpose_x_0"), val = bool(false)];
            bool Np_159_transpose_y_0 = const()[name = string("Np_159_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_159_cast_fp16 = matmul(transpose_x = Np_159_transpose_x_0, transpose_y = Np_159_transpose_y_0, x = Np_157_cast_fp16, y = Npa_105_cast_fp16)[name = string("Np_159_cast_fp16")];
            bool Npa_107_transpose_x_0 = const()[name = string("Npa_107_transpose_x_0"), val = bool(false)];
            bool Npa_107_transpose_y_0 = const()[name = string("Npa_107_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_107_cast_fp16 = matmul(transpose_x = Npa_107_transpose_x_0, transpose_y = Npa_107_transpose_y_0, x = Npa_105_cast_fp16, y = Np_157_cast_fp16)[name = string("Npa_107_cast_fp16")];
            bool var_6296_transpose_x_0 = const()[name = string("op_6296_transpose_x_0"), val = bool(false)];
            bool var_6296_transpose_y_0 = const()[name = string("op_6296_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6296_cast_fp16 = matmul(transpose_x = var_6296_transpose_x_0, transpose_y = var_6296_transpose_y_0, x = Minv_211_cast_fp16, y = Np_159_cast_fp16)[name = string("op_6296_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_213_cast_fp16 = add(x = Minv_211_cast_fp16, y = var_6296_cast_fp16)[name = string("Minv_213_cast_fp16")];
            bool Np_161_transpose_x_0 = const()[name = string("Np_161_transpose_x_0"), val = bool(false)];
            bool Np_161_transpose_y_0 = const()[name = string("Np_161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_161_cast_fp16 = matmul(transpose_x = Np_161_transpose_x_0, transpose_y = Np_161_transpose_y_0, x = Np_159_cast_fp16, y = Npa_107_cast_fp16)[name = string("Np_161_cast_fp16")];
            bool var_6300_transpose_x_0 = const()[name = string("op_6300_transpose_x_0"), val = bool(false)];
            bool var_6300_transpose_y_0 = const()[name = string("op_6300_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6300_cast_fp16 = matmul(transpose_x = var_6300_transpose_x_0, transpose_y = var_6300_transpose_y_0, x = Minv_213_cast_fp16, y = Np_161_cast_fp16)[name = string("op_6300_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_215_cast_fp16 = add(x = Minv_213_cast_fp16, y = var_6300_cast_fp16)[name = string("Minv_215_cast_fp16")];
            bool u_53_transpose_x_0 = const()[name = string("u_53_transpose_x_0"), val = bool(false)];
            bool u_53_transpose_y_0 = const()[name = string("u_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_53_cast_fp16 = matmul(transpose_x = u_53_transpose_x_0, transpose_y = u_53_transpose_y_0, x = Minv_215_cast_fp16, y = vb_63_cast_fp16)[name = string("u_53_cast_fp16")];
            bool var_6307_transpose_x_1 = const()[name = string("op_6307_transpose_x_1"), val = bool(false)];
            bool var_6307_transpose_y_1 = const()[name = string("op_6307_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_6307_cast_fp16 = matmul(transpose_x = var_6307_transpose_x_1, transpose_y = var_6307_transpose_y_1, x = q_95_cast_fp16, y = k_97_cast_fp16)[name = string("op_6307_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_53_cast_fp16 = mul(x = var_6307_cast_fp16, y = D_53_cast_fp16)[name = string("A_53_cast_fp16")];
            tensor<fp16, [16]> var_6309_bias_0_to_fp16 = const()[name = string("op_6309_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_6309_cast_fp16 = linear(bias = var_6309_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_63_cast_fp16)[name = string("op_6309_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_43_cast_fp16 = exp(x = var_6309_cast_fp16)[name = string("cum_43_cast_fp16")];
            tensor<int32, [1]> var_6312_axes_0 = const()[name = string("op_6312_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_6312_cast_fp16 = expand_dims(axes = var_6312_axes_0, x = cum_43_cast_fp16)[name = string("op_6312_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6313_cast_fp16 = mul(x = kb_53_cast_fp16, y = var_6312_cast_fp16)[name = string("op_6313_cast_fp16")];
            bool w_53_transpose_x_0 = const()[name = string("w_53_transpose_x_0"), val = bool(false)];
            bool w_53_transpose_y_0 = const()[name = string("w_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_53_cast_fp16 = matmul(transpose_x = w_53_transpose_x_0, transpose_y = w_53_transpose_y_0, x = Minv_215_cast_fp16, y = var_6313_cast_fp16)[name = string("w_53_cast_fp16")];
            bool var_6315_transpose_x_0 = const()[name = string("op_6315_transpose_x_0"), val = bool(false)];
            bool var_6315_transpose_y_0 = const()[name = string("op_6315_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6315_cast_fp16 = matmul(transpose_x = var_6315_transpose_x_0, transpose_y = var_6315_transpose_y_0, x = w_53_cast_fp16, y = S_43_cast_fp16)[name = string("op_6315_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_43_cast_fp16 = sub(x = u_53_cast_fp16, y = var_6315_cast_fp16)[name = string("vnew_43_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6320_cast_fp16 = mul(x = q_95_cast_fp16, y = var_6312_cast_fp16)[name = string("op_6320_cast_fp16")];
            bool var_6321_transpose_x_0 = const()[name = string("op_6321_transpose_x_0"), val = bool(false)];
            bool var_6321_transpose_y_0 = const()[name = string("op_6321_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6321_cast_fp16 = matmul(transpose_x = var_6321_transpose_x_0, transpose_y = var_6321_transpose_y_0, x = var_6320_cast_fp16, y = S_43_cast_fp16)[name = string("op_6321_cast_fp16")];
            bool var_6322_transpose_x_0 = const()[name = string("op_6322_transpose_x_0"), val = bool(false)];
            bool var_6322_transpose_y_0 = const()[name = string("op_6322_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6322_cast_fp16 = matmul(transpose_x = var_6322_transpose_x_0, transpose_y = var_6322_transpose_y_0, x = A_53_cast_fp16, y = vnew_43_cast_fp16)[name = string("op_6322_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6324_cast_fp16 = add(x = var_6321_cast_fp16, y = var_6322_cast_fp16)[name = string("op_6324_cast_fp16")];
            tensor<fp16, [16]> var_6325_bias_0_to_fp16 = const()[name = string("op_6325_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_6325_cast_fp16 = linear(bias = var_6325_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_63_cast_fp16)[name = string("op_6325_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_45_cast_fp16 = exp(x = var_6325_cast_fp16)[name = string("tail_45_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_89_to_fp16 = const()[name = string("transpose_89_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_6327_bias_0_to_fp16 = const()[name = string("op_6327_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_6327_cast_fp16 = linear(bias = var_6327_bias_0_to_fp16, weight = transpose_89_to_fp16, x = gdec_63_cast_fp16)[name = string("op_6327_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_6328_cast_fp16 = exp(x = var_6327_cast_fp16)[name = string("op_6328_cast_fp16")];
            tensor<int32, [1]> last_37_axes_0 = const()[name = string("last_37_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_37_cast_fp16 = expand_dims(axes = last_37_axes_0, x = var_6328_cast_fp16)[name = string("last_37_cast_fp16")];
            tensor<int32, [1]> var_6332_axes_0 = const()[name = string("op_6332_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_6332_cast_fp16 = expand_dims(axes = var_6332_axes_0, x = tail_45_cast_fp16)[name = string("op_6332_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6333_cast_fp16 = mul(x = k_97_cast_fp16, y = var_6332_cast_fp16)[name = string("op_6333_cast_fp16")];
            bool upd_35_transpose_x_1 = const()[name = string("upd_35_transpose_x_1"), val = bool(true)];
            bool upd_35_transpose_y_1 = const()[name = string("upd_35_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_35_cast_fp16 = matmul(transpose_x = upd_35_transpose_x_1, transpose_y = upd_35_transpose_y_1, x = var_6333_cast_fp16, y = vnew_43_cast_fp16)[name = string("upd_35_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_6338_cast_fp16 = mul(x = S_43_cast_fp16, y = last_37_cast_fp16)[name = string("op_6338_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_45_cast_fp16 = add(x = var_6338_cast_fp16, y = upd_35_cast_fp16)[name = string("S_45_cast_fp16")];
            tensor<int32, [4]> q_97_begin_0 = const()[name = string("q_97_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_97_end_0 = const()[name = string("q_97_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> q_97_end_mask_0 = const()[name = string("q_97_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, x = q_89_cast_fp16)[name = string("q_97_cast_fp16")];
            tensor<int32, [4]> k_99_begin_0 = const()[name = string("k_99_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_99_end_0 = const()[name = string("k_99_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> k_99_end_mask_0 = const()[name = string("k_99_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_99_cast_fp16 = slice_by_index(begin = k_99_begin_0, end = k_99_end_0, end_mask = k_99_end_mask_0, x = k_91_cast_fp16)[name = string("k_99_cast_fp16")];
            tensor<int32, [4]> kb_55_begin_0 = const()[name = string("kb_55_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_55_end_0 = const()[name = string("kb_55_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> kb_55_end_mask_0 = const()[name = string("kb_55_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_55_cast_fp16 = slice_by_index(begin = kb_55_begin_0, end = kb_55_end_0, end_mask = kb_55_end_mask_0, x = var_5933_cast_fp16)[name = string("kb_55_cast_fp16")];
            tensor<int32, [4]> vb_65_begin_0 = const()[name = string("vb_65_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_65_end_0 = const()[name = string("vb_65_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> vb_65_end_mask_0 = const()[name = string("vb_65_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_65_cast_fp16 = slice_by_index(begin = vb_65_begin_0, end = vb_65_end_0, end_mask = vb_65_end_mask_0, x = vb_57_cast_fp16)[name = string("vb_65_cast_fp16")];
            tensor<int32, [3]> gdec_65_begin_0 = const()[name = string("gdec_65_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_65_end_0 = const()[name = string("gdec_65_end_0"), val = tensor<int32, [3]>([1, 4, 64])];
            tensor<bool, [3]> gdec_65_end_mask_0 = const()[name = string("gdec_65_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_65_cast_fp16 = slice_by_index(begin = gdec_65_begin_0, end = gdec_65_end_0, end_mask = gdec_65_end_mask_0, x = gdec_57_cast_fp16)[name = string("gdec_65_cast_fp16")];
            tensor<int32, [2]> var_6421 = const()[name = string("op_6421"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_6422_cast_fp16 = reshape(shape = var_6421, x = gdec_65_cast_fp16)[name = string("op_6422_cast_fp16")];
            tensor<fp16, [4, 256]> var_6423_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_6422_cast_fp16)[name = string("op_6423_cast_fp16")];
            tensor<int32, [4]> var_6428 = const()[name = string("op_6428"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_55_cast_fp16 = reshape(shape = var_6428, x = var_6423_cast_fp16)[name = string("P_55_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6430_cast_fp16 = exp(x = P_55_cast_fp16)[name = string("op_6430_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_55_cast_fp16 = mul(x = var_6430_cast_fp16, y = tril_incl_to_fp16)[name = string("D_55_cast_fp16")];
            bool kk_55_transpose_x_1 = const()[name = string("kk_55_transpose_x_1"), val = bool(false)];
            bool kk_55_transpose_y_1 = const()[name = string("kk_55_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_55_cast_fp16 = matmul(transpose_x = kk_55_transpose_x_1, transpose_y = kk_55_transpose_y_1, x = kb_55_cast_fp16, y = k_99_cast_fp16)[name = string("kk_55_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_6436_cast_fp16 = mul(x = kk_55_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_6436_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6437_cast_fp16 = mul(x = var_6436_cast_fp16, y = D_55_cast_fp16)[name = string("op_6437_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_55_cast_fp16 = mul(x = var_6437_cast_fp16, y = cum_tail_to_fp16)[name = string("N_55_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6439_cast_fp16 = mul(x = D_55_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_6439_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_55_cast_fp16 = mul(x = kk_55_cast_fp16, y = var_6439_cast_fp16)[name = string("Na_55_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_217_cast_fp16 = add(x = eye_to_fp16, y = N_55_cast_fp16)[name = string("Minv_217_cast_fp16")];
            bool Np_163_transpose_x_0 = const()[name = string("Np_163_transpose_x_0"), val = bool(false)];
            bool Np_163_transpose_y_0 = const()[name = string("Np_163_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_163_cast_fp16 = matmul(transpose_x = Np_163_transpose_x_0, transpose_y = Np_163_transpose_y_0, x = N_55_cast_fp16, y = Na_55_cast_fp16)[name = string("Np_163_cast_fp16")];
            bool Npa_109_transpose_x_0 = const()[name = string("Npa_109_transpose_x_0"), val = bool(false)];
            bool Npa_109_transpose_y_0 = const()[name = string("Npa_109_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_109_cast_fp16 = matmul(transpose_x = Npa_109_transpose_x_0, transpose_y = Npa_109_transpose_y_0, x = Na_55_cast_fp16, y = N_55_cast_fp16)[name = string("Npa_109_cast_fp16")];
            bool var_6445_transpose_x_0 = const()[name = string("op_6445_transpose_x_0"), val = bool(false)];
            bool var_6445_transpose_y_0 = const()[name = string("op_6445_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6445_cast_fp16 = matmul(transpose_x = var_6445_transpose_x_0, transpose_y = var_6445_transpose_y_0, x = Minv_217_cast_fp16, y = Np_163_cast_fp16)[name = string("op_6445_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_219_cast_fp16 = add(x = Minv_217_cast_fp16, y = var_6445_cast_fp16)[name = string("Minv_219_cast_fp16")];
            bool Np_165_transpose_x_0 = const()[name = string("Np_165_transpose_x_0"), val = bool(false)];
            bool Np_165_transpose_y_0 = const()[name = string("Np_165_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_165_cast_fp16 = matmul(transpose_x = Np_165_transpose_x_0, transpose_y = Np_165_transpose_y_0, x = Np_163_cast_fp16, y = Npa_109_cast_fp16)[name = string("Np_165_cast_fp16")];
            bool Npa_111_transpose_x_0 = const()[name = string("Npa_111_transpose_x_0"), val = bool(false)];
            bool Npa_111_transpose_y_0 = const()[name = string("Npa_111_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_111_cast_fp16 = matmul(transpose_x = Npa_111_transpose_x_0, transpose_y = Npa_111_transpose_y_0, x = Npa_109_cast_fp16, y = Np_163_cast_fp16)[name = string("Npa_111_cast_fp16")];
            bool var_6450_transpose_x_0 = const()[name = string("op_6450_transpose_x_0"), val = bool(false)];
            bool var_6450_transpose_y_0 = const()[name = string("op_6450_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6450_cast_fp16 = matmul(transpose_x = var_6450_transpose_x_0, transpose_y = var_6450_transpose_y_0, x = Minv_219_cast_fp16, y = Np_165_cast_fp16)[name = string("op_6450_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_221_cast_fp16 = add(x = Minv_219_cast_fp16, y = var_6450_cast_fp16)[name = string("Minv_221_cast_fp16")];
            bool Np_167_transpose_x_0 = const()[name = string("Np_167_transpose_x_0"), val = bool(false)];
            bool Np_167_transpose_y_0 = const()[name = string("Np_167_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_167_cast_fp16 = matmul(transpose_x = Np_167_transpose_x_0, transpose_y = Np_167_transpose_y_0, x = Np_165_cast_fp16, y = Npa_111_cast_fp16)[name = string("Np_167_cast_fp16")];
            bool var_6454_transpose_x_0 = const()[name = string("op_6454_transpose_x_0"), val = bool(false)];
            bool var_6454_transpose_y_0 = const()[name = string("op_6454_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6454_cast_fp16 = matmul(transpose_x = var_6454_transpose_x_0, transpose_y = var_6454_transpose_y_0, x = Minv_221_cast_fp16, y = Np_167_cast_fp16)[name = string("op_6454_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_223_cast_fp16 = add(x = Minv_221_cast_fp16, y = var_6454_cast_fp16)[name = string("Minv_223_cast_fp16")];
            bool u_55_transpose_x_0 = const()[name = string("u_55_transpose_x_0"), val = bool(false)];
            bool u_55_transpose_y_0 = const()[name = string("u_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_55_cast_fp16 = matmul(transpose_x = u_55_transpose_x_0, transpose_y = u_55_transpose_y_0, x = Minv_223_cast_fp16, y = vb_65_cast_fp16)[name = string("u_55_cast_fp16")];
            bool var_6461_transpose_x_1 = const()[name = string("op_6461_transpose_x_1"), val = bool(false)];
            bool var_6461_transpose_y_1 = const()[name = string("op_6461_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_6461_cast_fp16 = matmul(transpose_x = var_6461_transpose_x_1, transpose_y = var_6461_transpose_y_1, x = q_97_cast_fp16, y = k_99_cast_fp16)[name = string("op_6461_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_55_cast_fp16 = mul(x = var_6461_cast_fp16, y = D_55_cast_fp16)[name = string("A_55_cast_fp16")];
            tensor<fp16, [16]> var_6463_bias_0_to_fp16 = const()[name = string("op_6463_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_6463_cast_fp16 = linear(bias = var_6463_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_65_cast_fp16)[name = string("op_6463_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_45_cast_fp16 = exp(x = var_6463_cast_fp16)[name = string("cum_45_cast_fp16")];
            tensor<int32, [1]> var_6466_axes_0 = const()[name = string("op_6466_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_6466_cast_fp16 = expand_dims(axes = var_6466_axes_0, x = cum_45_cast_fp16)[name = string("op_6466_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6467_cast_fp16 = mul(x = kb_55_cast_fp16, y = var_6466_cast_fp16)[name = string("op_6467_cast_fp16")];
            bool w_55_transpose_x_0 = const()[name = string("w_55_transpose_x_0"), val = bool(false)];
            bool w_55_transpose_y_0 = const()[name = string("w_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_55_cast_fp16 = matmul(transpose_x = w_55_transpose_x_0, transpose_y = w_55_transpose_y_0, x = Minv_223_cast_fp16, y = var_6467_cast_fp16)[name = string("w_55_cast_fp16")];
            bool var_6469_transpose_x_0 = const()[name = string("op_6469_transpose_x_0"), val = bool(false)];
            bool var_6469_transpose_y_0 = const()[name = string("op_6469_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6469_cast_fp16 = matmul(transpose_x = var_6469_transpose_x_0, transpose_y = var_6469_transpose_y_0, x = w_55_cast_fp16, y = S_45_cast_fp16)[name = string("op_6469_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_45_cast_fp16 = sub(x = u_55_cast_fp16, y = var_6469_cast_fp16)[name = string("vnew_45_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6474_cast_fp16 = mul(x = q_97_cast_fp16, y = var_6466_cast_fp16)[name = string("op_6474_cast_fp16")];
            bool var_6475_transpose_x_0 = const()[name = string("op_6475_transpose_x_0"), val = bool(false)];
            bool var_6475_transpose_y_0 = const()[name = string("op_6475_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6475_cast_fp16 = matmul(transpose_x = var_6475_transpose_x_0, transpose_y = var_6475_transpose_y_0, x = var_6474_cast_fp16, y = S_45_cast_fp16)[name = string("op_6475_cast_fp16")];
            bool var_6476_transpose_x_0 = const()[name = string("op_6476_transpose_x_0"), val = bool(false)];
            bool var_6476_transpose_y_0 = const()[name = string("op_6476_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6476_cast_fp16 = matmul(transpose_x = var_6476_transpose_x_0, transpose_y = var_6476_transpose_y_0, x = A_55_cast_fp16, y = vnew_45_cast_fp16)[name = string("op_6476_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6478_cast_fp16 = add(x = var_6475_cast_fp16, y = var_6476_cast_fp16)[name = string("op_6478_cast_fp16")];
            tensor<fp16, [16]> var_6479_bias_0_to_fp16 = const()[name = string("op_6479_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_6479_cast_fp16 = linear(bias = var_6479_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_65_cast_fp16)[name = string("op_6479_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_47_cast_fp16 = exp(x = var_6479_cast_fp16)[name = string("tail_47_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_93_to_fp16 = const()[name = string("transpose_93_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_6481_bias_0_to_fp16 = const()[name = string("op_6481_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_6481_cast_fp16 = linear(bias = var_6481_bias_0_to_fp16, weight = transpose_93_to_fp16, x = gdec_65_cast_fp16)[name = string("op_6481_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_6482_cast_fp16 = exp(x = var_6481_cast_fp16)[name = string("op_6482_cast_fp16")];
            tensor<int32, [1]> last_39_axes_0 = const()[name = string("last_39_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_39_cast_fp16 = expand_dims(axes = last_39_axes_0, x = var_6482_cast_fp16)[name = string("last_39_cast_fp16")];
            tensor<int32, [1]> var_6486_axes_0 = const()[name = string("op_6486_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_6486_cast_fp16 = expand_dims(axes = var_6486_axes_0, x = tail_47_cast_fp16)[name = string("op_6486_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6487_cast_fp16 = mul(x = k_99_cast_fp16, y = var_6486_cast_fp16)[name = string("op_6487_cast_fp16")];
            bool upd_37_transpose_x_1 = const()[name = string("upd_37_transpose_x_1"), val = bool(true)];
            bool upd_37_transpose_y_1 = const()[name = string("upd_37_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_37_cast_fp16 = matmul(transpose_x = upd_37_transpose_x_1, transpose_y = upd_37_transpose_y_1, x = var_6487_cast_fp16, y = vnew_45_cast_fp16)[name = string("upd_37_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_6492_cast_fp16 = mul(x = S_45_cast_fp16, y = last_39_cast_fp16)[name = string("op_6492_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_47_cast_fp16 = add(x = var_6492_cast_fp16, y = upd_37_cast_fp16)[name = string("S_47_cast_fp16")];
            tensor<int32, [4]> q_99_begin_0 = const()[name = string("q_99_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> q_99_end_0 = const()[name = string("q_99_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> q_99_end_mask_0 = const()[name = string("q_99_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_99_cast_fp16 = slice_by_index(begin = q_99_begin_0, end = q_99_end_0, end_mask = q_99_end_mask_0, x = q_89_cast_fp16)[name = string("q_99_cast_fp16")];
            tensor<int32, [4]> k_101_begin_0 = const()[name = string("k_101_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> k_101_end_0 = const()[name = string("k_101_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> k_101_end_mask_0 = const()[name = string("k_101_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_101_cast_fp16 = slice_by_index(begin = k_101_begin_0, end = k_101_end_0, end_mask = k_101_end_mask_0, x = k_91_cast_fp16)[name = string("k_101_cast_fp16")];
            tensor<int32, [4]> kb_57_begin_0 = const()[name = string("kb_57_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> kb_57_end_0 = const()[name = string("kb_57_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> kb_57_end_mask_0 = const()[name = string("kb_57_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_57_cast_fp16 = slice_by_index(begin = kb_57_begin_0, end = kb_57_end_0, end_mask = kb_57_end_mask_0, x = var_5933_cast_fp16)[name = string("kb_57_cast_fp16")];
            tensor<int32, [4]> vb_67_begin_0 = const()[name = string("vb_67_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> vb_67_end_0 = const()[name = string("vb_67_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> vb_67_end_mask_0 = const()[name = string("vb_67_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_67_cast_fp16 = slice_by_index(begin = vb_67_begin_0, end = vb_67_end_0, end_mask = vb_67_end_mask_0, x = vb_57_cast_fp16)[name = string("vb_67_cast_fp16")];
            tensor<int32, [3]> gdec_67_begin_0 = const()[name = string("gdec_67_begin_0"), val = tensor<int32, [3]>([0, 0, 64])];
            tensor<int32, [3]> gdec_67_end_0 = const()[name = string("gdec_67_end_0"), val = tensor<int32, [3]>([1, 4, 80])];
            tensor<bool, [3]> gdec_67_end_mask_0 = const()[name = string("gdec_67_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_67_cast_fp16 = slice_by_index(begin = gdec_67_begin_0, end = gdec_67_end_0, end_mask = gdec_67_end_mask_0, x = gdec_57_cast_fp16)[name = string("gdec_67_cast_fp16")];
            tensor<int32, [2]> var_6575 = const()[name = string("op_6575"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_6576_cast_fp16 = reshape(shape = var_6575, x = gdec_67_cast_fp16)[name = string("op_6576_cast_fp16")];
            tensor<fp16, [4, 256]> var_6577_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_6576_cast_fp16)[name = string("op_6577_cast_fp16")];
            tensor<int32, [4]> var_6582 = const()[name = string("op_6582"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_57_cast_fp16 = reshape(shape = var_6582, x = var_6577_cast_fp16)[name = string("P_57_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6584_cast_fp16 = exp(x = P_57_cast_fp16)[name = string("op_6584_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_57_cast_fp16 = mul(x = var_6584_cast_fp16, y = tril_incl_to_fp16)[name = string("D_57_cast_fp16")];
            bool kk_57_transpose_x_1 = const()[name = string("kk_57_transpose_x_1"), val = bool(false)];
            bool kk_57_transpose_y_1 = const()[name = string("kk_57_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_57_cast_fp16 = matmul(transpose_x = kk_57_transpose_x_1, transpose_y = kk_57_transpose_y_1, x = kb_57_cast_fp16, y = k_101_cast_fp16)[name = string("kk_57_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_6590_cast_fp16 = mul(x = kk_57_cast_fp16, y = const_35_promoted_to_fp16)[name = string("op_6590_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6591_cast_fp16 = mul(x = var_6590_cast_fp16, y = D_57_cast_fp16)[name = string("op_6591_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_57_cast_fp16 = mul(x = var_6591_cast_fp16, y = cum_tail_to_fp16)[name = string("N_57_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6593_cast_fp16 = mul(x = D_57_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_6593_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_57_cast_fp16 = mul(x = kk_57_cast_fp16, y = var_6593_cast_fp16)[name = string("Na_57_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_225_cast_fp16 = add(x = eye_to_fp16, y = N_57_cast_fp16)[name = string("Minv_225_cast_fp16")];
            bool Np_169_transpose_x_0 = const()[name = string("Np_169_transpose_x_0"), val = bool(false)];
            bool Np_169_transpose_y_0 = const()[name = string("Np_169_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_169_cast_fp16 = matmul(transpose_x = Np_169_transpose_x_0, transpose_y = Np_169_transpose_y_0, x = N_57_cast_fp16, y = Na_57_cast_fp16)[name = string("Np_169_cast_fp16")];
            bool Npa_113_transpose_x_0 = const()[name = string("Npa_113_transpose_x_0"), val = bool(false)];
            bool Npa_113_transpose_y_0 = const()[name = string("Npa_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_113_cast_fp16 = matmul(transpose_x = Npa_113_transpose_x_0, transpose_y = Npa_113_transpose_y_0, x = Na_57_cast_fp16, y = N_57_cast_fp16)[name = string("Npa_113_cast_fp16")];
            bool var_6599_transpose_x_0 = const()[name = string("op_6599_transpose_x_0"), val = bool(false)];
            bool var_6599_transpose_y_0 = const()[name = string("op_6599_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6599_cast_fp16 = matmul(transpose_x = var_6599_transpose_x_0, transpose_y = var_6599_transpose_y_0, x = Minv_225_cast_fp16, y = Np_169_cast_fp16)[name = string("op_6599_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_227_cast_fp16 = add(x = Minv_225_cast_fp16, y = var_6599_cast_fp16)[name = string("Minv_227_cast_fp16")];
            bool Np_171_transpose_x_0 = const()[name = string("Np_171_transpose_x_0"), val = bool(false)];
            bool Np_171_transpose_y_0 = const()[name = string("Np_171_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_171_cast_fp16 = matmul(transpose_x = Np_171_transpose_x_0, transpose_y = Np_171_transpose_y_0, x = Np_169_cast_fp16, y = Npa_113_cast_fp16)[name = string("Np_171_cast_fp16")];
            bool Npa_115_transpose_x_0 = const()[name = string("Npa_115_transpose_x_0"), val = bool(false)];
            bool Npa_115_transpose_y_0 = const()[name = string("Npa_115_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_115_cast_fp16 = matmul(transpose_x = Npa_115_transpose_x_0, transpose_y = Npa_115_transpose_y_0, x = Npa_113_cast_fp16, y = Np_169_cast_fp16)[name = string("Npa_115_cast_fp16")];
            bool var_6604_transpose_x_0 = const()[name = string("op_6604_transpose_x_0"), val = bool(false)];
            bool var_6604_transpose_y_0 = const()[name = string("op_6604_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6604_cast_fp16 = matmul(transpose_x = var_6604_transpose_x_0, transpose_y = var_6604_transpose_y_0, x = Minv_227_cast_fp16, y = Np_171_cast_fp16)[name = string("op_6604_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_229_cast_fp16 = add(x = Minv_227_cast_fp16, y = var_6604_cast_fp16)[name = string("Minv_229_cast_fp16")];
            bool Np_173_transpose_x_0 = const()[name = string("Np_173_transpose_x_0"), val = bool(false)];
            bool Np_173_transpose_y_0 = const()[name = string("Np_173_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_173_cast_fp16 = matmul(transpose_x = Np_173_transpose_x_0, transpose_y = Np_173_transpose_y_0, x = Np_171_cast_fp16, y = Npa_115_cast_fp16)[name = string("Np_173_cast_fp16")];
            bool var_6608_transpose_x_0 = const()[name = string("op_6608_transpose_x_0"), val = bool(false)];
            bool var_6608_transpose_y_0 = const()[name = string("op_6608_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6608_cast_fp16 = matmul(transpose_x = var_6608_transpose_x_0, transpose_y = var_6608_transpose_y_0, x = Minv_229_cast_fp16, y = Np_173_cast_fp16)[name = string("op_6608_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_231_cast_fp16 = add(x = Minv_229_cast_fp16, y = var_6608_cast_fp16)[name = string("Minv_231_cast_fp16")];
            bool u_57_transpose_x_0 = const()[name = string("u_57_transpose_x_0"), val = bool(false)];
            bool u_57_transpose_y_0 = const()[name = string("u_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_57_cast_fp16 = matmul(transpose_x = u_57_transpose_x_0, transpose_y = u_57_transpose_y_0, x = Minv_231_cast_fp16, y = vb_67_cast_fp16)[name = string("u_57_cast_fp16")];
            bool var_6615_transpose_x_1 = const()[name = string("op_6615_transpose_x_1"), val = bool(false)];
            bool var_6615_transpose_y_1 = const()[name = string("op_6615_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_6615_cast_fp16 = matmul(transpose_x = var_6615_transpose_x_1, transpose_y = var_6615_transpose_y_1, x = q_99_cast_fp16, y = k_101_cast_fp16)[name = string("op_6615_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_57_cast_fp16 = mul(x = var_6615_cast_fp16, y = D_57_cast_fp16)[name = string("A_57_cast_fp16")];
            tensor<fp16, [16]> var_6617_bias_0_to_fp16 = const()[name = string("op_6617_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_6617_cast_fp16 = linear(bias = var_6617_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_67_cast_fp16)[name = string("op_6617_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_47_cast_fp16 = exp(x = var_6617_cast_fp16)[name = string("cum_47_cast_fp16")];
            tensor<int32, [1]> var_6620_axes_0 = const()[name = string("op_6620_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_6620_cast_fp16 = expand_dims(axes = var_6620_axes_0, x = cum_47_cast_fp16)[name = string("op_6620_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6621_cast_fp16 = mul(x = kb_57_cast_fp16, y = var_6620_cast_fp16)[name = string("op_6621_cast_fp16")];
            bool w_57_transpose_x_0 = const()[name = string("w_57_transpose_x_0"), val = bool(false)];
            bool w_57_transpose_y_0 = const()[name = string("w_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_57_cast_fp16 = matmul(transpose_x = w_57_transpose_x_0, transpose_y = w_57_transpose_y_0, x = Minv_231_cast_fp16, y = var_6621_cast_fp16)[name = string("w_57_cast_fp16")];
            bool var_6623_transpose_x_0 = const()[name = string("op_6623_transpose_x_0"), val = bool(false)];
            bool var_6623_transpose_y_0 = const()[name = string("op_6623_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6623_cast_fp16 = matmul(transpose_x = var_6623_transpose_x_0, transpose_y = var_6623_transpose_y_0, x = w_57_cast_fp16, y = S_47_cast_fp16)[name = string("op_6623_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_47_cast_fp16 = sub(x = u_57_cast_fp16, y = var_6623_cast_fp16)[name = string("vnew_47_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6628_cast_fp16 = mul(x = q_99_cast_fp16, y = var_6620_cast_fp16)[name = string("op_6628_cast_fp16")];
            bool var_6629_transpose_x_0 = const()[name = string("op_6629_transpose_x_0"), val = bool(false)];
            bool var_6629_transpose_y_0 = const()[name = string("op_6629_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6629_cast_fp16 = matmul(transpose_x = var_6629_transpose_x_0, transpose_y = var_6629_transpose_y_0, x = var_6628_cast_fp16, y = S_47_cast_fp16)[name = string("op_6629_cast_fp16")];
            bool var_6630_transpose_x_0 = const()[name = string("op_6630_transpose_x_0"), val = bool(false)];
            bool var_6630_transpose_y_0 = const()[name = string("op_6630_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6630_cast_fp16 = matmul(transpose_x = var_6630_transpose_x_0, transpose_y = var_6630_transpose_y_0, x = A_57_cast_fp16, y = vnew_47_cast_fp16)[name = string("op_6630_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6632_cast_fp16 = add(x = var_6629_cast_fp16, y = var_6630_cast_fp16)[name = string("op_6632_cast_fp16")];
            tensor<fp16, [16]> var_6633_bias_0_to_fp16 = const()[name = string("op_6633_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_6633_cast_fp16 = linear(bias = var_6633_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_67_cast_fp16)[name = string("op_6633_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_49_cast_fp16 = exp(x = var_6633_cast_fp16)[name = string("tail_49_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_97_to_fp16 = const()[name = string("transpose_97_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_6635_bias_0_to_fp16 = const()[name = string("op_6635_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_6635_cast_fp16 = linear(bias = var_6635_bias_0_to_fp16, weight = transpose_97_to_fp16, x = gdec_67_cast_fp16)[name = string("op_6635_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_6636_cast_fp16 = exp(x = var_6635_cast_fp16)[name = string("op_6636_cast_fp16")];
            tensor<int32, [1]> last_41_axes_0 = const()[name = string("last_41_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_41_cast_fp16 = expand_dims(axes = last_41_axes_0, x = var_6636_cast_fp16)[name = string("last_41_cast_fp16")];
            tensor<int32, [1]> var_6640_axes_0 = const()[name = string("op_6640_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_6640_cast_fp16 = expand_dims(axes = var_6640_axes_0, x = tail_49_cast_fp16)[name = string("op_6640_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6641_cast_fp16 = mul(x = k_101_cast_fp16, y = var_6640_cast_fp16)[name = string("op_6641_cast_fp16")];
            bool upd_39_transpose_x_1 = const()[name = string("upd_39_transpose_x_1"), val = bool(true)];
            bool upd_39_transpose_y_1 = const()[name = string("upd_39_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_39_cast_fp16 = matmul(transpose_x = upd_39_transpose_x_1, transpose_y = upd_39_transpose_y_1, x = var_6641_cast_fp16, y = vnew_47_cast_fp16)[name = string("upd_39_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_6646_cast_fp16 = mul(x = S_47_cast_fp16, y = last_41_cast_fp16)[name = string("op_6646_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_49_cast_fp16 = add(x = var_6646_cast_fp16, y = upd_39_cast_fp16)[name = string("S_49_cast_fp16")];
            tensor<int32, [4]> q_101_begin_0 = const()[name = string("q_101_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> q_101_end_0 = const()[name = string("q_101_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_101_end_mask_0 = const()[name = string("q_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_101_cast_fp16 = slice_by_index(begin = q_101_begin_0, end = q_101_end_0, end_mask = q_101_end_mask_0, x = q_89_cast_fp16)[name = string("q_101_cast_fp16")];
            tensor<int32, [4]> k_103_begin_0 = const()[name = string("k_103_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> k_103_end_0 = const()[name = string("k_103_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_103_end_mask_0 = const()[name = string("k_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_103_cast_fp16 = slice_by_index(begin = k_103_begin_0, end = k_103_end_0, end_mask = k_103_end_mask_0, x = k_91_cast_fp16)[name = string("k_103_cast_fp16")];
            tensor<int32, [4]> kb_59_begin_0 = const()[name = string("kb_59_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> kb_59_end_0 = const()[name = string("kb_59_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_59_end_mask_0 = const()[name = string("kb_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_59_cast_fp16 = slice_by_index(begin = kb_59_begin_0, end = kb_59_end_0, end_mask = kb_59_end_mask_0, x = var_5933_cast_fp16)[name = string("kb_59_cast_fp16")];
            tensor<int32, [4]> vb_69_begin_0 = const()[name = string("vb_69_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> vb_69_end_0 = const()[name = string("vb_69_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_69_end_mask_0 = const()[name = string("vb_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_69_cast_fp16 = slice_by_index(begin = vb_69_begin_0, end = vb_69_end_0, end_mask = vb_69_end_mask_0, x = vb_57_cast_fp16)[name = string("vb_69_cast_fp16")];
            tensor<int32, [3]> gdec_69_begin_0 = const()[name = string("gdec_69_begin_0"), val = tensor<int32, [3]>([0, 0, 80])];
            tensor<int32, [3]> gdec_69_end_0 = const()[name = string("gdec_69_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_69_end_mask_0 = const()[name = string("gdec_69_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_69_cast_fp16 = slice_by_index(begin = gdec_69_begin_0, end = gdec_69_end_0, end_mask = gdec_69_end_mask_0, x = gdec_57_cast_fp16)[name = string("gdec_69_cast_fp16")];
            tensor<int32, [2]> var_6729 = const()[name = string("op_6729"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_6730_cast_fp16 = reshape(shape = var_6729, x = gdec_69_cast_fp16)[name = string("op_6730_cast_fp16")];
            tensor<fp16, [4, 256]> var_6731_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_6730_cast_fp16)[name = string("op_6731_cast_fp16")];
            tensor<int32, [4]> var_6736 = const()[name = string("op_6736"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_59_cast_fp16 = reshape(shape = var_6736, x = var_6731_cast_fp16)[name = string("P_59_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6738_cast_fp16 = exp(x = P_59_cast_fp16)[name = string("op_6738_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_59_cast_fp16 = mul(x = var_6738_cast_fp16, y = tril_incl_to_fp16)[name = string("D_59_cast_fp16")];
            bool kk_59_transpose_x_1 = const()[name = string("kk_59_transpose_x_1"), val = bool(false)];
            bool kk_59_transpose_y_1 = const()[name = string("kk_59_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_59_cast_fp16 = matmul(transpose_x = kk_59_transpose_x_1, transpose_y = kk_59_transpose_y_1, x = kb_59_cast_fp16, y = k_103_cast_fp16)[name = string("kk_59_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_6744_cast_fp16 = mul(x = kk_59_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_6744_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6745_cast_fp16 = mul(x = var_6744_cast_fp16, y = D_59_cast_fp16)[name = string("op_6745_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_59_cast_fp16 = mul(x = var_6745_cast_fp16, y = cum_tail_to_fp16)[name = string("N_59_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_6747_cast_fp16 = mul(x = D_59_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_6747_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_59_cast_fp16 = mul(x = kk_59_cast_fp16, y = var_6747_cast_fp16)[name = string("Na_59_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_233_cast_fp16 = add(x = eye_to_fp16, y = N_59_cast_fp16)[name = string("Minv_233_cast_fp16")];
            bool Np_175_transpose_x_0 = const()[name = string("Np_175_transpose_x_0"), val = bool(false)];
            bool Np_175_transpose_y_0 = const()[name = string("Np_175_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_175_cast_fp16 = matmul(transpose_x = Np_175_transpose_x_0, transpose_y = Np_175_transpose_y_0, x = N_59_cast_fp16, y = Na_59_cast_fp16)[name = string("Np_175_cast_fp16")];
            bool Npa_117_transpose_x_0 = const()[name = string("Npa_117_transpose_x_0"), val = bool(false)];
            bool Npa_117_transpose_y_0 = const()[name = string("Npa_117_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_117_cast_fp16 = matmul(transpose_x = Npa_117_transpose_x_0, transpose_y = Npa_117_transpose_y_0, x = Na_59_cast_fp16, y = N_59_cast_fp16)[name = string("Npa_117_cast_fp16")];
            bool var_6753_transpose_x_0 = const()[name = string("op_6753_transpose_x_0"), val = bool(false)];
            bool var_6753_transpose_y_0 = const()[name = string("op_6753_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6753_cast_fp16 = matmul(transpose_x = var_6753_transpose_x_0, transpose_y = var_6753_transpose_y_0, x = Minv_233_cast_fp16, y = Np_175_cast_fp16)[name = string("op_6753_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_235_cast_fp16 = add(x = Minv_233_cast_fp16, y = var_6753_cast_fp16)[name = string("Minv_235_cast_fp16")];
            bool Np_177_transpose_x_0 = const()[name = string("Np_177_transpose_x_0"), val = bool(false)];
            bool Np_177_transpose_y_0 = const()[name = string("Np_177_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_177_cast_fp16 = matmul(transpose_x = Np_177_transpose_x_0, transpose_y = Np_177_transpose_y_0, x = Np_175_cast_fp16, y = Npa_117_cast_fp16)[name = string("Np_177_cast_fp16")];
            bool Npa_119_transpose_x_0 = const()[name = string("Npa_119_transpose_x_0"), val = bool(false)];
            bool Npa_119_transpose_y_0 = const()[name = string("Npa_119_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_119_cast_fp16 = matmul(transpose_x = Npa_119_transpose_x_0, transpose_y = Npa_119_transpose_y_0, x = Npa_117_cast_fp16, y = Np_175_cast_fp16)[name = string("Npa_119_cast_fp16")];
            bool var_6758_transpose_x_0 = const()[name = string("op_6758_transpose_x_0"), val = bool(false)];
            bool var_6758_transpose_y_0 = const()[name = string("op_6758_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6758_cast_fp16 = matmul(transpose_x = var_6758_transpose_x_0, transpose_y = var_6758_transpose_y_0, x = Minv_235_cast_fp16, y = Np_177_cast_fp16)[name = string("op_6758_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_237_cast_fp16 = add(x = Minv_235_cast_fp16, y = var_6758_cast_fp16)[name = string("Minv_237_cast_fp16")];
            bool Np_179_transpose_x_0 = const()[name = string("Np_179_transpose_x_0"), val = bool(false)];
            bool Np_179_transpose_y_0 = const()[name = string("Np_179_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_179_cast_fp16 = matmul(transpose_x = Np_179_transpose_x_0, transpose_y = Np_179_transpose_y_0, x = Np_177_cast_fp16, y = Npa_119_cast_fp16)[name = string("Np_179_cast_fp16")];
            bool var_6762_transpose_x_0 = const()[name = string("op_6762_transpose_x_0"), val = bool(false)];
            bool var_6762_transpose_y_0 = const()[name = string("op_6762_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_6762_cast_fp16 = matmul(transpose_x = var_6762_transpose_x_0, transpose_y = var_6762_transpose_y_0, x = Minv_237_cast_fp16, y = Np_179_cast_fp16)[name = string("op_6762_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_239_cast_fp16 = add(x = Minv_237_cast_fp16, y = var_6762_cast_fp16)[name = string("Minv_239_cast_fp16")];
            bool u_59_transpose_x_0 = const()[name = string("u_59_transpose_x_0"), val = bool(false)];
            bool u_59_transpose_y_0 = const()[name = string("u_59_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_59_cast_fp16 = matmul(transpose_x = u_59_transpose_x_0, transpose_y = u_59_transpose_y_0, x = Minv_239_cast_fp16, y = vb_69_cast_fp16)[name = string("u_59_cast_fp16")];
            bool var_6769_transpose_x_1 = const()[name = string("op_6769_transpose_x_1"), val = bool(false)];
            bool var_6769_transpose_y_1 = const()[name = string("op_6769_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_6769_cast_fp16 = matmul(transpose_x = var_6769_transpose_x_1, transpose_y = var_6769_transpose_y_1, x = q_101_cast_fp16, y = k_103_cast_fp16)[name = string("op_6769_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_59_cast_fp16 = mul(x = var_6769_cast_fp16, y = D_59_cast_fp16)[name = string("A_59_cast_fp16")];
            tensor<fp16, [16]> var_6771_bias_0_to_fp16 = const()[name = string("op_6771_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_6771_cast_fp16 = linear(bias = var_6771_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_69_cast_fp16)[name = string("op_6771_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_49_cast_fp16 = exp(x = var_6771_cast_fp16)[name = string("cum_49_cast_fp16")];
            tensor<int32, [1]> var_6774_axes_0 = const()[name = string("op_6774_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_6774_cast_fp16 = expand_dims(axes = var_6774_axes_0, x = cum_49_cast_fp16)[name = string("op_6774_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6775_cast_fp16 = mul(x = kb_59_cast_fp16, y = var_6774_cast_fp16)[name = string("op_6775_cast_fp16")];
            bool w_59_transpose_x_0 = const()[name = string("w_59_transpose_x_0"), val = bool(false)];
            bool w_59_transpose_y_0 = const()[name = string("w_59_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_59_cast_fp16 = matmul(transpose_x = w_59_transpose_x_0, transpose_y = w_59_transpose_y_0, x = Minv_239_cast_fp16, y = var_6775_cast_fp16)[name = string("w_59_cast_fp16")];
            bool var_6777_transpose_x_0 = const()[name = string("op_6777_transpose_x_0"), val = bool(false)];
            bool var_6777_transpose_y_0 = const()[name = string("op_6777_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6777_cast_fp16 = matmul(transpose_x = var_6777_transpose_x_0, transpose_y = var_6777_transpose_y_0, x = w_59_cast_fp16, y = S_49_cast_fp16)[name = string("op_6777_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_49_cast_fp16 = sub(x = u_59_cast_fp16, y = var_6777_cast_fp16)[name = string("vnew_49_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_6782_cast_fp16 = mul(x = q_101_cast_fp16, y = var_6774_cast_fp16)[name = string("op_6782_cast_fp16")];
            bool var_6783_transpose_x_0 = const()[name = string("op_6783_transpose_x_0"), val = bool(false)];
            bool var_6783_transpose_y_0 = const()[name = string("op_6783_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6783_cast_fp16 = matmul(transpose_x = var_6783_transpose_x_0, transpose_y = var_6783_transpose_y_0, x = var_6782_cast_fp16, y = S_49_cast_fp16)[name = string("op_6783_cast_fp16")];
            bool var_6784_transpose_x_0 = const()[name = string("op_6784_transpose_x_0"), val = bool(false)];
            bool var_6784_transpose_y_0 = const()[name = string("op_6784_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_6784_cast_fp16 = matmul(transpose_x = var_6784_transpose_x_0, transpose_y = var_6784_transpose_y_0, x = A_59_cast_fp16, y = vnew_49_cast_fp16)[name = string("op_6784_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_9_cast_fp16 = add(x = var_6783_cast_fp16, y = var_6784_cast_fp16)[name = string("oc_9_cast_fp16")];
            int32 var_6788 = const()[name = string("op_6788"), val = int32(2)];
            bool o_27_interleave_0 = const()[name = string("o_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 96, 128]> o_27_cast_fp16 = concat(axis = var_6788, interleave = o_27_interleave_0, values = (var_6023_cast_fp16, var_6170_cast_fp16, var_6324_cast_fp16, var_6478_cast_fp16, var_6632_cast_fp16, oc_9_cast_fp16))[name = string("o_27_cast_fp16")];
            tensor<int32, [4]> var_6794 = const()[name = string("op_6794"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_6800 = const()[name = string("op_6800"), val = tensor<int32, [4]>([1, 96, 4, 128])];
            tensor<fp16, [1, 96, 4, 128]> input_81_cast_fp16 = reshape(shape = var_6800, x = z_9_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> o_29_cast_fp16 = transpose(perm = var_6794, x = o_27_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 96, 4, 128]> var_6802_cast_fp16 = mul(x = o_29_cast_fp16, y = o_29_cast_fp16)[name = string("op_6802_cast_fp16")];
            tensor<int32, [1]> var_6807_axes_0 = const()[name = string("op_6807_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6807_keep_dims_0 = const()[name = string("op_6807_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 4, 1]> var_6807_cast_fp16 = reduce_mean(axes = var_6807_axes_0, keep_dims = var_6807_keep_dims_0, x = var_6802_cast_fp16)[name = string("op_6807_cast_fp16")];
            fp16 var_6809_to_fp16 = const()[name = string("op_6809_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 96, 4, 1]> var_6810_cast_fp16 = add(x = var_6807_cast_fp16, y = var_6809_to_fp16)[name = string("op_6810_cast_fp16")];
            fp32 var_6811_epsilon_0 = const()[name = string("op_6811_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 4, 1]> var_6811_cast_fp16 = rsqrt(epsilon = var_6811_epsilon_0, x = var_6810_cast_fp16)[name = string("op_6811_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> var_6812_cast_fp16 = mul(x = o_29_cast_fp16, y = var_6811_cast_fp16)[name = string("op_6812_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> var_6814_cast_fp16 = silu(x = input_81_cast_fp16)[name = string("op_6814_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> o_31_cast_fp16 = mul(x = var_6812_cast_fp16, y = var_6814_cast_fp16)[name = string("o_31_cast_fp16")];
            tensor<int32, [3]> var_6819 = const()[name = string("op_6819"), val = tensor<int32, [3]>([1, 96, 512])];
            tensor<fp16, [1, 96, 512]> input_83_cast_fp16 = reshape(shape = var_6819, x = o_31_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [640, 512]> layers_5_out_weight_to_fp16 = const()[name = string("layers_5_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84536512)))];
            tensor<fp16, [1, 96, 640]> linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_out_weight_to_fp16, x = input_83_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [1, 96, 640]> x_57_cast_fp16 = add(x = x_53_cast_fp16, y = linear_21_cast_fp16)[name = string("x_57_cast_fp16")];
            fp16 var_6832_to_fp16 = const()[name = string("op_6832_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 96, 640]> x_59_cast_fp16 = mul(x = x_57_cast_fp16, y = var_6832_to_fp16)[name = string("x_59_cast_fp16")];
            fp16 var_6830_promoted_to_fp16 = const()[name = string("op_6830_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> var_6834_cast_fp16 = pow(x = x_59_cast_fp16, y = var_6830_promoted_to_fp16)[name = string("op_6834_cast_fp16")];
            tensor<int32, [1]> var_6836_axes_0 = const()[name = string("op_6836_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6836_keep_dims_0 = const()[name = string("op_6836_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 1]> var_6836_cast_fp16 = reduce_mean(axes = var_6836_axes_0, keep_dims = var_6836_keep_dims_0, x = var_6834_cast_fp16)[name = string("op_6836_cast_fp16")];
            fp16 var_6837_to_fp16 = const()[name = string("op_6837_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [1, 96, 1]> var_6838_cast_fp16 = add(x = var_6836_cast_fp16, y = var_6837_to_fp16)[name = string("op_6838_cast_fp16")];
            fp32 var_6839_epsilon_0 = const()[name = string("op_6839_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 1]> var_6839_cast_fp16 = rsqrt(epsilon = var_6839_epsilon_0, x = var_6838_cast_fp16)[name = string("op_6839_cast_fp16")];
            tensor<fp16, [1, 96, 640]> var_6840_cast_fp16 = mul(x = x_59_cast_fp16, y = var_6839_cast_fp16)[name = string("op_6840_cast_fp16")];
            tensor<fp16, [640]> layers_5_n2_w1_to_fp16 = const()[name = string("layers_5_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85191936)))];
            tensor<fp16, [1, 96, 640]> input_85_cast_fp16 = mul(x = var_6840_cast_fp16, y = layers_5_n2_w1_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_5_gate_up_weight_to_fp16 = const()[name = string("layers_5_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85193280)))];
            tensor<fp16, [1, 96, 5120]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_5_gate_up_weight_to_fp16, x = input_85_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [3]> input_87_begin_0 = const()[name = string("input_87_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_87_end_0 = const()[name = string("input_87_end_0"), val = tensor<int32, [3]>([1, 96, 2560])];
            tensor<bool, [3]> input_87_end_mask_0 = const()[name = string("input_87_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 2560]> input_87_cast_fp16 = slice_by_index(begin = input_87_begin_0, end = input_87_end_0, end_mask = input_87_end_mask_0, x = linear_22_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [1, 96, 2560]> var_6850_cast_fp16 = silu(x = input_87_cast_fp16)[name = string("op_6850_cast_fp16")];
            tensor<int32, [3]> var_6855_begin_0 = const()[name = string("op_6855_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_6855_end_0 = const()[name = string("op_6855_end_0"), val = tensor<int32, [3]>([1, 96, 5120])];
            tensor<bool, [3]> var_6855_end_mask_0 = const()[name = string("op_6855_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 2560]> var_6855_cast_fp16 = slice_by_index(begin = var_6855_begin_0, end = var_6855_end_0, end_mask = var_6855_end_mask_0, x = linear_22_cast_fp16)[name = string("op_6855_cast_fp16")];
            tensor<fp16, [1, 96, 2560]> input_89_cast_fp16 = mul(x = var_6850_cast_fp16, y = var_6855_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_5_down_weight_to_fp16 = const()[name = string("layers_5_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91746944)))];
            tensor<fp16, [1, 96, 640]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_down_weight_to_fp16, x = input_89_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<fp16, [1, 96, 640]> x_61_cast_fp16 = add(x = x_57_cast_fp16, y = linear_23_cast_fp16)[name = string("x_61_cast_fp16")];
            fp16 var_6868_to_fp16 = const()[name = string("op_6868_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [1, 96, 640]> x_63_cast_fp16 = mul(x = x_61_cast_fp16, y = var_6868_to_fp16)[name = string("x_63_cast_fp16")];
            fp16 var_6866_promoted_to_fp16 = const()[name = string("op_6866_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> var_6870_cast_fp16 = pow(x = x_63_cast_fp16, y = var_6866_promoted_to_fp16)[name = string("op_6870_cast_fp16")];
            tensor<int32, [1]> var_6872_axes_0 = const()[name = string("op_6872_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6872_keep_dims_0 = const()[name = string("op_6872_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 1]> var_6872_cast_fp16 = reduce_mean(axes = var_6872_axes_0, keep_dims = var_6872_keep_dims_0, x = var_6870_cast_fp16)[name = string("op_6872_cast_fp16")];
            fp16 var_6873_to_fp16 = const()[name = string("op_6873_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 96, 1]> var_6874_cast_fp16 = add(x = var_6872_cast_fp16, y = var_6873_to_fp16)[name = string("op_6874_cast_fp16")];
            fp32 var_6875_epsilon_0 = const()[name = string("op_6875_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 1]> var_6875_cast_fp16 = rsqrt(epsilon = var_6875_epsilon_0, x = var_6874_cast_fp16)[name = string("op_6875_cast_fp16")];
            tensor<fp16, [1, 96, 640]> var_6876_cast_fp16 = mul(x = x_63_cast_fp16, y = var_6875_cast_fp16)[name = string("op_6876_cast_fp16")];
            tensor<fp16, [640]> layers_6_n1_w1_to_fp16 = const()[name = string("layers_6_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95023808)))];
            tensor<fp16, [1, 96, 640]> input_91_cast_fp16 = mul(x = var_6876_cast_fp16, y = layers_6_n1_w1_to_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_6_qkvzba_weight_to_fp16 = const()[name = string("layers_6_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95025152)))];
            tensor<fp16, [1, 96, 1544]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_qkvzba_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<int32, [3]> qkv_11_begin_0 = const()[name = string("qkv_11_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_11_end_0 = const()[name = string("qkv_11_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> qkv_11_end_mask_0 = const()[name = string("qkv_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 1024]> qkv_11_cast_fp16 = slice_by_index(begin = qkv_11_begin_0, end = qkv_11_end_0, end_mask = qkv_11_end_mask_0, x = linear_24_cast_fp16)[name = string("qkv_11_cast_fp16")];
            tensor<int32, [3]> z_begin_0 = const()[name = string("z_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_end_0 = const()[name = string("z_end_0"), val = tensor<int32, [3]>([1, 96, 1536])];
            tensor<bool, [3]> z_end_mask_0 = const()[name = string("z_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 512]> z_cast_fp16 = slice_by_index(begin = z_begin_0, end = z_end_0, end_mask = z_end_mask_0, x = linear_24_cast_fp16)[name = string("z_cast_fp16")];
            tensor<int32, [3]> b_begin_0 = const()[name = string("b_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_end_0 = const()[name = string("b_end_0"), val = tensor<int32, [3]>([1, 96, 1540])];
            tensor<bool, [3]> b_end_mask_0 = const()[name = string("b_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 4]> b_cast_fp16 = slice_by_index(begin = b_begin_0, end = b_end_0, end_mask = b_end_mask_0, x = linear_24_cast_fp16)[name = string("b_cast_fp16")];
            tensor<int32, [3]> a_begin_0 = const()[name = string("a_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_end_0 = const()[name = string("a_end_0"), val = tensor<int32, [3]>([1, 96, 1])];
            tensor<bool, [3]> a_end_mask_0 = const()[name = string("a_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 4]> a_cast_fp16 = slice_by_index(begin = a_begin_0, end = a_end_0, end_mask = a_end_mask_0, x = linear_24_cast_fp16)[name = string("a_cast_fp16")];
            int32 var_6915 = const()[name = string("op_6915"), val = int32(1)];
            bool xp_interleave_0 = const()[name = string("xp_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 99, 1024]> xp_cast_fp16 = concat(axis = var_6915, interleave = xp_interleave_0, values = (zpad_to_fp16, qkv_11_cast_fp16))[name = string("xp_cast_fp16")];
            tensor<int32, [3]> var_6926_begin_0 = const()[name = string("op_6926_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_6926_end_0 = const()[name = string("op_6926_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_6926_end_mask_0 = const()[name = string("op_6926_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 1024]> var_6926_cast_fp16 = slice_by_index(begin = var_6926_begin_0, end = var_6926_end_0, end_mask = var_6926_end_mask_0, x = xp_cast_fp16)[name = string("op_6926_cast_fp16")];
            tensor<fp16, [1024]> var_6934_to_fp16 = const()[name = string("op_6934_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97001536)))];
            tensor<fp16, [1, 96, 1024]> var_6935_cast_fp16 = mul(x = var_6926_cast_fp16, y = var_6934_to_fp16)[name = string("op_6935_cast_fp16")];
            tensor<int32, [3]> var_6945_begin_0 = const()[name = string("op_6945_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_6945_end_0 = const()[name = string("op_6945_end_0"), val = tensor<int32, [3]>([1, 98, 1024])];
            tensor<bool, [3]> var_6945_end_mask_0 = const()[name = string("op_6945_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_6945_cast_fp16 = slice_by_index(begin = var_6945_begin_0, end = var_6945_end_0, end_mask = var_6945_end_mask_0, x = xp_cast_fp16)[name = string("op_6945_cast_fp16")];
            tensor<fp16, [1024]> var_6953_to_fp16 = const()[name = string("op_6953_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97003648)))];
            tensor<fp16, [1, 96, 1024]> var_6954_cast_fp16 = mul(x = var_6945_cast_fp16, y = var_6953_to_fp16)[name = string("op_6954_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> var_6956_cast_fp16 = add(x = var_6935_cast_fp16, y = var_6954_cast_fp16)[name = string("op_6956_cast_fp16")];
            tensor<int32, [3]> var_6966_begin_0 = const()[name = string("op_6966_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_6966_end_0 = const()[name = string("op_6966_end_0"), val = tensor<int32, [3]>([1, 97, 1024])];
            tensor<bool, [3]> var_6966_end_mask_0 = const()[name = string("op_6966_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_6966_cast_fp16 = slice_by_index(begin = var_6966_begin_0, end = var_6966_end_0, end_mask = var_6966_end_mask_0, x = xp_cast_fp16)[name = string("op_6966_cast_fp16")];
            tensor<fp16, [1024]> var_6974_to_fp16 = const()[name = string("op_6974_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97005760)))];
            tensor<fp16, [1, 96, 1024]> var_6975_cast_fp16 = mul(x = var_6966_cast_fp16, y = var_6974_to_fp16)[name = string("op_6975_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> var_6977_cast_fp16 = add(x = var_6956_cast_fp16, y = var_6975_cast_fp16)[name = string("op_6977_cast_fp16")];
            tensor<int32, [3]> var_6987_begin_0 = const()[name = string("op_6987_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_6987_end_0 = const()[name = string("op_6987_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> var_6987_end_mask_0 = const()[name = string("op_6987_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 96, 1024]> var_6987_cast_fp16 = slice_by_index(begin = var_6987_begin_0, end = var_6987_end_0, end_mask = var_6987_end_mask_0, x = xp_cast_fp16)[name = string("op_6987_cast_fp16")];
            tensor<fp16, [1024]> var_6995_to_fp16 = const()[name = string("op_6995_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97007872)))];
            tensor<fp16, [1, 96, 1024]> var_6996_cast_fp16 = mul(x = var_6987_cast_fp16, y = var_6995_to_fp16)[name = string("op_6996_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> input_93_cast_fp16 = add(x = var_6977_cast_fp16, y = var_6996_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [1, 96, 1024]> c_cast_fp16 = silu(x = input_93_cast_fp16)[name = string("c_cast_fp16")];
            tensor<int32, [3]> var_7004_begin_0 = const()[name = string("op_7004_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_7004_end_0 = const()[name = string("op_7004_end_0"), val = tensor<int32, [3]>([1, 96, 256])];
            tensor<bool, [3]> var_7004_end_mask_0 = const()[name = string("op_7004_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 256]> var_7004_cast_fp16 = slice_by_index(begin = var_7004_begin_0, end = var_7004_end_0, end_mask = var_7004_end_mask_0, x = c_cast_fp16)[name = string("op_7004_cast_fp16")];
            tensor<int32, [4]> var_7009 = const()[name = string("op_7009"), val = tensor<int32, [4]>([1, 96, 2, 128])];
            tensor<fp16, [1, 96, 2, 128]> q_103_cast_fp16 = reshape(shape = var_7009, x = var_7004_cast_fp16)[name = string("q_103_cast_fp16")];
            tensor<int32, [3]> var_7015_begin_0 = const()[name = string("op_7015_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_7015_end_0 = const()[name = string("op_7015_end_0"), val = tensor<int32, [3]>([1, 96, 512])];
            tensor<bool, [3]> var_7015_end_mask_0 = const()[name = string("op_7015_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 256]> var_7015_cast_fp16 = slice_by_index(begin = var_7015_begin_0, end = var_7015_end_0, end_mask = var_7015_end_mask_0, x = c_cast_fp16)[name = string("op_7015_cast_fp16")];
            tensor<int32, [4]> var_7020 = const()[name = string("op_7020"), val = tensor<int32, [4]>([1, 96, 2, 128])];
            tensor<fp16, [1, 96, 2, 128]> k_105_cast_fp16 = reshape(shape = var_7020, x = var_7015_cast_fp16)[name = string("k_105_cast_fp16")];
            tensor<int32, [3]> var_7026_begin_0 = const()[name = string("op_7026_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_7026_end_0 = const()[name = string("op_7026_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> var_7026_end_mask_0 = const()[name = string("op_7026_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 512]> var_7026_cast_fp16 = slice_by_index(begin = var_7026_begin_0, end = var_7026_end_0, end_mask = var_7026_end_mask_0, x = c_cast_fp16)[name = string("op_7026_cast_fp16")];
            tensor<int32, [4]> var_7031 = const()[name = string("op_7031"), val = tensor<int32, [4]>([1, 96, 4, 128])];
            tensor<fp16, [1, 96, 4, 128]> var_7032_cast_fp16 = reshape(shape = var_7031, x = var_7026_cast_fp16)[name = string("op_7032_cast_fp16")];
            tensor<int32, [4]> var_7037 = const()[name = string("op_7037"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7039_to_fp16 = const()[name = string("op_7039_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 96, 2, 128]> q_105_cast_fp16 = mul(x = q_103_cast_fp16, y = var_7039_to_fp16)[name = string("q_105_cast_fp16")];
            fp16 var_7041_to_fp16 = const()[name = string("op_7041_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [1, 96, 2, 128]> k_107_cast_fp16 = mul(x = k_105_cast_fp16, y = var_7041_to_fp16)[name = string("k_107_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_7043_cast_fp16 = mul(x = q_105_cast_fp16, y = q_105_cast_fp16)[name = string("op_7043_cast_fp16")];
            tensor<int32, [1]> var_7048_axes_0 = const()[name = string("op_7048_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7048_keep_dims_0 = const()[name = string("op_7048_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 2, 1]> var_7048_cast_fp16 = reduce_mean(axes = var_7048_axes_0, keep_dims = var_7048_keep_dims_0, x = var_7043_cast_fp16)[name = string("op_7048_cast_fp16")];
            fp16 var_7049_promoted_to_fp16 = const()[name = string("op_7049_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 96, 2, 1]> var_7050_cast_fp16 = mul(x = var_7048_cast_fp16, y = var_7049_promoted_to_fp16)[name = string("op_7050_cast_fp16")];
            fp16 var_7052_to_fp16 = const()[name = string("op_7052_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 96, 2, 1]> var_7053_cast_fp16 = add(x = var_7050_cast_fp16, y = var_7052_to_fp16)[name = string("op_7053_cast_fp16")];
            fp32 var_7054_epsilon_0 = const()[name = string("op_7054_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 2, 1]> var_7054_cast_fp16 = rsqrt(epsilon = var_7054_epsilon_0, x = var_7053_cast_fp16)[name = string("op_7054_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_7055_cast_fp16 = mul(x = q_105_cast_fp16, y = var_7054_cast_fp16)[name = string("op_7055_cast_fp16")];
            fp16 var_7056_to_fp16 = const()[name = string("op_7056_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [1, 96, 2, 128]> q_107_cast_fp16 = mul(x = var_7055_cast_fp16, y = var_7056_to_fp16)[name = string("q_107_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> var_7058_cast_fp16 = mul(x = k_107_cast_fp16, y = k_107_cast_fp16)[name = string("op_7058_cast_fp16")];
            tensor<int32, [1]> var_7063_axes_0 = const()[name = string("op_7063_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7063_keep_dims_0 = const()[name = string("op_7063_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 2, 1]> var_7063_cast_fp16 = reduce_mean(axes = var_7063_axes_0, keep_dims = var_7063_keep_dims_0, x = var_7058_cast_fp16)[name = string("op_7063_cast_fp16")];
            fp16 var_7064_promoted_to_fp16 = const()[name = string("op_7064_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [1, 96, 2, 1]> var_7065_cast_fp16 = mul(x = var_7063_cast_fp16, y = var_7064_promoted_to_fp16)[name = string("op_7065_cast_fp16")];
            fp16 var_7067_to_fp16 = const()[name = string("op_7067_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [1, 96, 2, 1]> var_7068_cast_fp16 = add(x = var_7065_cast_fp16, y = var_7067_to_fp16)[name = string("op_7068_cast_fp16")];
            fp32 var_7069_epsilon_0 = const()[name = string("op_7069_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 2, 1]> var_7069_cast_fp16 = rsqrt(epsilon = var_7069_epsilon_0, x = var_7068_cast_fp16)[name = string("op_7069_cast_fp16")];
            tensor<fp16, [1, 96, 2, 128]> k_109_cast_fp16 = mul(x = k_107_cast_fp16, y = var_7069_cast_fp16)[name = string("k_109_cast_fp16")];
            tensor<int32, [4]> var_7075 = const()[name = string("op_7075"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_7088_axes_0 = const()[name = string("op_7088_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 96, 128]> var_7076_cast_fp16 = transpose(perm = var_7075, x = q_107_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 2, 1, 96, 128]> var_7088_cast_fp16 = expand_dims(axes = var_7088_axes_0, x = var_7076_cast_fp16)[name = string("op_7088_cast_fp16")];
            tensor<int32, [5]> var_7096_reps_0 = const()[name = string("op_7096_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 96, 128]> var_7096_cast_fp16 = tile(reps = var_7096_reps_0, x = var_7088_cast_fp16)[name = string("op_7096_cast_fp16")];
            tensor<int32, [4]> var_7101 = const()[name = string("op_7101"), val = tensor<int32, [4]>([1, 4, 96, 128])];
            tensor<fp16, [1, 4, 96, 128]> q_109_cast_fp16 = reshape(shape = var_7101, x = var_7096_cast_fp16)[name = string("q_109_cast_fp16")];
            tensor<int32, [4]> var_7107 = const()[name = string("op_7107"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_7120_axes_0 = const()[name = string("op_7120_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 96, 128]> var_7108_cast_fp16 = transpose(perm = var_7107, x = k_109_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 2, 1, 96, 128]> var_7120_cast_fp16 = expand_dims(axes = var_7120_axes_0, x = var_7108_cast_fp16)[name = string("op_7120_cast_fp16")];
            tensor<int32, [5]> var_7128_reps_0 = const()[name = string("op_7128_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 2, 2, 96, 128]> var_7128_cast_fp16 = tile(reps = var_7128_reps_0, x = var_7120_cast_fp16)[name = string("op_7128_cast_fp16")];
            tensor<int32, [4]> var_7133 = const()[name = string("op_7133"), val = tensor<int32, [4]>([1, 4, 96, 128])];
            tensor<fp16, [1, 4, 96, 128]> k_111_cast_fp16 = reshape(shape = var_7133, x = var_7128_cast_fp16)[name = string("k_111_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_7135_cast_fp16 = sigmoid(x = b_cast_fp16)[name = string("op_7135_cast_fp16")];
            tensor<int32, [3]> var_7139 = const()[name = string("op_7139"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_6_dt_bias_to_fp16 = const()[name = string("layers_6_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [1, 96, 4]> input_95_cast_fp16 = add(x = a_cast_fp16, y = layers_6_dt_bias_to_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_7143_cast_fp16 = relu(x = input_95_cast_fp16)[name = string("op_7143_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_7144_cast_fp16 = abs(x = input_95_cast_fp16)[name = string("op_7144_cast_fp16")];
            fp16 const_37_promoted_to_fp16 = const()[name = string("const_37_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 96, 4]> var_7145_cast_fp16 = mul(x = var_7144_cast_fp16, y = const_37_promoted_to_fp16)[name = string("op_7145_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_7146_cast_fp16 = exp(x = var_7145_cast_fp16)[name = string("op_7146_cast_fp16")];
            fp16 var_7148_to_fp16 = const()[name = string("op_7148_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 96, 4]> var_7149_cast_fp16 = add(x = var_7146_cast_fp16, y = var_7148_to_fp16)[name = string("op_7149_cast_fp16")];
            fp32 var_7150_epsilon_0 = const()[name = string("op_7150_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 96, 4]> var_7150_cast_fp16 = log(epsilon = var_7150_epsilon_0, x = var_7149_cast_fp16)[name = string("op_7150_cast_fp16")];
            tensor<fp16, [1, 96, 4]> var_7152_cast_fp16 = add(x = var_7143_cast_fp16, y = var_7150_cast_fp16)[name = string("op_7152_cast_fp16")];
            tensor<fp16, [4]> layers_6_negA_to_fp16 = const()[name = string("layers_6_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.07cp+1, -0x1.eccp+3, -0x1.63p+3, -0x1.728p+1])];
            tensor<fp16, [1, 96, 4]> var_7153_cast_fp16 = mul(x = layers_6_negA_to_fp16, y = var_7152_cast_fp16)[name = string("op_7153_cast_fp16")];
            tensor<int32, [3]> var_7157 = const()[name = string("op_7157"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_7160_axes_0 = const()[name = string("op_7160_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 96]> beta_cast_fp16 = transpose(perm = var_7139, x = var_7135_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 4, 96, 1]> var_7160_cast_fp16 = expand_dims(axes = var_7160_axes_0, x = beta_cast_fp16)[name = string("op_7160_cast_fp16")];
            fp16 var_7161_to_fp16 = const()[name = string("op_7161_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [1, 4, 96, 1]> var_7162_cast_fp16 = mul(x = var_7160_cast_fp16, y = var_7161_to_fp16)[name = string("op_7162_cast_fp16")];
            tensor<fp16, [1, 4, 96, 128]> v_15_cast_fp16 = transpose(perm = var_7037, x = var_7032_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 4, 96, 128]> vb_71_cast_fp16 = mul(x = v_15_cast_fp16, y = var_7162_cast_fp16)[name = string("vb_71_cast_fp16")];
            tensor<int32, [4]> q_111_begin_0 = const()[name = string("q_111_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_111_end_0 = const()[name = string("q_111_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> q_111_end_mask_0 = const()[name = string("q_111_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_111_cast_fp16 = slice_by_index(begin = q_111_begin_0, end = q_111_end_0, end_mask = q_111_end_mask_0, x = q_109_cast_fp16)[name = string("q_111_cast_fp16")];
            tensor<int32, [4]> k_113_begin_0 = const()[name = string("k_113_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_113_end_0 = const()[name = string("k_113_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> k_113_end_mask_0 = const()[name = string("k_113_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_113_cast_fp16 = slice_by_index(begin = k_113_begin_0, end = k_113_end_0, end_mask = k_113_end_mask_0, x = k_111_cast_fp16)[name = string("k_113_cast_fp16")];
            tensor<fp16, [1, 4, 96, 128]> var_7196_cast_fp16 = mul(x = k_111_cast_fp16, y = var_7160_cast_fp16)[name = string("op_7196_cast_fp16")];
            tensor<int32, [4]> kb_61_begin_0 = const()[name = string("kb_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_61_end_0 = const()[name = string("kb_61_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> kb_61_end_mask_0 = const()[name = string("kb_61_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_61_cast_fp16 = slice_by_index(begin = kb_61_begin_0, end = kb_61_end_0, end_mask = kb_61_end_mask_0, x = var_7196_cast_fp16)[name = string("kb_61_cast_fp16")];
            tensor<int32, [4]> vb_73_begin_0 = const()[name = string("vb_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_73_end_0 = const()[name = string("vb_73_end_0"), val = tensor<int32, [4]>([1, 4, 16, 128])];
            tensor<bool, [4]> vb_73_end_mask_0 = const()[name = string("vb_73_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_73_cast_fp16 = slice_by_index(begin = vb_73_begin_0, end = vb_73_end_0, end_mask = vb_73_end_mask_0, x = vb_71_cast_fp16)[name = string("vb_73_cast_fp16")];
            tensor<int32, [3]> gdec_73_begin_0 = const()[name = string("gdec_73_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_73_end_0 = const()[name = string("gdec_73_end_0"), val = tensor<int32, [3]>([1, 4, 16])];
            tensor<bool, [3]> gdec_73_end_mask_0 = const()[name = string("gdec_73_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 96]> gdec_71_cast_fp16 = transpose(perm = var_7157, x = var_7153_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [1, 4, 16]> gdec_73_cast_fp16 = slice_by_index(begin = gdec_73_begin_0, end = gdec_73_end_0, end_mask = gdec_73_end_mask_0, x = gdec_71_cast_fp16)[name = string("gdec_73_cast_fp16")];
            tensor<int32, [2]> var_7244 = const()[name = string("op_7244"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_7245_cast_fp16 = reshape(shape = var_7244, x = gdec_73_cast_fp16)[name = string("op_7245_cast_fp16")];
            tensor<fp16, [4, 256]> var_7246_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_7245_cast_fp16)[name = string("op_7246_cast_fp16")];
            tensor<int32, [4]> var_7251 = const()[name = string("op_7251"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_61_cast_fp16 = reshape(shape = var_7251, x = var_7246_cast_fp16)[name = string("P_61_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_7253_cast_fp16 = exp(x = P_61_cast_fp16)[name = string("op_7253_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_61_cast_fp16 = mul(x = var_7253_cast_fp16, y = tril_incl_to_fp16)[name = string("D_61_cast_fp16")];
            bool kk_61_transpose_x_1 = const()[name = string("kk_61_transpose_x_1"), val = bool(false)];
            bool kk_61_transpose_y_1 = const()[name = string("kk_61_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_61_cast_fp16 = matmul(transpose_x = kk_61_transpose_x_1, transpose_y = kk_61_transpose_y_1, x = kb_61_cast_fp16, y = k_113_cast_fp16)[name = string("kk_61_cast_fp16")];
            fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_7259_cast_fp16 = mul(x = kk_61_cast_fp16, y = const_38_promoted_to_fp16)[name = string("op_7259_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_7260_cast_fp16 = mul(x = var_7259_cast_fp16, y = D_61_cast_fp16)[name = string("op_7260_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_61_cast_fp16 = mul(x = var_7260_cast_fp16, y = cum_tail_to_fp16)[name = string("N_61_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_7262_cast_fp16 = mul(x = D_61_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_7262_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_61_cast_fp16 = mul(x = kk_61_cast_fp16, y = var_7262_cast_fp16)[name = string("Na_61_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_241_cast_fp16 = add(x = eye_to_fp16, y = N_61_cast_fp16)[name = string("Minv_241_cast_fp16")];
            bool Np_181_transpose_x_0 = const()[name = string("Np_181_transpose_x_0"), val = bool(false)];
            bool Np_181_transpose_y_0 = const()[name = string("Np_181_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_181_cast_fp16 = matmul(transpose_x = Np_181_transpose_x_0, transpose_y = Np_181_transpose_y_0, x = N_61_cast_fp16, y = Na_61_cast_fp16)[name = string("Np_181_cast_fp16")];
            bool Npa_121_transpose_x_0 = const()[name = string("Npa_121_transpose_x_0"), val = bool(false)];
            bool Npa_121_transpose_y_0 = const()[name = string("Npa_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_121_cast_fp16 = matmul(transpose_x = Npa_121_transpose_x_0, transpose_y = Npa_121_transpose_y_0, x = Na_61_cast_fp16, y = N_61_cast_fp16)[name = string("Npa_121_cast_fp16")];
            bool var_7268_transpose_x_0 = const()[name = string("op_7268_transpose_x_0"), val = bool(false)];
            bool var_7268_transpose_y_0 = const()[name = string("op_7268_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_7268_cast_fp16 = matmul(transpose_x = var_7268_transpose_x_0, transpose_y = var_7268_transpose_y_0, x = Minv_241_cast_fp16, y = Np_181_cast_fp16)[name = string("op_7268_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_243_cast_fp16 = add(x = Minv_241_cast_fp16, y = var_7268_cast_fp16)[name = string("Minv_243_cast_fp16")];
            bool Np_183_transpose_x_0 = const()[name = string("Np_183_transpose_x_0"), val = bool(false)];
            bool Np_183_transpose_y_0 = const()[name = string("Np_183_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_183_cast_fp16 = matmul(transpose_x = Np_183_transpose_x_0, transpose_y = Np_183_transpose_y_0, x = Np_181_cast_fp16, y = Npa_121_cast_fp16)[name = string("Np_183_cast_fp16")];
            bool Npa_123_transpose_x_0 = const()[name = string("Npa_123_transpose_x_0"), val = bool(false)];
            bool Npa_123_transpose_y_0 = const()[name = string("Npa_123_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_123_cast_fp16 = matmul(transpose_x = Npa_123_transpose_x_0, transpose_y = Npa_123_transpose_y_0, x = Npa_121_cast_fp16, y = Np_181_cast_fp16)[name = string("Npa_123_cast_fp16")];
            bool var_7273_transpose_x_0 = const()[name = string("op_7273_transpose_x_0"), val = bool(false)];
            bool var_7273_transpose_y_0 = const()[name = string("op_7273_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_7273_cast_fp16 = matmul(transpose_x = var_7273_transpose_x_0, transpose_y = var_7273_transpose_y_0, x = Minv_243_cast_fp16, y = Np_183_cast_fp16)[name = string("op_7273_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_245_cast_fp16 = add(x = Minv_243_cast_fp16, y = var_7273_cast_fp16)[name = string("Minv_245_cast_fp16")];
            bool Np_185_transpose_x_0 = const()[name = string("Np_185_transpose_x_0"), val = bool(false)];
            bool Np_185_transpose_y_0 = const()[name = string("Np_185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_185_cast_fp16 = matmul(transpose_x = Np_185_transpose_x_0, transpose_y = Np_185_transpose_y_0, x = Np_183_cast_fp16, y = Npa_123_cast_fp16)[name = string("Np_185_cast_fp16")];
            bool var_7277_transpose_x_0 = const()[name = string("op_7277_transpose_x_0"), val = bool(false)];
            bool var_7277_transpose_y_0 = const()[name = string("op_7277_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_7277_cast_fp16 = matmul(transpose_x = var_7277_transpose_x_0, transpose_y = var_7277_transpose_y_0, x = Minv_245_cast_fp16, y = Np_185_cast_fp16)[name = string("op_7277_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_247_cast_fp16 = add(x = Minv_245_cast_fp16, y = var_7277_cast_fp16)[name = string("Minv_247_cast_fp16")];
            bool u_61_transpose_x_0 = const()[name = string("u_61_transpose_x_0"), val = bool(false)];
            bool u_61_transpose_y_0 = const()[name = string("u_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_61_cast_fp16 = matmul(transpose_x = u_61_transpose_x_0, transpose_y = u_61_transpose_y_0, x = Minv_247_cast_fp16, y = vb_73_cast_fp16)[name = string("u_61_cast_fp16")];
            bool var_7284_transpose_x_1 = const()[name = string("op_7284_transpose_x_1"), val = bool(false)];
            bool var_7284_transpose_y_1 = const()[name = string("op_7284_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_7284_cast_fp16 = matmul(transpose_x = var_7284_transpose_x_1, transpose_y = var_7284_transpose_y_1, x = q_111_cast_fp16, y = k_113_cast_fp16)[name = string("op_7284_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_61_cast_fp16 = mul(x = var_7284_cast_fp16, y = D_61_cast_fp16)[name = string("A_61_cast_fp16")];
            bool var_7286_transpose_x_0 = const()[name = string("op_7286_transpose_x_0"), val = bool(false)];
            bool var_7286_transpose_y_0 = const()[name = string("op_7286_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_7286_cast_fp16 = matmul(transpose_x = var_7286_transpose_x_0, transpose_y = var_7286_transpose_y_0, x = A_61_cast_fp16, y = u_61_cast_fp16)[name = string("op_7286_cast_fp16")];
            tensor<fp16, [16]> var_7287_bias_0_to_fp16 = const()[name = string("op_7287_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_7287_cast_fp16 = linear(bias = var_7287_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_73_cast_fp16)[name = string("op_7287_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_51_cast_fp16 = exp(x = var_7287_cast_fp16)[name = string("tail_51_cast_fp16")];
            tensor<int32, [1]> var_7290_axes_0 = const()[name = string("op_7290_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_7290_cast_fp16 = expand_dims(axes = var_7290_axes_0, x = tail_51_cast_fp16)[name = string("op_7290_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_7291_cast_fp16 = mul(x = k_113_cast_fp16, y = var_7290_cast_fp16)[name = string("op_7291_cast_fp16")];
            bool S_51_transpose_x_1 = const()[name = string("S_51_transpose_x_1"), val = bool(true)];
            bool S_51_transpose_y_1 = const()[name = string("S_51_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> S_51_cast_fp16 = matmul(transpose_x = S_51_transpose_x_1, transpose_y = S_51_transpose_y_1, x = var_7291_cast_fp16, y = u_61_cast_fp16)[name = string("S_51_cast_fp16")];
            tensor<int32, [4]> q_113_begin_0 = const()[name = string("q_113_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_113_end_0 = const()[name = string("q_113_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> q_113_end_mask_0 = const()[name = string("q_113_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_113_cast_fp16 = slice_by_index(begin = q_113_begin_0, end = q_113_end_0, end_mask = q_113_end_mask_0, x = q_109_cast_fp16)[name = string("q_113_cast_fp16")];
            tensor<int32, [4]> k_115_begin_0 = const()[name = string("k_115_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_115_end_0 = const()[name = string("k_115_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> k_115_end_mask_0 = const()[name = string("k_115_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_115_cast_fp16 = slice_by_index(begin = k_115_begin_0, end = k_115_end_0, end_mask = k_115_end_mask_0, x = k_111_cast_fp16)[name = string("k_115_cast_fp16")];
            tensor<int32, [4]> kb_63_begin_0 = const()[name = string("kb_63_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_63_end_0 = const()[name = string("kb_63_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> kb_63_end_mask_0 = const()[name = string("kb_63_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_63_cast_fp16 = slice_by_index(begin = kb_63_begin_0, end = kb_63_end_0, end_mask = kb_63_end_mask_0, x = var_7196_cast_fp16)[name = string("kb_63_cast_fp16")];
            tensor<int32, [4]> vb_75_begin_0 = const()[name = string("vb_75_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_75_end_0 = const()[name = string("vb_75_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> vb_75_end_mask_0 = const()[name = string("vb_75_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_75_cast_fp16 = slice_by_index(begin = vb_75_begin_0, end = vb_75_end_0, end_mask = vb_75_end_mask_0, x = vb_71_cast_fp16)[name = string("vb_75_cast_fp16")];
            tensor<int32, [3]> gdec_75_begin_0 = const()[name = string("gdec_75_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_75_end_0 = const()[name = string("gdec_75_end_0"), val = tensor<int32, [3]>([1, 4, 32])];
            tensor<bool, [3]> gdec_75_end_mask_0 = const()[name = string("gdec_75_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_75_cast_fp16 = slice_by_index(begin = gdec_75_begin_0, end = gdec_75_end_0, end_mask = gdec_75_end_mask_0, x = gdec_71_cast_fp16)[name = string("gdec_75_cast_fp16")];
            tensor<int32, [2]> var_7376 = const()[name = string("op_7376"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_7377_cast_fp16 = reshape(shape = var_7376, x = gdec_75_cast_fp16)[name = string("op_7377_cast_fp16")];
            tensor<fp16, [4, 256]> var_7378_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_7377_cast_fp16)[name = string("op_7378_cast_fp16")];
            tensor<int32, [4]> var_7383 = const()[name = string("op_7383"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_63_cast_fp16 = reshape(shape = var_7383, x = var_7378_cast_fp16)[name = string("P_63_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_7385_cast_fp16 = exp(x = P_63_cast_fp16)[name = string("op_7385_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_63_cast_fp16 = mul(x = var_7385_cast_fp16, y = tril_incl_to_fp16)[name = string("D_63_cast_fp16")];
            bool kk_63_transpose_x_1 = const()[name = string("kk_63_transpose_x_1"), val = bool(false)];
            bool kk_63_transpose_y_1 = const()[name = string("kk_63_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_63_cast_fp16 = matmul(transpose_x = kk_63_transpose_x_1, transpose_y = kk_63_transpose_y_1, x = kb_63_cast_fp16, y = k_115_cast_fp16)[name = string("kk_63_cast_fp16")];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_7391_cast_fp16 = mul(x = kk_63_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_7391_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_7392_cast_fp16 = mul(x = var_7391_cast_fp16, y = D_63_cast_fp16)[name = string("op_7392_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_63_cast_fp16 = mul(x = var_7392_cast_fp16, y = cum_tail_to_fp16)[name = string("N_63_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_7394_cast_fp16 = mul(x = D_63_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_7394_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_63_cast_fp16 = mul(x = kk_63_cast_fp16, y = var_7394_cast_fp16)[name = string("Na_63_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_249_cast_fp16 = add(x = eye_to_fp16, y = N_63_cast_fp16)[name = string("Minv_249_cast_fp16")];
            bool Np_187_transpose_x_0 = const()[name = string("Np_187_transpose_x_0"), val = bool(false)];
            bool Np_187_transpose_y_0 = const()[name = string("Np_187_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_187_cast_fp16 = matmul(transpose_x = Np_187_transpose_x_0, transpose_y = Np_187_transpose_y_0, x = N_63_cast_fp16, y = Na_63_cast_fp16)[name = string("Np_187_cast_fp16")];
            bool Npa_125_transpose_x_0 = const()[name = string("Npa_125_transpose_x_0"), val = bool(false)];
            bool Npa_125_transpose_y_0 = const()[name = string("Npa_125_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_125_cast_fp16 = matmul(transpose_x = Npa_125_transpose_x_0, transpose_y = Npa_125_transpose_y_0, x = Na_63_cast_fp16, y = N_63_cast_fp16)[name = string("Npa_125_cast_fp16")];
            bool var_7400_transpose_x_0 = const()[name = string("op_7400_transpose_x_0"), val = bool(false)];
            bool var_7400_transpose_y_0 = const()[name = string("op_7400_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_7400_cast_fp16 = matmul(transpose_x = var_7400_transpose_x_0, transpose_y = var_7400_transpose_y_0, x = Minv_249_cast_fp16, y = Np_187_cast_fp16)[name = string("op_7400_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_251_cast_fp16 = add(x = Minv_249_cast_fp16, y = var_7400_cast_fp16)[name = string("Minv_251_cast_fp16")];
            bool Np_189_transpose_x_0 = const()[name = string("Np_189_transpose_x_0"), val = bool(false)];
            bool Np_189_transpose_y_0 = const()[name = string("Np_189_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_189_cast_fp16 = matmul(transpose_x = Np_189_transpose_x_0, transpose_y = Np_189_transpose_y_0, x = Np_187_cast_fp16, y = Npa_125_cast_fp16)[name = string("Np_189_cast_fp16")];
            bool Npa_127_transpose_x_0 = const()[name = string("Npa_127_transpose_x_0"), val = bool(false)];
            bool Npa_127_transpose_y_0 = const()[name = string("Npa_127_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_127_cast_fp16 = matmul(transpose_x = Npa_127_transpose_x_0, transpose_y = Npa_127_transpose_y_0, x = Npa_125_cast_fp16, y = Np_187_cast_fp16)[name = string("Npa_127_cast_fp16")];
            bool var_7405_transpose_x_0 = const()[name = string("op_7405_transpose_x_0"), val = bool(false)];
            bool var_7405_transpose_y_0 = const()[name = string("op_7405_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_7405_cast_fp16 = matmul(transpose_x = var_7405_transpose_x_0, transpose_y = var_7405_transpose_y_0, x = Minv_251_cast_fp16, y = Np_189_cast_fp16)[name = string("op_7405_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_253_cast_fp16 = add(x = Minv_251_cast_fp16, y = var_7405_cast_fp16)[name = string("Minv_253_cast_fp16")];
            bool Np_191_transpose_x_0 = const()[name = string("Np_191_transpose_x_0"), val = bool(false)];
            bool Np_191_transpose_y_0 = const()[name = string("Np_191_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_191_cast_fp16 = matmul(transpose_x = Np_191_transpose_x_0, transpose_y = Np_191_transpose_y_0, x = Np_189_cast_fp16, y = Npa_127_cast_fp16)[name = string("Np_191_cast_fp16")];
            bool var_7409_transpose_x_0 = const()[name = string("op_7409_transpose_x_0"), val = bool(false)];
            bool var_7409_transpose_y_0 = const()[name = string("op_7409_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_7409_cast_fp16 = matmul(transpose_x = var_7409_transpose_x_0, transpose_y = var_7409_transpose_y_0, x = Minv_253_cast_fp16, y = Np_191_cast_fp16)[name = string("op_7409_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_255_cast_fp16 = add(x = Minv_253_cast_fp16, y = var_7409_cast_fp16)[name = string("Minv_255_cast_fp16")];
            bool u_63_transpose_x_0 = const()[name = string("u_63_transpose_x_0"), val = bool(false)];
            bool u_63_transpose_y_0 = const()[name = string("u_63_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_63_cast_fp16 = matmul(transpose_x = u_63_transpose_x_0, transpose_y = u_63_transpose_y_0, x = Minv_255_cast_fp16, y = vb_75_cast_fp16)[name = string("u_63_cast_fp16")];
            bool var_7416_transpose_x_1 = const()[name = string("op_7416_transpose_x_1"), val = bool(false)];
            bool var_7416_transpose_y_1 = const()[name = string("op_7416_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_7416_cast_fp16 = matmul(transpose_x = var_7416_transpose_x_1, transpose_y = var_7416_transpose_y_1, x = q_113_cast_fp16, y = k_115_cast_fp16)[name = string("op_7416_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_63_cast_fp16 = mul(x = var_7416_cast_fp16, y = D_63_cast_fp16)[name = string("A_63_cast_fp16")];
            tensor<fp16, [16]> var_7418_bias_0_to_fp16 = const()[name = string("op_7418_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_7418_cast_fp16 = linear(bias = var_7418_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_75_cast_fp16)[name = string("op_7418_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_51_cast_fp16 = exp(x = var_7418_cast_fp16)[name = string("cum_51_cast_fp16")];
            tensor<int32, [1]> var_7421_axes_0 = const()[name = string("op_7421_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_7421_cast_fp16 = expand_dims(axes = var_7421_axes_0, x = cum_51_cast_fp16)[name = string("op_7421_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_7422_cast_fp16 = mul(x = kb_63_cast_fp16, y = var_7421_cast_fp16)[name = string("op_7422_cast_fp16")];
            bool w_63_transpose_x_0 = const()[name = string("w_63_transpose_x_0"), val = bool(false)];
            bool w_63_transpose_y_0 = const()[name = string("w_63_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_63_cast_fp16 = matmul(transpose_x = w_63_transpose_x_0, transpose_y = w_63_transpose_y_0, x = Minv_255_cast_fp16, y = var_7422_cast_fp16)[name = string("w_63_cast_fp16")];
            bool var_7424_transpose_x_0 = const()[name = string("op_7424_transpose_x_0"), val = bool(false)];
            bool var_7424_transpose_y_0 = const()[name = string("op_7424_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_7424_cast_fp16 = matmul(transpose_x = var_7424_transpose_x_0, transpose_y = var_7424_transpose_y_0, x = w_63_cast_fp16, y = S_51_cast_fp16)[name = string("op_7424_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_51_cast_fp16 = sub(x = u_63_cast_fp16, y = var_7424_cast_fp16)[name = string("vnew_51_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_7429_cast_fp16 = mul(x = q_113_cast_fp16, y = var_7421_cast_fp16)[name = string("op_7429_cast_fp16")];
            bool var_7430_transpose_x_0 = const()[name = string("op_7430_transpose_x_0"), val = bool(false)];
            bool var_7430_transpose_y_0 = const()[name = string("op_7430_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_7430_cast_fp16 = matmul(transpose_x = var_7430_transpose_x_0, transpose_y = var_7430_transpose_y_0, x = var_7429_cast_fp16, y = S_51_cast_fp16)[name = string("op_7430_cast_fp16")];
            bool var_7431_transpose_x_0 = const()[name = string("op_7431_transpose_x_0"), val = bool(false)];
            bool var_7431_transpose_y_0 = const()[name = string("op_7431_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_7431_cast_fp16 = matmul(transpose_x = var_7431_transpose_x_0, transpose_y = var_7431_transpose_y_0, x = A_63_cast_fp16, y = vnew_51_cast_fp16)[name = string("op_7431_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_7433_cast_fp16 = add(x = var_7430_cast_fp16, y = var_7431_cast_fp16)[name = string("op_7433_cast_fp16")];
            tensor<fp16, [16]> var_7434_bias_0_to_fp16 = const()[name = string("op_7434_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_7434_cast_fp16 = linear(bias = var_7434_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_75_cast_fp16)[name = string("op_7434_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_53_cast_fp16 = exp(x = var_7434_cast_fp16)[name = string("tail_53_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_105_to_fp16 = const()[name = string("transpose_105_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_7436_bias_0_to_fp16 = const()[name = string("op_7436_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_7436_cast_fp16 = linear(bias = var_7436_bias_0_to_fp16, weight = transpose_105_to_fp16, x = gdec_75_cast_fp16)[name = string("op_7436_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_7437_cast_fp16 = exp(x = var_7436_cast_fp16)[name = string("op_7437_cast_fp16")];
            tensor<int32, [1]> last_43_axes_0 = const()[name = string("last_43_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_43_cast_fp16 = expand_dims(axes = last_43_axes_0, x = var_7437_cast_fp16)[name = string("last_43_cast_fp16")];
            tensor<int32, [1]> var_7441_axes_0 = const()[name = string("op_7441_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_7441_cast_fp16 = expand_dims(axes = var_7441_axes_0, x = tail_53_cast_fp16)[name = string("op_7441_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_7442_cast_fp16 = mul(x = k_115_cast_fp16, y = var_7441_cast_fp16)[name = string("op_7442_cast_fp16")];
            bool upd_41_transpose_x_1 = const()[name = string("upd_41_transpose_x_1"), val = bool(true)];
            bool upd_41_transpose_y_1 = const()[name = string("upd_41_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_41_cast_fp16 = matmul(transpose_x = upd_41_transpose_x_1, transpose_y = upd_41_transpose_y_1, x = var_7442_cast_fp16, y = vnew_51_cast_fp16)[name = string("upd_41_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_7447_cast_fp16 = mul(x = S_51_cast_fp16, y = last_43_cast_fp16)[name = string("op_7447_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_53_cast_fp16 = add(x = var_7447_cast_fp16, y = upd_41_cast_fp16)[name = string("S_53_cast_fp16")];
            tensor<int32, [4]> q_115_begin_0 = const()[name = string("q_115_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_115_end_0 = const()[name = string("q_115_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> q_115_end_mask_0 = const()[name = string("q_115_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_115_cast_fp16 = slice_by_index(begin = q_115_begin_0, end = q_115_end_0, end_mask = q_115_end_mask_0, x = q_109_cast_fp16)[name = string("q_115_cast_fp16")];
            tensor<int32, [4]> k_117_begin_0 = const()[name = string("k_117_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_117_end_0 = const()[name = string("k_117_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> k_117_end_mask_0 = const()[name = string("k_117_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_117_cast_fp16 = slice_by_index(begin = k_117_begin_0, end = k_117_end_0, end_mask = k_117_end_mask_0, x = k_111_cast_fp16)[name = string("k_117_cast_fp16")];
            tensor<int32, [4]> kb_65_begin_0 = const()[name = string("kb_65_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_65_end_0 = const()[name = string("kb_65_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> kb_65_end_mask_0 = const()[name = string("kb_65_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_65_cast_fp16 = slice_by_index(begin = kb_65_begin_0, end = kb_65_end_0, end_mask = kb_65_end_mask_0, x = var_7196_cast_fp16)[name = string("kb_65_cast_fp16")];
            tensor<int32, [4]> vb_77_begin_0 = const()[name = string("vb_77_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_77_end_0 = const()[name = string("vb_77_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> vb_77_end_mask_0 = const()[name = string("vb_77_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_77_cast_fp16 = slice_by_index(begin = vb_77_begin_0, end = vb_77_end_0, end_mask = vb_77_end_mask_0, x = vb_71_cast_fp16)[name = string("vb_77_cast_fp16")];
            tensor<int32, [3]> gdec_77_begin_0 = const()[name = string("gdec_77_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_77_end_0 = const()[name = string("gdec_77_end_0"), val = tensor<int32, [3]>([1, 4, 48])];
            tensor<bool, [3]> gdec_77_end_mask_0 = const()[name = string("gdec_77_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_77_cast_fp16 = slice_by_index(begin = gdec_77_begin_0, end = gdec_77_end_0, end_mask = gdec_77_end_mask_0, x = gdec_71_cast_fp16)[name = string("gdec_77_cast_fp16")];
            tensor<int32, [2]> var_7530 = const()[name = string("op_7530"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_7531_cast_fp16 = reshape(shape = var_7530, x = gdec_77_cast_fp16)[name = string("op_7531_cast_fp16")];
            tensor<fp16, [4, 256]> var_7532_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_7531_cast_fp16)[name = string("op_7532_cast_fp16")];
            tensor<int32, [4]> var_7537 = const()[name = string("op_7537"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_65_cast_fp16 = reshape(shape = var_7537, x = var_7532_cast_fp16)[name = string("P_65_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_7539_cast_fp16 = exp(x = P_65_cast_fp16)[name = string("op_7539_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_65_cast_fp16 = mul(x = var_7539_cast_fp16, y = tril_incl_to_fp16)[name = string("D_65_cast_fp16")];
            bool kk_65_transpose_x_1 = const()[name = string("kk_65_transpose_x_1"), val = bool(false)];
            bool kk_65_transpose_y_1 = const()[name = string("kk_65_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_65_cast_fp16 = matmul(transpose_x = kk_65_transpose_x_1, transpose_y = kk_65_transpose_y_1, x = kb_65_cast_fp16, y = k_117_cast_fp16)[name = string("kk_65_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_7545_cast_fp16 = mul(x = kk_65_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_7545_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_7546_cast_fp16 = mul(x = var_7545_cast_fp16, y = D_65_cast_fp16)[name = string("op_7546_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_65_cast_fp16 = mul(x = var_7546_cast_fp16, y = cum_tail_to_fp16)[name = string("N_65_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_7548_cast_fp16 = mul(x = D_65_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_7548_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_65_cast_fp16 = mul(x = kk_65_cast_fp16, y = var_7548_cast_fp16)[name = string("Na_65_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_257_cast_fp16 = add(x = eye_to_fp16, y = N_65_cast_fp16)[name = string("Minv_257_cast_fp16")];
            bool Np_193_transpose_x_0 = const()[name = string("Np_193_transpose_x_0"), val = bool(false)];
            bool Np_193_transpose_y_0 = const()[name = string("Np_193_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_193_cast_fp16 = matmul(transpose_x = Np_193_transpose_x_0, transpose_y = Np_193_transpose_y_0, x = N_65_cast_fp16, y = Na_65_cast_fp16)[name = string("Np_193_cast_fp16")];
            bool Npa_129_transpose_x_0 = const()[name = string("Npa_129_transpose_x_0"), val = bool(false)];
            bool Npa_129_transpose_y_0 = const()[name = string("Npa_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_129_cast_fp16 = matmul(transpose_x = Npa_129_transpose_x_0, transpose_y = Npa_129_transpose_y_0, x = Na_65_cast_fp16, y = N_65_cast_fp16)[name = string("Npa_129_cast_fp16")];
            bool var_7554_transpose_x_0 = const()[name = string("op_7554_transpose_x_0"), val = bool(false)];
            bool var_7554_transpose_y_0 = const()[name = string("op_7554_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_7554_cast_fp16 = matmul(transpose_x = var_7554_transpose_x_0, transpose_y = var_7554_transpose_y_0, x = Minv_257_cast_fp16, y = Np_193_cast_fp16)[name = string("op_7554_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_259_cast_fp16 = add(x = Minv_257_cast_fp16, y = var_7554_cast_fp16)[name = string("Minv_259_cast_fp16")];
            bool Np_195_transpose_x_0 = const()[name = string("Np_195_transpose_x_0"), val = bool(false)];
            bool Np_195_transpose_y_0 = const()[name = string("Np_195_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_195_cast_fp16 = matmul(transpose_x = Np_195_transpose_x_0, transpose_y = Np_195_transpose_y_0, x = Np_193_cast_fp16, y = Npa_129_cast_fp16)[name = string("Np_195_cast_fp16")];
            bool Npa_131_transpose_x_0 = const()[name = string("Npa_131_transpose_x_0"), val = bool(false)];
            bool Npa_131_transpose_y_0 = const()[name = string("Npa_131_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_131_cast_fp16 = matmul(transpose_x = Npa_131_transpose_x_0, transpose_y = Npa_131_transpose_y_0, x = Npa_129_cast_fp16, y = Np_193_cast_fp16)[name = string("Npa_131_cast_fp16")];
            bool var_7559_transpose_x_0 = const()[name = string("op_7559_transpose_x_0"), val = bool(false)];
            bool var_7559_transpose_y_0 = const()[name = string("op_7559_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_7559_cast_fp16 = matmul(transpose_x = var_7559_transpose_x_0, transpose_y = var_7559_transpose_y_0, x = Minv_259_cast_fp16, y = Np_195_cast_fp16)[name = string("op_7559_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_261_cast_fp16 = add(x = Minv_259_cast_fp16, y = var_7559_cast_fp16)[name = string("Minv_261_cast_fp16")];
            bool Np_197_transpose_x_0 = const()[name = string("Np_197_transpose_x_0"), val = bool(false)];
            bool Np_197_transpose_y_0 = const()[name = string("Np_197_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_197_cast_fp16 = matmul(transpose_x = Np_197_transpose_x_0, transpose_y = Np_197_transpose_y_0, x = Np_195_cast_fp16, y = Npa_131_cast_fp16)[name = string("Np_197_cast_fp16")];
            bool var_7563_transpose_x_0 = const()[name = string("op_7563_transpose_x_0"), val = bool(false)];
            bool var_7563_transpose_y_0 = const()[name = string("op_7563_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_7563_cast_fp16 = matmul(transpose_x = var_7563_transpose_x_0, transpose_y = var_7563_transpose_y_0, x = Minv_261_cast_fp16, y = Np_197_cast_fp16)[name = string("op_7563_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_263_cast_fp16 = add(x = Minv_261_cast_fp16, y = var_7563_cast_fp16)[name = string("Minv_263_cast_fp16")];
            bool u_65_transpose_x_0 = const()[name = string("u_65_transpose_x_0"), val = bool(false)];
            bool u_65_transpose_y_0 = const()[name = string("u_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_65_cast_fp16 = matmul(transpose_x = u_65_transpose_x_0, transpose_y = u_65_transpose_y_0, x = Minv_263_cast_fp16, y = vb_77_cast_fp16)[name = string("u_65_cast_fp16")];
            bool var_7570_transpose_x_1 = const()[name = string("op_7570_transpose_x_1"), val = bool(false)];
            bool var_7570_transpose_y_1 = const()[name = string("op_7570_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_7570_cast_fp16 = matmul(transpose_x = var_7570_transpose_x_1, transpose_y = var_7570_transpose_y_1, x = q_115_cast_fp16, y = k_117_cast_fp16)[name = string("op_7570_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_65_cast_fp16 = mul(x = var_7570_cast_fp16, y = D_65_cast_fp16)[name = string("A_65_cast_fp16")];
            tensor<fp16, [16]> var_7572_bias_0_to_fp16 = const()[name = string("op_7572_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_7572_cast_fp16 = linear(bias = var_7572_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_77_cast_fp16)[name = string("op_7572_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_53_cast_fp16 = exp(x = var_7572_cast_fp16)[name = string("cum_53_cast_fp16")];
            tensor<int32, [1]> var_7575_axes_0 = const()[name = string("op_7575_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_7575_cast_fp16 = expand_dims(axes = var_7575_axes_0, x = cum_53_cast_fp16)[name = string("op_7575_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_7576_cast_fp16 = mul(x = kb_65_cast_fp16, y = var_7575_cast_fp16)[name = string("op_7576_cast_fp16")];
            bool w_65_transpose_x_0 = const()[name = string("w_65_transpose_x_0"), val = bool(false)];
            bool w_65_transpose_y_0 = const()[name = string("w_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_65_cast_fp16 = matmul(transpose_x = w_65_transpose_x_0, transpose_y = w_65_transpose_y_0, x = Minv_263_cast_fp16, y = var_7576_cast_fp16)[name = string("w_65_cast_fp16")];
            bool var_7578_transpose_x_0 = const()[name = string("op_7578_transpose_x_0"), val = bool(false)];
            bool var_7578_transpose_y_0 = const()[name = string("op_7578_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_7578_cast_fp16 = matmul(transpose_x = var_7578_transpose_x_0, transpose_y = var_7578_transpose_y_0, x = w_65_cast_fp16, y = S_53_cast_fp16)[name = string("op_7578_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_53_cast_fp16 = sub(x = u_65_cast_fp16, y = var_7578_cast_fp16)[name = string("vnew_53_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_7583_cast_fp16 = mul(x = q_115_cast_fp16, y = var_7575_cast_fp16)[name = string("op_7583_cast_fp16")];
            bool var_7584_transpose_x_0 = const()[name = string("op_7584_transpose_x_0"), val = bool(false)];
            bool var_7584_transpose_y_0 = const()[name = string("op_7584_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_7584_cast_fp16 = matmul(transpose_x = var_7584_transpose_x_0, transpose_y = var_7584_transpose_y_0, x = var_7583_cast_fp16, y = S_53_cast_fp16)[name = string("op_7584_cast_fp16")];
            bool var_7585_transpose_x_0 = const()[name = string("op_7585_transpose_x_0"), val = bool(false)];
            bool var_7585_transpose_y_0 = const()[name = string("op_7585_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_7585_cast_fp16 = matmul(transpose_x = var_7585_transpose_x_0, transpose_y = var_7585_transpose_y_0, x = A_65_cast_fp16, y = vnew_53_cast_fp16)[name = string("op_7585_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_7587_cast_fp16 = add(x = var_7584_cast_fp16, y = var_7585_cast_fp16)[name = string("op_7587_cast_fp16")];
            tensor<fp16, [16]> var_7588_bias_0_to_fp16 = const()[name = string("op_7588_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_7588_cast_fp16 = linear(bias = var_7588_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_77_cast_fp16)[name = string("op_7588_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_55_cast_fp16 = exp(x = var_7588_cast_fp16)[name = string("tail_55_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_109_to_fp16 = const()[name = string("transpose_109_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_7590_bias_0_to_fp16 = const()[name = string("op_7590_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_7590_cast_fp16 = linear(bias = var_7590_bias_0_to_fp16, weight = transpose_109_to_fp16, x = gdec_77_cast_fp16)[name = string("op_7590_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_7591_cast_fp16 = exp(x = var_7590_cast_fp16)[name = string("op_7591_cast_fp16")];
            tensor<int32, [1]> last_45_axes_0 = const()[name = string("last_45_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_45_cast_fp16 = expand_dims(axes = last_45_axes_0, x = var_7591_cast_fp16)[name = string("last_45_cast_fp16")];
            tensor<int32, [1]> var_7595_axes_0 = const()[name = string("op_7595_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_7595_cast_fp16 = expand_dims(axes = var_7595_axes_0, x = tail_55_cast_fp16)[name = string("op_7595_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_7596_cast_fp16 = mul(x = k_117_cast_fp16, y = var_7595_cast_fp16)[name = string("op_7596_cast_fp16")];
            bool upd_43_transpose_x_1 = const()[name = string("upd_43_transpose_x_1"), val = bool(true)];
            bool upd_43_transpose_y_1 = const()[name = string("upd_43_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_43_cast_fp16 = matmul(transpose_x = upd_43_transpose_x_1, transpose_y = upd_43_transpose_y_1, x = var_7596_cast_fp16, y = vnew_53_cast_fp16)[name = string("upd_43_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_7601_cast_fp16 = mul(x = S_53_cast_fp16, y = last_45_cast_fp16)[name = string("op_7601_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_55_cast_fp16 = add(x = var_7601_cast_fp16, y = upd_43_cast_fp16)[name = string("S_55_cast_fp16")];
            tensor<int32, [4]> q_117_begin_0 = const()[name = string("q_117_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_117_end_0 = const()[name = string("q_117_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> q_117_end_mask_0 = const()[name = string("q_117_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_117_cast_fp16 = slice_by_index(begin = q_117_begin_0, end = q_117_end_0, end_mask = q_117_end_mask_0, x = q_109_cast_fp16)[name = string("q_117_cast_fp16")];
            tensor<int32, [4]> k_119_begin_0 = const()[name = string("k_119_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_119_end_0 = const()[name = string("k_119_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> k_119_end_mask_0 = const()[name = string("k_119_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_119_cast_fp16 = slice_by_index(begin = k_119_begin_0, end = k_119_end_0, end_mask = k_119_end_mask_0, x = k_111_cast_fp16)[name = string("k_119_cast_fp16")];
            tensor<int32, [4]> kb_67_begin_0 = const()[name = string("kb_67_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_67_end_0 = const()[name = string("kb_67_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> kb_67_end_mask_0 = const()[name = string("kb_67_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_67_cast_fp16 = slice_by_index(begin = kb_67_begin_0, end = kb_67_end_0, end_mask = kb_67_end_mask_0, x = var_7196_cast_fp16)[name = string("kb_67_cast_fp16")];
            tensor<int32, [4]> vb_79_begin_0 = const()[name = string("vb_79_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_79_end_0 = const()[name = string("vb_79_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> vb_79_end_mask_0 = const()[name = string("vb_79_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_79_cast_fp16 = slice_by_index(begin = vb_79_begin_0, end = vb_79_end_0, end_mask = vb_79_end_mask_0, x = vb_71_cast_fp16)[name = string("vb_79_cast_fp16")];
            tensor<int32, [3]> gdec_79_begin_0 = const()[name = string("gdec_79_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_79_end_0 = const()[name = string("gdec_79_end_0"), val = tensor<int32, [3]>([1, 4, 64])];
            tensor<bool, [3]> gdec_79_end_mask_0 = const()[name = string("gdec_79_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_79_cast_fp16 = slice_by_index(begin = gdec_79_begin_0, end = gdec_79_end_0, end_mask = gdec_79_end_mask_0, x = gdec_71_cast_fp16)[name = string("gdec_79_cast_fp16")];
            tensor<int32, [2]> var_7684 = const()[name = string("op_7684"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_7685_cast_fp16 = reshape(shape = var_7684, x = gdec_79_cast_fp16)[name = string("op_7685_cast_fp16")];
            tensor<fp16, [4, 256]> var_7686_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_7685_cast_fp16)[name = string("op_7686_cast_fp16")];
            tensor<int32, [4]> var_7691 = const()[name = string("op_7691"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_67_cast_fp16 = reshape(shape = var_7691, x = var_7686_cast_fp16)[name = string("P_67_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_7693_cast_fp16 = exp(x = P_67_cast_fp16)[name = string("op_7693_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_67_cast_fp16 = mul(x = var_7693_cast_fp16, y = tril_incl_to_fp16)[name = string("D_67_cast_fp16")];
            bool kk_67_transpose_x_1 = const()[name = string("kk_67_transpose_x_1"), val = bool(false)];
            bool kk_67_transpose_y_1 = const()[name = string("kk_67_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_67_cast_fp16 = matmul(transpose_x = kk_67_transpose_x_1, transpose_y = kk_67_transpose_y_1, x = kb_67_cast_fp16, y = k_119_cast_fp16)[name = string("kk_67_cast_fp16")];
            fp16 const_41_promoted_to_fp16 = const()[name = string("const_41_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_7699_cast_fp16 = mul(x = kk_67_cast_fp16, y = const_41_promoted_to_fp16)[name = string("op_7699_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_7700_cast_fp16 = mul(x = var_7699_cast_fp16, y = D_67_cast_fp16)[name = string("op_7700_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_67_cast_fp16 = mul(x = var_7700_cast_fp16, y = cum_tail_to_fp16)[name = string("N_67_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_7702_cast_fp16 = mul(x = D_67_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_7702_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_67_cast_fp16 = mul(x = kk_67_cast_fp16, y = var_7702_cast_fp16)[name = string("Na_67_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_265_cast_fp16 = add(x = eye_to_fp16, y = N_67_cast_fp16)[name = string("Minv_265_cast_fp16")];
            bool Np_199_transpose_x_0 = const()[name = string("Np_199_transpose_x_0"), val = bool(false)];
            bool Np_199_transpose_y_0 = const()[name = string("Np_199_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_199_cast_fp16 = matmul(transpose_x = Np_199_transpose_x_0, transpose_y = Np_199_transpose_y_0, x = N_67_cast_fp16, y = Na_67_cast_fp16)[name = string("Np_199_cast_fp16")];
            bool Npa_133_transpose_x_0 = const()[name = string("Npa_133_transpose_x_0"), val = bool(false)];
            bool Npa_133_transpose_y_0 = const()[name = string("Npa_133_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_133_cast_fp16 = matmul(transpose_x = Npa_133_transpose_x_0, transpose_y = Npa_133_transpose_y_0, x = Na_67_cast_fp16, y = N_67_cast_fp16)[name = string("Npa_133_cast_fp16")];
            bool var_7708_transpose_x_0 = const()[name = string("op_7708_transpose_x_0"), val = bool(false)];
            bool var_7708_transpose_y_0 = const()[name = string("op_7708_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_7708_cast_fp16 = matmul(transpose_x = var_7708_transpose_x_0, transpose_y = var_7708_transpose_y_0, x = Minv_265_cast_fp16, y = Np_199_cast_fp16)[name = string("op_7708_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_267_cast_fp16 = add(x = Minv_265_cast_fp16, y = var_7708_cast_fp16)[name = string("Minv_267_cast_fp16")];
            bool Np_201_transpose_x_0 = const()[name = string("Np_201_transpose_x_0"), val = bool(false)];
            bool Np_201_transpose_y_0 = const()[name = string("Np_201_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_201_cast_fp16 = matmul(transpose_x = Np_201_transpose_x_0, transpose_y = Np_201_transpose_y_0, x = Np_199_cast_fp16, y = Npa_133_cast_fp16)[name = string("Np_201_cast_fp16")];
            bool Npa_135_transpose_x_0 = const()[name = string("Npa_135_transpose_x_0"), val = bool(false)];
            bool Npa_135_transpose_y_0 = const()[name = string("Npa_135_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_135_cast_fp16 = matmul(transpose_x = Npa_135_transpose_x_0, transpose_y = Npa_135_transpose_y_0, x = Npa_133_cast_fp16, y = Np_199_cast_fp16)[name = string("Npa_135_cast_fp16")];
            bool var_7713_transpose_x_0 = const()[name = string("op_7713_transpose_x_0"), val = bool(false)];
            bool var_7713_transpose_y_0 = const()[name = string("op_7713_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_7713_cast_fp16 = matmul(transpose_x = var_7713_transpose_x_0, transpose_y = var_7713_transpose_y_0, x = Minv_267_cast_fp16, y = Np_201_cast_fp16)[name = string("op_7713_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_269_cast_fp16 = add(x = Minv_267_cast_fp16, y = var_7713_cast_fp16)[name = string("Minv_269_cast_fp16")];
            bool Np_203_transpose_x_0 = const()[name = string("Np_203_transpose_x_0"), val = bool(false)];
            bool Np_203_transpose_y_0 = const()[name = string("Np_203_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_203_cast_fp16 = matmul(transpose_x = Np_203_transpose_x_0, transpose_y = Np_203_transpose_y_0, x = Np_201_cast_fp16, y = Npa_135_cast_fp16)[name = string("Np_203_cast_fp16")];
            bool var_7717_transpose_x_0 = const()[name = string("op_7717_transpose_x_0"), val = bool(false)];
            bool var_7717_transpose_y_0 = const()[name = string("op_7717_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_7717_cast_fp16 = matmul(transpose_x = var_7717_transpose_x_0, transpose_y = var_7717_transpose_y_0, x = Minv_269_cast_fp16, y = Np_203_cast_fp16)[name = string("op_7717_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_271_cast_fp16 = add(x = Minv_269_cast_fp16, y = var_7717_cast_fp16)[name = string("Minv_271_cast_fp16")];
            bool u_67_transpose_x_0 = const()[name = string("u_67_transpose_x_0"), val = bool(false)];
            bool u_67_transpose_y_0 = const()[name = string("u_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_67_cast_fp16 = matmul(transpose_x = u_67_transpose_x_0, transpose_y = u_67_transpose_y_0, x = Minv_271_cast_fp16, y = vb_79_cast_fp16)[name = string("u_67_cast_fp16")];
            bool var_7724_transpose_x_1 = const()[name = string("op_7724_transpose_x_1"), val = bool(false)];
            bool var_7724_transpose_y_1 = const()[name = string("op_7724_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_7724_cast_fp16 = matmul(transpose_x = var_7724_transpose_x_1, transpose_y = var_7724_transpose_y_1, x = q_117_cast_fp16, y = k_119_cast_fp16)[name = string("op_7724_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_67_cast_fp16 = mul(x = var_7724_cast_fp16, y = D_67_cast_fp16)[name = string("A_67_cast_fp16")];
            tensor<fp16, [16]> var_7726_bias_0_to_fp16 = const()[name = string("op_7726_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_7726_cast_fp16 = linear(bias = var_7726_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_79_cast_fp16)[name = string("op_7726_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_55_cast_fp16 = exp(x = var_7726_cast_fp16)[name = string("cum_55_cast_fp16")];
            tensor<int32, [1]> var_7729_axes_0 = const()[name = string("op_7729_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_7729_cast_fp16 = expand_dims(axes = var_7729_axes_0, x = cum_55_cast_fp16)[name = string("op_7729_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_7730_cast_fp16 = mul(x = kb_67_cast_fp16, y = var_7729_cast_fp16)[name = string("op_7730_cast_fp16")];
            bool w_67_transpose_x_0 = const()[name = string("w_67_transpose_x_0"), val = bool(false)];
            bool w_67_transpose_y_0 = const()[name = string("w_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_67_cast_fp16 = matmul(transpose_x = w_67_transpose_x_0, transpose_y = w_67_transpose_y_0, x = Minv_271_cast_fp16, y = var_7730_cast_fp16)[name = string("w_67_cast_fp16")];
            bool var_7732_transpose_x_0 = const()[name = string("op_7732_transpose_x_0"), val = bool(false)];
            bool var_7732_transpose_y_0 = const()[name = string("op_7732_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_7732_cast_fp16 = matmul(transpose_x = var_7732_transpose_x_0, transpose_y = var_7732_transpose_y_0, x = w_67_cast_fp16, y = S_55_cast_fp16)[name = string("op_7732_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_55_cast_fp16 = sub(x = u_67_cast_fp16, y = var_7732_cast_fp16)[name = string("vnew_55_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_7737_cast_fp16 = mul(x = q_117_cast_fp16, y = var_7729_cast_fp16)[name = string("op_7737_cast_fp16")];
            bool var_7738_transpose_x_0 = const()[name = string("op_7738_transpose_x_0"), val = bool(false)];
            bool var_7738_transpose_y_0 = const()[name = string("op_7738_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_7738_cast_fp16 = matmul(transpose_x = var_7738_transpose_x_0, transpose_y = var_7738_transpose_y_0, x = var_7737_cast_fp16, y = S_55_cast_fp16)[name = string("op_7738_cast_fp16")];
            bool var_7739_transpose_x_0 = const()[name = string("op_7739_transpose_x_0"), val = bool(false)];
            bool var_7739_transpose_y_0 = const()[name = string("op_7739_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_7739_cast_fp16 = matmul(transpose_x = var_7739_transpose_x_0, transpose_y = var_7739_transpose_y_0, x = A_67_cast_fp16, y = vnew_55_cast_fp16)[name = string("op_7739_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_7741_cast_fp16 = add(x = var_7738_cast_fp16, y = var_7739_cast_fp16)[name = string("op_7741_cast_fp16")];
            tensor<fp16, [16]> var_7742_bias_0_to_fp16 = const()[name = string("op_7742_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_7742_cast_fp16 = linear(bias = var_7742_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_79_cast_fp16)[name = string("op_7742_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_57_cast_fp16 = exp(x = var_7742_cast_fp16)[name = string("tail_57_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_113_to_fp16 = const()[name = string("transpose_113_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_7744_bias_0_to_fp16 = const()[name = string("op_7744_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_7744_cast_fp16 = linear(bias = var_7744_bias_0_to_fp16, weight = transpose_113_to_fp16, x = gdec_79_cast_fp16)[name = string("op_7744_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_7745_cast_fp16 = exp(x = var_7744_cast_fp16)[name = string("op_7745_cast_fp16")];
            tensor<int32, [1]> last_47_axes_0 = const()[name = string("last_47_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_47_cast_fp16 = expand_dims(axes = last_47_axes_0, x = var_7745_cast_fp16)[name = string("last_47_cast_fp16")];
            tensor<int32, [1]> var_7749_axes_0 = const()[name = string("op_7749_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_7749_cast_fp16 = expand_dims(axes = var_7749_axes_0, x = tail_57_cast_fp16)[name = string("op_7749_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_7750_cast_fp16 = mul(x = k_119_cast_fp16, y = var_7749_cast_fp16)[name = string("op_7750_cast_fp16")];
            bool upd_45_transpose_x_1 = const()[name = string("upd_45_transpose_x_1"), val = bool(true)];
            bool upd_45_transpose_y_1 = const()[name = string("upd_45_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_45_cast_fp16 = matmul(transpose_x = upd_45_transpose_x_1, transpose_y = upd_45_transpose_y_1, x = var_7750_cast_fp16, y = vnew_55_cast_fp16)[name = string("upd_45_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_7755_cast_fp16 = mul(x = S_55_cast_fp16, y = last_47_cast_fp16)[name = string("op_7755_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_57_cast_fp16 = add(x = var_7755_cast_fp16, y = upd_45_cast_fp16)[name = string("S_57_cast_fp16")];
            tensor<int32, [4]> q_119_begin_0 = const()[name = string("q_119_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> q_119_end_0 = const()[name = string("q_119_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> q_119_end_mask_0 = const()[name = string("q_119_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> q_119_cast_fp16 = slice_by_index(begin = q_119_begin_0, end = q_119_end_0, end_mask = q_119_end_mask_0, x = q_109_cast_fp16)[name = string("q_119_cast_fp16")];
            tensor<int32, [4]> k_121_begin_0 = const()[name = string("k_121_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> k_121_end_0 = const()[name = string("k_121_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> k_121_end_mask_0 = const()[name = string("k_121_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> k_121_cast_fp16 = slice_by_index(begin = k_121_begin_0, end = k_121_end_0, end_mask = k_121_end_mask_0, x = k_111_cast_fp16)[name = string("k_121_cast_fp16")];
            tensor<int32, [4]> kb_69_begin_0 = const()[name = string("kb_69_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> kb_69_end_0 = const()[name = string("kb_69_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> kb_69_end_mask_0 = const()[name = string("kb_69_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_69_cast_fp16 = slice_by_index(begin = kb_69_begin_0, end = kb_69_end_0, end_mask = kb_69_end_mask_0, x = var_7196_cast_fp16)[name = string("kb_69_cast_fp16")];
            tensor<int32, [4]> vb_81_begin_0 = const()[name = string("vb_81_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> vb_81_end_0 = const()[name = string("vb_81_end_0"), val = tensor<int32, [4]>([1, 4, 80, 128])];
            tensor<bool, [4]> vb_81_end_mask_0 = const()[name = string("vb_81_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_81_cast_fp16 = slice_by_index(begin = vb_81_begin_0, end = vb_81_end_0, end_mask = vb_81_end_mask_0, x = vb_71_cast_fp16)[name = string("vb_81_cast_fp16")];
            tensor<int32, [3]> gdec_81_begin_0 = const()[name = string("gdec_81_begin_0"), val = tensor<int32, [3]>([0, 0, 64])];
            tensor<int32, [3]> gdec_81_end_0 = const()[name = string("gdec_81_end_0"), val = tensor<int32, [3]>([1, 4, 80])];
            tensor<bool, [3]> gdec_81_end_mask_0 = const()[name = string("gdec_81_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 4, 16]> gdec_81_cast_fp16 = slice_by_index(begin = gdec_81_begin_0, end = gdec_81_end_0, end_mask = gdec_81_end_mask_0, x = gdec_71_cast_fp16)[name = string("gdec_81_cast_fp16")];
            tensor<int32, [2]> var_7838 = const()[name = string("op_7838"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_7839_cast_fp16 = reshape(shape = var_7838, x = gdec_81_cast_fp16)[name = string("op_7839_cast_fp16")];
            tensor<fp16, [4, 256]> var_7840_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_7839_cast_fp16)[name = string("op_7840_cast_fp16")];
            tensor<int32, [4]> var_7845 = const()[name = string("op_7845"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_69_cast_fp16 = reshape(shape = var_7845, x = var_7840_cast_fp16)[name = string("P_69_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_7847_cast_fp16 = exp(x = P_69_cast_fp16)[name = string("op_7847_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_69_cast_fp16 = mul(x = var_7847_cast_fp16, y = tril_incl_to_fp16)[name = string("D_69_cast_fp16")];
            bool kk_69_transpose_x_1 = const()[name = string("kk_69_transpose_x_1"), val = bool(false)];
            bool kk_69_transpose_y_1 = const()[name = string("kk_69_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_69_cast_fp16 = matmul(transpose_x = kk_69_transpose_x_1, transpose_y = kk_69_transpose_y_1, x = kb_69_cast_fp16, y = k_121_cast_fp16)[name = string("kk_69_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_7853_cast_fp16 = mul(x = kk_69_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_7853_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_7854_cast_fp16 = mul(x = var_7853_cast_fp16, y = D_69_cast_fp16)[name = string("op_7854_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_69_cast_fp16 = mul(x = var_7854_cast_fp16, y = cum_tail_to_fp16)[name = string("N_69_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_7856_cast_fp16 = mul(x = D_69_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_7856_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_69_cast_fp16 = mul(x = kk_69_cast_fp16, y = var_7856_cast_fp16)[name = string("Na_69_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_273_cast_fp16 = add(x = eye_to_fp16, y = N_69_cast_fp16)[name = string("Minv_273_cast_fp16")];
            bool Np_205_transpose_x_0 = const()[name = string("Np_205_transpose_x_0"), val = bool(false)];
            bool Np_205_transpose_y_0 = const()[name = string("Np_205_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_205_cast_fp16 = matmul(transpose_x = Np_205_transpose_x_0, transpose_y = Np_205_transpose_y_0, x = N_69_cast_fp16, y = Na_69_cast_fp16)[name = string("Np_205_cast_fp16")];
            bool Npa_137_transpose_x_0 = const()[name = string("Npa_137_transpose_x_0"), val = bool(false)];
            bool Npa_137_transpose_y_0 = const()[name = string("Npa_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_137_cast_fp16 = matmul(transpose_x = Npa_137_transpose_x_0, transpose_y = Npa_137_transpose_y_0, x = Na_69_cast_fp16, y = N_69_cast_fp16)[name = string("Npa_137_cast_fp16")];
            bool var_7862_transpose_x_0 = const()[name = string("op_7862_transpose_x_0"), val = bool(false)];
            bool var_7862_transpose_y_0 = const()[name = string("op_7862_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_7862_cast_fp16 = matmul(transpose_x = var_7862_transpose_x_0, transpose_y = var_7862_transpose_y_0, x = Minv_273_cast_fp16, y = Np_205_cast_fp16)[name = string("op_7862_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_275_cast_fp16 = add(x = Minv_273_cast_fp16, y = var_7862_cast_fp16)[name = string("Minv_275_cast_fp16")];
            bool Np_207_transpose_x_0 = const()[name = string("Np_207_transpose_x_0"), val = bool(false)];
            bool Np_207_transpose_y_0 = const()[name = string("Np_207_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_207_cast_fp16 = matmul(transpose_x = Np_207_transpose_x_0, transpose_y = Np_207_transpose_y_0, x = Np_205_cast_fp16, y = Npa_137_cast_fp16)[name = string("Np_207_cast_fp16")];
            bool Npa_139_transpose_x_0 = const()[name = string("Npa_139_transpose_x_0"), val = bool(false)];
            bool Npa_139_transpose_y_0 = const()[name = string("Npa_139_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_139_cast_fp16 = matmul(transpose_x = Npa_139_transpose_x_0, transpose_y = Npa_139_transpose_y_0, x = Npa_137_cast_fp16, y = Np_205_cast_fp16)[name = string("Npa_139_cast_fp16")];
            bool var_7867_transpose_x_0 = const()[name = string("op_7867_transpose_x_0"), val = bool(false)];
            bool var_7867_transpose_y_0 = const()[name = string("op_7867_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_7867_cast_fp16 = matmul(transpose_x = var_7867_transpose_x_0, transpose_y = var_7867_transpose_y_0, x = Minv_275_cast_fp16, y = Np_207_cast_fp16)[name = string("op_7867_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_277_cast_fp16 = add(x = Minv_275_cast_fp16, y = var_7867_cast_fp16)[name = string("Minv_277_cast_fp16")];
            bool Np_209_transpose_x_0 = const()[name = string("Np_209_transpose_x_0"), val = bool(false)];
            bool Np_209_transpose_y_0 = const()[name = string("Np_209_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_209_cast_fp16 = matmul(transpose_x = Np_209_transpose_x_0, transpose_y = Np_209_transpose_y_0, x = Np_207_cast_fp16, y = Npa_139_cast_fp16)[name = string("Np_209_cast_fp16")];
            bool var_7871_transpose_x_0 = const()[name = string("op_7871_transpose_x_0"), val = bool(false)];
            bool var_7871_transpose_y_0 = const()[name = string("op_7871_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_7871_cast_fp16 = matmul(transpose_x = var_7871_transpose_x_0, transpose_y = var_7871_transpose_y_0, x = Minv_277_cast_fp16, y = Np_209_cast_fp16)[name = string("op_7871_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_279_cast_fp16 = add(x = Minv_277_cast_fp16, y = var_7871_cast_fp16)[name = string("Minv_279_cast_fp16")];
            bool u_69_transpose_x_0 = const()[name = string("u_69_transpose_x_0"), val = bool(false)];
            bool u_69_transpose_y_0 = const()[name = string("u_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_69_cast_fp16 = matmul(transpose_x = u_69_transpose_x_0, transpose_y = u_69_transpose_y_0, x = Minv_279_cast_fp16, y = vb_81_cast_fp16)[name = string("u_69_cast_fp16")];
            bool var_7878_transpose_x_1 = const()[name = string("op_7878_transpose_x_1"), val = bool(false)];
            bool var_7878_transpose_y_1 = const()[name = string("op_7878_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_7878_cast_fp16 = matmul(transpose_x = var_7878_transpose_x_1, transpose_y = var_7878_transpose_y_1, x = q_119_cast_fp16, y = k_121_cast_fp16)[name = string("op_7878_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_69_cast_fp16 = mul(x = var_7878_cast_fp16, y = D_69_cast_fp16)[name = string("A_69_cast_fp16")];
            tensor<fp16, [16]> var_7880_bias_0_to_fp16 = const()[name = string("op_7880_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_7880_cast_fp16 = linear(bias = var_7880_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_81_cast_fp16)[name = string("op_7880_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_57_cast_fp16 = exp(x = var_7880_cast_fp16)[name = string("cum_57_cast_fp16")];
            tensor<int32, [1]> var_7883_axes_0 = const()[name = string("op_7883_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_7883_cast_fp16 = expand_dims(axes = var_7883_axes_0, x = cum_57_cast_fp16)[name = string("op_7883_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_7884_cast_fp16 = mul(x = kb_69_cast_fp16, y = var_7883_cast_fp16)[name = string("op_7884_cast_fp16")];
            bool w_69_transpose_x_0 = const()[name = string("w_69_transpose_x_0"), val = bool(false)];
            bool w_69_transpose_y_0 = const()[name = string("w_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_69_cast_fp16 = matmul(transpose_x = w_69_transpose_x_0, transpose_y = w_69_transpose_y_0, x = Minv_279_cast_fp16, y = var_7884_cast_fp16)[name = string("w_69_cast_fp16")];
            bool var_7886_transpose_x_0 = const()[name = string("op_7886_transpose_x_0"), val = bool(false)];
            bool var_7886_transpose_y_0 = const()[name = string("op_7886_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_7886_cast_fp16 = matmul(transpose_x = var_7886_transpose_x_0, transpose_y = var_7886_transpose_y_0, x = w_69_cast_fp16, y = S_57_cast_fp16)[name = string("op_7886_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_57_cast_fp16 = sub(x = u_69_cast_fp16, y = var_7886_cast_fp16)[name = string("vnew_57_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_7891_cast_fp16 = mul(x = q_119_cast_fp16, y = var_7883_cast_fp16)[name = string("op_7891_cast_fp16")];
            bool var_7892_transpose_x_0 = const()[name = string("op_7892_transpose_x_0"), val = bool(false)];
            bool var_7892_transpose_y_0 = const()[name = string("op_7892_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_7892_cast_fp16 = matmul(transpose_x = var_7892_transpose_x_0, transpose_y = var_7892_transpose_y_0, x = var_7891_cast_fp16, y = S_57_cast_fp16)[name = string("op_7892_cast_fp16")];
            bool var_7893_transpose_x_0 = const()[name = string("op_7893_transpose_x_0"), val = bool(false)];
            bool var_7893_transpose_y_0 = const()[name = string("op_7893_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_7893_cast_fp16 = matmul(transpose_x = var_7893_transpose_x_0, transpose_y = var_7893_transpose_y_0, x = A_69_cast_fp16, y = vnew_57_cast_fp16)[name = string("op_7893_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_7895_cast_fp16 = add(x = var_7892_cast_fp16, y = var_7893_cast_fp16)[name = string("op_7895_cast_fp16")];
            tensor<fp16, [16]> var_7896_bias_0_to_fp16 = const()[name = string("op_7896_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_7896_cast_fp16 = linear(bias = var_7896_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_81_cast_fp16)[name = string("op_7896_cast_fp16")];
            tensor<fp16, [1, 4, 16]> tail_cast_fp16 = exp(x = var_7896_cast_fp16)[name = string("tail_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_117_to_fp16 = const()[name = string("transpose_117_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_7898_bias_0_to_fp16 = const()[name = string("op_7898_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 4, 1]> var_7898_cast_fp16 = linear(bias = var_7898_bias_0_to_fp16, weight = transpose_117_to_fp16, x = gdec_81_cast_fp16)[name = string("op_7898_cast_fp16")];
            tensor<fp16, [1, 4, 1]> var_7899_cast_fp16 = exp(x = var_7898_cast_fp16)[name = string("op_7899_cast_fp16")];
            tensor<int32, [1]> last_49_axes_0 = const()[name = string("last_49_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 1, 1]> last_49_cast_fp16 = expand_dims(axes = last_49_axes_0, x = var_7899_cast_fp16)[name = string("last_49_cast_fp16")];
            tensor<int32, [1]> var_7903_axes_0 = const()[name = string("op_7903_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_7903_cast_fp16 = expand_dims(axes = var_7903_axes_0, x = tail_cast_fp16)[name = string("op_7903_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_7904_cast_fp16 = mul(x = k_121_cast_fp16, y = var_7903_cast_fp16)[name = string("op_7904_cast_fp16")];
            bool upd_transpose_x_1 = const()[name = string("upd_transpose_x_1"), val = bool(true)];
            bool upd_transpose_y_1 = const()[name = string("upd_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 128]> upd_cast_fp16 = matmul(transpose_x = upd_transpose_x_1, transpose_y = upd_transpose_y_1, x = var_7904_cast_fp16, y = vnew_57_cast_fp16)[name = string("upd_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_7909_cast_fp16 = mul(x = S_57_cast_fp16, y = last_49_cast_fp16)[name = string("op_7909_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> S_cast_fp16 = add(x = var_7909_cast_fp16, y = upd_cast_fp16)[name = string("S_cast_fp16")];
            tensor<int32, [4]> q_121_begin_0 = const()[name = string("q_121_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> q_121_end_0 = const()[name = string("q_121_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> q_121_end_mask_0 = const()[name = string("q_121_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> q_121_cast_fp16 = slice_by_index(begin = q_121_begin_0, end = q_121_end_0, end_mask = q_121_end_mask_0, x = q_109_cast_fp16)[name = string("q_121_cast_fp16")];
            tensor<int32, [4]> k_123_begin_0 = const()[name = string("k_123_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> k_123_end_0 = const()[name = string("k_123_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> k_123_end_mask_0 = const()[name = string("k_123_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> k_123_cast_fp16 = slice_by_index(begin = k_123_begin_0, end = k_123_end_0, end_mask = k_123_end_mask_0, x = k_111_cast_fp16)[name = string("k_123_cast_fp16")];
            tensor<int32, [4]> kb_begin_0 = const()[name = string("kb_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> kb_end_0 = const()[name = string("kb_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> kb_end_mask_0 = const()[name = string("kb_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> kb_cast_fp16 = slice_by_index(begin = kb_begin_0, end = kb_end_0, end_mask = kb_end_mask_0, x = var_7196_cast_fp16)[name = string("kb_cast_fp16")];
            tensor<int32, [4]> vb_begin_0 = const()[name = string("vb_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> vb_end_0 = const()[name = string("vb_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> vb_end_mask_0 = const()[name = string("vb_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 16, 128]> vb_cast_fp16 = slice_by_index(begin = vb_begin_0, end = vb_end_0, end_mask = vb_end_mask_0, x = vb_71_cast_fp16)[name = string("vb_cast_fp16")];
            tensor<int32, [3]> gdec_begin_0 = const()[name = string("gdec_begin_0"), val = tensor<int32, [3]>([0, 0, 80])];
            tensor<int32, [3]> gdec_end_0 = const()[name = string("gdec_end_0"), val = tensor<int32, [3]>([1, 4, 1])];
            tensor<bool, [3]> gdec_end_mask_0 = const()[name = string("gdec_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 4, 16]> gdec_cast_fp16 = slice_by_index(begin = gdec_begin_0, end = gdec_end_0, end_mask = gdec_end_mask_0, x = gdec_71_cast_fp16)[name = string("gdec_cast_fp16")];
            tensor<int32, [2]> var_7992 = const()[name = string("op_7992"), val = tensor<int32, [2]>([4, 16])];
            tensor<fp16, [4, 16]> var_7993_cast_fp16 = reshape(shape = var_7992, x = gdec_cast_fp16)[name = string("op_7993_cast_fp16")];
            tensor<fp16, [4, 256]> var_7994_cast_fp16 = linear(bias = var_634_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_7993_cast_fp16)[name = string("op_7994_cast_fp16")];
            tensor<int32, [4]> var_7999 = const()[name = string("op_7999"), val = tensor<int32, [4]>([1, 4, 16, 16])];
            tensor<fp16, [1, 4, 16, 16]> P_cast_fp16 = reshape(shape = var_7999, x = var_7994_cast_fp16)[name = string("P_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_8001_cast_fp16 = exp(x = P_cast_fp16)[name = string("op_8001_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> D_cast_fp16 = mul(x = var_8001_cast_fp16, y = tril_incl_to_fp16)[name = string("D_cast_fp16")];
            bool kk_transpose_x_1 = const()[name = string("kk_transpose_x_1"), val = bool(false)];
            bool kk_transpose_y_1 = const()[name = string("kk_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> kk_cast_fp16 = matmul(transpose_x = kk_transpose_x_1, transpose_y = kk_transpose_y_1, x = kb_cast_fp16, y = k_123_cast_fp16)[name = string("kk_cast_fp16")];
            fp16 const_43_promoted_to_fp16 = const()[name = string("const_43_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 16, 16]> var_8007_cast_fp16 = mul(x = kk_cast_fp16, y = const_43_promoted_to_fp16)[name = string("op_8007_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_8008_cast_fp16 = mul(x = var_8007_cast_fp16, y = D_cast_fp16)[name = string("op_8008_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> N_cast_fp16 = mul(x = var_8008_cast_fp16, y = cum_tail_to_fp16)[name = string("N_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> var_8010_cast_fp16 = mul(x = D_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_8010_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Na_cast_fp16 = mul(x = kk_cast_fp16, y = var_8010_cast_fp16)[name = string("Na_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_281_cast_fp16 = add(x = eye_to_fp16, y = N_cast_fp16)[name = string("Minv_281_cast_fp16")];
            bool Np_211_transpose_x_0 = const()[name = string("Np_211_transpose_x_0"), val = bool(false)];
            bool Np_211_transpose_y_0 = const()[name = string("Np_211_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_211_cast_fp16 = matmul(transpose_x = Np_211_transpose_x_0, transpose_y = Np_211_transpose_y_0, x = N_cast_fp16, y = Na_cast_fp16)[name = string("Np_211_cast_fp16")];
            bool Npa_141_transpose_x_0 = const()[name = string("Npa_141_transpose_x_0"), val = bool(false)];
            bool Npa_141_transpose_y_0 = const()[name = string("Npa_141_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_141_cast_fp16 = matmul(transpose_x = Npa_141_transpose_x_0, transpose_y = Npa_141_transpose_y_0, x = Na_cast_fp16, y = N_cast_fp16)[name = string("Npa_141_cast_fp16")];
            bool var_8016_transpose_x_0 = const()[name = string("op_8016_transpose_x_0"), val = bool(false)];
            bool var_8016_transpose_y_0 = const()[name = string("op_8016_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_8016_cast_fp16 = matmul(transpose_x = var_8016_transpose_x_0, transpose_y = var_8016_transpose_y_0, x = Minv_281_cast_fp16, y = Np_211_cast_fp16)[name = string("op_8016_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_283_cast_fp16 = add(x = Minv_281_cast_fp16, y = var_8016_cast_fp16)[name = string("Minv_283_cast_fp16")];
            bool Np_213_transpose_x_0 = const()[name = string("Np_213_transpose_x_0"), val = bool(false)];
            bool Np_213_transpose_y_0 = const()[name = string("Np_213_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_213_cast_fp16 = matmul(transpose_x = Np_213_transpose_x_0, transpose_y = Np_213_transpose_y_0, x = Np_211_cast_fp16, y = Npa_141_cast_fp16)[name = string("Np_213_cast_fp16")];
            bool Npa_transpose_x_0 = const()[name = string("Npa_transpose_x_0"), val = bool(false)];
            bool Npa_transpose_y_0 = const()[name = string("Npa_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Npa_cast_fp16 = matmul(transpose_x = Npa_transpose_x_0, transpose_y = Npa_transpose_y_0, x = Npa_141_cast_fp16, y = Np_211_cast_fp16)[name = string("Npa_cast_fp16")];
            bool var_8021_transpose_x_0 = const()[name = string("op_8021_transpose_x_0"), val = bool(false)];
            bool var_8021_transpose_y_0 = const()[name = string("op_8021_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_8021_cast_fp16 = matmul(transpose_x = var_8021_transpose_x_0, transpose_y = var_8021_transpose_y_0, x = Minv_283_cast_fp16, y = Np_213_cast_fp16)[name = string("op_8021_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_285_cast_fp16 = add(x = Minv_283_cast_fp16, y = var_8021_cast_fp16)[name = string("Minv_285_cast_fp16")];
            bool Np_transpose_x_0 = const()[name = string("Np_transpose_x_0"), val = bool(false)];
            bool Np_transpose_y_0 = const()[name = string("Np_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> Np_cast_fp16 = matmul(transpose_x = Np_transpose_x_0, transpose_y = Np_transpose_y_0, x = Np_213_cast_fp16, y = Npa_cast_fp16)[name = string("Np_cast_fp16")];
            bool var_8025_transpose_x_0 = const()[name = string("op_8025_transpose_x_0"), val = bool(false)];
            bool var_8025_transpose_y_0 = const()[name = string("op_8025_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 16]> var_8025_cast_fp16 = matmul(transpose_x = var_8025_transpose_x_0, transpose_y = var_8025_transpose_y_0, x = Minv_285_cast_fp16, y = Np_cast_fp16)[name = string("op_8025_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> Minv_cast_fp16 = add(x = Minv_285_cast_fp16, y = var_8025_cast_fp16)[name = string("Minv_cast_fp16")];
            bool u_transpose_x_0 = const()[name = string("u_transpose_x_0"), val = bool(false)];
            bool u_transpose_y_0 = const()[name = string("u_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> u_cast_fp16 = matmul(transpose_x = u_transpose_x_0, transpose_y = u_transpose_y_0, x = Minv_cast_fp16, y = vb_cast_fp16)[name = string("u_cast_fp16")];
            bool var_8032_transpose_x_1 = const()[name = string("op_8032_transpose_x_1"), val = bool(false)];
            bool var_8032_transpose_y_1 = const()[name = string("op_8032_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 4, 16, 16]> var_8032_cast_fp16 = matmul(transpose_x = var_8032_transpose_x_1, transpose_y = var_8032_transpose_y_1, x = q_121_cast_fp16, y = k_123_cast_fp16)[name = string("op_8032_cast_fp16")];
            tensor<fp16, [1, 4, 16, 16]> A_cast_fp16 = mul(x = var_8032_cast_fp16, y = D_cast_fp16)[name = string("A_cast_fp16")];
            tensor<fp16, [16]> var_8034_bias_0_to_fp16 = const()[name = string("op_8034_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [1, 4, 16]> var_8034_cast_fp16 = linear(bias = var_8034_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_cast_fp16)[name = string("op_8034_cast_fp16")];
            tensor<fp16, [1, 4, 16]> cum_cast_fp16 = exp(x = var_8034_cast_fp16)[name = string("cum_cast_fp16")];
            tensor<int32, [1]> var_8037_axes_0 = const()[name = string("op_8037_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 4, 16, 1]> var_8037_cast_fp16 = expand_dims(axes = var_8037_axes_0, x = cum_cast_fp16)[name = string("op_8037_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_8038_cast_fp16 = mul(x = kb_cast_fp16, y = var_8037_cast_fp16)[name = string("op_8038_cast_fp16")];
            bool w_transpose_x_0 = const()[name = string("w_transpose_x_0"), val = bool(false)];
            bool w_transpose_y_0 = const()[name = string("w_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> w_cast_fp16 = matmul(transpose_x = w_transpose_x_0, transpose_y = w_transpose_y_0, x = Minv_cast_fp16, y = var_8038_cast_fp16)[name = string("w_cast_fp16")];
            bool var_8040_transpose_x_0 = const()[name = string("op_8040_transpose_x_0"), val = bool(false)];
            bool var_8040_transpose_y_0 = const()[name = string("op_8040_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_8040_cast_fp16 = matmul(transpose_x = var_8040_transpose_x_0, transpose_y = var_8040_transpose_y_0, x = w_cast_fp16, y = S_cast_fp16)[name = string("op_8040_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> vnew_cast_fp16 = sub(x = u_cast_fp16, y = var_8040_cast_fp16)[name = string("vnew_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> var_8045_cast_fp16 = mul(x = q_121_cast_fp16, y = var_8037_cast_fp16)[name = string("op_8045_cast_fp16")];
            bool var_8046_transpose_x_0 = const()[name = string("op_8046_transpose_x_0"), val = bool(false)];
            bool var_8046_transpose_y_0 = const()[name = string("op_8046_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_8046_cast_fp16 = matmul(transpose_x = var_8046_transpose_x_0, transpose_y = var_8046_transpose_y_0, x = var_8045_cast_fp16, y = S_cast_fp16)[name = string("op_8046_cast_fp16")];
            bool var_8047_transpose_x_0 = const()[name = string("op_8047_transpose_x_0"), val = bool(false)];
            bool var_8047_transpose_y_0 = const()[name = string("op_8047_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 16, 128]> var_8047_cast_fp16 = matmul(transpose_x = var_8047_transpose_x_0, transpose_y = var_8047_transpose_y_0, x = A_cast_fp16, y = vnew_cast_fp16)[name = string("op_8047_cast_fp16")];
            tensor<fp16, [1, 4, 16, 128]> oc_cast_fp16 = add(x = var_8046_cast_fp16, y = var_8047_cast_fp16)[name = string("oc_cast_fp16")];
            int32 var_8051 = const()[name = string("op_8051"), val = int32(2)];
            bool o_33_interleave_0 = const()[name = string("o_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 96, 128]> o_33_cast_fp16 = concat(axis = var_8051, interleave = o_33_interleave_0, values = (var_7286_cast_fp16, var_7433_cast_fp16, var_7587_cast_fp16, var_7741_cast_fp16, var_7895_cast_fp16, oc_cast_fp16))[name = string("o_33_cast_fp16")];
            tensor<int32, [4]> var_8057 = const()[name = string("op_8057"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_8063 = const()[name = string("op_8063"), val = tensor<int32, [4]>([1, 96, 4, 128])];
            tensor<fp16, [1, 96, 4, 128]> input_97_cast_fp16 = reshape(shape = var_8063, x = z_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> o_35_cast_fp16 = transpose(perm = var_8057, x = o_33_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [1, 96, 4, 128]> var_8065_cast_fp16 = mul(x = o_35_cast_fp16, y = o_35_cast_fp16)[name = string("op_8065_cast_fp16")];
            tensor<int32, [1]> var_8070_axes_0 = const()[name = string("op_8070_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8070_keep_dims_0 = const()[name = string("op_8070_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 4, 1]> var_8070_cast_fp16 = reduce_mean(axes = var_8070_axes_0, keep_dims = var_8070_keep_dims_0, x = var_8065_cast_fp16)[name = string("op_8070_cast_fp16")];
            fp16 var_8072_to_fp16 = const()[name = string("op_8072_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [1, 96, 4, 1]> var_8073_cast_fp16 = add(x = var_8070_cast_fp16, y = var_8072_to_fp16)[name = string("op_8073_cast_fp16")];
            fp32 var_8074_epsilon_0 = const()[name = string("op_8074_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 4, 1]> var_8074_cast_fp16 = rsqrt(epsilon = var_8074_epsilon_0, x = var_8073_cast_fp16)[name = string("op_8074_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> var_8075_cast_fp16 = mul(x = o_35_cast_fp16, y = var_8074_cast_fp16)[name = string("op_8075_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> var_8077_cast_fp16 = silu(x = input_97_cast_fp16)[name = string("op_8077_cast_fp16")];
            tensor<fp16, [1, 96, 4, 128]> o_37_cast_fp16 = mul(x = var_8075_cast_fp16, y = var_8077_cast_fp16)[name = string("o_37_cast_fp16")];
            tensor<int32, [3]> var_8082 = const()[name = string("op_8082"), val = tensor<int32, [3]>([1, 96, 512])];
            tensor<fp16, [1, 96, 512]> input_99_cast_fp16 = reshape(shape = var_8082, x = o_37_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [640, 512]> layers_6_out_weight_to_fp16 = const()[name = string("layers_6_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97009984)))];
            tensor<fp16, [1, 96, 640]> linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_out_weight_to_fp16, x = input_99_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 96, 640]> x_65_cast_fp16 = add(x = x_61_cast_fp16, y = linear_25_cast_fp16)[name = string("x_65_cast_fp16")];
            fp16 var_8095_to_fp16 = const()[name = string("op_8095_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [1, 96, 640]> x_67_cast_fp16 = mul(x = x_65_cast_fp16, y = var_8095_to_fp16)[name = string("x_67_cast_fp16")];
            fp16 var_8093_promoted_to_fp16 = const()[name = string("op_8093_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> var_8097_cast_fp16 = pow(x = x_67_cast_fp16, y = var_8093_promoted_to_fp16)[name = string("op_8097_cast_fp16")];
            tensor<int32, [1]> var_8099_axes_0 = const()[name = string("op_8099_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8099_keep_dims_0 = const()[name = string("op_8099_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 1]> var_8099_cast_fp16 = reduce_mean(axes = var_8099_axes_0, keep_dims = var_8099_keep_dims_0, x = var_8097_cast_fp16)[name = string("op_8099_cast_fp16")];
            fp16 var_8100_to_fp16 = const()[name = string("op_8100_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 96, 1]> var_8101_cast_fp16 = add(x = var_8099_cast_fp16, y = var_8100_to_fp16)[name = string("op_8101_cast_fp16")];
            fp32 var_8102_epsilon_0 = const()[name = string("op_8102_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 1]> var_8102_cast_fp16 = rsqrt(epsilon = var_8102_epsilon_0, x = var_8101_cast_fp16)[name = string("op_8102_cast_fp16")];
            tensor<fp16, [1, 96, 640]> var_8103_cast_fp16 = mul(x = x_67_cast_fp16, y = var_8102_cast_fp16)[name = string("op_8103_cast_fp16")];
            tensor<fp16, [640]> layers_6_n2_w1_to_fp16 = const()[name = string("layers_6_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97665408)))];
            tensor<fp16, [1, 96, 640]> input_101_cast_fp16 = mul(x = var_8103_cast_fp16, y = layers_6_n2_w1_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_6_gate_up_weight_to_fp16 = const()[name = string("layers_6_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97666752)))];
            tensor<fp16, [1, 96, 5120]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_6_gate_up_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<int32, [3]> input_103_begin_0 = const()[name = string("input_103_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_103_end_0 = const()[name = string("input_103_end_0"), val = tensor<int32, [3]>([1, 96, 2560])];
            tensor<bool, [3]> input_103_end_mask_0 = const()[name = string("input_103_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 2560]> input_103_cast_fp16 = slice_by_index(begin = input_103_begin_0, end = input_103_end_0, end_mask = input_103_end_mask_0, x = linear_26_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [1, 96, 2560]> var_8113_cast_fp16 = silu(x = input_103_cast_fp16)[name = string("op_8113_cast_fp16")];
            tensor<int32, [3]> var_8118_begin_0 = const()[name = string("op_8118_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_8118_end_0 = const()[name = string("op_8118_end_0"), val = tensor<int32, [3]>([1, 96, 5120])];
            tensor<bool, [3]> var_8118_end_mask_0 = const()[name = string("op_8118_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 2560]> var_8118_cast_fp16 = slice_by_index(begin = var_8118_begin_0, end = var_8118_end_0, end_mask = var_8118_end_mask_0, x = linear_26_cast_fp16)[name = string("op_8118_cast_fp16")];
            tensor<fp16, [1, 96, 2560]> input_105_cast_fp16 = mul(x = var_8113_cast_fp16, y = var_8118_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_6_down_weight_to_fp16 = const()[name = string("layers_6_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104220416)))];
            tensor<fp16, [1, 96, 640]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_down_weight_to_fp16, x = input_105_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 96, 640]> x_69_cast_fp16 = add(x = x_65_cast_fp16, y = linear_27_cast_fp16)[name = string("x_69_cast_fp16")];
            fp16 var_8131_to_fp16 = const()[name = string("op_8131_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [1, 96, 640]> x_71_cast_fp16 = mul(x = x_69_cast_fp16, y = var_8131_to_fp16)[name = string("x_71_cast_fp16")];
            fp16 var_8129_promoted_to_fp16 = const()[name = string("op_8129_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> var_8133_cast_fp16 = pow(x = x_71_cast_fp16, y = var_8129_promoted_to_fp16)[name = string("op_8133_cast_fp16")];
            tensor<int32, [1]> var_8135_axes_0 = const()[name = string("op_8135_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8135_keep_dims_0 = const()[name = string("op_8135_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 1]> var_8135_cast_fp16 = reduce_mean(axes = var_8135_axes_0, keep_dims = var_8135_keep_dims_0, x = var_8133_cast_fp16)[name = string("op_8135_cast_fp16")];
            fp16 var_8136_to_fp16 = const()[name = string("op_8136_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 96, 1]> var_8137_cast_fp16 = add(x = var_8135_cast_fp16, y = var_8136_to_fp16)[name = string("op_8137_cast_fp16")];
            fp32 var_8138_epsilon_0 = const()[name = string("op_8138_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 1]> var_8138_cast_fp16 = rsqrt(epsilon = var_8138_epsilon_0, x = var_8137_cast_fp16)[name = string("op_8138_cast_fp16")];
            tensor<fp16, [1, 96, 640]> var_8139_cast_fp16 = mul(x = x_71_cast_fp16, y = var_8138_cast_fp16)[name = string("op_8139_cast_fp16")];
            tensor<fp16, [640]> layers_7_n1_w1_to_fp16 = const()[name = string("layers_7_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107497280)))];
            tensor<fp16, [1, 96, 640]> input_107_cast_fp16 = mul(x = var_8139_cast_fp16, y = layers_7_n1_w1_to_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [1280, 640]> layers_7_qkv_weight_to_fp16 = const()[name = string("layers_7_qkv_weight_to_fp16"), val = tensor<fp16, [1280, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107498624)))];
            tensor<fp16, [1, 96, 1280]> linear_28_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = layers_7_qkv_weight_to_fp16, x = input_107_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [3]> var_8148_begin_0 = const()[name = string("op_8148_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_8148_end_0 = const()[name = string("op_8148_end_0"), val = tensor<int32, [3]>([1, 96, 1024])];
            tensor<bool, [3]> var_8148_end_mask_0 = const()[name = string("op_8148_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 1024]> var_8148_cast_fp16 = slice_by_index(begin = var_8148_begin_0, end = var_8148_end_0, end_mask = var_8148_end_mask_0, x = linear_28_cast_fp16)[name = string("op_8148_cast_fp16")];
            tensor<int32, [4]> var_8153 = const()[name = string("op_8153"), val = tensor<int32, [4]>([1, 96, 8, 128])];
            tensor<fp16, [1, 96, 8, 128]> qg_cast_fp16 = reshape(shape = var_8153, x = var_8148_cast_fp16)[name = string("qg_cast_fp16")];
            tensor<int32, [4]> x_73_begin_0 = const()[name = string("x_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_73_end_0 = const()[name = string("x_73_end_0"), val = tensor<int32, [4]>([1, 96, 8, 64])];
            tensor<bool, [4]> x_73_end_mask_0 = const()[name = string("x_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 96, 8, 64]> x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, x = qg_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<int32, [4]> var_8164_begin_0 = const()[name = string("op_8164_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> var_8164_end_0 = const()[name = string("op_8164_end_0"), val = tensor<int32, [4]>([1, 96, 8, 128])];
            tensor<bool, [4]> var_8164_end_mask_0 = const()[name = string("op_8164_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 96, 8, 64]> var_8164_cast_fp16 = slice_by_index(begin = var_8164_begin_0, end = var_8164_end_0, end_mask = var_8164_end_mask_0, x = qg_cast_fp16)[name = string("op_8164_cast_fp16")];
            tensor<int32, [3]> var_8168 = const()[name = string("op_8168"), val = tensor<int32, [3]>([1, 96, 512])];
            tensor<fp16, [1, 96, 512]> gate_cast_fp16 = reshape(shape = var_8168, x = var_8164_cast_fp16)[name = string("gate_cast_fp16")];
            tensor<int32, [3]> var_8174_begin_0 = const()[name = string("op_8174_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> var_8174_end_0 = const()[name = string("op_8174_end_0"), val = tensor<int32, [3]>([1, 96, 1152])];
            tensor<bool, [3]> var_8174_end_mask_0 = const()[name = string("op_8174_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 128]> var_8174_cast_fp16 = slice_by_index(begin = var_8174_begin_0, end = var_8174_end_0, end_mask = var_8174_end_mask_0, x = linear_28_cast_fp16)[name = string("op_8174_cast_fp16")];
            tensor<int32, [4]> var_8179 = const()[name = string("op_8179"), val = tensor<int32, [4]>([1, 96, 2, 64])];
            tensor<fp16, [1, 96, 2, 64]> x_79_cast_fp16 = reshape(shape = var_8179, x = var_8174_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<int32, [3]> var_8185_begin_0 = const()[name = string("op_8185_begin_0"), val = tensor<int32, [3]>([0, 0, 1152])];
            tensor<int32, [3]> var_8185_end_0 = const()[name = string("op_8185_end_0"), val = tensor<int32, [3]>([1, 96, 1280])];
            tensor<bool, [3]> var_8185_end_mask_0 = const()[name = string("op_8185_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 128]> var_8185_cast_fp16 = slice_by_index(begin = var_8185_begin_0, end = var_8185_end_0, end_mask = var_8185_end_mask_0, x = linear_28_cast_fp16)[name = string("op_8185_cast_fp16")];
            tensor<int32, [4]> var_8190 = const()[name = string("op_8190"), val = tensor<int32, [4]>([1, 96, 2, 64])];
            tensor<fp16, [1, 96, 2, 64]> var_8191_cast_fp16 = reshape(shape = var_8190, x = var_8185_cast_fp16)[name = string("op_8191_cast_fp16")];
            tensor<int32, [4]> var_8196 = const()[name = string("op_8196"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8202_promoted_to_fp16 = const()[name = string("op_8202_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 8, 64]> var_8206_cast_fp16 = pow(x = x_73_cast_fp16, y = var_8202_promoted_to_fp16)[name = string("op_8206_cast_fp16")];
            tensor<int32, [1]> var_8208_axes_0 = const()[name = string("op_8208_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8208_keep_dims_0 = const()[name = string("op_8208_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 8, 1]> var_8208_cast_fp16 = reduce_mean(axes = var_8208_axes_0, keep_dims = var_8208_keep_dims_0, x = var_8206_cast_fp16)[name = string("op_8208_cast_fp16")];
            fp16 var_8209_to_fp16 = const()[name = string("op_8209_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 96, 8, 1]> var_8210_cast_fp16 = add(x = var_8208_cast_fp16, y = var_8209_to_fp16)[name = string("op_8210_cast_fp16")];
            fp32 var_8211_epsilon_0 = const()[name = string("op_8211_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 8, 1]> var_8211_cast_fp16 = rsqrt(epsilon = var_8211_epsilon_0, x = var_8210_cast_fp16)[name = string("op_8211_cast_fp16")];
            tensor<fp16, [1, 96, 8, 64]> var_8212_cast_fp16 = mul(x = x_73_cast_fp16, y = var_8211_cast_fp16)[name = string("op_8212_cast_fp16")];
            tensor<fp16, [64]> layers_7_qn_w1_to_fp16 = const()[name = string("layers_7_qn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137088)))];
            tensor<fp16, [1, 96, 8, 64]> var_8213_cast_fp16 = mul(x = var_8212_cast_fp16, y = layers_7_qn_w1_to_fp16)[name = string("op_8213_cast_fp16")];
            tensor<int32, [4]> var_8218 = const()[name = string("op_8218"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_5_begin_0 = const()[name = string("r_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_5_end_0 = const()[name = string("r_5_end_0"), val = tensor<int32, [4]>([1, 8, 96, 16])];
            tensor<bool, [4]> r_5_end_mask_0 = const()[name = string("r_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 96, 64]> x_77_cast_fp16 = transpose(perm = var_8218, x = var_8213_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [1, 8, 96, 16]> r_5_cast_fp16 = slice_by_index(begin = r_5_begin_0, end = r_5_end_0, end_mask = r_5_end_mask_0, x = x_77_cast_fp16)[name = string("r_5_cast_fp16")];
            tensor<int32, [4]> p_5_begin_0 = const()[name = string("p_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_5_end_0 = const()[name = string("p_5_end_0"), val = tensor<int32, [4]>([1, 8, 96, 64])];
            tensor<bool, [4]> p_5_end_mask_0 = const()[name = string("p_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 96, 48]> p_5_cast_fp16 = slice_by_index(begin = p_5_begin_0, end = p_5_end_0, end_mask = p_5_end_mask_0, x = x_77_cast_fp16)[name = string("p_5_cast_fp16")];
            tensor<int32, [4]> var_8234_begin_0 = const()[name = string("op_8234_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_8234_end_0 = const()[name = string("op_8234_end_0"), val = tensor<int32, [4]>([1, 8, 96, 16])];
            tensor<bool, [4]> var_8234_end_mask_0 = const()[name = string("op_8234_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 96, 8]> var_8234_cast_fp16 = slice_by_index(begin = var_8234_begin_0, end = var_8234_end_0, end_mask = var_8234_end_mask_0, x = r_5_cast_fp16)[name = string("op_8234_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 96, 8]> var_8235_cast_fp16 = mul(x = var_8234_cast_fp16, y = const_44_promoted_to_fp16)[name = string("op_8235_cast_fp16")];
            tensor<int32, [4]> var_8240_begin_0 = const()[name = string("op_8240_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8240_end_0 = const()[name = string("op_8240_end_0"), val = tensor<int32, [4]>([1, 8, 96, 8])];
            tensor<bool, [4]> var_8240_end_mask_0 = const()[name = string("op_8240_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 96, 8]> var_8240_cast_fp16 = slice_by_index(begin = var_8240_begin_0, end = var_8240_end_0, end_mask = var_8240_end_mask_0, x = r_5_cast_fp16)[name = string("op_8240_cast_fp16")];
            int32 var_8242 = const()[name = string("op_8242"), val = int32(-1)];
            bool rh_5_interleave_0 = const()[name = string("rh_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 96, 16]> rh_5_cast_fp16 = concat(axis = var_8242, interleave = rh_5_interleave_0, values = (var_8235_cast_fp16, var_8240_cast_fp16))[name = string("rh_5_cast_fp16")];
            tensor<fp16, [1, 8, 96, 16]> var_8244_cast_fp16 = mul(x = r_5_cast_fp16, y = cos_to_fp16)[name = string("op_8244_cast_fp16")];
            tensor<fp16, [1, 8, 96, 16]> var_8245_cast_fp16 = mul(x = rh_5_cast_fp16, y = sin_to_fp16)[name = string("op_8245_cast_fp16")];
            tensor<fp16, [1, 8, 96, 16]> var_8247_cast_fp16 = add(x = var_8244_cast_fp16, y = var_8245_cast_fp16)[name = string("op_8247_cast_fp16")];
            int32 var_8249 = const()[name = string("op_8249"), val = int32(-1)];
            bool q_interleave_0 = const()[name = string("q_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 96, 64]> q_cast_fp16 = concat(axis = var_8249, interleave = q_interleave_0, values = (var_8247_cast_fp16, p_5_cast_fp16))[name = string("q_cast_fp16")];
            fp16 var_8255_promoted_to_fp16 = const()[name = string("op_8255_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 2, 64]> var_8259_cast_fp16 = pow(x = x_79_cast_fp16, y = var_8255_promoted_to_fp16)[name = string("op_8259_cast_fp16")];
            tensor<int32, [1]> var_8261_axes_0 = const()[name = string("op_8261_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8261_keep_dims_0 = const()[name = string("op_8261_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 2, 1]> var_8261_cast_fp16 = reduce_mean(axes = var_8261_axes_0, keep_dims = var_8261_keep_dims_0, x = var_8259_cast_fp16)[name = string("op_8261_cast_fp16")];
            fp16 var_8262_to_fp16 = const()[name = string("op_8262_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 96, 2, 1]> var_8263_cast_fp16 = add(x = var_8261_cast_fp16, y = var_8262_to_fp16)[name = string("op_8263_cast_fp16")];
            fp32 var_8264_epsilon_0 = const()[name = string("op_8264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 2, 1]> var_8264_cast_fp16 = rsqrt(epsilon = var_8264_epsilon_0, x = var_8263_cast_fp16)[name = string("op_8264_cast_fp16")];
            tensor<fp16, [1, 96, 2, 64]> var_8265_cast_fp16 = mul(x = x_79_cast_fp16, y = var_8264_cast_fp16)[name = string("op_8265_cast_fp16")];
            tensor<fp16, [64]> layers_7_kn_w1_to_fp16 = const()[name = string("layers_7_kn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137280)))];
            tensor<fp16, [1, 96, 2, 64]> var_8266_cast_fp16 = mul(x = var_8265_cast_fp16, y = layers_7_kn_w1_to_fp16)[name = string("op_8266_cast_fp16")];
            tensor<int32, [4]> var_8271 = const()[name = string("op_8271"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_begin_0 = const()[name = string("r_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_end_0 = const()[name = string("r_end_0"), val = tensor<int32, [4]>([1, 2, 96, 16])];
            tensor<bool, [4]> r_end_mask_0 = const()[name = string("r_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 96, 64]> x_83_cast_fp16 = transpose(perm = var_8271, x = var_8266_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 2, 96, 16]> r_cast_fp16 = slice_by_index(begin = r_begin_0, end = r_end_0, end_mask = r_end_mask_0, x = x_83_cast_fp16)[name = string("r_cast_fp16")];
            tensor<int32, [4]> p_begin_0 = const()[name = string("p_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_end_0 = const()[name = string("p_end_0"), val = tensor<int32, [4]>([1, 2, 96, 64])];
            tensor<bool, [4]> p_end_mask_0 = const()[name = string("p_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 96, 48]> p_cast_fp16 = slice_by_index(begin = p_begin_0, end = p_end_0, end_mask = p_end_mask_0, x = x_83_cast_fp16)[name = string("p_cast_fp16")];
            tensor<int32, [4]> var_8287_begin_0 = const()[name = string("op_8287_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_8287_end_0 = const()[name = string("op_8287_end_0"), val = tensor<int32, [4]>([1, 2, 96, 16])];
            tensor<bool, [4]> var_8287_end_mask_0 = const()[name = string("op_8287_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 96, 8]> var_8287_cast_fp16 = slice_by_index(begin = var_8287_begin_0, end = var_8287_end_0, end_mask = var_8287_end_mask_0, x = r_cast_fp16)[name = string("op_8287_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 96, 8]> var_8288_cast_fp16 = mul(x = var_8287_cast_fp16, y = const_45_promoted_to_fp16)[name = string("op_8288_cast_fp16")];
            tensor<int32, [4]> var_8293_begin_0 = const()[name = string("op_8293_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8293_end_0 = const()[name = string("op_8293_end_0"), val = tensor<int32, [4]>([1, 2, 96, 8])];
            tensor<bool, [4]> var_8293_end_mask_0 = const()[name = string("op_8293_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 96, 8]> var_8293_cast_fp16 = slice_by_index(begin = var_8293_begin_0, end = var_8293_end_0, end_mask = var_8293_end_mask_0, x = r_cast_fp16)[name = string("op_8293_cast_fp16")];
            int32 var_8295 = const()[name = string("op_8295"), val = int32(-1)];
            bool rh_interleave_0 = const()[name = string("rh_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 96, 16]> rh_cast_fp16 = concat(axis = var_8295, interleave = rh_interleave_0, values = (var_8288_cast_fp16, var_8293_cast_fp16))[name = string("rh_cast_fp16")];
            tensor<fp16, [1, 2, 96, 16]> var_8297_cast_fp16 = mul(x = r_cast_fp16, y = cos_to_fp16)[name = string("op_8297_cast_fp16")];
            tensor<fp16, [1, 2, 96, 16]> var_8298_cast_fp16 = mul(x = rh_cast_fp16, y = sin_to_fp16)[name = string("op_8298_cast_fp16")];
            tensor<fp16, [1, 2, 96, 16]> var_8300_cast_fp16 = add(x = var_8297_cast_fp16, y = var_8298_cast_fp16)[name = string("op_8300_cast_fp16")];
            int32 var_8302 = const()[name = string("op_8302"), val = int32(-1)];
            bool k_125_interleave_0 = const()[name = string("k_125_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 96, 64]> k_125_cast_fp16 = concat(axis = var_8302, interleave = k_125_interleave_0, values = (var_8300_cast_fp16, p_cast_fp16))[name = string("k_125_cast_fp16")];
            tensor<int32, [1]> var_8315_axes_0 = const()[name = string("op_8315_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 96, 64]> var_8315_cast_fp16 = expand_dims(axes = var_8315_axes_0, x = k_125_cast_fp16)[name = string("op_8315_cast_fp16")];
            tensor<int32, [5]> var_8323_reps_0 = const()[name = string("op_8323_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 2, 4, 96, 64]> var_8323_cast_fp16 = tile(reps = var_8323_reps_0, x = var_8315_cast_fp16)[name = string("op_8323_cast_fp16")];
            tensor<int32, [4]> var_8328 = const()[name = string("op_8328"), val = tensor<int32, [4]>([1, 8, 96, 64])];
            tensor<fp16, [1, 8, 96, 64]> k_cast_fp16 = reshape(shape = var_8328, x = var_8323_cast_fp16)[name = string("k_cast_fp16")];
            tensor<int32, [1]> var_8341_axes_0 = const()[name = string("op_8341_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 96, 64]> v_17_cast_fp16 = transpose(perm = var_8196, x = var_8191_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [1, 2, 1, 96, 64]> var_8341_cast_fp16 = expand_dims(axes = var_8341_axes_0, x = v_17_cast_fp16)[name = string("op_8341_cast_fp16")];
            tensor<int32, [5]> var_8349_reps_0 = const()[name = string("op_8349_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [1, 2, 4, 96, 64]> var_8349_cast_fp16 = tile(reps = var_8349_reps_0, x = var_8341_cast_fp16)[name = string("op_8349_cast_fp16")];
            tensor<int32, [4]> var_8354 = const()[name = string("op_8354"), val = tensor<int32, [4]>([1, 8, 96, 64])];
            tensor<fp16, [1, 8, 96, 64]> v_cast_fp16 = reshape(shape = var_8354, x = var_8349_cast_fp16)[name = string("v_cast_fp16")];
            bool var_8359_transpose_x_1 = const()[name = string("op_8359_transpose_x_1"), val = bool(false)];
            bool var_8359_transpose_y_1 = const()[name = string("op_8359_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 96, 96]> var_8359_cast_fp16 = matmul(transpose_x = var_8359_transpose_x_1, transpose_y = var_8359_transpose_y_1, x = q_cast_fp16, y = k_cast_fp16)[name = string("op_8359_cast_fp16")];
            fp16 var_8360_to_fp16 = const()[name = string("op_8360_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 8, 96, 96]> var_8361_cast_fp16 = mul(x = var_8359_cast_fp16, y = var_8360_to_fp16)[name = string("op_8361_cast_fp16")];
            tensor<fp16, [1, 8, 96, 96]> s_cast_fp16 = add(x = var_8361_cast_fp16, y = causal_add_to_fp16)[name = string("s_cast_fp16")];
            int32 var_8364 = const()[name = string("op_8364"), val = int32(-1)];
            tensor<fp16, [1, 8, 96, 96]> var_8366_cast_fp16 = softmax(axis = var_8364, x = s_cast_fp16)[name = string("op_8366_cast_fp16")];
            bool o_39_transpose_x_0 = const()[name = string("o_39_transpose_x_0"), val = bool(false)];
            bool o_39_transpose_y_0 = const()[name = string("o_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 96, 64]> o_39_cast_fp16 = matmul(transpose_x = o_39_transpose_x_0, transpose_y = o_39_transpose_y_0, x = var_8366_cast_fp16, y = v_cast_fp16)[name = string("o_39_cast_fp16")];
            tensor<int32, [4]> var_8372 = const()[name = string("op_8372"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8377 = const()[name = string("op_8377"), val = tensor<int32, [3]>([1, 96, 512])];
            tensor<fp16, [1, 96, 8, 64]> var_8373_cast_fp16 = transpose(perm = var_8372, x = o_39_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 96, 512]> var_8378_cast_fp16 = reshape(shape = var_8377, x = var_8373_cast_fp16)[name = string("op_8378_cast_fp16")];
            tensor<fp16, [1, 96, 512]> var_8379_cast_fp16 = sigmoid(x = gate_cast_fp16)[name = string("op_8379_cast_fp16")];
            tensor<fp16, [1, 96, 512]> input_109_cast_fp16 = mul(x = var_8378_cast_fp16, y = var_8379_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [640, 512]> layers_7_o_weight_to_fp16 = const()[name = string("layers_7_o_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137472)))];
            tensor<fp16, [1, 96, 640]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_o_weight_to_fp16, x = input_109_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<fp16, [1, 96, 640]> x_85_cast_fp16 = add(x = x_69_cast_fp16, y = linear_29_cast_fp16)[name = string("x_85_cast_fp16")];
            fp16 var_8392_to_fp16 = const()[name = string("op_8392_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [1, 96, 640]> x_87_cast_fp16 = mul(x = x_85_cast_fp16, y = var_8392_to_fp16)[name = string("x_87_cast_fp16")];
            fp16 var_8390_promoted_to_fp16 = const()[name = string("op_8390_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 96, 640]> var_8394_cast_fp16 = pow(x = x_87_cast_fp16, y = var_8390_promoted_to_fp16)[name = string("op_8394_cast_fp16")];
            tensor<int32, [1]> var_8396_axes_0 = const()[name = string("op_8396_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8396_keep_dims_0 = const()[name = string("op_8396_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 96, 1]> var_8396_cast_fp16 = reduce_mean(axes = var_8396_axes_0, keep_dims = var_8396_keep_dims_0, x = var_8394_cast_fp16)[name = string("op_8396_cast_fp16")];
            fp16 var_8397_to_fp16 = const()[name = string("op_8397_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 96, 1]> var_8398_cast_fp16 = add(x = var_8396_cast_fp16, y = var_8397_to_fp16)[name = string("op_8398_cast_fp16")];
            fp32 var_8399_epsilon_0 = const()[name = string("op_8399_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 96, 1]> var_8399_cast_fp16 = rsqrt(epsilon = var_8399_epsilon_0, x = var_8398_cast_fp16)[name = string("op_8399_cast_fp16")];
            tensor<fp16, [1, 96, 640]> var_8400_cast_fp16 = mul(x = x_87_cast_fp16, y = var_8399_cast_fp16)[name = string("op_8400_cast_fp16")];
            tensor<fp16, [640]> layers_7_n2_w1_to_fp16 = const()[name = string("layers_7_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109792896)))];
            tensor<fp16, [1, 96, 640]> input_111_cast_fp16 = mul(x = var_8400_cast_fp16, y = layers_7_n2_w1_to_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_7_gate_up_weight_to_fp16 = const()[name = string("layers_7_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109794240)))];
            tensor<fp16, [1, 96, 5120]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_7_gate_up_weight_to_fp16, x = input_111_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [3]> input_113_begin_0 = const()[name = string("input_113_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_113_end_0 = const()[name = string("input_113_end_0"), val = tensor<int32, [3]>([1, 96, 2560])];
            tensor<bool, [3]> input_113_end_mask_0 = const()[name = string("input_113_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [1, 96, 2560]> input_113_cast_fp16 = slice_by_index(begin = input_113_begin_0, end = input_113_end_0, end_mask = input_113_end_mask_0, x = linear_30_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [1, 96, 2560]> var_8410_cast_fp16 = silu(x = input_113_cast_fp16)[name = string("op_8410_cast_fp16")];
            tensor<int32, [3]> var_8415_begin_0 = const()[name = string("op_8415_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_8415_end_0 = const()[name = string("op_8415_end_0"), val = tensor<int32, [3]>([1, 96, 5120])];
            tensor<bool, [3]> var_8415_end_mask_0 = const()[name = string("op_8415_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 96, 2560]> var_8415_cast_fp16 = slice_by_index(begin = var_8415_begin_0, end = var_8415_end_0, end_mask = var_8415_end_mask_0, x = linear_30_cast_fp16)[name = string("op_8415_cast_fp16")];
            tensor<fp16, [1, 96, 2560]> input_cast_fp16 = mul(x = var_8410_cast_fp16, y = var_8415_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_7_down_weight_to_fp16 = const()[name = string("layers_7_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116347904)))];
            tensor<fp16, [1, 96, 640]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_down_weight_to_fp16, x = input_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 96, 640]> x_89_cast_fp16 = add(x = x_85_cast_fp16, y = linear_31_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<int32, [1]> var_8433_axes_0 = const()[name = string("op_8433_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_8433 = expand_dims(axes = var_8433_axes_0, x = last)[name = string("op_8433")];
            tensor<bool, [1, 96]> var_8434 = equal(x = pos, y = var_8433)[name = string("op_8434")];
            tensor<int32, [1]> var_8451_axes_0 = const()[name = string("op_8451_axes_0"), val = tensor<int32, [1]>([2])];
            string sel_to_fp16_dtype_0 = const()[name = string("sel_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 96]> var_8434_to_fp16 = cast(dtype = sel_to_fp16_dtype_0, x = var_8434)[name = string("cast_131")];
            tensor<fp16, [1, 96, 1]> var_8451_cast_fp16 = expand_dims(axes = var_8451_axes_0, x = var_8434_to_fp16)[name = string("op_8451_cast_fp16")];
            tensor<fp16, [1, 96, 640]> var_8452_cast_fp16 = mul(x = var_8451_cast_fp16, y = x_89_cast_fp16)[name = string("op_8452_cast_fp16")];
            tensor<int32, [1]> x_91_axes_0 = const()[name = string("x_91_axes_0"), val = tensor<int32, [1]>([1])];
            bool x_91_keep_dims_0 = const()[name = string("x_91_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 640]> x_91_cast_fp16 = reduce_sum(axes = x_91_axes_0, keep_dims = x_91_keep_dims_0, x = var_8452_cast_fp16)[name = string("x_91_cast_fp16")];
            fp16 var_8464_to_fp16 = const()[name = string("op_8464_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 640]> x_cast_fp16 = mul(x = x_91_cast_fp16, y = var_8464_to_fp16)[name = string("x_cast_fp16")];
            fp16 var_8462_promoted_to_fp16 = const()[name = string("op_8462_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 640]> var_8466_cast_fp16 = pow(x = x_cast_fp16, y = var_8462_promoted_to_fp16)[name = string("op_8466_cast_fp16")];
            tensor<int32, [1]> var_8468_axes_0 = const()[name = string("op_8468_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8468_keep_dims_0 = const()[name = string("op_8468_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_8468_cast_fp16 = reduce_mean(axes = var_8468_axes_0, keep_dims = var_8468_keep_dims_0, x = var_8466_cast_fp16)[name = string("op_8468_cast_fp16")];
            fp16 var_8469_to_fp16 = const()[name = string("op_8469_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 1]> var_8470_cast_fp16 = add(x = var_8468_cast_fp16, y = var_8469_to_fp16)[name = string("op_8470_cast_fp16")];
            fp32 var_8471_epsilon_0 = const()[name = string("op_8471_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1]> var_8471_cast_fp16 = rsqrt(epsilon = var_8471_epsilon_0, x = var_8470_cast_fp16)[name = string("op_8471_cast_fp16")];
            tensor<fp16, [1, 640]> var_8472_cast_fp16 = mul(x = x_cast_fp16, y = var_8471_cast_fp16)[name = string("op_8472_cast_fp16")];
            tensor<fp16, [640]> norm_w1_to_fp16 = const()[name = string("norm_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119624768)))];
            tensor<fp16, [1, 640]> h_cast_fp16 = mul(x = var_8472_cast_fp16, y = norm_w1_to_fp16)[name = string("h_cast_fp16")];
            tensor<fp16, [16000]> logits_bias_0_to_fp16 = const()[name = string("logits_bias_0_to_fp16"), val = tensor<fp16, [16000]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119626112)))];
            tensor<fp16, [1, 16000]> logits_cast_fp16 = linear(bias = logits_bias_0_to_fp16, weight = emb_to_fp16, x = h_cast_fp16)[name = string("logits_cast_fp16")];
            tensor<int32, [1]> reduce_max_0_axes_0 = const()[name = string("reduce_max_0_axes_0"), val = tensor<int32, [1]>([-1])];
            bool reduce_max_0_keep_dims_0 = const()[name = string("reduce_max_0_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> reduce_max_0_cast_fp16 = reduce_max(axes = reduce_max_0_axes_0, keep_dims = reduce_max_0_keep_dims_0, x = logits_cast_fp16)[name = string("reduce_max_0_cast_fp16")];
            tensor<fp16, [1, 16000]> var_8481_cast_fp16 = sub(x = logits_cast_fp16, y = reduce_max_0_cast_fp16)[name = string("op_8481_cast_fp16")];
            tensor<fp16, [1, 16000]> var_8482_cast_fp16 = exp(x = var_8481_cast_fp16)[name = string("op_8482_cast_fp16")];
            tensor<int32, [1]> var_8487_axes_0 = const()[name = string("op_8487_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8487_keep_dims_0 = const()[name = string("op_8487_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_8487_cast_fp16 = reduce_sum(axes = var_8487_axes_0, keep_dims = var_8487_keep_dims_0, x = var_8482_cast_fp16)[name = string("op_8487_cast_fp16")];
            fp32 var_8488_epsilon_0 = const()[name = string("op_8488_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [1, 1]> var_8488_cast_fp16 = log(epsilon = var_8488_epsilon_0, x = var_8487_cast_fp16)[name = string("op_8488_cast_fp16")];
            tensor<fp16, [1, 1]> var_8490_cast_fp16 = add(x = reduce_max_0_cast_fp16, y = var_8488_cast_fp16)[name = string("op_8490_cast_fp16")];
            tensor<fp16, [1, 16000]> var_8492_cast_fp16 = sub(x = logits_cast_fp16, y = var_8490_cast_fp16)[name = string("op_8492_cast_fp16")];
            string var_8492_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_8492_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 16000]> lp = cast(dtype = var_8492_cast_fp16_to_fp32_dtype_0, x = var_8492_cast_fp16)[name = string("cast_130")];
        } -> (lp);
    func t16<ios18>(tensor<int32, [3, 16]> ids) {
            int32 x_1_batch_dims_0 = const()[name = string("x_1_batch_dims_0"), val = int32(0)];
            bool x_1_validate_indices_0 = const()[name = string("x_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [16000, 640]> emb_to_fp16 = const()[name = string("emb_to_fp16"), val = tensor<fp16, [16000, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            string ids_to_int16_dtype_0 = const()[name = string("ids_to_int16_dtype_0"), val = string("int16")];
            string cast_70_dtype_0 = const()[name = string("cast_70_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int16, [3, 16]> ids_to_int16 = cast(dtype = ids_to_int16_dtype_0, x = ids)[name = string("cast_75")];
            tensor<int32, [3, 16]> cast_70 = cast(dtype = cast_70_dtype_0, x = ids_to_int16)[name = string("cast_74")];
            tensor<bool, [3, 16]> greater_equal_0 = greater_equal(x = cast_70, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(16000)];
            tensor<int32, [3, 16]> add_0 = add(x = cast_70, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [3, 16]> select_0 = select(a = cast_70, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 x_1_cast_fp16_cast_uint16_axis_0 = const()[name = string("x_1_cast_fp16_cast_uint16_axis_0"), val = int32(0)];
            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
            tensor<int16, [3, 16]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_73")];
            tensor<fp16, [3, 16, 640]> x_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = x_1_cast_fp16_cast_uint16_axis_0, batch_dims = x_1_batch_dims_0, indices = select_0_to_int16, validate_indices = x_1_validate_indices_0, x = emb_to_fp16)[name = string("x_1_cast_fp16_cast_uint16_cast_uint16")];
            fp16 var_251_to_fp16 = const()[name = string("op_251_to_fp16"), val = fp16(0x1p+5)];
            tensor<fp16, [3, 16, 640]> x_3_cast_fp16 = mul(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = var_251_to_fp16)[name = string("x_3_cast_fp16")];
            fp16 var_249_promoted_to_fp16 = const()[name = string("op_249_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> var_253_cast_fp16 = pow(x = x_3_cast_fp16, y = var_249_promoted_to_fp16)[name = string("op_253_cast_fp16")];
            tensor<int32, [1]> var_255_axes_0 = const()[name = string("op_255_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_255_keep_dims_0 = const()[name = string("op_255_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 1]> var_255_cast_fp16 = reduce_mean(axes = var_255_axes_0, keep_dims = var_255_keep_dims_0, x = var_253_cast_fp16)[name = string("op_255_cast_fp16")];
            fp16 var_256_to_fp16 = const()[name = string("op_256_to_fp16"), val = fp16(0x1.0c8p-10)];
            tensor<fp16, [3, 16, 1]> var_257_cast_fp16 = add(x = var_255_cast_fp16, y = var_256_to_fp16)[name = string("op_257_cast_fp16")];
            fp32 var_258_epsilon_0 = const()[name = string("op_258_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 1]> var_258_cast_fp16 = rsqrt(epsilon = var_258_epsilon_0, x = var_257_cast_fp16)[name = string("op_258_cast_fp16")];
            tensor<fp16, [3, 16, 640]> var_259_cast_fp16 = mul(x = x_3_cast_fp16, y = var_258_cast_fp16)[name = string("op_259_cast_fp16")];
            tensor<fp16, [640]> layers_0_n1_w1_to_fp16 = const()[name = string("layers_0_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20480128)))];
            tensor<fp16, [3, 16, 640]> input_1_cast_fp16 = mul(x = var_259_cast_fp16, y = layers_0_n1_w1_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_0_qkvzba_weight_to_fp16 = const()[name = string("layers_0_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20481472)))];
            tensor<fp16, [1544]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [1544]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22457856)))];
            tensor<fp16, [3, 16, 1544]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_qkvzba_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [3]> qkv_1_begin_0 = const()[name = string("qkv_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_1_end_0 = const()[name = string("qkv_1_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> qkv_1_end_mask_0 = const()[name = string("qkv_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 1024]> qkv_1_cast_fp16 = slice_by_index(begin = qkv_1_begin_0, end = qkv_1_end_0, end_mask = qkv_1_end_mask_0, x = linear_0_cast_fp16)[name = string("qkv_1_cast_fp16")];
            tensor<int32, [3]> z_1_begin_0 = const()[name = string("z_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_1_end_0 = const()[name = string("z_1_end_0"), val = tensor<int32, [3]>([3, 16, 1536])];
            tensor<bool, [3]> z_1_end_mask_0 = const()[name = string("z_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 512]> z_1_cast_fp16 = slice_by_index(begin = z_1_begin_0, end = z_1_end_0, end_mask = z_1_end_mask_0, x = linear_0_cast_fp16)[name = string("z_1_cast_fp16")];
            tensor<int32, [3]> b_1_begin_0 = const()[name = string("b_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_1_end_0 = const()[name = string("b_1_end_0"), val = tensor<int32, [3]>([3, 16, 1540])];
            tensor<bool, [3]> b_1_end_mask_0 = const()[name = string("b_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 4]> b_1_cast_fp16 = slice_by_index(begin = b_1_begin_0, end = b_1_end_0, end_mask = b_1_end_mask_0, x = linear_0_cast_fp16)[name = string("b_1_cast_fp16")];
            tensor<int32, [3]> a_1_begin_0 = const()[name = string("a_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_1_end_0 = const()[name = string("a_1_end_0"), val = tensor<int32, [3]>([3, 16, 1])];
            tensor<bool, [3]> a_1_end_mask_0 = const()[name = string("a_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 4]> a_1_cast_fp16 = slice_by_index(begin = a_1_begin_0, end = a_1_end_0, end_mask = a_1_end_mask_0, x = linear_0_cast_fp16)[name = string("a_1_cast_fp16")];
            int32 var_298 = const()[name = string("op_298"), val = int32(1)];
            bool xp_1_interleave_0 = const()[name = string("xp_1_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 3, 1024]> zpad_to_fp16 = const()[name = string("zpad_to_fp16"), val = tensor<fp16, [3, 3, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22461056)))];
            tensor<fp16, [3, 19, 1024]> xp_1_cast_fp16 = concat(axis = var_298, interleave = xp_1_interleave_0, values = (zpad_to_fp16, qkv_1_cast_fp16))[name = string("xp_1_cast_fp16")];
            tensor<int32, [3]> var_309_begin_0 = const()[name = string("op_309_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_309_end_0 = const()[name = string("op_309_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_309_end_mask_0 = const()[name = string("op_309_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 1024]> var_309_cast_fp16 = slice_by_index(begin = var_309_begin_0, end = var_309_end_0, end_mask = var_309_end_mask_0, x = xp_1_cast_fp16)[name = string("op_309_cast_fp16")];
            tensor<fp16, [1024]> var_317_to_fp16 = const()[name = string("op_317_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22479552)))];
            tensor<fp16, [3, 16, 1024]> var_318_cast_fp16 = mul(x = var_309_cast_fp16, y = var_317_to_fp16)[name = string("op_318_cast_fp16")];
            tensor<int32, [3]> var_328_begin_0 = const()[name = string("op_328_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_328_end_0 = const()[name = string("op_328_end_0"), val = tensor<int32, [3]>([3, 18, 1024])];
            tensor<bool, [3]> var_328_end_mask_0 = const()[name = string("op_328_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_328_cast_fp16 = slice_by_index(begin = var_328_begin_0, end = var_328_end_0, end_mask = var_328_end_mask_0, x = xp_1_cast_fp16)[name = string("op_328_cast_fp16")];
            tensor<fp16, [1024]> var_336_to_fp16 = const()[name = string("op_336_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22481664)))];
            tensor<fp16, [3, 16, 1024]> var_337_cast_fp16 = mul(x = var_328_cast_fp16, y = var_336_to_fp16)[name = string("op_337_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> var_339_cast_fp16 = add(x = var_318_cast_fp16, y = var_337_cast_fp16)[name = string("op_339_cast_fp16")];
            tensor<int32, [3]> var_349_begin_0 = const()[name = string("op_349_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_349_end_0 = const()[name = string("op_349_end_0"), val = tensor<int32, [3]>([3, 17, 1024])];
            tensor<bool, [3]> var_349_end_mask_0 = const()[name = string("op_349_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_349_cast_fp16 = slice_by_index(begin = var_349_begin_0, end = var_349_end_0, end_mask = var_349_end_mask_0, x = xp_1_cast_fp16)[name = string("op_349_cast_fp16")];
            tensor<fp16, [1024]> var_357_to_fp16 = const()[name = string("op_357_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22483776)))];
            tensor<fp16, [3, 16, 1024]> var_358_cast_fp16 = mul(x = var_349_cast_fp16, y = var_357_to_fp16)[name = string("op_358_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> var_360_cast_fp16 = add(x = var_339_cast_fp16, y = var_358_cast_fp16)[name = string("op_360_cast_fp16")];
            tensor<int32, [3]> var_370_begin_0 = const()[name = string("op_370_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_370_end_0 = const()[name = string("op_370_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> var_370_end_mask_0 = const()[name = string("op_370_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_370_cast_fp16 = slice_by_index(begin = var_370_begin_0, end = var_370_end_0, end_mask = var_370_end_mask_0, x = xp_1_cast_fp16)[name = string("op_370_cast_fp16")];
            tensor<fp16, [1024]> var_378_to_fp16 = const()[name = string("op_378_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22485888)))];
            tensor<fp16, [3, 16, 1024]> var_379_cast_fp16 = mul(x = var_370_cast_fp16, y = var_378_to_fp16)[name = string("op_379_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> input_3_cast_fp16 = add(x = var_360_cast_fp16, y = var_379_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> c_1_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("c_1_cast_fp16")];
            tensor<int32, [3]> var_387_begin_0 = const()[name = string("op_387_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_387_end_0 = const()[name = string("op_387_end_0"), val = tensor<int32, [3]>([3, 16, 256])];
            tensor<bool, [3]> var_387_end_mask_0 = const()[name = string("op_387_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 256]> var_387_cast_fp16 = slice_by_index(begin = var_387_begin_0, end = var_387_end_0, end_mask = var_387_end_mask_0, x = c_1_cast_fp16)[name = string("op_387_cast_fp16")];
            tensor<int32, [4]> var_392 = const()[name = string("op_392"), val = tensor<int32, [4]>([3, 16, 2, 128])];
            tensor<fp16, [3, 16, 2, 128]> q_1_cast_fp16 = reshape(shape = var_392, x = var_387_cast_fp16)[name = string("q_1_cast_fp16")];
            tensor<int32, [3]> var_398_begin_0 = const()[name = string("op_398_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_398_end_0 = const()[name = string("op_398_end_0"), val = tensor<int32, [3]>([3, 16, 512])];
            tensor<bool, [3]> var_398_end_mask_0 = const()[name = string("op_398_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 256]> var_398_cast_fp16 = slice_by_index(begin = var_398_begin_0, end = var_398_end_0, end_mask = var_398_end_mask_0, x = c_1_cast_fp16)[name = string("op_398_cast_fp16")];
            tensor<int32, [4]> var_403 = const()[name = string("op_403"), val = tensor<int32, [4]>([3, 16, 2, 128])];
            tensor<fp16, [3, 16, 2, 128]> k_1_cast_fp16 = reshape(shape = var_403, x = var_398_cast_fp16)[name = string("k_1_cast_fp16")];
            tensor<int32, [3]> var_409_begin_0 = const()[name = string("op_409_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_409_end_0 = const()[name = string("op_409_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> var_409_end_mask_0 = const()[name = string("op_409_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 512]> var_409_cast_fp16 = slice_by_index(begin = var_409_begin_0, end = var_409_end_0, end_mask = var_409_end_mask_0, x = c_1_cast_fp16)[name = string("op_409_cast_fp16")];
            tensor<int32, [4]> var_414 = const()[name = string("op_414"), val = tensor<int32, [4]>([3, 16, 4, 128])];
            tensor<fp16, [3, 16, 4, 128]> var_415_cast_fp16 = reshape(shape = var_414, x = var_409_cast_fp16)[name = string("op_415_cast_fp16")];
            tensor<int32, [4]> var_420 = const()[name = string("op_420"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_422_to_fp16 = const()[name = string("op_422_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 16, 2, 128]> q_3_cast_fp16 = mul(x = q_1_cast_fp16, y = var_422_to_fp16)[name = string("q_3_cast_fp16")];
            fp16 var_424_to_fp16 = const()[name = string("op_424_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 16, 2, 128]> k_3_cast_fp16 = mul(x = k_1_cast_fp16, y = var_424_to_fp16)[name = string("k_3_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_426_cast_fp16 = mul(x = q_3_cast_fp16, y = q_3_cast_fp16)[name = string("op_426_cast_fp16")];
            tensor<int32, [1]> var_431_axes_0 = const()[name = string("op_431_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_431_keep_dims_0 = const()[name = string("op_431_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 2, 1]> var_431_cast_fp16 = reduce_mean(axes = var_431_axes_0, keep_dims = var_431_keep_dims_0, x = var_426_cast_fp16)[name = string("op_431_cast_fp16")];
            fp16 var_432_promoted_to_fp16 = const()[name = string("op_432_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 16, 2, 1]> var_433_cast_fp16 = mul(x = var_431_cast_fp16, y = var_432_promoted_to_fp16)[name = string("op_433_cast_fp16")];
            fp16 var_435_to_fp16 = const()[name = string("op_435_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 16, 2, 1]> var_436_cast_fp16 = add(x = var_433_cast_fp16, y = var_435_to_fp16)[name = string("op_436_cast_fp16")];
            fp32 var_437_epsilon_0 = const()[name = string("op_437_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 2, 1]> var_437_cast_fp16 = rsqrt(epsilon = var_437_epsilon_0, x = var_436_cast_fp16)[name = string("op_437_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_438_cast_fp16 = mul(x = q_3_cast_fp16, y = var_437_cast_fp16)[name = string("op_438_cast_fp16")];
            fp16 var_439_to_fp16 = const()[name = string("op_439_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 16, 2, 128]> q_5_cast_fp16 = mul(x = var_438_cast_fp16, y = var_439_to_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_441_cast_fp16 = mul(x = k_3_cast_fp16, y = k_3_cast_fp16)[name = string("op_441_cast_fp16")];
            tensor<int32, [1]> var_446_axes_0 = const()[name = string("op_446_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_446_keep_dims_0 = const()[name = string("op_446_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 2, 1]> var_446_cast_fp16 = reduce_mean(axes = var_446_axes_0, keep_dims = var_446_keep_dims_0, x = var_441_cast_fp16)[name = string("op_446_cast_fp16")];
            fp16 var_447_promoted_to_fp16 = const()[name = string("op_447_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 16, 2, 1]> var_448_cast_fp16 = mul(x = var_446_cast_fp16, y = var_447_promoted_to_fp16)[name = string("op_448_cast_fp16")];
            fp16 var_450_to_fp16 = const()[name = string("op_450_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 16, 2, 1]> var_451_cast_fp16 = add(x = var_448_cast_fp16, y = var_450_to_fp16)[name = string("op_451_cast_fp16")];
            fp32 var_452_epsilon_0 = const()[name = string("op_452_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 2, 1]> var_452_cast_fp16 = rsqrt(epsilon = var_452_epsilon_0, x = var_451_cast_fp16)[name = string("op_452_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> k_5_cast_fp16 = mul(x = k_3_cast_fp16, y = var_452_cast_fp16)[name = string("k_5_cast_fp16")];
            tensor<int32, [4]> var_458 = const()[name = string("op_458"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_471_axes_0 = const()[name = string("op_471_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 16, 128]> var_459_cast_fp16 = transpose(perm = var_458, x = q_5_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [3, 2, 1, 16, 128]> var_471_cast_fp16 = expand_dims(axes = var_471_axes_0, x = var_459_cast_fp16)[name = string("op_471_cast_fp16")];
            tensor<int32, [5]> var_479_reps_0 = const()[name = string("op_479_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 16, 128]> var_479_cast_fp16 = tile(reps = var_479_reps_0, x = var_471_cast_fp16)[name = string("op_479_cast_fp16")];
            tensor<int32, [4]> var_484 = const()[name = string("op_484"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<fp16, [3, 4, 16, 128]> q_7_cast_fp16 = reshape(shape = var_484, x = var_479_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<int32, [4]> var_490 = const()[name = string("op_490"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_503_axes_0 = const()[name = string("op_503_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 16, 128]> var_491_cast_fp16 = transpose(perm = var_490, x = k_5_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [3, 2, 1, 16, 128]> var_503_cast_fp16 = expand_dims(axes = var_503_axes_0, x = var_491_cast_fp16)[name = string("op_503_cast_fp16")];
            tensor<int32, [5]> var_511_reps_0 = const()[name = string("op_511_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 16, 128]> var_511_cast_fp16 = tile(reps = var_511_reps_0, x = var_503_cast_fp16)[name = string("op_511_cast_fp16")];
            tensor<int32, [4]> var_516 = const()[name = string("op_516"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<fp16, [3, 4, 16, 128]> k_7_cast_fp16 = reshape(shape = var_516, x = var_511_cast_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_518_cast_fp16 = sigmoid(x = b_1_cast_fp16)[name = string("op_518_cast_fp16")];
            tensor<int32, [3]> var_522 = const()[name = string("op_522"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_0_dt_bias_to_fp16 = const()[name = string("layers_0_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 16, 4]> input_5_cast_fp16 = add(x = a_1_cast_fp16, y = layers_0_dt_bias_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_526_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("op_526_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_527_cast_fp16 = abs(x = input_5_cast_fp16)[name = string("op_527_cast_fp16")];
            fp16 const_0_promoted_to_fp16 = const()[name = string("const_0_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 16, 4]> var_528_cast_fp16 = mul(x = var_527_cast_fp16, y = const_0_promoted_to_fp16)[name = string("op_528_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_529_cast_fp16 = exp(x = var_528_cast_fp16)[name = string("op_529_cast_fp16")];
            fp16 var_531_to_fp16 = const()[name = string("op_531_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 16, 4]> var_532_cast_fp16 = add(x = var_529_cast_fp16, y = var_531_to_fp16)[name = string("op_532_cast_fp16")];
            fp32 var_533_epsilon_0 = const()[name = string("op_533_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 16, 4]> var_533_cast_fp16 = log(epsilon = var_533_epsilon_0, x = var_532_cast_fp16)[name = string("op_533_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_535_cast_fp16 = add(x = var_526_cast_fp16, y = var_533_cast_fp16)[name = string("op_535_cast_fp16")];
            tensor<fp16, [4]> layers_0_negA_to_fp16 = const()[name = string("layers_0_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.b3p+3, -0x1.a7cp+0, -0x1.ac4p+3, -0x1.704p+2])];
            tensor<fp16, [3, 16, 4]> var_536_cast_fp16 = mul(x = layers_0_negA_to_fp16, y = var_535_cast_fp16)[name = string("op_536_cast_fp16")];
            tensor<int32, [3]> var_540 = const()[name = string("op_540"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_543_axes_0 = const()[name = string("op_543_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16]> beta_1_cast_fp16 = transpose(perm = var_522, x = var_518_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [3, 4, 16, 1]> var_543_cast_fp16 = expand_dims(axes = var_543_axes_0, x = beta_1_cast_fp16)[name = string("op_543_cast_fp16")];
            fp16 var_544_to_fp16 = const()[name = string("op_544_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 16, 1]> var_545_cast_fp16 = mul(x = var_543_cast_fp16, y = var_544_to_fp16)[name = string("op_545_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> v_1_cast_fp16 = transpose(perm = var_420, x = var_415_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [3, 4, 16, 128]> vb_1_cast_fp16 = mul(x = v_1_cast_fp16, y = var_545_cast_fp16)[name = string("vb_1_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> kb_1_cast_fp16 = mul(x = k_7_cast_fp16, y = var_543_cast_fp16)[name = string("kb_1_cast_fp16")];
            tensor<int32, [2]> var_552 = const()[name = string("op_552"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [3, 4, 16]> gdec_1_cast_fp16 = transpose(perm = var_540, x = var_536_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [12, 16]> var_553_cast_fp16 = reshape(shape = var_552, x = gdec_1_cast_fp16)[name = string("op_553_cast_fp16")];
            tensor<fp16, [256, 16]> transpose_0_to_fp16 = const()[name = string("transpose_0_to_fp16"), val = tensor<fp16, [256, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22488000)))];
            tensor<fp16, [256]> var_554_bias_0_to_fp16 = const()[name = string("op_554_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22496256)))];
            tensor<fp16, [12, 256]> var_554_cast_fp16 = linear(bias = var_554_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_553_cast_fp16)[name = string("op_554_cast_fp16")];
            tensor<int32, [4]> var_559 = const()[name = string("op_559"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_1_cast_fp16 = reshape(shape = var_559, x = var_554_cast_fp16)[name = string("P_1_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_561_cast_fp16 = exp(x = P_1_cast_fp16)[name = string("op_561_cast_fp16")];
            tensor<fp16, [16, 16]> tril_incl_to_fp16 = const()[name = string("tril_incl_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22496832)))];
            tensor<fp16, [3, 4, 16, 16]> D_1_cast_fp16 = mul(x = var_561_cast_fp16, y = tril_incl_to_fp16)[name = string("D_1_cast_fp16")];
            bool kk_1_transpose_x_1 = const()[name = string("kk_1_transpose_x_1"), val = bool(false)];
            bool kk_1_transpose_y_1 = const()[name = string("kk_1_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_1_cast_fp16 = matmul(transpose_x = kk_1_transpose_x_1, transpose_y = kk_1_transpose_y_1, x = kb_1_cast_fp16, y = k_7_cast_fp16)[name = string("kk_1_cast_fp16")];
            fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_567_cast_fp16 = mul(x = kk_1_cast_fp16, y = const_1_promoted_to_fp16)[name = string("op_567_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_568_cast_fp16 = mul(x = var_567_cast_fp16, y = D_1_cast_fp16)[name = string("op_568_cast_fp16")];
            tensor<fp16, [16, 16]> tril_strict_to_fp16 = const()[name = string("tril_strict_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22497408)))];
            tensor<fp16, [3, 4, 16, 16]> N_1_cast_fp16 = mul(x = var_568_cast_fp16, y = tril_strict_to_fp16)[name = string("N_1_cast_fp16")];
            tensor<fp16, [16, 16]> neg_tril_strict_to_fp16 = const()[name = string("neg_tril_strict_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22497984)))];
            tensor<fp16, [3, 4, 16, 16]> var_570_cast_fp16 = mul(x = D_1_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_570_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_1_cast_fp16 = mul(x = kk_1_cast_fp16, y = var_570_cast_fp16)[name = string("Na_1_cast_fp16")];
            tensor<fp16, [16, 16]> eye_to_fp16 = const()[name = string("eye_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22498560)))];
            tensor<fp16, [3, 4, 16, 16]> Minv_1_cast_fp16 = add(x = eye_to_fp16, y = N_1_cast_fp16)[name = string("Minv_1_cast_fp16")];
            bool Np_1_transpose_x_0 = const()[name = string("Np_1_transpose_x_0"), val = bool(false)];
            bool Np_1_transpose_y_0 = const()[name = string("Np_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_1_cast_fp16 = matmul(transpose_x = Np_1_transpose_x_0, transpose_y = Np_1_transpose_y_0, x = N_1_cast_fp16, y = Na_1_cast_fp16)[name = string("Np_1_cast_fp16")];
            bool Npa_1_transpose_x_0 = const()[name = string("Npa_1_transpose_x_0"), val = bool(false)];
            bool Npa_1_transpose_y_0 = const()[name = string("Npa_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_1_cast_fp16 = matmul(transpose_x = Npa_1_transpose_x_0, transpose_y = Npa_1_transpose_y_0, x = Na_1_cast_fp16, y = N_1_cast_fp16)[name = string("Npa_1_cast_fp16")];
            bool var_576_transpose_x_0 = const()[name = string("op_576_transpose_x_0"), val = bool(false)];
            bool var_576_transpose_y_0 = const()[name = string("op_576_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_576_cast_fp16 = matmul(transpose_x = var_576_transpose_x_0, transpose_y = var_576_transpose_y_0, x = Minv_1_cast_fp16, y = Np_1_cast_fp16)[name = string("op_576_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_3_cast_fp16 = add(x = Minv_1_cast_fp16, y = var_576_cast_fp16)[name = string("Minv_3_cast_fp16")];
            bool Np_3_transpose_x_0 = const()[name = string("Np_3_transpose_x_0"), val = bool(false)];
            bool Np_3_transpose_y_0 = const()[name = string("Np_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_3_cast_fp16 = matmul(transpose_x = Np_3_transpose_x_0, transpose_y = Np_3_transpose_y_0, x = Np_1_cast_fp16, y = Npa_1_cast_fp16)[name = string("Np_3_cast_fp16")];
            bool Npa_3_transpose_x_0 = const()[name = string("Npa_3_transpose_x_0"), val = bool(false)];
            bool Npa_3_transpose_y_0 = const()[name = string("Npa_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_3_cast_fp16 = matmul(transpose_x = Npa_3_transpose_x_0, transpose_y = Npa_3_transpose_y_0, x = Npa_1_cast_fp16, y = Np_1_cast_fp16)[name = string("Npa_3_cast_fp16")];
            bool var_581_transpose_x_0 = const()[name = string("op_581_transpose_x_0"), val = bool(false)];
            bool var_581_transpose_y_0 = const()[name = string("op_581_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_581_cast_fp16 = matmul(transpose_x = var_581_transpose_x_0, transpose_y = var_581_transpose_y_0, x = Minv_3_cast_fp16, y = Np_3_cast_fp16)[name = string("op_581_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_5_cast_fp16 = add(x = Minv_3_cast_fp16, y = var_581_cast_fp16)[name = string("Minv_5_cast_fp16")];
            bool Np_5_transpose_x_0 = const()[name = string("Np_5_transpose_x_0"), val = bool(false)];
            bool Np_5_transpose_y_0 = const()[name = string("Np_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_5_cast_fp16 = matmul(transpose_x = Np_5_transpose_x_0, transpose_y = Np_5_transpose_y_0, x = Np_3_cast_fp16, y = Npa_3_cast_fp16)[name = string("Np_5_cast_fp16")];
            bool var_585_transpose_x_0 = const()[name = string("op_585_transpose_x_0"), val = bool(false)];
            bool var_585_transpose_y_0 = const()[name = string("op_585_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_585_cast_fp16 = matmul(transpose_x = var_585_transpose_x_0, transpose_y = var_585_transpose_y_0, x = Minv_5_cast_fp16, y = Np_5_cast_fp16)[name = string("op_585_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_7_cast_fp16 = add(x = Minv_5_cast_fp16, y = var_585_cast_fp16)[name = string("Minv_7_cast_fp16")];
            bool u_1_transpose_x_0 = const()[name = string("u_1_transpose_x_0"), val = bool(false)];
            bool u_1_transpose_y_0 = const()[name = string("u_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_1_cast_fp16 = matmul(transpose_x = u_1_transpose_x_0, transpose_y = u_1_transpose_y_0, x = Minv_7_cast_fp16, y = vb_1_cast_fp16)[name = string("u_1_cast_fp16")];
            bool var_592_transpose_x_1 = const()[name = string("op_592_transpose_x_1"), val = bool(false)];
            bool var_592_transpose_y_1 = const()[name = string("op_592_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_592_cast_fp16 = matmul(transpose_x = var_592_transpose_x_1, transpose_y = var_592_transpose_y_1, x = q_7_cast_fp16, y = k_7_cast_fp16)[name = string("op_592_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_1_cast_fp16 = mul(x = var_592_cast_fp16, y = D_1_cast_fp16)[name = string("A_1_cast_fp16")];
            bool o_1_transpose_x_0 = const()[name = string("o_1_transpose_x_0"), val = bool(false)];
            bool o_1_transpose_y_0 = const()[name = string("o_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> o_1_cast_fp16 = matmul(transpose_x = o_1_transpose_x_0, transpose_y = o_1_transpose_y_0, x = A_1_cast_fp16, y = u_1_cast_fp16)[name = string("o_1_cast_fp16")];
            tensor<int32, [4]> var_599 = const()[name = string("op_599"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_605 = const()[name = string("op_605"), val = tensor<int32, [4]>([3, 16, 4, 128])];
            tensor<fp16, [3, 16, 4, 128]> input_7_cast_fp16 = reshape(shape = var_605, x = z_1_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> o_3_cast_fp16 = transpose(perm = var_599, x = o_1_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [3, 16, 4, 128]> var_607_cast_fp16 = mul(x = o_3_cast_fp16, y = o_3_cast_fp16)[name = string("op_607_cast_fp16")];
            tensor<int32, [1]> var_612_axes_0 = const()[name = string("op_612_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_612_keep_dims_0 = const()[name = string("op_612_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 4, 1]> var_612_cast_fp16 = reduce_mean(axes = var_612_axes_0, keep_dims = var_612_keep_dims_0, x = var_607_cast_fp16)[name = string("op_612_cast_fp16")];
            fp16 var_614_to_fp16 = const()[name = string("op_614_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 16, 4, 1]> var_615_cast_fp16 = add(x = var_612_cast_fp16, y = var_614_to_fp16)[name = string("op_615_cast_fp16")];
            fp32 var_616_epsilon_0 = const()[name = string("op_616_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 4, 1]> var_616_cast_fp16 = rsqrt(epsilon = var_616_epsilon_0, x = var_615_cast_fp16)[name = string("op_616_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> var_617_cast_fp16 = mul(x = o_3_cast_fp16, y = var_616_cast_fp16)[name = string("op_617_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> var_619_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_619_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> o_5_cast_fp16 = mul(x = var_617_cast_fp16, y = var_619_cast_fp16)[name = string("o_5_cast_fp16")];
            tensor<int32, [3]> var_624 = const()[name = string("op_624"), val = tensor<int32, [3]>([3, 16, 512])];
            tensor<fp16, [3, 16, 512]> input_9_cast_fp16 = reshape(shape = var_624, x = o_5_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [640, 512]> layers_0_out_weight_to_fp16 = const()[name = string("layers_0_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22499136)))];
            tensor<fp16, [640]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23154560)))];
            tensor<fp16, [3, 16, 640]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_out_weight_to_fp16, x = input_9_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [3, 16, 640]> x_5_cast_fp16 = add(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = linear_1_cast_fp16)[name = string("x_5_cast_fp16")];
            fp16 var_637_to_fp16 = const()[name = string("op_637_to_fp16"), val = fp16(0x1p+5)];
            tensor<fp16, [3, 16, 640]> x_7_cast_fp16 = mul(x = x_5_cast_fp16, y = var_637_to_fp16)[name = string("x_7_cast_fp16")];
            fp16 var_635_promoted_to_fp16 = const()[name = string("op_635_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> var_639_cast_fp16 = pow(x = x_7_cast_fp16, y = var_635_promoted_to_fp16)[name = string("op_639_cast_fp16")];
            tensor<int32, [1]> var_641_axes_0 = const()[name = string("op_641_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_641_keep_dims_0 = const()[name = string("op_641_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 1]> var_641_cast_fp16 = reduce_mean(axes = var_641_axes_0, keep_dims = var_641_keep_dims_0, x = var_639_cast_fp16)[name = string("op_641_cast_fp16")];
            fp16 var_642_to_fp16 = const()[name = string("op_642_to_fp16"), val = fp16(0x1.0c8p-10)];
            tensor<fp16, [3, 16, 1]> var_643_cast_fp16 = add(x = var_641_cast_fp16, y = var_642_to_fp16)[name = string("op_643_cast_fp16")];
            fp32 var_644_epsilon_0 = const()[name = string("op_644_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 1]> var_644_cast_fp16 = rsqrt(epsilon = var_644_epsilon_0, x = var_643_cast_fp16)[name = string("op_644_cast_fp16")];
            tensor<fp16, [3, 16, 640]> var_645_cast_fp16 = mul(x = x_7_cast_fp16, y = var_644_cast_fp16)[name = string("op_645_cast_fp16")];
            tensor<fp16, [640]> layers_0_n2_w1_to_fp16 = const()[name = string("layers_0_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23155904)))];
            tensor<fp16, [3, 16, 640]> input_11_cast_fp16 = mul(x = var_645_cast_fp16, y = layers_0_n2_w1_to_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_0_gate_up_weight_to_fp16 = const()[name = string("layers_0_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23157248)))];
            tensor<fp16, [5120]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29710912)))];
            tensor<fp16, [3, 16, 5120]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_0_gate_up_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [3]> input_13_begin_0 = const()[name = string("input_13_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_13_end_0 = const()[name = string("input_13_end_0"), val = tensor<int32, [3]>([3, 16, 2560])];
            tensor<bool, [3]> input_13_end_mask_0 = const()[name = string("input_13_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 2560]> input_13_cast_fp16 = slice_by_index(begin = input_13_begin_0, end = input_13_end_0, end_mask = input_13_end_mask_0, x = linear_2_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [3, 16, 2560]> var_655_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_655_cast_fp16")];
            tensor<int32, [3]> var_660_begin_0 = const()[name = string("op_660_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_660_end_0 = const()[name = string("op_660_end_0"), val = tensor<int32, [3]>([3, 16, 5120])];
            tensor<bool, [3]> var_660_end_mask_0 = const()[name = string("op_660_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 2560]> var_660_cast_fp16 = slice_by_index(begin = var_660_begin_0, end = var_660_end_0, end_mask = var_660_end_mask_0, x = linear_2_cast_fp16)[name = string("op_660_cast_fp16")];
            tensor<fp16, [3, 16, 2560]> input_15_cast_fp16 = mul(x = var_655_cast_fp16, y = var_660_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_0_down_weight_to_fp16 = const()[name = string("layers_0_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29721216)))];
            tensor<fp16, [3, 16, 640]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_down_weight_to_fp16, x = input_15_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [3, 16, 640]> x_9_cast_fp16 = add(x = x_5_cast_fp16, y = linear_3_cast_fp16)[name = string("x_9_cast_fp16")];
            fp16 var_673_to_fp16 = const()[name = string("op_673_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> x_11_cast_fp16 = mul(x = x_9_cast_fp16, y = var_673_to_fp16)[name = string("x_11_cast_fp16")];
            fp16 var_671_promoted_to_fp16 = const()[name = string("op_671_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> var_675_cast_fp16 = pow(x = x_11_cast_fp16, y = var_671_promoted_to_fp16)[name = string("op_675_cast_fp16")];
            tensor<int32, [1]> var_677_axes_0 = const()[name = string("op_677_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_677_keep_dims_0 = const()[name = string("op_677_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 1]> var_677_cast_fp16 = reduce_mean(axes = var_677_axes_0, keep_dims = var_677_keep_dims_0, x = var_675_cast_fp16)[name = string("op_677_cast_fp16")];
            fp16 var_678_to_fp16 = const()[name = string("op_678_to_fp16"), val = fp16(0x1.0cp-18)];
            tensor<fp16, [3, 16, 1]> var_679_cast_fp16 = add(x = var_677_cast_fp16, y = var_678_to_fp16)[name = string("op_679_cast_fp16")];
            fp32 var_680_epsilon_0 = const()[name = string("op_680_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 1]> var_680_cast_fp16 = rsqrt(epsilon = var_680_epsilon_0, x = var_679_cast_fp16)[name = string("op_680_cast_fp16")];
            tensor<fp16, [3, 16, 640]> var_681_cast_fp16 = mul(x = x_11_cast_fp16, y = var_680_cast_fp16)[name = string("op_681_cast_fp16")];
            tensor<fp16, [640]> layers_1_n1_w1_to_fp16 = const()[name = string("layers_1_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32998080)))];
            tensor<fp16, [3, 16, 640]> input_17_cast_fp16 = mul(x = var_681_cast_fp16, y = layers_1_n1_w1_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_1_qkvzba_weight_to_fp16 = const()[name = string("layers_1_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32999424)))];
            tensor<fp16, [3, 16, 1544]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_qkvzba_weight_to_fp16, x = input_17_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<int32, [3]> qkv_3_begin_0 = const()[name = string("qkv_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_3_end_0 = const()[name = string("qkv_3_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> qkv_3_end_mask_0 = const()[name = string("qkv_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 1024]> qkv_3_cast_fp16 = slice_by_index(begin = qkv_3_begin_0, end = qkv_3_end_0, end_mask = qkv_3_end_mask_0, x = linear_4_cast_fp16)[name = string("qkv_3_cast_fp16")];
            tensor<int32, [3]> z_3_begin_0 = const()[name = string("z_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_3_end_0 = const()[name = string("z_3_end_0"), val = tensor<int32, [3]>([3, 16, 1536])];
            tensor<bool, [3]> z_3_end_mask_0 = const()[name = string("z_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 512]> z_3_cast_fp16 = slice_by_index(begin = z_3_begin_0, end = z_3_end_0, end_mask = z_3_end_mask_0, x = linear_4_cast_fp16)[name = string("z_3_cast_fp16")];
            tensor<int32, [3]> b_3_begin_0 = const()[name = string("b_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_3_end_0 = const()[name = string("b_3_end_0"), val = tensor<int32, [3]>([3, 16, 1540])];
            tensor<bool, [3]> b_3_end_mask_0 = const()[name = string("b_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 4]> b_3_cast_fp16 = slice_by_index(begin = b_3_begin_0, end = b_3_end_0, end_mask = b_3_end_mask_0, x = linear_4_cast_fp16)[name = string("b_3_cast_fp16")];
            tensor<int32, [3]> a_3_begin_0 = const()[name = string("a_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_3_end_0 = const()[name = string("a_3_end_0"), val = tensor<int32, [3]>([3, 16, 1])];
            tensor<bool, [3]> a_3_end_mask_0 = const()[name = string("a_3_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 4]> a_3_cast_fp16 = slice_by_index(begin = a_3_begin_0, end = a_3_end_0, end_mask = a_3_end_mask_0, x = linear_4_cast_fp16)[name = string("a_3_cast_fp16")];
            int32 var_720 = const()[name = string("op_720"), val = int32(1)];
            bool xp_3_interleave_0 = const()[name = string("xp_3_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 19, 1024]> xp_3_cast_fp16 = concat(axis = var_720, interleave = xp_3_interleave_0, values = (zpad_to_fp16, qkv_3_cast_fp16))[name = string("xp_3_cast_fp16")];
            tensor<int32, [3]> var_731_begin_0 = const()[name = string("op_731_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_731_end_0 = const()[name = string("op_731_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_731_end_mask_0 = const()[name = string("op_731_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 1024]> var_731_cast_fp16 = slice_by_index(begin = var_731_begin_0, end = var_731_end_0, end_mask = var_731_end_mask_0, x = xp_3_cast_fp16)[name = string("op_731_cast_fp16")];
            tensor<fp16, [1024]> var_739_to_fp16 = const()[name = string("op_739_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34975808)))];
            tensor<fp16, [3, 16, 1024]> var_740_cast_fp16 = mul(x = var_731_cast_fp16, y = var_739_to_fp16)[name = string("op_740_cast_fp16")];
            tensor<int32, [3]> var_750_begin_0 = const()[name = string("op_750_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_750_end_0 = const()[name = string("op_750_end_0"), val = tensor<int32, [3]>([3, 18, 1024])];
            tensor<bool, [3]> var_750_end_mask_0 = const()[name = string("op_750_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_750_cast_fp16 = slice_by_index(begin = var_750_begin_0, end = var_750_end_0, end_mask = var_750_end_mask_0, x = xp_3_cast_fp16)[name = string("op_750_cast_fp16")];
            tensor<fp16, [1024]> var_758_to_fp16 = const()[name = string("op_758_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34977920)))];
            tensor<fp16, [3, 16, 1024]> var_759_cast_fp16 = mul(x = var_750_cast_fp16, y = var_758_to_fp16)[name = string("op_759_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> var_761_cast_fp16 = add(x = var_740_cast_fp16, y = var_759_cast_fp16)[name = string("op_761_cast_fp16")];
            tensor<int32, [3]> var_771_begin_0 = const()[name = string("op_771_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_771_end_0 = const()[name = string("op_771_end_0"), val = tensor<int32, [3]>([3, 17, 1024])];
            tensor<bool, [3]> var_771_end_mask_0 = const()[name = string("op_771_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_771_cast_fp16 = slice_by_index(begin = var_771_begin_0, end = var_771_end_0, end_mask = var_771_end_mask_0, x = xp_3_cast_fp16)[name = string("op_771_cast_fp16")];
            tensor<fp16, [1024]> var_779_to_fp16 = const()[name = string("op_779_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34980032)))];
            tensor<fp16, [3, 16, 1024]> var_780_cast_fp16 = mul(x = var_771_cast_fp16, y = var_779_to_fp16)[name = string("op_780_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> var_782_cast_fp16 = add(x = var_761_cast_fp16, y = var_780_cast_fp16)[name = string("op_782_cast_fp16")];
            tensor<int32, [3]> var_792_begin_0 = const()[name = string("op_792_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_792_end_0 = const()[name = string("op_792_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> var_792_end_mask_0 = const()[name = string("op_792_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_792_cast_fp16 = slice_by_index(begin = var_792_begin_0, end = var_792_end_0, end_mask = var_792_end_mask_0, x = xp_3_cast_fp16)[name = string("op_792_cast_fp16")];
            tensor<fp16, [1024]> var_800_to_fp16 = const()[name = string("op_800_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34982144)))];
            tensor<fp16, [3, 16, 1024]> var_801_cast_fp16 = mul(x = var_792_cast_fp16, y = var_800_to_fp16)[name = string("op_801_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> input_19_cast_fp16 = add(x = var_782_cast_fp16, y = var_801_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> c_3_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("c_3_cast_fp16")];
            tensor<int32, [3]> var_809_begin_0 = const()[name = string("op_809_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_809_end_0 = const()[name = string("op_809_end_0"), val = tensor<int32, [3]>([3, 16, 256])];
            tensor<bool, [3]> var_809_end_mask_0 = const()[name = string("op_809_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 256]> var_809_cast_fp16 = slice_by_index(begin = var_809_begin_0, end = var_809_end_0, end_mask = var_809_end_mask_0, x = c_3_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<int32, [4]> var_814 = const()[name = string("op_814"), val = tensor<int32, [4]>([3, 16, 2, 128])];
            tensor<fp16, [3, 16, 2, 128]> q_9_cast_fp16 = reshape(shape = var_814, x = var_809_cast_fp16)[name = string("q_9_cast_fp16")];
            tensor<int32, [3]> var_820_begin_0 = const()[name = string("op_820_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_820_end_0 = const()[name = string("op_820_end_0"), val = tensor<int32, [3]>([3, 16, 512])];
            tensor<bool, [3]> var_820_end_mask_0 = const()[name = string("op_820_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 256]> var_820_cast_fp16 = slice_by_index(begin = var_820_begin_0, end = var_820_end_0, end_mask = var_820_end_mask_0, x = c_3_cast_fp16)[name = string("op_820_cast_fp16")];
            tensor<int32, [4]> var_825 = const()[name = string("op_825"), val = tensor<int32, [4]>([3, 16, 2, 128])];
            tensor<fp16, [3, 16, 2, 128]> k_9_cast_fp16 = reshape(shape = var_825, x = var_820_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<int32, [3]> var_831_begin_0 = const()[name = string("op_831_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_831_end_0 = const()[name = string("op_831_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> var_831_end_mask_0 = const()[name = string("op_831_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 512]> var_831_cast_fp16 = slice_by_index(begin = var_831_begin_0, end = var_831_end_0, end_mask = var_831_end_mask_0, x = c_3_cast_fp16)[name = string("op_831_cast_fp16")];
            tensor<int32, [4]> var_836 = const()[name = string("op_836"), val = tensor<int32, [4]>([3, 16, 4, 128])];
            tensor<fp16, [3, 16, 4, 128]> var_837_cast_fp16 = reshape(shape = var_836, x = var_831_cast_fp16)[name = string("op_837_cast_fp16")];
            tensor<int32, [4]> var_842 = const()[name = string("op_842"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_844_to_fp16 = const()[name = string("op_844_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 16, 2, 128]> q_11_cast_fp16 = mul(x = q_9_cast_fp16, y = var_844_to_fp16)[name = string("q_11_cast_fp16")];
            fp16 var_846_to_fp16 = const()[name = string("op_846_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 16, 2, 128]> k_11_cast_fp16 = mul(x = k_9_cast_fp16, y = var_846_to_fp16)[name = string("k_11_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_848_cast_fp16 = mul(x = q_11_cast_fp16, y = q_11_cast_fp16)[name = string("op_848_cast_fp16")];
            tensor<int32, [1]> var_853_axes_0 = const()[name = string("op_853_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_853_keep_dims_0 = const()[name = string("op_853_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 2, 1]> var_853_cast_fp16 = reduce_mean(axes = var_853_axes_0, keep_dims = var_853_keep_dims_0, x = var_848_cast_fp16)[name = string("op_853_cast_fp16")];
            fp16 var_854_promoted_to_fp16 = const()[name = string("op_854_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 16, 2, 1]> var_855_cast_fp16 = mul(x = var_853_cast_fp16, y = var_854_promoted_to_fp16)[name = string("op_855_cast_fp16")];
            fp16 var_857_to_fp16 = const()[name = string("op_857_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 16, 2, 1]> var_858_cast_fp16 = add(x = var_855_cast_fp16, y = var_857_to_fp16)[name = string("op_858_cast_fp16")];
            fp32 var_859_epsilon_0 = const()[name = string("op_859_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 2, 1]> var_859_cast_fp16 = rsqrt(epsilon = var_859_epsilon_0, x = var_858_cast_fp16)[name = string("op_859_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_860_cast_fp16 = mul(x = q_11_cast_fp16, y = var_859_cast_fp16)[name = string("op_860_cast_fp16")];
            fp16 var_861_to_fp16 = const()[name = string("op_861_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 16, 2, 128]> q_13_cast_fp16 = mul(x = var_860_cast_fp16, y = var_861_to_fp16)[name = string("q_13_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_863_cast_fp16 = mul(x = k_11_cast_fp16, y = k_11_cast_fp16)[name = string("op_863_cast_fp16")];
            tensor<int32, [1]> var_868_axes_0 = const()[name = string("op_868_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_868_keep_dims_0 = const()[name = string("op_868_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 2, 1]> var_868_cast_fp16 = reduce_mean(axes = var_868_axes_0, keep_dims = var_868_keep_dims_0, x = var_863_cast_fp16)[name = string("op_868_cast_fp16")];
            fp16 var_869_promoted_to_fp16 = const()[name = string("op_869_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 16, 2, 1]> var_870_cast_fp16 = mul(x = var_868_cast_fp16, y = var_869_promoted_to_fp16)[name = string("op_870_cast_fp16")];
            fp16 var_872_to_fp16 = const()[name = string("op_872_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 16, 2, 1]> var_873_cast_fp16 = add(x = var_870_cast_fp16, y = var_872_to_fp16)[name = string("op_873_cast_fp16")];
            fp32 var_874_epsilon_0 = const()[name = string("op_874_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 2, 1]> var_874_cast_fp16 = rsqrt(epsilon = var_874_epsilon_0, x = var_873_cast_fp16)[name = string("op_874_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> k_13_cast_fp16 = mul(x = k_11_cast_fp16, y = var_874_cast_fp16)[name = string("k_13_cast_fp16")];
            tensor<int32, [4]> var_880 = const()[name = string("op_880"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_893_axes_0 = const()[name = string("op_893_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 16, 128]> var_881_cast_fp16 = transpose(perm = var_880, x = q_13_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [3, 2, 1, 16, 128]> var_893_cast_fp16 = expand_dims(axes = var_893_axes_0, x = var_881_cast_fp16)[name = string("op_893_cast_fp16")];
            tensor<int32, [5]> var_901_reps_0 = const()[name = string("op_901_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 16, 128]> var_901_cast_fp16 = tile(reps = var_901_reps_0, x = var_893_cast_fp16)[name = string("op_901_cast_fp16")];
            tensor<int32, [4]> var_906 = const()[name = string("op_906"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<fp16, [3, 4, 16, 128]> q_15_cast_fp16 = reshape(shape = var_906, x = var_901_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<int32, [4]> var_912 = const()[name = string("op_912"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_925_axes_0 = const()[name = string("op_925_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 16, 128]> var_913_cast_fp16 = transpose(perm = var_912, x = k_13_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [3, 2, 1, 16, 128]> var_925_cast_fp16 = expand_dims(axes = var_925_axes_0, x = var_913_cast_fp16)[name = string("op_925_cast_fp16")];
            tensor<int32, [5]> var_933_reps_0 = const()[name = string("op_933_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 16, 128]> var_933_cast_fp16 = tile(reps = var_933_reps_0, x = var_925_cast_fp16)[name = string("op_933_cast_fp16")];
            tensor<int32, [4]> var_938 = const()[name = string("op_938"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<fp16, [3, 4, 16, 128]> k_15_cast_fp16 = reshape(shape = var_938, x = var_933_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_940_cast_fp16 = sigmoid(x = b_3_cast_fp16)[name = string("op_940_cast_fp16")];
            tensor<int32, [3]> var_944 = const()[name = string("op_944"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_1_dt_bias_to_fp16 = const()[name = string("layers_1_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 16, 4]> input_21_cast_fp16 = add(x = a_3_cast_fp16, y = layers_1_dt_bias_to_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_948_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("op_948_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_949_cast_fp16 = abs(x = input_21_cast_fp16)[name = string("op_949_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 16, 4]> var_950_cast_fp16 = mul(x = var_949_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_950_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_951_cast_fp16 = exp(x = var_950_cast_fp16)[name = string("op_951_cast_fp16")];
            fp16 var_953_to_fp16 = const()[name = string("op_953_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 16, 4]> var_954_cast_fp16 = add(x = var_951_cast_fp16, y = var_953_to_fp16)[name = string("op_954_cast_fp16")];
            fp32 var_955_epsilon_0 = const()[name = string("op_955_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 16, 4]> var_955_cast_fp16 = log(epsilon = var_955_epsilon_0, x = var_954_cast_fp16)[name = string("op_955_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_957_cast_fp16 = add(x = var_948_cast_fp16, y = var_955_cast_fp16)[name = string("op_957_cast_fp16")];
            tensor<fp16, [4]> layers_1_negA_to_fp16 = const()[name = string("layers_1_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.ff4p+2, -0x1.8cp+3, -0x1.c0cp+3, -0x1.104p+3])];
            tensor<fp16, [3, 16, 4]> var_958_cast_fp16 = mul(x = layers_1_negA_to_fp16, y = var_957_cast_fp16)[name = string("op_958_cast_fp16")];
            tensor<int32, [3]> var_962 = const()[name = string("op_962"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_965_axes_0 = const()[name = string("op_965_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16]> beta_3_cast_fp16 = transpose(perm = var_944, x = var_940_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [3, 4, 16, 1]> var_965_cast_fp16 = expand_dims(axes = var_965_axes_0, x = beta_3_cast_fp16)[name = string("op_965_cast_fp16")];
            fp16 var_966_to_fp16 = const()[name = string("op_966_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 16, 1]> var_967_cast_fp16 = mul(x = var_965_cast_fp16, y = var_966_to_fp16)[name = string("op_967_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> v_3_cast_fp16 = transpose(perm = var_842, x = var_837_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [3, 4, 16, 128]> vb_3_cast_fp16 = mul(x = v_3_cast_fp16, y = var_967_cast_fp16)[name = string("vb_3_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> kb_3_cast_fp16 = mul(x = k_15_cast_fp16, y = var_965_cast_fp16)[name = string("kb_3_cast_fp16")];
            tensor<int32, [2]> var_974 = const()[name = string("op_974"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [3, 4, 16]> gdec_3_cast_fp16 = transpose(perm = var_962, x = var_958_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [12, 16]> var_975_cast_fp16 = reshape(shape = var_974, x = gdec_3_cast_fp16)[name = string("op_975_cast_fp16")];
            tensor<fp16, [12, 256]> var_976_cast_fp16 = linear(bias = var_554_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_975_cast_fp16)[name = string("op_976_cast_fp16")];
            tensor<int32, [4]> var_981 = const()[name = string("op_981"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_3_cast_fp16 = reshape(shape = var_981, x = var_976_cast_fp16)[name = string("P_3_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_983_cast_fp16 = exp(x = P_3_cast_fp16)[name = string("op_983_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_3_cast_fp16 = mul(x = var_983_cast_fp16, y = tril_incl_to_fp16)[name = string("D_3_cast_fp16")];
            bool kk_3_transpose_x_1 = const()[name = string("kk_3_transpose_x_1"), val = bool(false)];
            bool kk_3_transpose_y_1 = const()[name = string("kk_3_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_3_cast_fp16 = matmul(transpose_x = kk_3_transpose_x_1, transpose_y = kk_3_transpose_y_1, x = kb_3_cast_fp16, y = k_15_cast_fp16)[name = string("kk_3_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_989_cast_fp16 = mul(x = kk_3_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_989_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_990_cast_fp16 = mul(x = var_989_cast_fp16, y = D_3_cast_fp16)[name = string("op_990_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_3_cast_fp16 = mul(x = var_990_cast_fp16, y = tril_strict_to_fp16)[name = string("N_3_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_992_cast_fp16 = mul(x = D_3_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_992_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_3_cast_fp16 = mul(x = kk_3_cast_fp16, y = var_992_cast_fp16)[name = string("Na_3_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_9_cast_fp16 = add(x = eye_to_fp16, y = N_3_cast_fp16)[name = string("Minv_9_cast_fp16")];
            bool Np_7_transpose_x_0 = const()[name = string("Np_7_transpose_x_0"), val = bool(false)];
            bool Np_7_transpose_y_0 = const()[name = string("Np_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_7_cast_fp16 = matmul(transpose_x = Np_7_transpose_x_0, transpose_y = Np_7_transpose_y_0, x = N_3_cast_fp16, y = Na_3_cast_fp16)[name = string("Np_7_cast_fp16")];
            bool Npa_5_transpose_x_0 = const()[name = string("Npa_5_transpose_x_0"), val = bool(false)];
            bool Npa_5_transpose_y_0 = const()[name = string("Npa_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_5_cast_fp16 = matmul(transpose_x = Npa_5_transpose_x_0, transpose_y = Npa_5_transpose_y_0, x = Na_3_cast_fp16, y = N_3_cast_fp16)[name = string("Npa_5_cast_fp16")];
            bool var_998_transpose_x_0 = const()[name = string("op_998_transpose_x_0"), val = bool(false)];
            bool var_998_transpose_y_0 = const()[name = string("op_998_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_998_cast_fp16 = matmul(transpose_x = var_998_transpose_x_0, transpose_y = var_998_transpose_y_0, x = Minv_9_cast_fp16, y = Np_7_cast_fp16)[name = string("op_998_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_11_cast_fp16 = add(x = Minv_9_cast_fp16, y = var_998_cast_fp16)[name = string("Minv_11_cast_fp16")];
            bool Np_9_transpose_x_0 = const()[name = string("Np_9_transpose_x_0"), val = bool(false)];
            bool Np_9_transpose_y_0 = const()[name = string("Np_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_9_cast_fp16 = matmul(transpose_x = Np_9_transpose_x_0, transpose_y = Np_9_transpose_y_0, x = Np_7_cast_fp16, y = Npa_5_cast_fp16)[name = string("Np_9_cast_fp16")];
            bool Npa_7_transpose_x_0 = const()[name = string("Npa_7_transpose_x_0"), val = bool(false)];
            bool Npa_7_transpose_y_0 = const()[name = string("Npa_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_7_cast_fp16 = matmul(transpose_x = Npa_7_transpose_x_0, transpose_y = Npa_7_transpose_y_0, x = Npa_5_cast_fp16, y = Np_7_cast_fp16)[name = string("Npa_7_cast_fp16")];
            bool var_1003_transpose_x_0 = const()[name = string("op_1003_transpose_x_0"), val = bool(false)];
            bool var_1003_transpose_y_0 = const()[name = string("op_1003_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1003_cast_fp16 = matmul(transpose_x = var_1003_transpose_x_0, transpose_y = var_1003_transpose_y_0, x = Minv_11_cast_fp16, y = Np_9_cast_fp16)[name = string("op_1003_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_13_cast_fp16 = add(x = Minv_11_cast_fp16, y = var_1003_cast_fp16)[name = string("Minv_13_cast_fp16")];
            bool Np_11_transpose_x_0 = const()[name = string("Np_11_transpose_x_0"), val = bool(false)];
            bool Np_11_transpose_y_0 = const()[name = string("Np_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_11_cast_fp16 = matmul(transpose_x = Np_11_transpose_x_0, transpose_y = Np_11_transpose_y_0, x = Np_9_cast_fp16, y = Npa_7_cast_fp16)[name = string("Np_11_cast_fp16")];
            bool var_1007_transpose_x_0 = const()[name = string("op_1007_transpose_x_0"), val = bool(false)];
            bool var_1007_transpose_y_0 = const()[name = string("op_1007_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1007_cast_fp16 = matmul(transpose_x = var_1007_transpose_x_0, transpose_y = var_1007_transpose_y_0, x = Minv_13_cast_fp16, y = Np_11_cast_fp16)[name = string("op_1007_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_15_cast_fp16 = add(x = Minv_13_cast_fp16, y = var_1007_cast_fp16)[name = string("Minv_15_cast_fp16")];
            bool u_3_transpose_x_0 = const()[name = string("u_3_transpose_x_0"), val = bool(false)];
            bool u_3_transpose_y_0 = const()[name = string("u_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_3_cast_fp16 = matmul(transpose_x = u_3_transpose_x_0, transpose_y = u_3_transpose_y_0, x = Minv_15_cast_fp16, y = vb_3_cast_fp16)[name = string("u_3_cast_fp16")];
            bool var_1014_transpose_x_1 = const()[name = string("op_1014_transpose_x_1"), val = bool(false)];
            bool var_1014_transpose_y_1 = const()[name = string("op_1014_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_1014_cast_fp16 = matmul(transpose_x = var_1014_transpose_x_1, transpose_y = var_1014_transpose_y_1, x = q_15_cast_fp16, y = k_15_cast_fp16)[name = string("op_1014_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_3_cast_fp16 = mul(x = var_1014_cast_fp16, y = D_3_cast_fp16)[name = string("A_3_cast_fp16")];
            bool o_7_transpose_x_0 = const()[name = string("o_7_transpose_x_0"), val = bool(false)];
            bool o_7_transpose_y_0 = const()[name = string("o_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> o_7_cast_fp16 = matmul(transpose_x = o_7_transpose_x_0, transpose_y = o_7_transpose_y_0, x = A_3_cast_fp16, y = u_3_cast_fp16)[name = string("o_7_cast_fp16")];
            tensor<int32, [4]> var_1021 = const()[name = string("op_1021"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1027 = const()[name = string("op_1027"), val = tensor<int32, [4]>([3, 16, 4, 128])];
            tensor<fp16, [3, 16, 4, 128]> input_23_cast_fp16 = reshape(shape = var_1027, x = z_3_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> o_9_cast_fp16 = transpose(perm = var_1021, x = o_7_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [3, 16, 4, 128]> var_1029_cast_fp16 = mul(x = o_9_cast_fp16, y = o_9_cast_fp16)[name = string("op_1029_cast_fp16")];
            tensor<int32, [1]> var_1034_axes_0 = const()[name = string("op_1034_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1034_keep_dims_0 = const()[name = string("op_1034_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 4, 1]> var_1034_cast_fp16 = reduce_mean(axes = var_1034_axes_0, keep_dims = var_1034_keep_dims_0, x = var_1029_cast_fp16)[name = string("op_1034_cast_fp16")];
            fp16 var_1036_to_fp16 = const()[name = string("op_1036_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 16, 4, 1]> var_1037_cast_fp16 = add(x = var_1034_cast_fp16, y = var_1036_to_fp16)[name = string("op_1037_cast_fp16")];
            fp32 var_1038_epsilon_0 = const()[name = string("op_1038_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 4, 1]> var_1038_cast_fp16 = rsqrt(epsilon = var_1038_epsilon_0, x = var_1037_cast_fp16)[name = string("op_1038_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> var_1039_cast_fp16 = mul(x = o_9_cast_fp16, y = var_1038_cast_fp16)[name = string("op_1039_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> var_1041_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_1041_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> o_11_cast_fp16 = mul(x = var_1039_cast_fp16, y = var_1041_cast_fp16)[name = string("o_11_cast_fp16")];
            tensor<int32, [3]> var_1046 = const()[name = string("op_1046"), val = tensor<int32, [3]>([3, 16, 512])];
            tensor<fp16, [3, 16, 512]> input_25_cast_fp16 = reshape(shape = var_1046, x = o_11_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [640, 512]> layers_1_out_weight_to_fp16 = const()[name = string("layers_1_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34984256)))];
            tensor<fp16, [3, 16, 640]> linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_out_weight_to_fp16, x = input_25_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [3, 16, 640]> x_13_cast_fp16 = add(x = x_9_cast_fp16, y = linear_5_cast_fp16)[name = string("x_13_cast_fp16")];
            fp16 var_1059_to_fp16 = const()[name = string("op_1059_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> x_15_cast_fp16 = mul(x = x_13_cast_fp16, y = var_1059_to_fp16)[name = string("x_15_cast_fp16")];
            fp16 var_1057_promoted_to_fp16 = const()[name = string("op_1057_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> var_1061_cast_fp16 = pow(x = x_15_cast_fp16, y = var_1057_promoted_to_fp16)[name = string("op_1061_cast_fp16")];
            tensor<int32, [1]> var_1063_axes_0 = const()[name = string("op_1063_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1063_keep_dims_0 = const()[name = string("op_1063_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 1]> var_1063_cast_fp16 = reduce_mean(axes = var_1063_axes_0, keep_dims = var_1063_keep_dims_0, x = var_1061_cast_fp16)[name = string("op_1063_cast_fp16")];
            fp16 var_1064_to_fp16 = const()[name = string("op_1064_to_fp16"), val = fp16(0x1.0cp-18)];
            tensor<fp16, [3, 16, 1]> var_1065_cast_fp16 = add(x = var_1063_cast_fp16, y = var_1064_to_fp16)[name = string("op_1065_cast_fp16")];
            fp32 var_1066_epsilon_0 = const()[name = string("op_1066_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 1]> var_1066_cast_fp16 = rsqrt(epsilon = var_1066_epsilon_0, x = var_1065_cast_fp16)[name = string("op_1066_cast_fp16")];
            tensor<fp16, [3, 16, 640]> var_1067_cast_fp16 = mul(x = x_15_cast_fp16, y = var_1066_cast_fp16)[name = string("op_1067_cast_fp16")];
            tensor<fp16, [640]> layers_1_n2_w1_to_fp16 = const()[name = string("layers_1_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35639680)))];
            tensor<fp16, [3, 16, 640]> input_27_cast_fp16 = mul(x = var_1067_cast_fp16, y = layers_1_n2_w1_to_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_1_gate_up_weight_to_fp16 = const()[name = string("layers_1_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35641024)))];
            tensor<fp16, [3, 16, 5120]> linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_1_gate_up_weight_to_fp16, x = input_27_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<int32, [3]> input_29_begin_0 = const()[name = string("input_29_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_29_end_0 = const()[name = string("input_29_end_0"), val = tensor<int32, [3]>([3, 16, 2560])];
            tensor<bool, [3]> input_29_end_mask_0 = const()[name = string("input_29_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 2560]> input_29_cast_fp16 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = linear_6_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [3, 16, 2560]> var_1077_cast_fp16 = silu(x = input_29_cast_fp16)[name = string("op_1077_cast_fp16")];
            tensor<int32, [3]> var_1082_begin_0 = const()[name = string("op_1082_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_1082_end_0 = const()[name = string("op_1082_end_0"), val = tensor<int32, [3]>([3, 16, 5120])];
            tensor<bool, [3]> var_1082_end_mask_0 = const()[name = string("op_1082_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 2560]> var_1082_cast_fp16 = slice_by_index(begin = var_1082_begin_0, end = var_1082_end_0, end_mask = var_1082_end_mask_0, x = linear_6_cast_fp16)[name = string("op_1082_cast_fp16")];
            tensor<fp16, [3, 16, 2560]> input_31_cast_fp16 = mul(x = var_1077_cast_fp16, y = var_1082_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_1_down_weight_to_fp16 = const()[name = string("layers_1_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42194688)))];
            tensor<fp16, [3, 16, 640]> linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_down_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [3, 16, 640]> x_17_cast_fp16 = add(x = x_13_cast_fp16, y = linear_7_cast_fp16)[name = string("x_17_cast_fp16")];
            fp16 var_1093_promoted_to_fp16 = const()[name = string("op_1093_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> var_1097_cast_fp16 = pow(x = x_17_cast_fp16, y = var_1093_promoted_to_fp16)[name = string("op_1097_cast_fp16")];
            tensor<int32, [1]> var_1099_axes_0 = const()[name = string("op_1099_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1099_keep_dims_0 = const()[name = string("op_1099_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 1]> var_1099_cast_fp16 = reduce_mean(axes = var_1099_axes_0, keep_dims = var_1099_keep_dims_0, x = var_1097_cast_fp16)[name = string("op_1099_cast_fp16")];
            fp16 var_1100_to_fp16 = const()[name = string("op_1100_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 16, 1]> var_1101_cast_fp16 = add(x = var_1099_cast_fp16, y = var_1100_to_fp16)[name = string("op_1101_cast_fp16")];
            fp32 var_1102_epsilon_0 = const()[name = string("op_1102_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 1]> var_1102_cast_fp16 = rsqrt(epsilon = var_1102_epsilon_0, x = var_1101_cast_fp16)[name = string("op_1102_cast_fp16")];
            tensor<fp16, [3, 16, 640]> var_1103_cast_fp16 = mul(x = x_17_cast_fp16, y = var_1102_cast_fp16)[name = string("op_1103_cast_fp16")];
            tensor<fp16, [640]> layers_2_n1_w1_to_fp16 = const()[name = string("layers_2_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45471552)))];
            tensor<fp16, [3, 16, 640]> input_33_cast_fp16 = mul(x = var_1103_cast_fp16, y = layers_2_n1_w1_to_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_2_qkvzba_weight_to_fp16 = const()[name = string("layers_2_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45472896)))];
            tensor<fp16, [3, 16, 1544]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_qkvzba_weight_to_fp16, x = input_33_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [3]> qkv_5_begin_0 = const()[name = string("qkv_5_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_5_end_0 = const()[name = string("qkv_5_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> qkv_5_end_mask_0 = const()[name = string("qkv_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 1024]> qkv_5_cast_fp16 = slice_by_index(begin = qkv_5_begin_0, end = qkv_5_end_0, end_mask = qkv_5_end_mask_0, x = linear_8_cast_fp16)[name = string("qkv_5_cast_fp16")];
            tensor<int32, [3]> z_5_begin_0 = const()[name = string("z_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_5_end_0 = const()[name = string("z_5_end_0"), val = tensor<int32, [3]>([3, 16, 1536])];
            tensor<bool, [3]> z_5_end_mask_0 = const()[name = string("z_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 512]> z_5_cast_fp16 = slice_by_index(begin = z_5_begin_0, end = z_5_end_0, end_mask = z_5_end_mask_0, x = linear_8_cast_fp16)[name = string("z_5_cast_fp16")];
            tensor<int32, [3]> b_5_begin_0 = const()[name = string("b_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_5_end_0 = const()[name = string("b_5_end_0"), val = tensor<int32, [3]>([3, 16, 1540])];
            tensor<bool, [3]> b_5_end_mask_0 = const()[name = string("b_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 4]> b_5_cast_fp16 = slice_by_index(begin = b_5_begin_0, end = b_5_end_0, end_mask = b_5_end_mask_0, x = linear_8_cast_fp16)[name = string("b_5_cast_fp16")];
            tensor<int32, [3]> a_5_begin_0 = const()[name = string("a_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_5_end_0 = const()[name = string("a_5_end_0"), val = tensor<int32, [3]>([3, 16, 1])];
            tensor<bool, [3]> a_5_end_mask_0 = const()[name = string("a_5_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 4]> a_5_cast_fp16 = slice_by_index(begin = a_5_begin_0, end = a_5_end_0, end_mask = a_5_end_mask_0, x = linear_8_cast_fp16)[name = string("a_5_cast_fp16")];
            int32 var_1142 = const()[name = string("op_1142"), val = int32(1)];
            bool xp_5_interleave_0 = const()[name = string("xp_5_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 19, 1024]> xp_5_cast_fp16 = concat(axis = var_1142, interleave = xp_5_interleave_0, values = (zpad_to_fp16, qkv_5_cast_fp16))[name = string("xp_5_cast_fp16")];
            tensor<int32, [3]> var_1153_begin_0 = const()[name = string("op_1153_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_1153_end_0 = const()[name = string("op_1153_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_1153_end_mask_0 = const()[name = string("op_1153_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 1024]> var_1153_cast_fp16 = slice_by_index(begin = var_1153_begin_0, end = var_1153_end_0, end_mask = var_1153_end_mask_0, x = xp_5_cast_fp16)[name = string("op_1153_cast_fp16")];
            tensor<fp16, [1024]> var_1161_to_fp16 = const()[name = string("op_1161_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47449280)))];
            tensor<fp16, [3, 16, 1024]> var_1162_cast_fp16 = mul(x = var_1153_cast_fp16, y = var_1161_to_fp16)[name = string("op_1162_cast_fp16")];
            tensor<int32, [3]> var_1172_begin_0 = const()[name = string("op_1172_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_1172_end_0 = const()[name = string("op_1172_end_0"), val = tensor<int32, [3]>([3, 18, 1024])];
            tensor<bool, [3]> var_1172_end_mask_0 = const()[name = string("op_1172_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_1172_cast_fp16 = slice_by_index(begin = var_1172_begin_0, end = var_1172_end_0, end_mask = var_1172_end_mask_0, x = xp_5_cast_fp16)[name = string("op_1172_cast_fp16")];
            tensor<fp16, [1024]> var_1180_to_fp16 = const()[name = string("op_1180_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47451392)))];
            tensor<fp16, [3, 16, 1024]> var_1181_cast_fp16 = mul(x = var_1172_cast_fp16, y = var_1180_to_fp16)[name = string("op_1181_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> var_1183_cast_fp16 = add(x = var_1162_cast_fp16, y = var_1181_cast_fp16)[name = string("op_1183_cast_fp16")];
            tensor<int32, [3]> var_1193_begin_0 = const()[name = string("op_1193_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1193_end_0 = const()[name = string("op_1193_end_0"), val = tensor<int32, [3]>([3, 17, 1024])];
            tensor<bool, [3]> var_1193_end_mask_0 = const()[name = string("op_1193_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_1193_cast_fp16 = slice_by_index(begin = var_1193_begin_0, end = var_1193_end_0, end_mask = var_1193_end_mask_0, x = xp_5_cast_fp16)[name = string("op_1193_cast_fp16")];
            tensor<fp16, [1024]> var_1201_to_fp16 = const()[name = string("op_1201_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47453504)))];
            tensor<fp16, [3, 16, 1024]> var_1202_cast_fp16 = mul(x = var_1193_cast_fp16, y = var_1201_to_fp16)[name = string("op_1202_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> var_1204_cast_fp16 = add(x = var_1183_cast_fp16, y = var_1202_cast_fp16)[name = string("op_1204_cast_fp16")];
            tensor<int32, [3]> var_1214_begin_0 = const()[name = string("op_1214_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1214_end_0 = const()[name = string("op_1214_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> var_1214_end_mask_0 = const()[name = string("op_1214_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_1214_cast_fp16 = slice_by_index(begin = var_1214_begin_0, end = var_1214_end_0, end_mask = var_1214_end_mask_0, x = xp_5_cast_fp16)[name = string("op_1214_cast_fp16")];
            tensor<fp16, [1024]> var_1222_to_fp16 = const()[name = string("op_1222_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47455616)))];
            tensor<fp16, [3, 16, 1024]> var_1223_cast_fp16 = mul(x = var_1214_cast_fp16, y = var_1222_to_fp16)[name = string("op_1223_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> input_35_cast_fp16 = add(x = var_1204_cast_fp16, y = var_1223_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> c_5_cast_fp16 = silu(x = input_35_cast_fp16)[name = string("c_5_cast_fp16")];
            tensor<int32, [3]> var_1231_begin_0 = const()[name = string("op_1231_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1231_end_0 = const()[name = string("op_1231_end_0"), val = tensor<int32, [3]>([3, 16, 256])];
            tensor<bool, [3]> var_1231_end_mask_0 = const()[name = string("op_1231_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 256]> var_1231_cast_fp16 = slice_by_index(begin = var_1231_begin_0, end = var_1231_end_0, end_mask = var_1231_end_mask_0, x = c_5_cast_fp16)[name = string("op_1231_cast_fp16")];
            tensor<int32, [4]> var_1236 = const()[name = string("op_1236"), val = tensor<int32, [4]>([3, 16, 2, 128])];
            tensor<fp16, [3, 16, 2, 128]> q_17_cast_fp16 = reshape(shape = var_1236, x = var_1231_cast_fp16)[name = string("q_17_cast_fp16")];
            tensor<int32, [3]> var_1242_begin_0 = const()[name = string("op_1242_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_1242_end_0 = const()[name = string("op_1242_end_0"), val = tensor<int32, [3]>([3, 16, 512])];
            tensor<bool, [3]> var_1242_end_mask_0 = const()[name = string("op_1242_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 256]> var_1242_cast_fp16 = slice_by_index(begin = var_1242_begin_0, end = var_1242_end_0, end_mask = var_1242_end_mask_0, x = c_5_cast_fp16)[name = string("op_1242_cast_fp16")];
            tensor<int32, [4]> var_1247 = const()[name = string("op_1247"), val = tensor<int32, [4]>([3, 16, 2, 128])];
            tensor<fp16, [3, 16, 2, 128]> k_17_cast_fp16 = reshape(shape = var_1247, x = var_1242_cast_fp16)[name = string("k_17_cast_fp16")];
            tensor<int32, [3]> var_1253_begin_0 = const()[name = string("op_1253_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_1253_end_0 = const()[name = string("op_1253_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> var_1253_end_mask_0 = const()[name = string("op_1253_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 512]> var_1253_cast_fp16 = slice_by_index(begin = var_1253_begin_0, end = var_1253_end_0, end_mask = var_1253_end_mask_0, x = c_5_cast_fp16)[name = string("op_1253_cast_fp16")];
            tensor<int32, [4]> var_1258 = const()[name = string("op_1258"), val = tensor<int32, [4]>([3, 16, 4, 128])];
            tensor<fp16, [3, 16, 4, 128]> var_1259_cast_fp16 = reshape(shape = var_1258, x = var_1253_cast_fp16)[name = string("op_1259_cast_fp16")];
            tensor<int32, [4]> var_1264 = const()[name = string("op_1264"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1266_to_fp16 = const()[name = string("op_1266_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 16, 2, 128]> q_19_cast_fp16 = mul(x = q_17_cast_fp16, y = var_1266_to_fp16)[name = string("q_19_cast_fp16")];
            fp16 var_1268_to_fp16 = const()[name = string("op_1268_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 16, 2, 128]> k_19_cast_fp16 = mul(x = k_17_cast_fp16, y = var_1268_to_fp16)[name = string("k_19_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_1270_cast_fp16 = mul(x = q_19_cast_fp16, y = q_19_cast_fp16)[name = string("op_1270_cast_fp16")];
            tensor<int32, [1]> var_1275_axes_0 = const()[name = string("op_1275_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1275_keep_dims_0 = const()[name = string("op_1275_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 2, 1]> var_1275_cast_fp16 = reduce_mean(axes = var_1275_axes_0, keep_dims = var_1275_keep_dims_0, x = var_1270_cast_fp16)[name = string("op_1275_cast_fp16")];
            fp16 var_1276_promoted_to_fp16 = const()[name = string("op_1276_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 16, 2, 1]> var_1277_cast_fp16 = mul(x = var_1275_cast_fp16, y = var_1276_promoted_to_fp16)[name = string("op_1277_cast_fp16")];
            fp16 var_1279_to_fp16 = const()[name = string("op_1279_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 16, 2, 1]> var_1280_cast_fp16 = add(x = var_1277_cast_fp16, y = var_1279_to_fp16)[name = string("op_1280_cast_fp16")];
            fp32 var_1281_epsilon_0 = const()[name = string("op_1281_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 2, 1]> var_1281_cast_fp16 = rsqrt(epsilon = var_1281_epsilon_0, x = var_1280_cast_fp16)[name = string("op_1281_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_1282_cast_fp16 = mul(x = q_19_cast_fp16, y = var_1281_cast_fp16)[name = string("op_1282_cast_fp16")];
            fp16 var_1283_to_fp16 = const()[name = string("op_1283_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 16, 2, 128]> q_21_cast_fp16 = mul(x = var_1282_cast_fp16, y = var_1283_to_fp16)[name = string("q_21_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_1285_cast_fp16 = mul(x = k_19_cast_fp16, y = k_19_cast_fp16)[name = string("op_1285_cast_fp16")];
            tensor<int32, [1]> var_1290_axes_0 = const()[name = string("op_1290_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1290_keep_dims_0 = const()[name = string("op_1290_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 2, 1]> var_1290_cast_fp16 = reduce_mean(axes = var_1290_axes_0, keep_dims = var_1290_keep_dims_0, x = var_1285_cast_fp16)[name = string("op_1290_cast_fp16")];
            fp16 var_1291_promoted_to_fp16 = const()[name = string("op_1291_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 16, 2, 1]> var_1292_cast_fp16 = mul(x = var_1290_cast_fp16, y = var_1291_promoted_to_fp16)[name = string("op_1292_cast_fp16")];
            fp16 var_1294_to_fp16 = const()[name = string("op_1294_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 16, 2, 1]> var_1295_cast_fp16 = add(x = var_1292_cast_fp16, y = var_1294_to_fp16)[name = string("op_1295_cast_fp16")];
            fp32 var_1296_epsilon_0 = const()[name = string("op_1296_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 2, 1]> var_1296_cast_fp16 = rsqrt(epsilon = var_1296_epsilon_0, x = var_1295_cast_fp16)[name = string("op_1296_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> k_21_cast_fp16 = mul(x = k_19_cast_fp16, y = var_1296_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<int32, [4]> var_1302 = const()[name = string("op_1302"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1315_axes_0 = const()[name = string("op_1315_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 16, 128]> var_1303_cast_fp16 = transpose(perm = var_1302, x = q_21_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [3, 2, 1, 16, 128]> var_1315_cast_fp16 = expand_dims(axes = var_1315_axes_0, x = var_1303_cast_fp16)[name = string("op_1315_cast_fp16")];
            tensor<int32, [5]> var_1323_reps_0 = const()[name = string("op_1323_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 16, 128]> var_1323_cast_fp16 = tile(reps = var_1323_reps_0, x = var_1315_cast_fp16)[name = string("op_1323_cast_fp16")];
            tensor<int32, [4]> var_1328 = const()[name = string("op_1328"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<fp16, [3, 4, 16, 128]> q_23_cast_fp16 = reshape(shape = var_1328, x = var_1323_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<int32, [4]> var_1334 = const()[name = string("op_1334"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1347_axes_0 = const()[name = string("op_1347_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 16, 128]> var_1335_cast_fp16 = transpose(perm = var_1334, x = k_21_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [3, 2, 1, 16, 128]> var_1347_cast_fp16 = expand_dims(axes = var_1347_axes_0, x = var_1335_cast_fp16)[name = string("op_1347_cast_fp16")];
            tensor<int32, [5]> var_1355_reps_0 = const()[name = string("op_1355_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 16, 128]> var_1355_cast_fp16 = tile(reps = var_1355_reps_0, x = var_1347_cast_fp16)[name = string("op_1355_cast_fp16")];
            tensor<int32, [4]> var_1360 = const()[name = string("op_1360"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<fp16, [3, 4, 16, 128]> k_23_cast_fp16 = reshape(shape = var_1360, x = var_1355_cast_fp16)[name = string("k_23_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_1362_cast_fp16 = sigmoid(x = b_5_cast_fp16)[name = string("op_1362_cast_fp16")];
            tensor<int32, [3]> var_1366 = const()[name = string("op_1366"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_2_dt_bias_to_fp16 = const()[name = string("layers_2_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 16, 4]> input_37_cast_fp16 = add(x = a_5_cast_fp16, y = layers_2_dt_bias_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_1370_cast_fp16 = relu(x = input_37_cast_fp16)[name = string("op_1370_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_1371_cast_fp16 = abs(x = input_37_cast_fp16)[name = string("op_1371_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 16, 4]> var_1372_cast_fp16 = mul(x = var_1371_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_1372_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_1373_cast_fp16 = exp(x = var_1372_cast_fp16)[name = string("op_1373_cast_fp16")];
            fp16 var_1375_to_fp16 = const()[name = string("op_1375_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 16, 4]> var_1376_cast_fp16 = add(x = var_1373_cast_fp16, y = var_1375_to_fp16)[name = string("op_1376_cast_fp16")];
            fp32 var_1377_epsilon_0 = const()[name = string("op_1377_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 16, 4]> var_1377_cast_fp16 = log(epsilon = var_1377_epsilon_0, x = var_1376_cast_fp16)[name = string("op_1377_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_1379_cast_fp16 = add(x = var_1370_cast_fp16, y = var_1377_cast_fp16)[name = string("op_1379_cast_fp16")];
            tensor<fp16, [4]> layers_2_negA_to_fp16 = const()[name = string("layers_2_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.c0cp+3, -0x1.3e4p+3, -0x1.568p+1, -0x1.9acp+2])];
            tensor<fp16, [3, 16, 4]> var_1380_cast_fp16 = mul(x = layers_2_negA_to_fp16, y = var_1379_cast_fp16)[name = string("op_1380_cast_fp16")];
            tensor<int32, [3]> var_1384 = const()[name = string("op_1384"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_1387_axes_0 = const()[name = string("op_1387_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16]> beta_5_cast_fp16 = transpose(perm = var_1366, x = var_1362_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [3, 4, 16, 1]> var_1387_cast_fp16 = expand_dims(axes = var_1387_axes_0, x = beta_5_cast_fp16)[name = string("op_1387_cast_fp16")];
            fp16 var_1388_to_fp16 = const()[name = string("op_1388_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 16, 1]> var_1389_cast_fp16 = mul(x = var_1387_cast_fp16, y = var_1388_to_fp16)[name = string("op_1389_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> v_5_cast_fp16 = transpose(perm = var_1264, x = var_1259_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [3, 4, 16, 128]> vb_5_cast_fp16 = mul(x = v_5_cast_fp16, y = var_1389_cast_fp16)[name = string("vb_5_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> kb_5_cast_fp16 = mul(x = k_23_cast_fp16, y = var_1387_cast_fp16)[name = string("kb_5_cast_fp16")];
            tensor<int32, [2]> var_1396 = const()[name = string("op_1396"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [3, 4, 16]> gdec_5_cast_fp16 = transpose(perm = var_1384, x = var_1380_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [12, 16]> var_1397_cast_fp16 = reshape(shape = var_1396, x = gdec_5_cast_fp16)[name = string("op_1397_cast_fp16")];
            tensor<fp16, [12, 256]> var_1398_cast_fp16 = linear(bias = var_554_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1397_cast_fp16)[name = string("op_1398_cast_fp16")];
            tensor<int32, [4]> var_1403 = const()[name = string("op_1403"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_5_cast_fp16 = reshape(shape = var_1403, x = var_1398_cast_fp16)[name = string("P_5_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1405_cast_fp16 = exp(x = P_5_cast_fp16)[name = string("op_1405_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_5_cast_fp16 = mul(x = var_1405_cast_fp16, y = tril_incl_to_fp16)[name = string("D_5_cast_fp16")];
            bool kk_5_transpose_x_1 = const()[name = string("kk_5_transpose_x_1"), val = bool(false)];
            bool kk_5_transpose_y_1 = const()[name = string("kk_5_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_5_cast_fp16 = matmul(transpose_x = kk_5_transpose_x_1, transpose_y = kk_5_transpose_y_1, x = kb_5_cast_fp16, y = k_23_cast_fp16)[name = string("kk_5_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_1411_cast_fp16 = mul(x = kk_5_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_1411_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1412_cast_fp16 = mul(x = var_1411_cast_fp16, y = D_5_cast_fp16)[name = string("op_1412_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_5_cast_fp16 = mul(x = var_1412_cast_fp16, y = tril_strict_to_fp16)[name = string("N_5_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1414_cast_fp16 = mul(x = D_5_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1414_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_5_cast_fp16 = mul(x = kk_5_cast_fp16, y = var_1414_cast_fp16)[name = string("Na_5_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_17_cast_fp16 = add(x = eye_to_fp16, y = N_5_cast_fp16)[name = string("Minv_17_cast_fp16")];
            bool Np_13_transpose_x_0 = const()[name = string("Np_13_transpose_x_0"), val = bool(false)];
            bool Np_13_transpose_y_0 = const()[name = string("Np_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_13_cast_fp16 = matmul(transpose_x = Np_13_transpose_x_0, transpose_y = Np_13_transpose_y_0, x = N_5_cast_fp16, y = Na_5_cast_fp16)[name = string("Np_13_cast_fp16")];
            bool Npa_9_transpose_x_0 = const()[name = string("Npa_9_transpose_x_0"), val = bool(false)];
            bool Npa_9_transpose_y_0 = const()[name = string("Npa_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_9_cast_fp16 = matmul(transpose_x = Npa_9_transpose_x_0, transpose_y = Npa_9_transpose_y_0, x = Na_5_cast_fp16, y = N_5_cast_fp16)[name = string("Npa_9_cast_fp16")];
            bool var_1420_transpose_x_0 = const()[name = string("op_1420_transpose_x_0"), val = bool(false)];
            bool var_1420_transpose_y_0 = const()[name = string("op_1420_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1420_cast_fp16 = matmul(transpose_x = var_1420_transpose_x_0, transpose_y = var_1420_transpose_y_0, x = Minv_17_cast_fp16, y = Np_13_cast_fp16)[name = string("op_1420_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_19_cast_fp16 = add(x = Minv_17_cast_fp16, y = var_1420_cast_fp16)[name = string("Minv_19_cast_fp16")];
            bool Np_15_transpose_x_0 = const()[name = string("Np_15_transpose_x_0"), val = bool(false)];
            bool Np_15_transpose_y_0 = const()[name = string("Np_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_15_cast_fp16 = matmul(transpose_x = Np_15_transpose_x_0, transpose_y = Np_15_transpose_y_0, x = Np_13_cast_fp16, y = Npa_9_cast_fp16)[name = string("Np_15_cast_fp16")];
            bool Npa_11_transpose_x_0 = const()[name = string("Npa_11_transpose_x_0"), val = bool(false)];
            bool Npa_11_transpose_y_0 = const()[name = string("Npa_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_11_cast_fp16 = matmul(transpose_x = Npa_11_transpose_x_0, transpose_y = Npa_11_transpose_y_0, x = Npa_9_cast_fp16, y = Np_13_cast_fp16)[name = string("Npa_11_cast_fp16")];
            bool var_1425_transpose_x_0 = const()[name = string("op_1425_transpose_x_0"), val = bool(false)];
            bool var_1425_transpose_y_0 = const()[name = string("op_1425_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1425_cast_fp16 = matmul(transpose_x = var_1425_transpose_x_0, transpose_y = var_1425_transpose_y_0, x = Minv_19_cast_fp16, y = Np_15_cast_fp16)[name = string("op_1425_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_21_cast_fp16 = add(x = Minv_19_cast_fp16, y = var_1425_cast_fp16)[name = string("Minv_21_cast_fp16")];
            bool Np_17_transpose_x_0 = const()[name = string("Np_17_transpose_x_0"), val = bool(false)];
            bool Np_17_transpose_y_0 = const()[name = string("Np_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_17_cast_fp16 = matmul(transpose_x = Np_17_transpose_x_0, transpose_y = Np_17_transpose_y_0, x = Np_15_cast_fp16, y = Npa_11_cast_fp16)[name = string("Np_17_cast_fp16")];
            bool var_1429_transpose_x_0 = const()[name = string("op_1429_transpose_x_0"), val = bool(false)];
            bool var_1429_transpose_y_0 = const()[name = string("op_1429_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1429_cast_fp16 = matmul(transpose_x = var_1429_transpose_x_0, transpose_y = var_1429_transpose_y_0, x = Minv_21_cast_fp16, y = Np_17_cast_fp16)[name = string("op_1429_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_23_cast_fp16 = add(x = Minv_21_cast_fp16, y = var_1429_cast_fp16)[name = string("Minv_23_cast_fp16")];
            bool u_5_transpose_x_0 = const()[name = string("u_5_transpose_x_0"), val = bool(false)];
            bool u_5_transpose_y_0 = const()[name = string("u_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_5_cast_fp16 = matmul(transpose_x = u_5_transpose_x_0, transpose_y = u_5_transpose_y_0, x = Minv_23_cast_fp16, y = vb_5_cast_fp16)[name = string("u_5_cast_fp16")];
            bool var_1436_transpose_x_1 = const()[name = string("op_1436_transpose_x_1"), val = bool(false)];
            bool var_1436_transpose_y_1 = const()[name = string("op_1436_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_1436_cast_fp16 = matmul(transpose_x = var_1436_transpose_x_1, transpose_y = var_1436_transpose_y_1, x = q_23_cast_fp16, y = k_23_cast_fp16)[name = string("op_1436_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_5_cast_fp16 = mul(x = var_1436_cast_fp16, y = D_5_cast_fp16)[name = string("A_5_cast_fp16")];
            bool o_13_transpose_x_0 = const()[name = string("o_13_transpose_x_0"), val = bool(false)];
            bool o_13_transpose_y_0 = const()[name = string("o_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> o_13_cast_fp16 = matmul(transpose_x = o_13_transpose_x_0, transpose_y = o_13_transpose_y_0, x = A_5_cast_fp16, y = u_5_cast_fp16)[name = string("o_13_cast_fp16")];
            tensor<int32, [4]> var_1443 = const()[name = string("op_1443"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1449 = const()[name = string("op_1449"), val = tensor<int32, [4]>([3, 16, 4, 128])];
            tensor<fp16, [3, 16, 4, 128]> input_39_cast_fp16 = reshape(shape = var_1449, x = z_5_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> o_15_cast_fp16 = transpose(perm = var_1443, x = o_13_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [3, 16, 4, 128]> var_1451_cast_fp16 = mul(x = o_15_cast_fp16, y = o_15_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, [3, 16, 4, 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")];
            fp16 var_1458_to_fp16 = const()[name = string("op_1458_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 16, 4, 1]> var_1459_cast_fp16 = add(x = var_1456_cast_fp16, y = var_1458_to_fp16)[name = string("op_1459_cast_fp16")];
            fp32 var_1460_epsilon_0 = const()[name = string("op_1460_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 4, 1]> var_1460_cast_fp16 = rsqrt(epsilon = var_1460_epsilon_0, x = var_1459_cast_fp16)[name = string("op_1460_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> var_1461_cast_fp16 = mul(x = o_15_cast_fp16, y = var_1460_cast_fp16)[name = string("op_1461_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> var_1463_cast_fp16 = silu(x = input_39_cast_fp16)[name = string("op_1463_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> o_17_cast_fp16 = mul(x = var_1461_cast_fp16, y = var_1463_cast_fp16)[name = string("o_17_cast_fp16")];
            tensor<int32, [3]> var_1468 = const()[name = string("op_1468"), val = tensor<int32, [3]>([3, 16, 512])];
            tensor<fp16, [3, 16, 512]> input_41_cast_fp16 = reshape(shape = var_1468, x = o_17_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [640, 512]> layers_2_out_weight_to_fp16 = const()[name = string("layers_2_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47457728)))];
            tensor<fp16, [3, 16, 640]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_out_weight_to_fp16, x = input_41_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [3, 16, 640]> x_21_cast_fp16 = add(x = x_17_cast_fp16, y = linear_9_cast_fp16)[name = string("x_21_cast_fp16")];
            fp16 var_1479_promoted_to_fp16 = const()[name = string("op_1479_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> var_1483_cast_fp16 = pow(x = x_21_cast_fp16, y = var_1479_promoted_to_fp16)[name = string("op_1483_cast_fp16")];
            tensor<int32, [1]> var_1485_axes_0 = const()[name = string("op_1485_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1485_keep_dims_0 = const()[name = string("op_1485_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 1]> var_1485_cast_fp16 = reduce_mean(axes = var_1485_axes_0, keep_dims = var_1485_keep_dims_0, x = var_1483_cast_fp16)[name = string("op_1485_cast_fp16")];
            fp16 var_1486_to_fp16 = const()[name = string("op_1486_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 16, 1]> var_1487_cast_fp16 = add(x = var_1485_cast_fp16, y = var_1486_to_fp16)[name = string("op_1487_cast_fp16")];
            fp32 var_1488_epsilon_0 = const()[name = string("op_1488_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 1]> var_1488_cast_fp16 = rsqrt(epsilon = var_1488_epsilon_0, x = var_1487_cast_fp16)[name = string("op_1488_cast_fp16")];
            tensor<fp16, [3, 16, 640]> var_1489_cast_fp16 = mul(x = x_21_cast_fp16, y = var_1488_cast_fp16)[name = string("op_1489_cast_fp16")];
            tensor<fp16, [640]> layers_2_n2_w1_to_fp16 = const()[name = string("layers_2_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48113152)))];
            tensor<fp16, [3, 16, 640]> input_43_cast_fp16 = mul(x = var_1489_cast_fp16, y = layers_2_n2_w1_to_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_2_gate_up_weight_to_fp16 = const()[name = string("layers_2_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48114496)))];
            tensor<fp16, [3, 16, 5120]> linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_2_gate_up_weight_to_fp16, x = input_43_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [3]> input_45_begin_0 = const()[name = string("input_45_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_45_end_0 = const()[name = string("input_45_end_0"), val = tensor<int32, [3]>([3, 16, 2560])];
            tensor<bool, [3]> input_45_end_mask_0 = const()[name = string("input_45_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 2560]> input_45_cast_fp16 = slice_by_index(begin = input_45_begin_0, end = input_45_end_0, end_mask = input_45_end_mask_0, x = linear_10_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [3, 16, 2560]> var_1499_cast_fp16 = silu(x = input_45_cast_fp16)[name = string("op_1499_cast_fp16")];
            tensor<int32, [3]> var_1504_begin_0 = const()[name = string("op_1504_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_1504_end_0 = const()[name = string("op_1504_end_0"), val = tensor<int32, [3]>([3, 16, 5120])];
            tensor<bool, [3]> var_1504_end_mask_0 = const()[name = string("op_1504_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 2560]> var_1504_cast_fp16 = slice_by_index(begin = var_1504_begin_0, end = var_1504_end_0, end_mask = var_1504_end_mask_0, x = linear_10_cast_fp16)[name = string("op_1504_cast_fp16")];
            tensor<fp16, [3, 16, 2560]> input_47_cast_fp16 = mul(x = var_1499_cast_fp16, y = var_1504_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_2_down_weight_to_fp16 = const()[name = string("layers_2_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54668160)))];
            tensor<fp16, [3, 16, 640]> linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_down_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [3, 16, 640]> x_25_cast_fp16 = add(x = x_21_cast_fp16, y = linear_11_cast_fp16)[name = string("x_25_cast_fp16")];
            fp16 var_1517_to_fp16 = const()[name = string("op_1517_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 16, 640]> x_27_cast_fp16 = mul(x = x_25_cast_fp16, y = var_1517_to_fp16)[name = string("x_27_cast_fp16")];
            fp16 var_1515_promoted_to_fp16 = const()[name = string("op_1515_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> var_1519_cast_fp16 = pow(x = x_27_cast_fp16, y = var_1515_promoted_to_fp16)[name = string("op_1519_cast_fp16")];
            tensor<int32, [1]> var_1521_axes_0 = const()[name = string("op_1521_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1521_keep_dims_0 = const()[name = string("op_1521_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 1]> var_1521_cast_fp16 = reduce_mean(axes = var_1521_axes_0, keep_dims = var_1521_keep_dims_0, x = var_1519_cast_fp16)[name = string("op_1521_cast_fp16")];
            fp16 var_1522_to_fp16 = const()[name = string("op_1522_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 16, 1]> var_1523_cast_fp16 = add(x = var_1521_cast_fp16, y = var_1522_to_fp16)[name = string("op_1523_cast_fp16")];
            fp32 var_1524_epsilon_0 = const()[name = string("op_1524_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 1]> var_1524_cast_fp16 = rsqrt(epsilon = var_1524_epsilon_0, x = var_1523_cast_fp16)[name = string("op_1524_cast_fp16")];
            tensor<fp16, [3, 16, 640]> var_1525_cast_fp16 = mul(x = x_27_cast_fp16, y = var_1524_cast_fp16)[name = string("op_1525_cast_fp16")];
            tensor<fp16, [640]> layers_3_n1_w1_to_fp16 = const()[name = string("layers_3_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57945024)))];
            tensor<fp16, [3, 16, 640]> input_49_cast_fp16 = mul(x = var_1525_cast_fp16, y = layers_3_n1_w1_to_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1280, 640]> layers_3_qkv_weight_to_fp16 = const()[name = string("layers_3_qkv_weight_to_fp16"), val = tensor<fp16, [1280, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57946368)))];
            tensor<fp16, [1280]> linear_12_bias_0_to_fp16 = const()[name = string("linear_12_bias_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59584832)))];
            tensor<fp16, [3, 16, 1280]> linear_12_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = layers_3_qkv_weight_to_fp16, x = input_49_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [3]> var_1534_begin_0 = const()[name = string("op_1534_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1534_end_0 = const()[name = string("op_1534_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> var_1534_end_mask_0 = const()[name = string("op_1534_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 1024]> var_1534_cast_fp16 = slice_by_index(begin = var_1534_begin_0, end = var_1534_end_0, end_mask = var_1534_end_mask_0, x = linear_12_cast_fp16)[name = string("op_1534_cast_fp16")];
            tensor<int32, [4]> var_1539 = const()[name = string("op_1539"), val = tensor<int32, [4]>([3, 16, 8, 128])];
            tensor<fp16, [3, 16, 8, 128]> qg_1_cast_fp16 = reshape(shape = var_1539, x = var_1534_cast_fp16)[name = string("qg_1_cast_fp16")];
            tensor<int32, [4]> x_29_begin_0 = const()[name = string("x_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_29_end_0 = const()[name = string("x_29_end_0"), val = tensor<int32, [4]>([3, 16, 8, 64])];
            tensor<bool, [4]> x_29_end_mask_0 = const()[name = string("x_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 16, 8, 64]> x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, x = qg_1_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [4]> var_1550_begin_0 = const()[name = string("op_1550_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> var_1550_end_0 = const()[name = string("op_1550_end_0"), val = tensor<int32, [4]>([3, 16, 8, 128])];
            tensor<bool, [4]> var_1550_end_mask_0 = const()[name = string("op_1550_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 16, 8, 64]> var_1550_cast_fp16 = slice_by_index(begin = var_1550_begin_0, end = var_1550_end_0, end_mask = var_1550_end_mask_0, x = qg_1_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<int32, [3]> var_1554 = const()[name = string("op_1554"), val = tensor<int32, [3]>([3, 16, 512])];
            tensor<fp16, [3, 16, 512]> gate_1_cast_fp16 = reshape(shape = var_1554, x = var_1550_cast_fp16)[name = string("gate_1_cast_fp16")];
            tensor<int32, [3]> var_1560_begin_0 = const()[name = string("op_1560_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> var_1560_end_0 = const()[name = string("op_1560_end_0"), val = tensor<int32, [3]>([3, 16, 1152])];
            tensor<bool, [3]> var_1560_end_mask_0 = const()[name = string("op_1560_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 128]> var_1560_cast_fp16 = slice_by_index(begin = var_1560_begin_0, end = var_1560_end_0, end_mask = var_1560_end_mask_0, x = linear_12_cast_fp16)[name = string("op_1560_cast_fp16")];
            tensor<int32, [4]> var_1565 = const()[name = string("op_1565"), val = tensor<int32, [4]>([3, 16, 2, 64])];
            tensor<fp16, [3, 16, 2, 64]> x_35_cast_fp16 = reshape(shape = var_1565, x = var_1560_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<int32, [3]> var_1571_begin_0 = const()[name = string("op_1571_begin_0"), val = tensor<int32, [3]>([0, 0, 1152])];
            tensor<int32, [3]> var_1571_end_0 = const()[name = string("op_1571_end_0"), val = tensor<int32, [3]>([3, 16, 1280])];
            tensor<bool, [3]> var_1571_end_mask_0 = const()[name = string("op_1571_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 128]> var_1571_cast_fp16 = slice_by_index(begin = var_1571_begin_0, end = var_1571_end_0, end_mask = var_1571_end_mask_0, x = linear_12_cast_fp16)[name = string("op_1571_cast_fp16")];
            tensor<int32, [4]> var_1576 = const()[name = string("op_1576"), val = tensor<int32, [4]>([3, 16, 2, 64])];
            tensor<fp16, [3, 16, 2, 64]> var_1577_cast_fp16 = reshape(shape = var_1576, x = var_1571_cast_fp16)[name = string("op_1577_cast_fp16")];
            tensor<int32, [4]> var_1582 = const()[name = string("op_1582"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1588_promoted_to_fp16 = const()[name = string("op_1588_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 8, 64]> var_1592_cast_fp16 = pow(x = x_29_cast_fp16, y = var_1588_promoted_to_fp16)[name = string("op_1592_cast_fp16")];
            tensor<int32, [1]> var_1594_axes_0 = const()[name = string("op_1594_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1594_keep_dims_0 = const()[name = string("op_1594_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 8, 1]> var_1594_cast_fp16 = reduce_mean(axes = var_1594_axes_0, keep_dims = var_1594_keep_dims_0, x = var_1592_cast_fp16)[name = string("op_1594_cast_fp16")];
            fp16 var_1595_to_fp16 = const()[name = string("op_1595_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 16, 8, 1]> var_1596_cast_fp16 = add(x = var_1594_cast_fp16, y = var_1595_to_fp16)[name = string("op_1596_cast_fp16")];
            fp32 var_1597_epsilon_0 = const()[name = string("op_1597_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 8, 1]> var_1597_cast_fp16 = rsqrt(epsilon = var_1597_epsilon_0, x = var_1596_cast_fp16)[name = string("op_1597_cast_fp16")];
            tensor<fp16, [3, 16, 8, 64]> var_1598_cast_fp16 = mul(x = x_29_cast_fp16, y = var_1597_cast_fp16)[name = string("op_1598_cast_fp16")];
            tensor<fp16, [64]> layers_3_qn_w1_to_fp16 = const()[name = string("layers_3_qn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59587456)))];
            tensor<fp16, [3, 16, 8, 64]> var_1599_cast_fp16 = mul(x = var_1598_cast_fp16, y = layers_3_qn_w1_to_fp16)[name = string("op_1599_cast_fp16")];
            tensor<int32, [4]> var_1604 = const()[name = string("op_1604"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_1_begin_0 = const()[name = string("r_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_1_end_0 = const()[name = string("r_1_end_0"), val = tensor<int32, [4]>([3, 8, 16, 16])];
            tensor<bool, [4]> r_1_end_mask_0 = const()[name = string("r_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 8, 16, 64]> x_33_cast_fp16 = transpose(perm = var_1604, x = var_1599_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [3, 8, 16, 16]> r_1_cast_fp16 = slice_by_index(begin = r_1_begin_0, end = r_1_end_0, end_mask = r_1_end_mask_0, x = x_33_cast_fp16)[name = string("r_1_cast_fp16")];
            tensor<int32, [4]> p_1_begin_0 = const()[name = string("p_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_1_end_0 = const()[name = string("p_1_end_0"), val = tensor<int32, [4]>([3, 8, 16, 64])];
            tensor<bool, [4]> p_1_end_mask_0 = const()[name = string("p_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 8, 16, 48]> p_1_cast_fp16 = slice_by_index(begin = p_1_begin_0, end = p_1_end_0, end_mask = p_1_end_mask_0, x = x_33_cast_fp16)[name = string("p_1_cast_fp16")];
            tensor<int32, [4]> var_1620_begin_0 = const()[name = string("op_1620_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_1620_end_0 = const()[name = string("op_1620_end_0"), val = tensor<int32, [4]>([3, 8, 16, 16])];
            tensor<bool, [4]> var_1620_end_mask_0 = const()[name = string("op_1620_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 8, 16, 8]> var_1620_cast_fp16 = slice_by_index(begin = var_1620_begin_0, end = var_1620_end_0, end_mask = var_1620_end_mask_0, x = r_1_cast_fp16)[name = string("op_1620_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 8, 16, 8]> var_1621_cast_fp16 = mul(x = var_1620_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_1621_cast_fp16")];
            tensor<int32, [4]> var_1626_begin_0 = const()[name = string("op_1626_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1626_end_0 = const()[name = string("op_1626_end_0"), val = tensor<int32, [4]>([3, 8, 16, 8])];
            tensor<bool, [4]> var_1626_end_mask_0 = const()[name = string("op_1626_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 8, 16, 8]> var_1626_cast_fp16 = slice_by_index(begin = var_1626_begin_0, end = var_1626_end_0, end_mask = var_1626_end_mask_0, x = r_1_cast_fp16)[name = string("op_1626_cast_fp16")];
            int32 var_1628 = const()[name = string("op_1628"), val = int32(-1)];
            bool rh_1_interleave_0 = const()[name = string("rh_1_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 8, 16, 16]> rh_1_cast_fp16 = concat(axis = var_1628, interleave = rh_1_interleave_0, values = (var_1621_cast_fp16, var_1626_cast_fp16))[name = string("rh_1_cast_fp16")];
            tensor<fp16, [16, 16]> cos_to_fp16 = const()[name = string("cos_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59587648)))];
            tensor<fp16, [3, 8, 16, 16]> var_1630_cast_fp16 = mul(x = r_1_cast_fp16, y = cos_to_fp16)[name = string("op_1630_cast_fp16")];
            tensor<fp16, [16, 16]> sin_to_fp16 = const()[name = string("sin_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59588224)))];
            tensor<fp16, [3, 8, 16, 16]> var_1631_cast_fp16 = mul(x = rh_1_cast_fp16, y = sin_to_fp16)[name = string("op_1631_cast_fp16")];
            tensor<fp16, [3, 8, 16, 16]> var_1633_cast_fp16 = add(x = var_1630_cast_fp16, y = var_1631_cast_fp16)[name = string("op_1633_cast_fp16")];
            int32 var_1635 = const()[name = string("op_1635"), val = int32(-1)];
            bool q_25_interleave_0 = const()[name = string("q_25_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 8, 16, 64]> q_25_cast_fp16 = concat(axis = var_1635, interleave = q_25_interleave_0, values = (var_1633_cast_fp16, p_1_cast_fp16))[name = string("q_25_cast_fp16")];
            fp16 var_1641_promoted_to_fp16 = const()[name = string("op_1641_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 2, 64]> var_1645_cast_fp16 = pow(x = x_35_cast_fp16, y = var_1641_promoted_to_fp16)[name = string("op_1645_cast_fp16")];
            tensor<int32, [1]> var_1647_axes_0 = const()[name = string("op_1647_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1647_keep_dims_0 = const()[name = string("op_1647_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 2, 1]> var_1647_cast_fp16 = reduce_mean(axes = var_1647_axes_0, keep_dims = var_1647_keep_dims_0, x = var_1645_cast_fp16)[name = string("op_1647_cast_fp16")];
            fp16 var_1648_to_fp16 = const()[name = string("op_1648_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 16, 2, 1]> var_1649_cast_fp16 = add(x = var_1647_cast_fp16, y = var_1648_to_fp16)[name = string("op_1649_cast_fp16")];
            fp32 var_1650_epsilon_0 = const()[name = string("op_1650_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 2, 1]> var_1650_cast_fp16 = rsqrt(epsilon = var_1650_epsilon_0, x = var_1649_cast_fp16)[name = string("op_1650_cast_fp16")];
            tensor<fp16, [3, 16, 2, 64]> var_1651_cast_fp16 = mul(x = x_35_cast_fp16, y = var_1650_cast_fp16)[name = string("op_1651_cast_fp16")];
            tensor<fp16, [64]> layers_3_kn_w1_to_fp16 = const()[name = string("layers_3_kn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59588800)))];
            tensor<fp16, [3, 16, 2, 64]> var_1652_cast_fp16 = mul(x = var_1651_cast_fp16, y = layers_3_kn_w1_to_fp16)[name = string("op_1652_cast_fp16")];
            tensor<int32, [4]> var_1657 = const()[name = string("op_1657"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_3_begin_0 = const()[name = string("r_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_3_end_0 = const()[name = string("r_3_end_0"), val = tensor<int32, [4]>([3, 2, 16, 16])];
            tensor<bool, [4]> r_3_end_mask_0 = const()[name = string("r_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 2, 16, 64]> x_39_cast_fp16 = transpose(perm = var_1657, x = var_1652_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [3, 2, 16, 16]> r_3_cast_fp16 = slice_by_index(begin = r_3_begin_0, end = r_3_end_0, end_mask = r_3_end_mask_0, x = x_39_cast_fp16)[name = string("r_3_cast_fp16")];
            tensor<int32, [4]> p_3_begin_0 = const()[name = string("p_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_3_end_0 = const()[name = string("p_3_end_0"), val = tensor<int32, [4]>([3, 2, 16, 64])];
            tensor<bool, [4]> p_3_end_mask_0 = const()[name = string("p_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 2, 16, 48]> p_3_cast_fp16 = slice_by_index(begin = p_3_begin_0, end = p_3_end_0, end_mask = p_3_end_mask_0, x = x_39_cast_fp16)[name = string("p_3_cast_fp16")];
            tensor<int32, [4]> var_1673_begin_0 = const()[name = string("op_1673_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_1673_end_0 = const()[name = string("op_1673_end_0"), val = tensor<int32, [4]>([3, 2, 16, 16])];
            tensor<bool, [4]> var_1673_end_mask_0 = const()[name = string("op_1673_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 2, 16, 8]> var_1673_cast_fp16 = slice_by_index(begin = var_1673_begin_0, end = var_1673_end_0, end_mask = var_1673_end_mask_0, x = r_3_cast_fp16)[name = string("op_1673_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 2, 16, 8]> var_1674_cast_fp16 = mul(x = var_1673_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_1674_cast_fp16")];
            tensor<int32, [4]> var_1679_begin_0 = const()[name = string("op_1679_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1679_end_0 = const()[name = string("op_1679_end_0"), val = tensor<int32, [4]>([3, 2, 16, 8])];
            tensor<bool, [4]> var_1679_end_mask_0 = const()[name = string("op_1679_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 2, 16, 8]> var_1679_cast_fp16 = slice_by_index(begin = var_1679_begin_0, end = var_1679_end_0, end_mask = var_1679_end_mask_0, x = r_3_cast_fp16)[name = string("op_1679_cast_fp16")];
            int32 var_1681 = const()[name = string("op_1681"), val = int32(-1)];
            bool rh_3_interleave_0 = const()[name = string("rh_3_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 2, 16, 16]> rh_3_cast_fp16 = concat(axis = var_1681, interleave = rh_3_interleave_0, values = (var_1674_cast_fp16, var_1679_cast_fp16))[name = string("rh_3_cast_fp16")];
            tensor<fp16, [3, 2, 16, 16]> var_1683_cast_fp16 = mul(x = r_3_cast_fp16, y = cos_to_fp16)[name = string("op_1683_cast_fp16")];
            tensor<fp16, [3, 2, 16, 16]> var_1684_cast_fp16 = mul(x = rh_3_cast_fp16, y = sin_to_fp16)[name = string("op_1684_cast_fp16")];
            tensor<fp16, [3, 2, 16, 16]> var_1686_cast_fp16 = add(x = var_1683_cast_fp16, y = var_1684_cast_fp16)[name = string("op_1686_cast_fp16")];
            int32 var_1688 = const()[name = string("op_1688"), val = int32(-1)];
            bool k_25_interleave_0 = const()[name = string("k_25_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 2, 16, 64]> k_25_cast_fp16 = concat(axis = var_1688, interleave = k_25_interleave_0, values = (var_1686_cast_fp16, p_3_cast_fp16))[name = string("k_25_cast_fp16")];
            tensor<int32, [1]> var_1701_axes_0 = const()[name = string("op_1701_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 1, 16, 64]> var_1701_cast_fp16 = expand_dims(axes = var_1701_axes_0, x = k_25_cast_fp16)[name = string("op_1701_cast_fp16")];
            tensor<int32, [5]> var_1709_reps_0 = const()[name = string("op_1709_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [3, 2, 4, 16, 64]> var_1709_cast_fp16 = tile(reps = var_1709_reps_0, x = var_1701_cast_fp16)[name = string("op_1709_cast_fp16")];
            tensor<int32, [4]> var_1714 = const()[name = string("op_1714"), val = tensor<int32, [4]>([3, 8, 16, 64])];
            tensor<fp16, [3, 8, 16, 64]> k_27_cast_fp16 = reshape(shape = var_1714, x = var_1709_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<int32, [1]> var_1727_axes_0 = const()[name = string("op_1727_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 16, 64]> v_7_cast_fp16 = transpose(perm = var_1582, x = var_1577_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [3, 2, 1, 16, 64]> var_1727_cast_fp16 = expand_dims(axes = var_1727_axes_0, x = v_7_cast_fp16)[name = string("op_1727_cast_fp16")];
            tensor<int32, [5]> var_1735_reps_0 = const()[name = string("op_1735_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [3, 2, 4, 16, 64]> var_1735_cast_fp16 = tile(reps = var_1735_reps_0, x = var_1727_cast_fp16)[name = string("op_1735_cast_fp16")];
            tensor<int32, [4]> var_1740 = const()[name = string("op_1740"), val = tensor<int32, [4]>([3, 8, 16, 64])];
            tensor<fp16, [3, 8, 16, 64]> v_9_cast_fp16 = reshape(shape = var_1740, x = var_1735_cast_fp16)[name = string("v_9_cast_fp16")];
            bool var_1745_transpose_x_1 = const()[name = string("op_1745_transpose_x_1"), val = bool(false)];
            bool var_1745_transpose_y_1 = const()[name = string("op_1745_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 8, 16, 16]> var_1745_cast_fp16 = matmul(transpose_x = var_1745_transpose_x_1, transpose_y = var_1745_transpose_y_1, x = q_25_cast_fp16, y = k_27_cast_fp16)[name = string("op_1745_cast_fp16")];
            fp16 var_1746_to_fp16 = const()[name = string("op_1746_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [3, 8, 16, 16]> var_1747_cast_fp16 = mul(x = var_1745_cast_fp16, y = var_1746_to_fp16)[name = string("op_1747_cast_fp16")];
            tensor<fp16, [16, 16]> causal_add_to_fp16 = const()[name = string("causal_add_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59588992)))];
            tensor<fp16, [3, 8, 16, 16]> s_1_cast_fp16 = add(x = var_1747_cast_fp16, y = causal_add_to_fp16)[name = string("s_1_cast_fp16")];
            int32 var_1750 = const()[name = string("op_1750"), val = int32(-1)];
            tensor<fp16, [3, 8, 16, 16]> var_1752_cast_fp16 = softmax(axis = var_1750, x = s_1_cast_fp16)[name = string("op_1752_cast_fp16")];
            bool o_19_transpose_x_0 = const()[name = string("o_19_transpose_x_0"), val = bool(false)];
            bool o_19_transpose_y_0 = const()[name = string("o_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 8, 16, 64]> o_19_cast_fp16 = matmul(transpose_x = o_19_transpose_x_0, transpose_y = o_19_transpose_y_0, x = var_1752_cast_fp16, y = v_9_cast_fp16)[name = string("o_19_cast_fp16")];
            tensor<int32, [4]> var_1758 = const()[name = string("op_1758"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1763 = const()[name = string("op_1763"), val = tensor<int32, [3]>([3, 16, 512])];
            tensor<fp16, [3, 16, 8, 64]> var_1759_cast_fp16 = transpose(perm = var_1758, x = o_19_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [3, 16, 512]> var_1764_cast_fp16 = reshape(shape = var_1763, x = var_1759_cast_fp16)[name = string("op_1764_cast_fp16")];
            tensor<fp16, [3, 16, 512]> var_1765_cast_fp16 = sigmoid(x = gate_1_cast_fp16)[name = string("op_1765_cast_fp16")];
            tensor<fp16, [3, 16, 512]> input_51_cast_fp16 = mul(x = var_1764_cast_fp16, y = var_1765_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [640, 512]> layers_3_o_weight_to_fp16 = const()[name = string("layers_3_o_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59589568)))];
            tensor<fp16, [3, 16, 640]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_o_weight_to_fp16, x = input_51_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [3, 16, 640]> x_41_cast_fp16 = add(x = x_25_cast_fp16, y = linear_13_cast_fp16)[name = string("x_41_cast_fp16")];
            fp16 var_1778_to_fp16 = const()[name = string("op_1778_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 16, 640]> x_43_cast_fp16 = mul(x = x_41_cast_fp16, y = var_1778_to_fp16)[name = string("x_43_cast_fp16")];
            fp16 var_1776_promoted_to_fp16 = const()[name = string("op_1776_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> var_1780_cast_fp16 = pow(x = x_43_cast_fp16, y = var_1776_promoted_to_fp16)[name = string("op_1780_cast_fp16")];
            tensor<int32, [1]> var_1782_axes_0 = const()[name = string("op_1782_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1782_keep_dims_0 = const()[name = string("op_1782_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 1]> var_1782_cast_fp16 = reduce_mean(axes = var_1782_axes_0, keep_dims = var_1782_keep_dims_0, x = var_1780_cast_fp16)[name = string("op_1782_cast_fp16")];
            fp16 var_1783_to_fp16 = const()[name = string("op_1783_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 16, 1]> var_1784_cast_fp16 = add(x = var_1782_cast_fp16, y = var_1783_to_fp16)[name = string("op_1784_cast_fp16")];
            fp32 var_1785_epsilon_0 = const()[name = string("op_1785_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 1]> var_1785_cast_fp16 = rsqrt(epsilon = var_1785_epsilon_0, x = var_1784_cast_fp16)[name = string("op_1785_cast_fp16")];
            tensor<fp16, [3, 16, 640]> var_1786_cast_fp16 = mul(x = x_43_cast_fp16, y = var_1785_cast_fp16)[name = string("op_1786_cast_fp16")];
            tensor<fp16, [640]> layers_3_n2_w1_to_fp16 = const()[name = string("layers_3_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60244992)))];
            tensor<fp16, [3, 16, 640]> input_53_cast_fp16 = mul(x = var_1786_cast_fp16, y = layers_3_n2_w1_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_3_gate_up_weight_to_fp16 = const()[name = string("layers_3_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60246336)))];
            tensor<fp16, [3, 16, 5120]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_3_gate_up_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [3]> input_55_begin_0 = const()[name = string("input_55_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_55_end_0 = const()[name = string("input_55_end_0"), val = tensor<int32, [3]>([3, 16, 2560])];
            tensor<bool, [3]> input_55_end_mask_0 = const()[name = string("input_55_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 2560]> input_55_cast_fp16 = slice_by_index(begin = input_55_begin_0, end = input_55_end_0, end_mask = input_55_end_mask_0, x = linear_14_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [3, 16, 2560]> var_1796_cast_fp16 = silu(x = input_55_cast_fp16)[name = string("op_1796_cast_fp16")];
            tensor<int32, [3]> var_1801_begin_0 = const()[name = string("op_1801_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_1801_end_0 = const()[name = string("op_1801_end_0"), val = tensor<int32, [3]>([3, 16, 5120])];
            tensor<bool, [3]> var_1801_end_mask_0 = const()[name = string("op_1801_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 2560]> var_1801_cast_fp16 = slice_by_index(begin = var_1801_begin_0, end = var_1801_end_0, end_mask = var_1801_end_mask_0, x = linear_14_cast_fp16)[name = string("op_1801_cast_fp16")];
            tensor<fp16, [3, 16, 2560]> input_57_cast_fp16 = mul(x = var_1796_cast_fp16, y = var_1801_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_3_down_weight_to_fp16 = const()[name = string("layers_3_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66800000)))];
            tensor<fp16, [3, 16, 640]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_down_weight_to_fp16, x = input_57_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [3, 16, 640]> x_45_cast_fp16 = add(x = x_41_cast_fp16, y = linear_15_cast_fp16)[name = string("x_45_cast_fp16")];
            fp16 var_1814_to_fp16 = const()[name = string("op_1814_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 16, 640]> x_47_cast_fp16 = mul(x = x_45_cast_fp16, y = var_1814_to_fp16)[name = string("x_47_cast_fp16")];
            fp16 var_1812_promoted_to_fp16 = const()[name = string("op_1812_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> var_1816_cast_fp16 = pow(x = x_47_cast_fp16, y = var_1812_promoted_to_fp16)[name = string("op_1816_cast_fp16")];
            tensor<int32, [1]> var_1818_axes_0 = const()[name = string("op_1818_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1818_keep_dims_0 = const()[name = string("op_1818_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 1]> var_1818_cast_fp16 = reduce_mean(axes = var_1818_axes_0, keep_dims = var_1818_keep_dims_0, x = var_1816_cast_fp16)[name = string("op_1818_cast_fp16")];
            fp16 var_1819_to_fp16 = const()[name = string("op_1819_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 16, 1]> var_1820_cast_fp16 = add(x = var_1818_cast_fp16, y = var_1819_to_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, [3, 16, 1]> var_1821_cast_fp16 = rsqrt(epsilon = var_1821_epsilon_0, x = var_1820_cast_fp16)[name = string("op_1821_cast_fp16")];
            tensor<fp16, [3, 16, 640]> var_1822_cast_fp16 = mul(x = x_47_cast_fp16, y = var_1821_cast_fp16)[name = string("op_1822_cast_fp16")];
            tensor<fp16, [640]> layers_4_n1_w1_to_fp16 = const()[name = string("layers_4_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70076864)))];
            tensor<fp16, [3, 16, 640]> input_59_cast_fp16 = mul(x = var_1822_cast_fp16, y = layers_4_n1_w1_to_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_4_qkvzba_weight_to_fp16 = const()[name = string("layers_4_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70078208)))];
            tensor<fp16, [3, 16, 1544]> linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_qkvzba_weight_to_fp16, x = input_59_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [3]> qkv_7_begin_0 = const()[name = string("qkv_7_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_7_end_0 = const()[name = string("qkv_7_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> qkv_7_end_mask_0 = const()[name = string("qkv_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 1024]> qkv_7_cast_fp16 = slice_by_index(begin = qkv_7_begin_0, end = qkv_7_end_0, end_mask = qkv_7_end_mask_0, x = linear_16_cast_fp16)[name = string("qkv_7_cast_fp16")];
            tensor<int32, [3]> z_7_begin_0 = const()[name = string("z_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_7_end_0 = const()[name = string("z_7_end_0"), val = tensor<int32, [3]>([3, 16, 1536])];
            tensor<bool, [3]> z_7_end_mask_0 = const()[name = string("z_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 512]> z_7_cast_fp16 = slice_by_index(begin = z_7_begin_0, end = z_7_end_0, end_mask = z_7_end_mask_0, x = linear_16_cast_fp16)[name = string("z_7_cast_fp16")];
            tensor<int32, [3]> b_7_begin_0 = const()[name = string("b_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_7_end_0 = const()[name = string("b_7_end_0"), val = tensor<int32, [3]>([3, 16, 1540])];
            tensor<bool, [3]> b_7_end_mask_0 = const()[name = string("b_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 4]> b_7_cast_fp16 = slice_by_index(begin = b_7_begin_0, end = b_7_end_0, end_mask = b_7_end_mask_0, x = linear_16_cast_fp16)[name = string("b_7_cast_fp16")];
            tensor<int32, [3]> a_7_begin_0 = const()[name = string("a_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_7_end_0 = const()[name = string("a_7_end_0"), val = tensor<int32, [3]>([3, 16, 1])];
            tensor<bool, [3]> a_7_end_mask_0 = const()[name = string("a_7_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 4]> a_7_cast_fp16 = slice_by_index(begin = a_7_begin_0, end = a_7_end_0, end_mask = a_7_end_mask_0, x = linear_16_cast_fp16)[name = string("a_7_cast_fp16")];
            int32 var_1861 = const()[name = string("op_1861"), val = int32(1)];
            bool xp_7_interleave_0 = const()[name = string("xp_7_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 19, 1024]> xp_7_cast_fp16 = concat(axis = var_1861, interleave = xp_7_interleave_0, values = (zpad_to_fp16, qkv_7_cast_fp16))[name = string("xp_7_cast_fp16")];
            tensor<int32, [3]> var_1872_begin_0 = const()[name = string("op_1872_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_1872_end_0 = const()[name = string("op_1872_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_1872_end_mask_0 = const()[name = string("op_1872_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 1024]> 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 = xp_7_cast_fp16)[name = string("op_1872_cast_fp16")];
            tensor<fp16, [1024]> var_1880_to_fp16 = const()[name = string("op_1880_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72054592)))];
            tensor<fp16, [3, 16, 1024]> var_1881_cast_fp16 = mul(x = var_1872_cast_fp16, y = var_1880_to_fp16)[name = string("op_1881_cast_fp16")];
            tensor<int32, [3]> var_1891_begin_0 = const()[name = string("op_1891_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_1891_end_0 = const()[name = string("op_1891_end_0"), val = tensor<int32, [3]>([3, 18, 1024])];
            tensor<bool, [3]> var_1891_end_mask_0 = const()[name = string("op_1891_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_1891_cast_fp16 = slice_by_index(begin = var_1891_begin_0, end = var_1891_end_0, end_mask = var_1891_end_mask_0, x = xp_7_cast_fp16)[name = string("op_1891_cast_fp16")];
            tensor<fp16, [1024]> var_1899_to_fp16 = const()[name = string("op_1899_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72056704)))];
            tensor<fp16, [3, 16, 1024]> var_1900_cast_fp16 = mul(x = var_1891_cast_fp16, y = var_1899_to_fp16)[name = string("op_1900_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> var_1902_cast_fp16 = add(x = var_1881_cast_fp16, y = var_1900_cast_fp16)[name = string("op_1902_cast_fp16")];
            tensor<int32, [3]> var_1912_begin_0 = const()[name = string("op_1912_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1912_end_0 = const()[name = string("op_1912_end_0"), val = tensor<int32, [3]>([3, 17, 1024])];
            tensor<bool, [3]> var_1912_end_mask_0 = const()[name = string("op_1912_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_1912_cast_fp16 = slice_by_index(begin = var_1912_begin_0, end = var_1912_end_0, end_mask = var_1912_end_mask_0, x = xp_7_cast_fp16)[name = string("op_1912_cast_fp16")];
            tensor<fp16, [1024]> var_1920_to_fp16 = const()[name = string("op_1920_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72058816)))];
            tensor<fp16, [3, 16, 1024]> var_1921_cast_fp16 = mul(x = var_1912_cast_fp16, y = var_1920_to_fp16)[name = string("op_1921_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> var_1923_cast_fp16 = add(x = var_1902_cast_fp16, y = var_1921_cast_fp16)[name = string("op_1923_cast_fp16")];
            tensor<int32, [3]> var_1933_begin_0 = const()[name = string("op_1933_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1933_end_0 = const()[name = string("op_1933_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> var_1933_end_mask_0 = const()[name = string("op_1933_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_1933_cast_fp16 = slice_by_index(begin = var_1933_begin_0, end = var_1933_end_0, end_mask = var_1933_end_mask_0, x = xp_7_cast_fp16)[name = string("op_1933_cast_fp16")];
            tensor<fp16, [1024]> var_1941_to_fp16 = const()[name = string("op_1941_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72060928)))];
            tensor<fp16, [3, 16, 1024]> var_1942_cast_fp16 = mul(x = var_1933_cast_fp16, y = var_1941_to_fp16)[name = string("op_1942_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> input_61_cast_fp16 = add(x = var_1923_cast_fp16, y = var_1942_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> c_7_cast_fp16 = silu(x = input_61_cast_fp16)[name = string("c_7_cast_fp16")];
            tensor<int32, [3]> var_1950_begin_0 = const()[name = string("op_1950_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1950_end_0 = const()[name = string("op_1950_end_0"), val = tensor<int32, [3]>([3, 16, 256])];
            tensor<bool, [3]> var_1950_end_mask_0 = const()[name = string("op_1950_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 256]> var_1950_cast_fp16 = slice_by_index(begin = var_1950_begin_0, end = var_1950_end_0, end_mask = var_1950_end_mask_0, x = c_7_cast_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [4]> var_1955 = const()[name = string("op_1955"), val = tensor<int32, [4]>([3, 16, 2, 128])];
            tensor<fp16, [3, 16, 2, 128]> q_27_cast_fp16 = reshape(shape = var_1955, x = var_1950_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<int32, [3]> var_1961_begin_0 = const()[name = string("op_1961_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_1961_end_0 = const()[name = string("op_1961_end_0"), val = tensor<int32, [3]>([3, 16, 512])];
            tensor<bool, [3]> var_1961_end_mask_0 = const()[name = string("op_1961_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 256]> var_1961_cast_fp16 = slice_by_index(begin = var_1961_begin_0, end = var_1961_end_0, end_mask = var_1961_end_mask_0, x = c_7_cast_fp16)[name = string("op_1961_cast_fp16")];
            tensor<int32, [4]> var_1966 = const()[name = string("op_1966"), val = tensor<int32, [4]>([3, 16, 2, 128])];
            tensor<fp16, [3, 16, 2, 128]> k_29_cast_fp16 = reshape(shape = var_1966, x = var_1961_cast_fp16)[name = string("k_29_cast_fp16")];
            tensor<int32, [3]> var_1972_begin_0 = const()[name = string("op_1972_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_1972_end_0 = const()[name = string("op_1972_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> var_1972_end_mask_0 = const()[name = string("op_1972_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 512]> var_1972_cast_fp16 = slice_by_index(begin = var_1972_begin_0, end = var_1972_end_0, end_mask = var_1972_end_mask_0, x = c_7_cast_fp16)[name = string("op_1972_cast_fp16")];
            tensor<int32, [4]> var_1977 = const()[name = string("op_1977"), val = tensor<int32, [4]>([3, 16, 4, 128])];
            tensor<fp16, [3, 16, 4, 128]> var_1978_cast_fp16 = reshape(shape = var_1977, x = var_1972_cast_fp16)[name = string("op_1978_cast_fp16")];
            tensor<int32, [4]> var_1983 = const()[name = string("op_1983"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1985_to_fp16 = const()[name = string("op_1985_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 16, 2, 128]> q_29_cast_fp16 = mul(x = q_27_cast_fp16, y = var_1985_to_fp16)[name = string("q_29_cast_fp16")];
            fp16 var_1987_to_fp16 = const()[name = string("op_1987_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 16, 2, 128]> k_31_cast_fp16 = mul(x = k_29_cast_fp16, y = var_1987_to_fp16)[name = string("k_31_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_1989_cast_fp16 = mul(x = q_29_cast_fp16, y = q_29_cast_fp16)[name = string("op_1989_cast_fp16")];
            tensor<int32, [1]> var_1994_axes_0 = const()[name = string("op_1994_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1994_keep_dims_0 = const()[name = string("op_1994_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 2, 1]> var_1994_cast_fp16 = reduce_mean(axes = var_1994_axes_0, keep_dims = var_1994_keep_dims_0, x = var_1989_cast_fp16)[name = string("op_1994_cast_fp16")];
            fp16 var_1995_promoted_to_fp16 = const()[name = string("op_1995_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 16, 2, 1]> var_1996_cast_fp16 = mul(x = var_1994_cast_fp16, y = var_1995_promoted_to_fp16)[name = string("op_1996_cast_fp16")];
            fp16 var_1998_to_fp16 = const()[name = string("op_1998_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 16, 2, 1]> var_1999_cast_fp16 = add(x = var_1996_cast_fp16, y = var_1998_to_fp16)[name = string("op_1999_cast_fp16")];
            fp32 var_2000_epsilon_0 = const()[name = string("op_2000_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 2, 1]> var_2000_cast_fp16 = rsqrt(epsilon = var_2000_epsilon_0, x = var_1999_cast_fp16)[name = string("op_2000_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_2001_cast_fp16 = mul(x = q_29_cast_fp16, y = var_2000_cast_fp16)[name = string("op_2001_cast_fp16")];
            fp16 var_2002_to_fp16 = const()[name = string("op_2002_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 16, 2, 128]> q_31_cast_fp16 = mul(x = var_2001_cast_fp16, y = var_2002_to_fp16)[name = string("q_31_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_2004_cast_fp16 = mul(x = k_31_cast_fp16, y = k_31_cast_fp16)[name = string("op_2004_cast_fp16")];
            tensor<int32, [1]> var_2009_axes_0 = const()[name = string("op_2009_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2009_keep_dims_0 = const()[name = string("op_2009_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 2, 1]> var_2009_cast_fp16 = reduce_mean(axes = var_2009_axes_0, keep_dims = var_2009_keep_dims_0, x = var_2004_cast_fp16)[name = string("op_2009_cast_fp16")];
            fp16 var_2010_promoted_to_fp16 = const()[name = string("op_2010_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 16, 2, 1]> var_2011_cast_fp16 = mul(x = var_2009_cast_fp16, y = var_2010_promoted_to_fp16)[name = string("op_2011_cast_fp16")];
            fp16 var_2013_to_fp16 = const()[name = string("op_2013_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 16, 2, 1]> var_2014_cast_fp16 = add(x = var_2011_cast_fp16, y = var_2013_to_fp16)[name = string("op_2014_cast_fp16")];
            fp32 var_2015_epsilon_0 = const()[name = string("op_2015_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 2, 1]> var_2015_cast_fp16 = rsqrt(epsilon = var_2015_epsilon_0, x = var_2014_cast_fp16)[name = string("op_2015_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> k_33_cast_fp16 = mul(x = k_31_cast_fp16, y = var_2015_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<int32, [4]> var_2021 = const()[name = string("op_2021"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2034_axes_0 = const()[name = string("op_2034_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 16, 128]> var_2022_cast_fp16 = transpose(perm = var_2021, x = q_31_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [3, 2, 1, 16, 128]> var_2034_cast_fp16 = expand_dims(axes = var_2034_axes_0, x = var_2022_cast_fp16)[name = string("op_2034_cast_fp16")];
            tensor<int32, [5]> var_2042_reps_0 = const()[name = string("op_2042_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 16, 128]> var_2042_cast_fp16 = tile(reps = var_2042_reps_0, x = var_2034_cast_fp16)[name = string("op_2042_cast_fp16")];
            tensor<int32, [4]> var_2047 = const()[name = string("op_2047"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<fp16, [3, 4, 16, 128]> q_33_cast_fp16 = reshape(shape = var_2047, x = var_2042_cast_fp16)[name = string("q_33_cast_fp16")];
            tensor<int32, [4]> var_2053 = const()[name = string("op_2053"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2066_axes_0 = const()[name = string("op_2066_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 16, 128]> var_2054_cast_fp16 = transpose(perm = var_2053, x = k_33_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [3, 2, 1, 16, 128]> var_2066_cast_fp16 = expand_dims(axes = var_2066_axes_0, x = var_2054_cast_fp16)[name = string("op_2066_cast_fp16")];
            tensor<int32, [5]> var_2074_reps_0 = const()[name = string("op_2074_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 16, 128]> var_2074_cast_fp16 = tile(reps = var_2074_reps_0, x = var_2066_cast_fp16)[name = string("op_2074_cast_fp16")];
            tensor<int32, [4]> var_2079 = const()[name = string("op_2079"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<fp16, [3, 4, 16, 128]> k_35_cast_fp16 = reshape(shape = var_2079, x = var_2074_cast_fp16)[name = string("k_35_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_2081_cast_fp16 = sigmoid(x = b_7_cast_fp16)[name = string("op_2081_cast_fp16")];
            tensor<int32, [3]> var_2085 = const()[name = string("op_2085"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_4_dt_bias_to_fp16 = const()[name = string("layers_4_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 16, 4]> input_63_cast_fp16 = add(x = a_7_cast_fp16, y = layers_4_dt_bias_to_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_2089_cast_fp16 = relu(x = input_63_cast_fp16)[name = string("op_2089_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_2090_cast_fp16 = abs(x = input_63_cast_fp16)[name = string("op_2090_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 16, 4]> var_2091_cast_fp16 = mul(x = var_2090_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_2091_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_2092_cast_fp16 = exp(x = var_2091_cast_fp16)[name = string("op_2092_cast_fp16")];
            fp16 var_2094_to_fp16 = const()[name = string("op_2094_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 16, 4]> var_2095_cast_fp16 = add(x = var_2092_cast_fp16, y = var_2094_to_fp16)[name = string("op_2095_cast_fp16")];
            fp32 var_2096_epsilon_0 = const()[name = string("op_2096_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 16, 4]> var_2096_cast_fp16 = log(epsilon = var_2096_epsilon_0, x = var_2095_cast_fp16)[name = string("op_2096_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_2098_cast_fp16 = add(x = var_2089_cast_fp16, y = var_2096_cast_fp16)[name = string("op_2098_cast_fp16")];
            tensor<fp16, [4]> layers_4_negA_to_fp16 = const()[name = string("layers_4_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.348p+3, -0x1.734p+2, -0x1.5a8p-1, -0x1.264p+3])];
            tensor<fp16, [3, 16, 4]> var_2099_cast_fp16 = mul(x = layers_4_negA_to_fp16, y = var_2098_cast_fp16)[name = string("op_2099_cast_fp16")];
            tensor<int32, [3]> var_2103 = const()[name = string("op_2103"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_2106_axes_0 = const()[name = string("op_2106_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16]> beta_7_cast_fp16 = transpose(perm = var_2085, x = var_2081_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [3, 4, 16, 1]> var_2106_cast_fp16 = expand_dims(axes = var_2106_axes_0, x = beta_7_cast_fp16)[name = string("op_2106_cast_fp16")];
            fp16 var_2107_to_fp16 = const()[name = string("op_2107_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 16, 1]> var_2108_cast_fp16 = mul(x = var_2106_cast_fp16, y = var_2107_to_fp16)[name = string("op_2108_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> v_11_cast_fp16 = transpose(perm = var_1983, x = var_1978_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [3, 4, 16, 128]> vb_7_cast_fp16 = mul(x = v_11_cast_fp16, y = var_2108_cast_fp16)[name = string("vb_7_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> kb_7_cast_fp16 = mul(x = k_35_cast_fp16, y = var_2106_cast_fp16)[name = string("kb_7_cast_fp16")];
            tensor<int32, [2]> var_2115 = const()[name = string("op_2115"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [3, 4, 16]> gdec_7_cast_fp16 = transpose(perm = var_2103, x = var_2099_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [12, 16]> var_2116_cast_fp16 = reshape(shape = var_2115, x = gdec_7_cast_fp16)[name = string("op_2116_cast_fp16")];
            tensor<fp16, [12, 256]> var_2117_cast_fp16 = linear(bias = var_554_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2116_cast_fp16)[name = string("op_2117_cast_fp16")];
            tensor<int32, [4]> var_2122 = const()[name = string("op_2122"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_7_cast_fp16 = reshape(shape = var_2122, x = var_2117_cast_fp16)[name = string("P_7_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2124_cast_fp16 = exp(x = P_7_cast_fp16)[name = string("op_2124_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_7_cast_fp16 = mul(x = var_2124_cast_fp16, y = tril_incl_to_fp16)[name = string("D_7_cast_fp16")];
            bool kk_7_transpose_x_1 = const()[name = string("kk_7_transpose_x_1"), val = bool(false)];
            bool kk_7_transpose_y_1 = const()[name = string("kk_7_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_7_cast_fp16 = matmul(transpose_x = kk_7_transpose_x_1, transpose_y = kk_7_transpose_y_1, x = kb_7_cast_fp16, y = k_35_cast_fp16)[name = string("kk_7_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_2130_cast_fp16 = mul(x = kk_7_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_2130_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2131_cast_fp16 = mul(x = var_2130_cast_fp16, y = D_7_cast_fp16)[name = string("op_2131_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_7_cast_fp16 = mul(x = var_2131_cast_fp16, y = tril_strict_to_fp16)[name = string("N_7_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2133_cast_fp16 = mul(x = D_7_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2133_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_7_cast_fp16 = mul(x = kk_7_cast_fp16, y = var_2133_cast_fp16)[name = string("Na_7_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_25_cast_fp16 = add(x = eye_to_fp16, y = N_7_cast_fp16)[name = string("Minv_25_cast_fp16")];
            bool Np_19_transpose_x_0 = const()[name = string("Np_19_transpose_x_0"), val = bool(false)];
            bool Np_19_transpose_y_0 = const()[name = string("Np_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_19_cast_fp16 = matmul(transpose_x = Np_19_transpose_x_0, transpose_y = Np_19_transpose_y_0, x = N_7_cast_fp16, y = Na_7_cast_fp16)[name = string("Np_19_cast_fp16")];
            bool Npa_13_transpose_x_0 = const()[name = string("Npa_13_transpose_x_0"), val = bool(false)];
            bool Npa_13_transpose_y_0 = const()[name = string("Npa_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_13_cast_fp16 = matmul(transpose_x = Npa_13_transpose_x_0, transpose_y = Npa_13_transpose_y_0, x = Na_7_cast_fp16, y = N_7_cast_fp16)[name = string("Npa_13_cast_fp16")];
            bool var_2139_transpose_x_0 = const()[name = string("op_2139_transpose_x_0"), val = bool(false)];
            bool var_2139_transpose_y_0 = const()[name = string("op_2139_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2139_cast_fp16 = matmul(transpose_x = var_2139_transpose_x_0, transpose_y = var_2139_transpose_y_0, x = Minv_25_cast_fp16, y = Np_19_cast_fp16)[name = string("op_2139_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_27_cast_fp16 = add(x = Minv_25_cast_fp16, y = var_2139_cast_fp16)[name = string("Minv_27_cast_fp16")];
            bool Np_21_transpose_x_0 = const()[name = string("Np_21_transpose_x_0"), val = bool(false)];
            bool Np_21_transpose_y_0 = const()[name = string("Np_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_21_cast_fp16 = matmul(transpose_x = Np_21_transpose_x_0, transpose_y = Np_21_transpose_y_0, x = Np_19_cast_fp16, y = Npa_13_cast_fp16)[name = string("Np_21_cast_fp16")];
            bool Npa_15_transpose_x_0 = const()[name = string("Npa_15_transpose_x_0"), val = bool(false)];
            bool Npa_15_transpose_y_0 = const()[name = string("Npa_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_15_cast_fp16 = matmul(transpose_x = Npa_15_transpose_x_0, transpose_y = Npa_15_transpose_y_0, x = Npa_13_cast_fp16, y = Np_19_cast_fp16)[name = string("Npa_15_cast_fp16")];
            bool var_2144_transpose_x_0 = const()[name = string("op_2144_transpose_x_0"), val = bool(false)];
            bool var_2144_transpose_y_0 = const()[name = string("op_2144_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2144_cast_fp16 = matmul(transpose_x = var_2144_transpose_x_0, transpose_y = var_2144_transpose_y_0, x = Minv_27_cast_fp16, y = Np_21_cast_fp16)[name = string("op_2144_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_29_cast_fp16 = add(x = Minv_27_cast_fp16, y = var_2144_cast_fp16)[name = string("Minv_29_cast_fp16")];
            bool Np_23_transpose_x_0 = const()[name = string("Np_23_transpose_x_0"), val = bool(false)];
            bool Np_23_transpose_y_0 = const()[name = string("Np_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_23_cast_fp16 = matmul(transpose_x = Np_23_transpose_x_0, transpose_y = Np_23_transpose_y_0, x = Np_21_cast_fp16, y = Npa_15_cast_fp16)[name = string("Np_23_cast_fp16")];
            bool var_2148_transpose_x_0 = const()[name = string("op_2148_transpose_x_0"), val = bool(false)];
            bool var_2148_transpose_y_0 = const()[name = string("op_2148_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2148_cast_fp16 = matmul(transpose_x = var_2148_transpose_x_0, transpose_y = var_2148_transpose_y_0, x = Minv_29_cast_fp16, y = Np_23_cast_fp16)[name = string("op_2148_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_31_cast_fp16 = add(x = Minv_29_cast_fp16, y = var_2148_cast_fp16)[name = string("Minv_31_cast_fp16")];
            bool u_7_transpose_x_0 = const()[name = string("u_7_transpose_x_0"), val = bool(false)];
            bool u_7_transpose_y_0 = const()[name = string("u_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_7_cast_fp16 = matmul(transpose_x = u_7_transpose_x_0, transpose_y = u_7_transpose_y_0, x = Minv_31_cast_fp16, y = vb_7_cast_fp16)[name = string("u_7_cast_fp16")];
            bool var_2155_transpose_x_1 = const()[name = string("op_2155_transpose_x_1"), val = bool(false)];
            bool var_2155_transpose_y_1 = const()[name = string("op_2155_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_2155_cast_fp16 = matmul(transpose_x = var_2155_transpose_x_1, transpose_y = var_2155_transpose_y_1, x = q_33_cast_fp16, y = k_35_cast_fp16)[name = string("op_2155_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_7_cast_fp16 = mul(x = var_2155_cast_fp16, y = D_7_cast_fp16)[name = string("A_7_cast_fp16")];
            bool o_21_transpose_x_0 = const()[name = string("o_21_transpose_x_0"), val = bool(false)];
            bool o_21_transpose_y_0 = const()[name = string("o_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> o_21_cast_fp16 = matmul(transpose_x = o_21_transpose_x_0, transpose_y = o_21_transpose_y_0, x = A_7_cast_fp16, y = u_7_cast_fp16)[name = string("o_21_cast_fp16")];
            tensor<int32, [4]> var_2162 = const()[name = string("op_2162"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2168 = const()[name = string("op_2168"), val = tensor<int32, [4]>([3, 16, 4, 128])];
            tensor<fp16, [3, 16, 4, 128]> input_65_cast_fp16 = reshape(shape = var_2168, x = z_7_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> o_23_cast_fp16 = transpose(perm = var_2162, x = o_21_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [3, 16, 4, 128]> var_2170_cast_fp16 = mul(x = o_23_cast_fp16, y = o_23_cast_fp16)[name = string("op_2170_cast_fp16")];
            tensor<int32, [1]> var_2175_axes_0 = const()[name = string("op_2175_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2175_keep_dims_0 = const()[name = string("op_2175_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 4, 1]> var_2175_cast_fp16 = reduce_mean(axes = var_2175_axes_0, keep_dims = var_2175_keep_dims_0, x = var_2170_cast_fp16)[name = string("op_2175_cast_fp16")];
            fp16 var_2177_to_fp16 = const()[name = string("op_2177_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 16, 4, 1]> var_2178_cast_fp16 = add(x = var_2175_cast_fp16, y = var_2177_to_fp16)[name = string("op_2178_cast_fp16")];
            fp32 var_2179_epsilon_0 = const()[name = string("op_2179_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 4, 1]> var_2179_cast_fp16 = rsqrt(epsilon = var_2179_epsilon_0, x = var_2178_cast_fp16)[name = string("op_2179_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> var_2180_cast_fp16 = mul(x = o_23_cast_fp16, y = var_2179_cast_fp16)[name = string("op_2180_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> var_2182_cast_fp16 = silu(x = input_65_cast_fp16)[name = string("op_2182_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> o_25_cast_fp16 = mul(x = var_2180_cast_fp16, y = var_2182_cast_fp16)[name = string("o_25_cast_fp16")];
            tensor<int32, [3]> var_2187 = const()[name = string("op_2187"), val = tensor<int32, [3]>([3, 16, 512])];
            tensor<fp16, [3, 16, 512]> input_67_cast_fp16 = reshape(shape = var_2187, x = o_25_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [640, 512]> layers_4_out_weight_to_fp16 = const()[name = string("layers_4_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72063040)))];
            tensor<fp16, [3, 16, 640]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_out_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [3, 16, 640]> x_49_cast_fp16 = add(x = x_45_cast_fp16, y = linear_17_cast_fp16)[name = string("x_49_cast_fp16")];
            fp16 var_2200_to_fp16 = const()[name = string("op_2200_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 16, 640]> x_51_cast_fp16 = mul(x = x_49_cast_fp16, y = var_2200_to_fp16)[name = string("x_51_cast_fp16")];
            fp16 var_2198_promoted_to_fp16 = const()[name = string("op_2198_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> var_2202_cast_fp16 = pow(x = x_51_cast_fp16, y = var_2198_promoted_to_fp16)[name = string("op_2202_cast_fp16")];
            tensor<int32, [1]> var_2204_axes_0 = const()[name = string("op_2204_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2204_keep_dims_0 = const()[name = string("op_2204_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 1]> var_2204_cast_fp16 = reduce_mean(axes = var_2204_axes_0, keep_dims = var_2204_keep_dims_0, x = var_2202_cast_fp16)[name = string("op_2204_cast_fp16")];
            fp16 var_2205_to_fp16 = const()[name = string("op_2205_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 16, 1]> var_2206_cast_fp16 = add(x = var_2204_cast_fp16, y = var_2205_to_fp16)[name = string("op_2206_cast_fp16")];
            fp32 var_2207_epsilon_0 = const()[name = string("op_2207_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 1]> var_2207_cast_fp16 = rsqrt(epsilon = var_2207_epsilon_0, x = var_2206_cast_fp16)[name = string("op_2207_cast_fp16")];
            tensor<fp16, [3, 16, 640]> var_2208_cast_fp16 = mul(x = x_51_cast_fp16, y = var_2207_cast_fp16)[name = string("op_2208_cast_fp16")];
            tensor<fp16, [640]> layers_4_n2_w1_to_fp16 = const()[name = string("layers_4_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72718464)))];
            tensor<fp16, [3, 16, 640]> input_69_cast_fp16 = mul(x = var_2208_cast_fp16, y = layers_4_n2_w1_to_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_4_gate_up_weight_to_fp16 = const()[name = string("layers_4_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72719808)))];
            tensor<fp16, [3, 16, 5120]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_4_gate_up_weight_to_fp16, x = input_69_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<int32, [3]> input_71_begin_0 = const()[name = string("input_71_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_71_end_0 = const()[name = string("input_71_end_0"), val = tensor<int32, [3]>([3, 16, 2560])];
            tensor<bool, [3]> input_71_end_mask_0 = const()[name = string("input_71_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 2560]> input_71_cast_fp16 = slice_by_index(begin = input_71_begin_0, end = input_71_end_0, end_mask = input_71_end_mask_0, x = linear_18_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [3, 16, 2560]> var_2218_cast_fp16 = silu(x = input_71_cast_fp16)[name = string("op_2218_cast_fp16")];
            tensor<int32, [3]> var_2223_begin_0 = const()[name = string("op_2223_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_2223_end_0 = const()[name = string("op_2223_end_0"), val = tensor<int32, [3]>([3, 16, 5120])];
            tensor<bool, [3]> var_2223_end_mask_0 = const()[name = string("op_2223_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 2560]> var_2223_cast_fp16 = slice_by_index(begin = var_2223_begin_0, end = var_2223_end_0, end_mask = var_2223_end_mask_0, x = linear_18_cast_fp16)[name = string("op_2223_cast_fp16")];
            tensor<fp16, [3, 16, 2560]> input_73_cast_fp16 = mul(x = var_2218_cast_fp16, y = var_2223_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_4_down_weight_to_fp16 = const()[name = string("layers_4_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79273472)))];
            tensor<fp16, [3, 16, 640]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_down_weight_to_fp16, x = input_73_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [3, 16, 640]> x_53_cast_fp16 = add(x = x_49_cast_fp16, y = linear_19_cast_fp16)[name = string("x_53_cast_fp16")];
            fp16 var_2236_to_fp16 = const()[name = string("op_2236_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 16, 640]> x_55_cast_fp16 = mul(x = x_53_cast_fp16, y = var_2236_to_fp16)[name = string("x_55_cast_fp16")];
            fp16 var_2234_promoted_to_fp16 = const()[name = string("op_2234_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> var_2238_cast_fp16 = pow(x = x_55_cast_fp16, y = var_2234_promoted_to_fp16)[name = string("op_2238_cast_fp16")];
            tensor<int32, [1]> var_2240_axes_0 = const()[name = string("op_2240_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2240_keep_dims_0 = const()[name = string("op_2240_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 1]> var_2240_cast_fp16 = reduce_mean(axes = var_2240_axes_0, keep_dims = var_2240_keep_dims_0, x = var_2238_cast_fp16)[name = string("op_2240_cast_fp16")];
            fp16 var_2241_to_fp16 = const()[name = string("op_2241_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 16, 1]> var_2242_cast_fp16 = add(x = var_2240_cast_fp16, y = var_2241_to_fp16)[name = string("op_2242_cast_fp16")];
            fp32 var_2243_epsilon_0 = const()[name = string("op_2243_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 1]> var_2243_cast_fp16 = rsqrt(epsilon = var_2243_epsilon_0, x = var_2242_cast_fp16)[name = string("op_2243_cast_fp16")];
            tensor<fp16, [3, 16, 640]> var_2244_cast_fp16 = mul(x = x_55_cast_fp16, y = var_2243_cast_fp16)[name = string("op_2244_cast_fp16")];
            tensor<fp16, [640]> layers_5_n1_w1_to_fp16 = const()[name = string("layers_5_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82550336)))];
            tensor<fp16, [3, 16, 640]> input_75_cast_fp16 = mul(x = var_2244_cast_fp16, y = layers_5_n1_w1_to_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_5_qkvzba_weight_to_fp16 = const()[name = string("layers_5_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82551680)))];
            tensor<fp16, [3, 16, 1544]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_qkvzba_weight_to_fp16, x = input_75_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [3]> qkv_9_begin_0 = const()[name = string("qkv_9_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_9_end_0 = const()[name = string("qkv_9_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> qkv_9_end_mask_0 = const()[name = string("qkv_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 1024]> qkv_9_cast_fp16 = slice_by_index(begin = qkv_9_begin_0, end = qkv_9_end_0, end_mask = qkv_9_end_mask_0, x = linear_20_cast_fp16)[name = string("qkv_9_cast_fp16")];
            tensor<int32, [3]> z_9_begin_0 = const()[name = string("z_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_9_end_0 = const()[name = string("z_9_end_0"), val = tensor<int32, [3]>([3, 16, 1536])];
            tensor<bool, [3]> z_9_end_mask_0 = const()[name = string("z_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 512]> z_9_cast_fp16 = slice_by_index(begin = z_9_begin_0, end = z_9_end_0, end_mask = z_9_end_mask_0, x = linear_20_cast_fp16)[name = string("z_9_cast_fp16")];
            tensor<int32, [3]> b_9_begin_0 = const()[name = string("b_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_9_end_0 = const()[name = string("b_9_end_0"), val = tensor<int32, [3]>([3, 16, 1540])];
            tensor<bool, [3]> b_9_end_mask_0 = const()[name = string("b_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 4]> b_9_cast_fp16 = slice_by_index(begin = b_9_begin_0, end = b_9_end_0, end_mask = b_9_end_mask_0, x = linear_20_cast_fp16)[name = string("b_9_cast_fp16")];
            tensor<int32, [3]> a_9_begin_0 = const()[name = string("a_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_9_end_0 = const()[name = string("a_9_end_0"), val = tensor<int32, [3]>([3, 16, 1])];
            tensor<bool, [3]> a_9_end_mask_0 = const()[name = string("a_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 4]> a_9_cast_fp16 = slice_by_index(begin = a_9_begin_0, end = a_9_end_0, end_mask = a_9_end_mask_0, x = linear_20_cast_fp16)[name = string("a_9_cast_fp16")];
            int32 var_2283 = const()[name = string("op_2283"), val = int32(1)];
            bool xp_9_interleave_0 = const()[name = string("xp_9_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 19, 1024]> xp_9_cast_fp16 = concat(axis = var_2283, interleave = xp_9_interleave_0, values = (zpad_to_fp16, qkv_9_cast_fp16))[name = string("xp_9_cast_fp16")];
            tensor<int32, [3]> var_2294_begin_0 = const()[name = string("op_2294_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_2294_end_0 = const()[name = string("op_2294_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_2294_end_mask_0 = const()[name = string("op_2294_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 1024]> var_2294_cast_fp16 = slice_by_index(begin = var_2294_begin_0, end = var_2294_end_0, end_mask = var_2294_end_mask_0, x = xp_9_cast_fp16)[name = string("op_2294_cast_fp16")];
            tensor<fp16, [1024]> var_2302_to_fp16 = const()[name = string("op_2302_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84528064)))];
            tensor<fp16, [3, 16, 1024]> var_2303_cast_fp16 = mul(x = var_2294_cast_fp16, y = var_2302_to_fp16)[name = string("op_2303_cast_fp16")];
            tensor<int32, [3]> var_2313_begin_0 = const()[name = string("op_2313_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_2313_end_0 = const()[name = string("op_2313_end_0"), val = tensor<int32, [3]>([3, 18, 1024])];
            tensor<bool, [3]> var_2313_end_mask_0 = const()[name = string("op_2313_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_2313_cast_fp16 = slice_by_index(begin = var_2313_begin_0, end = var_2313_end_0, end_mask = var_2313_end_mask_0, x = xp_9_cast_fp16)[name = string("op_2313_cast_fp16")];
            tensor<fp16, [1024]> var_2321_to_fp16 = const()[name = string("op_2321_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84530176)))];
            tensor<fp16, [3, 16, 1024]> var_2322_cast_fp16 = mul(x = var_2313_cast_fp16, y = var_2321_to_fp16)[name = string("op_2322_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> var_2324_cast_fp16 = add(x = var_2303_cast_fp16, y = var_2322_cast_fp16)[name = string("op_2324_cast_fp16")];
            tensor<int32, [3]> var_2334_begin_0 = const()[name = string("op_2334_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2334_end_0 = const()[name = string("op_2334_end_0"), val = tensor<int32, [3]>([3, 17, 1024])];
            tensor<bool, [3]> var_2334_end_mask_0 = const()[name = string("op_2334_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_2334_cast_fp16 = slice_by_index(begin = var_2334_begin_0, end = var_2334_end_0, end_mask = var_2334_end_mask_0, x = xp_9_cast_fp16)[name = string("op_2334_cast_fp16")];
            tensor<fp16, [1024]> var_2342_to_fp16 = const()[name = string("op_2342_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84532288)))];
            tensor<fp16, [3, 16, 1024]> var_2343_cast_fp16 = mul(x = var_2334_cast_fp16, y = var_2342_to_fp16)[name = string("op_2343_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> var_2345_cast_fp16 = add(x = var_2324_cast_fp16, y = var_2343_cast_fp16)[name = string("op_2345_cast_fp16")];
            tensor<int32, [3]> var_2355_begin_0 = const()[name = string("op_2355_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2355_end_0 = const()[name = string("op_2355_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> var_2355_end_mask_0 = const()[name = string("op_2355_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_2355_cast_fp16 = slice_by_index(begin = var_2355_begin_0, end = var_2355_end_0, end_mask = var_2355_end_mask_0, x = xp_9_cast_fp16)[name = string("op_2355_cast_fp16")];
            tensor<fp16, [1024]> var_2363_to_fp16 = const()[name = string("op_2363_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84534400)))];
            tensor<fp16, [3, 16, 1024]> var_2364_cast_fp16 = mul(x = var_2355_cast_fp16, y = var_2363_to_fp16)[name = string("op_2364_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> input_77_cast_fp16 = add(x = var_2345_cast_fp16, y = var_2364_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> c_9_cast_fp16 = silu(x = input_77_cast_fp16)[name = string("c_9_cast_fp16")];
            tensor<int32, [3]> var_2372_begin_0 = const()[name = string("op_2372_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2372_end_0 = const()[name = string("op_2372_end_0"), val = tensor<int32, [3]>([3, 16, 256])];
            tensor<bool, [3]> var_2372_end_mask_0 = const()[name = string("op_2372_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 256]> var_2372_cast_fp16 = slice_by_index(begin = var_2372_begin_0, end = var_2372_end_0, end_mask = var_2372_end_mask_0, x = c_9_cast_fp16)[name = string("op_2372_cast_fp16")];
            tensor<int32, [4]> var_2377 = const()[name = string("op_2377"), val = tensor<int32, [4]>([3, 16, 2, 128])];
            tensor<fp16, [3, 16, 2, 128]> q_35_cast_fp16 = reshape(shape = var_2377, x = var_2372_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<int32, [3]> var_2383_begin_0 = const()[name = string("op_2383_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_2383_end_0 = const()[name = string("op_2383_end_0"), val = tensor<int32, [3]>([3, 16, 512])];
            tensor<bool, [3]> var_2383_end_mask_0 = const()[name = string("op_2383_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 256]> var_2383_cast_fp16 = slice_by_index(begin = var_2383_begin_0, end = var_2383_end_0, end_mask = var_2383_end_mask_0, x = c_9_cast_fp16)[name = string("op_2383_cast_fp16")];
            tensor<int32, [4]> var_2388 = const()[name = string("op_2388"), val = tensor<int32, [4]>([3, 16, 2, 128])];
            tensor<fp16, [3, 16, 2, 128]> k_37_cast_fp16 = reshape(shape = var_2388, x = var_2383_cast_fp16)[name = string("k_37_cast_fp16")];
            tensor<int32, [3]> var_2394_begin_0 = const()[name = string("op_2394_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_2394_end_0 = const()[name = string("op_2394_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> var_2394_end_mask_0 = const()[name = string("op_2394_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 512]> var_2394_cast_fp16 = slice_by_index(begin = var_2394_begin_0, end = var_2394_end_0, end_mask = var_2394_end_mask_0, x = c_9_cast_fp16)[name = string("op_2394_cast_fp16")];
            tensor<int32, [4]> var_2399 = const()[name = string("op_2399"), val = tensor<int32, [4]>([3, 16, 4, 128])];
            tensor<fp16, [3, 16, 4, 128]> var_2400_cast_fp16 = reshape(shape = var_2399, x = var_2394_cast_fp16)[name = string("op_2400_cast_fp16")];
            tensor<int32, [4]> var_2405 = const()[name = string("op_2405"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2407_to_fp16 = const()[name = string("op_2407_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 16, 2, 128]> q_37_cast_fp16 = mul(x = q_35_cast_fp16, y = var_2407_to_fp16)[name = string("q_37_cast_fp16")];
            fp16 var_2409_to_fp16 = const()[name = string("op_2409_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 16, 2, 128]> k_39_cast_fp16 = mul(x = k_37_cast_fp16, y = var_2409_to_fp16)[name = string("k_39_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_2411_cast_fp16 = mul(x = q_37_cast_fp16, y = q_37_cast_fp16)[name = string("op_2411_cast_fp16")];
            tensor<int32, [1]> var_2416_axes_0 = const()[name = string("op_2416_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2416_keep_dims_0 = const()[name = string("op_2416_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 2, 1]> var_2416_cast_fp16 = reduce_mean(axes = var_2416_axes_0, keep_dims = var_2416_keep_dims_0, x = var_2411_cast_fp16)[name = string("op_2416_cast_fp16")];
            fp16 var_2417_promoted_to_fp16 = const()[name = string("op_2417_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 16, 2, 1]> var_2418_cast_fp16 = mul(x = var_2416_cast_fp16, y = var_2417_promoted_to_fp16)[name = string("op_2418_cast_fp16")];
            fp16 var_2420_to_fp16 = const()[name = string("op_2420_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 16, 2, 1]> var_2421_cast_fp16 = add(x = var_2418_cast_fp16, y = var_2420_to_fp16)[name = string("op_2421_cast_fp16")];
            fp32 var_2422_epsilon_0 = const()[name = string("op_2422_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 2, 1]> var_2422_cast_fp16 = rsqrt(epsilon = var_2422_epsilon_0, x = var_2421_cast_fp16)[name = string("op_2422_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_2423_cast_fp16 = mul(x = q_37_cast_fp16, y = var_2422_cast_fp16)[name = string("op_2423_cast_fp16")];
            fp16 var_2424_to_fp16 = const()[name = string("op_2424_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 16, 2, 128]> q_39_cast_fp16 = mul(x = var_2423_cast_fp16, y = var_2424_to_fp16)[name = string("q_39_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_2426_cast_fp16 = mul(x = k_39_cast_fp16, y = k_39_cast_fp16)[name = string("op_2426_cast_fp16")];
            tensor<int32, [1]> var_2431_axes_0 = const()[name = string("op_2431_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2431_keep_dims_0 = const()[name = string("op_2431_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 2, 1]> var_2431_cast_fp16 = reduce_mean(axes = var_2431_axes_0, keep_dims = var_2431_keep_dims_0, x = var_2426_cast_fp16)[name = string("op_2431_cast_fp16")];
            fp16 var_2432_promoted_to_fp16 = const()[name = string("op_2432_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 16, 2, 1]> var_2433_cast_fp16 = mul(x = var_2431_cast_fp16, y = var_2432_promoted_to_fp16)[name = string("op_2433_cast_fp16")];
            fp16 var_2435_to_fp16 = const()[name = string("op_2435_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 16, 2, 1]> var_2436_cast_fp16 = add(x = var_2433_cast_fp16, y = var_2435_to_fp16)[name = string("op_2436_cast_fp16")];
            fp32 var_2437_epsilon_0 = const()[name = string("op_2437_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 2, 1]> var_2437_cast_fp16 = rsqrt(epsilon = var_2437_epsilon_0, x = var_2436_cast_fp16)[name = string("op_2437_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> k_41_cast_fp16 = mul(x = k_39_cast_fp16, y = var_2437_cast_fp16)[name = string("k_41_cast_fp16")];
            tensor<int32, [4]> var_2443 = const()[name = string("op_2443"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2456_axes_0 = const()[name = string("op_2456_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 16, 128]> var_2444_cast_fp16 = transpose(perm = var_2443, x = q_39_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [3, 2, 1, 16, 128]> var_2456_cast_fp16 = expand_dims(axes = var_2456_axes_0, x = var_2444_cast_fp16)[name = string("op_2456_cast_fp16")];
            tensor<int32, [5]> var_2464_reps_0 = const()[name = string("op_2464_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 16, 128]> var_2464_cast_fp16 = tile(reps = var_2464_reps_0, x = var_2456_cast_fp16)[name = string("op_2464_cast_fp16")];
            tensor<int32, [4]> var_2469 = const()[name = string("op_2469"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<fp16, [3, 4, 16, 128]> q_41_cast_fp16 = reshape(shape = var_2469, x = var_2464_cast_fp16)[name = string("q_41_cast_fp16")];
            tensor<int32, [4]> var_2475 = const()[name = string("op_2475"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2488_axes_0 = const()[name = string("op_2488_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 16, 128]> var_2476_cast_fp16 = transpose(perm = var_2475, x = k_41_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [3, 2, 1, 16, 128]> var_2488_cast_fp16 = expand_dims(axes = var_2488_axes_0, x = var_2476_cast_fp16)[name = string("op_2488_cast_fp16")];
            tensor<int32, [5]> var_2496_reps_0 = const()[name = string("op_2496_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 16, 128]> var_2496_cast_fp16 = tile(reps = var_2496_reps_0, x = var_2488_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<int32, [4]> var_2501 = const()[name = string("op_2501"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<fp16, [3, 4, 16, 128]> k_43_cast_fp16 = reshape(shape = var_2501, x = var_2496_cast_fp16)[name = string("k_43_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_2503_cast_fp16 = sigmoid(x = b_9_cast_fp16)[name = string("op_2503_cast_fp16")];
            tensor<int32, [3]> var_2507 = const()[name = string("op_2507"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_5_dt_bias_to_fp16 = const()[name = string("layers_5_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 16, 4]> input_79_cast_fp16 = add(x = a_9_cast_fp16, y = layers_5_dt_bias_to_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_2511_cast_fp16 = relu(x = input_79_cast_fp16)[name = string("op_2511_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_2512_cast_fp16 = abs(x = input_79_cast_fp16)[name = string("op_2512_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 16, 4]> var_2513_cast_fp16 = mul(x = var_2512_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_2513_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_2514_cast_fp16 = exp(x = var_2513_cast_fp16)[name = string("op_2514_cast_fp16")];
            fp16 var_2516_to_fp16 = const()[name = string("op_2516_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 16, 4]> var_2517_cast_fp16 = add(x = var_2514_cast_fp16, y = var_2516_to_fp16)[name = string("op_2517_cast_fp16")];
            fp32 var_2518_epsilon_0 = const()[name = string("op_2518_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 16, 4]> var_2518_cast_fp16 = log(epsilon = var_2518_epsilon_0, x = var_2517_cast_fp16)[name = string("op_2518_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_2520_cast_fp16 = add(x = var_2511_cast_fp16, y = var_2518_cast_fp16)[name = string("op_2520_cast_fp16")];
            tensor<fp16, [4]> layers_5_negA_to_fp16 = const()[name = string("layers_5_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.2f8p+3, -0x1.ae8p+2, -0x1.548p+2, -0x1.63p+3])];
            tensor<fp16, [3, 16, 4]> var_2521_cast_fp16 = mul(x = layers_5_negA_to_fp16, y = var_2520_cast_fp16)[name = string("op_2521_cast_fp16")];
            tensor<int32, [3]> var_2525 = const()[name = string("op_2525"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_2528_axes_0 = const()[name = string("op_2528_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16]> beta_9_cast_fp16 = transpose(perm = var_2507, x = var_2503_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [3, 4, 16, 1]> var_2528_cast_fp16 = expand_dims(axes = var_2528_axes_0, x = beta_9_cast_fp16)[name = string("op_2528_cast_fp16")];
            fp16 var_2529_to_fp16 = const()[name = string("op_2529_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 16, 1]> var_2530_cast_fp16 = mul(x = var_2528_cast_fp16, y = var_2529_to_fp16)[name = string("op_2530_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> v_13_cast_fp16 = transpose(perm = var_2405, x = var_2400_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [3, 4, 16, 128]> vb_9_cast_fp16 = mul(x = v_13_cast_fp16, y = var_2530_cast_fp16)[name = string("vb_9_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> kb_9_cast_fp16 = mul(x = k_43_cast_fp16, y = var_2528_cast_fp16)[name = string("kb_9_cast_fp16")];
            tensor<int32, [2]> var_2537 = const()[name = string("op_2537"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [3, 4, 16]> gdec_9_cast_fp16 = transpose(perm = var_2525, x = var_2521_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [12, 16]> var_2538_cast_fp16 = reshape(shape = var_2537, x = gdec_9_cast_fp16)[name = string("op_2538_cast_fp16")];
            tensor<fp16, [12, 256]> var_2539_cast_fp16 = linear(bias = var_554_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2538_cast_fp16)[name = string("op_2539_cast_fp16")];
            tensor<int32, [4]> var_2544 = const()[name = string("op_2544"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_9_cast_fp16 = reshape(shape = var_2544, x = var_2539_cast_fp16)[name = string("P_9_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2546_cast_fp16 = exp(x = P_9_cast_fp16)[name = string("op_2546_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_9_cast_fp16 = mul(x = var_2546_cast_fp16, y = tril_incl_to_fp16)[name = string("D_9_cast_fp16")];
            bool kk_9_transpose_x_1 = const()[name = string("kk_9_transpose_x_1"), val = bool(false)];
            bool kk_9_transpose_y_1 = const()[name = string("kk_9_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_9_cast_fp16 = matmul(transpose_x = kk_9_transpose_x_1, transpose_y = kk_9_transpose_y_1, x = kb_9_cast_fp16, y = k_43_cast_fp16)[name = string("kk_9_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_2552_cast_fp16 = mul(x = kk_9_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_2552_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2553_cast_fp16 = mul(x = var_2552_cast_fp16, y = D_9_cast_fp16)[name = string("op_2553_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_9_cast_fp16 = mul(x = var_2553_cast_fp16, y = tril_strict_to_fp16)[name = string("N_9_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2555_cast_fp16 = mul(x = D_9_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2555_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_9_cast_fp16 = mul(x = kk_9_cast_fp16, y = var_2555_cast_fp16)[name = string("Na_9_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_33_cast_fp16 = add(x = eye_to_fp16, y = N_9_cast_fp16)[name = string("Minv_33_cast_fp16")];
            bool Np_25_transpose_x_0 = const()[name = string("Np_25_transpose_x_0"), val = bool(false)];
            bool Np_25_transpose_y_0 = const()[name = string("Np_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_25_cast_fp16 = matmul(transpose_x = Np_25_transpose_x_0, transpose_y = Np_25_transpose_y_0, x = N_9_cast_fp16, y = Na_9_cast_fp16)[name = string("Np_25_cast_fp16")];
            bool Npa_17_transpose_x_0 = const()[name = string("Npa_17_transpose_x_0"), val = bool(false)];
            bool Npa_17_transpose_y_0 = const()[name = string("Npa_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_17_cast_fp16 = matmul(transpose_x = Npa_17_transpose_x_0, transpose_y = Npa_17_transpose_y_0, x = Na_9_cast_fp16, y = N_9_cast_fp16)[name = string("Npa_17_cast_fp16")];
            bool var_2561_transpose_x_0 = const()[name = string("op_2561_transpose_x_0"), val = bool(false)];
            bool var_2561_transpose_y_0 = const()[name = string("op_2561_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2561_cast_fp16 = matmul(transpose_x = var_2561_transpose_x_0, transpose_y = var_2561_transpose_y_0, x = Minv_33_cast_fp16, y = Np_25_cast_fp16)[name = string("op_2561_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_35_cast_fp16 = add(x = Minv_33_cast_fp16, y = var_2561_cast_fp16)[name = string("Minv_35_cast_fp16")];
            bool Np_27_transpose_x_0 = const()[name = string("Np_27_transpose_x_0"), val = bool(false)];
            bool Np_27_transpose_y_0 = const()[name = string("Np_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_27_cast_fp16 = matmul(transpose_x = Np_27_transpose_x_0, transpose_y = Np_27_transpose_y_0, x = Np_25_cast_fp16, y = Npa_17_cast_fp16)[name = string("Np_27_cast_fp16")];
            bool Npa_19_transpose_x_0 = const()[name = string("Npa_19_transpose_x_0"), val = bool(false)];
            bool Npa_19_transpose_y_0 = const()[name = string("Npa_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_19_cast_fp16 = matmul(transpose_x = Npa_19_transpose_x_0, transpose_y = Npa_19_transpose_y_0, x = Npa_17_cast_fp16, y = Np_25_cast_fp16)[name = string("Npa_19_cast_fp16")];
            bool var_2566_transpose_x_0 = const()[name = string("op_2566_transpose_x_0"), val = bool(false)];
            bool var_2566_transpose_y_0 = const()[name = string("op_2566_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2566_cast_fp16 = matmul(transpose_x = var_2566_transpose_x_0, transpose_y = var_2566_transpose_y_0, x = Minv_35_cast_fp16, y = Np_27_cast_fp16)[name = string("op_2566_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_37_cast_fp16 = add(x = Minv_35_cast_fp16, y = var_2566_cast_fp16)[name = string("Minv_37_cast_fp16")];
            bool Np_29_transpose_x_0 = const()[name = string("Np_29_transpose_x_0"), val = bool(false)];
            bool Np_29_transpose_y_0 = const()[name = string("Np_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_29_cast_fp16 = matmul(transpose_x = Np_29_transpose_x_0, transpose_y = Np_29_transpose_y_0, x = Np_27_cast_fp16, y = Npa_19_cast_fp16)[name = string("Np_29_cast_fp16")];
            bool var_2570_transpose_x_0 = const()[name = string("op_2570_transpose_x_0"), val = bool(false)];
            bool var_2570_transpose_y_0 = const()[name = string("op_2570_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2570_cast_fp16 = matmul(transpose_x = var_2570_transpose_x_0, transpose_y = var_2570_transpose_y_0, x = Minv_37_cast_fp16, y = Np_29_cast_fp16)[name = string("op_2570_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_39_cast_fp16 = add(x = Minv_37_cast_fp16, y = var_2570_cast_fp16)[name = string("Minv_39_cast_fp16")];
            bool u_9_transpose_x_0 = const()[name = string("u_9_transpose_x_0"), val = bool(false)];
            bool u_9_transpose_y_0 = const()[name = string("u_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_9_cast_fp16 = matmul(transpose_x = u_9_transpose_x_0, transpose_y = u_9_transpose_y_0, x = Minv_39_cast_fp16, y = vb_9_cast_fp16)[name = string("u_9_cast_fp16")];
            bool var_2577_transpose_x_1 = const()[name = string("op_2577_transpose_x_1"), val = bool(false)];
            bool var_2577_transpose_y_1 = const()[name = string("op_2577_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_2577_cast_fp16 = matmul(transpose_x = var_2577_transpose_x_1, transpose_y = var_2577_transpose_y_1, x = q_41_cast_fp16, y = k_43_cast_fp16)[name = string("op_2577_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_9_cast_fp16 = mul(x = var_2577_cast_fp16, y = D_9_cast_fp16)[name = string("A_9_cast_fp16")];
            bool o_27_transpose_x_0 = const()[name = string("o_27_transpose_x_0"), val = bool(false)];
            bool o_27_transpose_y_0 = const()[name = string("o_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> o_27_cast_fp16 = matmul(transpose_x = o_27_transpose_x_0, transpose_y = o_27_transpose_y_0, x = A_9_cast_fp16, y = u_9_cast_fp16)[name = string("o_27_cast_fp16")];
            tensor<int32, [4]> var_2584 = const()[name = string("op_2584"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2590 = const()[name = string("op_2590"), val = tensor<int32, [4]>([3, 16, 4, 128])];
            tensor<fp16, [3, 16, 4, 128]> input_81_cast_fp16 = reshape(shape = var_2590, x = z_9_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> o_29_cast_fp16 = transpose(perm = var_2584, x = o_27_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [3, 16, 4, 128]> var_2592_cast_fp16 = mul(x = o_29_cast_fp16, y = o_29_cast_fp16)[name = string("op_2592_cast_fp16")];
            tensor<int32, [1]> var_2597_axes_0 = const()[name = string("op_2597_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2597_keep_dims_0 = const()[name = string("op_2597_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 4, 1]> var_2597_cast_fp16 = reduce_mean(axes = var_2597_axes_0, keep_dims = var_2597_keep_dims_0, x = var_2592_cast_fp16)[name = string("op_2597_cast_fp16")];
            fp16 var_2599_to_fp16 = const()[name = string("op_2599_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 16, 4, 1]> var_2600_cast_fp16 = add(x = var_2597_cast_fp16, y = var_2599_to_fp16)[name = string("op_2600_cast_fp16")];
            fp32 var_2601_epsilon_0 = const()[name = string("op_2601_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 4, 1]> var_2601_cast_fp16 = rsqrt(epsilon = var_2601_epsilon_0, x = var_2600_cast_fp16)[name = string("op_2601_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> var_2602_cast_fp16 = mul(x = o_29_cast_fp16, y = var_2601_cast_fp16)[name = string("op_2602_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> var_2604_cast_fp16 = silu(x = input_81_cast_fp16)[name = string("op_2604_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> o_31_cast_fp16 = mul(x = var_2602_cast_fp16, y = var_2604_cast_fp16)[name = string("o_31_cast_fp16")];
            tensor<int32, [3]> var_2609 = const()[name = string("op_2609"), val = tensor<int32, [3]>([3, 16, 512])];
            tensor<fp16, [3, 16, 512]> input_83_cast_fp16 = reshape(shape = var_2609, x = o_31_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [640, 512]> layers_5_out_weight_to_fp16 = const()[name = string("layers_5_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84536512)))];
            tensor<fp16, [3, 16, 640]> linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_out_weight_to_fp16, x = input_83_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [3, 16, 640]> x_57_cast_fp16 = add(x = x_53_cast_fp16, y = linear_21_cast_fp16)[name = string("x_57_cast_fp16")];
            fp16 var_2622_to_fp16 = const()[name = string("op_2622_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 16, 640]> x_59_cast_fp16 = mul(x = x_57_cast_fp16, y = var_2622_to_fp16)[name = string("x_59_cast_fp16")];
            fp16 var_2620_promoted_to_fp16 = const()[name = string("op_2620_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> var_2624_cast_fp16 = pow(x = x_59_cast_fp16, y = var_2620_promoted_to_fp16)[name = string("op_2624_cast_fp16")];
            tensor<int32, [1]> var_2626_axes_0 = const()[name = string("op_2626_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2626_keep_dims_0 = const()[name = string("op_2626_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 1]> var_2626_cast_fp16 = reduce_mean(axes = var_2626_axes_0, keep_dims = var_2626_keep_dims_0, x = var_2624_cast_fp16)[name = string("op_2626_cast_fp16")];
            fp16 var_2627_to_fp16 = const()[name = string("op_2627_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 16, 1]> var_2628_cast_fp16 = add(x = var_2626_cast_fp16, y = var_2627_to_fp16)[name = string("op_2628_cast_fp16")];
            fp32 var_2629_epsilon_0 = const()[name = string("op_2629_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 1]> var_2629_cast_fp16 = rsqrt(epsilon = var_2629_epsilon_0, x = var_2628_cast_fp16)[name = string("op_2629_cast_fp16")];
            tensor<fp16, [3, 16, 640]> var_2630_cast_fp16 = mul(x = x_59_cast_fp16, y = var_2629_cast_fp16)[name = string("op_2630_cast_fp16")];
            tensor<fp16, [640]> layers_5_n2_w1_to_fp16 = const()[name = string("layers_5_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85191936)))];
            tensor<fp16, [3, 16, 640]> input_85_cast_fp16 = mul(x = var_2630_cast_fp16, y = layers_5_n2_w1_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_5_gate_up_weight_to_fp16 = const()[name = string("layers_5_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85193280)))];
            tensor<fp16, [3, 16, 5120]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_5_gate_up_weight_to_fp16, x = input_85_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [3]> input_87_begin_0 = const()[name = string("input_87_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_87_end_0 = const()[name = string("input_87_end_0"), val = tensor<int32, [3]>([3, 16, 2560])];
            tensor<bool, [3]> input_87_end_mask_0 = const()[name = string("input_87_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 2560]> input_87_cast_fp16 = slice_by_index(begin = input_87_begin_0, end = input_87_end_0, end_mask = input_87_end_mask_0, x = linear_22_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [3, 16, 2560]> var_2640_cast_fp16 = silu(x = input_87_cast_fp16)[name = string("op_2640_cast_fp16")];
            tensor<int32, [3]> var_2645_begin_0 = const()[name = string("op_2645_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_2645_end_0 = const()[name = string("op_2645_end_0"), val = tensor<int32, [3]>([3, 16, 5120])];
            tensor<bool, [3]> var_2645_end_mask_0 = const()[name = string("op_2645_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 2560]> var_2645_cast_fp16 = slice_by_index(begin = var_2645_begin_0, end = var_2645_end_0, end_mask = var_2645_end_mask_0, x = linear_22_cast_fp16)[name = string("op_2645_cast_fp16")];
            tensor<fp16, [3, 16, 2560]> input_89_cast_fp16 = mul(x = var_2640_cast_fp16, y = var_2645_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_5_down_weight_to_fp16 = const()[name = string("layers_5_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91746944)))];
            tensor<fp16, [3, 16, 640]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_down_weight_to_fp16, x = input_89_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<fp16, [3, 16, 640]> x_61_cast_fp16 = add(x = x_57_cast_fp16, y = linear_23_cast_fp16)[name = string("x_61_cast_fp16")];
            fp16 var_2658_to_fp16 = const()[name = string("op_2658_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [3, 16, 640]> x_63_cast_fp16 = mul(x = x_61_cast_fp16, y = var_2658_to_fp16)[name = string("x_63_cast_fp16")];
            fp16 var_2656_promoted_to_fp16 = const()[name = string("op_2656_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> var_2660_cast_fp16 = pow(x = x_63_cast_fp16, y = var_2656_promoted_to_fp16)[name = string("op_2660_cast_fp16")];
            tensor<int32, [1]> var_2662_axes_0 = const()[name = string("op_2662_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2662_keep_dims_0 = const()[name = string("op_2662_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 1]> var_2662_cast_fp16 = reduce_mean(axes = var_2662_axes_0, keep_dims = var_2662_keep_dims_0, x = var_2660_cast_fp16)[name = string("op_2662_cast_fp16")];
            fp16 var_2663_to_fp16 = const()[name = string("op_2663_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 16, 1]> var_2664_cast_fp16 = add(x = var_2662_cast_fp16, y = var_2663_to_fp16)[name = string("op_2664_cast_fp16")];
            fp32 var_2665_epsilon_0 = const()[name = string("op_2665_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 1]> var_2665_cast_fp16 = rsqrt(epsilon = var_2665_epsilon_0, x = var_2664_cast_fp16)[name = string("op_2665_cast_fp16")];
            tensor<fp16, [3, 16, 640]> var_2666_cast_fp16 = mul(x = x_63_cast_fp16, y = var_2665_cast_fp16)[name = string("op_2666_cast_fp16")];
            tensor<fp16, [640]> layers_6_n1_w1_to_fp16 = const()[name = string("layers_6_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95023808)))];
            tensor<fp16, [3, 16, 640]> input_91_cast_fp16 = mul(x = var_2666_cast_fp16, y = layers_6_n1_w1_to_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_6_qkvzba_weight_to_fp16 = const()[name = string("layers_6_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95025152)))];
            tensor<fp16, [3, 16, 1544]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_qkvzba_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<int32, [3]> qkv_11_begin_0 = const()[name = string("qkv_11_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_11_end_0 = const()[name = string("qkv_11_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> qkv_11_end_mask_0 = const()[name = string("qkv_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 1024]> qkv_11_cast_fp16 = slice_by_index(begin = qkv_11_begin_0, end = qkv_11_end_0, end_mask = qkv_11_end_mask_0, x = linear_24_cast_fp16)[name = string("qkv_11_cast_fp16")];
            tensor<int32, [3]> z_begin_0 = const()[name = string("z_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_end_0 = const()[name = string("z_end_0"), val = tensor<int32, [3]>([3, 16, 1536])];
            tensor<bool, [3]> z_end_mask_0 = const()[name = string("z_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 512]> z_cast_fp16 = slice_by_index(begin = z_begin_0, end = z_end_0, end_mask = z_end_mask_0, x = linear_24_cast_fp16)[name = string("z_cast_fp16")];
            tensor<int32, [3]> b_begin_0 = const()[name = string("b_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_end_0 = const()[name = string("b_end_0"), val = tensor<int32, [3]>([3, 16, 1540])];
            tensor<bool, [3]> b_end_mask_0 = const()[name = string("b_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 4]> b_cast_fp16 = slice_by_index(begin = b_begin_0, end = b_end_0, end_mask = b_end_mask_0, x = linear_24_cast_fp16)[name = string("b_cast_fp16")];
            tensor<int32, [3]> a_begin_0 = const()[name = string("a_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_end_0 = const()[name = string("a_end_0"), val = tensor<int32, [3]>([3, 16, 1])];
            tensor<bool, [3]> a_end_mask_0 = const()[name = string("a_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 4]> a_cast_fp16 = slice_by_index(begin = a_begin_0, end = a_end_0, end_mask = a_end_mask_0, x = linear_24_cast_fp16)[name = string("a_cast_fp16")];
            int32 var_2705 = const()[name = string("op_2705"), val = int32(1)];
            bool xp_interleave_0 = const()[name = string("xp_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 19, 1024]> xp_cast_fp16 = concat(axis = var_2705, interleave = xp_interleave_0, values = (zpad_to_fp16, qkv_11_cast_fp16))[name = string("xp_cast_fp16")];
            tensor<int32, [3]> var_2716_begin_0 = const()[name = string("op_2716_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_2716_end_0 = const()[name = string("op_2716_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_2716_end_mask_0 = const()[name = string("op_2716_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 1024]> var_2716_cast_fp16 = slice_by_index(begin = var_2716_begin_0, end = var_2716_end_0, end_mask = var_2716_end_mask_0, x = xp_cast_fp16)[name = string("op_2716_cast_fp16")];
            tensor<fp16, [1024]> var_2724_to_fp16 = const()[name = string("op_2724_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97001536)))];
            tensor<fp16, [3, 16, 1024]> var_2725_cast_fp16 = mul(x = var_2716_cast_fp16, y = var_2724_to_fp16)[name = string("op_2725_cast_fp16")];
            tensor<int32, [3]> var_2735_begin_0 = const()[name = string("op_2735_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_2735_end_0 = const()[name = string("op_2735_end_0"), val = tensor<int32, [3]>([3, 18, 1024])];
            tensor<bool, [3]> var_2735_end_mask_0 = const()[name = string("op_2735_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_2735_cast_fp16 = slice_by_index(begin = var_2735_begin_0, end = var_2735_end_0, end_mask = var_2735_end_mask_0, x = xp_cast_fp16)[name = string("op_2735_cast_fp16")];
            tensor<fp16, [1024]> var_2743_to_fp16 = const()[name = string("op_2743_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97003648)))];
            tensor<fp16, [3, 16, 1024]> var_2744_cast_fp16 = mul(x = var_2735_cast_fp16, y = var_2743_to_fp16)[name = string("op_2744_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> var_2746_cast_fp16 = add(x = var_2725_cast_fp16, y = var_2744_cast_fp16)[name = string("op_2746_cast_fp16")];
            tensor<int32, [3]> var_2756_begin_0 = const()[name = string("op_2756_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2756_end_0 = const()[name = string("op_2756_end_0"), val = tensor<int32, [3]>([3, 17, 1024])];
            tensor<bool, [3]> var_2756_end_mask_0 = const()[name = string("op_2756_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_2756_cast_fp16 = slice_by_index(begin = var_2756_begin_0, end = var_2756_end_0, end_mask = var_2756_end_mask_0, x = xp_cast_fp16)[name = string("op_2756_cast_fp16")];
            tensor<fp16, [1024]> var_2764_to_fp16 = const()[name = string("op_2764_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97005760)))];
            tensor<fp16, [3, 16, 1024]> var_2765_cast_fp16 = mul(x = var_2756_cast_fp16, y = var_2764_to_fp16)[name = string("op_2765_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> var_2767_cast_fp16 = add(x = var_2746_cast_fp16, y = var_2765_cast_fp16)[name = string("op_2767_cast_fp16")];
            tensor<int32, [3]> var_2777_begin_0 = const()[name = string("op_2777_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2777_end_0 = const()[name = string("op_2777_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> var_2777_end_mask_0 = const()[name = string("op_2777_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 16, 1024]> var_2777_cast_fp16 = slice_by_index(begin = var_2777_begin_0, end = var_2777_end_0, end_mask = var_2777_end_mask_0, x = xp_cast_fp16)[name = string("op_2777_cast_fp16")];
            tensor<fp16, [1024]> var_2785_to_fp16 = const()[name = string("op_2785_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97007872)))];
            tensor<fp16, [3, 16, 1024]> var_2786_cast_fp16 = mul(x = var_2777_cast_fp16, y = var_2785_to_fp16)[name = string("op_2786_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> input_93_cast_fp16 = add(x = var_2767_cast_fp16, y = var_2786_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [3, 16, 1024]> c_cast_fp16 = silu(x = input_93_cast_fp16)[name = string("c_cast_fp16")];
            tensor<int32, [3]> var_2794_begin_0 = const()[name = string("op_2794_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2794_end_0 = const()[name = string("op_2794_end_0"), val = tensor<int32, [3]>([3, 16, 256])];
            tensor<bool, [3]> var_2794_end_mask_0 = const()[name = string("op_2794_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 256]> var_2794_cast_fp16 = slice_by_index(begin = var_2794_begin_0, end = var_2794_end_0, end_mask = var_2794_end_mask_0, x = c_cast_fp16)[name = string("op_2794_cast_fp16")];
            tensor<int32, [4]> var_2799 = const()[name = string("op_2799"), val = tensor<int32, [4]>([3, 16, 2, 128])];
            tensor<fp16, [3, 16, 2, 128]> q_43_cast_fp16 = reshape(shape = var_2799, x = var_2794_cast_fp16)[name = string("q_43_cast_fp16")];
            tensor<int32, [3]> var_2805_begin_0 = const()[name = string("op_2805_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_2805_end_0 = const()[name = string("op_2805_end_0"), val = tensor<int32, [3]>([3, 16, 512])];
            tensor<bool, [3]> var_2805_end_mask_0 = const()[name = string("op_2805_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 256]> var_2805_cast_fp16 = slice_by_index(begin = var_2805_begin_0, end = var_2805_end_0, end_mask = var_2805_end_mask_0, x = c_cast_fp16)[name = string("op_2805_cast_fp16")];
            tensor<int32, [4]> var_2810 = const()[name = string("op_2810"), val = tensor<int32, [4]>([3, 16, 2, 128])];
            tensor<fp16, [3, 16, 2, 128]> k_45_cast_fp16 = reshape(shape = var_2810, x = var_2805_cast_fp16)[name = string("k_45_cast_fp16")];
            tensor<int32, [3]> var_2816_begin_0 = const()[name = string("op_2816_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_2816_end_0 = const()[name = string("op_2816_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> var_2816_end_mask_0 = const()[name = string("op_2816_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 512]> var_2816_cast_fp16 = slice_by_index(begin = var_2816_begin_0, end = var_2816_end_0, end_mask = var_2816_end_mask_0, x = c_cast_fp16)[name = string("op_2816_cast_fp16")];
            tensor<int32, [4]> var_2821 = const()[name = string("op_2821"), val = tensor<int32, [4]>([3, 16, 4, 128])];
            tensor<fp16, [3, 16, 4, 128]> var_2822_cast_fp16 = reshape(shape = var_2821, x = var_2816_cast_fp16)[name = string("op_2822_cast_fp16")];
            tensor<int32, [4]> var_2827 = const()[name = string("op_2827"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2829_to_fp16 = const()[name = string("op_2829_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 16, 2, 128]> q_45_cast_fp16 = mul(x = q_43_cast_fp16, y = var_2829_to_fp16)[name = string("q_45_cast_fp16")];
            fp16 var_2831_to_fp16 = const()[name = string("op_2831_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 16, 2, 128]> k_47_cast_fp16 = mul(x = k_45_cast_fp16, y = var_2831_to_fp16)[name = string("k_47_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_2833_cast_fp16 = mul(x = q_45_cast_fp16, y = q_45_cast_fp16)[name = string("op_2833_cast_fp16")];
            tensor<int32, [1]> var_2838_axes_0 = const()[name = string("op_2838_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2838_keep_dims_0 = const()[name = string("op_2838_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 2, 1]> var_2838_cast_fp16 = reduce_mean(axes = var_2838_axes_0, keep_dims = var_2838_keep_dims_0, x = var_2833_cast_fp16)[name = string("op_2838_cast_fp16")];
            fp16 var_2839_promoted_to_fp16 = const()[name = string("op_2839_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 16, 2, 1]> var_2840_cast_fp16 = mul(x = var_2838_cast_fp16, y = var_2839_promoted_to_fp16)[name = string("op_2840_cast_fp16")];
            fp16 var_2842_to_fp16 = const()[name = string("op_2842_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 16, 2, 1]> var_2843_cast_fp16 = add(x = var_2840_cast_fp16, y = var_2842_to_fp16)[name = string("op_2843_cast_fp16")];
            fp32 var_2844_epsilon_0 = const()[name = string("op_2844_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 2, 1]> var_2844_cast_fp16 = rsqrt(epsilon = var_2844_epsilon_0, x = var_2843_cast_fp16)[name = string("op_2844_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_2845_cast_fp16 = mul(x = q_45_cast_fp16, y = var_2844_cast_fp16)[name = string("op_2845_cast_fp16")];
            fp16 var_2846_to_fp16 = const()[name = string("op_2846_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 16, 2, 128]> q_47_cast_fp16 = mul(x = var_2845_cast_fp16, y = var_2846_to_fp16)[name = string("q_47_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> var_2848_cast_fp16 = mul(x = k_47_cast_fp16, y = k_47_cast_fp16)[name = string("op_2848_cast_fp16")];
            tensor<int32, [1]> var_2853_axes_0 = const()[name = string("op_2853_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2853_keep_dims_0 = const()[name = string("op_2853_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 2, 1]> var_2853_cast_fp16 = reduce_mean(axes = var_2853_axes_0, keep_dims = var_2853_keep_dims_0, x = var_2848_cast_fp16)[name = string("op_2853_cast_fp16")];
            fp16 var_2854_promoted_to_fp16 = const()[name = string("op_2854_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 16, 2, 1]> var_2855_cast_fp16 = mul(x = var_2853_cast_fp16, y = var_2854_promoted_to_fp16)[name = string("op_2855_cast_fp16")];
            fp16 var_2857_to_fp16 = const()[name = string("op_2857_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 16, 2, 1]> var_2858_cast_fp16 = add(x = var_2855_cast_fp16, y = var_2857_to_fp16)[name = string("op_2858_cast_fp16")];
            fp32 var_2859_epsilon_0 = const()[name = string("op_2859_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 2, 1]> var_2859_cast_fp16 = rsqrt(epsilon = var_2859_epsilon_0, x = var_2858_cast_fp16)[name = string("op_2859_cast_fp16")];
            tensor<fp16, [3, 16, 2, 128]> k_49_cast_fp16 = mul(x = k_47_cast_fp16, y = var_2859_cast_fp16)[name = string("k_49_cast_fp16")];
            tensor<int32, [4]> var_2865 = const()[name = string("op_2865"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2878_axes_0 = const()[name = string("op_2878_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 16, 128]> var_2866_cast_fp16 = transpose(perm = var_2865, x = q_47_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [3, 2, 1, 16, 128]> var_2878_cast_fp16 = expand_dims(axes = var_2878_axes_0, x = var_2866_cast_fp16)[name = string("op_2878_cast_fp16")];
            tensor<int32, [5]> var_2886_reps_0 = const()[name = string("op_2886_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 16, 128]> var_2886_cast_fp16 = tile(reps = var_2886_reps_0, x = var_2878_cast_fp16)[name = string("op_2886_cast_fp16")];
            tensor<int32, [4]> var_2891 = const()[name = string("op_2891"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<fp16, [3, 4, 16, 128]> q_49_cast_fp16 = reshape(shape = var_2891, x = var_2886_cast_fp16)[name = string("q_49_cast_fp16")];
            tensor<int32, [4]> var_2897 = const()[name = string("op_2897"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2910_axes_0 = const()[name = string("op_2910_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 16, 128]> var_2898_cast_fp16 = transpose(perm = var_2897, x = k_49_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [3, 2, 1, 16, 128]> var_2910_cast_fp16 = expand_dims(axes = var_2910_axes_0, x = var_2898_cast_fp16)[name = string("op_2910_cast_fp16")];
            tensor<int32, [5]> var_2918_reps_0 = const()[name = string("op_2918_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 16, 128]> var_2918_cast_fp16 = tile(reps = var_2918_reps_0, x = var_2910_cast_fp16)[name = string("op_2918_cast_fp16")];
            tensor<int32, [4]> var_2923 = const()[name = string("op_2923"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<fp16, [3, 4, 16, 128]> k_51_cast_fp16 = reshape(shape = var_2923, x = var_2918_cast_fp16)[name = string("k_51_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_2925_cast_fp16 = sigmoid(x = b_cast_fp16)[name = string("op_2925_cast_fp16")];
            tensor<int32, [3]> var_2929 = const()[name = string("op_2929"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_6_dt_bias_to_fp16 = const()[name = string("layers_6_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 16, 4]> input_95_cast_fp16 = add(x = a_cast_fp16, y = layers_6_dt_bias_to_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_2933_cast_fp16 = relu(x = input_95_cast_fp16)[name = string("op_2933_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_2934_cast_fp16 = abs(x = input_95_cast_fp16)[name = string("op_2934_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 16, 4]> var_2935_cast_fp16 = mul(x = var_2934_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_2935_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_2936_cast_fp16 = exp(x = var_2935_cast_fp16)[name = string("op_2936_cast_fp16")];
            fp16 var_2938_to_fp16 = const()[name = string("op_2938_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 16, 4]> var_2939_cast_fp16 = add(x = var_2936_cast_fp16, y = var_2938_to_fp16)[name = string("op_2939_cast_fp16")];
            fp32 var_2940_epsilon_0 = const()[name = string("op_2940_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 16, 4]> var_2940_cast_fp16 = log(epsilon = var_2940_epsilon_0, x = var_2939_cast_fp16)[name = string("op_2940_cast_fp16")];
            tensor<fp16, [3, 16, 4]> var_2942_cast_fp16 = add(x = var_2933_cast_fp16, y = var_2940_cast_fp16)[name = string("op_2942_cast_fp16")];
            tensor<fp16, [4]> layers_6_negA_to_fp16 = const()[name = string("layers_6_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.07cp+1, -0x1.eccp+3, -0x1.63p+3, -0x1.728p+1])];
            tensor<fp16, [3, 16, 4]> var_2943_cast_fp16 = mul(x = layers_6_negA_to_fp16, y = var_2942_cast_fp16)[name = string("op_2943_cast_fp16")];
            tensor<int32, [3]> var_2947 = const()[name = string("op_2947"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_2950_axes_0 = const()[name = string("op_2950_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16]> beta_cast_fp16 = transpose(perm = var_2929, x = var_2925_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [3, 4, 16, 1]> var_2950_cast_fp16 = expand_dims(axes = var_2950_axes_0, x = beta_cast_fp16)[name = string("op_2950_cast_fp16")];
            fp16 var_2951_to_fp16 = const()[name = string("op_2951_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 16, 1]> var_2952_cast_fp16 = mul(x = var_2950_cast_fp16, y = var_2951_to_fp16)[name = string("op_2952_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> v_15_cast_fp16 = transpose(perm = var_2827, x = var_2822_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [3, 4, 16, 128]> vb_cast_fp16 = mul(x = v_15_cast_fp16, y = var_2952_cast_fp16)[name = string("vb_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> kb_cast_fp16 = mul(x = k_51_cast_fp16, y = var_2950_cast_fp16)[name = string("kb_cast_fp16")];
            tensor<int32, [2]> var_2959 = const()[name = string("op_2959"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [3, 4, 16]> gdec_cast_fp16 = transpose(perm = var_2947, x = var_2943_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [12, 16]> var_2960_cast_fp16 = reshape(shape = var_2959, x = gdec_cast_fp16)[name = string("op_2960_cast_fp16")];
            tensor<fp16, [12, 256]> var_2961_cast_fp16 = linear(bias = var_554_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2960_cast_fp16)[name = string("op_2961_cast_fp16")];
            tensor<int32, [4]> var_2966 = const()[name = string("op_2966"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_cast_fp16 = reshape(shape = var_2966, x = var_2961_cast_fp16)[name = string("P_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2968_cast_fp16 = exp(x = P_cast_fp16)[name = string("op_2968_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_cast_fp16 = mul(x = var_2968_cast_fp16, y = tril_incl_to_fp16)[name = string("D_cast_fp16")];
            bool kk_transpose_x_1 = const()[name = string("kk_transpose_x_1"), val = bool(false)];
            bool kk_transpose_y_1 = const()[name = string("kk_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_cast_fp16 = matmul(transpose_x = kk_transpose_x_1, transpose_y = kk_transpose_y_1, x = kb_cast_fp16, y = k_51_cast_fp16)[name = string("kk_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_2974_cast_fp16 = mul(x = kk_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_2974_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2975_cast_fp16 = mul(x = var_2974_cast_fp16, y = D_cast_fp16)[name = string("op_2975_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_cast_fp16 = mul(x = var_2975_cast_fp16, y = tril_strict_to_fp16)[name = string("N_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2977_cast_fp16 = mul(x = D_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2977_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_cast_fp16 = mul(x = kk_cast_fp16, y = var_2977_cast_fp16)[name = string("Na_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_41_cast_fp16 = add(x = eye_to_fp16, y = N_cast_fp16)[name = string("Minv_41_cast_fp16")];
            bool Np_31_transpose_x_0 = const()[name = string("Np_31_transpose_x_0"), val = bool(false)];
            bool Np_31_transpose_y_0 = const()[name = string("Np_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_31_cast_fp16 = matmul(transpose_x = Np_31_transpose_x_0, transpose_y = Np_31_transpose_y_0, x = N_cast_fp16, y = Na_cast_fp16)[name = string("Np_31_cast_fp16")];
            bool Npa_21_transpose_x_0 = const()[name = string("Npa_21_transpose_x_0"), val = bool(false)];
            bool Npa_21_transpose_y_0 = const()[name = string("Npa_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_21_cast_fp16 = matmul(transpose_x = Npa_21_transpose_x_0, transpose_y = Npa_21_transpose_y_0, x = Na_cast_fp16, y = N_cast_fp16)[name = string("Npa_21_cast_fp16")];
            bool var_2983_transpose_x_0 = const()[name = string("op_2983_transpose_x_0"), val = bool(false)];
            bool var_2983_transpose_y_0 = const()[name = string("op_2983_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2983_cast_fp16 = matmul(transpose_x = var_2983_transpose_x_0, transpose_y = var_2983_transpose_y_0, x = Minv_41_cast_fp16, y = Np_31_cast_fp16)[name = string("op_2983_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_43_cast_fp16 = add(x = Minv_41_cast_fp16, y = var_2983_cast_fp16)[name = string("Minv_43_cast_fp16")];
            bool Np_33_transpose_x_0 = const()[name = string("Np_33_transpose_x_0"), val = bool(false)];
            bool Np_33_transpose_y_0 = const()[name = string("Np_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_33_cast_fp16 = matmul(transpose_x = Np_33_transpose_x_0, transpose_y = Np_33_transpose_y_0, x = Np_31_cast_fp16, y = Npa_21_cast_fp16)[name = string("Np_33_cast_fp16")];
            bool Npa_transpose_x_0 = const()[name = string("Npa_transpose_x_0"), val = bool(false)];
            bool Npa_transpose_y_0 = const()[name = string("Npa_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_cast_fp16 = matmul(transpose_x = Npa_transpose_x_0, transpose_y = Npa_transpose_y_0, x = Npa_21_cast_fp16, y = Np_31_cast_fp16)[name = string("Npa_cast_fp16")];
            bool var_2988_transpose_x_0 = const()[name = string("op_2988_transpose_x_0"), val = bool(false)];
            bool var_2988_transpose_y_0 = const()[name = string("op_2988_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2988_cast_fp16 = matmul(transpose_x = var_2988_transpose_x_0, transpose_y = var_2988_transpose_y_0, x = Minv_43_cast_fp16, y = Np_33_cast_fp16)[name = string("op_2988_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_45_cast_fp16 = add(x = Minv_43_cast_fp16, y = var_2988_cast_fp16)[name = string("Minv_45_cast_fp16")];
            bool Np_transpose_x_0 = const()[name = string("Np_transpose_x_0"), val = bool(false)];
            bool Np_transpose_y_0 = const()[name = string("Np_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_cast_fp16 = matmul(transpose_x = Np_transpose_x_0, transpose_y = Np_transpose_y_0, x = Np_33_cast_fp16, y = Npa_cast_fp16)[name = string("Np_cast_fp16")];
            bool var_2992_transpose_x_0 = const()[name = string("op_2992_transpose_x_0"), val = bool(false)];
            bool var_2992_transpose_y_0 = const()[name = string("op_2992_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2992_cast_fp16 = matmul(transpose_x = var_2992_transpose_x_0, transpose_y = var_2992_transpose_y_0, x = Minv_45_cast_fp16, y = Np_cast_fp16)[name = string("op_2992_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_cast_fp16 = add(x = Minv_45_cast_fp16, y = var_2992_cast_fp16)[name = string("Minv_cast_fp16")];
            bool u_transpose_x_0 = const()[name = string("u_transpose_x_0"), val = bool(false)];
            bool u_transpose_y_0 = const()[name = string("u_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_cast_fp16 = matmul(transpose_x = u_transpose_x_0, transpose_y = u_transpose_y_0, x = Minv_cast_fp16, y = vb_cast_fp16)[name = string("u_cast_fp16")];
            bool var_2999_transpose_x_1 = const()[name = string("op_2999_transpose_x_1"), val = bool(false)];
            bool var_2999_transpose_y_1 = const()[name = string("op_2999_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_2999_cast_fp16 = matmul(transpose_x = var_2999_transpose_x_1, transpose_y = var_2999_transpose_y_1, x = q_49_cast_fp16, y = k_51_cast_fp16)[name = string("op_2999_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_cast_fp16 = mul(x = var_2999_cast_fp16, y = D_cast_fp16)[name = string("A_cast_fp16")];
            bool o_33_transpose_x_0 = const()[name = string("o_33_transpose_x_0"), val = bool(false)];
            bool o_33_transpose_y_0 = const()[name = string("o_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> o_33_cast_fp16 = matmul(transpose_x = o_33_transpose_x_0, transpose_y = o_33_transpose_y_0, x = A_cast_fp16, y = u_cast_fp16)[name = string("o_33_cast_fp16")];
            tensor<int32, [4]> var_3006 = const()[name = string("op_3006"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3012 = const()[name = string("op_3012"), val = tensor<int32, [4]>([3, 16, 4, 128])];
            tensor<fp16, [3, 16, 4, 128]> input_97_cast_fp16 = reshape(shape = var_3012, x = z_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> o_35_cast_fp16 = transpose(perm = var_3006, x = o_33_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [3, 16, 4, 128]> var_3014_cast_fp16 = mul(x = o_35_cast_fp16, y = o_35_cast_fp16)[name = string("op_3014_cast_fp16")];
            tensor<int32, [1]> var_3019_axes_0 = const()[name = string("op_3019_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3019_keep_dims_0 = const()[name = string("op_3019_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 4, 1]> var_3019_cast_fp16 = reduce_mean(axes = var_3019_axes_0, keep_dims = var_3019_keep_dims_0, x = var_3014_cast_fp16)[name = string("op_3019_cast_fp16")];
            fp16 var_3021_to_fp16 = const()[name = string("op_3021_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 16, 4, 1]> var_3022_cast_fp16 = add(x = var_3019_cast_fp16, y = var_3021_to_fp16)[name = string("op_3022_cast_fp16")];
            fp32 var_3023_epsilon_0 = const()[name = string("op_3023_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 4, 1]> var_3023_cast_fp16 = rsqrt(epsilon = var_3023_epsilon_0, x = var_3022_cast_fp16)[name = string("op_3023_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> var_3024_cast_fp16 = mul(x = o_35_cast_fp16, y = var_3023_cast_fp16)[name = string("op_3024_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> var_3026_cast_fp16 = silu(x = input_97_cast_fp16)[name = string("op_3026_cast_fp16")];
            tensor<fp16, [3, 16, 4, 128]> o_37_cast_fp16 = mul(x = var_3024_cast_fp16, y = var_3026_cast_fp16)[name = string("o_37_cast_fp16")];
            tensor<int32, [3]> var_3031 = const()[name = string("op_3031"), val = tensor<int32, [3]>([3, 16, 512])];
            tensor<fp16, [3, 16, 512]> input_99_cast_fp16 = reshape(shape = var_3031, x = o_37_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [640, 512]> layers_6_out_weight_to_fp16 = const()[name = string("layers_6_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97009984)))];
            tensor<fp16, [3, 16, 640]> linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_out_weight_to_fp16, x = input_99_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [3, 16, 640]> x_65_cast_fp16 = add(x = x_61_cast_fp16, y = linear_25_cast_fp16)[name = string("x_65_cast_fp16")];
            fp16 var_3044_to_fp16 = const()[name = string("op_3044_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [3, 16, 640]> x_67_cast_fp16 = mul(x = x_65_cast_fp16, y = var_3044_to_fp16)[name = string("x_67_cast_fp16")];
            fp16 var_3042_promoted_to_fp16 = const()[name = string("op_3042_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> var_3046_cast_fp16 = pow(x = x_67_cast_fp16, y = var_3042_promoted_to_fp16)[name = string("op_3046_cast_fp16")];
            tensor<int32, [1]> var_3048_axes_0 = const()[name = string("op_3048_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3048_keep_dims_0 = const()[name = string("op_3048_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 1]> var_3048_cast_fp16 = reduce_mean(axes = var_3048_axes_0, keep_dims = var_3048_keep_dims_0, x = var_3046_cast_fp16)[name = string("op_3048_cast_fp16")];
            fp16 var_3049_to_fp16 = const()[name = string("op_3049_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 16, 1]> var_3050_cast_fp16 = add(x = var_3048_cast_fp16, y = var_3049_to_fp16)[name = string("op_3050_cast_fp16")];
            fp32 var_3051_epsilon_0 = const()[name = string("op_3051_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 1]> var_3051_cast_fp16 = rsqrt(epsilon = var_3051_epsilon_0, x = var_3050_cast_fp16)[name = string("op_3051_cast_fp16")];
            tensor<fp16, [3, 16, 640]> var_3052_cast_fp16 = mul(x = x_67_cast_fp16, y = var_3051_cast_fp16)[name = string("op_3052_cast_fp16")];
            tensor<fp16, [640]> layers_6_n2_w1_to_fp16 = const()[name = string("layers_6_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97665408)))];
            tensor<fp16, [3, 16, 640]> input_101_cast_fp16 = mul(x = var_3052_cast_fp16, y = layers_6_n2_w1_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_6_gate_up_weight_to_fp16 = const()[name = string("layers_6_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97666752)))];
            tensor<fp16, [3, 16, 5120]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_6_gate_up_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<int32, [3]> input_103_begin_0 = const()[name = string("input_103_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_103_end_0 = const()[name = string("input_103_end_0"), val = tensor<int32, [3]>([3, 16, 2560])];
            tensor<bool, [3]> input_103_end_mask_0 = const()[name = string("input_103_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 2560]> input_103_cast_fp16 = slice_by_index(begin = input_103_begin_0, end = input_103_end_0, end_mask = input_103_end_mask_0, x = linear_26_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [3, 16, 2560]> var_3062_cast_fp16 = silu(x = input_103_cast_fp16)[name = string("op_3062_cast_fp16")];
            tensor<int32, [3]> var_3067_begin_0 = const()[name = string("op_3067_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_3067_end_0 = const()[name = string("op_3067_end_0"), val = tensor<int32, [3]>([3, 16, 5120])];
            tensor<bool, [3]> var_3067_end_mask_0 = const()[name = string("op_3067_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 2560]> var_3067_cast_fp16 = slice_by_index(begin = var_3067_begin_0, end = var_3067_end_0, end_mask = var_3067_end_mask_0, x = linear_26_cast_fp16)[name = string("op_3067_cast_fp16")];
            tensor<fp16, [3, 16, 2560]> input_105_cast_fp16 = mul(x = var_3062_cast_fp16, y = var_3067_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_6_down_weight_to_fp16 = const()[name = string("layers_6_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104220416)))];
            tensor<fp16, [3, 16, 640]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_down_weight_to_fp16, x = input_105_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [3, 16, 640]> x_69_cast_fp16 = add(x = x_65_cast_fp16, y = linear_27_cast_fp16)[name = string("x_69_cast_fp16")];
            fp16 var_3080_to_fp16 = const()[name = string("op_3080_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [3, 16, 640]> x_71_cast_fp16 = mul(x = x_69_cast_fp16, y = var_3080_to_fp16)[name = string("x_71_cast_fp16")];
            fp16 var_3078_promoted_to_fp16 = const()[name = string("op_3078_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> var_3082_cast_fp16 = pow(x = x_71_cast_fp16, y = var_3078_promoted_to_fp16)[name = string("op_3082_cast_fp16")];
            tensor<int32, [1]> var_3084_axes_0 = const()[name = string("op_3084_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3084_keep_dims_0 = const()[name = string("op_3084_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 1]> var_3084_cast_fp16 = reduce_mean(axes = var_3084_axes_0, keep_dims = var_3084_keep_dims_0, x = var_3082_cast_fp16)[name = string("op_3084_cast_fp16")];
            fp16 var_3085_to_fp16 = const()[name = string("op_3085_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 16, 1]> var_3086_cast_fp16 = add(x = var_3084_cast_fp16, y = var_3085_to_fp16)[name = string("op_3086_cast_fp16")];
            fp32 var_3087_epsilon_0 = const()[name = string("op_3087_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 1]> var_3087_cast_fp16 = rsqrt(epsilon = var_3087_epsilon_0, x = var_3086_cast_fp16)[name = string("op_3087_cast_fp16")];
            tensor<fp16, [3, 16, 640]> var_3088_cast_fp16 = mul(x = x_71_cast_fp16, y = var_3087_cast_fp16)[name = string("op_3088_cast_fp16")];
            tensor<fp16, [640]> layers_7_n1_w1_to_fp16 = const()[name = string("layers_7_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107497280)))];
            tensor<fp16, [3, 16, 640]> input_107_cast_fp16 = mul(x = var_3088_cast_fp16, y = layers_7_n1_w1_to_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [1280, 640]> layers_7_qkv_weight_to_fp16 = const()[name = string("layers_7_qkv_weight_to_fp16"), val = tensor<fp16, [1280, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107498624)))];
            tensor<fp16, [3, 16, 1280]> linear_28_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = layers_7_qkv_weight_to_fp16, x = input_107_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [3]> var_3097_begin_0 = const()[name = string("op_3097_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3097_end_0 = const()[name = string("op_3097_end_0"), val = tensor<int32, [3]>([3, 16, 1024])];
            tensor<bool, [3]> var_3097_end_mask_0 = const()[name = string("op_3097_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 1024]> var_3097_cast_fp16 = slice_by_index(begin = var_3097_begin_0, end = var_3097_end_0, end_mask = var_3097_end_mask_0, x = linear_28_cast_fp16)[name = string("op_3097_cast_fp16")];
            tensor<int32, [4]> var_3102 = const()[name = string("op_3102"), val = tensor<int32, [4]>([3, 16, 8, 128])];
            tensor<fp16, [3, 16, 8, 128]> qg_cast_fp16 = reshape(shape = var_3102, x = var_3097_cast_fp16)[name = string("qg_cast_fp16")];
            tensor<int32, [4]> x_73_begin_0 = const()[name = string("x_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_73_end_0 = const()[name = string("x_73_end_0"), val = tensor<int32, [4]>([3, 16, 8, 64])];
            tensor<bool, [4]> x_73_end_mask_0 = const()[name = string("x_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 16, 8, 64]> x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, x = qg_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<int32, [4]> var_3113_begin_0 = const()[name = string("op_3113_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> var_3113_end_0 = const()[name = string("op_3113_end_0"), val = tensor<int32, [4]>([3, 16, 8, 128])];
            tensor<bool, [4]> var_3113_end_mask_0 = const()[name = string("op_3113_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 16, 8, 64]> var_3113_cast_fp16 = slice_by_index(begin = var_3113_begin_0, end = var_3113_end_0, end_mask = var_3113_end_mask_0, x = qg_cast_fp16)[name = string("op_3113_cast_fp16")];
            tensor<int32, [3]> var_3117 = const()[name = string("op_3117"), val = tensor<int32, [3]>([3, 16, 512])];
            tensor<fp16, [3, 16, 512]> gate_cast_fp16 = reshape(shape = var_3117, x = var_3113_cast_fp16)[name = string("gate_cast_fp16")];
            tensor<int32, [3]> var_3123_begin_0 = const()[name = string("op_3123_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> var_3123_end_0 = const()[name = string("op_3123_end_0"), val = tensor<int32, [3]>([3, 16, 1152])];
            tensor<bool, [3]> var_3123_end_mask_0 = const()[name = string("op_3123_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 128]> var_3123_cast_fp16 = slice_by_index(begin = var_3123_begin_0, end = var_3123_end_0, end_mask = var_3123_end_mask_0, x = linear_28_cast_fp16)[name = string("op_3123_cast_fp16")];
            tensor<int32, [4]> var_3128 = const()[name = string("op_3128"), val = tensor<int32, [4]>([3, 16, 2, 64])];
            tensor<fp16, [3, 16, 2, 64]> x_79_cast_fp16 = reshape(shape = var_3128, x = var_3123_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<int32, [3]> var_3134_begin_0 = const()[name = string("op_3134_begin_0"), val = tensor<int32, [3]>([0, 0, 1152])];
            tensor<int32, [3]> var_3134_end_0 = const()[name = string("op_3134_end_0"), val = tensor<int32, [3]>([3, 16, 1280])];
            tensor<bool, [3]> var_3134_end_mask_0 = const()[name = string("op_3134_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 128]> var_3134_cast_fp16 = slice_by_index(begin = var_3134_begin_0, end = var_3134_end_0, end_mask = var_3134_end_mask_0, x = linear_28_cast_fp16)[name = string("op_3134_cast_fp16")];
            tensor<int32, [4]> var_3139 = const()[name = string("op_3139"), val = tensor<int32, [4]>([3, 16, 2, 64])];
            tensor<fp16, [3, 16, 2, 64]> var_3140_cast_fp16 = reshape(shape = var_3139, x = var_3134_cast_fp16)[name = string("op_3140_cast_fp16")];
            tensor<int32, [4]> var_3145 = const()[name = string("op_3145"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3151_promoted_to_fp16 = const()[name = string("op_3151_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 8, 64]> var_3155_cast_fp16 = pow(x = x_73_cast_fp16, y = var_3151_promoted_to_fp16)[name = string("op_3155_cast_fp16")];
            tensor<int32, [1]> var_3157_axes_0 = const()[name = string("op_3157_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3157_keep_dims_0 = const()[name = string("op_3157_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 8, 1]> var_3157_cast_fp16 = reduce_mean(axes = var_3157_axes_0, keep_dims = var_3157_keep_dims_0, x = var_3155_cast_fp16)[name = string("op_3157_cast_fp16")];
            fp16 var_3158_to_fp16 = const()[name = string("op_3158_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 16, 8, 1]> var_3159_cast_fp16 = add(x = var_3157_cast_fp16, y = var_3158_to_fp16)[name = string("op_3159_cast_fp16")];
            fp32 var_3160_epsilon_0 = const()[name = string("op_3160_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 8, 1]> var_3160_cast_fp16 = rsqrt(epsilon = var_3160_epsilon_0, x = var_3159_cast_fp16)[name = string("op_3160_cast_fp16")];
            tensor<fp16, [3, 16, 8, 64]> var_3161_cast_fp16 = mul(x = x_73_cast_fp16, y = var_3160_cast_fp16)[name = string("op_3161_cast_fp16")];
            tensor<fp16, [64]> layers_7_qn_w1_to_fp16 = const()[name = string("layers_7_qn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137088)))];
            tensor<fp16, [3, 16, 8, 64]> var_3162_cast_fp16 = mul(x = var_3161_cast_fp16, y = layers_7_qn_w1_to_fp16)[name = string("op_3162_cast_fp16")];
            tensor<int32, [4]> var_3167 = const()[name = string("op_3167"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_5_begin_0 = const()[name = string("r_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_5_end_0 = const()[name = string("r_5_end_0"), val = tensor<int32, [4]>([3, 8, 16, 16])];
            tensor<bool, [4]> r_5_end_mask_0 = const()[name = string("r_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 8, 16, 64]> x_77_cast_fp16 = transpose(perm = var_3167, x = var_3162_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [3, 8, 16, 16]> r_5_cast_fp16 = slice_by_index(begin = r_5_begin_0, end = r_5_end_0, end_mask = r_5_end_mask_0, x = x_77_cast_fp16)[name = string("r_5_cast_fp16")];
            tensor<int32, [4]> p_5_begin_0 = const()[name = string("p_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_5_end_0 = const()[name = string("p_5_end_0"), val = tensor<int32, [4]>([3, 8, 16, 64])];
            tensor<bool, [4]> p_5_end_mask_0 = const()[name = string("p_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 8, 16, 48]> p_5_cast_fp16 = slice_by_index(begin = p_5_begin_0, end = p_5_end_0, end_mask = p_5_end_mask_0, x = x_77_cast_fp16)[name = string("p_5_cast_fp16")];
            tensor<int32, [4]> var_3183_begin_0 = const()[name = string("op_3183_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_3183_end_0 = const()[name = string("op_3183_end_0"), val = tensor<int32, [4]>([3, 8, 16, 16])];
            tensor<bool, [4]> var_3183_end_mask_0 = const()[name = string("op_3183_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 8, 16, 8]> var_3183_cast_fp16 = slice_by_index(begin = var_3183_begin_0, end = var_3183_end_0, end_mask = var_3183_end_mask_0, x = r_5_cast_fp16)[name = string("op_3183_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 8, 16, 8]> var_3184_cast_fp16 = mul(x = var_3183_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_3184_cast_fp16")];
            tensor<int32, [4]> var_3189_begin_0 = const()[name = string("op_3189_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3189_end_0 = const()[name = string("op_3189_end_0"), val = tensor<int32, [4]>([3, 8, 16, 8])];
            tensor<bool, [4]> var_3189_end_mask_0 = const()[name = string("op_3189_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 8, 16, 8]> var_3189_cast_fp16 = slice_by_index(begin = var_3189_begin_0, end = var_3189_end_0, end_mask = var_3189_end_mask_0, x = r_5_cast_fp16)[name = string("op_3189_cast_fp16")];
            int32 var_3191 = const()[name = string("op_3191"), val = int32(-1)];
            bool rh_5_interleave_0 = const()[name = string("rh_5_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 8, 16, 16]> rh_5_cast_fp16 = concat(axis = var_3191, interleave = rh_5_interleave_0, values = (var_3184_cast_fp16, var_3189_cast_fp16))[name = string("rh_5_cast_fp16")];
            tensor<fp16, [3, 8, 16, 16]> var_3193_cast_fp16 = mul(x = r_5_cast_fp16, y = cos_to_fp16)[name = string("op_3193_cast_fp16")];
            tensor<fp16, [3, 8, 16, 16]> var_3194_cast_fp16 = mul(x = rh_5_cast_fp16, y = sin_to_fp16)[name = string("op_3194_cast_fp16")];
            tensor<fp16, [3, 8, 16, 16]> var_3196_cast_fp16 = add(x = var_3193_cast_fp16, y = var_3194_cast_fp16)[name = string("op_3196_cast_fp16")];
            int32 var_3198 = const()[name = string("op_3198"), val = int32(-1)];
            bool q_interleave_0 = const()[name = string("q_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 8, 16, 64]> q_cast_fp16 = concat(axis = var_3198, interleave = q_interleave_0, values = (var_3196_cast_fp16, p_5_cast_fp16))[name = string("q_cast_fp16")];
            fp16 var_3204_promoted_to_fp16 = const()[name = string("op_3204_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 2, 64]> var_3208_cast_fp16 = pow(x = x_79_cast_fp16, y = var_3204_promoted_to_fp16)[name = string("op_3208_cast_fp16")];
            tensor<int32, [1]> var_3210_axes_0 = const()[name = string("op_3210_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3210_keep_dims_0 = const()[name = string("op_3210_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 2, 1]> var_3210_cast_fp16 = reduce_mean(axes = var_3210_axes_0, keep_dims = var_3210_keep_dims_0, x = var_3208_cast_fp16)[name = string("op_3210_cast_fp16")];
            fp16 var_3211_to_fp16 = const()[name = string("op_3211_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 16, 2, 1]> var_3212_cast_fp16 = add(x = var_3210_cast_fp16, y = var_3211_to_fp16)[name = string("op_3212_cast_fp16")];
            fp32 var_3213_epsilon_0 = const()[name = string("op_3213_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 2, 1]> var_3213_cast_fp16 = rsqrt(epsilon = var_3213_epsilon_0, x = var_3212_cast_fp16)[name = string("op_3213_cast_fp16")];
            tensor<fp16, [3, 16, 2, 64]> var_3214_cast_fp16 = mul(x = x_79_cast_fp16, y = var_3213_cast_fp16)[name = string("op_3214_cast_fp16")];
            tensor<fp16, [64]> layers_7_kn_w1_to_fp16 = const()[name = string("layers_7_kn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137280)))];
            tensor<fp16, [3, 16, 2, 64]> var_3215_cast_fp16 = mul(x = var_3214_cast_fp16, y = layers_7_kn_w1_to_fp16)[name = string("op_3215_cast_fp16")];
            tensor<int32, [4]> var_3220 = const()[name = string("op_3220"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_begin_0 = const()[name = string("r_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_end_0 = const()[name = string("r_end_0"), val = tensor<int32, [4]>([3, 2, 16, 16])];
            tensor<bool, [4]> r_end_mask_0 = const()[name = string("r_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 2, 16, 64]> x_83_cast_fp16 = transpose(perm = var_3220, x = var_3215_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [3, 2, 16, 16]> r_cast_fp16 = slice_by_index(begin = r_begin_0, end = r_end_0, end_mask = r_end_mask_0, x = x_83_cast_fp16)[name = string("r_cast_fp16")];
            tensor<int32, [4]> p_begin_0 = const()[name = string("p_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_end_0 = const()[name = string("p_end_0"), val = tensor<int32, [4]>([3, 2, 16, 64])];
            tensor<bool, [4]> p_end_mask_0 = const()[name = string("p_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 2, 16, 48]> p_cast_fp16 = slice_by_index(begin = p_begin_0, end = p_end_0, end_mask = p_end_mask_0, x = x_83_cast_fp16)[name = string("p_cast_fp16")];
            tensor<int32, [4]> var_3236_begin_0 = const()[name = string("op_3236_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_3236_end_0 = const()[name = string("op_3236_end_0"), val = tensor<int32, [4]>([3, 2, 16, 16])];
            tensor<bool, [4]> var_3236_end_mask_0 = const()[name = string("op_3236_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 2, 16, 8]> var_3236_cast_fp16 = slice_by_index(begin = var_3236_begin_0, end = var_3236_end_0, end_mask = var_3236_end_mask_0, x = r_cast_fp16)[name = string("op_3236_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 2, 16, 8]> var_3237_cast_fp16 = mul(x = var_3236_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_3237_cast_fp16")];
            tensor<int32, [4]> var_3242_begin_0 = const()[name = string("op_3242_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3242_end_0 = const()[name = string("op_3242_end_0"), val = tensor<int32, [4]>([3, 2, 16, 8])];
            tensor<bool, [4]> var_3242_end_mask_0 = const()[name = string("op_3242_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 2, 16, 8]> var_3242_cast_fp16 = slice_by_index(begin = var_3242_begin_0, end = var_3242_end_0, end_mask = var_3242_end_mask_0, x = r_cast_fp16)[name = string("op_3242_cast_fp16")];
            int32 var_3244 = const()[name = string("op_3244"), val = int32(-1)];
            bool rh_interleave_0 = const()[name = string("rh_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 2, 16, 16]> rh_cast_fp16 = concat(axis = var_3244, interleave = rh_interleave_0, values = (var_3237_cast_fp16, var_3242_cast_fp16))[name = string("rh_cast_fp16")];
            tensor<fp16, [3, 2, 16, 16]> var_3246_cast_fp16 = mul(x = r_cast_fp16, y = cos_to_fp16)[name = string("op_3246_cast_fp16")];
            tensor<fp16, [3, 2, 16, 16]> var_3247_cast_fp16 = mul(x = rh_cast_fp16, y = sin_to_fp16)[name = string("op_3247_cast_fp16")];
            tensor<fp16, [3, 2, 16, 16]> var_3249_cast_fp16 = add(x = var_3246_cast_fp16, y = var_3247_cast_fp16)[name = string("op_3249_cast_fp16")];
            int32 var_3251 = const()[name = string("op_3251"), val = int32(-1)];
            bool k_53_interleave_0 = const()[name = string("k_53_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 2, 16, 64]> k_53_cast_fp16 = concat(axis = var_3251, interleave = k_53_interleave_0, values = (var_3249_cast_fp16, p_cast_fp16))[name = string("k_53_cast_fp16")];
            tensor<int32, [1]> var_3264_axes_0 = const()[name = string("op_3264_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 1, 16, 64]> var_3264_cast_fp16 = expand_dims(axes = var_3264_axes_0, x = k_53_cast_fp16)[name = string("op_3264_cast_fp16")];
            tensor<int32, [5]> var_3272_reps_0 = const()[name = string("op_3272_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [3, 2, 4, 16, 64]> var_3272_cast_fp16 = tile(reps = var_3272_reps_0, x = var_3264_cast_fp16)[name = string("op_3272_cast_fp16")];
            tensor<int32, [4]> var_3277 = const()[name = string("op_3277"), val = tensor<int32, [4]>([3, 8, 16, 64])];
            tensor<fp16, [3, 8, 16, 64]> k_cast_fp16 = reshape(shape = var_3277, x = var_3272_cast_fp16)[name = string("k_cast_fp16")];
            tensor<int32, [1]> var_3290_axes_0 = const()[name = string("op_3290_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 16, 64]> v_17_cast_fp16 = transpose(perm = var_3145, x = var_3140_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [3, 2, 1, 16, 64]> var_3290_cast_fp16 = expand_dims(axes = var_3290_axes_0, x = v_17_cast_fp16)[name = string("op_3290_cast_fp16")];
            tensor<int32, [5]> var_3298_reps_0 = const()[name = string("op_3298_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [3, 2, 4, 16, 64]> var_3298_cast_fp16 = tile(reps = var_3298_reps_0, x = var_3290_cast_fp16)[name = string("op_3298_cast_fp16")];
            tensor<int32, [4]> var_3303 = const()[name = string("op_3303"), val = tensor<int32, [4]>([3, 8, 16, 64])];
            tensor<fp16, [3, 8, 16, 64]> v_cast_fp16 = reshape(shape = var_3303, x = var_3298_cast_fp16)[name = string("v_cast_fp16")];
            bool var_3308_transpose_x_1 = const()[name = string("op_3308_transpose_x_1"), val = bool(false)];
            bool var_3308_transpose_y_1 = const()[name = string("op_3308_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 8, 16, 16]> var_3308_cast_fp16 = matmul(transpose_x = var_3308_transpose_x_1, transpose_y = var_3308_transpose_y_1, x = q_cast_fp16, y = k_cast_fp16)[name = string("op_3308_cast_fp16")];
            fp16 var_3309_to_fp16 = const()[name = string("op_3309_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [3, 8, 16, 16]> var_3310_cast_fp16 = mul(x = var_3308_cast_fp16, y = var_3309_to_fp16)[name = string("op_3310_cast_fp16")];
            tensor<fp16, [3, 8, 16, 16]> s_cast_fp16 = add(x = var_3310_cast_fp16, y = causal_add_to_fp16)[name = string("s_cast_fp16")];
            int32 var_3313 = const()[name = string("op_3313"), val = int32(-1)];
            tensor<fp16, [3, 8, 16, 16]> var_3315_cast_fp16 = softmax(axis = var_3313, x = s_cast_fp16)[name = string("op_3315_cast_fp16")];
            bool o_39_transpose_x_0 = const()[name = string("o_39_transpose_x_0"), val = bool(false)];
            bool o_39_transpose_y_0 = const()[name = string("o_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 8, 16, 64]> o_39_cast_fp16 = matmul(transpose_x = o_39_transpose_x_0, transpose_y = o_39_transpose_y_0, x = var_3315_cast_fp16, y = v_cast_fp16)[name = string("o_39_cast_fp16")];
            tensor<int32, [4]> var_3321 = const()[name = string("op_3321"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3326 = const()[name = string("op_3326"), val = tensor<int32, [3]>([3, 16, 512])];
            tensor<fp16, [3, 16, 8, 64]> var_3322_cast_fp16 = transpose(perm = var_3321, x = o_39_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [3, 16, 512]> var_3327_cast_fp16 = reshape(shape = var_3326, x = var_3322_cast_fp16)[name = string("op_3327_cast_fp16")];
            tensor<fp16, [3, 16, 512]> var_3328_cast_fp16 = sigmoid(x = gate_cast_fp16)[name = string("op_3328_cast_fp16")];
            tensor<fp16, [3, 16, 512]> input_109_cast_fp16 = mul(x = var_3327_cast_fp16, y = var_3328_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [640, 512]> layers_7_o_weight_to_fp16 = const()[name = string("layers_7_o_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137472)))];
            tensor<fp16, [3, 16, 640]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_o_weight_to_fp16, x = input_109_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<fp16, [3, 16, 640]> x_85_cast_fp16 = add(x = x_69_cast_fp16, y = linear_29_cast_fp16)[name = string("x_85_cast_fp16")];
            fp16 var_3341_to_fp16 = const()[name = string("op_3341_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [3, 16, 640]> x_87_cast_fp16 = mul(x = x_85_cast_fp16, y = var_3341_to_fp16)[name = string("x_87_cast_fp16")];
            fp16 var_3339_promoted_to_fp16 = const()[name = string("op_3339_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> var_3343_cast_fp16 = pow(x = x_87_cast_fp16, y = var_3339_promoted_to_fp16)[name = string("op_3343_cast_fp16")];
            tensor<int32, [1]> var_3345_axes_0 = const()[name = string("op_3345_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3345_keep_dims_0 = const()[name = string("op_3345_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 1]> var_3345_cast_fp16 = reduce_mean(axes = var_3345_axes_0, keep_dims = var_3345_keep_dims_0, x = var_3343_cast_fp16)[name = string("op_3345_cast_fp16")];
            fp16 var_3346_to_fp16 = const()[name = string("op_3346_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 16, 1]> var_3347_cast_fp16 = add(x = var_3345_cast_fp16, y = var_3346_to_fp16)[name = string("op_3347_cast_fp16")];
            fp32 var_3348_epsilon_0 = const()[name = string("op_3348_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 1]> var_3348_cast_fp16 = rsqrt(epsilon = var_3348_epsilon_0, x = var_3347_cast_fp16)[name = string("op_3348_cast_fp16")];
            tensor<fp16, [3, 16, 640]> var_3349_cast_fp16 = mul(x = x_87_cast_fp16, y = var_3348_cast_fp16)[name = string("op_3349_cast_fp16")];
            tensor<fp16, [640]> layers_7_n2_w1_to_fp16 = const()[name = string("layers_7_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109792896)))];
            tensor<fp16, [3, 16, 640]> input_111_cast_fp16 = mul(x = var_3349_cast_fp16, y = layers_7_n2_w1_to_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_7_gate_up_weight_to_fp16 = const()[name = string("layers_7_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109794240)))];
            tensor<fp16, [3, 16, 5120]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_7_gate_up_weight_to_fp16, x = input_111_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [3]> input_113_begin_0 = const()[name = string("input_113_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_113_end_0 = const()[name = string("input_113_end_0"), val = tensor<int32, [3]>([3, 16, 2560])];
            tensor<bool, [3]> input_113_end_mask_0 = const()[name = string("input_113_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 16, 2560]> input_113_cast_fp16 = slice_by_index(begin = input_113_begin_0, end = input_113_end_0, end_mask = input_113_end_mask_0, x = linear_30_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [3, 16, 2560]> var_3359_cast_fp16 = silu(x = input_113_cast_fp16)[name = string("op_3359_cast_fp16")];
            tensor<int32, [3]> var_3364_begin_0 = const()[name = string("op_3364_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_3364_end_0 = const()[name = string("op_3364_end_0"), val = tensor<int32, [3]>([3, 16, 5120])];
            tensor<bool, [3]> var_3364_end_mask_0 = const()[name = string("op_3364_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 16, 2560]> var_3364_cast_fp16 = slice_by_index(begin = var_3364_begin_0, end = var_3364_end_0, end_mask = var_3364_end_mask_0, x = linear_30_cast_fp16)[name = string("op_3364_cast_fp16")];
            tensor<fp16, [3, 16, 2560]> input_115_cast_fp16 = mul(x = var_3359_cast_fp16, y = var_3364_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_7_down_weight_to_fp16 = const()[name = string("layers_7_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116347904)))];
            tensor<fp16, [3, 16, 640]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_down_weight_to_fp16, x = input_115_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [3, 16, 640]> x_89_cast_fp16 = add(x = x_85_cast_fp16, y = linear_31_cast_fp16)[name = string("x_89_cast_fp16")];
            fp16 var_3377_to_fp16 = const()[name = string("op_3377_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [3, 16, 640]> x_cast_fp16 = mul(x = x_89_cast_fp16, y = var_3377_to_fp16)[name = string("x_cast_fp16")];
            fp16 var_3375_promoted_to_fp16 = const()[name = string("op_3375_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 16, 640]> var_3379_cast_fp16 = pow(x = x_cast_fp16, y = var_3375_promoted_to_fp16)[name = string("op_3379_cast_fp16")];
            tensor<int32, [1]> var_3381_axes_0 = const()[name = string("op_3381_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3381_keep_dims_0 = const()[name = string("op_3381_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 16, 1]> var_3381_cast_fp16 = reduce_mean(axes = var_3381_axes_0, keep_dims = var_3381_keep_dims_0, x = var_3379_cast_fp16)[name = string("op_3381_cast_fp16")];
            fp16 var_3382_to_fp16 = const()[name = string("op_3382_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 16, 1]> var_3383_cast_fp16 = add(x = var_3381_cast_fp16, y = var_3382_to_fp16)[name = string("op_3383_cast_fp16")];
            fp32 var_3384_epsilon_0 = const()[name = string("op_3384_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 16, 1]> var_3384_cast_fp16 = rsqrt(epsilon = var_3384_epsilon_0, x = var_3383_cast_fp16)[name = string("op_3384_cast_fp16")];
            tensor<fp16, [3, 16, 640]> var_3385_cast_fp16 = mul(x = x_cast_fp16, y = var_3384_cast_fp16)[name = string("op_3385_cast_fp16")];
            tensor<fp16, [640]> norm_w1_to_fp16 = const()[name = string("norm_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119624768)))];
            tensor<fp16, [3, 16, 640]> var_3386_cast_fp16 = mul(x = var_3385_cast_fp16, y = norm_w1_to_fp16)[name = string("op_3386_cast_fp16")];
            tensor<int32, [3]> h_begin_0 = const()[name = string("h_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> h_end_0 = const()[name = string("h_end_0"), val = tensor<int32, [3]>([3, 15, 640])];
            tensor<bool, [3]> h_end_mask_0 = const()[name = string("h_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 15, 640]> h_cast_fp16 = slice_by_index(begin = h_begin_0, end = h_end_0, end_mask = h_end_mask_0, x = var_3386_cast_fp16)[name = string("h_cast_fp16")];
            tensor<fp16, [16000]> logits_bias_0_to_fp16 = const()[name = string("logits_bias_0_to_fp16"), val = tensor<fp16, [16000]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119626112)))];
            tensor<fp16, [3, 15, 16000]> logits_cast_fp16 = linear(bias = logits_bias_0_to_fp16, weight = emb_to_fp16, x = h_cast_fp16)[name = string("logits_cast_fp16")];
            tensor<int32, [1]> reduce_max_0_axes_0 = const()[name = string("reduce_max_0_axes_0"), val = tensor<int32, [1]>([-1])];
            bool reduce_max_0_keep_dims_0 = const()[name = string("reduce_max_0_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 15, 1]> reduce_max_0_cast_fp16 = reduce_max(axes = reduce_max_0_axes_0, keep_dims = reduce_max_0_keep_dims_0, x = logits_cast_fp16)[name = string("reduce_max_0_cast_fp16")];
            tensor<fp16, [3, 15, 16000]> var_3404_cast_fp16 = sub(x = logits_cast_fp16, y = reduce_max_0_cast_fp16)[name = string("op_3404_cast_fp16")];
            tensor<fp16, [3, 15, 16000]> var_3405_cast_fp16 = exp(x = var_3404_cast_fp16)[name = string("op_3405_cast_fp16")];
            tensor<int32, [1]> var_3410_axes_0 = const()[name = string("op_3410_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3410_keep_dims_0 = const()[name = string("op_3410_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 15, 1]> var_3410_cast_fp16 = reduce_sum(axes = var_3410_axes_0, keep_dims = var_3410_keep_dims_0, x = var_3405_cast_fp16)[name = string("op_3410_cast_fp16")];
            fp32 var_3411_epsilon_0 = const()[name = string("op_3411_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 15, 1]> var_3411_cast_fp16 = log(epsilon = var_3411_epsilon_0, x = var_3410_cast_fp16)[name = string("op_3411_cast_fp16")];
            tensor<fp16, [3, 15, 1]> var_3413_cast_fp16 = add(x = reduce_max_0_cast_fp16, y = var_3411_cast_fp16)[name = string("op_3413_cast_fp16")];
            tensor<int32, [3]> lse_begin_0 = const()[name = string("lse_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> lse_end_0 = const()[name = string("lse_end_0"), val = tensor<int32, [3]>([3, 15, 1])];
            tensor<bool, [3]> lse_end_mask_0 = const()[name = string("lse_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> lse_squeeze_mask_0 = const()[name = string("lse_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<fp16, [3, 15]> lse_cast_fp16 = slice_by_index(begin = lse_begin_0, end = lse_end_0, end_mask = lse_end_mask_0, squeeze_mask = lse_squeeze_mask_0, x = var_3413_cast_fp16)[name = string("lse_cast_fp16")];
            tensor<int32, [2]> input_begin_0 = const()[name = string("input_begin_0"), val = tensor<int32, [2]>([0, 1])];
            tensor<int32, [2]> input_end_0 = const()[name = string("input_end_0"), val = tensor<int32, [2]>([3, 16])];
            tensor<bool, [2]> input_end_mask_0 = const()[name = string("input_end_mask_0"), val = tensor<bool, [2]>([true, true])];
            tensor<int32, [3, 15]> input = slice_by_index(begin = input_begin_0, end = input_end_0, end_mask = input_end_mask_0, x = ids)[name = string("input")];
            int32 var_3430_axis_0 = const()[name = string("op_3430_axis_0"), val = int32(0)];
            int32 var_3430_batch_dims_0 = const()[name = string("op_3430_batch_dims_0"), val = int32(0)];
            bool var_3430_validate_indices_0 = const()[name = string("op_3430_validate_indices_0"), val = bool(false)];
            string input_to_uint16_dtype_0 = const()[name = string("input_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [3, 15]> input_to_uint16 = cast(dtype = input_to_uint16_dtype_0, x = input)[name = string("cast_72")];
            tensor<fp16, [3, 15, 640]> var_3430_cast_fp16_cast_uint16 = gather(axis = var_3430_axis_0, batch_dims = var_3430_batch_dims_0, indices = input_to_uint16, validate_indices = var_3430_validate_indices_0, x = emb_to_fp16)[name = string("op_3430_cast_fp16_cast_uint16")];
            tensor<fp16, [3, 15, 640]> var_3431_cast_fp16 = mul(x = h_cast_fp16, y = var_3430_cast_fp16_cast_uint16)[name = string("op_3431_cast_fp16")];
            tensor<int32, [1]> tgt_axes_0 = const()[name = string("tgt_axes_0"), val = tensor<int32, [1]>([-1])];
            bool tgt_keep_dims_0 = const()[name = string("tgt_keep_dims_0"), val = bool(false)];
            tensor<fp16, [3, 15]> tgt_cast_fp16 = reduce_sum(axes = tgt_axes_0, keep_dims = tgt_keep_dims_0, x = var_3431_cast_fp16)[name = string("tgt_cast_fp16")];
            tensor<fp16, [3, 15]> var_3438_cast_fp16 = sub(x = tgt_cast_fp16, y = lse_cast_fp16)[name = string("op_3438_cast_fp16")];
            string var_3438_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_3438_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [3, 15]> lp = cast(dtype = var_3438_cast_fp16_to_fp32_dtype_0, x = var_3438_cast_fp16)[name = string("cast_71")];
        } -> (lp);
    func t32<ios18>(tensor<int32, [3, 32]> ids) {
            int32 x_1_batch_dims_0 = const()[name = string("x_1_batch_dims_0"), val = int32(0)];
            bool x_1_validate_indices_0 = const()[name = string("x_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [16000, 640]> emb_to_fp16 = const()[name = string("emb_to_fp16"), val = tensor<fp16, [16000, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            string ids_to_int16_dtype_0 = const()[name = string("ids_to_int16_dtype_0"), val = string("int16")];
            string cast_82_dtype_0 = const()[name = string("cast_82_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int16, [3, 32]> ids_to_int16 = cast(dtype = ids_to_int16_dtype_0, x = ids)[name = string("cast_87")];
            tensor<int32, [3, 32]> cast_82 = cast(dtype = cast_82_dtype_0, x = ids_to_int16)[name = string("cast_86")];
            tensor<bool, [3, 32]> greater_equal_0 = greater_equal(x = cast_82, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(16000)];
            tensor<int32, [3, 32]> add_0 = add(x = cast_82, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [3, 32]> select_0 = select(a = cast_82, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 x_1_cast_fp16_cast_uint16_axis_0 = const()[name = string("x_1_cast_fp16_cast_uint16_axis_0"), val = int32(0)];
            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
            tensor<int16, [3, 32]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_85")];
            tensor<fp16, [3, 32, 640]> x_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = x_1_cast_fp16_cast_uint16_axis_0, batch_dims = x_1_batch_dims_0, indices = select_0_to_int16, validate_indices = x_1_validate_indices_0, x = emb_to_fp16)[name = string("x_1_cast_fp16_cast_uint16_cast_uint16")];
            fp16 var_253_to_fp16 = const()[name = string("op_253_to_fp16"), val = fp16(0x1p+5)];
            tensor<fp16, [3, 32, 640]> x_3_cast_fp16 = mul(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = var_253_to_fp16)[name = string("x_3_cast_fp16")];
            fp16 var_251_promoted_to_fp16 = const()[name = string("op_251_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> var_255_cast_fp16 = pow(x = x_3_cast_fp16, y = var_251_promoted_to_fp16)[name = string("op_255_cast_fp16")];
            tensor<int32, [1]> var_257_axes_0 = const()[name = string("op_257_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_257_keep_dims_0 = const()[name = string("op_257_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 1]> var_257_cast_fp16 = reduce_mean(axes = var_257_axes_0, keep_dims = var_257_keep_dims_0, x = var_255_cast_fp16)[name = string("op_257_cast_fp16")];
            fp16 var_258_to_fp16 = const()[name = string("op_258_to_fp16"), val = fp16(0x1.0c8p-10)];
            tensor<fp16, [3, 32, 1]> var_259_cast_fp16 = add(x = var_257_cast_fp16, y = var_258_to_fp16)[name = string("op_259_cast_fp16")];
            fp32 var_260_epsilon_0 = const()[name = string("op_260_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 1]> var_260_cast_fp16 = rsqrt(epsilon = var_260_epsilon_0, x = var_259_cast_fp16)[name = string("op_260_cast_fp16")];
            tensor<fp16, [3, 32, 640]> var_261_cast_fp16 = mul(x = x_3_cast_fp16, y = var_260_cast_fp16)[name = string("op_261_cast_fp16")];
            tensor<fp16, [640]> layers_0_n1_w1_to_fp16 = const()[name = string("layers_0_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20480128)))];
            tensor<fp16, [3, 32, 640]> input_1_cast_fp16 = mul(x = var_261_cast_fp16, y = layers_0_n1_w1_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_0_qkvzba_weight_to_fp16 = const()[name = string("layers_0_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20481472)))];
            tensor<fp16, [1544]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [1544]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22457856)))];
            tensor<fp16, [3, 32, 1544]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_qkvzba_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [3]> qkv_1_begin_0 = const()[name = string("qkv_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_1_end_0 = const()[name = string("qkv_1_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> qkv_1_end_mask_0 = const()[name = string("qkv_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 1024]> qkv_1_cast_fp16 = slice_by_index(begin = qkv_1_begin_0, end = qkv_1_end_0, end_mask = qkv_1_end_mask_0, x = linear_0_cast_fp16)[name = string("qkv_1_cast_fp16")];
            tensor<int32, [3]> z_1_begin_0 = const()[name = string("z_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_1_end_0 = const()[name = string("z_1_end_0"), val = tensor<int32, [3]>([3, 32, 1536])];
            tensor<bool, [3]> z_1_end_mask_0 = const()[name = string("z_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 512]> z_1_cast_fp16 = slice_by_index(begin = z_1_begin_0, end = z_1_end_0, end_mask = z_1_end_mask_0, x = linear_0_cast_fp16)[name = string("z_1_cast_fp16")];
            tensor<int32, [3]> b_1_begin_0 = const()[name = string("b_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_1_end_0 = const()[name = string("b_1_end_0"), val = tensor<int32, [3]>([3, 32, 1540])];
            tensor<bool, [3]> b_1_end_mask_0 = const()[name = string("b_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 4]> b_1_cast_fp16 = slice_by_index(begin = b_1_begin_0, end = b_1_end_0, end_mask = b_1_end_mask_0, x = linear_0_cast_fp16)[name = string("b_1_cast_fp16")];
            tensor<int32, [3]> a_1_begin_0 = const()[name = string("a_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_1_end_0 = const()[name = string("a_1_end_0"), val = tensor<int32, [3]>([3, 32, 1])];
            tensor<bool, [3]> a_1_end_mask_0 = const()[name = string("a_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 4]> a_1_cast_fp16 = slice_by_index(begin = a_1_begin_0, end = a_1_end_0, end_mask = a_1_end_mask_0, x = linear_0_cast_fp16)[name = string("a_1_cast_fp16")];
            int32 var_300 = const()[name = string("op_300"), val = int32(1)];
            bool xp_1_interleave_0 = const()[name = string("xp_1_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 3, 1024]> zpad_to_fp16 = const()[name = string("zpad_to_fp16"), val = tensor<fp16, [3, 3, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22461056)))];
            tensor<fp16, [3, 35, 1024]> xp_1_cast_fp16 = concat(axis = var_300, interleave = xp_1_interleave_0, values = (zpad_to_fp16, qkv_1_cast_fp16))[name = string("xp_1_cast_fp16")];
            tensor<int32, [3]> var_311_begin_0 = const()[name = string("op_311_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_311_end_0 = const()[name = string("op_311_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_311_end_mask_0 = const()[name = string("op_311_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 1024]> var_311_cast_fp16 = slice_by_index(begin = var_311_begin_0, end = var_311_end_0, end_mask = var_311_end_mask_0, x = xp_1_cast_fp16)[name = string("op_311_cast_fp16")];
            tensor<fp16, [1024]> var_319_to_fp16 = const()[name = string("op_319_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22479552)))];
            tensor<fp16, [3, 32, 1024]> var_320_cast_fp16 = mul(x = var_311_cast_fp16, y = var_319_to_fp16)[name = string("op_320_cast_fp16")];
            tensor<int32, [3]> var_330_begin_0 = const()[name = string("op_330_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_330_end_0 = const()[name = string("op_330_end_0"), val = tensor<int32, [3]>([3, 34, 1024])];
            tensor<bool, [3]> var_330_end_mask_0 = const()[name = string("op_330_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_330_cast_fp16 = slice_by_index(begin = var_330_begin_0, end = var_330_end_0, end_mask = var_330_end_mask_0, x = xp_1_cast_fp16)[name = string("op_330_cast_fp16")];
            tensor<fp16, [1024]> var_338_to_fp16 = const()[name = string("op_338_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22481664)))];
            tensor<fp16, [3, 32, 1024]> var_339_cast_fp16 = mul(x = var_330_cast_fp16, y = var_338_to_fp16)[name = string("op_339_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> var_341_cast_fp16 = add(x = var_320_cast_fp16, y = var_339_cast_fp16)[name = string("op_341_cast_fp16")];
            tensor<int32, [3]> var_351_begin_0 = const()[name = string("op_351_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_351_end_0 = const()[name = string("op_351_end_0"), val = tensor<int32, [3]>([3, 33, 1024])];
            tensor<bool, [3]> var_351_end_mask_0 = const()[name = string("op_351_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_351_cast_fp16 = slice_by_index(begin = var_351_begin_0, end = var_351_end_0, end_mask = var_351_end_mask_0, x = xp_1_cast_fp16)[name = string("op_351_cast_fp16")];
            tensor<fp16, [1024]> var_359_to_fp16 = const()[name = string("op_359_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22483776)))];
            tensor<fp16, [3, 32, 1024]> var_360_cast_fp16 = mul(x = var_351_cast_fp16, y = var_359_to_fp16)[name = string("op_360_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> var_362_cast_fp16 = add(x = var_341_cast_fp16, y = var_360_cast_fp16)[name = string("op_362_cast_fp16")];
            tensor<int32, [3]> var_372_begin_0 = const()[name = string("op_372_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_372_end_0 = const()[name = string("op_372_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> var_372_end_mask_0 = const()[name = string("op_372_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_372_cast_fp16 = slice_by_index(begin = var_372_begin_0, end = var_372_end_0, end_mask = var_372_end_mask_0, x = xp_1_cast_fp16)[name = string("op_372_cast_fp16")];
            tensor<fp16, [1024]> var_380_to_fp16 = const()[name = string("op_380_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22485888)))];
            tensor<fp16, [3, 32, 1024]> var_381_cast_fp16 = mul(x = var_372_cast_fp16, y = var_380_to_fp16)[name = string("op_381_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> input_3_cast_fp16 = add(x = var_362_cast_fp16, y = var_381_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> c_1_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("c_1_cast_fp16")];
            tensor<int32, [3]> var_389_begin_0 = const()[name = string("op_389_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_389_end_0 = const()[name = string("op_389_end_0"), val = tensor<int32, [3]>([3, 32, 256])];
            tensor<bool, [3]> var_389_end_mask_0 = const()[name = string("op_389_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 256]> var_389_cast_fp16 = slice_by_index(begin = var_389_begin_0, end = var_389_end_0, end_mask = var_389_end_mask_0, x = c_1_cast_fp16)[name = string("op_389_cast_fp16")];
            tensor<int32, [4]> var_394 = const()[name = string("op_394"), val = tensor<int32, [4]>([3, 32, 2, 128])];
            tensor<fp16, [3, 32, 2, 128]> q_1_cast_fp16 = reshape(shape = var_394, x = var_389_cast_fp16)[name = string("q_1_cast_fp16")];
            tensor<int32, [3]> var_400_begin_0 = const()[name = string("op_400_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_400_end_0 = const()[name = string("op_400_end_0"), val = tensor<int32, [3]>([3, 32, 512])];
            tensor<bool, [3]> var_400_end_mask_0 = const()[name = string("op_400_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 256]> var_400_cast_fp16 = slice_by_index(begin = var_400_begin_0, end = var_400_end_0, end_mask = var_400_end_mask_0, x = c_1_cast_fp16)[name = string("op_400_cast_fp16")];
            tensor<int32, [4]> var_405 = const()[name = string("op_405"), val = tensor<int32, [4]>([3, 32, 2, 128])];
            tensor<fp16, [3, 32, 2, 128]> k_1_cast_fp16 = reshape(shape = var_405, x = var_400_cast_fp16)[name = string("k_1_cast_fp16")];
            tensor<int32, [3]> var_411_begin_0 = const()[name = string("op_411_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_411_end_0 = const()[name = string("op_411_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> var_411_end_mask_0 = const()[name = string("op_411_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 512]> var_411_cast_fp16 = slice_by_index(begin = var_411_begin_0, end = var_411_end_0, end_mask = var_411_end_mask_0, x = c_1_cast_fp16)[name = string("op_411_cast_fp16")];
            tensor<int32, [4]> var_416 = const()[name = string("op_416"), val = tensor<int32, [4]>([3, 32, 4, 128])];
            tensor<fp16, [3, 32, 4, 128]> var_417_cast_fp16 = reshape(shape = var_416, x = var_411_cast_fp16)[name = string("op_417_cast_fp16")];
            tensor<int32, [4]> var_422 = const()[name = string("op_422"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_424_to_fp16 = const()[name = string("op_424_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 32, 2, 128]> q_3_cast_fp16 = mul(x = q_1_cast_fp16, y = var_424_to_fp16)[name = string("q_3_cast_fp16")];
            fp16 var_426_to_fp16 = const()[name = string("op_426_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 32, 2, 128]> k_3_cast_fp16 = mul(x = k_1_cast_fp16, y = var_426_to_fp16)[name = string("k_3_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_428_cast_fp16 = mul(x = q_3_cast_fp16, y = q_3_cast_fp16)[name = string("op_428_cast_fp16")];
            tensor<int32, [1]> var_433_axes_0 = const()[name = string("op_433_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_433_keep_dims_0 = const()[name = string("op_433_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 2, 1]> var_433_cast_fp16 = reduce_mean(axes = var_433_axes_0, keep_dims = var_433_keep_dims_0, x = var_428_cast_fp16)[name = string("op_433_cast_fp16")];
            fp16 var_434_promoted_to_fp16 = const()[name = string("op_434_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 32, 2, 1]> var_435_cast_fp16 = mul(x = var_433_cast_fp16, y = var_434_promoted_to_fp16)[name = string("op_435_cast_fp16")];
            fp16 var_437_to_fp16 = const()[name = string("op_437_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 32, 2, 1]> var_438_cast_fp16 = add(x = var_435_cast_fp16, y = var_437_to_fp16)[name = string("op_438_cast_fp16")];
            fp32 var_439_epsilon_0 = const()[name = string("op_439_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 2, 1]> var_439_cast_fp16 = rsqrt(epsilon = var_439_epsilon_0, x = var_438_cast_fp16)[name = string("op_439_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_440_cast_fp16 = mul(x = q_3_cast_fp16, y = var_439_cast_fp16)[name = string("op_440_cast_fp16")];
            fp16 var_441_to_fp16 = const()[name = string("op_441_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 32, 2, 128]> q_5_cast_fp16 = mul(x = var_440_cast_fp16, y = var_441_to_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_443_cast_fp16 = mul(x = k_3_cast_fp16, y = k_3_cast_fp16)[name = string("op_443_cast_fp16")];
            tensor<int32, [1]> var_448_axes_0 = const()[name = string("op_448_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_448_keep_dims_0 = const()[name = string("op_448_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 2, 1]> var_448_cast_fp16 = reduce_mean(axes = var_448_axes_0, keep_dims = var_448_keep_dims_0, x = var_443_cast_fp16)[name = string("op_448_cast_fp16")];
            fp16 var_449_promoted_to_fp16 = const()[name = string("op_449_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 32, 2, 1]> var_450_cast_fp16 = mul(x = var_448_cast_fp16, y = var_449_promoted_to_fp16)[name = string("op_450_cast_fp16")];
            fp16 var_452_to_fp16 = const()[name = string("op_452_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 32, 2, 1]> var_453_cast_fp16 = add(x = var_450_cast_fp16, y = var_452_to_fp16)[name = string("op_453_cast_fp16")];
            fp32 var_454_epsilon_0 = const()[name = string("op_454_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 2, 1]> var_454_cast_fp16 = rsqrt(epsilon = var_454_epsilon_0, x = var_453_cast_fp16)[name = string("op_454_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> k_5_cast_fp16 = mul(x = k_3_cast_fp16, y = var_454_cast_fp16)[name = string("k_5_cast_fp16")];
            tensor<int32, [4]> var_460 = const()[name = string("op_460"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_473_axes_0 = const()[name = string("op_473_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 32, 128]> var_461_cast_fp16 = transpose(perm = var_460, x = q_5_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [3, 2, 1, 32, 128]> var_473_cast_fp16 = expand_dims(axes = var_473_axes_0, x = var_461_cast_fp16)[name = string("op_473_cast_fp16")];
            tensor<int32, [5]> var_481_reps_0 = const()[name = string("op_481_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 32, 128]> var_481_cast_fp16 = tile(reps = var_481_reps_0, x = var_473_cast_fp16)[name = string("op_481_cast_fp16")];
            tensor<int32, [4]> var_486 = const()[name = string("op_486"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<fp16, [3, 4, 32, 128]> q_7_cast_fp16 = reshape(shape = var_486, x = var_481_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<int32, [4]> var_492 = const()[name = string("op_492"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_505_axes_0 = const()[name = string("op_505_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 32, 128]> var_493_cast_fp16 = transpose(perm = var_492, x = k_5_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [3, 2, 1, 32, 128]> var_505_cast_fp16 = expand_dims(axes = var_505_axes_0, x = var_493_cast_fp16)[name = string("op_505_cast_fp16")];
            tensor<int32, [5]> var_513_reps_0 = const()[name = string("op_513_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 32, 128]> var_513_cast_fp16 = tile(reps = var_513_reps_0, x = var_505_cast_fp16)[name = string("op_513_cast_fp16")];
            tensor<int32, [4]> var_518 = const()[name = string("op_518"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<fp16, [3, 4, 32, 128]> k_7_cast_fp16 = reshape(shape = var_518, x = var_513_cast_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_520_cast_fp16 = sigmoid(x = b_1_cast_fp16)[name = string("op_520_cast_fp16")];
            tensor<int32, [3]> var_524 = const()[name = string("op_524"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_0_dt_bias_to_fp16 = const()[name = string("layers_0_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 32, 4]> input_5_cast_fp16 = add(x = a_1_cast_fp16, y = layers_0_dt_bias_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_528_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("op_528_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_529_cast_fp16 = abs(x = input_5_cast_fp16)[name = string("op_529_cast_fp16")];
            fp16 const_0_promoted_to_fp16 = const()[name = string("const_0_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 32, 4]> var_530_cast_fp16 = mul(x = var_529_cast_fp16, y = const_0_promoted_to_fp16)[name = string("op_530_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_531_cast_fp16 = exp(x = var_530_cast_fp16)[name = string("op_531_cast_fp16")];
            fp16 var_533_to_fp16 = const()[name = string("op_533_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 32, 4]> var_534_cast_fp16 = add(x = var_531_cast_fp16, y = var_533_to_fp16)[name = string("op_534_cast_fp16")];
            fp32 var_535_epsilon_0 = const()[name = string("op_535_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 32, 4]> var_535_cast_fp16 = log(epsilon = var_535_epsilon_0, x = var_534_cast_fp16)[name = string("op_535_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_537_cast_fp16 = add(x = var_528_cast_fp16, y = var_535_cast_fp16)[name = string("op_537_cast_fp16")];
            tensor<fp16, [4]> layers_0_negA_to_fp16 = const()[name = string("layers_0_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.b3p+3, -0x1.a7cp+0, -0x1.ac4p+3, -0x1.704p+2])];
            tensor<fp16, [3, 32, 4]> var_538_cast_fp16 = mul(x = layers_0_negA_to_fp16, y = var_537_cast_fp16)[name = string("op_538_cast_fp16")];
            tensor<int32, [3]> var_542 = const()[name = string("op_542"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_545_axes_0 = const()[name = string("op_545_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 32]> beta_1_cast_fp16 = transpose(perm = var_524, x = var_520_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [3, 4, 32, 1]> var_545_cast_fp16 = expand_dims(axes = var_545_axes_0, x = beta_1_cast_fp16)[name = string("op_545_cast_fp16")];
            fp16 var_546_to_fp16 = const()[name = string("op_546_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 32, 1]> var_547_cast_fp16 = mul(x = var_545_cast_fp16, y = var_546_to_fp16)[name = string("op_547_cast_fp16")];
            tensor<fp16, [3, 4, 32, 128]> v_1_cast_fp16 = transpose(perm = var_422, x = var_417_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [3, 4, 32, 128]> vb_1_cast_fp16 = mul(x = v_1_cast_fp16, y = var_547_cast_fp16)[name = string("vb_1_cast_fp16")];
            tensor<int32, [4]> q_9_begin_0 = const()[name = string("q_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_9_end_0 = const()[name = string("q_9_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_9_end_mask_0 = const()[name = string("q_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_9_cast_fp16 = slice_by_index(begin = q_9_begin_0, end = q_9_end_0, end_mask = q_9_end_mask_0, x = q_7_cast_fp16)[name = string("q_9_cast_fp16")];
            tensor<int32, [4]> k_9_begin_0 = const()[name = string("k_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_9_end_0 = const()[name = string("k_9_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_9_end_mask_0 = const()[name = string("k_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, x = k_7_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<fp16, [3, 4, 32, 128]> var_581_cast_fp16 = mul(x = k_7_cast_fp16, y = var_545_cast_fp16)[name = string("op_581_cast_fp16")];
            tensor<int32, [4]> kb_1_begin_0 = const()[name = string("kb_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_1_end_0 = const()[name = string("kb_1_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_1_end_mask_0 = const()[name = string("kb_1_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_1_cast_fp16 = slice_by_index(begin = kb_1_begin_0, end = kb_1_end_0, end_mask = kb_1_end_mask_0, x = var_581_cast_fp16)[name = string("kb_1_cast_fp16")];
            tensor<int32, [4]> vb_3_begin_0 = const()[name = string("vb_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_3_end_0 = const()[name = string("vb_3_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_3_end_mask_0 = const()[name = string("vb_3_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_3_cast_fp16 = slice_by_index(begin = vb_3_begin_0, end = vb_3_end_0, end_mask = vb_3_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_3_cast_fp16")];
            tensor<int32, [3]> gdec_3_begin_0 = const()[name = string("gdec_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_3_end_0 = const()[name = string("gdec_3_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_3_end_mask_0 = const()[name = string("gdec_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 32]> gdec_1_cast_fp16 = transpose(perm = var_542, x = var_538_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [3, 4, 16]> gdec_3_cast_fp16 = slice_by_index(begin = gdec_3_begin_0, end = gdec_3_end_0, end_mask = gdec_3_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_3_cast_fp16")];
            tensor<int32, [2]> var_629 = const()[name = string("op_629"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_630_cast_fp16 = reshape(shape = var_629, x = gdec_3_cast_fp16)[name = string("op_630_cast_fp16")];
            tensor<fp16, [256, 16]> transpose_0_to_fp16 = const()[name = string("transpose_0_to_fp16"), val = tensor<fp16, [256, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22488000)))];
            tensor<fp16, [256]> var_631_bias_0_to_fp16 = const()[name = string("op_631_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22496256)))];
            tensor<fp16, [12, 256]> var_631_cast_fp16 = linear(bias = var_631_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_630_cast_fp16)[name = string("op_631_cast_fp16")];
            tensor<int32, [4]> var_636 = const()[name = string("op_636"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_1_cast_fp16 = reshape(shape = var_636, x = var_631_cast_fp16)[name = string("P_1_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_638_cast_fp16 = exp(x = P_1_cast_fp16)[name = string("op_638_cast_fp16")];
            tensor<fp16, [16, 16]> tril_incl_to_fp16 = const()[name = string("tril_incl_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22496832)))];
            tensor<fp16, [3, 4, 16, 16]> D_1_cast_fp16 = mul(x = var_638_cast_fp16, y = tril_incl_to_fp16)[name = string("D_1_cast_fp16")];
            bool kk_1_transpose_x_1 = const()[name = string("kk_1_transpose_x_1"), val = bool(false)];
            bool kk_1_transpose_y_1 = const()[name = string("kk_1_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_1_cast_fp16 = matmul(transpose_x = kk_1_transpose_x_1, transpose_y = kk_1_transpose_y_1, x = kb_1_cast_fp16, y = k_9_cast_fp16)[name = string("kk_1_cast_fp16")];
            fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_644_cast_fp16 = mul(x = kk_1_cast_fp16, y = const_1_promoted_to_fp16)[name = string("op_644_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_645_cast_fp16 = mul(x = var_644_cast_fp16, y = D_1_cast_fp16)[name = string("op_645_cast_fp16")];
            tensor<fp16, [16, 16]> cum_tail_to_fp16 = const()[name = string("cum_tail_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22497408)))];
            tensor<fp16, [3, 4, 16, 16]> N_1_cast_fp16 = mul(x = var_645_cast_fp16, y = cum_tail_to_fp16)[name = string("N_1_cast_fp16")];
            tensor<fp16, [16, 16]> neg_tril_strict_to_fp16 = const()[name = string("neg_tril_strict_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22497984)))];
            tensor<fp16, [3, 4, 16, 16]> var_647_cast_fp16 = mul(x = D_1_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_647_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_1_cast_fp16 = mul(x = kk_1_cast_fp16, y = var_647_cast_fp16)[name = string("Na_1_cast_fp16")];
            tensor<fp16, [16, 16]> eye_to_fp16 = const()[name = string("eye_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22498560)))];
            tensor<fp16, [3, 4, 16, 16]> Minv_1_cast_fp16 = add(x = eye_to_fp16, y = N_1_cast_fp16)[name = string("Minv_1_cast_fp16")];
            bool Np_1_transpose_x_0 = const()[name = string("Np_1_transpose_x_0"), val = bool(false)];
            bool Np_1_transpose_y_0 = const()[name = string("Np_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_1_cast_fp16 = matmul(transpose_x = Np_1_transpose_x_0, transpose_y = Np_1_transpose_y_0, x = N_1_cast_fp16, y = Na_1_cast_fp16)[name = string("Np_1_cast_fp16")];
            bool Npa_1_transpose_x_0 = const()[name = string("Npa_1_transpose_x_0"), val = bool(false)];
            bool Npa_1_transpose_y_0 = const()[name = string("Npa_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_1_cast_fp16 = matmul(transpose_x = Npa_1_transpose_x_0, transpose_y = Npa_1_transpose_y_0, x = Na_1_cast_fp16, y = N_1_cast_fp16)[name = string("Npa_1_cast_fp16")];
            bool var_653_transpose_x_0 = const()[name = string("op_653_transpose_x_0"), val = bool(false)];
            bool var_653_transpose_y_0 = const()[name = string("op_653_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_653_cast_fp16 = matmul(transpose_x = var_653_transpose_x_0, transpose_y = var_653_transpose_y_0, x = Minv_1_cast_fp16, y = Np_1_cast_fp16)[name = string("op_653_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_3_cast_fp16 = add(x = Minv_1_cast_fp16, y = var_653_cast_fp16)[name = string("Minv_3_cast_fp16")];
            bool Np_3_transpose_x_0 = const()[name = string("Np_3_transpose_x_0"), val = bool(false)];
            bool Np_3_transpose_y_0 = const()[name = string("Np_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_3_cast_fp16 = matmul(transpose_x = Np_3_transpose_x_0, transpose_y = Np_3_transpose_y_0, x = Np_1_cast_fp16, y = Npa_1_cast_fp16)[name = string("Np_3_cast_fp16")];
            bool Npa_3_transpose_x_0 = const()[name = string("Npa_3_transpose_x_0"), val = bool(false)];
            bool Npa_3_transpose_y_0 = const()[name = string("Npa_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_3_cast_fp16 = matmul(transpose_x = Npa_3_transpose_x_0, transpose_y = Npa_3_transpose_y_0, x = Npa_1_cast_fp16, y = Np_1_cast_fp16)[name = string("Npa_3_cast_fp16")];
            bool var_658_transpose_x_0 = const()[name = string("op_658_transpose_x_0"), val = bool(false)];
            bool var_658_transpose_y_0 = const()[name = string("op_658_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_658_cast_fp16 = matmul(transpose_x = var_658_transpose_x_0, transpose_y = var_658_transpose_y_0, x = Minv_3_cast_fp16, y = Np_3_cast_fp16)[name = string("op_658_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_5_cast_fp16 = add(x = Minv_3_cast_fp16, y = var_658_cast_fp16)[name = string("Minv_5_cast_fp16")];
            bool Np_5_transpose_x_0 = const()[name = string("Np_5_transpose_x_0"), val = bool(false)];
            bool Np_5_transpose_y_0 = const()[name = string("Np_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_5_cast_fp16 = matmul(transpose_x = Np_5_transpose_x_0, transpose_y = Np_5_transpose_y_0, x = Np_3_cast_fp16, y = Npa_3_cast_fp16)[name = string("Np_5_cast_fp16")];
            bool var_662_transpose_x_0 = const()[name = string("op_662_transpose_x_0"), val = bool(false)];
            bool var_662_transpose_y_0 = const()[name = string("op_662_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_662_cast_fp16 = matmul(transpose_x = var_662_transpose_x_0, transpose_y = var_662_transpose_y_0, x = Minv_5_cast_fp16, y = Np_5_cast_fp16)[name = string("op_662_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_7_cast_fp16 = add(x = Minv_5_cast_fp16, y = var_662_cast_fp16)[name = string("Minv_7_cast_fp16")];
            bool u_1_transpose_x_0 = const()[name = string("u_1_transpose_x_0"), val = bool(false)];
            bool u_1_transpose_y_0 = const()[name = string("u_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_1_cast_fp16 = matmul(transpose_x = u_1_transpose_x_0, transpose_y = u_1_transpose_y_0, x = Minv_7_cast_fp16, y = vb_3_cast_fp16)[name = string("u_1_cast_fp16")];
            bool var_669_transpose_x_1 = const()[name = string("op_669_transpose_x_1"), val = bool(false)];
            bool var_669_transpose_y_1 = const()[name = string("op_669_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_669_cast_fp16 = matmul(transpose_x = var_669_transpose_x_1, transpose_y = var_669_transpose_y_1, x = q_9_cast_fp16, y = k_9_cast_fp16)[name = string("op_669_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_1_cast_fp16 = mul(x = var_669_cast_fp16, y = D_1_cast_fp16)[name = string("A_1_cast_fp16")];
            bool var_671_transpose_x_0 = const()[name = string("op_671_transpose_x_0"), val = bool(false)];
            bool var_671_transpose_y_0 = const()[name = string("op_671_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_671_cast_fp16 = matmul(transpose_x = var_671_transpose_x_0, transpose_y = var_671_transpose_y_0, x = A_1_cast_fp16, y = u_1_cast_fp16)[name = string("op_671_cast_fp16")];
            tensor<fp16, [16, 16]> transpose_1_to_fp16 = const()[name = string("transpose_1_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658176)))];
            tensor<fp16, [16]> var_672_bias_0_to_fp16 = const()[name = string("op_672_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_672_cast_fp16 = linear(bias = var_672_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_3_cast_fp16)[name = string("op_672_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_1_cast_fp16 = exp(x = var_672_cast_fp16)[name = string("tail_1_cast_fp16")];
            tensor<int32, [1]> var_675_axes_0 = const()[name = string("op_675_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_675_cast_fp16 = expand_dims(axes = var_675_axes_0, x = tail_1_cast_fp16)[name = string("op_675_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_676_cast_fp16 = mul(x = k_9_cast_fp16, y = var_675_cast_fp16)[name = string("op_676_cast_fp16")];
            bool S_1_transpose_x_1 = const()[name = string("S_1_transpose_x_1"), val = bool(true)];
            bool S_1_transpose_y_1 = const()[name = string("S_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_1_cast_fp16 = matmul(transpose_x = S_1_transpose_x_1, transpose_y = S_1_transpose_y_1, x = var_676_cast_fp16, y = u_1_cast_fp16)[name = string("S_1_cast_fp16")];
            tensor<int32, [4]> q_11_begin_0 = const()[name = string("q_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_11_end_0 = const()[name = string("q_11_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_11_end_mask_0 = const()[name = string("q_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_11_cast_fp16 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, x = q_7_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<int32, [4]> k_11_begin_0 = const()[name = string("k_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_11_end_0 = const()[name = string("k_11_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_11_end_mask_0 = const()[name = string("k_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_11_cast_fp16 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, x = k_7_cast_fp16)[name = string("k_11_cast_fp16")];
            tensor<int32, [4]> kb_3_begin_0 = const()[name = string("kb_3_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_3_end_0 = const()[name = string("kb_3_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_3_end_mask_0 = const()[name = string("kb_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_3_cast_fp16 = slice_by_index(begin = kb_3_begin_0, end = kb_3_end_0, end_mask = kb_3_end_mask_0, x = var_581_cast_fp16)[name = string("kb_3_cast_fp16")];
            tensor<int32, [4]> vb_5_begin_0 = const()[name = string("vb_5_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_5_end_0 = const()[name = string("vb_5_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_5_end_mask_0 = const()[name = string("vb_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_5_cast_fp16 = slice_by_index(begin = vb_5_begin_0, end = vb_5_end_0, end_mask = vb_5_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_5_cast_fp16")];
            tensor<int32, [3]> gdec_5_begin_0 = const()[name = string("gdec_5_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_5_end_0 = const()[name = string("gdec_5_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_5_end_mask_0 = const()[name = string("gdec_5_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_5_cast_fp16 = slice_by_index(begin = gdec_5_begin_0, end = gdec_5_end_0, end_mask = gdec_5_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_5_cast_fp16")];
            tensor<int32, [2]> var_761 = const()[name = string("op_761"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_762_cast_fp16 = reshape(shape = var_761, x = gdec_5_cast_fp16)[name = string("op_762_cast_fp16")];
            tensor<fp16, [12, 256]> var_763_cast_fp16 = linear(bias = var_631_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_762_cast_fp16)[name = string("op_763_cast_fp16")];
            tensor<int32, [4]> var_768 = const()[name = string("op_768"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_3_cast_fp16 = reshape(shape = var_768, x = var_763_cast_fp16)[name = string("P_3_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_770_cast_fp16 = exp(x = P_3_cast_fp16)[name = string("op_770_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_3_cast_fp16 = mul(x = var_770_cast_fp16, y = tril_incl_to_fp16)[name = string("D_3_cast_fp16")];
            bool kk_3_transpose_x_1 = const()[name = string("kk_3_transpose_x_1"), val = bool(false)];
            bool kk_3_transpose_y_1 = const()[name = string("kk_3_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_3_cast_fp16 = matmul(transpose_x = kk_3_transpose_x_1, transpose_y = kk_3_transpose_y_1, x = kb_3_cast_fp16, y = k_11_cast_fp16)[name = string("kk_3_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_776_cast_fp16 = mul(x = kk_3_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_776_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_777_cast_fp16 = mul(x = var_776_cast_fp16, y = D_3_cast_fp16)[name = string("op_777_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_3_cast_fp16 = mul(x = var_777_cast_fp16, y = cum_tail_to_fp16)[name = string("N_3_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_779_cast_fp16 = mul(x = D_3_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_779_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_3_cast_fp16 = mul(x = kk_3_cast_fp16, y = var_779_cast_fp16)[name = string("Na_3_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_9_cast_fp16 = add(x = eye_to_fp16, y = N_3_cast_fp16)[name = string("Minv_9_cast_fp16")];
            bool Np_7_transpose_x_0 = const()[name = string("Np_7_transpose_x_0"), val = bool(false)];
            bool Np_7_transpose_y_0 = const()[name = string("Np_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_7_cast_fp16 = matmul(transpose_x = Np_7_transpose_x_0, transpose_y = Np_7_transpose_y_0, x = N_3_cast_fp16, y = Na_3_cast_fp16)[name = string("Np_7_cast_fp16")];
            bool Npa_5_transpose_x_0 = const()[name = string("Npa_5_transpose_x_0"), val = bool(false)];
            bool Npa_5_transpose_y_0 = const()[name = string("Npa_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_5_cast_fp16 = matmul(transpose_x = Npa_5_transpose_x_0, transpose_y = Npa_5_transpose_y_0, x = Na_3_cast_fp16, y = N_3_cast_fp16)[name = string("Npa_5_cast_fp16")];
            bool var_785_transpose_x_0 = const()[name = string("op_785_transpose_x_0"), val = bool(false)];
            bool var_785_transpose_y_0 = const()[name = string("op_785_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_785_cast_fp16 = matmul(transpose_x = var_785_transpose_x_0, transpose_y = var_785_transpose_y_0, x = Minv_9_cast_fp16, y = Np_7_cast_fp16)[name = string("op_785_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_11_cast_fp16 = add(x = Minv_9_cast_fp16, y = var_785_cast_fp16)[name = string("Minv_11_cast_fp16")];
            bool Np_9_transpose_x_0 = const()[name = string("Np_9_transpose_x_0"), val = bool(false)];
            bool Np_9_transpose_y_0 = const()[name = string("Np_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_9_cast_fp16 = matmul(transpose_x = Np_9_transpose_x_0, transpose_y = Np_9_transpose_y_0, x = Np_7_cast_fp16, y = Npa_5_cast_fp16)[name = string("Np_9_cast_fp16")];
            bool Npa_7_transpose_x_0 = const()[name = string("Npa_7_transpose_x_0"), val = bool(false)];
            bool Npa_7_transpose_y_0 = const()[name = string("Npa_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_7_cast_fp16 = matmul(transpose_x = Npa_7_transpose_x_0, transpose_y = Npa_7_transpose_y_0, x = Npa_5_cast_fp16, y = Np_7_cast_fp16)[name = string("Npa_7_cast_fp16")];
            bool var_790_transpose_x_0 = const()[name = string("op_790_transpose_x_0"), val = bool(false)];
            bool var_790_transpose_y_0 = const()[name = string("op_790_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_790_cast_fp16 = matmul(transpose_x = var_790_transpose_x_0, transpose_y = var_790_transpose_y_0, x = Minv_11_cast_fp16, y = Np_9_cast_fp16)[name = string("op_790_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_13_cast_fp16 = add(x = Minv_11_cast_fp16, y = var_790_cast_fp16)[name = string("Minv_13_cast_fp16")];
            bool Np_11_transpose_x_0 = const()[name = string("Np_11_transpose_x_0"), val = bool(false)];
            bool Np_11_transpose_y_0 = const()[name = string("Np_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_11_cast_fp16 = matmul(transpose_x = Np_11_transpose_x_0, transpose_y = Np_11_transpose_y_0, x = Np_9_cast_fp16, y = Npa_7_cast_fp16)[name = string("Np_11_cast_fp16")];
            bool var_794_transpose_x_0 = const()[name = string("op_794_transpose_x_0"), val = bool(false)];
            bool var_794_transpose_y_0 = const()[name = string("op_794_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_794_cast_fp16 = matmul(transpose_x = var_794_transpose_x_0, transpose_y = var_794_transpose_y_0, x = Minv_13_cast_fp16, y = Np_11_cast_fp16)[name = string("op_794_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_15_cast_fp16 = add(x = Minv_13_cast_fp16, y = var_794_cast_fp16)[name = string("Minv_15_cast_fp16")];
            bool u_3_transpose_x_0 = const()[name = string("u_3_transpose_x_0"), val = bool(false)];
            bool u_3_transpose_y_0 = const()[name = string("u_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_3_cast_fp16 = matmul(transpose_x = u_3_transpose_x_0, transpose_y = u_3_transpose_y_0, x = Minv_15_cast_fp16, y = vb_5_cast_fp16)[name = string("u_3_cast_fp16")];
            bool var_801_transpose_x_1 = const()[name = string("op_801_transpose_x_1"), val = bool(false)];
            bool var_801_transpose_y_1 = const()[name = string("op_801_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_801_cast_fp16 = matmul(transpose_x = var_801_transpose_x_1, transpose_y = var_801_transpose_y_1, x = q_11_cast_fp16, y = k_11_cast_fp16)[name = string("op_801_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_3_cast_fp16 = mul(x = var_801_cast_fp16, y = D_3_cast_fp16)[name = string("A_3_cast_fp16")];
            tensor<fp16, [16]> var_803_bias_0_to_fp16 = const()[name = string("op_803_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_803_cast_fp16 = linear(bias = var_803_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_5_cast_fp16)[name = string("op_803_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_1_cast_fp16 = exp(x = var_803_cast_fp16)[name = string("cum_1_cast_fp16")];
            tensor<int32, [1]> var_806_axes_0 = const()[name = string("op_806_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_806_cast_fp16 = expand_dims(axes = var_806_axes_0, x = cum_1_cast_fp16)[name = string("op_806_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_807_cast_fp16 = mul(x = kb_3_cast_fp16, y = var_806_cast_fp16)[name = string("op_807_cast_fp16")];
            bool w_3_transpose_x_0 = const()[name = string("w_3_transpose_x_0"), val = bool(false)];
            bool w_3_transpose_y_0 = const()[name = string("w_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_3_cast_fp16 = matmul(transpose_x = w_3_transpose_x_0, transpose_y = w_3_transpose_y_0, x = Minv_15_cast_fp16, y = var_807_cast_fp16)[name = string("w_3_cast_fp16")];
            bool var_809_transpose_x_0 = const()[name = string("op_809_transpose_x_0"), val = bool(false)];
            bool var_809_transpose_y_0 = const()[name = string("op_809_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_809_cast_fp16 = matmul(transpose_x = var_809_transpose_x_0, transpose_y = var_809_transpose_y_0, x = w_3_cast_fp16, y = S_1_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_1_cast_fp16 = sub(x = u_3_cast_fp16, y = var_809_cast_fp16)[name = string("vnew_1_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_814_cast_fp16 = mul(x = q_11_cast_fp16, y = var_806_cast_fp16)[name = string("op_814_cast_fp16")];
            bool var_815_transpose_x_0 = const()[name = string("op_815_transpose_x_0"), val = bool(false)];
            bool var_815_transpose_y_0 = const()[name = string("op_815_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_815_cast_fp16 = matmul(transpose_x = var_815_transpose_x_0, transpose_y = var_815_transpose_y_0, x = var_814_cast_fp16, y = S_1_cast_fp16)[name = string("op_815_cast_fp16")];
            bool var_816_transpose_x_0 = const()[name = string("op_816_transpose_x_0"), val = bool(false)];
            bool var_816_transpose_y_0 = const()[name = string("op_816_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_816_cast_fp16 = matmul(transpose_x = var_816_transpose_x_0, transpose_y = var_816_transpose_y_0, x = A_3_cast_fp16, y = vnew_1_cast_fp16)[name = string("op_816_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_1_cast_fp16 = add(x = var_815_cast_fp16, y = var_816_cast_fp16)[name = string("oc_1_cast_fp16")];
            int32 var_820 = const()[name = string("op_820"), val = int32(2)];
            bool o_1_interleave_0 = const()[name = string("o_1_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 32, 128]> o_1_cast_fp16 = concat(axis = var_820, interleave = o_1_interleave_0, values = (var_671_cast_fp16, oc_1_cast_fp16))[name = string("o_1_cast_fp16")];
            tensor<int32, [4]> var_826 = const()[name = string("op_826"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_832 = const()[name = string("op_832"), val = tensor<int32, [4]>([3, 32, 4, 128])];
            tensor<fp16, [3, 32, 4, 128]> input_7_cast_fp16 = reshape(shape = var_832, x = z_1_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> o_3_cast_fp16 = transpose(perm = var_826, x = o_1_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [3, 32, 4, 128]> var_834_cast_fp16 = mul(x = o_3_cast_fp16, y = o_3_cast_fp16)[name = string("op_834_cast_fp16")];
            tensor<int32, [1]> var_839_axes_0 = const()[name = string("op_839_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_839_keep_dims_0 = const()[name = string("op_839_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 4, 1]> var_839_cast_fp16 = reduce_mean(axes = var_839_axes_0, keep_dims = var_839_keep_dims_0, x = var_834_cast_fp16)[name = string("op_839_cast_fp16")];
            fp16 var_841_to_fp16 = const()[name = string("op_841_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 32, 4, 1]> var_842_cast_fp16 = add(x = var_839_cast_fp16, y = var_841_to_fp16)[name = string("op_842_cast_fp16")];
            fp32 var_843_epsilon_0 = const()[name = string("op_843_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 4, 1]> var_843_cast_fp16 = rsqrt(epsilon = var_843_epsilon_0, x = var_842_cast_fp16)[name = string("op_843_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> var_844_cast_fp16 = mul(x = o_3_cast_fp16, y = var_843_cast_fp16)[name = string("op_844_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> var_846_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_846_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> o_5_cast_fp16 = mul(x = var_844_cast_fp16, y = var_846_cast_fp16)[name = string("o_5_cast_fp16")];
            tensor<int32, [3]> var_851 = const()[name = string("op_851"), val = tensor<int32, [3]>([3, 32, 512])];
            tensor<fp16, [3, 32, 512]> input_9_cast_fp16 = reshape(shape = var_851, x = o_5_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [640, 512]> layers_0_out_weight_to_fp16 = const()[name = string("layers_0_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22499136)))];
            tensor<fp16, [640]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23154560)))];
            tensor<fp16, [3, 32, 640]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_out_weight_to_fp16, x = input_9_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [3, 32, 640]> x_5_cast_fp16 = add(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = linear_1_cast_fp16)[name = string("x_5_cast_fp16")];
            fp16 var_864_to_fp16 = const()[name = string("op_864_to_fp16"), val = fp16(0x1p+5)];
            tensor<fp16, [3, 32, 640]> x_7_cast_fp16 = mul(x = x_5_cast_fp16, y = var_864_to_fp16)[name = string("x_7_cast_fp16")];
            fp16 var_862_promoted_to_fp16 = const()[name = string("op_862_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> var_866_cast_fp16 = pow(x = x_7_cast_fp16, y = var_862_promoted_to_fp16)[name = string("op_866_cast_fp16")];
            tensor<int32, [1]> var_868_axes_0 = const()[name = string("op_868_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_868_keep_dims_0 = const()[name = string("op_868_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 1]> var_868_cast_fp16 = reduce_mean(axes = var_868_axes_0, keep_dims = var_868_keep_dims_0, x = var_866_cast_fp16)[name = string("op_868_cast_fp16")];
            fp16 var_869_to_fp16 = const()[name = string("op_869_to_fp16"), val = fp16(0x1.0c8p-10)];
            tensor<fp16, [3, 32, 1]> var_870_cast_fp16 = add(x = var_868_cast_fp16, y = var_869_to_fp16)[name = string("op_870_cast_fp16")];
            fp32 var_871_epsilon_0 = const()[name = string("op_871_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 1]> var_871_cast_fp16 = rsqrt(epsilon = var_871_epsilon_0, x = var_870_cast_fp16)[name = string("op_871_cast_fp16")];
            tensor<fp16, [3, 32, 640]> var_872_cast_fp16 = mul(x = x_7_cast_fp16, y = var_871_cast_fp16)[name = string("op_872_cast_fp16")];
            tensor<fp16, [640]> layers_0_n2_w1_to_fp16 = const()[name = string("layers_0_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23155904)))];
            tensor<fp16, [3, 32, 640]> input_11_cast_fp16 = mul(x = var_872_cast_fp16, y = layers_0_n2_w1_to_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_0_gate_up_weight_to_fp16 = const()[name = string("layers_0_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23157248)))];
            tensor<fp16, [5120]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29710912)))];
            tensor<fp16, [3, 32, 5120]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_0_gate_up_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [3]> input_13_begin_0 = const()[name = string("input_13_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_13_end_0 = const()[name = string("input_13_end_0"), val = tensor<int32, [3]>([3, 32, 2560])];
            tensor<bool, [3]> input_13_end_mask_0 = const()[name = string("input_13_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 2560]> input_13_cast_fp16 = slice_by_index(begin = input_13_begin_0, end = input_13_end_0, end_mask = input_13_end_mask_0, x = linear_2_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [3, 32, 2560]> var_882_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_882_cast_fp16")];
            tensor<int32, [3]> var_887_begin_0 = const()[name = string("op_887_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_887_end_0 = const()[name = string("op_887_end_0"), val = tensor<int32, [3]>([3, 32, 5120])];
            tensor<bool, [3]> var_887_end_mask_0 = const()[name = string("op_887_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 2560]> var_887_cast_fp16 = slice_by_index(begin = var_887_begin_0, end = var_887_end_0, end_mask = var_887_end_mask_0, x = linear_2_cast_fp16)[name = string("op_887_cast_fp16")];
            tensor<fp16, [3, 32, 2560]> input_15_cast_fp16 = mul(x = var_882_cast_fp16, y = var_887_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_0_down_weight_to_fp16 = const()[name = string("layers_0_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29721216)))];
            tensor<fp16, [3, 32, 640]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_down_weight_to_fp16, x = input_15_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [3, 32, 640]> x_9_cast_fp16 = add(x = x_5_cast_fp16, y = linear_3_cast_fp16)[name = string("x_9_cast_fp16")];
            fp16 var_900_to_fp16 = const()[name = string("op_900_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> x_11_cast_fp16 = mul(x = x_9_cast_fp16, y = var_900_to_fp16)[name = string("x_11_cast_fp16")];
            fp16 var_898_promoted_to_fp16 = const()[name = string("op_898_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> var_902_cast_fp16 = pow(x = x_11_cast_fp16, y = var_898_promoted_to_fp16)[name = string("op_902_cast_fp16")];
            tensor<int32, [1]> var_904_axes_0 = const()[name = string("op_904_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_904_keep_dims_0 = const()[name = string("op_904_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 1]> var_904_cast_fp16 = reduce_mean(axes = var_904_axes_0, keep_dims = var_904_keep_dims_0, x = var_902_cast_fp16)[name = string("op_904_cast_fp16")];
            fp16 var_905_to_fp16 = const()[name = string("op_905_to_fp16"), val = fp16(0x1.0cp-18)];
            tensor<fp16, [3, 32, 1]> var_906_cast_fp16 = add(x = var_904_cast_fp16, y = var_905_to_fp16)[name = string("op_906_cast_fp16")];
            fp32 var_907_epsilon_0 = const()[name = string("op_907_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 1]> var_907_cast_fp16 = rsqrt(epsilon = var_907_epsilon_0, x = var_906_cast_fp16)[name = string("op_907_cast_fp16")];
            tensor<fp16, [3, 32, 640]> var_908_cast_fp16 = mul(x = x_11_cast_fp16, y = var_907_cast_fp16)[name = string("op_908_cast_fp16")];
            tensor<fp16, [640]> layers_1_n1_w1_to_fp16 = const()[name = string("layers_1_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32998080)))];
            tensor<fp16, [3, 32, 640]> input_17_cast_fp16 = mul(x = var_908_cast_fp16, y = layers_1_n1_w1_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_1_qkvzba_weight_to_fp16 = const()[name = string("layers_1_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32999424)))];
            tensor<fp16, [3, 32, 1544]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_qkvzba_weight_to_fp16, x = input_17_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<int32, [3]> qkv_3_begin_0 = const()[name = string("qkv_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_3_end_0 = const()[name = string("qkv_3_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> qkv_3_end_mask_0 = const()[name = string("qkv_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 1024]> qkv_3_cast_fp16 = slice_by_index(begin = qkv_3_begin_0, end = qkv_3_end_0, end_mask = qkv_3_end_mask_0, x = linear_4_cast_fp16)[name = string("qkv_3_cast_fp16")];
            tensor<int32, [3]> z_3_begin_0 = const()[name = string("z_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_3_end_0 = const()[name = string("z_3_end_0"), val = tensor<int32, [3]>([3, 32, 1536])];
            tensor<bool, [3]> z_3_end_mask_0 = const()[name = string("z_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 512]> z_3_cast_fp16 = slice_by_index(begin = z_3_begin_0, end = z_3_end_0, end_mask = z_3_end_mask_0, x = linear_4_cast_fp16)[name = string("z_3_cast_fp16")];
            tensor<int32, [3]> b_3_begin_0 = const()[name = string("b_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_3_end_0 = const()[name = string("b_3_end_0"), val = tensor<int32, [3]>([3, 32, 1540])];
            tensor<bool, [3]> b_3_end_mask_0 = const()[name = string("b_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 4]> b_3_cast_fp16 = slice_by_index(begin = b_3_begin_0, end = b_3_end_0, end_mask = b_3_end_mask_0, x = linear_4_cast_fp16)[name = string("b_3_cast_fp16")];
            tensor<int32, [3]> a_3_begin_0 = const()[name = string("a_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_3_end_0 = const()[name = string("a_3_end_0"), val = tensor<int32, [3]>([3, 32, 1])];
            tensor<bool, [3]> a_3_end_mask_0 = const()[name = string("a_3_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 4]> a_3_cast_fp16 = slice_by_index(begin = a_3_begin_0, end = a_3_end_0, end_mask = a_3_end_mask_0, x = linear_4_cast_fp16)[name = string("a_3_cast_fp16")];
            int32 var_947 = const()[name = string("op_947"), val = int32(1)];
            bool xp_3_interleave_0 = const()[name = string("xp_3_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 35, 1024]> xp_3_cast_fp16 = concat(axis = var_947, interleave = xp_3_interleave_0, values = (zpad_to_fp16, qkv_3_cast_fp16))[name = string("xp_3_cast_fp16")];
            tensor<int32, [3]> var_958_begin_0 = const()[name = string("op_958_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_958_end_0 = const()[name = string("op_958_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_958_end_mask_0 = const()[name = string("op_958_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 1024]> var_958_cast_fp16 = slice_by_index(begin = var_958_begin_0, end = var_958_end_0, end_mask = var_958_end_mask_0, x = xp_3_cast_fp16)[name = string("op_958_cast_fp16")];
            tensor<fp16, [1024]> var_966_to_fp16 = const()[name = string("op_966_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34975808)))];
            tensor<fp16, [3, 32, 1024]> var_967_cast_fp16 = mul(x = var_958_cast_fp16, y = var_966_to_fp16)[name = string("op_967_cast_fp16")];
            tensor<int32, [3]> var_977_begin_0 = const()[name = string("op_977_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_977_end_0 = const()[name = string("op_977_end_0"), val = tensor<int32, [3]>([3, 34, 1024])];
            tensor<bool, [3]> var_977_end_mask_0 = const()[name = string("op_977_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_977_cast_fp16 = slice_by_index(begin = var_977_begin_0, end = var_977_end_0, end_mask = var_977_end_mask_0, x = xp_3_cast_fp16)[name = string("op_977_cast_fp16")];
            tensor<fp16, [1024]> var_985_to_fp16 = const()[name = string("op_985_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34977920)))];
            tensor<fp16, [3, 32, 1024]> var_986_cast_fp16 = mul(x = var_977_cast_fp16, y = var_985_to_fp16)[name = string("op_986_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> var_988_cast_fp16 = add(x = var_967_cast_fp16, y = var_986_cast_fp16)[name = string("op_988_cast_fp16")];
            tensor<int32, [3]> var_998_begin_0 = const()[name = string("op_998_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_998_end_0 = const()[name = string("op_998_end_0"), val = tensor<int32, [3]>([3, 33, 1024])];
            tensor<bool, [3]> var_998_end_mask_0 = const()[name = string("op_998_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_998_cast_fp16 = slice_by_index(begin = var_998_begin_0, end = var_998_end_0, end_mask = var_998_end_mask_0, x = xp_3_cast_fp16)[name = string("op_998_cast_fp16")];
            tensor<fp16, [1024]> var_1006_to_fp16 = const()[name = string("op_1006_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34980032)))];
            tensor<fp16, [3, 32, 1024]> var_1007_cast_fp16 = mul(x = var_998_cast_fp16, y = var_1006_to_fp16)[name = string("op_1007_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> var_1009_cast_fp16 = add(x = var_988_cast_fp16, y = var_1007_cast_fp16)[name = string("op_1009_cast_fp16")];
            tensor<int32, [3]> var_1019_begin_0 = const()[name = string("op_1019_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1019_end_0 = const()[name = string("op_1019_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> var_1019_end_mask_0 = const()[name = string("op_1019_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_1019_cast_fp16 = slice_by_index(begin = var_1019_begin_0, end = var_1019_end_0, end_mask = var_1019_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1019_cast_fp16")];
            tensor<fp16, [1024]> var_1027_to_fp16 = const()[name = string("op_1027_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34982144)))];
            tensor<fp16, [3, 32, 1024]> var_1028_cast_fp16 = mul(x = var_1019_cast_fp16, y = var_1027_to_fp16)[name = string("op_1028_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> input_19_cast_fp16 = add(x = var_1009_cast_fp16, y = var_1028_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> c_3_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("c_3_cast_fp16")];
            tensor<int32, [3]> var_1036_begin_0 = const()[name = string("op_1036_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1036_end_0 = const()[name = string("op_1036_end_0"), val = tensor<int32, [3]>([3, 32, 256])];
            tensor<bool, [3]> var_1036_end_mask_0 = const()[name = string("op_1036_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 256]> var_1036_cast_fp16 = slice_by_index(begin = var_1036_begin_0, end = var_1036_end_0, end_mask = var_1036_end_mask_0, x = c_3_cast_fp16)[name = string("op_1036_cast_fp16")];
            tensor<int32, [4]> var_1041 = const()[name = string("op_1041"), val = tensor<int32, [4]>([3, 32, 2, 128])];
            tensor<fp16, [3, 32, 2, 128]> q_13_cast_fp16 = reshape(shape = var_1041, x = var_1036_cast_fp16)[name = string("q_13_cast_fp16")];
            tensor<int32, [3]> var_1047_begin_0 = const()[name = string("op_1047_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_1047_end_0 = const()[name = string("op_1047_end_0"), val = tensor<int32, [3]>([3, 32, 512])];
            tensor<bool, [3]> var_1047_end_mask_0 = const()[name = string("op_1047_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 256]> var_1047_cast_fp16 = slice_by_index(begin = var_1047_begin_0, end = var_1047_end_0, end_mask = var_1047_end_mask_0, x = c_3_cast_fp16)[name = string("op_1047_cast_fp16")];
            tensor<int32, [4]> var_1052 = const()[name = string("op_1052"), val = tensor<int32, [4]>([3, 32, 2, 128])];
            tensor<fp16, [3, 32, 2, 128]> k_13_cast_fp16 = reshape(shape = var_1052, x = var_1047_cast_fp16)[name = string("k_13_cast_fp16")];
            tensor<int32, [3]> var_1058_begin_0 = const()[name = string("op_1058_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_1058_end_0 = const()[name = string("op_1058_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> var_1058_end_mask_0 = const()[name = string("op_1058_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 512]> var_1058_cast_fp16 = slice_by_index(begin = var_1058_begin_0, end = var_1058_end_0, end_mask = var_1058_end_mask_0, x = c_3_cast_fp16)[name = string("op_1058_cast_fp16")];
            tensor<int32, [4]> var_1063 = const()[name = string("op_1063"), val = tensor<int32, [4]>([3, 32, 4, 128])];
            tensor<fp16, [3, 32, 4, 128]> var_1064_cast_fp16 = reshape(shape = var_1063, x = var_1058_cast_fp16)[name = string("op_1064_cast_fp16")];
            tensor<int32, [4]> var_1069 = const()[name = string("op_1069"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1071_to_fp16 = const()[name = string("op_1071_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 32, 2, 128]> q_15_cast_fp16 = mul(x = q_13_cast_fp16, y = var_1071_to_fp16)[name = string("q_15_cast_fp16")];
            fp16 var_1073_to_fp16 = const()[name = string("op_1073_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 32, 2, 128]> k_15_cast_fp16 = mul(x = k_13_cast_fp16, y = var_1073_to_fp16)[name = string("k_15_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_1075_cast_fp16 = mul(x = q_15_cast_fp16, y = q_15_cast_fp16)[name = string("op_1075_cast_fp16")];
            tensor<int32, [1]> var_1080_axes_0 = const()[name = string("op_1080_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1080_keep_dims_0 = const()[name = string("op_1080_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 2, 1]> var_1080_cast_fp16 = reduce_mean(axes = var_1080_axes_0, keep_dims = var_1080_keep_dims_0, x = var_1075_cast_fp16)[name = string("op_1080_cast_fp16")];
            fp16 var_1081_promoted_to_fp16 = const()[name = string("op_1081_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 32, 2, 1]> var_1082_cast_fp16 = mul(x = var_1080_cast_fp16, y = var_1081_promoted_to_fp16)[name = string("op_1082_cast_fp16")];
            fp16 var_1084_to_fp16 = const()[name = string("op_1084_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 32, 2, 1]> var_1085_cast_fp16 = add(x = var_1082_cast_fp16, y = var_1084_to_fp16)[name = string("op_1085_cast_fp16")];
            fp32 var_1086_epsilon_0 = const()[name = string("op_1086_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 2, 1]> var_1086_cast_fp16 = rsqrt(epsilon = var_1086_epsilon_0, x = var_1085_cast_fp16)[name = string("op_1086_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_1087_cast_fp16 = mul(x = q_15_cast_fp16, y = var_1086_cast_fp16)[name = string("op_1087_cast_fp16")];
            fp16 var_1088_to_fp16 = const()[name = string("op_1088_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 32, 2, 128]> q_17_cast_fp16 = mul(x = var_1087_cast_fp16, y = var_1088_to_fp16)[name = string("q_17_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_1090_cast_fp16 = mul(x = k_15_cast_fp16, y = k_15_cast_fp16)[name = string("op_1090_cast_fp16")];
            tensor<int32, [1]> var_1095_axes_0 = const()[name = string("op_1095_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1095_keep_dims_0 = const()[name = string("op_1095_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 2, 1]> var_1095_cast_fp16 = reduce_mean(axes = var_1095_axes_0, keep_dims = var_1095_keep_dims_0, x = var_1090_cast_fp16)[name = string("op_1095_cast_fp16")];
            fp16 var_1096_promoted_to_fp16 = const()[name = string("op_1096_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 32, 2, 1]> var_1097_cast_fp16 = mul(x = var_1095_cast_fp16, y = var_1096_promoted_to_fp16)[name = string("op_1097_cast_fp16")];
            fp16 var_1099_to_fp16 = const()[name = string("op_1099_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 32, 2, 1]> var_1100_cast_fp16 = add(x = var_1097_cast_fp16, y = var_1099_to_fp16)[name = string("op_1100_cast_fp16")];
            fp32 var_1101_epsilon_0 = const()[name = string("op_1101_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 2, 1]> var_1101_cast_fp16 = rsqrt(epsilon = var_1101_epsilon_0, x = var_1100_cast_fp16)[name = string("op_1101_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> k_17_cast_fp16 = mul(x = k_15_cast_fp16, y = var_1101_cast_fp16)[name = string("k_17_cast_fp16")];
            tensor<int32, [4]> var_1107 = const()[name = string("op_1107"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1120_axes_0 = const()[name = string("op_1120_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 32, 128]> var_1108_cast_fp16 = transpose(perm = var_1107, x = q_17_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [3, 2, 1, 32, 128]> var_1120_cast_fp16 = expand_dims(axes = var_1120_axes_0, x = var_1108_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [5]> var_1128_reps_0 = const()[name = string("op_1128_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 32, 128]> var_1128_cast_fp16 = tile(reps = var_1128_reps_0, x = var_1120_cast_fp16)[name = string("op_1128_cast_fp16")];
            tensor<int32, [4]> var_1133 = const()[name = string("op_1133"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<fp16, [3, 4, 32, 128]> q_19_cast_fp16 = reshape(shape = var_1133, x = var_1128_cast_fp16)[name = string("q_19_cast_fp16")];
            tensor<int32, [4]> var_1139 = const()[name = string("op_1139"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1152_axes_0 = const()[name = string("op_1152_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 32, 128]> var_1140_cast_fp16 = transpose(perm = var_1139, x = k_17_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [3, 2, 1, 32, 128]> var_1152_cast_fp16 = expand_dims(axes = var_1152_axes_0, x = var_1140_cast_fp16)[name = string("op_1152_cast_fp16")];
            tensor<int32, [5]> var_1160_reps_0 = const()[name = string("op_1160_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 32, 128]> var_1160_cast_fp16 = tile(reps = var_1160_reps_0, x = var_1152_cast_fp16)[name = string("op_1160_cast_fp16")];
            tensor<int32, [4]> var_1165 = const()[name = string("op_1165"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<fp16, [3, 4, 32, 128]> k_19_cast_fp16 = reshape(shape = var_1165, x = var_1160_cast_fp16)[name = string("k_19_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_1167_cast_fp16 = sigmoid(x = b_3_cast_fp16)[name = string("op_1167_cast_fp16")];
            tensor<int32, [3]> var_1171 = const()[name = string("op_1171"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_1_dt_bias_to_fp16 = const()[name = string("layers_1_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 32, 4]> input_21_cast_fp16 = add(x = a_3_cast_fp16, y = layers_1_dt_bias_to_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_1175_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("op_1175_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_1176_cast_fp16 = abs(x = input_21_cast_fp16)[name = string("op_1176_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 32, 4]> var_1177_cast_fp16 = mul(x = var_1176_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_1177_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_1178_cast_fp16 = exp(x = var_1177_cast_fp16)[name = string("op_1178_cast_fp16")];
            fp16 var_1180_to_fp16 = const()[name = string("op_1180_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 32, 4]> var_1181_cast_fp16 = add(x = var_1178_cast_fp16, y = var_1180_to_fp16)[name = string("op_1181_cast_fp16")];
            fp32 var_1182_epsilon_0 = const()[name = string("op_1182_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 32, 4]> var_1182_cast_fp16 = log(epsilon = var_1182_epsilon_0, x = var_1181_cast_fp16)[name = string("op_1182_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_1184_cast_fp16 = add(x = var_1175_cast_fp16, y = var_1182_cast_fp16)[name = string("op_1184_cast_fp16")];
            tensor<fp16, [4]> layers_1_negA_to_fp16 = const()[name = string("layers_1_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.ff4p+2, -0x1.8cp+3, -0x1.c0cp+3, -0x1.104p+3])];
            tensor<fp16, [3, 32, 4]> var_1185_cast_fp16 = mul(x = layers_1_negA_to_fp16, y = var_1184_cast_fp16)[name = string("op_1185_cast_fp16")];
            tensor<int32, [3]> var_1189 = const()[name = string("op_1189"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_1192_axes_0 = const()[name = string("op_1192_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 32]> beta_3_cast_fp16 = transpose(perm = var_1171, x = var_1167_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [3, 4, 32, 1]> var_1192_cast_fp16 = expand_dims(axes = var_1192_axes_0, x = beta_3_cast_fp16)[name = string("op_1192_cast_fp16")];
            fp16 var_1193_to_fp16 = const()[name = string("op_1193_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 32, 1]> var_1194_cast_fp16 = mul(x = var_1192_cast_fp16, y = var_1193_to_fp16)[name = string("op_1194_cast_fp16")];
            tensor<fp16, [3, 4, 32, 128]> v_3_cast_fp16 = transpose(perm = var_1069, x = var_1064_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [3, 4, 32, 128]> vb_7_cast_fp16 = mul(x = v_3_cast_fp16, y = var_1194_cast_fp16)[name = string("vb_7_cast_fp16")];
            tensor<int32, [4]> q_21_begin_0 = const()[name = string("q_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_21_end_0 = const()[name = string("q_21_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_21_end_mask_0 = const()[name = string("q_21_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_21_cast_fp16 = slice_by_index(begin = q_21_begin_0, end = q_21_end_0, end_mask = q_21_end_mask_0, x = q_19_cast_fp16)[name = string("q_21_cast_fp16")];
            tensor<int32, [4]> k_21_begin_0 = const()[name = string("k_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_21_end_0 = const()[name = string("k_21_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_21_end_mask_0 = const()[name = string("k_21_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, x = k_19_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<fp16, [3, 4, 32, 128]> var_1228_cast_fp16 = mul(x = k_19_cast_fp16, y = var_1192_cast_fp16)[name = string("op_1228_cast_fp16")];
            tensor<int32, [4]> kb_5_begin_0 = const()[name = string("kb_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_5_end_0 = const()[name = string("kb_5_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_5_end_mask_0 = const()[name = string("kb_5_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_5_cast_fp16 = slice_by_index(begin = kb_5_begin_0, end = kb_5_end_0, end_mask = kb_5_end_mask_0, x = var_1228_cast_fp16)[name = string("kb_5_cast_fp16")];
            tensor<int32, [4]> vb_9_begin_0 = const()[name = string("vb_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_9_end_0 = const()[name = string("vb_9_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_9_end_mask_0 = const()[name = string("vb_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_9_cast_fp16 = slice_by_index(begin = vb_9_begin_0, end = vb_9_end_0, end_mask = vb_9_end_mask_0, x = vb_7_cast_fp16)[name = string("vb_9_cast_fp16")];
            tensor<int32, [3]> gdec_9_begin_0 = const()[name = string("gdec_9_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_9_end_0 = const()[name = string("gdec_9_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_9_end_mask_0 = const()[name = string("gdec_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 32]> gdec_7_cast_fp16 = transpose(perm = var_1189, x = var_1185_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [3, 4, 16]> gdec_9_cast_fp16 = slice_by_index(begin = gdec_9_begin_0, end = gdec_9_end_0, end_mask = gdec_9_end_mask_0, x = gdec_7_cast_fp16)[name = string("gdec_9_cast_fp16")];
            tensor<int32, [2]> var_1276 = const()[name = string("op_1276"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_1277_cast_fp16 = reshape(shape = var_1276, x = gdec_9_cast_fp16)[name = string("op_1277_cast_fp16")];
            tensor<fp16, [12, 256]> var_1278_cast_fp16 = linear(bias = var_631_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1277_cast_fp16)[name = string("op_1278_cast_fp16")];
            tensor<int32, [4]> var_1283 = const()[name = string("op_1283"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_5_cast_fp16 = reshape(shape = var_1283, x = var_1278_cast_fp16)[name = string("P_5_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1285_cast_fp16 = exp(x = P_5_cast_fp16)[name = string("op_1285_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_5_cast_fp16 = mul(x = var_1285_cast_fp16, y = tril_incl_to_fp16)[name = string("D_5_cast_fp16")];
            bool kk_5_transpose_x_1 = const()[name = string("kk_5_transpose_x_1"), val = bool(false)];
            bool kk_5_transpose_y_1 = const()[name = string("kk_5_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_5_cast_fp16 = matmul(transpose_x = kk_5_transpose_x_1, transpose_y = kk_5_transpose_y_1, x = kb_5_cast_fp16, y = k_21_cast_fp16)[name = string("kk_5_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_1291_cast_fp16 = mul(x = kk_5_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_1291_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1292_cast_fp16 = mul(x = var_1291_cast_fp16, y = D_5_cast_fp16)[name = string("op_1292_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_5_cast_fp16 = mul(x = var_1292_cast_fp16, y = cum_tail_to_fp16)[name = string("N_5_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1294_cast_fp16 = mul(x = D_5_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1294_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_5_cast_fp16 = mul(x = kk_5_cast_fp16, y = var_1294_cast_fp16)[name = string("Na_5_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_17_cast_fp16 = add(x = eye_to_fp16, y = N_5_cast_fp16)[name = string("Minv_17_cast_fp16")];
            bool Np_13_transpose_x_0 = const()[name = string("Np_13_transpose_x_0"), val = bool(false)];
            bool Np_13_transpose_y_0 = const()[name = string("Np_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_13_cast_fp16 = matmul(transpose_x = Np_13_transpose_x_0, transpose_y = Np_13_transpose_y_0, x = N_5_cast_fp16, y = Na_5_cast_fp16)[name = string("Np_13_cast_fp16")];
            bool Npa_9_transpose_x_0 = const()[name = string("Npa_9_transpose_x_0"), val = bool(false)];
            bool Npa_9_transpose_y_0 = const()[name = string("Npa_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_9_cast_fp16 = matmul(transpose_x = Npa_9_transpose_x_0, transpose_y = Npa_9_transpose_y_0, x = Na_5_cast_fp16, y = N_5_cast_fp16)[name = string("Npa_9_cast_fp16")];
            bool var_1300_transpose_x_0 = const()[name = string("op_1300_transpose_x_0"), val = bool(false)];
            bool var_1300_transpose_y_0 = const()[name = string("op_1300_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1300_cast_fp16 = matmul(transpose_x = var_1300_transpose_x_0, transpose_y = var_1300_transpose_y_0, x = Minv_17_cast_fp16, y = Np_13_cast_fp16)[name = string("op_1300_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_19_cast_fp16 = add(x = Minv_17_cast_fp16, y = var_1300_cast_fp16)[name = string("Minv_19_cast_fp16")];
            bool Np_15_transpose_x_0 = const()[name = string("Np_15_transpose_x_0"), val = bool(false)];
            bool Np_15_transpose_y_0 = const()[name = string("Np_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_15_cast_fp16 = matmul(transpose_x = Np_15_transpose_x_0, transpose_y = Np_15_transpose_y_0, x = Np_13_cast_fp16, y = Npa_9_cast_fp16)[name = string("Np_15_cast_fp16")];
            bool Npa_11_transpose_x_0 = const()[name = string("Npa_11_transpose_x_0"), val = bool(false)];
            bool Npa_11_transpose_y_0 = const()[name = string("Npa_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_11_cast_fp16 = matmul(transpose_x = Npa_11_transpose_x_0, transpose_y = Npa_11_transpose_y_0, x = Npa_9_cast_fp16, y = Np_13_cast_fp16)[name = string("Npa_11_cast_fp16")];
            bool var_1305_transpose_x_0 = const()[name = string("op_1305_transpose_x_0"), val = bool(false)];
            bool var_1305_transpose_y_0 = const()[name = string("op_1305_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1305_cast_fp16 = matmul(transpose_x = var_1305_transpose_x_0, transpose_y = var_1305_transpose_y_0, x = Minv_19_cast_fp16, y = Np_15_cast_fp16)[name = string("op_1305_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_21_cast_fp16 = add(x = Minv_19_cast_fp16, y = var_1305_cast_fp16)[name = string("Minv_21_cast_fp16")];
            bool Np_17_transpose_x_0 = const()[name = string("Np_17_transpose_x_0"), val = bool(false)];
            bool Np_17_transpose_y_0 = const()[name = string("Np_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_17_cast_fp16 = matmul(transpose_x = Np_17_transpose_x_0, transpose_y = Np_17_transpose_y_0, x = Np_15_cast_fp16, y = Npa_11_cast_fp16)[name = string("Np_17_cast_fp16")];
            bool var_1309_transpose_x_0 = const()[name = string("op_1309_transpose_x_0"), val = bool(false)];
            bool var_1309_transpose_y_0 = const()[name = string("op_1309_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1309_cast_fp16 = matmul(transpose_x = var_1309_transpose_x_0, transpose_y = var_1309_transpose_y_0, x = Minv_21_cast_fp16, y = Np_17_cast_fp16)[name = string("op_1309_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_23_cast_fp16 = add(x = Minv_21_cast_fp16, y = var_1309_cast_fp16)[name = string("Minv_23_cast_fp16")];
            bool u_5_transpose_x_0 = const()[name = string("u_5_transpose_x_0"), val = bool(false)];
            bool u_5_transpose_y_0 = const()[name = string("u_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_5_cast_fp16 = matmul(transpose_x = u_5_transpose_x_0, transpose_y = u_5_transpose_y_0, x = Minv_23_cast_fp16, y = vb_9_cast_fp16)[name = string("u_5_cast_fp16")];
            bool var_1316_transpose_x_1 = const()[name = string("op_1316_transpose_x_1"), val = bool(false)];
            bool var_1316_transpose_y_1 = const()[name = string("op_1316_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_1316_cast_fp16 = matmul(transpose_x = var_1316_transpose_x_1, transpose_y = var_1316_transpose_y_1, x = q_21_cast_fp16, y = k_21_cast_fp16)[name = string("op_1316_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_5_cast_fp16 = mul(x = var_1316_cast_fp16, y = D_5_cast_fp16)[name = string("A_5_cast_fp16")];
            bool var_1318_transpose_x_0 = const()[name = string("op_1318_transpose_x_0"), val = bool(false)];
            bool var_1318_transpose_y_0 = const()[name = string("op_1318_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1318_cast_fp16 = matmul(transpose_x = var_1318_transpose_x_0, transpose_y = var_1318_transpose_y_0, x = A_5_cast_fp16, y = u_5_cast_fp16)[name = string("op_1318_cast_fp16")];
            tensor<fp16, [16]> var_1319_bias_0_to_fp16 = const()[name = string("op_1319_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_1319_cast_fp16 = linear(bias = var_1319_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_9_cast_fp16)[name = string("op_1319_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_3_cast_fp16 = exp(x = var_1319_cast_fp16)[name = string("tail_3_cast_fp16")];
            tensor<int32, [1]> var_1322_axes_0 = const()[name = string("op_1322_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_1322_cast_fp16 = expand_dims(axes = var_1322_axes_0, x = tail_3_cast_fp16)[name = string("op_1322_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1323_cast_fp16 = mul(x = k_21_cast_fp16, y = var_1322_cast_fp16)[name = string("op_1323_cast_fp16")];
            bool S_3_transpose_x_1 = const()[name = string("S_3_transpose_x_1"), val = bool(true)];
            bool S_3_transpose_y_1 = const()[name = string("S_3_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_3_cast_fp16 = matmul(transpose_x = S_3_transpose_x_1, transpose_y = S_3_transpose_y_1, x = var_1323_cast_fp16, y = u_5_cast_fp16)[name = string("S_3_cast_fp16")];
            tensor<int32, [4]> q_23_begin_0 = const()[name = string("q_23_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_23_end_0 = const()[name = string("q_23_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_23_end_mask_0 = const()[name = string("q_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_23_cast_fp16 = slice_by_index(begin = q_23_begin_0, end = q_23_end_0, end_mask = q_23_end_mask_0, x = q_19_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<int32, [4]> k_23_begin_0 = const()[name = string("k_23_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_23_end_0 = const()[name = string("k_23_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_23_end_mask_0 = const()[name = string("k_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_23_cast_fp16 = slice_by_index(begin = k_23_begin_0, end = k_23_end_0, end_mask = k_23_end_mask_0, x = k_19_cast_fp16)[name = string("k_23_cast_fp16")];
            tensor<int32, [4]> kb_7_begin_0 = const()[name = string("kb_7_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_7_end_0 = const()[name = string("kb_7_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_7_end_mask_0 = const()[name = string("kb_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_7_cast_fp16 = slice_by_index(begin = kb_7_begin_0, end = kb_7_end_0, end_mask = kb_7_end_mask_0, x = var_1228_cast_fp16)[name = string("kb_7_cast_fp16")];
            tensor<int32, [4]> vb_11_begin_0 = const()[name = string("vb_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_11_end_0 = const()[name = string("vb_11_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_11_end_mask_0 = const()[name = string("vb_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_11_cast_fp16 = slice_by_index(begin = vb_11_begin_0, end = vb_11_end_0, end_mask = vb_11_end_mask_0, x = vb_7_cast_fp16)[name = string("vb_11_cast_fp16")];
            tensor<int32, [3]> gdec_11_begin_0 = const()[name = string("gdec_11_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_11_end_0 = const()[name = string("gdec_11_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_11_end_mask_0 = const()[name = string("gdec_11_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_11_cast_fp16 = slice_by_index(begin = gdec_11_begin_0, end = gdec_11_end_0, end_mask = gdec_11_end_mask_0, x = gdec_7_cast_fp16)[name = string("gdec_11_cast_fp16")];
            tensor<int32, [2]> var_1408 = const()[name = string("op_1408"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_1409_cast_fp16 = reshape(shape = var_1408, x = gdec_11_cast_fp16)[name = string("op_1409_cast_fp16")];
            tensor<fp16, [12, 256]> var_1410_cast_fp16 = linear(bias = var_631_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1409_cast_fp16)[name = string("op_1410_cast_fp16")];
            tensor<int32, [4]> var_1415 = const()[name = string("op_1415"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_7_cast_fp16 = reshape(shape = var_1415, x = var_1410_cast_fp16)[name = string("P_7_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1417_cast_fp16 = exp(x = P_7_cast_fp16)[name = string("op_1417_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_7_cast_fp16 = mul(x = var_1417_cast_fp16, y = tril_incl_to_fp16)[name = string("D_7_cast_fp16")];
            bool kk_7_transpose_x_1 = const()[name = string("kk_7_transpose_x_1"), val = bool(false)];
            bool kk_7_transpose_y_1 = const()[name = string("kk_7_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_7_cast_fp16 = matmul(transpose_x = kk_7_transpose_x_1, transpose_y = kk_7_transpose_y_1, x = kb_7_cast_fp16, y = k_23_cast_fp16)[name = string("kk_7_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_1423_cast_fp16 = mul(x = kk_7_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_1423_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1424_cast_fp16 = mul(x = var_1423_cast_fp16, y = D_7_cast_fp16)[name = string("op_1424_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_7_cast_fp16 = mul(x = var_1424_cast_fp16, y = cum_tail_to_fp16)[name = string("N_7_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1426_cast_fp16 = mul(x = D_7_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1426_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_7_cast_fp16 = mul(x = kk_7_cast_fp16, y = var_1426_cast_fp16)[name = string("Na_7_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_25_cast_fp16 = add(x = eye_to_fp16, y = N_7_cast_fp16)[name = string("Minv_25_cast_fp16")];
            bool Np_19_transpose_x_0 = const()[name = string("Np_19_transpose_x_0"), val = bool(false)];
            bool Np_19_transpose_y_0 = const()[name = string("Np_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_19_cast_fp16 = matmul(transpose_x = Np_19_transpose_x_0, transpose_y = Np_19_transpose_y_0, x = N_7_cast_fp16, y = Na_7_cast_fp16)[name = string("Np_19_cast_fp16")];
            bool Npa_13_transpose_x_0 = const()[name = string("Npa_13_transpose_x_0"), val = bool(false)];
            bool Npa_13_transpose_y_0 = const()[name = string("Npa_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_13_cast_fp16 = matmul(transpose_x = Npa_13_transpose_x_0, transpose_y = Npa_13_transpose_y_0, x = Na_7_cast_fp16, y = N_7_cast_fp16)[name = string("Npa_13_cast_fp16")];
            bool var_1432_transpose_x_0 = const()[name = string("op_1432_transpose_x_0"), val = bool(false)];
            bool var_1432_transpose_y_0 = const()[name = string("op_1432_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1432_cast_fp16 = matmul(transpose_x = var_1432_transpose_x_0, transpose_y = var_1432_transpose_y_0, x = Minv_25_cast_fp16, y = Np_19_cast_fp16)[name = string("op_1432_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_27_cast_fp16 = add(x = Minv_25_cast_fp16, y = var_1432_cast_fp16)[name = string("Minv_27_cast_fp16")];
            bool Np_21_transpose_x_0 = const()[name = string("Np_21_transpose_x_0"), val = bool(false)];
            bool Np_21_transpose_y_0 = const()[name = string("Np_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_21_cast_fp16 = matmul(transpose_x = Np_21_transpose_x_0, transpose_y = Np_21_transpose_y_0, x = Np_19_cast_fp16, y = Npa_13_cast_fp16)[name = string("Np_21_cast_fp16")];
            bool Npa_15_transpose_x_0 = const()[name = string("Npa_15_transpose_x_0"), val = bool(false)];
            bool Npa_15_transpose_y_0 = const()[name = string("Npa_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_15_cast_fp16 = matmul(transpose_x = Npa_15_transpose_x_0, transpose_y = Npa_15_transpose_y_0, x = Npa_13_cast_fp16, y = Np_19_cast_fp16)[name = string("Npa_15_cast_fp16")];
            bool var_1437_transpose_x_0 = const()[name = string("op_1437_transpose_x_0"), val = bool(false)];
            bool var_1437_transpose_y_0 = const()[name = string("op_1437_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1437_cast_fp16 = matmul(transpose_x = var_1437_transpose_x_0, transpose_y = var_1437_transpose_y_0, x = Minv_27_cast_fp16, y = Np_21_cast_fp16)[name = string("op_1437_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_29_cast_fp16 = add(x = Minv_27_cast_fp16, y = var_1437_cast_fp16)[name = string("Minv_29_cast_fp16")];
            bool Np_23_transpose_x_0 = const()[name = string("Np_23_transpose_x_0"), val = bool(false)];
            bool Np_23_transpose_y_0 = const()[name = string("Np_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_23_cast_fp16 = matmul(transpose_x = Np_23_transpose_x_0, transpose_y = Np_23_transpose_y_0, x = Np_21_cast_fp16, y = Npa_15_cast_fp16)[name = string("Np_23_cast_fp16")];
            bool var_1441_transpose_x_0 = const()[name = string("op_1441_transpose_x_0"), val = bool(false)];
            bool var_1441_transpose_y_0 = const()[name = string("op_1441_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1441_cast_fp16 = matmul(transpose_x = var_1441_transpose_x_0, transpose_y = var_1441_transpose_y_0, x = Minv_29_cast_fp16, y = Np_23_cast_fp16)[name = string("op_1441_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_31_cast_fp16 = add(x = Minv_29_cast_fp16, y = var_1441_cast_fp16)[name = string("Minv_31_cast_fp16")];
            bool u_7_transpose_x_0 = const()[name = string("u_7_transpose_x_0"), val = bool(false)];
            bool u_7_transpose_y_0 = const()[name = string("u_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_7_cast_fp16 = matmul(transpose_x = u_7_transpose_x_0, transpose_y = u_7_transpose_y_0, x = Minv_31_cast_fp16, y = vb_11_cast_fp16)[name = string("u_7_cast_fp16")];
            bool var_1448_transpose_x_1 = const()[name = string("op_1448_transpose_x_1"), val = bool(false)];
            bool var_1448_transpose_y_1 = const()[name = string("op_1448_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_1448_cast_fp16 = matmul(transpose_x = var_1448_transpose_x_1, transpose_y = var_1448_transpose_y_1, x = q_23_cast_fp16, y = k_23_cast_fp16)[name = string("op_1448_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_7_cast_fp16 = mul(x = var_1448_cast_fp16, y = D_7_cast_fp16)[name = string("A_7_cast_fp16")];
            tensor<fp16, [16]> var_1450_bias_0_to_fp16 = const()[name = string("op_1450_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_1450_cast_fp16 = linear(bias = var_1450_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_11_cast_fp16)[name = string("op_1450_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_3_cast_fp16 = exp(x = var_1450_cast_fp16)[name = string("cum_3_cast_fp16")];
            tensor<int32, [1]> var_1453_axes_0 = const()[name = string("op_1453_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_1453_cast_fp16 = expand_dims(axes = var_1453_axes_0, x = cum_3_cast_fp16)[name = string("op_1453_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1454_cast_fp16 = mul(x = kb_7_cast_fp16, y = var_1453_cast_fp16)[name = string("op_1454_cast_fp16")];
            bool w_7_transpose_x_0 = const()[name = string("w_7_transpose_x_0"), val = bool(false)];
            bool w_7_transpose_y_0 = const()[name = string("w_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_7_cast_fp16 = matmul(transpose_x = w_7_transpose_x_0, transpose_y = w_7_transpose_y_0, x = Minv_31_cast_fp16, y = var_1454_cast_fp16)[name = string("w_7_cast_fp16")];
            bool var_1456_transpose_x_0 = const()[name = string("op_1456_transpose_x_0"), val = bool(false)];
            bool var_1456_transpose_y_0 = const()[name = string("op_1456_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1456_cast_fp16 = matmul(transpose_x = var_1456_transpose_x_0, transpose_y = var_1456_transpose_y_0, x = w_7_cast_fp16, y = S_3_cast_fp16)[name = string("op_1456_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_3_cast_fp16 = sub(x = u_7_cast_fp16, y = var_1456_cast_fp16)[name = string("vnew_3_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1461_cast_fp16 = mul(x = q_23_cast_fp16, y = var_1453_cast_fp16)[name = string("op_1461_cast_fp16")];
            bool var_1462_transpose_x_0 = const()[name = string("op_1462_transpose_x_0"), val = bool(false)];
            bool var_1462_transpose_y_0 = const()[name = string("op_1462_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1462_cast_fp16 = matmul(transpose_x = var_1462_transpose_x_0, transpose_y = var_1462_transpose_y_0, x = var_1461_cast_fp16, y = S_3_cast_fp16)[name = string("op_1462_cast_fp16")];
            bool var_1463_transpose_x_0 = const()[name = string("op_1463_transpose_x_0"), val = bool(false)];
            bool var_1463_transpose_y_0 = const()[name = string("op_1463_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1463_cast_fp16 = matmul(transpose_x = var_1463_transpose_x_0, transpose_y = var_1463_transpose_y_0, x = A_7_cast_fp16, y = vnew_3_cast_fp16)[name = string("op_1463_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_3_cast_fp16 = add(x = var_1462_cast_fp16, y = var_1463_cast_fp16)[name = string("oc_3_cast_fp16")];
            int32 var_1467 = const()[name = string("op_1467"), val = int32(2)];
            bool o_7_interleave_0 = const()[name = string("o_7_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 32, 128]> o_7_cast_fp16 = concat(axis = var_1467, interleave = o_7_interleave_0, values = (var_1318_cast_fp16, oc_3_cast_fp16))[name = string("o_7_cast_fp16")];
            tensor<int32, [4]> var_1473 = const()[name = string("op_1473"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1479 = const()[name = string("op_1479"), val = tensor<int32, [4]>([3, 32, 4, 128])];
            tensor<fp16, [3, 32, 4, 128]> input_23_cast_fp16 = reshape(shape = var_1479, x = z_3_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> o_9_cast_fp16 = transpose(perm = var_1473, x = o_7_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [3, 32, 4, 128]> var_1481_cast_fp16 = mul(x = o_9_cast_fp16, y = o_9_cast_fp16)[name = string("op_1481_cast_fp16")];
            tensor<int32, [1]> var_1486_axes_0 = const()[name = string("op_1486_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1486_keep_dims_0 = const()[name = string("op_1486_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 4, 1]> var_1486_cast_fp16 = reduce_mean(axes = var_1486_axes_0, keep_dims = var_1486_keep_dims_0, x = var_1481_cast_fp16)[name = string("op_1486_cast_fp16")];
            fp16 var_1488_to_fp16 = const()[name = string("op_1488_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 32, 4, 1]> var_1489_cast_fp16 = add(x = var_1486_cast_fp16, y = var_1488_to_fp16)[name = string("op_1489_cast_fp16")];
            fp32 var_1490_epsilon_0 = const()[name = string("op_1490_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 4, 1]> var_1490_cast_fp16 = rsqrt(epsilon = var_1490_epsilon_0, x = var_1489_cast_fp16)[name = string("op_1490_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> var_1491_cast_fp16 = mul(x = o_9_cast_fp16, y = var_1490_cast_fp16)[name = string("op_1491_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> var_1493_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_1493_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> o_11_cast_fp16 = mul(x = var_1491_cast_fp16, y = var_1493_cast_fp16)[name = string("o_11_cast_fp16")];
            tensor<int32, [3]> var_1498 = const()[name = string("op_1498"), val = tensor<int32, [3]>([3, 32, 512])];
            tensor<fp16, [3, 32, 512]> input_25_cast_fp16 = reshape(shape = var_1498, x = o_11_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [640, 512]> layers_1_out_weight_to_fp16 = const()[name = string("layers_1_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34984256)))];
            tensor<fp16, [3, 32, 640]> linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_out_weight_to_fp16, x = input_25_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [3, 32, 640]> x_13_cast_fp16 = add(x = x_9_cast_fp16, y = linear_5_cast_fp16)[name = string("x_13_cast_fp16")];
            fp16 var_1511_to_fp16 = const()[name = string("op_1511_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> x_15_cast_fp16 = mul(x = x_13_cast_fp16, y = var_1511_to_fp16)[name = string("x_15_cast_fp16")];
            fp16 var_1509_promoted_to_fp16 = const()[name = string("op_1509_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> var_1513_cast_fp16 = pow(x = x_15_cast_fp16, y = var_1509_promoted_to_fp16)[name = string("op_1513_cast_fp16")];
            tensor<int32, [1]> var_1515_axes_0 = const()[name = string("op_1515_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1515_keep_dims_0 = const()[name = string("op_1515_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 1]> var_1515_cast_fp16 = reduce_mean(axes = var_1515_axes_0, keep_dims = var_1515_keep_dims_0, x = var_1513_cast_fp16)[name = string("op_1515_cast_fp16")];
            fp16 var_1516_to_fp16 = const()[name = string("op_1516_to_fp16"), val = fp16(0x1.0cp-18)];
            tensor<fp16, [3, 32, 1]> var_1517_cast_fp16 = add(x = var_1515_cast_fp16, y = var_1516_to_fp16)[name = string("op_1517_cast_fp16")];
            fp32 var_1518_epsilon_0 = const()[name = string("op_1518_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 1]> var_1518_cast_fp16 = rsqrt(epsilon = var_1518_epsilon_0, x = var_1517_cast_fp16)[name = string("op_1518_cast_fp16")];
            tensor<fp16, [3, 32, 640]> var_1519_cast_fp16 = mul(x = x_15_cast_fp16, y = var_1518_cast_fp16)[name = string("op_1519_cast_fp16")];
            tensor<fp16, [640]> layers_1_n2_w1_to_fp16 = const()[name = string("layers_1_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35639680)))];
            tensor<fp16, [3, 32, 640]> input_27_cast_fp16 = mul(x = var_1519_cast_fp16, y = layers_1_n2_w1_to_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_1_gate_up_weight_to_fp16 = const()[name = string("layers_1_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35641024)))];
            tensor<fp16, [3, 32, 5120]> linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_1_gate_up_weight_to_fp16, x = input_27_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<int32, [3]> input_29_begin_0 = const()[name = string("input_29_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_29_end_0 = const()[name = string("input_29_end_0"), val = tensor<int32, [3]>([3, 32, 2560])];
            tensor<bool, [3]> input_29_end_mask_0 = const()[name = string("input_29_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 2560]> input_29_cast_fp16 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = linear_6_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [3, 32, 2560]> var_1529_cast_fp16 = silu(x = input_29_cast_fp16)[name = string("op_1529_cast_fp16")];
            tensor<int32, [3]> var_1534_begin_0 = const()[name = string("op_1534_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_1534_end_0 = const()[name = string("op_1534_end_0"), val = tensor<int32, [3]>([3, 32, 5120])];
            tensor<bool, [3]> var_1534_end_mask_0 = const()[name = string("op_1534_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 2560]> var_1534_cast_fp16 = slice_by_index(begin = var_1534_begin_0, end = var_1534_end_0, end_mask = var_1534_end_mask_0, x = linear_6_cast_fp16)[name = string("op_1534_cast_fp16")];
            tensor<fp16, [3, 32, 2560]> input_31_cast_fp16 = mul(x = var_1529_cast_fp16, y = var_1534_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_1_down_weight_to_fp16 = const()[name = string("layers_1_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42194688)))];
            tensor<fp16, [3, 32, 640]> linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_down_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [3, 32, 640]> x_17_cast_fp16 = add(x = x_13_cast_fp16, y = linear_7_cast_fp16)[name = string("x_17_cast_fp16")];
            fp16 var_1545_promoted_to_fp16 = const()[name = string("op_1545_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> var_1549_cast_fp16 = pow(x = x_17_cast_fp16, y = var_1545_promoted_to_fp16)[name = string("op_1549_cast_fp16")];
            tensor<int32, [1]> var_1551_axes_0 = const()[name = string("op_1551_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1551_keep_dims_0 = const()[name = string("op_1551_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 1]> var_1551_cast_fp16 = reduce_mean(axes = var_1551_axes_0, keep_dims = var_1551_keep_dims_0, x = var_1549_cast_fp16)[name = string("op_1551_cast_fp16")];
            fp16 var_1552_to_fp16 = const()[name = string("op_1552_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 32, 1]> var_1553_cast_fp16 = add(x = var_1551_cast_fp16, y = var_1552_to_fp16)[name = string("op_1553_cast_fp16")];
            fp32 var_1554_epsilon_0 = const()[name = string("op_1554_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 1]> var_1554_cast_fp16 = rsqrt(epsilon = var_1554_epsilon_0, x = var_1553_cast_fp16)[name = string("op_1554_cast_fp16")];
            tensor<fp16, [3, 32, 640]> var_1555_cast_fp16 = mul(x = x_17_cast_fp16, y = var_1554_cast_fp16)[name = string("op_1555_cast_fp16")];
            tensor<fp16, [640]> layers_2_n1_w1_to_fp16 = const()[name = string("layers_2_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45471552)))];
            tensor<fp16, [3, 32, 640]> input_33_cast_fp16 = mul(x = var_1555_cast_fp16, y = layers_2_n1_w1_to_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_2_qkvzba_weight_to_fp16 = const()[name = string("layers_2_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45472896)))];
            tensor<fp16, [3, 32, 1544]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_qkvzba_weight_to_fp16, x = input_33_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [3]> qkv_5_begin_0 = const()[name = string("qkv_5_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_5_end_0 = const()[name = string("qkv_5_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> qkv_5_end_mask_0 = const()[name = string("qkv_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 1024]> qkv_5_cast_fp16 = slice_by_index(begin = qkv_5_begin_0, end = qkv_5_end_0, end_mask = qkv_5_end_mask_0, x = linear_8_cast_fp16)[name = string("qkv_5_cast_fp16")];
            tensor<int32, [3]> z_5_begin_0 = const()[name = string("z_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_5_end_0 = const()[name = string("z_5_end_0"), val = tensor<int32, [3]>([3, 32, 1536])];
            tensor<bool, [3]> z_5_end_mask_0 = const()[name = string("z_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 512]> z_5_cast_fp16 = slice_by_index(begin = z_5_begin_0, end = z_5_end_0, end_mask = z_5_end_mask_0, x = linear_8_cast_fp16)[name = string("z_5_cast_fp16")];
            tensor<int32, [3]> b_5_begin_0 = const()[name = string("b_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_5_end_0 = const()[name = string("b_5_end_0"), val = tensor<int32, [3]>([3, 32, 1540])];
            tensor<bool, [3]> b_5_end_mask_0 = const()[name = string("b_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 4]> b_5_cast_fp16 = slice_by_index(begin = b_5_begin_0, end = b_5_end_0, end_mask = b_5_end_mask_0, x = linear_8_cast_fp16)[name = string("b_5_cast_fp16")];
            tensor<int32, [3]> a_5_begin_0 = const()[name = string("a_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_5_end_0 = const()[name = string("a_5_end_0"), val = tensor<int32, [3]>([3, 32, 1])];
            tensor<bool, [3]> a_5_end_mask_0 = const()[name = string("a_5_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 4]> a_5_cast_fp16 = slice_by_index(begin = a_5_begin_0, end = a_5_end_0, end_mask = a_5_end_mask_0, x = linear_8_cast_fp16)[name = string("a_5_cast_fp16")];
            int32 var_1594 = const()[name = string("op_1594"), val = int32(1)];
            bool xp_5_interleave_0 = const()[name = string("xp_5_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 35, 1024]> xp_5_cast_fp16 = concat(axis = var_1594, interleave = xp_5_interleave_0, values = (zpad_to_fp16, qkv_5_cast_fp16))[name = string("xp_5_cast_fp16")];
            tensor<int32, [3]> var_1605_begin_0 = const()[name = string("op_1605_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_1605_end_0 = const()[name = string("op_1605_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_1605_end_mask_0 = const()[name = string("op_1605_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 1024]> var_1605_cast_fp16 = slice_by_index(begin = var_1605_begin_0, end = var_1605_end_0, end_mask = var_1605_end_mask_0, x = xp_5_cast_fp16)[name = string("op_1605_cast_fp16")];
            tensor<fp16, [1024]> var_1613_to_fp16 = const()[name = string("op_1613_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47449280)))];
            tensor<fp16, [3, 32, 1024]> var_1614_cast_fp16 = mul(x = var_1605_cast_fp16, y = var_1613_to_fp16)[name = string("op_1614_cast_fp16")];
            tensor<int32, [3]> var_1624_begin_0 = const()[name = string("op_1624_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_1624_end_0 = const()[name = string("op_1624_end_0"), val = tensor<int32, [3]>([3, 34, 1024])];
            tensor<bool, [3]> var_1624_end_mask_0 = const()[name = string("op_1624_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_1624_cast_fp16 = slice_by_index(begin = var_1624_begin_0, end = var_1624_end_0, end_mask = var_1624_end_mask_0, x = xp_5_cast_fp16)[name = string("op_1624_cast_fp16")];
            tensor<fp16, [1024]> var_1632_to_fp16 = const()[name = string("op_1632_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47451392)))];
            tensor<fp16, [3, 32, 1024]> var_1633_cast_fp16 = mul(x = var_1624_cast_fp16, y = var_1632_to_fp16)[name = string("op_1633_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> var_1635_cast_fp16 = add(x = var_1614_cast_fp16, y = var_1633_cast_fp16)[name = string("op_1635_cast_fp16")];
            tensor<int32, [3]> var_1645_begin_0 = const()[name = string("op_1645_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1645_end_0 = const()[name = string("op_1645_end_0"), val = tensor<int32, [3]>([3, 33, 1024])];
            tensor<bool, [3]> var_1645_end_mask_0 = const()[name = string("op_1645_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_1645_cast_fp16 = slice_by_index(begin = var_1645_begin_0, end = var_1645_end_0, end_mask = var_1645_end_mask_0, x = xp_5_cast_fp16)[name = string("op_1645_cast_fp16")];
            tensor<fp16, [1024]> var_1653_to_fp16 = const()[name = string("op_1653_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47453504)))];
            tensor<fp16, [3, 32, 1024]> var_1654_cast_fp16 = mul(x = var_1645_cast_fp16, y = var_1653_to_fp16)[name = string("op_1654_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> var_1656_cast_fp16 = add(x = var_1635_cast_fp16, y = var_1654_cast_fp16)[name = string("op_1656_cast_fp16")];
            tensor<int32, [3]> var_1666_begin_0 = const()[name = string("op_1666_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1666_end_0 = const()[name = string("op_1666_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> var_1666_end_mask_0 = const()[name = string("op_1666_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_1666_cast_fp16 = slice_by_index(begin = var_1666_begin_0, end = var_1666_end_0, end_mask = var_1666_end_mask_0, x = xp_5_cast_fp16)[name = string("op_1666_cast_fp16")];
            tensor<fp16, [1024]> var_1674_to_fp16 = const()[name = string("op_1674_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47455616)))];
            tensor<fp16, [3, 32, 1024]> var_1675_cast_fp16 = mul(x = var_1666_cast_fp16, y = var_1674_to_fp16)[name = string("op_1675_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> input_35_cast_fp16 = add(x = var_1656_cast_fp16, y = var_1675_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> c_5_cast_fp16 = silu(x = input_35_cast_fp16)[name = string("c_5_cast_fp16")];
            tensor<int32, [3]> var_1683_begin_0 = const()[name = string("op_1683_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1683_end_0 = const()[name = string("op_1683_end_0"), val = tensor<int32, [3]>([3, 32, 256])];
            tensor<bool, [3]> var_1683_end_mask_0 = const()[name = string("op_1683_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 256]> var_1683_cast_fp16 = slice_by_index(begin = var_1683_begin_0, end = var_1683_end_0, end_mask = var_1683_end_mask_0, x = c_5_cast_fp16)[name = string("op_1683_cast_fp16")];
            tensor<int32, [4]> var_1688 = const()[name = string("op_1688"), val = tensor<int32, [4]>([3, 32, 2, 128])];
            tensor<fp16, [3, 32, 2, 128]> q_25_cast_fp16 = reshape(shape = var_1688, x = var_1683_cast_fp16)[name = string("q_25_cast_fp16")];
            tensor<int32, [3]> var_1694_begin_0 = const()[name = string("op_1694_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_1694_end_0 = const()[name = string("op_1694_end_0"), val = tensor<int32, [3]>([3, 32, 512])];
            tensor<bool, [3]> var_1694_end_mask_0 = const()[name = string("op_1694_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 256]> var_1694_cast_fp16 = slice_by_index(begin = var_1694_begin_0, end = var_1694_end_0, end_mask = var_1694_end_mask_0, x = c_5_cast_fp16)[name = string("op_1694_cast_fp16")];
            tensor<int32, [4]> var_1699 = const()[name = string("op_1699"), val = tensor<int32, [4]>([3, 32, 2, 128])];
            tensor<fp16, [3, 32, 2, 128]> k_25_cast_fp16 = reshape(shape = var_1699, x = var_1694_cast_fp16)[name = string("k_25_cast_fp16")];
            tensor<int32, [3]> var_1705_begin_0 = const()[name = string("op_1705_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_1705_end_0 = const()[name = string("op_1705_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> var_1705_end_mask_0 = const()[name = string("op_1705_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 512]> var_1705_cast_fp16 = slice_by_index(begin = var_1705_begin_0, end = var_1705_end_0, end_mask = var_1705_end_mask_0, x = c_5_cast_fp16)[name = string("op_1705_cast_fp16")];
            tensor<int32, [4]> var_1710 = const()[name = string("op_1710"), val = tensor<int32, [4]>([3, 32, 4, 128])];
            tensor<fp16, [3, 32, 4, 128]> var_1711_cast_fp16 = reshape(shape = var_1710, x = var_1705_cast_fp16)[name = string("op_1711_cast_fp16")];
            tensor<int32, [4]> var_1716 = const()[name = string("op_1716"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1718_to_fp16 = const()[name = string("op_1718_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 32, 2, 128]> q_27_cast_fp16 = mul(x = q_25_cast_fp16, y = var_1718_to_fp16)[name = string("q_27_cast_fp16")];
            fp16 var_1720_to_fp16 = const()[name = string("op_1720_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 32, 2, 128]> k_27_cast_fp16 = mul(x = k_25_cast_fp16, y = var_1720_to_fp16)[name = string("k_27_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_1722_cast_fp16 = mul(x = q_27_cast_fp16, y = q_27_cast_fp16)[name = string("op_1722_cast_fp16")];
            tensor<int32, [1]> var_1727_axes_0 = const()[name = string("op_1727_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1727_keep_dims_0 = const()[name = string("op_1727_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 2, 1]> var_1727_cast_fp16 = reduce_mean(axes = var_1727_axes_0, keep_dims = var_1727_keep_dims_0, x = var_1722_cast_fp16)[name = string("op_1727_cast_fp16")];
            fp16 var_1728_promoted_to_fp16 = const()[name = string("op_1728_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 32, 2, 1]> var_1729_cast_fp16 = mul(x = var_1727_cast_fp16, y = var_1728_promoted_to_fp16)[name = string("op_1729_cast_fp16")];
            fp16 var_1731_to_fp16 = const()[name = string("op_1731_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 32, 2, 1]> var_1732_cast_fp16 = add(x = var_1729_cast_fp16, y = var_1731_to_fp16)[name = string("op_1732_cast_fp16")];
            fp32 var_1733_epsilon_0 = const()[name = string("op_1733_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 2, 1]> var_1733_cast_fp16 = rsqrt(epsilon = var_1733_epsilon_0, x = var_1732_cast_fp16)[name = string("op_1733_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_1734_cast_fp16 = mul(x = q_27_cast_fp16, y = var_1733_cast_fp16)[name = string("op_1734_cast_fp16")];
            fp16 var_1735_to_fp16 = const()[name = string("op_1735_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 32, 2, 128]> q_29_cast_fp16 = mul(x = var_1734_cast_fp16, y = var_1735_to_fp16)[name = string("q_29_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_1737_cast_fp16 = mul(x = k_27_cast_fp16, y = k_27_cast_fp16)[name = string("op_1737_cast_fp16")];
            tensor<int32, [1]> var_1742_axes_0 = const()[name = string("op_1742_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1742_keep_dims_0 = const()[name = string("op_1742_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 2, 1]> var_1742_cast_fp16 = reduce_mean(axes = var_1742_axes_0, keep_dims = var_1742_keep_dims_0, x = var_1737_cast_fp16)[name = string("op_1742_cast_fp16")];
            fp16 var_1743_promoted_to_fp16 = const()[name = string("op_1743_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 32, 2, 1]> var_1744_cast_fp16 = mul(x = var_1742_cast_fp16, y = var_1743_promoted_to_fp16)[name = string("op_1744_cast_fp16")];
            fp16 var_1746_to_fp16 = const()[name = string("op_1746_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 32, 2, 1]> var_1747_cast_fp16 = add(x = var_1744_cast_fp16, y = var_1746_to_fp16)[name = string("op_1747_cast_fp16")];
            fp32 var_1748_epsilon_0 = const()[name = string("op_1748_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 2, 1]> var_1748_cast_fp16 = rsqrt(epsilon = var_1748_epsilon_0, x = var_1747_cast_fp16)[name = string("op_1748_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> k_29_cast_fp16 = mul(x = k_27_cast_fp16, y = var_1748_cast_fp16)[name = string("k_29_cast_fp16")];
            tensor<int32, [4]> var_1754 = const()[name = string("op_1754"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1767_axes_0 = const()[name = string("op_1767_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 32, 128]> var_1755_cast_fp16 = transpose(perm = var_1754, x = q_29_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [3, 2, 1, 32, 128]> var_1767_cast_fp16 = expand_dims(axes = var_1767_axes_0, x = var_1755_cast_fp16)[name = string("op_1767_cast_fp16")];
            tensor<int32, [5]> var_1775_reps_0 = const()[name = string("op_1775_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 32, 128]> var_1775_cast_fp16 = tile(reps = var_1775_reps_0, x = var_1767_cast_fp16)[name = string("op_1775_cast_fp16")];
            tensor<int32, [4]> var_1780 = const()[name = string("op_1780"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<fp16, [3, 4, 32, 128]> q_31_cast_fp16 = reshape(shape = var_1780, x = var_1775_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<int32, [4]> var_1786 = const()[name = string("op_1786"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1799_axes_0 = const()[name = string("op_1799_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 32, 128]> var_1787_cast_fp16 = transpose(perm = var_1786, x = k_29_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [3, 2, 1, 32, 128]> var_1799_cast_fp16 = expand_dims(axes = var_1799_axes_0, x = var_1787_cast_fp16)[name = string("op_1799_cast_fp16")];
            tensor<int32, [5]> var_1807_reps_0 = const()[name = string("op_1807_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 32, 128]> var_1807_cast_fp16 = tile(reps = var_1807_reps_0, x = var_1799_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<int32, [4]> var_1812 = const()[name = string("op_1812"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<fp16, [3, 4, 32, 128]> k_31_cast_fp16 = reshape(shape = var_1812, x = var_1807_cast_fp16)[name = string("k_31_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_1814_cast_fp16 = sigmoid(x = b_5_cast_fp16)[name = string("op_1814_cast_fp16")];
            tensor<int32, [3]> var_1818 = const()[name = string("op_1818"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_2_dt_bias_to_fp16 = const()[name = string("layers_2_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 32, 4]> input_37_cast_fp16 = add(x = a_5_cast_fp16, y = layers_2_dt_bias_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_1822_cast_fp16 = relu(x = input_37_cast_fp16)[name = string("op_1822_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_1823_cast_fp16 = abs(x = input_37_cast_fp16)[name = string("op_1823_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 32, 4]> var_1824_cast_fp16 = mul(x = var_1823_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_1824_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_1825_cast_fp16 = exp(x = var_1824_cast_fp16)[name = string("op_1825_cast_fp16")];
            fp16 var_1827_to_fp16 = const()[name = string("op_1827_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 32, 4]> var_1828_cast_fp16 = add(x = var_1825_cast_fp16, y = var_1827_to_fp16)[name = string("op_1828_cast_fp16")];
            fp32 var_1829_epsilon_0 = const()[name = string("op_1829_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 32, 4]> var_1829_cast_fp16 = log(epsilon = var_1829_epsilon_0, x = var_1828_cast_fp16)[name = string("op_1829_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_1831_cast_fp16 = add(x = var_1822_cast_fp16, y = var_1829_cast_fp16)[name = string("op_1831_cast_fp16")];
            tensor<fp16, [4]> layers_2_negA_to_fp16 = const()[name = string("layers_2_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.c0cp+3, -0x1.3e4p+3, -0x1.568p+1, -0x1.9acp+2])];
            tensor<fp16, [3, 32, 4]> var_1832_cast_fp16 = mul(x = layers_2_negA_to_fp16, y = var_1831_cast_fp16)[name = string("op_1832_cast_fp16")];
            tensor<int32, [3]> var_1836 = const()[name = string("op_1836"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_1839_axes_0 = const()[name = string("op_1839_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 32]> beta_5_cast_fp16 = transpose(perm = var_1818, x = var_1814_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [3, 4, 32, 1]> var_1839_cast_fp16 = expand_dims(axes = var_1839_axes_0, x = beta_5_cast_fp16)[name = string("op_1839_cast_fp16")];
            fp16 var_1840_to_fp16 = const()[name = string("op_1840_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 32, 1]> var_1841_cast_fp16 = mul(x = var_1839_cast_fp16, y = var_1840_to_fp16)[name = string("op_1841_cast_fp16")];
            tensor<fp16, [3, 4, 32, 128]> v_5_cast_fp16 = transpose(perm = var_1716, x = var_1711_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [3, 4, 32, 128]> vb_13_cast_fp16 = mul(x = v_5_cast_fp16, y = var_1841_cast_fp16)[name = string("vb_13_cast_fp16")];
            tensor<int32, [4]> q_33_begin_0 = const()[name = string("q_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_33_end_0 = const()[name = string("q_33_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_33_end_mask_0 = const()[name = string("q_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_33_cast_fp16 = slice_by_index(begin = q_33_begin_0, end = q_33_end_0, end_mask = q_33_end_mask_0, x = q_31_cast_fp16)[name = string("q_33_cast_fp16")];
            tensor<int32, [4]> k_33_begin_0 = const()[name = string("k_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_33_end_0 = const()[name = string("k_33_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_33_end_mask_0 = const()[name = string("k_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, x = k_31_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<fp16, [3, 4, 32, 128]> var_1875_cast_fp16 = mul(x = k_31_cast_fp16, y = var_1839_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [4]> kb_9_begin_0 = const()[name = string("kb_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_9_end_0 = const()[name = string("kb_9_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_9_end_mask_0 = const()[name = string("kb_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_9_cast_fp16 = slice_by_index(begin = kb_9_begin_0, end = kb_9_end_0, end_mask = kb_9_end_mask_0, x = var_1875_cast_fp16)[name = string("kb_9_cast_fp16")];
            tensor<int32, [4]> vb_15_begin_0 = const()[name = string("vb_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_15_end_0 = const()[name = string("vb_15_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_15_end_mask_0 = const()[name = string("vb_15_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_15_cast_fp16 = slice_by_index(begin = vb_15_begin_0, end = vb_15_end_0, end_mask = vb_15_end_mask_0, x = vb_13_cast_fp16)[name = string("vb_15_cast_fp16")];
            tensor<int32, [3]> gdec_15_begin_0 = const()[name = string("gdec_15_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_15_end_0 = const()[name = string("gdec_15_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_15_end_mask_0 = const()[name = string("gdec_15_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 32]> gdec_13_cast_fp16 = transpose(perm = var_1836, x = var_1832_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [3, 4, 16]> gdec_15_cast_fp16 = slice_by_index(begin = gdec_15_begin_0, end = gdec_15_end_0, end_mask = gdec_15_end_mask_0, x = gdec_13_cast_fp16)[name = string("gdec_15_cast_fp16")];
            tensor<int32, [2]> var_1923 = const()[name = string("op_1923"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_1924_cast_fp16 = reshape(shape = var_1923, x = gdec_15_cast_fp16)[name = string("op_1924_cast_fp16")];
            tensor<fp16, [12, 256]> var_1925_cast_fp16 = linear(bias = var_631_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1924_cast_fp16)[name = string("op_1925_cast_fp16")];
            tensor<int32, [4]> var_1930 = const()[name = string("op_1930"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_9_cast_fp16 = reshape(shape = var_1930, x = var_1925_cast_fp16)[name = string("P_9_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1932_cast_fp16 = exp(x = P_9_cast_fp16)[name = string("op_1932_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_9_cast_fp16 = mul(x = var_1932_cast_fp16, y = tril_incl_to_fp16)[name = string("D_9_cast_fp16")];
            bool kk_9_transpose_x_1 = const()[name = string("kk_9_transpose_x_1"), val = bool(false)];
            bool kk_9_transpose_y_1 = const()[name = string("kk_9_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_9_cast_fp16 = matmul(transpose_x = kk_9_transpose_x_1, transpose_y = kk_9_transpose_y_1, x = kb_9_cast_fp16, y = k_33_cast_fp16)[name = string("kk_9_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_1938_cast_fp16 = mul(x = kk_9_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_1938_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1939_cast_fp16 = mul(x = var_1938_cast_fp16, y = D_9_cast_fp16)[name = string("op_1939_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_9_cast_fp16 = mul(x = var_1939_cast_fp16, y = cum_tail_to_fp16)[name = string("N_9_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1941_cast_fp16 = mul(x = D_9_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1941_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_9_cast_fp16 = mul(x = kk_9_cast_fp16, y = var_1941_cast_fp16)[name = string("Na_9_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_33_cast_fp16 = add(x = eye_to_fp16, y = N_9_cast_fp16)[name = string("Minv_33_cast_fp16")];
            bool Np_25_transpose_x_0 = const()[name = string("Np_25_transpose_x_0"), val = bool(false)];
            bool Np_25_transpose_y_0 = const()[name = string("Np_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_25_cast_fp16 = matmul(transpose_x = Np_25_transpose_x_0, transpose_y = Np_25_transpose_y_0, x = N_9_cast_fp16, y = Na_9_cast_fp16)[name = string("Np_25_cast_fp16")];
            bool Npa_17_transpose_x_0 = const()[name = string("Npa_17_transpose_x_0"), val = bool(false)];
            bool Npa_17_transpose_y_0 = const()[name = string("Npa_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_17_cast_fp16 = matmul(transpose_x = Npa_17_transpose_x_0, transpose_y = Npa_17_transpose_y_0, x = Na_9_cast_fp16, y = N_9_cast_fp16)[name = string("Npa_17_cast_fp16")];
            bool var_1947_transpose_x_0 = const()[name = string("op_1947_transpose_x_0"), val = bool(false)];
            bool var_1947_transpose_y_0 = const()[name = string("op_1947_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1947_cast_fp16 = matmul(transpose_x = var_1947_transpose_x_0, transpose_y = var_1947_transpose_y_0, x = Minv_33_cast_fp16, y = Np_25_cast_fp16)[name = string("op_1947_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_35_cast_fp16 = add(x = Minv_33_cast_fp16, y = var_1947_cast_fp16)[name = string("Minv_35_cast_fp16")];
            bool Np_27_transpose_x_0 = const()[name = string("Np_27_transpose_x_0"), val = bool(false)];
            bool Np_27_transpose_y_0 = const()[name = string("Np_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_27_cast_fp16 = matmul(transpose_x = Np_27_transpose_x_0, transpose_y = Np_27_transpose_y_0, x = Np_25_cast_fp16, y = Npa_17_cast_fp16)[name = string("Np_27_cast_fp16")];
            bool Npa_19_transpose_x_0 = const()[name = string("Npa_19_transpose_x_0"), val = bool(false)];
            bool Npa_19_transpose_y_0 = const()[name = string("Npa_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_19_cast_fp16 = matmul(transpose_x = Npa_19_transpose_x_0, transpose_y = Npa_19_transpose_y_0, x = Npa_17_cast_fp16, y = Np_25_cast_fp16)[name = string("Npa_19_cast_fp16")];
            bool var_1952_transpose_x_0 = const()[name = string("op_1952_transpose_x_0"), val = bool(false)];
            bool var_1952_transpose_y_0 = const()[name = string("op_1952_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1952_cast_fp16 = matmul(transpose_x = var_1952_transpose_x_0, transpose_y = var_1952_transpose_y_0, x = Minv_35_cast_fp16, y = Np_27_cast_fp16)[name = string("op_1952_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_37_cast_fp16 = add(x = Minv_35_cast_fp16, y = var_1952_cast_fp16)[name = string("Minv_37_cast_fp16")];
            bool Np_29_transpose_x_0 = const()[name = string("Np_29_transpose_x_0"), val = bool(false)];
            bool Np_29_transpose_y_0 = const()[name = string("Np_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_29_cast_fp16 = matmul(transpose_x = Np_29_transpose_x_0, transpose_y = Np_29_transpose_y_0, x = Np_27_cast_fp16, y = Npa_19_cast_fp16)[name = string("Np_29_cast_fp16")];
            bool var_1956_transpose_x_0 = const()[name = string("op_1956_transpose_x_0"), val = bool(false)];
            bool var_1956_transpose_y_0 = const()[name = string("op_1956_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1956_cast_fp16 = matmul(transpose_x = var_1956_transpose_x_0, transpose_y = var_1956_transpose_y_0, x = Minv_37_cast_fp16, y = Np_29_cast_fp16)[name = string("op_1956_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_39_cast_fp16 = add(x = Minv_37_cast_fp16, y = var_1956_cast_fp16)[name = string("Minv_39_cast_fp16")];
            bool u_9_transpose_x_0 = const()[name = string("u_9_transpose_x_0"), val = bool(false)];
            bool u_9_transpose_y_0 = const()[name = string("u_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_9_cast_fp16 = matmul(transpose_x = u_9_transpose_x_0, transpose_y = u_9_transpose_y_0, x = Minv_39_cast_fp16, y = vb_15_cast_fp16)[name = string("u_9_cast_fp16")];
            bool var_1963_transpose_x_1 = const()[name = string("op_1963_transpose_x_1"), val = bool(false)];
            bool var_1963_transpose_y_1 = const()[name = string("op_1963_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_1963_cast_fp16 = matmul(transpose_x = var_1963_transpose_x_1, transpose_y = var_1963_transpose_y_1, x = q_33_cast_fp16, y = k_33_cast_fp16)[name = string("op_1963_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_9_cast_fp16 = mul(x = var_1963_cast_fp16, y = D_9_cast_fp16)[name = string("A_9_cast_fp16")];
            bool var_1965_transpose_x_0 = const()[name = string("op_1965_transpose_x_0"), val = bool(false)];
            bool var_1965_transpose_y_0 = const()[name = string("op_1965_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1965_cast_fp16 = matmul(transpose_x = var_1965_transpose_x_0, transpose_y = var_1965_transpose_y_0, x = A_9_cast_fp16, y = u_9_cast_fp16)[name = string("op_1965_cast_fp16")];
            tensor<fp16, [16]> var_1966_bias_0_to_fp16 = const()[name = string("op_1966_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_1966_cast_fp16 = linear(bias = var_1966_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_15_cast_fp16)[name = string("op_1966_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_5_cast_fp16 = exp(x = var_1966_cast_fp16)[name = string("tail_5_cast_fp16")];
            tensor<int32, [1]> var_1969_axes_0 = const()[name = string("op_1969_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_1969_cast_fp16 = expand_dims(axes = var_1969_axes_0, x = tail_5_cast_fp16)[name = string("op_1969_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1970_cast_fp16 = mul(x = k_33_cast_fp16, y = var_1969_cast_fp16)[name = string("op_1970_cast_fp16")];
            bool S_5_transpose_x_1 = const()[name = string("S_5_transpose_x_1"), val = bool(true)];
            bool S_5_transpose_y_1 = const()[name = string("S_5_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_5_cast_fp16 = matmul(transpose_x = S_5_transpose_x_1, transpose_y = S_5_transpose_y_1, x = var_1970_cast_fp16, y = u_9_cast_fp16)[name = string("S_5_cast_fp16")];
            tensor<int32, [4]> q_35_begin_0 = const()[name = string("q_35_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_35_end_0 = const()[name = string("q_35_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_35_end_mask_0 = const()[name = string("q_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_35_cast_fp16 = slice_by_index(begin = q_35_begin_0, end = q_35_end_0, end_mask = q_35_end_mask_0, x = q_31_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<int32, [4]> k_35_begin_0 = const()[name = string("k_35_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_35_end_0 = const()[name = string("k_35_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_35_end_mask_0 = const()[name = string("k_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_35_cast_fp16 = slice_by_index(begin = k_35_begin_0, end = k_35_end_0, end_mask = k_35_end_mask_0, x = k_31_cast_fp16)[name = string("k_35_cast_fp16")];
            tensor<int32, [4]> kb_11_begin_0 = const()[name = string("kb_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_11_end_0 = const()[name = string("kb_11_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_11_end_mask_0 = const()[name = string("kb_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_11_cast_fp16 = slice_by_index(begin = kb_11_begin_0, end = kb_11_end_0, end_mask = kb_11_end_mask_0, x = var_1875_cast_fp16)[name = string("kb_11_cast_fp16")];
            tensor<int32, [4]> vb_17_begin_0 = const()[name = string("vb_17_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_17_end_0 = const()[name = string("vb_17_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_17_end_mask_0 = const()[name = string("vb_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_17_cast_fp16 = slice_by_index(begin = vb_17_begin_0, end = vb_17_end_0, end_mask = vb_17_end_mask_0, x = vb_13_cast_fp16)[name = string("vb_17_cast_fp16")];
            tensor<int32, [3]> gdec_17_begin_0 = const()[name = string("gdec_17_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_17_end_0 = const()[name = string("gdec_17_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_17_end_mask_0 = const()[name = string("gdec_17_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_17_cast_fp16 = slice_by_index(begin = gdec_17_begin_0, end = gdec_17_end_0, end_mask = gdec_17_end_mask_0, x = gdec_13_cast_fp16)[name = string("gdec_17_cast_fp16")];
            tensor<int32, [2]> var_2055 = const()[name = string("op_2055"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_2056_cast_fp16 = reshape(shape = var_2055, x = gdec_17_cast_fp16)[name = string("op_2056_cast_fp16")];
            tensor<fp16, [12, 256]> var_2057_cast_fp16 = linear(bias = var_631_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2056_cast_fp16)[name = string("op_2057_cast_fp16")];
            tensor<int32, [4]> var_2062 = const()[name = string("op_2062"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_11_cast_fp16 = reshape(shape = var_2062, x = var_2057_cast_fp16)[name = string("P_11_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2064_cast_fp16 = exp(x = P_11_cast_fp16)[name = string("op_2064_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_11_cast_fp16 = mul(x = var_2064_cast_fp16, y = tril_incl_to_fp16)[name = string("D_11_cast_fp16")];
            bool kk_11_transpose_x_1 = const()[name = string("kk_11_transpose_x_1"), val = bool(false)];
            bool kk_11_transpose_y_1 = const()[name = string("kk_11_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_11_cast_fp16 = matmul(transpose_x = kk_11_transpose_x_1, transpose_y = kk_11_transpose_y_1, x = kb_11_cast_fp16, y = k_35_cast_fp16)[name = string("kk_11_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_2070_cast_fp16 = mul(x = kk_11_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_2070_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2071_cast_fp16 = mul(x = var_2070_cast_fp16, y = D_11_cast_fp16)[name = string("op_2071_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_11_cast_fp16 = mul(x = var_2071_cast_fp16, y = cum_tail_to_fp16)[name = string("N_11_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2073_cast_fp16 = mul(x = D_11_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2073_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_11_cast_fp16 = mul(x = kk_11_cast_fp16, y = var_2073_cast_fp16)[name = string("Na_11_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_41_cast_fp16 = add(x = eye_to_fp16, y = N_11_cast_fp16)[name = string("Minv_41_cast_fp16")];
            bool Np_31_transpose_x_0 = const()[name = string("Np_31_transpose_x_0"), val = bool(false)];
            bool Np_31_transpose_y_0 = const()[name = string("Np_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_31_cast_fp16 = matmul(transpose_x = Np_31_transpose_x_0, transpose_y = Np_31_transpose_y_0, x = N_11_cast_fp16, y = Na_11_cast_fp16)[name = string("Np_31_cast_fp16")];
            bool Npa_21_transpose_x_0 = const()[name = string("Npa_21_transpose_x_0"), val = bool(false)];
            bool Npa_21_transpose_y_0 = const()[name = string("Npa_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_21_cast_fp16 = matmul(transpose_x = Npa_21_transpose_x_0, transpose_y = Npa_21_transpose_y_0, x = Na_11_cast_fp16, y = N_11_cast_fp16)[name = string("Npa_21_cast_fp16")];
            bool var_2079_transpose_x_0 = const()[name = string("op_2079_transpose_x_0"), val = bool(false)];
            bool var_2079_transpose_y_0 = const()[name = string("op_2079_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2079_cast_fp16 = matmul(transpose_x = var_2079_transpose_x_0, transpose_y = var_2079_transpose_y_0, x = Minv_41_cast_fp16, y = Np_31_cast_fp16)[name = string("op_2079_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_43_cast_fp16 = add(x = Minv_41_cast_fp16, y = var_2079_cast_fp16)[name = string("Minv_43_cast_fp16")];
            bool Np_33_transpose_x_0 = const()[name = string("Np_33_transpose_x_0"), val = bool(false)];
            bool Np_33_transpose_y_0 = const()[name = string("Np_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_33_cast_fp16 = matmul(transpose_x = Np_33_transpose_x_0, transpose_y = Np_33_transpose_y_0, x = Np_31_cast_fp16, y = Npa_21_cast_fp16)[name = string("Np_33_cast_fp16")];
            bool Npa_23_transpose_x_0 = const()[name = string("Npa_23_transpose_x_0"), val = bool(false)];
            bool Npa_23_transpose_y_0 = const()[name = string("Npa_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_23_cast_fp16 = matmul(transpose_x = Npa_23_transpose_x_0, transpose_y = Npa_23_transpose_y_0, x = Npa_21_cast_fp16, y = Np_31_cast_fp16)[name = string("Npa_23_cast_fp16")];
            bool var_2084_transpose_x_0 = const()[name = string("op_2084_transpose_x_0"), val = bool(false)];
            bool var_2084_transpose_y_0 = const()[name = string("op_2084_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2084_cast_fp16 = matmul(transpose_x = var_2084_transpose_x_0, transpose_y = var_2084_transpose_y_0, x = Minv_43_cast_fp16, y = Np_33_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_45_cast_fp16 = add(x = Minv_43_cast_fp16, y = var_2084_cast_fp16)[name = string("Minv_45_cast_fp16")];
            bool Np_35_transpose_x_0 = const()[name = string("Np_35_transpose_x_0"), val = bool(false)];
            bool Np_35_transpose_y_0 = const()[name = string("Np_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_35_cast_fp16 = matmul(transpose_x = Np_35_transpose_x_0, transpose_y = Np_35_transpose_y_0, x = Np_33_cast_fp16, y = Npa_23_cast_fp16)[name = string("Np_35_cast_fp16")];
            bool var_2088_transpose_x_0 = const()[name = string("op_2088_transpose_x_0"), val = bool(false)];
            bool var_2088_transpose_y_0 = const()[name = string("op_2088_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2088_cast_fp16 = matmul(transpose_x = var_2088_transpose_x_0, transpose_y = var_2088_transpose_y_0, x = Minv_45_cast_fp16, y = Np_35_cast_fp16)[name = string("op_2088_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_47_cast_fp16 = add(x = Minv_45_cast_fp16, y = var_2088_cast_fp16)[name = string("Minv_47_cast_fp16")];
            bool u_11_transpose_x_0 = const()[name = string("u_11_transpose_x_0"), val = bool(false)];
            bool u_11_transpose_y_0 = const()[name = string("u_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_11_cast_fp16 = matmul(transpose_x = u_11_transpose_x_0, transpose_y = u_11_transpose_y_0, x = Minv_47_cast_fp16, y = vb_17_cast_fp16)[name = string("u_11_cast_fp16")];
            bool var_2095_transpose_x_1 = const()[name = string("op_2095_transpose_x_1"), val = bool(false)];
            bool var_2095_transpose_y_1 = const()[name = string("op_2095_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_2095_cast_fp16 = matmul(transpose_x = var_2095_transpose_x_1, transpose_y = var_2095_transpose_y_1, x = q_35_cast_fp16, y = k_35_cast_fp16)[name = string("op_2095_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_11_cast_fp16 = mul(x = var_2095_cast_fp16, y = D_11_cast_fp16)[name = string("A_11_cast_fp16")];
            tensor<fp16, [16]> var_2097_bias_0_to_fp16 = const()[name = string("op_2097_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_2097_cast_fp16 = linear(bias = var_2097_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_17_cast_fp16)[name = string("op_2097_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_5_cast_fp16 = exp(x = var_2097_cast_fp16)[name = string("cum_5_cast_fp16")];
            tensor<int32, [1]> var_2100_axes_0 = const()[name = string("op_2100_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_2100_cast_fp16 = expand_dims(axes = var_2100_axes_0, x = cum_5_cast_fp16)[name = string("op_2100_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2101_cast_fp16 = mul(x = kb_11_cast_fp16, y = var_2100_cast_fp16)[name = string("op_2101_cast_fp16")];
            bool w_11_transpose_x_0 = const()[name = string("w_11_transpose_x_0"), val = bool(false)];
            bool w_11_transpose_y_0 = const()[name = string("w_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_11_cast_fp16 = matmul(transpose_x = w_11_transpose_x_0, transpose_y = w_11_transpose_y_0, x = Minv_47_cast_fp16, y = var_2101_cast_fp16)[name = string("w_11_cast_fp16")];
            bool var_2103_transpose_x_0 = const()[name = string("op_2103_transpose_x_0"), val = bool(false)];
            bool var_2103_transpose_y_0 = const()[name = string("op_2103_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2103_cast_fp16 = matmul(transpose_x = var_2103_transpose_x_0, transpose_y = var_2103_transpose_y_0, x = w_11_cast_fp16, y = S_5_cast_fp16)[name = string("op_2103_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_5_cast_fp16 = sub(x = u_11_cast_fp16, y = var_2103_cast_fp16)[name = string("vnew_5_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2108_cast_fp16 = mul(x = q_35_cast_fp16, y = var_2100_cast_fp16)[name = string("op_2108_cast_fp16")];
            bool var_2109_transpose_x_0 = const()[name = string("op_2109_transpose_x_0"), val = bool(false)];
            bool var_2109_transpose_y_0 = const()[name = string("op_2109_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2109_cast_fp16 = matmul(transpose_x = var_2109_transpose_x_0, transpose_y = var_2109_transpose_y_0, x = var_2108_cast_fp16, y = S_5_cast_fp16)[name = string("op_2109_cast_fp16")];
            bool var_2110_transpose_x_0 = const()[name = string("op_2110_transpose_x_0"), val = bool(false)];
            bool var_2110_transpose_y_0 = const()[name = string("op_2110_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2110_cast_fp16 = matmul(transpose_x = var_2110_transpose_x_0, transpose_y = var_2110_transpose_y_0, x = A_11_cast_fp16, y = vnew_5_cast_fp16)[name = string("op_2110_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_5_cast_fp16 = add(x = var_2109_cast_fp16, y = var_2110_cast_fp16)[name = string("oc_5_cast_fp16")];
            int32 var_2114 = const()[name = string("op_2114"), val = int32(2)];
            bool o_13_interleave_0 = const()[name = string("o_13_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 32, 128]> o_13_cast_fp16 = concat(axis = var_2114, interleave = o_13_interleave_0, values = (var_1965_cast_fp16, oc_5_cast_fp16))[name = string("o_13_cast_fp16")];
            tensor<int32, [4]> var_2120 = const()[name = string("op_2120"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2126 = const()[name = string("op_2126"), val = tensor<int32, [4]>([3, 32, 4, 128])];
            tensor<fp16, [3, 32, 4, 128]> input_39_cast_fp16 = reshape(shape = var_2126, x = z_5_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> o_15_cast_fp16 = transpose(perm = var_2120, x = o_13_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [3, 32, 4, 128]> var_2128_cast_fp16 = mul(x = o_15_cast_fp16, y = o_15_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<int32, [1]> var_2133_axes_0 = const()[name = string("op_2133_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2133_keep_dims_0 = const()[name = string("op_2133_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 4, 1]> var_2133_cast_fp16 = reduce_mean(axes = var_2133_axes_0, keep_dims = var_2133_keep_dims_0, x = var_2128_cast_fp16)[name = string("op_2133_cast_fp16")];
            fp16 var_2135_to_fp16 = const()[name = string("op_2135_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 32, 4, 1]> var_2136_cast_fp16 = add(x = var_2133_cast_fp16, y = var_2135_to_fp16)[name = string("op_2136_cast_fp16")];
            fp32 var_2137_epsilon_0 = const()[name = string("op_2137_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 4, 1]> var_2137_cast_fp16 = rsqrt(epsilon = var_2137_epsilon_0, x = var_2136_cast_fp16)[name = string("op_2137_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> var_2138_cast_fp16 = mul(x = o_15_cast_fp16, y = var_2137_cast_fp16)[name = string("op_2138_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> var_2140_cast_fp16 = silu(x = input_39_cast_fp16)[name = string("op_2140_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> o_17_cast_fp16 = mul(x = var_2138_cast_fp16, y = var_2140_cast_fp16)[name = string("o_17_cast_fp16")];
            tensor<int32, [3]> var_2145 = const()[name = string("op_2145"), val = tensor<int32, [3]>([3, 32, 512])];
            tensor<fp16, [3, 32, 512]> input_41_cast_fp16 = reshape(shape = var_2145, x = o_17_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [640, 512]> layers_2_out_weight_to_fp16 = const()[name = string("layers_2_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47457728)))];
            tensor<fp16, [3, 32, 640]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_out_weight_to_fp16, x = input_41_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [3, 32, 640]> x_21_cast_fp16 = add(x = x_17_cast_fp16, y = linear_9_cast_fp16)[name = string("x_21_cast_fp16")];
            fp16 var_2156_promoted_to_fp16 = const()[name = string("op_2156_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> var_2160_cast_fp16 = pow(x = x_21_cast_fp16, y = var_2156_promoted_to_fp16)[name = string("op_2160_cast_fp16")];
            tensor<int32, [1]> var_2162_axes_0 = const()[name = string("op_2162_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2162_keep_dims_0 = const()[name = string("op_2162_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 1]> var_2162_cast_fp16 = reduce_mean(axes = var_2162_axes_0, keep_dims = var_2162_keep_dims_0, x = var_2160_cast_fp16)[name = string("op_2162_cast_fp16")];
            fp16 var_2163_to_fp16 = const()[name = string("op_2163_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 32, 1]> var_2164_cast_fp16 = add(x = var_2162_cast_fp16, y = var_2163_to_fp16)[name = string("op_2164_cast_fp16")];
            fp32 var_2165_epsilon_0 = const()[name = string("op_2165_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 1]> var_2165_cast_fp16 = rsqrt(epsilon = var_2165_epsilon_0, x = var_2164_cast_fp16)[name = string("op_2165_cast_fp16")];
            tensor<fp16, [3, 32, 640]> var_2166_cast_fp16 = mul(x = x_21_cast_fp16, y = var_2165_cast_fp16)[name = string("op_2166_cast_fp16")];
            tensor<fp16, [640]> layers_2_n2_w1_to_fp16 = const()[name = string("layers_2_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48113152)))];
            tensor<fp16, [3, 32, 640]> input_43_cast_fp16 = mul(x = var_2166_cast_fp16, y = layers_2_n2_w1_to_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_2_gate_up_weight_to_fp16 = const()[name = string("layers_2_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48114496)))];
            tensor<fp16, [3, 32, 5120]> linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_2_gate_up_weight_to_fp16, x = input_43_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [3]> input_45_begin_0 = const()[name = string("input_45_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_45_end_0 = const()[name = string("input_45_end_0"), val = tensor<int32, [3]>([3, 32, 2560])];
            tensor<bool, [3]> input_45_end_mask_0 = const()[name = string("input_45_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 2560]> input_45_cast_fp16 = slice_by_index(begin = input_45_begin_0, end = input_45_end_0, end_mask = input_45_end_mask_0, x = linear_10_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [3, 32, 2560]> var_2176_cast_fp16 = silu(x = input_45_cast_fp16)[name = string("op_2176_cast_fp16")];
            tensor<int32, [3]> var_2181_begin_0 = const()[name = string("op_2181_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_2181_end_0 = const()[name = string("op_2181_end_0"), val = tensor<int32, [3]>([3, 32, 5120])];
            tensor<bool, [3]> var_2181_end_mask_0 = const()[name = string("op_2181_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 2560]> var_2181_cast_fp16 = slice_by_index(begin = var_2181_begin_0, end = var_2181_end_0, end_mask = var_2181_end_mask_0, x = linear_10_cast_fp16)[name = string("op_2181_cast_fp16")];
            tensor<fp16, [3, 32, 2560]> input_47_cast_fp16 = mul(x = var_2176_cast_fp16, y = var_2181_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_2_down_weight_to_fp16 = const()[name = string("layers_2_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54668160)))];
            tensor<fp16, [3, 32, 640]> linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_down_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [3, 32, 640]> x_25_cast_fp16 = add(x = x_21_cast_fp16, y = linear_11_cast_fp16)[name = string("x_25_cast_fp16")];
            fp16 var_2194_to_fp16 = const()[name = string("op_2194_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 32, 640]> x_27_cast_fp16 = mul(x = x_25_cast_fp16, y = var_2194_to_fp16)[name = string("x_27_cast_fp16")];
            fp16 var_2192_promoted_to_fp16 = const()[name = string("op_2192_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> var_2196_cast_fp16 = pow(x = x_27_cast_fp16, y = var_2192_promoted_to_fp16)[name = string("op_2196_cast_fp16")];
            tensor<int32, [1]> var_2198_axes_0 = const()[name = string("op_2198_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2198_keep_dims_0 = const()[name = string("op_2198_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 1]> var_2198_cast_fp16 = reduce_mean(axes = var_2198_axes_0, keep_dims = var_2198_keep_dims_0, x = var_2196_cast_fp16)[name = string("op_2198_cast_fp16")];
            fp16 var_2199_to_fp16 = const()[name = string("op_2199_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 32, 1]> var_2200_cast_fp16 = add(x = var_2198_cast_fp16, y = var_2199_to_fp16)[name = string("op_2200_cast_fp16")];
            fp32 var_2201_epsilon_0 = const()[name = string("op_2201_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 1]> var_2201_cast_fp16 = rsqrt(epsilon = var_2201_epsilon_0, x = var_2200_cast_fp16)[name = string("op_2201_cast_fp16")];
            tensor<fp16, [3, 32, 640]> var_2202_cast_fp16 = mul(x = x_27_cast_fp16, y = var_2201_cast_fp16)[name = string("op_2202_cast_fp16")];
            tensor<fp16, [640]> layers_3_n1_w1_to_fp16 = const()[name = string("layers_3_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57945024)))];
            tensor<fp16, [3, 32, 640]> input_49_cast_fp16 = mul(x = var_2202_cast_fp16, y = layers_3_n1_w1_to_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1280, 640]> layers_3_qkv_weight_to_fp16 = const()[name = string("layers_3_qkv_weight_to_fp16"), val = tensor<fp16, [1280, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57946368)))];
            tensor<fp16, [1280]> linear_12_bias_0_to_fp16 = const()[name = string("linear_12_bias_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59584832)))];
            tensor<fp16, [3, 32, 1280]> linear_12_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = layers_3_qkv_weight_to_fp16, x = input_49_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [3]> var_2211_begin_0 = const()[name = string("op_2211_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2211_end_0 = const()[name = string("op_2211_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> var_2211_end_mask_0 = const()[name = string("op_2211_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 1024]> var_2211_cast_fp16 = slice_by_index(begin = var_2211_begin_0, end = var_2211_end_0, end_mask = var_2211_end_mask_0, x = linear_12_cast_fp16)[name = string("op_2211_cast_fp16")];
            tensor<int32, [4]> var_2216 = const()[name = string("op_2216"), val = tensor<int32, [4]>([3, 32, 8, 128])];
            tensor<fp16, [3, 32, 8, 128]> qg_1_cast_fp16 = reshape(shape = var_2216, x = var_2211_cast_fp16)[name = string("qg_1_cast_fp16")];
            tensor<int32, [4]> x_29_begin_0 = const()[name = string("x_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_29_end_0 = const()[name = string("x_29_end_0"), val = tensor<int32, [4]>([3, 32, 8, 64])];
            tensor<bool, [4]> x_29_end_mask_0 = const()[name = string("x_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 32, 8, 64]> x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, x = qg_1_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [4]> var_2227_begin_0 = const()[name = string("op_2227_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> var_2227_end_0 = const()[name = string("op_2227_end_0"), val = tensor<int32, [4]>([3, 32, 8, 128])];
            tensor<bool, [4]> var_2227_end_mask_0 = const()[name = string("op_2227_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 32, 8, 64]> var_2227_cast_fp16 = slice_by_index(begin = var_2227_begin_0, end = var_2227_end_0, end_mask = var_2227_end_mask_0, x = qg_1_cast_fp16)[name = string("op_2227_cast_fp16")];
            tensor<int32, [3]> var_2231 = const()[name = string("op_2231"), val = tensor<int32, [3]>([3, 32, 512])];
            tensor<fp16, [3, 32, 512]> gate_1_cast_fp16 = reshape(shape = var_2231, x = var_2227_cast_fp16)[name = string("gate_1_cast_fp16")];
            tensor<int32, [3]> var_2237_begin_0 = const()[name = string("op_2237_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> var_2237_end_0 = const()[name = string("op_2237_end_0"), val = tensor<int32, [3]>([3, 32, 1152])];
            tensor<bool, [3]> var_2237_end_mask_0 = const()[name = string("op_2237_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 128]> var_2237_cast_fp16 = slice_by_index(begin = var_2237_begin_0, end = var_2237_end_0, end_mask = var_2237_end_mask_0, x = linear_12_cast_fp16)[name = string("op_2237_cast_fp16")];
            tensor<int32, [4]> var_2242 = const()[name = string("op_2242"), val = tensor<int32, [4]>([3, 32, 2, 64])];
            tensor<fp16, [3, 32, 2, 64]> x_35_cast_fp16 = reshape(shape = var_2242, x = var_2237_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<int32, [3]> var_2248_begin_0 = const()[name = string("op_2248_begin_0"), val = tensor<int32, [3]>([0, 0, 1152])];
            tensor<int32, [3]> var_2248_end_0 = const()[name = string("op_2248_end_0"), val = tensor<int32, [3]>([3, 32, 1280])];
            tensor<bool, [3]> var_2248_end_mask_0 = const()[name = string("op_2248_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 128]> var_2248_cast_fp16 = slice_by_index(begin = var_2248_begin_0, end = var_2248_end_0, end_mask = var_2248_end_mask_0, x = linear_12_cast_fp16)[name = string("op_2248_cast_fp16")];
            tensor<int32, [4]> var_2253 = const()[name = string("op_2253"), val = tensor<int32, [4]>([3, 32, 2, 64])];
            tensor<fp16, [3, 32, 2, 64]> var_2254_cast_fp16 = reshape(shape = var_2253, x = var_2248_cast_fp16)[name = string("op_2254_cast_fp16")];
            tensor<int32, [4]> var_2259 = const()[name = string("op_2259"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2265_promoted_to_fp16 = const()[name = string("op_2265_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 8, 64]> var_2269_cast_fp16 = pow(x = x_29_cast_fp16, y = var_2265_promoted_to_fp16)[name = string("op_2269_cast_fp16")];
            tensor<int32, [1]> var_2271_axes_0 = const()[name = string("op_2271_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2271_keep_dims_0 = const()[name = string("op_2271_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 8, 1]> var_2271_cast_fp16 = reduce_mean(axes = var_2271_axes_0, keep_dims = var_2271_keep_dims_0, x = var_2269_cast_fp16)[name = string("op_2271_cast_fp16")];
            fp16 var_2272_to_fp16 = const()[name = string("op_2272_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 32, 8, 1]> var_2273_cast_fp16 = add(x = var_2271_cast_fp16, y = var_2272_to_fp16)[name = string("op_2273_cast_fp16")];
            fp32 var_2274_epsilon_0 = const()[name = string("op_2274_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 8, 1]> var_2274_cast_fp16 = rsqrt(epsilon = var_2274_epsilon_0, x = var_2273_cast_fp16)[name = string("op_2274_cast_fp16")];
            tensor<fp16, [3, 32, 8, 64]> var_2275_cast_fp16 = mul(x = x_29_cast_fp16, y = var_2274_cast_fp16)[name = string("op_2275_cast_fp16")];
            tensor<fp16, [64]> layers_3_qn_w1_to_fp16 = const()[name = string("layers_3_qn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59587456)))];
            tensor<fp16, [3, 32, 8, 64]> var_2276_cast_fp16 = mul(x = var_2275_cast_fp16, y = layers_3_qn_w1_to_fp16)[name = string("op_2276_cast_fp16")];
            tensor<int32, [4]> var_2281 = const()[name = string("op_2281"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_1_begin_0 = const()[name = string("r_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_1_end_0 = const()[name = string("r_1_end_0"), val = tensor<int32, [4]>([3, 8, 32, 16])];
            tensor<bool, [4]> r_1_end_mask_0 = const()[name = string("r_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 8, 32, 64]> x_33_cast_fp16 = transpose(perm = var_2281, x = var_2276_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [3, 8, 32, 16]> r_1_cast_fp16 = slice_by_index(begin = r_1_begin_0, end = r_1_end_0, end_mask = r_1_end_mask_0, x = x_33_cast_fp16)[name = string("r_1_cast_fp16")];
            tensor<int32, [4]> p_1_begin_0 = const()[name = string("p_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_1_end_0 = const()[name = string("p_1_end_0"), val = tensor<int32, [4]>([3, 8, 32, 64])];
            tensor<bool, [4]> p_1_end_mask_0 = const()[name = string("p_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 8, 32, 48]> p_1_cast_fp16 = slice_by_index(begin = p_1_begin_0, end = p_1_end_0, end_mask = p_1_end_mask_0, x = x_33_cast_fp16)[name = string("p_1_cast_fp16")];
            tensor<int32, [4]> var_2297_begin_0 = const()[name = string("op_2297_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_2297_end_0 = const()[name = string("op_2297_end_0"), val = tensor<int32, [4]>([3, 8, 32, 16])];
            tensor<bool, [4]> var_2297_end_mask_0 = const()[name = string("op_2297_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 8, 32, 8]> var_2297_cast_fp16 = slice_by_index(begin = var_2297_begin_0, end = var_2297_end_0, end_mask = var_2297_end_mask_0, x = r_1_cast_fp16)[name = string("op_2297_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 8, 32, 8]> var_2298_cast_fp16 = mul(x = var_2297_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_2298_cast_fp16")];
            tensor<int32, [4]> var_2303_begin_0 = const()[name = string("op_2303_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2303_end_0 = const()[name = string("op_2303_end_0"), val = tensor<int32, [4]>([3, 8, 32, 8])];
            tensor<bool, [4]> var_2303_end_mask_0 = const()[name = string("op_2303_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 8, 32, 8]> var_2303_cast_fp16 = slice_by_index(begin = var_2303_begin_0, end = var_2303_end_0, end_mask = var_2303_end_mask_0, x = r_1_cast_fp16)[name = string("op_2303_cast_fp16")];
            int32 var_2305 = const()[name = string("op_2305"), val = int32(-1)];
            bool rh_1_interleave_0 = const()[name = string("rh_1_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 8, 32, 16]> rh_1_cast_fp16 = concat(axis = var_2305, interleave = rh_1_interleave_0, values = (var_2298_cast_fp16, var_2303_cast_fp16))[name = string("rh_1_cast_fp16")];
            tensor<fp16, [32, 16]> cos_to_fp16 = const()[name = string("cos_to_fp16"), val = tensor<fp16, [32, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658880)))];
            tensor<fp16, [3, 8, 32, 16]> var_2307_cast_fp16 = mul(x = r_1_cast_fp16, y = cos_to_fp16)[name = string("op_2307_cast_fp16")];
            tensor<fp16, [32, 16]> sin_to_fp16 = const()[name = string("sin_to_fp16"), val = tensor<fp16, [32, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119659968)))];
            tensor<fp16, [3, 8, 32, 16]> var_2308_cast_fp16 = mul(x = rh_1_cast_fp16, y = sin_to_fp16)[name = string("op_2308_cast_fp16")];
            tensor<fp16, [3, 8, 32, 16]> var_2310_cast_fp16 = add(x = var_2307_cast_fp16, y = var_2308_cast_fp16)[name = string("op_2310_cast_fp16")];
            int32 var_2312 = const()[name = string("op_2312"), val = int32(-1)];
            bool q_37_interleave_0 = const()[name = string("q_37_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 8, 32, 64]> q_37_cast_fp16 = concat(axis = var_2312, interleave = q_37_interleave_0, values = (var_2310_cast_fp16, p_1_cast_fp16))[name = string("q_37_cast_fp16")];
            fp16 var_2318_promoted_to_fp16 = const()[name = string("op_2318_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 2, 64]> var_2322_cast_fp16 = pow(x = x_35_cast_fp16, y = var_2318_promoted_to_fp16)[name = string("op_2322_cast_fp16")];
            tensor<int32, [1]> var_2324_axes_0 = const()[name = string("op_2324_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2324_keep_dims_0 = const()[name = string("op_2324_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 2, 1]> var_2324_cast_fp16 = reduce_mean(axes = var_2324_axes_0, keep_dims = var_2324_keep_dims_0, x = var_2322_cast_fp16)[name = string("op_2324_cast_fp16")];
            fp16 var_2325_to_fp16 = const()[name = string("op_2325_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 32, 2, 1]> var_2326_cast_fp16 = add(x = var_2324_cast_fp16, y = var_2325_to_fp16)[name = string("op_2326_cast_fp16")];
            fp32 var_2327_epsilon_0 = const()[name = string("op_2327_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 2, 1]> var_2327_cast_fp16 = rsqrt(epsilon = var_2327_epsilon_0, x = var_2326_cast_fp16)[name = string("op_2327_cast_fp16")];
            tensor<fp16, [3, 32, 2, 64]> var_2328_cast_fp16 = mul(x = x_35_cast_fp16, y = var_2327_cast_fp16)[name = string("op_2328_cast_fp16")];
            tensor<fp16, [64]> layers_3_kn_w1_to_fp16 = const()[name = string("layers_3_kn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59588800)))];
            tensor<fp16, [3, 32, 2, 64]> var_2329_cast_fp16 = mul(x = var_2328_cast_fp16, y = layers_3_kn_w1_to_fp16)[name = string("op_2329_cast_fp16")];
            tensor<int32, [4]> var_2334 = const()[name = string("op_2334"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_3_begin_0 = const()[name = string("r_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_3_end_0 = const()[name = string("r_3_end_0"), val = tensor<int32, [4]>([3, 2, 32, 16])];
            tensor<bool, [4]> r_3_end_mask_0 = const()[name = string("r_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 2, 32, 64]> x_39_cast_fp16 = transpose(perm = var_2334, x = var_2329_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [3, 2, 32, 16]> r_3_cast_fp16 = slice_by_index(begin = r_3_begin_0, end = r_3_end_0, end_mask = r_3_end_mask_0, x = x_39_cast_fp16)[name = string("r_3_cast_fp16")];
            tensor<int32, [4]> p_3_begin_0 = const()[name = string("p_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_3_end_0 = const()[name = string("p_3_end_0"), val = tensor<int32, [4]>([3, 2, 32, 64])];
            tensor<bool, [4]> p_3_end_mask_0 = const()[name = string("p_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 2, 32, 48]> p_3_cast_fp16 = slice_by_index(begin = p_3_begin_0, end = p_3_end_0, end_mask = p_3_end_mask_0, x = x_39_cast_fp16)[name = string("p_3_cast_fp16")];
            tensor<int32, [4]> var_2350_begin_0 = const()[name = string("op_2350_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_2350_end_0 = const()[name = string("op_2350_end_0"), val = tensor<int32, [4]>([3, 2, 32, 16])];
            tensor<bool, [4]> var_2350_end_mask_0 = const()[name = string("op_2350_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 2, 32, 8]> var_2350_cast_fp16 = slice_by_index(begin = var_2350_begin_0, end = var_2350_end_0, end_mask = var_2350_end_mask_0, x = r_3_cast_fp16)[name = string("op_2350_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 2, 32, 8]> var_2351_cast_fp16 = mul(x = var_2350_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_2351_cast_fp16")];
            tensor<int32, [4]> var_2356_begin_0 = const()[name = string("op_2356_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2356_end_0 = const()[name = string("op_2356_end_0"), val = tensor<int32, [4]>([3, 2, 32, 8])];
            tensor<bool, [4]> var_2356_end_mask_0 = const()[name = string("op_2356_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 2, 32, 8]> var_2356_cast_fp16 = slice_by_index(begin = var_2356_begin_0, end = var_2356_end_0, end_mask = var_2356_end_mask_0, x = r_3_cast_fp16)[name = string("op_2356_cast_fp16")];
            int32 var_2358 = const()[name = string("op_2358"), val = int32(-1)];
            bool rh_3_interleave_0 = const()[name = string("rh_3_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 2, 32, 16]> rh_3_cast_fp16 = concat(axis = var_2358, interleave = rh_3_interleave_0, values = (var_2351_cast_fp16, var_2356_cast_fp16))[name = string("rh_3_cast_fp16")];
            tensor<fp16, [3, 2, 32, 16]> var_2360_cast_fp16 = mul(x = r_3_cast_fp16, y = cos_to_fp16)[name = string("op_2360_cast_fp16")];
            tensor<fp16, [3, 2, 32, 16]> var_2361_cast_fp16 = mul(x = rh_3_cast_fp16, y = sin_to_fp16)[name = string("op_2361_cast_fp16")];
            tensor<fp16, [3, 2, 32, 16]> var_2363_cast_fp16 = add(x = var_2360_cast_fp16, y = var_2361_cast_fp16)[name = string("op_2363_cast_fp16")];
            int32 var_2365 = const()[name = string("op_2365"), val = int32(-1)];
            bool k_37_interleave_0 = const()[name = string("k_37_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 2, 32, 64]> k_37_cast_fp16 = concat(axis = var_2365, interleave = k_37_interleave_0, values = (var_2363_cast_fp16, p_3_cast_fp16))[name = string("k_37_cast_fp16")];
            tensor<int32, [1]> var_2378_axes_0 = const()[name = string("op_2378_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 1, 32, 64]> var_2378_cast_fp16 = expand_dims(axes = var_2378_axes_0, x = k_37_cast_fp16)[name = string("op_2378_cast_fp16")];
            tensor<int32, [5]> var_2386_reps_0 = const()[name = string("op_2386_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [3, 2, 4, 32, 64]> var_2386_cast_fp16 = tile(reps = var_2386_reps_0, x = var_2378_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<int32, [4]> var_2391 = const()[name = string("op_2391"), val = tensor<int32, [4]>([3, 8, 32, 64])];
            tensor<fp16, [3, 8, 32, 64]> k_39_cast_fp16 = reshape(shape = var_2391, x = var_2386_cast_fp16)[name = string("k_39_cast_fp16")];
            tensor<int32, [1]> var_2404_axes_0 = const()[name = string("op_2404_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 32, 64]> v_7_cast_fp16 = transpose(perm = var_2259, x = var_2254_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [3, 2, 1, 32, 64]> var_2404_cast_fp16 = expand_dims(axes = var_2404_axes_0, x = v_7_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<int32, [5]> var_2412_reps_0 = const()[name = string("op_2412_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [3, 2, 4, 32, 64]> var_2412_cast_fp16 = tile(reps = var_2412_reps_0, x = var_2404_cast_fp16)[name = string("op_2412_cast_fp16")];
            tensor<int32, [4]> var_2417 = const()[name = string("op_2417"), val = tensor<int32, [4]>([3, 8, 32, 64])];
            tensor<fp16, [3, 8, 32, 64]> v_9_cast_fp16 = reshape(shape = var_2417, x = var_2412_cast_fp16)[name = string("v_9_cast_fp16")];
            bool var_2422_transpose_x_1 = const()[name = string("op_2422_transpose_x_1"), val = bool(false)];
            bool var_2422_transpose_y_1 = const()[name = string("op_2422_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 8, 32, 32]> var_2422_cast_fp16 = matmul(transpose_x = var_2422_transpose_x_1, transpose_y = var_2422_transpose_y_1, x = q_37_cast_fp16, y = k_39_cast_fp16)[name = string("op_2422_cast_fp16")];
            fp16 var_2423_to_fp16 = const()[name = string("op_2423_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [3, 8, 32, 32]> var_2424_cast_fp16 = mul(x = var_2422_cast_fp16, y = var_2423_to_fp16)[name = string("op_2424_cast_fp16")];
            tensor<fp16, [32, 32]> causal_add_to_fp16 = const()[name = string("causal_add_to_fp16"), val = tensor<fp16, [32, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119661056)))];
            tensor<fp16, [3, 8, 32, 32]> s_1_cast_fp16 = add(x = var_2424_cast_fp16, y = causal_add_to_fp16)[name = string("s_1_cast_fp16")];
            int32 var_2427 = const()[name = string("op_2427"), val = int32(-1)];
            tensor<fp16, [3, 8, 32, 32]> var_2429_cast_fp16 = softmax(axis = var_2427, x = s_1_cast_fp16)[name = string("op_2429_cast_fp16")];
            bool o_19_transpose_x_0 = const()[name = string("o_19_transpose_x_0"), val = bool(false)];
            bool o_19_transpose_y_0 = const()[name = string("o_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 8, 32, 64]> o_19_cast_fp16 = matmul(transpose_x = o_19_transpose_x_0, transpose_y = o_19_transpose_y_0, x = var_2429_cast_fp16, y = v_9_cast_fp16)[name = string("o_19_cast_fp16")];
            tensor<int32, [4]> var_2435 = const()[name = string("op_2435"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2440 = const()[name = string("op_2440"), val = tensor<int32, [3]>([3, 32, 512])];
            tensor<fp16, [3, 32, 8, 64]> var_2436_cast_fp16 = transpose(perm = var_2435, x = o_19_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [3, 32, 512]> var_2441_cast_fp16 = reshape(shape = var_2440, x = var_2436_cast_fp16)[name = string("op_2441_cast_fp16")];
            tensor<fp16, [3, 32, 512]> var_2442_cast_fp16 = sigmoid(x = gate_1_cast_fp16)[name = string("op_2442_cast_fp16")];
            tensor<fp16, [3, 32, 512]> input_51_cast_fp16 = mul(x = var_2441_cast_fp16, y = var_2442_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [640, 512]> layers_3_o_weight_to_fp16 = const()[name = string("layers_3_o_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59589568)))];
            tensor<fp16, [3, 32, 640]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_o_weight_to_fp16, x = input_51_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [3, 32, 640]> x_41_cast_fp16 = add(x = x_25_cast_fp16, y = linear_13_cast_fp16)[name = string("x_41_cast_fp16")];
            fp16 var_2455_to_fp16 = const()[name = string("op_2455_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 32, 640]> x_43_cast_fp16 = mul(x = x_41_cast_fp16, y = var_2455_to_fp16)[name = string("x_43_cast_fp16")];
            fp16 var_2453_promoted_to_fp16 = const()[name = string("op_2453_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> var_2457_cast_fp16 = pow(x = x_43_cast_fp16, y = var_2453_promoted_to_fp16)[name = string("op_2457_cast_fp16")];
            tensor<int32, [1]> var_2459_axes_0 = const()[name = string("op_2459_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2459_keep_dims_0 = const()[name = string("op_2459_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 1]> var_2459_cast_fp16 = reduce_mean(axes = var_2459_axes_0, keep_dims = var_2459_keep_dims_0, x = var_2457_cast_fp16)[name = string("op_2459_cast_fp16")];
            fp16 var_2460_to_fp16 = const()[name = string("op_2460_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 32, 1]> var_2461_cast_fp16 = add(x = var_2459_cast_fp16, y = var_2460_to_fp16)[name = string("op_2461_cast_fp16")];
            fp32 var_2462_epsilon_0 = const()[name = string("op_2462_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 1]> var_2462_cast_fp16 = rsqrt(epsilon = var_2462_epsilon_0, x = var_2461_cast_fp16)[name = string("op_2462_cast_fp16")];
            tensor<fp16, [3, 32, 640]> var_2463_cast_fp16 = mul(x = x_43_cast_fp16, y = var_2462_cast_fp16)[name = string("op_2463_cast_fp16")];
            tensor<fp16, [640]> layers_3_n2_w1_to_fp16 = const()[name = string("layers_3_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60244992)))];
            tensor<fp16, [3, 32, 640]> input_53_cast_fp16 = mul(x = var_2463_cast_fp16, y = layers_3_n2_w1_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_3_gate_up_weight_to_fp16 = const()[name = string("layers_3_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60246336)))];
            tensor<fp16, [3, 32, 5120]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_3_gate_up_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [3]> input_55_begin_0 = const()[name = string("input_55_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_55_end_0 = const()[name = string("input_55_end_0"), val = tensor<int32, [3]>([3, 32, 2560])];
            tensor<bool, [3]> input_55_end_mask_0 = const()[name = string("input_55_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 2560]> input_55_cast_fp16 = slice_by_index(begin = input_55_begin_0, end = input_55_end_0, end_mask = input_55_end_mask_0, x = linear_14_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [3, 32, 2560]> var_2473_cast_fp16 = silu(x = input_55_cast_fp16)[name = string("op_2473_cast_fp16")];
            tensor<int32, [3]> var_2478_begin_0 = const()[name = string("op_2478_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_2478_end_0 = const()[name = string("op_2478_end_0"), val = tensor<int32, [3]>([3, 32, 5120])];
            tensor<bool, [3]> var_2478_end_mask_0 = const()[name = string("op_2478_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 2560]> var_2478_cast_fp16 = slice_by_index(begin = var_2478_begin_0, end = var_2478_end_0, end_mask = var_2478_end_mask_0, x = linear_14_cast_fp16)[name = string("op_2478_cast_fp16")];
            tensor<fp16, [3, 32, 2560]> input_57_cast_fp16 = mul(x = var_2473_cast_fp16, y = var_2478_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_3_down_weight_to_fp16 = const()[name = string("layers_3_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66800000)))];
            tensor<fp16, [3, 32, 640]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_down_weight_to_fp16, x = input_57_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [3, 32, 640]> x_45_cast_fp16 = add(x = x_41_cast_fp16, y = linear_15_cast_fp16)[name = string("x_45_cast_fp16")];
            fp16 var_2491_to_fp16 = const()[name = string("op_2491_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 32, 640]> x_47_cast_fp16 = mul(x = x_45_cast_fp16, y = var_2491_to_fp16)[name = string("x_47_cast_fp16")];
            fp16 var_2489_promoted_to_fp16 = const()[name = string("op_2489_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> var_2493_cast_fp16 = pow(x = x_47_cast_fp16, y = var_2489_promoted_to_fp16)[name = string("op_2493_cast_fp16")];
            tensor<int32, [1]> var_2495_axes_0 = const()[name = string("op_2495_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2495_keep_dims_0 = const()[name = string("op_2495_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 1]> var_2495_cast_fp16 = reduce_mean(axes = var_2495_axes_0, keep_dims = var_2495_keep_dims_0, x = var_2493_cast_fp16)[name = string("op_2495_cast_fp16")];
            fp16 var_2496_to_fp16 = const()[name = string("op_2496_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 32, 1]> var_2497_cast_fp16 = add(x = var_2495_cast_fp16, y = var_2496_to_fp16)[name = string("op_2497_cast_fp16")];
            fp32 var_2498_epsilon_0 = const()[name = string("op_2498_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 1]> var_2498_cast_fp16 = rsqrt(epsilon = var_2498_epsilon_0, x = var_2497_cast_fp16)[name = string("op_2498_cast_fp16")];
            tensor<fp16, [3, 32, 640]> var_2499_cast_fp16 = mul(x = x_47_cast_fp16, y = var_2498_cast_fp16)[name = string("op_2499_cast_fp16")];
            tensor<fp16, [640]> layers_4_n1_w1_to_fp16 = const()[name = string("layers_4_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70076864)))];
            tensor<fp16, [3, 32, 640]> input_59_cast_fp16 = mul(x = var_2499_cast_fp16, y = layers_4_n1_w1_to_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_4_qkvzba_weight_to_fp16 = const()[name = string("layers_4_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70078208)))];
            tensor<fp16, [3, 32, 1544]> linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_qkvzba_weight_to_fp16, x = input_59_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [3]> qkv_7_begin_0 = const()[name = string("qkv_7_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_7_end_0 = const()[name = string("qkv_7_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> qkv_7_end_mask_0 = const()[name = string("qkv_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 1024]> qkv_7_cast_fp16 = slice_by_index(begin = qkv_7_begin_0, end = qkv_7_end_0, end_mask = qkv_7_end_mask_0, x = linear_16_cast_fp16)[name = string("qkv_7_cast_fp16")];
            tensor<int32, [3]> z_7_begin_0 = const()[name = string("z_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_7_end_0 = const()[name = string("z_7_end_0"), val = tensor<int32, [3]>([3, 32, 1536])];
            tensor<bool, [3]> z_7_end_mask_0 = const()[name = string("z_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 512]> z_7_cast_fp16 = slice_by_index(begin = z_7_begin_0, end = z_7_end_0, end_mask = z_7_end_mask_0, x = linear_16_cast_fp16)[name = string("z_7_cast_fp16")];
            tensor<int32, [3]> b_7_begin_0 = const()[name = string("b_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_7_end_0 = const()[name = string("b_7_end_0"), val = tensor<int32, [3]>([3, 32, 1540])];
            tensor<bool, [3]> b_7_end_mask_0 = const()[name = string("b_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 4]> b_7_cast_fp16 = slice_by_index(begin = b_7_begin_0, end = b_7_end_0, end_mask = b_7_end_mask_0, x = linear_16_cast_fp16)[name = string("b_7_cast_fp16")];
            tensor<int32, [3]> a_7_begin_0 = const()[name = string("a_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_7_end_0 = const()[name = string("a_7_end_0"), val = tensor<int32, [3]>([3, 32, 1])];
            tensor<bool, [3]> a_7_end_mask_0 = const()[name = string("a_7_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 4]> a_7_cast_fp16 = slice_by_index(begin = a_7_begin_0, end = a_7_end_0, end_mask = a_7_end_mask_0, x = linear_16_cast_fp16)[name = string("a_7_cast_fp16")];
            int32 var_2538 = const()[name = string("op_2538"), val = int32(1)];
            bool xp_7_interleave_0 = const()[name = string("xp_7_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 35, 1024]> xp_7_cast_fp16 = concat(axis = var_2538, interleave = xp_7_interleave_0, values = (zpad_to_fp16, qkv_7_cast_fp16))[name = string("xp_7_cast_fp16")];
            tensor<int32, [3]> var_2549_begin_0 = const()[name = string("op_2549_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_2549_end_0 = const()[name = string("op_2549_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_2549_end_mask_0 = const()[name = string("op_2549_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 1024]> var_2549_cast_fp16 = slice_by_index(begin = var_2549_begin_0, end = var_2549_end_0, end_mask = var_2549_end_mask_0, x = xp_7_cast_fp16)[name = string("op_2549_cast_fp16")];
            tensor<fp16, [1024]> var_2557_to_fp16 = const()[name = string("op_2557_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72054592)))];
            tensor<fp16, [3, 32, 1024]> var_2558_cast_fp16 = mul(x = var_2549_cast_fp16, y = var_2557_to_fp16)[name = string("op_2558_cast_fp16")];
            tensor<int32, [3]> var_2568_begin_0 = const()[name = string("op_2568_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_2568_end_0 = const()[name = string("op_2568_end_0"), val = tensor<int32, [3]>([3, 34, 1024])];
            tensor<bool, [3]> var_2568_end_mask_0 = const()[name = string("op_2568_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_2568_cast_fp16 = slice_by_index(begin = var_2568_begin_0, end = var_2568_end_0, end_mask = var_2568_end_mask_0, x = xp_7_cast_fp16)[name = string("op_2568_cast_fp16")];
            tensor<fp16, [1024]> var_2576_to_fp16 = const()[name = string("op_2576_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72056704)))];
            tensor<fp16, [3, 32, 1024]> var_2577_cast_fp16 = mul(x = var_2568_cast_fp16, y = var_2576_to_fp16)[name = string("op_2577_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> var_2579_cast_fp16 = add(x = var_2558_cast_fp16, y = var_2577_cast_fp16)[name = string("op_2579_cast_fp16")];
            tensor<int32, [3]> var_2589_begin_0 = const()[name = string("op_2589_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2589_end_0 = const()[name = string("op_2589_end_0"), val = tensor<int32, [3]>([3, 33, 1024])];
            tensor<bool, [3]> var_2589_end_mask_0 = const()[name = string("op_2589_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_2589_cast_fp16 = slice_by_index(begin = var_2589_begin_0, end = var_2589_end_0, end_mask = var_2589_end_mask_0, x = xp_7_cast_fp16)[name = string("op_2589_cast_fp16")];
            tensor<fp16, [1024]> var_2597_to_fp16 = const()[name = string("op_2597_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72058816)))];
            tensor<fp16, [3, 32, 1024]> var_2598_cast_fp16 = mul(x = var_2589_cast_fp16, y = var_2597_to_fp16)[name = string("op_2598_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> var_2600_cast_fp16 = add(x = var_2579_cast_fp16, y = var_2598_cast_fp16)[name = string("op_2600_cast_fp16")];
            tensor<int32, [3]> var_2610_begin_0 = const()[name = string("op_2610_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2610_end_0 = const()[name = string("op_2610_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> var_2610_end_mask_0 = const()[name = string("op_2610_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_2610_cast_fp16 = slice_by_index(begin = var_2610_begin_0, end = var_2610_end_0, end_mask = var_2610_end_mask_0, x = xp_7_cast_fp16)[name = string("op_2610_cast_fp16")];
            tensor<fp16, [1024]> var_2618_to_fp16 = const()[name = string("op_2618_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72060928)))];
            tensor<fp16, [3, 32, 1024]> var_2619_cast_fp16 = mul(x = var_2610_cast_fp16, y = var_2618_to_fp16)[name = string("op_2619_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> input_61_cast_fp16 = add(x = var_2600_cast_fp16, y = var_2619_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> c_7_cast_fp16 = silu(x = input_61_cast_fp16)[name = string("c_7_cast_fp16")];
            tensor<int32, [3]> var_2627_begin_0 = const()[name = string("op_2627_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2627_end_0 = const()[name = string("op_2627_end_0"), val = tensor<int32, [3]>([3, 32, 256])];
            tensor<bool, [3]> var_2627_end_mask_0 = const()[name = string("op_2627_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 256]> var_2627_cast_fp16 = slice_by_index(begin = var_2627_begin_0, end = var_2627_end_0, end_mask = var_2627_end_mask_0, x = c_7_cast_fp16)[name = string("op_2627_cast_fp16")];
            tensor<int32, [4]> var_2632 = const()[name = string("op_2632"), val = tensor<int32, [4]>([3, 32, 2, 128])];
            tensor<fp16, [3, 32, 2, 128]> q_39_cast_fp16 = reshape(shape = var_2632, x = var_2627_cast_fp16)[name = string("q_39_cast_fp16")];
            tensor<int32, [3]> var_2638_begin_0 = const()[name = string("op_2638_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_2638_end_0 = const()[name = string("op_2638_end_0"), val = tensor<int32, [3]>([3, 32, 512])];
            tensor<bool, [3]> var_2638_end_mask_0 = const()[name = string("op_2638_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 256]> var_2638_cast_fp16 = slice_by_index(begin = var_2638_begin_0, end = var_2638_end_0, end_mask = var_2638_end_mask_0, x = c_7_cast_fp16)[name = string("op_2638_cast_fp16")];
            tensor<int32, [4]> var_2643 = const()[name = string("op_2643"), val = tensor<int32, [4]>([3, 32, 2, 128])];
            tensor<fp16, [3, 32, 2, 128]> k_41_cast_fp16 = reshape(shape = var_2643, x = var_2638_cast_fp16)[name = string("k_41_cast_fp16")];
            tensor<int32, [3]> var_2649_begin_0 = const()[name = string("op_2649_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_2649_end_0 = const()[name = string("op_2649_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> var_2649_end_mask_0 = const()[name = string("op_2649_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 512]> var_2649_cast_fp16 = slice_by_index(begin = var_2649_begin_0, end = var_2649_end_0, end_mask = var_2649_end_mask_0, x = c_7_cast_fp16)[name = string("op_2649_cast_fp16")];
            tensor<int32, [4]> var_2654 = const()[name = string("op_2654"), val = tensor<int32, [4]>([3, 32, 4, 128])];
            tensor<fp16, [3, 32, 4, 128]> var_2655_cast_fp16 = reshape(shape = var_2654, x = var_2649_cast_fp16)[name = string("op_2655_cast_fp16")];
            tensor<int32, [4]> var_2660 = const()[name = string("op_2660"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2662_to_fp16 = const()[name = string("op_2662_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 32, 2, 128]> q_41_cast_fp16 = mul(x = q_39_cast_fp16, y = var_2662_to_fp16)[name = string("q_41_cast_fp16")];
            fp16 var_2664_to_fp16 = const()[name = string("op_2664_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 32, 2, 128]> k_43_cast_fp16 = mul(x = k_41_cast_fp16, y = var_2664_to_fp16)[name = string("k_43_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_2666_cast_fp16 = mul(x = q_41_cast_fp16, y = q_41_cast_fp16)[name = string("op_2666_cast_fp16")];
            tensor<int32, [1]> var_2671_axes_0 = const()[name = string("op_2671_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2671_keep_dims_0 = const()[name = string("op_2671_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 2, 1]> var_2671_cast_fp16 = reduce_mean(axes = var_2671_axes_0, keep_dims = var_2671_keep_dims_0, x = var_2666_cast_fp16)[name = string("op_2671_cast_fp16")];
            fp16 var_2672_promoted_to_fp16 = const()[name = string("op_2672_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 32, 2, 1]> var_2673_cast_fp16 = mul(x = var_2671_cast_fp16, y = var_2672_promoted_to_fp16)[name = string("op_2673_cast_fp16")];
            fp16 var_2675_to_fp16 = const()[name = string("op_2675_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 32, 2, 1]> var_2676_cast_fp16 = add(x = var_2673_cast_fp16, y = var_2675_to_fp16)[name = string("op_2676_cast_fp16")];
            fp32 var_2677_epsilon_0 = const()[name = string("op_2677_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 2, 1]> var_2677_cast_fp16 = rsqrt(epsilon = var_2677_epsilon_0, x = var_2676_cast_fp16)[name = string("op_2677_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_2678_cast_fp16 = mul(x = q_41_cast_fp16, y = var_2677_cast_fp16)[name = string("op_2678_cast_fp16")];
            fp16 var_2679_to_fp16 = const()[name = string("op_2679_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 32, 2, 128]> q_43_cast_fp16 = mul(x = var_2678_cast_fp16, y = var_2679_to_fp16)[name = string("q_43_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_2681_cast_fp16 = mul(x = k_43_cast_fp16, y = k_43_cast_fp16)[name = string("op_2681_cast_fp16")];
            tensor<int32, [1]> var_2686_axes_0 = const()[name = string("op_2686_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2686_keep_dims_0 = const()[name = string("op_2686_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 2, 1]> var_2686_cast_fp16 = reduce_mean(axes = var_2686_axes_0, keep_dims = var_2686_keep_dims_0, x = var_2681_cast_fp16)[name = string("op_2686_cast_fp16")];
            fp16 var_2687_promoted_to_fp16 = const()[name = string("op_2687_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 32, 2, 1]> var_2688_cast_fp16 = mul(x = var_2686_cast_fp16, y = var_2687_promoted_to_fp16)[name = string("op_2688_cast_fp16")];
            fp16 var_2690_to_fp16 = const()[name = string("op_2690_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 32, 2, 1]> var_2691_cast_fp16 = add(x = var_2688_cast_fp16, y = var_2690_to_fp16)[name = string("op_2691_cast_fp16")];
            fp32 var_2692_epsilon_0 = const()[name = string("op_2692_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 2, 1]> var_2692_cast_fp16 = rsqrt(epsilon = var_2692_epsilon_0, x = var_2691_cast_fp16)[name = string("op_2692_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> k_45_cast_fp16 = mul(x = k_43_cast_fp16, y = var_2692_cast_fp16)[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> var_2698 = const()[name = string("op_2698"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2711_axes_0 = const()[name = string("op_2711_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 32, 128]> var_2699_cast_fp16 = transpose(perm = var_2698, x = q_43_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [3, 2, 1, 32, 128]> var_2711_cast_fp16 = expand_dims(axes = var_2711_axes_0, x = var_2699_cast_fp16)[name = string("op_2711_cast_fp16")];
            tensor<int32, [5]> var_2719_reps_0 = const()[name = string("op_2719_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 32, 128]> var_2719_cast_fp16 = tile(reps = var_2719_reps_0, x = var_2711_cast_fp16)[name = string("op_2719_cast_fp16")];
            tensor<int32, [4]> var_2724 = const()[name = string("op_2724"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<fp16, [3, 4, 32, 128]> q_45_cast_fp16 = reshape(shape = var_2724, x = var_2719_cast_fp16)[name = string("q_45_cast_fp16")];
            tensor<int32, [4]> var_2730 = const()[name = string("op_2730"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2743_axes_0 = const()[name = string("op_2743_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 32, 128]> var_2731_cast_fp16 = transpose(perm = var_2730, x = k_45_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [3, 2, 1, 32, 128]> var_2743_cast_fp16 = expand_dims(axes = var_2743_axes_0, x = var_2731_cast_fp16)[name = string("op_2743_cast_fp16")];
            tensor<int32, [5]> var_2751_reps_0 = const()[name = string("op_2751_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 32, 128]> var_2751_cast_fp16 = tile(reps = var_2751_reps_0, x = var_2743_cast_fp16)[name = string("op_2751_cast_fp16")];
            tensor<int32, [4]> var_2756 = const()[name = string("op_2756"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<fp16, [3, 4, 32, 128]> k_47_cast_fp16 = reshape(shape = var_2756, x = var_2751_cast_fp16)[name = string("k_47_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_2758_cast_fp16 = sigmoid(x = b_7_cast_fp16)[name = string("op_2758_cast_fp16")];
            tensor<int32, [3]> var_2762 = const()[name = string("op_2762"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_4_dt_bias_to_fp16 = const()[name = string("layers_4_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 32, 4]> input_63_cast_fp16 = add(x = a_7_cast_fp16, y = layers_4_dt_bias_to_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_2766_cast_fp16 = relu(x = input_63_cast_fp16)[name = string("op_2766_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_2767_cast_fp16 = abs(x = input_63_cast_fp16)[name = string("op_2767_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 32, 4]> var_2768_cast_fp16 = mul(x = var_2767_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_2768_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_2769_cast_fp16 = exp(x = var_2768_cast_fp16)[name = string("op_2769_cast_fp16")];
            fp16 var_2771_to_fp16 = const()[name = string("op_2771_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 32, 4]> var_2772_cast_fp16 = add(x = var_2769_cast_fp16, y = var_2771_to_fp16)[name = string("op_2772_cast_fp16")];
            fp32 var_2773_epsilon_0 = const()[name = string("op_2773_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 32, 4]> var_2773_cast_fp16 = log(epsilon = var_2773_epsilon_0, x = var_2772_cast_fp16)[name = string("op_2773_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_2775_cast_fp16 = add(x = var_2766_cast_fp16, y = var_2773_cast_fp16)[name = string("op_2775_cast_fp16")];
            tensor<fp16, [4]> layers_4_negA_to_fp16 = const()[name = string("layers_4_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.348p+3, -0x1.734p+2, -0x1.5a8p-1, -0x1.264p+3])];
            tensor<fp16, [3, 32, 4]> var_2776_cast_fp16 = mul(x = layers_4_negA_to_fp16, y = var_2775_cast_fp16)[name = string("op_2776_cast_fp16")];
            tensor<int32, [3]> var_2780 = const()[name = string("op_2780"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_2783_axes_0 = const()[name = string("op_2783_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 32]> beta_7_cast_fp16 = transpose(perm = var_2762, x = var_2758_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [3, 4, 32, 1]> var_2783_cast_fp16 = expand_dims(axes = var_2783_axes_0, x = beta_7_cast_fp16)[name = string("op_2783_cast_fp16")];
            fp16 var_2784_to_fp16 = const()[name = string("op_2784_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 32, 1]> var_2785_cast_fp16 = mul(x = var_2783_cast_fp16, y = var_2784_to_fp16)[name = string("op_2785_cast_fp16")];
            tensor<fp16, [3, 4, 32, 128]> v_11_cast_fp16 = transpose(perm = var_2660, x = var_2655_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [3, 4, 32, 128]> vb_19_cast_fp16 = mul(x = v_11_cast_fp16, y = var_2785_cast_fp16)[name = string("vb_19_cast_fp16")];
            tensor<int32, [4]> q_47_begin_0 = const()[name = string("q_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_47_end_0 = const()[name = string("q_47_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_47_end_mask_0 = const()[name = string("q_47_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_47_cast_fp16 = slice_by_index(begin = q_47_begin_0, end = q_47_end_0, end_mask = q_47_end_mask_0, x = q_45_cast_fp16)[name = string("q_47_cast_fp16")];
            tensor<int32, [4]> k_49_begin_0 = const()[name = string("k_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_49_end_0 = const()[name = string("k_49_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_49_end_mask_0 = const()[name = string("k_49_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, x = k_47_cast_fp16)[name = string("k_49_cast_fp16")];
            tensor<fp16, [3, 4, 32, 128]> var_2819_cast_fp16 = mul(x = k_47_cast_fp16, y = var_2783_cast_fp16)[name = string("op_2819_cast_fp16")];
            tensor<int32, [4]> kb_13_begin_0 = const()[name = string("kb_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_13_end_0 = const()[name = string("kb_13_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_13_end_mask_0 = const()[name = string("kb_13_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_13_cast_fp16 = slice_by_index(begin = kb_13_begin_0, end = kb_13_end_0, end_mask = kb_13_end_mask_0, x = var_2819_cast_fp16)[name = string("kb_13_cast_fp16")];
            tensor<int32, [4]> vb_21_begin_0 = const()[name = string("vb_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_21_end_0 = const()[name = string("vb_21_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_21_end_mask_0 = const()[name = string("vb_21_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_21_cast_fp16 = slice_by_index(begin = vb_21_begin_0, end = vb_21_end_0, end_mask = vb_21_end_mask_0, x = vb_19_cast_fp16)[name = string("vb_21_cast_fp16")];
            tensor<int32, [3]> gdec_21_begin_0 = const()[name = string("gdec_21_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_21_end_0 = const()[name = string("gdec_21_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_21_end_mask_0 = const()[name = string("gdec_21_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 32]> gdec_19_cast_fp16 = transpose(perm = var_2780, x = var_2776_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [3, 4, 16]> gdec_21_cast_fp16 = slice_by_index(begin = gdec_21_begin_0, end = gdec_21_end_0, end_mask = gdec_21_end_mask_0, x = gdec_19_cast_fp16)[name = string("gdec_21_cast_fp16")];
            tensor<int32, [2]> var_2867 = const()[name = string("op_2867"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_2868_cast_fp16 = reshape(shape = var_2867, x = gdec_21_cast_fp16)[name = string("op_2868_cast_fp16")];
            tensor<fp16, [12, 256]> var_2869_cast_fp16 = linear(bias = var_631_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2868_cast_fp16)[name = string("op_2869_cast_fp16")];
            tensor<int32, [4]> var_2874 = const()[name = string("op_2874"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_13_cast_fp16 = reshape(shape = var_2874, x = var_2869_cast_fp16)[name = string("P_13_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2876_cast_fp16 = exp(x = P_13_cast_fp16)[name = string("op_2876_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_13_cast_fp16 = mul(x = var_2876_cast_fp16, y = tril_incl_to_fp16)[name = string("D_13_cast_fp16")];
            bool kk_13_transpose_x_1 = const()[name = string("kk_13_transpose_x_1"), val = bool(false)];
            bool kk_13_transpose_y_1 = const()[name = string("kk_13_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_13_cast_fp16 = matmul(transpose_x = kk_13_transpose_x_1, transpose_y = kk_13_transpose_y_1, x = kb_13_cast_fp16, y = k_49_cast_fp16)[name = string("kk_13_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_2882_cast_fp16 = mul(x = kk_13_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_2882_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2883_cast_fp16 = mul(x = var_2882_cast_fp16, y = D_13_cast_fp16)[name = string("op_2883_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_13_cast_fp16 = mul(x = var_2883_cast_fp16, y = cum_tail_to_fp16)[name = string("N_13_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2885_cast_fp16 = mul(x = D_13_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2885_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_13_cast_fp16 = mul(x = kk_13_cast_fp16, y = var_2885_cast_fp16)[name = string("Na_13_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_49_cast_fp16 = add(x = eye_to_fp16, y = N_13_cast_fp16)[name = string("Minv_49_cast_fp16")];
            bool Np_37_transpose_x_0 = const()[name = string("Np_37_transpose_x_0"), val = bool(false)];
            bool Np_37_transpose_y_0 = const()[name = string("Np_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_37_cast_fp16 = matmul(transpose_x = Np_37_transpose_x_0, transpose_y = Np_37_transpose_y_0, x = N_13_cast_fp16, y = Na_13_cast_fp16)[name = string("Np_37_cast_fp16")];
            bool Npa_25_transpose_x_0 = const()[name = string("Npa_25_transpose_x_0"), val = bool(false)];
            bool Npa_25_transpose_y_0 = const()[name = string("Npa_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_25_cast_fp16 = matmul(transpose_x = Npa_25_transpose_x_0, transpose_y = Npa_25_transpose_y_0, x = Na_13_cast_fp16, y = N_13_cast_fp16)[name = string("Npa_25_cast_fp16")];
            bool var_2891_transpose_x_0 = const()[name = string("op_2891_transpose_x_0"), val = bool(false)];
            bool var_2891_transpose_y_0 = const()[name = string("op_2891_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2891_cast_fp16 = matmul(transpose_x = var_2891_transpose_x_0, transpose_y = var_2891_transpose_y_0, x = Minv_49_cast_fp16, y = Np_37_cast_fp16)[name = string("op_2891_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_51_cast_fp16 = add(x = Minv_49_cast_fp16, y = var_2891_cast_fp16)[name = string("Minv_51_cast_fp16")];
            bool Np_39_transpose_x_0 = const()[name = string("Np_39_transpose_x_0"), val = bool(false)];
            bool Np_39_transpose_y_0 = const()[name = string("Np_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_39_cast_fp16 = matmul(transpose_x = Np_39_transpose_x_0, transpose_y = Np_39_transpose_y_0, x = Np_37_cast_fp16, y = Npa_25_cast_fp16)[name = string("Np_39_cast_fp16")];
            bool Npa_27_transpose_x_0 = const()[name = string("Npa_27_transpose_x_0"), val = bool(false)];
            bool Npa_27_transpose_y_0 = const()[name = string("Npa_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_27_cast_fp16 = matmul(transpose_x = Npa_27_transpose_x_0, transpose_y = Npa_27_transpose_y_0, x = Npa_25_cast_fp16, y = Np_37_cast_fp16)[name = string("Npa_27_cast_fp16")];
            bool var_2896_transpose_x_0 = const()[name = string("op_2896_transpose_x_0"), val = bool(false)];
            bool var_2896_transpose_y_0 = const()[name = string("op_2896_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2896_cast_fp16 = matmul(transpose_x = var_2896_transpose_x_0, transpose_y = var_2896_transpose_y_0, x = Minv_51_cast_fp16, y = Np_39_cast_fp16)[name = string("op_2896_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_53_cast_fp16 = add(x = Minv_51_cast_fp16, y = var_2896_cast_fp16)[name = string("Minv_53_cast_fp16")];
            bool Np_41_transpose_x_0 = const()[name = string("Np_41_transpose_x_0"), val = bool(false)];
            bool Np_41_transpose_y_0 = const()[name = string("Np_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_41_cast_fp16 = matmul(transpose_x = Np_41_transpose_x_0, transpose_y = Np_41_transpose_y_0, x = Np_39_cast_fp16, y = Npa_27_cast_fp16)[name = string("Np_41_cast_fp16")];
            bool var_2900_transpose_x_0 = const()[name = string("op_2900_transpose_x_0"), val = bool(false)];
            bool var_2900_transpose_y_0 = const()[name = string("op_2900_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2900_cast_fp16 = matmul(transpose_x = var_2900_transpose_x_0, transpose_y = var_2900_transpose_y_0, x = Minv_53_cast_fp16, y = Np_41_cast_fp16)[name = string("op_2900_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_55_cast_fp16 = add(x = Minv_53_cast_fp16, y = var_2900_cast_fp16)[name = string("Minv_55_cast_fp16")];
            bool u_13_transpose_x_0 = const()[name = string("u_13_transpose_x_0"), val = bool(false)];
            bool u_13_transpose_y_0 = const()[name = string("u_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_13_cast_fp16 = matmul(transpose_x = u_13_transpose_x_0, transpose_y = u_13_transpose_y_0, x = Minv_55_cast_fp16, y = vb_21_cast_fp16)[name = string("u_13_cast_fp16")];
            bool var_2907_transpose_x_1 = const()[name = string("op_2907_transpose_x_1"), val = bool(false)];
            bool var_2907_transpose_y_1 = const()[name = string("op_2907_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_2907_cast_fp16 = matmul(transpose_x = var_2907_transpose_x_1, transpose_y = var_2907_transpose_y_1, x = q_47_cast_fp16, y = k_49_cast_fp16)[name = string("op_2907_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_13_cast_fp16 = mul(x = var_2907_cast_fp16, y = D_13_cast_fp16)[name = string("A_13_cast_fp16")];
            bool var_2909_transpose_x_0 = const()[name = string("op_2909_transpose_x_0"), val = bool(false)];
            bool var_2909_transpose_y_0 = const()[name = string("op_2909_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2909_cast_fp16 = matmul(transpose_x = var_2909_transpose_x_0, transpose_y = var_2909_transpose_y_0, x = A_13_cast_fp16, y = u_13_cast_fp16)[name = string("op_2909_cast_fp16")];
            tensor<fp16, [16]> var_2910_bias_0_to_fp16 = const()[name = string("op_2910_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_2910_cast_fp16 = linear(bias = var_2910_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_21_cast_fp16)[name = string("op_2910_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_7_cast_fp16 = exp(x = var_2910_cast_fp16)[name = string("tail_7_cast_fp16")];
            tensor<int32, [1]> var_2913_axes_0 = const()[name = string("op_2913_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_2913_cast_fp16 = expand_dims(axes = var_2913_axes_0, x = tail_7_cast_fp16)[name = string("op_2913_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2914_cast_fp16 = mul(x = k_49_cast_fp16, y = var_2913_cast_fp16)[name = string("op_2914_cast_fp16")];
            bool S_7_transpose_x_1 = const()[name = string("S_7_transpose_x_1"), val = bool(true)];
            bool S_7_transpose_y_1 = const()[name = string("S_7_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_7_cast_fp16 = matmul(transpose_x = S_7_transpose_x_1, transpose_y = S_7_transpose_y_1, x = var_2914_cast_fp16, y = u_13_cast_fp16)[name = string("S_7_cast_fp16")];
            tensor<int32, [4]> q_49_begin_0 = const()[name = string("q_49_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_49_end_0 = const()[name = string("q_49_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_49_end_mask_0 = const()[name = string("q_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, x = q_45_cast_fp16)[name = string("q_49_cast_fp16")];
            tensor<int32, [4]> k_51_begin_0 = const()[name = string("k_51_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_51_end_0 = const()[name = string("k_51_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_51_end_mask_0 = const()[name = string("k_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_51_cast_fp16 = slice_by_index(begin = k_51_begin_0, end = k_51_end_0, end_mask = k_51_end_mask_0, x = k_47_cast_fp16)[name = string("k_51_cast_fp16")];
            tensor<int32, [4]> kb_15_begin_0 = const()[name = string("kb_15_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_15_end_0 = const()[name = string("kb_15_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_15_end_mask_0 = const()[name = string("kb_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_15_cast_fp16 = slice_by_index(begin = kb_15_begin_0, end = kb_15_end_0, end_mask = kb_15_end_mask_0, x = var_2819_cast_fp16)[name = string("kb_15_cast_fp16")];
            tensor<int32, [4]> vb_23_begin_0 = const()[name = string("vb_23_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_23_end_0 = const()[name = string("vb_23_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_23_end_mask_0 = const()[name = string("vb_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_23_cast_fp16 = slice_by_index(begin = vb_23_begin_0, end = vb_23_end_0, end_mask = vb_23_end_mask_0, x = vb_19_cast_fp16)[name = string("vb_23_cast_fp16")];
            tensor<int32, [3]> gdec_23_begin_0 = const()[name = string("gdec_23_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_23_end_0 = const()[name = string("gdec_23_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_23_end_mask_0 = const()[name = string("gdec_23_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_23_cast_fp16 = slice_by_index(begin = gdec_23_begin_0, end = gdec_23_end_0, end_mask = gdec_23_end_mask_0, x = gdec_19_cast_fp16)[name = string("gdec_23_cast_fp16")];
            tensor<int32, [2]> var_2999 = const()[name = string("op_2999"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_3000_cast_fp16 = reshape(shape = var_2999, x = gdec_23_cast_fp16)[name = string("op_3000_cast_fp16")];
            tensor<fp16, [12, 256]> var_3001_cast_fp16 = linear(bias = var_631_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3000_cast_fp16)[name = string("op_3001_cast_fp16")];
            tensor<int32, [4]> var_3006 = const()[name = string("op_3006"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_15_cast_fp16 = reshape(shape = var_3006, x = var_3001_cast_fp16)[name = string("P_15_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3008_cast_fp16 = exp(x = P_15_cast_fp16)[name = string("op_3008_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_15_cast_fp16 = mul(x = var_3008_cast_fp16, y = tril_incl_to_fp16)[name = string("D_15_cast_fp16")];
            bool kk_15_transpose_x_1 = const()[name = string("kk_15_transpose_x_1"), val = bool(false)];
            bool kk_15_transpose_y_1 = const()[name = string("kk_15_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_15_cast_fp16 = matmul(transpose_x = kk_15_transpose_x_1, transpose_y = kk_15_transpose_y_1, x = kb_15_cast_fp16, y = k_51_cast_fp16)[name = string("kk_15_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_3014_cast_fp16 = mul(x = kk_15_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_3014_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3015_cast_fp16 = mul(x = var_3014_cast_fp16, y = D_15_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_15_cast_fp16 = mul(x = var_3015_cast_fp16, y = cum_tail_to_fp16)[name = string("N_15_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3017_cast_fp16 = mul(x = D_15_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3017_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_15_cast_fp16 = mul(x = kk_15_cast_fp16, y = var_3017_cast_fp16)[name = string("Na_15_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_57_cast_fp16 = add(x = eye_to_fp16, y = N_15_cast_fp16)[name = string("Minv_57_cast_fp16")];
            bool Np_43_transpose_x_0 = const()[name = string("Np_43_transpose_x_0"), val = bool(false)];
            bool Np_43_transpose_y_0 = const()[name = string("Np_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_43_cast_fp16 = matmul(transpose_x = Np_43_transpose_x_0, transpose_y = Np_43_transpose_y_0, x = N_15_cast_fp16, y = Na_15_cast_fp16)[name = string("Np_43_cast_fp16")];
            bool Npa_29_transpose_x_0 = const()[name = string("Npa_29_transpose_x_0"), val = bool(false)];
            bool Npa_29_transpose_y_0 = const()[name = string("Npa_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_29_cast_fp16 = matmul(transpose_x = Npa_29_transpose_x_0, transpose_y = Npa_29_transpose_y_0, x = Na_15_cast_fp16, y = N_15_cast_fp16)[name = string("Npa_29_cast_fp16")];
            bool var_3023_transpose_x_0 = const()[name = string("op_3023_transpose_x_0"), val = bool(false)];
            bool var_3023_transpose_y_0 = const()[name = string("op_3023_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3023_cast_fp16 = matmul(transpose_x = var_3023_transpose_x_0, transpose_y = var_3023_transpose_y_0, x = Minv_57_cast_fp16, y = Np_43_cast_fp16)[name = string("op_3023_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_59_cast_fp16 = add(x = Minv_57_cast_fp16, y = var_3023_cast_fp16)[name = string("Minv_59_cast_fp16")];
            bool Np_45_transpose_x_0 = const()[name = string("Np_45_transpose_x_0"), val = bool(false)];
            bool Np_45_transpose_y_0 = const()[name = string("Np_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_45_cast_fp16 = matmul(transpose_x = Np_45_transpose_x_0, transpose_y = Np_45_transpose_y_0, x = Np_43_cast_fp16, y = Npa_29_cast_fp16)[name = string("Np_45_cast_fp16")];
            bool Npa_31_transpose_x_0 = const()[name = string("Npa_31_transpose_x_0"), val = bool(false)];
            bool Npa_31_transpose_y_0 = const()[name = string("Npa_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_31_cast_fp16 = matmul(transpose_x = Npa_31_transpose_x_0, transpose_y = Npa_31_transpose_y_0, x = Npa_29_cast_fp16, y = Np_43_cast_fp16)[name = string("Npa_31_cast_fp16")];
            bool var_3028_transpose_x_0 = const()[name = string("op_3028_transpose_x_0"), val = bool(false)];
            bool var_3028_transpose_y_0 = const()[name = string("op_3028_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3028_cast_fp16 = matmul(transpose_x = var_3028_transpose_x_0, transpose_y = var_3028_transpose_y_0, x = Minv_59_cast_fp16, y = Np_45_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_61_cast_fp16 = add(x = Minv_59_cast_fp16, y = var_3028_cast_fp16)[name = string("Minv_61_cast_fp16")];
            bool Np_47_transpose_x_0 = const()[name = string("Np_47_transpose_x_0"), val = bool(false)];
            bool Np_47_transpose_y_0 = const()[name = string("Np_47_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_47_cast_fp16 = matmul(transpose_x = Np_47_transpose_x_0, transpose_y = Np_47_transpose_y_0, x = Np_45_cast_fp16, y = Npa_31_cast_fp16)[name = string("Np_47_cast_fp16")];
            bool var_3032_transpose_x_0 = const()[name = string("op_3032_transpose_x_0"), val = bool(false)];
            bool var_3032_transpose_y_0 = const()[name = string("op_3032_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3032_cast_fp16 = matmul(transpose_x = var_3032_transpose_x_0, transpose_y = var_3032_transpose_y_0, x = Minv_61_cast_fp16, y = Np_47_cast_fp16)[name = string("op_3032_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_63_cast_fp16 = add(x = Minv_61_cast_fp16, y = var_3032_cast_fp16)[name = string("Minv_63_cast_fp16")];
            bool u_15_transpose_x_0 = const()[name = string("u_15_transpose_x_0"), val = bool(false)];
            bool u_15_transpose_y_0 = const()[name = string("u_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_15_cast_fp16 = matmul(transpose_x = u_15_transpose_x_0, transpose_y = u_15_transpose_y_0, x = Minv_63_cast_fp16, y = vb_23_cast_fp16)[name = string("u_15_cast_fp16")];
            bool var_3039_transpose_x_1 = const()[name = string("op_3039_transpose_x_1"), val = bool(false)];
            bool var_3039_transpose_y_1 = const()[name = string("op_3039_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_3039_cast_fp16 = matmul(transpose_x = var_3039_transpose_x_1, transpose_y = var_3039_transpose_y_1, x = q_49_cast_fp16, y = k_51_cast_fp16)[name = string("op_3039_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_15_cast_fp16 = mul(x = var_3039_cast_fp16, y = D_15_cast_fp16)[name = string("A_15_cast_fp16")];
            tensor<fp16, [16]> var_3041_bias_0_to_fp16 = const()[name = string("op_3041_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_3041_cast_fp16 = linear(bias = var_3041_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_23_cast_fp16)[name = string("op_3041_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_7_cast_fp16 = exp(x = var_3041_cast_fp16)[name = string("cum_7_cast_fp16")];
            tensor<int32, [1]> var_3044_axes_0 = const()[name = string("op_3044_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_3044_cast_fp16 = expand_dims(axes = var_3044_axes_0, x = cum_7_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3045_cast_fp16 = mul(x = kb_15_cast_fp16, y = var_3044_cast_fp16)[name = string("op_3045_cast_fp16")];
            bool w_15_transpose_x_0 = const()[name = string("w_15_transpose_x_0"), val = bool(false)];
            bool w_15_transpose_y_0 = const()[name = string("w_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_15_cast_fp16 = matmul(transpose_x = w_15_transpose_x_0, transpose_y = w_15_transpose_y_0, x = Minv_63_cast_fp16, y = var_3045_cast_fp16)[name = string("w_15_cast_fp16")];
            bool var_3047_transpose_x_0 = const()[name = string("op_3047_transpose_x_0"), val = bool(false)];
            bool var_3047_transpose_y_0 = const()[name = string("op_3047_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3047_cast_fp16 = matmul(transpose_x = var_3047_transpose_x_0, transpose_y = var_3047_transpose_y_0, x = w_15_cast_fp16, y = S_7_cast_fp16)[name = string("op_3047_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_7_cast_fp16 = sub(x = u_15_cast_fp16, y = var_3047_cast_fp16)[name = string("vnew_7_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3052_cast_fp16 = mul(x = q_49_cast_fp16, y = var_3044_cast_fp16)[name = string("op_3052_cast_fp16")];
            bool var_3053_transpose_x_0 = const()[name = string("op_3053_transpose_x_0"), val = bool(false)];
            bool var_3053_transpose_y_0 = const()[name = string("op_3053_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3053_cast_fp16 = matmul(transpose_x = var_3053_transpose_x_0, transpose_y = var_3053_transpose_y_0, x = var_3052_cast_fp16, y = S_7_cast_fp16)[name = string("op_3053_cast_fp16")];
            bool var_3054_transpose_x_0 = const()[name = string("op_3054_transpose_x_0"), val = bool(false)];
            bool var_3054_transpose_y_0 = const()[name = string("op_3054_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3054_cast_fp16 = matmul(transpose_x = var_3054_transpose_x_0, transpose_y = var_3054_transpose_y_0, x = A_15_cast_fp16, y = vnew_7_cast_fp16)[name = string("op_3054_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_7_cast_fp16 = add(x = var_3053_cast_fp16, y = var_3054_cast_fp16)[name = string("oc_7_cast_fp16")];
            int32 var_3058 = const()[name = string("op_3058"), val = int32(2)];
            bool o_21_interleave_0 = const()[name = string("o_21_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 32, 128]> o_21_cast_fp16 = concat(axis = var_3058, interleave = o_21_interleave_0, values = (var_2909_cast_fp16, oc_7_cast_fp16))[name = string("o_21_cast_fp16")];
            tensor<int32, [4]> var_3064 = const()[name = string("op_3064"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3070 = const()[name = string("op_3070"), val = tensor<int32, [4]>([3, 32, 4, 128])];
            tensor<fp16, [3, 32, 4, 128]> input_65_cast_fp16 = reshape(shape = var_3070, x = z_7_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> o_23_cast_fp16 = transpose(perm = var_3064, x = o_21_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [3, 32, 4, 128]> var_3072_cast_fp16 = mul(x = o_23_cast_fp16, y = o_23_cast_fp16)[name = string("op_3072_cast_fp16")];
            tensor<int32, [1]> var_3077_axes_0 = const()[name = string("op_3077_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3077_keep_dims_0 = const()[name = string("op_3077_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 4, 1]> var_3077_cast_fp16 = reduce_mean(axes = var_3077_axes_0, keep_dims = var_3077_keep_dims_0, x = var_3072_cast_fp16)[name = string("op_3077_cast_fp16")];
            fp16 var_3079_to_fp16 = const()[name = string("op_3079_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 32, 4, 1]> var_3080_cast_fp16 = add(x = var_3077_cast_fp16, y = var_3079_to_fp16)[name = string("op_3080_cast_fp16")];
            fp32 var_3081_epsilon_0 = const()[name = string("op_3081_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 4, 1]> var_3081_cast_fp16 = rsqrt(epsilon = var_3081_epsilon_0, x = var_3080_cast_fp16)[name = string("op_3081_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> var_3082_cast_fp16 = mul(x = o_23_cast_fp16, y = var_3081_cast_fp16)[name = string("op_3082_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> var_3084_cast_fp16 = silu(x = input_65_cast_fp16)[name = string("op_3084_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> o_25_cast_fp16 = mul(x = var_3082_cast_fp16, y = var_3084_cast_fp16)[name = string("o_25_cast_fp16")];
            tensor<int32, [3]> var_3089 = const()[name = string("op_3089"), val = tensor<int32, [3]>([3, 32, 512])];
            tensor<fp16, [3, 32, 512]> input_67_cast_fp16 = reshape(shape = var_3089, x = o_25_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [640, 512]> layers_4_out_weight_to_fp16 = const()[name = string("layers_4_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72063040)))];
            tensor<fp16, [3, 32, 640]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_out_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [3, 32, 640]> x_49_cast_fp16 = add(x = x_45_cast_fp16, y = linear_17_cast_fp16)[name = string("x_49_cast_fp16")];
            fp16 var_3102_to_fp16 = const()[name = string("op_3102_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 32, 640]> x_51_cast_fp16 = mul(x = x_49_cast_fp16, y = var_3102_to_fp16)[name = string("x_51_cast_fp16")];
            fp16 var_3100_promoted_to_fp16 = const()[name = string("op_3100_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> var_3104_cast_fp16 = pow(x = x_51_cast_fp16, y = var_3100_promoted_to_fp16)[name = string("op_3104_cast_fp16")];
            tensor<int32, [1]> var_3106_axes_0 = const()[name = string("op_3106_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3106_keep_dims_0 = const()[name = string("op_3106_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 1]> var_3106_cast_fp16 = reduce_mean(axes = var_3106_axes_0, keep_dims = var_3106_keep_dims_0, x = var_3104_cast_fp16)[name = string("op_3106_cast_fp16")];
            fp16 var_3107_to_fp16 = const()[name = string("op_3107_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 32, 1]> var_3108_cast_fp16 = add(x = var_3106_cast_fp16, y = var_3107_to_fp16)[name = string("op_3108_cast_fp16")];
            fp32 var_3109_epsilon_0 = const()[name = string("op_3109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 1]> var_3109_cast_fp16 = rsqrt(epsilon = var_3109_epsilon_0, x = var_3108_cast_fp16)[name = string("op_3109_cast_fp16")];
            tensor<fp16, [3, 32, 640]> var_3110_cast_fp16 = mul(x = x_51_cast_fp16, y = var_3109_cast_fp16)[name = string("op_3110_cast_fp16")];
            tensor<fp16, [640]> layers_4_n2_w1_to_fp16 = const()[name = string("layers_4_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72718464)))];
            tensor<fp16, [3, 32, 640]> input_69_cast_fp16 = mul(x = var_3110_cast_fp16, y = layers_4_n2_w1_to_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_4_gate_up_weight_to_fp16 = const()[name = string("layers_4_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72719808)))];
            tensor<fp16, [3, 32, 5120]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_4_gate_up_weight_to_fp16, x = input_69_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<int32, [3]> input_71_begin_0 = const()[name = string("input_71_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_71_end_0 = const()[name = string("input_71_end_0"), val = tensor<int32, [3]>([3, 32, 2560])];
            tensor<bool, [3]> input_71_end_mask_0 = const()[name = string("input_71_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 2560]> input_71_cast_fp16 = slice_by_index(begin = input_71_begin_0, end = input_71_end_0, end_mask = input_71_end_mask_0, x = linear_18_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [3, 32, 2560]> var_3120_cast_fp16 = silu(x = input_71_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<int32, [3]> var_3125_begin_0 = const()[name = string("op_3125_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_3125_end_0 = const()[name = string("op_3125_end_0"), val = tensor<int32, [3]>([3, 32, 5120])];
            tensor<bool, [3]> var_3125_end_mask_0 = const()[name = string("op_3125_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 2560]> var_3125_cast_fp16 = slice_by_index(begin = var_3125_begin_0, end = var_3125_end_0, end_mask = var_3125_end_mask_0, x = linear_18_cast_fp16)[name = string("op_3125_cast_fp16")];
            tensor<fp16, [3, 32, 2560]> input_73_cast_fp16 = mul(x = var_3120_cast_fp16, y = var_3125_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_4_down_weight_to_fp16 = const()[name = string("layers_4_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79273472)))];
            tensor<fp16, [3, 32, 640]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_down_weight_to_fp16, x = input_73_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [3, 32, 640]> x_53_cast_fp16 = add(x = x_49_cast_fp16, y = linear_19_cast_fp16)[name = string("x_53_cast_fp16")];
            fp16 var_3138_to_fp16 = const()[name = string("op_3138_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 32, 640]> x_55_cast_fp16 = mul(x = x_53_cast_fp16, y = var_3138_to_fp16)[name = string("x_55_cast_fp16")];
            fp16 var_3136_promoted_to_fp16 = const()[name = string("op_3136_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> var_3140_cast_fp16 = pow(x = x_55_cast_fp16, y = var_3136_promoted_to_fp16)[name = string("op_3140_cast_fp16")];
            tensor<int32, [1]> var_3142_axes_0 = const()[name = string("op_3142_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3142_keep_dims_0 = const()[name = string("op_3142_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 1]> var_3142_cast_fp16 = reduce_mean(axes = var_3142_axes_0, keep_dims = var_3142_keep_dims_0, x = var_3140_cast_fp16)[name = string("op_3142_cast_fp16")];
            fp16 var_3143_to_fp16 = const()[name = string("op_3143_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 32, 1]> var_3144_cast_fp16 = add(x = var_3142_cast_fp16, y = var_3143_to_fp16)[name = string("op_3144_cast_fp16")];
            fp32 var_3145_epsilon_0 = const()[name = string("op_3145_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 1]> var_3145_cast_fp16 = rsqrt(epsilon = var_3145_epsilon_0, x = var_3144_cast_fp16)[name = string("op_3145_cast_fp16")];
            tensor<fp16, [3, 32, 640]> var_3146_cast_fp16 = mul(x = x_55_cast_fp16, y = var_3145_cast_fp16)[name = string("op_3146_cast_fp16")];
            tensor<fp16, [640]> layers_5_n1_w1_to_fp16 = const()[name = string("layers_5_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82550336)))];
            tensor<fp16, [3, 32, 640]> input_75_cast_fp16 = mul(x = var_3146_cast_fp16, y = layers_5_n1_w1_to_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_5_qkvzba_weight_to_fp16 = const()[name = string("layers_5_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82551680)))];
            tensor<fp16, [3, 32, 1544]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_qkvzba_weight_to_fp16, x = input_75_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [3]> qkv_9_begin_0 = const()[name = string("qkv_9_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_9_end_0 = const()[name = string("qkv_9_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> qkv_9_end_mask_0 = const()[name = string("qkv_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 1024]> qkv_9_cast_fp16 = slice_by_index(begin = qkv_9_begin_0, end = qkv_9_end_0, end_mask = qkv_9_end_mask_0, x = linear_20_cast_fp16)[name = string("qkv_9_cast_fp16")];
            tensor<int32, [3]> z_9_begin_0 = const()[name = string("z_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_9_end_0 = const()[name = string("z_9_end_0"), val = tensor<int32, [3]>([3, 32, 1536])];
            tensor<bool, [3]> z_9_end_mask_0 = const()[name = string("z_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 512]> z_9_cast_fp16 = slice_by_index(begin = z_9_begin_0, end = z_9_end_0, end_mask = z_9_end_mask_0, x = linear_20_cast_fp16)[name = string("z_9_cast_fp16")];
            tensor<int32, [3]> b_9_begin_0 = const()[name = string("b_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_9_end_0 = const()[name = string("b_9_end_0"), val = tensor<int32, [3]>([3, 32, 1540])];
            tensor<bool, [3]> b_9_end_mask_0 = const()[name = string("b_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 4]> b_9_cast_fp16 = slice_by_index(begin = b_9_begin_0, end = b_9_end_0, end_mask = b_9_end_mask_0, x = linear_20_cast_fp16)[name = string("b_9_cast_fp16")];
            tensor<int32, [3]> a_9_begin_0 = const()[name = string("a_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_9_end_0 = const()[name = string("a_9_end_0"), val = tensor<int32, [3]>([3, 32, 1])];
            tensor<bool, [3]> a_9_end_mask_0 = const()[name = string("a_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 4]> a_9_cast_fp16 = slice_by_index(begin = a_9_begin_0, end = a_9_end_0, end_mask = a_9_end_mask_0, x = linear_20_cast_fp16)[name = string("a_9_cast_fp16")];
            int32 var_3185 = const()[name = string("op_3185"), val = int32(1)];
            bool xp_9_interleave_0 = const()[name = string("xp_9_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 35, 1024]> xp_9_cast_fp16 = concat(axis = var_3185, interleave = xp_9_interleave_0, values = (zpad_to_fp16, qkv_9_cast_fp16))[name = string("xp_9_cast_fp16")];
            tensor<int32, [3]> var_3196_begin_0 = const()[name = string("op_3196_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_3196_end_0 = const()[name = string("op_3196_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_3196_end_mask_0 = const()[name = string("op_3196_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 1024]> var_3196_cast_fp16 = slice_by_index(begin = var_3196_begin_0, end = var_3196_end_0, end_mask = var_3196_end_mask_0, x = xp_9_cast_fp16)[name = string("op_3196_cast_fp16")];
            tensor<fp16, [1024]> var_3204_to_fp16 = const()[name = string("op_3204_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84528064)))];
            tensor<fp16, [3, 32, 1024]> var_3205_cast_fp16 = mul(x = var_3196_cast_fp16, y = var_3204_to_fp16)[name = string("op_3205_cast_fp16")];
            tensor<int32, [3]> var_3215_begin_0 = const()[name = string("op_3215_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_3215_end_0 = const()[name = string("op_3215_end_0"), val = tensor<int32, [3]>([3, 34, 1024])];
            tensor<bool, [3]> var_3215_end_mask_0 = const()[name = string("op_3215_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_3215_cast_fp16 = slice_by_index(begin = var_3215_begin_0, end = var_3215_end_0, end_mask = var_3215_end_mask_0, x = xp_9_cast_fp16)[name = string("op_3215_cast_fp16")];
            tensor<fp16, [1024]> var_3223_to_fp16 = const()[name = string("op_3223_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84530176)))];
            tensor<fp16, [3, 32, 1024]> var_3224_cast_fp16 = mul(x = var_3215_cast_fp16, y = var_3223_to_fp16)[name = string("op_3224_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> var_3226_cast_fp16 = add(x = var_3205_cast_fp16, y = var_3224_cast_fp16)[name = string("op_3226_cast_fp16")];
            tensor<int32, [3]> var_3236_begin_0 = const()[name = string("op_3236_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_3236_end_0 = const()[name = string("op_3236_end_0"), val = tensor<int32, [3]>([3, 33, 1024])];
            tensor<bool, [3]> var_3236_end_mask_0 = const()[name = string("op_3236_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_3236_cast_fp16 = slice_by_index(begin = var_3236_begin_0, end = var_3236_end_0, end_mask = var_3236_end_mask_0, x = xp_9_cast_fp16)[name = string("op_3236_cast_fp16")];
            tensor<fp16, [1024]> var_3244_to_fp16 = const()[name = string("op_3244_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84532288)))];
            tensor<fp16, [3, 32, 1024]> var_3245_cast_fp16 = mul(x = var_3236_cast_fp16, y = var_3244_to_fp16)[name = string("op_3245_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> var_3247_cast_fp16 = add(x = var_3226_cast_fp16, y = var_3245_cast_fp16)[name = string("op_3247_cast_fp16")];
            tensor<int32, [3]> var_3257_begin_0 = const()[name = string("op_3257_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3257_end_0 = const()[name = string("op_3257_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> var_3257_end_mask_0 = const()[name = string("op_3257_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_3257_cast_fp16 = slice_by_index(begin = var_3257_begin_0, end = var_3257_end_0, end_mask = var_3257_end_mask_0, x = xp_9_cast_fp16)[name = string("op_3257_cast_fp16")];
            tensor<fp16, [1024]> var_3265_to_fp16 = const()[name = string("op_3265_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84534400)))];
            tensor<fp16, [3, 32, 1024]> var_3266_cast_fp16 = mul(x = var_3257_cast_fp16, y = var_3265_to_fp16)[name = string("op_3266_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> input_77_cast_fp16 = add(x = var_3247_cast_fp16, y = var_3266_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> c_9_cast_fp16 = silu(x = input_77_cast_fp16)[name = string("c_9_cast_fp16")];
            tensor<int32, [3]> var_3274_begin_0 = const()[name = string("op_3274_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3274_end_0 = const()[name = string("op_3274_end_0"), val = tensor<int32, [3]>([3, 32, 256])];
            tensor<bool, [3]> var_3274_end_mask_0 = const()[name = string("op_3274_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 256]> var_3274_cast_fp16 = slice_by_index(begin = var_3274_begin_0, end = var_3274_end_0, end_mask = var_3274_end_mask_0, x = c_9_cast_fp16)[name = string("op_3274_cast_fp16")];
            tensor<int32, [4]> var_3279 = const()[name = string("op_3279"), val = tensor<int32, [4]>([3, 32, 2, 128])];
            tensor<fp16, [3, 32, 2, 128]> q_51_cast_fp16 = reshape(shape = var_3279, x = var_3274_cast_fp16)[name = string("q_51_cast_fp16")];
            tensor<int32, [3]> var_3285_begin_0 = const()[name = string("op_3285_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_3285_end_0 = const()[name = string("op_3285_end_0"), val = tensor<int32, [3]>([3, 32, 512])];
            tensor<bool, [3]> var_3285_end_mask_0 = const()[name = string("op_3285_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 256]> var_3285_cast_fp16 = slice_by_index(begin = var_3285_begin_0, end = var_3285_end_0, end_mask = var_3285_end_mask_0, x = c_9_cast_fp16)[name = string("op_3285_cast_fp16")];
            tensor<int32, [4]> var_3290 = const()[name = string("op_3290"), val = tensor<int32, [4]>([3, 32, 2, 128])];
            tensor<fp16, [3, 32, 2, 128]> k_53_cast_fp16 = reshape(shape = var_3290, x = var_3285_cast_fp16)[name = string("k_53_cast_fp16")];
            tensor<int32, [3]> var_3296_begin_0 = const()[name = string("op_3296_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_3296_end_0 = const()[name = string("op_3296_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> var_3296_end_mask_0 = const()[name = string("op_3296_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 512]> var_3296_cast_fp16 = slice_by_index(begin = var_3296_begin_0, end = var_3296_end_0, end_mask = var_3296_end_mask_0, x = c_9_cast_fp16)[name = string("op_3296_cast_fp16")];
            tensor<int32, [4]> var_3301 = const()[name = string("op_3301"), val = tensor<int32, [4]>([3, 32, 4, 128])];
            tensor<fp16, [3, 32, 4, 128]> var_3302_cast_fp16 = reshape(shape = var_3301, x = var_3296_cast_fp16)[name = string("op_3302_cast_fp16")];
            tensor<int32, [4]> var_3307 = const()[name = string("op_3307"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3309_to_fp16 = const()[name = string("op_3309_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 32, 2, 128]> q_53_cast_fp16 = mul(x = q_51_cast_fp16, y = var_3309_to_fp16)[name = string("q_53_cast_fp16")];
            fp16 var_3311_to_fp16 = const()[name = string("op_3311_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 32, 2, 128]> k_55_cast_fp16 = mul(x = k_53_cast_fp16, y = var_3311_to_fp16)[name = string("k_55_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_3313_cast_fp16 = mul(x = q_53_cast_fp16, y = q_53_cast_fp16)[name = string("op_3313_cast_fp16")];
            tensor<int32, [1]> var_3318_axes_0 = const()[name = string("op_3318_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3318_keep_dims_0 = const()[name = string("op_3318_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 2, 1]> var_3318_cast_fp16 = reduce_mean(axes = var_3318_axes_0, keep_dims = var_3318_keep_dims_0, x = var_3313_cast_fp16)[name = string("op_3318_cast_fp16")];
            fp16 var_3319_promoted_to_fp16 = const()[name = string("op_3319_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 32, 2, 1]> var_3320_cast_fp16 = mul(x = var_3318_cast_fp16, y = var_3319_promoted_to_fp16)[name = string("op_3320_cast_fp16")];
            fp16 var_3322_to_fp16 = const()[name = string("op_3322_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 32, 2, 1]> var_3323_cast_fp16 = add(x = var_3320_cast_fp16, y = var_3322_to_fp16)[name = string("op_3323_cast_fp16")];
            fp32 var_3324_epsilon_0 = const()[name = string("op_3324_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 2, 1]> var_3324_cast_fp16 = rsqrt(epsilon = var_3324_epsilon_0, x = var_3323_cast_fp16)[name = string("op_3324_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_3325_cast_fp16 = mul(x = q_53_cast_fp16, y = var_3324_cast_fp16)[name = string("op_3325_cast_fp16")];
            fp16 var_3326_to_fp16 = const()[name = string("op_3326_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 32, 2, 128]> q_55_cast_fp16 = mul(x = var_3325_cast_fp16, y = var_3326_to_fp16)[name = string("q_55_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_3328_cast_fp16 = mul(x = k_55_cast_fp16, y = k_55_cast_fp16)[name = string("op_3328_cast_fp16")];
            tensor<int32, [1]> var_3333_axes_0 = const()[name = string("op_3333_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3333_keep_dims_0 = const()[name = string("op_3333_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 2, 1]> var_3333_cast_fp16 = reduce_mean(axes = var_3333_axes_0, keep_dims = var_3333_keep_dims_0, x = var_3328_cast_fp16)[name = string("op_3333_cast_fp16")];
            fp16 var_3334_promoted_to_fp16 = const()[name = string("op_3334_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 32, 2, 1]> var_3335_cast_fp16 = mul(x = var_3333_cast_fp16, y = var_3334_promoted_to_fp16)[name = string("op_3335_cast_fp16")];
            fp16 var_3337_to_fp16 = const()[name = string("op_3337_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 32, 2, 1]> var_3338_cast_fp16 = add(x = var_3335_cast_fp16, y = var_3337_to_fp16)[name = string("op_3338_cast_fp16")];
            fp32 var_3339_epsilon_0 = const()[name = string("op_3339_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 2, 1]> var_3339_cast_fp16 = rsqrt(epsilon = var_3339_epsilon_0, x = var_3338_cast_fp16)[name = string("op_3339_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> k_57_cast_fp16 = mul(x = k_55_cast_fp16, y = var_3339_cast_fp16)[name = string("k_57_cast_fp16")];
            tensor<int32, [4]> var_3345 = const()[name = string("op_3345"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_3358_axes_0 = const()[name = string("op_3358_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 32, 128]> var_3346_cast_fp16 = transpose(perm = var_3345, x = q_55_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [3, 2, 1, 32, 128]> var_3358_cast_fp16 = expand_dims(axes = var_3358_axes_0, x = var_3346_cast_fp16)[name = string("op_3358_cast_fp16")];
            tensor<int32, [5]> var_3366_reps_0 = const()[name = string("op_3366_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 32, 128]> var_3366_cast_fp16 = tile(reps = var_3366_reps_0, x = var_3358_cast_fp16)[name = string("op_3366_cast_fp16")];
            tensor<int32, [4]> var_3371 = const()[name = string("op_3371"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<fp16, [3, 4, 32, 128]> q_57_cast_fp16 = reshape(shape = var_3371, x = var_3366_cast_fp16)[name = string("q_57_cast_fp16")];
            tensor<int32, [4]> var_3377 = const()[name = string("op_3377"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_3390_axes_0 = const()[name = string("op_3390_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 32, 128]> var_3378_cast_fp16 = transpose(perm = var_3377, x = k_57_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [3, 2, 1, 32, 128]> var_3390_cast_fp16 = expand_dims(axes = var_3390_axes_0, x = var_3378_cast_fp16)[name = string("op_3390_cast_fp16")];
            tensor<int32, [5]> var_3398_reps_0 = const()[name = string("op_3398_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 32, 128]> var_3398_cast_fp16 = tile(reps = var_3398_reps_0, x = var_3390_cast_fp16)[name = string("op_3398_cast_fp16")];
            tensor<int32, [4]> var_3403 = const()[name = string("op_3403"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<fp16, [3, 4, 32, 128]> k_59_cast_fp16 = reshape(shape = var_3403, x = var_3398_cast_fp16)[name = string("k_59_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_3405_cast_fp16 = sigmoid(x = b_9_cast_fp16)[name = string("op_3405_cast_fp16")];
            tensor<int32, [3]> var_3409 = const()[name = string("op_3409"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_5_dt_bias_to_fp16 = const()[name = string("layers_5_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 32, 4]> input_79_cast_fp16 = add(x = a_9_cast_fp16, y = layers_5_dt_bias_to_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_3413_cast_fp16 = relu(x = input_79_cast_fp16)[name = string("op_3413_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_3414_cast_fp16 = abs(x = input_79_cast_fp16)[name = string("op_3414_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 32, 4]> var_3415_cast_fp16 = mul(x = var_3414_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_3415_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_3416_cast_fp16 = exp(x = var_3415_cast_fp16)[name = string("op_3416_cast_fp16")];
            fp16 var_3418_to_fp16 = const()[name = string("op_3418_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 32, 4]> var_3419_cast_fp16 = add(x = var_3416_cast_fp16, y = var_3418_to_fp16)[name = string("op_3419_cast_fp16")];
            fp32 var_3420_epsilon_0 = const()[name = string("op_3420_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 32, 4]> var_3420_cast_fp16 = log(epsilon = var_3420_epsilon_0, x = var_3419_cast_fp16)[name = string("op_3420_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_3422_cast_fp16 = add(x = var_3413_cast_fp16, y = var_3420_cast_fp16)[name = string("op_3422_cast_fp16")];
            tensor<fp16, [4]> layers_5_negA_to_fp16 = const()[name = string("layers_5_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.2f8p+3, -0x1.ae8p+2, -0x1.548p+2, -0x1.63p+3])];
            tensor<fp16, [3, 32, 4]> var_3423_cast_fp16 = mul(x = layers_5_negA_to_fp16, y = var_3422_cast_fp16)[name = string("op_3423_cast_fp16")];
            tensor<int32, [3]> var_3427 = const()[name = string("op_3427"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_3430_axes_0 = const()[name = string("op_3430_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 32]> beta_9_cast_fp16 = transpose(perm = var_3409, x = var_3405_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [3, 4, 32, 1]> var_3430_cast_fp16 = expand_dims(axes = var_3430_axes_0, x = beta_9_cast_fp16)[name = string("op_3430_cast_fp16")];
            fp16 var_3431_to_fp16 = const()[name = string("op_3431_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 32, 1]> var_3432_cast_fp16 = mul(x = var_3430_cast_fp16, y = var_3431_to_fp16)[name = string("op_3432_cast_fp16")];
            tensor<fp16, [3, 4, 32, 128]> v_13_cast_fp16 = transpose(perm = var_3307, x = var_3302_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [3, 4, 32, 128]> vb_25_cast_fp16 = mul(x = v_13_cast_fp16, y = var_3432_cast_fp16)[name = string("vb_25_cast_fp16")];
            tensor<int32, [4]> q_59_begin_0 = const()[name = string("q_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_59_end_0 = const()[name = string("q_59_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_59_end_mask_0 = const()[name = string("q_59_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_59_cast_fp16 = slice_by_index(begin = q_59_begin_0, end = q_59_end_0, end_mask = q_59_end_mask_0, x = q_57_cast_fp16)[name = string("q_59_cast_fp16")];
            tensor<int32, [4]> k_61_begin_0 = const()[name = string("k_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_61_end_0 = const()[name = string("k_61_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_61_end_mask_0 = const()[name = string("k_61_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, x = k_59_cast_fp16)[name = string("k_61_cast_fp16")];
            tensor<fp16, [3, 4, 32, 128]> var_3466_cast_fp16 = mul(x = k_59_cast_fp16, y = var_3430_cast_fp16)[name = string("op_3466_cast_fp16")];
            tensor<int32, [4]> kb_17_begin_0 = const()[name = string("kb_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_17_end_0 = const()[name = string("kb_17_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_17_end_mask_0 = const()[name = string("kb_17_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_17_cast_fp16 = slice_by_index(begin = kb_17_begin_0, end = kb_17_end_0, end_mask = kb_17_end_mask_0, x = var_3466_cast_fp16)[name = string("kb_17_cast_fp16")];
            tensor<int32, [4]> vb_27_begin_0 = const()[name = string("vb_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_27_end_0 = const()[name = string("vb_27_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_27_end_mask_0 = const()[name = string("vb_27_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_27_cast_fp16 = slice_by_index(begin = vb_27_begin_0, end = vb_27_end_0, end_mask = vb_27_end_mask_0, x = vb_25_cast_fp16)[name = string("vb_27_cast_fp16")];
            tensor<int32, [3]> gdec_27_begin_0 = const()[name = string("gdec_27_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_27_end_0 = const()[name = string("gdec_27_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_27_end_mask_0 = const()[name = string("gdec_27_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 32]> gdec_25_cast_fp16 = transpose(perm = var_3427, x = var_3423_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [3, 4, 16]> gdec_27_cast_fp16 = slice_by_index(begin = gdec_27_begin_0, end = gdec_27_end_0, end_mask = gdec_27_end_mask_0, x = gdec_25_cast_fp16)[name = string("gdec_27_cast_fp16")];
            tensor<int32, [2]> var_3514 = const()[name = string("op_3514"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_3515_cast_fp16 = reshape(shape = var_3514, x = gdec_27_cast_fp16)[name = string("op_3515_cast_fp16")];
            tensor<fp16, [12, 256]> var_3516_cast_fp16 = linear(bias = var_631_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3515_cast_fp16)[name = string("op_3516_cast_fp16")];
            tensor<int32, [4]> var_3521 = const()[name = string("op_3521"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_17_cast_fp16 = reshape(shape = var_3521, x = var_3516_cast_fp16)[name = string("P_17_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3523_cast_fp16 = exp(x = P_17_cast_fp16)[name = string("op_3523_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_17_cast_fp16 = mul(x = var_3523_cast_fp16, y = tril_incl_to_fp16)[name = string("D_17_cast_fp16")];
            bool kk_17_transpose_x_1 = const()[name = string("kk_17_transpose_x_1"), val = bool(false)];
            bool kk_17_transpose_y_1 = const()[name = string("kk_17_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_17_cast_fp16 = matmul(transpose_x = kk_17_transpose_x_1, transpose_y = kk_17_transpose_y_1, x = kb_17_cast_fp16, y = k_61_cast_fp16)[name = string("kk_17_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_3529_cast_fp16 = mul(x = kk_17_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_3529_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3530_cast_fp16 = mul(x = var_3529_cast_fp16, y = D_17_cast_fp16)[name = string("op_3530_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_17_cast_fp16 = mul(x = var_3530_cast_fp16, y = cum_tail_to_fp16)[name = string("N_17_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3532_cast_fp16 = mul(x = D_17_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3532_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_17_cast_fp16 = mul(x = kk_17_cast_fp16, y = var_3532_cast_fp16)[name = string("Na_17_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_65_cast_fp16 = add(x = eye_to_fp16, y = N_17_cast_fp16)[name = string("Minv_65_cast_fp16")];
            bool Np_49_transpose_x_0 = const()[name = string("Np_49_transpose_x_0"), val = bool(false)];
            bool Np_49_transpose_y_0 = const()[name = string("Np_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_49_cast_fp16 = matmul(transpose_x = Np_49_transpose_x_0, transpose_y = Np_49_transpose_y_0, x = N_17_cast_fp16, y = Na_17_cast_fp16)[name = string("Np_49_cast_fp16")];
            bool Npa_33_transpose_x_0 = const()[name = string("Npa_33_transpose_x_0"), val = bool(false)];
            bool Npa_33_transpose_y_0 = const()[name = string("Npa_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_33_cast_fp16 = matmul(transpose_x = Npa_33_transpose_x_0, transpose_y = Npa_33_transpose_y_0, x = Na_17_cast_fp16, y = N_17_cast_fp16)[name = string("Npa_33_cast_fp16")];
            bool var_3538_transpose_x_0 = const()[name = string("op_3538_transpose_x_0"), val = bool(false)];
            bool var_3538_transpose_y_0 = const()[name = string("op_3538_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3538_cast_fp16 = matmul(transpose_x = var_3538_transpose_x_0, transpose_y = var_3538_transpose_y_0, x = Minv_65_cast_fp16, y = Np_49_cast_fp16)[name = string("op_3538_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_67_cast_fp16 = add(x = Minv_65_cast_fp16, y = var_3538_cast_fp16)[name = string("Minv_67_cast_fp16")];
            bool Np_51_transpose_x_0 = const()[name = string("Np_51_transpose_x_0"), val = bool(false)];
            bool Np_51_transpose_y_0 = const()[name = string("Np_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_51_cast_fp16 = matmul(transpose_x = Np_51_transpose_x_0, transpose_y = Np_51_transpose_y_0, x = Np_49_cast_fp16, y = Npa_33_cast_fp16)[name = string("Np_51_cast_fp16")];
            bool Npa_35_transpose_x_0 = const()[name = string("Npa_35_transpose_x_0"), val = bool(false)];
            bool Npa_35_transpose_y_0 = const()[name = string("Npa_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_35_cast_fp16 = matmul(transpose_x = Npa_35_transpose_x_0, transpose_y = Npa_35_transpose_y_0, x = Npa_33_cast_fp16, y = Np_49_cast_fp16)[name = string("Npa_35_cast_fp16")];
            bool var_3543_transpose_x_0 = const()[name = string("op_3543_transpose_x_0"), val = bool(false)];
            bool var_3543_transpose_y_0 = const()[name = string("op_3543_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3543_cast_fp16 = matmul(transpose_x = var_3543_transpose_x_0, transpose_y = var_3543_transpose_y_0, x = Minv_67_cast_fp16, y = Np_51_cast_fp16)[name = string("op_3543_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_69_cast_fp16 = add(x = Minv_67_cast_fp16, y = var_3543_cast_fp16)[name = string("Minv_69_cast_fp16")];
            bool Np_53_transpose_x_0 = const()[name = string("Np_53_transpose_x_0"), val = bool(false)];
            bool Np_53_transpose_y_0 = const()[name = string("Np_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_53_cast_fp16 = matmul(transpose_x = Np_53_transpose_x_0, transpose_y = Np_53_transpose_y_0, x = Np_51_cast_fp16, y = Npa_35_cast_fp16)[name = string("Np_53_cast_fp16")];
            bool var_3547_transpose_x_0 = const()[name = string("op_3547_transpose_x_0"), val = bool(false)];
            bool var_3547_transpose_y_0 = const()[name = string("op_3547_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3547_cast_fp16 = matmul(transpose_x = var_3547_transpose_x_0, transpose_y = var_3547_transpose_y_0, x = Minv_69_cast_fp16, y = Np_53_cast_fp16)[name = string("op_3547_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_71_cast_fp16 = add(x = Minv_69_cast_fp16, y = var_3547_cast_fp16)[name = string("Minv_71_cast_fp16")];
            bool u_17_transpose_x_0 = const()[name = string("u_17_transpose_x_0"), val = bool(false)];
            bool u_17_transpose_y_0 = const()[name = string("u_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_17_cast_fp16 = matmul(transpose_x = u_17_transpose_x_0, transpose_y = u_17_transpose_y_0, x = Minv_71_cast_fp16, y = vb_27_cast_fp16)[name = string("u_17_cast_fp16")];
            bool var_3554_transpose_x_1 = const()[name = string("op_3554_transpose_x_1"), val = bool(false)];
            bool var_3554_transpose_y_1 = const()[name = string("op_3554_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_3554_cast_fp16 = matmul(transpose_x = var_3554_transpose_x_1, transpose_y = var_3554_transpose_y_1, x = q_59_cast_fp16, y = k_61_cast_fp16)[name = string("op_3554_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_17_cast_fp16 = mul(x = var_3554_cast_fp16, y = D_17_cast_fp16)[name = string("A_17_cast_fp16")];
            bool var_3556_transpose_x_0 = const()[name = string("op_3556_transpose_x_0"), val = bool(false)];
            bool var_3556_transpose_y_0 = const()[name = string("op_3556_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3556_cast_fp16 = matmul(transpose_x = var_3556_transpose_x_0, transpose_y = var_3556_transpose_y_0, x = A_17_cast_fp16, y = u_17_cast_fp16)[name = string("op_3556_cast_fp16")];
            tensor<fp16, [16]> var_3557_bias_0_to_fp16 = const()[name = string("op_3557_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_3557_cast_fp16 = linear(bias = var_3557_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_27_cast_fp16)[name = string("op_3557_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_9_cast_fp16 = exp(x = var_3557_cast_fp16)[name = string("tail_9_cast_fp16")];
            tensor<int32, [1]> var_3560_axes_0 = const()[name = string("op_3560_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_3560_cast_fp16 = expand_dims(axes = var_3560_axes_0, x = tail_9_cast_fp16)[name = string("op_3560_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3561_cast_fp16 = mul(x = k_61_cast_fp16, y = var_3560_cast_fp16)[name = string("op_3561_cast_fp16")];
            bool S_9_transpose_x_1 = const()[name = string("S_9_transpose_x_1"), val = bool(true)];
            bool S_9_transpose_y_1 = const()[name = string("S_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_9_cast_fp16 = matmul(transpose_x = S_9_transpose_x_1, transpose_y = S_9_transpose_y_1, x = var_3561_cast_fp16, y = u_17_cast_fp16)[name = string("S_9_cast_fp16")];
            tensor<int32, [4]> q_61_begin_0 = const()[name = string("q_61_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_61_end_0 = const()[name = string("q_61_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_61_end_mask_0 = const()[name = string("q_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, x = q_57_cast_fp16)[name = string("q_61_cast_fp16")];
            tensor<int32, [4]> k_63_begin_0 = const()[name = string("k_63_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_63_end_0 = const()[name = string("k_63_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_63_end_mask_0 = const()[name = string("k_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_63_cast_fp16 = slice_by_index(begin = k_63_begin_0, end = k_63_end_0, end_mask = k_63_end_mask_0, x = k_59_cast_fp16)[name = string("k_63_cast_fp16")];
            tensor<int32, [4]> kb_19_begin_0 = const()[name = string("kb_19_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_19_end_0 = const()[name = string("kb_19_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_19_end_mask_0 = const()[name = string("kb_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_19_cast_fp16 = slice_by_index(begin = kb_19_begin_0, end = kb_19_end_0, end_mask = kb_19_end_mask_0, x = var_3466_cast_fp16)[name = string("kb_19_cast_fp16")];
            tensor<int32, [4]> vb_29_begin_0 = const()[name = string("vb_29_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_29_end_0 = const()[name = string("vb_29_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_29_end_mask_0 = const()[name = string("vb_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_29_cast_fp16 = slice_by_index(begin = vb_29_begin_0, end = vb_29_end_0, end_mask = vb_29_end_mask_0, x = vb_25_cast_fp16)[name = string("vb_29_cast_fp16")];
            tensor<int32, [3]> gdec_29_begin_0 = const()[name = string("gdec_29_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_29_end_0 = const()[name = string("gdec_29_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_29_end_mask_0 = const()[name = string("gdec_29_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_29_cast_fp16 = slice_by_index(begin = gdec_29_begin_0, end = gdec_29_end_0, end_mask = gdec_29_end_mask_0, x = gdec_25_cast_fp16)[name = string("gdec_29_cast_fp16")];
            tensor<int32, [2]> var_3646 = const()[name = string("op_3646"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_3647_cast_fp16 = reshape(shape = var_3646, x = gdec_29_cast_fp16)[name = string("op_3647_cast_fp16")];
            tensor<fp16, [12, 256]> var_3648_cast_fp16 = linear(bias = var_631_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3647_cast_fp16)[name = string("op_3648_cast_fp16")];
            tensor<int32, [4]> var_3653 = const()[name = string("op_3653"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_19_cast_fp16 = reshape(shape = var_3653, x = var_3648_cast_fp16)[name = string("P_19_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3655_cast_fp16 = exp(x = P_19_cast_fp16)[name = string("op_3655_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_19_cast_fp16 = mul(x = var_3655_cast_fp16, y = tril_incl_to_fp16)[name = string("D_19_cast_fp16")];
            bool kk_19_transpose_x_1 = const()[name = string("kk_19_transpose_x_1"), val = bool(false)];
            bool kk_19_transpose_y_1 = const()[name = string("kk_19_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_19_cast_fp16 = matmul(transpose_x = kk_19_transpose_x_1, transpose_y = kk_19_transpose_y_1, x = kb_19_cast_fp16, y = k_63_cast_fp16)[name = string("kk_19_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_3661_cast_fp16 = mul(x = kk_19_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_3661_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3662_cast_fp16 = mul(x = var_3661_cast_fp16, y = D_19_cast_fp16)[name = string("op_3662_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_19_cast_fp16 = mul(x = var_3662_cast_fp16, y = cum_tail_to_fp16)[name = string("N_19_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3664_cast_fp16 = mul(x = D_19_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3664_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_19_cast_fp16 = mul(x = kk_19_cast_fp16, y = var_3664_cast_fp16)[name = string("Na_19_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_73_cast_fp16 = add(x = eye_to_fp16, y = N_19_cast_fp16)[name = string("Minv_73_cast_fp16")];
            bool Np_55_transpose_x_0 = const()[name = string("Np_55_transpose_x_0"), val = bool(false)];
            bool Np_55_transpose_y_0 = const()[name = string("Np_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_55_cast_fp16 = matmul(transpose_x = Np_55_transpose_x_0, transpose_y = Np_55_transpose_y_0, x = N_19_cast_fp16, y = Na_19_cast_fp16)[name = string("Np_55_cast_fp16")];
            bool Npa_37_transpose_x_0 = const()[name = string("Npa_37_transpose_x_0"), val = bool(false)];
            bool Npa_37_transpose_y_0 = const()[name = string("Npa_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_37_cast_fp16 = matmul(transpose_x = Npa_37_transpose_x_0, transpose_y = Npa_37_transpose_y_0, x = Na_19_cast_fp16, y = N_19_cast_fp16)[name = string("Npa_37_cast_fp16")];
            bool var_3670_transpose_x_0 = const()[name = string("op_3670_transpose_x_0"), val = bool(false)];
            bool var_3670_transpose_y_0 = const()[name = string("op_3670_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3670_cast_fp16 = matmul(transpose_x = var_3670_transpose_x_0, transpose_y = var_3670_transpose_y_0, x = Minv_73_cast_fp16, y = Np_55_cast_fp16)[name = string("op_3670_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_75_cast_fp16 = add(x = Minv_73_cast_fp16, y = var_3670_cast_fp16)[name = string("Minv_75_cast_fp16")];
            bool Np_57_transpose_x_0 = const()[name = string("Np_57_transpose_x_0"), val = bool(false)];
            bool Np_57_transpose_y_0 = const()[name = string("Np_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_57_cast_fp16 = matmul(transpose_x = Np_57_transpose_x_0, transpose_y = Np_57_transpose_y_0, x = Np_55_cast_fp16, y = Npa_37_cast_fp16)[name = string("Np_57_cast_fp16")];
            bool Npa_39_transpose_x_0 = const()[name = string("Npa_39_transpose_x_0"), val = bool(false)];
            bool Npa_39_transpose_y_0 = const()[name = string("Npa_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_39_cast_fp16 = matmul(transpose_x = Npa_39_transpose_x_0, transpose_y = Npa_39_transpose_y_0, x = Npa_37_cast_fp16, y = Np_55_cast_fp16)[name = string("Npa_39_cast_fp16")];
            bool var_3675_transpose_x_0 = const()[name = string("op_3675_transpose_x_0"), val = bool(false)];
            bool var_3675_transpose_y_0 = const()[name = string("op_3675_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3675_cast_fp16 = matmul(transpose_x = var_3675_transpose_x_0, transpose_y = var_3675_transpose_y_0, x = Minv_75_cast_fp16, y = Np_57_cast_fp16)[name = string("op_3675_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_77_cast_fp16 = add(x = Minv_75_cast_fp16, y = var_3675_cast_fp16)[name = string("Minv_77_cast_fp16")];
            bool Np_59_transpose_x_0 = const()[name = string("Np_59_transpose_x_0"), val = bool(false)];
            bool Np_59_transpose_y_0 = const()[name = string("Np_59_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_59_cast_fp16 = matmul(transpose_x = Np_59_transpose_x_0, transpose_y = Np_59_transpose_y_0, x = Np_57_cast_fp16, y = Npa_39_cast_fp16)[name = string("Np_59_cast_fp16")];
            bool var_3679_transpose_x_0 = const()[name = string("op_3679_transpose_x_0"), val = bool(false)];
            bool var_3679_transpose_y_0 = const()[name = string("op_3679_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3679_cast_fp16 = matmul(transpose_x = var_3679_transpose_x_0, transpose_y = var_3679_transpose_y_0, x = Minv_77_cast_fp16, y = Np_59_cast_fp16)[name = string("op_3679_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_79_cast_fp16 = add(x = Minv_77_cast_fp16, y = var_3679_cast_fp16)[name = string("Minv_79_cast_fp16")];
            bool u_19_transpose_x_0 = const()[name = string("u_19_transpose_x_0"), val = bool(false)];
            bool u_19_transpose_y_0 = const()[name = string("u_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_19_cast_fp16 = matmul(transpose_x = u_19_transpose_x_0, transpose_y = u_19_transpose_y_0, x = Minv_79_cast_fp16, y = vb_29_cast_fp16)[name = string("u_19_cast_fp16")];
            bool var_3686_transpose_x_1 = const()[name = string("op_3686_transpose_x_1"), val = bool(false)];
            bool var_3686_transpose_y_1 = const()[name = string("op_3686_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_3686_cast_fp16 = matmul(transpose_x = var_3686_transpose_x_1, transpose_y = var_3686_transpose_y_1, x = q_61_cast_fp16, y = k_63_cast_fp16)[name = string("op_3686_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_19_cast_fp16 = mul(x = var_3686_cast_fp16, y = D_19_cast_fp16)[name = string("A_19_cast_fp16")];
            tensor<fp16, [16]> var_3688_bias_0_to_fp16 = const()[name = string("op_3688_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_3688_cast_fp16 = linear(bias = var_3688_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_29_cast_fp16)[name = string("op_3688_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_9_cast_fp16 = exp(x = var_3688_cast_fp16)[name = string("cum_9_cast_fp16")];
            tensor<int32, [1]> var_3691_axes_0 = const()[name = string("op_3691_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_3691_cast_fp16 = expand_dims(axes = var_3691_axes_0, x = cum_9_cast_fp16)[name = string("op_3691_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3692_cast_fp16 = mul(x = kb_19_cast_fp16, y = var_3691_cast_fp16)[name = string("op_3692_cast_fp16")];
            bool w_19_transpose_x_0 = const()[name = string("w_19_transpose_x_0"), val = bool(false)];
            bool w_19_transpose_y_0 = const()[name = string("w_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_19_cast_fp16 = matmul(transpose_x = w_19_transpose_x_0, transpose_y = w_19_transpose_y_0, x = Minv_79_cast_fp16, y = var_3692_cast_fp16)[name = string("w_19_cast_fp16")];
            bool var_3694_transpose_x_0 = const()[name = string("op_3694_transpose_x_0"), val = bool(false)];
            bool var_3694_transpose_y_0 = const()[name = string("op_3694_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3694_cast_fp16 = matmul(transpose_x = var_3694_transpose_x_0, transpose_y = var_3694_transpose_y_0, x = w_19_cast_fp16, y = S_9_cast_fp16)[name = string("op_3694_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_9_cast_fp16 = sub(x = u_19_cast_fp16, y = var_3694_cast_fp16)[name = string("vnew_9_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3699_cast_fp16 = mul(x = q_61_cast_fp16, y = var_3691_cast_fp16)[name = string("op_3699_cast_fp16")];
            bool var_3700_transpose_x_0 = const()[name = string("op_3700_transpose_x_0"), val = bool(false)];
            bool var_3700_transpose_y_0 = const()[name = string("op_3700_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3700_cast_fp16 = matmul(transpose_x = var_3700_transpose_x_0, transpose_y = var_3700_transpose_y_0, x = var_3699_cast_fp16, y = S_9_cast_fp16)[name = string("op_3700_cast_fp16")];
            bool var_3701_transpose_x_0 = const()[name = string("op_3701_transpose_x_0"), val = bool(false)];
            bool var_3701_transpose_y_0 = const()[name = string("op_3701_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3701_cast_fp16 = matmul(transpose_x = var_3701_transpose_x_0, transpose_y = var_3701_transpose_y_0, x = A_19_cast_fp16, y = vnew_9_cast_fp16)[name = string("op_3701_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_9_cast_fp16 = add(x = var_3700_cast_fp16, y = var_3701_cast_fp16)[name = string("oc_9_cast_fp16")];
            int32 var_3705 = const()[name = string("op_3705"), val = int32(2)];
            bool o_27_interleave_0 = const()[name = string("o_27_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 32, 128]> o_27_cast_fp16 = concat(axis = var_3705, interleave = o_27_interleave_0, values = (var_3556_cast_fp16, oc_9_cast_fp16))[name = string("o_27_cast_fp16")];
            tensor<int32, [4]> var_3711 = const()[name = string("op_3711"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3717 = const()[name = string("op_3717"), val = tensor<int32, [4]>([3, 32, 4, 128])];
            tensor<fp16, [3, 32, 4, 128]> input_81_cast_fp16 = reshape(shape = var_3717, x = z_9_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> o_29_cast_fp16 = transpose(perm = var_3711, x = o_27_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [3, 32, 4, 128]> var_3719_cast_fp16 = mul(x = o_29_cast_fp16, y = o_29_cast_fp16)[name = string("op_3719_cast_fp16")];
            tensor<int32, [1]> var_3724_axes_0 = const()[name = string("op_3724_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3724_keep_dims_0 = const()[name = string("op_3724_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 4, 1]> var_3724_cast_fp16 = reduce_mean(axes = var_3724_axes_0, keep_dims = var_3724_keep_dims_0, x = var_3719_cast_fp16)[name = string("op_3724_cast_fp16")];
            fp16 var_3726_to_fp16 = const()[name = string("op_3726_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 32, 4, 1]> var_3727_cast_fp16 = add(x = var_3724_cast_fp16, y = var_3726_to_fp16)[name = string("op_3727_cast_fp16")];
            fp32 var_3728_epsilon_0 = const()[name = string("op_3728_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 4, 1]> var_3728_cast_fp16 = rsqrt(epsilon = var_3728_epsilon_0, x = var_3727_cast_fp16)[name = string("op_3728_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> var_3729_cast_fp16 = mul(x = o_29_cast_fp16, y = var_3728_cast_fp16)[name = string("op_3729_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> var_3731_cast_fp16 = silu(x = input_81_cast_fp16)[name = string("op_3731_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> o_31_cast_fp16 = mul(x = var_3729_cast_fp16, y = var_3731_cast_fp16)[name = string("o_31_cast_fp16")];
            tensor<int32, [3]> var_3736 = const()[name = string("op_3736"), val = tensor<int32, [3]>([3, 32, 512])];
            tensor<fp16, [3, 32, 512]> input_83_cast_fp16 = reshape(shape = var_3736, x = o_31_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [640, 512]> layers_5_out_weight_to_fp16 = const()[name = string("layers_5_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84536512)))];
            tensor<fp16, [3, 32, 640]> linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_out_weight_to_fp16, x = input_83_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [3, 32, 640]> x_57_cast_fp16 = add(x = x_53_cast_fp16, y = linear_21_cast_fp16)[name = string("x_57_cast_fp16")];
            fp16 var_3749_to_fp16 = const()[name = string("op_3749_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 32, 640]> x_59_cast_fp16 = mul(x = x_57_cast_fp16, y = var_3749_to_fp16)[name = string("x_59_cast_fp16")];
            fp16 var_3747_promoted_to_fp16 = const()[name = string("op_3747_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> var_3751_cast_fp16 = pow(x = x_59_cast_fp16, y = var_3747_promoted_to_fp16)[name = string("op_3751_cast_fp16")];
            tensor<int32, [1]> var_3753_axes_0 = const()[name = string("op_3753_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3753_keep_dims_0 = const()[name = string("op_3753_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 1]> var_3753_cast_fp16 = reduce_mean(axes = var_3753_axes_0, keep_dims = var_3753_keep_dims_0, x = var_3751_cast_fp16)[name = string("op_3753_cast_fp16")];
            fp16 var_3754_to_fp16 = const()[name = string("op_3754_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 32, 1]> var_3755_cast_fp16 = add(x = var_3753_cast_fp16, y = var_3754_to_fp16)[name = string("op_3755_cast_fp16")];
            fp32 var_3756_epsilon_0 = const()[name = string("op_3756_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 1]> var_3756_cast_fp16 = rsqrt(epsilon = var_3756_epsilon_0, x = var_3755_cast_fp16)[name = string("op_3756_cast_fp16")];
            tensor<fp16, [3, 32, 640]> var_3757_cast_fp16 = mul(x = x_59_cast_fp16, y = var_3756_cast_fp16)[name = string("op_3757_cast_fp16")];
            tensor<fp16, [640]> layers_5_n2_w1_to_fp16 = const()[name = string("layers_5_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85191936)))];
            tensor<fp16, [3, 32, 640]> input_85_cast_fp16 = mul(x = var_3757_cast_fp16, y = layers_5_n2_w1_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_5_gate_up_weight_to_fp16 = const()[name = string("layers_5_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85193280)))];
            tensor<fp16, [3, 32, 5120]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_5_gate_up_weight_to_fp16, x = input_85_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [3]> input_87_begin_0 = const()[name = string("input_87_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_87_end_0 = const()[name = string("input_87_end_0"), val = tensor<int32, [3]>([3, 32, 2560])];
            tensor<bool, [3]> input_87_end_mask_0 = const()[name = string("input_87_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 2560]> input_87_cast_fp16 = slice_by_index(begin = input_87_begin_0, end = input_87_end_0, end_mask = input_87_end_mask_0, x = linear_22_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [3, 32, 2560]> var_3767_cast_fp16 = silu(x = input_87_cast_fp16)[name = string("op_3767_cast_fp16")];
            tensor<int32, [3]> var_3772_begin_0 = const()[name = string("op_3772_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_3772_end_0 = const()[name = string("op_3772_end_0"), val = tensor<int32, [3]>([3, 32, 5120])];
            tensor<bool, [3]> var_3772_end_mask_0 = const()[name = string("op_3772_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 2560]> var_3772_cast_fp16 = slice_by_index(begin = var_3772_begin_0, end = var_3772_end_0, end_mask = var_3772_end_mask_0, x = linear_22_cast_fp16)[name = string("op_3772_cast_fp16")];
            tensor<fp16, [3, 32, 2560]> input_89_cast_fp16 = mul(x = var_3767_cast_fp16, y = var_3772_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_5_down_weight_to_fp16 = const()[name = string("layers_5_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91746944)))];
            tensor<fp16, [3, 32, 640]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_down_weight_to_fp16, x = input_89_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<fp16, [3, 32, 640]> x_61_cast_fp16 = add(x = x_57_cast_fp16, y = linear_23_cast_fp16)[name = string("x_61_cast_fp16")];
            fp16 var_3785_to_fp16 = const()[name = string("op_3785_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [3, 32, 640]> x_63_cast_fp16 = mul(x = x_61_cast_fp16, y = var_3785_to_fp16)[name = string("x_63_cast_fp16")];
            fp16 var_3783_promoted_to_fp16 = const()[name = string("op_3783_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> var_3787_cast_fp16 = pow(x = x_63_cast_fp16, y = var_3783_promoted_to_fp16)[name = string("op_3787_cast_fp16")];
            tensor<int32, [1]> var_3789_axes_0 = const()[name = string("op_3789_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3789_keep_dims_0 = const()[name = string("op_3789_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 1]> var_3789_cast_fp16 = reduce_mean(axes = var_3789_axes_0, keep_dims = var_3789_keep_dims_0, x = var_3787_cast_fp16)[name = string("op_3789_cast_fp16")];
            fp16 var_3790_to_fp16 = const()[name = string("op_3790_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 32, 1]> var_3791_cast_fp16 = add(x = var_3789_cast_fp16, y = var_3790_to_fp16)[name = string("op_3791_cast_fp16")];
            fp32 var_3792_epsilon_0 = const()[name = string("op_3792_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 1]> var_3792_cast_fp16 = rsqrt(epsilon = var_3792_epsilon_0, x = var_3791_cast_fp16)[name = string("op_3792_cast_fp16")];
            tensor<fp16, [3, 32, 640]> var_3793_cast_fp16 = mul(x = x_63_cast_fp16, y = var_3792_cast_fp16)[name = string("op_3793_cast_fp16")];
            tensor<fp16, [640]> layers_6_n1_w1_to_fp16 = const()[name = string("layers_6_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95023808)))];
            tensor<fp16, [3, 32, 640]> input_91_cast_fp16 = mul(x = var_3793_cast_fp16, y = layers_6_n1_w1_to_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_6_qkvzba_weight_to_fp16 = const()[name = string("layers_6_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95025152)))];
            tensor<fp16, [3, 32, 1544]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_qkvzba_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<int32, [3]> qkv_11_begin_0 = const()[name = string("qkv_11_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_11_end_0 = const()[name = string("qkv_11_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> qkv_11_end_mask_0 = const()[name = string("qkv_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 1024]> qkv_11_cast_fp16 = slice_by_index(begin = qkv_11_begin_0, end = qkv_11_end_0, end_mask = qkv_11_end_mask_0, x = linear_24_cast_fp16)[name = string("qkv_11_cast_fp16")];
            tensor<int32, [3]> z_begin_0 = const()[name = string("z_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_end_0 = const()[name = string("z_end_0"), val = tensor<int32, [3]>([3, 32, 1536])];
            tensor<bool, [3]> z_end_mask_0 = const()[name = string("z_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 512]> z_cast_fp16 = slice_by_index(begin = z_begin_0, end = z_end_0, end_mask = z_end_mask_0, x = linear_24_cast_fp16)[name = string("z_cast_fp16")];
            tensor<int32, [3]> b_begin_0 = const()[name = string("b_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_end_0 = const()[name = string("b_end_0"), val = tensor<int32, [3]>([3, 32, 1540])];
            tensor<bool, [3]> b_end_mask_0 = const()[name = string("b_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 4]> b_cast_fp16 = slice_by_index(begin = b_begin_0, end = b_end_0, end_mask = b_end_mask_0, x = linear_24_cast_fp16)[name = string("b_cast_fp16")];
            tensor<int32, [3]> a_begin_0 = const()[name = string("a_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_end_0 = const()[name = string("a_end_0"), val = tensor<int32, [3]>([3, 32, 1])];
            tensor<bool, [3]> a_end_mask_0 = const()[name = string("a_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 4]> a_cast_fp16 = slice_by_index(begin = a_begin_0, end = a_end_0, end_mask = a_end_mask_0, x = linear_24_cast_fp16)[name = string("a_cast_fp16")];
            int32 var_3832 = const()[name = string("op_3832"), val = int32(1)];
            bool xp_interleave_0 = const()[name = string("xp_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 35, 1024]> xp_cast_fp16 = concat(axis = var_3832, interleave = xp_interleave_0, values = (zpad_to_fp16, qkv_11_cast_fp16))[name = string("xp_cast_fp16")];
            tensor<int32, [3]> var_3843_begin_0 = const()[name = string("op_3843_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_3843_end_0 = const()[name = string("op_3843_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_3843_end_mask_0 = const()[name = string("op_3843_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 1024]> var_3843_cast_fp16 = slice_by_index(begin = var_3843_begin_0, end = var_3843_end_0, end_mask = var_3843_end_mask_0, x = xp_cast_fp16)[name = string("op_3843_cast_fp16")];
            tensor<fp16, [1024]> var_3851_to_fp16 = const()[name = string("op_3851_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97001536)))];
            tensor<fp16, [3, 32, 1024]> var_3852_cast_fp16 = mul(x = var_3843_cast_fp16, y = var_3851_to_fp16)[name = string("op_3852_cast_fp16")];
            tensor<int32, [3]> var_3862_begin_0 = const()[name = string("op_3862_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_3862_end_0 = const()[name = string("op_3862_end_0"), val = tensor<int32, [3]>([3, 34, 1024])];
            tensor<bool, [3]> var_3862_end_mask_0 = const()[name = string("op_3862_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_3862_cast_fp16 = slice_by_index(begin = var_3862_begin_0, end = var_3862_end_0, end_mask = var_3862_end_mask_0, x = xp_cast_fp16)[name = string("op_3862_cast_fp16")];
            tensor<fp16, [1024]> var_3870_to_fp16 = const()[name = string("op_3870_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97003648)))];
            tensor<fp16, [3, 32, 1024]> var_3871_cast_fp16 = mul(x = var_3862_cast_fp16, y = var_3870_to_fp16)[name = string("op_3871_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> var_3873_cast_fp16 = add(x = var_3852_cast_fp16, y = var_3871_cast_fp16)[name = string("op_3873_cast_fp16")];
            tensor<int32, [3]> var_3883_begin_0 = const()[name = string("op_3883_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_3883_end_0 = const()[name = string("op_3883_end_0"), val = tensor<int32, [3]>([3, 33, 1024])];
            tensor<bool, [3]> var_3883_end_mask_0 = const()[name = string("op_3883_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_3883_cast_fp16 = slice_by_index(begin = var_3883_begin_0, end = var_3883_end_0, end_mask = var_3883_end_mask_0, x = xp_cast_fp16)[name = string("op_3883_cast_fp16")];
            tensor<fp16, [1024]> var_3891_to_fp16 = const()[name = string("op_3891_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97005760)))];
            tensor<fp16, [3, 32, 1024]> var_3892_cast_fp16 = mul(x = var_3883_cast_fp16, y = var_3891_to_fp16)[name = string("op_3892_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> var_3894_cast_fp16 = add(x = var_3873_cast_fp16, y = var_3892_cast_fp16)[name = string("op_3894_cast_fp16")];
            tensor<int32, [3]> var_3904_begin_0 = const()[name = string("op_3904_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3904_end_0 = const()[name = string("op_3904_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> var_3904_end_mask_0 = const()[name = string("op_3904_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 32, 1024]> var_3904_cast_fp16 = slice_by_index(begin = var_3904_begin_0, end = var_3904_end_0, end_mask = var_3904_end_mask_0, x = xp_cast_fp16)[name = string("op_3904_cast_fp16")];
            tensor<fp16, [1024]> var_3912_to_fp16 = const()[name = string("op_3912_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97007872)))];
            tensor<fp16, [3, 32, 1024]> var_3913_cast_fp16 = mul(x = var_3904_cast_fp16, y = var_3912_to_fp16)[name = string("op_3913_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> input_93_cast_fp16 = add(x = var_3894_cast_fp16, y = var_3913_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [3, 32, 1024]> c_cast_fp16 = silu(x = input_93_cast_fp16)[name = string("c_cast_fp16")];
            tensor<int32, [3]> var_3921_begin_0 = const()[name = string("op_3921_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3921_end_0 = const()[name = string("op_3921_end_0"), val = tensor<int32, [3]>([3, 32, 256])];
            tensor<bool, [3]> var_3921_end_mask_0 = const()[name = string("op_3921_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 256]> var_3921_cast_fp16 = slice_by_index(begin = var_3921_begin_0, end = var_3921_end_0, end_mask = var_3921_end_mask_0, x = c_cast_fp16)[name = string("op_3921_cast_fp16")];
            tensor<int32, [4]> var_3926 = const()[name = string("op_3926"), val = tensor<int32, [4]>([3, 32, 2, 128])];
            tensor<fp16, [3, 32, 2, 128]> q_63_cast_fp16 = reshape(shape = var_3926, x = var_3921_cast_fp16)[name = string("q_63_cast_fp16")];
            tensor<int32, [3]> var_3932_begin_0 = const()[name = string("op_3932_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_3932_end_0 = const()[name = string("op_3932_end_0"), val = tensor<int32, [3]>([3, 32, 512])];
            tensor<bool, [3]> var_3932_end_mask_0 = const()[name = string("op_3932_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 256]> var_3932_cast_fp16 = slice_by_index(begin = var_3932_begin_0, end = var_3932_end_0, end_mask = var_3932_end_mask_0, x = c_cast_fp16)[name = string("op_3932_cast_fp16")];
            tensor<int32, [4]> var_3937 = const()[name = string("op_3937"), val = tensor<int32, [4]>([3, 32, 2, 128])];
            tensor<fp16, [3, 32, 2, 128]> k_65_cast_fp16 = reshape(shape = var_3937, x = var_3932_cast_fp16)[name = string("k_65_cast_fp16")];
            tensor<int32, [3]> var_3943_begin_0 = const()[name = string("op_3943_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_3943_end_0 = const()[name = string("op_3943_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> var_3943_end_mask_0 = const()[name = string("op_3943_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 512]> var_3943_cast_fp16 = slice_by_index(begin = var_3943_begin_0, end = var_3943_end_0, end_mask = var_3943_end_mask_0, x = c_cast_fp16)[name = string("op_3943_cast_fp16")];
            tensor<int32, [4]> var_3948 = const()[name = string("op_3948"), val = tensor<int32, [4]>([3, 32, 4, 128])];
            tensor<fp16, [3, 32, 4, 128]> var_3949_cast_fp16 = reshape(shape = var_3948, x = var_3943_cast_fp16)[name = string("op_3949_cast_fp16")];
            tensor<int32, [4]> var_3954 = const()[name = string("op_3954"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3956_to_fp16 = const()[name = string("op_3956_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 32, 2, 128]> q_65_cast_fp16 = mul(x = q_63_cast_fp16, y = var_3956_to_fp16)[name = string("q_65_cast_fp16")];
            fp16 var_3958_to_fp16 = const()[name = string("op_3958_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 32, 2, 128]> k_67_cast_fp16 = mul(x = k_65_cast_fp16, y = var_3958_to_fp16)[name = string("k_67_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_3960_cast_fp16 = mul(x = q_65_cast_fp16, y = q_65_cast_fp16)[name = string("op_3960_cast_fp16")];
            tensor<int32, [1]> var_3965_axes_0 = const()[name = string("op_3965_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3965_keep_dims_0 = const()[name = string("op_3965_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 2, 1]> var_3965_cast_fp16 = reduce_mean(axes = var_3965_axes_0, keep_dims = var_3965_keep_dims_0, x = var_3960_cast_fp16)[name = string("op_3965_cast_fp16")];
            fp16 var_3966_promoted_to_fp16 = const()[name = string("op_3966_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 32, 2, 1]> var_3967_cast_fp16 = mul(x = var_3965_cast_fp16, y = var_3966_promoted_to_fp16)[name = string("op_3967_cast_fp16")];
            fp16 var_3969_to_fp16 = const()[name = string("op_3969_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 32, 2, 1]> var_3970_cast_fp16 = add(x = var_3967_cast_fp16, y = var_3969_to_fp16)[name = string("op_3970_cast_fp16")];
            fp32 var_3971_epsilon_0 = const()[name = string("op_3971_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 2, 1]> var_3971_cast_fp16 = rsqrt(epsilon = var_3971_epsilon_0, x = var_3970_cast_fp16)[name = string("op_3971_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_3972_cast_fp16 = mul(x = q_65_cast_fp16, y = var_3971_cast_fp16)[name = string("op_3972_cast_fp16")];
            fp16 var_3973_to_fp16 = const()[name = string("op_3973_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 32, 2, 128]> q_67_cast_fp16 = mul(x = var_3972_cast_fp16, y = var_3973_to_fp16)[name = string("q_67_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> var_3975_cast_fp16 = mul(x = k_67_cast_fp16, y = k_67_cast_fp16)[name = string("op_3975_cast_fp16")];
            tensor<int32, [1]> var_3980_axes_0 = const()[name = string("op_3980_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3980_keep_dims_0 = const()[name = string("op_3980_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 2, 1]> var_3980_cast_fp16 = reduce_mean(axes = var_3980_axes_0, keep_dims = var_3980_keep_dims_0, x = var_3975_cast_fp16)[name = string("op_3980_cast_fp16")];
            fp16 var_3981_promoted_to_fp16 = const()[name = string("op_3981_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 32, 2, 1]> var_3982_cast_fp16 = mul(x = var_3980_cast_fp16, y = var_3981_promoted_to_fp16)[name = string("op_3982_cast_fp16")];
            fp16 var_3984_to_fp16 = const()[name = string("op_3984_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 32, 2, 1]> var_3985_cast_fp16 = add(x = var_3982_cast_fp16, y = var_3984_to_fp16)[name = string("op_3985_cast_fp16")];
            fp32 var_3986_epsilon_0 = const()[name = string("op_3986_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 2, 1]> var_3986_cast_fp16 = rsqrt(epsilon = var_3986_epsilon_0, x = var_3985_cast_fp16)[name = string("op_3986_cast_fp16")];
            tensor<fp16, [3, 32, 2, 128]> k_69_cast_fp16 = mul(x = k_67_cast_fp16, y = var_3986_cast_fp16)[name = string("k_69_cast_fp16")];
            tensor<int32, [4]> var_3992 = const()[name = string("op_3992"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_4005_axes_0 = const()[name = string("op_4005_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 32, 128]> var_3993_cast_fp16 = transpose(perm = var_3992, x = q_67_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [3, 2, 1, 32, 128]> var_4005_cast_fp16 = expand_dims(axes = var_4005_axes_0, x = var_3993_cast_fp16)[name = string("op_4005_cast_fp16")];
            tensor<int32, [5]> var_4013_reps_0 = const()[name = string("op_4013_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 32, 128]> var_4013_cast_fp16 = tile(reps = var_4013_reps_0, x = var_4005_cast_fp16)[name = string("op_4013_cast_fp16")];
            tensor<int32, [4]> var_4018 = const()[name = string("op_4018"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<fp16, [3, 4, 32, 128]> q_69_cast_fp16 = reshape(shape = var_4018, x = var_4013_cast_fp16)[name = string("q_69_cast_fp16")];
            tensor<int32, [4]> var_4024 = const()[name = string("op_4024"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_4037_axes_0 = const()[name = string("op_4037_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 32, 128]> var_4025_cast_fp16 = transpose(perm = var_4024, x = k_69_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [3, 2, 1, 32, 128]> var_4037_cast_fp16 = expand_dims(axes = var_4037_axes_0, x = var_4025_cast_fp16)[name = string("op_4037_cast_fp16")];
            tensor<int32, [5]> var_4045_reps_0 = const()[name = string("op_4045_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 32, 128]> var_4045_cast_fp16 = tile(reps = var_4045_reps_0, x = var_4037_cast_fp16)[name = string("op_4045_cast_fp16")];
            tensor<int32, [4]> var_4050 = const()[name = string("op_4050"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<fp16, [3, 4, 32, 128]> k_71_cast_fp16 = reshape(shape = var_4050, x = var_4045_cast_fp16)[name = string("k_71_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_4052_cast_fp16 = sigmoid(x = b_cast_fp16)[name = string("op_4052_cast_fp16")];
            tensor<int32, [3]> var_4056 = const()[name = string("op_4056"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_6_dt_bias_to_fp16 = const()[name = string("layers_6_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 32, 4]> input_95_cast_fp16 = add(x = a_cast_fp16, y = layers_6_dt_bias_to_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_4060_cast_fp16 = relu(x = input_95_cast_fp16)[name = string("op_4060_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_4061_cast_fp16 = abs(x = input_95_cast_fp16)[name = string("op_4061_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 32, 4]> var_4062_cast_fp16 = mul(x = var_4061_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_4062_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_4063_cast_fp16 = exp(x = var_4062_cast_fp16)[name = string("op_4063_cast_fp16")];
            fp16 var_4065_to_fp16 = const()[name = string("op_4065_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 32, 4]> var_4066_cast_fp16 = add(x = var_4063_cast_fp16, y = var_4065_to_fp16)[name = string("op_4066_cast_fp16")];
            fp32 var_4067_epsilon_0 = const()[name = string("op_4067_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 32, 4]> var_4067_cast_fp16 = log(epsilon = var_4067_epsilon_0, x = var_4066_cast_fp16)[name = string("op_4067_cast_fp16")];
            tensor<fp16, [3, 32, 4]> var_4069_cast_fp16 = add(x = var_4060_cast_fp16, y = var_4067_cast_fp16)[name = string("op_4069_cast_fp16")];
            tensor<fp16, [4]> layers_6_negA_to_fp16 = const()[name = string("layers_6_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.07cp+1, -0x1.eccp+3, -0x1.63p+3, -0x1.728p+1])];
            tensor<fp16, [3, 32, 4]> var_4070_cast_fp16 = mul(x = layers_6_negA_to_fp16, y = var_4069_cast_fp16)[name = string("op_4070_cast_fp16")];
            tensor<int32, [3]> var_4074 = const()[name = string("op_4074"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_4077_axes_0 = const()[name = string("op_4077_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 32]> beta_cast_fp16 = transpose(perm = var_4056, x = var_4052_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [3, 4, 32, 1]> var_4077_cast_fp16 = expand_dims(axes = var_4077_axes_0, x = beta_cast_fp16)[name = string("op_4077_cast_fp16")];
            fp16 var_4078_to_fp16 = const()[name = string("op_4078_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 32, 1]> var_4079_cast_fp16 = mul(x = var_4077_cast_fp16, y = var_4078_to_fp16)[name = string("op_4079_cast_fp16")];
            tensor<fp16, [3, 4, 32, 128]> v_15_cast_fp16 = transpose(perm = var_3954, x = var_3949_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [3, 4, 32, 128]> vb_31_cast_fp16 = mul(x = v_15_cast_fp16, y = var_4079_cast_fp16)[name = string("vb_31_cast_fp16")];
            tensor<int32, [4]> q_71_begin_0 = const()[name = string("q_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_71_end_0 = const()[name = string("q_71_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_71_end_mask_0 = const()[name = string("q_71_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_71_cast_fp16 = slice_by_index(begin = q_71_begin_0, end = q_71_end_0, end_mask = q_71_end_mask_0, x = q_69_cast_fp16)[name = string("q_71_cast_fp16")];
            tensor<int32, [4]> k_73_begin_0 = const()[name = string("k_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_73_end_0 = const()[name = string("k_73_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_73_end_mask_0 = const()[name = string("k_73_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, x = k_71_cast_fp16)[name = string("k_73_cast_fp16")];
            tensor<fp16, [3, 4, 32, 128]> var_4113_cast_fp16 = mul(x = k_71_cast_fp16, y = var_4077_cast_fp16)[name = string("op_4113_cast_fp16")];
            tensor<int32, [4]> kb_21_begin_0 = const()[name = string("kb_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_21_end_0 = const()[name = string("kb_21_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_21_end_mask_0 = const()[name = string("kb_21_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_21_cast_fp16 = slice_by_index(begin = kb_21_begin_0, end = kb_21_end_0, end_mask = kb_21_end_mask_0, x = var_4113_cast_fp16)[name = string("kb_21_cast_fp16")];
            tensor<int32, [4]> vb_33_begin_0 = const()[name = string("vb_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_33_end_0 = const()[name = string("vb_33_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_33_end_mask_0 = const()[name = string("vb_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_33_cast_fp16 = slice_by_index(begin = vb_33_begin_0, end = vb_33_end_0, end_mask = vb_33_end_mask_0, x = vb_31_cast_fp16)[name = string("vb_33_cast_fp16")];
            tensor<int32, [3]> gdec_33_begin_0 = const()[name = string("gdec_33_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_33_end_0 = const()[name = string("gdec_33_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_33_end_mask_0 = const()[name = string("gdec_33_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 32]> gdec_31_cast_fp16 = transpose(perm = var_4074, x = var_4070_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [3, 4, 16]> gdec_33_cast_fp16 = slice_by_index(begin = gdec_33_begin_0, end = gdec_33_end_0, end_mask = gdec_33_end_mask_0, x = gdec_31_cast_fp16)[name = string("gdec_33_cast_fp16")];
            tensor<int32, [2]> var_4161 = const()[name = string("op_4161"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_4162_cast_fp16 = reshape(shape = var_4161, x = gdec_33_cast_fp16)[name = string("op_4162_cast_fp16")];
            tensor<fp16, [12, 256]> var_4163_cast_fp16 = linear(bias = var_631_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_4162_cast_fp16)[name = string("op_4163_cast_fp16")];
            tensor<int32, [4]> var_4168 = const()[name = string("op_4168"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_21_cast_fp16 = reshape(shape = var_4168, x = var_4163_cast_fp16)[name = string("P_21_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4170_cast_fp16 = exp(x = P_21_cast_fp16)[name = string("op_4170_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_21_cast_fp16 = mul(x = var_4170_cast_fp16, y = tril_incl_to_fp16)[name = string("D_21_cast_fp16")];
            bool kk_21_transpose_x_1 = const()[name = string("kk_21_transpose_x_1"), val = bool(false)];
            bool kk_21_transpose_y_1 = const()[name = string("kk_21_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_21_cast_fp16 = matmul(transpose_x = kk_21_transpose_x_1, transpose_y = kk_21_transpose_y_1, x = kb_21_cast_fp16, y = k_73_cast_fp16)[name = string("kk_21_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_4176_cast_fp16 = mul(x = kk_21_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_4176_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4177_cast_fp16 = mul(x = var_4176_cast_fp16, y = D_21_cast_fp16)[name = string("op_4177_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_21_cast_fp16 = mul(x = var_4177_cast_fp16, y = cum_tail_to_fp16)[name = string("N_21_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4179_cast_fp16 = mul(x = D_21_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_4179_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_21_cast_fp16 = mul(x = kk_21_cast_fp16, y = var_4179_cast_fp16)[name = string("Na_21_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_81_cast_fp16 = add(x = eye_to_fp16, y = N_21_cast_fp16)[name = string("Minv_81_cast_fp16")];
            bool Np_61_transpose_x_0 = const()[name = string("Np_61_transpose_x_0"), val = bool(false)];
            bool Np_61_transpose_y_0 = const()[name = string("Np_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_61_cast_fp16 = matmul(transpose_x = Np_61_transpose_x_0, transpose_y = Np_61_transpose_y_0, x = N_21_cast_fp16, y = Na_21_cast_fp16)[name = string("Np_61_cast_fp16")];
            bool Npa_41_transpose_x_0 = const()[name = string("Npa_41_transpose_x_0"), val = bool(false)];
            bool Npa_41_transpose_y_0 = const()[name = string("Npa_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_41_cast_fp16 = matmul(transpose_x = Npa_41_transpose_x_0, transpose_y = Npa_41_transpose_y_0, x = Na_21_cast_fp16, y = N_21_cast_fp16)[name = string("Npa_41_cast_fp16")];
            bool var_4185_transpose_x_0 = const()[name = string("op_4185_transpose_x_0"), val = bool(false)];
            bool var_4185_transpose_y_0 = const()[name = string("op_4185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4185_cast_fp16 = matmul(transpose_x = var_4185_transpose_x_0, transpose_y = var_4185_transpose_y_0, x = Minv_81_cast_fp16, y = Np_61_cast_fp16)[name = string("op_4185_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_83_cast_fp16 = add(x = Minv_81_cast_fp16, y = var_4185_cast_fp16)[name = string("Minv_83_cast_fp16")];
            bool Np_63_transpose_x_0 = const()[name = string("Np_63_transpose_x_0"), val = bool(false)];
            bool Np_63_transpose_y_0 = const()[name = string("Np_63_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_63_cast_fp16 = matmul(transpose_x = Np_63_transpose_x_0, transpose_y = Np_63_transpose_y_0, x = Np_61_cast_fp16, y = Npa_41_cast_fp16)[name = string("Np_63_cast_fp16")];
            bool Npa_43_transpose_x_0 = const()[name = string("Npa_43_transpose_x_0"), val = bool(false)];
            bool Npa_43_transpose_y_0 = const()[name = string("Npa_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_43_cast_fp16 = matmul(transpose_x = Npa_43_transpose_x_0, transpose_y = Npa_43_transpose_y_0, x = Npa_41_cast_fp16, y = Np_61_cast_fp16)[name = string("Npa_43_cast_fp16")];
            bool var_4190_transpose_x_0 = const()[name = string("op_4190_transpose_x_0"), val = bool(false)];
            bool var_4190_transpose_y_0 = const()[name = string("op_4190_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4190_cast_fp16 = matmul(transpose_x = var_4190_transpose_x_0, transpose_y = var_4190_transpose_y_0, x = Minv_83_cast_fp16, y = Np_63_cast_fp16)[name = string("op_4190_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_85_cast_fp16 = add(x = Minv_83_cast_fp16, y = var_4190_cast_fp16)[name = string("Minv_85_cast_fp16")];
            bool Np_65_transpose_x_0 = const()[name = string("Np_65_transpose_x_0"), val = bool(false)];
            bool Np_65_transpose_y_0 = const()[name = string("Np_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_65_cast_fp16 = matmul(transpose_x = Np_65_transpose_x_0, transpose_y = Np_65_transpose_y_0, x = Np_63_cast_fp16, y = Npa_43_cast_fp16)[name = string("Np_65_cast_fp16")];
            bool var_4194_transpose_x_0 = const()[name = string("op_4194_transpose_x_0"), val = bool(false)];
            bool var_4194_transpose_y_0 = const()[name = string("op_4194_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4194_cast_fp16 = matmul(transpose_x = var_4194_transpose_x_0, transpose_y = var_4194_transpose_y_0, x = Minv_85_cast_fp16, y = Np_65_cast_fp16)[name = string("op_4194_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_87_cast_fp16 = add(x = Minv_85_cast_fp16, y = var_4194_cast_fp16)[name = string("Minv_87_cast_fp16")];
            bool u_21_transpose_x_0 = const()[name = string("u_21_transpose_x_0"), val = bool(false)];
            bool u_21_transpose_y_0 = const()[name = string("u_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_21_cast_fp16 = matmul(transpose_x = u_21_transpose_x_0, transpose_y = u_21_transpose_y_0, x = Minv_87_cast_fp16, y = vb_33_cast_fp16)[name = string("u_21_cast_fp16")];
            bool var_4201_transpose_x_1 = const()[name = string("op_4201_transpose_x_1"), val = bool(false)];
            bool var_4201_transpose_y_1 = const()[name = string("op_4201_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_4201_cast_fp16 = matmul(transpose_x = var_4201_transpose_x_1, transpose_y = var_4201_transpose_y_1, x = q_71_cast_fp16, y = k_73_cast_fp16)[name = string("op_4201_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_21_cast_fp16 = mul(x = var_4201_cast_fp16, y = D_21_cast_fp16)[name = string("A_21_cast_fp16")];
            bool var_4203_transpose_x_0 = const()[name = string("op_4203_transpose_x_0"), val = bool(false)];
            bool var_4203_transpose_y_0 = const()[name = string("op_4203_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4203_cast_fp16 = matmul(transpose_x = var_4203_transpose_x_0, transpose_y = var_4203_transpose_y_0, x = A_21_cast_fp16, y = u_21_cast_fp16)[name = string("op_4203_cast_fp16")];
            tensor<fp16, [16]> var_4204_bias_0_to_fp16 = const()[name = string("op_4204_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_4204_cast_fp16 = linear(bias = var_4204_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_33_cast_fp16)[name = string("op_4204_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_cast_fp16 = exp(x = var_4204_cast_fp16)[name = string("tail_cast_fp16")];
            tensor<int32, [1]> var_4207_axes_0 = const()[name = string("op_4207_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_4207_cast_fp16 = expand_dims(axes = var_4207_axes_0, x = tail_cast_fp16)[name = string("op_4207_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4208_cast_fp16 = mul(x = k_73_cast_fp16, y = var_4207_cast_fp16)[name = string("op_4208_cast_fp16")];
            bool S_transpose_x_1 = const()[name = string("S_transpose_x_1"), val = bool(true)];
            bool S_transpose_y_1 = const()[name = string("S_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_cast_fp16 = matmul(transpose_x = S_transpose_x_1, transpose_y = S_transpose_y_1, x = var_4208_cast_fp16, y = u_21_cast_fp16)[name = string("S_cast_fp16")];
            tensor<int32, [4]> q_73_begin_0 = const()[name = string("q_73_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_73_end_0 = const()[name = string("q_73_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_73_end_mask_0 = const()[name = string("q_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, x = q_69_cast_fp16)[name = string("q_73_cast_fp16")];
            tensor<int32, [4]> k_75_begin_0 = const()[name = string("k_75_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_75_end_0 = const()[name = string("k_75_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_75_end_mask_0 = const()[name = string("k_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_75_cast_fp16 = slice_by_index(begin = k_75_begin_0, end = k_75_end_0, end_mask = k_75_end_mask_0, x = k_71_cast_fp16)[name = string("k_75_cast_fp16")];
            tensor<int32, [4]> kb_begin_0 = const()[name = string("kb_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_end_0 = const()[name = string("kb_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_end_mask_0 = const()[name = string("kb_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_cast_fp16 = slice_by_index(begin = kb_begin_0, end = kb_end_0, end_mask = kb_end_mask_0, x = var_4113_cast_fp16)[name = string("kb_cast_fp16")];
            tensor<int32, [4]> vb_begin_0 = const()[name = string("vb_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_end_0 = const()[name = string("vb_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_end_mask_0 = const()[name = string("vb_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_cast_fp16 = slice_by_index(begin = vb_begin_0, end = vb_end_0, end_mask = vb_end_mask_0, x = vb_31_cast_fp16)[name = string("vb_cast_fp16")];
            tensor<int32, [3]> gdec_begin_0 = const()[name = string("gdec_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_end_0 = const()[name = string("gdec_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_end_mask_0 = const()[name = string("gdec_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_cast_fp16 = slice_by_index(begin = gdec_begin_0, end = gdec_end_0, end_mask = gdec_end_mask_0, x = gdec_31_cast_fp16)[name = string("gdec_cast_fp16")];
            tensor<int32, [2]> var_4293 = const()[name = string("op_4293"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_4294_cast_fp16 = reshape(shape = var_4293, x = gdec_cast_fp16)[name = string("op_4294_cast_fp16")];
            tensor<fp16, [12, 256]> var_4295_cast_fp16 = linear(bias = var_631_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_4294_cast_fp16)[name = string("op_4295_cast_fp16")];
            tensor<int32, [4]> var_4300 = const()[name = string("op_4300"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_cast_fp16 = reshape(shape = var_4300, x = var_4295_cast_fp16)[name = string("P_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4302_cast_fp16 = exp(x = P_cast_fp16)[name = string("op_4302_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_cast_fp16 = mul(x = var_4302_cast_fp16, y = tril_incl_to_fp16)[name = string("D_cast_fp16")];
            bool kk_transpose_x_1 = const()[name = string("kk_transpose_x_1"), val = bool(false)];
            bool kk_transpose_y_1 = const()[name = string("kk_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_cast_fp16 = matmul(transpose_x = kk_transpose_x_1, transpose_y = kk_transpose_y_1, x = kb_cast_fp16, y = k_75_cast_fp16)[name = string("kk_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_4308_cast_fp16 = mul(x = kk_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_4308_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4309_cast_fp16 = mul(x = var_4308_cast_fp16, y = D_cast_fp16)[name = string("op_4309_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_cast_fp16 = mul(x = var_4309_cast_fp16, y = cum_tail_to_fp16)[name = string("N_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4311_cast_fp16 = mul(x = D_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_4311_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_cast_fp16 = mul(x = kk_cast_fp16, y = var_4311_cast_fp16)[name = string("Na_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_89_cast_fp16 = add(x = eye_to_fp16, y = N_cast_fp16)[name = string("Minv_89_cast_fp16")];
            bool Np_67_transpose_x_0 = const()[name = string("Np_67_transpose_x_0"), val = bool(false)];
            bool Np_67_transpose_y_0 = const()[name = string("Np_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_67_cast_fp16 = matmul(transpose_x = Np_67_transpose_x_0, transpose_y = Np_67_transpose_y_0, x = N_cast_fp16, y = Na_cast_fp16)[name = string("Np_67_cast_fp16")];
            bool Npa_45_transpose_x_0 = const()[name = string("Npa_45_transpose_x_0"), val = bool(false)];
            bool Npa_45_transpose_y_0 = const()[name = string("Npa_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_45_cast_fp16 = matmul(transpose_x = Npa_45_transpose_x_0, transpose_y = Npa_45_transpose_y_0, x = Na_cast_fp16, y = N_cast_fp16)[name = string("Npa_45_cast_fp16")];
            bool var_4317_transpose_x_0 = const()[name = string("op_4317_transpose_x_0"), val = bool(false)];
            bool var_4317_transpose_y_0 = const()[name = string("op_4317_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4317_cast_fp16 = matmul(transpose_x = var_4317_transpose_x_0, transpose_y = var_4317_transpose_y_0, x = Minv_89_cast_fp16, y = Np_67_cast_fp16)[name = string("op_4317_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_91_cast_fp16 = add(x = Minv_89_cast_fp16, y = var_4317_cast_fp16)[name = string("Minv_91_cast_fp16")];
            bool Np_69_transpose_x_0 = const()[name = string("Np_69_transpose_x_0"), val = bool(false)];
            bool Np_69_transpose_y_0 = const()[name = string("Np_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_69_cast_fp16 = matmul(transpose_x = Np_69_transpose_x_0, transpose_y = Np_69_transpose_y_0, x = Np_67_cast_fp16, y = Npa_45_cast_fp16)[name = string("Np_69_cast_fp16")];
            bool Npa_transpose_x_0 = const()[name = string("Npa_transpose_x_0"), val = bool(false)];
            bool Npa_transpose_y_0 = const()[name = string("Npa_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_cast_fp16 = matmul(transpose_x = Npa_transpose_x_0, transpose_y = Npa_transpose_y_0, x = Npa_45_cast_fp16, y = Np_67_cast_fp16)[name = string("Npa_cast_fp16")];
            bool var_4322_transpose_x_0 = const()[name = string("op_4322_transpose_x_0"), val = bool(false)];
            bool var_4322_transpose_y_0 = const()[name = string("op_4322_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4322_cast_fp16 = matmul(transpose_x = var_4322_transpose_x_0, transpose_y = var_4322_transpose_y_0, x = Minv_91_cast_fp16, y = Np_69_cast_fp16)[name = string("op_4322_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_93_cast_fp16 = add(x = Minv_91_cast_fp16, y = var_4322_cast_fp16)[name = string("Minv_93_cast_fp16")];
            bool Np_transpose_x_0 = const()[name = string("Np_transpose_x_0"), val = bool(false)];
            bool Np_transpose_y_0 = const()[name = string("Np_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_cast_fp16 = matmul(transpose_x = Np_transpose_x_0, transpose_y = Np_transpose_y_0, x = Np_69_cast_fp16, y = Npa_cast_fp16)[name = string("Np_cast_fp16")];
            bool var_4326_transpose_x_0 = const()[name = string("op_4326_transpose_x_0"), val = bool(false)];
            bool var_4326_transpose_y_0 = const()[name = string("op_4326_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4326_cast_fp16 = matmul(transpose_x = var_4326_transpose_x_0, transpose_y = var_4326_transpose_y_0, x = Minv_93_cast_fp16, y = Np_cast_fp16)[name = string("op_4326_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_cast_fp16 = add(x = Minv_93_cast_fp16, y = var_4326_cast_fp16)[name = string("Minv_cast_fp16")];
            bool u_transpose_x_0 = const()[name = string("u_transpose_x_0"), val = bool(false)];
            bool u_transpose_y_0 = const()[name = string("u_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_cast_fp16 = matmul(transpose_x = u_transpose_x_0, transpose_y = u_transpose_y_0, x = Minv_cast_fp16, y = vb_cast_fp16)[name = string("u_cast_fp16")];
            bool var_4333_transpose_x_1 = const()[name = string("op_4333_transpose_x_1"), val = bool(false)];
            bool var_4333_transpose_y_1 = const()[name = string("op_4333_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_4333_cast_fp16 = matmul(transpose_x = var_4333_transpose_x_1, transpose_y = var_4333_transpose_y_1, x = q_73_cast_fp16, y = k_75_cast_fp16)[name = string("op_4333_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_cast_fp16 = mul(x = var_4333_cast_fp16, y = D_cast_fp16)[name = string("A_cast_fp16")];
            tensor<fp16, [16]> var_4335_bias_0_to_fp16 = const()[name = string("op_4335_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_4335_cast_fp16 = linear(bias = var_4335_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_cast_fp16)[name = string("op_4335_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_cast_fp16 = exp(x = var_4335_cast_fp16)[name = string("cum_cast_fp16")];
            tensor<int32, [1]> var_4338_axes_0 = const()[name = string("op_4338_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_4338_cast_fp16 = expand_dims(axes = var_4338_axes_0, x = cum_cast_fp16)[name = string("op_4338_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4339_cast_fp16 = mul(x = kb_cast_fp16, y = var_4338_cast_fp16)[name = string("op_4339_cast_fp16")];
            bool w_transpose_x_0 = const()[name = string("w_transpose_x_0"), val = bool(false)];
            bool w_transpose_y_0 = const()[name = string("w_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_cast_fp16 = matmul(transpose_x = w_transpose_x_0, transpose_y = w_transpose_y_0, x = Minv_cast_fp16, y = var_4339_cast_fp16)[name = string("w_cast_fp16")];
            bool var_4341_transpose_x_0 = const()[name = string("op_4341_transpose_x_0"), val = bool(false)];
            bool var_4341_transpose_y_0 = const()[name = string("op_4341_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4341_cast_fp16 = matmul(transpose_x = var_4341_transpose_x_0, transpose_y = var_4341_transpose_y_0, x = w_cast_fp16, y = S_cast_fp16)[name = string("op_4341_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_cast_fp16 = sub(x = u_cast_fp16, y = var_4341_cast_fp16)[name = string("vnew_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4346_cast_fp16 = mul(x = q_73_cast_fp16, y = var_4338_cast_fp16)[name = string("op_4346_cast_fp16")];
            bool var_4347_transpose_x_0 = const()[name = string("op_4347_transpose_x_0"), val = bool(false)];
            bool var_4347_transpose_y_0 = const()[name = string("op_4347_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4347_cast_fp16 = matmul(transpose_x = var_4347_transpose_x_0, transpose_y = var_4347_transpose_y_0, x = var_4346_cast_fp16, y = S_cast_fp16)[name = string("op_4347_cast_fp16")];
            bool var_4348_transpose_x_0 = const()[name = string("op_4348_transpose_x_0"), val = bool(false)];
            bool var_4348_transpose_y_0 = const()[name = string("op_4348_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4348_cast_fp16 = matmul(transpose_x = var_4348_transpose_x_0, transpose_y = var_4348_transpose_y_0, x = A_cast_fp16, y = vnew_cast_fp16)[name = string("op_4348_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_cast_fp16 = add(x = var_4347_cast_fp16, y = var_4348_cast_fp16)[name = string("oc_cast_fp16")];
            int32 var_4352 = const()[name = string("op_4352"), val = int32(2)];
            bool o_33_interleave_0 = const()[name = string("o_33_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 32, 128]> o_33_cast_fp16 = concat(axis = var_4352, interleave = o_33_interleave_0, values = (var_4203_cast_fp16, oc_cast_fp16))[name = string("o_33_cast_fp16")];
            tensor<int32, [4]> var_4358 = const()[name = string("op_4358"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_4364 = const()[name = string("op_4364"), val = tensor<int32, [4]>([3, 32, 4, 128])];
            tensor<fp16, [3, 32, 4, 128]> input_97_cast_fp16 = reshape(shape = var_4364, x = z_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> o_35_cast_fp16 = transpose(perm = var_4358, x = o_33_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [3, 32, 4, 128]> var_4366_cast_fp16 = mul(x = o_35_cast_fp16, y = o_35_cast_fp16)[name = string("op_4366_cast_fp16")];
            tensor<int32, [1]> var_4371_axes_0 = const()[name = string("op_4371_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4371_keep_dims_0 = const()[name = string("op_4371_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 4, 1]> var_4371_cast_fp16 = reduce_mean(axes = var_4371_axes_0, keep_dims = var_4371_keep_dims_0, x = var_4366_cast_fp16)[name = string("op_4371_cast_fp16")];
            fp16 var_4373_to_fp16 = const()[name = string("op_4373_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 32, 4, 1]> var_4374_cast_fp16 = add(x = var_4371_cast_fp16, y = var_4373_to_fp16)[name = string("op_4374_cast_fp16")];
            fp32 var_4375_epsilon_0 = const()[name = string("op_4375_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 4, 1]> var_4375_cast_fp16 = rsqrt(epsilon = var_4375_epsilon_0, x = var_4374_cast_fp16)[name = string("op_4375_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> var_4376_cast_fp16 = mul(x = o_35_cast_fp16, y = var_4375_cast_fp16)[name = string("op_4376_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> var_4378_cast_fp16 = silu(x = input_97_cast_fp16)[name = string("op_4378_cast_fp16")];
            tensor<fp16, [3, 32, 4, 128]> o_37_cast_fp16 = mul(x = var_4376_cast_fp16, y = var_4378_cast_fp16)[name = string("o_37_cast_fp16")];
            tensor<int32, [3]> var_4383 = const()[name = string("op_4383"), val = tensor<int32, [3]>([3, 32, 512])];
            tensor<fp16, [3, 32, 512]> input_99_cast_fp16 = reshape(shape = var_4383, x = o_37_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [640, 512]> layers_6_out_weight_to_fp16 = const()[name = string("layers_6_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97009984)))];
            tensor<fp16, [3, 32, 640]> linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_out_weight_to_fp16, x = input_99_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [3, 32, 640]> x_65_cast_fp16 = add(x = x_61_cast_fp16, y = linear_25_cast_fp16)[name = string("x_65_cast_fp16")];
            fp16 var_4396_to_fp16 = const()[name = string("op_4396_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [3, 32, 640]> x_67_cast_fp16 = mul(x = x_65_cast_fp16, y = var_4396_to_fp16)[name = string("x_67_cast_fp16")];
            fp16 var_4394_promoted_to_fp16 = const()[name = string("op_4394_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> var_4398_cast_fp16 = pow(x = x_67_cast_fp16, y = var_4394_promoted_to_fp16)[name = string("op_4398_cast_fp16")];
            tensor<int32, [1]> var_4400_axes_0 = const()[name = string("op_4400_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4400_keep_dims_0 = const()[name = string("op_4400_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 1]> var_4400_cast_fp16 = reduce_mean(axes = var_4400_axes_0, keep_dims = var_4400_keep_dims_0, x = var_4398_cast_fp16)[name = string("op_4400_cast_fp16")];
            fp16 var_4401_to_fp16 = const()[name = string("op_4401_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 32, 1]> var_4402_cast_fp16 = add(x = var_4400_cast_fp16, y = var_4401_to_fp16)[name = string("op_4402_cast_fp16")];
            fp32 var_4403_epsilon_0 = const()[name = string("op_4403_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 1]> var_4403_cast_fp16 = rsqrt(epsilon = var_4403_epsilon_0, x = var_4402_cast_fp16)[name = string("op_4403_cast_fp16")];
            tensor<fp16, [3, 32, 640]> var_4404_cast_fp16 = mul(x = x_67_cast_fp16, y = var_4403_cast_fp16)[name = string("op_4404_cast_fp16")];
            tensor<fp16, [640]> layers_6_n2_w1_to_fp16 = const()[name = string("layers_6_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97665408)))];
            tensor<fp16, [3, 32, 640]> input_101_cast_fp16 = mul(x = var_4404_cast_fp16, y = layers_6_n2_w1_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_6_gate_up_weight_to_fp16 = const()[name = string("layers_6_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97666752)))];
            tensor<fp16, [3, 32, 5120]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_6_gate_up_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<int32, [3]> input_103_begin_0 = const()[name = string("input_103_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_103_end_0 = const()[name = string("input_103_end_0"), val = tensor<int32, [3]>([3, 32, 2560])];
            tensor<bool, [3]> input_103_end_mask_0 = const()[name = string("input_103_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 2560]> input_103_cast_fp16 = slice_by_index(begin = input_103_begin_0, end = input_103_end_0, end_mask = input_103_end_mask_0, x = linear_26_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [3, 32, 2560]> var_4414_cast_fp16 = silu(x = input_103_cast_fp16)[name = string("op_4414_cast_fp16")];
            tensor<int32, [3]> var_4419_begin_0 = const()[name = string("op_4419_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_4419_end_0 = const()[name = string("op_4419_end_0"), val = tensor<int32, [3]>([3, 32, 5120])];
            tensor<bool, [3]> var_4419_end_mask_0 = const()[name = string("op_4419_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 2560]> var_4419_cast_fp16 = slice_by_index(begin = var_4419_begin_0, end = var_4419_end_0, end_mask = var_4419_end_mask_0, x = linear_26_cast_fp16)[name = string("op_4419_cast_fp16")];
            tensor<fp16, [3, 32, 2560]> input_105_cast_fp16 = mul(x = var_4414_cast_fp16, y = var_4419_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_6_down_weight_to_fp16 = const()[name = string("layers_6_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104220416)))];
            tensor<fp16, [3, 32, 640]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_down_weight_to_fp16, x = input_105_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [3, 32, 640]> x_69_cast_fp16 = add(x = x_65_cast_fp16, y = linear_27_cast_fp16)[name = string("x_69_cast_fp16")];
            fp16 var_4432_to_fp16 = const()[name = string("op_4432_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [3, 32, 640]> x_71_cast_fp16 = mul(x = x_69_cast_fp16, y = var_4432_to_fp16)[name = string("x_71_cast_fp16")];
            fp16 var_4430_promoted_to_fp16 = const()[name = string("op_4430_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> var_4434_cast_fp16 = pow(x = x_71_cast_fp16, y = var_4430_promoted_to_fp16)[name = string("op_4434_cast_fp16")];
            tensor<int32, [1]> var_4436_axes_0 = const()[name = string("op_4436_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4436_keep_dims_0 = const()[name = string("op_4436_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 1]> var_4436_cast_fp16 = reduce_mean(axes = var_4436_axes_0, keep_dims = var_4436_keep_dims_0, x = var_4434_cast_fp16)[name = string("op_4436_cast_fp16")];
            fp16 var_4437_to_fp16 = const()[name = string("op_4437_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 32, 1]> var_4438_cast_fp16 = add(x = var_4436_cast_fp16, y = var_4437_to_fp16)[name = string("op_4438_cast_fp16")];
            fp32 var_4439_epsilon_0 = const()[name = string("op_4439_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 1]> var_4439_cast_fp16 = rsqrt(epsilon = var_4439_epsilon_0, x = var_4438_cast_fp16)[name = string("op_4439_cast_fp16")];
            tensor<fp16, [3, 32, 640]> var_4440_cast_fp16 = mul(x = x_71_cast_fp16, y = var_4439_cast_fp16)[name = string("op_4440_cast_fp16")];
            tensor<fp16, [640]> layers_7_n1_w1_to_fp16 = const()[name = string("layers_7_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107497280)))];
            tensor<fp16, [3, 32, 640]> input_107_cast_fp16 = mul(x = var_4440_cast_fp16, y = layers_7_n1_w1_to_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [1280, 640]> layers_7_qkv_weight_to_fp16 = const()[name = string("layers_7_qkv_weight_to_fp16"), val = tensor<fp16, [1280, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107498624)))];
            tensor<fp16, [3, 32, 1280]> linear_28_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = layers_7_qkv_weight_to_fp16, x = input_107_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [3]> var_4449_begin_0 = const()[name = string("op_4449_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_4449_end_0 = const()[name = string("op_4449_end_0"), val = tensor<int32, [3]>([3, 32, 1024])];
            tensor<bool, [3]> var_4449_end_mask_0 = const()[name = string("op_4449_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 1024]> var_4449_cast_fp16 = slice_by_index(begin = var_4449_begin_0, end = var_4449_end_0, end_mask = var_4449_end_mask_0, x = linear_28_cast_fp16)[name = string("op_4449_cast_fp16")];
            tensor<int32, [4]> var_4454 = const()[name = string("op_4454"), val = tensor<int32, [4]>([3, 32, 8, 128])];
            tensor<fp16, [3, 32, 8, 128]> qg_cast_fp16 = reshape(shape = var_4454, x = var_4449_cast_fp16)[name = string("qg_cast_fp16")];
            tensor<int32, [4]> x_73_begin_0 = const()[name = string("x_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_73_end_0 = const()[name = string("x_73_end_0"), val = tensor<int32, [4]>([3, 32, 8, 64])];
            tensor<bool, [4]> x_73_end_mask_0 = const()[name = string("x_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 32, 8, 64]> x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, x = qg_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<int32, [4]> var_4465_begin_0 = const()[name = string("op_4465_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> var_4465_end_0 = const()[name = string("op_4465_end_0"), val = tensor<int32, [4]>([3, 32, 8, 128])];
            tensor<bool, [4]> var_4465_end_mask_0 = const()[name = string("op_4465_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 32, 8, 64]> var_4465_cast_fp16 = slice_by_index(begin = var_4465_begin_0, end = var_4465_end_0, end_mask = var_4465_end_mask_0, x = qg_cast_fp16)[name = string("op_4465_cast_fp16")];
            tensor<int32, [3]> var_4469 = const()[name = string("op_4469"), val = tensor<int32, [3]>([3, 32, 512])];
            tensor<fp16, [3, 32, 512]> gate_cast_fp16 = reshape(shape = var_4469, x = var_4465_cast_fp16)[name = string("gate_cast_fp16")];
            tensor<int32, [3]> var_4475_begin_0 = const()[name = string("op_4475_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> var_4475_end_0 = const()[name = string("op_4475_end_0"), val = tensor<int32, [3]>([3, 32, 1152])];
            tensor<bool, [3]> var_4475_end_mask_0 = const()[name = string("op_4475_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 128]> var_4475_cast_fp16 = slice_by_index(begin = var_4475_begin_0, end = var_4475_end_0, end_mask = var_4475_end_mask_0, x = linear_28_cast_fp16)[name = string("op_4475_cast_fp16")];
            tensor<int32, [4]> var_4480 = const()[name = string("op_4480"), val = tensor<int32, [4]>([3, 32, 2, 64])];
            tensor<fp16, [3, 32, 2, 64]> x_79_cast_fp16 = reshape(shape = var_4480, x = var_4475_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<int32, [3]> var_4486_begin_0 = const()[name = string("op_4486_begin_0"), val = tensor<int32, [3]>([0, 0, 1152])];
            tensor<int32, [3]> var_4486_end_0 = const()[name = string("op_4486_end_0"), val = tensor<int32, [3]>([3, 32, 1280])];
            tensor<bool, [3]> var_4486_end_mask_0 = const()[name = string("op_4486_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 128]> var_4486_cast_fp16 = slice_by_index(begin = var_4486_begin_0, end = var_4486_end_0, end_mask = var_4486_end_mask_0, x = linear_28_cast_fp16)[name = string("op_4486_cast_fp16")];
            tensor<int32, [4]> var_4491 = const()[name = string("op_4491"), val = tensor<int32, [4]>([3, 32, 2, 64])];
            tensor<fp16, [3, 32, 2, 64]> var_4492_cast_fp16 = reshape(shape = var_4491, x = var_4486_cast_fp16)[name = string("op_4492_cast_fp16")];
            tensor<int32, [4]> var_4497 = const()[name = string("op_4497"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4503_promoted_to_fp16 = const()[name = string("op_4503_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 8, 64]> var_4507_cast_fp16 = pow(x = x_73_cast_fp16, y = var_4503_promoted_to_fp16)[name = string("op_4507_cast_fp16")];
            tensor<int32, [1]> var_4509_axes_0 = const()[name = string("op_4509_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4509_keep_dims_0 = const()[name = string("op_4509_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 8, 1]> var_4509_cast_fp16 = reduce_mean(axes = var_4509_axes_0, keep_dims = var_4509_keep_dims_0, x = var_4507_cast_fp16)[name = string("op_4509_cast_fp16")];
            fp16 var_4510_to_fp16 = const()[name = string("op_4510_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 32, 8, 1]> var_4511_cast_fp16 = add(x = var_4509_cast_fp16, y = var_4510_to_fp16)[name = string("op_4511_cast_fp16")];
            fp32 var_4512_epsilon_0 = const()[name = string("op_4512_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 8, 1]> var_4512_cast_fp16 = rsqrt(epsilon = var_4512_epsilon_0, x = var_4511_cast_fp16)[name = string("op_4512_cast_fp16")];
            tensor<fp16, [3, 32, 8, 64]> var_4513_cast_fp16 = mul(x = x_73_cast_fp16, y = var_4512_cast_fp16)[name = string("op_4513_cast_fp16")];
            tensor<fp16, [64]> layers_7_qn_w1_to_fp16 = const()[name = string("layers_7_qn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137088)))];
            tensor<fp16, [3, 32, 8, 64]> var_4514_cast_fp16 = mul(x = var_4513_cast_fp16, y = layers_7_qn_w1_to_fp16)[name = string("op_4514_cast_fp16")];
            tensor<int32, [4]> var_4519 = const()[name = string("op_4519"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_5_begin_0 = const()[name = string("r_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_5_end_0 = const()[name = string("r_5_end_0"), val = tensor<int32, [4]>([3, 8, 32, 16])];
            tensor<bool, [4]> r_5_end_mask_0 = const()[name = string("r_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 8, 32, 64]> x_77_cast_fp16 = transpose(perm = var_4519, x = var_4514_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [3, 8, 32, 16]> r_5_cast_fp16 = slice_by_index(begin = r_5_begin_0, end = r_5_end_0, end_mask = r_5_end_mask_0, x = x_77_cast_fp16)[name = string("r_5_cast_fp16")];
            tensor<int32, [4]> p_5_begin_0 = const()[name = string("p_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_5_end_0 = const()[name = string("p_5_end_0"), val = tensor<int32, [4]>([3, 8, 32, 64])];
            tensor<bool, [4]> p_5_end_mask_0 = const()[name = string("p_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 8, 32, 48]> p_5_cast_fp16 = slice_by_index(begin = p_5_begin_0, end = p_5_end_0, end_mask = p_5_end_mask_0, x = x_77_cast_fp16)[name = string("p_5_cast_fp16")];
            tensor<int32, [4]> var_4535_begin_0 = const()[name = string("op_4535_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_4535_end_0 = const()[name = string("op_4535_end_0"), val = tensor<int32, [4]>([3, 8, 32, 16])];
            tensor<bool, [4]> var_4535_end_mask_0 = const()[name = string("op_4535_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 8, 32, 8]> var_4535_cast_fp16 = slice_by_index(begin = var_4535_begin_0, end = var_4535_end_0, end_mask = var_4535_end_mask_0, x = r_5_cast_fp16)[name = string("op_4535_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 8, 32, 8]> var_4536_cast_fp16 = mul(x = var_4535_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_4536_cast_fp16")];
            tensor<int32, [4]> var_4541_begin_0 = const()[name = string("op_4541_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4541_end_0 = const()[name = string("op_4541_end_0"), val = tensor<int32, [4]>([3, 8, 32, 8])];
            tensor<bool, [4]> var_4541_end_mask_0 = const()[name = string("op_4541_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 8, 32, 8]> var_4541_cast_fp16 = slice_by_index(begin = var_4541_begin_0, end = var_4541_end_0, end_mask = var_4541_end_mask_0, x = r_5_cast_fp16)[name = string("op_4541_cast_fp16")];
            int32 var_4543 = const()[name = string("op_4543"), val = int32(-1)];
            bool rh_5_interleave_0 = const()[name = string("rh_5_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 8, 32, 16]> rh_5_cast_fp16 = concat(axis = var_4543, interleave = rh_5_interleave_0, values = (var_4536_cast_fp16, var_4541_cast_fp16))[name = string("rh_5_cast_fp16")];
            tensor<fp16, [3, 8, 32, 16]> var_4545_cast_fp16 = mul(x = r_5_cast_fp16, y = cos_to_fp16)[name = string("op_4545_cast_fp16")];
            tensor<fp16, [3, 8, 32, 16]> var_4546_cast_fp16 = mul(x = rh_5_cast_fp16, y = sin_to_fp16)[name = string("op_4546_cast_fp16")];
            tensor<fp16, [3, 8, 32, 16]> var_4548_cast_fp16 = add(x = var_4545_cast_fp16, y = var_4546_cast_fp16)[name = string("op_4548_cast_fp16")];
            int32 var_4550 = const()[name = string("op_4550"), val = int32(-1)];
            bool q_interleave_0 = const()[name = string("q_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 8, 32, 64]> q_cast_fp16 = concat(axis = var_4550, interleave = q_interleave_0, values = (var_4548_cast_fp16, p_5_cast_fp16))[name = string("q_cast_fp16")];
            fp16 var_4556_promoted_to_fp16 = const()[name = string("op_4556_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 2, 64]> var_4560_cast_fp16 = pow(x = x_79_cast_fp16, y = var_4556_promoted_to_fp16)[name = string("op_4560_cast_fp16")];
            tensor<int32, [1]> var_4562_axes_0 = const()[name = string("op_4562_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4562_keep_dims_0 = const()[name = string("op_4562_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 2, 1]> var_4562_cast_fp16 = reduce_mean(axes = var_4562_axes_0, keep_dims = var_4562_keep_dims_0, x = var_4560_cast_fp16)[name = string("op_4562_cast_fp16")];
            fp16 var_4563_to_fp16 = const()[name = string("op_4563_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 32, 2, 1]> var_4564_cast_fp16 = add(x = var_4562_cast_fp16, y = var_4563_to_fp16)[name = string("op_4564_cast_fp16")];
            fp32 var_4565_epsilon_0 = const()[name = string("op_4565_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 2, 1]> var_4565_cast_fp16 = rsqrt(epsilon = var_4565_epsilon_0, x = var_4564_cast_fp16)[name = string("op_4565_cast_fp16")];
            tensor<fp16, [3, 32, 2, 64]> var_4566_cast_fp16 = mul(x = x_79_cast_fp16, y = var_4565_cast_fp16)[name = string("op_4566_cast_fp16")];
            tensor<fp16, [64]> layers_7_kn_w1_to_fp16 = const()[name = string("layers_7_kn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137280)))];
            tensor<fp16, [3, 32, 2, 64]> var_4567_cast_fp16 = mul(x = var_4566_cast_fp16, y = layers_7_kn_w1_to_fp16)[name = string("op_4567_cast_fp16")];
            tensor<int32, [4]> var_4572 = const()[name = string("op_4572"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_begin_0 = const()[name = string("r_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_end_0 = const()[name = string("r_end_0"), val = tensor<int32, [4]>([3, 2, 32, 16])];
            tensor<bool, [4]> r_end_mask_0 = const()[name = string("r_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 2, 32, 64]> x_83_cast_fp16 = transpose(perm = var_4572, x = var_4567_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [3, 2, 32, 16]> r_cast_fp16 = slice_by_index(begin = r_begin_0, end = r_end_0, end_mask = r_end_mask_0, x = x_83_cast_fp16)[name = string("r_cast_fp16")];
            tensor<int32, [4]> p_begin_0 = const()[name = string("p_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_end_0 = const()[name = string("p_end_0"), val = tensor<int32, [4]>([3, 2, 32, 64])];
            tensor<bool, [4]> p_end_mask_0 = const()[name = string("p_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 2, 32, 48]> p_cast_fp16 = slice_by_index(begin = p_begin_0, end = p_end_0, end_mask = p_end_mask_0, x = x_83_cast_fp16)[name = string("p_cast_fp16")];
            tensor<int32, [4]> var_4588_begin_0 = const()[name = string("op_4588_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_4588_end_0 = const()[name = string("op_4588_end_0"), val = tensor<int32, [4]>([3, 2, 32, 16])];
            tensor<bool, [4]> var_4588_end_mask_0 = const()[name = string("op_4588_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 2, 32, 8]> var_4588_cast_fp16 = slice_by_index(begin = var_4588_begin_0, end = var_4588_end_0, end_mask = var_4588_end_mask_0, x = r_cast_fp16)[name = string("op_4588_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 2, 32, 8]> var_4589_cast_fp16 = mul(x = var_4588_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_4589_cast_fp16")];
            tensor<int32, [4]> var_4594_begin_0 = const()[name = string("op_4594_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4594_end_0 = const()[name = string("op_4594_end_0"), val = tensor<int32, [4]>([3, 2, 32, 8])];
            tensor<bool, [4]> var_4594_end_mask_0 = const()[name = string("op_4594_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 2, 32, 8]> var_4594_cast_fp16 = slice_by_index(begin = var_4594_begin_0, end = var_4594_end_0, end_mask = var_4594_end_mask_0, x = r_cast_fp16)[name = string("op_4594_cast_fp16")];
            int32 var_4596 = const()[name = string("op_4596"), val = int32(-1)];
            bool rh_interleave_0 = const()[name = string("rh_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 2, 32, 16]> rh_cast_fp16 = concat(axis = var_4596, interleave = rh_interleave_0, values = (var_4589_cast_fp16, var_4594_cast_fp16))[name = string("rh_cast_fp16")];
            tensor<fp16, [3, 2, 32, 16]> var_4598_cast_fp16 = mul(x = r_cast_fp16, y = cos_to_fp16)[name = string("op_4598_cast_fp16")];
            tensor<fp16, [3, 2, 32, 16]> var_4599_cast_fp16 = mul(x = rh_cast_fp16, y = sin_to_fp16)[name = string("op_4599_cast_fp16")];
            tensor<fp16, [3, 2, 32, 16]> var_4601_cast_fp16 = add(x = var_4598_cast_fp16, y = var_4599_cast_fp16)[name = string("op_4601_cast_fp16")];
            int32 var_4603 = const()[name = string("op_4603"), val = int32(-1)];
            bool k_77_interleave_0 = const()[name = string("k_77_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 2, 32, 64]> k_77_cast_fp16 = concat(axis = var_4603, interleave = k_77_interleave_0, values = (var_4601_cast_fp16, p_cast_fp16))[name = string("k_77_cast_fp16")];
            tensor<int32, [1]> var_4616_axes_0 = const()[name = string("op_4616_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 1, 32, 64]> var_4616_cast_fp16 = expand_dims(axes = var_4616_axes_0, x = k_77_cast_fp16)[name = string("op_4616_cast_fp16")];
            tensor<int32, [5]> var_4624_reps_0 = const()[name = string("op_4624_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [3, 2, 4, 32, 64]> var_4624_cast_fp16 = tile(reps = var_4624_reps_0, x = var_4616_cast_fp16)[name = string("op_4624_cast_fp16")];
            tensor<int32, [4]> var_4629 = const()[name = string("op_4629"), val = tensor<int32, [4]>([3, 8, 32, 64])];
            tensor<fp16, [3, 8, 32, 64]> k_cast_fp16 = reshape(shape = var_4629, x = var_4624_cast_fp16)[name = string("k_cast_fp16")];
            tensor<int32, [1]> var_4642_axes_0 = const()[name = string("op_4642_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 32, 64]> v_17_cast_fp16 = transpose(perm = var_4497, x = var_4492_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [3, 2, 1, 32, 64]> var_4642_cast_fp16 = expand_dims(axes = var_4642_axes_0, x = v_17_cast_fp16)[name = string("op_4642_cast_fp16")];
            tensor<int32, [5]> var_4650_reps_0 = const()[name = string("op_4650_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [3, 2, 4, 32, 64]> var_4650_cast_fp16 = tile(reps = var_4650_reps_0, x = var_4642_cast_fp16)[name = string("op_4650_cast_fp16")];
            tensor<int32, [4]> var_4655 = const()[name = string("op_4655"), val = tensor<int32, [4]>([3, 8, 32, 64])];
            tensor<fp16, [3, 8, 32, 64]> v_cast_fp16 = reshape(shape = var_4655, x = var_4650_cast_fp16)[name = string("v_cast_fp16")];
            bool var_4660_transpose_x_1 = const()[name = string("op_4660_transpose_x_1"), val = bool(false)];
            bool var_4660_transpose_y_1 = const()[name = string("op_4660_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 8, 32, 32]> var_4660_cast_fp16 = matmul(transpose_x = var_4660_transpose_x_1, transpose_y = var_4660_transpose_y_1, x = q_cast_fp16, y = k_cast_fp16)[name = string("op_4660_cast_fp16")];
            fp16 var_4661_to_fp16 = const()[name = string("op_4661_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [3, 8, 32, 32]> var_4662_cast_fp16 = mul(x = var_4660_cast_fp16, y = var_4661_to_fp16)[name = string("op_4662_cast_fp16")];
            tensor<fp16, [3, 8, 32, 32]> s_cast_fp16 = add(x = var_4662_cast_fp16, y = causal_add_to_fp16)[name = string("s_cast_fp16")];
            int32 var_4665 = const()[name = string("op_4665"), val = int32(-1)];
            tensor<fp16, [3, 8, 32, 32]> var_4667_cast_fp16 = softmax(axis = var_4665, x = s_cast_fp16)[name = string("op_4667_cast_fp16")];
            bool o_39_transpose_x_0 = const()[name = string("o_39_transpose_x_0"), val = bool(false)];
            bool o_39_transpose_y_0 = const()[name = string("o_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 8, 32, 64]> o_39_cast_fp16 = matmul(transpose_x = o_39_transpose_x_0, transpose_y = o_39_transpose_y_0, x = var_4667_cast_fp16, y = v_cast_fp16)[name = string("o_39_cast_fp16")];
            tensor<int32, [4]> var_4673 = const()[name = string("op_4673"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4678 = const()[name = string("op_4678"), val = tensor<int32, [3]>([3, 32, 512])];
            tensor<fp16, [3, 32, 8, 64]> var_4674_cast_fp16 = transpose(perm = var_4673, x = o_39_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [3, 32, 512]> var_4679_cast_fp16 = reshape(shape = var_4678, x = var_4674_cast_fp16)[name = string("op_4679_cast_fp16")];
            tensor<fp16, [3, 32, 512]> var_4680_cast_fp16 = sigmoid(x = gate_cast_fp16)[name = string("op_4680_cast_fp16")];
            tensor<fp16, [3, 32, 512]> input_109_cast_fp16 = mul(x = var_4679_cast_fp16, y = var_4680_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [640, 512]> layers_7_o_weight_to_fp16 = const()[name = string("layers_7_o_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137472)))];
            tensor<fp16, [3, 32, 640]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_o_weight_to_fp16, x = input_109_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<fp16, [3, 32, 640]> x_85_cast_fp16 = add(x = x_69_cast_fp16, y = linear_29_cast_fp16)[name = string("x_85_cast_fp16")];
            fp16 var_4693_to_fp16 = const()[name = string("op_4693_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [3, 32, 640]> x_87_cast_fp16 = mul(x = x_85_cast_fp16, y = var_4693_to_fp16)[name = string("x_87_cast_fp16")];
            fp16 var_4691_promoted_to_fp16 = const()[name = string("op_4691_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> var_4695_cast_fp16 = pow(x = x_87_cast_fp16, y = var_4691_promoted_to_fp16)[name = string("op_4695_cast_fp16")];
            tensor<int32, [1]> var_4697_axes_0 = const()[name = string("op_4697_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4697_keep_dims_0 = const()[name = string("op_4697_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 1]> var_4697_cast_fp16 = reduce_mean(axes = var_4697_axes_0, keep_dims = var_4697_keep_dims_0, x = var_4695_cast_fp16)[name = string("op_4697_cast_fp16")];
            fp16 var_4698_to_fp16 = const()[name = string("op_4698_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 32, 1]> var_4699_cast_fp16 = add(x = var_4697_cast_fp16, y = var_4698_to_fp16)[name = string("op_4699_cast_fp16")];
            fp32 var_4700_epsilon_0 = const()[name = string("op_4700_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 1]> var_4700_cast_fp16 = rsqrt(epsilon = var_4700_epsilon_0, x = var_4699_cast_fp16)[name = string("op_4700_cast_fp16")];
            tensor<fp16, [3, 32, 640]> var_4701_cast_fp16 = mul(x = x_87_cast_fp16, y = var_4700_cast_fp16)[name = string("op_4701_cast_fp16")];
            tensor<fp16, [640]> layers_7_n2_w1_to_fp16 = const()[name = string("layers_7_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109792896)))];
            tensor<fp16, [3, 32, 640]> input_111_cast_fp16 = mul(x = var_4701_cast_fp16, y = layers_7_n2_w1_to_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_7_gate_up_weight_to_fp16 = const()[name = string("layers_7_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109794240)))];
            tensor<fp16, [3, 32, 5120]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_7_gate_up_weight_to_fp16, x = input_111_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [3]> input_113_begin_0 = const()[name = string("input_113_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_113_end_0 = const()[name = string("input_113_end_0"), val = tensor<int32, [3]>([3, 32, 2560])];
            tensor<bool, [3]> input_113_end_mask_0 = const()[name = string("input_113_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 32, 2560]> input_113_cast_fp16 = slice_by_index(begin = input_113_begin_0, end = input_113_end_0, end_mask = input_113_end_mask_0, x = linear_30_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [3, 32, 2560]> var_4711_cast_fp16 = silu(x = input_113_cast_fp16)[name = string("op_4711_cast_fp16")];
            tensor<int32, [3]> var_4716_begin_0 = const()[name = string("op_4716_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_4716_end_0 = const()[name = string("op_4716_end_0"), val = tensor<int32, [3]>([3, 32, 5120])];
            tensor<bool, [3]> var_4716_end_mask_0 = const()[name = string("op_4716_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 32, 2560]> var_4716_cast_fp16 = slice_by_index(begin = var_4716_begin_0, end = var_4716_end_0, end_mask = var_4716_end_mask_0, x = linear_30_cast_fp16)[name = string("op_4716_cast_fp16")];
            tensor<fp16, [3, 32, 2560]> input_115_cast_fp16 = mul(x = var_4711_cast_fp16, y = var_4716_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_7_down_weight_to_fp16 = const()[name = string("layers_7_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116347904)))];
            tensor<fp16, [3, 32, 640]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_down_weight_to_fp16, x = input_115_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [3, 32, 640]> x_89_cast_fp16 = add(x = x_85_cast_fp16, y = linear_31_cast_fp16)[name = string("x_89_cast_fp16")];
            fp16 var_4729_to_fp16 = const()[name = string("op_4729_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [3, 32, 640]> x_cast_fp16 = mul(x = x_89_cast_fp16, y = var_4729_to_fp16)[name = string("x_cast_fp16")];
            fp16 var_4727_promoted_to_fp16 = const()[name = string("op_4727_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 32, 640]> var_4731_cast_fp16 = pow(x = x_cast_fp16, y = var_4727_promoted_to_fp16)[name = string("op_4731_cast_fp16")];
            tensor<int32, [1]> var_4733_axes_0 = const()[name = string("op_4733_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4733_keep_dims_0 = const()[name = string("op_4733_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 32, 1]> var_4733_cast_fp16 = reduce_mean(axes = var_4733_axes_0, keep_dims = var_4733_keep_dims_0, x = var_4731_cast_fp16)[name = string("op_4733_cast_fp16")];
            fp16 var_4734_to_fp16 = const()[name = string("op_4734_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 32, 1]> var_4735_cast_fp16 = add(x = var_4733_cast_fp16, y = var_4734_to_fp16)[name = string("op_4735_cast_fp16")];
            fp32 var_4736_epsilon_0 = const()[name = string("op_4736_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 32, 1]> var_4736_cast_fp16 = rsqrt(epsilon = var_4736_epsilon_0, x = var_4735_cast_fp16)[name = string("op_4736_cast_fp16")];
            tensor<fp16, [3, 32, 640]> var_4737_cast_fp16 = mul(x = x_cast_fp16, y = var_4736_cast_fp16)[name = string("op_4737_cast_fp16")];
            tensor<fp16, [640]> norm_w1_to_fp16 = const()[name = string("norm_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119624768)))];
            tensor<fp16, [3, 32, 640]> var_4738_cast_fp16 = mul(x = var_4737_cast_fp16, y = norm_w1_to_fp16)[name = string("op_4738_cast_fp16")];
            tensor<int32, [3]> h_begin_0 = const()[name = string("h_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> h_end_0 = const()[name = string("h_end_0"), val = tensor<int32, [3]>([3, 31, 640])];
            tensor<bool, [3]> h_end_mask_0 = const()[name = string("h_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 31, 640]> h_cast_fp16 = slice_by_index(begin = h_begin_0, end = h_end_0, end_mask = h_end_mask_0, x = var_4738_cast_fp16)[name = string("h_cast_fp16")];
            tensor<fp16, [16000]> logits_bias_0_to_fp16 = const()[name = string("logits_bias_0_to_fp16"), val = tensor<fp16, [16000]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119626112)))];
            tensor<fp16, [3, 31, 16000]> logits_cast_fp16 = linear(bias = logits_bias_0_to_fp16, weight = emb_to_fp16, x = h_cast_fp16)[name = string("logits_cast_fp16")];
            tensor<int32, [1]> reduce_max_0_axes_0 = const()[name = string("reduce_max_0_axes_0"), val = tensor<int32, [1]>([-1])];
            bool reduce_max_0_keep_dims_0 = const()[name = string("reduce_max_0_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 31, 1]> reduce_max_0_cast_fp16 = reduce_max(axes = reduce_max_0_axes_0, keep_dims = reduce_max_0_keep_dims_0, x = logits_cast_fp16)[name = string("reduce_max_0_cast_fp16")];
            tensor<fp16, [3, 31, 16000]> var_4756_cast_fp16 = sub(x = logits_cast_fp16, y = reduce_max_0_cast_fp16)[name = string("op_4756_cast_fp16")];
            tensor<fp16, [3, 31, 16000]> var_4757_cast_fp16 = exp(x = var_4756_cast_fp16)[name = string("op_4757_cast_fp16")];
            tensor<int32, [1]> var_4762_axes_0 = const()[name = string("op_4762_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4762_keep_dims_0 = const()[name = string("op_4762_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 31, 1]> var_4762_cast_fp16 = reduce_sum(axes = var_4762_axes_0, keep_dims = var_4762_keep_dims_0, x = var_4757_cast_fp16)[name = string("op_4762_cast_fp16")];
            fp32 var_4763_epsilon_0 = const()[name = string("op_4763_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 31, 1]> var_4763_cast_fp16 = log(epsilon = var_4763_epsilon_0, x = var_4762_cast_fp16)[name = string("op_4763_cast_fp16")];
            tensor<fp16, [3, 31, 1]> var_4765_cast_fp16 = add(x = reduce_max_0_cast_fp16, y = var_4763_cast_fp16)[name = string("op_4765_cast_fp16")];
            tensor<int32, [3]> lse_begin_0 = const()[name = string("lse_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> lse_end_0 = const()[name = string("lse_end_0"), val = tensor<int32, [3]>([3, 31, 1])];
            tensor<bool, [3]> lse_end_mask_0 = const()[name = string("lse_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> lse_squeeze_mask_0 = const()[name = string("lse_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<fp16, [3, 31]> lse_cast_fp16 = slice_by_index(begin = lse_begin_0, end = lse_end_0, end_mask = lse_end_mask_0, squeeze_mask = lse_squeeze_mask_0, x = var_4765_cast_fp16)[name = string("lse_cast_fp16")];
            tensor<int32, [2]> input_begin_0 = const()[name = string("input_begin_0"), val = tensor<int32, [2]>([0, 1])];
            tensor<int32, [2]> input_end_0 = const()[name = string("input_end_0"), val = tensor<int32, [2]>([3, 32])];
            tensor<bool, [2]> input_end_mask_0 = const()[name = string("input_end_mask_0"), val = tensor<bool, [2]>([true, true])];
            tensor<int32, [3, 31]> input = slice_by_index(begin = input_begin_0, end = input_end_0, end_mask = input_end_mask_0, x = ids)[name = string("input")];
            int32 var_4782_axis_0 = const()[name = string("op_4782_axis_0"), val = int32(0)];
            int32 var_4782_batch_dims_0 = const()[name = string("op_4782_batch_dims_0"), val = int32(0)];
            bool var_4782_validate_indices_0 = const()[name = string("op_4782_validate_indices_0"), val = bool(false)];
            string input_to_uint16_dtype_0 = const()[name = string("input_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [3, 31]> input_to_uint16 = cast(dtype = input_to_uint16_dtype_0, x = input)[name = string("cast_84")];
            tensor<fp16, [3, 31, 640]> var_4782_cast_fp16_cast_uint16 = gather(axis = var_4782_axis_0, batch_dims = var_4782_batch_dims_0, indices = input_to_uint16, validate_indices = var_4782_validate_indices_0, x = emb_to_fp16)[name = string("op_4782_cast_fp16_cast_uint16")];
            tensor<fp16, [3, 31, 640]> var_4783_cast_fp16 = mul(x = h_cast_fp16, y = var_4782_cast_fp16_cast_uint16)[name = string("op_4783_cast_fp16")];
            tensor<int32, [1]> tgt_axes_0 = const()[name = string("tgt_axes_0"), val = tensor<int32, [1]>([-1])];
            bool tgt_keep_dims_0 = const()[name = string("tgt_keep_dims_0"), val = bool(false)];
            tensor<fp16, [3, 31]> tgt_cast_fp16 = reduce_sum(axes = tgt_axes_0, keep_dims = tgt_keep_dims_0, x = var_4783_cast_fp16)[name = string("tgt_cast_fp16")];
            tensor<fp16, [3, 31]> var_4790_cast_fp16 = sub(x = tgt_cast_fp16, y = lse_cast_fp16)[name = string("op_4790_cast_fp16")];
            string var_4790_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_4790_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [3, 31]> lp = cast(dtype = var_4790_cast_fp16_to_fp32_dtype_0, x = var_4790_cast_fp16)[name = string("cast_83")];
        } -> (lp);
    func t64<ios18>(tensor<int32, [3, 64]> ids) {
            int32 x_1_batch_dims_0 = const()[name = string("x_1_batch_dims_0"), val = int32(0)];
            bool x_1_validate_indices_0 = const()[name = string("x_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [16000, 640]> emb_to_fp16 = const()[name = string("emb_to_fp16"), val = tensor<fp16, [16000, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            string ids_to_int16_dtype_0 = const()[name = string("ids_to_int16_dtype_0"), val = string("int16")];
            string cast_106_dtype_0 = const()[name = string("cast_106_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int16, [3, 64]> ids_to_int16 = cast(dtype = ids_to_int16_dtype_0, x = ids)[name = string("cast_111")];
            tensor<int32, [3, 64]> cast_106 = cast(dtype = cast_106_dtype_0, x = ids_to_int16)[name = string("cast_110")];
            tensor<bool, [3, 64]> greater_equal_0 = greater_equal(x = cast_106, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(16000)];
            tensor<int32, [3, 64]> add_0 = add(x = cast_106, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [3, 64]> select_0 = select(a = cast_106, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 x_1_cast_fp16_cast_uint16_axis_0 = const()[name = string("x_1_cast_fp16_cast_uint16_axis_0"), val = int32(0)];
            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
            tensor<int16, [3, 64]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_109")];
            tensor<fp16, [3, 64, 640]> x_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = x_1_cast_fp16_cast_uint16_axis_0, batch_dims = x_1_batch_dims_0, indices = select_0_to_int16, validate_indices = x_1_validate_indices_0, x = emb_to_fp16)[name = string("x_1_cast_fp16_cast_uint16_cast_uint16")];
            fp16 var_254_to_fp16 = const()[name = string("op_254_to_fp16"), val = fp16(0x1p+5)];
            tensor<fp16, [3, 64, 640]> x_3_cast_fp16 = mul(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = var_254_to_fp16)[name = string("x_3_cast_fp16")];
            fp16 var_252_promoted_to_fp16 = const()[name = string("op_252_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> var_256_cast_fp16 = pow(x = x_3_cast_fp16, y = var_252_promoted_to_fp16)[name = string("op_256_cast_fp16")];
            tensor<int32, [1]> var_258_axes_0 = const()[name = string("op_258_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_258_keep_dims_0 = const()[name = string("op_258_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 1]> var_258_cast_fp16 = reduce_mean(axes = var_258_axes_0, keep_dims = var_258_keep_dims_0, x = var_256_cast_fp16)[name = string("op_258_cast_fp16")];
            fp16 var_259_to_fp16 = const()[name = string("op_259_to_fp16"), val = fp16(0x1.0c8p-10)];
            tensor<fp16, [3, 64, 1]> var_260_cast_fp16 = add(x = var_258_cast_fp16, y = var_259_to_fp16)[name = string("op_260_cast_fp16")];
            fp32 var_261_epsilon_0 = const()[name = string("op_261_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 1]> var_261_cast_fp16 = rsqrt(epsilon = var_261_epsilon_0, x = var_260_cast_fp16)[name = string("op_261_cast_fp16")];
            tensor<fp16, [3, 64, 640]> var_262_cast_fp16 = mul(x = x_3_cast_fp16, y = var_261_cast_fp16)[name = string("op_262_cast_fp16")];
            tensor<fp16, [640]> layers_0_n1_w1_to_fp16 = const()[name = string("layers_0_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20480128)))];
            tensor<fp16, [3, 64, 640]> input_1_cast_fp16 = mul(x = var_262_cast_fp16, y = layers_0_n1_w1_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_0_qkvzba_weight_to_fp16 = const()[name = string("layers_0_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20481472)))];
            tensor<fp16, [1544]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [1544]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22457856)))];
            tensor<fp16, [3, 64, 1544]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_qkvzba_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [3]> qkv_1_begin_0 = const()[name = string("qkv_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_1_end_0 = const()[name = string("qkv_1_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> qkv_1_end_mask_0 = const()[name = string("qkv_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 1024]> qkv_1_cast_fp16 = slice_by_index(begin = qkv_1_begin_0, end = qkv_1_end_0, end_mask = qkv_1_end_mask_0, x = linear_0_cast_fp16)[name = string("qkv_1_cast_fp16")];
            tensor<int32, [3]> z_1_begin_0 = const()[name = string("z_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_1_end_0 = const()[name = string("z_1_end_0"), val = tensor<int32, [3]>([3, 64, 1536])];
            tensor<bool, [3]> z_1_end_mask_0 = const()[name = string("z_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 512]> z_1_cast_fp16 = slice_by_index(begin = z_1_begin_0, end = z_1_end_0, end_mask = z_1_end_mask_0, x = linear_0_cast_fp16)[name = string("z_1_cast_fp16")];
            tensor<int32, [3]> b_1_begin_0 = const()[name = string("b_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_1_end_0 = const()[name = string("b_1_end_0"), val = tensor<int32, [3]>([3, 64, 1540])];
            tensor<bool, [3]> b_1_end_mask_0 = const()[name = string("b_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 4]> b_1_cast_fp16 = slice_by_index(begin = b_1_begin_0, end = b_1_end_0, end_mask = b_1_end_mask_0, x = linear_0_cast_fp16)[name = string("b_1_cast_fp16")];
            tensor<int32, [3]> a_1_begin_0 = const()[name = string("a_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_1_end_0 = const()[name = string("a_1_end_0"), val = tensor<int32, [3]>([3, 64, 1])];
            tensor<bool, [3]> a_1_end_mask_0 = const()[name = string("a_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 4]> a_1_cast_fp16 = slice_by_index(begin = a_1_begin_0, end = a_1_end_0, end_mask = a_1_end_mask_0, x = linear_0_cast_fp16)[name = string("a_1_cast_fp16")];
            int32 var_301 = const()[name = string("op_301"), val = int32(1)];
            bool xp_1_interleave_0 = const()[name = string("xp_1_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 3, 1024]> zpad_to_fp16 = const()[name = string("zpad_to_fp16"), val = tensor<fp16, [3, 3, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22461056)))];
            tensor<fp16, [3, 67, 1024]> xp_1_cast_fp16 = concat(axis = var_301, interleave = xp_1_interleave_0, values = (zpad_to_fp16, qkv_1_cast_fp16))[name = string("xp_1_cast_fp16")];
            tensor<int32, [3]> var_312_begin_0 = const()[name = string("op_312_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_312_end_0 = const()[name = string("op_312_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_312_end_mask_0 = const()[name = string("op_312_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 1024]> var_312_cast_fp16 = slice_by_index(begin = var_312_begin_0, end = var_312_end_0, end_mask = var_312_end_mask_0, x = xp_1_cast_fp16)[name = string("op_312_cast_fp16")];
            tensor<fp16, [1024]> var_320_to_fp16 = const()[name = string("op_320_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22479552)))];
            tensor<fp16, [3, 64, 1024]> var_321_cast_fp16 = mul(x = var_312_cast_fp16, y = var_320_to_fp16)[name = string("op_321_cast_fp16")];
            tensor<int32, [3]> var_331_begin_0 = const()[name = string("op_331_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_331_end_0 = const()[name = string("op_331_end_0"), val = tensor<int32, [3]>([3, 66, 1024])];
            tensor<bool, [3]> var_331_end_mask_0 = const()[name = string("op_331_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_331_cast_fp16 = slice_by_index(begin = var_331_begin_0, end = var_331_end_0, end_mask = var_331_end_mask_0, x = xp_1_cast_fp16)[name = string("op_331_cast_fp16")];
            tensor<fp16, [1024]> var_339_to_fp16 = const()[name = string("op_339_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22481664)))];
            tensor<fp16, [3, 64, 1024]> var_340_cast_fp16 = mul(x = var_331_cast_fp16, y = var_339_to_fp16)[name = string("op_340_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> var_342_cast_fp16 = add(x = var_321_cast_fp16, y = var_340_cast_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [3]> var_352_begin_0 = const()[name = string("op_352_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_352_end_0 = const()[name = string("op_352_end_0"), val = tensor<int32, [3]>([3, 65, 1024])];
            tensor<bool, [3]> var_352_end_mask_0 = const()[name = string("op_352_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_352_cast_fp16 = slice_by_index(begin = var_352_begin_0, end = var_352_end_0, end_mask = var_352_end_mask_0, x = xp_1_cast_fp16)[name = string("op_352_cast_fp16")];
            tensor<fp16, [1024]> var_360_to_fp16 = const()[name = string("op_360_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22483776)))];
            tensor<fp16, [3, 64, 1024]> var_361_cast_fp16 = mul(x = var_352_cast_fp16, y = var_360_to_fp16)[name = string("op_361_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> var_363_cast_fp16 = add(x = var_342_cast_fp16, y = var_361_cast_fp16)[name = string("op_363_cast_fp16")];
            tensor<int32, [3]> var_373_begin_0 = const()[name = string("op_373_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_373_end_0 = const()[name = string("op_373_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> var_373_end_mask_0 = const()[name = string("op_373_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_373_cast_fp16 = slice_by_index(begin = var_373_begin_0, end = var_373_end_0, end_mask = var_373_end_mask_0, x = xp_1_cast_fp16)[name = string("op_373_cast_fp16")];
            tensor<fp16, [1024]> var_381_to_fp16 = const()[name = string("op_381_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22485888)))];
            tensor<fp16, [3, 64, 1024]> var_382_cast_fp16 = mul(x = var_373_cast_fp16, y = var_381_to_fp16)[name = string("op_382_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> input_3_cast_fp16 = add(x = var_363_cast_fp16, y = var_382_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> c_1_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("c_1_cast_fp16")];
            tensor<int32, [3]> var_390_begin_0 = const()[name = string("op_390_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_390_end_0 = const()[name = string("op_390_end_0"), val = tensor<int32, [3]>([3, 64, 256])];
            tensor<bool, [3]> var_390_end_mask_0 = const()[name = string("op_390_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 256]> var_390_cast_fp16 = slice_by_index(begin = var_390_begin_0, end = var_390_end_0, end_mask = var_390_end_mask_0, x = c_1_cast_fp16)[name = string("op_390_cast_fp16")];
            tensor<int32, [4]> var_395 = const()[name = string("op_395"), val = tensor<int32, [4]>([3, 64, 2, 128])];
            tensor<fp16, [3, 64, 2, 128]> q_1_cast_fp16 = reshape(shape = var_395, x = var_390_cast_fp16)[name = string("q_1_cast_fp16")];
            tensor<int32, [3]> var_401_begin_0 = const()[name = string("op_401_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_401_end_0 = const()[name = string("op_401_end_0"), val = tensor<int32, [3]>([3, 64, 512])];
            tensor<bool, [3]> var_401_end_mask_0 = const()[name = string("op_401_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 256]> var_401_cast_fp16 = slice_by_index(begin = var_401_begin_0, end = var_401_end_0, end_mask = var_401_end_mask_0, x = c_1_cast_fp16)[name = string("op_401_cast_fp16")];
            tensor<int32, [4]> var_406 = const()[name = string("op_406"), val = tensor<int32, [4]>([3, 64, 2, 128])];
            tensor<fp16, [3, 64, 2, 128]> k_1_cast_fp16 = reshape(shape = var_406, x = var_401_cast_fp16)[name = string("k_1_cast_fp16")];
            tensor<int32, [3]> var_412_begin_0 = const()[name = string("op_412_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_412_end_0 = const()[name = string("op_412_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> var_412_end_mask_0 = const()[name = string("op_412_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 512]> var_412_cast_fp16 = slice_by_index(begin = var_412_begin_0, end = var_412_end_0, end_mask = var_412_end_mask_0, x = c_1_cast_fp16)[name = string("op_412_cast_fp16")];
            tensor<int32, [4]> var_417 = const()[name = string("op_417"), val = tensor<int32, [4]>([3, 64, 4, 128])];
            tensor<fp16, [3, 64, 4, 128]> var_418_cast_fp16 = reshape(shape = var_417, x = var_412_cast_fp16)[name = string("op_418_cast_fp16")];
            tensor<int32, [4]> var_423 = const()[name = string("op_423"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_425_to_fp16 = const()[name = string("op_425_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 64, 2, 128]> q_3_cast_fp16 = mul(x = q_1_cast_fp16, y = var_425_to_fp16)[name = string("q_3_cast_fp16")];
            fp16 var_427_to_fp16 = const()[name = string("op_427_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 64, 2, 128]> k_3_cast_fp16 = mul(x = k_1_cast_fp16, y = var_427_to_fp16)[name = string("k_3_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_429_cast_fp16 = mul(x = q_3_cast_fp16, y = q_3_cast_fp16)[name = string("op_429_cast_fp16")];
            tensor<int32, [1]> var_434_axes_0 = const()[name = string("op_434_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_434_keep_dims_0 = const()[name = string("op_434_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 2, 1]> var_434_cast_fp16 = reduce_mean(axes = var_434_axes_0, keep_dims = var_434_keep_dims_0, x = var_429_cast_fp16)[name = string("op_434_cast_fp16")];
            fp16 var_435_promoted_to_fp16 = const()[name = string("op_435_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 64, 2, 1]> var_436_cast_fp16 = mul(x = var_434_cast_fp16, y = var_435_promoted_to_fp16)[name = string("op_436_cast_fp16")];
            fp16 var_438_to_fp16 = const()[name = string("op_438_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 64, 2, 1]> var_439_cast_fp16 = add(x = var_436_cast_fp16, y = var_438_to_fp16)[name = string("op_439_cast_fp16")];
            fp32 var_440_epsilon_0 = const()[name = string("op_440_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 2, 1]> var_440_cast_fp16 = rsqrt(epsilon = var_440_epsilon_0, x = var_439_cast_fp16)[name = string("op_440_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_441_cast_fp16 = mul(x = q_3_cast_fp16, y = var_440_cast_fp16)[name = string("op_441_cast_fp16")];
            fp16 var_442_to_fp16 = const()[name = string("op_442_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 64, 2, 128]> q_5_cast_fp16 = mul(x = var_441_cast_fp16, y = var_442_to_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_444_cast_fp16 = mul(x = k_3_cast_fp16, y = k_3_cast_fp16)[name = string("op_444_cast_fp16")];
            tensor<int32, [1]> var_449_axes_0 = const()[name = string("op_449_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_449_keep_dims_0 = const()[name = string("op_449_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 2, 1]> var_449_cast_fp16 = reduce_mean(axes = var_449_axes_0, keep_dims = var_449_keep_dims_0, x = var_444_cast_fp16)[name = string("op_449_cast_fp16")];
            fp16 var_450_promoted_to_fp16 = const()[name = string("op_450_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 64, 2, 1]> var_451_cast_fp16 = mul(x = var_449_cast_fp16, y = var_450_promoted_to_fp16)[name = string("op_451_cast_fp16")];
            fp16 var_453_to_fp16 = const()[name = string("op_453_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 64, 2, 1]> var_454_cast_fp16 = add(x = var_451_cast_fp16, y = var_453_to_fp16)[name = string("op_454_cast_fp16")];
            fp32 var_455_epsilon_0 = const()[name = string("op_455_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 2, 1]> var_455_cast_fp16 = rsqrt(epsilon = var_455_epsilon_0, x = var_454_cast_fp16)[name = string("op_455_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> k_5_cast_fp16 = mul(x = k_3_cast_fp16, y = var_455_cast_fp16)[name = string("k_5_cast_fp16")];
            tensor<int32, [4]> var_461 = const()[name = string("op_461"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_474_axes_0 = const()[name = string("op_474_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 64, 128]> var_462_cast_fp16 = transpose(perm = var_461, x = q_5_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [3, 2, 1, 64, 128]> var_474_cast_fp16 = expand_dims(axes = var_474_axes_0, x = var_462_cast_fp16)[name = string("op_474_cast_fp16")];
            tensor<int32, [5]> var_482_reps_0 = const()[name = string("op_482_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 64, 128]> var_482_cast_fp16 = tile(reps = var_482_reps_0, x = var_474_cast_fp16)[name = string("op_482_cast_fp16")];
            tensor<int32, [4]> var_487 = const()[name = string("op_487"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<fp16, [3, 4, 64, 128]> q_7_cast_fp16 = reshape(shape = var_487, x = var_482_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<int32, [4]> var_493 = const()[name = string("op_493"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_506_axes_0 = const()[name = string("op_506_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 64, 128]> var_494_cast_fp16 = transpose(perm = var_493, x = k_5_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [3, 2, 1, 64, 128]> var_506_cast_fp16 = expand_dims(axes = var_506_axes_0, x = var_494_cast_fp16)[name = string("op_506_cast_fp16")];
            tensor<int32, [5]> var_514_reps_0 = const()[name = string("op_514_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 64, 128]> var_514_cast_fp16 = tile(reps = var_514_reps_0, x = var_506_cast_fp16)[name = string("op_514_cast_fp16")];
            tensor<int32, [4]> var_519 = const()[name = string("op_519"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<fp16, [3, 4, 64, 128]> k_7_cast_fp16 = reshape(shape = var_519, x = var_514_cast_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_521_cast_fp16 = sigmoid(x = b_1_cast_fp16)[name = string("op_521_cast_fp16")];
            tensor<int32, [3]> var_525 = const()[name = string("op_525"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_0_dt_bias_to_fp16 = const()[name = string("layers_0_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 64, 4]> input_5_cast_fp16 = add(x = a_1_cast_fp16, y = layers_0_dt_bias_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_529_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("op_529_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_530_cast_fp16 = abs(x = input_5_cast_fp16)[name = string("op_530_cast_fp16")];
            fp16 const_0_promoted_to_fp16 = const()[name = string("const_0_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 64, 4]> var_531_cast_fp16 = mul(x = var_530_cast_fp16, y = const_0_promoted_to_fp16)[name = string("op_531_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_532_cast_fp16 = exp(x = var_531_cast_fp16)[name = string("op_532_cast_fp16")];
            fp16 var_534_to_fp16 = const()[name = string("op_534_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 64, 4]> var_535_cast_fp16 = add(x = var_532_cast_fp16, y = var_534_to_fp16)[name = string("op_535_cast_fp16")];
            fp32 var_536_epsilon_0 = const()[name = string("op_536_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 64, 4]> var_536_cast_fp16 = log(epsilon = var_536_epsilon_0, x = var_535_cast_fp16)[name = string("op_536_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_538_cast_fp16 = add(x = var_529_cast_fp16, y = var_536_cast_fp16)[name = string("op_538_cast_fp16")];
            tensor<fp16, [4]> layers_0_negA_to_fp16 = const()[name = string("layers_0_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.b3p+3, -0x1.a7cp+0, -0x1.ac4p+3, -0x1.704p+2])];
            tensor<fp16, [3, 64, 4]> var_539_cast_fp16 = mul(x = layers_0_negA_to_fp16, y = var_538_cast_fp16)[name = string("op_539_cast_fp16")];
            tensor<int32, [3]> var_543 = const()[name = string("op_543"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_546_axes_0 = const()[name = string("op_546_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 64]> beta_1_cast_fp16 = transpose(perm = var_525, x = var_521_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [3, 4, 64, 1]> var_546_cast_fp16 = expand_dims(axes = var_546_axes_0, x = beta_1_cast_fp16)[name = string("op_546_cast_fp16")];
            fp16 var_547_to_fp16 = const()[name = string("op_547_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 64, 1]> var_548_cast_fp16 = mul(x = var_546_cast_fp16, y = var_547_to_fp16)[name = string("op_548_cast_fp16")];
            tensor<fp16, [3, 4, 64, 128]> v_1_cast_fp16 = transpose(perm = var_423, x = var_418_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [3, 4, 64, 128]> vb_1_cast_fp16 = mul(x = v_1_cast_fp16, y = var_548_cast_fp16)[name = string("vb_1_cast_fp16")];
            tensor<int32, [4]> q_9_begin_0 = const()[name = string("q_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_9_end_0 = const()[name = string("q_9_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_9_end_mask_0 = const()[name = string("q_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_9_cast_fp16 = slice_by_index(begin = q_9_begin_0, end = q_9_end_0, end_mask = q_9_end_mask_0, x = q_7_cast_fp16)[name = string("q_9_cast_fp16")];
            tensor<int32, [4]> k_9_begin_0 = const()[name = string("k_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_9_end_0 = const()[name = string("k_9_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_9_end_mask_0 = const()[name = string("k_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, x = k_7_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<fp16, [3, 4, 64, 128]> var_582_cast_fp16 = mul(x = k_7_cast_fp16, y = var_546_cast_fp16)[name = string("op_582_cast_fp16")];
            tensor<int32, [4]> kb_1_begin_0 = const()[name = string("kb_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_1_end_0 = const()[name = string("kb_1_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_1_end_mask_0 = const()[name = string("kb_1_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_1_cast_fp16 = slice_by_index(begin = kb_1_begin_0, end = kb_1_end_0, end_mask = kb_1_end_mask_0, x = var_582_cast_fp16)[name = string("kb_1_cast_fp16")];
            tensor<int32, [4]> vb_3_begin_0 = const()[name = string("vb_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_3_end_0 = const()[name = string("vb_3_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_3_end_mask_0 = const()[name = string("vb_3_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_3_cast_fp16 = slice_by_index(begin = vb_3_begin_0, end = vb_3_end_0, end_mask = vb_3_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_3_cast_fp16")];
            tensor<int32, [3]> gdec_3_begin_0 = const()[name = string("gdec_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_3_end_0 = const()[name = string("gdec_3_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_3_end_mask_0 = const()[name = string("gdec_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 64]> gdec_1_cast_fp16 = transpose(perm = var_543, x = var_539_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [3, 4, 16]> gdec_3_cast_fp16 = slice_by_index(begin = gdec_3_begin_0, end = gdec_3_end_0, end_mask = gdec_3_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_3_cast_fp16")];
            tensor<int32, [2]> var_630 = const()[name = string("op_630"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_631_cast_fp16 = reshape(shape = var_630, x = gdec_3_cast_fp16)[name = string("op_631_cast_fp16")];
            tensor<fp16, [256, 16]> transpose_0_to_fp16 = const()[name = string("transpose_0_to_fp16"), val = tensor<fp16, [256, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22488000)))];
            tensor<fp16, [256]> var_632_bias_0_to_fp16 = const()[name = string("op_632_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22496256)))];
            tensor<fp16, [12, 256]> var_632_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_631_cast_fp16)[name = string("op_632_cast_fp16")];
            tensor<int32, [4]> var_637 = const()[name = string("op_637"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_1_cast_fp16 = reshape(shape = var_637, x = var_632_cast_fp16)[name = string("P_1_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_639_cast_fp16 = exp(x = P_1_cast_fp16)[name = string("op_639_cast_fp16")];
            tensor<fp16, [16, 16]> tril_incl_to_fp16 = const()[name = string("tril_incl_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22496832)))];
            tensor<fp16, [3, 4, 16, 16]> D_1_cast_fp16 = mul(x = var_639_cast_fp16, y = tril_incl_to_fp16)[name = string("D_1_cast_fp16")];
            bool kk_1_transpose_x_1 = const()[name = string("kk_1_transpose_x_1"), val = bool(false)];
            bool kk_1_transpose_y_1 = const()[name = string("kk_1_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_1_cast_fp16 = matmul(transpose_x = kk_1_transpose_x_1, transpose_y = kk_1_transpose_y_1, x = kb_1_cast_fp16, y = k_9_cast_fp16)[name = string("kk_1_cast_fp16")];
            fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_645_cast_fp16 = mul(x = kk_1_cast_fp16, y = const_1_promoted_to_fp16)[name = string("op_645_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_646_cast_fp16 = mul(x = var_645_cast_fp16, y = D_1_cast_fp16)[name = string("op_646_cast_fp16")];
            tensor<fp16, [16, 16]> cum_tail_to_fp16 = const()[name = string("cum_tail_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22497408)))];
            tensor<fp16, [3, 4, 16, 16]> N_1_cast_fp16 = mul(x = var_646_cast_fp16, y = cum_tail_to_fp16)[name = string("N_1_cast_fp16")];
            tensor<fp16, [16, 16]> neg_tril_strict_to_fp16 = const()[name = string("neg_tril_strict_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22497984)))];
            tensor<fp16, [3, 4, 16, 16]> var_648_cast_fp16 = mul(x = D_1_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_648_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_1_cast_fp16 = mul(x = kk_1_cast_fp16, y = var_648_cast_fp16)[name = string("Na_1_cast_fp16")];
            tensor<fp16, [16, 16]> eye_to_fp16 = const()[name = string("eye_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22498560)))];
            tensor<fp16, [3, 4, 16, 16]> Minv_1_cast_fp16 = add(x = eye_to_fp16, y = N_1_cast_fp16)[name = string("Minv_1_cast_fp16")];
            bool Np_1_transpose_x_0 = const()[name = string("Np_1_transpose_x_0"), val = bool(false)];
            bool Np_1_transpose_y_0 = const()[name = string("Np_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_1_cast_fp16 = matmul(transpose_x = Np_1_transpose_x_0, transpose_y = Np_1_transpose_y_0, x = N_1_cast_fp16, y = Na_1_cast_fp16)[name = string("Np_1_cast_fp16")];
            bool Npa_1_transpose_x_0 = const()[name = string("Npa_1_transpose_x_0"), val = bool(false)];
            bool Npa_1_transpose_y_0 = const()[name = string("Npa_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_1_cast_fp16 = matmul(transpose_x = Npa_1_transpose_x_0, transpose_y = Npa_1_transpose_y_0, x = Na_1_cast_fp16, y = N_1_cast_fp16)[name = string("Npa_1_cast_fp16")];
            bool var_654_transpose_x_0 = const()[name = string("op_654_transpose_x_0"), val = bool(false)];
            bool var_654_transpose_y_0 = const()[name = string("op_654_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_654_cast_fp16 = matmul(transpose_x = var_654_transpose_x_0, transpose_y = var_654_transpose_y_0, x = Minv_1_cast_fp16, y = Np_1_cast_fp16)[name = string("op_654_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_3_cast_fp16 = add(x = Minv_1_cast_fp16, y = var_654_cast_fp16)[name = string("Minv_3_cast_fp16")];
            bool Np_3_transpose_x_0 = const()[name = string("Np_3_transpose_x_0"), val = bool(false)];
            bool Np_3_transpose_y_0 = const()[name = string("Np_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_3_cast_fp16 = matmul(transpose_x = Np_3_transpose_x_0, transpose_y = Np_3_transpose_y_0, x = Np_1_cast_fp16, y = Npa_1_cast_fp16)[name = string("Np_3_cast_fp16")];
            bool Npa_3_transpose_x_0 = const()[name = string("Npa_3_transpose_x_0"), val = bool(false)];
            bool Npa_3_transpose_y_0 = const()[name = string("Npa_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_3_cast_fp16 = matmul(transpose_x = Npa_3_transpose_x_0, transpose_y = Npa_3_transpose_y_0, x = Npa_1_cast_fp16, y = Np_1_cast_fp16)[name = string("Npa_3_cast_fp16")];
            bool var_659_transpose_x_0 = const()[name = string("op_659_transpose_x_0"), val = bool(false)];
            bool var_659_transpose_y_0 = const()[name = string("op_659_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_659_cast_fp16 = matmul(transpose_x = var_659_transpose_x_0, transpose_y = var_659_transpose_y_0, x = Minv_3_cast_fp16, y = Np_3_cast_fp16)[name = string("op_659_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_5_cast_fp16 = add(x = Minv_3_cast_fp16, y = var_659_cast_fp16)[name = string("Minv_5_cast_fp16")];
            bool Np_5_transpose_x_0 = const()[name = string("Np_5_transpose_x_0"), val = bool(false)];
            bool Np_5_transpose_y_0 = const()[name = string("Np_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_5_cast_fp16 = matmul(transpose_x = Np_5_transpose_x_0, transpose_y = Np_5_transpose_y_0, x = Np_3_cast_fp16, y = Npa_3_cast_fp16)[name = string("Np_5_cast_fp16")];
            bool var_663_transpose_x_0 = const()[name = string("op_663_transpose_x_0"), val = bool(false)];
            bool var_663_transpose_y_0 = const()[name = string("op_663_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_663_cast_fp16 = matmul(transpose_x = var_663_transpose_x_0, transpose_y = var_663_transpose_y_0, x = Minv_5_cast_fp16, y = Np_5_cast_fp16)[name = string("op_663_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_7_cast_fp16 = add(x = Minv_5_cast_fp16, y = var_663_cast_fp16)[name = string("Minv_7_cast_fp16")];
            bool u_1_transpose_x_0 = const()[name = string("u_1_transpose_x_0"), val = bool(false)];
            bool u_1_transpose_y_0 = const()[name = string("u_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_1_cast_fp16 = matmul(transpose_x = u_1_transpose_x_0, transpose_y = u_1_transpose_y_0, x = Minv_7_cast_fp16, y = vb_3_cast_fp16)[name = string("u_1_cast_fp16")];
            bool var_670_transpose_x_1 = const()[name = string("op_670_transpose_x_1"), val = bool(false)];
            bool var_670_transpose_y_1 = const()[name = string("op_670_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_670_cast_fp16 = matmul(transpose_x = var_670_transpose_x_1, transpose_y = var_670_transpose_y_1, x = q_9_cast_fp16, y = k_9_cast_fp16)[name = string("op_670_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_1_cast_fp16 = mul(x = var_670_cast_fp16, y = D_1_cast_fp16)[name = string("A_1_cast_fp16")];
            bool var_672_transpose_x_0 = const()[name = string("op_672_transpose_x_0"), val = bool(false)];
            bool var_672_transpose_y_0 = const()[name = string("op_672_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_672_cast_fp16 = matmul(transpose_x = var_672_transpose_x_0, transpose_y = var_672_transpose_y_0, x = A_1_cast_fp16, y = u_1_cast_fp16)[name = string("op_672_cast_fp16")];
            tensor<fp16, [16, 16]> transpose_1_to_fp16 = const()[name = string("transpose_1_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658176)))];
            tensor<fp16, [16]> var_673_bias_0_to_fp16 = const()[name = string("op_673_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_673_cast_fp16 = linear(bias = var_673_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_3_cast_fp16)[name = string("op_673_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_1_cast_fp16 = exp(x = var_673_cast_fp16)[name = string("tail_1_cast_fp16")];
            tensor<int32, [1]> var_676_axes_0 = const()[name = string("op_676_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_676_cast_fp16 = expand_dims(axes = var_676_axes_0, x = tail_1_cast_fp16)[name = string("op_676_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_677_cast_fp16 = mul(x = k_9_cast_fp16, y = var_676_cast_fp16)[name = string("op_677_cast_fp16")];
            bool S_1_transpose_x_1 = const()[name = string("S_1_transpose_x_1"), val = bool(true)];
            bool S_1_transpose_y_1 = const()[name = string("S_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_1_cast_fp16 = matmul(transpose_x = S_1_transpose_x_1, transpose_y = S_1_transpose_y_1, x = var_677_cast_fp16, y = u_1_cast_fp16)[name = string("S_1_cast_fp16")];
            tensor<int32, [4]> q_11_begin_0 = const()[name = string("q_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_11_end_0 = const()[name = string("q_11_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> q_11_end_mask_0 = const()[name = string("q_11_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_11_cast_fp16 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, x = q_7_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<int32, [4]> k_11_begin_0 = const()[name = string("k_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_11_end_0 = const()[name = string("k_11_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> k_11_end_mask_0 = const()[name = string("k_11_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_11_cast_fp16 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, x = k_7_cast_fp16)[name = string("k_11_cast_fp16")];
            tensor<int32, [4]> kb_3_begin_0 = const()[name = string("kb_3_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_3_end_0 = const()[name = string("kb_3_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> kb_3_end_mask_0 = const()[name = string("kb_3_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_3_cast_fp16 = slice_by_index(begin = kb_3_begin_0, end = kb_3_end_0, end_mask = kb_3_end_mask_0, x = var_582_cast_fp16)[name = string("kb_3_cast_fp16")];
            tensor<int32, [4]> vb_5_begin_0 = const()[name = string("vb_5_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_5_end_0 = const()[name = string("vb_5_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> vb_5_end_mask_0 = const()[name = string("vb_5_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_5_cast_fp16 = slice_by_index(begin = vb_5_begin_0, end = vb_5_end_0, end_mask = vb_5_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_5_cast_fp16")];
            tensor<int32, [3]> gdec_5_begin_0 = const()[name = string("gdec_5_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_5_end_0 = const()[name = string("gdec_5_end_0"), val = tensor<int32, [3]>([3, 4, 32])];
            tensor<bool, [3]> gdec_5_end_mask_0 = const()[name = string("gdec_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_5_cast_fp16 = slice_by_index(begin = gdec_5_begin_0, end = gdec_5_end_0, end_mask = gdec_5_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_5_cast_fp16")];
            tensor<int32, [2]> var_762 = const()[name = string("op_762"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_763_cast_fp16 = reshape(shape = var_762, x = gdec_5_cast_fp16)[name = string("op_763_cast_fp16")];
            tensor<fp16, [12, 256]> var_764_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_763_cast_fp16)[name = string("op_764_cast_fp16")];
            tensor<int32, [4]> var_769 = const()[name = string("op_769"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_3_cast_fp16 = reshape(shape = var_769, x = var_764_cast_fp16)[name = string("P_3_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_771_cast_fp16 = exp(x = P_3_cast_fp16)[name = string("op_771_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_3_cast_fp16 = mul(x = var_771_cast_fp16, y = tril_incl_to_fp16)[name = string("D_3_cast_fp16")];
            bool kk_3_transpose_x_1 = const()[name = string("kk_3_transpose_x_1"), val = bool(false)];
            bool kk_3_transpose_y_1 = const()[name = string("kk_3_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_3_cast_fp16 = matmul(transpose_x = kk_3_transpose_x_1, transpose_y = kk_3_transpose_y_1, x = kb_3_cast_fp16, y = k_11_cast_fp16)[name = string("kk_3_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_777_cast_fp16 = mul(x = kk_3_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_777_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_778_cast_fp16 = mul(x = var_777_cast_fp16, y = D_3_cast_fp16)[name = string("op_778_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_3_cast_fp16 = mul(x = var_778_cast_fp16, y = cum_tail_to_fp16)[name = string("N_3_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_780_cast_fp16 = mul(x = D_3_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_780_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_3_cast_fp16 = mul(x = kk_3_cast_fp16, y = var_780_cast_fp16)[name = string("Na_3_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_9_cast_fp16 = add(x = eye_to_fp16, y = N_3_cast_fp16)[name = string("Minv_9_cast_fp16")];
            bool Np_7_transpose_x_0 = const()[name = string("Np_7_transpose_x_0"), val = bool(false)];
            bool Np_7_transpose_y_0 = const()[name = string("Np_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_7_cast_fp16 = matmul(transpose_x = Np_7_transpose_x_0, transpose_y = Np_7_transpose_y_0, x = N_3_cast_fp16, y = Na_3_cast_fp16)[name = string("Np_7_cast_fp16")];
            bool Npa_5_transpose_x_0 = const()[name = string("Npa_5_transpose_x_0"), val = bool(false)];
            bool Npa_5_transpose_y_0 = const()[name = string("Npa_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_5_cast_fp16 = matmul(transpose_x = Npa_5_transpose_x_0, transpose_y = Npa_5_transpose_y_0, x = Na_3_cast_fp16, y = N_3_cast_fp16)[name = string("Npa_5_cast_fp16")];
            bool var_786_transpose_x_0 = const()[name = string("op_786_transpose_x_0"), val = bool(false)];
            bool var_786_transpose_y_0 = const()[name = string("op_786_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_786_cast_fp16 = matmul(transpose_x = var_786_transpose_x_0, transpose_y = var_786_transpose_y_0, x = Minv_9_cast_fp16, y = Np_7_cast_fp16)[name = string("op_786_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_11_cast_fp16 = add(x = Minv_9_cast_fp16, y = var_786_cast_fp16)[name = string("Minv_11_cast_fp16")];
            bool Np_9_transpose_x_0 = const()[name = string("Np_9_transpose_x_0"), val = bool(false)];
            bool Np_9_transpose_y_0 = const()[name = string("Np_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_9_cast_fp16 = matmul(transpose_x = Np_9_transpose_x_0, transpose_y = Np_9_transpose_y_0, x = Np_7_cast_fp16, y = Npa_5_cast_fp16)[name = string("Np_9_cast_fp16")];
            bool Npa_7_transpose_x_0 = const()[name = string("Npa_7_transpose_x_0"), val = bool(false)];
            bool Npa_7_transpose_y_0 = const()[name = string("Npa_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_7_cast_fp16 = matmul(transpose_x = Npa_7_transpose_x_0, transpose_y = Npa_7_transpose_y_0, x = Npa_5_cast_fp16, y = Np_7_cast_fp16)[name = string("Npa_7_cast_fp16")];
            bool var_791_transpose_x_0 = const()[name = string("op_791_transpose_x_0"), val = bool(false)];
            bool var_791_transpose_y_0 = const()[name = string("op_791_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_791_cast_fp16 = matmul(transpose_x = var_791_transpose_x_0, transpose_y = var_791_transpose_y_0, x = Minv_11_cast_fp16, y = Np_9_cast_fp16)[name = string("op_791_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_13_cast_fp16 = add(x = Minv_11_cast_fp16, y = var_791_cast_fp16)[name = string("Minv_13_cast_fp16")];
            bool Np_11_transpose_x_0 = const()[name = string("Np_11_transpose_x_0"), val = bool(false)];
            bool Np_11_transpose_y_0 = const()[name = string("Np_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_11_cast_fp16 = matmul(transpose_x = Np_11_transpose_x_0, transpose_y = Np_11_transpose_y_0, x = Np_9_cast_fp16, y = Npa_7_cast_fp16)[name = string("Np_11_cast_fp16")];
            bool var_795_transpose_x_0 = const()[name = string("op_795_transpose_x_0"), val = bool(false)];
            bool var_795_transpose_y_0 = const()[name = string("op_795_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_795_cast_fp16 = matmul(transpose_x = var_795_transpose_x_0, transpose_y = var_795_transpose_y_0, x = Minv_13_cast_fp16, y = Np_11_cast_fp16)[name = string("op_795_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_15_cast_fp16 = add(x = Minv_13_cast_fp16, y = var_795_cast_fp16)[name = string("Minv_15_cast_fp16")];
            bool u_3_transpose_x_0 = const()[name = string("u_3_transpose_x_0"), val = bool(false)];
            bool u_3_transpose_y_0 = const()[name = string("u_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_3_cast_fp16 = matmul(transpose_x = u_3_transpose_x_0, transpose_y = u_3_transpose_y_0, x = Minv_15_cast_fp16, y = vb_5_cast_fp16)[name = string("u_3_cast_fp16")];
            bool var_802_transpose_x_1 = const()[name = string("op_802_transpose_x_1"), val = bool(false)];
            bool var_802_transpose_y_1 = const()[name = string("op_802_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_802_cast_fp16 = matmul(transpose_x = var_802_transpose_x_1, transpose_y = var_802_transpose_y_1, x = q_11_cast_fp16, y = k_11_cast_fp16)[name = string("op_802_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_3_cast_fp16 = mul(x = var_802_cast_fp16, y = D_3_cast_fp16)[name = string("A_3_cast_fp16")];
            tensor<fp16, [16]> var_804_bias_0_to_fp16 = const()[name = string("op_804_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_804_cast_fp16 = linear(bias = var_804_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_5_cast_fp16)[name = string("op_804_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_1_cast_fp16 = exp(x = var_804_cast_fp16)[name = string("cum_1_cast_fp16")];
            tensor<int32, [1]> var_807_axes_0 = const()[name = string("op_807_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_807_cast_fp16 = expand_dims(axes = var_807_axes_0, x = cum_1_cast_fp16)[name = string("op_807_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_808_cast_fp16 = mul(x = kb_3_cast_fp16, y = var_807_cast_fp16)[name = string("op_808_cast_fp16")];
            bool w_3_transpose_x_0 = const()[name = string("w_3_transpose_x_0"), val = bool(false)];
            bool w_3_transpose_y_0 = const()[name = string("w_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_3_cast_fp16 = matmul(transpose_x = w_3_transpose_x_0, transpose_y = w_3_transpose_y_0, x = Minv_15_cast_fp16, y = var_808_cast_fp16)[name = string("w_3_cast_fp16")];
            bool var_810_transpose_x_0 = const()[name = string("op_810_transpose_x_0"), val = bool(false)];
            bool var_810_transpose_y_0 = const()[name = string("op_810_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_810_cast_fp16 = matmul(transpose_x = var_810_transpose_x_0, transpose_y = var_810_transpose_y_0, x = w_3_cast_fp16, y = S_1_cast_fp16)[name = string("op_810_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_1_cast_fp16 = sub(x = u_3_cast_fp16, y = var_810_cast_fp16)[name = string("vnew_1_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_815_cast_fp16 = mul(x = q_11_cast_fp16, y = var_807_cast_fp16)[name = string("op_815_cast_fp16")];
            bool var_816_transpose_x_0 = const()[name = string("op_816_transpose_x_0"), val = bool(false)];
            bool var_816_transpose_y_0 = const()[name = string("op_816_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_816_cast_fp16 = matmul(transpose_x = var_816_transpose_x_0, transpose_y = var_816_transpose_y_0, x = var_815_cast_fp16, y = S_1_cast_fp16)[name = string("op_816_cast_fp16")];
            bool var_817_transpose_x_0 = const()[name = string("op_817_transpose_x_0"), val = bool(false)];
            bool var_817_transpose_y_0 = const()[name = string("op_817_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_817_cast_fp16 = matmul(transpose_x = var_817_transpose_x_0, transpose_y = var_817_transpose_y_0, x = A_3_cast_fp16, y = vnew_1_cast_fp16)[name = string("op_817_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_819_cast_fp16 = add(x = var_816_cast_fp16, y = var_817_cast_fp16)[name = string("op_819_cast_fp16")];
            tensor<fp16, [16]> var_820_bias_0_to_fp16 = const()[name = string("op_820_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_820_cast_fp16 = linear(bias = var_820_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_5_cast_fp16)[name = string("op_820_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_3_cast_fp16 = exp(x = var_820_cast_fp16)[name = string("tail_3_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_5_to_fp16 = const()[name = string("transpose_5_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_822_bias_0_to_fp16 = const()[name = string("op_822_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_822_cast_fp16 = linear(bias = var_822_bias_0_to_fp16, weight = transpose_5_to_fp16, x = gdec_5_cast_fp16)[name = string("op_822_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_823_cast_fp16 = exp(x = var_822_cast_fp16)[name = string("op_823_cast_fp16")];
            tensor<int32, [1]> last_1_axes_0 = const()[name = string("last_1_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_1_cast_fp16 = expand_dims(axes = last_1_axes_0, x = var_823_cast_fp16)[name = string("last_1_cast_fp16")];
            tensor<int32, [1]> var_827_axes_0 = const()[name = string("op_827_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_827_cast_fp16 = expand_dims(axes = var_827_axes_0, x = tail_3_cast_fp16)[name = string("op_827_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_828_cast_fp16 = mul(x = k_11_cast_fp16, y = var_827_cast_fp16)[name = string("op_828_cast_fp16")];
            bool upd_1_transpose_x_1 = const()[name = string("upd_1_transpose_x_1"), val = bool(true)];
            bool upd_1_transpose_y_1 = const()[name = string("upd_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_1_cast_fp16 = matmul(transpose_x = upd_1_transpose_x_1, transpose_y = upd_1_transpose_y_1, x = var_828_cast_fp16, y = vnew_1_cast_fp16)[name = string("upd_1_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_833_cast_fp16 = mul(x = S_1_cast_fp16, y = last_1_cast_fp16)[name = string("op_833_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_3_cast_fp16 = add(x = var_833_cast_fp16, y = upd_1_cast_fp16)[name = string("S_3_cast_fp16")];
            tensor<int32, [4]> q_13_begin_0 = const()[name = string("q_13_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_13_end_0 = const()[name = string("q_13_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> q_13_end_mask_0 = const()[name = string("q_13_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, x = q_7_cast_fp16)[name = string("q_13_cast_fp16")];
            tensor<int32, [4]> k_13_begin_0 = const()[name = string("k_13_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_13_end_0 = const()[name = string("k_13_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> k_13_end_mask_0 = const()[name = string("k_13_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, x = k_7_cast_fp16)[name = string("k_13_cast_fp16")];
            tensor<int32, [4]> kb_5_begin_0 = const()[name = string("kb_5_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_5_end_0 = const()[name = string("kb_5_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> kb_5_end_mask_0 = const()[name = string("kb_5_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_5_cast_fp16 = slice_by_index(begin = kb_5_begin_0, end = kb_5_end_0, end_mask = kb_5_end_mask_0, x = var_582_cast_fp16)[name = string("kb_5_cast_fp16")];
            tensor<int32, [4]> vb_7_begin_0 = const()[name = string("vb_7_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_7_end_0 = const()[name = string("vb_7_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> vb_7_end_mask_0 = const()[name = string("vb_7_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_7_cast_fp16 = slice_by_index(begin = vb_7_begin_0, end = vb_7_end_0, end_mask = vb_7_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_7_cast_fp16")];
            tensor<int32, [3]> gdec_7_begin_0 = const()[name = string("gdec_7_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_7_end_0 = const()[name = string("gdec_7_end_0"), val = tensor<int32, [3]>([3, 4, 48])];
            tensor<bool, [3]> gdec_7_end_mask_0 = const()[name = string("gdec_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_7_cast_fp16 = slice_by_index(begin = gdec_7_begin_0, end = gdec_7_end_0, end_mask = gdec_7_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_7_cast_fp16")];
            tensor<int32, [2]> var_916 = const()[name = string("op_916"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_917_cast_fp16 = reshape(shape = var_916, x = gdec_7_cast_fp16)[name = string("op_917_cast_fp16")];
            tensor<fp16, [12, 256]> var_918_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_917_cast_fp16)[name = string("op_918_cast_fp16")];
            tensor<int32, [4]> var_923 = const()[name = string("op_923"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_5_cast_fp16 = reshape(shape = var_923, x = var_918_cast_fp16)[name = string("P_5_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_925_cast_fp16 = exp(x = P_5_cast_fp16)[name = string("op_925_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_5_cast_fp16 = mul(x = var_925_cast_fp16, y = tril_incl_to_fp16)[name = string("D_5_cast_fp16")];
            bool kk_5_transpose_x_1 = const()[name = string("kk_5_transpose_x_1"), val = bool(false)];
            bool kk_5_transpose_y_1 = const()[name = string("kk_5_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_5_cast_fp16 = matmul(transpose_x = kk_5_transpose_x_1, transpose_y = kk_5_transpose_y_1, x = kb_5_cast_fp16, y = k_13_cast_fp16)[name = string("kk_5_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_931_cast_fp16 = mul(x = kk_5_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_931_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_932_cast_fp16 = mul(x = var_931_cast_fp16, y = D_5_cast_fp16)[name = string("op_932_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_5_cast_fp16 = mul(x = var_932_cast_fp16, y = cum_tail_to_fp16)[name = string("N_5_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_934_cast_fp16 = mul(x = D_5_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_934_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_5_cast_fp16 = mul(x = kk_5_cast_fp16, y = var_934_cast_fp16)[name = string("Na_5_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_17_cast_fp16 = add(x = eye_to_fp16, y = N_5_cast_fp16)[name = string("Minv_17_cast_fp16")];
            bool Np_13_transpose_x_0 = const()[name = string("Np_13_transpose_x_0"), val = bool(false)];
            bool Np_13_transpose_y_0 = const()[name = string("Np_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_13_cast_fp16 = matmul(transpose_x = Np_13_transpose_x_0, transpose_y = Np_13_transpose_y_0, x = N_5_cast_fp16, y = Na_5_cast_fp16)[name = string("Np_13_cast_fp16")];
            bool Npa_9_transpose_x_0 = const()[name = string("Npa_9_transpose_x_0"), val = bool(false)];
            bool Npa_9_transpose_y_0 = const()[name = string("Npa_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_9_cast_fp16 = matmul(transpose_x = Npa_9_transpose_x_0, transpose_y = Npa_9_transpose_y_0, x = Na_5_cast_fp16, y = N_5_cast_fp16)[name = string("Npa_9_cast_fp16")];
            bool var_940_transpose_x_0 = const()[name = string("op_940_transpose_x_0"), val = bool(false)];
            bool var_940_transpose_y_0 = const()[name = string("op_940_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_940_cast_fp16 = matmul(transpose_x = var_940_transpose_x_0, transpose_y = var_940_transpose_y_0, x = Minv_17_cast_fp16, y = Np_13_cast_fp16)[name = string("op_940_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_19_cast_fp16 = add(x = Minv_17_cast_fp16, y = var_940_cast_fp16)[name = string("Minv_19_cast_fp16")];
            bool Np_15_transpose_x_0 = const()[name = string("Np_15_transpose_x_0"), val = bool(false)];
            bool Np_15_transpose_y_0 = const()[name = string("Np_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_15_cast_fp16 = matmul(transpose_x = Np_15_transpose_x_0, transpose_y = Np_15_transpose_y_0, x = Np_13_cast_fp16, y = Npa_9_cast_fp16)[name = string("Np_15_cast_fp16")];
            bool Npa_11_transpose_x_0 = const()[name = string("Npa_11_transpose_x_0"), val = bool(false)];
            bool Npa_11_transpose_y_0 = const()[name = string("Npa_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_11_cast_fp16 = matmul(transpose_x = Npa_11_transpose_x_0, transpose_y = Npa_11_transpose_y_0, x = Npa_9_cast_fp16, y = Np_13_cast_fp16)[name = string("Npa_11_cast_fp16")];
            bool var_945_transpose_x_0 = const()[name = string("op_945_transpose_x_0"), val = bool(false)];
            bool var_945_transpose_y_0 = const()[name = string("op_945_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_945_cast_fp16 = matmul(transpose_x = var_945_transpose_x_0, transpose_y = var_945_transpose_y_0, x = Minv_19_cast_fp16, y = Np_15_cast_fp16)[name = string("op_945_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_21_cast_fp16 = add(x = Minv_19_cast_fp16, y = var_945_cast_fp16)[name = string("Minv_21_cast_fp16")];
            bool Np_17_transpose_x_0 = const()[name = string("Np_17_transpose_x_0"), val = bool(false)];
            bool Np_17_transpose_y_0 = const()[name = string("Np_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_17_cast_fp16 = matmul(transpose_x = Np_17_transpose_x_0, transpose_y = Np_17_transpose_y_0, x = Np_15_cast_fp16, y = Npa_11_cast_fp16)[name = string("Np_17_cast_fp16")];
            bool var_949_transpose_x_0 = const()[name = string("op_949_transpose_x_0"), val = bool(false)];
            bool var_949_transpose_y_0 = const()[name = string("op_949_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_949_cast_fp16 = matmul(transpose_x = var_949_transpose_x_0, transpose_y = var_949_transpose_y_0, x = Minv_21_cast_fp16, y = Np_17_cast_fp16)[name = string("op_949_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_23_cast_fp16 = add(x = Minv_21_cast_fp16, y = var_949_cast_fp16)[name = string("Minv_23_cast_fp16")];
            bool u_5_transpose_x_0 = const()[name = string("u_5_transpose_x_0"), val = bool(false)];
            bool u_5_transpose_y_0 = const()[name = string("u_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_5_cast_fp16 = matmul(transpose_x = u_5_transpose_x_0, transpose_y = u_5_transpose_y_0, x = Minv_23_cast_fp16, y = vb_7_cast_fp16)[name = string("u_5_cast_fp16")];
            bool var_956_transpose_x_1 = const()[name = string("op_956_transpose_x_1"), val = bool(false)];
            bool var_956_transpose_y_1 = const()[name = string("op_956_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_956_cast_fp16 = matmul(transpose_x = var_956_transpose_x_1, transpose_y = var_956_transpose_y_1, x = q_13_cast_fp16, y = k_13_cast_fp16)[name = string("op_956_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_5_cast_fp16 = mul(x = var_956_cast_fp16, y = D_5_cast_fp16)[name = string("A_5_cast_fp16")];
            tensor<fp16, [16]> var_958_bias_0_to_fp16 = const()[name = string("op_958_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_958_cast_fp16 = linear(bias = var_958_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_7_cast_fp16)[name = string("op_958_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_3_cast_fp16 = exp(x = var_958_cast_fp16)[name = string("cum_3_cast_fp16")];
            tensor<int32, [1]> var_961_axes_0 = const()[name = string("op_961_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_961_cast_fp16 = expand_dims(axes = var_961_axes_0, x = cum_3_cast_fp16)[name = string("op_961_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_962_cast_fp16 = mul(x = kb_5_cast_fp16, y = var_961_cast_fp16)[name = string("op_962_cast_fp16")];
            bool w_5_transpose_x_0 = const()[name = string("w_5_transpose_x_0"), val = bool(false)];
            bool w_5_transpose_y_0 = const()[name = string("w_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_5_cast_fp16 = matmul(transpose_x = w_5_transpose_x_0, transpose_y = w_5_transpose_y_0, x = Minv_23_cast_fp16, y = var_962_cast_fp16)[name = string("w_5_cast_fp16")];
            bool var_964_transpose_x_0 = const()[name = string("op_964_transpose_x_0"), val = bool(false)];
            bool var_964_transpose_y_0 = const()[name = string("op_964_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_964_cast_fp16 = matmul(transpose_x = var_964_transpose_x_0, transpose_y = var_964_transpose_y_0, x = w_5_cast_fp16, y = S_3_cast_fp16)[name = string("op_964_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_3_cast_fp16 = sub(x = u_5_cast_fp16, y = var_964_cast_fp16)[name = string("vnew_3_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_969_cast_fp16 = mul(x = q_13_cast_fp16, y = var_961_cast_fp16)[name = string("op_969_cast_fp16")];
            bool var_970_transpose_x_0 = const()[name = string("op_970_transpose_x_0"), val = bool(false)];
            bool var_970_transpose_y_0 = const()[name = string("op_970_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_970_cast_fp16 = matmul(transpose_x = var_970_transpose_x_0, transpose_y = var_970_transpose_y_0, x = var_969_cast_fp16, y = S_3_cast_fp16)[name = string("op_970_cast_fp16")];
            bool var_971_transpose_x_0 = const()[name = string("op_971_transpose_x_0"), val = bool(false)];
            bool var_971_transpose_y_0 = const()[name = string("op_971_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_971_cast_fp16 = matmul(transpose_x = var_971_transpose_x_0, transpose_y = var_971_transpose_y_0, x = A_5_cast_fp16, y = vnew_3_cast_fp16)[name = string("op_971_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_973_cast_fp16 = add(x = var_970_cast_fp16, y = var_971_cast_fp16)[name = string("op_973_cast_fp16")];
            tensor<fp16, [16]> var_974_bias_0_to_fp16 = const()[name = string("op_974_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_974_cast_fp16 = linear(bias = var_974_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_7_cast_fp16)[name = string("op_974_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_5_cast_fp16 = exp(x = var_974_cast_fp16)[name = string("tail_5_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_9_to_fp16 = const()[name = string("transpose_9_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_976_bias_0_to_fp16 = const()[name = string("op_976_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_976_cast_fp16 = linear(bias = var_976_bias_0_to_fp16, weight = transpose_9_to_fp16, x = gdec_7_cast_fp16)[name = string("op_976_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_977_cast_fp16 = exp(x = var_976_cast_fp16)[name = string("op_977_cast_fp16")];
            tensor<int32, [1]> last_3_axes_0 = const()[name = string("last_3_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_3_cast_fp16 = expand_dims(axes = last_3_axes_0, x = var_977_cast_fp16)[name = string("last_3_cast_fp16")];
            tensor<int32, [1]> var_981_axes_0 = const()[name = string("op_981_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_981_cast_fp16 = expand_dims(axes = var_981_axes_0, x = tail_5_cast_fp16)[name = string("op_981_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_982_cast_fp16 = mul(x = k_13_cast_fp16, y = var_981_cast_fp16)[name = string("op_982_cast_fp16")];
            bool upd_3_transpose_x_1 = const()[name = string("upd_3_transpose_x_1"), val = bool(true)];
            bool upd_3_transpose_y_1 = const()[name = string("upd_3_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_3_cast_fp16 = matmul(transpose_x = upd_3_transpose_x_1, transpose_y = upd_3_transpose_y_1, x = var_982_cast_fp16, y = vnew_3_cast_fp16)[name = string("upd_3_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_987_cast_fp16 = mul(x = S_3_cast_fp16, y = last_3_cast_fp16)[name = string("op_987_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_5_cast_fp16 = add(x = var_987_cast_fp16, y = upd_3_cast_fp16)[name = string("S_5_cast_fp16")];
            tensor<int32, [4]> q_15_begin_0 = const()[name = string("q_15_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_15_end_0 = const()[name = string("q_15_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_15_end_mask_0 = const()[name = string("q_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_15_cast_fp16 = slice_by_index(begin = q_15_begin_0, end = q_15_end_0, end_mask = q_15_end_mask_0, x = q_7_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<int32, [4]> k_15_begin_0 = const()[name = string("k_15_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_15_end_0 = const()[name = string("k_15_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_15_end_mask_0 = const()[name = string("k_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_15_cast_fp16 = slice_by_index(begin = k_15_begin_0, end = k_15_end_0, end_mask = k_15_end_mask_0, x = k_7_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> kb_7_begin_0 = const()[name = string("kb_7_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_7_end_0 = const()[name = string("kb_7_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_7_end_mask_0 = const()[name = string("kb_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_7_cast_fp16 = slice_by_index(begin = kb_7_begin_0, end = kb_7_end_0, end_mask = kb_7_end_mask_0, x = var_582_cast_fp16)[name = string("kb_7_cast_fp16")];
            tensor<int32, [4]> vb_9_begin_0 = const()[name = string("vb_9_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_9_end_0 = const()[name = string("vb_9_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_9_end_mask_0 = const()[name = string("vb_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_9_cast_fp16 = slice_by_index(begin = vb_9_begin_0, end = vb_9_end_0, end_mask = vb_9_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_9_cast_fp16")];
            tensor<int32, [3]> gdec_9_begin_0 = const()[name = string("gdec_9_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_9_end_0 = const()[name = string("gdec_9_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_9_end_mask_0 = const()[name = string("gdec_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_9_cast_fp16 = slice_by_index(begin = gdec_9_begin_0, end = gdec_9_end_0, end_mask = gdec_9_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_9_cast_fp16")];
            tensor<int32, [2]> var_1070 = const()[name = string("op_1070"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_1071_cast_fp16 = reshape(shape = var_1070, x = gdec_9_cast_fp16)[name = string("op_1071_cast_fp16")];
            tensor<fp16, [12, 256]> var_1072_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1071_cast_fp16)[name = string("op_1072_cast_fp16")];
            tensor<int32, [4]> var_1077 = const()[name = string("op_1077"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_7_cast_fp16 = reshape(shape = var_1077, x = var_1072_cast_fp16)[name = string("P_7_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1079_cast_fp16 = exp(x = P_7_cast_fp16)[name = string("op_1079_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_7_cast_fp16 = mul(x = var_1079_cast_fp16, y = tril_incl_to_fp16)[name = string("D_7_cast_fp16")];
            bool kk_7_transpose_x_1 = const()[name = string("kk_7_transpose_x_1"), val = bool(false)];
            bool kk_7_transpose_y_1 = const()[name = string("kk_7_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_7_cast_fp16 = matmul(transpose_x = kk_7_transpose_x_1, transpose_y = kk_7_transpose_y_1, x = kb_7_cast_fp16, y = k_15_cast_fp16)[name = string("kk_7_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_1085_cast_fp16 = mul(x = kk_7_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_1085_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1086_cast_fp16 = mul(x = var_1085_cast_fp16, y = D_7_cast_fp16)[name = string("op_1086_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_7_cast_fp16 = mul(x = var_1086_cast_fp16, y = cum_tail_to_fp16)[name = string("N_7_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1088_cast_fp16 = mul(x = D_7_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1088_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_7_cast_fp16 = mul(x = kk_7_cast_fp16, y = var_1088_cast_fp16)[name = string("Na_7_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_25_cast_fp16 = add(x = eye_to_fp16, y = N_7_cast_fp16)[name = string("Minv_25_cast_fp16")];
            bool Np_19_transpose_x_0 = const()[name = string("Np_19_transpose_x_0"), val = bool(false)];
            bool Np_19_transpose_y_0 = const()[name = string("Np_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_19_cast_fp16 = matmul(transpose_x = Np_19_transpose_x_0, transpose_y = Np_19_transpose_y_0, x = N_7_cast_fp16, y = Na_7_cast_fp16)[name = string("Np_19_cast_fp16")];
            bool Npa_13_transpose_x_0 = const()[name = string("Npa_13_transpose_x_0"), val = bool(false)];
            bool Npa_13_transpose_y_0 = const()[name = string("Npa_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_13_cast_fp16 = matmul(transpose_x = Npa_13_transpose_x_0, transpose_y = Npa_13_transpose_y_0, x = Na_7_cast_fp16, y = N_7_cast_fp16)[name = string("Npa_13_cast_fp16")];
            bool var_1094_transpose_x_0 = const()[name = string("op_1094_transpose_x_0"), val = bool(false)];
            bool var_1094_transpose_y_0 = const()[name = string("op_1094_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1094_cast_fp16 = matmul(transpose_x = var_1094_transpose_x_0, transpose_y = var_1094_transpose_y_0, x = Minv_25_cast_fp16, y = Np_19_cast_fp16)[name = string("op_1094_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_27_cast_fp16 = add(x = Minv_25_cast_fp16, y = var_1094_cast_fp16)[name = string("Minv_27_cast_fp16")];
            bool Np_21_transpose_x_0 = const()[name = string("Np_21_transpose_x_0"), val = bool(false)];
            bool Np_21_transpose_y_0 = const()[name = string("Np_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_21_cast_fp16 = matmul(transpose_x = Np_21_transpose_x_0, transpose_y = Np_21_transpose_y_0, x = Np_19_cast_fp16, y = Npa_13_cast_fp16)[name = string("Np_21_cast_fp16")];
            bool Npa_15_transpose_x_0 = const()[name = string("Npa_15_transpose_x_0"), val = bool(false)];
            bool Npa_15_transpose_y_0 = const()[name = string("Npa_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_15_cast_fp16 = matmul(transpose_x = Npa_15_transpose_x_0, transpose_y = Npa_15_transpose_y_0, x = Npa_13_cast_fp16, y = Np_19_cast_fp16)[name = string("Npa_15_cast_fp16")];
            bool var_1099_transpose_x_0 = const()[name = string("op_1099_transpose_x_0"), val = bool(false)];
            bool var_1099_transpose_y_0 = const()[name = string("op_1099_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1099_cast_fp16 = matmul(transpose_x = var_1099_transpose_x_0, transpose_y = var_1099_transpose_y_0, x = Minv_27_cast_fp16, y = Np_21_cast_fp16)[name = string("op_1099_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_29_cast_fp16 = add(x = Minv_27_cast_fp16, y = var_1099_cast_fp16)[name = string("Minv_29_cast_fp16")];
            bool Np_23_transpose_x_0 = const()[name = string("Np_23_transpose_x_0"), val = bool(false)];
            bool Np_23_transpose_y_0 = const()[name = string("Np_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_23_cast_fp16 = matmul(transpose_x = Np_23_transpose_x_0, transpose_y = Np_23_transpose_y_0, x = Np_21_cast_fp16, y = Npa_15_cast_fp16)[name = string("Np_23_cast_fp16")];
            bool var_1103_transpose_x_0 = const()[name = string("op_1103_transpose_x_0"), val = bool(false)];
            bool var_1103_transpose_y_0 = const()[name = string("op_1103_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1103_cast_fp16 = matmul(transpose_x = var_1103_transpose_x_0, transpose_y = var_1103_transpose_y_0, x = Minv_29_cast_fp16, y = Np_23_cast_fp16)[name = string("op_1103_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_31_cast_fp16 = add(x = Minv_29_cast_fp16, y = var_1103_cast_fp16)[name = string("Minv_31_cast_fp16")];
            bool u_7_transpose_x_0 = const()[name = string("u_7_transpose_x_0"), val = bool(false)];
            bool u_7_transpose_y_0 = const()[name = string("u_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_7_cast_fp16 = matmul(transpose_x = u_7_transpose_x_0, transpose_y = u_7_transpose_y_0, x = Minv_31_cast_fp16, y = vb_9_cast_fp16)[name = string("u_7_cast_fp16")];
            bool var_1110_transpose_x_1 = const()[name = string("op_1110_transpose_x_1"), val = bool(false)];
            bool var_1110_transpose_y_1 = const()[name = string("op_1110_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_1110_cast_fp16 = matmul(transpose_x = var_1110_transpose_x_1, transpose_y = var_1110_transpose_y_1, x = q_15_cast_fp16, y = k_15_cast_fp16)[name = string("op_1110_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_7_cast_fp16 = mul(x = var_1110_cast_fp16, y = D_7_cast_fp16)[name = string("A_7_cast_fp16")];
            tensor<fp16, [16]> var_1112_bias_0_to_fp16 = const()[name = string("op_1112_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_1112_cast_fp16 = linear(bias = var_1112_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_9_cast_fp16)[name = string("op_1112_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_5_cast_fp16 = exp(x = var_1112_cast_fp16)[name = string("cum_5_cast_fp16")];
            tensor<int32, [1]> var_1115_axes_0 = const()[name = string("op_1115_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_1115_cast_fp16 = expand_dims(axes = var_1115_axes_0, x = cum_5_cast_fp16)[name = string("op_1115_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1116_cast_fp16 = mul(x = kb_7_cast_fp16, y = var_1115_cast_fp16)[name = string("op_1116_cast_fp16")];
            bool w_7_transpose_x_0 = const()[name = string("w_7_transpose_x_0"), val = bool(false)];
            bool w_7_transpose_y_0 = const()[name = string("w_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_7_cast_fp16 = matmul(transpose_x = w_7_transpose_x_0, transpose_y = w_7_transpose_y_0, x = Minv_31_cast_fp16, y = var_1116_cast_fp16)[name = string("w_7_cast_fp16")];
            bool var_1118_transpose_x_0 = const()[name = string("op_1118_transpose_x_0"), val = bool(false)];
            bool var_1118_transpose_y_0 = const()[name = string("op_1118_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1118_cast_fp16 = matmul(transpose_x = var_1118_transpose_x_0, transpose_y = var_1118_transpose_y_0, x = w_7_cast_fp16, y = S_5_cast_fp16)[name = string("op_1118_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_5_cast_fp16 = sub(x = u_7_cast_fp16, y = var_1118_cast_fp16)[name = string("vnew_5_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1123_cast_fp16 = mul(x = q_15_cast_fp16, y = var_1115_cast_fp16)[name = string("op_1123_cast_fp16")];
            bool var_1124_transpose_x_0 = const()[name = string("op_1124_transpose_x_0"), val = bool(false)];
            bool var_1124_transpose_y_0 = const()[name = string("op_1124_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1124_cast_fp16 = matmul(transpose_x = var_1124_transpose_x_0, transpose_y = var_1124_transpose_y_0, x = var_1123_cast_fp16, y = S_5_cast_fp16)[name = string("op_1124_cast_fp16")];
            bool var_1125_transpose_x_0 = const()[name = string("op_1125_transpose_x_0"), val = bool(false)];
            bool var_1125_transpose_y_0 = const()[name = string("op_1125_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1125_cast_fp16 = matmul(transpose_x = var_1125_transpose_x_0, transpose_y = var_1125_transpose_y_0, x = A_7_cast_fp16, y = vnew_5_cast_fp16)[name = string("op_1125_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_1_cast_fp16 = add(x = var_1124_cast_fp16, y = var_1125_cast_fp16)[name = string("oc_1_cast_fp16")];
            int32 var_1129 = const()[name = string("op_1129"), val = int32(2)];
            bool o_1_interleave_0 = const()[name = string("o_1_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 64, 128]> o_1_cast_fp16 = concat(axis = var_1129, interleave = o_1_interleave_0, values = (var_672_cast_fp16, var_819_cast_fp16, var_973_cast_fp16, oc_1_cast_fp16))[name = string("o_1_cast_fp16")];
            tensor<int32, [4]> var_1135 = const()[name = string("op_1135"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1141 = const()[name = string("op_1141"), val = tensor<int32, [4]>([3, 64, 4, 128])];
            tensor<fp16, [3, 64, 4, 128]> input_7_cast_fp16 = reshape(shape = var_1141, x = z_1_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> o_3_cast_fp16 = transpose(perm = var_1135, x = o_1_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [3, 64, 4, 128]> var_1143_cast_fp16 = mul(x = o_3_cast_fp16, y = o_3_cast_fp16)[name = string("op_1143_cast_fp16")];
            tensor<int32, [1]> var_1148_axes_0 = const()[name = string("op_1148_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1148_keep_dims_0 = const()[name = string("op_1148_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 4, 1]> var_1148_cast_fp16 = reduce_mean(axes = var_1148_axes_0, keep_dims = var_1148_keep_dims_0, x = var_1143_cast_fp16)[name = string("op_1148_cast_fp16")];
            fp16 var_1150_to_fp16 = const()[name = string("op_1150_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 64, 4, 1]> var_1151_cast_fp16 = add(x = var_1148_cast_fp16, y = var_1150_to_fp16)[name = string("op_1151_cast_fp16")];
            fp32 var_1152_epsilon_0 = const()[name = string("op_1152_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 4, 1]> var_1152_cast_fp16 = rsqrt(epsilon = var_1152_epsilon_0, x = var_1151_cast_fp16)[name = string("op_1152_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> var_1153_cast_fp16 = mul(x = o_3_cast_fp16, y = var_1152_cast_fp16)[name = string("op_1153_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> var_1155_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1155_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> o_5_cast_fp16 = mul(x = var_1153_cast_fp16, y = var_1155_cast_fp16)[name = string("o_5_cast_fp16")];
            tensor<int32, [3]> var_1160 = const()[name = string("op_1160"), val = tensor<int32, [3]>([3, 64, 512])];
            tensor<fp16, [3, 64, 512]> input_9_cast_fp16 = reshape(shape = var_1160, x = o_5_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [640, 512]> layers_0_out_weight_to_fp16 = const()[name = string("layers_0_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22499136)))];
            tensor<fp16, [640]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23154560)))];
            tensor<fp16, [3, 64, 640]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_out_weight_to_fp16, x = input_9_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [3, 64, 640]> x_5_cast_fp16 = add(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = linear_1_cast_fp16)[name = string("x_5_cast_fp16")];
            fp16 var_1173_to_fp16 = const()[name = string("op_1173_to_fp16"), val = fp16(0x1p+5)];
            tensor<fp16, [3, 64, 640]> x_7_cast_fp16 = mul(x = x_5_cast_fp16, y = var_1173_to_fp16)[name = string("x_7_cast_fp16")];
            fp16 var_1171_promoted_to_fp16 = const()[name = string("op_1171_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> var_1175_cast_fp16 = pow(x = x_7_cast_fp16, y = var_1171_promoted_to_fp16)[name = string("op_1175_cast_fp16")];
            tensor<int32, [1]> var_1177_axes_0 = const()[name = string("op_1177_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1177_keep_dims_0 = const()[name = string("op_1177_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 1]> var_1177_cast_fp16 = reduce_mean(axes = var_1177_axes_0, keep_dims = var_1177_keep_dims_0, x = var_1175_cast_fp16)[name = string("op_1177_cast_fp16")];
            fp16 var_1178_to_fp16 = const()[name = string("op_1178_to_fp16"), val = fp16(0x1.0c8p-10)];
            tensor<fp16, [3, 64, 1]> var_1179_cast_fp16 = add(x = var_1177_cast_fp16, y = var_1178_to_fp16)[name = string("op_1179_cast_fp16")];
            fp32 var_1180_epsilon_0 = const()[name = string("op_1180_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 1]> var_1180_cast_fp16 = rsqrt(epsilon = var_1180_epsilon_0, x = var_1179_cast_fp16)[name = string("op_1180_cast_fp16")];
            tensor<fp16, [3, 64, 640]> var_1181_cast_fp16 = mul(x = x_7_cast_fp16, y = var_1180_cast_fp16)[name = string("op_1181_cast_fp16")];
            tensor<fp16, [640]> layers_0_n2_w1_to_fp16 = const()[name = string("layers_0_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23155904)))];
            tensor<fp16, [3, 64, 640]> input_11_cast_fp16 = mul(x = var_1181_cast_fp16, y = layers_0_n2_w1_to_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_0_gate_up_weight_to_fp16 = const()[name = string("layers_0_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23157248)))];
            tensor<fp16, [5120]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29710912)))];
            tensor<fp16, [3, 64, 5120]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_0_gate_up_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [3]> input_13_begin_0 = const()[name = string("input_13_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_13_end_0 = const()[name = string("input_13_end_0"), val = tensor<int32, [3]>([3, 64, 2560])];
            tensor<bool, [3]> input_13_end_mask_0 = const()[name = string("input_13_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 2560]> input_13_cast_fp16 = slice_by_index(begin = input_13_begin_0, end = input_13_end_0, end_mask = input_13_end_mask_0, x = linear_2_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [3, 64, 2560]> var_1191_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_1191_cast_fp16")];
            tensor<int32, [3]> var_1196_begin_0 = const()[name = string("op_1196_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_1196_end_0 = const()[name = string("op_1196_end_0"), val = tensor<int32, [3]>([3, 64, 5120])];
            tensor<bool, [3]> var_1196_end_mask_0 = const()[name = string("op_1196_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 2560]> var_1196_cast_fp16 = slice_by_index(begin = var_1196_begin_0, end = var_1196_end_0, end_mask = var_1196_end_mask_0, x = linear_2_cast_fp16)[name = string("op_1196_cast_fp16")];
            tensor<fp16, [3, 64, 2560]> input_15_cast_fp16 = mul(x = var_1191_cast_fp16, y = var_1196_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_0_down_weight_to_fp16 = const()[name = string("layers_0_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29721216)))];
            tensor<fp16, [3, 64, 640]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_down_weight_to_fp16, x = input_15_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [3, 64, 640]> x_9_cast_fp16 = add(x = x_5_cast_fp16, y = linear_3_cast_fp16)[name = string("x_9_cast_fp16")];
            fp16 var_1209_to_fp16 = const()[name = string("op_1209_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> x_11_cast_fp16 = mul(x = x_9_cast_fp16, y = var_1209_to_fp16)[name = string("x_11_cast_fp16")];
            fp16 var_1207_promoted_to_fp16 = const()[name = string("op_1207_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> var_1211_cast_fp16 = pow(x = x_11_cast_fp16, y = var_1207_promoted_to_fp16)[name = string("op_1211_cast_fp16")];
            tensor<int32, [1]> var_1213_axes_0 = const()[name = string("op_1213_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1213_keep_dims_0 = const()[name = string("op_1213_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 1]> var_1213_cast_fp16 = reduce_mean(axes = var_1213_axes_0, keep_dims = var_1213_keep_dims_0, x = var_1211_cast_fp16)[name = string("op_1213_cast_fp16")];
            fp16 var_1214_to_fp16 = const()[name = string("op_1214_to_fp16"), val = fp16(0x1.0cp-18)];
            tensor<fp16, [3, 64, 1]> var_1215_cast_fp16 = add(x = var_1213_cast_fp16, y = var_1214_to_fp16)[name = string("op_1215_cast_fp16")];
            fp32 var_1216_epsilon_0 = const()[name = string("op_1216_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 1]> var_1216_cast_fp16 = rsqrt(epsilon = var_1216_epsilon_0, x = var_1215_cast_fp16)[name = string("op_1216_cast_fp16")];
            tensor<fp16, [3, 64, 640]> var_1217_cast_fp16 = mul(x = x_11_cast_fp16, y = var_1216_cast_fp16)[name = string("op_1217_cast_fp16")];
            tensor<fp16, [640]> layers_1_n1_w1_to_fp16 = const()[name = string("layers_1_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32998080)))];
            tensor<fp16, [3, 64, 640]> input_17_cast_fp16 = mul(x = var_1217_cast_fp16, y = layers_1_n1_w1_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_1_qkvzba_weight_to_fp16 = const()[name = string("layers_1_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32999424)))];
            tensor<fp16, [3, 64, 1544]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_qkvzba_weight_to_fp16, x = input_17_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<int32, [3]> qkv_3_begin_0 = const()[name = string("qkv_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_3_end_0 = const()[name = string("qkv_3_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> qkv_3_end_mask_0 = const()[name = string("qkv_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 1024]> qkv_3_cast_fp16 = slice_by_index(begin = qkv_3_begin_0, end = qkv_3_end_0, end_mask = qkv_3_end_mask_0, x = linear_4_cast_fp16)[name = string("qkv_3_cast_fp16")];
            tensor<int32, [3]> z_3_begin_0 = const()[name = string("z_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_3_end_0 = const()[name = string("z_3_end_0"), val = tensor<int32, [3]>([3, 64, 1536])];
            tensor<bool, [3]> z_3_end_mask_0 = const()[name = string("z_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 512]> z_3_cast_fp16 = slice_by_index(begin = z_3_begin_0, end = z_3_end_0, end_mask = z_3_end_mask_0, x = linear_4_cast_fp16)[name = string("z_3_cast_fp16")];
            tensor<int32, [3]> b_3_begin_0 = const()[name = string("b_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_3_end_0 = const()[name = string("b_3_end_0"), val = tensor<int32, [3]>([3, 64, 1540])];
            tensor<bool, [3]> b_3_end_mask_0 = const()[name = string("b_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 4]> b_3_cast_fp16 = slice_by_index(begin = b_3_begin_0, end = b_3_end_0, end_mask = b_3_end_mask_0, x = linear_4_cast_fp16)[name = string("b_3_cast_fp16")];
            tensor<int32, [3]> a_3_begin_0 = const()[name = string("a_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_3_end_0 = const()[name = string("a_3_end_0"), val = tensor<int32, [3]>([3, 64, 1])];
            tensor<bool, [3]> a_3_end_mask_0 = const()[name = string("a_3_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 4]> a_3_cast_fp16 = slice_by_index(begin = a_3_begin_0, end = a_3_end_0, end_mask = a_3_end_mask_0, x = linear_4_cast_fp16)[name = string("a_3_cast_fp16")];
            int32 var_1256 = const()[name = string("op_1256"), val = int32(1)];
            bool xp_3_interleave_0 = const()[name = string("xp_3_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 67, 1024]> xp_3_cast_fp16 = concat(axis = var_1256, interleave = xp_3_interleave_0, values = (zpad_to_fp16, qkv_3_cast_fp16))[name = string("xp_3_cast_fp16")];
            tensor<int32, [3]> var_1267_begin_0 = const()[name = string("op_1267_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_1267_end_0 = const()[name = string("op_1267_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_1267_end_mask_0 = const()[name = string("op_1267_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 1024]> var_1267_cast_fp16 = slice_by_index(begin = var_1267_begin_0, end = var_1267_end_0, end_mask = var_1267_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1267_cast_fp16")];
            tensor<fp16, [1024]> var_1275_to_fp16 = const()[name = string("op_1275_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34975808)))];
            tensor<fp16, [3, 64, 1024]> var_1276_cast_fp16 = mul(x = var_1267_cast_fp16, y = var_1275_to_fp16)[name = string("op_1276_cast_fp16")];
            tensor<int32, [3]> var_1286_begin_0 = const()[name = string("op_1286_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_1286_end_0 = const()[name = string("op_1286_end_0"), val = tensor<int32, [3]>([3, 66, 1024])];
            tensor<bool, [3]> var_1286_end_mask_0 = const()[name = string("op_1286_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_1286_cast_fp16 = slice_by_index(begin = var_1286_begin_0, end = var_1286_end_0, end_mask = var_1286_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1286_cast_fp16")];
            tensor<fp16, [1024]> var_1294_to_fp16 = const()[name = string("op_1294_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34977920)))];
            tensor<fp16, [3, 64, 1024]> var_1295_cast_fp16 = mul(x = var_1286_cast_fp16, y = var_1294_to_fp16)[name = string("op_1295_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> var_1297_cast_fp16 = add(x = var_1276_cast_fp16, y = var_1295_cast_fp16)[name = string("op_1297_cast_fp16")];
            tensor<int32, [3]> var_1307_begin_0 = const()[name = string("op_1307_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1307_end_0 = const()[name = string("op_1307_end_0"), val = tensor<int32, [3]>([3, 65, 1024])];
            tensor<bool, [3]> var_1307_end_mask_0 = const()[name = string("op_1307_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_1307_cast_fp16 = slice_by_index(begin = var_1307_begin_0, end = var_1307_end_0, end_mask = var_1307_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1307_cast_fp16")];
            tensor<fp16, [1024]> var_1315_to_fp16 = const()[name = string("op_1315_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34980032)))];
            tensor<fp16, [3, 64, 1024]> var_1316_cast_fp16 = mul(x = var_1307_cast_fp16, y = var_1315_to_fp16)[name = string("op_1316_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> var_1318_cast_fp16 = add(x = var_1297_cast_fp16, y = var_1316_cast_fp16)[name = string("op_1318_cast_fp16")];
            tensor<int32, [3]> var_1328_begin_0 = const()[name = string("op_1328_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1328_end_0 = const()[name = string("op_1328_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> var_1328_end_mask_0 = const()[name = string("op_1328_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_1328_cast_fp16 = slice_by_index(begin = var_1328_begin_0, end = var_1328_end_0, end_mask = var_1328_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1328_cast_fp16")];
            tensor<fp16, [1024]> var_1336_to_fp16 = const()[name = string("op_1336_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34982144)))];
            tensor<fp16, [3, 64, 1024]> var_1337_cast_fp16 = mul(x = var_1328_cast_fp16, y = var_1336_to_fp16)[name = string("op_1337_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> input_19_cast_fp16 = add(x = var_1318_cast_fp16, y = var_1337_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> c_3_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("c_3_cast_fp16")];
            tensor<int32, [3]> var_1345_begin_0 = const()[name = string("op_1345_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1345_end_0 = const()[name = string("op_1345_end_0"), val = tensor<int32, [3]>([3, 64, 256])];
            tensor<bool, [3]> var_1345_end_mask_0 = const()[name = string("op_1345_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 256]> var_1345_cast_fp16 = slice_by_index(begin = var_1345_begin_0, end = var_1345_end_0, end_mask = var_1345_end_mask_0, x = c_3_cast_fp16)[name = string("op_1345_cast_fp16")];
            tensor<int32, [4]> var_1350 = const()[name = string("op_1350"), val = tensor<int32, [4]>([3, 64, 2, 128])];
            tensor<fp16, [3, 64, 2, 128]> q_17_cast_fp16 = reshape(shape = var_1350, x = var_1345_cast_fp16)[name = string("q_17_cast_fp16")];
            tensor<int32, [3]> var_1356_begin_0 = const()[name = string("op_1356_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_1356_end_0 = const()[name = string("op_1356_end_0"), val = tensor<int32, [3]>([3, 64, 512])];
            tensor<bool, [3]> var_1356_end_mask_0 = const()[name = string("op_1356_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 256]> var_1356_cast_fp16 = slice_by_index(begin = var_1356_begin_0, end = var_1356_end_0, end_mask = var_1356_end_mask_0, x = c_3_cast_fp16)[name = string("op_1356_cast_fp16")];
            tensor<int32, [4]> var_1361 = const()[name = string("op_1361"), val = tensor<int32, [4]>([3, 64, 2, 128])];
            tensor<fp16, [3, 64, 2, 128]> k_17_cast_fp16 = reshape(shape = var_1361, x = var_1356_cast_fp16)[name = string("k_17_cast_fp16")];
            tensor<int32, [3]> var_1367_begin_0 = const()[name = string("op_1367_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_1367_end_0 = const()[name = string("op_1367_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> var_1367_end_mask_0 = const()[name = string("op_1367_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 512]> var_1367_cast_fp16 = slice_by_index(begin = var_1367_begin_0, end = var_1367_end_0, end_mask = var_1367_end_mask_0, x = c_3_cast_fp16)[name = string("op_1367_cast_fp16")];
            tensor<int32, [4]> var_1372 = const()[name = string("op_1372"), val = tensor<int32, [4]>([3, 64, 4, 128])];
            tensor<fp16, [3, 64, 4, 128]> var_1373_cast_fp16 = reshape(shape = var_1372, x = var_1367_cast_fp16)[name = string("op_1373_cast_fp16")];
            tensor<int32, [4]> var_1378 = const()[name = string("op_1378"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1380_to_fp16 = const()[name = string("op_1380_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 64, 2, 128]> q_19_cast_fp16 = mul(x = q_17_cast_fp16, y = var_1380_to_fp16)[name = string("q_19_cast_fp16")];
            fp16 var_1382_to_fp16 = const()[name = string("op_1382_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 64, 2, 128]> k_19_cast_fp16 = mul(x = k_17_cast_fp16, y = var_1382_to_fp16)[name = string("k_19_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_1384_cast_fp16 = mul(x = q_19_cast_fp16, y = q_19_cast_fp16)[name = string("op_1384_cast_fp16")];
            tensor<int32, [1]> var_1389_axes_0 = const()[name = string("op_1389_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1389_keep_dims_0 = const()[name = string("op_1389_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 2, 1]> var_1389_cast_fp16 = reduce_mean(axes = var_1389_axes_0, keep_dims = var_1389_keep_dims_0, x = var_1384_cast_fp16)[name = string("op_1389_cast_fp16")];
            fp16 var_1390_promoted_to_fp16 = const()[name = string("op_1390_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 64, 2, 1]> var_1391_cast_fp16 = mul(x = var_1389_cast_fp16, y = var_1390_promoted_to_fp16)[name = string("op_1391_cast_fp16")];
            fp16 var_1393_to_fp16 = const()[name = string("op_1393_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 64, 2, 1]> var_1394_cast_fp16 = add(x = var_1391_cast_fp16, y = var_1393_to_fp16)[name = string("op_1394_cast_fp16")];
            fp32 var_1395_epsilon_0 = const()[name = string("op_1395_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 2, 1]> var_1395_cast_fp16 = rsqrt(epsilon = var_1395_epsilon_0, x = var_1394_cast_fp16)[name = string("op_1395_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_1396_cast_fp16 = mul(x = q_19_cast_fp16, y = var_1395_cast_fp16)[name = string("op_1396_cast_fp16")];
            fp16 var_1397_to_fp16 = const()[name = string("op_1397_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 64, 2, 128]> q_21_cast_fp16 = mul(x = var_1396_cast_fp16, y = var_1397_to_fp16)[name = string("q_21_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_1399_cast_fp16 = mul(x = k_19_cast_fp16, y = k_19_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [1]> var_1404_axes_0 = const()[name = string("op_1404_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1404_keep_dims_0 = const()[name = string("op_1404_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 2, 1]> var_1404_cast_fp16 = reduce_mean(axes = var_1404_axes_0, keep_dims = var_1404_keep_dims_0, x = var_1399_cast_fp16)[name = string("op_1404_cast_fp16")];
            fp16 var_1405_promoted_to_fp16 = const()[name = string("op_1405_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 64, 2, 1]> var_1406_cast_fp16 = mul(x = var_1404_cast_fp16, y = var_1405_promoted_to_fp16)[name = string("op_1406_cast_fp16")];
            fp16 var_1408_to_fp16 = const()[name = string("op_1408_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 64, 2, 1]> var_1409_cast_fp16 = add(x = var_1406_cast_fp16, y = var_1408_to_fp16)[name = string("op_1409_cast_fp16")];
            fp32 var_1410_epsilon_0 = const()[name = string("op_1410_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 2, 1]> var_1410_cast_fp16 = rsqrt(epsilon = var_1410_epsilon_0, x = var_1409_cast_fp16)[name = string("op_1410_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> k_21_cast_fp16 = mul(x = k_19_cast_fp16, y = var_1410_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<int32, [4]> var_1416 = const()[name = string("op_1416"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1429_axes_0 = const()[name = string("op_1429_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 64, 128]> var_1417_cast_fp16 = transpose(perm = var_1416, x = q_21_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [3, 2, 1, 64, 128]> var_1429_cast_fp16 = expand_dims(axes = var_1429_axes_0, x = var_1417_cast_fp16)[name = string("op_1429_cast_fp16")];
            tensor<int32, [5]> var_1437_reps_0 = const()[name = string("op_1437_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 64, 128]> var_1437_cast_fp16 = tile(reps = var_1437_reps_0, x = var_1429_cast_fp16)[name = string("op_1437_cast_fp16")];
            tensor<int32, [4]> var_1442 = const()[name = string("op_1442"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<fp16, [3, 4, 64, 128]> q_23_cast_fp16 = reshape(shape = var_1442, x = var_1437_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<int32, [4]> var_1448 = const()[name = string("op_1448"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1461_axes_0 = const()[name = string("op_1461_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 64, 128]> var_1449_cast_fp16 = transpose(perm = var_1448, x = k_21_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [3, 2, 1, 64, 128]> var_1461_cast_fp16 = expand_dims(axes = var_1461_axes_0, x = var_1449_cast_fp16)[name = string("op_1461_cast_fp16")];
            tensor<int32, [5]> var_1469_reps_0 = const()[name = string("op_1469_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 64, 128]> var_1469_cast_fp16 = tile(reps = var_1469_reps_0, x = var_1461_cast_fp16)[name = string("op_1469_cast_fp16")];
            tensor<int32, [4]> var_1474 = const()[name = string("op_1474"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<fp16, [3, 4, 64, 128]> k_23_cast_fp16 = reshape(shape = var_1474, x = var_1469_cast_fp16)[name = string("k_23_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_1476_cast_fp16 = sigmoid(x = b_3_cast_fp16)[name = string("op_1476_cast_fp16")];
            tensor<int32, [3]> var_1480 = const()[name = string("op_1480"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_1_dt_bias_to_fp16 = const()[name = string("layers_1_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 64, 4]> input_21_cast_fp16 = add(x = a_3_cast_fp16, y = layers_1_dt_bias_to_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_1484_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("op_1484_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_1485_cast_fp16 = abs(x = input_21_cast_fp16)[name = string("op_1485_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 64, 4]> var_1486_cast_fp16 = mul(x = var_1485_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_1486_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_1487_cast_fp16 = exp(x = var_1486_cast_fp16)[name = string("op_1487_cast_fp16")];
            fp16 var_1489_to_fp16 = const()[name = string("op_1489_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 64, 4]> var_1490_cast_fp16 = add(x = var_1487_cast_fp16, y = var_1489_to_fp16)[name = string("op_1490_cast_fp16")];
            fp32 var_1491_epsilon_0 = const()[name = string("op_1491_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 64, 4]> var_1491_cast_fp16 = log(epsilon = var_1491_epsilon_0, x = var_1490_cast_fp16)[name = string("op_1491_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_1493_cast_fp16 = add(x = var_1484_cast_fp16, y = var_1491_cast_fp16)[name = string("op_1493_cast_fp16")];
            tensor<fp16, [4]> layers_1_negA_to_fp16 = const()[name = string("layers_1_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.ff4p+2, -0x1.8cp+3, -0x1.c0cp+3, -0x1.104p+3])];
            tensor<fp16, [3, 64, 4]> var_1494_cast_fp16 = mul(x = layers_1_negA_to_fp16, y = var_1493_cast_fp16)[name = string("op_1494_cast_fp16")];
            tensor<int32, [3]> var_1498 = const()[name = string("op_1498"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_1501_axes_0 = const()[name = string("op_1501_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 64]> beta_3_cast_fp16 = transpose(perm = var_1480, x = var_1476_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [3, 4, 64, 1]> var_1501_cast_fp16 = expand_dims(axes = var_1501_axes_0, x = beta_3_cast_fp16)[name = string("op_1501_cast_fp16")];
            fp16 var_1502_to_fp16 = const()[name = string("op_1502_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 64, 1]> var_1503_cast_fp16 = mul(x = var_1501_cast_fp16, y = var_1502_to_fp16)[name = string("op_1503_cast_fp16")];
            tensor<fp16, [3, 4, 64, 128]> v_3_cast_fp16 = transpose(perm = var_1378, x = var_1373_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [3, 4, 64, 128]> vb_11_cast_fp16 = mul(x = v_3_cast_fp16, y = var_1503_cast_fp16)[name = string("vb_11_cast_fp16")];
            tensor<int32, [4]> q_25_begin_0 = const()[name = string("q_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_25_end_0 = const()[name = string("q_25_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_25_end_mask_0 = const()[name = string("q_25_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, x = q_23_cast_fp16)[name = string("q_25_cast_fp16")];
            tensor<int32, [4]> k_25_begin_0 = const()[name = string("k_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_25_end_0 = const()[name = string("k_25_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_25_end_mask_0 = const()[name = string("k_25_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, x = k_23_cast_fp16)[name = string("k_25_cast_fp16")];
            tensor<fp16, [3, 4, 64, 128]> var_1537_cast_fp16 = mul(x = k_23_cast_fp16, y = var_1501_cast_fp16)[name = string("op_1537_cast_fp16")];
            tensor<int32, [4]> kb_9_begin_0 = const()[name = string("kb_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_9_end_0 = const()[name = string("kb_9_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_9_end_mask_0 = const()[name = string("kb_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_9_cast_fp16 = slice_by_index(begin = kb_9_begin_0, end = kb_9_end_0, end_mask = kb_9_end_mask_0, x = var_1537_cast_fp16)[name = string("kb_9_cast_fp16")];
            tensor<int32, [4]> vb_13_begin_0 = const()[name = string("vb_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_13_end_0 = const()[name = string("vb_13_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_13_end_mask_0 = const()[name = string("vb_13_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_13_cast_fp16 = slice_by_index(begin = vb_13_begin_0, end = vb_13_end_0, end_mask = vb_13_end_mask_0, x = vb_11_cast_fp16)[name = string("vb_13_cast_fp16")];
            tensor<int32, [3]> gdec_13_begin_0 = const()[name = string("gdec_13_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_13_end_0 = const()[name = string("gdec_13_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_13_end_mask_0 = const()[name = string("gdec_13_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 64]> gdec_11_cast_fp16 = transpose(perm = var_1498, x = var_1494_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [3, 4, 16]> gdec_13_cast_fp16 = slice_by_index(begin = gdec_13_begin_0, end = gdec_13_end_0, end_mask = gdec_13_end_mask_0, x = gdec_11_cast_fp16)[name = string("gdec_13_cast_fp16")];
            tensor<int32, [2]> var_1585 = const()[name = string("op_1585"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_1586_cast_fp16 = reshape(shape = var_1585, x = gdec_13_cast_fp16)[name = string("op_1586_cast_fp16")];
            tensor<fp16, [12, 256]> var_1587_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1586_cast_fp16)[name = string("op_1587_cast_fp16")];
            tensor<int32, [4]> var_1592 = const()[name = string("op_1592"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_9_cast_fp16 = reshape(shape = var_1592, x = var_1587_cast_fp16)[name = string("P_9_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1594_cast_fp16 = exp(x = P_9_cast_fp16)[name = string("op_1594_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_9_cast_fp16 = mul(x = var_1594_cast_fp16, y = tril_incl_to_fp16)[name = string("D_9_cast_fp16")];
            bool kk_9_transpose_x_1 = const()[name = string("kk_9_transpose_x_1"), val = bool(false)];
            bool kk_9_transpose_y_1 = const()[name = string("kk_9_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_9_cast_fp16 = matmul(transpose_x = kk_9_transpose_x_1, transpose_y = kk_9_transpose_y_1, x = kb_9_cast_fp16, y = k_25_cast_fp16)[name = string("kk_9_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_1600_cast_fp16 = mul(x = kk_9_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_1600_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1601_cast_fp16 = mul(x = var_1600_cast_fp16, y = D_9_cast_fp16)[name = string("op_1601_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_9_cast_fp16 = mul(x = var_1601_cast_fp16, y = cum_tail_to_fp16)[name = string("N_9_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1603_cast_fp16 = mul(x = D_9_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1603_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_9_cast_fp16 = mul(x = kk_9_cast_fp16, y = var_1603_cast_fp16)[name = string("Na_9_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_33_cast_fp16 = add(x = eye_to_fp16, y = N_9_cast_fp16)[name = string("Minv_33_cast_fp16")];
            bool Np_25_transpose_x_0 = const()[name = string("Np_25_transpose_x_0"), val = bool(false)];
            bool Np_25_transpose_y_0 = const()[name = string("Np_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_25_cast_fp16 = matmul(transpose_x = Np_25_transpose_x_0, transpose_y = Np_25_transpose_y_0, x = N_9_cast_fp16, y = Na_9_cast_fp16)[name = string("Np_25_cast_fp16")];
            bool Npa_17_transpose_x_0 = const()[name = string("Npa_17_transpose_x_0"), val = bool(false)];
            bool Npa_17_transpose_y_0 = const()[name = string("Npa_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_17_cast_fp16 = matmul(transpose_x = Npa_17_transpose_x_0, transpose_y = Npa_17_transpose_y_0, x = Na_9_cast_fp16, y = N_9_cast_fp16)[name = string("Npa_17_cast_fp16")];
            bool var_1609_transpose_x_0 = const()[name = string("op_1609_transpose_x_0"), val = bool(false)];
            bool var_1609_transpose_y_0 = const()[name = string("op_1609_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1609_cast_fp16 = matmul(transpose_x = var_1609_transpose_x_0, transpose_y = var_1609_transpose_y_0, x = Minv_33_cast_fp16, y = Np_25_cast_fp16)[name = string("op_1609_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_35_cast_fp16 = add(x = Minv_33_cast_fp16, y = var_1609_cast_fp16)[name = string("Minv_35_cast_fp16")];
            bool Np_27_transpose_x_0 = const()[name = string("Np_27_transpose_x_0"), val = bool(false)];
            bool Np_27_transpose_y_0 = const()[name = string("Np_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_27_cast_fp16 = matmul(transpose_x = Np_27_transpose_x_0, transpose_y = Np_27_transpose_y_0, x = Np_25_cast_fp16, y = Npa_17_cast_fp16)[name = string("Np_27_cast_fp16")];
            bool Npa_19_transpose_x_0 = const()[name = string("Npa_19_transpose_x_0"), val = bool(false)];
            bool Npa_19_transpose_y_0 = const()[name = string("Npa_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_19_cast_fp16 = matmul(transpose_x = Npa_19_transpose_x_0, transpose_y = Npa_19_transpose_y_0, x = Npa_17_cast_fp16, y = Np_25_cast_fp16)[name = string("Npa_19_cast_fp16")];
            bool var_1614_transpose_x_0 = const()[name = string("op_1614_transpose_x_0"), val = bool(false)];
            bool var_1614_transpose_y_0 = const()[name = string("op_1614_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1614_cast_fp16 = matmul(transpose_x = var_1614_transpose_x_0, transpose_y = var_1614_transpose_y_0, x = Minv_35_cast_fp16, y = Np_27_cast_fp16)[name = string("op_1614_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_37_cast_fp16 = add(x = Minv_35_cast_fp16, y = var_1614_cast_fp16)[name = string("Minv_37_cast_fp16")];
            bool Np_29_transpose_x_0 = const()[name = string("Np_29_transpose_x_0"), val = bool(false)];
            bool Np_29_transpose_y_0 = const()[name = string("Np_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_29_cast_fp16 = matmul(transpose_x = Np_29_transpose_x_0, transpose_y = Np_29_transpose_y_0, x = Np_27_cast_fp16, y = Npa_19_cast_fp16)[name = string("Np_29_cast_fp16")];
            bool var_1618_transpose_x_0 = const()[name = string("op_1618_transpose_x_0"), val = bool(false)];
            bool var_1618_transpose_y_0 = const()[name = string("op_1618_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1618_cast_fp16 = matmul(transpose_x = var_1618_transpose_x_0, transpose_y = var_1618_transpose_y_0, x = Minv_37_cast_fp16, y = Np_29_cast_fp16)[name = string("op_1618_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_39_cast_fp16 = add(x = Minv_37_cast_fp16, y = var_1618_cast_fp16)[name = string("Minv_39_cast_fp16")];
            bool u_9_transpose_x_0 = const()[name = string("u_9_transpose_x_0"), val = bool(false)];
            bool u_9_transpose_y_0 = const()[name = string("u_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_9_cast_fp16 = matmul(transpose_x = u_9_transpose_x_0, transpose_y = u_9_transpose_y_0, x = Minv_39_cast_fp16, y = vb_13_cast_fp16)[name = string("u_9_cast_fp16")];
            bool var_1625_transpose_x_1 = const()[name = string("op_1625_transpose_x_1"), val = bool(false)];
            bool var_1625_transpose_y_1 = const()[name = string("op_1625_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_1625_cast_fp16 = matmul(transpose_x = var_1625_transpose_x_1, transpose_y = var_1625_transpose_y_1, x = q_25_cast_fp16, y = k_25_cast_fp16)[name = string("op_1625_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_9_cast_fp16 = mul(x = var_1625_cast_fp16, y = D_9_cast_fp16)[name = string("A_9_cast_fp16")];
            bool var_1627_transpose_x_0 = const()[name = string("op_1627_transpose_x_0"), val = bool(false)];
            bool var_1627_transpose_y_0 = const()[name = string("op_1627_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1627_cast_fp16 = matmul(transpose_x = var_1627_transpose_x_0, transpose_y = var_1627_transpose_y_0, x = A_9_cast_fp16, y = u_9_cast_fp16)[name = string("op_1627_cast_fp16")];
            tensor<fp16, [16]> var_1628_bias_0_to_fp16 = const()[name = string("op_1628_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_1628_cast_fp16 = linear(bias = var_1628_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_13_cast_fp16)[name = string("op_1628_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_7_cast_fp16 = exp(x = var_1628_cast_fp16)[name = string("tail_7_cast_fp16")];
            tensor<int32, [1]> var_1631_axes_0 = const()[name = string("op_1631_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_1631_cast_fp16 = expand_dims(axes = var_1631_axes_0, x = tail_7_cast_fp16)[name = string("op_1631_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1632_cast_fp16 = mul(x = k_25_cast_fp16, y = var_1631_cast_fp16)[name = string("op_1632_cast_fp16")];
            bool S_7_transpose_x_1 = const()[name = string("S_7_transpose_x_1"), val = bool(true)];
            bool S_7_transpose_y_1 = const()[name = string("S_7_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_7_cast_fp16 = matmul(transpose_x = S_7_transpose_x_1, transpose_y = S_7_transpose_y_1, x = var_1632_cast_fp16, y = u_9_cast_fp16)[name = string("S_7_cast_fp16")];
            tensor<int32, [4]> q_27_begin_0 = const()[name = string("q_27_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_27_end_0 = const()[name = string("q_27_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> q_27_end_mask_0 = const()[name = string("q_27_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_27_cast_fp16 = slice_by_index(begin = q_27_begin_0, end = q_27_end_0, end_mask = q_27_end_mask_0, x = q_23_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<int32, [4]> k_27_begin_0 = const()[name = string("k_27_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_27_end_0 = const()[name = string("k_27_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> k_27_end_mask_0 = const()[name = string("k_27_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_27_cast_fp16 = slice_by_index(begin = k_27_begin_0, end = k_27_end_0, end_mask = k_27_end_mask_0, x = k_23_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<int32, [4]> kb_11_begin_0 = const()[name = string("kb_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_11_end_0 = const()[name = string("kb_11_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> kb_11_end_mask_0 = const()[name = string("kb_11_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_11_cast_fp16 = slice_by_index(begin = kb_11_begin_0, end = kb_11_end_0, end_mask = kb_11_end_mask_0, x = var_1537_cast_fp16)[name = string("kb_11_cast_fp16")];
            tensor<int32, [4]> vb_15_begin_0 = const()[name = string("vb_15_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_15_end_0 = const()[name = string("vb_15_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> vb_15_end_mask_0 = const()[name = string("vb_15_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_15_cast_fp16 = slice_by_index(begin = vb_15_begin_0, end = vb_15_end_0, end_mask = vb_15_end_mask_0, x = vb_11_cast_fp16)[name = string("vb_15_cast_fp16")];
            tensor<int32, [3]> gdec_15_begin_0 = const()[name = string("gdec_15_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_15_end_0 = const()[name = string("gdec_15_end_0"), val = tensor<int32, [3]>([3, 4, 32])];
            tensor<bool, [3]> gdec_15_end_mask_0 = const()[name = string("gdec_15_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_15_cast_fp16 = slice_by_index(begin = gdec_15_begin_0, end = gdec_15_end_0, end_mask = gdec_15_end_mask_0, x = gdec_11_cast_fp16)[name = string("gdec_15_cast_fp16")];
            tensor<int32, [2]> var_1717 = const()[name = string("op_1717"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_1718_cast_fp16 = reshape(shape = var_1717, x = gdec_15_cast_fp16)[name = string("op_1718_cast_fp16")];
            tensor<fp16, [12, 256]> var_1719_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1718_cast_fp16)[name = string("op_1719_cast_fp16")];
            tensor<int32, [4]> var_1724 = const()[name = string("op_1724"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_11_cast_fp16 = reshape(shape = var_1724, x = var_1719_cast_fp16)[name = string("P_11_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1726_cast_fp16 = exp(x = P_11_cast_fp16)[name = string("op_1726_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_11_cast_fp16 = mul(x = var_1726_cast_fp16, y = tril_incl_to_fp16)[name = string("D_11_cast_fp16")];
            bool kk_11_transpose_x_1 = const()[name = string("kk_11_transpose_x_1"), val = bool(false)];
            bool kk_11_transpose_y_1 = const()[name = string("kk_11_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_11_cast_fp16 = matmul(transpose_x = kk_11_transpose_x_1, transpose_y = kk_11_transpose_y_1, x = kb_11_cast_fp16, y = k_27_cast_fp16)[name = string("kk_11_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_1732_cast_fp16 = mul(x = kk_11_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_1732_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1733_cast_fp16 = mul(x = var_1732_cast_fp16, y = D_11_cast_fp16)[name = string("op_1733_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_11_cast_fp16 = mul(x = var_1733_cast_fp16, y = cum_tail_to_fp16)[name = string("N_11_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1735_cast_fp16 = mul(x = D_11_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1735_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_11_cast_fp16 = mul(x = kk_11_cast_fp16, y = var_1735_cast_fp16)[name = string("Na_11_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_41_cast_fp16 = add(x = eye_to_fp16, y = N_11_cast_fp16)[name = string("Minv_41_cast_fp16")];
            bool Np_31_transpose_x_0 = const()[name = string("Np_31_transpose_x_0"), val = bool(false)];
            bool Np_31_transpose_y_0 = const()[name = string("Np_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_31_cast_fp16 = matmul(transpose_x = Np_31_transpose_x_0, transpose_y = Np_31_transpose_y_0, x = N_11_cast_fp16, y = Na_11_cast_fp16)[name = string("Np_31_cast_fp16")];
            bool Npa_21_transpose_x_0 = const()[name = string("Npa_21_transpose_x_0"), val = bool(false)];
            bool Npa_21_transpose_y_0 = const()[name = string("Npa_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_21_cast_fp16 = matmul(transpose_x = Npa_21_transpose_x_0, transpose_y = Npa_21_transpose_y_0, x = Na_11_cast_fp16, y = N_11_cast_fp16)[name = string("Npa_21_cast_fp16")];
            bool var_1741_transpose_x_0 = const()[name = string("op_1741_transpose_x_0"), val = bool(false)];
            bool var_1741_transpose_y_0 = const()[name = string("op_1741_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1741_cast_fp16 = matmul(transpose_x = var_1741_transpose_x_0, transpose_y = var_1741_transpose_y_0, x = Minv_41_cast_fp16, y = Np_31_cast_fp16)[name = string("op_1741_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_43_cast_fp16 = add(x = Minv_41_cast_fp16, y = var_1741_cast_fp16)[name = string("Minv_43_cast_fp16")];
            bool Np_33_transpose_x_0 = const()[name = string("Np_33_transpose_x_0"), val = bool(false)];
            bool Np_33_transpose_y_0 = const()[name = string("Np_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_33_cast_fp16 = matmul(transpose_x = Np_33_transpose_x_0, transpose_y = Np_33_transpose_y_0, x = Np_31_cast_fp16, y = Npa_21_cast_fp16)[name = string("Np_33_cast_fp16")];
            bool Npa_23_transpose_x_0 = const()[name = string("Npa_23_transpose_x_0"), val = bool(false)];
            bool Npa_23_transpose_y_0 = const()[name = string("Npa_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_23_cast_fp16 = matmul(transpose_x = Npa_23_transpose_x_0, transpose_y = Npa_23_transpose_y_0, x = Npa_21_cast_fp16, y = Np_31_cast_fp16)[name = string("Npa_23_cast_fp16")];
            bool var_1746_transpose_x_0 = const()[name = string("op_1746_transpose_x_0"), val = bool(false)];
            bool var_1746_transpose_y_0 = const()[name = string("op_1746_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1746_cast_fp16 = matmul(transpose_x = var_1746_transpose_x_0, transpose_y = var_1746_transpose_y_0, x = Minv_43_cast_fp16, y = Np_33_cast_fp16)[name = string("op_1746_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_45_cast_fp16 = add(x = Minv_43_cast_fp16, y = var_1746_cast_fp16)[name = string("Minv_45_cast_fp16")];
            bool Np_35_transpose_x_0 = const()[name = string("Np_35_transpose_x_0"), val = bool(false)];
            bool Np_35_transpose_y_0 = const()[name = string("Np_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_35_cast_fp16 = matmul(transpose_x = Np_35_transpose_x_0, transpose_y = Np_35_transpose_y_0, x = Np_33_cast_fp16, y = Npa_23_cast_fp16)[name = string("Np_35_cast_fp16")];
            bool var_1750_transpose_x_0 = const()[name = string("op_1750_transpose_x_0"), val = bool(false)];
            bool var_1750_transpose_y_0 = const()[name = string("op_1750_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1750_cast_fp16 = matmul(transpose_x = var_1750_transpose_x_0, transpose_y = var_1750_transpose_y_0, x = Minv_45_cast_fp16, y = Np_35_cast_fp16)[name = string("op_1750_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_47_cast_fp16 = add(x = Minv_45_cast_fp16, y = var_1750_cast_fp16)[name = string("Minv_47_cast_fp16")];
            bool u_11_transpose_x_0 = const()[name = string("u_11_transpose_x_0"), val = bool(false)];
            bool u_11_transpose_y_0 = const()[name = string("u_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_11_cast_fp16 = matmul(transpose_x = u_11_transpose_x_0, transpose_y = u_11_transpose_y_0, x = Minv_47_cast_fp16, y = vb_15_cast_fp16)[name = string("u_11_cast_fp16")];
            bool var_1757_transpose_x_1 = const()[name = string("op_1757_transpose_x_1"), val = bool(false)];
            bool var_1757_transpose_y_1 = const()[name = string("op_1757_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_1757_cast_fp16 = matmul(transpose_x = var_1757_transpose_x_1, transpose_y = var_1757_transpose_y_1, x = q_27_cast_fp16, y = k_27_cast_fp16)[name = string("op_1757_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_11_cast_fp16 = mul(x = var_1757_cast_fp16, y = D_11_cast_fp16)[name = string("A_11_cast_fp16")];
            tensor<fp16, [16]> var_1759_bias_0_to_fp16 = const()[name = string("op_1759_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_1759_cast_fp16 = linear(bias = var_1759_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_15_cast_fp16)[name = string("op_1759_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_7_cast_fp16 = exp(x = var_1759_cast_fp16)[name = string("cum_7_cast_fp16")];
            tensor<int32, [1]> var_1762_axes_0 = const()[name = string("op_1762_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_1762_cast_fp16 = expand_dims(axes = var_1762_axes_0, x = cum_7_cast_fp16)[name = string("op_1762_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1763_cast_fp16 = mul(x = kb_11_cast_fp16, y = var_1762_cast_fp16)[name = string("op_1763_cast_fp16")];
            bool w_11_transpose_x_0 = const()[name = string("w_11_transpose_x_0"), val = bool(false)];
            bool w_11_transpose_y_0 = const()[name = string("w_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_11_cast_fp16 = matmul(transpose_x = w_11_transpose_x_0, transpose_y = w_11_transpose_y_0, x = Minv_47_cast_fp16, y = var_1763_cast_fp16)[name = string("w_11_cast_fp16")];
            bool var_1765_transpose_x_0 = const()[name = string("op_1765_transpose_x_0"), val = bool(false)];
            bool var_1765_transpose_y_0 = const()[name = string("op_1765_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1765_cast_fp16 = matmul(transpose_x = var_1765_transpose_x_0, transpose_y = var_1765_transpose_y_0, x = w_11_cast_fp16, y = S_7_cast_fp16)[name = string("op_1765_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_7_cast_fp16 = sub(x = u_11_cast_fp16, y = var_1765_cast_fp16)[name = string("vnew_7_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1770_cast_fp16 = mul(x = q_27_cast_fp16, y = var_1762_cast_fp16)[name = string("op_1770_cast_fp16")];
            bool var_1771_transpose_x_0 = const()[name = string("op_1771_transpose_x_0"), val = bool(false)];
            bool var_1771_transpose_y_0 = const()[name = string("op_1771_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1771_cast_fp16 = matmul(transpose_x = var_1771_transpose_x_0, transpose_y = var_1771_transpose_y_0, x = var_1770_cast_fp16, y = S_7_cast_fp16)[name = string("op_1771_cast_fp16")];
            bool var_1772_transpose_x_0 = const()[name = string("op_1772_transpose_x_0"), val = bool(false)];
            bool var_1772_transpose_y_0 = const()[name = string("op_1772_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1772_cast_fp16 = matmul(transpose_x = var_1772_transpose_x_0, transpose_y = var_1772_transpose_y_0, x = A_11_cast_fp16, y = vnew_7_cast_fp16)[name = string("op_1772_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1774_cast_fp16 = add(x = var_1771_cast_fp16, y = var_1772_cast_fp16)[name = string("op_1774_cast_fp16")];
            tensor<fp16, [16]> var_1775_bias_0_to_fp16 = const()[name = string("op_1775_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_1775_cast_fp16 = linear(bias = var_1775_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_15_cast_fp16)[name = string("op_1775_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_9_cast_fp16 = exp(x = var_1775_cast_fp16)[name = string("tail_9_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_17_to_fp16 = const()[name = string("transpose_17_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_1777_bias_0_to_fp16 = const()[name = string("op_1777_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_1777_cast_fp16 = linear(bias = var_1777_bias_0_to_fp16, weight = transpose_17_to_fp16, x = gdec_15_cast_fp16)[name = string("op_1777_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_1778_cast_fp16 = exp(x = var_1777_cast_fp16)[name = string("op_1778_cast_fp16")];
            tensor<int32, [1]> last_5_axes_0 = const()[name = string("last_5_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_5_cast_fp16 = expand_dims(axes = last_5_axes_0, x = var_1778_cast_fp16)[name = string("last_5_cast_fp16")];
            tensor<int32, [1]> var_1782_axes_0 = const()[name = string("op_1782_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_1782_cast_fp16 = expand_dims(axes = var_1782_axes_0, x = tail_9_cast_fp16)[name = string("op_1782_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1783_cast_fp16 = mul(x = k_27_cast_fp16, y = var_1782_cast_fp16)[name = string("op_1783_cast_fp16")];
            bool upd_5_transpose_x_1 = const()[name = string("upd_5_transpose_x_1"), val = bool(true)];
            bool upd_5_transpose_y_1 = const()[name = string("upd_5_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_5_cast_fp16 = matmul(transpose_x = upd_5_transpose_x_1, transpose_y = upd_5_transpose_y_1, x = var_1783_cast_fp16, y = vnew_7_cast_fp16)[name = string("upd_5_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_1788_cast_fp16 = mul(x = S_7_cast_fp16, y = last_5_cast_fp16)[name = string("op_1788_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_9_cast_fp16 = add(x = var_1788_cast_fp16, y = upd_5_cast_fp16)[name = string("S_9_cast_fp16")];
            tensor<int32, [4]> q_29_begin_0 = const()[name = string("q_29_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_29_end_0 = const()[name = string("q_29_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> q_29_end_mask_0 = const()[name = string("q_29_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_29_cast_fp16 = slice_by_index(begin = q_29_begin_0, end = q_29_end_0, end_mask = q_29_end_mask_0, x = q_23_cast_fp16)[name = string("q_29_cast_fp16")];
            tensor<int32, [4]> k_29_begin_0 = const()[name = string("k_29_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_29_end_0 = const()[name = string("k_29_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> k_29_end_mask_0 = const()[name = string("k_29_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, x = k_23_cast_fp16)[name = string("k_29_cast_fp16")];
            tensor<int32, [4]> kb_13_begin_0 = const()[name = string("kb_13_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_13_end_0 = const()[name = string("kb_13_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> kb_13_end_mask_0 = const()[name = string("kb_13_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_13_cast_fp16 = slice_by_index(begin = kb_13_begin_0, end = kb_13_end_0, end_mask = kb_13_end_mask_0, x = var_1537_cast_fp16)[name = string("kb_13_cast_fp16")];
            tensor<int32, [4]> vb_17_begin_0 = const()[name = string("vb_17_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_17_end_0 = const()[name = string("vb_17_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> vb_17_end_mask_0 = const()[name = string("vb_17_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_17_cast_fp16 = slice_by_index(begin = vb_17_begin_0, end = vb_17_end_0, end_mask = vb_17_end_mask_0, x = vb_11_cast_fp16)[name = string("vb_17_cast_fp16")];
            tensor<int32, [3]> gdec_17_begin_0 = const()[name = string("gdec_17_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_17_end_0 = const()[name = string("gdec_17_end_0"), val = tensor<int32, [3]>([3, 4, 48])];
            tensor<bool, [3]> gdec_17_end_mask_0 = const()[name = string("gdec_17_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_17_cast_fp16 = slice_by_index(begin = gdec_17_begin_0, end = gdec_17_end_0, end_mask = gdec_17_end_mask_0, x = gdec_11_cast_fp16)[name = string("gdec_17_cast_fp16")];
            tensor<int32, [2]> var_1871 = const()[name = string("op_1871"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_1872_cast_fp16 = reshape(shape = var_1871, x = gdec_17_cast_fp16)[name = string("op_1872_cast_fp16")];
            tensor<fp16, [12, 256]> var_1873_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1872_cast_fp16)[name = string("op_1873_cast_fp16")];
            tensor<int32, [4]> var_1878 = const()[name = string("op_1878"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_13_cast_fp16 = reshape(shape = var_1878, x = var_1873_cast_fp16)[name = string("P_13_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1880_cast_fp16 = exp(x = P_13_cast_fp16)[name = string("op_1880_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_13_cast_fp16 = mul(x = var_1880_cast_fp16, y = tril_incl_to_fp16)[name = string("D_13_cast_fp16")];
            bool kk_13_transpose_x_1 = const()[name = string("kk_13_transpose_x_1"), val = bool(false)];
            bool kk_13_transpose_y_1 = const()[name = string("kk_13_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_13_cast_fp16 = matmul(transpose_x = kk_13_transpose_x_1, transpose_y = kk_13_transpose_y_1, x = kb_13_cast_fp16, y = k_29_cast_fp16)[name = string("kk_13_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_1886_cast_fp16 = mul(x = kk_13_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_1886_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1887_cast_fp16 = mul(x = var_1886_cast_fp16, y = D_13_cast_fp16)[name = string("op_1887_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_13_cast_fp16 = mul(x = var_1887_cast_fp16, y = cum_tail_to_fp16)[name = string("N_13_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1889_cast_fp16 = mul(x = D_13_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1889_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_13_cast_fp16 = mul(x = kk_13_cast_fp16, y = var_1889_cast_fp16)[name = string("Na_13_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_49_cast_fp16 = add(x = eye_to_fp16, y = N_13_cast_fp16)[name = string("Minv_49_cast_fp16")];
            bool Np_37_transpose_x_0 = const()[name = string("Np_37_transpose_x_0"), val = bool(false)];
            bool Np_37_transpose_y_0 = const()[name = string("Np_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_37_cast_fp16 = matmul(transpose_x = Np_37_transpose_x_0, transpose_y = Np_37_transpose_y_0, x = N_13_cast_fp16, y = Na_13_cast_fp16)[name = string("Np_37_cast_fp16")];
            bool Npa_25_transpose_x_0 = const()[name = string("Npa_25_transpose_x_0"), val = bool(false)];
            bool Npa_25_transpose_y_0 = const()[name = string("Npa_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_25_cast_fp16 = matmul(transpose_x = Npa_25_transpose_x_0, transpose_y = Npa_25_transpose_y_0, x = Na_13_cast_fp16, y = N_13_cast_fp16)[name = string("Npa_25_cast_fp16")];
            bool var_1895_transpose_x_0 = const()[name = string("op_1895_transpose_x_0"), val = bool(false)];
            bool var_1895_transpose_y_0 = const()[name = string("op_1895_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1895_cast_fp16 = matmul(transpose_x = var_1895_transpose_x_0, transpose_y = var_1895_transpose_y_0, x = Minv_49_cast_fp16, y = Np_37_cast_fp16)[name = string("op_1895_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_51_cast_fp16 = add(x = Minv_49_cast_fp16, y = var_1895_cast_fp16)[name = string("Minv_51_cast_fp16")];
            bool Np_39_transpose_x_0 = const()[name = string("Np_39_transpose_x_0"), val = bool(false)];
            bool Np_39_transpose_y_0 = const()[name = string("Np_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_39_cast_fp16 = matmul(transpose_x = Np_39_transpose_x_0, transpose_y = Np_39_transpose_y_0, x = Np_37_cast_fp16, y = Npa_25_cast_fp16)[name = string("Np_39_cast_fp16")];
            bool Npa_27_transpose_x_0 = const()[name = string("Npa_27_transpose_x_0"), val = bool(false)];
            bool Npa_27_transpose_y_0 = const()[name = string("Npa_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_27_cast_fp16 = matmul(transpose_x = Npa_27_transpose_x_0, transpose_y = Npa_27_transpose_y_0, x = Npa_25_cast_fp16, y = Np_37_cast_fp16)[name = string("Npa_27_cast_fp16")];
            bool var_1900_transpose_x_0 = const()[name = string("op_1900_transpose_x_0"), val = bool(false)];
            bool var_1900_transpose_y_0 = const()[name = string("op_1900_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1900_cast_fp16 = matmul(transpose_x = var_1900_transpose_x_0, transpose_y = var_1900_transpose_y_0, x = Minv_51_cast_fp16, y = Np_39_cast_fp16)[name = string("op_1900_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_53_cast_fp16 = add(x = Minv_51_cast_fp16, y = var_1900_cast_fp16)[name = string("Minv_53_cast_fp16")];
            bool Np_41_transpose_x_0 = const()[name = string("Np_41_transpose_x_0"), val = bool(false)];
            bool Np_41_transpose_y_0 = const()[name = string("Np_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_41_cast_fp16 = matmul(transpose_x = Np_41_transpose_x_0, transpose_y = Np_41_transpose_y_0, x = Np_39_cast_fp16, y = Npa_27_cast_fp16)[name = string("Np_41_cast_fp16")];
            bool var_1904_transpose_x_0 = const()[name = string("op_1904_transpose_x_0"), val = bool(false)];
            bool var_1904_transpose_y_0 = const()[name = string("op_1904_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1904_cast_fp16 = matmul(transpose_x = var_1904_transpose_x_0, transpose_y = var_1904_transpose_y_0, x = Minv_53_cast_fp16, y = Np_41_cast_fp16)[name = string("op_1904_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_55_cast_fp16 = add(x = Minv_53_cast_fp16, y = var_1904_cast_fp16)[name = string("Minv_55_cast_fp16")];
            bool u_13_transpose_x_0 = const()[name = string("u_13_transpose_x_0"), val = bool(false)];
            bool u_13_transpose_y_0 = const()[name = string("u_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_13_cast_fp16 = matmul(transpose_x = u_13_transpose_x_0, transpose_y = u_13_transpose_y_0, x = Minv_55_cast_fp16, y = vb_17_cast_fp16)[name = string("u_13_cast_fp16")];
            bool var_1911_transpose_x_1 = const()[name = string("op_1911_transpose_x_1"), val = bool(false)];
            bool var_1911_transpose_y_1 = const()[name = string("op_1911_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_1911_cast_fp16 = matmul(transpose_x = var_1911_transpose_x_1, transpose_y = var_1911_transpose_y_1, x = q_29_cast_fp16, y = k_29_cast_fp16)[name = string("op_1911_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_13_cast_fp16 = mul(x = var_1911_cast_fp16, y = D_13_cast_fp16)[name = string("A_13_cast_fp16")];
            tensor<fp16, [16]> var_1913_bias_0_to_fp16 = const()[name = string("op_1913_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_1913_cast_fp16 = linear(bias = var_1913_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_17_cast_fp16)[name = string("op_1913_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_9_cast_fp16 = exp(x = var_1913_cast_fp16)[name = string("cum_9_cast_fp16")];
            tensor<int32, [1]> var_1916_axes_0 = const()[name = string("op_1916_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_1916_cast_fp16 = expand_dims(axes = var_1916_axes_0, x = cum_9_cast_fp16)[name = string("op_1916_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1917_cast_fp16 = mul(x = kb_13_cast_fp16, y = var_1916_cast_fp16)[name = string("op_1917_cast_fp16")];
            bool w_13_transpose_x_0 = const()[name = string("w_13_transpose_x_0"), val = bool(false)];
            bool w_13_transpose_y_0 = const()[name = string("w_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_13_cast_fp16 = matmul(transpose_x = w_13_transpose_x_0, transpose_y = w_13_transpose_y_0, x = Minv_55_cast_fp16, y = var_1917_cast_fp16)[name = string("w_13_cast_fp16")];
            bool var_1919_transpose_x_0 = const()[name = string("op_1919_transpose_x_0"), val = bool(false)];
            bool var_1919_transpose_y_0 = const()[name = string("op_1919_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1919_cast_fp16 = matmul(transpose_x = var_1919_transpose_x_0, transpose_y = var_1919_transpose_y_0, x = w_13_cast_fp16, y = S_9_cast_fp16)[name = string("op_1919_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_9_cast_fp16 = sub(x = u_13_cast_fp16, y = var_1919_cast_fp16)[name = string("vnew_9_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1924_cast_fp16 = mul(x = q_29_cast_fp16, y = var_1916_cast_fp16)[name = string("op_1924_cast_fp16")];
            bool var_1925_transpose_x_0 = const()[name = string("op_1925_transpose_x_0"), val = bool(false)];
            bool var_1925_transpose_y_0 = const()[name = string("op_1925_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1925_cast_fp16 = matmul(transpose_x = var_1925_transpose_x_0, transpose_y = var_1925_transpose_y_0, x = var_1924_cast_fp16, y = S_9_cast_fp16)[name = string("op_1925_cast_fp16")];
            bool var_1926_transpose_x_0 = const()[name = string("op_1926_transpose_x_0"), val = bool(false)];
            bool var_1926_transpose_y_0 = const()[name = string("op_1926_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1926_cast_fp16 = matmul(transpose_x = var_1926_transpose_x_0, transpose_y = var_1926_transpose_y_0, x = A_13_cast_fp16, y = vnew_9_cast_fp16)[name = string("op_1926_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1928_cast_fp16 = add(x = var_1925_cast_fp16, y = var_1926_cast_fp16)[name = string("op_1928_cast_fp16")];
            tensor<fp16, [16]> var_1929_bias_0_to_fp16 = const()[name = string("op_1929_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_1929_cast_fp16 = linear(bias = var_1929_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_17_cast_fp16)[name = string("op_1929_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_11_cast_fp16 = exp(x = var_1929_cast_fp16)[name = string("tail_11_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_21_to_fp16 = const()[name = string("transpose_21_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_1931_bias_0_to_fp16 = const()[name = string("op_1931_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_1931_cast_fp16 = linear(bias = var_1931_bias_0_to_fp16, weight = transpose_21_to_fp16, x = gdec_17_cast_fp16)[name = string("op_1931_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_1932_cast_fp16 = exp(x = var_1931_cast_fp16)[name = string("op_1932_cast_fp16")];
            tensor<int32, [1]> last_7_axes_0 = const()[name = string("last_7_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_7_cast_fp16 = expand_dims(axes = last_7_axes_0, x = var_1932_cast_fp16)[name = string("last_7_cast_fp16")];
            tensor<int32, [1]> var_1936_axes_0 = const()[name = string("op_1936_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_1936_cast_fp16 = expand_dims(axes = var_1936_axes_0, x = tail_11_cast_fp16)[name = string("op_1936_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1937_cast_fp16 = mul(x = k_29_cast_fp16, y = var_1936_cast_fp16)[name = string("op_1937_cast_fp16")];
            bool upd_7_transpose_x_1 = const()[name = string("upd_7_transpose_x_1"), val = bool(true)];
            bool upd_7_transpose_y_1 = const()[name = string("upd_7_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_7_cast_fp16 = matmul(transpose_x = upd_7_transpose_x_1, transpose_y = upd_7_transpose_y_1, x = var_1937_cast_fp16, y = vnew_9_cast_fp16)[name = string("upd_7_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_1942_cast_fp16 = mul(x = S_9_cast_fp16, y = last_7_cast_fp16)[name = string("op_1942_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_11_cast_fp16 = add(x = var_1942_cast_fp16, y = upd_7_cast_fp16)[name = string("S_11_cast_fp16")];
            tensor<int32, [4]> q_31_begin_0 = const()[name = string("q_31_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_31_end_0 = const()[name = string("q_31_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_31_end_mask_0 = const()[name = string("q_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, x = q_23_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<int32, [4]> k_31_begin_0 = const()[name = string("k_31_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_31_end_0 = const()[name = string("k_31_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_31_end_mask_0 = const()[name = string("k_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_31_cast_fp16 = slice_by_index(begin = k_31_begin_0, end = k_31_end_0, end_mask = k_31_end_mask_0, x = k_23_cast_fp16)[name = string("k_31_cast_fp16")];
            tensor<int32, [4]> kb_15_begin_0 = const()[name = string("kb_15_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_15_end_0 = const()[name = string("kb_15_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_15_end_mask_0 = const()[name = string("kb_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_15_cast_fp16 = slice_by_index(begin = kb_15_begin_0, end = kb_15_end_0, end_mask = kb_15_end_mask_0, x = var_1537_cast_fp16)[name = string("kb_15_cast_fp16")];
            tensor<int32, [4]> vb_19_begin_0 = const()[name = string("vb_19_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_19_end_0 = const()[name = string("vb_19_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_19_end_mask_0 = const()[name = string("vb_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_19_cast_fp16 = slice_by_index(begin = vb_19_begin_0, end = vb_19_end_0, end_mask = vb_19_end_mask_0, x = vb_11_cast_fp16)[name = string("vb_19_cast_fp16")];
            tensor<int32, [3]> gdec_19_begin_0 = const()[name = string("gdec_19_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_19_end_0 = const()[name = string("gdec_19_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_19_end_mask_0 = const()[name = string("gdec_19_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_19_cast_fp16 = slice_by_index(begin = gdec_19_begin_0, end = gdec_19_end_0, end_mask = gdec_19_end_mask_0, x = gdec_11_cast_fp16)[name = string("gdec_19_cast_fp16")];
            tensor<int32, [2]> var_2025 = const()[name = string("op_2025"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_2026_cast_fp16 = reshape(shape = var_2025, x = gdec_19_cast_fp16)[name = string("op_2026_cast_fp16")];
            tensor<fp16, [12, 256]> var_2027_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2026_cast_fp16)[name = string("op_2027_cast_fp16")];
            tensor<int32, [4]> var_2032 = const()[name = string("op_2032"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_15_cast_fp16 = reshape(shape = var_2032, x = var_2027_cast_fp16)[name = string("P_15_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2034_cast_fp16 = exp(x = P_15_cast_fp16)[name = string("op_2034_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_15_cast_fp16 = mul(x = var_2034_cast_fp16, y = tril_incl_to_fp16)[name = string("D_15_cast_fp16")];
            bool kk_15_transpose_x_1 = const()[name = string("kk_15_transpose_x_1"), val = bool(false)];
            bool kk_15_transpose_y_1 = const()[name = string("kk_15_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_15_cast_fp16 = matmul(transpose_x = kk_15_transpose_x_1, transpose_y = kk_15_transpose_y_1, x = kb_15_cast_fp16, y = k_31_cast_fp16)[name = string("kk_15_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_2040_cast_fp16 = mul(x = kk_15_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_2040_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2041_cast_fp16 = mul(x = var_2040_cast_fp16, y = D_15_cast_fp16)[name = string("op_2041_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_15_cast_fp16 = mul(x = var_2041_cast_fp16, y = cum_tail_to_fp16)[name = string("N_15_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2043_cast_fp16 = mul(x = D_15_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2043_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_15_cast_fp16 = mul(x = kk_15_cast_fp16, y = var_2043_cast_fp16)[name = string("Na_15_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_57_cast_fp16 = add(x = eye_to_fp16, y = N_15_cast_fp16)[name = string("Minv_57_cast_fp16")];
            bool Np_43_transpose_x_0 = const()[name = string("Np_43_transpose_x_0"), val = bool(false)];
            bool Np_43_transpose_y_0 = const()[name = string("Np_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_43_cast_fp16 = matmul(transpose_x = Np_43_transpose_x_0, transpose_y = Np_43_transpose_y_0, x = N_15_cast_fp16, y = Na_15_cast_fp16)[name = string("Np_43_cast_fp16")];
            bool Npa_29_transpose_x_0 = const()[name = string("Npa_29_transpose_x_0"), val = bool(false)];
            bool Npa_29_transpose_y_0 = const()[name = string("Npa_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_29_cast_fp16 = matmul(transpose_x = Npa_29_transpose_x_0, transpose_y = Npa_29_transpose_y_0, x = Na_15_cast_fp16, y = N_15_cast_fp16)[name = string("Npa_29_cast_fp16")];
            bool var_2049_transpose_x_0 = const()[name = string("op_2049_transpose_x_0"), val = bool(false)];
            bool var_2049_transpose_y_0 = const()[name = string("op_2049_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2049_cast_fp16 = matmul(transpose_x = var_2049_transpose_x_0, transpose_y = var_2049_transpose_y_0, x = Minv_57_cast_fp16, y = Np_43_cast_fp16)[name = string("op_2049_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_59_cast_fp16 = add(x = Minv_57_cast_fp16, y = var_2049_cast_fp16)[name = string("Minv_59_cast_fp16")];
            bool Np_45_transpose_x_0 = const()[name = string("Np_45_transpose_x_0"), val = bool(false)];
            bool Np_45_transpose_y_0 = const()[name = string("Np_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_45_cast_fp16 = matmul(transpose_x = Np_45_transpose_x_0, transpose_y = Np_45_transpose_y_0, x = Np_43_cast_fp16, y = Npa_29_cast_fp16)[name = string("Np_45_cast_fp16")];
            bool Npa_31_transpose_x_0 = const()[name = string("Npa_31_transpose_x_0"), val = bool(false)];
            bool Npa_31_transpose_y_0 = const()[name = string("Npa_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_31_cast_fp16 = matmul(transpose_x = Npa_31_transpose_x_0, transpose_y = Npa_31_transpose_y_0, x = Npa_29_cast_fp16, y = Np_43_cast_fp16)[name = string("Npa_31_cast_fp16")];
            bool var_2054_transpose_x_0 = const()[name = string("op_2054_transpose_x_0"), val = bool(false)];
            bool var_2054_transpose_y_0 = const()[name = string("op_2054_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2054_cast_fp16 = matmul(transpose_x = var_2054_transpose_x_0, transpose_y = var_2054_transpose_y_0, x = Minv_59_cast_fp16, y = Np_45_cast_fp16)[name = string("op_2054_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_61_cast_fp16 = add(x = Minv_59_cast_fp16, y = var_2054_cast_fp16)[name = string("Minv_61_cast_fp16")];
            bool Np_47_transpose_x_0 = const()[name = string("Np_47_transpose_x_0"), val = bool(false)];
            bool Np_47_transpose_y_0 = const()[name = string("Np_47_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_47_cast_fp16 = matmul(transpose_x = Np_47_transpose_x_0, transpose_y = Np_47_transpose_y_0, x = Np_45_cast_fp16, y = Npa_31_cast_fp16)[name = string("Np_47_cast_fp16")];
            bool var_2058_transpose_x_0 = const()[name = string("op_2058_transpose_x_0"), val = bool(false)];
            bool var_2058_transpose_y_0 = const()[name = string("op_2058_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2058_cast_fp16 = matmul(transpose_x = var_2058_transpose_x_0, transpose_y = var_2058_transpose_y_0, x = Minv_61_cast_fp16, y = Np_47_cast_fp16)[name = string("op_2058_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_63_cast_fp16 = add(x = Minv_61_cast_fp16, y = var_2058_cast_fp16)[name = string("Minv_63_cast_fp16")];
            bool u_15_transpose_x_0 = const()[name = string("u_15_transpose_x_0"), val = bool(false)];
            bool u_15_transpose_y_0 = const()[name = string("u_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_15_cast_fp16 = matmul(transpose_x = u_15_transpose_x_0, transpose_y = u_15_transpose_y_0, x = Minv_63_cast_fp16, y = vb_19_cast_fp16)[name = string("u_15_cast_fp16")];
            bool var_2065_transpose_x_1 = const()[name = string("op_2065_transpose_x_1"), val = bool(false)];
            bool var_2065_transpose_y_1 = const()[name = string("op_2065_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_2065_cast_fp16 = matmul(transpose_x = var_2065_transpose_x_1, transpose_y = var_2065_transpose_y_1, x = q_31_cast_fp16, y = k_31_cast_fp16)[name = string("op_2065_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_15_cast_fp16 = mul(x = var_2065_cast_fp16, y = D_15_cast_fp16)[name = string("A_15_cast_fp16")];
            tensor<fp16, [16]> var_2067_bias_0_to_fp16 = const()[name = string("op_2067_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_2067_cast_fp16 = linear(bias = var_2067_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_19_cast_fp16)[name = string("op_2067_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_11_cast_fp16 = exp(x = var_2067_cast_fp16)[name = string("cum_11_cast_fp16")];
            tensor<int32, [1]> var_2070_axes_0 = const()[name = string("op_2070_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_2070_cast_fp16 = expand_dims(axes = var_2070_axes_0, x = cum_11_cast_fp16)[name = string("op_2070_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2071_cast_fp16 = mul(x = kb_15_cast_fp16, y = var_2070_cast_fp16)[name = string("op_2071_cast_fp16")];
            bool w_15_transpose_x_0 = const()[name = string("w_15_transpose_x_0"), val = bool(false)];
            bool w_15_transpose_y_0 = const()[name = string("w_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_15_cast_fp16 = matmul(transpose_x = w_15_transpose_x_0, transpose_y = w_15_transpose_y_0, x = Minv_63_cast_fp16, y = var_2071_cast_fp16)[name = string("w_15_cast_fp16")];
            bool var_2073_transpose_x_0 = const()[name = string("op_2073_transpose_x_0"), val = bool(false)];
            bool var_2073_transpose_y_0 = const()[name = string("op_2073_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2073_cast_fp16 = matmul(transpose_x = var_2073_transpose_x_0, transpose_y = var_2073_transpose_y_0, x = w_15_cast_fp16, y = S_11_cast_fp16)[name = string("op_2073_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_11_cast_fp16 = sub(x = u_15_cast_fp16, y = var_2073_cast_fp16)[name = string("vnew_11_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2078_cast_fp16 = mul(x = q_31_cast_fp16, y = var_2070_cast_fp16)[name = string("op_2078_cast_fp16")];
            bool var_2079_transpose_x_0 = const()[name = string("op_2079_transpose_x_0"), val = bool(false)];
            bool var_2079_transpose_y_0 = const()[name = string("op_2079_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2079_cast_fp16 = matmul(transpose_x = var_2079_transpose_x_0, transpose_y = var_2079_transpose_y_0, x = var_2078_cast_fp16, y = S_11_cast_fp16)[name = string("op_2079_cast_fp16")];
            bool var_2080_transpose_x_0 = const()[name = string("op_2080_transpose_x_0"), val = bool(false)];
            bool var_2080_transpose_y_0 = const()[name = string("op_2080_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2080_cast_fp16 = matmul(transpose_x = var_2080_transpose_x_0, transpose_y = var_2080_transpose_y_0, x = A_15_cast_fp16, y = vnew_11_cast_fp16)[name = string("op_2080_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_3_cast_fp16 = add(x = var_2079_cast_fp16, y = var_2080_cast_fp16)[name = string("oc_3_cast_fp16")];
            int32 var_2084 = const()[name = string("op_2084"), val = int32(2)];
            bool o_7_interleave_0 = const()[name = string("o_7_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 64, 128]> o_7_cast_fp16 = concat(axis = var_2084, interleave = o_7_interleave_0, values = (var_1627_cast_fp16, var_1774_cast_fp16, var_1928_cast_fp16, oc_3_cast_fp16))[name = string("o_7_cast_fp16")];
            tensor<int32, [4]> var_2090 = const()[name = string("op_2090"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2096 = const()[name = string("op_2096"), val = tensor<int32, [4]>([3, 64, 4, 128])];
            tensor<fp16, [3, 64, 4, 128]> input_23_cast_fp16 = reshape(shape = var_2096, x = z_3_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> o_9_cast_fp16 = transpose(perm = var_2090, x = o_7_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [3, 64, 4, 128]> var_2098_cast_fp16 = mul(x = o_9_cast_fp16, y = o_9_cast_fp16)[name = string("op_2098_cast_fp16")];
            tensor<int32, [1]> var_2103_axes_0 = const()[name = string("op_2103_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2103_keep_dims_0 = const()[name = string("op_2103_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 4, 1]> var_2103_cast_fp16 = reduce_mean(axes = var_2103_axes_0, keep_dims = var_2103_keep_dims_0, x = var_2098_cast_fp16)[name = string("op_2103_cast_fp16")];
            fp16 var_2105_to_fp16 = const()[name = string("op_2105_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 64, 4, 1]> var_2106_cast_fp16 = add(x = var_2103_cast_fp16, y = var_2105_to_fp16)[name = string("op_2106_cast_fp16")];
            fp32 var_2107_epsilon_0 = const()[name = string("op_2107_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 4, 1]> var_2107_cast_fp16 = rsqrt(epsilon = var_2107_epsilon_0, x = var_2106_cast_fp16)[name = string("op_2107_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> var_2108_cast_fp16 = mul(x = o_9_cast_fp16, y = var_2107_cast_fp16)[name = string("op_2108_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> var_2110_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_2110_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> o_11_cast_fp16 = mul(x = var_2108_cast_fp16, y = var_2110_cast_fp16)[name = string("o_11_cast_fp16")];
            tensor<int32, [3]> var_2115 = const()[name = string("op_2115"), val = tensor<int32, [3]>([3, 64, 512])];
            tensor<fp16, [3, 64, 512]> input_25_cast_fp16 = reshape(shape = var_2115, x = o_11_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [640, 512]> layers_1_out_weight_to_fp16 = const()[name = string("layers_1_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34984256)))];
            tensor<fp16, [3, 64, 640]> linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_out_weight_to_fp16, x = input_25_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [3, 64, 640]> x_13_cast_fp16 = add(x = x_9_cast_fp16, y = linear_5_cast_fp16)[name = string("x_13_cast_fp16")];
            fp16 var_2128_to_fp16 = const()[name = string("op_2128_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> x_15_cast_fp16 = mul(x = x_13_cast_fp16, y = var_2128_to_fp16)[name = string("x_15_cast_fp16")];
            fp16 var_2126_promoted_to_fp16 = const()[name = string("op_2126_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> var_2130_cast_fp16 = pow(x = x_15_cast_fp16, y = var_2126_promoted_to_fp16)[name = string("op_2130_cast_fp16")];
            tensor<int32, [1]> var_2132_axes_0 = const()[name = string("op_2132_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2132_keep_dims_0 = const()[name = string("op_2132_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 1]> var_2132_cast_fp16 = reduce_mean(axes = var_2132_axes_0, keep_dims = var_2132_keep_dims_0, x = var_2130_cast_fp16)[name = string("op_2132_cast_fp16")];
            fp16 var_2133_to_fp16 = const()[name = string("op_2133_to_fp16"), val = fp16(0x1.0cp-18)];
            tensor<fp16, [3, 64, 1]> var_2134_cast_fp16 = add(x = var_2132_cast_fp16, y = var_2133_to_fp16)[name = string("op_2134_cast_fp16")];
            fp32 var_2135_epsilon_0 = const()[name = string("op_2135_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 1]> var_2135_cast_fp16 = rsqrt(epsilon = var_2135_epsilon_0, x = var_2134_cast_fp16)[name = string("op_2135_cast_fp16")];
            tensor<fp16, [3, 64, 640]> var_2136_cast_fp16 = mul(x = x_15_cast_fp16, y = var_2135_cast_fp16)[name = string("op_2136_cast_fp16")];
            tensor<fp16, [640]> layers_1_n2_w1_to_fp16 = const()[name = string("layers_1_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35639680)))];
            tensor<fp16, [3, 64, 640]> input_27_cast_fp16 = mul(x = var_2136_cast_fp16, y = layers_1_n2_w1_to_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_1_gate_up_weight_to_fp16 = const()[name = string("layers_1_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35641024)))];
            tensor<fp16, [3, 64, 5120]> linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_1_gate_up_weight_to_fp16, x = input_27_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<int32, [3]> input_29_begin_0 = const()[name = string("input_29_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_29_end_0 = const()[name = string("input_29_end_0"), val = tensor<int32, [3]>([3, 64, 2560])];
            tensor<bool, [3]> input_29_end_mask_0 = const()[name = string("input_29_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 2560]> input_29_cast_fp16 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = linear_6_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [3, 64, 2560]> var_2146_cast_fp16 = silu(x = input_29_cast_fp16)[name = string("op_2146_cast_fp16")];
            tensor<int32, [3]> var_2151_begin_0 = const()[name = string("op_2151_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_2151_end_0 = const()[name = string("op_2151_end_0"), val = tensor<int32, [3]>([3, 64, 5120])];
            tensor<bool, [3]> var_2151_end_mask_0 = const()[name = string("op_2151_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 2560]> var_2151_cast_fp16 = slice_by_index(begin = var_2151_begin_0, end = var_2151_end_0, end_mask = var_2151_end_mask_0, x = linear_6_cast_fp16)[name = string("op_2151_cast_fp16")];
            tensor<fp16, [3, 64, 2560]> input_31_cast_fp16 = mul(x = var_2146_cast_fp16, y = var_2151_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_1_down_weight_to_fp16 = const()[name = string("layers_1_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42194688)))];
            tensor<fp16, [3, 64, 640]> linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_down_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [3, 64, 640]> x_17_cast_fp16 = add(x = x_13_cast_fp16, y = linear_7_cast_fp16)[name = string("x_17_cast_fp16")];
            fp16 var_2162_promoted_to_fp16 = const()[name = string("op_2162_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> var_2166_cast_fp16 = pow(x = x_17_cast_fp16, y = var_2162_promoted_to_fp16)[name = string("op_2166_cast_fp16")];
            tensor<int32, [1]> var_2168_axes_0 = const()[name = string("op_2168_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2168_keep_dims_0 = const()[name = string("op_2168_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 1]> var_2168_cast_fp16 = reduce_mean(axes = var_2168_axes_0, keep_dims = var_2168_keep_dims_0, x = var_2166_cast_fp16)[name = string("op_2168_cast_fp16")];
            fp16 var_2169_to_fp16 = const()[name = string("op_2169_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 64, 1]> var_2170_cast_fp16 = add(x = var_2168_cast_fp16, y = var_2169_to_fp16)[name = string("op_2170_cast_fp16")];
            fp32 var_2171_epsilon_0 = const()[name = string("op_2171_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 1]> var_2171_cast_fp16 = rsqrt(epsilon = var_2171_epsilon_0, x = var_2170_cast_fp16)[name = string("op_2171_cast_fp16")];
            tensor<fp16, [3, 64, 640]> var_2172_cast_fp16 = mul(x = x_17_cast_fp16, y = var_2171_cast_fp16)[name = string("op_2172_cast_fp16")];
            tensor<fp16, [640]> layers_2_n1_w1_to_fp16 = const()[name = string("layers_2_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45471552)))];
            tensor<fp16, [3, 64, 640]> input_33_cast_fp16 = mul(x = var_2172_cast_fp16, y = layers_2_n1_w1_to_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_2_qkvzba_weight_to_fp16 = const()[name = string("layers_2_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45472896)))];
            tensor<fp16, [3, 64, 1544]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_qkvzba_weight_to_fp16, x = input_33_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [3]> qkv_5_begin_0 = const()[name = string("qkv_5_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_5_end_0 = const()[name = string("qkv_5_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> qkv_5_end_mask_0 = const()[name = string("qkv_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 1024]> qkv_5_cast_fp16 = slice_by_index(begin = qkv_5_begin_0, end = qkv_5_end_0, end_mask = qkv_5_end_mask_0, x = linear_8_cast_fp16)[name = string("qkv_5_cast_fp16")];
            tensor<int32, [3]> z_5_begin_0 = const()[name = string("z_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_5_end_0 = const()[name = string("z_5_end_0"), val = tensor<int32, [3]>([3, 64, 1536])];
            tensor<bool, [3]> z_5_end_mask_0 = const()[name = string("z_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 512]> z_5_cast_fp16 = slice_by_index(begin = z_5_begin_0, end = z_5_end_0, end_mask = z_5_end_mask_0, x = linear_8_cast_fp16)[name = string("z_5_cast_fp16")];
            tensor<int32, [3]> b_5_begin_0 = const()[name = string("b_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_5_end_0 = const()[name = string("b_5_end_0"), val = tensor<int32, [3]>([3, 64, 1540])];
            tensor<bool, [3]> b_5_end_mask_0 = const()[name = string("b_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 4]> b_5_cast_fp16 = slice_by_index(begin = b_5_begin_0, end = b_5_end_0, end_mask = b_5_end_mask_0, x = linear_8_cast_fp16)[name = string("b_5_cast_fp16")];
            tensor<int32, [3]> a_5_begin_0 = const()[name = string("a_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_5_end_0 = const()[name = string("a_5_end_0"), val = tensor<int32, [3]>([3, 64, 1])];
            tensor<bool, [3]> a_5_end_mask_0 = const()[name = string("a_5_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 4]> a_5_cast_fp16 = slice_by_index(begin = a_5_begin_0, end = a_5_end_0, end_mask = a_5_end_mask_0, x = linear_8_cast_fp16)[name = string("a_5_cast_fp16")];
            int32 var_2211 = const()[name = string("op_2211"), val = int32(1)];
            bool xp_5_interleave_0 = const()[name = string("xp_5_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 67, 1024]> xp_5_cast_fp16 = concat(axis = var_2211, interleave = xp_5_interleave_0, values = (zpad_to_fp16, qkv_5_cast_fp16))[name = string("xp_5_cast_fp16")];
            tensor<int32, [3]> var_2222_begin_0 = const()[name = string("op_2222_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_2222_end_0 = const()[name = string("op_2222_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_2222_end_mask_0 = const()[name = string("op_2222_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 1024]> var_2222_cast_fp16 = slice_by_index(begin = var_2222_begin_0, end = var_2222_end_0, end_mask = var_2222_end_mask_0, x = xp_5_cast_fp16)[name = string("op_2222_cast_fp16")];
            tensor<fp16, [1024]> var_2230_to_fp16 = const()[name = string("op_2230_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47449280)))];
            tensor<fp16, [3, 64, 1024]> var_2231_cast_fp16 = mul(x = var_2222_cast_fp16, y = var_2230_to_fp16)[name = string("op_2231_cast_fp16")];
            tensor<int32, [3]> var_2241_begin_0 = const()[name = string("op_2241_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_2241_end_0 = const()[name = string("op_2241_end_0"), val = tensor<int32, [3]>([3, 66, 1024])];
            tensor<bool, [3]> var_2241_end_mask_0 = const()[name = string("op_2241_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_2241_cast_fp16 = slice_by_index(begin = var_2241_begin_0, end = var_2241_end_0, end_mask = var_2241_end_mask_0, x = xp_5_cast_fp16)[name = string("op_2241_cast_fp16")];
            tensor<fp16, [1024]> var_2249_to_fp16 = const()[name = string("op_2249_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47451392)))];
            tensor<fp16, [3, 64, 1024]> var_2250_cast_fp16 = mul(x = var_2241_cast_fp16, y = var_2249_to_fp16)[name = string("op_2250_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> var_2252_cast_fp16 = add(x = var_2231_cast_fp16, y = var_2250_cast_fp16)[name = string("op_2252_cast_fp16")];
            tensor<int32, [3]> var_2262_begin_0 = const()[name = string("op_2262_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2262_end_0 = const()[name = string("op_2262_end_0"), val = tensor<int32, [3]>([3, 65, 1024])];
            tensor<bool, [3]> var_2262_end_mask_0 = const()[name = string("op_2262_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_2262_cast_fp16 = slice_by_index(begin = var_2262_begin_0, end = var_2262_end_0, end_mask = var_2262_end_mask_0, x = xp_5_cast_fp16)[name = string("op_2262_cast_fp16")];
            tensor<fp16, [1024]> var_2270_to_fp16 = const()[name = string("op_2270_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47453504)))];
            tensor<fp16, [3, 64, 1024]> var_2271_cast_fp16 = mul(x = var_2262_cast_fp16, y = var_2270_to_fp16)[name = string("op_2271_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> var_2273_cast_fp16 = add(x = var_2252_cast_fp16, y = var_2271_cast_fp16)[name = string("op_2273_cast_fp16")];
            tensor<int32, [3]> var_2283_begin_0 = const()[name = string("op_2283_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2283_end_0 = const()[name = string("op_2283_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> var_2283_end_mask_0 = const()[name = string("op_2283_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_2283_cast_fp16 = slice_by_index(begin = var_2283_begin_0, end = var_2283_end_0, end_mask = var_2283_end_mask_0, x = xp_5_cast_fp16)[name = string("op_2283_cast_fp16")];
            tensor<fp16, [1024]> var_2291_to_fp16 = const()[name = string("op_2291_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47455616)))];
            tensor<fp16, [3, 64, 1024]> var_2292_cast_fp16 = mul(x = var_2283_cast_fp16, y = var_2291_to_fp16)[name = string("op_2292_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> input_35_cast_fp16 = add(x = var_2273_cast_fp16, y = var_2292_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> c_5_cast_fp16 = silu(x = input_35_cast_fp16)[name = string("c_5_cast_fp16")];
            tensor<int32, [3]> var_2300_begin_0 = const()[name = string("op_2300_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2300_end_0 = const()[name = string("op_2300_end_0"), val = tensor<int32, [3]>([3, 64, 256])];
            tensor<bool, [3]> var_2300_end_mask_0 = const()[name = string("op_2300_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 256]> var_2300_cast_fp16 = slice_by_index(begin = var_2300_begin_0, end = var_2300_end_0, end_mask = var_2300_end_mask_0, x = c_5_cast_fp16)[name = string("op_2300_cast_fp16")];
            tensor<int32, [4]> var_2305 = const()[name = string("op_2305"), val = tensor<int32, [4]>([3, 64, 2, 128])];
            tensor<fp16, [3, 64, 2, 128]> q_33_cast_fp16 = reshape(shape = var_2305, x = var_2300_cast_fp16)[name = string("q_33_cast_fp16")];
            tensor<int32, [3]> var_2311_begin_0 = const()[name = string("op_2311_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_2311_end_0 = const()[name = string("op_2311_end_0"), val = tensor<int32, [3]>([3, 64, 512])];
            tensor<bool, [3]> var_2311_end_mask_0 = const()[name = string("op_2311_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 256]> var_2311_cast_fp16 = slice_by_index(begin = var_2311_begin_0, end = var_2311_end_0, end_mask = var_2311_end_mask_0, x = c_5_cast_fp16)[name = string("op_2311_cast_fp16")];
            tensor<int32, [4]> var_2316 = const()[name = string("op_2316"), val = tensor<int32, [4]>([3, 64, 2, 128])];
            tensor<fp16, [3, 64, 2, 128]> k_33_cast_fp16 = reshape(shape = var_2316, x = var_2311_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<int32, [3]> var_2322_begin_0 = const()[name = string("op_2322_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_2322_end_0 = const()[name = string("op_2322_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> var_2322_end_mask_0 = const()[name = string("op_2322_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 512]> var_2322_cast_fp16 = slice_by_index(begin = var_2322_begin_0, end = var_2322_end_0, end_mask = var_2322_end_mask_0, x = c_5_cast_fp16)[name = string("op_2322_cast_fp16")];
            tensor<int32, [4]> var_2327 = const()[name = string("op_2327"), val = tensor<int32, [4]>([3, 64, 4, 128])];
            tensor<fp16, [3, 64, 4, 128]> var_2328_cast_fp16 = reshape(shape = var_2327, x = var_2322_cast_fp16)[name = string("op_2328_cast_fp16")];
            tensor<int32, [4]> var_2333 = const()[name = string("op_2333"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2335_to_fp16 = const()[name = string("op_2335_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 64, 2, 128]> q_35_cast_fp16 = mul(x = q_33_cast_fp16, y = var_2335_to_fp16)[name = string("q_35_cast_fp16")];
            fp16 var_2337_to_fp16 = const()[name = string("op_2337_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 64, 2, 128]> k_35_cast_fp16 = mul(x = k_33_cast_fp16, y = var_2337_to_fp16)[name = string("k_35_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_2339_cast_fp16 = mul(x = q_35_cast_fp16, y = q_35_cast_fp16)[name = string("op_2339_cast_fp16")];
            tensor<int32, [1]> var_2344_axes_0 = const()[name = string("op_2344_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2344_keep_dims_0 = const()[name = string("op_2344_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 2, 1]> var_2344_cast_fp16 = reduce_mean(axes = var_2344_axes_0, keep_dims = var_2344_keep_dims_0, x = var_2339_cast_fp16)[name = string("op_2344_cast_fp16")];
            fp16 var_2345_promoted_to_fp16 = const()[name = string("op_2345_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 64, 2, 1]> var_2346_cast_fp16 = mul(x = var_2344_cast_fp16, y = var_2345_promoted_to_fp16)[name = string("op_2346_cast_fp16")];
            fp16 var_2348_to_fp16 = const()[name = string("op_2348_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 64, 2, 1]> var_2349_cast_fp16 = add(x = var_2346_cast_fp16, y = var_2348_to_fp16)[name = string("op_2349_cast_fp16")];
            fp32 var_2350_epsilon_0 = const()[name = string("op_2350_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 2, 1]> var_2350_cast_fp16 = rsqrt(epsilon = var_2350_epsilon_0, x = var_2349_cast_fp16)[name = string("op_2350_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_2351_cast_fp16 = mul(x = q_35_cast_fp16, y = var_2350_cast_fp16)[name = string("op_2351_cast_fp16")];
            fp16 var_2352_to_fp16 = const()[name = string("op_2352_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 64, 2, 128]> q_37_cast_fp16 = mul(x = var_2351_cast_fp16, y = var_2352_to_fp16)[name = string("q_37_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_2354_cast_fp16 = mul(x = k_35_cast_fp16, y = k_35_cast_fp16)[name = string("op_2354_cast_fp16")];
            tensor<int32, [1]> var_2359_axes_0 = const()[name = string("op_2359_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2359_keep_dims_0 = const()[name = string("op_2359_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 2, 1]> var_2359_cast_fp16 = reduce_mean(axes = var_2359_axes_0, keep_dims = var_2359_keep_dims_0, x = var_2354_cast_fp16)[name = string("op_2359_cast_fp16")];
            fp16 var_2360_promoted_to_fp16 = const()[name = string("op_2360_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 64, 2, 1]> var_2361_cast_fp16 = mul(x = var_2359_cast_fp16, y = var_2360_promoted_to_fp16)[name = string("op_2361_cast_fp16")];
            fp16 var_2363_to_fp16 = const()[name = string("op_2363_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 64, 2, 1]> var_2364_cast_fp16 = add(x = var_2361_cast_fp16, y = var_2363_to_fp16)[name = string("op_2364_cast_fp16")];
            fp32 var_2365_epsilon_0 = const()[name = string("op_2365_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 2, 1]> var_2365_cast_fp16 = rsqrt(epsilon = var_2365_epsilon_0, x = var_2364_cast_fp16)[name = string("op_2365_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> k_37_cast_fp16 = mul(x = k_35_cast_fp16, y = var_2365_cast_fp16)[name = string("k_37_cast_fp16")];
            tensor<int32, [4]> var_2371 = const()[name = string("op_2371"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2384_axes_0 = const()[name = string("op_2384_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 64, 128]> var_2372_cast_fp16 = transpose(perm = var_2371, x = q_37_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [3, 2, 1, 64, 128]> var_2384_cast_fp16 = expand_dims(axes = var_2384_axes_0, x = var_2372_cast_fp16)[name = string("op_2384_cast_fp16")];
            tensor<int32, [5]> var_2392_reps_0 = const()[name = string("op_2392_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 64, 128]> var_2392_cast_fp16 = tile(reps = var_2392_reps_0, x = var_2384_cast_fp16)[name = string("op_2392_cast_fp16")];
            tensor<int32, [4]> var_2397 = const()[name = string("op_2397"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<fp16, [3, 4, 64, 128]> q_39_cast_fp16 = reshape(shape = var_2397, x = var_2392_cast_fp16)[name = string("q_39_cast_fp16")];
            tensor<int32, [4]> var_2403 = const()[name = string("op_2403"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_2416_axes_0 = const()[name = string("op_2416_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 64, 128]> var_2404_cast_fp16 = transpose(perm = var_2403, x = k_37_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [3, 2, 1, 64, 128]> var_2416_cast_fp16 = expand_dims(axes = var_2416_axes_0, x = var_2404_cast_fp16)[name = string("op_2416_cast_fp16")];
            tensor<int32, [5]> var_2424_reps_0 = const()[name = string("op_2424_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 64, 128]> var_2424_cast_fp16 = tile(reps = var_2424_reps_0, x = var_2416_cast_fp16)[name = string("op_2424_cast_fp16")];
            tensor<int32, [4]> var_2429 = const()[name = string("op_2429"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<fp16, [3, 4, 64, 128]> k_39_cast_fp16 = reshape(shape = var_2429, x = var_2424_cast_fp16)[name = string("k_39_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_2431_cast_fp16 = sigmoid(x = b_5_cast_fp16)[name = string("op_2431_cast_fp16")];
            tensor<int32, [3]> var_2435 = const()[name = string("op_2435"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_2_dt_bias_to_fp16 = const()[name = string("layers_2_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 64, 4]> input_37_cast_fp16 = add(x = a_5_cast_fp16, y = layers_2_dt_bias_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_2439_cast_fp16 = relu(x = input_37_cast_fp16)[name = string("op_2439_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_2440_cast_fp16 = abs(x = input_37_cast_fp16)[name = string("op_2440_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 64, 4]> var_2441_cast_fp16 = mul(x = var_2440_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_2441_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_2442_cast_fp16 = exp(x = var_2441_cast_fp16)[name = string("op_2442_cast_fp16")];
            fp16 var_2444_to_fp16 = const()[name = string("op_2444_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 64, 4]> var_2445_cast_fp16 = add(x = var_2442_cast_fp16, y = var_2444_to_fp16)[name = string("op_2445_cast_fp16")];
            fp32 var_2446_epsilon_0 = const()[name = string("op_2446_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 64, 4]> var_2446_cast_fp16 = log(epsilon = var_2446_epsilon_0, x = var_2445_cast_fp16)[name = string("op_2446_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_2448_cast_fp16 = add(x = var_2439_cast_fp16, y = var_2446_cast_fp16)[name = string("op_2448_cast_fp16")];
            tensor<fp16, [4]> layers_2_negA_to_fp16 = const()[name = string("layers_2_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.c0cp+3, -0x1.3e4p+3, -0x1.568p+1, -0x1.9acp+2])];
            tensor<fp16, [3, 64, 4]> var_2449_cast_fp16 = mul(x = layers_2_negA_to_fp16, y = var_2448_cast_fp16)[name = string("op_2449_cast_fp16")];
            tensor<int32, [3]> var_2453 = const()[name = string("op_2453"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_2456_axes_0 = const()[name = string("op_2456_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 64]> beta_5_cast_fp16 = transpose(perm = var_2435, x = var_2431_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [3, 4, 64, 1]> var_2456_cast_fp16 = expand_dims(axes = var_2456_axes_0, x = beta_5_cast_fp16)[name = string("op_2456_cast_fp16")];
            fp16 var_2457_to_fp16 = const()[name = string("op_2457_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 64, 1]> var_2458_cast_fp16 = mul(x = var_2456_cast_fp16, y = var_2457_to_fp16)[name = string("op_2458_cast_fp16")];
            tensor<fp16, [3, 4, 64, 128]> v_5_cast_fp16 = transpose(perm = var_2333, x = var_2328_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [3, 4, 64, 128]> vb_21_cast_fp16 = mul(x = v_5_cast_fp16, y = var_2458_cast_fp16)[name = string("vb_21_cast_fp16")];
            tensor<int32, [4]> q_41_begin_0 = const()[name = string("q_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_41_end_0 = const()[name = string("q_41_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_41_end_mask_0 = const()[name = string("q_41_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_41_cast_fp16 = slice_by_index(begin = q_41_begin_0, end = q_41_end_0, end_mask = q_41_end_mask_0, x = q_39_cast_fp16)[name = string("q_41_cast_fp16")];
            tensor<int32, [4]> k_41_begin_0 = const()[name = string("k_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_41_end_0 = const()[name = string("k_41_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_41_end_mask_0 = const()[name = string("k_41_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, x = k_39_cast_fp16)[name = string("k_41_cast_fp16")];
            tensor<fp16, [3, 4, 64, 128]> var_2492_cast_fp16 = mul(x = k_39_cast_fp16, y = var_2456_cast_fp16)[name = string("op_2492_cast_fp16")];
            tensor<int32, [4]> kb_17_begin_0 = const()[name = string("kb_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_17_end_0 = const()[name = string("kb_17_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_17_end_mask_0 = const()[name = string("kb_17_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_17_cast_fp16 = slice_by_index(begin = kb_17_begin_0, end = kb_17_end_0, end_mask = kb_17_end_mask_0, x = var_2492_cast_fp16)[name = string("kb_17_cast_fp16")];
            tensor<int32, [4]> vb_23_begin_0 = const()[name = string("vb_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_23_end_0 = const()[name = string("vb_23_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_23_end_mask_0 = const()[name = string("vb_23_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_23_cast_fp16 = slice_by_index(begin = vb_23_begin_0, end = vb_23_end_0, end_mask = vb_23_end_mask_0, x = vb_21_cast_fp16)[name = string("vb_23_cast_fp16")];
            tensor<int32, [3]> gdec_23_begin_0 = const()[name = string("gdec_23_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_23_end_0 = const()[name = string("gdec_23_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_23_end_mask_0 = const()[name = string("gdec_23_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 64]> gdec_21_cast_fp16 = transpose(perm = var_2453, x = var_2449_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [3, 4, 16]> gdec_23_cast_fp16 = slice_by_index(begin = gdec_23_begin_0, end = gdec_23_end_0, end_mask = gdec_23_end_mask_0, x = gdec_21_cast_fp16)[name = string("gdec_23_cast_fp16")];
            tensor<int32, [2]> var_2540 = const()[name = string("op_2540"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_2541_cast_fp16 = reshape(shape = var_2540, x = gdec_23_cast_fp16)[name = string("op_2541_cast_fp16")];
            tensor<fp16, [12, 256]> var_2542_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2541_cast_fp16)[name = string("op_2542_cast_fp16")];
            tensor<int32, [4]> var_2547 = const()[name = string("op_2547"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_17_cast_fp16 = reshape(shape = var_2547, x = var_2542_cast_fp16)[name = string("P_17_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2549_cast_fp16 = exp(x = P_17_cast_fp16)[name = string("op_2549_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_17_cast_fp16 = mul(x = var_2549_cast_fp16, y = tril_incl_to_fp16)[name = string("D_17_cast_fp16")];
            bool kk_17_transpose_x_1 = const()[name = string("kk_17_transpose_x_1"), val = bool(false)];
            bool kk_17_transpose_y_1 = const()[name = string("kk_17_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_17_cast_fp16 = matmul(transpose_x = kk_17_transpose_x_1, transpose_y = kk_17_transpose_y_1, x = kb_17_cast_fp16, y = k_41_cast_fp16)[name = string("kk_17_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_2555_cast_fp16 = mul(x = kk_17_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_2555_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2556_cast_fp16 = mul(x = var_2555_cast_fp16, y = D_17_cast_fp16)[name = string("op_2556_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_17_cast_fp16 = mul(x = var_2556_cast_fp16, y = cum_tail_to_fp16)[name = string("N_17_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2558_cast_fp16 = mul(x = D_17_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2558_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_17_cast_fp16 = mul(x = kk_17_cast_fp16, y = var_2558_cast_fp16)[name = string("Na_17_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_65_cast_fp16 = add(x = eye_to_fp16, y = N_17_cast_fp16)[name = string("Minv_65_cast_fp16")];
            bool Np_49_transpose_x_0 = const()[name = string("Np_49_transpose_x_0"), val = bool(false)];
            bool Np_49_transpose_y_0 = const()[name = string("Np_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_49_cast_fp16 = matmul(transpose_x = Np_49_transpose_x_0, transpose_y = Np_49_transpose_y_0, x = N_17_cast_fp16, y = Na_17_cast_fp16)[name = string("Np_49_cast_fp16")];
            bool Npa_33_transpose_x_0 = const()[name = string("Npa_33_transpose_x_0"), val = bool(false)];
            bool Npa_33_transpose_y_0 = const()[name = string("Npa_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_33_cast_fp16 = matmul(transpose_x = Npa_33_transpose_x_0, transpose_y = Npa_33_transpose_y_0, x = Na_17_cast_fp16, y = N_17_cast_fp16)[name = string("Npa_33_cast_fp16")];
            bool var_2564_transpose_x_0 = const()[name = string("op_2564_transpose_x_0"), val = bool(false)];
            bool var_2564_transpose_y_0 = const()[name = string("op_2564_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2564_cast_fp16 = matmul(transpose_x = var_2564_transpose_x_0, transpose_y = var_2564_transpose_y_0, x = Minv_65_cast_fp16, y = Np_49_cast_fp16)[name = string("op_2564_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_67_cast_fp16 = add(x = Minv_65_cast_fp16, y = var_2564_cast_fp16)[name = string("Minv_67_cast_fp16")];
            bool Np_51_transpose_x_0 = const()[name = string("Np_51_transpose_x_0"), val = bool(false)];
            bool Np_51_transpose_y_0 = const()[name = string("Np_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_51_cast_fp16 = matmul(transpose_x = Np_51_transpose_x_0, transpose_y = Np_51_transpose_y_0, x = Np_49_cast_fp16, y = Npa_33_cast_fp16)[name = string("Np_51_cast_fp16")];
            bool Npa_35_transpose_x_0 = const()[name = string("Npa_35_transpose_x_0"), val = bool(false)];
            bool Npa_35_transpose_y_0 = const()[name = string("Npa_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_35_cast_fp16 = matmul(transpose_x = Npa_35_transpose_x_0, transpose_y = Npa_35_transpose_y_0, x = Npa_33_cast_fp16, y = Np_49_cast_fp16)[name = string("Npa_35_cast_fp16")];
            bool var_2569_transpose_x_0 = const()[name = string("op_2569_transpose_x_0"), val = bool(false)];
            bool var_2569_transpose_y_0 = const()[name = string("op_2569_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2569_cast_fp16 = matmul(transpose_x = var_2569_transpose_x_0, transpose_y = var_2569_transpose_y_0, x = Minv_67_cast_fp16, y = Np_51_cast_fp16)[name = string("op_2569_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_69_cast_fp16 = add(x = Minv_67_cast_fp16, y = var_2569_cast_fp16)[name = string("Minv_69_cast_fp16")];
            bool Np_53_transpose_x_0 = const()[name = string("Np_53_transpose_x_0"), val = bool(false)];
            bool Np_53_transpose_y_0 = const()[name = string("Np_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_53_cast_fp16 = matmul(transpose_x = Np_53_transpose_x_0, transpose_y = Np_53_transpose_y_0, x = Np_51_cast_fp16, y = Npa_35_cast_fp16)[name = string("Np_53_cast_fp16")];
            bool var_2573_transpose_x_0 = const()[name = string("op_2573_transpose_x_0"), val = bool(false)];
            bool var_2573_transpose_y_0 = const()[name = string("op_2573_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2573_cast_fp16 = matmul(transpose_x = var_2573_transpose_x_0, transpose_y = var_2573_transpose_y_0, x = Minv_69_cast_fp16, y = Np_53_cast_fp16)[name = string("op_2573_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_71_cast_fp16 = add(x = Minv_69_cast_fp16, y = var_2573_cast_fp16)[name = string("Minv_71_cast_fp16")];
            bool u_17_transpose_x_0 = const()[name = string("u_17_transpose_x_0"), val = bool(false)];
            bool u_17_transpose_y_0 = const()[name = string("u_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_17_cast_fp16 = matmul(transpose_x = u_17_transpose_x_0, transpose_y = u_17_transpose_y_0, x = Minv_71_cast_fp16, y = vb_23_cast_fp16)[name = string("u_17_cast_fp16")];
            bool var_2580_transpose_x_1 = const()[name = string("op_2580_transpose_x_1"), val = bool(false)];
            bool var_2580_transpose_y_1 = const()[name = string("op_2580_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_2580_cast_fp16 = matmul(transpose_x = var_2580_transpose_x_1, transpose_y = var_2580_transpose_y_1, x = q_41_cast_fp16, y = k_41_cast_fp16)[name = string("op_2580_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_17_cast_fp16 = mul(x = var_2580_cast_fp16, y = D_17_cast_fp16)[name = string("A_17_cast_fp16")];
            bool var_2582_transpose_x_0 = const()[name = string("op_2582_transpose_x_0"), val = bool(false)];
            bool var_2582_transpose_y_0 = const()[name = string("op_2582_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2582_cast_fp16 = matmul(transpose_x = var_2582_transpose_x_0, transpose_y = var_2582_transpose_y_0, x = A_17_cast_fp16, y = u_17_cast_fp16)[name = string("op_2582_cast_fp16")];
            tensor<fp16, [16]> var_2583_bias_0_to_fp16 = const()[name = string("op_2583_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_2583_cast_fp16 = linear(bias = var_2583_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_23_cast_fp16)[name = string("op_2583_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_13_cast_fp16 = exp(x = var_2583_cast_fp16)[name = string("tail_13_cast_fp16")];
            tensor<int32, [1]> var_2586_axes_0 = const()[name = string("op_2586_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_2586_cast_fp16 = expand_dims(axes = var_2586_axes_0, x = tail_13_cast_fp16)[name = string("op_2586_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2587_cast_fp16 = mul(x = k_41_cast_fp16, y = var_2586_cast_fp16)[name = string("op_2587_cast_fp16")];
            bool S_13_transpose_x_1 = const()[name = string("S_13_transpose_x_1"), val = bool(true)];
            bool S_13_transpose_y_1 = const()[name = string("S_13_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_13_cast_fp16 = matmul(transpose_x = S_13_transpose_x_1, transpose_y = S_13_transpose_y_1, x = var_2587_cast_fp16, y = u_17_cast_fp16)[name = string("S_13_cast_fp16")];
            tensor<int32, [4]> q_43_begin_0 = const()[name = string("q_43_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_43_end_0 = const()[name = string("q_43_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> q_43_end_mask_0 = const()[name = string("q_43_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, x = q_39_cast_fp16)[name = string("q_43_cast_fp16")];
            tensor<int32, [4]> k_43_begin_0 = const()[name = string("k_43_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_43_end_0 = const()[name = string("k_43_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> k_43_end_mask_0 = const()[name = string("k_43_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_43_cast_fp16 = slice_by_index(begin = k_43_begin_0, end = k_43_end_0, end_mask = k_43_end_mask_0, x = k_39_cast_fp16)[name = string("k_43_cast_fp16")];
            tensor<int32, [4]> kb_19_begin_0 = const()[name = string("kb_19_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_19_end_0 = const()[name = string("kb_19_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> kb_19_end_mask_0 = const()[name = string("kb_19_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_19_cast_fp16 = slice_by_index(begin = kb_19_begin_0, end = kb_19_end_0, end_mask = kb_19_end_mask_0, x = var_2492_cast_fp16)[name = string("kb_19_cast_fp16")];
            tensor<int32, [4]> vb_25_begin_0 = const()[name = string("vb_25_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_25_end_0 = const()[name = string("vb_25_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> vb_25_end_mask_0 = const()[name = string("vb_25_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_25_cast_fp16 = slice_by_index(begin = vb_25_begin_0, end = vb_25_end_0, end_mask = vb_25_end_mask_0, x = vb_21_cast_fp16)[name = string("vb_25_cast_fp16")];
            tensor<int32, [3]> gdec_25_begin_0 = const()[name = string("gdec_25_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_25_end_0 = const()[name = string("gdec_25_end_0"), val = tensor<int32, [3]>([3, 4, 32])];
            tensor<bool, [3]> gdec_25_end_mask_0 = const()[name = string("gdec_25_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_25_cast_fp16 = slice_by_index(begin = gdec_25_begin_0, end = gdec_25_end_0, end_mask = gdec_25_end_mask_0, x = gdec_21_cast_fp16)[name = string("gdec_25_cast_fp16")];
            tensor<int32, [2]> var_2672 = const()[name = string("op_2672"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_2673_cast_fp16 = reshape(shape = var_2672, x = gdec_25_cast_fp16)[name = string("op_2673_cast_fp16")];
            tensor<fp16, [12, 256]> var_2674_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2673_cast_fp16)[name = string("op_2674_cast_fp16")];
            tensor<int32, [4]> var_2679 = const()[name = string("op_2679"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_19_cast_fp16 = reshape(shape = var_2679, x = var_2674_cast_fp16)[name = string("P_19_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2681_cast_fp16 = exp(x = P_19_cast_fp16)[name = string("op_2681_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_19_cast_fp16 = mul(x = var_2681_cast_fp16, y = tril_incl_to_fp16)[name = string("D_19_cast_fp16")];
            bool kk_19_transpose_x_1 = const()[name = string("kk_19_transpose_x_1"), val = bool(false)];
            bool kk_19_transpose_y_1 = const()[name = string("kk_19_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_19_cast_fp16 = matmul(transpose_x = kk_19_transpose_x_1, transpose_y = kk_19_transpose_y_1, x = kb_19_cast_fp16, y = k_43_cast_fp16)[name = string("kk_19_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_2687_cast_fp16 = mul(x = kk_19_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_2687_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2688_cast_fp16 = mul(x = var_2687_cast_fp16, y = D_19_cast_fp16)[name = string("op_2688_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_19_cast_fp16 = mul(x = var_2688_cast_fp16, y = cum_tail_to_fp16)[name = string("N_19_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2690_cast_fp16 = mul(x = D_19_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2690_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_19_cast_fp16 = mul(x = kk_19_cast_fp16, y = var_2690_cast_fp16)[name = string("Na_19_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_73_cast_fp16 = add(x = eye_to_fp16, y = N_19_cast_fp16)[name = string("Minv_73_cast_fp16")];
            bool Np_55_transpose_x_0 = const()[name = string("Np_55_transpose_x_0"), val = bool(false)];
            bool Np_55_transpose_y_0 = const()[name = string("Np_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_55_cast_fp16 = matmul(transpose_x = Np_55_transpose_x_0, transpose_y = Np_55_transpose_y_0, x = N_19_cast_fp16, y = Na_19_cast_fp16)[name = string("Np_55_cast_fp16")];
            bool Npa_37_transpose_x_0 = const()[name = string("Npa_37_transpose_x_0"), val = bool(false)];
            bool Npa_37_transpose_y_0 = const()[name = string("Npa_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_37_cast_fp16 = matmul(transpose_x = Npa_37_transpose_x_0, transpose_y = Npa_37_transpose_y_0, x = Na_19_cast_fp16, y = N_19_cast_fp16)[name = string("Npa_37_cast_fp16")];
            bool var_2696_transpose_x_0 = const()[name = string("op_2696_transpose_x_0"), val = bool(false)];
            bool var_2696_transpose_y_0 = const()[name = string("op_2696_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2696_cast_fp16 = matmul(transpose_x = var_2696_transpose_x_0, transpose_y = var_2696_transpose_y_0, x = Minv_73_cast_fp16, y = Np_55_cast_fp16)[name = string("op_2696_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_75_cast_fp16 = add(x = Minv_73_cast_fp16, y = var_2696_cast_fp16)[name = string("Minv_75_cast_fp16")];
            bool Np_57_transpose_x_0 = const()[name = string("Np_57_transpose_x_0"), val = bool(false)];
            bool Np_57_transpose_y_0 = const()[name = string("Np_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_57_cast_fp16 = matmul(transpose_x = Np_57_transpose_x_0, transpose_y = Np_57_transpose_y_0, x = Np_55_cast_fp16, y = Npa_37_cast_fp16)[name = string("Np_57_cast_fp16")];
            bool Npa_39_transpose_x_0 = const()[name = string("Npa_39_transpose_x_0"), val = bool(false)];
            bool Npa_39_transpose_y_0 = const()[name = string("Npa_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_39_cast_fp16 = matmul(transpose_x = Npa_39_transpose_x_0, transpose_y = Npa_39_transpose_y_0, x = Npa_37_cast_fp16, y = Np_55_cast_fp16)[name = string("Npa_39_cast_fp16")];
            bool var_2701_transpose_x_0 = const()[name = string("op_2701_transpose_x_0"), val = bool(false)];
            bool var_2701_transpose_y_0 = const()[name = string("op_2701_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2701_cast_fp16 = matmul(transpose_x = var_2701_transpose_x_0, transpose_y = var_2701_transpose_y_0, x = Minv_75_cast_fp16, y = Np_57_cast_fp16)[name = string("op_2701_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_77_cast_fp16 = add(x = Minv_75_cast_fp16, y = var_2701_cast_fp16)[name = string("Minv_77_cast_fp16")];
            bool Np_59_transpose_x_0 = const()[name = string("Np_59_transpose_x_0"), val = bool(false)];
            bool Np_59_transpose_y_0 = const()[name = string("Np_59_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_59_cast_fp16 = matmul(transpose_x = Np_59_transpose_x_0, transpose_y = Np_59_transpose_y_0, x = Np_57_cast_fp16, y = Npa_39_cast_fp16)[name = string("Np_59_cast_fp16")];
            bool var_2705_transpose_x_0 = const()[name = string("op_2705_transpose_x_0"), val = bool(false)];
            bool var_2705_transpose_y_0 = const()[name = string("op_2705_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2705_cast_fp16 = matmul(transpose_x = var_2705_transpose_x_0, transpose_y = var_2705_transpose_y_0, x = Minv_77_cast_fp16, y = Np_59_cast_fp16)[name = string("op_2705_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_79_cast_fp16 = add(x = Minv_77_cast_fp16, y = var_2705_cast_fp16)[name = string("Minv_79_cast_fp16")];
            bool u_19_transpose_x_0 = const()[name = string("u_19_transpose_x_0"), val = bool(false)];
            bool u_19_transpose_y_0 = const()[name = string("u_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_19_cast_fp16 = matmul(transpose_x = u_19_transpose_x_0, transpose_y = u_19_transpose_y_0, x = Minv_79_cast_fp16, y = vb_25_cast_fp16)[name = string("u_19_cast_fp16")];
            bool var_2712_transpose_x_1 = const()[name = string("op_2712_transpose_x_1"), val = bool(false)];
            bool var_2712_transpose_y_1 = const()[name = string("op_2712_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_2712_cast_fp16 = matmul(transpose_x = var_2712_transpose_x_1, transpose_y = var_2712_transpose_y_1, x = q_43_cast_fp16, y = k_43_cast_fp16)[name = string("op_2712_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_19_cast_fp16 = mul(x = var_2712_cast_fp16, y = D_19_cast_fp16)[name = string("A_19_cast_fp16")];
            tensor<fp16, [16]> var_2714_bias_0_to_fp16 = const()[name = string("op_2714_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_2714_cast_fp16 = linear(bias = var_2714_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_25_cast_fp16)[name = string("op_2714_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_13_cast_fp16 = exp(x = var_2714_cast_fp16)[name = string("cum_13_cast_fp16")];
            tensor<int32, [1]> var_2717_axes_0 = const()[name = string("op_2717_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_2717_cast_fp16 = expand_dims(axes = var_2717_axes_0, x = cum_13_cast_fp16)[name = string("op_2717_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2718_cast_fp16 = mul(x = kb_19_cast_fp16, y = var_2717_cast_fp16)[name = string("op_2718_cast_fp16")];
            bool w_19_transpose_x_0 = const()[name = string("w_19_transpose_x_0"), val = bool(false)];
            bool w_19_transpose_y_0 = const()[name = string("w_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_19_cast_fp16 = matmul(transpose_x = w_19_transpose_x_0, transpose_y = w_19_transpose_y_0, x = Minv_79_cast_fp16, y = var_2718_cast_fp16)[name = string("w_19_cast_fp16")];
            bool var_2720_transpose_x_0 = const()[name = string("op_2720_transpose_x_0"), val = bool(false)];
            bool var_2720_transpose_y_0 = const()[name = string("op_2720_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2720_cast_fp16 = matmul(transpose_x = var_2720_transpose_x_0, transpose_y = var_2720_transpose_y_0, x = w_19_cast_fp16, y = S_13_cast_fp16)[name = string("op_2720_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_13_cast_fp16 = sub(x = u_19_cast_fp16, y = var_2720_cast_fp16)[name = string("vnew_13_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2725_cast_fp16 = mul(x = q_43_cast_fp16, y = var_2717_cast_fp16)[name = string("op_2725_cast_fp16")];
            bool var_2726_transpose_x_0 = const()[name = string("op_2726_transpose_x_0"), val = bool(false)];
            bool var_2726_transpose_y_0 = const()[name = string("op_2726_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2726_cast_fp16 = matmul(transpose_x = var_2726_transpose_x_0, transpose_y = var_2726_transpose_y_0, x = var_2725_cast_fp16, y = S_13_cast_fp16)[name = string("op_2726_cast_fp16")];
            bool var_2727_transpose_x_0 = const()[name = string("op_2727_transpose_x_0"), val = bool(false)];
            bool var_2727_transpose_y_0 = const()[name = string("op_2727_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2727_cast_fp16 = matmul(transpose_x = var_2727_transpose_x_0, transpose_y = var_2727_transpose_y_0, x = A_19_cast_fp16, y = vnew_13_cast_fp16)[name = string("op_2727_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2729_cast_fp16 = add(x = var_2726_cast_fp16, y = var_2727_cast_fp16)[name = string("op_2729_cast_fp16")];
            tensor<fp16, [16]> var_2730_bias_0_to_fp16 = const()[name = string("op_2730_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_2730_cast_fp16 = linear(bias = var_2730_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_25_cast_fp16)[name = string("op_2730_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_15_cast_fp16 = exp(x = var_2730_cast_fp16)[name = string("tail_15_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_29_to_fp16 = const()[name = string("transpose_29_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_2732_bias_0_to_fp16 = const()[name = string("op_2732_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_2732_cast_fp16 = linear(bias = var_2732_bias_0_to_fp16, weight = transpose_29_to_fp16, x = gdec_25_cast_fp16)[name = string("op_2732_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_2733_cast_fp16 = exp(x = var_2732_cast_fp16)[name = string("op_2733_cast_fp16")];
            tensor<int32, [1]> last_9_axes_0 = const()[name = string("last_9_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_9_cast_fp16 = expand_dims(axes = last_9_axes_0, x = var_2733_cast_fp16)[name = string("last_9_cast_fp16")];
            tensor<int32, [1]> var_2737_axes_0 = const()[name = string("op_2737_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_2737_cast_fp16 = expand_dims(axes = var_2737_axes_0, x = tail_15_cast_fp16)[name = string("op_2737_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2738_cast_fp16 = mul(x = k_43_cast_fp16, y = var_2737_cast_fp16)[name = string("op_2738_cast_fp16")];
            bool upd_9_transpose_x_1 = const()[name = string("upd_9_transpose_x_1"), val = bool(true)];
            bool upd_9_transpose_y_1 = const()[name = string("upd_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_9_cast_fp16 = matmul(transpose_x = upd_9_transpose_x_1, transpose_y = upd_9_transpose_y_1, x = var_2738_cast_fp16, y = vnew_13_cast_fp16)[name = string("upd_9_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_2743_cast_fp16 = mul(x = S_13_cast_fp16, y = last_9_cast_fp16)[name = string("op_2743_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_15_cast_fp16 = add(x = var_2743_cast_fp16, y = upd_9_cast_fp16)[name = string("S_15_cast_fp16")];
            tensor<int32, [4]> q_45_begin_0 = const()[name = string("q_45_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_45_end_0 = const()[name = string("q_45_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> q_45_end_mask_0 = const()[name = string("q_45_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_45_cast_fp16 = slice_by_index(begin = q_45_begin_0, end = q_45_end_0, end_mask = q_45_end_mask_0, x = q_39_cast_fp16)[name = string("q_45_cast_fp16")];
            tensor<int32, [4]> k_45_begin_0 = const()[name = string("k_45_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_45_end_0 = const()[name = string("k_45_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> k_45_end_mask_0 = const()[name = string("k_45_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, x = k_39_cast_fp16)[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> kb_21_begin_0 = const()[name = string("kb_21_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_21_end_0 = const()[name = string("kb_21_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> kb_21_end_mask_0 = const()[name = string("kb_21_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_21_cast_fp16 = slice_by_index(begin = kb_21_begin_0, end = kb_21_end_0, end_mask = kb_21_end_mask_0, x = var_2492_cast_fp16)[name = string("kb_21_cast_fp16")];
            tensor<int32, [4]> vb_27_begin_0 = const()[name = string("vb_27_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_27_end_0 = const()[name = string("vb_27_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> vb_27_end_mask_0 = const()[name = string("vb_27_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_27_cast_fp16 = slice_by_index(begin = vb_27_begin_0, end = vb_27_end_0, end_mask = vb_27_end_mask_0, x = vb_21_cast_fp16)[name = string("vb_27_cast_fp16")];
            tensor<int32, [3]> gdec_27_begin_0 = const()[name = string("gdec_27_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_27_end_0 = const()[name = string("gdec_27_end_0"), val = tensor<int32, [3]>([3, 4, 48])];
            tensor<bool, [3]> gdec_27_end_mask_0 = const()[name = string("gdec_27_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_27_cast_fp16 = slice_by_index(begin = gdec_27_begin_0, end = gdec_27_end_0, end_mask = gdec_27_end_mask_0, x = gdec_21_cast_fp16)[name = string("gdec_27_cast_fp16")];
            tensor<int32, [2]> var_2826 = const()[name = string("op_2826"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_2827_cast_fp16 = reshape(shape = var_2826, x = gdec_27_cast_fp16)[name = string("op_2827_cast_fp16")];
            tensor<fp16, [12, 256]> var_2828_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2827_cast_fp16)[name = string("op_2828_cast_fp16")];
            tensor<int32, [4]> var_2833 = const()[name = string("op_2833"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_21_cast_fp16 = reshape(shape = var_2833, x = var_2828_cast_fp16)[name = string("P_21_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2835_cast_fp16 = exp(x = P_21_cast_fp16)[name = string("op_2835_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_21_cast_fp16 = mul(x = var_2835_cast_fp16, y = tril_incl_to_fp16)[name = string("D_21_cast_fp16")];
            bool kk_21_transpose_x_1 = const()[name = string("kk_21_transpose_x_1"), val = bool(false)];
            bool kk_21_transpose_y_1 = const()[name = string("kk_21_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_21_cast_fp16 = matmul(transpose_x = kk_21_transpose_x_1, transpose_y = kk_21_transpose_y_1, x = kb_21_cast_fp16, y = k_45_cast_fp16)[name = string("kk_21_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_2841_cast_fp16 = mul(x = kk_21_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_2841_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2842_cast_fp16 = mul(x = var_2841_cast_fp16, y = D_21_cast_fp16)[name = string("op_2842_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_21_cast_fp16 = mul(x = var_2842_cast_fp16, y = cum_tail_to_fp16)[name = string("N_21_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2844_cast_fp16 = mul(x = D_21_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2844_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_21_cast_fp16 = mul(x = kk_21_cast_fp16, y = var_2844_cast_fp16)[name = string("Na_21_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_81_cast_fp16 = add(x = eye_to_fp16, y = N_21_cast_fp16)[name = string("Minv_81_cast_fp16")];
            bool Np_61_transpose_x_0 = const()[name = string("Np_61_transpose_x_0"), val = bool(false)];
            bool Np_61_transpose_y_0 = const()[name = string("Np_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_61_cast_fp16 = matmul(transpose_x = Np_61_transpose_x_0, transpose_y = Np_61_transpose_y_0, x = N_21_cast_fp16, y = Na_21_cast_fp16)[name = string("Np_61_cast_fp16")];
            bool Npa_41_transpose_x_0 = const()[name = string("Npa_41_transpose_x_0"), val = bool(false)];
            bool Npa_41_transpose_y_0 = const()[name = string("Npa_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_41_cast_fp16 = matmul(transpose_x = Npa_41_transpose_x_0, transpose_y = Npa_41_transpose_y_0, x = Na_21_cast_fp16, y = N_21_cast_fp16)[name = string("Npa_41_cast_fp16")];
            bool var_2850_transpose_x_0 = const()[name = string("op_2850_transpose_x_0"), val = bool(false)];
            bool var_2850_transpose_y_0 = const()[name = string("op_2850_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2850_cast_fp16 = matmul(transpose_x = var_2850_transpose_x_0, transpose_y = var_2850_transpose_y_0, x = Minv_81_cast_fp16, y = Np_61_cast_fp16)[name = string("op_2850_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_83_cast_fp16 = add(x = Minv_81_cast_fp16, y = var_2850_cast_fp16)[name = string("Minv_83_cast_fp16")];
            bool Np_63_transpose_x_0 = const()[name = string("Np_63_transpose_x_0"), val = bool(false)];
            bool Np_63_transpose_y_0 = const()[name = string("Np_63_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_63_cast_fp16 = matmul(transpose_x = Np_63_transpose_x_0, transpose_y = Np_63_transpose_y_0, x = Np_61_cast_fp16, y = Npa_41_cast_fp16)[name = string("Np_63_cast_fp16")];
            bool Npa_43_transpose_x_0 = const()[name = string("Npa_43_transpose_x_0"), val = bool(false)];
            bool Npa_43_transpose_y_0 = const()[name = string("Npa_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_43_cast_fp16 = matmul(transpose_x = Npa_43_transpose_x_0, transpose_y = Npa_43_transpose_y_0, x = Npa_41_cast_fp16, y = Np_61_cast_fp16)[name = string("Npa_43_cast_fp16")];
            bool var_2855_transpose_x_0 = const()[name = string("op_2855_transpose_x_0"), val = bool(false)];
            bool var_2855_transpose_y_0 = const()[name = string("op_2855_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2855_cast_fp16 = matmul(transpose_x = var_2855_transpose_x_0, transpose_y = var_2855_transpose_y_0, x = Minv_83_cast_fp16, y = Np_63_cast_fp16)[name = string("op_2855_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_85_cast_fp16 = add(x = Minv_83_cast_fp16, y = var_2855_cast_fp16)[name = string("Minv_85_cast_fp16")];
            bool Np_65_transpose_x_0 = const()[name = string("Np_65_transpose_x_0"), val = bool(false)];
            bool Np_65_transpose_y_0 = const()[name = string("Np_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_65_cast_fp16 = matmul(transpose_x = Np_65_transpose_x_0, transpose_y = Np_65_transpose_y_0, x = Np_63_cast_fp16, y = Npa_43_cast_fp16)[name = string("Np_65_cast_fp16")];
            bool var_2859_transpose_x_0 = const()[name = string("op_2859_transpose_x_0"), val = bool(false)];
            bool var_2859_transpose_y_0 = const()[name = string("op_2859_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2859_cast_fp16 = matmul(transpose_x = var_2859_transpose_x_0, transpose_y = var_2859_transpose_y_0, x = Minv_85_cast_fp16, y = Np_65_cast_fp16)[name = string("op_2859_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_87_cast_fp16 = add(x = Minv_85_cast_fp16, y = var_2859_cast_fp16)[name = string("Minv_87_cast_fp16")];
            bool u_21_transpose_x_0 = const()[name = string("u_21_transpose_x_0"), val = bool(false)];
            bool u_21_transpose_y_0 = const()[name = string("u_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_21_cast_fp16 = matmul(transpose_x = u_21_transpose_x_0, transpose_y = u_21_transpose_y_0, x = Minv_87_cast_fp16, y = vb_27_cast_fp16)[name = string("u_21_cast_fp16")];
            bool var_2866_transpose_x_1 = const()[name = string("op_2866_transpose_x_1"), val = bool(false)];
            bool var_2866_transpose_y_1 = const()[name = string("op_2866_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_2866_cast_fp16 = matmul(transpose_x = var_2866_transpose_x_1, transpose_y = var_2866_transpose_y_1, x = q_45_cast_fp16, y = k_45_cast_fp16)[name = string("op_2866_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_21_cast_fp16 = mul(x = var_2866_cast_fp16, y = D_21_cast_fp16)[name = string("A_21_cast_fp16")];
            tensor<fp16, [16]> var_2868_bias_0_to_fp16 = const()[name = string("op_2868_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_2868_cast_fp16 = linear(bias = var_2868_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_27_cast_fp16)[name = string("op_2868_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_15_cast_fp16 = exp(x = var_2868_cast_fp16)[name = string("cum_15_cast_fp16")];
            tensor<int32, [1]> var_2871_axes_0 = const()[name = string("op_2871_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_2871_cast_fp16 = expand_dims(axes = var_2871_axes_0, x = cum_15_cast_fp16)[name = string("op_2871_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2872_cast_fp16 = mul(x = kb_21_cast_fp16, y = var_2871_cast_fp16)[name = string("op_2872_cast_fp16")];
            bool w_21_transpose_x_0 = const()[name = string("w_21_transpose_x_0"), val = bool(false)];
            bool w_21_transpose_y_0 = const()[name = string("w_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_21_cast_fp16 = matmul(transpose_x = w_21_transpose_x_0, transpose_y = w_21_transpose_y_0, x = Minv_87_cast_fp16, y = var_2872_cast_fp16)[name = string("w_21_cast_fp16")];
            bool var_2874_transpose_x_0 = const()[name = string("op_2874_transpose_x_0"), val = bool(false)];
            bool var_2874_transpose_y_0 = const()[name = string("op_2874_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2874_cast_fp16 = matmul(transpose_x = var_2874_transpose_x_0, transpose_y = var_2874_transpose_y_0, x = w_21_cast_fp16, y = S_15_cast_fp16)[name = string("op_2874_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_15_cast_fp16 = sub(x = u_21_cast_fp16, y = var_2874_cast_fp16)[name = string("vnew_15_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2879_cast_fp16 = mul(x = q_45_cast_fp16, y = var_2871_cast_fp16)[name = string("op_2879_cast_fp16")];
            bool var_2880_transpose_x_0 = const()[name = string("op_2880_transpose_x_0"), val = bool(false)];
            bool var_2880_transpose_y_0 = const()[name = string("op_2880_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2880_cast_fp16 = matmul(transpose_x = var_2880_transpose_x_0, transpose_y = var_2880_transpose_y_0, x = var_2879_cast_fp16, y = S_15_cast_fp16)[name = string("op_2880_cast_fp16")];
            bool var_2881_transpose_x_0 = const()[name = string("op_2881_transpose_x_0"), val = bool(false)];
            bool var_2881_transpose_y_0 = const()[name = string("op_2881_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2881_cast_fp16 = matmul(transpose_x = var_2881_transpose_x_0, transpose_y = var_2881_transpose_y_0, x = A_21_cast_fp16, y = vnew_15_cast_fp16)[name = string("op_2881_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2883_cast_fp16 = add(x = var_2880_cast_fp16, y = var_2881_cast_fp16)[name = string("op_2883_cast_fp16")];
            tensor<fp16, [16]> var_2884_bias_0_to_fp16 = const()[name = string("op_2884_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_2884_cast_fp16 = linear(bias = var_2884_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_27_cast_fp16)[name = string("op_2884_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_17_cast_fp16 = exp(x = var_2884_cast_fp16)[name = string("tail_17_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_33_to_fp16 = const()[name = string("transpose_33_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_2886_bias_0_to_fp16 = const()[name = string("op_2886_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_2886_cast_fp16 = linear(bias = var_2886_bias_0_to_fp16, weight = transpose_33_to_fp16, x = gdec_27_cast_fp16)[name = string("op_2886_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_2887_cast_fp16 = exp(x = var_2886_cast_fp16)[name = string("op_2887_cast_fp16")];
            tensor<int32, [1]> last_11_axes_0 = const()[name = string("last_11_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_11_cast_fp16 = expand_dims(axes = last_11_axes_0, x = var_2887_cast_fp16)[name = string("last_11_cast_fp16")];
            tensor<int32, [1]> var_2891_axes_0 = const()[name = string("op_2891_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_2891_cast_fp16 = expand_dims(axes = var_2891_axes_0, x = tail_17_cast_fp16)[name = string("op_2891_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2892_cast_fp16 = mul(x = k_45_cast_fp16, y = var_2891_cast_fp16)[name = string("op_2892_cast_fp16")];
            bool upd_11_transpose_x_1 = const()[name = string("upd_11_transpose_x_1"), val = bool(true)];
            bool upd_11_transpose_y_1 = const()[name = string("upd_11_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_11_cast_fp16 = matmul(transpose_x = upd_11_transpose_x_1, transpose_y = upd_11_transpose_y_1, x = var_2892_cast_fp16, y = vnew_15_cast_fp16)[name = string("upd_11_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_2897_cast_fp16 = mul(x = S_15_cast_fp16, y = last_11_cast_fp16)[name = string("op_2897_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_17_cast_fp16 = add(x = var_2897_cast_fp16, y = upd_11_cast_fp16)[name = string("S_17_cast_fp16")];
            tensor<int32, [4]> q_47_begin_0 = const()[name = string("q_47_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_47_end_0 = const()[name = string("q_47_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_47_end_mask_0 = const()[name = string("q_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_47_cast_fp16 = slice_by_index(begin = q_47_begin_0, end = q_47_end_0, end_mask = q_47_end_mask_0, x = q_39_cast_fp16)[name = string("q_47_cast_fp16")];
            tensor<int32, [4]> k_47_begin_0 = const()[name = string("k_47_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_47_end_0 = const()[name = string("k_47_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_47_end_mask_0 = const()[name = string("k_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_47_cast_fp16 = slice_by_index(begin = k_47_begin_0, end = k_47_end_0, end_mask = k_47_end_mask_0, x = k_39_cast_fp16)[name = string("k_47_cast_fp16")];
            tensor<int32, [4]> kb_23_begin_0 = const()[name = string("kb_23_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_23_end_0 = const()[name = string("kb_23_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_23_end_mask_0 = const()[name = string("kb_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_23_cast_fp16 = slice_by_index(begin = kb_23_begin_0, end = kb_23_end_0, end_mask = kb_23_end_mask_0, x = var_2492_cast_fp16)[name = string("kb_23_cast_fp16")];
            tensor<int32, [4]> vb_29_begin_0 = const()[name = string("vb_29_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_29_end_0 = const()[name = string("vb_29_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_29_end_mask_0 = const()[name = string("vb_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_29_cast_fp16 = slice_by_index(begin = vb_29_begin_0, end = vb_29_end_0, end_mask = vb_29_end_mask_0, x = vb_21_cast_fp16)[name = string("vb_29_cast_fp16")];
            tensor<int32, [3]> gdec_29_begin_0 = const()[name = string("gdec_29_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_29_end_0 = const()[name = string("gdec_29_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_29_end_mask_0 = const()[name = string("gdec_29_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_29_cast_fp16 = slice_by_index(begin = gdec_29_begin_0, end = gdec_29_end_0, end_mask = gdec_29_end_mask_0, x = gdec_21_cast_fp16)[name = string("gdec_29_cast_fp16")];
            tensor<int32, [2]> var_2980 = const()[name = string("op_2980"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_2981_cast_fp16 = reshape(shape = var_2980, x = gdec_29_cast_fp16)[name = string("op_2981_cast_fp16")];
            tensor<fp16, [12, 256]> var_2982_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2981_cast_fp16)[name = string("op_2982_cast_fp16")];
            tensor<int32, [4]> var_2987 = const()[name = string("op_2987"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_23_cast_fp16 = reshape(shape = var_2987, x = var_2982_cast_fp16)[name = string("P_23_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2989_cast_fp16 = exp(x = P_23_cast_fp16)[name = string("op_2989_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_23_cast_fp16 = mul(x = var_2989_cast_fp16, y = tril_incl_to_fp16)[name = string("D_23_cast_fp16")];
            bool kk_23_transpose_x_1 = const()[name = string("kk_23_transpose_x_1"), val = bool(false)];
            bool kk_23_transpose_y_1 = const()[name = string("kk_23_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_23_cast_fp16 = matmul(transpose_x = kk_23_transpose_x_1, transpose_y = kk_23_transpose_y_1, x = kb_23_cast_fp16, y = k_47_cast_fp16)[name = string("kk_23_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_2995_cast_fp16 = mul(x = kk_23_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_2995_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2996_cast_fp16 = mul(x = var_2995_cast_fp16, y = D_23_cast_fp16)[name = string("op_2996_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_23_cast_fp16 = mul(x = var_2996_cast_fp16, y = cum_tail_to_fp16)[name = string("N_23_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2998_cast_fp16 = mul(x = D_23_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2998_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_23_cast_fp16 = mul(x = kk_23_cast_fp16, y = var_2998_cast_fp16)[name = string("Na_23_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_89_cast_fp16 = add(x = eye_to_fp16, y = N_23_cast_fp16)[name = string("Minv_89_cast_fp16")];
            bool Np_67_transpose_x_0 = const()[name = string("Np_67_transpose_x_0"), val = bool(false)];
            bool Np_67_transpose_y_0 = const()[name = string("Np_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_67_cast_fp16 = matmul(transpose_x = Np_67_transpose_x_0, transpose_y = Np_67_transpose_y_0, x = N_23_cast_fp16, y = Na_23_cast_fp16)[name = string("Np_67_cast_fp16")];
            bool Npa_45_transpose_x_0 = const()[name = string("Npa_45_transpose_x_0"), val = bool(false)];
            bool Npa_45_transpose_y_0 = const()[name = string("Npa_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_45_cast_fp16 = matmul(transpose_x = Npa_45_transpose_x_0, transpose_y = Npa_45_transpose_y_0, x = Na_23_cast_fp16, y = N_23_cast_fp16)[name = string("Npa_45_cast_fp16")];
            bool var_3004_transpose_x_0 = const()[name = string("op_3004_transpose_x_0"), val = bool(false)];
            bool var_3004_transpose_y_0 = const()[name = string("op_3004_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3004_cast_fp16 = matmul(transpose_x = var_3004_transpose_x_0, transpose_y = var_3004_transpose_y_0, x = Minv_89_cast_fp16, y = Np_67_cast_fp16)[name = string("op_3004_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_91_cast_fp16 = add(x = Minv_89_cast_fp16, y = var_3004_cast_fp16)[name = string("Minv_91_cast_fp16")];
            bool Np_69_transpose_x_0 = const()[name = string("Np_69_transpose_x_0"), val = bool(false)];
            bool Np_69_transpose_y_0 = const()[name = string("Np_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_69_cast_fp16 = matmul(transpose_x = Np_69_transpose_x_0, transpose_y = Np_69_transpose_y_0, x = Np_67_cast_fp16, y = Npa_45_cast_fp16)[name = string("Np_69_cast_fp16")];
            bool Npa_47_transpose_x_0 = const()[name = string("Npa_47_transpose_x_0"), val = bool(false)];
            bool Npa_47_transpose_y_0 = const()[name = string("Npa_47_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_47_cast_fp16 = matmul(transpose_x = Npa_47_transpose_x_0, transpose_y = Npa_47_transpose_y_0, x = Npa_45_cast_fp16, y = Np_67_cast_fp16)[name = string("Npa_47_cast_fp16")];
            bool var_3009_transpose_x_0 = const()[name = string("op_3009_transpose_x_0"), val = bool(false)];
            bool var_3009_transpose_y_0 = const()[name = string("op_3009_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3009_cast_fp16 = matmul(transpose_x = var_3009_transpose_x_0, transpose_y = var_3009_transpose_y_0, x = Minv_91_cast_fp16, y = Np_69_cast_fp16)[name = string("op_3009_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_93_cast_fp16 = add(x = Minv_91_cast_fp16, y = var_3009_cast_fp16)[name = string("Minv_93_cast_fp16")];
            bool Np_71_transpose_x_0 = const()[name = string("Np_71_transpose_x_0"), val = bool(false)];
            bool Np_71_transpose_y_0 = const()[name = string("Np_71_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_71_cast_fp16 = matmul(transpose_x = Np_71_transpose_x_0, transpose_y = Np_71_transpose_y_0, x = Np_69_cast_fp16, y = Npa_47_cast_fp16)[name = string("Np_71_cast_fp16")];
            bool var_3013_transpose_x_0 = const()[name = string("op_3013_transpose_x_0"), val = bool(false)];
            bool var_3013_transpose_y_0 = const()[name = string("op_3013_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3013_cast_fp16 = matmul(transpose_x = var_3013_transpose_x_0, transpose_y = var_3013_transpose_y_0, x = Minv_93_cast_fp16, y = Np_71_cast_fp16)[name = string("op_3013_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_95_cast_fp16 = add(x = Minv_93_cast_fp16, y = var_3013_cast_fp16)[name = string("Minv_95_cast_fp16")];
            bool u_23_transpose_x_0 = const()[name = string("u_23_transpose_x_0"), val = bool(false)];
            bool u_23_transpose_y_0 = const()[name = string("u_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_23_cast_fp16 = matmul(transpose_x = u_23_transpose_x_0, transpose_y = u_23_transpose_y_0, x = Minv_95_cast_fp16, y = vb_29_cast_fp16)[name = string("u_23_cast_fp16")];
            bool var_3020_transpose_x_1 = const()[name = string("op_3020_transpose_x_1"), val = bool(false)];
            bool var_3020_transpose_y_1 = const()[name = string("op_3020_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_3020_cast_fp16 = matmul(transpose_x = var_3020_transpose_x_1, transpose_y = var_3020_transpose_y_1, x = q_47_cast_fp16, y = k_47_cast_fp16)[name = string("op_3020_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_23_cast_fp16 = mul(x = var_3020_cast_fp16, y = D_23_cast_fp16)[name = string("A_23_cast_fp16")];
            tensor<fp16, [16]> var_3022_bias_0_to_fp16 = const()[name = string("op_3022_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_3022_cast_fp16 = linear(bias = var_3022_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_29_cast_fp16)[name = string("op_3022_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_17_cast_fp16 = exp(x = var_3022_cast_fp16)[name = string("cum_17_cast_fp16")];
            tensor<int32, [1]> var_3025_axes_0 = const()[name = string("op_3025_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_3025_cast_fp16 = expand_dims(axes = var_3025_axes_0, x = cum_17_cast_fp16)[name = string("op_3025_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3026_cast_fp16 = mul(x = kb_23_cast_fp16, y = var_3025_cast_fp16)[name = string("op_3026_cast_fp16")];
            bool w_23_transpose_x_0 = const()[name = string("w_23_transpose_x_0"), val = bool(false)];
            bool w_23_transpose_y_0 = const()[name = string("w_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_23_cast_fp16 = matmul(transpose_x = w_23_transpose_x_0, transpose_y = w_23_transpose_y_0, x = Minv_95_cast_fp16, y = var_3026_cast_fp16)[name = string("w_23_cast_fp16")];
            bool var_3028_transpose_x_0 = const()[name = string("op_3028_transpose_x_0"), val = bool(false)];
            bool var_3028_transpose_y_0 = const()[name = string("op_3028_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3028_cast_fp16 = matmul(transpose_x = var_3028_transpose_x_0, transpose_y = var_3028_transpose_y_0, x = w_23_cast_fp16, y = S_17_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_17_cast_fp16 = sub(x = u_23_cast_fp16, y = var_3028_cast_fp16)[name = string("vnew_17_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3033_cast_fp16 = mul(x = q_47_cast_fp16, y = var_3025_cast_fp16)[name = string("op_3033_cast_fp16")];
            bool var_3034_transpose_x_0 = const()[name = string("op_3034_transpose_x_0"), val = bool(false)];
            bool var_3034_transpose_y_0 = const()[name = string("op_3034_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3034_cast_fp16 = matmul(transpose_x = var_3034_transpose_x_0, transpose_y = var_3034_transpose_y_0, x = var_3033_cast_fp16, y = S_17_cast_fp16)[name = string("op_3034_cast_fp16")];
            bool var_3035_transpose_x_0 = const()[name = string("op_3035_transpose_x_0"), val = bool(false)];
            bool var_3035_transpose_y_0 = const()[name = string("op_3035_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3035_cast_fp16 = matmul(transpose_x = var_3035_transpose_x_0, transpose_y = var_3035_transpose_y_0, x = A_23_cast_fp16, y = vnew_17_cast_fp16)[name = string("op_3035_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_5_cast_fp16 = add(x = var_3034_cast_fp16, y = var_3035_cast_fp16)[name = string("oc_5_cast_fp16")];
            int32 var_3039 = const()[name = string("op_3039"), val = int32(2)];
            bool o_13_interleave_0 = const()[name = string("o_13_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 64, 128]> o_13_cast_fp16 = concat(axis = var_3039, interleave = o_13_interleave_0, values = (var_2582_cast_fp16, var_2729_cast_fp16, var_2883_cast_fp16, oc_5_cast_fp16))[name = string("o_13_cast_fp16")];
            tensor<int32, [4]> var_3045 = const()[name = string("op_3045"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3051 = const()[name = string("op_3051"), val = tensor<int32, [4]>([3, 64, 4, 128])];
            tensor<fp16, [3, 64, 4, 128]> input_39_cast_fp16 = reshape(shape = var_3051, x = z_5_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> o_15_cast_fp16 = transpose(perm = var_3045, x = o_13_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [3, 64, 4, 128]> var_3053_cast_fp16 = mul(x = o_15_cast_fp16, y = o_15_cast_fp16)[name = string("op_3053_cast_fp16")];
            tensor<int32, [1]> var_3058_axes_0 = const()[name = string("op_3058_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3058_keep_dims_0 = const()[name = string("op_3058_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 4, 1]> var_3058_cast_fp16 = reduce_mean(axes = var_3058_axes_0, keep_dims = var_3058_keep_dims_0, x = var_3053_cast_fp16)[name = string("op_3058_cast_fp16")];
            fp16 var_3060_to_fp16 = const()[name = string("op_3060_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 64, 4, 1]> var_3061_cast_fp16 = add(x = var_3058_cast_fp16, y = var_3060_to_fp16)[name = string("op_3061_cast_fp16")];
            fp32 var_3062_epsilon_0 = const()[name = string("op_3062_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 4, 1]> var_3062_cast_fp16 = rsqrt(epsilon = var_3062_epsilon_0, x = var_3061_cast_fp16)[name = string("op_3062_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> var_3063_cast_fp16 = mul(x = o_15_cast_fp16, y = var_3062_cast_fp16)[name = string("op_3063_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> var_3065_cast_fp16 = silu(x = input_39_cast_fp16)[name = string("op_3065_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> o_17_cast_fp16 = mul(x = var_3063_cast_fp16, y = var_3065_cast_fp16)[name = string("o_17_cast_fp16")];
            tensor<int32, [3]> var_3070 = const()[name = string("op_3070"), val = tensor<int32, [3]>([3, 64, 512])];
            tensor<fp16, [3, 64, 512]> input_41_cast_fp16 = reshape(shape = var_3070, x = o_17_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [640, 512]> layers_2_out_weight_to_fp16 = const()[name = string("layers_2_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47457728)))];
            tensor<fp16, [3, 64, 640]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_out_weight_to_fp16, x = input_41_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [3, 64, 640]> x_21_cast_fp16 = add(x = x_17_cast_fp16, y = linear_9_cast_fp16)[name = string("x_21_cast_fp16")];
            fp16 var_3081_promoted_to_fp16 = const()[name = string("op_3081_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> var_3085_cast_fp16 = pow(x = x_21_cast_fp16, y = var_3081_promoted_to_fp16)[name = string("op_3085_cast_fp16")];
            tensor<int32, [1]> var_3087_axes_0 = const()[name = string("op_3087_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3087_keep_dims_0 = const()[name = string("op_3087_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 1]> var_3087_cast_fp16 = reduce_mean(axes = var_3087_axes_0, keep_dims = var_3087_keep_dims_0, x = var_3085_cast_fp16)[name = string("op_3087_cast_fp16")];
            fp16 var_3088_to_fp16 = const()[name = string("op_3088_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 64, 1]> var_3089_cast_fp16 = add(x = var_3087_cast_fp16, y = var_3088_to_fp16)[name = string("op_3089_cast_fp16")];
            fp32 var_3090_epsilon_0 = const()[name = string("op_3090_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 1]> var_3090_cast_fp16 = rsqrt(epsilon = var_3090_epsilon_0, x = var_3089_cast_fp16)[name = string("op_3090_cast_fp16")];
            tensor<fp16, [3, 64, 640]> var_3091_cast_fp16 = mul(x = x_21_cast_fp16, y = var_3090_cast_fp16)[name = string("op_3091_cast_fp16")];
            tensor<fp16, [640]> layers_2_n2_w1_to_fp16 = const()[name = string("layers_2_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48113152)))];
            tensor<fp16, [3, 64, 640]> input_43_cast_fp16 = mul(x = var_3091_cast_fp16, y = layers_2_n2_w1_to_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_2_gate_up_weight_to_fp16 = const()[name = string("layers_2_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48114496)))];
            tensor<fp16, [3, 64, 5120]> linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_2_gate_up_weight_to_fp16, x = input_43_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [3]> input_45_begin_0 = const()[name = string("input_45_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_45_end_0 = const()[name = string("input_45_end_0"), val = tensor<int32, [3]>([3, 64, 2560])];
            tensor<bool, [3]> input_45_end_mask_0 = const()[name = string("input_45_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 2560]> input_45_cast_fp16 = slice_by_index(begin = input_45_begin_0, end = input_45_end_0, end_mask = input_45_end_mask_0, x = linear_10_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [3, 64, 2560]> var_3101_cast_fp16 = silu(x = input_45_cast_fp16)[name = string("op_3101_cast_fp16")];
            tensor<int32, [3]> var_3106_begin_0 = const()[name = string("op_3106_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_3106_end_0 = const()[name = string("op_3106_end_0"), val = tensor<int32, [3]>([3, 64, 5120])];
            tensor<bool, [3]> var_3106_end_mask_0 = const()[name = string("op_3106_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 2560]> var_3106_cast_fp16 = slice_by_index(begin = var_3106_begin_0, end = var_3106_end_0, end_mask = var_3106_end_mask_0, x = linear_10_cast_fp16)[name = string("op_3106_cast_fp16")];
            tensor<fp16, [3, 64, 2560]> input_47_cast_fp16 = mul(x = var_3101_cast_fp16, y = var_3106_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_2_down_weight_to_fp16 = const()[name = string("layers_2_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54668160)))];
            tensor<fp16, [3, 64, 640]> linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_down_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [3, 64, 640]> x_25_cast_fp16 = add(x = x_21_cast_fp16, y = linear_11_cast_fp16)[name = string("x_25_cast_fp16")];
            fp16 var_3119_to_fp16 = const()[name = string("op_3119_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 64, 640]> x_27_cast_fp16 = mul(x = x_25_cast_fp16, y = var_3119_to_fp16)[name = string("x_27_cast_fp16")];
            fp16 var_3117_promoted_to_fp16 = const()[name = string("op_3117_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> var_3121_cast_fp16 = pow(x = x_27_cast_fp16, y = var_3117_promoted_to_fp16)[name = string("op_3121_cast_fp16")];
            tensor<int32, [1]> var_3123_axes_0 = const()[name = string("op_3123_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3123_keep_dims_0 = const()[name = string("op_3123_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 1]> var_3123_cast_fp16 = reduce_mean(axes = var_3123_axes_0, keep_dims = var_3123_keep_dims_0, x = var_3121_cast_fp16)[name = string("op_3123_cast_fp16")];
            fp16 var_3124_to_fp16 = const()[name = string("op_3124_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 64, 1]> var_3125_cast_fp16 = add(x = var_3123_cast_fp16, y = var_3124_to_fp16)[name = string("op_3125_cast_fp16")];
            fp32 var_3126_epsilon_0 = const()[name = string("op_3126_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 1]> var_3126_cast_fp16 = rsqrt(epsilon = var_3126_epsilon_0, x = var_3125_cast_fp16)[name = string("op_3126_cast_fp16")];
            tensor<fp16, [3, 64, 640]> var_3127_cast_fp16 = mul(x = x_27_cast_fp16, y = var_3126_cast_fp16)[name = string("op_3127_cast_fp16")];
            tensor<fp16, [640]> layers_3_n1_w1_to_fp16 = const()[name = string("layers_3_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57945024)))];
            tensor<fp16, [3, 64, 640]> input_49_cast_fp16 = mul(x = var_3127_cast_fp16, y = layers_3_n1_w1_to_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1280, 640]> layers_3_qkv_weight_to_fp16 = const()[name = string("layers_3_qkv_weight_to_fp16"), val = tensor<fp16, [1280, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57946368)))];
            tensor<fp16, [1280]> linear_12_bias_0_to_fp16 = const()[name = string("linear_12_bias_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59584832)))];
            tensor<fp16, [3, 64, 1280]> linear_12_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = layers_3_qkv_weight_to_fp16, x = input_49_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [3]> var_3136_begin_0 = const()[name = string("op_3136_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3136_end_0 = const()[name = string("op_3136_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> var_3136_end_mask_0 = const()[name = string("op_3136_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 1024]> var_3136_cast_fp16 = slice_by_index(begin = var_3136_begin_0, end = var_3136_end_0, end_mask = var_3136_end_mask_0, x = linear_12_cast_fp16)[name = string("op_3136_cast_fp16")];
            tensor<int32, [4]> var_3141 = const()[name = string("op_3141"), val = tensor<int32, [4]>([3, 64, 8, 128])];
            tensor<fp16, [3, 64, 8, 128]> qg_1_cast_fp16 = reshape(shape = var_3141, x = var_3136_cast_fp16)[name = string("qg_1_cast_fp16")];
            tensor<int32, [4]> x_29_begin_0 = const()[name = string("x_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_29_end_0 = const()[name = string("x_29_end_0"), val = tensor<int32, [4]>([3, 64, 8, 64])];
            tensor<bool, [4]> x_29_end_mask_0 = const()[name = string("x_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 64, 8, 64]> x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, x = qg_1_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [4]> var_3152_begin_0 = const()[name = string("op_3152_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> var_3152_end_0 = const()[name = string("op_3152_end_0"), val = tensor<int32, [4]>([3, 64, 8, 128])];
            tensor<bool, [4]> var_3152_end_mask_0 = const()[name = string("op_3152_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 64, 8, 64]> var_3152_cast_fp16 = slice_by_index(begin = var_3152_begin_0, end = var_3152_end_0, end_mask = var_3152_end_mask_0, x = qg_1_cast_fp16)[name = string("op_3152_cast_fp16")];
            tensor<int32, [3]> var_3156 = const()[name = string("op_3156"), val = tensor<int32, [3]>([3, 64, 512])];
            tensor<fp16, [3, 64, 512]> gate_1_cast_fp16 = reshape(shape = var_3156, x = var_3152_cast_fp16)[name = string("gate_1_cast_fp16")];
            tensor<int32, [3]> var_3162_begin_0 = const()[name = string("op_3162_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> var_3162_end_0 = const()[name = string("op_3162_end_0"), val = tensor<int32, [3]>([3, 64, 1152])];
            tensor<bool, [3]> var_3162_end_mask_0 = const()[name = string("op_3162_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 128]> var_3162_cast_fp16 = slice_by_index(begin = var_3162_begin_0, end = var_3162_end_0, end_mask = var_3162_end_mask_0, x = linear_12_cast_fp16)[name = string("op_3162_cast_fp16")];
            tensor<int32, [4]> var_3167 = const()[name = string("op_3167"), val = tensor<int32, [4]>([3, 64, 2, 64])];
            tensor<fp16, [3, 64, 2, 64]> x_35_cast_fp16 = reshape(shape = var_3167, x = var_3162_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<int32, [3]> var_3173_begin_0 = const()[name = string("op_3173_begin_0"), val = tensor<int32, [3]>([0, 0, 1152])];
            tensor<int32, [3]> var_3173_end_0 = const()[name = string("op_3173_end_0"), val = tensor<int32, [3]>([3, 64, 1280])];
            tensor<bool, [3]> var_3173_end_mask_0 = const()[name = string("op_3173_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 128]> var_3173_cast_fp16 = slice_by_index(begin = var_3173_begin_0, end = var_3173_end_0, end_mask = var_3173_end_mask_0, x = linear_12_cast_fp16)[name = string("op_3173_cast_fp16")];
            tensor<int32, [4]> var_3178 = const()[name = string("op_3178"), val = tensor<int32, [4]>([3, 64, 2, 64])];
            tensor<fp16, [3, 64, 2, 64]> var_3179_cast_fp16 = reshape(shape = var_3178, x = var_3173_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<int32, [4]> var_3184 = const()[name = string("op_3184"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3190_promoted_to_fp16 = const()[name = string("op_3190_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 8, 64]> var_3194_cast_fp16 = pow(x = x_29_cast_fp16, y = var_3190_promoted_to_fp16)[name = string("op_3194_cast_fp16")];
            tensor<int32, [1]> var_3196_axes_0 = const()[name = string("op_3196_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3196_keep_dims_0 = const()[name = string("op_3196_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 8, 1]> var_3196_cast_fp16 = reduce_mean(axes = var_3196_axes_0, keep_dims = var_3196_keep_dims_0, x = var_3194_cast_fp16)[name = string("op_3196_cast_fp16")];
            fp16 var_3197_to_fp16 = const()[name = string("op_3197_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 64, 8, 1]> var_3198_cast_fp16 = add(x = var_3196_cast_fp16, y = var_3197_to_fp16)[name = string("op_3198_cast_fp16")];
            fp32 var_3199_epsilon_0 = const()[name = string("op_3199_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 8, 1]> var_3199_cast_fp16 = rsqrt(epsilon = var_3199_epsilon_0, x = var_3198_cast_fp16)[name = string("op_3199_cast_fp16")];
            tensor<fp16, [3, 64, 8, 64]> var_3200_cast_fp16 = mul(x = x_29_cast_fp16, y = var_3199_cast_fp16)[name = string("op_3200_cast_fp16")];
            tensor<fp16, [64]> layers_3_qn_w1_to_fp16 = const()[name = string("layers_3_qn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59587456)))];
            tensor<fp16, [3, 64, 8, 64]> var_3201_cast_fp16 = mul(x = var_3200_cast_fp16, y = layers_3_qn_w1_to_fp16)[name = string("op_3201_cast_fp16")];
            tensor<int32, [4]> var_3206 = const()[name = string("op_3206"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_1_begin_0 = const()[name = string("r_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_1_end_0 = const()[name = string("r_1_end_0"), val = tensor<int32, [4]>([3, 8, 64, 16])];
            tensor<bool, [4]> r_1_end_mask_0 = const()[name = string("r_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 8, 64, 64]> x_33_cast_fp16 = transpose(perm = var_3206, x = var_3201_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [3, 8, 64, 16]> r_1_cast_fp16 = slice_by_index(begin = r_1_begin_0, end = r_1_end_0, end_mask = r_1_end_mask_0, x = x_33_cast_fp16)[name = string("r_1_cast_fp16")];
            tensor<int32, [4]> p_1_begin_0 = const()[name = string("p_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_1_end_0 = const()[name = string("p_1_end_0"), val = tensor<int32, [4]>([3, 8, 64, 64])];
            tensor<bool, [4]> p_1_end_mask_0 = const()[name = string("p_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 8, 64, 48]> p_1_cast_fp16 = slice_by_index(begin = p_1_begin_0, end = p_1_end_0, end_mask = p_1_end_mask_0, x = x_33_cast_fp16)[name = string("p_1_cast_fp16")];
            tensor<int32, [4]> var_3222_begin_0 = const()[name = string("op_3222_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_3222_end_0 = const()[name = string("op_3222_end_0"), val = tensor<int32, [4]>([3, 8, 64, 16])];
            tensor<bool, [4]> var_3222_end_mask_0 = const()[name = string("op_3222_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 8, 64, 8]> var_3222_cast_fp16 = slice_by_index(begin = var_3222_begin_0, end = var_3222_end_0, end_mask = var_3222_end_mask_0, x = r_1_cast_fp16)[name = string("op_3222_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 8, 64, 8]> var_3223_cast_fp16 = mul(x = var_3222_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_3223_cast_fp16")];
            tensor<int32, [4]> var_3228_begin_0 = const()[name = string("op_3228_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3228_end_0 = const()[name = string("op_3228_end_0"), val = tensor<int32, [4]>([3, 8, 64, 8])];
            tensor<bool, [4]> var_3228_end_mask_0 = const()[name = string("op_3228_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 8, 64, 8]> var_3228_cast_fp16 = slice_by_index(begin = var_3228_begin_0, end = var_3228_end_0, end_mask = var_3228_end_mask_0, x = r_1_cast_fp16)[name = string("op_3228_cast_fp16")];
            int32 var_3230 = const()[name = string("op_3230"), val = int32(-1)];
            bool rh_1_interleave_0 = const()[name = string("rh_1_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 8, 64, 16]> rh_1_cast_fp16 = concat(axis = var_3230, interleave = rh_1_interleave_0, values = (var_3223_cast_fp16, var_3228_cast_fp16))[name = string("rh_1_cast_fp16")];
            tensor<fp16, [64, 16]> cos_to_fp16 = const()[name = string("cos_to_fp16"), val = tensor<fp16, [64, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663296)))];
            tensor<fp16, [3, 8, 64, 16]> var_3232_cast_fp16 = mul(x = r_1_cast_fp16, y = cos_to_fp16)[name = string("op_3232_cast_fp16")];
            tensor<fp16, [64, 16]> sin_to_fp16 = const()[name = string("sin_to_fp16"), val = tensor<fp16, [64, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119665408)))];
            tensor<fp16, [3, 8, 64, 16]> var_3233_cast_fp16 = mul(x = rh_1_cast_fp16, y = sin_to_fp16)[name = string("op_3233_cast_fp16")];
            tensor<fp16, [3, 8, 64, 16]> var_3235_cast_fp16 = add(x = var_3232_cast_fp16, y = var_3233_cast_fp16)[name = string("op_3235_cast_fp16")];
            int32 var_3237 = const()[name = string("op_3237"), val = int32(-1)];
            bool q_49_interleave_0 = const()[name = string("q_49_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 8, 64, 64]> q_49_cast_fp16 = concat(axis = var_3237, interleave = q_49_interleave_0, values = (var_3235_cast_fp16, p_1_cast_fp16))[name = string("q_49_cast_fp16")];
            fp16 var_3243_promoted_to_fp16 = const()[name = string("op_3243_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 2, 64]> var_3247_cast_fp16 = pow(x = x_35_cast_fp16, y = var_3243_promoted_to_fp16)[name = string("op_3247_cast_fp16")];
            tensor<int32, [1]> var_3249_axes_0 = const()[name = string("op_3249_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3249_keep_dims_0 = const()[name = string("op_3249_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 2, 1]> var_3249_cast_fp16 = reduce_mean(axes = var_3249_axes_0, keep_dims = var_3249_keep_dims_0, x = var_3247_cast_fp16)[name = string("op_3249_cast_fp16")];
            fp16 var_3250_to_fp16 = const()[name = string("op_3250_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 64, 2, 1]> var_3251_cast_fp16 = add(x = var_3249_cast_fp16, y = var_3250_to_fp16)[name = string("op_3251_cast_fp16")];
            fp32 var_3252_epsilon_0 = const()[name = string("op_3252_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 2, 1]> var_3252_cast_fp16 = rsqrt(epsilon = var_3252_epsilon_0, x = var_3251_cast_fp16)[name = string("op_3252_cast_fp16")];
            tensor<fp16, [3, 64, 2, 64]> var_3253_cast_fp16 = mul(x = x_35_cast_fp16, y = var_3252_cast_fp16)[name = string("op_3253_cast_fp16")];
            tensor<fp16, [64]> layers_3_kn_w1_to_fp16 = const()[name = string("layers_3_kn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59588800)))];
            tensor<fp16, [3, 64, 2, 64]> var_3254_cast_fp16 = mul(x = var_3253_cast_fp16, y = layers_3_kn_w1_to_fp16)[name = string("op_3254_cast_fp16")];
            tensor<int32, [4]> var_3259 = const()[name = string("op_3259"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_3_begin_0 = const()[name = string("r_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_3_end_0 = const()[name = string("r_3_end_0"), val = tensor<int32, [4]>([3, 2, 64, 16])];
            tensor<bool, [4]> r_3_end_mask_0 = const()[name = string("r_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 2, 64, 64]> x_39_cast_fp16 = transpose(perm = var_3259, x = var_3254_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [3, 2, 64, 16]> r_3_cast_fp16 = slice_by_index(begin = r_3_begin_0, end = r_3_end_0, end_mask = r_3_end_mask_0, x = x_39_cast_fp16)[name = string("r_3_cast_fp16")];
            tensor<int32, [4]> p_3_begin_0 = const()[name = string("p_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_3_end_0 = const()[name = string("p_3_end_0"), val = tensor<int32, [4]>([3, 2, 64, 64])];
            tensor<bool, [4]> p_3_end_mask_0 = const()[name = string("p_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 2, 64, 48]> p_3_cast_fp16 = slice_by_index(begin = p_3_begin_0, end = p_3_end_0, end_mask = p_3_end_mask_0, x = x_39_cast_fp16)[name = string("p_3_cast_fp16")];
            tensor<int32, [4]> var_3275_begin_0 = const()[name = string("op_3275_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_3275_end_0 = const()[name = string("op_3275_end_0"), val = tensor<int32, [4]>([3, 2, 64, 16])];
            tensor<bool, [4]> var_3275_end_mask_0 = const()[name = string("op_3275_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 2, 64, 8]> var_3275_cast_fp16 = slice_by_index(begin = var_3275_begin_0, end = var_3275_end_0, end_mask = var_3275_end_mask_0, x = r_3_cast_fp16)[name = string("op_3275_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 2, 64, 8]> var_3276_cast_fp16 = mul(x = var_3275_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_3276_cast_fp16")];
            tensor<int32, [4]> var_3281_begin_0 = const()[name = string("op_3281_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3281_end_0 = const()[name = string("op_3281_end_0"), val = tensor<int32, [4]>([3, 2, 64, 8])];
            tensor<bool, [4]> var_3281_end_mask_0 = const()[name = string("op_3281_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 2, 64, 8]> var_3281_cast_fp16 = slice_by_index(begin = var_3281_begin_0, end = var_3281_end_0, end_mask = var_3281_end_mask_0, x = r_3_cast_fp16)[name = string("op_3281_cast_fp16")];
            int32 var_3283 = const()[name = string("op_3283"), val = int32(-1)];
            bool rh_3_interleave_0 = const()[name = string("rh_3_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 2, 64, 16]> rh_3_cast_fp16 = concat(axis = var_3283, interleave = rh_3_interleave_0, values = (var_3276_cast_fp16, var_3281_cast_fp16))[name = string("rh_3_cast_fp16")];
            tensor<fp16, [3, 2, 64, 16]> var_3285_cast_fp16 = mul(x = r_3_cast_fp16, y = cos_to_fp16)[name = string("op_3285_cast_fp16")];
            tensor<fp16, [3, 2, 64, 16]> var_3286_cast_fp16 = mul(x = rh_3_cast_fp16, y = sin_to_fp16)[name = string("op_3286_cast_fp16")];
            tensor<fp16, [3, 2, 64, 16]> var_3288_cast_fp16 = add(x = var_3285_cast_fp16, y = var_3286_cast_fp16)[name = string("op_3288_cast_fp16")];
            int32 var_3290 = const()[name = string("op_3290"), val = int32(-1)];
            bool k_49_interleave_0 = const()[name = string("k_49_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 2, 64, 64]> k_49_cast_fp16 = concat(axis = var_3290, interleave = k_49_interleave_0, values = (var_3288_cast_fp16, p_3_cast_fp16))[name = string("k_49_cast_fp16")];
            tensor<int32, [1]> var_3303_axes_0 = const()[name = string("op_3303_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 1, 64, 64]> var_3303_cast_fp16 = expand_dims(axes = var_3303_axes_0, x = k_49_cast_fp16)[name = string("op_3303_cast_fp16")];
            tensor<int32, [5]> var_3311_reps_0 = const()[name = string("op_3311_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [3, 2, 4, 64, 64]> var_3311_cast_fp16 = tile(reps = var_3311_reps_0, x = var_3303_cast_fp16)[name = string("op_3311_cast_fp16")];
            tensor<int32, [4]> var_3316 = const()[name = string("op_3316"), val = tensor<int32, [4]>([3, 8, 64, 64])];
            tensor<fp16, [3, 8, 64, 64]> k_51_cast_fp16 = reshape(shape = var_3316, x = var_3311_cast_fp16)[name = string("k_51_cast_fp16")];
            tensor<int32, [1]> var_3329_axes_0 = const()[name = string("op_3329_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 64, 64]> v_7_cast_fp16 = transpose(perm = var_3184, x = var_3179_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [3, 2, 1, 64, 64]> var_3329_cast_fp16 = expand_dims(axes = var_3329_axes_0, x = v_7_cast_fp16)[name = string("op_3329_cast_fp16")];
            tensor<int32, [5]> var_3337_reps_0 = const()[name = string("op_3337_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [3, 2, 4, 64, 64]> var_3337_cast_fp16 = tile(reps = var_3337_reps_0, x = var_3329_cast_fp16)[name = string("op_3337_cast_fp16")];
            tensor<int32, [4]> var_3342 = const()[name = string("op_3342"), val = tensor<int32, [4]>([3, 8, 64, 64])];
            tensor<fp16, [3, 8, 64, 64]> v_9_cast_fp16 = reshape(shape = var_3342, x = var_3337_cast_fp16)[name = string("v_9_cast_fp16")];
            bool var_3347_transpose_x_1 = const()[name = string("op_3347_transpose_x_1"), val = bool(false)];
            bool var_3347_transpose_y_1 = const()[name = string("op_3347_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 8, 64, 64]> var_3347_cast_fp16 = matmul(transpose_x = var_3347_transpose_x_1, transpose_y = var_3347_transpose_y_1, x = q_49_cast_fp16, y = k_51_cast_fp16)[name = string("op_3347_cast_fp16")];
            fp16 var_3348_to_fp16 = const()[name = string("op_3348_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [3, 8, 64, 64]> var_3349_cast_fp16 = mul(x = var_3347_cast_fp16, y = var_3348_to_fp16)[name = string("op_3349_cast_fp16")];
            tensor<fp16, [64, 64]> causal_add_to_fp16 = const()[name = string("causal_add_to_fp16"), val = tensor<fp16, [64, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119667520)))];
            tensor<fp16, [3, 8, 64, 64]> s_1_cast_fp16 = add(x = var_3349_cast_fp16, y = causal_add_to_fp16)[name = string("s_1_cast_fp16")];
            int32 var_3352 = const()[name = string("op_3352"), val = int32(-1)];
            tensor<fp16, [3, 8, 64, 64]> var_3354_cast_fp16 = softmax(axis = var_3352, x = s_1_cast_fp16)[name = string("op_3354_cast_fp16")];
            bool o_19_transpose_x_0 = const()[name = string("o_19_transpose_x_0"), val = bool(false)];
            bool o_19_transpose_y_0 = const()[name = string("o_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 8, 64, 64]> o_19_cast_fp16 = matmul(transpose_x = o_19_transpose_x_0, transpose_y = o_19_transpose_y_0, x = var_3354_cast_fp16, y = v_9_cast_fp16)[name = string("o_19_cast_fp16")];
            tensor<int32, [4]> var_3360 = const()[name = string("op_3360"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3365 = const()[name = string("op_3365"), val = tensor<int32, [3]>([3, 64, 512])];
            tensor<fp16, [3, 64, 8, 64]> var_3361_cast_fp16 = transpose(perm = var_3360, x = o_19_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [3, 64, 512]> var_3366_cast_fp16 = reshape(shape = var_3365, x = var_3361_cast_fp16)[name = string("op_3366_cast_fp16")];
            tensor<fp16, [3, 64, 512]> var_3367_cast_fp16 = sigmoid(x = gate_1_cast_fp16)[name = string("op_3367_cast_fp16")];
            tensor<fp16, [3, 64, 512]> input_51_cast_fp16 = mul(x = var_3366_cast_fp16, y = var_3367_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [640, 512]> layers_3_o_weight_to_fp16 = const()[name = string("layers_3_o_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59589568)))];
            tensor<fp16, [3, 64, 640]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_o_weight_to_fp16, x = input_51_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [3, 64, 640]> x_41_cast_fp16 = add(x = x_25_cast_fp16, y = linear_13_cast_fp16)[name = string("x_41_cast_fp16")];
            fp16 var_3380_to_fp16 = const()[name = string("op_3380_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 64, 640]> x_43_cast_fp16 = mul(x = x_41_cast_fp16, y = var_3380_to_fp16)[name = string("x_43_cast_fp16")];
            fp16 var_3378_promoted_to_fp16 = const()[name = string("op_3378_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> var_3382_cast_fp16 = pow(x = x_43_cast_fp16, y = var_3378_promoted_to_fp16)[name = string("op_3382_cast_fp16")];
            tensor<int32, [1]> var_3384_axes_0 = const()[name = string("op_3384_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3384_keep_dims_0 = const()[name = string("op_3384_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 1]> var_3384_cast_fp16 = reduce_mean(axes = var_3384_axes_0, keep_dims = var_3384_keep_dims_0, x = var_3382_cast_fp16)[name = string("op_3384_cast_fp16")];
            fp16 var_3385_to_fp16 = const()[name = string("op_3385_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 64, 1]> var_3386_cast_fp16 = add(x = var_3384_cast_fp16, y = var_3385_to_fp16)[name = string("op_3386_cast_fp16")];
            fp32 var_3387_epsilon_0 = const()[name = string("op_3387_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 1]> var_3387_cast_fp16 = rsqrt(epsilon = var_3387_epsilon_0, x = var_3386_cast_fp16)[name = string("op_3387_cast_fp16")];
            tensor<fp16, [3, 64, 640]> var_3388_cast_fp16 = mul(x = x_43_cast_fp16, y = var_3387_cast_fp16)[name = string("op_3388_cast_fp16")];
            tensor<fp16, [640]> layers_3_n2_w1_to_fp16 = const()[name = string("layers_3_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60244992)))];
            tensor<fp16, [3, 64, 640]> input_53_cast_fp16 = mul(x = var_3388_cast_fp16, y = layers_3_n2_w1_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_3_gate_up_weight_to_fp16 = const()[name = string("layers_3_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60246336)))];
            tensor<fp16, [3, 64, 5120]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_3_gate_up_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [3]> input_55_begin_0 = const()[name = string("input_55_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_55_end_0 = const()[name = string("input_55_end_0"), val = tensor<int32, [3]>([3, 64, 2560])];
            tensor<bool, [3]> input_55_end_mask_0 = const()[name = string("input_55_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 2560]> input_55_cast_fp16 = slice_by_index(begin = input_55_begin_0, end = input_55_end_0, end_mask = input_55_end_mask_0, x = linear_14_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [3, 64, 2560]> var_3398_cast_fp16 = silu(x = input_55_cast_fp16)[name = string("op_3398_cast_fp16")];
            tensor<int32, [3]> var_3403_begin_0 = const()[name = string("op_3403_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_3403_end_0 = const()[name = string("op_3403_end_0"), val = tensor<int32, [3]>([3, 64, 5120])];
            tensor<bool, [3]> var_3403_end_mask_0 = const()[name = string("op_3403_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 2560]> var_3403_cast_fp16 = slice_by_index(begin = var_3403_begin_0, end = var_3403_end_0, end_mask = var_3403_end_mask_0, x = linear_14_cast_fp16)[name = string("op_3403_cast_fp16")];
            tensor<fp16, [3, 64, 2560]> input_57_cast_fp16 = mul(x = var_3398_cast_fp16, y = var_3403_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_3_down_weight_to_fp16 = const()[name = string("layers_3_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66800000)))];
            tensor<fp16, [3, 64, 640]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_down_weight_to_fp16, x = input_57_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [3, 64, 640]> x_45_cast_fp16 = add(x = x_41_cast_fp16, y = linear_15_cast_fp16)[name = string("x_45_cast_fp16")];
            fp16 var_3416_to_fp16 = const()[name = string("op_3416_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 64, 640]> x_47_cast_fp16 = mul(x = x_45_cast_fp16, y = var_3416_to_fp16)[name = string("x_47_cast_fp16")];
            fp16 var_3414_promoted_to_fp16 = const()[name = string("op_3414_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> var_3418_cast_fp16 = pow(x = x_47_cast_fp16, y = var_3414_promoted_to_fp16)[name = string("op_3418_cast_fp16")];
            tensor<int32, [1]> var_3420_axes_0 = const()[name = string("op_3420_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3420_keep_dims_0 = const()[name = string("op_3420_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 1]> var_3420_cast_fp16 = reduce_mean(axes = var_3420_axes_0, keep_dims = var_3420_keep_dims_0, x = var_3418_cast_fp16)[name = string("op_3420_cast_fp16")];
            fp16 var_3421_to_fp16 = const()[name = string("op_3421_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 64, 1]> var_3422_cast_fp16 = add(x = var_3420_cast_fp16, y = var_3421_to_fp16)[name = string("op_3422_cast_fp16")];
            fp32 var_3423_epsilon_0 = const()[name = string("op_3423_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 1]> var_3423_cast_fp16 = rsqrt(epsilon = var_3423_epsilon_0, x = var_3422_cast_fp16)[name = string("op_3423_cast_fp16")];
            tensor<fp16, [3, 64, 640]> var_3424_cast_fp16 = mul(x = x_47_cast_fp16, y = var_3423_cast_fp16)[name = string("op_3424_cast_fp16")];
            tensor<fp16, [640]> layers_4_n1_w1_to_fp16 = const()[name = string("layers_4_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70076864)))];
            tensor<fp16, [3, 64, 640]> input_59_cast_fp16 = mul(x = var_3424_cast_fp16, y = layers_4_n1_w1_to_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_4_qkvzba_weight_to_fp16 = const()[name = string("layers_4_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70078208)))];
            tensor<fp16, [3, 64, 1544]> linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_qkvzba_weight_to_fp16, x = input_59_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [3]> qkv_7_begin_0 = const()[name = string("qkv_7_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_7_end_0 = const()[name = string("qkv_7_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> qkv_7_end_mask_0 = const()[name = string("qkv_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 1024]> qkv_7_cast_fp16 = slice_by_index(begin = qkv_7_begin_0, end = qkv_7_end_0, end_mask = qkv_7_end_mask_0, x = linear_16_cast_fp16)[name = string("qkv_7_cast_fp16")];
            tensor<int32, [3]> z_7_begin_0 = const()[name = string("z_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_7_end_0 = const()[name = string("z_7_end_0"), val = tensor<int32, [3]>([3, 64, 1536])];
            tensor<bool, [3]> z_7_end_mask_0 = const()[name = string("z_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 512]> z_7_cast_fp16 = slice_by_index(begin = z_7_begin_0, end = z_7_end_0, end_mask = z_7_end_mask_0, x = linear_16_cast_fp16)[name = string("z_7_cast_fp16")];
            tensor<int32, [3]> b_7_begin_0 = const()[name = string("b_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_7_end_0 = const()[name = string("b_7_end_0"), val = tensor<int32, [3]>([3, 64, 1540])];
            tensor<bool, [3]> b_7_end_mask_0 = const()[name = string("b_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 4]> b_7_cast_fp16 = slice_by_index(begin = b_7_begin_0, end = b_7_end_0, end_mask = b_7_end_mask_0, x = linear_16_cast_fp16)[name = string("b_7_cast_fp16")];
            tensor<int32, [3]> a_7_begin_0 = const()[name = string("a_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_7_end_0 = const()[name = string("a_7_end_0"), val = tensor<int32, [3]>([3, 64, 1])];
            tensor<bool, [3]> a_7_end_mask_0 = const()[name = string("a_7_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 4]> a_7_cast_fp16 = slice_by_index(begin = a_7_begin_0, end = a_7_end_0, end_mask = a_7_end_mask_0, x = linear_16_cast_fp16)[name = string("a_7_cast_fp16")];
            int32 var_3463 = const()[name = string("op_3463"), val = int32(1)];
            bool xp_7_interleave_0 = const()[name = string("xp_7_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 67, 1024]> xp_7_cast_fp16 = concat(axis = var_3463, interleave = xp_7_interleave_0, values = (zpad_to_fp16, qkv_7_cast_fp16))[name = string("xp_7_cast_fp16")];
            tensor<int32, [3]> var_3474_begin_0 = const()[name = string("op_3474_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_3474_end_0 = const()[name = string("op_3474_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_3474_end_mask_0 = const()[name = string("op_3474_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 1024]> var_3474_cast_fp16 = slice_by_index(begin = var_3474_begin_0, end = var_3474_end_0, end_mask = var_3474_end_mask_0, x = xp_7_cast_fp16)[name = string("op_3474_cast_fp16")];
            tensor<fp16, [1024]> var_3482_to_fp16 = const()[name = string("op_3482_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72054592)))];
            tensor<fp16, [3, 64, 1024]> var_3483_cast_fp16 = mul(x = var_3474_cast_fp16, y = var_3482_to_fp16)[name = string("op_3483_cast_fp16")];
            tensor<int32, [3]> var_3493_begin_0 = const()[name = string("op_3493_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_3493_end_0 = const()[name = string("op_3493_end_0"), val = tensor<int32, [3]>([3, 66, 1024])];
            tensor<bool, [3]> var_3493_end_mask_0 = const()[name = string("op_3493_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_3493_cast_fp16 = slice_by_index(begin = var_3493_begin_0, end = var_3493_end_0, end_mask = var_3493_end_mask_0, x = xp_7_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<fp16, [1024]> var_3501_to_fp16 = const()[name = string("op_3501_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72056704)))];
            tensor<fp16, [3, 64, 1024]> var_3502_cast_fp16 = mul(x = var_3493_cast_fp16, y = var_3501_to_fp16)[name = string("op_3502_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> var_3504_cast_fp16 = add(x = var_3483_cast_fp16, y = var_3502_cast_fp16)[name = string("op_3504_cast_fp16")];
            tensor<int32, [3]> var_3514_begin_0 = const()[name = string("op_3514_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_3514_end_0 = const()[name = string("op_3514_end_0"), val = tensor<int32, [3]>([3, 65, 1024])];
            tensor<bool, [3]> var_3514_end_mask_0 = const()[name = string("op_3514_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_3514_cast_fp16 = slice_by_index(begin = var_3514_begin_0, end = var_3514_end_0, end_mask = var_3514_end_mask_0, x = xp_7_cast_fp16)[name = string("op_3514_cast_fp16")];
            tensor<fp16, [1024]> var_3522_to_fp16 = const()[name = string("op_3522_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72058816)))];
            tensor<fp16, [3, 64, 1024]> var_3523_cast_fp16 = mul(x = var_3514_cast_fp16, y = var_3522_to_fp16)[name = string("op_3523_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> var_3525_cast_fp16 = add(x = var_3504_cast_fp16, y = var_3523_cast_fp16)[name = string("op_3525_cast_fp16")];
            tensor<int32, [3]> var_3535_begin_0 = const()[name = string("op_3535_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3535_end_0 = const()[name = string("op_3535_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> var_3535_end_mask_0 = const()[name = string("op_3535_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_3535_cast_fp16 = slice_by_index(begin = var_3535_begin_0, end = var_3535_end_0, end_mask = var_3535_end_mask_0, x = xp_7_cast_fp16)[name = string("op_3535_cast_fp16")];
            tensor<fp16, [1024]> var_3543_to_fp16 = const()[name = string("op_3543_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72060928)))];
            tensor<fp16, [3, 64, 1024]> var_3544_cast_fp16 = mul(x = var_3535_cast_fp16, y = var_3543_to_fp16)[name = string("op_3544_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> input_61_cast_fp16 = add(x = var_3525_cast_fp16, y = var_3544_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> c_7_cast_fp16 = silu(x = input_61_cast_fp16)[name = string("c_7_cast_fp16")];
            tensor<int32, [3]> var_3552_begin_0 = const()[name = string("op_3552_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3552_end_0 = const()[name = string("op_3552_end_0"), val = tensor<int32, [3]>([3, 64, 256])];
            tensor<bool, [3]> var_3552_end_mask_0 = const()[name = string("op_3552_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 256]> var_3552_cast_fp16 = slice_by_index(begin = var_3552_begin_0, end = var_3552_end_0, end_mask = var_3552_end_mask_0, x = c_7_cast_fp16)[name = string("op_3552_cast_fp16")];
            tensor<int32, [4]> var_3557 = const()[name = string("op_3557"), val = tensor<int32, [4]>([3, 64, 2, 128])];
            tensor<fp16, [3, 64, 2, 128]> q_51_cast_fp16 = reshape(shape = var_3557, x = var_3552_cast_fp16)[name = string("q_51_cast_fp16")];
            tensor<int32, [3]> var_3563_begin_0 = const()[name = string("op_3563_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_3563_end_0 = const()[name = string("op_3563_end_0"), val = tensor<int32, [3]>([3, 64, 512])];
            tensor<bool, [3]> var_3563_end_mask_0 = const()[name = string("op_3563_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 256]> var_3563_cast_fp16 = slice_by_index(begin = var_3563_begin_0, end = var_3563_end_0, end_mask = var_3563_end_mask_0, x = c_7_cast_fp16)[name = string("op_3563_cast_fp16")];
            tensor<int32, [4]> var_3568 = const()[name = string("op_3568"), val = tensor<int32, [4]>([3, 64, 2, 128])];
            tensor<fp16, [3, 64, 2, 128]> k_53_cast_fp16 = reshape(shape = var_3568, x = var_3563_cast_fp16)[name = string("k_53_cast_fp16")];
            tensor<int32, [3]> var_3574_begin_0 = const()[name = string("op_3574_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_3574_end_0 = const()[name = string("op_3574_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> var_3574_end_mask_0 = const()[name = string("op_3574_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 512]> var_3574_cast_fp16 = slice_by_index(begin = var_3574_begin_0, end = var_3574_end_0, end_mask = var_3574_end_mask_0, x = c_7_cast_fp16)[name = string("op_3574_cast_fp16")];
            tensor<int32, [4]> var_3579 = const()[name = string("op_3579"), val = tensor<int32, [4]>([3, 64, 4, 128])];
            tensor<fp16, [3, 64, 4, 128]> var_3580_cast_fp16 = reshape(shape = var_3579, x = var_3574_cast_fp16)[name = string("op_3580_cast_fp16")];
            tensor<int32, [4]> var_3585 = const()[name = string("op_3585"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3587_to_fp16 = const()[name = string("op_3587_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 64, 2, 128]> q_53_cast_fp16 = mul(x = q_51_cast_fp16, y = var_3587_to_fp16)[name = string("q_53_cast_fp16")];
            fp16 var_3589_to_fp16 = const()[name = string("op_3589_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 64, 2, 128]> k_55_cast_fp16 = mul(x = k_53_cast_fp16, y = var_3589_to_fp16)[name = string("k_55_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_3591_cast_fp16 = mul(x = q_53_cast_fp16, y = q_53_cast_fp16)[name = string("op_3591_cast_fp16")];
            tensor<int32, [1]> var_3596_axes_0 = const()[name = string("op_3596_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3596_keep_dims_0 = const()[name = string("op_3596_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 2, 1]> var_3596_cast_fp16 = reduce_mean(axes = var_3596_axes_0, keep_dims = var_3596_keep_dims_0, x = var_3591_cast_fp16)[name = string("op_3596_cast_fp16")];
            fp16 var_3597_promoted_to_fp16 = const()[name = string("op_3597_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 64, 2, 1]> var_3598_cast_fp16 = mul(x = var_3596_cast_fp16, y = var_3597_promoted_to_fp16)[name = string("op_3598_cast_fp16")];
            fp16 var_3600_to_fp16 = const()[name = string("op_3600_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 64, 2, 1]> var_3601_cast_fp16 = add(x = var_3598_cast_fp16, y = var_3600_to_fp16)[name = string("op_3601_cast_fp16")];
            fp32 var_3602_epsilon_0 = const()[name = string("op_3602_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 2, 1]> var_3602_cast_fp16 = rsqrt(epsilon = var_3602_epsilon_0, x = var_3601_cast_fp16)[name = string("op_3602_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_3603_cast_fp16 = mul(x = q_53_cast_fp16, y = var_3602_cast_fp16)[name = string("op_3603_cast_fp16")];
            fp16 var_3604_to_fp16 = const()[name = string("op_3604_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 64, 2, 128]> q_55_cast_fp16 = mul(x = var_3603_cast_fp16, y = var_3604_to_fp16)[name = string("q_55_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_3606_cast_fp16 = mul(x = k_55_cast_fp16, y = k_55_cast_fp16)[name = string("op_3606_cast_fp16")];
            tensor<int32, [1]> var_3611_axes_0 = const()[name = string("op_3611_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3611_keep_dims_0 = const()[name = string("op_3611_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 2, 1]> var_3611_cast_fp16 = reduce_mean(axes = var_3611_axes_0, keep_dims = var_3611_keep_dims_0, x = var_3606_cast_fp16)[name = string("op_3611_cast_fp16")];
            fp16 var_3612_promoted_to_fp16 = const()[name = string("op_3612_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 64, 2, 1]> var_3613_cast_fp16 = mul(x = var_3611_cast_fp16, y = var_3612_promoted_to_fp16)[name = string("op_3613_cast_fp16")];
            fp16 var_3615_to_fp16 = const()[name = string("op_3615_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 64, 2, 1]> var_3616_cast_fp16 = add(x = var_3613_cast_fp16, y = var_3615_to_fp16)[name = string("op_3616_cast_fp16")];
            fp32 var_3617_epsilon_0 = const()[name = string("op_3617_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 2, 1]> var_3617_cast_fp16 = rsqrt(epsilon = var_3617_epsilon_0, x = var_3616_cast_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> k_57_cast_fp16 = mul(x = k_55_cast_fp16, y = var_3617_cast_fp16)[name = string("k_57_cast_fp16")];
            tensor<int32, [4]> var_3623 = const()[name = string("op_3623"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_3636_axes_0 = const()[name = string("op_3636_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 64, 128]> var_3624_cast_fp16 = transpose(perm = var_3623, x = q_55_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [3, 2, 1, 64, 128]> var_3636_cast_fp16 = expand_dims(axes = var_3636_axes_0, x = var_3624_cast_fp16)[name = string("op_3636_cast_fp16")];
            tensor<int32, [5]> var_3644_reps_0 = const()[name = string("op_3644_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 64, 128]> var_3644_cast_fp16 = tile(reps = var_3644_reps_0, x = var_3636_cast_fp16)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> var_3649 = const()[name = string("op_3649"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<fp16, [3, 4, 64, 128]> q_57_cast_fp16 = reshape(shape = var_3649, x = var_3644_cast_fp16)[name = string("q_57_cast_fp16")];
            tensor<int32, [4]> var_3655 = const()[name = string("op_3655"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_3668_axes_0 = const()[name = string("op_3668_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 64, 128]> var_3656_cast_fp16 = transpose(perm = var_3655, x = k_57_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [3, 2, 1, 64, 128]> var_3668_cast_fp16 = expand_dims(axes = var_3668_axes_0, x = var_3656_cast_fp16)[name = string("op_3668_cast_fp16")];
            tensor<int32, [5]> var_3676_reps_0 = const()[name = string("op_3676_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 64, 128]> var_3676_cast_fp16 = tile(reps = var_3676_reps_0, x = var_3668_cast_fp16)[name = string("op_3676_cast_fp16")];
            tensor<int32, [4]> var_3681 = const()[name = string("op_3681"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<fp16, [3, 4, 64, 128]> k_59_cast_fp16 = reshape(shape = var_3681, x = var_3676_cast_fp16)[name = string("k_59_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_3683_cast_fp16 = sigmoid(x = b_7_cast_fp16)[name = string("op_3683_cast_fp16")];
            tensor<int32, [3]> var_3687 = const()[name = string("op_3687"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_4_dt_bias_to_fp16 = const()[name = string("layers_4_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 64, 4]> input_63_cast_fp16 = add(x = a_7_cast_fp16, y = layers_4_dt_bias_to_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_3691_cast_fp16 = relu(x = input_63_cast_fp16)[name = string("op_3691_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_3692_cast_fp16 = abs(x = input_63_cast_fp16)[name = string("op_3692_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 64, 4]> var_3693_cast_fp16 = mul(x = var_3692_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_3693_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_3694_cast_fp16 = exp(x = var_3693_cast_fp16)[name = string("op_3694_cast_fp16")];
            fp16 var_3696_to_fp16 = const()[name = string("op_3696_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 64, 4]> var_3697_cast_fp16 = add(x = var_3694_cast_fp16, y = var_3696_to_fp16)[name = string("op_3697_cast_fp16")];
            fp32 var_3698_epsilon_0 = const()[name = string("op_3698_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 64, 4]> var_3698_cast_fp16 = log(epsilon = var_3698_epsilon_0, x = var_3697_cast_fp16)[name = string("op_3698_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_3700_cast_fp16 = add(x = var_3691_cast_fp16, y = var_3698_cast_fp16)[name = string("op_3700_cast_fp16")];
            tensor<fp16, [4]> layers_4_negA_to_fp16 = const()[name = string("layers_4_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.348p+3, -0x1.734p+2, -0x1.5a8p-1, -0x1.264p+3])];
            tensor<fp16, [3, 64, 4]> var_3701_cast_fp16 = mul(x = layers_4_negA_to_fp16, y = var_3700_cast_fp16)[name = string("op_3701_cast_fp16")];
            tensor<int32, [3]> var_3705 = const()[name = string("op_3705"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_3708_axes_0 = const()[name = string("op_3708_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 64]> beta_7_cast_fp16 = transpose(perm = var_3687, x = var_3683_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [3, 4, 64, 1]> var_3708_cast_fp16 = expand_dims(axes = var_3708_axes_0, x = beta_7_cast_fp16)[name = string("op_3708_cast_fp16")];
            fp16 var_3709_to_fp16 = const()[name = string("op_3709_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 64, 1]> var_3710_cast_fp16 = mul(x = var_3708_cast_fp16, y = var_3709_to_fp16)[name = string("op_3710_cast_fp16")];
            tensor<fp16, [3, 4, 64, 128]> v_11_cast_fp16 = transpose(perm = var_3585, x = var_3580_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [3, 4, 64, 128]> vb_31_cast_fp16 = mul(x = v_11_cast_fp16, y = var_3710_cast_fp16)[name = string("vb_31_cast_fp16")];
            tensor<int32, [4]> q_59_begin_0 = const()[name = string("q_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_59_end_0 = const()[name = string("q_59_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_59_end_mask_0 = const()[name = string("q_59_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_59_cast_fp16 = slice_by_index(begin = q_59_begin_0, end = q_59_end_0, end_mask = q_59_end_mask_0, x = q_57_cast_fp16)[name = string("q_59_cast_fp16")];
            tensor<int32, [4]> k_61_begin_0 = const()[name = string("k_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_61_end_0 = const()[name = string("k_61_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_61_end_mask_0 = const()[name = string("k_61_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, x = k_59_cast_fp16)[name = string("k_61_cast_fp16")];
            tensor<fp16, [3, 4, 64, 128]> var_3744_cast_fp16 = mul(x = k_59_cast_fp16, y = var_3708_cast_fp16)[name = string("op_3744_cast_fp16")];
            tensor<int32, [4]> kb_25_begin_0 = const()[name = string("kb_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_25_end_0 = const()[name = string("kb_25_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_25_end_mask_0 = const()[name = string("kb_25_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_25_cast_fp16 = slice_by_index(begin = kb_25_begin_0, end = kb_25_end_0, end_mask = kb_25_end_mask_0, x = var_3744_cast_fp16)[name = string("kb_25_cast_fp16")];
            tensor<int32, [4]> vb_33_begin_0 = const()[name = string("vb_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_33_end_0 = const()[name = string("vb_33_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_33_end_mask_0 = const()[name = string("vb_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_33_cast_fp16 = slice_by_index(begin = vb_33_begin_0, end = vb_33_end_0, end_mask = vb_33_end_mask_0, x = vb_31_cast_fp16)[name = string("vb_33_cast_fp16")];
            tensor<int32, [3]> gdec_33_begin_0 = const()[name = string("gdec_33_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_33_end_0 = const()[name = string("gdec_33_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_33_end_mask_0 = const()[name = string("gdec_33_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 64]> gdec_31_cast_fp16 = transpose(perm = var_3705, x = var_3701_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [3, 4, 16]> gdec_33_cast_fp16 = slice_by_index(begin = gdec_33_begin_0, end = gdec_33_end_0, end_mask = gdec_33_end_mask_0, x = gdec_31_cast_fp16)[name = string("gdec_33_cast_fp16")];
            tensor<int32, [2]> var_3792 = const()[name = string("op_3792"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_3793_cast_fp16 = reshape(shape = var_3792, x = gdec_33_cast_fp16)[name = string("op_3793_cast_fp16")];
            tensor<fp16, [12, 256]> var_3794_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3793_cast_fp16)[name = string("op_3794_cast_fp16")];
            tensor<int32, [4]> var_3799 = const()[name = string("op_3799"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_25_cast_fp16 = reshape(shape = var_3799, x = var_3794_cast_fp16)[name = string("P_25_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3801_cast_fp16 = exp(x = P_25_cast_fp16)[name = string("op_3801_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_25_cast_fp16 = mul(x = var_3801_cast_fp16, y = tril_incl_to_fp16)[name = string("D_25_cast_fp16")];
            bool kk_25_transpose_x_1 = const()[name = string("kk_25_transpose_x_1"), val = bool(false)];
            bool kk_25_transpose_y_1 = const()[name = string("kk_25_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_25_cast_fp16 = matmul(transpose_x = kk_25_transpose_x_1, transpose_y = kk_25_transpose_y_1, x = kb_25_cast_fp16, y = k_61_cast_fp16)[name = string("kk_25_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_3807_cast_fp16 = mul(x = kk_25_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_3807_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3808_cast_fp16 = mul(x = var_3807_cast_fp16, y = D_25_cast_fp16)[name = string("op_3808_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_25_cast_fp16 = mul(x = var_3808_cast_fp16, y = cum_tail_to_fp16)[name = string("N_25_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3810_cast_fp16 = mul(x = D_25_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3810_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_25_cast_fp16 = mul(x = kk_25_cast_fp16, y = var_3810_cast_fp16)[name = string("Na_25_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_97_cast_fp16 = add(x = eye_to_fp16, y = N_25_cast_fp16)[name = string("Minv_97_cast_fp16")];
            bool Np_73_transpose_x_0 = const()[name = string("Np_73_transpose_x_0"), val = bool(false)];
            bool Np_73_transpose_y_0 = const()[name = string("Np_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_73_cast_fp16 = matmul(transpose_x = Np_73_transpose_x_0, transpose_y = Np_73_transpose_y_0, x = N_25_cast_fp16, y = Na_25_cast_fp16)[name = string("Np_73_cast_fp16")];
            bool Npa_49_transpose_x_0 = const()[name = string("Npa_49_transpose_x_0"), val = bool(false)];
            bool Npa_49_transpose_y_0 = const()[name = string("Npa_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_49_cast_fp16 = matmul(transpose_x = Npa_49_transpose_x_0, transpose_y = Npa_49_transpose_y_0, x = Na_25_cast_fp16, y = N_25_cast_fp16)[name = string("Npa_49_cast_fp16")];
            bool var_3816_transpose_x_0 = const()[name = string("op_3816_transpose_x_0"), val = bool(false)];
            bool var_3816_transpose_y_0 = const()[name = string("op_3816_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3816_cast_fp16 = matmul(transpose_x = var_3816_transpose_x_0, transpose_y = var_3816_transpose_y_0, x = Minv_97_cast_fp16, y = Np_73_cast_fp16)[name = string("op_3816_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_99_cast_fp16 = add(x = Minv_97_cast_fp16, y = var_3816_cast_fp16)[name = string("Minv_99_cast_fp16")];
            bool Np_75_transpose_x_0 = const()[name = string("Np_75_transpose_x_0"), val = bool(false)];
            bool Np_75_transpose_y_0 = const()[name = string("Np_75_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_75_cast_fp16 = matmul(transpose_x = Np_75_transpose_x_0, transpose_y = Np_75_transpose_y_0, x = Np_73_cast_fp16, y = Npa_49_cast_fp16)[name = string("Np_75_cast_fp16")];
            bool Npa_51_transpose_x_0 = const()[name = string("Npa_51_transpose_x_0"), val = bool(false)];
            bool Npa_51_transpose_y_0 = const()[name = string("Npa_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_51_cast_fp16 = matmul(transpose_x = Npa_51_transpose_x_0, transpose_y = Npa_51_transpose_y_0, x = Npa_49_cast_fp16, y = Np_73_cast_fp16)[name = string("Npa_51_cast_fp16")];
            bool var_3821_transpose_x_0 = const()[name = string("op_3821_transpose_x_0"), val = bool(false)];
            bool var_3821_transpose_y_0 = const()[name = string("op_3821_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3821_cast_fp16 = matmul(transpose_x = var_3821_transpose_x_0, transpose_y = var_3821_transpose_y_0, x = Minv_99_cast_fp16, y = Np_75_cast_fp16)[name = string("op_3821_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_101_cast_fp16 = add(x = Minv_99_cast_fp16, y = var_3821_cast_fp16)[name = string("Minv_101_cast_fp16")];
            bool Np_77_transpose_x_0 = const()[name = string("Np_77_transpose_x_0"), val = bool(false)];
            bool Np_77_transpose_y_0 = const()[name = string("Np_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_77_cast_fp16 = matmul(transpose_x = Np_77_transpose_x_0, transpose_y = Np_77_transpose_y_0, x = Np_75_cast_fp16, y = Npa_51_cast_fp16)[name = string("Np_77_cast_fp16")];
            bool var_3825_transpose_x_0 = const()[name = string("op_3825_transpose_x_0"), val = bool(false)];
            bool var_3825_transpose_y_0 = const()[name = string("op_3825_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3825_cast_fp16 = matmul(transpose_x = var_3825_transpose_x_0, transpose_y = var_3825_transpose_y_0, x = Minv_101_cast_fp16, y = Np_77_cast_fp16)[name = string("op_3825_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_103_cast_fp16 = add(x = Minv_101_cast_fp16, y = var_3825_cast_fp16)[name = string("Minv_103_cast_fp16")];
            bool u_25_transpose_x_0 = const()[name = string("u_25_transpose_x_0"), val = bool(false)];
            bool u_25_transpose_y_0 = const()[name = string("u_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_25_cast_fp16 = matmul(transpose_x = u_25_transpose_x_0, transpose_y = u_25_transpose_y_0, x = Minv_103_cast_fp16, y = vb_33_cast_fp16)[name = string("u_25_cast_fp16")];
            bool var_3832_transpose_x_1 = const()[name = string("op_3832_transpose_x_1"), val = bool(false)];
            bool var_3832_transpose_y_1 = const()[name = string("op_3832_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_3832_cast_fp16 = matmul(transpose_x = var_3832_transpose_x_1, transpose_y = var_3832_transpose_y_1, x = q_59_cast_fp16, y = k_61_cast_fp16)[name = string("op_3832_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_25_cast_fp16 = mul(x = var_3832_cast_fp16, y = D_25_cast_fp16)[name = string("A_25_cast_fp16")];
            bool var_3834_transpose_x_0 = const()[name = string("op_3834_transpose_x_0"), val = bool(false)];
            bool var_3834_transpose_y_0 = const()[name = string("op_3834_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3834_cast_fp16 = matmul(transpose_x = var_3834_transpose_x_0, transpose_y = var_3834_transpose_y_0, x = A_25_cast_fp16, y = u_25_cast_fp16)[name = string("op_3834_cast_fp16")];
            tensor<fp16, [16]> var_3835_bias_0_to_fp16 = const()[name = string("op_3835_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_3835_cast_fp16 = linear(bias = var_3835_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_33_cast_fp16)[name = string("op_3835_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_19_cast_fp16 = exp(x = var_3835_cast_fp16)[name = string("tail_19_cast_fp16")];
            tensor<int32, [1]> var_3838_axes_0 = const()[name = string("op_3838_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_3838_cast_fp16 = expand_dims(axes = var_3838_axes_0, x = tail_19_cast_fp16)[name = string("op_3838_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3839_cast_fp16 = mul(x = k_61_cast_fp16, y = var_3838_cast_fp16)[name = string("op_3839_cast_fp16")];
            bool S_19_transpose_x_1 = const()[name = string("S_19_transpose_x_1"), val = bool(true)];
            bool S_19_transpose_y_1 = const()[name = string("S_19_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_19_cast_fp16 = matmul(transpose_x = S_19_transpose_x_1, transpose_y = S_19_transpose_y_1, x = var_3839_cast_fp16, y = u_25_cast_fp16)[name = string("S_19_cast_fp16")];
            tensor<int32, [4]> q_61_begin_0 = const()[name = string("q_61_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_61_end_0 = const()[name = string("q_61_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> q_61_end_mask_0 = const()[name = string("q_61_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, x = q_57_cast_fp16)[name = string("q_61_cast_fp16")];
            tensor<int32, [4]> k_63_begin_0 = const()[name = string("k_63_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_63_end_0 = const()[name = string("k_63_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> k_63_end_mask_0 = const()[name = string("k_63_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_63_cast_fp16 = slice_by_index(begin = k_63_begin_0, end = k_63_end_0, end_mask = k_63_end_mask_0, x = k_59_cast_fp16)[name = string("k_63_cast_fp16")];
            tensor<int32, [4]> kb_27_begin_0 = const()[name = string("kb_27_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_27_end_0 = const()[name = string("kb_27_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> kb_27_end_mask_0 = const()[name = string("kb_27_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_27_cast_fp16 = slice_by_index(begin = kb_27_begin_0, end = kb_27_end_0, end_mask = kb_27_end_mask_0, x = var_3744_cast_fp16)[name = string("kb_27_cast_fp16")];
            tensor<int32, [4]> vb_35_begin_0 = const()[name = string("vb_35_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_35_end_0 = const()[name = string("vb_35_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> vb_35_end_mask_0 = const()[name = string("vb_35_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_35_cast_fp16 = slice_by_index(begin = vb_35_begin_0, end = vb_35_end_0, end_mask = vb_35_end_mask_0, x = vb_31_cast_fp16)[name = string("vb_35_cast_fp16")];
            tensor<int32, [3]> gdec_35_begin_0 = const()[name = string("gdec_35_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_35_end_0 = const()[name = string("gdec_35_end_0"), val = tensor<int32, [3]>([3, 4, 32])];
            tensor<bool, [3]> gdec_35_end_mask_0 = const()[name = string("gdec_35_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_35_cast_fp16 = slice_by_index(begin = gdec_35_begin_0, end = gdec_35_end_0, end_mask = gdec_35_end_mask_0, x = gdec_31_cast_fp16)[name = string("gdec_35_cast_fp16")];
            tensor<int32, [2]> var_3924 = const()[name = string("op_3924"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_3925_cast_fp16 = reshape(shape = var_3924, x = gdec_35_cast_fp16)[name = string("op_3925_cast_fp16")];
            tensor<fp16, [12, 256]> var_3926_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3925_cast_fp16)[name = string("op_3926_cast_fp16")];
            tensor<int32, [4]> var_3931 = const()[name = string("op_3931"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_27_cast_fp16 = reshape(shape = var_3931, x = var_3926_cast_fp16)[name = string("P_27_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3933_cast_fp16 = exp(x = P_27_cast_fp16)[name = string("op_3933_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_27_cast_fp16 = mul(x = var_3933_cast_fp16, y = tril_incl_to_fp16)[name = string("D_27_cast_fp16")];
            bool kk_27_transpose_x_1 = const()[name = string("kk_27_transpose_x_1"), val = bool(false)];
            bool kk_27_transpose_y_1 = const()[name = string("kk_27_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_27_cast_fp16 = matmul(transpose_x = kk_27_transpose_x_1, transpose_y = kk_27_transpose_y_1, x = kb_27_cast_fp16, y = k_63_cast_fp16)[name = string("kk_27_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_3939_cast_fp16 = mul(x = kk_27_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_3939_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3940_cast_fp16 = mul(x = var_3939_cast_fp16, y = D_27_cast_fp16)[name = string("op_3940_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_27_cast_fp16 = mul(x = var_3940_cast_fp16, y = cum_tail_to_fp16)[name = string("N_27_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3942_cast_fp16 = mul(x = D_27_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3942_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_27_cast_fp16 = mul(x = kk_27_cast_fp16, y = var_3942_cast_fp16)[name = string("Na_27_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_105_cast_fp16 = add(x = eye_to_fp16, y = N_27_cast_fp16)[name = string("Minv_105_cast_fp16")];
            bool Np_79_transpose_x_0 = const()[name = string("Np_79_transpose_x_0"), val = bool(false)];
            bool Np_79_transpose_y_0 = const()[name = string("Np_79_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_79_cast_fp16 = matmul(transpose_x = Np_79_transpose_x_0, transpose_y = Np_79_transpose_y_0, x = N_27_cast_fp16, y = Na_27_cast_fp16)[name = string("Np_79_cast_fp16")];
            bool Npa_53_transpose_x_0 = const()[name = string("Npa_53_transpose_x_0"), val = bool(false)];
            bool Npa_53_transpose_y_0 = const()[name = string("Npa_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_53_cast_fp16 = matmul(transpose_x = Npa_53_transpose_x_0, transpose_y = Npa_53_transpose_y_0, x = Na_27_cast_fp16, y = N_27_cast_fp16)[name = string("Npa_53_cast_fp16")];
            bool var_3948_transpose_x_0 = const()[name = string("op_3948_transpose_x_0"), val = bool(false)];
            bool var_3948_transpose_y_0 = const()[name = string("op_3948_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3948_cast_fp16 = matmul(transpose_x = var_3948_transpose_x_0, transpose_y = var_3948_transpose_y_0, x = Minv_105_cast_fp16, y = Np_79_cast_fp16)[name = string("op_3948_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_107_cast_fp16 = add(x = Minv_105_cast_fp16, y = var_3948_cast_fp16)[name = string("Minv_107_cast_fp16")];
            bool Np_81_transpose_x_0 = const()[name = string("Np_81_transpose_x_0"), val = bool(false)];
            bool Np_81_transpose_y_0 = const()[name = string("Np_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_81_cast_fp16 = matmul(transpose_x = Np_81_transpose_x_0, transpose_y = Np_81_transpose_y_0, x = Np_79_cast_fp16, y = Npa_53_cast_fp16)[name = string("Np_81_cast_fp16")];
            bool Npa_55_transpose_x_0 = const()[name = string("Npa_55_transpose_x_0"), val = bool(false)];
            bool Npa_55_transpose_y_0 = const()[name = string("Npa_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_55_cast_fp16 = matmul(transpose_x = Npa_55_transpose_x_0, transpose_y = Npa_55_transpose_y_0, x = Npa_53_cast_fp16, y = Np_79_cast_fp16)[name = string("Npa_55_cast_fp16")];
            bool var_3953_transpose_x_0 = const()[name = string("op_3953_transpose_x_0"), val = bool(false)];
            bool var_3953_transpose_y_0 = const()[name = string("op_3953_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3953_cast_fp16 = matmul(transpose_x = var_3953_transpose_x_0, transpose_y = var_3953_transpose_y_0, x = Minv_107_cast_fp16, y = Np_81_cast_fp16)[name = string("op_3953_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_109_cast_fp16 = add(x = Minv_107_cast_fp16, y = var_3953_cast_fp16)[name = string("Minv_109_cast_fp16")];
            bool Np_83_transpose_x_0 = const()[name = string("Np_83_transpose_x_0"), val = bool(false)];
            bool Np_83_transpose_y_0 = const()[name = string("Np_83_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_83_cast_fp16 = matmul(transpose_x = Np_83_transpose_x_0, transpose_y = Np_83_transpose_y_0, x = Np_81_cast_fp16, y = Npa_55_cast_fp16)[name = string("Np_83_cast_fp16")];
            bool var_3957_transpose_x_0 = const()[name = string("op_3957_transpose_x_0"), val = bool(false)];
            bool var_3957_transpose_y_0 = const()[name = string("op_3957_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3957_cast_fp16 = matmul(transpose_x = var_3957_transpose_x_0, transpose_y = var_3957_transpose_y_0, x = Minv_109_cast_fp16, y = Np_83_cast_fp16)[name = string("op_3957_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_111_cast_fp16 = add(x = Minv_109_cast_fp16, y = var_3957_cast_fp16)[name = string("Minv_111_cast_fp16")];
            bool u_27_transpose_x_0 = const()[name = string("u_27_transpose_x_0"), val = bool(false)];
            bool u_27_transpose_y_0 = const()[name = string("u_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_27_cast_fp16 = matmul(transpose_x = u_27_transpose_x_0, transpose_y = u_27_transpose_y_0, x = Minv_111_cast_fp16, y = vb_35_cast_fp16)[name = string("u_27_cast_fp16")];
            bool var_3964_transpose_x_1 = const()[name = string("op_3964_transpose_x_1"), val = bool(false)];
            bool var_3964_transpose_y_1 = const()[name = string("op_3964_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_3964_cast_fp16 = matmul(transpose_x = var_3964_transpose_x_1, transpose_y = var_3964_transpose_y_1, x = q_61_cast_fp16, y = k_63_cast_fp16)[name = string("op_3964_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_27_cast_fp16 = mul(x = var_3964_cast_fp16, y = D_27_cast_fp16)[name = string("A_27_cast_fp16")];
            tensor<fp16, [16]> var_3966_bias_0_to_fp16 = const()[name = string("op_3966_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_3966_cast_fp16 = linear(bias = var_3966_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_35_cast_fp16)[name = string("op_3966_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_19_cast_fp16 = exp(x = var_3966_cast_fp16)[name = string("cum_19_cast_fp16")];
            tensor<int32, [1]> var_3969_axes_0 = const()[name = string("op_3969_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_3969_cast_fp16 = expand_dims(axes = var_3969_axes_0, x = cum_19_cast_fp16)[name = string("op_3969_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3970_cast_fp16 = mul(x = kb_27_cast_fp16, y = var_3969_cast_fp16)[name = string("op_3970_cast_fp16")];
            bool w_27_transpose_x_0 = const()[name = string("w_27_transpose_x_0"), val = bool(false)];
            bool w_27_transpose_y_0 = const()[name = string("w_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_27_cast_fp16 = matmul(transpose_x = w_27_transpose_x_0, transpose_y = w_27_transpose_y_0, x = Minv_111_cast_fp16, y = var_3970_cast_fp16)[name = string("w_27_cast_fp16")];
            bool var_3972_transpose_x_0 = const()[name = string("op_3972_transpose_x_0"), val = bool(false)];
            bool var_3972_transpose_y_0 = const()[name = string("op_3972_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3972_cast_fp16 = matmul(transpose_x = var_3972_transpose_x_0, transpose_y = var_3972_transpose_y_0, x = w_27_cast_fp16, y = S_19_cast_fp16)[name = string("op_3972_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_19_cast_fp16 = sub(x = u_27_cast_fp16, y = var_3972_cast_fp16)[name = string("vnew_19_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3977_cast_fp16 = mul(x = q_61_cast_fp16, y = var_3969_cast_fp16)[name = string("op_3977_cast_fp16")];
            bool var_3978_transpose_x_0 = const()[name = string("op_3978_transpose_x_0"), val = bool(false)];
            bool var_3978_transpose_y_0 = const()[name = string("op_3978_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3978_cast_fp16 = matmul(transpose_x = var_3978_transpose_x_0, transpose_y = var_3978_transpose_y_0, x = var_3977_cast_fp16, y = S_19_cast_fp16)[name = string("op_3978_cast_fp16")];
            bool var_3979_transpose_x_0 = const()[name = string("op_3979_transpose_x_0"), val = bool(false)];
            bool var_3979_transpose_y_0 = const()[name = string("op_3979_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3979_cast_fp16 = matmul(transpose_x = var_3979_transpose_x_0, transpose_y = var_3979_transpose_y_0, x = A_27_cast_fp16, y = vnew_19_cast_fp16)[name = string("op_3979_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3981_cast_fp16 = add(x = var_3978_cast_fp16, y = var_3979_cast_fp16)[name = string("op_3981_cast_fp16")];
            tensor<fp16, [16]> var_3982_bias_0_to_fp16 = const()[name = string("op_3982_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_3982_cast_fp16 = linear(bias = var_3982_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_35_cast_fp16)[name = string("op_3982_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_21_cast_fp16 = exp(x = var_3982_cast_fp16)[name = string("tail_21_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_41_to_fp16 = const()[name = string("transpose_41_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_3984_bias_0_to_fp16 = const()[name = string("op_3984_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_3984_cast_fp16 = linear(bias = var_3984_bias_0_to_fp16, weight = transpose_41_to_fp16, x = gdec_35_cast_fp16)[name = string("op_3984_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_3985_cast_fp16 = exp(x = var_3984_cast_fp16)[name = string("op_3985_cast_fp16")];
            tensor<int32, [1]> last_13_axes_0 = const()[name = string("last_13_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_13_cast_fp16 = expand_dims(axes = last_13_axes_0, x = var_3985_cast_fp16)[name = string("last_13_cast_fp16")];
            tensor<int32, [1]> var_3989_axes_0 = const()[name = string("op_3989_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_3989_cast_fp16 = expand_dims(axes = var_3989_axes_0, x = tail_21_cast_fp16)[name = string("op_3989_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3990_cast_fp16 = mul(x = k_63_cast_fp16, y = var_3989_cast_fp16)[name = string("op_3990_cast_fp16")];
            bool upd_13_transpose_x_1 = const()[name = string("upd_13_transpose_x_1"), val = bool(true)];
            bool upd_13_transpose_y_1 = const()[name = string("upd_13_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_13_cast_fp16 = matmul(transpose_x = upd_13_transpose_x_1, transpose_y = upd_13_transpose_y_1, x = var_3990_cast_fp16, y = vnew_19_cast_fp16)[name = string("upd_13_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_3995_cast_fp16 = mul(x = S_19_cast_fp16, y = last_13_cast_fp16)[name = string("op_3995_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_21_cast_fp16 = add(x = var_3995_cast_fp16, y = upd_13_cast_fp16)[name = string("S_21_cast_fp16")];
            tensor<int32, [4]> q_63_begin_0 = const()[name = string("q_63_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_63_end_0 = const()[name = string("q_63_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> q_63_end_mask_0 = const()[name = string("q_63_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_63_cast_fp16 = slice_by_index(begin = q_63_begin_0, end = q_63_end_0, end_mask = q_63_end_mask_0, x = q_57_cast_fp16)[name = string("q_63_cast_fp16")];
            tensor<int32, [4]> k_65_begin_0 = const()[name = string("k_65_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_65_end_0 = const()[name = string("k_65_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> k_65_end_mask_0 = const()[name = string("k_65_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_65_cast_fp16 = slice_by_index(begin = k_65_begin_0, end = k_65_end_0, end_mask = k_65_end_mask_0, x = k_59_cast_fp16)[name = string("k_65_cast_fp16")];
            tensor<int32, [4]> kb_29_begin_0 = const()[name = string("kb_29_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_29_end_0 = const()[name = string("kb_29_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> kb_29_end_mask_0 = const()[name = string("kb_29_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_29_cast_fp16 = slice_by_index(begin = kb_29_begin_0, end = kb_29_end_0, end_mask = kb_29_end_mask_0, x = var_3744_cast_fp16)[name = string("kb_29_cast_fp16")];
            tensor<int32, [4]> vb_37_begin_0 = const()[name = string("vb_37_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_37_end_0 = const()[name = string("vb_37_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> vb_37_end_mask_0 = const()[name = string("vb_37_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_37_cast_fp16 = slice_by_index(begin = vb_37_begin_0, end = vb_37_end_0, end_mask = vb_37_end_mask_0, x = vb_31_cast_fp16)[name = string("vb_37_cast_fp16")];
            tensor<int32, [3]> gdec_37_begin_0 = const()[name = string("gdec_37_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_37_end_0 = const()[name = string("gdec_37_end_0"), val = tensor<int32, [3]>([3, 4, 48])];
            tensor<bool, [3]> gdec_37_end_mask_0 = const()[name = string("gdec_37_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_37_cast_fp16 = slice_by_index(begin = gdec_37_begin_0, end = gdec_37_end_0, end_mask = gdec_37_end_mask_0, x = gdec_31_cast_fp16)[name = string("gdec_37_cast_fp16")];
            tensor<int32, [2]> var_4078 = const()[name = string("op_4078"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_4079_cast_fp16 = reshape(shape = var_4078, x = gdec_37_cast_fp16)[name = string("op_4079_cast_fp16")];
            tensor<fp16, [12, 256]> var_4080_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_4079_cast_fp16)[name = string("op_4080_cast_fp16")];
            tensor<int32, [4]> var_4085 = const()[name = string("op_4085"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_29_cast_fp16 = reshape(shape = var_4085, x = var_4080_cast_fp16)[name = string("P_29_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4087_cast_fp16 = exp(x = P_29_cast_fp16)[name = string("op_4087_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_29_cast_fp16 = mul(x = var_4087_cast_fp16, y = tril_incl_to_fp16)[name = string("D_29_cast_fp16")];
            bool kk_29_transpose_x_1 = const()[name = string("kk_29_transpose_x_1"), val = bool(false)];
            bool kk_29_transpose_y_1 = const()[name = string("kk_29_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_29_cast_fp16 = matmul(transpose_x = kk_29_transpose_x_1, transpose_y = kk_29_transpose_y_1, x = kb_29_cast_fp16, y = k_65_cast_fp16)[name = string("kk_29_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_4093_cast_fp16 = mul(x = kk_29_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_4093_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4094_cast_fp16 = mul(x = var_4093_cast_fp16, y = D_29_cast_fp16)[name = string("op_4094_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_29_cast_fp16 = mul(x = var_4094_cast_fp16, y = cum_tail_to_fp16)[name = string("N_29_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4096_cast_fp16 = mul(x = D_29_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_4096_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_29_cast_fp16 = mul(x = kk_29_cast_fp16, y = var_4096_cast_fp16)[name = string("Na_29_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_113_cast_fp16 = add(x = eye_to_fp16, y = N_29_cast_fp16)[name = string("Minv_113_cast_fp16")];
            bool Np_85_transpose_x_0 = const()[name = string("Np_85_transpose_x_0"), val = bool(false)];
            bool Np_85_transpose_y_0 = const()[name = string("Np_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_85_cast_fp16 = matmul(transpose_x = Np_85_transpose_x_0, transpose_y = Np_85_transpose_y_0, x = N_29_cast_fp16, y = Na_29_cast_fp16)[name = string("Np_85_cast_fp16")];
            bool Npa_57_transpose_x_0 = const()[name = string("Npa_57_transpose_x_0"), val = bool(false)];
            bool Npa_57_transpose_y_0 = const()[name = string("Npa_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_57_cast_fp16 = matmul(transpose_x = Npa_57_transpose_x_0, transpose_y = Npa_57_transpose_y_0, x = Na_29_cast_fp16, y = N_29_cast_fp16)[name = string("Npa_57_cast_fp16")];
            bool var_4102_transpose_x_0 = const()[name = string("op_4102_transpose_x_0"), val = bool(false)];
            bool var_4102_transpose_y_0 = const()[name = string("op_4102_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4102_cast_fp16 = matmul(transpose_x = var_4102_transpose_x_0, transpose_y = var_4102_transpose_y_0, x = Minv_113_cast_fp16, y = Np_85_cast_fp16)[name = string("op_4102_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_115_cast_fp16 = add(x = Minv_113_cast_fp16, y = var_4102_cast_fp16)[name = string("Minv_115_cast_fp16")];
            bool Np_87_transpose_x_0 = const()[name = string("Np_87_transpose_x_0"), val = bool(false)];
            bool Np_87_transpose_y_0 = const()[name = string("Np_87_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_87_cast_fp16 = matmul(transpose_x = Np_87_transpose_x_0, transpose_y = Np_87_transpose_y_0, x = Np_85_cast_fp16, y = Npa_57_cast_fp16)[name = string("Np_87_cast_fp16")];
            bool Npa_59_transpose_x_0 = const()[name = string("Npa_59_transpose_x_0"), val = bool(false)];
            bool Npa_59_transpose_y_0 = const()[name = string("Npa_59_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_59_cast_fp16 = matmul(transpose_x = Npa_59_transpose_x_0, transpose_y = Npa_59_transpose_y_0, x = Npa_57_cast_fp16, y = Np_85_cast_fp16)[name = string("Npa_59_cast_fp16")];
            bool var_4107_transpose_x_0 = const()[name = string("op_4107_transpose_x_0"), val = bool(false)];
            bool var_4107_transpose_y_0 = const()[name = string("op_4107_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4107_cast_fp16 = matmul(transpose_x = var_4107_transpose_x_0, transpose_y = var_4107_transpose_y_0, x = Minv_115_cast_fp16, y = Np_87_cast_fp16)[name = string("op_4107_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_117_cast_fp16 = add(x = Minv_115_cast_fp16, y = var_4107_cast_fp16)[name = string("Minv_117_cast_fp16")];
            bool Np_89_transpose_x_0 = const()[name = string("Np_89_transpose_x_0"), val = bool(false)];
            bool Np_89_transpose_y_0 = const()[name = string("Np_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_89_cast_fp16 = matmul(transpose_x = Np_89_transpose_x_0, transpose_y = Np_89_transpose_y_0, x = Np_87_cast_fp16, y = Npa_59_cast_fp16)[name = string("Np_89_cast_fp16")];
            bool var_4111_transpose_x_0 = const()[name = string("op_4111_transpose_x_0"), val = bool(false)];
            bool var_4111_transpose_y_0 = const()[name = string("op_4111_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4111_cast_fp16 = matmul(transpose_x = var_4111_transpose_x_0, transpose_y = var_4111_transpose_y_0, x = Minv_117_cast_fp16, y = Np_89_cast_fp16)[name = string("op_4111_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_119_cast_fp16 = add(x = Minv_117_cast_fp16, y = var_4111_cast_fp16)[name = string("Minv_119_cast_fp16")];
            bool u_29_transpose_x_0 = const()[name = string("u_29_transpose_x_0"), val = bool(false)];
            bool u_29_transpose_y_0 = const()[name = string("u_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_29_cast_fp16 = matmul(transpose_x = u_29_transpose_x_0, transpose_y = u_29_transpose_y_0, x = Minv_119_cast_fp16, y = vb_37_cast_fp16)[name = string("u_29_cast_fp16")];
            bool var_4118_transpose_x_1 = const()[name = string("op_4118_transpose_x_1"), val = bool(false)];
            bool var_4118_transpose_y_1 = const()[name = string("op_4118_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_4118_cast_fp16 = matmul(transpose_x = var_4118_transpose_x_1, transpose_y = var_4118_transpose_y_1, x = q_63_cast_fp16, y = k_65_cast_fp16)[name = string("op_4118_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_29_cast_fp16 = mul(x = var_4118_cast_fp16, y = D_29_cast_fp16)[name = string("A_29_cast_fp16")];
            tensor<fp16, [16]> var_4120_bias_0_to_fp16 = const()[name = string("op_4120_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_4120_cast_fp16 = linear(bias = var_4120_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_37_cast_fp16)[name = string("op_4120_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_21_cast_fp16 = exp(x = var_4120_cast_fp16)[name = string("cum_21_cast_fp16")];
            tensor<int32, [1]> var_4123_axes_0 = const()[name = string("op_4123_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_4123_cast_fp16 = expand_dims(axes = var_4123_axes_0, x = cum_21_cast_fp16)[name = string("op_4123_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4124_cast_fp16 = mul(x = kb_29_cast_fp16, y = var_4123_cast_fp16)[name = string("op_4124_cast_fp16")];
            bool w_29_transpose_x_0 = const()[name = string("w_29_transpose_x_0"), val = bool(false)];
            bool w_29_transpose_y_0 = const()[name = string("w_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_29_cast_fp16 = matmul(transpose_x = w_29_transpose_x_0, transpose_y = w_29_transpose_y_0, x = Minv_119_cast_fp16, y = var_4124_cast_fp16)[name = string("w_29_cast_fp16")];
            bool var_4126_transpose_x_0 = const()[name = string("op_4126_transpose_x_0"), val = bool(false)];
            bool var_4126_transpose_y_0 = const()[name = string("op_4126_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4126_cast_fp16 = matmul(transpose_x = var_4126_transpose_x_0, transpose_y = var_4126_transpose_y_0, x = w_29_cast_fp16, y = S_21_cast_fp16)[name = string("op_4126_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_21_cast_fp16 = sub(x = u_29_cast_fp16, y = var_4126_cast_fp16)[name = string("vnew_21_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4131_cast_fp16 = mul(x = q_63_cast_fp16, y = var_4123_cast_fp16)[name = string("op_4131_cast_fp16")];
            bool var_4132_transpose_x_0 = const()[name = string("op_4132_transpose_x_0"), val = bool(false)];
            bool var_4132_transpose_y_0 = const()[name = string("op_4132_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4132_cast_fp16 = matmul(transpose_x = var_4132_transpose_x_0, transpose_y = var_4132_transpose_y_0, x = var_4131_cast_fp16, y = S_21_cast_fp16)[name = string("op_4132_cast_fp16")];
            bool var_4133_transpose_x_0 = const()[name = string("op_4133_transpose_x_0"), val = bool(false)];
            bool var_4133_transpose_y_0 = const()[name = string("op_4133_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4133_cast_fp16 = matmul(transpose_x = var_4133_transpose_x_0, transpose_y = var_4133_transpose_y_0, x = A_29_cast_fp16, y = vnew_21_cast_fp16)[name = string("op_4133_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4135_cast_fp16 = add(x = var_4132_cast_fp16, y = var_4133_cast_fp16)[name = string("op_4135_cast_fp16")];
            tensor<fp16, [16]> var_4136_bias_0_to_fp16 = const()[name = string("op_4136_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_4136_cast_fp16 = linear(bias = var_4136_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_37_cast_fp16)[name = string("op_4136_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_23_cast_fp16 = exp(x = var_4136_cast_fp16)[name = string("tail_23_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_45_to_fp16 = const()[name = string("transpose_45_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_4138_bias_0_to_fp16 = const()[name = string("op_4138_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_4138_cast_fp16 = linear(bias = var_4138_bias_0_to_fp16, weight = transpose_45_to_fp16, x = gdec_37_cast_fp16)[name = string("op_4138_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_4139_cast_fp16 = exp(x = var_4138_cast_fp16)[name = string("op_4139_cast_fp16")];
            tensor<int32, [1]> last_15_axes_0 = const()[name = string("last_15_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_15_cast_fp16 = expand_dims(axes = last_15_axes_0, x = var_4139_cast_fp16)[name = string("last_15_cast_fp16")];
            tensor<int32, [1]> var_4143_axes_0 = const()[name = string("op_4143_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_4143_cast_fp16 = expand_dims(axes = var_4143_axes_0, x = tail_23_cast_fp16)[name = string("op_4143_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4144_cast_fp16 = mul(x = k_65_cast_fp16, y = var_4143_cast_fp16)[name = string("op_4144_cast_fp16")];
            bool upd_15_transpose_x_1 = const()[name = string("upd_15_transpose_x_1"), val = bool(true)];
            bool upd_15_transpose_y_1 = const()[name = string("upd_15_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_15_cast_fp16 = matmul(transpose_x = upd_15_transpose_x_1, transpose_y = upd_15_transpose_y_1, x = var_4144_cast_fp16, y = vnew_21_cast_fp16)[name = string("upd_15_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_4149_cast_fp16 = mul(x = S_21_cast_fp16, y = last_15_cast_fp16)[name = string("op_4149_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_23_cast_fp16 = add(x = var_4149_cast_fp16, y = upd_15_cast_fp16)[name = string("S_23_cast_fp16")];
            tensor<int32, [4]> q_65_begin_0 = const()[name = string("q_65_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_65_end_0 = const()[name = string("q_65_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_65_end_mask_0 = const()[name = string("q_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_65_cast_fp16 = slice_by_index(begin = q_65_begin_0, end = q_65_end_0, end_mask = q_65_end_mask_0, x = q_57_cast_fp16)[name = string("q_65_cast_fp16")];
            tensor<int32, [4]> k_67_begin_0 = const()[name = string("k_67_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_67_end_0 = const()[name = string("k_67_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_67_end_mask_0 = const()[name = string("k_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_67_cast_fp16 = slice_by_index(begin = k_67_begin_0, end = k_67_end_0, end_mask = k_67_end_mask_0, x = k_59_cast_fp16)[name = string("k_67_cast_fp16")];
            tensor<int32, [4]> kb_31_begin_0 = const()[name = string("kb_31_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_31_end_0 = const()[name = string("kb_31_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_31_end_mask_0 = const()[name = string("kb_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_31_cast_fp16 = slice_by_index(begin = kb_31_begin_0, end = kb_31_end_0, end_mask = kb_31_end_mask_0, x = var_3744_cast_fp16)[name = string("kb_31_cast_fp16")];
            tensor<int32, [4]> vb_39_begin_0 = const()[name = string("vb_39_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_39_end_0 = const()[name = string("vb_39_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_39_end_mask_0 = const()[name = string("vb_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_39_cast_fp16 = slice_by_index(begin = vb_39_begin_0, end = vb_39_end_0, end_mask = vb_39_end_mask_0, x = vb_31_cast_fp16)[name = string("vb_39_cast_fp16")];
            tensor<int32, [3]> gdec_39_begin_0 = const()[name = string("gdec_39_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_39_end_0 = const()[name = string("gdec_39_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_39_end_mask_0 = const()[name = string("gdec_39_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_39_cast_fp16 = slice_by_index(begin = gdec_39_begin_0, end = gdec_39_end_0, end_mask = gdec_39_end_mask_0, x = gdec_31_cast_fp16)[name = string("gdec_39_cast_fp16")];
            tensor<int32, [2]> var_4232 = const()[name = string("op_4232"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_4233_cast_fp16 = reshape(shape = var_4232, x = gdec_39_cast_fp16)[name = string("op_4233_cast_fp16")];
            tensor<fp16, [12, 256]> var_4234_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_4233_cast_fp16)[name = string("op_4234_cast_fp16")];
            tensor<int32, [4]> var_4239 = const()[name = string("op_4239"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_31_cast_fp16 = reshape(shape = var_4239, x = var_4234_cast_fp16)[name = string("P_31_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4241_cast_fp16 = exp(x = P_31_cast_fp16)[name = string("op_4241_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_31_cast_fp16 = mul(x = var_4241_cast_fp16, y = tril_incl_to_fp16)[name = string("D_31_cast_fp16")];
            bool kk_31_transpose_x_1 = const()[name = string("kk_31_transpose_x_1"), val = bool(false)];
            bool kk_31_transpose_y_1 = const()[name = string("kk_31_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_31_cast_fp16 = matmul(transpose_x = kk_31_transpose_x_1, transpose_y = kk_31_transpose_y_1, x = kb_31_cast_fp16, y = k_67_cast_fp16)[name = string("kk_31_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_4247_cast_fp16 = mul(x = kk_31_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_4247_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4248_cast_fp16 = mul(x = var_4247_cast_fp16, y = D_31_cast_fp16)[name = string("op_4248_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_31_cast_fp16 = mul(x = var_4248_cast_fp16, y = cum_tail_to_fp16)[name = string("N_31_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4250_cast_fp16 = mul(x = D_31_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_4250_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_31_cast_fp16 = mul(x = kk_31_cast_fp16, y = var_4250_cast_fp16)[name = string("Na_31_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_121_cast_fp16 = add(x = eye_to_fp16, y = N_31_cast_fp16)[name = string("Minv_121_cast_fp16")];
            bool Np_91_transpose_x_0 = const()[name = string("Np_91_transpose_x_0"), val = bool(false)];
            bool Np_91_transpose_y_0 = const()[name = string("Np_91_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_91_cast_fp16 = matmul(transpose_x = Np_91_transpose_x_0, transpose_y = Np_91_transpose_y_0, x = N_31_cast_fp16, y = Na_31_cast_fp16)[name = string("Np_91_cast_fp16")];
            bool Npa_61_transpose_x_0 = const()[name = string("Npa_61_transpose_x_0"), val = bool(false)];
            bool Npa_61_transpose_y_0 = const()[name = string("Npa_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_61_cast_fp16 = matmul(transpose_x = Npa_61_transpose_x_0, transpose_y = Npa_61_transpose_y_0, x = Na_31_cast_fp16, y = N_31_cast_fp16)[name = string("Npa_61_cast_fp16")];
            bool var_4256_transpose_x_0 = const()[name = string("op_4256_transpose_x_0"), val = bool(false)];
            bool var_4256_transpose_y_0 = const()[name = string("op_4256_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4256_cast_fp16 = matmul(transpose_x = var_4256_transpose_x_0, transpose_y = var_4256_transpose_y_0, x = Minv_121_cast_fp16, y = Np_91_cast_fp16)[name = string("op_4256_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_123_cast_fp16 = add(x = Minv_121_cast_fp16, y = var_4256_cast_fp16)[name = string("Minv_123_cast_fp16")];
            bool Np_93_transpose_x_0 = const()[name = string("Np_93_transpose_x_0"), val = bool(false)];
            bool Np_93_transpose_y_0 = const()[name = string("Np_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_93_cast_fp16 = matmul(transpose_x = Np_93_transpose_x_0, transpose_y = Np_93_transpose_y_0, x = Np_91_cast_fp16, y = Npa_61_cast_fp16)[name = string("Np_93_cast_fp16")];
            bool Npa_63_transpose_x_0 = const()[name = string("Npa_63_transpose_x_0"), val = bool(false)];
            bool Npa_63_transpose_y_0 = const()[name = string("Npa_63_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_63_cast_fp16 = matmul(transpose_x = Npa_63_transpose_x_0, transpose_y = Npa_63_transpose_y_0, x = Npa_61_cast_fp16, y = Np_91_cast_fp16)[name = string("Npa_63_cast_fp16")];
            bool var_4261_transpose_x_0 = const()[name = string("op_4261_transpose_x_0"), val = bool(false)];
            bool var_4261_transpose_y_0 = const()[name = string("op_4261_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4261_cast_fp16 = matmul(transpose_x = var_4261_transpose_x_0, transpose_y = var_4261_transpose_y_0, x = Minv_123_cast_fp16, y = Np_93_cast_fp16)[name = string("op_4261_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_125_cast_fp16 = add(x = Minv_123_cast_fp16, y = var_4261_cast_fp16)[name = string("Minv_125_cast_fp16")];
            bool Np_95_transpose_x_0 = const()[name = string("Np_95_transpose_x_0"), val = bool(false)];
            bool Np_95_transpose_y_0 = const()[name = string("Np_95_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_95_cast_fp16 = matmul(transpose_x = Np_95_transpose_x_0, transpose_y = Np_95_transpose_y_0, x = Np_93_cast_fp16, y = Npa_63_cast_fp16)[name = string("Np_95_cast_fp16")];
            bool var_4265_transpose_x_0 = const()[name = string("op_4265_transpose_x_0"), val = bool(false)];
            bool var_4265_transpose_y_0 = const()[name = string("op_4265_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4265_cast_fp16 = matmul(transpose_x = var_4265_transpose_x_0, transpose_y = var_4265_transpose_y_0, x = Minv_125_cast_fp16, y = Np_95_cast_fp16)[name = string("op_4265_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_127_cast_fp16 = add(x = Minv_125_cast_fp16, y = var_4265_cast_fp16)[name = string("Minv_127_cast_fp16")];
            bool u_31_transpose_x_0 = const()[name = string("u_31_transpose_x_0"), val = bool(false)];
            bool u_31_transpose_y_0 = const()[name = string("u_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_31_cast_fp16 = matmul(transpose_x = u_31_transpose_x_0, transpose_y = u_31_transpose_y_0, x = Minv_127_cast_fp16, y = vb_39_cast_fp16)[name = string("u_31_cast_fp16")];
            bool var_4272_transpose_x_1 = const()[name = string("op_4272_transpose_x_1"), val = bool(false)];
            bool var_4272_transpose_y_1 = const()[name = string("op_4272_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_4272_cast_fp16 = matmul(transpose_x = var_4272_transpose_x_1, transpose_y = var_4272_transpose_y_1, x = q_65_cast_fp16, y = k_67_cast_fp16)[name = string("op_4272_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_31_cast_fp16 = mul(x = var_4272_cast_fp16, y = D_31_cast_fp16)[name = string("A_31_cast_fp16")];
            tensor<fp16, [16]> var_4274_bias_0_to_fp16 = const()[name = string("op_4274_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_4274_cast_fp16 = linear(bias = var_4274_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_39_cast_fp16)[name = string("op_4274_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_23_cast_fp16 = exp(x = var_4274_cast_fp16)[name = string("cum_23_cast_fp16")];
            tensor<int32, [1]> var_4277_axes_0 = const()[name = string("op_4277_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_4277_cast_fp16 = expand_dims(axes = var_4277_axes_0, x = cum_23_cast_fp16)[name = string("op_4277_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4278_cast_fp16 = mul(x = kb_31_cast_fp16, y = var_4277_cast_fp16)[name = string("op_4278_cast_fp16")];
            bool w_31_transpose_x_0 = const()[name = string("w_31_transpose_x_0"), val = bool(false)];
            bool w_31_transpose_y_0 = const()[name = string("w_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_31_cast_fp16 = matmul(transpose_x = w_31_transpose_x_0, transpose_y = w_31_transpose_y_0, x = Minv_127_cast_fp16, y = var_4278_cast_fp16)[name = string("w_31_cast_fp16")];
            bool var_4280_transpose_x_0 = const()[name = string("op_4280_transpose_x_0"), val = bool(false)];
            bool var_4280_transpose_y_0 = const()[name = string("op_4280_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4280_cast_fp16 = matmul(transpose_x = var_4280_transpose_x_0, transpose_y = var_4280_transpose_y_0, x = w_31_cast_fp16, y = S_23_cast_fp16)[name = string("op_4280_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_23_cast_fp16 = sub(x = u_31_cast_fp16, y = var_4280_cast_fp16)[name = string("vnew_23_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4285_cast_fp16 = mul(x = q_65_cast_fp16, y = var_4277_cast_fp16)[name = string("op_4285_cast_fp16")];
            bool var_4286_transpose_x_0 = const()[name = string("op_4286_transpose_x_0"), val = bool(false)];
            bool var_4286_transpose_y_0 = const()[name = string("op_4286_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4286_cast_fp16 = matmul(transpose_x = var_4286_transpose_x_0, transpose_y = var_4286_transpose_y_0, x = var_4285_cast_fp16, y = S_23_cast_fp16)[name = string("op_4286_cast_fp16")];
            bool var_4287_transpose_x_0 = const()[name = string("op_4287_transpose_x_0"), val = bool(false)];
            bool var_4287_transpose_y_0 = const()[name = string("op_4287_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4287_cast_fp16 = matmul(transpose_x = var_4287_transpose_x_0, transpose_y = var_4287_transpose_y_0, x = A_31_cast_fp16, y = vnew_23_cast_fp16)[name = string("op_4287_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_7_cast_fp16 = add(x = var_4286_cast_fp16, y = var_4287_cast_fp16)[name = string("oc_7_cast_fp16")];
            int32 var_4291 = const()[name = string("op_4291"), val = int32(2)];
            bool o_21_interleave_0 = const()[name = string("o_21_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 64, 128]> o_21_cast_fp16 = concat(axis = var_4291, interleave = o_21_interleave_0, values = (var_3834_cast_fp16, var_3981_cast_fp16, var_4135_cast_fp16, oc_7_cast_fp16))[name = string("o_21_cast_fp16")];
            tensor<int32, [4]> var_4297 = const()[name = string("op_4297"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_4303 = const()[name = string("op_4303"), val = tensor<int32, [4]>([3, 64, 4, 128])];
            tensor<fp16, [3, 64, 4, 128]> input_65_cast_fp16 = reshape(shape = var_4303, x = z_7_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> o_23_cast_fp16 = transpose(perm = var_4297, x = o_21_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [3, 64, 4, 128]> var_4305_cast_fp16 = mul(x = o_23_cast_fp16, y = o_23_cast_fp16)[name = string("op_4305_cast_fp16")];
            tensor<int32, [1]> var_4310_axes_0 = const()[name = string("op_4310_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4310_keep_dims_0 = const()[name = string("op_4310_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 4, 1]> var_4310_cast_fp16 = reduce_mean(axes = var_4310_axes_0, keep_dims = var_4310_keep_dims_0, x = var_4305_cast_fp16)[name = string("op_4310_cast_fp16")];
            fp16 var_4312_to_fp16 = const()[name = string("op_4312_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 64, 4, 1]> var_4313_cast_fp16 = add(x = var_4310_cast_fp16, y = var_4312_to_fp16)[name = string("op_4313_cast_fp16")];
            fp32 var_4314_epsilon_0 = const()[name = string("op_4314_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 4, 1]> var_4314_cast_fp16 = rsqrt(epsilon = var_4314_epsilon_0, x = var_4313_cast_fp16)[name = string("op_4314_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> var_4315_cast_fp16 = mul(x = o_23_cast_fp16, y = var_4314_cast_fp16)[name = string("op_4315_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> var_4317_cast_fp16 = silu(x = input_65_cast_fp16)[name = string("op_4317_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> o_25_cast_fp16 = mul(x = var_4315_cast_fp16, y = var_4317_cast_fp16)[name = string("o_25_cast_fp16")];
            tensor<int32, [3]> var_4322 = const()[name = string("op_4322"), val = tensor<int32, [3]>([3, 64, 512])];
            tensor<fp16, [3, 64, 512]> input_67_cast_fp16 = reshape(shape = var_4322, x = o_25_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [640, 512]> layers_4_out_weight_to_fp16 = const()[name = string("layers_4_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72063040)))];
            tensor<fp16, [3, 64, 640]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_out_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [3, 64, 640]> x_49_cast_fp16 = add(x = x_45_cast_fp16, y = linear_17_cast_fp16)[name = string("x_49_cast_fp16")];
            fp16 var_4335_to_fp16 = const()[name = string("op_4335_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 64, 640]> x_51_cast_fp16 = mul(x = x_49_cast_fp16, y = var_4335_to_fp16)[name = string("x_51_cast_fp16")];
            fp16 var_4333_promoted_to_fp16 = const()[name = string("op_4333_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> var_4337_cast_fp16 = pow(x = x_51_cast_fp16, y = var_4333_promoted_to_fp16)[name = string("op_4337_cast_fp16")];
            tensor<int32, [1]> var_4339_axes_0 = const()[name = string("op_4339_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4339_keep_dims_0 = const()[name = string("op_4339_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 1]> var_4339_cast_fp16 = reduce_mean(axes = var_4339_axes_0, keep_dims = var_4339_keep_dims_0, x = var_4337_cast_fp16)[name = string("op_4339_cast_fp16")];
            fp16 var_4340_to_fp16 = const()[name = string("op_4340_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 64, 1]> var_4341_cast_fp16 = add(x = var_4339_cast_fp16, y = var_4340_to_fp16)[name = string("op_4341_cast_fp16")];
            fp32 var_4342_epsilon_0 = const()[name = string("op_4342_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 1]> var_4342_cast_fp16 = rsqrt(epsilon = var_4342_epsilon_0, x = var_4341_cast_fp16)[name = string("op_4342_cast_fp16")];
            tensor<fp16, [3, 64, 640]> var_4343_cast_fp16 = mul(x = x_51_cast_fp16, y = var_4342_cast_fp16)[name = string("op_4343_cast_fp16")];
            tensor<fp16, [640]> layers_4_n2_w1_to_fp16 = const()[name = string("layers_4_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72718464)))];
            tensor<fp16, [3, 64, 640]> input_69_cast_fp16 = mul(x = var_4343_cast_fp16, y = layers_4_n2_w1_to_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_4_gate_up_weight_to_fp16 = const()[name = string("layers_4_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72719808)))];
            tensor<fp16, [3, 64, 5120]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_4_gate_up_weight_to_fp16, x = input_69_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<int32, [3]> input_71_begin_0 = const()[name = string("input_71_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_71_end_0 = const()[name = string("input_71_end_0"), val = tensor<int32, [3]>([3, 64, 2560])];
            tensor<bool, [3]> input_71_end_mask_0 = const()[name = string("input_71_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 2560]> input_71_cast_fp16 = slice_by_index(begin = input_71_begin_0, end = input_71_end_0, end_mask = input_71_end_mask_0, x = linear_18_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [3, 64, 2560]> var_4353_cast_fp16 = silu(x = input_71_cast_fp16)[name = string("op_4353_cast_fp16")];
            tensor<int32, [3]> var_4358_begin_0 = const()[name = string("op_4358_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_4358_end_0 = const()[name = string("op_4358_end_0"), val = tensor<int32, [3]>([3, 64, 5120])];
            tensor<bool, [3]> var_4358_end_mask_0 = const()[name = string("op_4358_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 2560]> var_4358_cast_fp16 = slice_by_index(begin = var_4358_begin_0, end = var_4358_end_0, end_mask = var_4358_end_mask_0, x = linear_18_cast_fp16)[name = string("op_4358_cast_fp16")];
            tensor<fp16, [3, 64, 2560]> input_73_cast_fp16 = mul(x = var_4353_cast_fp16, y = var_4358_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_4_down_weight_to_fp16 = const()[name = string("layers_4_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79273472)))];
            tensor<fp16, [3, 64, 640]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_down_weight_to_fp16, x = input_73_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [3, 64, 640]> x_53_cast_fp16 = add(x = x_49_cast_fp16, y = linear_19_cast_fp16)[name = string("x_53_cast_fp16")];
            fp16 var_4371_to_fp16 = const()[name = string("op_4371_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 64, 640]> x_55_cast_fp16 = mul(x = x_53_cast_fp16, y = var_4371_to_fp16)[name = string("x_55_cast_fp16")];
            fp16 var_4369_promoted_to_fp16 = const()[name = string("op_4369_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> var_4373_cast_fp16 = pow(x = x_55_cast_fp16, y = var_4369_promoted_to_fp16)[name = string("op_4373_cast_fp16")];
            tensor<int32, [1]> var_4375_axes_0 = const()[name = string("op_4375_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4375_keep_dims_0 = const()[name = string("op_4375_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 1]> var_4375_cast_fp16 = reduce_mean(axes = var_4375_axes_0, keep_dims = var_4375_keep_dims_0, x = var_4373_cast_fp16)[name = string("op_4375_cast_fp16")];
            fp16 var_4376_to_fp16 = const()[name = string("op_4376_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 64, 1]> var_4377_cast_fp16 = add(x = var_4375_cast_fp16, y = var_4376_to_fp16)[name = string("op_4377_cast_fp16")];
            fp32 var_4378_epsilon_0 = const()[name = string("op_4378_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 1]> var_4378_cast_fp16 = rsqrt(epsilon = var_4378_epsilon_0, x = var_4377_cast_fp16)[name = string("op_4378_cast_fp16")];
            tensor<fp16, [3, 64, 640]> var_4379_cast_fp16 = mul(x = x_55_cast_fp16, y = var_4378_cast_fp16)[name = string("op_4379_cast_fp16")];
            tensor<fp16, [640]> layers_5_n1_w1_to_fp16 = const()[name = string("layers_5_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82550336)))];
            tensor<fp16, [3, 64, 640]> input_75_cast_fp16 = mul(x = var_4379_cast_fp16, y = layers_5_n1_w1_to_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_5_qkvzba_weight_to_fp16 = const()[name = string("layers_5_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82551680)))];
            tensor<fp16, [3, 64, 1544]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_qkvzba_weight_to_fp16, x = input_75_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [3]> qkv_9_begin_0 = const()[name = string("qkv_9_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_9_end_0 = const()[name = string("qkv_9_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> qkv_9_end_mask_0 = const()[name = string("qkv_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 1024]> qkv_9_cast_fp16 = slice_by_index(begin = qkv_9_begin_0, end = qkv_9_end_0, end_mask = qkv_9_end_mask_0, x = linear_20_cast_fp16)[name = string("qkv_9_cast_fp16")];
            tensor<int32, [3]> z_9_begin_0 = const()[name = string("z_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_9_end_0 = const()[name = string("z_9_end_0"), val = tensor<int32, [3]>([3, 64, 1536])];
            tensor<bool, [3]> z_9_end_mask_0 = const()[name = string("z_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 512]> z_9_cast_fp16 = slice_by_index(begin = z_9_begin_0, end = z_9_end_0, end_mask = z_9_end_mask_0, x = linear_20_cast_fp16)[name = string("z_9_cast_fp16")];
            tensor<int32, [3]> b_9_begin_0 = const()[name = string("b_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_9_end_0 = const()[name = string("b_9_end_0"), val = tensor<int32, [3]>([3, 64, 1540])];
            tensor<bool, [3]> b_9_end_mask_0 = const()[name = string("b_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 4]> b_9_cast_fp16 = slice_by_index(begin = b_9_begin_0, end = b_9_end_0, end_mask = b_9_end_mask_0, x = linear_20_cast_fp16)[name = string("b_9_cast_fp16")];
            tensor<int32, [3]> a_9_begin_0 = const()[name = string("a_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_9_end_0 = const()[name = string("a_9_end_0"), val = tensor<int32, [3]>([3, 64, 1])];
            tensor<bool, [3]> a_9_end_mask_0 = const()[name = string("a_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 4]> a_9_cast_fp16 = slice_by_index(begin = a_9_begin_0, end = a_9_end_0, end_mask = a_9_end_mask_0, x = linear_20_cast_fp16)[name = string("a_9_cast_fp16")];
            int32 var_4418 = const()[name = string("op_4418"), val = int32(1)];
            bool xp_9_interleave_0 = const()[name = string("xp_9_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 67, 1024]> xp_9_cast_fp16 = concat(axis = var_4418, interleave = xp_9_interleave_0, values = (zpad_to_fp16, qkv_9_cast_fp16))[name = string("xp_9_cast_fp16")];
            tensor<int32, [3]> var_4429_begin_0 = const()[name = string("op_4429_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_4429_end_0 = const()[name = string("op_4429_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_4429_end_mask_0 = const()[name = string("op_4429_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 1024]> var_4429_cast_fp16 = slice_by_index(begin = var_4429_begin_0, end = var_4429_end_0, end_mask = var_4429_end_mask_0, x = xp_9_cast_fp16)[name = string("op_4429_cast_fp16")];
            tensor<fp16, [1024]> var_4437_to_fp16 = const()[name = string("op_4437_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84528064)))];
            tensor<fp16, [3, 64, 1024]> var_4438_cast_fp16 = mul(x = var_4429_cast_fp16, y = var_4437_to_fp16)[name = string("op_4438_cast_fp16")];
            tensor<int32, [3]> var_4448_begin_0 = const()[name = string("op_4448_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_4448_end_0 = const()[name = string("op_4448_end_0"), val = tensor<int32, [3]>([3, 66, 1024])];
            tensor<bool, [3]> var_4448_end_mask_0 = const()[name = string("op_4448_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_4448_cast_fp16 = slice_by_index(begin = var_4448_begin_0, end = var_4448_end_0, end_mask = var_4448_end_mask_0, x = xp_9_cast_fp16)[name = string("op_4448_cast_fp16")];
            tensor<fp16, [1024]> var_4456_to_fp16 = const()[name = string("op_4456_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84530176)))];
            tensor<fp16, [3, 64, 1024]> var_4457_cast_fp16 = mul(x = var_4448_cast_fp16, y = var_4456_to_fp16)[name = string("op_4457_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> var_4459_cast_fp16 = add(x = var_4438_cast_fp16, y = var_4457_cast_fp16)[name = string("op_4459_cast_fp16")];
            tensor<int32, [3]> var_4469_begin_0 = const()[name = string("op_4469_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_4469_end_0 = const()[name = string("op_4469_end_0"), val = tensor<int32, [3]>([3, 65, 1024])];
            tensor<bool, [3]> var_4469_end_mask_0 = const()[name = string("op_4469_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_4469_cast_fp16 = slice_by_index(begin = var_4469_begin_0, end = var_4469_end_0, end_mask = var_4469_end_mask_0, x = xp_9_cast_fp16)[name = string("op_4469_cast_fp16")];
            tensor<fp16, [1024]> var_4477_to_fp16 = const()[name = string("op_4477_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84532288)))];
            tensor<fp16, [3, 64, 1024]> var_4478_cast_fp16 = mul(x = var_4469_cast_fp16, y = var_4477_to_fp16)[name = string("op_4478_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> var_4480_cast_fp16 = add(x = var_4459_cast_fp16, y = var_4478_cast_fp16)[name = string("op_4480_cast_fp16")];
            tensor<int32, [3]> var_4490_begin_0 = const()[name = string("op_4490_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_4490_end_0 = const()[name = string("op_4490_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> var_4490_end_mask_0 = const()[name = string("op_4490_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_4490_cast_fp16 = slice_by_index(begin = var_4490_begin_0, end = var_4490_end_0, end_mask = var_4490_end_mask_0, x = xp_9_cast_fp16)[name = string("op_4490_cast_fp16")];
            tensor<fp16, [1024]> var_4498_to_fp16 = const()[name = string("op_4498_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84534400)))];
            tensor<fp16, [3, 64, 1024]> var_4499_cast_fp16 = mul(x = var_4490_cast_fp16, y = var_4498_to_fp16)[name = string("op_4499_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> input_77_cast_fp16 = add(x = var_4480_cast_fp16, y = var_4499_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> c_9_cast_fp16 = silu(x = input_77_cast_fp16)[name = string("c_9_cast_fp16")];
            tensor<int32, [3]> var_4507_begin_0 = const()[name = string("op_4507_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_4507_end_0 = const()[name = string("op_4507_end_0"), val = tensor<int32, [3]>([3, 64, 256])];
            tensor<bool, [3]> var_4507_end_mask_0 = const()[name = string("op_4507_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 256]> var_4507_cast_fp16 = slice_by_index(begin = var_4507_begin_0, end = var_4507_end_0, end_mask = var_4507_end_mask_0, x = c_9_cast_fp16)[name = string("op_4507_cast_fp16")];
            tensor<int32, [4]> var_4512 = const()[name = string("op_4512"), val = tensor<int32, [4]>([3, 64, 2, 128])];
            tensor<fp16, [3, 64, 2, 128]> q_67_cast_fp16 = reshape(shape = var_4512, x = var_4507_cast_fp16)[name = string("q_67_cast_fp16")];
            tensor<int32, [3]> var_4518_begin_0 = const()[name = string("op_4518_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_4518_end_0 = const()[name = string("op_4518_end_0"), val = tensor<int32, [3]>([3, 64, 512])];
            tensor<bool, [3]> var_4518_end_mask_0 = const()[name = string("op_4518_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 256]> var_4518_cast_fp16 = slice_by_index(begin = var_4518_begin_0, end = var_4518_end_0, end_mask = var_4518_end_mask_0, x = c_9_cast_fp16)[name = string("op_4518_cast_fp16")];
            tensor<int32, [4]> var_4523 = const()[name = string("op_4523"), val = tensor<int32, [4]>([3, 64, 2, 128])];
            tensor<fp16, [3, 64, 2, 128]> k_69_cast_fp16 = reshape(shape = var_4523, x = var_4518_cast_fp16)[name = string("k_69_cast_fp16")];
            tensor<int32, [3]> var_4529_begin_0 = const()[name = string("op_4529_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_4529_end_0 = const()[name = string("op_4529_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> var_4529_end_mask_0 = const()[name = string("op_4529_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 512]> var_4529_cast_fp16 = slice_by_index(begin = var_4529_begin_0, end = var_4529_end_0, end_mask = var_4529_end_mask_0, x = c_9_cast_fp16)[name = string("op_4529_cast_fp16")];
            tensor<int32, [4]> var_4534 = const()[name = string("op_4534"), val = tensor<int32, [4]>([3, 64, 4, 128])];
            tensor<fp16, [3, 64, 4, 128]> var_4535_cast_fp16 = reshape(shape = var_4534, x = var_4529_cast_fp16)[name = string("op_4535_cast_fp16")];
            tensor<int32, [4]> var_4540 = const()[name = string("op_4540"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4542_to_fp16 = const()[name = string("op_4542_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 64, 2, 128]> q_69_cast_fp16 = mul(x = q_67_cast_fp16, y = var_4542_to_fp16)[name = string("q_69_cast_fp16")];
            fp16 var_4544_to_fp16 = const()[name = string("op_4544_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 64, 2, 128]> k_71_cast_fp16 = mul(x = k_69_cast_fp16, y = var_4544_to_fp16)[name = string("k_71_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_4546_cast_fp16 = mul(x = q_69_cast_fp16, y = q_69_cast_fp16)[name = string("op_4546_cast_fp16")];
            tensor<int32, [1]> var_4551_axes_0 = const()[name = string("op_4551_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4551_keep_dims_0 = const()[name = string("op_4551_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 2, 1]> var_4551_cast_fp16 = reduce_mean(axes = var_4551_axes_0, keep_dims = var_4551_keep_dims_0, x = var_4546_cast_fp16)[name = string("op_4551_cast_fp16")];
            fp16 var_4552_promoted_to_fp16 = const()[name = string("op_4552_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 64, 2, 1]> var_4553_cast_fp16 = mul(x = var_4551_cast_fp16, y = var_4552_promoted_to_fp16)[name = string("op_4553_cast_fp16")];
            fp16 var_4555_to_fp16 = const()[name = string("op_4555_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 64, 2, 1]> var_4556_cast_fp16 = add(x = var_4553_cast_fp16, y = var_4555_to_fp16)[name = string("op_4556_cast_fp16")];
            fp32 var_4557_epsilon_0 = const()[name = string("op_4557_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 2, 1]> var_4557_cast_fp16 = rsqrt(epsilon = var_4557_epsilon_0, x = var_4556_cast_fp16)[name = string("op_4557_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_4558_cast_fp16 = mul(x = q_69_cast_fp16, y = var_4557_cast_fp16)[name = string("op_4558_cast_fp16")];
            fp16 var_4559_to_fp16 = const()[name = string("op_4559_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 64, 2, 128]> q_71_cast_fp16 = mul(x = var_4558_cast_fp16, y = var_4559_to_fp16)[name = string("q_71_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_4561_cast_fp16 = mul(x = k_71_cast_fp16, y = k_71_cast_fp16)[name = string("op_4561_cast_fp16")];
            tensor<int32, [1]> var_4566_axes_0 = const()[name = string("op_4566_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4566_keep_dims_0 = const()[name = string("op_4566_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 2, 1]> var_4566_cast_fp16 = reduce_mean(axes = var_4566_axes_0, keep_dims = var_4566_keep_dims_0, x = var_4561_cast_fp16)[name = string("op_4566_cast_fp16")];
            fp16 var_4567_promoted_to_fp16 = const()[name = string("op_4567_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 64, 2, 1]> var_4568_cast_fp16 = mul(x = var_4566_cast_fp16, y = var_4567_promoted_to_fp16)[name = string("op_4568_cast_fp16")];
            fp16 var_4570_to_fp16 = const()[name = string("op_4570_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 64, 2, 1]> var_4571_cast_fp16 = add(x = var_4568_cast_fp16, y = var_4570_to_fp16)[name = string("op_4571_cast_fp16")];
            fp32 var_4572_epsilon_0 = const()[name = string("op_4572_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 2, 1]> var_4572_cast_fp16 = rsqrt(epsilon = var_4572_epsilon_0, x = var_4571_cast_fp16)[name = string("op_4572_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> k_73_cast_fp16 = mul(x = k_71_cast_fp16, y = var_4572_cast_fp16)[name = string("k_73_cast_fp16")];
            tensor<int32, [4]> var_4578 = const()[name = string("op_4578"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_4591_axes_0 = const()[name = string("op_4591_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 64, 128]> var_4579_cast_fp16 = transpose(perm = var_4578, x = q_71_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [3, 2, 1, 64, 128]> var_4591_cast_fp16 = expand_dims(axes = var_4591_axes_0, x = var_4579_cast_fp16)[name = string("op_4591_cast_fp16")];
            tensor<int32, [5]> var_4599_reps_0 = const()[name = string("op_4599_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 64, 128]> var_4599_cast_fp16 = tile(reps = var_4599_reps_0, x = var_4591_cast_fp16)[name = string("op_4599_cast_fp16")];
            tensor<int32, [4]> var_4604 = const()[name = string("op_4604"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<fp16, [3, 4, 64, 128]> q_73_cast_fp16 = reshape(shape = var_4604, x = var_4599_cast_fp16)[name = string("q_73_cast_fp16")];
            tensor<int32, [4]> var_4610 = const()[name = string("op_4610"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_4623_axes_0 = const()[name = string("op_4623_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 64, 128]> var_4611_cast_fp16 = transpose(perm = var_4610, x = k_73_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [3, 2, 1, 64, 128]> var_4623_cast_fp16 = expand_dims(axes = var_4623_axes_0, x = var_4611_cast_fp16)[name = string("op_4623_cast_fp16")];
            tensor<int32, [5]> var_4631_reps_0 = const()[name = string("op_4631_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 64, 128]> var_4631_cast_fp16 = tile(reps = var_4631_reps_0, x = var_4623_cast_fp16)[name = string("op_4631_cast_fp16")];
            tensor<int32, [4]> var_4636 = const()[name = string("op_4636"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<fp16, [3, 4, 64, 128]> k_75_cast_fp16 = reshape(shape = var_4636, x = var_4631_cast_fp16)[name = string("k_75_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_4638_cast_fp16 = sigmoid(x = b_9_cast_fp16)[name = string("op_4638_cast_fp16")];
            tensor<int32, [3]> var_4642 = const()[name = string("op_4642"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_5_dt_bias_to_fp16 = const()[name = string("layers_5_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 64, 4]> input_79_cast_fp16 = add(x = a_9_cast_fp16, y = layers_5_dt_bias_to_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_4646_cast_fp16 = relu(x = input_79_cast_fp16)[name = string("op_4646_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_4647_cast_fp16 = abs(x = input_79_cast_fp16)[name = string("op_4647_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 64, 4]> var_4648_cast_fp16 = mul(x = var_4647_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_4648_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_4649_cast_fp16 = exp(x = var_4648_cast_fp16)[name = string("op_4649_cast_fp16")];
            fp16 var_4651_to_fp16 = const()[name = string("op_4651_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 64, 4]> var_4652_cast_fp16 = add(x = var_4649_cast_fp16, y = var_4651_to_fp16)[name = string("op_4652_cast_fp16")];
            fp32 var_4653_epsilon_0 = const()[name = string("op_4653_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 64, 4]> var_4653_cast_fp16 = log(epsilon = var_4653_epsilon_0, x = var_4652_cast_fp16)[name = string("op_4653_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_4655_cast_fp16 = add(x = var_4646_cast_fp16, y = var_4653_cast_fp16)[name = string("op_4655_cast_fp16")];
            tensor<fp16, [4]> layers_5_negA_to_fp16 = const()[name = string("layers_5_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.2f8p+3, -0x1.ae8p+2, -0x1.548p+2, -0x1.63p+3])];
            tensor<fp16, [3, 64, 4]> var_4656_cast_fp16 = mul(x = layers_5_negA_to_fp16, y = var_4655_cast_fp16)[name = string("op_4656_cast_fp16")];
            tensor<int32, [3]> var_4660 = const()[name = string("op_4660"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_4663_axes_0 = const()[name = string("op_4663_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 64]> beta_9_cast_fp16 = transpose(perm = var_4642, x = var_4638_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [3, 4, 64, 1]> var_4663_cast_fp16 = expand_dims(axes = var_4663_axes_0, x = beta_9_cast_fp16)[name = string("op_4663_cast_fp16")];
            fp16 var_4664_to_fp16 = const()[name = string("op_4664_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 64, 1]> var_4665_cast_fp16 = mul(x = var_4663_cast_fp16, y = var_4664_to_fp16)[name = string("op_4665_cast_fp16")];
            tensor<fp16, [3, 4, 64, 128]> v_13_cast_fp16 = transpose(perm = var_4540, x = var_4535_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [3, 4, 64, 128]> vb_41_cast_fp16 = mul(x = v_13_cast_fp16, y = var_4665_cast_fp16)[name = string("vb_41_cast_fp16")];
            tensor<int32, [4]> q_75_begin_0 = const()[name = string("q_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_75_end_0 = const()[name = string("q_75_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_75_end_mask_0 = const()[name = string("q_75_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_75_cast_fp16 = slice_by_index(begin = q_75_begin_0, end = q_75_end_0, end_mask = q_75_end_mask_0, x = q_73_cast_fp16)[name = string("q_75_cast_fp16")];
            tensor<int32, [4]> k_77_begin_0 = const()[name = string("k_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_77_end_0 = const()[name = string("k_77_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_77_end_mask_0 = const()[name = string("k_77_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, x = k_75_cast_fp16)[name = string("k_77_cast_fp16")];
            tensor<fp16, [3, 4, 64, 128]> var_4699_cast_fp16 = mul(x = k_75_cast_fp16, y = var_4663_cast_fp16)[name = string("op_4699_cast_fp16")];
            tensor<int32, [4]> kb_33_begin_0 = const()[name = string("kb_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_33_end_0 = const()[name = string("kb_33_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_33_end_mask_0 = const()[name = string("kb_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_33_cast_fp16 = slice_by_index(begin = kb_33_begin_0, end = kb_33_end_0, end_mask = kb_33_end_mask_0, x = var_4699_cast_fp16)[name = string("kb_33_cast_fp16")];
            tensor<int32, [4]> vb_43_begin_0 = const()[name = string("vb_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_43_end_0 = const()[name = string("vb_43_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_43_end_mask_0 = const()[name = string("vb_43_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_43_cast_fp16 = slice_by_index(begin = vb_43_begin_0, end = vb_43_end_0, end_mask = vb_43_end_mask_0, x = vb_41_cast_fp16)[name = string("vb_43_cast_fp16")];
            tensor<int32, [3]> gdec_43_begin_0 = const()[name = string("gdec_43_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_43_end_0 = const()[name = string("gdec_43_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_43_end_mask_0 = const()[name = string("gdec_43_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 64]> gdec_41_cast_fp16 = transpose(perm = var_4660, x = var_4656_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [3, 4, 16]> gdec_43_cast_fp16 = slice_by_index(begin = gdec_43_begin_0, end = gdec_43_end_0, end_mask = gdec_43_end_mask_0, x = gdec_41_cast_fp16)[name = string("gdec_43_cast_fp16")];
            tensor<int32, [2]> var_4747 = const()[name = string("op_4747"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_4748_cast_fp16 = reshape(shape = var_4747, x = gdec_43_cast_fp16)[name = string("op_4748_cast_fp16")];
            tensor<fp16, [12, 256]> var_4749_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_4748_cast_fp16)[name = string("op_4749_cast_fp16")];
            tensor<int32, [4]> var_4754 = const()[name = string("op_4754"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_33_cast_fp16 = reshape(shape = var_4754, x = var_4749_cast_fp16)[name = string("P_33_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4756_cast_fp16 = exp(x = P_33_cast_fp16)[name = string("op_4756_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_33_cast_fp16 = mul(x = var_4756_cast_fp16, y = tril_incl_to_fp16)[name = string("D_33_cast_fp16")];
            bool kk_33_transpose_x_1 = const()[name = string("kk_33_transpose_x_1"), val = bool(false)];
            bool kk_33_transpose_y_1 = const()[name = string("kk_33_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_33_cast_fp16 = matmul(transpose_x = kk_33_transpose_x_1, transpose_y = kk_33_transpose_y_1, x = kb_33_cast_fp16, y = k_77_cast_fp16)[name = string("kk_33_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_4762_cast_fp16 = mul(x = kk_33_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_4762_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4763_cast_fp16 = mul(x = var_4762_cast_fp16, y = D_33_cast_fp16)[name = string("op_4763_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_33_cast_fp16 = mul(x = var_4763_cast_fp16, y = cum_tail_to_fp16)[name = string("N_33_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4765_cast_fp16 = mul(x = D_33_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_4765_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_33_cast_fp16 = mul(x = kk_33_cast_fp16, y = var_4765_cast_fp16)[name = string("Na_33_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_129_cast_fp16 = add(x = eye_to_fp16, y = N_33_cast_fp16)[name = string("Minv_129_cast_fp16")];
            bool Np_97_transpose_x_0 = const()[name = string("Np_97_transpose_x_0"), val = bool(false)];
            bool Np_97_transpose_y_0 = const()[name = string("Np_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_97_cast_fp16 = matmul(transpose_x = Np_97_transpose_x_0, transpose_y = Np_97_transpose_y_0, x = N_33_cast_fp16, y = Na_33_cast_fp16)[name = string("Np_97_cast_fp16")];
            bool Npa_65_transpose_x_0 = const()[name = string("Npa_65_transpose_x_0"), val = bool(false)];
            bool Npa_65_transpose_y_0 = const()[name = string("Npa_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_65_cast_fp16 = matmul(transpose_x = Npa_65_transpose_x_0, transpose_y = Npa_65_transpose_y_0, x = Na_33_cast_fp16, y = N_33_cast_fp16)[name = string("Npa_65_cast_fp16")];
            bool var_4771_transpose_x_0 = const()[name = string("op_4771_transpose_x_0"), val = bool(false)];
            bool var_4771_transpose_y_0 = const()[name = string("op_4771_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4771_cast_fp16 = matmul(transpose_x = var_4771_transpose_x_0, transpose_y = var_4771_transpose_y_0, x = Minv_129_cast_fp16, y = Np_97_cast_fp16)[name = string("op_4771_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_131_cast_fp16 = add(x = Minv_129_cast_fp16, y = var_4771_cast_fp16)[name = string("Minv_131_cast_fp16")];
            bool Np_99_transpose_x_0 = const()[name = string("Np_99_transpose_x_0"), val = bool(false)];
            bool Np_99_transpose_y_0 = const()[name = string("Np_99_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_99_cast_fp16 = matmul(transpose_x = Np_99_transpose_x_0, transpose_y = Np_99_transpose_y_0, x = Np_97_cast_fp16, y = Npa_65_cast_fp16)[name = string("Np_99_cast_fp16")];
            bool Npa_67_transpose_x_0 = const()[name = string("Npa_67_transpose_x_0"), val = bool(false)];
            bool Npa_67_transpose_y_0 = const()[name = string("Npa_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_67_cast_fp16 = matmul(transpose_x = Npa_67_transpose_x_0, transpose_y = Npa_67_transpose_y_0, x = Npa_65_cast_fp16, y = Np_97_cast_fp16)[name = string("Npa_67_cast_fp16")];
            bool var_4776_transpose_x_0 = const()[name = string("op_4776_transpose_x_0"), val = bool(false)];
            bool var_4776_transpose_y_0 = const()[name = string("op_4776_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4776_cast_fp16 = matmul(transpose_x = var_4776_transpose_x_0, transpose_y = var_4776_transpose_y_0, x = Minv_131_cast_fp16, y = Np_99_cast_fp16)[name = string("op_4776_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_133_cast_fp16 = add(x = Minv_131_cast_fp16, y = var_4776_cast_fp16)[name = string("Minv_133_cast_fp16")];
            bool Np_101_transpose_x_0 = const()[name = string("Np_101_transpose_x_0"), val = bool(false)];
            bool Np_101_transpose_y_0 = const()[name = string("Np_101_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_101_cast_fp16 = matmul(transpose_x = Np_101_transpose_x_0, transpose_y = Np_101_transpose_y_0, x = Np_99_cast_fp16, y = Npa_67_cast_fp16)[name = string("Np_101_cast_fp16")];
            bool var_4780_transpose_x_0 = const()[name = string("op_4780_transpose_x_0"), val = bool(false)];
            bool var_4780_transpose_y_0 = const()[name = string("op_4780_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4780_cast_fp16 = matmul(transpose_x = var_4780_transpose_x_0, transpose_y = var_4780_transpose_y_0, x = Minv_133_cast_fp16, y = Np_101_cast_fp16)[name = string("op_4780_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_135_cast_fp16 = add(x = Minv_133_cast_fp16, y = var_4780_cast_fp16)[name = string("Minv_135_cast_fp16")];
            bool u_33_transpose_x_0 = const()[name = string("u_33_transpose_x_0"), val = bool(false)];
            bool u_33_transpose_y_0 = const()[name = string("u_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_33_cast_fp16 = matmul(transpose_x = u_33_transpose_x_0, transpose_y = u_33_transpose_y_0, x = Minv_135_cast_fp16, y = vb_43_cast_fp16)[name = string("u_33_cast_fp16")];
            bool var_4787_transpose_x_1 = const()[name = string("op_4787_transpose_x_1"), val = bool(false)];
            bool var_4787_transpose_y_1 = const()[name = string("op_4787_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_4787_cast_fp16 = matmul(transpose_x = var_4787_transpose_x_1, transpose_y = var_4787_transpose_y_1, x = q_75_cast_fp16, y = k_77_cast_fp16)[name = string("op_4787_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_33_cast_fp16 = mul(x = var_4787_cast_fp16, y = D_33_cast_fp16)[name = string("A_33_cast_fp16")];
            bool var_4789_transpose_x_0 = const()[name = string("op_4789_transpose_x_0"), val = bool(false)];
            bool var_4789_transpose_y_0 = const()[name = string("op_4789_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4789_cast_fp16 = matmul(transpose_x = var_4789_transpose_x_0, transpose_y = var_4789_transpose_y_0, x = A_33_cast_fp16, y = u_33_cast_fp16)[name = string("op_4789_cast_fp16")];
            tensor<fp16, [16]> var_4790_bias_0_to_fp16 = const()[name = string("op_4790_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_4790_cast_fp16 = linear(bias = var_4790_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_43_cast_fp16)[name = string("op_4790_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_25_cast_fp16 = exp(x = var_4790_cast_fp16)[name = string("tail_25_cast_fp16")];
            tensor<int32, [1]> var_4793_axes_0 = const()[name = string("op_4793_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_4793_cast_fp16 = expand_dims(axes = var_4793_axes_0, x = tail_25_cast_fp16)[name = string("op_4793_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4794_cast_fp16 = mul(x = k_77_cast_fp16, y = var_4793_cast_fp16)[name = string("op_4794_cast_fp16")];
            bool S_25_transpose_x_1 = const()[name = string("S_25_transpose_x_1"), val = bool(true)];
            bool S_25_transpose_y_1 = const()[name = string("S_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_25_cast_fp16 = matmul(transpose_x = S_25_transpose_x_1, transpose_y = S_25_transpose_y_1, x = var_4794_cast_fp16, y = u_33_cast_fp16)[name = string("S_25_cast_fp16")];
            tensor<int32, [4]> q_77_begin_0 = const()[name = string("q_77_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_77_end_0 = const()[name = string("q_77_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> q_77_end_mask_0 = const()[name = string("q_77_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_77_cast_fp16 = slice_by_index(begin = q_77_begin_0, end = q_77_end_0, end_mask = q_77_end_mask_0, x = q_73_cast_fp16)[name = string("q_77_cast_fp16")];
            tensor<int32, [4]> k_79_begin_0 = const()[name = string("k_79_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_79_end_0 = const()[name = string("k_79_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> k_79_end_mask_0 = const()[name = string("k_79_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_79_cast_fp16 = slice_by_index(begin = k_79_begin_0, end = k_79_end_0, end_mask = k_79_end_mask_0, x = k_75_cast_fp16)[name = string("k_79_cast_fp16")];
            tensor<int32, [4]> kb_35_begin_0 = const()[name = string("kb_35_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_35_end_0 = const()[name = string("kb_35_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> kb_35_end_mask_0 = const()[name = string("kb_35_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_35_cast_fp16 = slice_by_index(begin = kb_35_begin_0, end = kb_35_end_0, end_mask = kb_35_end_mask_0, x = var_4699_cast_fp16)[name = string("kb_35_cast_fp16")];
            tensor<int32, [4]> vb_45_begin_0 = const()[name = string("vb_45_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_45_end_0 = const()[name = string("vb_45_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> vb_45_end_mask_0 = const()[name = string("vb_45_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_45_cast_fp16 = slice_by_index(begin = vb_45_begin_0, end = vb_45_end_0, end_mask = vb_45_end_mask_0, x = vb_41_cast_fp16)[name = string("vb_45_cast_fp16")];
            tensor<int32, [3]> gdec_45_begin_0 = const()[name = string("gdec_45_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_45_end_0 = const()[name = string("gdec_45_end_0"), val = tensor<int32, [3]>([3, 4, 32])];
            tensor<bool, [3]> gdec_45_end_mask_0 = const()[name = string("gdec_45_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_45_cast_fp16 = slice_by_index(begin = gdec_45_begin_0, end = gdec_45_end_0, end_mask = gdec_45_end_mask_0, x = gdec_41_cast_fp16)[name = string("gdec_45_cast_fp16")];
            tensor<int32, [2]> var_4879 = const()[name = string("op_4879"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_4880_cast_fp16 = reshape(shape = var_4879, x = gdec_45_cast_fp16)[name = string("op_4880_cast_fp16")];
            tensor<fp16, [12, 256]> var_4881_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_4880_cast_fp16)[name = string("op_4881_cast_fp16")];
            tensor<int32, [4]> var_4886 = const()[name = string("op_4886"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_35_cast_fp16 = reshape(shape = var_4886, x = var_4881_cast_fp16)[name = string("P_35_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4888_cast_fp16 = exp(x = P_35_cast_fp16)[name = string("op_4888_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_35_cast_fp16 = mul(x = var_4888_cast_fp16, y = tril_incl_to_fp16)[name = string("D_35_cast_fp16")];
            bool kk_35_transpose_x_1 = const()[name = string("kk_35_transpose_x_1"), val = bool(false)];
            bool kk_35_transpose_y_1 = const()[name = string("kk_35_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_35_cast_fp16 = matmul(transpose_x = kk_35_transpose_x_1, transpose_y = kk_35_transpose_y_1, x = kb_35_cast_fp16, y = k_79_cast_fp16)[name = string("kk_35_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_4894_cast_fp16 = mul(x = kk_35_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_4894_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4895_cast_fp16 = mul(x = var_4894_cast_fp16, y = D_35_cast_fp16)[name = string("op_4895_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_35_cast_fp16 = mul(x = var_4895_cast_fp16, y = cum_tail_to_fp16)[name = string("N_35_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4897_cast_fp16 = mul(x = D_35_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_4897_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_35_cast_fp16 = mul(x = kk_35_cast_fp16, y = var_4897_cast_fp16)[name = string("Na_35_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_137_cast_fp16 = add(x = eye_to_fp16, y = N_35_cast_fp16)[name = string("Minv_137_cast_fp16")];
            bool Np_103_transpose_x_0 = const()[name = string("Np_103_transpose_x_0"), val = bool(false)];
            bool Np_103_transpose_y_0 = const()[name = string("Np_103_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_103_cast_fp16 = matmul(transpose_x = Np_103_transpose_x_0, transpose_y = Np_103_transpose_y_0, x = N_35_cast_fp16, y = Na_35_cast_fp16)[name = string("Np_103_cast_fp16")];
            bool Npa_69_transpose_x_0 = const()[name = string("Npa_69_transpose_x_0"), val = bool(false)];
            bool Npa_69_transpose_y_0 = const()[name = string("Npa_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_69_cast_fp16 = matmul(transpose_x = Npa_69_transpose_x_0, transpose_y = Npa_69_transpose_y_0, x = Na_35_cast_fp16, y = N_35_cast_fp16)[name = string("Npa_69_cast_fp16")];
            bool var_4903_transpose_x_0 = const()[name = string("op_4903_transpose_x_0"), val = bool(false)];
            bool var_4903_transpose_y_0 = const()[name = string("op_4903_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4903_cast_fp16 = matmul(transpose_x = var_4903_transpose_x_0, transpose_y = var_4903_transpose_y_0, x = Minv_137_cast_fp16, y = Np_103_cast_fp16)[name = string("op_4903_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_139_cast_fp16 = add(x = Minv_137_cast_fp16, y = var_4903_cast_fp16)[name = string("Minv_139_cast_fp16")];
            bool Np_105_transpose_x_0 = const()[name = string("Np_105_transpose_x_0"), val = bool(false)];
            bool Np_105_transpose_y_0 = const()[name = string("Np_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_105_cast_fp16 = matmul(transpose_x = Np_105_transpose_x_0, transpose_y = Np_105_transpose_y_0, x = Np_103_cast_fp16, y = Npa_69_cast_fp16)[name = string("Np_105_cast_fp16")];
            bool Npa_71_transpose_x_0 = const()[name = string("Npa_71_transpose_x_0"), val = bool(false)];
            bool Npa_71_transpose_y_0 = const()[name = string("Npa_71_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_71_cast_fp16 = matmul(transpose_x = Npa_71_transpose_x_0, transpose_y = Npa_71_transpose_y_0, x = Npa_69_cast_fp16, y = Np_103_cast_fp16)[name = string("Npa_71_cast_fp16")];
            bool var_4908_transpose_x_0 = const()[name = string("op_4908_transpose_x_0"), val = bool(false)];
            bool var_4908_transpose_y_0 = const()[name = string("op_4908_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4908_cast_fp16 = matmul(transpose_x = var_4908_transpose_x_0, transpose_y = var_4908_transpose_y_0, x = Minv_139_cast_fp16, y = Np_105_cast_fp16)[name = string("op_4908_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_141_cast_fp16 = add(x = Minv_139_cast_fp16, y = var_4908_cast_fp16)[name = string("Minv_141_cast_fp16")];
            bool Np_107_transpose_x_0 = const()[name = string("Np_107_transpose_x_0"), val = bool(false)];
            bool Np_107_transpose_y_0 = const()[name = string("Np_107_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_107_cast_fp16 = matmul(transpose_x = Np_107_transpose_x_0, transpose_y = Np_107_transpose_y_0, x = Np_105_cast_fp16, y = Npa_71_cast_fp16)[name = string("Np_107_cast_fp16")];
            bool var_4912_transpose_x_0 = const()[name = string("op_4912_transpose_x_0"), val = bool(false)];
            bool var_4912_transpose_y_0 = const()[name = string("op_4912_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4912_cast_fp16 = matmul(transpose_x = var_4912_transpose_x_0, transpose_y = var_4912_transpose_y_0, x = Minv_141_cast_fp16, y = Np_107_cast_fp16)[name = string("op_4912_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_143_cast_fp16 = add(x = Minv_141_cast_fp16, y = var_4912_cast_fp16)[name = string("Minv_143_cast_fp16")];
            bool u_35_transpose_x_0 = const()[name = string("u_35_transpose_x_0"), val = bool(false)];
            bool u_35_transpose_y_0 = const()[name = string("u_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_35_cast_fp16 = matmul(transpose_x = u_35_transpose_x_0, transpose_y = u_35_transpose_y_0, x = Minv_143_cast_fp16, y = vb_45_cast_fp16)[name = string("u_35_cast_fp16")];
            bool var_4919_transpose_x_1 = const()[name = string("op_4919_transpose_x_1"), val = bool(false)];
            bool var_4919_transpose_y_1 = const()[name = string("op_4919_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_4919_cast_fp16 = matmul(transpose_x = var_4919_transpose_x_1, transpose_y = var_4919_transpose_y_1, x = q_77_cast_fp16, y = k_79_cast_fp16)[name = string("op_4919_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_35_cast_fp16 = mul(x = var_4919_cast_fp16, y = D_35_cast_fp16)[name = string("A_35_cast_fp16")];
            tensor<fp16, [16]> var_4921_bias_0_to_fp16 = const()[name = string("op_4921_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_4921_cast_fp16 = linear(bias = var_4921_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_45_cast_fp16)[name = string("op_4921_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_25_cast_fp16 = exp(x = var_4921_cast_fp16)[name = string("cum_25_cast_fp16")];
            tensor<int32, [1]> var_4924_axes_0 = const()[name = string("op_4924_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_4924_cast_fp16 = expand_dims(axes = var_4924_axes_0, x = cum_25_cast_fp16)[name = string("op_4924_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4925_cast_fp16 = mul(x = kb_35_cast_fp16, y = var_4924_cast_fp16)[name = string("op_4925_cast_fp16")];
            bool w_35_transpose_x_0 = const()[name = string("w_35_transpose_x_0"), val = bool(false)];
            bool w_35_transpose_y_0 = const()[name = string("w_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_35_cast_fp16 = matmul(transpose_x = w_35_transpose_x_0, transpose_y = w_35_transpose_y_0, x = Minv_143_cast_fp16, y = var_4925_cast_fp16)[name = string("w_35_cast_fp16")];
            bool var_4927_transpose_x_0 = const()[name = string("op_4927_transpose_x_0"), val = bool(false)];
            bool var_4927_transpose_y_0 = const()[name = string("op_4927_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4927_cast_fp16 = matmul(transpose_x = var_4927_transpose_x_0, transpose_y = var_4927_transpose_y_0, x = w_35_cast_fp16, y = S_25_cast_fp16)[name = string("op_4927_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_25_cast_fp16 = sub(x = u_35_cast_fp16, y = var_4927_cast_fp16)[name = string("vnew_25_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4932_cast_fp16 = mul(x = q_77_cast_fp16, y = var_4924_cast_fp16)[name = string("op_4932_cast_fp16")];
            bool var_4933_transpose_x_0 = const()[name = string("op_4933_transpose_x_0"), val = bool(false)];
            bool var_4933_transpose_y_0 = const()[name = string("op_4933_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4933_cast_fp16 = matmul(transpose_x = var_4933_transpose_x_0, transpose_y = var_4933_transpose_y_0, x = var_4932_cast_fp16, y = S_25_cast_fp16)[name = string("op_4933_cast_fp16")];
            bool var_4934_transpose_x_0 = const()[name = string("op_4934_transpose_x_0"), val = bool(false)];
            bool var_4934_transpose_y_0 = const()[name = string("op_4934_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4934_cast_fp16 = matmul(transpose_x = var_4934_transpose_x_0, transpose_y = var_4934_transpose_y_0, x = A_35_cast_fp16, y = vnew_25_cast_fp16)[name = string("op_4934_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4936_cast_fp16 = add(x = var_4933_cast_fp16, y = var_4934_cast_fp16)[name = string("op_4936_cast_fp16")];
            tensor<fp16, [16]> var_4937_bias_0_to_fp16 = const()[name = string("op_4937_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_4937_cast_fp16 = linear(bias = var_4937_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_45_cast_fp16)[name = string("op_4937_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_27_cast_fp16 = exp(x = var_4937_cast_fp16)[name = string("tail_27_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_53_to_fp16 = const()[name = string("transpose_53_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_4939_bias_0_to_fp16 = const()[name = string("op_4939_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_4939_cast_fp16 = linear(bias = var_4939_bias_0_to_fp16, weight = transpose_53_to_fp16, x = gdec_45_cast_fp16)[name = string("op_4939_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_4940_cast_fp16 = exp(x = var_4939_cast_fp16)[name = string("op_4940_cast_fp16")];
            tensor<int32, [1]> last_17_axes_0 = const()[name = string("last_17_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_17_cast_fp16 = expand_dims(axes = last_17_axes_0, x = var_4940_cast_fp16)[name = string("last_17_cast_fp16")];
            tensor<int32, [1]> var_4944_axes_0 = const()[name = string("op_4944_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_4944_cast_fp16 = expand_dims(axes = var_4944_axes_0, x = tail_27_cast_fp16)[name = string("op_4944_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4945_cast_fp16 = mul(x = k_79_cast_fp16, y = var_4944_cast_fp16)[name = string("op_4945_cast_fp16")];
            bool upd_17_transpose_x_1 = const()[name = string("upd_17_transpose_x_1"), val = bool(true)];
            bool upd_17_transpose_y_1 = const()[name = string("upd_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_17_cast_fp16 = matmul(transpose_x = upd_17_transpose_x_1, transpose_y = upd_17_transpose_y_1, x = var_4945_cast_fp16, y = vnew_25_cast_fp16)[name = string("upd_17_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_4950_cast_fp16 = mul(x = S_25_cast_fp16, y = last_17_cast_fp16)[name = string("op_4950_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_27_cast_fp16 = add(x = var_4950_cast_fp16, y = upd_17_cast_fp16)[name = string("S_27_cast_fp16")];
            tensor<int32, [4]> q_79_begin_0 = const()[name = string("q_79_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_79_end_0 = const()[name = string("q_79_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> q_79_end_mask_0 = const()[name = string("q_79_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, x = q_73_cast_fp16)[name = string("q_79_cast_fp16")];
            tensor<int32, [4]> k_81_begin_0 = const()[name = string("k_81_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_81_end_0 = const()[name = string("k_81_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> k_81_end_mask_0 = const()[name = string("k_81_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, x = k_75_cast_fp16)[name = string("k_81_cast_fp16")];
            tensor<int32, [4]> kb_37_begin_0 = const()[name = string("kb_37_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_37_end_0 = const()[name = string("kb_37_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> kb_37_end_mask_0 = const()[name = string("kb_37_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_37_cast_fp16 = slice_by_index(begin = kb_37_begin_0, end = kb_37_end_0, end_mask = kb_37_end_mask_0, x = var_4699_cast_fp16)[name = string("kb_37_cast_fp16")];
            tensor<int32, [4]> vb_47_begin_0 = const()[name = string("vb_47_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_47_end_0 = const()[name = string("vb_47_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> vb_47_end_mask_0 = const()[name = string("vb_47_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_47_cast_fp16 = slice_by_index(begin = vb_47_begin_0, end = vb_47_end_0, end_mask = vb_47_end_mask_0, x = vb_41_cast_fp16)[name = string("vb_47_cast_fp16")];
            tensor<int32, [3]> gdec_47_begin_0 = const()[name = string("gdec_47_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_47_end_0 = const()[name = string("gdec_47_end_0"), val = tensor<int32, [3]>([3, 4, 48])];
            tensor<bool, [3]> gdec_47_end_mask_0 = const()[name = string("gdec_47_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_47_cast_fp16 = slice_by_index(begin = gdec_47_begin_0, end = gdec_47_end_0, end_mask = gdec_47_end_mask_0, x = gdec_41_cast_fp16)[name = string("gdec_47_cast_fp16")];
            tensor<int32, [2]> var_5033 = const()[name = string("op_5033"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_5034_cast_fp16 = reshape(shape = var_5033, x = gdec_47_cast_fp16)[name = string("op_5034_cast_fp16")];
            tensor<fp16, [12, 256]> var_5035_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5034_cast_fp16)[name = string("op_5035_cast_fp16")];
            tensor<int32, [4]> var_5040 = const()[name = string("op_5040"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_37_cast_fp16 = reshape(shape = var_5040, x = var_5035_cast_fp16)[name = string("P_37_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5042_cast_fp16 = exp(x = P_37_cast_fp16)[name = string("op_5042_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_37_cast_fp16 = mul(x = var_5042_cast_fp16, y = tril_incl_to_fp16)[name = string("D_37_cast_fp16")];
            bool kk_37_transpose_x_1 = const()[name = string("kk_37_transpose_x_1"), val = bool(false)];
            bool kk_37_transpose_y_1 = const()[name = string("kk_37_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_37_cast_fp16 = matmul(transpose_x = kk_37_transpose_x_1, transpose_y = kk_37_transpose_y_1, x = kb_37_cast_fp16, y = k_81_cast_fp16)[name = string("kk_37_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_5048_cast_fp16 = mul(x = kk_37_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_5048_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5049_cast_fp16 = mul(x = var_5048_cast_fp16, y = D_37_cast_fp16)[name = string("op_5049_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_37_cast_fp16 = mul(x = var_5049_cast_fp16, y = cum_tail_to_fp16)[name = string("N_37_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5051_cast_fp16 = mul(x = D_37_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5051_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_37_cast_fp16 = mul(x = kk_37_cast_fp16, y = var_5051_cast_fp16)[name = string("Na_37_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_145_cast_fp16 = add(x = eye_to_fp16, y = N_37_cast_fp16)[name = string("Minv_145_cast_fp16")];
            bool Np_109_transpose_x_0 = const()[name = string("Np_109_transpose_x_0"), val = bool(false)];
            bool Np_109_transpose_y_0 = const()[name = string("Np_109_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_109_cast_fp16 = matmul(transpose_x = Np_109_transpose_x_0, transpose_y = Np_109_transpose_y_0, x = N_37_cast_fp16, y = Na_37_cast_fp16)[name = string("Np_109_cast_fp16")];
            bool Npa_73_transpose_x_0 = const()[name = string("Npa_73_transpose_x_0"), val = bool(false)];
            bool Npa_73_transpose_y_0 = const()[name = string("Npa_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_73_cast_fp16 = matmul(transpose_x = Npa_73_transpose_x_0, transpose_y = Npa_73_transpose_y_0, x = Na_37_cast_fp16, y = N_37_cast_fp16)[name = string("Npa_73_cast_fp16")];
            bool var_5057_transpose_x_0 = const()[name = string("op_5057_transpose_x_0"), val = bool(false)];
            bool var_5057_transpose_y_0 = const()[name = string("op_5057_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5057_cast_fp16 = matmul(transpose_x = var_5057_transpose_x_0, transpose_y = var_5057_transpose_y_0, x = Minv_145_cast_fp16, y = Np_109_cast_fp16)[name = string("op_5057_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_147_cast_fp16 = add(x = Minv_145_cast_fp16, y = var_5057_cast_fp16)[name = string("Minv_147_cast_fp16")];
            bool Np_111_transpose_x_0 = const()[name = string("Np_111_transpose_x_0"), val = bool(false)];
            bool Np_111_transpose_y_0 = const()[name = string("Np_111_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_111_cast_fp16 = matmul(transpose_x = Np_111_transpose_x_0, transpose_y = Np_111_transpose_y_0, x = Np_109_cast_fp16, y = Npa_73_cast_fp16)[name = string("Np_111_cast_fp16")];
            bool Npa_75_transpose_x_0 = const()[name = string("Npa_75_transpose_x_0"), val = bool(false)];
            bool Npa_75_transpose_y_0 = const()[name = string("Npa_75_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_75_cast_fp16 = matmul(transpose_x = Npa_75_transpose_x_0, transpose_y = Npa_75_transpose_y_0, x = Npa_73_cast_fp16, y = Np_109_cast_fp16)[name = string("Npa_75_cast_fp16")];
            bool var_5062_transpose_x_0 = const()[name = string("op_5062_transpose_x_0"), val = bool(false)];
            bool var_5062_transpose_y_0 = const()[name = string("op_5062_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5062_cast_fp16 = matmul(transpose_x = var_5062_transpose_x_0, transpose_y = var_5062_transpose_y_0, x = Minv_147_cast_fp16, y = Np_111_cast_fp16)[name = string("op_5062_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_149_cast_fp16 = add(x = Minv_147_cast_fp16, y = var_5062_cast_fp16)[name = string("Minv_149_cast_fp16")];
            bool Np_113_transpose_x_0 = const()[name = string("Np_113_transpose_x_0"), val = bool(false)];
            bool Np_113_transpose_y_0 = const()[name = string("Np_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_113_cast_fp16 = matmul(transpose_x = Np_113_transpose_x_0, transpose_y = Np_113_transpose_y_0, x = Np_111_cast_fp16, y = Npa_75_cast_fp16)[name = string("Np_113_cast_fp16")];
            bool var_5066_transpose_x_0 = const()[name = string("op_5066_transpose_x_0"), val = bool(false)];
            bool var_5066_transpose_y_0 = const()[name = string("op_5066_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5066_cast_fp16 = matmul(transpose_x = var_5066_transpose_x_0, transpose_y = var_5066_transpose_y_0, x = Minv_149_cast_fp16, y = Np_113_cast_fp16)[name = string("op_5066_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_151_cast_fp16 = add(x = Minv_149_cast_fp16, y = var_5066_cast_fp16)[name = string("Minv_151_cast_fp16")];
            bool u_37_transpose_x_0 = const()[name = string("u_37_transpose_x_0"), val = bool(false)];
            bool u_37_transpose_y_0 = const()[name = string("u_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_37_cast_fp16 = matmul(transpose_x = u_37_transpose_x_0, transpose_y = u_37_transpose_y_0, x = Minv_151_cast_fp16, y = vb_47_cast_fp16)[name = string("u_37_cast_fp16")];
            bool var_5073_transpose_x_1 = const()[name = string("op_5073_transpose_x_1"), val = bool(false)];
            bool var_5073_transpose_y_1 = const()[name = string("op_5073_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_5073_cast_fp16 = matmul(transpose_x = var_5073_transpose_x_1, transpose_y = var_5073_transpose_y_1, x = q_79_cast_fp16, y = k_81_cast_fp16)[name = string("op_5073_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_37_cast_fp16 = mul(x = var_5073_cast_fp16, y = D_37_cast_fp16)[name = string("A_37_cast_fp16")];
            tensor<fp16, [16]> var_5075_bias_0_to_fp16 = const()[name = string("op_5075_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_5075_cast_fp16 = linear(bias = var_5075_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_47_cast_fp16)[name = string("op_5075_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_27_cast_fp16 = exp(x = var_5075_cast_fp16)[name = string("cum_27_cast_fp16")];
            tensor<int32, [1]> var_5078_axes_0 = const()[name = string("op_5078_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_5078_cast_fp16 = expand_dims(axes = var_5078_axes_0, x = cum_27_cast_fp16)[name = string("op_5078_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5079_cast_fp16 = mul(x = kb_37_cast_fp16, y = var_5078_cast_fp16)[name = string("op_5079_cast_fp16")];
            bool w_37_transpose_x_0 = const()[name = string("w_37_transpose_x_0"), val = bool(false)];
            bool w_37_transpose_y_0 = const()[name = string("w_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_37_cast_fp16 = matmul(transpose_x = w_37_transpose_x_0, transpose_y = w_37_transpose_y_0, x = Minv_151_cast_fp16, y = var_5079_cast_fp16)[name = string("w_37_cast_fp16")];
            bool var_5081_transpose_x_0 = const()[name = string("op_5081_transpose_x_0"), val = bool(false)];
            bool var_5081_transpose_y_0 = const()[name = string("op_5081_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5081_cast_fp16 = matmul(transpose_x = var_5081_transpose_x_0, transpose_y = var_5081_transpose_y_0, x = w_37_cast_fp16, y = S_27_cast_fp16)[name = string("op_5081_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_27_cast_fp16 = sub(x = u_37_cast_fp16, y = var_5081_cast_fp16)[name = string("vnew_27_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5086_cast_fp16 = mul(x = q_79_cast_fp16, y = var_5078_cast_fp16)[name = string("op_5086_cast_fp16")];
            bool var_5087_transpose_x_0 = const()[name = string("op_5087_transpose_x_0"), val = bool(false)];
            bool var_5087_transpose_y_0 = const()[name = string("op_5087_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5087_cast_fp16 = matmul(transpose_x = var_5087_transpose_x_0, transpose_y = var_5087_transpose_y_0, x = var_5086_cast_fp16, y = S_27_cast_fp16)[name = string("op_5087_cast_fp16")];
            bool var_5088_transpose_x_0 = const()[name = string("op_5088_transpose_x_0"), val = bool(false)];
            bool var_5088_transpose_y_0 = const()[name = string("op_5088_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5088_cast_fp16 = matmul(transpose_x = var_5088_transpose_x_0, transpose_y = var_5088_transpose_y_0, x = A_37_cast_fp16, y = vnew_27_cast_fp16)[name = string("op_5088_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5090_cast_fp16 = add(x = var_5087_cast_fp16, y = var_5088_cast_fp16)[name = string("op_5090_cast_fp16")];
            tensor<fp16, [16]> var_5091_bias_0_to_fp16 = const()[name = string("op_5091_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_5091_cast_fp16 = linear(bias = var_5091_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_47_cast_fp16)[name = string("op_5091_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_29_cast_fp16 = exp(x = var_5091_cast_fp16)[name = string("tail_29_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_57_to_fp16 = const()[name = string("transpose_57_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_5093_bias_0_to_fp16 = const()[name = string("op_5093_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_5093_cast_fp16 = linear(bias = var_5093_bias_0_to_fp16, weight = transpose_57_to_fp16, x = gdec_47_cast_fp16)[name = string("op_5093_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_5094_cast_fp16 = exp(x = var_5093_cast_fp16)[name = string("op_5094_cast_fp16")];
            tensor<int32, [1]> last_19_axes_0 = const()[name = string("last_19_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_19_cast_fp16 = expand_dims(axes = last_19_axes_0, x = var_5094_cast_fp16)[name = string("last_19_cast_fp16")];
            tensor<int32, [1]> var_5098_axes_0 = const()[name = string("op_5098_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_5098_cast_fp16 = expand_dims(axes = var_5098_axes_0, x = tail_29_cast_fp16)[name = string("op_5098_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5099_cast_fp16 = mul(x = k_81_cast_fp16, y = var_5098_cast_fp16)[name = string("op_5099_cast_fp16")];
            bool upd_19_transpose_x_1 = const()[name = string("upd_19_transpose_x_1"), val = bool(true)];
            bool upd_19_transpose_y_1 = const()[name = string("upd_19_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_19_cast_fp16 = matmul(transpose_x = upd_19_transpose_x_1, transpose_y = upd_19_transpose_y_1, x = var_5099_cast_fp16, y = vnew_27_cast_fp16)[name = string("upd_19_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_5104_cast_fp16 = mul(x = S_27_cast_fp16, y = last_19_cast_fp16)[name = string("op_5104_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_29_cast_fp16 = add(x = var_5104_cast_fp16, y = upd_19_cast_fp16)[name = string("S_29_cast_fp16")];
            tensor<int32, [4]> q_81_begin_0 = const()[name = string("q_81_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_81_end_0 = const()[name = string("q_81_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_81_end_mask_0 = const()[name = string("q_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_81_cast_fp16 = slice_by_index(begin = q_81_begin_0, end = q_81_end_0, end_mask = q_81_end_mask_0, x = q_73_cast_fp16)[name = string("q_81_cast_fp16")];
            tensor<int32, [4]> k_83_begin_0 = const()[name = string("k_83_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_83_end_0 = const()[name = string("k_83_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_83_end_mask_0 = const()[name = string("k_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_83_cast_fp16 = slice_by_index(begin = k_83_begin_0, end = k_83_end_0, end_mask = k_83_end_mask_0, x = k_75_cast_fp16)[name = string("k_83_cast_fp16")];
            tensor<int32, [4]> kb_39_begin_0 = const()[name = string("kb_39_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_39_end_0 = const()[name = string("kb_39_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_39_end_mask_0 = const()[name = string("kb_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_39_cast_fp16 = slice_by_index(begin = kb_39_begin_0, end = kb_39_end_0, end_mask = kb_39_end_mask_0, x = var_4699_cast_fp16)[name = string("kb_39_cast_fp16")];
            tensor<int32, [4]> vb_49_begin_0 = const()[name = string("vb_49_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_49_end_0 = const()[name = string("vb_49_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_49_end_mask_0 = const()[name = string("vb_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_49_cast_fp16 = slice_by_index(begin = vb_49_begin_0, end = vb_49_end_0, end_mask = vb_49_end_mask_0, x = vb_41_cast_fp16)[name = string("vb_49_cast_fp16")];
            tensor<int32, [3]> gdec_49_begin_0 = const()[name = string("gdec_49_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_49_end_0 = const()[name = string("gdec_49_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_49_end_mask_0 = const()[name = string("gdec_49_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_49_cast_fp16 = slice_by_index(begin = gdec_49_begin_0, end = gdec_49_end_0, end_mask = gdec_49_end_mask_0, x = gdec_41_cast_fp16)[name = string("gdec_49_cast_fp16")];
            tensor<int32, [2]> var_5187 = const()[name = string("op_5187"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_5188_cast_fp16 = reshape(shape = var_5187, x = gdec_49_cast_fp16)[name = string("op_5188_cast_fp16")];
            tensor<fp16, [12, 256]> var_5189_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5188_cast_fp16)[name = string("op_5189_cast_fp16")];
            tensor<int32, [4]> var_5194 = const()[name = string("op_5194"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_39_cast_fp16 = reshape(shape = var_5194, x = var_5189_cast_fp16)[name = string("P_39_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5196_cast_fp16 = exp(x = P_39_cast_fp16)[name = string("op_5196_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_39_cast_fp16 = mul(x = var_5196_cast_fp16, y = tril_incl_to_fp16)[name = string("D_39_cast_fp16")];
            bool kk_39_transpose_x_1 = const()[name = string("kk_39_transpose_x_1"), val = bool(false)];
            bool kk_39_transpose_y_1 = const()[name = string("kk_39_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_39_cast_fp16 = matmul(transpose_x = kk_39_transpose_x_1, transpose_y = kk_39_transpose_y_1, x = kb_39_cast_fp16, y = k_83_cast_fp16)[name = string("kk_39_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_5202_cast_fp16 = mul(x = kk_39_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_5202_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5203_cast_fp16 = mul(x = var_5202_cast_fp16, y = D_39_cast_fp16)[name = string("op_5203_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_39_cast_fp16 = mul(x = var_5203_cast_fp16, y = cum_tail_to_fp16)[name = string("N_39_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5205_cast_fp16 = mul(x = D_39_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5205_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_39_cast_fp16 = mul(x = kk_39_cast_fp16, y = var_5205_cast_fp16)[name = string("Na_39_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_153_cast_fp16 = add(x = eye_to_fp16, y = N_39_cast_fp16)[name = string("Minv_153_cast_fp16")];
            bool Np_115_transpose_x_0 = const()[name = string("Np_115_transpose_x_0"), val = bool(false)];
            bool Np_115_transpose_y_0 = const()[name = string("Np_115_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_115_cast_fp16 = matmul(transpose_x = Np_115_transpose_x_0, transpose_y = Np_115_transpose_y_0, x = N_39_cast_fp16, y = Na_39_cast_fp16)[name = string("Np_115_cast_fp16")];
            bool Npa_77_transpose_x_0 = const()[name = string("Npa_77_transpose_x_0"), val = bool(false)];
            bool Npa_77_transpose_y_0 = const()[name = string("Npa_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_77_cast_fp16 = matmul(transpose_x = Npa_77_transpose_x_0, transpose_y = Npa_77_transpose_y_0, x = Na_39_cast_fp16, y = N_39_cast_fp16)[name = string("Npa_77_cast_fp16")];
            bool var_5211_transpose_x_0 = const()[name = string("op_5211_transpose_x_0"), val = bool(false)];
            bool var_5211_transpose_y_0 = const()[name = string("op_5211_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5211_cast_fp16 = matmul(transpose_x = var_5211_transpose_x_0, transpose_y = var_5211_transpose_y_0, x = Minv_153_cast_fp16, y = Np_115_cast_fp16)[name = string("op_5211_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_155_cast_fp16 = add(x = Minv_153_cast_fp16, y = var_5211_cast_fp16)[name = string("Minv_155_cast_fp16")];
            bool Np_117_transpose_x_0 = const()[name = string("Np_117_transpose_x_0"), val = bool(false)];
            bool Np_117_transpose_y_0 = const()[name = string("Np_117_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_117_cast_fp16 = matmul(transpose_x = Np_117_transpose_x_0, transpose_y = Np_117_transpose_y_0, x = Np_115_cast_fp16, y = Npa_77_cast_fp16)[name = string("Np_117_cast_fp16")];
            bool Npa_79_transpose_x_0 = const()[name = string("Npa_79_transpose_x_0"), val = bool(false)];
            bool Npa_79_transpose_y_0 = const()[name = string("Npa_79_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_79_cast_fp16 = matmul(transpose_x = Npa_79_transpose_x_0, transpose_y = Npa_79_transpose_y_0, x = Npa_77_cast_fp16, y = Np_115_cast_fp16)[name = string("Npa_79_cast_fp16")];
            bool var_5216_transpose_x_0 = const()[name = string("op_5216_transpose_x_0"), val = bool(false)];
            bool var_5216_transpose_y_0 = const()[name = string("op_5216_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5216_cast_fp16 = matmul(transpose_x = var_5216_transpose_x_0, transpose_y = var_5216_transpose_y_0, x = Minv_155_cast_fp16, y = Np_117_cast_fp16)[name = string("op_5216_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_157_cast_fp16 = add(x = Minv_155_cast_fp16, y = var_5216_cast_fp16)[name = string("Minv_157_cast_fp16")];
            bool Np_119_transpose_x_0 = const()[name = string("Np_119_transpose_x_0"), val = bool(false)];
            bool Np_119_transpose_y_0 = const()[name = string("Np_119_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_119_cast_fp16 = matmul(transpose_x = Np_119_transpose_x_0, transpose_y = Np_119_transpose_y_0, x = Np_117_cast_fp16, y = Npa_79_cast_fp16)[name = string("Np_119_cast_fp16")];
            bool var_5220_transpose_x_0 = const()[name = string("op_5220_transpose_x_0"), val = bool(false)];
            bool var_5220_transpose_y_0 = const()[name = string("op_5220_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5220_cast_fp16 = matmul(transpose_x = var_5220_transpose_x_0, transpose_y = var_5220_transpose_y_0, x = Minv_157_cast_fp16, y = Np_119_cast_fp16)[name = string("op_5220_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_159_cast_fp16 = add(x = Minv_157_cast_fp16, y = var_5220_cast_fp16)[name = string("Minv_159_cast_fp16")];
            bool u_39_transpose_x_0 = const()[name = string("u_39_transpose_x_0"), val = bool(false)];
            bool u_39_transpose_y_0 = const()[name = string("u_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_39_cast_fp16 = matmul(transpose_x = u_39_transpose_x_0, transpose_y = u_39_transpose_y_0, x = Minv_159_cast_fp16, y = vb_49_cast_fp16)[name = string("u_39_cast_fp16")];
            bool var_5227_transpose_x_1 = const()[name = string("op_5227_transpose_x_1"), val = bool(false)];
            bool var_5227_transpose_y_1 = const()[name = string("op_5227_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_5227_cast_fp16 = matmul(transpose_x = var_5227_transpose_x_1, transpose_y = var_5227_transpose_y_1, x = q_81_cast_fp16, y = k_83_cast_fp16)[name = string("op_5227_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_39_cast_fp16 = mul(x = var_5227_cast_fp16, y = D_39_cast_fp16)[name = string("A_39_cast_fp16")];
            tensor<fp16, [16]> var_5229_bias_0_to_fp16 = const()[name = string("op_5229_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_5229_cast_fp16 = linear(bias = var_5229_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_49_cast_fp16)[name = string("op_5229_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_29_cast_fp16 = exp(x = var_5229_cast_fp16)[name = string("cum_29_cast_fp16")];
            tensor<int32, [1]> var_5232_axes_0 = const()[name = string("op_5232_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_5232_cast_fp16 = expand_dims(axes = var_5232_axes_0, x = cum_29_cast_fp16)[name = string("op_5232_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5233_cast_fp16 = mul(x = kb_39_cast_fp16, y = var_5232_cast_fp16)[name = string("op_5233_cast_fp16")];
            bool w_39_transpose_x_0 = const()[name = string("w_39_transpose_x_0"), val = bool(false)];
            bool w_39_transpose_y_0 = const()[name = string("w_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_39_cast_fp16 = matmul(transpose_x = w_39_transpose_x_0, transpose_y = w_39_transpose_y_0, x = Minv_159_cast_fp16, y = var_5233_cast_fp16)[name = string("w_39_cast_fp16")];
            bool var_5235_transpose_x_0 = const()[name = string("op_5235_transpose_x_0"), val = bool(false)];
            bool var_5235_transpose_y_0 = const()[name = string("op_5235_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5235_cast_fp16 = matmul(transpose_x = var_5235_transpose_x_0, transpose_y = var_5235_transpose_y_0, x = w_39_cast_fp16, y = S_29_cast_fp16)[name = string("op_5235_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_29_cast_fp16 = sub(x = u_39_cast_fp16, y = var_5235_cast_fp16)[name = string("vnew_29_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5240_cast_fp16 = mul(x = q_81_cast_fp16, y = var_5232_cast_fp16)[name = string("op_5240_cast_fp16")];
            bool var_5241_transpose_x_0 = const()[name = string("op_5241_transpose_x_0"), val = bool(false)];
            bool var_5241_transpose_y_0 = const()[name = string("op_5241_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5241_cast_fp16 = matmul(transpose_x = var_5241_transpose_x_0, transpose_y = var_5241_transpose_y_0, x = var_5240_cast_fp16, y = S_29_cast_fp16)[name = string("op_5241_cast_fp16")];
            bool var_5242_transpose_x_0 = const()[name = string("op_5242_transpose_x_0"), val = bool(false)];
            bool var_5242_transpose_y_0 = const()[name = string("op_5242_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5242_cast_fp16 = matmul(transpose_x = var_5242_transpose_x_0, transpose_y = var_5242_transpose_y_0, x = A_39_cast_fp16, y = vnew_29_cast_fp16)[name = string("op_5242_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_9_cast_fp16 = add(x = var_5241_cast_fp16, y = var_5242_cast_fp16)[name = string("oc_9_cast_fp16")];
            int32 var_5246 = const()[name = string("op_5246"), val = int32(2)];
            bool o_27_interleave_0 = const()[name = string("o_27_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 64, 128]> o_27_cast_fp16 = concat(axis = var_5246, interleave = o_27_interleave_0, values = (var_4789_cast_fp16, var_4936_cast_fp16, var_5090_cast_fp16, oc_9_cast_fp16))[name = string("o_27_cast_fp16")];
            tensor<int32, [4]> var_5252 = const()[name = string("op_5252"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_5258 = const()[name = string("op_5258"), val = tensor<int32, [4]>([3, 64, 4, 128])];
            tensor<fp16, [3, 64, 4, 128]> input_81_cast_fp16 = reshape(shape = var_5258, x = z_9_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> o_29_cast_fp16 = transpose(perm = var_5252, x = o_27_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [3, 64, 4, 128]> var_5260_cast_fp16 = mul(x = o_29_cast_fp16, y = o_29_cast_fp16)[name = string("op_5260_cast_fp16")];
            tensor<int32, [1]> var_5265_axes_0 = const()[name = string("op_5265_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5265_keep_dims_0 = const()[name = string("op_5265_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 4, 1]> var_5265_cast_fp16 = reduce_mean(axes = var_5265_axes_0, keep_dims = var_5265_keep_dims_0, x = var_5260_cast_fp16)[name = string("op_5265_cast_fp16")];
            fp16 var_5267_to_fp16 = const()[name = string("op_5267_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 64, 4, 1]> var_5268_cast_fp16 = add(x = var_5265_cast_fp16, y = var_5267_to_fp16)[name = string("op_5268_cast_fp16")];
            fp32 var_5269_epsilon_0 = const()[name = string("op_5269_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 4, 1]> var_5269_cast_fp16 = rsqrt(epsilon = var_5269_epsilon_0, x = var_5268_cast_fp16)[name = string("op_5269_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> var_5270_cast_fp16 = mul(x = o_29_cast_fp16, y = var_5269_cast_fp16)[name = string("op_5270_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> var_5272_cast_fp16 = silu(x = input_81_cast_fp16)[name = string("op_5272_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> o_31_cast_fp16 = mul(x = var_5270_cast_fp16, y = var_5272_cast_fp16)[name = string("o_31_cast_fp16")];
            tensor<int32, [3]> var_5277 = const()[name = string("op_5277"), val = tensor<int32, [3]>([3, 64, 512])];
            tensor<fp16, [3, 64, 512]> input_83_cast_fp16 = reshape(shape = var_5277, x = o_31_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [640, 512]> layers_5_out_weight_to_fp16 = const()[name = string("layers_5_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84536512)))];
            tensor<fp16, [3, 64, 640]> linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_out_weight_to_fp16, x = input_83_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [3, 64, 640]> x_57_cast_fp16 = add(x = x_53_cast_fp16, y = linear_21_cast_fp16)[name = string("x_57_cast_fp16")];
            fp16 var_5290_to_fp16 = const()[name = string("op_5290_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 64, 640]> x_59_cast_fp16 = mul(x = x_57_cast_fp16, y = var_5290_to_fp16)[name = string("x_59_cast_fp16")];
            fp16 var_5288_promoted_to_fp16 = const()[name = string("op_5288_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> var_5292_cast_fp16 = pow(x = x_59_cast_fp16, y = var_5288_promoted_to_fp16)[name = string("op_5292_cast_fp16")];
            tensor<int32, [1]> var_5294_axes_0 = const()[name = string("op_5294_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5294_keep_dims_0 = const()[name = string("op_5294_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 1]> var_5294_cast_fp16 = reduce_mean(axes = var_5294_axes_0, keep_dims = var_5294_keep_dims_0, x = var_5292_cast_fp16)[name = string("op_5294_cast_fp16")];
            fp16 var_5295_to_fp16 = const()[name = string("op_5295_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 64, 1]> var_5296_cast_fp16 = add(x = var_5294_cast_fp16, y = var_5295_to_fp16)[name = string("op_5296_cast_fp16")];
            fp32 var_5297_epsilon_0 = const()[name = string("op_5297_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 1]> var_5297_cast_fp16 = rsqrt(epsilon = var_5297_epsilon_0, x = var_5296_cast_fp16)[name = string("op_5297_cast_fp16")];
            tensor<fp16, [3, 64, 640]> var_5298_cast_fp16 = mul(x = x_59_cast_fp16, y = var_5297_cast_fp16)[name = string("op_5298_cast_fp16")];
            tensor<fp16, [640]> layers_5_n2_w1_to_fp16 = const()[name = string("layers_5_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85191936)))];
            tensor<fp16, [3, 64, 640]> input_85_cast_fp16 = mul(x = var_5298_cast_fp16, y = layers_5_n2_w1_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_5_gate_up_weight_to_fp16 = const()[name = string("layers_5_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85193280)))];
            tensor<fp16, [3, 64, 5120]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_5_gate_up_weight_to_fp16, x = input_85_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [3]> input_87_begin_0 = const()[name = string("input_87_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_87_end_0 = const()[name = string("input_87_end_0"), val = tensor<int32, [3]>([3, 64, 2560])];
            tensor<bool, [3]> input_87_end_mask_0 = const()[name = string("input_87_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 2560]> input_87_cast_fp16 = slice_by_index(begin = input_87_begin_0, end = input_87_end_0, end_mask = input_87_end_mask_0, x = linear_22_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [3, 64, 2560]> var_5308_cast_fp16 = silu(x = input_87_cast_fp16)[name = string("op_5308_cast_fp16")];
            tensor<int32, [3]> var_5313_begin_0 = const()[name = string("op_5313_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_5313_end_0 = const()[name = string("op_5313_end_0"), val = tensor<int32, [3]>([3, 64, 5120])];
            tensor<bool, [3]> var_5313_end_mask_0 = const()[name = string("op_5313_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 2560]> var_5313_cast_fp16 = slice_by_index(begin = var_5313_begin_0, end = var_5313_end_0, end_mask = var_5313_end_mask_0, x = linear_22_cast_fp16)[name = string("op_5313_cast_fp16")];
            tensor<fp16, [3, 64, 2560]> input_89_cast_fp16 = mul(x = var_5308_cast_fp16, y = var_5313_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_5_down_weight_to_fp16 = const()[name = string("layers_5_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91746944)))];
            tensor<fp16, [3, 64, 640]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_down_weight_to_fp16, x = input_89_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<fp16, [3, 64, 640]> x_61_cast_fp16 = add(x = x_57_cast_fp16, y = linear_23_cast_fp16)[name = string("x_61_cast_fp16")];
            fp16 var_5326_to_fp16 = const()[name = string("op_5326_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [3, 64, 640]> x_63_cast_fp16 = mul(x = x_61_cast_fp16, y = var_5326_to_fp16)[name = string("x_63_cast_fp16")];
            fp16 var_5324_promoted_to_fp16 = const()[name = string("op_5324_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> var_5328_cast_fp16 = pow(x = x_63_cast_fp16, y = var_5324_promoted_to_fp16)[name = string("op_5328_cast_fp16")];
            tensor<int32, [1]> var_5330_axes_0 = const()[name = string("op_5330_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5330_keep_dims_0 = const()[name = string("op_5330_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 1]> var_5330_cast_fp16 = reduce_mean(axes = var_5330_axes_0, keep_dims = var_5330_keep_dims_0, x = var_5328_cast_fp16)[name = string("op_5330_cast_fp16")];
            fp16 var_5331_to_fp16 = const()[name = string("op_5331_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 64, 1]> var_5332_cast_fp16 = add(x = var_5330_cast_fp16, y = var_5331_to_fp16)[name = string("op_5332_cast_fp16")];
            fp32 var_5333_epsilon_0 = const()[name = string("op_5333_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 1]> var_5333_cast_fp16 = rsqrt(epsilon = var_5333_epsilon_0, x = var_5332_cast_fp16)[name = string("op_5333_cast_fp16")];
            tensor<fp16, [3, 64, 640]> var_5334_cast_fp16 = mul(x = x_63_cast_fp16, y = var_5333_cast_fp16)[name = string("op_5334_cast_fp16")];
            tensor<fp16, [640]> layers_6_n1_w1_to_fp16 = const()[name = string("layers_6_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95023808)))];
            tensor<fp16, [3, 64, 640]> input_91_cast_fp16 = mul(x = var_5334_cast_fp16, y = layers_6_n1_w1_to_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_6_qkvzba_weight_to_fp16 = const()[name = string("layers_6_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95025152)))];
            tensor<fp16, [3, 64, 1544]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_qkvzba_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<int32, [3]> qkv_11_begin_0 = const()[name = string("qkv_11_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_11_end_0 = const()[name = string("qkv_11_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> qkv_11_end_mask_0 = const()[name = string("qkv_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 1024]> qkv_11_cast_fp16 = slice_by_index(begin = qkv_11_begin_0, end = qkv_11_end_0, end_mask = qkv_11_end_mask_0, x = linear_24_cast_fp16)[name = string("qkv_11_cast_fp16")];
            tensor<int32, [3]> z_begin_0 = const()[name = string("z_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_end_0 = const()[name = string("z_end_0"), val = tensor<int32, [3]>([3, 64, 1536])];
            tensor<bool, [3]> z_end_mask_0 = const()[name = string("z_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 512]> z_cast_fp16 = slice_by_index(begin = z_begin_0, end = z_end_0, end_mask = z_end_mask_0, x = linear_24_cast_fp16)[name = string("z_cast_fp16")];
            tensor<int32, [3]> b_begin_0 = const()[name = string("b_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_end_0 = const()[name = string("b_end_0"), val = tensor<int32, [3]>([3, 64, 1540])];
            tensor<bool, [3]> b_end_mask_0 = const()[name = string("b_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 4]> b_cast_fp16 = slice_by_index(begin = b_begin_0, end = b_end_0, end_mask = b_end_mask_0, x = linear_24_cast_fp16)[name = string("b_cast_fp16")];
            tensor<int32, [3]> a_begin_0 = const()[name = string("a_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_end_0 = const()[name = string("a_end_0"), val = tensor<int32, [3]>([3, 64, 1])];
            tensor<bool, [3]> a_end_mask_0 = const()[name = string("a_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 4]> a_cast_fp16 = slice_by_index(begin = a_begin_0, end = a_end_0, end_mask = a_end_mask_0, x = linear_24_cast_fp16)[name = string("a_cast_fp16")];
            int32 var_5373 = const()[name = string("op_5373"), val = int32(1)];
            bool xp_interleave_0 = const()[name = string("xp_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 67, 1024]> xp_cast_fp16 = concat(axis = var_5373, interleave = xp_interleave_0, values = (zpad_to_fp16, qkv_11_cast_fp16))[name = string("xp_cast_fp16")];
            tensor<int32, [3]> var_5384_begin_0 = const()[name = string("op_5384_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_5384_end_0 = const()[name = string("op_5384_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_5384_end_mask_0 = const()[name = string("op_5384_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 1024]> var_5384_cast_fp16 = slice_by_index(begin = var_5384_begin_0, end = var_5384_end_0, end_mask = var_5384_end_mask_0, x = xp_cast_fp16)[name = string("op_5384_cast_fp16")];
            tensor<fp16, [1024]> var_5392_to_fp16 = const()[name = string("op_5392_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97001536)))];
            tensor<fp16, [3, 64, 1024]> var_5393_cast_fp16 = mul(x = var_5384_cast_fp16, y = var_5392_to_fp16)[name = string("op_5393_cast_fp16")];
            tensor<int32, [3]> var_5403_begin_0 = const()[name = string("op_5403_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_5403_end_0 = const()[name = string("op_5403_end_0"), val = tensor<int32, [3]>([3, 66, 1024])];
            tensor<bool, [3]> var_5403_end_mask_0 = const()[name = string("op_5403_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_5403_cast_fp16 = slice_by_index(begin = var_5403_begin_0, end = var_5403_end_0, end_mask = var_5403_end_mask_0, x = xp_cast_fp16)[name = string("op_5403_cast_fp16")];
            tensor<fp16, [1024]> var_5411_to_fp16 = const()[name = string("op_5411_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97003648)))];
            tensor<fp16, [3, 64, 1024]> var_5412_cast_fp16 = mul(x = var_5403_cast_fp16, y = var_5411_to_fp16)[name = string("op_5412_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> var_5414_cast_fp16 = add(x = var_5393_cast_fp16, y = var_5412_cast_fp16)[name = string("op_5414_cast_fp16")];
            tensor<int32, [3]> var_5424_begin_0 = const()[name = string("op_5424_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_5424_end_0 = const()[name = string("op_5424_end_0"), val = tensor<int32, [3]>([3, 65, 1024])];
            tensor<bool, [3]> var_5424_end_mask_0 = const()[name = string("op_5424_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_5424_cast_fp16 = slice_by_index(begin = var_5424_begin_0, end = var_5424_end_0, end_mask = var_5424_end_mask_0, x = xp_cast_fp16)[name = string("op_5424_cast_fp16")];
            tensor<fp16, [1024]> var_5432_to_fp16 = const()[name = string("op_5432_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97005760)))];
            tensor<fp16, [3, 64, 1024]> var_5433_cast_fp16 = mul(x = var_5424_cast_fp16, y = var_5432_to_fp16)[name = string("op_5433_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> var_5435_cast_fp16 = add(x = var_5414_cast_fp16, y = var_5433_cast_fp16)[name = string("op_5435_cast_fp16")];
            tensor<int32, [3]> var_5445_begin_0 = const()[name = string("op_5445_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_5445_end_0 = const()[name = string("op_5445_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> var_5445_end_mask_0 = const()[name = string("op_5445_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 64, 1024]> var_5445_cast_fp16 = slice_by_index(begin = var_5445_begin_0, end = var_5445_end_0, end_mask = var_5445_end_mask_0, x = xp_cast_fp16)[name = string("op_5445_cast_fp16")];
            tensor<fp16, [1024]> var_5453_to_fp16 = const()[name = string("op_5453_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97007872)))];
            tensor<fp16, [3, 64, 1024]> var_5454_cast_fp16 = mul(x = var_5445_cast_fp16, y = var_5453_to_fp16)[name = string("op_5454_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> input_93_cast_fp16 = add(x = var_5435_cast_fp16, y = var_5454_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [3, 64, 1024]> c_cast_fp16 = silu(x = input_93_cast_fp16)[name = string("c_cast_fp16")];
            tensor<int32, [3]> var_5462_begin_0 = const()[name = string("op_5462_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_5462_end_0 = const()[name = string("op_5462_end_0"), val = tensor<int32, [3]>([3, 64, 256])];
            tensor<bool, [3]> var_5462_end_mask_0 = const()[name = string("op_5462_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 256]> var_5462_cast_fp16 = slice_by_index(begin = var_5462_begin_0, end = var_5462_end_0, end_mask = var_5462_end_mask_0, x = c_cast_fp16)[name = string("op_5462_cast_fp16")];
            tensor<int32, [4]> var_5467 = const()[name = string("op_5467"), val = tensor<int32, [4]>([3, 64, 2, 128])];
            tensor<fp16, [3, 64, 2, 128]> q_83_cast_fp16 = reshape(shape = var_5467, x = var_5462_cast_fp16)[name = string("q_83_cast_fp16")];
            tensor<int32, [3]> var_5473_begin_0 = const()[name = string("op_5473_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_5473_end_0 = const()[name = string("op_5473_end_0"), val = tensor<int32, [3]>([3, 64, 512])];
            tensor<bool, [3]> var_5473_end_mask_0 = const()[name = string("op_5473_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 256]> var_5473_cast_fp16 = slice_by_index(begin = var_5473_begin_0, end = var_5473_end_0, end_mask = var_5473_end_mask_0, x = c_cast_fp16)[name = string("op_5473_cast_fp16")];
            tensor<int32, [4]> var_5478 = const()[name = string("op_5478"), val = tensor<int32, [4]>([3, 64, 2, 128])];
            tensor<fp16, [3, 64, 2, 128]> k_85_cast_fp16 = reshape(shape = var_5478, x = var_5473_cast_fp16)[name = string("k_85_cast_fp16")];
            tensor<int32, [3]> var_5484_begin_0 = const()[name = string("op_5484_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_5484_end_0 = const()[name = string("op_5484_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> var_5484_end_mask_0 = const()[name = string("op_5484_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 512]> var_5484_cast_fp16 = slice_by_index(begin = var_5484_begin_0, end = var_5484_end_0, end_mask = var_5484_end_mask_0, x = c_cast_fp16)[name = string("op_5484_cast_fp16")];
            tensor<int32, [4]> var_5489 = const()[name = string("op_5489"), val = tensor<int32, [4]>([3, 64, 4, 128])];
            tensor<fp16, [3, 64, 4, 128]> var_5490_cast_fp16 = reshape(shape = var_5489, x = var_5484_cast_fp16)[name = string("op_5490_cast_fp16")];
            tensor<int32, [4]> var_5495 = const()[name = string("op_5495"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5497_to_fp16 = const()[name = string("op_5497_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 64, 2, 128]> q_85_cast_fp16 = mul(x = q_83_cast_fp16, y = var_5497_to_fp16)[name = string("q_85_cast_fp16")];
            fp16 var_5499_to_fp16 = const()[name = string("op_5499_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 64, 2, 128]> k_87_cast_fp16 = mul(x = k_85_cast_fp16, y = var_5499_to_fp16)[name = string("k_87_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_5501_cast_fp16 = mul(x = q_85_cast_fp16, y = q_85_cast_fp16)[name = string("op_5501_cast_fp16")];
            tensor<int32, [1]> var_5506_axes_0 = const()[name = string("op_5506_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5506_keep_dims_0 = const()[name = string("op_5506_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 2, 1]> var_5506_cast_fp16 = reduce_mean(axes = var_5506_axes_0, keep_dims = var_5506_keep_dims_0, x = var_5501_cast_fp16)[name = string("op_5506_cast_fp16")];
            fp16 var_5507_promoted_to_fp16 = const()[name = string("op_5507_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 64, 2, 1]> var_5508_cast_fp16 = mul(x = var_5506_cast_fp16, y = var_5507_promoted_to_fp16)[name = string("op_5508_cast_fp16")];
            fp16 var_5510_to_fp16 = const()[name = string("op_5510_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 64, 2, 1]> var_5511_cast_fp16 = add(x = var_5508_cast_fp16, y = var_5510_to_fp16)[name = string("op_5511_cast_fp16")];
            fp32 var_5512_epsilon_0 = const()[name = string("op_5512_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 2, 1]> var_5512_cast_fp16 = rsqrt(epsilon = var_5512_epsilon_0, x = var_5511_cast_fp16)[name = string("op_5512_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_5513_cast_fp16 = mul(x = q_85_cast_fp16, y = var_5512_cast_fp16)[name = string("op_5513_cast_fp16")];
            fp16 var_5514_to_fp16 = const()[name = string("op_5514_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 64, 2, 128]> q_87_cast_fp16 = mul(x = var_5513_cast_fp16, y = var_5514_to_fp16)[name = string("q_87_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> var_5516_cast_fp16 = mul(x = k_87_cast_fp16, y = k_87_cast_fp16)[name = string("op_5516_cast_fp16")];
            tensor<int32, [1]> var_5521_axes_0 = const()[name = string("op_5521_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5521_keep_dims_0 = const()[name = string("op_5521_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 2, 1]> var_5521_cast_fp16 = reduce_mean(axes = var_5521_axes_0, keep_dims = var_5521_keep_dims_0, x = var_5516_cast_fp16)[name = string("op_5521_cast_fp16")];
            fp16 var_5522_promoted_to_fp16 = const()[name = string("op_5522_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 64, 2, 1]> var_5523_cast_fp16 = mul(x = var_5521_cast_fp16, y = var_5522_promoted_to_fp16)[name = string("op_5523_cast_fp16")];
            fp16 var_5525_to_fp16 = const()[name = string("op_5525_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 64, 2, 1]> var_5526_cast_fp16 = add(x = var_5523_cast_fp16, y = var_5525_to_fp16)[name = string("op_5526_cast_fp16")];
            fp32 var_5527_epsilon_0 = const()[name = string("op_5527_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 2, 1]> var_5527_cast_fp16 = rsqrt(epsilon = var_5527_epsilon_0, x = var_5526_cast_fp16)[name = string("op_5527_cast_fp16")];
            tensor<fp16, [3, 64, 2, 128]> k_89_cast_fp16 = mul(x = k_87_cast_fp16, y = var_5527_cast_fp16)[name = string("k_89_cast_fp16")];
            tensor<int32, [4]> var_5533 = const()[name = string("op_5533"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_5546_axes_0 = const()[name = string("op_5546_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 64, 128]> var_5534_cast_fp16 = transpose(perm = var_5533, x = q_87_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [3, 2, 1, 64, 128]> var_5546_cast_fp16 = expand_dims(axes = var_5546_axes_0, x = var_5534_cast_fp16)[name = string("op_5546_cast_fp16")];
            tensor<int32, [5]> var_5554_reps_0 = const()[name = string("op_5554_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 64, 128]> var_5554_cast_fp16 = tile(reps = var_5554_reps_0, x = var_5546_cast_fp16)[name = string("op_5554_cast_fp16")];
            tensor<int32, [4]> var_5559 = const()[name = string("op_5559"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<fp16, [3, 4, 64, 128]> q_89_cast_fp16 = reshape(shape = var_5559, x = var_5554_cast_fp16)[name = string("q_89_cast_fp16")];
            tensor<int32, [4]> var_5565 = const()[name = string("op_5565"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_5578_axes_0 = const()[name = string("op_5578_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 64, 128]> var_5566_cast_fp16 = transpose(perm = var_5565, x = k_89_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [3, 2, 1, 64, 128]> var_5578_cast_fp16 = expand_dims(axes = var_5578_axes_0, x = var_5566_cast_fp16)[name = string("op_5578_cast_fp16")];
            tensor<int32, [5]> var_5586_reps_0 = const()[name = string("op_5586_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 64, 128]> var_5586_cast_fp16 = tile(reps = var_5586_reps_0, x = var_5578_cast_fp16)[name = string("op_5586_cast_fp16")];
            tensor<int32, [4]> var_5591 = const()[name = string("op_5591"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<fp16, [3, 4, 64, 128]> k_91_cast_fp16 = reshape(shape = var_5591, x = var_5586_cast_fp16)[name = string("k_91_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_5593_cast_fp16 = sigmoid(x = b_cast_fp16)[name = string("op_5593_cast_fp16")];
            tensor<int32, [3]> var_5597 = const()[name = string("op_5597"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_6_dt_bias_to_fp16 = const()[name = string("layers_6_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 64, 4]> input_95_cast_fp16 = add(x = a_cast_fp16, y = layers_6_dt_bias_to_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_5601_cast_fp16 = relu(x = input_95_cast_fp16)[name = string("op_5601_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_5602_cast_fp16 = abs(x = input_95_cast_fp16)[name = string("op_5602_cast_fp16")];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 64, 4]> var_5603_cast_fp16 = mul(x = var_5602_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_5603_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_5604_cast_fp16 = exp(x = var_5603_cast_fp16)[name = string("op_5604_cast_fp16")];
            fp16 var_5606_to_fp16 = const()[name = string("op_5606_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 64, 4]> var_5607_cast_fp16 = add(x = var_5604_cast_fp16, y = var_5606_to_fp16)[name = string("op_5607_cast_fp16")];
            fp32 var_5608_epsilon_0 = const()[name = string("op_5608_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 64, 4]> var_5608_cast_fp16 = log(epsilon = var_5608_epsilon_0, x = var_5607_cast_fp16)[name = string("op_5608_cast_fp16")];
            tensor<fp16, [3, 64, 4]> var_5610_cast_fp16 = add(x = var_5601_cast_fp16, y = var_5608_cast_fp16)[name = string("op_5610_cast_fp16")];
            tensor<fp16, [4]> layers_6_negA_to_fp16 = const()[name = string("layers_6_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.07cp+1, -0x1.eccp+3, -0x1.63p+3, -0x1.728p+1])];
            tensor<fp16, [3, 64, 4]> var_5611_cast_fp16 = mul(x = layers_6_negA_to_fp16, y = var_5610_cast_fp16)[name = string("op_5611_cast_fp16")];
            tensor<int32, [3]> var_5615 = const()[name = string("op_5615"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_5618_axes_0 = const()[name = string("op_5618_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 64]> beta_cast_fp16 = transpose(perm = var_5597, x = var_5593_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [3, 4, 64, 1]> var_5618_cast_fp16 = expand_dims(axes = var_5618_axes_0, x = beta_cast_fp16)[name = string("op_5618_cast_fp16")];
            fp16 var_5619_to_fp16 = const()[name = string("op_5619_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 64, 1]> var_5620_cast_fp16 = mul(x = var_5618_cast_fp16, y = var_5619_to_fp16)[name = string("op_5620_cast_fp16")];
            tensor<fp16, [3, 4, 64, 128]> v_15_cast_fp16 = transpose(perm = var_5495, x = var_5490_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [3, 4, 64, 128]> vb_51_cast_fp16 = mul(x = v_15_cast_fp16, y = var_5620_cast_fp16)[name = string("vb_51_cast_fp16")];
            tensor<int32, [4]> q_91_begin_0 = const()[name = string("q_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_91_end_0 = const()[name = string("q_91_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_91_end_mask_0 = const()[name = string("q_91_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, x = q_89_cast_fp16)[name = string("q_91_cast_fp16")];
            tensor<int32, [4]> k_93_begin_0 = const()[name = string("k_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_93_end_0 = const()[name = string("k_93_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_93_end_mask_0 = const()[name = string("k_93_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, x = k_91_cast_fp16)[name = string("k_93_cast_fp16")];
            tensor<fp16, [3, 4, 64, 128]> var_5654_cast_fp16 = mul(x = k_91_cast_fp16, y = var_5618_cast_fp16)[name = string("op_5654_cast_fp16")];
            tensor<int32, [4]> kb_41_begin_0 = const()[name = string("kb_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_41_end_0 = const()[name = string("kb_41_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_41_end_mask_0 = const()[name = string("kb_41_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_41_cast_fp16 = slice_by_index(begin = kb_41_begin_0, end = kb_41_end_0, end_mask = kb_41_end_mask_0, x = var_5654_cast_fp16)[name = string("kb_41_cast_fp16")];
            tensor<int32, [4]> vb_53_begin_0 = const()[name = string("vb_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_53_end_0 = const()[name = string("vb_53_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_53_end_mask_0 = const()[name = string("vb_53_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_53_cast_fp16 = slice_by_index(begin = vb_53_begin_0, end = vb_53_end_0, end_mask = vb_53_end_mask_0, x = vb_51_cast_fp16)[name = string("vb_53_cast_fp16")];
            tensor<int32, [3]> gdec_53_begin_0 = const()[name = string("gdec_53_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_53_end_0 = const()[name = string("gdec_53_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_53_end_mask_0 = const()[name = string("gdec_53_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 64]> gdec_51_cast_fp16 = transpose(perm = var_5615, x = var_5611_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [3, 4, 16]> gdec_53_cast_fp16 = slice_by_index(begin = gdec_53_begin_0, end = gdec_53_end_0, end_mask = gdec_53_end_mask_0, x = gdec_51_cast_fp16)[name = string("gdec_53_cast_fp16")];
            tensor<int32, [2]> var_5702 = const()[name = string("op_5702"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_5703_cast_fp16 = reshape(shape = var_5702, x = gdec_53_cast_fp16)[name = string("op_5703_cast_fp16")];
            tensor<fp16, [12, 256]> var_5704_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5703_cast_fp16)[name = string("op_5704_cast_fp16")];
            tensor<int32, [4]> var_5709 = const()[name = string("op_5709"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_41_cast_fp16 = reshape(shape = var_5709, x = var_5704_cast_fp16)[name = string("P_41_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5711_cast_fp16 = exp(x = P_41_cast_fp16)[name = string("op_5711_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_41_cast_fp16 = mul(x = var_5711_cast_fp16, y = tril_incl_to_fp16)[name = string("D_41_cast_fp16")];
            bool kk_41_transpose_x_1 = const()[name = string("kk_41_transpose_x_1"), val = bool(false)];
            bool kk_41_transpose_y_1 = const()[name = string("kk_41_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_41_cast_fp16 = matmul(transpose_x = kk_41_transpose_x_1, transpose_y = kk_41_transpose_y_1, x = kb_41_cast_fp16, y = k_93_cast_fp16)[name = string("kk_41_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_5717_cast_fp16 = mul(x = kk_41_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_5717_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5718_cast_fp16 = mul(x = var_5717_cast_fp16, y = D_41_cast_fp16)[name = string("op_5718_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_41_cast_fp16 = mul(x = var_5718_cast_fp16, y = cum_tail_to_fp16)[name = string("N_41_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5720_cast_fp16 = mul(x = D_41_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5720_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_41_cast_fp16 = mul(x = kk_41_cast_fp16, y = var_5720_cast_fp16)[name = string("Na_41_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_161_cast_fp16 = add(x = eye_to_fp16, y = N_41_cast_fp16)[name = string("Minv_161_cast_fp16")];
            bool Np_121_transpose_x_0 = const()[name = string("Np_121_transpose_x_0"), val = bool(false)];
            bool Np_121_transpose_y_0 = const()[name = string("Np_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_121_cast_fp16 = matmul(transpose_x = Np_121_transpose_x_0, transpose_y = Np_121_transpose_y_0, x = N_41_cast_fp16, y = Na_41_cast_fp16)[name = string("Np_121_cast_fp16")];
            bool Npa_81_transpose_x_0 = const()[name = string("Npa_81_transpose_x_0"), val = bool(false)];
            bool Npa_81_transpose_y_0 = const()[name = string("Npa_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_81_cast_fp16 = matmul(transpose_x = Npa_81_transpose_x_0, transpose_y = Npa_81_transpose_y_0, x = Na_41_cast_fp16, y = N_41_cast_fp16)[name = string("Npa_81_cast_fp16")];
            bool var_5726_transpose_x_0 = const()[name = string("op_5726_transpose_x_0"), val = bool(false)];
            bool var_5726_transpose_y_0 = const()[name = string("op_5726_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5726_cast_fp16 = matmul(transpose_x = var_5726_transpose_x_0, transpose_y = var_5726_transpose_y_0, x = Minv_161_cast_fp16, y = Np_121_cast_fp16)[name = string("op_5726_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_163_cast_fp16 = add(x = Minv_161_cast_fp16, y = var_5726_cast_fp16)[name = string("Minv_163_cast_fp16")];
            bool Np_123_transpose_x_0 = const()[name = string("Np_123_transpose_x_0"), val = bool(false)];
            bool Np_123_transpose_y_0 = const()[name = string("Np_123_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_123_cast_fp16 = matmul(transpose_x = Np_123_transpose_x_0, transpose_y = Np_123_transpose_y_0, x = Np_121_cast_fp16, y = Npa_81_cast_fp16)[name = string("Np_123_cast_fp16")];
            bool Npa_83_transpose_x_0 = const()[name = string("Npa_83_transpose_x_0"), val = bool(false)];
            bool Npa_83_transpose_y_0 = const()[name = string("Npa_83_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_83_cast_fp16 = matmul(transpose_x = Npa_83_transpose_x_0, transpose_y = Npa_83_transpose_y_0, x = Npa_81_cast_fp16, y = Np_121_cast_fp16)[name = string("Npa_83_cast_fp16")];
            bool var_5731_transpose_x_0 = const()[name = string("op_5731_transpose_x_0"), val = bool(false)];
            bool var_5731_transpose_y_0 = const()[name = string("op_5731_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5731_cast_fp16 = matmul(transpose_x = var_5731_transpose_x_0, transpose_y = var_5731_transpose_y_0, x = Minv_163_cast_fp16, y = Np_123_cast_fp16)[name = string("op_5731_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_165_cast_fp16 = add(x = Minv_163_cast_fp16, y = var_5731_cast_fp16)[name = string("Minv_165_cast_fp16")];
            bool Np_125_transpose_x_0 = const()[name = string("Np_125_transpose_x_0"), val = bool(false)];
            bool Np_125_transpose_y_0 = const()[name = string("Np_125_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_125_cast_fp16 = matmul(transpose_x = Np_125_transpose_x_0, transpose_y = Np_125_transpose_y_0, x = Np_123_cast_fp16, y = Npa_83_cast_fp16)[name = string("Np_125_cast_fp16")];
            bool var_5735_transpose_x_0 = const()[name = string("op_5735_transpose_x_0"), val = bool(false)];
            bool var_5735_transpose_y_0 = const()[name = string("op_5735_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5735_cast_fp16 = matmul(transpose_x = var_5735_transpose_x_0, transpose_y = var_5735_transpose_y_0, x = Minv_165_cast_fp16, y = Np_125_cast_fp16)[name = string("op_5735_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_167_cast_fp16 = add(x = Minv_165_cast_fp16, y = var_5735_cast_fp16)[name = string("Minv_167_cast_fp16")];
            bool u_41_transpose_x_0 = const()[name = string("u_41_transpose_x_0"), val = bool(false)];
            bool u_41_transpose_y_0 = const()[name = string("u_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_41_cast_fp16 = matmul(transpose_x = u_41_transpose_x_0, transpose_y = u_41_transpose_y_0, x = Minv_167_cast_fp16, y = vb_53_cast_fp16)[name = string("u_41_cast_fp16")];
            bool var_5742_transpose_x_1 = const()[name = string("op_5742_transpose_x_1"), val = bool(false)];
            bool var_5742_transpose_y_1 = const()[name = string("op_5742_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_5742_cast_fp16 = matmul(transpose_x = var_5742_transpose_x_1, transpose_y = var_5742_transpose_y_1, x = q_91_cast_fp16, y = k_93_cast_fp16)[name = string("op_5742_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_41_cast_fp16 = mul(x = var_5742_cast_fp16, y = D_41_cast_fp16)[name = string("A_41_cast_fp16")];
            bool var_5744_transpose_x_0 = const()[name = string("op_5744_transpose_x_0"), val = bool(false)];
            bool var_5744_transpose_y_0 = const()[name = string("op_5744_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5744_cast_fp16 = matmul(transpose_x = var_5744_transpose_x_0, transpose_y = var_5744_transpose_y_0, x = A_41_cast_fp16, y = u_41_cast_fp16)[name = string("op_5744_cast_fp16")];
            tensor<fp16, [16]> var_5745_bias_0_to_fp16 = const()[name = string("op_5745_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_5745_cast_fp16 = linear(bias = var_5745_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_53_cast_fp16)[name = string("op_5745_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_31_cast_fp16 = exp(x = var_5745_cast_fp16)[name = string("tail_31_cast_fp16")];
            tensor<int32, [1]> var_5748_axes_0 = const()[name = string("op_5748_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_5748_cast_fp16 = expand_dims(axes = var_5748_axes_0, x = tail_31_cast_fp16)[name = string("op_5748_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5749_cast_fp16 = mul(x = k_93_cast_fp16, y = var_5748_cast_fp16)[name = string("op_5749_cast_fp16")];
            bool S_31_transpose_x_1 = const()[name = string("S_31_transpose_x_1"), val = bool(true)];
            bool S_31_transpose_y_1 = const()[name = string("S_31_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_31_cast_fp16 = matmul(transpose_x = S_31_transpose_x_1, transpose_y = S_31_transpose_y_1, x = var_5749_cast_fp16, y = u_41_cast_fp16)[name = string("S_31_cast_fp16")];
            tensor<int32, [4]> q_93_begin_0 = const()[name = string("q_93_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_93_end_0 = const()[name = string("q_93_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> q_93_end_mask_0 = const()[name = string("q_93_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_93_cast_fp16 = slice_by_index(begin = q_93_begin_0, end = q_93_end_0, end_mask = q_93_end_mask_0, x = q_89_cast_fp16)[name = string("q_93_cast_fp16")];
            tensor<int32, [4]> k_95_begin_0 = const()[name = string("k_95_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_95_end_0 = const()[name = string("k_95_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> k_95_end_mask_0 = const()[name = string("k_95_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_95_cast_fp16 = slice_by_index(begin = k_95_begin_0, end = k_95_end_0, end_mask = k_95_end_mask_0, x = k_91_cast_fp16)[name = string("k_95_cast_fp16")];
            tensor<int32, [4]> kb_43_begin_0 = const()[name = string("kb_43_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_43_end_0 = const()[name = string("kb_43_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> kb_43_end_mask_0 = const()[name = string("kb_43_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_43_cast_fp16 = slice_by_index(begin = kb_43_begin_0, end = kb_43_end_0, end_mask = kb_43_end_mask_0, x = var_5654_cast_fp16)[name = string("kb_43_cast_fp16")];
            tensor<int32, [4]> vb_55_begin_0 = const()[name = string("vb_55_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_55_end_0 = const()[name = string("vb_55_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> vb_55_end_mask_0 = const()[name = string("vb_55_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_55_cast_fp16 = slice_by_index(begin = vb_55_begin_0, end = vb_55_end_0, end_mask = vb_55_end_mask_0, x = vb_51_cast_fp16)[name = string("vb_55_cast_fp16")];
            tensor<int32, [3]> gdec_55_begin_0 = const()[name = string("gdec_55_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_55_end_0 = const()[name = string("gdec_55_end_0"), val = tensor<int32, [3]>([3, 4, 32])];
            tensor<bool, [3]> gdec_55_end_mask_0 = const()[name = string("gdec_55_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_55_cast_fp16 = slice_by_index(begin = gdec_55_begin_0, end = gdec_55_end_0, end_mask = gdec_55_end_mask_0, x = gdec_51_cast_fp16)[name = string("gdec_55_cast_fp16")];
            tensor<int32, [2]> var_5834 = const()[name = string("op_5834"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_5835_cast_fp16 = reshape(shape = var_5834, x = gdec_55_cast_fp16)[name = string("op_5835_cast_fp16")];
            tensor<fp16, [12, 256]> var_5836_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5835_cast_fp16)[name = string("op_5836_cast_fp16")];
            tensor<int32, [4]> var_5841 = const()[name = string("op_5841"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_43_cast_fp16 = reshape(shape = var_5841, x = var_5836_cast_fp16)[name = string("P_43_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5843_cast_fp16 = exp(x = P_43_cast_fp16)[name = string("op_5843_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_43_cast_fp16 = mul(x = var_5843_cast_fp16, y = tril_incl_to_fp16)[name = string("D_43_cast_fp16")];
            bool kk_43_transpose_x_1 = const()[name = string("kk_43_transpose_x_1"), val = bool(false)];
            bool kk_43_transpose_y_1 = const()[name = string("kk_43_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_43_cast_fp16 = matmul(transpose_x = kk_43_transpose_x_1, transpose_y = kk_43_transpose_y_1, x = kb_43_cast_fp16, y = k_95_cast_fp16)[name = string("kk_43_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_5849_cast_fp16 = mul(x = kk_43_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_5849_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5850_cast_fp16 = mul(x = var_5849_cast_fp16, y = D_43_cast_fp16)[name = string("op_5850_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_43_cast_fp16 = mul(x = var_5850_cast_fp16, y = cum_tail_to_fp16)[name = string("N_43_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5852_cast_fp16 = mul(x = D_43_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5852_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_43_cast_fp16 = mul(x = kk_43_cast_fp16, y = var_5852_cast_fp16)[name = string("Na_43_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_169_cast_fp16 = add(x = eye_to_fp16, y = N_43_cast_fp16)[name = string("Minv_169_cast_fp16")];
            bool Np_127_transpose_x_0 = const()[name = string("Np_127_transpose_x_0"), val = bool(false)];
            bool Np_127_transpose_y_0 = const()[name = string("Np_127_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_127_cast_fp16 = matmul(transpose_x = Np_127_transpose_x_0, transpose_y = Np_127_transpose_y_0, x = N_43_cast_fp16, y = Na_43_cast_fp16)[name = string("Np_127_cast_fp16")];
            bool Npa_85_transpose_x_0 = const()[name = string("Npa_85_transpose_x_0"), val = bool(false)];
            bool Npa_85_transpose_y_0 = const()[name = string("Npa_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_85_cast_fp16 = matmul(transpose_x = Npa_85_transpose_x_0, transpose_y = Npa_85_transpose_y_0, x = Na_43_cast_fp16, y = N_43_cast_fp16)[name = string("Npa_85_cast_fp16")];
            bool var_5858_transpose_x_0 = const()[name = string("op_5858_transpose_x_0"), val = bool(false)];
            bool var_5858_transpose_y_0 = const()[name = string("op_5858_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5858_cast_fp16 = matmul(transpose_x = var_5858_transpose_x_0, transpose_y = var_5858_transpose_y_0, x = Minv_169_cast_fp16, y = Np_127_cast_fp16)[name = string("op_5858_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_171_cast_fp16 = add(x = Minv_169_cast_fp16, y = var_5858_cast_fp16)[name = string("Minv_171_cast_fp16")];
            bool Np_129_transpose_x_0 = const()[name = string("Np_129_transpose_x_0"), val = bool(false)];
            bool Np_129_transpose_y_0 = const()[name = string("Np_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_129_cast_fp16 = matmul(transpose_x = Np_129_transpose_x_0, transpose_y = Np_129_transpose_y_0, x = Np_127_cast_fp16, y = Npa_85_cast_fp16)[name = string("Np_129_cast_fp16")];
            bool Npa_87_transpose_x_0 = const()[name = string("Npa_87_transpose_x_0"), val = bool(false)];
            bool Npa_87_transpose_y_0 = const()[name = string("Npa_87_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_87_cast_fp16 = matmul(transpose_x = Npa_87_transpose_x_0, transpose_y = Npa_87_transpose_y_0, x = Npa_85_cast_fp16, y = Np_127_cast_fp16)[name = string("Npa_87_cast_fp16")];
            bool var_5863_transpose_x_0 = const()[name = string("op_5863_transpose_x_0"), val = bool(false)];
            bool var_5863_transpose_y_0 = const()[name = string("op_5863_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5863_cast_fp16 = matmul(transpose_x = var_5863_transpose_x_0, transpose_y = var_5863_transpose_y_0, x = Minv_171_cast_fp16, y = Np_129_cast_fp16)[name = string("op_5863_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_173_cast_fp16 = add(x = Minv_171_cast_fp16, y = var_5863_cast_fp16)[name = string("Minv_173_cast_fp16")];
            bool Np_131_transpose_x_0 = const()[name = string("Np_131_transpose_x_0"), val = bool(false)];
            bool Np_131_transpose_y_0 = const()[name = string("Np_131_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_131_cast_fp16 = matmul(transpose_x = Np_131_transpose_x_0, transpose_y = Np_131_transpose_y_0, x = Np_129_cast_fp16, y = Npa_87_cast_fp16)[name = string("Np_131_cast_fp16")];
            bool var_5867_transpose_x_0 = const()[name = string("op_5867_transpose_x_0"), val = bool(false)];
            bool var_5867_transpose_y_0 = const()[name = string("op_5867_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5867_cast_fp16 = matmul(transpose_x = var_5867_transpose_x_0, transpose_y = var_5867_transpose_y_0, x = Minv_173_cast_fp16, y = Np_131_cast_fp16)[name = string("op_5867_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_175_cast_fp16 = add(x = Minv_173_cast_fp16, y = var_5867_cast_fp16)[name = string("Minv_175_cast_fp16")];
            bool u_43_transpose_x_0 = const()[name = string("u_43_transpose_x_0"), val = bool(false)];
            bool u_43_transpose_y_0 = const()[name = string("u_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_43_cast_fp16 = matmul(transpose_x = u_43_transpose_x_0, transpose_y = u_43_transpose_y_0, x = Minv_175_cast_fp16, y = vb_55_cast_fp16)[name = string("u_43_cast_fp16")];
            bool var_5874_transpose_x_1 = const()[name = string("op_5874_transpose_x_1"), val = bool(false)];
            bool var_5874_transpose_y_1 = const()[name = string("op_5874_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_5874_cast_fp16 = matmul(transpose_x = var_5874_transpose_x_1, transpose_y = var_5874_transpose_y_1, x = q_93_cast_fp16, y = k_95_cast_fp16)[name = string("op_5874_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_43_cast_fp16 = mul(x = var_5874_cast_fp16, y = D_43_cast_fp16)[name = string("A_43_cast_fp16")];
            tensor<fp16, [16]> var_5876_bias_0_to_fp16 = const()[name = string("op_5876_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_5876_cast_fp16 = linear(bias = var_5876_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_55_cast_fp16)[name = string("op_5876_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_31_cast_fp16 = exp(x = var_5876_cast_fp16)[name = string("cum_31_cast_fp16")];
            tensor<int32, [1]> var_5879_axes_0 = const()[name = string("op_5879_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_5879_cast_fp16 = expand_dims(axes = var_5879_axes_0, x = cum_31_cast_fp16)[name = string("op_5879_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5880_cast_fp16 = mul(x = kb_43_cast_fp16, y = var_5879_cast_fp16)[name = string("op_5880_cast_fp16")];
            bool w_43_transpose_x_0 = const()[name = string("w_43_transpose_x_0"), val = bool(false)];
            bool w_43_transpose_y_0 = const()[name = string("w_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_43_cast_fp16 = matmul(transpose_x = w_43_transpose_x_0, transpose_y = w_43_transpose_y_0, x = Minv_175_cast_fp16, y = var_5880_cast_fp16)[name = string("w_43_cast_fp16")];
            bool var_5882_transpose_x_0 = const()[name = string("op_5882_transpose_x_0"), val = bool(false)];
            bool var_5882_transpose_y_0 = const()[name = string("op_5882_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5882_cast_fp16 = matmul(transpose_x = var_5882_transpose_x_0, transpose_y = var_5882_transpose_y_0, x = w_43_cast_fp16, y = S_31_cast_fp16)[name = string("op_5882_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_31_cast_fp16 = sub(x = u_43_cast_fp16, y = var_5882_cast_fp16)[name = string("vnew_31_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5887_cast_fp16 = mul(x = q_93_cast_fp16, y = var_5879_cast_fp16)[name = string("op_5887_cast_fp16")];
            bool var_5888_transpose_x_0 = const()[name = string("op_5888_transpose_x_0"), val = bool(false)];
            bool var_5888_transpose_y_0 = const()[name = string("op_5888_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5888_cast_fp16 = matmul(transpose_x = var_5888_transpose_x_0, transpose_y = var_5888_transpose_y_0, x = var_5887_cast_fp16, y = S_31_cast_fp16)[name = string("op_5888_cast_fp16")];
            bool var_5889_transpose_x_0 = const()[name = string("op_5889_transpose_x_0"), val = bool(false)];
            bool var_5889_transpose_y_0 = const()[name = string("op_5889_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5889_cast_fp16 = matmul(transpose_x = var_5889_transpose_x_0, transpose_y = var_5889_transpose_y_0, x = A_43_cast_fp16, y = vnew_31_cast_fp16)[name = string("op_5889_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5891_cast_fp16 = add(x = var_5888_cast_fp16, y = var_5889_cast_fp16)[name = string("op_5891_cast_fp16")];
            tensor<fp16, [16]> var_5892_bias_0_to_fp16 = const()[name = string("op_5892_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_5892_cast_fp16 = linear(bias = var_5892_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_55_cast_fp16)[name = string("op_5892_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_33_cast_fp16 = exp(x = var_5892_cast_fp16)[name = string("tail_33_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_65_to_fp16 = const()[name = string("transpose_65_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_5894_bias_0_to_fp16 = const()[name = string("op_5894_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_5894_cast_fp16 = linear(bias = var_5894_bias_0_to_fp16, weight = transpose_65_to_fp16, x = gdec_55_cast_fp16)[name = string("op_5894_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_5895_cast_fp16 = exp(x = var_5894_cast_fp16)[name = string("op_5895_cast_fp16")];
            tensor<int32, [1]> last_21_axes_0 = const()[name = string("last_21_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_21_cast_fp16 = expand_dims(axes = last_21_axes_0, x = var_5895_cast_fp16)[name = string("last_21_cast_fp16")];
            tensor<int32, [1]> var_5899_axes_0 = const()[name = string("op_5899_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_5899_cast_fp16 = expand_dims(axes = var_5899_axes_0, x = tail_33_cast_fp16)[name = string("op_5899_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5900_cast_fp16 = mul(x = k_95_cast_fp16, y = var_5899_cast_fp16)[name = string("op_5900_cast_fp16")];
            bool upd_21_transpose_x_1 = const()[name = string("upd_21_transpose_x_1"), val = bool(true)];
            bool upd_21_transpose_y_1 = const()[name = string("upd_21_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_21_cast_fp16 = matmul(transpose_x = upd_21_transpose_x_1, transpose_y = upd_21_transpose_y_1, x = var_5900_cast_fp16, y = vnew_31_cast_fp16)[name = string("upd_21_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_5905_cast_fp16 = mul(x = S_31_cast_fp16, y = last_21_cast_fp16)[name = string("op_5905_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_33_cast_fp16 = add(x = var_5905_cast_fp16, y = upd_21_cast_fp16)[name = string("S_33_cast_fp16")];
            tensor<int32, [4]> q_95_begin_0 = const()[name = string("q_95_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_95_end_0 = const()[name = string("q_95_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> q_95_end_mask_0 = const()[name = string("q_95_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_95_cast_fp16 = slice_by_index(begin = q_95_begin_0, end = q_95_end_0, end_mask = q_95_end_mask_0, x = q_89_cast_fp16)[name = string("q_95_cast_fp16")];
            tensor<int32, [4]> k_97_begin_0 = const()[name = string("k_97_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_97_end_0 = const()[name = string("k_97_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> k_97_end_mask_0 = const()[name = string("k_97_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_97_cast_fp16 = slice_by_index(begin = k_97_begin_0, end = k_97_end_0, end_mask = k_97_end_mask_0, x = k_91_cast_fp16)[name = string("k_97_cast_fp16")];
            tensor<int32, [4]> kb_45_begin_0 = const()[name = string("kb_45_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_45_end_0 = const()[name = string("kb_45_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> kb_45_end_mask_0 = const()[name = string("kb_45_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_45_cast_fp16 = slice_by_index(begin = kb_45_begin_0, end = kb_45_end_0, end_mask = kb_45_end_mask_0, x = var_5654_cast_fp16)[name = string("kb_45_cast_fp16")];
            tensor<int32, [4]> vb_57_begin_0 = const()[name = string("vb_57_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_57_end_0 = const()[name = string("vb_57_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> vb_57_end_mask_0 = const()[name = string("vb_57_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_57_cast_fp16 = slice_by_index(begin = vb_57_begin_0, end = vb_57_end_0, end_mask = vb_57_end_mask_0, x = vb_51_cast_fp16)[name = string("vb_57_cast_fp16")];
            tensor<int32, [3]> gdec_57_begin_0 = const()[name = string("gdec_57_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_57_end_0 = const()[name = string("gdec_57_end_0"), val = tensor<int32, [3]>([3, 4, 48])];
            tensor<bool, [3]> gdec_57_end_mask_0 = const()[name = string("gdec_57_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_57_cast_fp16 = slice_by_index(begin = gdec_57_begin_0, end = gdec_57_end_0, end_mask = gdec_57_end_mask_0, x = gdec_51_cast_fp16)[name = string("gdec_57_cast_fp16")];
            tensor<int32, [2]> var_5988 = const()[name = string("op_5988"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_5989_cast_fp16 = reshape(shape = var_5988, x = gdec_57_cast_fp16)[name = string("op_5989_cast_fp16")];
            tensor<fp16, [12, 256]> var_5990_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5989_cast_fp16)[name = string("op_5990_cast_fp16")];
            tensor<int32, [4]> var_5995 = const()[name = string("op_5995"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_45_cast_fp16 = reshape(shape = var_5995, x = var_5990_cast_fp16)[name = string("P_45_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5997_cast_fp16 = exp(x = P_45_cast_fp16)[name = string("op_5997_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_45_cast_fp16 = mul(x = var_5997_cast_fp16, y = tril_incl_to_fp16)[name = string("D_45_cast_fp16")];
            bool kk_45_transpose_x_1 = const()[name = string("kk_45_transpose_x_1"), val = bool(false)];
            bool kk_45_transpose_y_1 = const()[name = string("kk_45_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_45_cast_fp16 = matmul(transpose_x = kk_45_transpose_x_1, transpose_y = kk_45_transpose_y_1, x = kb_45_cast_fp16, y = k_97_cast_fp16)[name = string("kk_45_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_6003_cast_fp16 = mul(x = kk_45_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_6003_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6004_cast_fp16 = mul(x = var_6003_cast_fp16, y = D_45_cast_fp16)[name = string("op_6004_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_45_cast_fp16 = mul(x = var_6004_cast_fp16, y = cum_tail_to_fp16)[name = string("N_45_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6006_cast_fp16 = mul(x = D_45_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_6006_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_45_cast_fp16 = mul(x = kk_45_cast_fp16, y = var_6006_cast_fp16)[name = string("Na_45_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_177_cast_fp16 = add(x = eye_to_fp16, y = N_45_cast_fp16)[name = string("Minv_177_cast_fp16")];
            bool Np_133_transpose_x_0 = const()[name = string("Np_133_transpose_x_0"), val = bool(false)];
            bool Np_133_transpose_y_0 = const()[name = string("Np_133_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_133_cast_fp16 = matmul(transpose_x = Np_133_transpose_x_0, transpose_y = Np_133_transpose_y_0, x = N_45_cast_fp16, y = Na_45_cast_fp16)[name = string("Np_133_cast_fp16")];
            bool Npa_89_transpose_x_0 = const()[name = string("Npa_89_transpose_x_0"), val = bool(false)];
            bool Npa_89_transpose_y_0 = const()[name = string("Npa_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_89_cast_fp16 = matmul(transpose_x = Npa_89_transpose_x_0, transpose_y = Npa_89_transpose_y_0, x = Na_45_cast_fp16, y = N_45_cast_fp16)[name = string("Npa_89_cast_fp16")];
            bool var_6012_transpose_x_0 = const()[name = string("op_6012_transpose_x_0"), val = bool(false)];
            bool var_6012_transpose_y_0 = const()[name = string("op_6012_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6012_cast_fp16 = matmul(transpose_x = var_6012_transpose_x_0, transpose_y = var_6012_transpose_y_0, x = Minv_177_cast_fp16, y = Np_133_cast_fp16)[name = string("op_6012_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_179_cast_fp16 = add(x = Minv_177_cast_fp16, y = var_6012_cast_fp16)[name = string("Minv_179_cast_fp16")];
            bool Np_135_transpose_x_0 = const()[name = string("Np_135_transpose_x_0"), val = bool(false)];
            bool Np_135_transpose_y_0 = const()[name = string("Np_135_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_135_cast_fp16 = matmul(transpose_x = Np_135_transpose_x_0, transpose_y = Np_135_transpose_y_0, x = Np_133_cast_fp16, y = Npa_89_cast_fp16)[name = string("Np_135_cast_fp16")];
            bool Npa_91_transpose_x_0 = const()[name = string("Npa_91_transpose_x_0"), val = bool(false)];
            bool Npa_91_transpose_y_0 = const()[name = string("Npa_91_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_91_cast_fp16 = matmul(transpose_x = Npa_91_transpose_x_0, transpose_y = Npa_91_transpose_y_0, x = Npa_89_cast_fp16, y = Np_133_cast_fp16)[name = string("Npa_91_cast_fp16")];
            bool var_6017_transpose_x_0 = const()[name = string("op_6017_transpose_x_0"), val = bool(false)];
            bool var_6017_transpose_y_0 = const()[name = string("op_6017_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6017_cast_fp16 = matmul(transpose_x = var_6017_transpose_x_0, transpose_y = var_6017_transpose_y_0, x = Minv_179_cast_fp16, y = Np_135_cast_fp16)[name = string("op_6017_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_181_cast_fp16 = add(x = Minv_179_cast_fp16, y = var_6017_cast_fp16)[name = string("Minv_181_cast_fp16")];
            bool Np_137_transpose_x_0 = const()[name = string("Np_137_transpose_x_0"), val = bool(false)];
            bool Np_137_transpose_y_0 = const()[name = string("Np_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_137_cast_fp16 = matmul(transpose_x = Np_137_transpose_x_0, transpose_y = Np_137_transpose_y_0, x = Np_135_cast_fp16, y = Npa_91_cast_fp16)[name = string("Np_137_cast_fp16")];
            bool var_6021_transpose_x_0 = const()[name = string("op_6021_transpose_x_0"), val = bool(false)];
            bool var_6021_transpose_y_0 = const()[name = string("op_6021_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6021_cast_fp16 = matmul(transpose_x = var_6021_transpose_x_0, transpose_y = var_6021_transpose_y_0, x = Minv_181_cast_fp16, y = Np_137_cast_fp16)[name = string("op_6021_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_183_cast_fp16 = add(x = Minv_181_cast_fp16, y = var_6021_cast_fp16)[name = string("Minv_183_cast_fp16")];
            bool u_45_transpose_x_0 = const()[name = string("u_45_transpose_x_0"), val = bool(false)];
            bool u_45_transpose_y_0 = const()[name = string("u_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_45_cast_fp16 = matmul(transpose_x = u_45_transpose_x_0, transpose_y = u_45_transpose_y_0, x = Minv_183_cast_fp16, y = vb_57_cast_fp16)[name = string("u_45_cast_fp16")];
            bool var_6028_transpose_x_1 = const()[name = string("op_6028_transpose_x_1"), val = bool(false)];
            bool var_6028_transpose_y_1 = const()[name = string("op_6028_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_6028_cast_fp16 = matmul(transpose_x = var_6028_transpose_x_1, transpose_y = var_6028_transpose_y_1, x = q_95_cast_fp16, y = k_97_cast_fp16)[name = string("op_6028_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_45_cast_fp16 = mul(x = var_6028_cast_fp16, y = D_45_cast_fp16)[name = string("A_45_cast_fp16")];
            tensor<fp16, [16]> var_6030_bias_0_to_fp16 = const()[name = string("op_6030_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_6030_cast_fp16 = linear(bias = var_6030_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_57_cast_fp16)[name = string("op_6030_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_33_cast_fp16 = exp(x = var_6030_cast_fp16)[name = string("cum_33_cast_fp16")];
            tensor<int32, [1]> var_6033_axes_0 = const()[name = string("op_6033_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_6033_cast_fp16 = expand_dims(axes = var_6033_axes_0, x = cum_33_cast_fp16)[name = string("op_6033_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6034_cast_fp16 = mul(x = kb_45_cast_fp16, y = var_6033_cast_fp16)[name = string("op_6034_cast_fp16")];
            bool w_45_transpose_x_0 = const()[name = string("w_45_transpose_x_0"), val = bool(false)];
            bool w_45_transpose_y_0 = const()[name = string("w_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_45_cast_fp16 = matmul(transpose_x = w_45_transpose_x_0, transpose_y = w_45_transpose_y_0, x = Minv_183_cast_fp16, y = var_6034_cast_fp16)[name = string("w_45_cast_fp16")];
            bool var_6036_transpose_x_0 = const()[name = string("op_6036_transpose_x_0"), val = bool(false)];
            bool var_6036_transpose_y_0 = const()[name = string("op_6036_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6036_cast_fp16 = matmul(transpose_x = var_6036_transpose_x_0, transpose_y = var_6036_transpose_y_0, x = w_45_cast_fp16, y = S_33_cast_fp16)[name = string("op_6036_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_33_cast_fp16 = sub(x = u_45_cast_fp16, y = var_6036_cast_fp16)[name = string("vnew_33_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6041_cast_fp16 = mul(x = q_95_cast_fp16, y = var_6033_cast_fp16)[name = string("op_6041_cast_fp16")];
            bool var_6042_transpose_x_0 = const()[name = string("op_6042_transpose_x_0"), val = bool(false)];
            bool var_6042_transpose_y_0 = const()[name = string("op_6042_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6042_cast_fp16 = matmul(transpose_x = var_6042_transpose_x_0, transpose_y = var_6042_transpose_y_0, x = var_6041_cast_fp16, y = S_33_cast_fp16)[name = string("op_6042_cast_fp16")];
            bool var_6043_transpose_x_0 = const()[name = string("op_6043_transpose_x_0"), val = bool(false)];
            bool var_6043_transpose_y_0 = const()[name = string("op_6043_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6043_cast_fp16 = matmul(transpose_x = var_6043_transpose_x_0, transpose_y = var_6043_transpose_y_0, x = A_45_cast_fp16, y = vnew_33_cast_fp16)[name = string("op_6043_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6045_cast_fp16 = add(x = var_6042_cast_fp16, y = var_6043_cast_fp16)[name = string("op_6045_cast_fp16")];
            tensor<fp16, [16]> var_6046_bias_0_to_fp16 = const()[name = string("op_6046_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_6046_cast_fp16 = linear(bias = var_6046_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_57_cast_fp16)[name = string("op_6046_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_cast_fp16 = exp(x = var_6046_cast_fp16)[name = string("tail_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_69_to_fp16 = const()[name = string("transpose_69_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_6048_bias_0_to_fp16 = const()[name = string("op_6048_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_6048_cast_fp16 = linear(bias = var_6048_bias_0_to_fp16, weight = transpose_69_to_fp16, x = gdec_57_cast_fp16)[name = string("op_6048_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_6049_cast_fp16 = exp(x = var_6048_cast_fp16)[name = string("op_6049_cast_fp16")];
            tensor<int32, [1]> last_axes_0 = const()[name = string("last_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_cast_fp16 = expand_dims(axes = last_axes_0, x = var_6049_cast_fp16)[name = string("last_cast_fp16")];
            tensor<int32, [1]> var_6053_axes_0 = const()[name = string("op_6053_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_6053_cast_fp16 = expand_dims(axes = var_6053_axes_0, x = tail_cast_fp16)[name = string("op_6053_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6054_cast_fp16 = mul(x = k_97_cast_fp16, y = var_6053_cast_fp16)[name = string("op_6054_cast_fp16")];
            bool upd_transpose_x_1 = const()[name = string("upd_transpose_x_1"), val = bool(true)];
            bool upd_transpose_y_1 = const()[name = string("upd_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_cast_fp16 = matmul(transpose_x = upd_transpose_x_1, transpose_y = upd_transpose_y_1, x = var_6054_cast_fp16, y = vnew_33_cast_fp16)[name = string("upd_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_6059_cast_fp16 = mul(x = S_33_cast_fp16, y = last_cast_fp16)[name = string("op_6059_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_cast_fp16 = add(x = var_6059_cast_fp16, y = upd_cast_fp16)[name = string("S_cast_fp16")];
            tensor<int32, [4]> q_97_begin_0 = const()[name = string("q_97_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_97_end_0 = const()[name = string("q_97_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_97_end_mask_0 = const()[name = string("q_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, x = q_89_cast_fp16)[name = string("q_97_cast_fp16")];
            tensor<int32, [4]> k_99_begin_0 = const()[name = string("k_99_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_99_end_0 = const()[name = string("k_99_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_99_end_mask_0 = const()[name = string("k_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_99_cast_fp16 = slice_by_index(begin = k_99_begin_0, end = k_99_end_0, end_mask = k_99_end_mask_0, x = k_91_cast_fp16)[name = string("k_99_cast_fp16")];
            tensor<int32, [4]> kb_begin_0 = const()[name = string("kb_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_end_0 = const()[name = string("kb_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_end_mask_0 = const()[name = string("kb_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_cast_fp16 = slice_by_index(begin = kb_begin_0, end = kb_end_0, end_mask = kb_end_mask_0, x = var_5654_cast_fp16)[name = string("kb_cast_fp16")];
            tensor<int32, [4]> vb_begin_0 = const()[name = string("vb_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_end_0 = const()[name = string("vb_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_end_mask_0 = const()[name = string("vb_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_cast_fp16 = slice_by_index(begin = vb_begin_0, end = vb_end_0, end_mask = vb_end_mask_0, x = vb_51_cast_fp16)[name = string("vb_cast_fp16")];
            tensor<int32, [3]> gdec_begin_0 = const()[name = string("gdec_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_end_0 = const()[name = string("gdec_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_end_mask_0 = const()[name = string("gdec_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_cast_fp16 = slice_by_index(begin = gdec_begin_0, end = gdec_end_0, end_mask = gdec_end_mask_0, x = gdec_51_cast_fp16)[name = string("gdec_cast_fp16")];
            tensor<int32, [2]> var_6142 = const()[name = string("op_6142"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_6143_cast_fp16 = reshape(shape = var_6142, x = gdec_cast_fp16)[name = string("op_6143_cast_fp16")];
            tensor<fp16, [12, 256]> var_6144_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_6143_cast_fp16)[name = string("op_6144_cast_fp16")];
            tensor<int32, [4]> var_6149 = const()[name = string("op_6149"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_cast_fp16 = reshape(shape = var_6149, x = var_6144_cast_fp16)[name = string("P_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6151_cast_fp16 = exp(x = P_cast_fp16)[name = string("op_6151_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_cast_fp16 = mul(x = var_6151_cast_fp16, y = tril_incl_to_fp16)[name = string("D_cast_fp16")];
            bool kk_transpose_x_1 = const()[name = string("kk_transpose_x_1"), val = bool(false)];
            bool kk_transpose_y_1 = const()[name = string("kk_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_cast_fp16 = matmul(transpose_x = kk_transpose_x_1, transpose_y = kk_transpose_y_1, x = kb_cast_fp16, y = k_99_cast_fp16)[name = string("kk_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_6157_cast_fp16 = mul(x = kk_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_6157_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6158_cast_fp16 = mul(x = var_6157_cast_fp16, y = D_cast_fp16)[name = string("op_6158_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_cast_fp16 = mul(x = var_6158_cast_fp16, y = cum_tail_to_fp16)[name = string("N_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6160_cast_fp16 = mul(x = D_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_6160_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_cast_fp16 = mul(x = kk_cast_fp16, y = var_6160_cast_fp16)[name = string("Na_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_185_cast_fp16 = add(x = eye_to_fp16, y = N_cast_fp16)[name = string("Minv_185_cast_fp16")];
            bool Np_139_transpose_x_0 = const()[name = string("Np_139_transpose_x_0"), val = bool(false)];
            bool Np_139_transpose_y_0 = const()[name = string("Np_139_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_139_cast_fp16 = matmul(transpose_x = Np_139_transpose_x_0, transpose_y = Np_139_transpose_y_0, x = N_cast_fp16, y = Na_cast_fp16)[name = string("Np_139_cast_fp16")];
            bool Npa_93_transpose_x_0 = const()[name = string("Npa_93_transpose_x_0"), val = bool(false)];
            bool Npa_93_transpose_y_0 = const()[name = string("Npa_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_93_cast_fp16 = matmul(transpose_x = Npa_93_transpose_x_0, transpose_y = Npa_93_transpose_y_0, x = Na_cast_fp16, y = N_cast_fp16)[name = string("Npa_93_cast_fp16")];
            bool var_6166_transpose_x_0 = const()[name = string("op_6166_transpose_x_0"), val = bool(false)];
            bool var_6166_transpose_y_0 = const()[name = string("op_6166_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6166_cast_fp16 = matmul(transpose_x = var_6166_transpose_x_0, transpose_y = var_6166_transpose_y_0, x = Minv_185_cast_fp16, y = Np_139_cast_fp16)[name = string("op_6166_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_187_cast_fp16 = add(x = Minv_185_cast_fp16, y = var_6166_cast_fp16)[name = string("Minv_187_cast_fp16")];
            bool Np_141_transpose_x_0 = const()[name = string("Np_141_transpose_x_0"), val = bool(false)];
            bool Np_141_transpose_y_0 = const()[name = string("Np_141_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_141_cast_fp16 = matmul(transpose_x = Np_141_transpose_x_0, transpose_y = Np_141_transpose_y_0, x = Np_139_cast_fp16, y = Npa_93_cast_fp16)[name = string("Np_141_cast_fp16")];
            bool Npa_transpose_x_0 = const()[name = string("Npa_transpose_x_0"), val = bool(false)];
            bool Npa_transpose_y_0 = const()[name = string("Npa_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_cast_fp16 = matmul(transpose_x = Npa_transpose_x_0, transpose_y = Npa_transpose_y_0, x = Npa_93_cast_fp16, y = Np_139_cast_fp16)[name = string("Npa_cast_fp16")];
            bool var_6171_transpose_x_0 = const()[name = string("op_6171_transpose_x_0"), val = bool(false)];
            bool var_6171_transpose_y_0 = const()[name = string("op_6171_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6171_cast_fp16 = matmul(transpose_x = var_6171_transpose_x_0, transpose_y = var_6171_transpose_y_0, x = Minv_187_cast_fp16, y = Np_141_cast_fp16)[name = string("op_6171_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_189_cast_fp16 = add(x = Minv_187_cast_fp16, y = var_6171_cast_fp16)[name = string("Minv_189_cast_fp16")];
            bool Np_transpose_x_0 = const()[name = string("Np_transpose_x_0"), val = bool(false)];
            bool Np_transpose_y_0 = const()[name = string("Np_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_cast_fp16 = matmul(transpose_x = Np_transpose_x_0, transpose_y = Np_transpose_y_0, x = Np_141_cast_fp16, y = Npa_cast_fp16)[name = string("Np_cast_fp16")];
            bool var_6175_transpose_x_0 = const()[name = string("op_6175_transpose_x_0"), val = bool(false)];
            bool var_6175_transpose_y_0 = const()[name = string("op_6175_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6175_cast_fp16 = matmul(transpose_x = var_6175_transpose_x_0, transpose_y = var_6175_transpose_y_0, x = Minv_189_cast_fp16, y = Np_cast_fp16)[name = string("op_6175_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_cast_fp16 = add(x = Minv_189_cast_fp16, y = var_6175_cast_fp16)[name = string("Minv_cast_fp16")];
            bool u_transpose_x_0 = const()[name = string("u_transpose_x_0"), val = bool(false)];
            bool u_transpose_y_0 = const()[name = string("u_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_cast_fp16 = matmul(transpose_x = u_transpose_x_0, transpose_y = u_transpose_y_0, x = Minv_cast_fp16, y = vb_cast_fp16)[name = string("u_cast_fp16")];
            bool var_6182_transpose_x_1 = const()[name = string("op_6182_transpose_x_1"), val = bool(false)];
            bool var_6182_transpose_y_1 = const()[name = string("op_6182_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_6182_cast_fp16 = matmul(transpose_x = var_6182_transpose_x_1, transpose_y = var_6182_transpose_y_1, x = q_97_cast_fp16, y = k_99_cast_fp16)[name = string("op_6182_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_cast_fp16 = mul(x = var_6182_cast_fp16, y = D_cast_fp16)[name = string("A_cast_fp16")];
            tensor<fp16, [16]> var_6184_bias_0_to_fp16 = const()[name = string("op_6184_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_6184_cast_fp16 = linear(bias = var_6184_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_cast_fp16)[name = string("op_6184_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_cast_fp16 = exp(x = var_6184_cast_fp16)[name = string("cum_cast_fp16")];
            tensor<int32, [1]> var_6187_axes_0 = const()[name = string("op_6187_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_6187_cast_fp16 = expand_dims(axes = var_6187_axes_0, x = cum_cast_fp16)[name = string("op_6187_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6188_cast_fp16 = mul(x = kb_cast_fp16, y = var_6187_cast_fp16)[name = string("op_6188_cast_fp16")];
            bool w_transpose_x_0 = const()[name = string("w_transpose_x_0"), val = bool(false)];
            bool w_transpose_y_0 = const()[name = string("w_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_cast_fp16 = matmul(transpose_x = w_transpose_x_0, transpose_y = w_transpose_y_0, x = Minv_cast_fp16, y = var_6188_cast_fp16)[name = string("w_cast_fp16")];
            bool var_6190_transpose_x_0 = const()[name = string("op_6190_transpose_x_0"), val = bool(false)];
            bool var_6190_transpose_y_0 = const()[name = string("op_6190_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6190_cast_fp16 = matmul(transpose_x = var_6190_transpose_x_0, transpose_y = var_6190_transpose_y_0, x = w_cast_fp16, y = S_cast_fp16)[name = string("op_6190_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_cast_fp16 = sub(x = u_cast_fp16, y = var_6190_cast_fp16)[name = string("vnew_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6195_cast_fp16 = mul(x = q_97_cast_fp16, y = var_6187_cast_fp16)[name = string("op_6195_cast_fp16")];
            bool var_6196_transpose_x_0 = const()[name = string("op_6196_transpose_x_0"), val = bool(false)];
            bool var_6196_transpose_y_0 = const()[name = string("op_6196_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6196_cast_fp16 = matmul(transpose_x = var_6196_transpose_x_0, transpose_y = var_6196_transpose_y_0, x = var_6195_cast_fp16, y = S_cast_fp16)[name = string("op_6196_cast_fp16")];
            bool var_6197_transpose_x_0 = const()[name = string("op_6197_transpose_x_0"), val = bool(false)];
            bool var_6197_transpose_y_0 = const()[name = string("op_6197_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6197_cast_fp16 = matmul(transpose_x = var_6197_transpose_x_0, transpose_y = var_6197_transpose_y_0, x = A_cast_fp16, y = vnew_cast_fp16)[name = string("op_6197_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_cast_fp16 = add(x = var_6196_cast_fp16, y = var_6197_cast_fp16)[name = string("oc_cast_fp16")];
            int32 var_6201 = const()[name = string("op_6201"), val = int32(2)];
            bool o_33_interleave_0 = const()[name = string("o_33_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 64, 128]> o_33_cast_fp16 = concat(axis = var_6201, interleave = o_33_interleave_0, values = (var_5744_cast_fp16, var_5891_cast_fp16, var_6045_cast_fp16, oc_cast_fp16))[name = string("o_33_cast_fp16")];
            tensor<int32, [4]> var_6207 = const()[name = string("op_6207"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_6213 = const()[name = string("op_6213"), val = tensor<int32, [4]>([3, 64, 4, 128])];
            tensor<fp16, [3, 64, 4, 128]> input_97_cast_fp16 = reshape(shape = var_6213, x = z_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> o_35_cast_fp16 = transpose(perm = var_6207, x = o_33_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [3, 64, 4, 128]> var_6215_cast_fp16 = mul(x = o_35_cast_fp16, y = o_35_cast_fp16)[name = string("op_6215_cast_fp16")];
            tensor<int32, [1]> var_6220_axes_0 = const()[name = string("op_6220_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6220_keep_dims_0 = const()[name = string("op_6220_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 4, 1]> var_6220_cast_fp16 = reduce_mean(axes = var_6220_axes_0, keep_dims = var_6220_keep_dims_0, x = var_6215_cast_fp16)[name = string("op_6220_cast_fp16")];
            fp16 var_6222_to_fp16 = const()[name = string("op_6222_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 64, 4, 1]> var_6223_cast_fp16 = add(x = var_6220_cast_fp16, y = var_6222_to_fp16)[name = string("op_6223_cast_fp16")];
            fp32 var_6224_epsilon_0 = const()[name = string("op_6224_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 4, 1]> var_6224_cast_fp16 = rsqrt(epsilon = var_6224_epsilon_0, x = var_6223_cast_fp16)[name = string("op_6224_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> var_6225_cast_fp16 = mul(x = o_35_cast_fp16, y = var_6224_cast_fp16)[name = string("op_6225_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> var_6227_cast_fp16 = silu(x = input_97_cast_fp16)[name = string("op_6227_cast_fp16")];
            tensor<fp16, [3, 64, 4, 128]> o_37_cast_fp16 = mul(x = var_6225_cast_fp16, y = var_6227_cast_fp16)[name = string("o_37_cast_fp16")];
            tensor<int32, [3]> var_6232 = const()[name = string("op_6232"), val = tensor<int32, [3]>([3, 64, 512])];
            tensor<fp16, [3, 64, 512]> input_99_cast_fp16 = reshape(shape = var_6232, x = o_37_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [640, 512]> layers_6_out_weight_to_fp16 = const()[name = string("layers_6_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97009984)))];
            tensor<fp16, [3, 64, 640]> linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_out_weight_to_fp16, x = input_99_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [3, 64, 640]> x_65_cast_fp16 = add(x = x_61_cast_fp16, y = linear_25_cast_fp16)[name = string("x_65_cast_fp16")];
            fp16 var_6245_to_fp16 = const()[name = string("op_6245_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [3, 64, 640]> x_67_cast_fp16 = mul(x = x_65_cast_fp16, y = var_6245_to_fp16)[name = string("x_67_cast_fp16")];
            fp16 var_6243_promoted_to_fp16 = const()[name = string("op_6243_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> var_6247_cast_fp16 = pow(x = x_67_cast_fp16, y = var_6243_promoted_to_fp16)[name = string("op_6247_cast_fp16")];
            tensor<int32, [1]> var_6249_axes_0 = const()[name = string("op_6249_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6249_keep_dims_0 = const()[name = string("op_6249_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 1]> var_6249_cast_fp16 = reduce_mean(axes = var_6249_axes_0, keep_dims = var_6249_keep_dims_0, x = var_6247_cast_fp16)[name = string("op_6249_cast_fp16")];
            fp16 var_6250_to_fp16 = const()[name = string("op_6250_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 64, 1]> var_6251_cast_fp16 = add(x = var_6249_cast_fp16, y = var_6250_to_fp16)[name = string("op_6251_cast_fp16")];
            fp32 var_6252_epsilon_0 = const()[name = string("op_6252_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 1]> var_6252_cast_fp16 = rsqrt(epsilon = var_6252_epsilon_0, x = var_6251_cast_fp16)[name = string("op_6252_cast_fp16")];
            tensor<fp16, [3, 64, 640]> var_6253_cast_fp16 = mul(x = x_67_cast_fp16, y = var_6252_cast_fp16)[name = string("op_6253_cast_fp16")];
            tensor<fp16, [640]> layers_6_n2_w1_to_fp16 = const()[name = string("layers_6_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97665408)))];
            tensor<fp16, [3, 64, 640]> input_101_cast_fp16 = mul(x = var_6253_cast_fp16, y = layers_6_n2_w1_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_6_gate_up_weight_to_fp16 = const()[name = string("layers_6_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97666752)))];
            tensor<fp16, [3, 64, 5120]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_6_gate_up_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<int32, [3]> input_103_begin_0 = const()[name = string("input_103_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_103_end_0 = const()[name = string("input_103_end_0"), val = tensor<int32, [3]>([3, 64, 2560])];
            tensor<bool, [3]> input_103_end_mask_0 = const()[name = string("input_103_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 2560]> input_103_cast_fp16 = slice_by_index(begin = input_103_begin_0, end = input_103_end_0, end_mask = input_103_end_mask_0, x = linear_26_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [3, 64, 2560]> var_6263_cast_fp16 = silu(x = input_103_cast_fp16)[name = string("op_6263_cast_fp16")];
            tensor<int32, [3]> var_6268_begin_0 = const()[name = string("op_6268_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_6268_end_0 = const()[name = string("op_6268_end_0"), val = tensor<int32, [3]>([3, 64, 5120])];
            tensor<bool, [3]> var_6268_end_mask_0 = const()[name = string("op_6268_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 2560]> var_6268_cast_fp16 = slice_by_index(begin = var_6268_begin_0, end = var_6268_end_0, end_mask = var_6268_end_mask_0, x = linear_26_cast_fp16)[name = string("op_6268_cast_fp16")];
            tensor<fp16, [3, 64, 2560]> input_105_cast_fp16 = mul(x = var_6263_cast_fp16, y = var_6268_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_6_down_weight_to_fp16 = const()[name = string("layers_6_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104220416)))];
            tensor<fp16, [3, 64, 640]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_down_weight_to_fp16, x = input_105_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [3, 64, 640]> x_69_cast_fp16 = add(x = x_65_cast_fp16, y = linear_27_cast_fp16)[name = string("x_69_cast_fp16")];
            fp16 var_6281_to_fp16 = const()[name = string("op_6281_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [3, 64, 640]> x_71_cast_fp16 = mul(x = x_69_cast_fp16, y = var_6281_to_fp16)[name = string("x_71_cast_fp16")];
            fp16 var_6279_promoted_to_fp16 = const()[name = string("op_6279_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> var_6283_cast_fp16 = pow(x = x_71_cast_fp16, y = var_6279_promoted_to_fp16)[name = string("op_6283_cast_fp16")];
            tensor<int32, [1]> var_6285_axes_0 = const()[name = string("op_6285_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6285_keep_dims_0 = const()[name = string("op_6285_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 1]> var_6285_cast_fp16 = reduce_mean(axes = var_6285_axes_0, keep_dims = var_6285_keep_dims_0, x = var_6283_cast_fp16)[name = string("op_6285_cast_fp16")];
            fp16 var_6286_to_fp16 = const()[name = string("op_6286_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 64, 1]> var_6287_cast_fp16 = add(x = var_6285_cast_fp16, y = var_6286_to_fp16)[name = string("op_6287_cast_fp16")];
            fp32 var_6288_epsilon_0 = const()[name = string("op_6288_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 1]> var_6288_cast_fp16 = rsqrt(epsilon = var_6288_epsilon_0, x = var_6287_cast_fp16)[name = string("op_6288_cast_fp16")];
            tensor<fp16, [3, 64, 640]> var_6289_cast_fp16 = mul(x = x_71_cast_fp16, y = var_6288_cast_fp16)[name = string("op_6289_cast_fp16")];
            tensor<fp16, [640]> layers_7_n1_w1_to_fp16 = const()[name = string("layers_7_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107497280)))];
            tensor<fp16, [3, 64, 640]> input_107_cast_fp16 = mul(x = var_6289_cast_fp16, y = layers_7_n1_w1_to_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [1280, 640]> layers_7_qkv_weight_to_fp16 = const()[name = string("layers_7_qkv_weight_to_fp16"), val = tensor<fp16, [1280, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107498624)))];
            tensor<fp16, [3, 64, 1280]> linear_28_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = layers_7_qkv_weight_to_fp16, x = input_107_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [3]> var_6298_begin_0 = const()[name = string("op_6298_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_6298_end_0 = const()[name = string("op_6298_end_0"), val = tensor<int32, [3]>([3, 64, 1024])];
            tensor<bool, [3]> var_6298_end_mask_0 = const()[name = string("op_6298_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 1024]> var_6298_cast_fp16 = slice_by_index(begin = var_6298_begin_0, end = var_6298_end_0, end_mask = var_6298_end_mask_0, x = linear_28_cast_fp16)[name = string("op_6298_cast_fp16")];
            tensor<int32, [4]> var_6303 = const()[name = string("op_6303"), val = tensor<int32, [4]>([3, 64, 8, 128])];
            tensor<fp16, [3, 64, 8, 128]> qg_cast_fp16 = reshape(shape = var_6303, x = var_6298_cast_fp16)[name = string("qg_cast_fp16")];
            tensor<int32, [4]> x_73_begin_0 = const()[name = string("x_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_73_end_0 = const()[name = string("x_73_end_0"), val = tensor<int32, [4]>([3, 64, 8, 64])];
            tensor<bool, [4]> x_73_end_mask_0 = const()[name = string("x_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 64, 8, 64]> x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, x = qg_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<int32, [4]> var_6314_begin_0 = const()[name = string("op_6314_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> var_6314_end_0 = const()[name = string("op_6314_end_0"), val = tensor<int32, [4]>([3, 64, 8, 128])];
            tensor<bool, [4]> var_6314_end_mask_0 = const()[name = string("op_6314_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 64, 8, 64]> var_6314_cast_fp16 = slice_by_index(begin = var_6314_begin_0, end = var_6314_end_0, end_mask = var_6314_end_mask_0, x = qg_cast_fp16)[name = string("op_6314_cast_fp16")];
            tensor<int32, [3]> var_6318 = const()[name = string("op_6318"), val = tensor<int32, [3]>([3, 64, 512])];
            tensor<fp16, [3, 64, 512]> gate_cast_fp16 = reshape(shape = var_6318, x = var_6314_cast_fp16)[name = string("gate_cast_fp16")];
            tensor<int32, [3]> var_6324_begin_0 = const()[name = string("op_6324_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> var_6324_end_0 = const()[name = string("op_6324_end_0"), val = tensor<int32, [3]>([3, 64, 1152])];
            tensor<bool, [3]> var_6324_end_mask_0 = const()[name = string("op_6324_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 128]> var_6324_cast_fp16 = slice_by_index(begin = var_6324_begin_0, end = var_6324_end_0, end_mask = var_6324_end_mask_0, x = linear_28_cast_fp16)[name = string("op_6324_cast_fp16")];
            tensor<int32, [4]> var_6329 = const()[name = string("op_6329"), val = tensor<int32, [4]>([3, 64, 2, 64])];
            tensor<fp16, [3, 64, 2, 64]> x_79_cast_fp16 = reshape(shape = var_6329, x = var_6324_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<int32, [3]> var_6335_begin_0 = const()[name = string("op_6335_begin_0"), val = tensor<int32, [3]>([0, 0, 1152])];
            tensor<int32, [3]> var_6335_end_0 = const()[name = string("op_6335_end_0"), val = tensor<int32, [3]>([3, 64, 1280])];
            tensor<bool, [3]> var_6335_end_mask_0 = const()[name = string("op_6335_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 128]> var_6335_cast_fp16 = slice_by_index(begin = var_6335_begin_0, end = var_6335_end_0, end_mask = var_6335_end_mask_0, x = linear_28_cast_fp16)[name = string("op_6335_cast_fp16")];
            tensor<int32, [4]> var_6340 = const()[name = string("op_6340"), val = tensor<int32, [4]>([3, 64, 2, 64])];
            tensor<fp16, [3, 64, 2, 64]> var_6341_cast_fp16 = reshape(shape = var_6340, x = var_6335_cast_fp16)[name = string("op_6341_cast_fp16")];
            tensor<int32, [4]> var_6346 = const()[name = string("op_6346"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6352_promoted_to_fp16 = const()[name = string("op_6352_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 8, 64]> var_6356_cast_fp16 = pow(x = x_73_cast_fp16, y = var_6352_promoted_to_fp16)[name = string("op_6356_cast_fp16")];
            tensor<int32, [1]> var_6358_axes_0 = const()[name = string("op_6358_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6358_keep_dims_0 = const()[name = string("op_6358_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 8, 1]> var_6358_cast_fp16 = reduce_mean(axes = var_6358_axes_0, keep_dims = var_6358_keep_dims_0, x = var_6356_cast_fp16)[name = string("op_6358_cast_fp16")];
            fp16 var_6359_to_fp16 = const()[name = string("op_6359_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 64, 8, 1]> var_6360_cast_fp16 = add(x = var_6358_cast_fp16, y = var_6359_to_fp16)[name = string("op_6360_cast_fp16")];
            fp32 var_6361_epsilon_0 = const()[name = string("op_6361_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 8, 1]> var_6361_cast_fp16 = rsqrt(epsilon = var_6361_epsilon_0, x = var_6360_cast_fp16)[name = string("op_6361_cast_fp16")];
            tensor<fp16, [3, 64, 8, 64]> var_6362_cast_fp16 = mul(x = x_73_cast_fp16, y = var_6361_cast_fp16)[name = string("op_6362_cast_fp16")];
            tensor<fp16, [64]> layers_7_qn_w1_to_fp16 = const()[name = string("layers_7_qn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137088)))];
            tensor<fp16, [3, 64, 8, 64]> var_6363_cast_fp16 = mul(x = var_6362_cast_fp16, y = layers_7_qn_w1_to_fp16)[name = string("op_6363_cast_fp16")];
            tensor<int32, [4]> var_6368 = const()[name = string("op_6368"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_5_begin_0 = const()[name = string("r_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_5_end_0 = const()[name = string("r_5_end_0"), val = tensor<int32, [4]>([3, 8, 64, 16])];
            tensor<bool, [4]> r_5_end_mask_0 = const()[name = string("r_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 8, 64, 64]> x_77_cast_fp16 = transpose(perm = var_6368, x = var_6363_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [3, 8, 64, 16]> r_5_cast_fp16 = slice_by_index(begin = r_5_begin_0, end = r_5_end_0, end_mask = r_5_end_mask_0, x = x_77_cast_fp16)[name = string("r_5_cast_fp16")];
            tensor<int32, [4]> p_5_begin_0 = const()[name = string("p_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_5_end_0 = const()[name = string("p_5_end_0"), val = tensor<int32, [4]>([3, 8, 64, 64])];
            tensor<bool, [4]> p_5_end_mask_0 = const()[name = string("p_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 8, 64, 48]> p_5_cast_fp16 = slice_by_index(begin = p_5_begin_0, end = p_5_end_0, end_mask = p_5_end_mask_0, x = x_77_cast_fp16)[name = string("p_5_cast_fp16")];
            tensor<int32, [4]> var_6384_begin_0 = const()[name = string("op_6384_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_6384_end_0 = const()[name = string("op_6384_end_0"), val = tensor<int32, [4]>([3, 8, 64, 16])];
            tensor<bool, [4]> var_6384_end_mask_0 = const()[name = string("op_6384_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 8, 64, 8]> var_6384_cast_fp16 = slice_by_index(begin = var_6384_begin_0, end = var_6384_end_0, end_mask = var_6384_end_mask_0, x = r_5_cast_fp16)[name = string("op_6384_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 8, 64, 8]> var_6385_cast_fp16 = mul(x = var_6384_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_6385_cast_fp16")];
            tensor<int32, [4]> var_6390_begin_0 = const()[name = string("op_6390_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6390_end_0 = const()[name = string("op_6390_end_0"), val = tensor<int32, [4]>([3, 8, 64, 8])];
            tensor<bool, [4]> var_6390_end_mask_0 = const()[name = string("op_6390_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 8, 64, 8]> var_6390_cast_fp16 = slice_by_index(begin = var_6390_begin_0, end = var_6390_end_0, end_mask = var_6390_end_mask_0, x = r_5_cast_fp16)[name = string("op_6390_cast_fp16")];
            int32 var_6392 = const()[name = string("op_6392"), val = int32(-1)];
            bool rh_5_interleave_0 = const()[name = string("rh_5_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 8, 64, 16]> rh_5_cast_fp16 = concat(axis = var_6392, interleave = rh_5_interleave_0, values = (var_6385_cast_fp16, var_6390_cast_fp16))[name = string("rh_5_cast_fp16")];
            tensor<fp16, [3, 8, 64, 16]> var_6394_cast_fp16 = mul(x = r_5_cast_fp16, y = cos_to_fp16)[name = string("op_6394_cast_fp16")];
            tensor<fp16, [3, 8, 64, 16]> var_6395_cast_fp16 = mul(x = rh_5_cast_fp16, y = sin_to_fp16)[name = string("op_6395_cast_fp16")];
            tensor<fp16, [3, 8, 64, 16]> var_6397_cast_fp16 = add(x = var_6394_cast_fp16, y = var_6395_cast_fp16)[name = string("op_6397_cast_fp16")];
            int32 var_6399 = const()[name = string("op_6399"), val = int32(-1)];
            bool q_interleave_0 = const()[name = string("q_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 8, 64, 64]> q_cast_fp16 = concat(axis = var_6399, interleave = q_interleave_0, values = (var_6397_cast_fp16, p_5_cast_fp16))[name = string("q_cast_fp16")];
            fp16 var_6405_promoted_to_fp16 = const()[name = string("op_6405_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 2, 64]> var_6409_cast_fp16 = pow(x = x_79_cast_fp16, y = var_6405_promoted_to_fp16)[name = string("op_6409_cast_fp16")];
            tensor<int32, [1]> var_6411_axes_0 = const()[name = string("op_6411_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6411_keep_dims_0 = const()[name = string("op_6411_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 2, 1]> var_6411_cast_fp16 = reduce_mean(axes = var_6411_axes_0, keep_dims = var_6411_keep_dims_0, x = var_6409_cast_fp16)[name = string("op_6411_cast_fp16")];
            fp16 var_6412_to_fp16 = const()[name = string("op_6412_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 64, 2, 1]> var_6413_cast_fp16 = add(x = var_6411_cast_fp16, y = var_6412_to_fp16)[name = string("op_6413_cast_fp16")];
            fp32 var_6414_epsilon_0 = const()[name = string("op_6414_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 2, 1]> var_6414_cast_fp16 = rsqrt(epsilon = var_6414_epsilon_0, x = var_6413_cast_fp16)[name = string("op_6414_cast_fp16")];
            tensor<fp16, [3, 64, 2, 64]> var_6415_cast_fp16 = mul(x = x_79_cast_fp16, y = var_6414_cast_fp16)[name = string("op_6415_cast_fp16")];
            tensor<fp16, [64]> layers_7_kn_w1_to_fp16 = const()[name = string("layers_7_kn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137280)))];
            tensor<fp16, [3, 64, 2, 64]> var_6416_cast_fp16 = mul(x = var_6415_cast_fp16, y = layers_7_kn_w1_to_fp16)[name = string("op_6416_cast_fp16")];
            tensor<int32, [4]> var_6421 = const()[name = string("op_6421"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_begin_0 = const()[name = string("r_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_end_0 = const()[name = string("r_end_0"), val = tensor<int32, [4]>([3, 2, 64, 16])];
            tensor<bool, [4]> r_end_mask_0 = const()[name = string("r_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 2, 64, 64]> x_83_cast_fp16 = transpose(perm = var_6421, x = var_6416_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [3, 2, 64, 16]> r_cast_fp16 = slice_by_index(begin = r_begin_0, end = r_end_0, end_mask = r_end_mask_0, x = x_83_cast_fp16)[name = string("r_cast_fp16")];
            tensor<int32, [4]> p_begin_0 = const()[name = string("p_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_end_0 = const()[name = string("p_end_0"), val = tensor<int32, [4]>([3, 2, 64, 64])];
            tensor<bool, [4]> p_end_mask_0 = const()[name = string("p_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 2, 64, 48]> p_cast_fp16 = slice_by_index(begin = p_begin_0, end = p_end_0, end_mask = p_end_mask_0, x = x_83_cast_fp16)[name = string("p_cast_fp16")];
            tensor<int32, [4]> var_6437_begin_0 = const()[name = string("op_6437_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_6437_end_0 = const()[name = string("op_6437_end_0"), val = tensor<int32, [4]>([3, 2, 64, 16])];
            tensor<bool, [4]> var_6437_end_mask_0 = const()[name = string("op_6437_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 2, 64, 8]> var_6437_cast_fp16 = slice_by_index(begin = var_6437_begin_0, end = var_6437_end_0, end_mask = var_6437_end_mask_0, x = r_cast_fp16)[name = string("op_6437_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 2, 64, 8]> var_6438_cast_fp16 = mul(x = var_6437_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_6438_cast_fp16")];
            tensor<int32, [4]> var_6443_begin_0 = const()[name = string("op_6443_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6443_end_0 = const()[name = string("op_6443_end_0"), val = tensor<int32, [4]>([3, 2, 64, 8])];
            tensor<bool, [4]> var_6443_end_mask_0 = const()[name = string("op_6443_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 2, 64, 8]> var_6443_cast_fp16 = slice_by_index(begin = var_6443_begin_0, end = var_6443_end_0, end_mask = var_6443_end_mask_0, x = r_cast_fp16)[name = string("op_6443_cast_fp16")];
            int32 var_6445 = const()[name = string("op_6445"), val = int32(-1)];
            bool rh_interleave_0 = const()[name = string("rh_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 2, 64, 16]> rh_cast_fp16 = concat(axis = var_6445, interleave = rh_interleave_0, values = (var_6438_cast_fp16, var_6443_cast_fp16))[name = string("rh_cast_fp16")];
            tensor<fp16, [3, 2, 64, 16]> var_6447_cast_fp16 = mul(x = r_cast_fp16, y = cos_to_fp16)[name = string("op_6447_cast_fp16")];
            tensor<fp16, [3, 2, 64, 16]> var_6448_cast_fp16 = mul(x = rh_cast_fp16, y = sin_to_fp16)[name = string("op_6448_cast_fp16")];
            tensor<fp16, [3, 2, 64, 16]> var_6450_cast_fp16 = add(x = var_6447_cast_fp16, y = var_6448_cast_fp16)[name = string("op_6450_cast_fp16")];
            int32 var_6452 = const()[name = string("op_6452"), val = int32(-1)];
            bool k_101_interleave_0 = const()[name = string("k_101_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 2, 64, 64]> k_101_cast_fp16 = concat(axis = var_6452, interleave = k_101_interleave_0, values = (var_6450_cast_fp16, p_cast_fp16))[name = string("k_101_cast_fp16")];
            tensor<int32, [1]> var_6465_axes_0 = const()[name = string("op_6465_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 1, 64, 64]> var_6465_cast_fp16 = expand_dims(axes = var_6465_axes_0, x = k_101_cast_fp16)[name = string("op_6465_cast_fp16")];
            tensor<int32, [5]> var_6473_reps_0 = const()[name = string("op_6473_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [3, 2, 4, 64, 64]> var_6473_cast_fp16 = tile(reps = var_6473_reps_0, x = var_6465_cast_fp16)[name = string("op_6473_cast_fp16")];
            tensor<int32, [4]> var_6478 = const()[name = string("op_6478"), val = tensor<int32, [4]>([3, 8, 64, 64])];
            tensor<fp16, [3, 8, 64, 64]> k_cast_fp16 = reshape(shape = var_6478, x = var_6473_cast_fp16)[name = string("k_cast_fp16")];
            tensor<int32, [1]> var_6491_axes_0 = const()[name = string("op_6491_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 64, 64]> v_17_cast_fp16 = transpose(perm = var_6346, x = var_6341_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [3, 2, 1, 64, 64]> var_6491_cast_fp16 = expand_dims(axes = var_6491_axes_0, x = v_17_cast_fp16)[name = string("op_6491_cast_fp16")];
            tensor<int32, [5]> var_6499_reps_0 = const()[name = string("op_6499_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [3, 2, 4, 64, 64]> var_6499_cast_fp16 = tile(reps = var_6499_reps_0, x = var_6491_cast_fp16)[name = string("op_6499_cast_fp16")];
            tensor<int32, [4]> var_6504 = const()[name = string("op_6504"), val = tensor<int32, [4]>([3, 8, 64, 64])];
            tensor<fp16, [3, 8, 64, 64]> v_cast_fp16 = reshape(shape = var_6504, x = var_6499_cast_fp16)[name = string("v_cast_fp16")];
            bool var_6509_transpose_x_1 = const()[name = string("op_6509_transpose_x_1"), val = bool(false)];
            bool var_6509_transpose_y_1 = const()[name = string("op_6509_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 8, 64, 64]> var_6509_cast_fp16 = matmul(transpose_x = var_6509_transpose_x_1, transpose_y = var_6509_transpose_y_1, x = q_cast_fp16, y = k_cast_fp16)[name = string("op_6509_cast_fp16")];
            fp16 var_6510_to_fp16 = const()[name = string("op_6510_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [3, 8, 64, 64]> var_6511_cast_fp16 = mul(x = var_6509_cast_fp16, y = var_6510_to_fp16)[name = string("op_6511_cast_fp16")];
            tensor<fp16, [3, 8, 64, 64]> s_cast_fp16 = add(x = var_6511_cast_fp16, y = causal_add_to_fp16)[name = string("s_cast_fp16")];
            int32 var_6514 = const()[name = string("op_6514"), val = int32(-1)];
            tensor<fp16, [3, 8, 64, 64]> var_6516_cast_fp16 = softmax(axis = var_6514, x = s_cast_fp16)[name = string("op_6516_cast_fp16")];
            bool o_39_transpose_x_0 = const()[name = string("o_39_transpose_x_0"), val = bool(false)];
            bool o_39_transpose_y_0 = const()[name = string("o_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 8, 64, 64]> o_39_cast_fp16 = matmul(transpose_x = o_39_transpose_x_0, transpose_y = o_39_transpose_y_0, x = var_6516_cast_fp16, y = v_cast_fp16)[name = string("o_39_cast_fp16")];
            tensor<int32, [4]> var_6522 = const()[name = string("op_6522"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6527 = const()[name = string("op_6527"), val = tensor<int32, [3]>([3, 64, 512])];
            tensor<fp16, [3, 64, 8, 64]> var_6523_cast_fp16 = transpose(perm = var_6522, x = o_39_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [3, 64, 512]> var_6528_cast_fp16 = reshape(shape = var_6527, x = var_6523_cast_fp16)[name = string("op_6528_cast_fp16")];
            tensor<fp16, [3, 64, 512]> var_6529_cast_fp16 = sigmoid(x = gate_cast_fp16)[name = string("op_6529_cast_fp16")];
            tensor<fp16, [3, 64, 512]> input_109_cast_fp16 = mul(x = var_6528_cast_fp16, y = var_6529_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [640, 512]> layers_7_o_weight_to_fp16 = const()[name = string("layers_7_o_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137472)))];
            tensor<fp16, [3, 64, 640]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_o_weight_to_fp16, x = input_109_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<fp16, [3, 64, 640]> x_85_cast_fp16 = add(x = x_69_cast_fp16, y = linear_29_cast_fp16)[name = string("x_85_cast_fp16")];
            fp16 var_6542_to_fp16 = const()[name = string("op_6542_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [3, 64, 640]> x_87_cast_fp16 = mul(x = x_85_cast_fp16, y = var_6542_to_fp16)[name = string("x_87_cast_fp16")];
            fp16 var_6540_promoted_to_fp16 = const()[name = string("op_6540_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> var_6544_cast_fp16 = pow(x = x_87_cast_fp16, y = var_6540_promoted_to_fp16)[name = string("op_6544_cast_fp16")];
            tensor<int32, [1]> var_6546_axes_0 = const()[name = string("op_6546_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6546_keep_dims_0 = const()[name = string("op_6546_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 1]> var_6546_cast_fp16 = reduce_mean(axes = var_6546_axes_0, keep_dims = var_6546_keep_dims_0, x = var_6544_cast_fp16)[name = string("op_6546_cast_fp16")];
            fp16 var_6547_to_fp16 = const()[name = string("op_6547_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 64, 1]> var_6548_cast_fp16 = add(x = var_6546_cast_fp16, y = var_6547_to_fp16)[name = string("op_6548_cast_fp16")];
            fp32 var_6549_epsilon_0 = const()[name = string("op_6549_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 1]> var_6549_cast_fp16 = rsqrt(epsilon = var_6549_epsilon_0, x = var_6548_cast_fp16)[name = string("op_6549_cast_fp16")];
            tensor<fp16, [3, 64, 640]> var_6550_cast_fp16 = mul(x = x_87_cast_fp16, y = var_6549_cast_fp16)[name = string("op_6550_cast_fp16")];
            tensor<fp16, [640]> layers_7_n2_w1_to_fp16 = const()[name = string("layers_7_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109792896)))];
            tensor<fp16, [3, 64, 640]> input_111_cast_fp16 = mul(x = var_6550_cast_fp16, y = layers_7_n2_w1_to_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_7_gate_up_weight_to_fp16 = const()[name = string("layers_7_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109794240)))];
            tensor<fp16, [3, 64, 5120]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_7_gate_up_weight_to_fp16, x = input_111_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [3]> input_113_begin_0 = const()[name = string("input_113_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_113_end_0 = const()[name = string("input_113_end_0"), val = tensor<int32, [3]>([3, 64, 2560])];
            tensor<bool, [3]> input_113_end_mask_0 = const()[name = string("input_113_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 64, 2560]> input_113_cast_fp16 = slice_by_index(begin = input_113_begin_0, end = input_113_end_0, end_mask = input_113_end_mask_0, x = linear_30_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [3, 64, 2560]> var_6560_cast_fp16 = silu(x = input_113_cast_fp16)[name = string("op_6560_cast_fp16")];
            tensor<int32, [3]> var_6565_begin_0 = const()[name = string("op_6565_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_6565_end_0 = const()[name = string("op_6565_end_0"), val = tensor<int32, [3]>([3, 64, 5120])];
            tensor<bool, [3]> var_6565_end_mask_0 = const()[name = string("op_6565_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 64, 2560]> var_6565_cast_fp16 = slice_by_index(begin = var_6565_begin_0, end = var_6565_end_0, end_mask = var_6565_end_mask_0, x = linear_30_cast_fp16)[name = string("op_6565_cast_fp16")];
            tensor<fp16, [3, 64, 2560]> input_115_cast_fp16 = mul(x = var_6560_cast_fp16, y = var_6565_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_7_down_weight_to_fp16 = const()[name = string("layers_7_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116347904)))];
            tensor<fp16, [3, 64, 640]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_down_weight_to_fp16, x = input_115_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [3, 64, 640]> x_89_cast_fp16 = add(x = x_85_cast_fp16, y = linear_31_cast_fp16)[name = string("x_89_cast_fp16")];
            fp16 var_6578_to_fp16 = const()[name = string("op_6578_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [3, 64, 640]> x_cast_fp16 = mul(x = x_89_cast_fp16, y = var_6578_to_fp16)[name = string("x_cast_fp16")];
            fp16 var_6576_promoted_to_fp16 = const()[name = string("op_6576_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 64, 640]> var_6580_cast_fp16 = pow(x = x_cast_fp16, y = var_6576_promoted_to_fp16)[name = string("op_6580_cast_fp16")];
            tensor<int32, [1]> var_6582_axes_0 = const()[name = string("op_6582_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6582_keep_dims_0 = const()[name = string("op_6582_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 64, 1]> var_6582_cast_fp16 = reduce_mean(axes = var_6582_axes_0, keep_dims = var_6582_keep_dims_0, x = var_6580_cast_fp16)[name = string("op_6582_cast_fp16")];
            fp16 var_6583_to_fp16 = const()[name = string("op_6583_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 64, 1]> var_6584_cast_fp16 = add(x = var_6582_cast_fp16, y = var_6583_to_fp16)[name = string("op_6584_cast_fp16")];
            fp32 var_6585_epsilon_0 = const()[name = string("op_6585_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 64, 1]> var_6585_cast_fp16 = rsqrt(epsilon = var_6585_epsilon_0, x = var_6584_cast_fp16)[name = string("op_6585_cast_fp16")];
            tensor<fp16, [3, 64, 640]> var_6586_cast_fp16 = mul(x = x_cast_fp16, y = var_6585_cast_fp16)[name = string("op_6586_cast_fp16")];
            tensor<fp16, [640]> norm_w1_to_fp16 = const()[name = string("norm_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119624768)))];
            tensor<fp16, [3, 64, 640]> var_6587_cast_fp16 = mul(x = var_6586_cast_fp16, y = norm_w1_to_fp16)[name = string("op_6587_cast_fp16")];
            tensor<int32, [3]> h_begin_0 = const()[name = string("h_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> h_end_0 = const()[name = string("h_end_0"), val = tensor<int32, [3]>([3, 63, 640])];
            tensor<bool, [3]> h_end_mask_0 = const()[name = string("h_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 63, 640]> h_cast_fp16 = slice_by_index(begin = h_begin_0, end = h_end_0, end_mask = h_end_mask_0, x = var_6587_cast_fp16)[name = string("h_cast_fp16")];
            tensor<fp16, [16000]> logits_bias_0_to_fp16 = const()[name = string("logits_bias_0_to_fp16"), val = tensor<fp16, [16000]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119626112)))];
            tensor<fp16, [3, 63, 16000]> logits_cast_fp16 = linear(bias = logits_bias_0_to_fp16, weight = emb_to_fp16, x = h_cast_fp16)[name = string("logits_cast_fp16")];
            tensor<int32, [1]> reduce_max_0_axes_0 = const()[name = string("reduce_max_0_axes_0"), val = tensor<int32, [1]>([-1])];
            bool reduce_max_0_keep_dims_0 = const()[name = string("reduce_max_0_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 63, 1]> reduce_max_0_cast_fp16 = reduce_max(axes = reduce_max_0_axes_0, keep_dims = reduce_max_0_keep_dims_0, x = logits_cast_fp16)[name = string("reduce_max_0_cast_fp16")];
            tensor<fp16, [3, 63, 16000]> var_6605_cast_fp16 = sub(x = logits_cast_fp16, y = reduce_max_0_cast_fp16)[name = string("op_6605_cast_fp16")];
            tensor<fp16, [3, 63, 16000]> var_6606_cast_fp16 = exp(x = var_6605_cast_fp16)[name = string("op_6606_cast_fp16")];
            tensor<int32, [1]> var_6611_axes_0 = const()[name = string("op_6611_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6611_keep_dims_0 = const()[name = string("op_6611_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 63, 1]> var_6611_cast_fp16 = reduce_sum(axes = var_6611_axes_0, keep_dims = var_6611_keep_dims_0, x = var_6606_cast_fp16)[name = string("op_6611_cast_fp16")];
            fp32 var_6612_epsilon_0 = const()[name = string("op_6612_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 63, 1]> var_6612_cast_fp16 = log(epsilon = var_6612_epsilon_0, x = var_6611_cast_fp16)[name = string("op_6612_cast_fp16")];
            tensor<fp16, [3, 63, 1]> var_6614_cast_fp16 = add(x = reduce_max_0_cast_fp16, y = var_6612_cast_fp16)[name = string("op_6614_cast_fp16")];
            tensor<int32, [3]> lse_begin_0 = const()[name = string("lse_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> lse_end_0 = const()[name = string("lse_end_0"), val = tensor<int32, [3]>([3, 63, 1])];
            tensor<bool, [3]> lse_end_mask_0 = const()[name = string("lse_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> lse_squeeze_mask_0 = const()[name = string("lse_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<fp16, [3, 63]> lse_cast_fp16 = slice_by_index(begin = lse_begin_0, end = lse_end_0, end_mask = lse_end_mask_0, squeeze_mask = lse_squeeze_mask_0, x = var_6614_cast_fp16)[name = string("lse_cast_fp16")];
            tensor<int32, [2]> input_begin_0 = const()[name = string("input_begin_0"), val = tensor<int32, [2]>([0, 1])];
            tensor<int32, [2]> input_end_0 = const()[name = string("input_end_0"), val = tensor<int32, [2]>([3, 64])];
            tensor<bool, [2]> input_end_mask_0 = const()[name = string("input_end_mask_0"), val = tensor<bool, [2]>([true, true])];
            tensor<int32, [3, 63]> input = slice_by_index(begin = input_begin_0, end = input_end_0, end_mask = input_end_mask_0, x = ids)[name = string("input")];
            int32 var_6631_axis_0 = const()[name = string("op_6631_axis_0"), val = int32(0)];
            int32 var_6631_batch_dims_0 = const()[name = string("op_6631_batch_dims_0"), val = int32(0)];
            bool var_6631_validate_indices_0 = const()[name = string("op_6631_validate_indices_0"), val = bool(false)];
            string input_to_uint16_dtype_0 = const()[name = string("input_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [3, 63]> input_to_uint16 = cast(dtype = input_to_uint16_dtype_0, x = input)[name = string("cast_108")];
            tensor<fp16, [3, 63, 640]> var_6631_cast_fp16_cast_uint16 = gather(axis = var_6631_axis_0, batch_dims = var_6631_batch_dims_0, indices = input_to_uint16, validate_indices = var_6631_validate_indices_0, x = emb_to_fp16)[name = string("op_6631_cast_fp16_cast_uint16")];
            tensor<fp16, [3, 63, 640]> var_6632_cast_fp16 = mul(x = h_cast_fp16, y = var_6631_cast_fp16_cast_uint16)[name = string("op_6632_cast_fp16")];
            tensor<int32, [1]> tgt_axes_0 = const()[name = string("tgt_axes_0"), val = tensor<int32, [1]>([-1])];
            bool tgt_keep_dims_0 = const()[name = string("tgt_keep_dims_0"), val = bool(false)];
            tensor<fp16, [3, 63]> tgt_cast_fp16 = reduce_sum(axes = tgt_axes_0, keep_dims = tgt_keep_dims_0, x = var_6632_cast_fp16)[name = string("tgt_cast_fp16")];
            tensor<fp16, [3, 63]> var_6639_cast_fp16 = sub(x = tgt_cast_fp16, y = lse_cast_fp16)[name = string("op_6639_cast_fp16")];
            string var_6639_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_6639_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [3, 63]> lp = cast(dtype = var_6639_cast_fp16_to_fp32_dtype_0, x = var_6639_cast_fp16)[name = string("cast_107")];
        } -> (lp);
    func t96<ios18>(tensor<int32, [3, 96]> ids) {
            int32 x_1_batch_dims_0 = const()[name = string("x_1_batch_dims_0"), val = int32(0)];
            bool x_1_validate_indices_0 = const()[name = string("x_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [16000, 640]> emb_to_fp16 = const()[name = string("emb_to_fp16"), val = tensor<fp16, [16000, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            string ids_to_int16_dtype_0 = const()[name = string("ids_to_int16_dtype_0"), val = string("int16")];
            string cast_130_dtype_0 = const()[name = string("cast_130_dtype_0"), val = string("int32")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<int16, [3, 96]> ids_to_int16 = cast(dtype = ids_to_int16_dtype_0, x = ids)[name = string("cast_135")];
            tensor<int32, [3, 96]> cast_130 = cast(dtype = cast_130_dtype_0, x = ids_to_int16)[name = string("cast_134")];
            tensor<bool, [3, 96]> greater_equal_0 = greater_equal(x = cast_130, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(16000)];
            tensor<int32, [3, 96]> add_0 = add(x = cast_130, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [3, 96]> select_0 = select(a = cast_130, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 x_1_cast_fp16_cast_uint16_axis_0 = const()[name = string("x_1_cast_fp16_cast_uint16_axis_0"), val = int32(0)];
            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
            tensor<int16, [3, 96]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_133")];
            tensor<fp16, [3, 96, 640]> x_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = x_1_cast_fp16_cast_uint16_axis_0, batch_dims = x_1_batch_dims_0, indices = select_0_to_int16, validate_indices = x_1_validate_indices_0, x = emb_to_fp16)[name = string("x_1_cast_fp16_cast_uint16_cast_uint16")];
            fp16 var_254_to_fp16 = const()[name = string("op_254_to_fp16"), val = fp16(0x1p+5)];
            tensor<fp16, [3, 96, 640]> x_3_cast_fp16 = mul(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = var_254_to_fp16)[name = string("x_3_cast_fp16")];
            fp16 var_252_promoted_to_fp16 = const()[name = string("op_252_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> var_256_cast_fp16 = pow(x = x_3_cast_fp16, y = var_252_promoted_to_fp16)[name = string("op_256_cast_fp16")];
            tensor<int32, [1]> var_258_axes_0 = const()[name = string("op_258_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_258_keep_dims_0 = const()[name = string("op_258_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 1]> var_258_cast_fp16 = reduce_mean(axes = var_258_axes_0, keep_dims = var_258_keep_dims_0, x = var_256_cast_fp16)[name = string("op_258_cast_fp16")];
            fp16 var_259_to_fp16 = const()[name = string("op_259_to_fp16"), val = fp16(0x1.0c8p-10)];
            tensor<fp16, [3, 96, 1]> var_260_cast_fp16 = add(x = var_258_cast_fp16, y = var_259_to_fp16)[name = string("op_260_cast_fp16")];
            fp32 var_261_epsilon_0 = const()[name = string("op_261_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 1]> var_261_cast_fp16 = rsqrt(epsilon = var_261_epsilon_0, x = var_260_cast_fp16)[name = string("op_261_cast_fp16")];
            tensor<fp16, [3, 96, 640]> var_262_cast_fp16 = mul(x = x_3_cast_fp16, y = var_261_cast_fp16)[name = string("op_262_cast_fp16")];
            tensor<fp16, [640]> layers_0_n1_w1_to_fp16 = const()[name = string("layers_0_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20480128)))];
            tensor<fp16, [3, 96, 640]> input_1_cast_fp16 = mul(x = var_262_cast_fp16, y = layers_0_n1_w1_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_0_qkvzba_weight_to_fp16 = const()[name = string("layers_0_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20481472)))];
            tensor<fp16, [1544]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [1544]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22457856)))];
            tensor<fp16, [3, 96, 1544]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_qkvzba_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [3]> qkv_1_begin_0 = const()[name = string("qkv_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_1_end_0 = const()[name = string("qkv_1_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> qkv_1_end_mask_0 = const()[name = string("qkv_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 1024]> qkv_1_cast_fp16 = slice_by_index(begin = qkv_1_begin_0, end = qkv_1_end_0, end_mask = qkv_1_end_mask_0, x = linear_0_cast_fp16)[name = string("qkv_1_cast_fp16")];
            tensor<int32, [3]> z_1_begin_0 = const()[name = string("z_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_1_end_0 = const()[name = string("z_1_end_0"), val = tensor<int32, [3]>([3, 96, 1536])];
            tensor<bool, [3]> z_1_end_mask_0 = const()[name = string("z_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 512]> z_1_cast_fp16 = slice_by_index(begin = z_1_begin_0, end = z_1_end_0, end_mask = z_1_end_mask_0, x = linear_0_cast_fp16)[name = string("z_1_cast_fp16")];
            tensor<int32, [3]> b_1_begin_0 = const()[name = string("b_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_1_end_0 = const()[name = string("b_1_end_0"), val = tensor<int32, [3]>([3, 96, 1540])];
            tensor<bool, [3]> b_1_end_mask_0 = const()[name = string("b_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 4]> b_1_cast_fp16 = slice_by_index(begin = b_1_begin_0, end = b_1_end_0, end_mask = b_1_end_mask_0, x = linear_0_cast_fp16)[name = string("b_1_cast_fp16")];
            tensor<int32, [3]> a_1_begin_0 = const()[name = string("a_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_1_end_0 = const()[name = string("a_1_end_0"), val = tensor<int32, [3]>([3, 96, 1])];
            tensor<bool, [3]> a_1_end_mask_0 = const()[name = string("a_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 4]> a_1_cast_fp16 = slice_by_index(begin = a_1_begin_0, end = a_1_end_0, end_mask = a_1_end_mask_0, x = linear_0_cast_fp16)[name = string("a_1_cast_fp16")];
            int32 var_301 = const()[name = string("op_301"), val = int32(1)];
            bool xp_1_interleave_0 = const()[name = string("xp_1_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 3, 1024]> zpad_to_fp16 = const()[name = string("zpad_to_fp16"), val = tensor<fp16, [3, 3, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22461056)))];
            tensor<fp16, [3, 99, 1024]> xp_1_cast_fp16 = concat(axis = var_301, interleave = xp_1_interleave_0, values = (zpad_to_fp16, qkv_1_cast_fp16))[name = string("xp_1_cast_fp16")];
            tensor<int32, [3]> var_312_begin_0 = const()[name = string("op_312_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_312_end_0 = const()[name = string("op_312_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_312_end_mask_0 = const()[name = string("op_312_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 1024]> var_312_cast_fp16 = slice_by_index(begin = var_312_begin_0, end = var_312_end_0, end_mask = var_312_end_mask_0, x = xp_1_cast_fp16)[name = string("op_312_cast_fp16")];
            tensor<fp16, [1024]> var_320_to_fp16 = const()[name = string("op_320_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22479552)))];
            tensor<fp16, [3, 96, 1024]> var_321_cast_fp16 = mul(x = var_312_cast_fp16, y = var_320_to_fp16)[name = string("op_321_cast_fp16")];
            tensor<int32, [3]> var_331_begin_0 = const()[name = string("op_331_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_331_end_0 = const()[name = string("op_331_end_0"), val = tensor<int32, [3]>([3, 98, 1024])];
            tensor<bool, [3]> var_331_end_mask_0 = const()[name = string("op_331_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_331_cast_fp16 = slice_by_index(begin = var_331_begin_0, end = var_331_end_0, end_mask = var_331_end_mask_0, x = xp_1_cast_fp16)[name = string("op_331_cast_fp16")];
            tensor<fp16, [1024]> var_339_to_fp16 = const()[name = string("op_339_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22481664)))];
            tensor<fp16, [3, 96, 1024]> var_340_cast_fp16 = mul(x = var_331_cast_fp16, y = var_339_to_fp16)[name = string("op_340_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> var_342_cast_fp16 = add(x = var_321_cast_fp16, y = var_340_cast_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [3]> var_352_begin_0 = const()[name = string("op_352_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_352_end_0 = const()[name = string("op_352_end_0"), val = tensor<int32, [3]>([3, 97, 1024])];
            tensor<bool, [3]> var_352_end_mask_0 = const()[name = string("op_352_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_352_cast_fp16 = slice_by_index(begin = var_352_begin_0, end = var_352_end_0, end_mask = var_352_end_mask_0, x = xp_1_cast_fp16)[name = string("op_352_cast_fp16")];
            tensor<fp16, [1024]> var_360_to_fp16 = const()[name = string("op_360_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22483776)))];
            tensor<fp16, [3, 96, 1024]> var_361_cast_fp16 = mul(x = var_352_cast_fp16, y = var_360_to_fp16)[name = string("op_361_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> var_363_cast_fp16 = add(x = var_342_cast_fp16, y = var_361_cast_fp16)[name = string("op_363_cast_fp16")];
            tensor<int32, [3]> var_373_begin_0 = const()[name = string("op_373_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_373_end_0 = const()[name = string("op_373_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> var_373_end_mask_0 = const()[name = string("op_373_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_373_cast_fp16 = slice_by_index(begin = var_373_begin_0, end = var_373_end_0, end_mask = var_373_end_mask_0, x = xp_1_cast_fp16)[name = string("op_373_cast_fp16")];
            tensor<fp16, [1024]> var_381_to_fp16 = const()[name = string("op_381_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22485888)))];
            tensor<fp16, [3, 96, 1024]> var_382_cast_fp16 = mul(x = var_373_cast_fp16, y = var_381_to_fp16)[name = string("op_382_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> input_3_cast_fp16 = add(x = var_363_cast_fp16, y = var_382_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> c_1_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("c_1_cast_fp16")];
            tensor<int32, [3]> var_390_begin_0 = const()[name = string("op_390_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_390_end_0 = const()[name = string("op_390_end_0"), val = tensor<int32, [3]>([3, 96, 256])];
            tensor<bool, [3]> var_390_end_mask_0 = const()[name = string("op_390_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 256]> var_390_cast_fp16 = slice_by_index(begin = var_390_begin_0, end = var_390_end_0, end_mask = var_390_end_mask_0, x = c_1_cast_fp16)[name = string("op_390_cast_fp16")];
            tensor<int32, [4]> var_395 = const()[name = string("op_395"), val = tensor<int32, [4]>([3, 96, 2, 128])];
            tensor<fp16, [3, 96, 2, 128]> q_1_cast_fp16 = reshape(shape = var_395, x = var_390_cast_fp16)[name = string("q_1_cast_fp16")];
            tensor<int32, [3]> var_401_begin_0 = const()[name = string("op_401_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_401_end_0 = const()[name = string("op_401_end_0"), val = tensor<int32, [3]>([3, 96, 512])];
            tensor<bool, [3]> var_401_end_mask_0 = const()[name = string("op_401_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 256]> var_401_cast_fp16 = slice_by_index(begin = var_401_begin_0, end = var_401_end_0, end_mask = var_401_end_mask_0, x = c_1_cast_fp16)[name = string("op_401_cast_fp16")];
            tensor<int32, [4]> var_406 = const()[name = string("op_406"), val = tensor<int32, [4]>([3, 96, 2, 128])];
            tensor<fp16, [3, 96, 2, 128]> k_1_cast_fp16 = reshape(shape = var_406, x = var_401_cast_fp16)[name = string("k_1_cast_fp16")];
            tensor<int32, [3]> var_412_begin_0 = const()[name = string("op_412_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_412_end_0 = const()[name = string("op_412_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> var_412_end_mask_0 = const()[name = string("op_412_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 512]> var_412_cast_fp16 = slice_by_index(begin = var_412_begin_0, end = var_412_end_0, end_mask = var_412_end_mask_0, x = c_1_cast_fp16)[name = string("op_412_cast_fp16")];
            tensor<int32, [4]> var_417 = const()[name = string("op_417"), val = tensor<int32, [4]>([3, 96, 4, 128])];
            tensor<fp16, [3, 96, 4, 128]> var_418_cast_fp16 = reshape(shape = var_417, x = var_412_cast_fp16)[name = string("op_418_cast_fp16")];
            tensor<int32, [4]> var_423 = const()[name = string("op_423"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_425_to_fp16 = const()[name = string("op_425_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 96, 2, 128]> q_3_cast_fp16 = mul(x = q_1_cast_fp16, y = var_425_to_fp16)[name = string("q_3_cast_fp16")];
            fp16 var_427_to_fp16 = const()[name = string("op_427_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 96, 2, 128]> k_3_cast_fp16 = mul(x = k_1_cast_fp16, y = var_427_to_fp16)[name = string("k_3_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_429_cast_fp16 = mul(x = q_3_cast_fp16, y = q_3_cast_fp16)[name = string("op_429_cast_fp16")];
            tensor<int32, [1]> var_434_axes_0 = const()[name = string("op_434_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_434_keep_dims_0 = const()[name = string("op_434_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 2, 1]> var_434_cast_fp16 = reduce_mean(axes = var_434_axes_0, keep_dims = var_434_keep_dims_0, x = var_429_cast_fp16)[name = string("op_434_cast_fp16")];
            fp16 var_435_promoted_to_fp16 = const()[name = string("op_435_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 96, 2, 1]> var_436_cast_fp16 = mul(x = var_434_cast_fp16, y = var_435_promoted_to_fp16)[name = string("op_436_cast_fp16")];
            fp16 var_438_to_fp16 = const()[name = string("op_438_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 96, 2, 1]> var_439_cast_fp16 = add(x = var_436_cast_fp16, y = var_438_to_fp16)[name = string("op_439_cast_fp16")];
            fp32 var_440_epsilon_0 = const()[name = string("op_440_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 2, 1]> var_440_cast_fp16 = rsqrt(epsilon = var_440_epsilon_0, x = var_439_cast_fp16)[name = string("op_440_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_441_cast_fp16 = mul(x = q_3_cast_fp16, y = var_440_cast_fp16)[name = string("op_441_cast_fp16")];
            fp16 var_442_to_fp16 = const()[name = string("op_442_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 96, 2, 128]> q_5_cast_fp16 = mul(x = var_441_cast_fp16, y = var_442_to_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_444_cast_fp16 = mul(x = k_3_cast_fp16, y = k_3_cast_fp16)[name = string("op_444_cast_fp16")];
            tensor<int32, [1]> var_449_axes_0 = const()[name = string("op_449_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_449_keep_dims_0 = const()[name = string("op_449_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 2, 1]> var_449_cast_fp16 = reduce_mean(axes = var_449_axes_0, keep_dims = var_449_keep_dims_0, x = var_444_cast_fp16)[name = string("op_449_cast_fp16")];
            fp16 var_450_promoted_to_fp16 = const()[name = string("op_450_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 96, 2, 1]> var_451_cast_fp16 = mul(x = var_449_cast_fp16, y = var_450_promoted_to_fp16)[name = string("op_451_cast_fp16")];
            fp16 var_453_to_fp16 = const()[name = string("op_453_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 96, 2, 1]> var_454_cast_fp16 = add(x = var_451_cast_fp16, y = var_453_to_fp16)[name = string("op_454_cast_fp16")];
            fp32 var_455_epsilon_0 = const()[name = string("op_455_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 2, 1]> var_455_cast_fp16 = rsqrt(epsilon = var_455_epsilon_0, x = var_454_cast_fp16)[name = string("op_455_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> k_5_cast_fp16 = mul(x = k_3_cast_fp16, y = var_455_cast_fp16)[name = string("k_5_cast_fp16")];
            tensor<int32, [4]> var_461 = const()[name = string("op_461"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_474_axes_0 = const()[name = string("op_474_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 96, 128]> var_462_cast_fp16 = transpose(perm = var_461, x = q_5_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [3, 2, 1, 96, 128]> var_474_cast_fp16 = expand_dims(axes = var_474_axes_0, x = var_462_cast_fp16)[name = string("op_474_cast_fp16")];
            tensor<int32, [5]> var_482_reps_0 = const()[name = string("op_482_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 96, 128]> var_482_cast_fp16 = tile(reps = var_482_reps_0, x = var_474_cast_fp16)[name = string("op_482_cast_fp16")];
            tensor<int32, [4]> var_487 = const()[name = string("op_487"), val = tensor<int32, [4]>([3, 4, 96, 128])];
            tensor<fp16, [3, 4, 96, 128]> q_7_cast_fp16 = reshape(shape = var_487, x = var_482_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<int32, [4]> var_493 = const()[name = string("op_493"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_506_axes_0 = const()[name = string("op_506_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 96, 128]> var_494_cast_fp16 = transpose(perm = var_493, x = k_5_cast_fp16)[name = string("transpose_162")];
            tensor<fp16, [3, 2, 1, 96, 128]> var_506_cast_fp16 = expand_dims(axes = var_506_axes_0, x = var_494_cast_fp16)[name = string("op_506_cast_fp16")];
            tensor<int32, [5]> var_514_reps_0 = const()[name = string("op_514_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 96, 128]> var_514_cast_fp16 = tile(reps = var_514_reps_0, x = var_506_cast_fp16)[name = string("op_514_cast_fp16")];
            tensor<int32, [4]> var_519 = const()[name = string("op_519"), val = tensor<int32, [4]>([3, 4, 96, 128])];
            tensor<fp16, [3, 4, 96, 128]> k_7_cast_fp16 = reshape(shape = var_519, x = var_514_cast_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_521_cast_fp16 = sigmoid(x = b_1_cast_fp16)[name = string("op_521_cast_fp16")];
            tensor<int32, [3]> var_525 = const()[name = string("op_525"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_0_dt_bias_to_fp16 = const()[name = string("layers_0_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 96, 4]> input_5_cast_fp16 = add(x = a_1_cast_fp16, y = layers_0_dt_bias_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_529_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("op_529_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_530_cast_fp16 = abs(x = input_5_cast_fp16)[name = string("op_530_cast_fp16")];
            fp16 const_0_promoted_to_fp16 = const()[name = string("const_0_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 96, 4]> var_531_cast_fp16 = mul(x = var_530_cast_fp16, y = const_0_promoted_to_fp16)[name = string("op_531_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_532_cast_fp16 = exp(x = var_531_cast_fp16)[name = string("op_532_cast_fp16")];
            fp16 var_534_to_fp16 = const()[name = string("op_534_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 96, 4]> var_535_cast_fp16 = add(x = var_532_cast_fp16, y = var_534_to_fp16)[name = string("op_535_cast_fp16")];
            fp32 var_536_epsilon_0 = const()[name = string("op_536_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 96, 4]> var_536_cast_fp16 = log(epsilon = var_536_epsilon_0, x = var_535_cast_fp16)[name = string("op_536_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_538_cast_fp16 = add(x = var_529_cast_fp16, y = var_536_cast_fp16)[name = string("op_538_cast_fp16")];
            tensor<fp16, [4]> layers_0_negA_to_fp16 = const()[name = string("layers_0_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.b3p+3, -0x1.a7cp+0, -0x1.ac4p+3, -0x1.704p+2])];
            tensor<fp16, [3, 96, 4]> var_539_cast_fp16 = mul(x = layers_0_negA_to_fp16, y = var_538_cast_fp16)[name = string("op_539_cast_fp16")];
            tensor<int32, [3]> var_543 = const()[name = string("op_543"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_546_axes_0 = const()[name = string("op_546_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 96]> beta_1_cast_fp16 = transpose(perm = var_525, x = var_521_cast_fp16)[name = string("transpose_161")];
            tensor<fp16, [3, 4, 96, 1]> var_546_cast_fp16 = expand_dims(axes = var_546_axes_0, x = beta_1_cast_fp16)[name = string("op_546_cast_fp16")];
            fp16 var_547_to_fp16 = const()[name = string("op_547_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 96, 1]> var_548_cast_fp16 = mul(x = var_546_cast_fp16, y = var_547_to_fp16)[name = string("op_548_cast_fp16")];
            tensor<fp16, [3, 4, 96, 128]> v_1_cast_fp16 = transpose(perm = var_423, x = var_418_cast_fp16)[name = string("transpose_164")];
            tensor<fp16, [3, 4, 96, 128]> vb_1_cast_fp16 = mul(x = v_1_cast_fp16, y = var_548_cast_fp16)[name = string("vb_1_cast_fp16")];
            tensor<int32, [4]> q_9_begin_0 = const()[name = string("q_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_9_end_0 = const()[name = string("q_9_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_9_end_mask_0 = const()[name = string("q_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_9_cast_fp16 = slice_by_index(begin = q_9_begin_0, end = q_9_end_0, end_mask = q_9_end_mask_0, x = q_7_cast_fp16)[name = string("q_9_cast_fp16")];
            tensor<int32, [4]> k_9_begin_0 = const()[name = string("k_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_9_end_0 = const()[name = string("k_9_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_9_end_mask_0 = const()[name = string("k_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, x = k_7_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<fp16, [3, 4, 96, 128]> var_582_cast_fp16 = mul(x = k_7_cast_fp16, y = var_546_cast_fp16)[name = string("op_582_cast_fp16")];
            tensor<int32, [4]> kb_1_begin_0 = const()[name = string("kb_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_1_end_0 = const()[name = string("kb_1_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_1_end_mask_0 = const()[name = string("kb_1_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_1_cast_fp16 = slice_by_index(begin = kb_1_begin_0, end = kb_1_end_0, end_mask = kb_1_end_mask_0, x = var_582_cast_fp16)[name = string("kb_1_cast_fp16")];
            tensor<int32, [4]> vb_3_begin_0 = const()[name = string("vb_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_3_end_0 = const()[name = string("vb_3_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_3_end_mask_0 = const()[name = string("vb_3_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_3_cast_fp16 = slice_by_index(begin = vb_3_begin_0, end = vb_3_end_0, end_mask = vb_3_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_3_cast_fp16")];
            tensor<int32, [3]> gdec_3_begin_0 = const()[name = string("gdec_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_3_end_0 = const()[name = string("gdec_3_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_3_end_mask_0 = const()[name = string("gdec_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 96]> gdec_1_cast_fp16 = transpose(perm = var_543, x = var_539_cast_fp16)[name = string("transpose_160")];
            tensor<fp16, [3, 4, 16]> gdec_3_cast_fp16 = slice_by_index(begin = gdec_3_begin_0, end = gdec_3_end_0, end_mask = gdec_3_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_3_cast_fp16")];
            tensor<int32, [2]> var_630 = const()[name = string("op_630"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_631_cast_fp16 = reshape(shape = var_630, x = gdec_3_cast_fp16)[name = string("op_631_cast_fp16")];
            tensor<fp16, [256, 16]> transpose_0_to_fp16 = const()[name = string("transpose_0_to_fp16"), val = tensor<fp16, [256, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22488000)))];
            tensor<fp16, [256]> var_632_bias_0_to_fp16 = const()[name = string("op_632_bias_0_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22496256)))];
            tensor<fp16, [12, 256]> var_632_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_631_cast_fp16)[name = string("op_632_cast_fp16")];
            tensor<int32, [4]> var_637 = const()[name = string("op_637"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_1_cast_fp16 = reshape(shape = var_637, x = var_632_cast_fp16)[name = string("P_1_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_639_cast_fp16 = exp(x = P_1_cast_fp16)[name = string("op_639_cast_fp16")];
            tensor<fp16, [16, 16]> tril_incl_to_fp16 = const()[name = string("tril_incl_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22496832)))];
            tensor<fp16, [3, 4, 16, 16]> D_1_cast_fp16 = mul(x = var_639_cast_fp16, y = tril_incl_to_fp16)[name = string("D_1_cast_fp16")];
            bool kk_1_transpose_x_1 = const()[name = string("kk_1_transpose_x_1"), val = bool(false)];
            bool kk_1_transpose_y_1 = const()[name = string("kk_1_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_1_cast_fp16 = matmul(transpose_x = kk_1_transpose_x_1, transpose_y = kk_1_transpose_y_1, x = kb_1_cast_fp16, y = k_9_cast_fp16)[name = string("kk_1_cast_fp16")];
            fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_645_cast_fp16 = mul(x = kk_1_cast_fp16, y = const_1_promoted_to_fp16)[name = string("op_645_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_646_cast_fp16 = mul(x = var_645_cast_fp16, y = D_1_cast_fp16)[name = string("op_646_cast_fp16")];
            tensor<fp16, [16, 16]> cum_tail_to_fp16 = const()[name = string("cum_tail_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22497408)))];
            tensor<fp16, [3, 4, 16, 16]> N_1_cast_fp16 = mul(x = var_646_cast_fp16, y = cum_tail_to_fp16)[name = string("N_1_cast_fp16")];
            tensor<fp16, [16, 16]> neg_tril_strict_to_fp16 = const()[name = string("neg_tril_strict_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22497984)))];
            tensor<fp16, [3, 4, 16, 16]> var_648_cast_fp16 = mul(x = D_1_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_648_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_1_cast_fp16 = mul(x = kk_1_cast_fp16, y = var_648_cast_fp16)[name = string("Na_1_cast_fp16")];
            tensor<fp16, [16, 16]> eye_to_fp16 = const()[name = string("eye_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22498560)))];
            tensor<fp16, [3, 4, 16, 16]> Minv_1_cast_fp16 = add(x = eye_to_fp16, y = N_1_cast_fp16)[name = string("Minv_1_cast_fp16")];
            bool Np_1_transpose_x_0 = const()[name = string("Np_1_transpose_x_0"), val = bool(false)];
            bool Np_1_transpose_y_0 = const()[name = string("Np_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_1_cast_fp16 = matmul(transpose_x = Np_1_transpose_x_0, transpose_y = Np_1_transpose_y_0, x = N_1_cast_fp16, y = Na_1_cast_fp16)[name = string("Np_1_cast_fp16")];
            bool Npa_1_transpose_x_0 = const()[name = string("Npa_1_transpose_x_0"), val = bool(false)];
            bool Npa_1_transpose_y_0 = const()[name = string("Npa_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_1_cast_fp16 = matmul(transpose_x = Npa_1_transpose_x_0, transpose_y = Npa_1_transpose_y_0, x = Na_1_cast_fp16, y = N_1_cast_fp16)[name = string("Npa_1_cast_fp16")];
            bool var_654_transpose_x_0 = const()[name = string("op_654_transpose_x_0"), val = bool(false)];
            bool var_654_transpose_y_0 = const()[name = string("op_654_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_654_cast_fp16 = matmul(transpose_x = var_654_transpose_x_0, transpose_y = var_654_transpose_y_0, x = Minv_1_cast_fp16, y = Np_1_cast_fp16)[name = string("op_654_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_3_cast_fp16 = add(x = Minv_1_cast_fp16, y = var_654_cast_fp16)[name = string("Minv_3_cast_fp16")];
            bool Np_3_transpose_x_0 = const()[name = string("Np_3_transpose_x_0"), val = bool(false)];
            bool Np_3_transpose_y_0 = const()[name = string("Np_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_3_cast_fp16 = matmul(transpose_x = Np_3_transpose_x_0, transpose_y = Np_3_transpose_y_0, x = Np_1_cast_fp16, y = Npa_1_cast_fp16)[name = string("Np_3_cast_fp16")];
            bool Npa_3_transpose_x_0 = const()[name = string("Npa_3_transpose_x_0"), val = bool(false)];
            bool Npa_3_transpose_y_0 = const()[name = string("Npa_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_3_cast_fp16 = matmul(transpose_x = Npa_3_transpose_x_0, transpose_y = Npa_3_transpose_y_0, x = Npa_1_cast_fp16, y = Np_1_cast_fp16)[name = string("Npa_3_cast_fp16")];
            bool var_659_transpose_x_0 = const()[name = string("op_659_transpose_x_0"), val = bool(false)];
            bool var_659_transpose_y_0 = const()[name = string("op_659_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_659_cast_fp16 = matmul(transpose_x = var_659_transpose_x_0, transpose_y = var_659_transpose_y_0, x = Minv_3_cast_fp16, y = Np_3_cast_fp16)[name = string("op_659_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_5_cast_fp16 = add(x = Minv_3_cast_fp16, y = var_659_cast_fp16)[name = string("Minv_5_cast_fp16")];
            bool Np_5_transpose_x_0 = const()[name = string("Np_5_transpose_x_0"), val = bool(false)];
            bool Np_5_transpose_y_0 = const()[name = string("Np_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_5_cast_fp16 = matmul(transpose_x = Np_5_transpose_x_0, transpose_y = Np_5_transpose_y_0, x = Np_3_cast_fp16, y = Npa_3_cast_fp16)[name = string("Np_5_cast_fp16")];
            bool var_663_transpose_x_0 = const()[name = string("op_663_transpose_x_0"), val = bool(false)];
            bool var_663_transpose_y_0 = const()[name = string("op_663_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_663_cast_fp16 = matmul(transpose_x = var_663_transpose_x_0, transpose_y = var_663_transpose_y_0, x = Minv_5_cast_fp16, y = Np_5_cast_fp16)[name = string("op_663_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_7_cast_fp16 = add(x = Minv_5_cast_fp16, y = var_663_cast_fp16)[name = string("Minv_7_cast_fp16")];
            bool u_1_transpose_x_0 = const()[name = string("u_1_transpose_x_0"), val = bool(false)];
            bool u_1_transpose_y_0 = const()[name = string("u_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_1_cast_fp16 = matmul(transpose_x = u_1_transpose_x_0, transpose_y = u_1_transpose_y_0, x = Minv_7_cast_fp16, y = vb_3_cast_fp16)[name = string("u_1_cast_fp16")];
            bool var_670_transpose_x_1 = const()[name = string("op_670_transpose_x_1"), val = bool(false)];
            bool var_670_transpose_y_1 = const()[name = string("op_670_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_670_cast_fp16 = matmul(transpose_x = var_670_transpose_x_1, transpose_y = var_670_transpose_y_1, x = q_9_cast_fp16, y = k_9_cast_fp16)[name = string("op_670_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_1_cast_fp16 = mul(x = var_670_cast_fp16, y = D_1_cast_fp16)[name = string("A_1_cast_fp16")];
            bool var_672_transpose_x_0 = const()[name = string("op_672_transpose_x_0"), val = bool(false)];
            bool var_672_transpose_y_0 = const()[name = string("op_672_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_672_cast_fp16 = matmul(transpose_x = var_672_transpose_x_0, transpose_y = var_672_transpose_y_0, x = A_1_cast_fp16, y = u_1_cast_fp16)[name = string("op_672_cast_fp16")];
            tensor<fp16, [16, 16]> transpose_1_to_fp16 = const()[name = string("transpose_1_to_fp16"), val = tensor<fp16, [16, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658176)))];
            tensor<fp16, [16]> var_673_bias_0_to_fp16 = const()[name = string("op_673_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_673_cast_fp16 = linear(bias = var_673_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_3_cast_fp16)[name = string("op_673_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_1_cast_fp16 = exp(x = var_673_cast_fp16)[name = string("tail_1_cast_fp16")];
            tensor<int32, [1]> var_676_axes_0 = const()[name = string("op_676_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_676_cast_fp16 = expand_dims(axes = var_676_axes_0, x = tail_1_cast_fp16)[name = string("op_676_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_677_cast_fp16 = mul(x = k_9_cast_fp16, y = var_676_cast_fp16)[name = string("op_677_cast_fp16")];
            bool S_1_transpose_x_1 = const()[name = string("S_1_transpose_x_1"), val = bool(true)];
            bool S_1_transpose_y_1 = const()[name = string("S_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_1_cast_fp16 = matmul(transpose_x = S_1_transpose_x_1, transpose_y = S_1_transpose_y_1, x = var_677_cast_fp16, y = u_1_cast_fp16)[name = string("S_1_cast_fp16")];
            tensor<int32, [4]> q_11_begin_0 = const()[name = string("q_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_11_end_0 = const()[name = string("q_11_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> q_11_end_mask_0 = const()[name = string("q_11_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_11_cast_fp16 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, x = q_7_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<int32, [4]> k_11_begin_0 = const()[name = string("k_11_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_11_end_0 = const()[name = string("k_11_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> k_11_end_mask_0 = const()[name = string("k_11_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_11_cast_fp16 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, x = k_7_cast_fp16)[name = string("k_11_cast_fp16")];
            tensor<int32, [4]> kb_3_begin_0 = const()[name = string("kb_3_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_3_end_0 = const()[name = string("kb_3_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> kb_3_end_mask_0 = const()[name = string("kb_3_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_3_cast_fp16 = slice_by_index(begin = kb_3_begin_0, end = kb_3_end_0, end_mask = kb_3_end_mask_0, x = var_582_cast_fp16)[name = string("kb_3_cast_fp16")];
            tensor<int32, [4]> vb_5_begin_0 = const()[name = string("vb_5_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_5_end_0 = const()[name = string("vb_5_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> vb_5_end_mask_0 = const()[name = string("vb_5_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_5_cast_fp16 = slice_by_index(begin = vb_5_begin_0, end = vb_5_end_0, end_mask = vb_5_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_5_cast_fp16")];
            tensor<int32, [3]> gdec_5_begin_0 = const()[name = string("gdec_5_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_5_end_0 = const()[name = string("gdec_5_end_0"), val = tensor<int32, [3]>([3, 4, 32])];
            tensor<bool, [3]> gdec_5_end_mask_0 = const()[name = string("gdec_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_5_cast_fp16 = slice_by_index(begin = gdec_5_begin_0, end = gdec_5_end_0, end_mask = gdec_5_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_5_cast_fp16")];
            tensor<int32, [2]> var_762 = const()[name = string("op_762"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_763_cast_fp16 = reshape(shape = var_762, x = gdec_5_cast_fp16)[name = string("op_763_cast_fp16")];
            tensor<fp16, [12, 256]> var_764_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_763_cast_fp16)[name = string("op_764_cast_fp16")];
            tensor<int32, [4]> var_769 = const()[name = string("op_769"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_3_cast_fp16 = reshape(shape = var_769, x = var_764_cast_fp16)[name = string("P_3_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_771_cast_fp16 = exp(x = P_3_cast_fp16)[name = string("op_771_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_3_cast_fp16 = mul(x = var_771_cast_fp16, y = tril_incl_to_fp16)[name = string("D_3_cast_fp16")];
            bool kk_3_transpose_x_1 = const()[name = string("kk_3_transpose_x_1"), val = bool(false)];
            bool kk_3_transpose_y_1 = const()[name = string("kk_3_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_3_cast_fp16 = matmul(transpose_x = kk_3_transpose_x_1, transpose_y = kk_3_transpose_y_1, x = kb_3_cast_fp16, y = k_11_cast_fp16)[name = string("kk_3_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_777_cast_fp16 = mul(x = kk_3_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_777_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_778_cast_fp16 = mul(x = var_777_cast_fp16, y = D_3_cast_fp16)[name = string("op_778_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_3_cast_fp16 = mul(x = var_778_cast_fp16, y = cum_tail_to_fp16)[name = string("N_3_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_780_cast_fp16 = mul(x = D_3_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_780_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_3_cast_fp16 = mul(x = kk_3_cast_fp16, y = var_780_cast_fp16)[name = string("Na_3_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_9_cast_fp16 = add(x = eye_to_fp16, y = N_3_cast_fp16)[name = string("Minv_9_cast_fp16")];
            bool Np_7_transpose_x_0 = const()[name = string("Np_7_transpose_x_0"), val = bool(false)];
            bool Np_7_transpose_y_0 = const()[name = string("Np_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_7_cast_fp16 = matmul(transpose_x = Np_7_transpose_x_0, transpose_y = Np_7_transpose_y_0, x = N_3_cast_fp16, y = Na_3_cast_fp16)[name = string("Np_7_cast_fp16")];
            bool Npa_5_transpose_x_0 = const()[name = string("Npa_5_transpose_x_0"), val = bool(false)];
            bool Npa_5_transpose_y_0 = const()[name = string("Npa_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_5_cast_fp16 = matmul(transpose_x = Npa_5_transpose_x_0, transpose_y = Npa_5_transpose_y_0, x = Na_3_cast_fp16, y = N_3_cast_fp16)[name = string("Npa_5_cast_fp16")];
            bool var_786_transpose_x_0 = const()[name = string("op_786_transpose_x_0"), val = bool(false)];
            bool var_786_transpose_y_0 = const()[name = string("op_786_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_786_cast_fp16 = matmul(transpose_x = var_786_transpose_x_0, transpose_y = var_786_transpose_y_0, x = Minv_9_cast_fp16, y = Np_7_cast_fp16)[name = string("op_786_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_11_cast_fp16 = add(x = Minv_9_cast_fp16, y = var_786_cast_fp16)[name = string("Minv_11_cast_fp16")];
            bool Np_9_transpose_x_0 = const()[name = string("Np_9_transpose_x_0"), val = bool(false)];
            bool Np_9_transpose_y_0 = const()[name = string("Np_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_9_cast_fp16 = matmul(transpose_x = Np_9_transpose_x_0, transpose_y = Np_9_transpose_y_0, x = Np_7_cast_fp16, y = Npa_5_cast_fp16)[name = string("Np_9_cast_fp16")];
            bool Npa_7_transpose_x_0 = const()[name = string("Npa_7_transpose_x_0"), val = bool(false)];
            bool Npa_7_transpose_y_0 = const()[name = string("Npa_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_7_cast_fp16 = matmul(transpose_x = Npa_7_transpose_x_0, transpose_y = Npa_7_transpose_y_0, x = Npa_5_cast_fp16, y = Np_7_cast_fp16)[name = string("Npa_7_cast_fp16")];
            bool var_791_transpose_x_0 = const()[name = string("op_791_transpose_x_0"), val = bool(false)];
            bool var_791_transpose_y_0 = const()[name = string("op_791_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_791_cast_fp16 = matmul(transpose_x = var_791_transpose_x_0, transpose_y = var_791_transpose_y_0, x = Minv_11_cast_fp16, y = Np_9_cast_fp16)[name = string("op_791_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_13_cast_fp16 = add(x = Minv_11_cast_fp16, y = var_791_cast_fp16)[name = string("Minv_13_cast_fp16")];
            bool Np_11_transpose_x_0 = const()[name = string("Np_11_transpose_x_0"), val = bool(false)];
            bool Np_11_transpose_y_0 = const()[name = string("Np_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_11_cast_fp16 = matmul(transpose_x = Np_11_transpose_x_0, transpose_y = Np_11_transpose_y_0, x = Np_9_cast_fp16, y = Npa_7_cast_fp16)[name = string("Np_11_cast_fp16")];
            bool var_795_transpose_x_0 = const()[name = string("op_795_transpose_x_0"), val = bool(false)];
            bool var_795_transpose_y_0 = const()[name = string("op_795_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_795_cast_fp16 = matmul(transpose_x = var_795_transpose_x_0, transpose_y = var_795_transpose_y_0, x = Minv_13_cast_fp16, y = Np_11_cast_fp16)[name = string("op_795_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_15_cast_fp16 = add(x = Minv_13_cast_fp16, y = var_795_cast_fp16)[name = string("Minv_15_cast_fp16")];
            bool u_3_transpose_x_0 = const()[name = string("u_3_transpose_x_0"), val = bool(false)];
            bool u_3_transpose_y_0 = const()[name = string("u_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_3_cast_fp16 = matmul(transpose_x = u_3_transpose_x_0, transpose_y = u_3_transpose_y_0, x = Minv_15_cast_fp16, y = vb_5_cast_fp16)[name = string("u_3_cast_fp16")];
            bool var_802_transpose_x_1 = const()[name = string("op_802_transpose_x_1"), val = bool(false)];
            bool var_802_transpose_y_1 = const()[name = string("op_802_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_802_cast_fp16 = matmul(transpose_x = var_802_transpose_x_1, transpose_y = var_802_transpose_y_1, x = q_11_cast_fp16, y = k_11_cast_fp16)[name = string("op_802_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_3_cast_fp16 = mul(x = var_802_cast_fp16, y = D_3_cast_fp16)[name = string("A_3_cast_fp16")];
            tensor<fp16, [16]> var_804_bias_0_to_fp16 = const()[name = string("op_804_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_804_cast_fp16 = linear(bias = var_804_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_5_cast_fp16)[name = string("op_804_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_1_cast_fp16 = exp(x = var_804_cast_fp16)[name = string("cum_1_cast_fp16")];
            tensor<int32, [1]> var_807_axes_0 = const()[name = string("op_807_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_807_cast_fp16 = expand_dims(axes = var_807_axes_0, x = cum_1_cast_fp16)[name = string("op_807_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_808_cast_fp16 = mul(x = kb_3_cast_fp16, y = var_807_cast_fp16)[name = string("op_808_cast_fp16")];
            bool w_3_transpose_x_0 = const()[name = string("w_3_transpose_x_0"), val = bool(false)];
            bool w_3_transpose_y_0 = const()[name = string("w_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_3_cast_fp16 = matmul(transpose_x = w_3_transpose_x_0, transpose_y = w_3_transpose_y_0, x = Minv_15_cast_fp16, y = var_808_cast_fp16)[name = string("w_3_cast_fp16")];
            bool var_810_transpose_x_0 = const()[name = string("op_810_transpose_x_0"), val = bool(false)];
            bool var_810_transpose_y_0 = const()[name = string("op_810_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_810_cast_fp16 = matmul(transpose_x = var_810_transpose_x_0, transpose_y = var_810_transpose_y_0, x = w_3_cast_fp16, y = S_1_cast_fp16)[name = string("op_810_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_1_cast_fp16 = sub(x = u_3_cast_fp16, y = var_810_cast_fp16)[name = string("vnew_1_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_815_cast_fp16 = mul(x = q_11_cast_fp16, y = var_807_cast_fp16)[name = string("op_815_cast_fp16")];
            bool var_816_transpose_x_0 = const()[name = string("op_816_transpose_x_0"), val = bool(false)];
            bool var_816_transpose_y_0 = const()[name = string("op_816_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_816_cast_fp16 = matmul(transpose_x = var_816_transpose_x_0, transpose_y = var_816_transpose_y_0, x = var_815_cast_fp16, y = S_1_cast_fp16)[name = string("op_816_cast_fp16")];
            bool var_817_transpose_x_0 = const()[name = string("op_817_transpose_x_0"), val = bool(false)];
            bool var_817_transpose_y_0 = const()[name = string("op_817_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_817_cast_fp16 = matmul(transpose_x = var_817_transpose_x_0, transpose_y = var_817_transpose_y_0, x = A_3_cast_fp16, y = vnew_1_cast_fp16)[name = string("op_817_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_819_cast_fp16 = add(x = var_816_cast_fp16, y = var_817_cast_fp16)[name = string("op_819_cast_fp16")];
            tensor<fp16, [16]> var_820_bias_0_to_fp16 = const()[name = string("op_820_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_820_cast_fp16 = linear(bias = var_820_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_5_cast_fp16)[name = string("op_820_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_3_cast_fp16 = exp(x = var_820_cast_fp16)[name = string("tail_3_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_5_to_fp16 = const()[name = string("transpose_5_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_822_bias_0_to_fp16 = const()[name = string("op_822_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_822_cast_fp16 = linear(bias = var_822_bias_0_to_fp16, weight = transpose_5_to_fp16, x = gdec_5_cast_fp16)[name = string("op_822_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_823_cast_fp16 = exp(x = var_822_cast_fp16)[name = string("op_823_cast_fp16")];
            tensor<int32, [1]> last_1_axes_0 = const()[name = string("last_1_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_1_cast_fp16 = expand_dims(axes = last_1_axes_0, x = var_823_cast_fp16)[name = string("last_1_cast_fp16")];
            tensor<int32, [1]> var_827_axes_0 = const()[name = string("op_827_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_827_cast_fp16 = expand_dims(axes = var_827_axes_0, x = tail_3_cast_fp16)[name = string("op_827_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_828_cast_fp16 = mul(x = k_11_cast_fp16, y = var_827_cast_fp16)[name = string("op_828_cast_fp16")];
            bool upd_1_transpose_x_1 = const()[name = string("upd_1_transpose_x_1"), val = bool(true)];
            bool upd_1_transpose_y_1 = const()[name = string("upd_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_1_cast_fp16 = matmul(transpose_x = upd_1_transpose_x_1, transpose_y = upd_1_transpose_y_1, x = var_828_cast_fp16, y = vnew_1_cast_fp16)[name = string("upd_1_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_833_cast_fp16 = mul(x = S_1_cast_fp16, y = last_1_cast_fp16)[name = string("op_833_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_3_cast_fp16 = add(x = var_833_cast_fp16, y = upd_1_cast_fp16)[name = string("S_3_cast_fp16")];
            tensor<int32, [4]> q_13_begin_0 = const()[name = string("q_13_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_13_end_0 = const()[name = string("q_13_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> q_13_end_mask_0 = const()[name = string("q_13_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, x = q_7_cast_fp16)[name = string("q_13_cast_fp16")];
            tensor<int32, [4]> k_13_begin_0 = const()[name = string("k_13_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_13_end_0 = const()[name = string("k_13_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> k_13_end_mask_0 = const()[name = string("k_13_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, x = k_7_cast_fp16)[name = string("k_13_cast_fp16")];
            tensor<int32, [4]> kb_5_begin_0 = const()[name = string("kb_5_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_5_end_0 = const()[name = string("kb_5_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> kb_5_end_mask_0 = const()[name = string("kb_5_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_5_cast_fp16 = slice_by_index(begin = kb_5_begin_0, end = kb_5_end_0, end_mask = kb_5_end_mask_0, x = var_582_cast_fp16)[name = string("kb_5_cast_fp16")];
            tensor<int32, [4]> vb_7_begin_0 = const()[name = string("vb_7_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_7_end_0 = const()[name = string("vb_7_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> vb_7_end_mask_0 = const()[name = string("vb_7_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_7_cast_fp16 = slice_by_index(begin = vb_7_begin_0, end = vb_7_end_0, end_mask = vb_7_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_7_cast_fp16")];
            tensor<int32, [3]> gdec_7_begin_0 = const()[name = string("gdec_7_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_7_end_0 = const()[name = string("gdec_7_end_0"), val = tensor<int32, [3]>([3, 4, 48])];
            tensor<bool, [3]> gdec_7_end_mask_0 = const()[name = string("gdec_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_7_cast_fp16 = slice_by_index(begin = gdec_7_begin_0, end = gdec_7_end_0, end_mask = gdec_7_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_7_cast_fp16")];
            tensor<int32, [2]> var_916 = const()[name = string("op_916"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_917_cast_fp16 = reshape(shape = var_916, x = gdec_7_cast_fp16)[name = string("op_917_cast_fp16")];
            tensor<fp16, [12, 256]> var_918_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_917_cast_fp16)[name = string("op_918_cast_fp16")];
            tensor<int32, [4]> var_923 = const()[name = string("op_923"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_5_cast_fp16 = reshape(shape = var_923, x = var_918_cast_fp16)[name = string("P_5_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_925_cast_fp16 = exp(x = P_5_cast_fp16)[name = string("op_925_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_5_cast_fp16 = mul(x = var_925_cast_fp16, y = tril_incl_to_fp16)[name = string("D_5_cast_fp16")];
            bool kk_5_transpose_x_1 = const()[name = string("kk_5_transpose_x_1"), val = bool(false)];
            bool kk_5_transpose_y_1 = const()[name = string("kk_5_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_5_cast_fp16 = matmul(transpose_x = kk_5_transpose_x_1, transpose_y = kk_5_transpose_y_1, x = kb_5_cast_fp16, y = k_13_cast_fp16)[name = string("kk_5_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_931_cast_fp16 = mul(x = kk_5_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_931_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_932_cast_fp16 = mul(x = var_931_cast_fp16, y = D_5_cast_fp16)[name = string("op_932_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_5_cast_fp16 = mul(x = var_932_cast_fp16, y = cum_tail_to_fp16)[name = string("N_5_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_934_cast_fp16 = mul(x = D_5_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_934_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_5_cast_fp16 = mul(x = kk_5_cast_fp16, y = var_934_cast_fp16)[name = string("Na_5_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_17_cast_fp16 = add(x = eye_to_fp16, y = N_5_cast_fp16)[name = string("Minv_17_cast_fp16")];
            bool Np_13_transpose_x_0 = const()[name = string("Np_13_transpose_x_0"), val = bool(false)];
            bool Np_13_transpose_y_0 = const()[name = string("Np_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_13_cast_fp16 = matmul(transpose_x = Np_13_transpose_x_0, transpose_y = Np_13_transpose_y_0, x = N_5_cast_fp16, y = Na_5_cast_fp16)[name = string("Np_13_cast_fp16")];
            bool Npa_9_transpose_x_0 = const()[name = string("Npa_9_transpose_x_0"), val = bool(false)];
            bool Npa_9_transpose_y_0 = const()[name = string("Npa_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_9_cast_fp16 = matmul(transpose_x = Npa_9_transpose_x_0, transpose_y = Npa_9_transpose_y_0, x = Na_5_cast_fp16, y = N_5_cast_fp16)[name = string("Npa_9_cast_fp16")];
            bool var_940_transpose_x_0 = const()[name = string("op_940_transpose_x_0"), val = bool(false)];
            bool var_940_transpose_y_0 = const()[name = string("op_940_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_940_cast_fp16 = matmul(transpose_x = var_940_transpose_x_0, transpose_y = var_940_transpose_y_0, x = Minv_17_cast_fp16, y = Np_13_cast_fp16)[name = string("op_940_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_19_cast_fp16 = add(x = Minv_17_cast_fp16, y = var_940_cast_fp16)[name = string("Minv_19_cast_fp16")];
            bool Np_15_transpose_x_0 = const()[name = string("Np_15_transpose_x_0"), val = bool(false)];
            bool Np_15_transpose_y_0 = const()[name = string("Np_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_15_cast_fp16 = matmul(transpose_x = Np_15_transpose_x_0, transpose_y = Np_15_transpose_y_0, x = Np_13_cast_fp16, y = Npa_9_cast_fp16)[name = string("Np_15_cast_fp16")];
            bool Npa_11_transpose_x_0 = const()[name = string("Npa_11_transpose_x_0"), val = bool(false)];
            bool Npa_11_transpose_y_0 = const()[name = string("Npa_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_11_cast_fp16 = matmul(transpose_x = Npa_11_transpose_x_0, transpose_y = Npa_11_transpose_y_0, x = Npa_9_cast_fp16, y = Np_13_cast_fp16)[name = string("Npa_11_cast_fp16")];
            bool var_945_transpose_x_0 = const()[name = string("op_945_transpose_x_0"), val = bool(false)];
            bool var_945_transpose_y_0 = const()[name = string("op_945_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_945_cast_fp16 = matmul(transpose_x = var_945_transpose_x_0, transpose_y = var_945_transpose_y_0, x = Minv_19_cast_fp16, y = Np_15_cast_fp16)[name = string("op_945_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_21_cast_fp16 = add(x = Minv_19_cast_fp16, y = var_945_cast_fp16)[name = string("Minv_21_cast_fp16")];
            bool Np_17_transpose_x_0 = const()[name = string("Np_17_transpose_x_0"), val = bool(false)];
            bool Np_17_transpose_y_0 = const()[name = string("Np_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_17_cast_fp16 = matmul(transpose_x = Np_17_transpose_x_0, transpose_y = Np_17_transpose_y_0, x = Np_15_cast_fp16, y = Npa_11_cast_fp16)[name = string("Np_17_cast_fp16")];
            bool var_949_transpose_x_0 = const()[name = string("op_949_transpose_x_0"), val = bool(false)];
            bool var_949_transpose_y_0 = const()[name = string("op_949_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_949_cast_fp16 = matmul(transpose_x = var_949_transpose_x_0, transpose_y = var_949_transpose_y_0, x = Minv_21_cast_fp16, y = Np_17_cast_fp16)[name = string("op_949_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_23_cast_fp16 = add(x = Minv_21_cast_fp16, y = var_949_cast_fp16)[name = string("Minv_23_cast_fp16")];
            bool u_5_transpose_x_0 = const()[name = string("u_5_transpose_x_0"), val = bool(false)];
            bool u_5_transpose_y_0 = const()[name = string("u_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_5_cast_fp16 = matmul(transpose_x = u_5_transpose_x_0, transpose_y = u_5_transpose_y_0, x = Minv_23_cast_fp16, y = vb_7_cast_fp16)[name = string("u_5_cast_fp16")];
            bool var_956_transpose_x_1 = const()[name = string("op_956_transpose_x_1"), val = bool(false)];
            bool var_956_transpose_y_1 = const()[name = string("op_956_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_956_cast_fp16 = matmul(transpose_x = var_956_transpose_x_1, transpose_y = var_956_transpose_y_1, x = q_13_cast_fp16, y = k_13_cast_fp16)[name = string("op_956_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_5_cast_fp16 = mul(x = var_956_cast_fp16, y = D_5_cast_fp16)[name = string("A_5_cast_fp16")];
            tensor<fp16, [16]> var_958_bias_0_to_fp16 = const()[name = string("op_958_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_958_cast_fp16 = linear(bias = var_958_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_7_cast_fp16)[name = string("op_958_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_3_cast_fp16 = exp(x = var_958_cast_fp16)[name = string("cum_3_cast_fp16")];
            tensor<int32, [1]> var_961_axes_0 = const()[name = string("op_961_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_961_cast_fp16 = expand_dims(axes = var_961_axes_0, x = cum_3_cast_fp16)[name = string("op_961_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_962_cast_fp16 = mul(x = kb_5_cast_fp16, y = var_961_cast_fp16)[name = string("op_962_cast_fp16")];
            bool w_5_transpose_x_0 = const()[name = string("w_5_transpose_x_0"), val = bool(false)];
            bool w_5_transpose_y_0 = const()[name = string("w_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_5_cast_fp16 = matmul(transpose_x = w_5_transpose_x_0, transpose_y = w_5_transpose_y_0, x = Minv_23_cast_fp16, y = var_962_cast_fp16)[name = string("w_5_cast_fp16")];
            bool var_964_transpose_x_0 = const()[name = string("op_964_transpose_x_0"), val = bool(false)];
            bool var_964_transpose_y_0 = const()[name = string("op_964_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_964_cast_fp16 = matmul(transpose_x = var_964_transpose_x_0, transpose_y = var_964_transpose_y_0, x = w_5_cast_fp16, y = S_3_cast_fp16)[name = string("op_964_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_3_cast_fp16 = sub(x = u_5_cast_fp16, y = var_964_cast_fp16)[name = string("vnew_3_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_969_cast_fp16 = mul(x = q_13_cast_fp16, y = var_961_cast_fp16)[name = string("op_969_cast_fp16")];
            bool var_970_transpose_x_0 = const()[name = string("op_970_transpose_x_0"), val = bool(false)];
            bool var_970_transpose_y_0 = const()[name = string("op_970_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_970_cast_fp16 = matmul(transpose_x = var_970_transpose_x_0, transpose_y = var_970_transpose_y_0, x = var_969_cast_fp16, y = S_3_cast_fp16)[name = string("op_970_cast_fp16")];
            bool var_971_transpose_x_0 = const()[name = string("op_971_transpose_x_0"), val = bool(false)];
            bool var_971_transpose_y_0 = const()[name = string("op_971_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_971_cast_fp16 = matmul(transpose_x = var_971_transpose_x_0, transpose_y = var_971_transpose_y_0, x = A_5_cast_fp16, y = vnew_3_cast_fp16)[name = string("op_971_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_973_cast_fp16 = add(x = var_970_cast_fp16, y = var_971_cast_fp16)[name = string("op_973_cast_fp16")];
            tensor<fp16, [16]> var_974_bias_0_to_fp16 = const()[name = string("op_974_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_974_cast_fp16 = linear(bias = var_974_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_7_cast_fp16)[name = string("op_974_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_5_cast_fp16 = exp(x = var_974_cast_fp16)[name = string("tail_5_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_9_to_fp16 = const()[name = string("transpose_9_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_976_bias_0_to_fp16 = const()[name = string("op_976_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_976_cast_fp16 = linear(bias = var_976_bias_0_to_fp16, weight = transpose_9_to_fp16, x = gdec_7_cast_fp16)[name = string("op_976_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_977_cast_fp16 = exp(x = var_976_cast_fp16)[name = string("op_977_cast_fp16")];
            tensor<int32, [1]> last_3_axes_0 = const()[name = string("last_3_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_3_cast_fp16 = expand_dims(axes = last_3_axes_0, x = var_977_cast_fp16)[name = string("last_3_cast_fp16")];
            tensor<int32, [1]> var_981_axes_0 = const()[name = string("op_981_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_981_cast_fp16 = expand_dims(axes = var_981_axes_0, x = tail_5_cast_fp16)[name = string("op_981_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_982_cast_fp16 = mul(x = k_13_cast_fp16, y = var_981_cast_fp16)[name = string("op_982_cast_fp16")];
            bool upd_3_transpose_x_1 = const()[name = string("upd_3_transpose_x_1"), val = bool(true)];
            bool upd_3_transpose_y_1 = const()[name = string("upd_3_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_3_cast_fp16 = matmul(transpose_x = upd_3_transpose_x_1, transpose_y = upd_3_transpose_y_1, x = var_982_cast_fp16, y = vnew_3_cast_fp16)[name = string("upd_3_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_987_cast_fp16 = mul(x = S_3_cast_fp16, y = last_3_cast_fp16)[name = string("op_987_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_5_cast_fp16 = add(x = var_987_cast_fp16, y = upd_3_cast_fp16)[name = string("S_5_cast_fp16")];
            tensor<int32, [4]> q_15_begin_0 = const()[name = string("q_15_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_15_end_0 = const()[name = string("q_15_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> q_15_end_mask_0 = const()[name = string("q_15_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_15_cast_fp16 = slice_by_index(begin = q_15_begin_0, end = q_15_end_0, end_mask = q_15_end_mask_0, x = q_7_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<int32, [4]> k_15_begin_0 = const()[name = string("k_15_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_15_end_0 = const()[name = string("k_15_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> k_15_end_mask_0 = const()[name = string("k_15_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_15_cast_fp16 = slice_by_index(begin = k_15_begin_0, end = k_15_end_0, end_mask = k_15_end_mask_0, x = k_7_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> kb_7_begin_0 = const()[name = string("kb_7_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_7_end_0 = const()[name = string("kb_7_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> kb_7_end_mask_0 = const()[name = string("kb_7_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_7_cast_fp16 = slice_by_index(begin = kb_7_begin_0, end = kb_7_end_0, end_mask = kb_7_end_mask_0, x = var_582_cast_fp16)[name = string("kb_7_cast_fp16")];
            tensor<int32, [4]> vb_9_begin_0 = const()[name = string("vb_9_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_9_end_0 = const()[name = string("vb_9_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> vb_9_end_mask_0 = const()[name = string("vb_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_9_cast_fp16 = slice_by_index(begin = vb_9_begin_0, end = vb_9_end_0, end_mask = vb_9_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_9_cast_fp16")];
            tensor<int32, [3]> gdec_9_begin_0 = const()[name = string("gdec_9_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_9_end_0 = const()[name = string("gdec_9_end_0"), val = tensor<int32, [3]>([3, 4, 64])];
            tensor<bool, [3]> gdec_9_end_mask_0 = const()[name = string("gdec_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_9_cast_fp16 = slice_by_index(begin = gdec_9_begin_0, end = gdec_9_end_0, end_mask = gdec_9_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_9_cast_fp16")];
            tensor<int32, [2]> var_1070 = const()[name = string("op_1070"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_1071_cast_fp16 = reshape(shape = var_1070, x = gdec_9_cast_fp16)[name = string("op_1071_cast_fp16")];
            tensor<fp16, [12, 256]> var_1072_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1071_cast_fp16)[name = string("op_1072_cast_fp16")];
            tensor<int32, [4]> var_1077 = const()[name = string("op_1077"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_7_cast_fp16 = reshape(shape = var_1077, x = var_1072_cast_fp16)[name = string("P_7_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1079_cast_fp16 = exp(x = P_7_cast_fp16)[name = string("op_1079_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_7_cast_fp16 = mul(x = var_1079_cast_fp16, y = tril_incl_to_fp16)[name = string("D_7_cast_fp16")];
            bool kk_7_transpose_x_1 = const()[name = string("kk_7_transpose_x_1"), val = bool(false)];
            bool kk_7_transpose_y_1 = const()[name = string("kk_7_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_7_cast_fp16 = matmul(transpose_x = kk_7_transpose_x_1, transpose_y = kk_7_transpose_y_1, x = kb_7_cast_fp16, y = k_15_cast_fp16)[name = string("kk_7_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_1085_cast_fp16 = mul(x = kk_7_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_1085_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1086_cast_fp16 = mul(x = var_1085_cast_fp16, y = D_7_cast_fp16)[name = string("op_1086_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_7_cast_fp16 = mul(x = var_1086_cast_fp16, y = cum_tail_to_fp16)[name = string("N_7_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1088_cast_fp16 = mul(x = D_7_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1088_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_7_cast_fp16 = mul(x = kk_7_cast_fp16, y = var_1088_cast_fp16)[name = string("Na_7_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_25_cast_fp16 = add(x = eye_to_fp16, y = N_7_cast_fp16)[name = string("Minv_25_cast_fp16")];
            bool Np_19_transpose_x_0 = const()[name = string("Np_19_transpose_x_0"), val = bool(false)];
            bool Np_19_transpose_y_0 = const()[name = string("Np_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_19_cast_fp16 = matmul(transpose_x = Np_19_transpose_x_0, transpose_y = Np_19_transpose_y_0, x = N_7_cast_fp16, y = Na_7_cast_fp16)[name = string("Np_19_cast_fp16")];
            bool Npa_13_transpose_x_0 = const()[name = string("Npa_13_transpose_x_0"), val = bool(false)];
            bool Npa_13_transpose_y_0 = const()[name = string("Npa_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_13_cast_fp16 = matmul(transpose_x = Npa_13_transpose_x_0, transpose_y = Npa_13_transpose_y_0, x = Na_7_cast_fp16, y = N_7_cast_fp16)[name = string("Npa_13_cast_fp16")];
            bool var_1094_transpose_x_0 = const()[name = string("op_1094_transpose_x_0"), val = bool(false)];
            bool var_1094_transpose_y_0 = const()[name = string("op_1094_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1094_cast_fp16 = matmul(transpose_x = var_1094_transpose_x_0, transpose_y = var_1094_transpose_y_0, x = Minv_25_cast_fp16, y = Np_19_cast_fp16)[name = string("op_1094_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_27_cast_fp16 = add(x = Minv_25_cast_fp16, y = var_1094_cast_fp16)[name = string("Minv_27_cast_fp16")];
            bool Np_21_transpose_x_0 = const()[name = string("Np_21_transpose_x_0"), val = bool(false)];
            bool Np_21_transpose_y_0 = const()[name = string("Np_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_21_cast_fp16 = matmul(transpose_x = Np_21_transpose_x_0, transpose_y = Np_21_transpose_y_0, x = Np_19_cast_fp16, y = Npa_13_cast_fp16)[name = string("Np_21_cast_fp16")];
            bool Npa_15_transpose_x_0 = const()[name = string("Npa_15_transpose_x_0"), val = bool(false)];
            bool Npa_15_transpose_y_0 = const()[name = string("Npa_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_15_cast_fp16 = matmul(transpose_x = Npa_15_transpose_x_0, transpose_y = Npa_15_transpose_y_0, x = Npa_13_cast_fp16, y = Np_19_cast_fp16)[name = string("Npa_15_cast_fp16")];
            bool var_1099_transpose_x_0 = const()[name = string("op_1099_transpose_x_0"), val = bool(false)];
            bool var_1099_transpose_y_0 = const()[name = string("op_1099_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1099_cast_fp16 = matmul(transpose_x = var_1099_transpose_x_0, transpose_y = var_1099_transpose_y_0, x = Minv_27_cast_fp16, y = Np_21_cast_fp16)[name = string("op_1099_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_29_cast_fp16 = add(x = Minv_27_cast_fp16, y = var_1099_cast_fp16)[name = string("Minv_29_cast_fp16")];
            bool Np_23_transpose_x_0 = const()[name = string("Np_23_transpose_x_0"), val = bool(false)];
            bool Np_23_transpose_y_0 = const()[name = string("Np_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_23_cast_fp16 = matmul(transpose_x = Np_23_transpose_x_0, transpose_y = Np_23_transpose_y_0, x = Np_21_cast_fp16, y = Npa_15_cast_fp16)[name = string("Np_23_cast_fp16")];
            bool var_1103_transpose_x_0 = const()[name = string("op_1103_transpose_x_0"), val = bool(false)];
            bool var_1103_transpose_y_0 = const()[name = string("op_1103_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1103_cast_fp16 = matmul(transpose_x = var_1103_transpose_x_0, transpose_y = var_1103_transpose_y_0, x = Minv_29_cast_fp16, y = Np_23_cast_fp16)[name = string("op_1103_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_31_cast_fp16 = add(x = Minv_29_cast_fp16, y = var_1103_cast_fp16)[name = string("Minv_31_cast_fp16")];
            bool u_7_transpose_x_0 = const()[name = string("u_7_transpose_x_0"), val = bool(false)];
            bool u_7_transpose_y_0 = const()[name = string("u_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_7_cast_fp16 = matmul(transpose_x = u_7_transpose_x_0, transpose_y = u_7_transpose_y_0, x = Minv_31_cast_fp16, y = vb_9_cast_fp16)[name = string("u_7_cast_fp16")];
            bool var_1110_transpose_x_1 = const()[name = string("op_1110_transpose_x_1"), val = bool(false)];
            bool var_1110_transpose_y_1 = const()[name = string("op_1110_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_1110_cast_fp16 = matmul(transpose_x = var_1110_transpose_x_1, transpose_y = var_1110_transpose_y_1, x = q_15_cast_fp16, y = k_15_cast_fp16)[name = string("op_1110_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_7_cast_fp16 = mul(x = var_1110_cast_fp16, y = D_7_cast_fp16)[name = string("A_7_cast_fp16")];
            tensor<fp16, [16]> var_1112_bias_0_to_fp16 = const()[name = string("op_1112_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_1112_cast_fp16 = linear(bias = var_1112_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_9_cast_fp16)[name = string("op_1112_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_5_cast_fp16 = exp(x = var_1112_cast_fp16)[name = string("cum_5_cast_fp16")];
            tensor<int32, [1]> var_1115_axes_0 = const()[name = string("op_1115_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_1115_cast_fp16 = expand_dims(axes = var_1115_axes_0, x = cum_5_cast_fp16)[name = string("op_1115_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1116_cast_fp16 = mul(x = kb_7_cast_fp16, y = var_1115_cast_fp16)[name = string("op_1116_cast_fp16")];
            bool w_7_transpose_x_0 = const()[name = string("w_7_transpose_x_0"), val = bool(false)];
            bool w_7_transpose_y_0 = const()[name = string("w_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_7_cast_fp16 = matmul(transpose_x = w_7_transpose_x_0, transpose_y = w_7_transpose_y_0, x = Minv_31_cast_fp16, y = var_1116_cast_fp16)[name = string("w_7_cast_fp16")];
            bool var_1118_transpose_x_0 = const()[name = string("op_1118_transpose_x_0"), val = bool(false)];
            bool var_1118_transpose_y_0 = const()[name = string("op_1118_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1118_cast_fp16 = matmul(transpose_x = var_1118_transpose_x_0, transpose_y = var_1118_transpose_y_0, x = w_7_cast_fp16, y = S_5_cast_fp16)[name = string("op_1118_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_5_cast_fp16 = sub(x = u_7_cast_fp16, y = var_1118_cast_fp16)[name = string("vnew_5_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1123_cast_fp16 = mul(x = q_15_cast_fp16, y = var_1115_cast_fp16)[name = string("op_1123_cast_fp16")];
            bool var_1124_transpose_x_0 = const()[name = string("op_1124_transpose_x_0"), val = bool(false)];
            bool var_1124_transpose_y_0 = const()[name = string("op_1124_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1124_cast_fp16 = matmul(transpose_x = var_1124_transpose_x_0, transpose_y = var_1124_transpose_y_0, x = var_1123_cast_fp16, y = S_5_cast_fp16)[name = string("op_1124_cast_fp16")];
            bool var_1125_transpose_x_0 = const()[name = string("op_1125_transpose_x_0"), val = bool(false)];
            bool var_1125_transpose_y_0 = const()[name = string("op_1125_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1125_cast_fp16 = matmul(transpose_x = var_1125_transpose_x_0, transpose_y = var_1125_transpose_y_0, x = A_7_cast_fp16, y = vnew_5_cast_fp16)[name = string("op_1125_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1127_cast_fp16 = add(x = var_1124_cast_fp16, y = var_1125_cast_fp16)[name = string("op_1127_cast_fp16")];
            tensor<fp16, [16]> var_1128_bias_0_to_fp16 = const()[name = string("op_1128_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_1128_cast_fp16 = linear(bias = var_1128_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_9_cast_fp16)[name = string("op_1128_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_7_cast_fp16 = exp(x = var_1128_cast_fp16)[name = string("tail_7_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_13_to_fp16 = const()[name = string("transpose_13_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_1130_bias_0_to_fp16 = const()[name = string("op_1130_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_1130_cast_fp16 = linear(bias = var_1130_bias_0_to_fp16, weight = transpose_13_to_fp16, x = gdec_9_cast_fp16)[name = string("op_1130_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_1131_cast_fp16 = exp(x = var_1130_cast_fp16)[name = string("op_1131_cast_fp16")];
            tensor<int32, [1]> last_5_axes_0 = const()[name = string("last_5_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_5_cast_fp16 = expand_dims(axes = last_5_axes_0, x = var_1131_cast_fp16)[name = string("last_5_cast_fp16")];
            tensor<int32, [1]> var_1135_axes_0 = const()[name = string("op_1135_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_1135_cast_fp16 = expand_dims(axes = var_1135_axes_0, x = tail_7_cast_fp16)[name = string("op_1135_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1136_cast_fp16 = mul(x = k_15_cast_fp16, y = var_1135_cast_fp16)[name = string("op_1136_cast_fp16")];
            bool upd_5_transpose_x_1 = const()[name = string("upd_5_transpose_x_1"), val = bool(true)];
            bool upd_5_transpose_y_1 = const()[name = string("upd_5_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_5_cast_fp16 = matmul(transpose_x = upd_5_transpose_x_1, transpose_y = upd_5_transpose_y_1, x = var_1136_cast_fp16, y = vnew_5_cast_fp16)[name = string("upd_5_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_1141_cast_fp16 = mul(x = S_5_cast_fp16, y = last_5_cast_fp16)[name = string("op_1141_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_7_cast_fp16 = add(x = var_1141_cast_fp16, y = upd_5_cast_fp16)[name = string("S_7_cast_fp16")];
            tensor<int32, [4]> q_17_begin_0 = const()[name = string("q_17_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> q_17_end_0 = const()[name = string("q_17_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> q_17_end_mask_0 = const()[name = string("q_17_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_17_cast_fp16 = slice_by_index(begin = q_17_begin_0, end = q_17_end_0, end_mask = q_17_end_mask_0, x = q_7_cast_fp16)[name = string("q_17_cast_fp16")];
            tensor<int32, [4]> k_17_begin_0 = const()[name = string("k_17_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> k_17_end_0 = const()[name = string("k_17_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> k_17_end_mask_0 = const()[name = string("k_17_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, x = k_7_cast_fp16)[name = string("k_17_cast_fp16")];
            tensor<int32, [4]> kb_9_begin_0 = const()[name = string("kb_9_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> kb_9_end_0 = const()[name = string("kb_9_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> kb_9_end_mask_0 = const()[name = string("kb_9_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_9_cast_fp16 = slice_by_index(begin = kb_9_begin_0, end = kb_9_end_0, end_mask = kb_9_end_mask_0, x = var_582_cast_fp16)[name = string("kb_9_cast_fp16")];
            tensor<int32, [4]> vb_11_begin_0 = const()[name = string("vb_11_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> vb_11_end_0 = const()[name = string("vb_11_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> vb_11_end_mask_0 = const()[name = string("vb_11_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_11_cast_fp16 = slice_by_index(begin = vb_11_begin_0, end = vb_11_end_0, end_mask = vb_11_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_11_cast_fp16")];
            tensor<int32, [3]> gdec_11_begin_0 = const()[name = string("gdec_11_begin_0"), val = tensor<int32, [3]>([0, 0, 64])];
            tensor<int32, [3]> gdec_11_end_0 = const()[name = string("gdec_11_end_0"), val = tensor<int32, [3]>([3, 4, 80])];
            tensor<bool, [3]> gdec_11_end_mask_0 = const()[name = string("gdec_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_11_cast_fp16 = slice_by_index(begin = gdec_11_begin_0, end = gdec_11_end_0, end_mask = gdec_11_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_11_cast_fp16")];
            tensor<int32, [2]> var_1224 = const()[name = string("op_1224"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_1225_cast_fp16 = reshape(shape = var_1224, x = gdec_11_cast_fp16)[name = string("op_1225_cast_fp16")];
            tensor<fp16, [12, 256]> var_1226_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1225_cast_fp16)[name = string("op_1226_cast_fp16")];
            tensor<int32, [4]> var_1231 = const()[name = string("op_1231"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_9_cast_fp16 = reshape(shape = var_1231, x = var_1226_cast_fp16)[name = string("P_9_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1233_cast_fp16 = exp(x = P_9_cast_fp16)[name = string("op_1233_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_9_cast_fp16 = mul(x = var_1233_cast_fp16, y = tril_incl_to_fp16)[name = string("D_9_cast_fp16")];
            bool kk_9_transpose_x_1 = const()[name = string("kk_9_transpose_x_1"), val = bool(false)];
            bool kk_9_transpose_y_1 = const()[name = string("kk_9_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_9_cast_fp16 = matmul(transpose_x = kk_9_transpose_x_1, transpose_y = kk_9_transpose_y_1, x = kb_9_cast_fp16, y = k_17_cast_fp16)[name = string("kk_9_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_1239_cast_fp16 = mul(x = kk_9_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_1239_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1240_cast_fp16 = mul(x = var_1239_cast_fp16, y = D_9_cast_fp16)[name = string("op_1240_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_9_cast_fp16 = mul(x = var_1240_cast_fp16, y = cum_tail_to_fp16)[name = string("N_9_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1242_cast_fp16 = mul(x = D_9_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1242_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_9_cast_fp16 = mul(x = kk_9_cast_fp16, y = var_1242_cast_fp16)[name = string("Na_9_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_33_cast_fp16 = add(x = eye_to_fp16, y = N_9_cast_fp16)[name = string("Minv_33_cast_fp16")];
            bool Np_25_transpose_x_0 = const()[name = string("Np_25_transpose_x_0"), val = bool(false)];
            bool Np_25_transpose_y_0 = const()[name = string("Np_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_25_cast_fp16 = matmul(transpose_x = Np_25_transpose_x_0, transpose_y = Np_25_transpose_y_0, x = N_9_cast_fp16, y = Na_9_cast_fp16)[name = string("Np_25_cast_fp16")];
            bool Npa_17_transpose_x_0 = const()[name = string("Npa_17_transpose_x_0"), val = bool(false)];
            bool Npa_17_transpose_y_0 = const()[name = string("Npa_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_17_cast_fp16 = matmul(transpose_x = Npa_17_transpose_x_0, transpose_y = Npa_17_transpose_y_0, x = Na_9_cast_fp16, y = N_9_cast_fp16)[name = string("Npa_17_cast_fp16")];
            bool var_1248_transpose_x_0 = const()[name = string("op_1248_transpose_x_0"), val = bool(false)];
            bool var_1248_transpose_y_0 = const()[name = string("op_1248_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1248_cast_fp16 = matmul(transpose_x = var_1248_transpose_x_0, transpose_y = var_1248_transpose_y_0, x = Minv_33_cast_fp16, y = Np_25_cast_fp16)[name = string("op_1248_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_35_cast_fp16 = add(x = Minv_33_cast_fp16, y = var_1248_cast_fp16)[name = string("Minv_35_cast_fp16")];
            bool Np_27_transpose_x_0 = const()[name = string("Np_27_transpose_x_0"), val = bool(false)];
            bool Np_27_transpose_y_0 = const()[name = string("Np_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_27_cast_fp16 = matmul(transpose_x = Np_27_transpose_x_0, transpose_y = Np_27_transpose_y_0, x = Np_25_cast_fp16, y = Npa_17_cast_fp16)[name = string("Np_27_cast_fp16")];
            bool Npa_19_transpose_x_0 = const()[name = string("Npa_19_transpose_x_0"), val = bool(false)];
            bool Npa_19_transpose_y_0 = const()[name = string("Npa_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_19_cast_fp16 = matmul(transpose_x = Npa_19_transpose_x_0, transpose_y = Npa_19_transpose_y_0, x = Npa_17_cast_fp16, y = Np_25_cast_fp16)[name = string("Npa_19_cast_fp16")];
            bool var_1253_transpose_x_0 = const()[name = string("op_1253_transpose_x_0"), val = bool(false)];
            bool var_1253_transpose_y_0 = const()[name = string("op_1253_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1253_cast_fp16 = matmul(transpose_x = var_1253_transpose_x_0, transpose_y = var_1253_transpose_y_0, x = Minv_35_cast_fp16, y = Np_27_cast_fp16)[name = string("op_1253_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_37_cast_fp16 = add(x = Minv_35_cast_fp16, y = var_1253_cast_fp16)[name = string("Minv_37_cast_fp16")];
            bool Np_29_transpose_x_0 = const()[name = string("Np_29_transpose_x_0"), val = bool(false)];
            bool Np_29_transpose_y_0 = const()[name = string("Np_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_29_cast_fp16 = matmul(transpose_x = Np_29_transpose_x_0, transpose_y = Np_29_transpose_y_0, x = Np_27_cast_fp16, y = Npa_19_cast_fp16)[name = string("Np_29_cast_fp16")];
            bool var_1257_transpose_x_0 = const()[name = string("op_1257_transpose_x_0"), val = bool(false)];
            bool var_1257_transpose_y_0 = const()[name = string("op_1257_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1257_cast_fp16 = matmul(transpose_x = var_1257_transpose_x_0, transpose_y = var_1257_transpose_y_0, x = Minv_37_cast_fp16, y = Np_29_cast_fp16)[name = string("op_1257_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_39_cast_fp16 = add(x = Minv_37_cast_fp16, y = var_1257_cast_fp16)[name = string("Minv_39_cast_fp16")];
            bool u_9_transpose_x_0 = const()[name = string("u_9_transpose_x_0"), val = bool(false)];
            bool u_9_transpose_y_0 = const()[name = string("u_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_9_cast_fp16 = matmul(transpose_x = u_9_transpose_x_0, transpose_y = u_9_transpose_y_0, x = Minv_39_cast_fp16, y = vb_11_cast_fp16)[name = string("u_9_cast_fp16")];
            bool var_1264_transpose_x_1 = const()[name = string("op_1264_transpose_x_1"), val = bool(false)];
            bool var_1264_transpose_y_1 = const()[name = string("op_1264_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_1264_cast_fp16 = matmul(transpose_x = var_1264_transpose_x_1, transpose_y = var_1264_transpose_y_1, x = q_17_cast_fp16, y = k_17_cast_fp16)[name = string("op_1264_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_9_cast_fp16 = mul(x = var_1264_cast_fp16, y = D_9_cast_fp16)[name = string("A_9_cast_fp16")];
            tensor<fp16, [16]> var_1266_bias_0_to_fp16 = const()[name = string("op_1266_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_1266_cast_fp16 = linear(bias = var_1266_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_11_cast_fp16)[name = string("op_1266_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_7_cast_fp16 = exp(x = var_1266_cast_fp16)[name = string("cum_7_cast_fp16")];
            tensor<int32, [1]> var_1269_axes_0 = const()[name = string("op_1269_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_1269_cast_fp16 = expand_dims(axes = var_1269_axes_0, x = cum_7_cast_fp16)[name = string("op_1269_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1270_cast_fp16 = mul(x = kb_9_cast_fp16, y = var_1269_cast_fp16)[name = string("op_1270_cast_fp16")];
            bool w_9_transpose_x_0 = const()[name = string("w_9_transpose_x_0"), val = bool(false)];
            bool w_9_transpose_y_0 = const()[name = string("w_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_9_cast_fp16 = matmul(transpose_x = w_9_transpose_x_0, transpose_y = w_9_transpose_y_0, x = Minv_39_cast_fp16, y = var_1270_cast_fp16)[name = string("w_9_cast_fp16")];
            bool var_1272_transpose_x_0 = const()[name = string("op_1272_transpose_x_0"), val = bool(false)];
            bool var_1272_transpose_y_0 = const()[name = string("op_1272_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1272_cast_fp16 = matmul(transpose_x = var_1272_transpose_x_0, transpose_y = var_1272_transpose_y_0, x = w_9_cast_fp16, y = S_7_cast_fp16)[name = string("op_1272_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_7_cast_fp16 = sub(x = u_9_cast_fp16, y = var_1272_cast_fp16)[name = string("vnew_7_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1277_cast_fp16 = mul(x = q_17_cast_fp16, y = var_1269_cast_fp16)[name = string("op_1277_cast_fp16")];
            bool var_1278_transpose_x_0 = const()[name = string("op_1278_transpose_x_0"), val = bool(false)];
            bool var_1278_transpose_y_0 = const()[name = string("op_1278_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1278_cast_fp16 = matmul(transpose_x = var_1278_transpose_x_0, transpose_y = var_1278_transpose_y_0, x = var_1277_cast_fp16, y = S_7_cast_fp16)[name = string("op_1278_cast_fp16")];
            bool var_1279_transpose_x_0 = const()[name = string("op_1279_transpose_x_0"), val = bool(false)];
            bool var_1279_transpose_y_0 = const()[name = string("op_1279_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1279_cast_fp16 = matmul(transpose_x = var_1279_transpose_x_0, transpose_y = var_1279_transpose_y_0, x = A_9_cast_fp16, y = vnew_7_cast_fp16)[name = string("op_1279_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1281_cast_fp16 = add(x = var_1278_cast_fp16, y = var_1279_cast_fp16)[name = string("op_1281_cast_fp16")];
            tensor<fp16, [16]> var_1282_bias_0_to_fp16 = const()[name = string("op_1282_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_1282_cast_fp16 = linear(bias = var_1282_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_11_cast_fp16)[name = string("op_1282_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_9_cast_fp16 = exp(x = var_1282_cast_fp16)[name = string("tail_9_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_17_to_fp16 = const()[name = string("transpose_17_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_1284_bias_0_to_fp16 = const()[name = string("op_1284_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_1284_cast_fp16 = linear(bias = var_1284_bias_0_to_fp16, weight = transpose_17_to_fp16, x = gdec_11_cast_fp16)[name = string("op_1284_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_1285_cast_fp16 = exp(x = var_1284_cast_fp16)[name = string("op_1285_cast_fp16")];
            tensor<int32, [1]> last_7_axes_0 = const()[name = string("last_7_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_7_cast_fp16 = expand_dims(axes = last_7_axes_0, x = var_1285_cast_fp16)[name = string("last_7_cast_fp16")];
            tensor<int32, [1]> var_1289_axes_0 = const()[name = string("op_1289_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_1289_cast_fp16 = expand_dims(axes = var_1289_axes_0, x = tail_9_cast_fp16)[name = string("op_1289_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1290_cast_fp16 = mul(x = k_17_cast_fp16, y = var_1289_cast_fp16)[name = string("op_1290_cast_fp16")];
            bool upd_7_transpose_x_1 = const()[name = string("upd_7_transpose_x_1"), val = bool(true)];
            bool upd_7_transpose_y_1 = const()[name = string("upd_7_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_7_cast_fp16 = matmul(transpose_x = upd_7_transpose_x_1, transpose_y = upd_7_transpose_y_1, x = var_1290_cast_fp16, y = vnew_7_cast_fp16)[name = string("upd_7_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_1295_cast_fp16 = mul(x = S_7_cast_fp16, y = last_7_cast_fp16)[name = string("op_1295_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_9_cast_fp16 = add(x = var_1295_cast_fp16, y = upd_7_cast_fp16)[name = string("S_9_cast_fp16")];
            tensor<int32, [4]> q_19_begin_0 = const()[name = string("q_19_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> q_19_end_0 = const()[name = string("q_19_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_19_end_mask_0 = const()[name = string("q_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, x = q_7_cast_fp16)[name = string("q_19_cast_fp16")];
            tensor<int32, [4]> k_19_begin_0 = const()[name = string("k_19_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> k_19_end_0 = const()[name = string("k_19_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_19_end_mask_0 = const()[name = string("k_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_19_cast_fp16 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, x = k_7_cast_fp16)[name = string("k_19_cast_fp16")];
            tensor<int32, [4]> kb_11_begin_0 = const()[name = string("kb_11_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> kb_11_end_0 = const()[name = string("kb_11_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_11_end_mask_0 = const()[name = string("kb_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_11_cast_fp16 = slice_by_index(begin = kb_11_begin_0, end = kb_11_end_0, end_mask = kb_11_end_mask_0, x = var_582_cast_fp16)[name = string("kb_11_cast_fp16")];
            tensor<int32, [4]> vb_13_begin_0 = const()[name = string("vb_13_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> vb_13_end_0 = const()[name = string("vb_13_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_13_end_mask_0 = const()[name = string("vb_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_13_cast_fp16 = slice_by_index(begin = vb_13_begin_0, end = vb_13_end_0, end_mask = vb_13_end_mask_0, x = vb_1_cast_fp16)[name = string("vb_13_cast_fp16")];
            tensor<int32, [3]> gdec_13_begin_0 = const()[name = string("gdec_13_begin_0"), val = tensor<int32, [3]>([0, 0, 80])];
            tensor<int32, [3]> gdec_13_end_0 = const()[name = string("gdec_13_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_13_end_mask_0 = const()[name = string("gdec_13_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_13_cast_fp16 = slice_by_index(begin = gdec_13_begin_0, end = gdec_13_end_0, end_mask = gdec_13_end_mask_0, x = gdec_1_cast_fp16)[name = string("gdec_13_cast_fp16")];
            tensor<int32, [2]> var_1378 = const()[name = string("op_1378"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_1379_cast_fp16 = reshape(shape = var_1378, x = gdec_13_cast_fp16)[name = string("op_1379_cast_fp16")];
            tensor<fp16, [12, 256]> var_1380_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1379_cast_fp16)[name = string("op_1380_cast_fp16")];
            tensor<int32, [4]> var_1385 = const()[name = string("op_1385"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_11_cast_fp16 = reshape(shape = var_1385, x = var_1380_cast_fp16)[name = string("P_11_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1387_cast_fp16 = exp(x = P_11_cast_fp16)[name = string("op_1387_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_11_cast_fp16 = mul(x = var_1387_cast_fp16, y = tril_incl_to_fp16)[name = string("D_11_cast_fp16")];
            bool kk_11_transpose_x_1 = const()[name = string("kk_11_transpose_x_1"), val = bool(false)];
            bool kk_11_transpose_y_1 = const()[name = string("kk_11_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_11_cast_fp16 = matmul(transpose_x = kk_11_transpose_x_1, transpose_y = kk_11_transpose_y_1, x = kb_11_cast_fp16, y = k_19_cast_fp16)[name = string("kk_11_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_1393_cast_fp16 = mul(x = kk_11_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_1393_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1394_cast_fp16 = mul(x = var_1393_cast_fp16, y = D_11_cast_fp16)[name = string("op_1394_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_11_cast_fp16 = mul(x = var_1394_cast_fp16, y = cum_tail_to_fp16)[name = string("N_11_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1396_cast_fp16 = mul(x = D_11_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1396_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_11_cast_fp16 = mul(x = kk_11_cast_fp16, y = var_1396_cast_fp16)[name = string("Na_11_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_41_cast_fp16 = add(x = eye_to_fp16, y = N_11_cast_fp16)[name = string("Minv_41_cast_fp16")];
            bool Np_31_transpose_x_0 = const()[name = string("Np_31_transpose_x_0"), val = bool(false)];
            bool Np_31_transpose_y_0 = const()[name = string("Np_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_31_cast_fp16 = matmul(transpose_x = Np_31_transpose_x_0, transpose_y = Np_31_transpose_y_0, x = N_11_cast_fp16, y = Na_11_cast_fp16)[name = string("Np_31_cast_fp16")];
            bool Npa_21_transpose_x_0 = const()[name = string("Npa_21_transpose_x_0"), val = bool(false)];
            bool Npa_21_transpose_y_0 = const()[name = string("Npa_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_21_cast_fp16 = matmul(transpose_x = Npa_21_transpose_x_0, transpose_y = Npa_21_transpose_y_0, x = Na_11_cast_fp16, y = N_11_cast_fp16)[name = string("Npa_21_cast_fp16")];
            bool var_1402_transpose_x_0 = const()[name = string("op_1402_transpose_x_0"), val = bool(false)];
            bool var_1402_transpose_y_0 = const()[name = string("op_1402_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1402_cast_fp16 = matmul(transpose_x = var_1402_transpose_x_0, transpose_y = var_1402_transpose_y_0, x = Minv_41_cast_fp16, y = Np_31_cast_fp16)[name = string("op_1402_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_43_cast_fp16 = add(x = Minv_41_cast_fp16, y = var_1402_cast_fp16)[name = string("Minv_43_cast_fp16")];
            bool Np_33_transpose_x_0 = const()[name = string("Np_33_transpose_x_0"), val = bool(false)];
            bool Np_33_transpose_y_0 = const()[name = string("Np_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_33_cast_fp16 = matmul(transpose_x = Np_33_transpose_x_0, transpose_y = Np_33_transpose_y_0, x = Np_31_cast_fp16, y = Npa_21_cast_fp16)[name = string("Np_33_cast_fp16")];
            bool Npa_23_transpose_x_0 = const()[name = string("Npa_23_transpose_x_0"), val = bool(false)];
            bool Npa_23_transpose_y_0 = const()[name = string("Npa_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_23_cast_fp16 = matmul(transpose_x = Npa_23_transpose_x_0, transpose_y = Npa_23_transpose_y_0, x = Npa_21_cast_fp16, y = Np_31_cast_fp16)[name = string("Npa_23_cast_fp16")];
            bool var_1407_transpose_x_0 = const()[name = string("op_1407_transpose_x_0"), val = bool(false)];
            bool var_1407_transpose_y_0 = const()[name = string("op_1407_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1407_cast_fp16 = matmul(transpose_x = var_1407_transpose_x_0, transpose_y = var_1407_transpose_y_0, x = Minv_43_cast_fp16, y = Np_33_cast_fp16)[name = string("op_1407_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_45_cast_fp16 = add(x = Minv_43_cast_fp16, y = var_1407_cast_fp16)[name = string("Minv_45_cast_fp16")];
            bool Np_35_transpose_x_0 = const()[name = string("Np_35_transpose_x_0"), val = bool(false)];
            bool Np_35_transpose_y_0 = const()[name = string("Np_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_35_cast_fp16 = matmul(transpose_x = Np_35_transpose_x_0, transpose_y = Np_35_transpose_y_0, x = Np_33_cast_fp16, y = Npa_23_cast_fp16)[name = string("Np_35_cast_fp16")];
            bool var_1411_transpose_x_0 = const()[name = string("op_1411_transpose_x_0"), val = bool(false)];
            bool var_1411_transpose_y_0 = const()[name = string("op_1411_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1411_cast_fp16 = matmul(transpose_x = var_1411_transpose_x_0, transpose_y = var_1411_transpose_y_0, x = Minv_45_cast_fp16, y = Np_35_cast_fp16)[name = string("op_1411_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_47_cast_fp16 = add(x = Minv_45_cast_fp16, y = var_1411_cast_fp16)[name = string("Minv_47_cast_fp16")];
            bool u_11_transpose_x_0 = const()[name = string("u_11_transpose_x_0"), val = bool(false)];
            bool u_11_transpose_y_0 = const()[name = string("u_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_11_cast_fp16 = matmul(transpose_x = u_11_transpose_x_0, transpose_y = u_11_transpose_y_0, x = Minv_47_cast_fp16, y = vb_13_cast_fp16)[name = string("u_11_cast_fp16")];
            bool var_1418_transpose_x_1 = const()[name = string("op_1418_transpose_x_1"), val = bool(false)];
            bool var_1418_transpose_y_1 = const()[name = string("op_1418_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_1418_cast_fp16 = matmul(transpose_x = var_1418_transpose_x_1, transpose_y = var_1418_transpose_y_1, x = q_19_cast_fp16, y = k_19_cast_fp16)[name = string("op_1418_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_11_cast_fp16 = mul(x = var_1418_cast_fp16, y = D_11_cast_fp16)[name = string("A_11_cast_fp16")];
            tensor<fp16, [16]> var_1420_bias_0_to_fp16 = const()[name = string("op_1420_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_1420_cast_fp16 = linear(bias = var_1420_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_13_cast_fp16)[name = string("op_1420_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_9_cast_fp16 = exp(x = var_1420_cast_fp16)[name = string("cum_9_cast_fp16")];
            tensor<int32, [1]> var_1423_axes_0 = const()[name = string("op_1423_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_1423_cast_fp16 = expand_dims(axes = var_1423_axes_0, x = cum_9_cast_fp16)[name = string("op_1423_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1424_cast_fp16 = mul(x = kb_11_cast_fp16, y = var_1423_cast_fp16)[name = string("op_1424_cast_fp16")];
            bool w_11_transpose_x_0 = const()[name = string("w_11_transpose_x_0"), val = bool(false)];
            bool w_11_transpose_y_0 = const()[name = string("w_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_11_cast_fp16 = matmul(transpose_x = w_11_transpose_x_0, transpose_y = w_11_transpose_y_0, x = Minv_47_cast_fp16, y = var_1424_cast_fp16)[name = string("w_11_cast_fp16")];
            bool var_1426_transpose_x_0 = const()[name = string("op_1426_transpose_x_0"), val = bool(false)];
            bool var_1426_transpose_y_0 = const()[name = string("op_1426_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1426_cast_fp16 = matmul(transpose_x = var_1426_transpose_x_0, transpose_y = var_1426_transpose_y_0, x = w_11_cast_fp16, y = S_9_cast_fp16)[name = string("op_1426_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_9_cast_fp16 = sub(x = u_11_cast_fp16, y = var_1426_cast_fp16)[name = string("vnew_9_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1431_cast_fp16 = mul(x = q_19_cast_fp16, y = var_1423_cast_fp16)[name = string("op_1431_cast_fp16")];
            bool var_1432_transpose_x_0 = const()[name = string("op_1432_transpose_x_0"), val = bool(false)];
            bool var_1432_transpose_y_0 = const()[name = string("op_1432_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1432_cast_fp16 = matmul(transpose_x = var_1432_transpose_x_0, transpose_y = var_1432_transpose_y_0, x = var_1431_cast_fp16, y = S_9_cast_fp16)[name = string("op_1432_cast_fp16")];
            bool var_1433_transpose_x_0 = const()[name = string("op_1433_transpose_x_0"), val = bool(false)];
            bool var_1433_transpose_y_0 = const()[name = string("op_1433_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1433_cast_fp16 = matmul(transpose_x = var_1433_transpose_x_0, transpose_y = var_1433_transpose_y_0, x = A_11_cast_fp16, y = vnew_9_cast_fp16)[name = string("op_1433_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_1_cast_fp16 = add(x = var_1432_cast_fp16, y = var_1433_cast_fp16)[name = string("oc_1_cast_fp16")];
            int32 var_1437 = const()[name = string("op_1437"), val = int32(2)];
            bool o_1_interleave_0 = const()[name = string("o_1_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 96, 128]> o_1_cast_fp16 = concat(axis = var_1437, interleave = o_1_interleave_0, values = (var_672_cast_fp16, var_819_cast_fp16, var_973_cast_fp16, var_1127_cast_fp16, var_1281_cast_fp16, oc_1_cast_fp16))[name = string("o_1_cast_fp16")];
            tensor<int32, [4]> var_1443 = const()[name = string("op_1443"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1449 = const()[name = string("op_1449"), val = tensor<int32, [4]>([3, 96, 4, 128])];
            tensor<fp16, [3, 96, 4, 128]> input_7_cast_fp16 = reshape(shape = var_1449, x = z_1_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> o_3_cast_fp16 = transpose(perm = var_1443, x = o_1_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [3, 96, 4, 128]> var_1451_cast_fp16 = mul(x = o_3_cast_fp16, y = o_3_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, [3, 96, 4, 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")];
            fp16 var_1458_to_fp16 = const()[name = string("op_1458_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 96, 4, 1]> var_1459_cast_fp16 = add(x = var_1456_cast_fp16, y = var_1458_to_fp16)[name = string("op_1459_cast_fp16")];
            fp32 var_1460_epsilon_0 = const()[name = string("op_1460_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 4, 1]> var_1460_cast_fp16 = rsqrt(epsilon = var_1460_epsilon_0, x = var_1459_cast_fp16)[name = string("op_1460_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> var_1461_cast_fp16 = mul(x = o_3_cast_fp16, y = var_1460_cast_fp16)[name = string("op_1461_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> var_1463_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_1463_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> o_5_cast_fp16 = mul(x = var_1461_cast_fp16, y = var_1463_cast_fp16)[name = string("o_5_cast_fp16")];
            tensor<int32, [3]> var_1468 = const()[name = string("op_1468"), val = tensor<int32, [3]>([3, 96, 512])];
            tensor<fp16, [3, 96, 512]> input_9_cast_fp16 = reshape(shape = var_1468, x = o_5_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [640, 512]> layers_0_out_weight_to_fp16 = const()[name = string("layers_0_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22499136)))];
            tensor<fp16, [640]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23154560)))];
            tensor<fp16, [3, 96, 640]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_out_weight_to_fp16, x = input_9_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [3, 96, 640]> x_5_cast_fp16 = add(x = x_1_cast_fp16_cast_uint16_cast_uint16, y = linear_1_cast_fp16)[name = string("x_5_cast_fp16")];
            fp16 var_1481_to_fp16 = const()[name = string("op_1481_to_fp16"), val = fp16(0x1p+5)];
            tensor<fp16, [3, 96, 640]> x_7_cast_fp16 = mul(x = x_5_cast_fp16, y = var_1481_to_fp16)[name = string("x_7_cast_fp16")];
            fp16 var_1479_promoted_to_fp16 = const()[name = string("op_1479_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> var_1483_cast_fp16 = pow(x = x_7_cast_fp16, y = var_1479_promoted_to_fp16)[name = string("op_1483_cast_fp16")];
            tensor<int32, [1]> var_1485_axes_0 = const()[name = string("op_1485_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1485_keep_dims_0 = const()[name = string("op_1485_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 1]> var_1485_cast_fp16 = reduce_mean(axes = var_1485_axes_0, keep_dims = var_1485_keep_dims_0, x = var_1483_cast_fp16)[name = string("op_1485_cast_fp16")];
            fp16 var_1486_to_fp16 = const()[name = string("op_1486_to_fp16"), val = fp16(0x1.0c8p-10)];
            tensor<fp16, [3, 96, 1]> var_1487_cast_fp16 = add(x = var_1485_cast_fp16, y = var_1486_to_fp16)[name = string("op_1487_cast_fp16")];
            fp32 var_1488_epsilon_0 = const()[name = string("op_1488_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 1]> var_1488_cast_fp16 = rsqrt(epsilon = var_1488_epsilon_0, x = var_1487_cast_fp16)[name = string("op_1488_cast_fp16")];
            tensor<fp16, [3, 96, 640]> var_1489_cast_fp16 = mul(x = x_7_cast_fp16, y = var_1488_cast_fp16)[name = string("op_1489_cast_fp16")];
            tensor<fp16, [640]> layers_0_n2_w1_to_fp16 = const()[name = string("layers_0_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23155904)))];
            tensor<fp16, [3, 96, 640]> input_11_cast_fp16 = mul(x = var_1489_cast_fp16, y = layers_0_n2_w1_to_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_0_gate_up_weight_to_fp16 = const()[name = string("layers_0_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23157248)))];
            tensor<fp16, [5120]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29710912)))];
            tensor<fp16, [3, 96, 5120]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_0_gate_up_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [3]> input_13_begin_0 = const()[name = string("input_13_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_13_end_0 = const()[name = string("input_13_end_0"), val = tensor<int32, [3]>([3, 96, 2560])];
            tensor<bool, [3]> input_13_end_mask_0 = const()[name = string("input_13_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 2560]> input_13_cast_fp16 = slice_by_index(begin = input_13_begin_0, end = input_13_end_0, end_mask = input_13_end_mask_0, x = linear_2_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [3, 96, 2560]> var_1499_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_1499_cast_fp16")];
            tensor<int32, [3]> var_1504_begin_0 = const()[name = string("op_1504_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_1504_end_0 = const()[name = string("op_1504_end_0"), val = tensor<int32, [3]>([3, 96, 5120])];
            tensor<bool, [3]> var_1504_end_mask_0 = const()[name = string("op_1504_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 2560]> var_1504_cast_fp16 = slice_by_index(begin = var_1504_begin_0, end = var_1504_end_0, end_mask = var_1504_end_mask_0, x = linear_2_cast_fp16)[name = string("op_1504_cast_fp16")];
            tensor<fp16, [3, 96, 2560]> input_15_cast_fp16 = mul(x = var_1499_cast_fp16, y = var_1504_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_0_down_weight_to_fp16 = const()[name = string("layers_0_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29721216)))];
            tensor<fp16, [3, 96, 640]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_down_weight_to_fp16, x = input_15_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [3, 96, 640]> x_9_cast_fp16 = add(x = x_5_cast_fp16, y = linear_3_cast_fp16)[name = string("x_9_cast_fp16")];
            fp16 var_1517_to_fp16 = const()[name = string("op_1517_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> x_11_cast_fp16 = mul(x = x_9_cast_fp16, y = var_1517_to_fp16)[name = string("x_11_cast_fp16")];
            fp16 var_1515_promoted_to_fp16 = const()[name = string("op_1515_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> var_1519_cast_fp16 = pow(x = x_11_cast_fp16, y = var_1515_promoted_to_fp16)[name = string("op_1519_cast_fp16")];
            tensor<int32, [1]> var_1521_axes_0 = const()[name = string("op_1521_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1521_keep_dims_0 = const()[name = string("op_1521_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 1]> var_1521_cast_fp16 = reduce_mean(axes = var_1521_axes_0, keep_dims = var_1521_keep_dims_0, x = var_1519_cast_fp16)[name = string("op_1521_cast_fp16")];
            fp16 var_1522_to_fp16 = const()[name = string("op_1522_to_fp16"), val = fp16(0x1.0cp-18)];
            tensor<fp16, [3, 96, 1]> var_1523_cast_fp16 = add(x = var_1521_cast_fp16, y = var_1522_to_fp16)[name = string("op_1523_cast_fp16")];
            fp32 var_1524_epsilon_0 = const()[name = string("op_1524_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 1]> var_1524_cast_fp16 = rsqrt(epsilon = var_1524_epsilon_0, x = var_1523_cast_fp16)[name = string("op_1524_cast_fp16")];
            tensor<fp16, [3, 96, 640]> var_1525_cast_fp16 = mul(x = x_11_cast_fp16, y = var_1524_cast_fp16)[name = string("op_1525_cast_fp16")];
            tensor<fp16, [640]> layers_1_n1_w1_to_fp16 = const()[name = string("layers_1_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32998080)))];
            tensor<fp16, [3, 96, 640]> input_17_cast_fp16 = mul(x = var_1525_cast_fp16, y = layers_1_n1_w1_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_1_qkvzba_weight_to_fp16 = const()[name = string("layers_1_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32999424)))];
            tensor<fp16, [3, 96, 1544]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_qkvzba_weight_to_fp16, x = input_17_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<int32, [3]> qkv_3_begin_0 = const()[name = string("qkv_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_3_end_0 = const()[name = string("qkv_3_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> qkv_3_end_mask_0 = const()[name = string("qkv_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 1024]> qkv_3_cast_fp16 = slice_by_index(begin = qkv_3_begin_0, end = qkv_3_end_0, end_mask = qkv_3_end_mask_0, x = linear_4_cast_fp16)[name = string("qkv_3_cast_fp16")];
            tensor<int32, [3]> z_3_begin_0 = const()[name = string("z_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_3_end_0 = const()[name = string("z_3_end_0"), val = tensor<int32, [3]>([3, 96, 1536])];
            tensor<bool, [3]> z_3_end_mask_0 = const()[name = string("z_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 512]> z_3_cast_fp16 = slice_by_index(begin = z_3_begin_0, end = z_3_end_0, end_mask = z_3_end_mask_0, x = linear_4_cast_fp16)[name = string("z_3_cast_fp16")];
            tensor<int32, [3]> b_3_begin_0 = const()[name = string("b_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_3_end_0 = const()[name = string("b_3_end_0"), val = tensor<int32, [3]>([3, 96, 1540])];
            tensor<bool, [3]> b_3_end_mask_0 = const()[name = string("b_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 4]> b_3_cast_fp16 = slice_by_index(begin = b_3_begin_0, end = b_3_end_0, end_mask = b_3_end_mask_0, x = linear_4_cast_fp16)[name = string("b_3_cast_fp16")];
            tensor<int32, [3]> a_3_begin_0 = const()[name = string("a_3_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_3_end_0 = const()[name = string("a_3_end_0"), val = tensor<int32, [3]>([3, 96, 1])];
            tensor<bool, [3]> a_3_end_mask_0 = const()[name = string("a_3_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 4]> a_3_cast_fp16 = slice_by_index(begin = a_3_begin_0, end = a_3_end_0, end_mask = a_3_end_mask_0, x = linear_4_cast_fp16)[name = string("a_3_cast_fp16")];
            int32 var_1564 = const()[name = string("op_1564"), val = int32(1)];
            bool xp_3_interleave_0 = const()[name = string("xp_3_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 99, 1024]> xp_3_cast_fp16 = concat(axis = var_1564, interleave = xp_3_interleave_0, values = (zpad_to_fp16, qkv_3_cast_fp16))[name = string("xp_3_cast_fp16")];
            tensor<int32, [3]> var_1575_begin_0 = const()[name = string("op_1575_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_1575_end_0 = const()[name = string("op_1575_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_1575_end_mask_0 = const()[name = string("op_1575_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 1024]> var_1575_cast_fp16 = slice_by_index(begin = var_1575_begin_0, end = var_1575_end_0, end_mask = var_1575_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1575_cast_fp16")];
            tensor<fp16, [1024]> var_1583_to_fp16 = const()[name = string("op_1583_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34975808)))];
            tensor<fp16, [3, 96, 1024]> var_1584_cast_fp16 = mul(x = var_1575_cast_fp16, y = var_1583_to_fp16)[name = string("op_1584_cast_fp16")];
            tensor<int32, [3]> var_1594_begin_0 = const()[name = string("op_1594_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_1594_end_0 = const()[name = string("op_1594_end_0"), val = tensor<int32, [3]>([3, 98, 1024])];
            tensor<bool, [3]> var_1594_end_mask_0 = const()[name = string("op_1594_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_1594_cast_fp16 = slice_by_index(begin = var_1594_begin_0, end = var_1594_end_0, end_mask = var_1594_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1594_cast_fp16")];
            tensor<fp16, [1024]> var_1602_to_fp16 = const()[name = string("op_1602_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34977920)))];
            tensor<fp16, [3, 96, 1024]> var_1603_cast_fp16 = mul(x = var_1594_cast_fp16, y = var_1602_to_fp16)[name = string("op_1603_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> var_1605_cast_fp16 = add(x = var_1584_cast_fp16, y = var_1603_cast_fp16)[name = string("op_1605_cast_fp16")];
            tensor<int32, [3]> var_1615_begin_0 = const()[name = string("op_1615_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1615_end_0 = const()[name = string("op_1615_end_0"), val = tensor<int32, [3]>([3, 97, 1024])];
            tensor<bool, [3]> var_1615_end_mask_0 = const()[name = string("op_1615_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_1615_cast_fp16 = slice_by_index(begin = var_1615_begin_0, end = var_1615_end_0, end_mask = var_1615_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1615_cast_fp16")];
            tensor<fp16, [1024]> var_1623_to_fp16 = const()[name = string("op_1623_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34980032)))];
            tensor<fp16, [3, 96, 1024]> var_1624_cast_fp16 = mul(x = var_1615_cast_fp16, y = var_1623_to_fp16)[name = string("op_1624_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> var_1626_cast_fp16 = add(x = var_1605_cast_fp16, y = var_1624_cast_fp16)[name = string("op_1626_cast_fp16")];
            tensor<int32, [3]> var_1636_begin_0 = const()[name = string("op_1636_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1636_end_0 = const()[name = string("op_1636_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> var_1636_end_mask_0 = const()[name = string("op_1636_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_1636_cast_fp16 = slice_by_index(begin = var_1636_begin_0, end = var_1636_end_0, end_mask = var_1636_end_mask_0, x = xp_3_cast_fp16)[name = string("op_1636_cast_fp16")];
            tensor<fp16, [1024]> var_1644_to_fp16 = const()[name = string("op_1644_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34982144)))];
            tensor<fp16, [3, 96, 1024]> var_1645_cast_fp16 = mul(x = var_1636_cast_fp16, y = var_1644_to_fp16)[name = string("op_1645_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> input_19_cast_fp16 = add(x = var_1626_cast_fp16, y = var_1645_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> c_3_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("c_3_cast_fp16")];
            tensor<int32, [3]> var_1653_begin_0 = const()[name = string("op_1653_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1653_end_0 = const()[name = string("op_1653_end_0"), val = tensor<int32, [3]>([3, 96, 256])];
            tensor<bool, [3]> var_1653_end_mask_0 = const()[name = string("op_1653_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 256]> var_1653_cast_fp16 = slice_by_index(begin = var_1653_begin_0, end = var_1653_end_0, end_mask = var_1653_end_mask_0, x = c_3_cast_fp16)[name = string("op_1653_cast_fp16")];
            tensor<int32, [4]> var_1658 = const()[name = string("op_1658"), val = tensor<int32, [4]>([3, 96, 2, 128])];
            tensor<fp16, [3, 96, 2, 128]> q_21_cast_fp16 = reshape(shape = var_1658, x = var_1653_cast_fp16)[name = string("q_21_cast_fp16")];
            tensor<int32, [3]> var_1664_begin_0 = const()[name = string("op_1664_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_1664_end_0 = const()[name = string("op_1664_end_0"), val = tensor<int32, [3]>([3, 96, 512])];
            tensor<bool, [3]> var_1664_end_mask_0 = const()[name = string("op_1664_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 256]> 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 = c_3_cast_fp16)[name = string("op_1664_cast_fp16")];
            tensor<int32, [4]> var_1669 = const()[name = string("op_1669"), val = tensor<int32, [4]>([3, 96, 2, 128])];
            tensor<fp16, [3, 96, 2, 128]> k_21_cast_fp16 = reshape(shape = var_1669, x = var_1664_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<int32, [3]> var_1675_begin_0 = const()[name = string("op_1675_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_1675_end_0 = const()[name = string("op_1675_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> var_1675_end_mask_0 = const()[name = string("op_1675_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 512]> var_1675_cast_fp16 = slice_by_index(begin = var_1675_begin_0, end = var_1675_end_0, end_mask = var_1675_end_mask_0, x = c_3_cast_fp16)[name = string("op_1675_cast_fp16")];
            tensor<int32, [4]> var_1680 = const()[name = string("op_1680"), val = tensor<int32, [4]>([3, 96, 4, 128])];
            tensor<fp16, [3, 96, 4, 128]> var_1681_cast_fp16 = reshape(shape = var_1680, x = var_1675_cast_fp16)[name = string("op_1681_cast_fp16")];
            tensor<int32, [4]> var_1686 = const()[name = string("op_1686"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1688_to_fp16 = const()[name = string("op_1688_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 96, 2, 128]> q_23_cast_fp16 = mul(x = q_21_cast_fp16, y = var_1688_to_fp16)[name = string("q_23_cast_fp16")];
            fp16 var_1690_to_fp16 = const()[name = string("op_1690_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 96, 2, 128]> k_23_cast_fp16 = mul(x = k_21_cast_fp16, y = var_1690_to_fp16)[name = string("k_23_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_1692_cast_fp16 = mul(x = q_23_cast_fp16, y = q_23_cast_fp16)[name = string("op_1692_cast_fp16")];
            tensor<int32, [1]> var_1697_axes_0 = const()[name = string("op_1697_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1697_keep_dims_0 = const()[name = string("op_1697_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 2, 1]> var_1697_cast_fp16 = reduce_mean(axes = var_1697_axes_0, keep_dims = var_1697_keep_dims_0, x = var_1692_cast_fp16)[name = string("op_1697_cast_fp16")];
            fp16 var_1698_promoted_to_fp16 = const()[name = string("op_1698_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 96, 2, 1]> var_1699_cast_fp16 = mul(x = var_1697_cast_fp16, y = var_1698_promoted_to_fp16)[name = string("op_1699_cast_fp16")];
            fp16 var_1701_to_fp16 = const()[name = string("op_1701_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 96, 2, 1]> var_1702_cast_fp16 = add(x = var_1699_cast_fp16, y = var_1701_to_fp16)[name = string("op_1702_cast_fp16")];
            fp32 var_1703_epsilon_0 = const()[name = string("op_1703_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 2, 1]> var_1703_cast_fp16 = rsqrt(epsilon = var_1703_epsilon_0, x = var_1702_cast_fp16)[name = string("op_1703_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_1704_cast_fp16 = mul(x = q_23_cast_fp16, y = var_1703_cast_fp16)[name = string("op_1704_cast_fp16")];
            fp16 var_1705_to_fp16 = const()[name = string("op_1705_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 96, 2, 128]> q_25_cast_fp16 = mul(x = var_1704_cast_fp16, y = var_1705_to_fp16)[name = string("q_25_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_1707_cast_fp16 = mul(x = k_23_cast_fp16, y = k_23_cast_fp16)[name = string("op_1707_cast_fp16")];
            tensor<int32, [1]> var_1712_axes_0 = const()[name = string("op_1712_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1712_keep_dims_0 = const()[name = string("op_1712_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 2, 1]> var_1712_cast_fp16 = reduce_mean(axes = var_1712_axes_0, keep_dims = var_1712_keep_dims_0, x = var_1707_cast_fp16)[name = string("op_1712_cast_fp16")];
            fp16 var_1713_promoted_to_fp16 = const()[name = string("op_1713_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 96, 2, 1]> var_1714_cast_fp16 = mul(x = var_1712_cast_fp16, y = var_1713_promoted_to_fp16)[name = string("op_1714_cast_fp16")];
            fp16 var_1716_to_fp16 = const()[name = string("op_1716_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 96, 2, 1]> var_1717_cast_fp16 = add(x = var_1714_cast_fp16, y = var_1716_to_fp16)[name = string("op_1717_cast_fp16")];
            fp32 var_1718_epsilon_0 = const()[name = string("op_1718_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 2, 1]> var_1718_cast_fp16 = rsqrt(epsilon = var_1718_epsilon_0, x = var_1717_cast_fp16)[name = string("op_1718_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> k_25_cast_fp16 = mul(x = k_23_cast_fp16, y = var_1718_cast_fp16)[name = string("k_25_cast_fp16")];
            tensor<int32, [4]> var_1724 = const()[name = string("op_1724"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1737_axes_0 = const()[name = string("op_1737_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 96, 128]> var_1725_cast_fp16 = transpose(perm = var_1724, x = q_25_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [3, 2, 1, 96, 128]> var_1737_cast_fp16 = expand_dims(axes = var_1737_axes_0, x = var_1725_cast_fp16)[name = string("op_1737_cast_fp16")];
            tensor<int32, [5]> var_1745_reps_0 = const()[name = string("op_1745_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 96, 128]> var_1745_cast_fp16 = tile(reps = var_1745_reps_0, x = var_1737_cast_fp16)[name = string("op_1745_cast_fp16")];
            tensor<int32, [4]> var_1750 = const()[name = string("op_1750"), val = tensor<int32, [4]>([3, 4, 96, 128])];
            tensor<fp16, [3, 4, 96, 128]> q_27_cast_fp16 = reshape(shape = var_1750, x = var_1745_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<int32, [4]> var_1756 = const()[name = string("op_1756"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_1769_axes_0 = const()[name = string("op_1769_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 96, 128]> var_1757_cast_fp16 = transpose(perm = var_1756, x = k_25_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [3, 2, 1, 96, 128]> var_1769_cast_fp16 = expand_dims(axes = var_1769_axes_0, x = var_1757_cast_fp16)[name = string("op_1769_cast_fp16")];
            tensor<int32, [5]> var_1777_reps_0 = const()[name = string("op_1777_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 96, 128]> var_1777_cast_fp16 = tile(reps = var_1777_reps_0, x = var_1769_cast_fp16)[name = string("op_1777_cast_fp16")];
            tensor<int32, [4]> var_1782 = const()[name = string("op_1782"), val = tensor<int32, [4]>([3, 4, 96, 128])];
            tensor<fp16, [3, 4, 96, 128]> k_27_cast_fp16 = reshape(shape = var_1782, x = var_1777_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_1784_cast_fp16 = sigmoid(x = b_3_cast_fp16)[name = string("op_1784_cast_fp16")];
            tensor<int32, [3]> var_1788 = const()[name = string("op_1788"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_1_dt_bias_to_fp16 = const()[name = string("layers_1_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 96, 4]> input_21_cast_fp16 = add(x = a_3_cast_fp16, y = layers_1_dt_bias_to_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_1792_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("op_1792_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_1793_cast_fp16 = abs(x = input_21_cast_fp16)[name = string("op_1793_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 96, 4]> var_1794_cast_fp16 = mul(x = var_1793_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_1794_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_1795_cast_fp16 = exp(x = var_1794_cast_fp16)[name = string("op_1795_cast_fp16")];
            fp16 var_1797_to_fp16 = const()[name = string("op_1797_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 96, 4]> var_1798_cast_fp16 = add(x = var_1795_cast_fp16, y = var_1797_to_fp16)[name = string("op_1798_cast_fp16")];
            fp32 var_1799_epsilon_0 = const()[name = string("op_1799_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 96, 4]> var_1799_cast_fp16 = log(epsilon = var_1799_epsilon_0, x = var_1798_cast_fp16)[name = string("op_1799_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_1801_cast_fp16 = add(x = var_1792_cast_fp16, y = var_1799_cast_fp16)[name = string("op_1801_cast_fp16")];
            tensor<fp16, [4]> layers_1_negA_to_fp16 = const()[name = string("layers_1_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.ff4p+2, -0x1.8cp+3, -0x1.c0cp+3, -0x1.104p+3])];
            tensor<fp16, [3, 96, 4]> var_1802_cast_fp16 = mul(x = layers_1_negA_to_fp16, y = var_1801_cast_fp16)[name = string("op_1802_cast_fp16")];
            tensor<int32, [3]> var_1806 = const()[name = string("op_1806"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_1809_axes_0 = const()[name = string("op_1809_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 96]> beta_3_cast_fp16 = transpose(perm = var_1788, x = var_1784_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [3, 4, 96, 1]> var_1809_cast_fp16 = expand_dims(axes = var_1809_axes_0, x = beta_3_cast_fp16)[name = string("op_1809_cast_fp16")];
            fp16 var_1810_to_fp16 = const()[name = string("op_1810_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 96, 1]> var_1811_cast_fp16 = mul(x = var_1809_cast_fp16, y = var_1810_to_fp16)[name = string("op_1811_cast_fp16")];
            tensor<fp16, [3, 4, 96, 128]> v_3_cast_fp16 = transpose(perm = var_1686, x = var_1681_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [3, 4, 96, 128]> vb_15_cast_fp16 = mul(x = v_3_cast_fp16, y = var_1811_cast_fp16)[name = string("vb_15_cast_fp16")];
            tensor<int32, [4]> q_29_begin_0 = const()[name = string("q_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_29_end_0 = const()[name = string("q_29_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_29_end_mask_0 = const()[name = string("q_29_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_29_cast_fp16 = slice_by_index(begin = q_29_begin_0, end = q_29_end_0, end_mask = q_29_end_mask_0, x = q_27_cast_fp16)[name = string("q_29_cast_fp16")];
            tensor<int32, [4]> k_29_begin_0 = const()[name = string("k_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_29_end_0 = const()[name = string("k_29_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_29_end_mask_0 = const()[name = string("k_29_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, x = k_27_cast_fp16)[name = string("k_29_cast_fp16")];
            tensor<fp16, [3, 4, 96, 128]> var_1845_cast_fp16 = mul(x = k_27_cast_fp16, y = var_1809_cast_fp16)[name = string("op_1845_cast_fp16")];
            tensor<int32, [4]> kb_13_begin_0 = const()[name = string("kb_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_13_end_0 = const()[name = string("kb_13_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_13_end_mask_0 = const()[name = string("kb_13_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_13_cast_fp16 = slice_by_index(begin = kb_13_begin_0, end = kb_13_end_0, end_mask = kb_13_end_mask_0, x = var_1845_cast_fp16)[name = string("kb_13_cast_fp16")];
            tensor<int32, [4]> vb_17_begin_0 = const()[name = string("vb_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_17_end_0 = const()[name = string("vb_17_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_17_end_mask_0 = const()[name = string("vb_17_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_17_cast_fp16 = slice_by_index(begin = vb_17_begin_0, end = vb_17_end_0, end_mask = vb_17_end_mask_0, x = vb_15_cast_fp16)[name = string("vb_17_cast_fp16")];
            tensor<int32, [3]> gdec_17_begin_0 = const()[name = string("gdec_17_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_17_end_0 = const()[name = string("gdec_17_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_17_end_mask_0 = const()[name = string("gdec_17_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 96]> gdec_15_cast_fp16 = transpose(perm = var_1806, x = var_1802_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [3, 4, 16]> gdec_17_cast_fp16 = slice_by_index(begin = gdec_17_begin_0, end = gdec_17_end_0, end_mask = gdec_17_end_mask_0, x = gdec_15_cast_fp16)[name = string("gdec_17_cast_fp16")];
            tensor<int32, [2]> var_1893 = const()[name = string("op_1893"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_1894_cast_fp16 = reshape(shape = var_1893, x = gdec_17_cast_fp16)[name = string("op_1894_cast_fp16")];
            tensor<fp16, [12, 256]> var_1895_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_1894_cast_fp16)[name = string("op_1895_cast_fp16")];
            tensor<int32, [4]> var_1900 = const()[name = string("op_1900"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_13_cast_fp16 = reshape(shape = var_1900, x = var_1895_cast_fp16)[name = string("P_13_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1902_cast_fp16 = exp(x = P_13_cast_fp16)[name = string("op_1902_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_13_cast_fp16 = mul(x = var_1902_cast_fp16, y = tril_incl_to_fp16)[name = string("D_13_cast_fp16")];
            bool kk_13_transpose_x_1 = const()[name = string("kk_13_transpose_x_1"), val = bool(false)];
            bool kk_13_transpose_y_1 = const()[name = string("kk_13_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_13_cast_fp16 = matmul(transpose_x = kk_13_transpose_x_1, transpose_y = kk_13_transpose_y_1, x = kb_13_cast_fp16, y = k_29_cast_fp16)[name = string("kk_13_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_1908_cast_fp16 = mul(x = kk_13_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_1908_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1909_cast_fp16 = mul(x = var_1908_cast_fp16, y = D_13_cast_fp16)[name = string("op_1909_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_13_cast_fp16 = mul(x = var_1909_cast_fp16, y = cum_tail_to_fp16)[name = string("N_13_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_1911_cast_fp16 = mul(x = D_13_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_1911_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_13_cast_fp16 = mul(x = kk_13_cast_fp16, y = var_1911_cast_fp16)[name = string("Na_13_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_49_cast_fp16 = add(x = eye_to_fp16, y = N_13_cast_fp16)[name = string("Minv_49_cast_fp16")];
            bool Np_37_transpose_x_0 = const()[name = string("Np_37_transpose_x_0"), val = bool(false)];
            bool Np_37_transpose_y_0 = const()[name = string("Np_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_37_cast_fp16 = matmul(transpose_x = Np_37_transpose_x_0, transpose_y = Np_37_transpose_y_0, x = N_13_cast_fp16, y = Na_13_cast_fp16)[name = string("Np_37_cast_fp16")];
            bool Npa_25_transpose_x_0 = const()[name = string("Npa_25_transpose_x_0"), val = bool(false)];
            bool Npa_25_transpose_y_0 = const()[name = string("Npa_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_25_cast_fp16 = matmul(transpose_x = Npa_25_transpose_x_0, transpose_y = Npa_25_transpose_y_0, x = Na_13_cast_fp16, y = N_13_cast_fp16)[name = string("Npa_25_cast_fp16")];
            bool var_1917_transpose_x_0 = const()[name = string("op_1917_transpose_x_0"), val = bool(false)];
            bool var_1917_transpose_y_0 = const()[name = string("op_1917_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1917_cast_fp16 = matmul(transpose_x = var_1917_transpose_x_0, transpose_y = var_1917_transpose_y_0, x = Minv_49_cast_fp16, y = Np_37_cast_fp16)[name = string("op_1917_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_51_cast_fp16 = add(x = Minv_49_cast_fp16, y = var_1917_cast_fp16)[name = string("Minv_51_cast_fp16")];
            bool Np_39_transpose_x_0 = const()[name = string("Np_39_transpose_x_0"), val = bool(false)];
            bool Np_39_transpose_y_0 = const()[name = string("Np_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_39_cast_fp16 = matmul(transpose_x = Np_39_transpose_x_0, transpose_y = Np_39_transpose_y_0, x = Np_37_cast_fp16, y = Npa_25_cast_fp16)[name = string("Np_39_cast_fp16")];
            bool Npa_27_transpose_x_0 = const()[name = string("Npa_27_transpose_x_0"), val = bool(false)];
            bool Npa_27_transpose_y_0 = const()[name = string("Npa_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_27_cast_fp16 = matmul(transpose_x = Npa_27_transpose_x_0, transpose_y = Npa_27_transpose_y_0, x = Npa_25_cast_fp16, y = Np_37_cast_fp16)[name = string("Npa_27_cast_fp16")];
            bool var_1922_transpose_x_0 = const()[name = string("op_1922_transpose_x_0"), val = bool(false)];
            bool var_1922_transpose_y_0 = const()[name = string("op_1922_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1922_cast_fp16 = matmul(transpose_x = var_1922_transpose_x_0, transpose_y = var_1922_transpose_y_0, x = Minv_51_cast_fp16, y = Np_39_cast_fp16)[name = string("op_1922_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_53_cast_fp16 = add(x = Minv_51_cast_fp16, y = var_1922_cast_fp16)[name = string("Minv_53_cast_fp16")];
            bool Np_41_transpose_x_0 = const()[name = string("Np_41_transpose_x_0"), val = bool(false)];
            bool Np_41_transpose_y_0 = const()[name = string("Np_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_41_cast_fp16 = matmul(transpose_x = Np_41_transpose_x_0, transpose_y = Np_41_transpose_y_0, x = Np_39_cast_fp16, y = Npa_27_cast_fp16)[name = string("Np_41_cast_fp16")];
            bool var_1926_transpose_x_0 = const()[name = string("op_1926_transpose_x_0"), val = bool(false)];
            bool var_1926_transpose_y_0 = const()[name = string("op_1926_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_1926_cast_fp16 = matmul(transpose_x = var_1926_transpose_x_0, transpose_y = var_1926_transpose_y_0, x = Minv_53_cast_fp16, y = Np_41_cast_fp16)[name = string("op_1926_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_55_cast_fp16 = add(x = Minv_53_cast_fp16, y = var_1926_cast_fp16)[name = string("Minv_55_cast_fp16")];
            bool u_13_transpose_x_0 = const()[name = string("u_13_transpose_x_0"), val = bool(false)];
            bool u_13_transpose_y_0 = const()[name = string("u_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_13_cast_fp16 = matmul(transpose_x = u_13_transpose_x_0, transpose_y = u_13_transpose_y_0, x = Minv_55_cast_fp16, y = vb_17_cast_fp16)[name = string("u_13_cast_fp16")];
            bool var_1933_transpose_x_1 = const()[name = string("op_1933_transpose_x_1"), val = bool(false)];
            bool var_1933_transpose_y_1 = const()[name = string("op_1933_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_1933_cast_fp16 = matmul(transpose_x = var_1933_transpose_x_1, transpose_y = var_1933_transpose_y_1, x = q_29_cast_fp16, y = k_29_cast_fp16)[name = string("op_1933_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_13_cast_fp16 = mul(x = var_1933_cast_fp16, y = D_13_cast_fp16)[name = string("A_13_cast_fp16")];
            bool var_1935_transpose_x_0 = const()[name = string("op_1935_transpose_x_0"), val = bool(false)];
            bool var_1935_transpose_y_0 = const()[name = string("op_1935_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_1935_cast_fp16 = matmul(transpose_x = var_1935_transpose_x_0, transpose_y = var_1935_transpose_y_0, x = A_13_cast_fp16, y = u_13_cast_fp16)[name = string("op_1935_cast_fp16")];
            tensor<fp16, [16]> var_1936_bias_0_to_fp16 = const()[name = string("op_1936_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_1936_cast_fp16 = linear(bias = var_1936_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_17_cast_fp16)[name = string("op_1936_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_11_cast_fp16 = exp(x = var_1936_cast_fp16)[name = string("tail_11_cast_fp16")];
            tensor<int32, [1]> var_1939_axes_0 = const()[name = string("op_1939_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_1939_cast_fp16 = expand_dims(axes = var_1939_axes_0, x = tail_11_cast_fp16)[name = string("op_1939_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_1940_cast_fp16 = mul(x = k_29_cast_fp16, y = var_1939_cast_fp16)[name = string("op_1940_cast_fp16")];
            bool S_11_transpose_x_1 = const()[name = string("S_11_transpose_x_1"), val = bool(true)];
            bool S_11_transpose_y_1 = const()[name = string("S_11_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_11_cast_fp16 = matmul(transpose_x = S_11_transpose_x_1, transpose_y = S_11_transpose_y_1, x = var_1940_cast_fp16, y = u_13_cast_fp16)[name = string("S_11_cast_fp16")];
            tensor<int32, [4]> q_31_begin_0 = const()[name = string("q_31_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_31_end_0 = const()[name = string("q_31_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> q_31_end_mask_0 = const()[name = string("q_31_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, x = q_27_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<int32, [4]> k_31_begin_0 = const()[name = string("k_31_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_31_end_0 = const()[name = string("k_31_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> k_31_end_mask_0 = const()[name = string("k_31_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_31_cast_fp16 = slice_by_index(begin = k_31_begin_0, end = k_31_end_0, end_mask = k_31_end_mask_0, x = k_27_cast_fp16)[name = string("k_31_cast_fp16")];
            tensor<int32, [4]> kb_15_begin_0 = const()[name = string("kb_15_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_15_end_0 = const()[name = string("kb_15_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> kb_15_end_mask_0 = const()[name = string("kb_15_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_15_cast_fp16 = slice_by_index(begin = kb_15_begin_0, end = kb_15_end_0, end_mask = kb_15_end_mask_0, x = var_1845_cast_fp16)[name = string("kb_15_cast_fp16")];
            tensor<int32, [4]> vb_19_begin_0 = const()[name = string("vb_19_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_19_end_0 = const()[name = string("vb_19_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> vb_19_end_mask_0 = const()[name = string("vb_19_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_19_cast_fp16 = slice_by_index(begin = vb_19_begin_0, end = vb_19_end_0, end_mask = vb_19_end_mask_0, x = vb_15_cast_fp16)[name = string("vb_19_cast_fp16")];
            tensor<int32, [3]> gdec_19_begin_0 = const()[name = string("gdec_19_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_19_end_0 = const()[name = string("gdec_19_end_0"), val = tensor<int32, [3]>([3, 4, 32])];
            tensor<bool, [3]> gdec_19_end_mask_0 = const()[name = string("gdec_19_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_19_cast_fp16 = slice_by_index(begin = gdec_19_begin_0, end = gdec_19_end_0, end_mask = gdec_19_end_mask_0, x = gdec_15_cast_fp16)[name = string("gdec_19_cast_fp16")];
            tensor<int32, [2]> var_2025 = const()[name = string("op_2025"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_2026_cast_fp16 = reshape(shape = var_2025, x = gdec_19_cast_fp16)[name = string("op_2026_cast_fp16")];
            tensor<fp16, [12, 256]> var_2027_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2026_cast_fp16)[name = string("op_2027_cast_fp16")];
            tensor<int32, [4]> var_2032 = const()[name = string("op_2032"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_15_cast_fp16 = reshape(shape = var_2032, x = var_2027_cast_fp16)[name = string("P_15_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2034_cast_fp16 = exp(x = P_15_cast_fp16)[name = string("op_2034_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_15_cast_fp16 = mul(x = var_2034_cast_fp16, y = tril_incl_to_fp16)[name = string("D_15_cast_fp16")];
            bool kk_15_transpose_x_1 = const()[name = string("kk_15_transpose_x_1"), val = bool(false)];
            bool kk_15_transpose_y_1 = const()[name = string("kk_15_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_15_cast_fp16 = matmul(transpose_x = kk_15_transpose_x_1, transpose_y = kk_15_transpose_y_1, x = kb_15_cast_fp16, y = k_31_cast_fp16)[name = string("kk_15_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_2040_cast_fp16 = mul(x = kk_15_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_2040_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2041_cast_fp16 = mul(x = var_2040_cast_fp16, y = D_15_cast_fp16)[name = string("op_2041_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_15_cast_fp16 = mul(x = var_2041_cast_fp16, y = cum_tail_to_fp16)[name = string("N_15_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2043_cast_fp16 = mul(x = D_15_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2043_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_15_cast_fp16 = mul(x = kk_15_cast_fp16, y = var_2043_cast_fp16)[name = string("Na_15_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_57_cast_fp16 = add(x = eye_to_fp16, y = N_15_cast_fp16)[name = string("Minv_57_cast_fp16")];
            bool Np_43_transpose_x_0 = const()[name = string("Np_43_transpose_x_0"), val = bool(false)];
            bool Np_43_transpose_y_0 = const()[name = string("Np_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_43_cast_fp16 = matmul(transpose_x = Np_43_transpose_x_0, transpose_y = Np_43_transpose_y_0, x = N_15_cast_fp16, y = Na_15_cast_fp16)[name = string("Np_43_cast_fp16")];
            bool Npa_29_transpose_x_0 = const()[name = string("Npa_29_transpose_x_0"), val = bool(false)];
            bool Npa_29_transpose_y_0 = const()[name = string("Npa_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_29_cast_fp16 = matmul(transpose_x = Npa_29_transpose_x_0, transpose_y = Npa_29_transpose_y_0, x = Na_15_cast_fp16, y = N_15_cast_fp16)[name = string("Npa_29_cast_fp16")];
            bool var_2049_transpose_x_0 = const()[name = string("op_2049_transpose_x_0"), val = bool(false)];
            bool var_2049_transpose_y_0 = const()[name = string("op_2049_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2049_cast_fp16 = matmul(transpose_x = var_2049_transpose_x_0, transpose_y = var_2049_transpose_y_0, x = Minv_57_cast_fp16, y = Np_43_cast_fp16)[name = string("op_2049_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_59_cast_fp16 = add(x = Minv_57_cast_fp16, y = var_2049_cast_fp16)[name = string("Minv_59_cast_fp16")];
            bool Np_45_transpose_x_0 = const()[name = string("Np_45_transpose_x_0"), val = bool(false)];
            bool Np_45_transpose_y_0 = const()[name = string("Np_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_45_cast_fp16 = matmul(transpose_x = Np_45_transpose_x_0, transpose_y = Np_45_transpose_y_0, x = Np_43_cast_fp16, y = Npa_29_cast_fp16)[name = string("Np_45_cast_fp16")];
            bool Npa_31_transpose_x_0 = const()[name = string("Npa_31_transpose_x_0"), val = bool(false)];
            bool Npa_31_transpose_y_0 = const()[name = string("Npa_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_31_cast_fp16 = matmul(transpose_x = Npa_31_transpose_x_0, transpose_y = Npa_31_transpose_y_0, x = Npa_29_cast_fp16, y = Np_43_cast_fp16)[name = string("Npa_31_cast_fp16")];
            bool var_2054_transpose_x_0 = const()[name = string("op_2054_transpose_x_0"), val = bool(false)];
            bool var_2054_transpose_y_0 = const()[name = string("op_2054_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2054_cast_fp16 = matmul(transpose_x = var_2054_transpose_x_0, transpose_y = var_2054_transpose_y_0, x = Minv_59_cast_fp16, y = Np_45_cast_fp16)[name = string("op_2054_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_61_cast_fp16 = add(x = Minv_59_cast_fp16, y = var_2054_cast_fp16)[name = string("Minv_61_cast_fp16")];
            bool Np_47_transpose_x_0 = const()[name = string("Np_47_transpose_x_0"), val = bool(false)];
            bool Np_47_transpose_y_0 = const()[name = string("Np_47_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_47_cast_fp16 = matmul(transpose_x = Np_47_transpose_x_0, transpose_y = Np_47_transpose_y_0, x = Np_45_cast_fp16, y = Npa_31_cast_fp16)[name = string("Np_47_cast_fp16")];
            bool var_2058_transpose_x_0 = const()[name = string("op_2058_transpose_x_0"), val = bool(false)];
            bool var_2058_transpose_y_0 = const()[name = string("op_2058_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2058_cast_fp16 = matmul(transpose_x = var_2058_transpose_x_0, transpose_y = var_2058_transpose_y_0, x = Minv_61_cast_fp16, y = Np_47_cast_fp16)[name = string("op_2058_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_63_cast_fp16 = add(x = Minv_61_cast_fp16, y = var_2058_cast_fp16)[name = string("Minv_63_cast_fp16")];
            bool u_15_transpose_x_0 = const()[name = string("u_15_transpose_x_0"), val = bool(false)];
            bool u_15_transpose_y_0 = const()[name = string("u_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_15_cast_fp16 = matmul(transpose_x = u_15_transpose_x_0, transpose_y = u_15_transpose_y_0, x = Minv_63_cast_fp16, y = vb_19_cast_fp16)[name = string("u_15_cast_fp16")];
            bool var_2065_transpose_x_1 = const()[name = string("op_2065_transpose_x_1"), val = bool(false)];
            bool var_2065_transpose_y_1 = const()[name = string("op_2065_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_2065_cast_fp16 = matmul(transpose_x = var_2065_transpose_x_1, transpose_y = var_2065_transpose_y_1, x = q_31_cast_fp16, y = k_31_cast_fp16)[name = string("op_2065_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_15_cast_fp16 = mul(x = var_2065_cast_fp16, y = D_15_cast_fp16)[name = string("A_15_cast_fp16")];
            tensor<fp16, [16]> var_2067_bias_0_to_fp16 = const()[name = string("op_2067_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_2067_cast_fp16 = linear(bias = var_2067_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_19_cast_fp16)[name = string("op_2067_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_11_cast_fp16 = exp(x = var_2067_cast_fp16)[name = string("cum_11_cast_fp16")];
            tensor<int32, [1]> var_2070_axes_0 = const()[name = string("op_2070_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_2070_cast_fp16 = expand_dims(axes = var_2070_axes_0, x = cum_11_cast_fp16)[name = string("op_2070_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2071_cast_fp16 = mul(x = kb_15_cast_fp16, y = var_2070_cast_fp16)[name = string("op_2071_cast_fp16")];
            bool w_15_transpose_x_0 = const()[name = string("w_15_transpose_x_0"), val = bool(false)];
            bool w_15_transpose_y_0 = const()[name = string("w_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_15_cast_fp16 = matmul(transpose_x = w_15_transpose_x_0, transpose_y = w_15_transpose_y_0, x = Minv_63_cast_fp16, y = var_2071_cast_fp16)[name = string("w_15_cast_fp16")];
            bool var_2073_transpose_x_0 = const()[name = string("op_2073_transpose_x_0"), val = bool(false)];
            bool var_2073_transpose_y_0 = const()[name = string("op_2073_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2073_cast_fp16 = matmul(transpose_x = var_2073_transpose_x_0, transpose_y = var_2073_transpose_y_0, x = w_15_cast_fp16, y = S_11_cast_fp16)[name = string("op_2073_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_11_cast_fp16 = sub(x = u_15_cast_fp16, y = var_2073_cast_fp16)[name = string("vnew_11_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2078_cast_fp16 = mul(x = q_31_cast_fp16, y = var_2070_cast_fp16)[name = string("op_2078_cast_fp16")];
            bool var_2079_transpose_x_0 = const()[name = string("op_2079_transpose_x_0"), val = bool(false)];
            bool var_2079_transpose_y_0 = const()[name = string("op_2079_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2079_cast_fp16 = matmul(transpose_x = var_2079_transpose_x_0, transpose_y = var_2079_transpose_y_0, x = var_2078_cast_fp16, y = S_11_cast_fp16)[name = string("op_2079_cast_fp16")];
            bool var_2080_transpose_x_0 = const()[name = string("op_2080_transpose_x_0"), val = bool(false)];
            bool var_2080_transpose_y_0 = const()[name = string("op_2080_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2080_cast_fp16 = matmul(transpose_x = var_2080_transpose_x_0, transpose_y = var_2080_transpose_y_0, x = A_15_cast_fp16, y = vnew_11_cast_fp16)[name = string("op_2080_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2082_cast_fp16 = add(x = var_2079_cast_fp16, y = var_2080_cast_fp16)[name = string("op_2082_cast_fp16")];
            tensor<fp16, [16]> var_2083_bias_0_to_fp16 = const()[name = string("op_2083_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_2083_cast_fp16 = linear(bias = var_2083_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_19_cast_fp16)[name = string("op_2083_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_13_cast_fp16 = exp(x = var_2083_cast_fp16)[name = string("tail_13_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_25_to_fp16 = const()[name = string("transpose_25_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_2085_bias_0_to_fp16 = const()[name = string("op_2085_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_2085_cast_fp16 = linear(bias = var_2085_bias_0_to_fp16, weight = transpose_25_to_fp16, x = gdec_19_cast_fp16)[name = string("op_2085_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_2086_cast_fp16 = exp(x = var_2085_cast_fp16)[name = string("op_2086_cast_fp16")];
            tensor<int32, [1]> last_9_axes_0 = const()[name = string("last_9_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_9_cast_fp16 = expand_dims(axes = last_9_axes_0, x = var_2086_cast_fp16)[name = string("last_9_cast_fp16")];
            tensor<int32, [1]> var_2090_axes_0 = const()[name = string("op_2090_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_2090_cast_fp16 = expand_dims(axes = var_2090_axes_0, x = tail_13_cast_fp16)[name = string("op_2090_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2091_cast_fp16 = mul(x = k_31_cast_fp16, y = var_2090_cast_fp16)[name = string("op_2091_cast_fp16")];
            bool upd_9_transpose_x_1 = const()[name = string("upd_9_transpose_x_1"), val = bool(true)];
            bool upd_9_transpose_y_1 = const()[name = string("upd_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_9_cast_fp16 = matmul(transpose_x = upd_9_transpose_x_1, transpose_y = upd_9_transpose_y_1, x = var_2091_cast_fp16, y = vnew_11_cast_fp16)[name = string("upd_9_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_2096_cast_fp16 = mul(x = S_11_cast_fp16, y = last_9_cast_fp16)[name = string("op_2096_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_13_cast_fp16 = add(x = var_2096_cast_fp16, y = upd_9_cast_fp16)[name = string("S_13_cast_fp16")];
            tensor<int32, [4]> q_33_begin_0 = const()[name = string("q_33_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_33_end_0 = const()[name = string("q_33_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> q_33_end_mask_0 = const()[name = string("q_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_33_cast_fp16 = slice_by_index(begin = q_33_begin_0, end = q_33_end_0, end_mask = q_33_end_mask_0, x = q_27_cast_fp16)[name = string("q_33_cast_fp16")];
            tensor<int32, [4]> k_33_begin_0 = const()[name = string("k_33_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_33_end_0 = const()[name = string("k_33_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> k_33_end_mask_0 = const()[name = string("k_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, x = k_27_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<int32, [4]> kb_17_begin_0 = const()[name = string("kb_17_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_17_end_0 = const()[name = string("kb_17_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> kb_17_end_mask_0 = const()[name = string("kb_17_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_17_cast_fp16 = slice_by_index(begin = kb_17_begin_0, end = kb_17_end_0, end_mask = kb_17_end_mask_0, x = var_1845_cast_fp16)[name = string("kb_17_cast_fp16")];
            tensor<int32, [4]> vb_21_begin_0 = const()[name = string("vb_21_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_21_end_0 = const()[name = string("vb_21_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> vb_21_end_mask_0 = const()[name = string("vb_21_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_21_cast_fp16 = slice_by_index(begin = vb_21_begin_0, end = vb_21_end_0, end_mask = vb_21_end_mask_0, x = vb_15_cast_fp16)[name = string("vb_21_cast_fp16")];
            tensor<int32, [3]> gdec_21_begin_0 = const()[name = string("gdec_21_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_21_end_0 = const()[name = string("gdec_21_end_0"), val = tensor<int32, [3]>([3, 4, 48])];
            tensor<bool, [3]> gdec_21_end_mask_0 = const()[name = string("gdec_21_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_21_cast_fp16 = slice_by_index(begin = gdec_21_begin_0, end = gdec_21_end_0, end_mask = gdec_21_end_mask_0, x = gdec_15_cast_fp16)[name = string("gdec_21_cast_fp16")];
            tensor<int32, [2]> var_2179 = const()[name = string("op_2179"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_2180_cast_fp16 = reshape(shape = var_2179, x = gdec_21_cast_fp16)[name = string("op_2180_cast_fp16")];
            tensor<fp16, [12, 256]> var_2181_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2180_cast_fp16)[name = string("op_2181_cast_fp16")];
            tensor<int32, [4]> var_2186 = const()[name = string("op_2186"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_17_cast_fp16 = reshape(shape = var_2186, x = var_2181_cast_fp16)[name = string("P_17_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2188_cast_fp16 = exp(x = P_17_cast_fp16)[name = string("op_2188_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_17_cast_fp16 = mul(x = var_2188_cast_fp16, y = tril_incl_to_fp16)[name = string("D_17_cast_fp16")];
            bool kk_17_transpose_x_1 = const()[name = string("kk_17_transpose_x_1"), val = bool(false)];
            bool kk_17_transpose_y_1 = const()[name = string("kk_17_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_17_cast_fp16 = matmul(transpose_x = kk_17_transpose_x_1, transpose_y = kk_17_transpose_y_1, x = kb_17_cast_fp16, y = k_33_cast_fp16)[name = string("kk_17_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_2194_cast_fp16 = mul(x = kk_17_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_2194_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2195_cast_fp16 = mul(x = var_2194_cast_fp16, y = D_17_cast_fp16)[name = string("op_2195_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_17_cast_fp16 = mul(x = var_2195_cast_fp16, y = cum_tail_to_fp16)[name = string("N_17_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2197_cast_fp16 = mul(x = D_17_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2197_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_17_cast_fp16 = mul(x = kk_17_cast_fp16, y = var_2197_cast_fp16)[name = string("Na_17_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_65_cast_fp16 = add(x = eye_to_fp16, y = N_17_cast_fp16)[name = string("Minv_65_cast_fp16")];
            bool Np_49_transpose_x_0 = const()[name = string("Np_49_transpose_x_0"), val = bool(false)];
            bool Np_49_transpose_y_0 = const()[name = string("Np_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_49_cast_fp16 = matmul(transpose_x = Np_49_transpose_x_0, transpose_y = Np_49_transpose_y_0, x = N_17_cast_fp16, y = Na_17_cast_fp16)[name = string("Np_49_cast_fp16")];
            bool Npa_33_transpose_x_0 = const()[name = string("Npa_33_transpose_x_0"), val = bool(false)];
            bool Npa_33_transpose_y_0 = const()[name = string("Npa_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_33_cast_fp16 = matmul(transpose_x = Npa_33_transpose_x_0, transpose_y = Npa_33_transpose_y_0, x = Na_17_cast_fp16, y = N_17_cast_fp16)[name = string("Npa_33_cast_fp16")];
            bool var_2203_transpose_x_0 = const()[name = string("op_2203_transpose_x_0"), val = bool(false)];
            bool var_2203_transpose_y_0 = const()[name = string("op_2203_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2203_cast_fp16 = matmul(transpose_x = var_2203_transpose_x_0, transpose_y = var_2203_transpose_y_0, x = Minv_65_cast_fp16, y = Np_49_cast_fp16)[name = string("op_2203_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_67_cast_fp16 = add(x = Minv_65_cast_fp16, y = var_2203_cast_fp16)[name = string("Minv_67_cast_fp16")];
            bool Np_51_transpose_x_0 = const()[name = string("Np_51_transpose_x_0"), val = bool(false)];
            bool Np_51_transpose_y_0 = const()[name = string("Np_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_51_cast_fp16 = matmul(transpose_x = Np_51_transpose_x_0, transpose_y = Np_51_transpose_y_0, x = Np_49_cast_fp16, y = Npa_33_cast_fp16)[name = string("Np_51_cast_fp16")];
            bool Npa_35_transpose_x_0 = const()[name = string("Npa_35_transpose_x_0"), val = bool(false)];
            bool Npa_35_transpose_y_0 = const()[name = string("Npa_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_35_cast_fp16 = matmul(transpose_x = Npa_35_transpose_x_0, transpose_y = Npa_35_transpose_y_0, x = Npa_33_cast_fp16, y = Np_49_cast_fp16)[name = string("Npa_35_cast_fp16")];
            bool var_2208_transpose_x_0 = const()[name = string("op_2208_transpose_x_0"), val = bool(false)];
            bool var_2208_transpose_y_0 = const()[name = string("op_2208_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2208_cast_fp16 = matmul(transpose_x = var_2208_transpose_x_0, transpose_y = var_2208_transpose_y_0, x = Minv_67_cast_fp16, y = Np_51_cast_fp16)[name = string("op_2208_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_69_cast_fp16 = add(x = Minv_67_cast_fp16, y = var_2208_cast_fp16)[name = string("Minv_69_cast_fp16")];
            bool Np_53_transpose_x_0 = const()[name = string("Np_53_transpose_x_0"), val = bool(false)];
            bool Np_53_transpose_y_0 = const()[name = string("Np_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_53_cast_fp16 = matmul(transpose_x = Np_53_transpose_x_0, transpose_y = Np_53_transpose_y_0, x = Np_51_cast_fp16, y = Npa_35_cast_fp16)[name = string("Np_53_cast_fp16")];
            bool var_2212_transpose_x_0 = const()[name = string("op_2212_transpose_x_0"), val = bool(false)];
            bool var_2212_transpose_y_0 = const()[name = string("op_2212_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2212_cast_fp16 = matmul(transpose_x = var_2212_transpose_x_0, transpose_y = var_2212_transpose_y_0, x = Minv_69_cast_fp16, y = Np_53_cast_fp16)[name = string("op_2212_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_71_cast_fp16 = add(x = Minv_69_cast_fp16, y = var_2212_cast_fp16)[name = string("Minv_71_cast_fp16")];
            bool u_17_transpose_x_0 = const()[name = string("u_17_transpose_x_0"), val = bool(false)];
            bool u_17_transpose_y_0 = const()[name = string("u_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_17_cast_fp16 = matmul(transpose_x = u_17_transpose_x_0, transpose_y = u_17_transpose_y_0, x = Minv_71_cast_fp16, y = vb_21_cast_fp16)[name = string("u_17_cast_fp16")];
            bool var_2219_transpose_x_1 = const()[name = string("op_2219_transpose_x_1"), val = bool(false)];
            bool var_2219_transpose_y_1 = const()[name = string("op_2219_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_2219_cast_fp16 = matmul(transpose_x = var_2219_transpose_x_1, transpose_y = var_2219_transpose_y_1, x = q_33_cast_fp16, y = k_33_cast_fp16)[name = string("op_2219_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_17_cast_fp16 = mul(x = var_2219_cast_fp16, y = D_17_cast_fp16)[name = string("A_17_cast_fp16")];
            tensor<fp16, [16]> var_2221_bias_0_to_fp16 = const()[name = string("op_2221_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_2221_cast_fp16 = linear(bias = var_2221_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_21_cast_fp16)[name = string("op_2221_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_13_cast_fp16 = exp(x = var_2221_cast_fp16)[name = string("cum_13_cast_fp16")];
            tensor<int32, [1]> var_2224_axes_0 = const()[name = string("op_2224_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_2224_cast_fp16 = expand_dims(axes = var_2224_axes_0, x = cum_13_cast_fp16)[name = string("op_2224_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2225_cast_fp16 = mul(x = kb_17_cast_fp16, y = var_2224_cast_fp16)[name = string("op_2225_cast_fp16")];
            bool w_17_transpose_x_0 = const()[name = string("w_17_transpose_x_0"), val = bool(false)];
            bool w_17_transpose_y_0 = const()[name = string("w_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_17_cast_fp16 = matmul(transpose_x = w_17_transpose_x_0, transpose_y = w_17_transpose_y_0, x = Minv_71_cast_fp16, y = var_2225_cast_fp16)[name = string("w_17_cast_fp16")];
            bool var_2227_transpose_x_0 = const()[name = string("op_2227_transpose_x_0"), val = bool(false)];
            bool var_2227_transpose_y_0 = const()[name = string("op_2227_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2227_cast_fp16 = matmul(transpose_x = var_2227_transpose_x_0, transpose_y = var_2227_transpose_y_0, x = w_17_cast_fp16, y = S_13_cast_fp16)[name = string("op_2227_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_13_cast_fp16 = sub(x = u_17_cast_fp16, y = var_2227_cast_fp16)[name = string("vnew_13_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2232_cast_fp16 = mul(x = q_33_cast_fp16, y = var_2224_cast_fp16)[name = string("op_2232_cast_fp16")];
            bool var_2233_transpose_x_0 = const()[name = string("op_2233_transpose_x_0"), val = bool(false)];
            bool var_2233_transpose_y_0 = const()[name = string("op_2233_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2233_cast_fp16 = matmul(transpose_x = var_2233_transpose_x_0, transpose_y = var_2233_transpose_y_0, x = var_2232_cast_fp16, y = S_13_cast_fp16)[name = string("op_2233_cast_fp16")];
            bool var_2234_transpose_x_0 = const()[name = string("op_2234_transpose_x_0"), val = bool(false)];
            bool var_2234_transpose_y_0 = const()[name = string("op_2234_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2234_cast_fp16 = matmul(transpose_x = var_2234_transpose_x_0, transpose_y = var_2234_transpose_y_0, x = A_17_cast_fp16, y = vnew_13_cast_fp16)[name = string("op_2234_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2236_cast_fp16 = add(x = var_2233_cast_fp16, y = var_2234_cast_fp16)[name = string("op_2236_cast_fp16")];
            tensor<fp16, [16]> var_2237_bias_0_to_fp16 = const()[name = string("op_2237_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_2237_cast_fp16 = linear(bias = var_2237_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_21_cast_fp16)[name = string("op_2237_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_15_cast_fp16 = exp(x = var_2237_cast_fp16)[name = string("tail_15_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_29_to_fp16 = const()[name = string("transpose_29_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_2239_bias_0_to_fp16 = const()[name = string("op_2239_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_2239_cast_fp16 = linear(bias = var_2239_bias_0_to_fp16, weight = transpose_29_to_fp16, x = gdec_21_cast_fp16)[name = string("op_2239_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_2240_cast_fp16 = exp(x = var_2239_cast_fp16)[name = string("op_2240_cast_fp16")];
            tensor<int32, [1]> last_11_axes_0 = const()[name = string("last_11_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_11_cast_fp16 = expand_dims(axes = last_11_axes_0, x = var_2240_cast_fp16)[name = string("last_11_cast_fp16")];
            tensor<int32, [1]> var_2244_axes_0 = const()[name = string("op_2244_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_2244_cast_fp16 = expand_dims(axes = var_2244_axes_0, x = tail_15_cast_fp16)[name = string("op_2244_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2245_cast_fp16 = mul(x = k_33_cast_fp16, y = var_2244_cast_fp16)[name = string("op_2245_cast_fp16")];
            bool upd_11_transpose_x_1 = const()[name = string("upd_11_transpose_x_1"), val = bool(true)];
            bool upd_11_transpose_y_1 = const()[name = string("upd_11_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_11_cast_fp16 = matmul(transpose_x = upd_11_transpose_x_1, transpose_y = upd_11_transpose_y_1, x = var_2245_cast_fp16, y = vnew_13_cast_fp16)[name = string("upd_11_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_2250_cast_fp16 = mul(x = S_13_cast_fp16, y = last_11_cast_fp16)[name = string("op_2250_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_15_cast_fp16 = add(x = var_2250_cast_fp16, y = upd_11_cast_fp16)[name = string("S_15_cast_fp16")];
            tensor<int32, [4]> q_35_begin_0 = const()[name = string("q_35_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_35_end_0 = const()[name = string("q_35_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> q_35_end_mask_0 = const()[name = string("q_35_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_35_cast_fp16 = slice_by_index(begin = q_35_begin_0, end = q_35_end_0, end_mask = q_35_end_mask_0, x = q_27_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<int32, [4]> k_35_begin_0 = const()[name = string("k_35_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_35_end_0 = const()[name = string("k_35_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> k_35_end_mask_0 = const()[name = string("k_35_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_35_cast_fp16 = slice_by_index(begin = k_35_begin_0, end = k_35_end_0, end_mask = k_35_end_mask_0, x = k_27_cast_fp16)[name = string("k_35_cast_fp16")];
            tensor<int32, [4]> kb_19_begin_0 = const()[name = string("kb_19_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_19_end_0 = const()[name = string("kb_19_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> kb_19_end_mask_0 = const()[name = string("kb_19_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_19_cast_fp16 = slice_by_index(begin = kb_19_begin_0, end = kb_19_end_0, end_mask = kb_19_end_mask_0, x = var_1845_cast_fp16)[name = string("kb_19_cast_fp16")];
            tensor<int32, [4]> vb_23_begin_0 = const()[name = string("vb_23_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_23_end_0 = const()[name = string("vb_23_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> vb_23_end_mask_0 = const()[name = string("vb_23_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_23_cast_fp16 = slice_by_index(begin = vb_23_begin_0, end = vb_23_end_0, end_mask = vb_23_end_mask_0, x = vb_15_cast_fp16)[name = string("vb_23_cast_fp16")];
            tensor<int32, [3]> gdec_23_begin_0 = const()[name = string("gdec_23_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_23_end_0 = const()[name = string("gdec_23_end_0"), val = tensor<int32, [3]>([3, 4, 64])];
            tensor<bool, [3]> gdec_23_end_mask_0 = const()[name = string("gdec_23_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_23_cast_fp16 = slice_by_index(begin = gdec_23_begin_0, end = gdec_23_end_0, end_mask = gdec_23_end_mask_0, x = gdec_15_cast_fp16)[name = string("gdec_23_cast_fp16")];
            tensor<int32, [2]> var_2333 = const()[name = string("op_2333"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_2334_cast_fp16 = reshape(shape = var_2333, x = gdec_23_cast_fp16)[name = string("op_2334_cast_fp16")];
            tensor<fp16, [12, 256]> var_2335_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2334_cast_fp16)[name = string("op_2335_cast_fp16")];
            tensor<int32, [4]> var_2340 = const()[name = string("op_2340"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_19_cast_fp16 = reshape(shape = var_2340, x = var_2335_cast_fp16)[name = string("P_19_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2342_cast_fp16 = exp(x = P_19_cast_fp16)[name = string("op_2342_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_19_cast_fp16 = mul(x = var_2342_cast_fp16, y = tril_incl_to_fp16)[name = string("D_19_cast_fp16")];
            bool kk_19_transpose_x_1 = const()[name = string("kk_19_transpose_x_1"), val = bool(false)];
            bool kk_19_transpose_y_1 = const()[name = string("kk_19_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_19_cast_fp16 = matmul(transpose_x = kk_19_transpose_x_1, transpose_y = kk_19_transpose_y_1, x = kb_19_cast_fp16, y = k_35_cast_fp16)[name = string("kk_19_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_2348_cast_fp16 = mul(x = kk_19_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_2348_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2349_cast_fp16 = mul(x = var_2348_cast_fp16, y = D_19_cast_fp16)[name = string("op_2349_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_19_cast_fp16 = mul(x = var_2349_cast_fp16, y = cum_tail_to_fp16)[name = string("N_19_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2351_cast_fp16 = mul(x = D_19_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2351_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_19_cast_fp16 = mul(x = kk_19_cast_fp16, y = var_2351_cast_fp16)[name = string("Na_19_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_73_cast_fp16 = add(x = eye_to_fp16, y = N_19_cast_fp16)[name = string("Minv_73_cast_fp16")];
            bool Np_55_transpose_x_0 = const()[name = string("Np_55_transpose_x_0"), val = bool(false)];
            bool Np_55_transpose_y_0 = const()[name = string("Np_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_55_cast_fp16 = matmul(transpose_x = Np_55_transpose_x_0, transpose_y = Np_55_transpose_y_0, x = N_19_cast_fp16, y = Na_19_cast_fp16)[name = string("Np_55_cast_fp16")];
            bool Npa_37_transpose_x_0 = const()[name = string("Npa_37_transpose_x_0"), val = bool(false)];
            bool Npa_37_transpose_y_0 = const()[name = string("Npa_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_37_cast_fp16 = matmul(transpose_x = Npa_37_transpose_x_0, transpose_y = Npa_37_transpose_y_0, x = Na_19_cast_fp16, y = N_19_cast_fp16)[name = string("Npa_37_cast_fp16")];
            bool var_2357_transpose_x_0 = const()[name = string("op_2357_transpose_x_0"), val = bool(false)];
            bool var_2357_transpose_y_0 = const()[name = string("op_2357_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2357_cast_fp16 = matmul(transpose_x = var_2357_transpose_x_0, transpose_y = var_2357_transpose_y_0, x = Minv_73_cast_fp16, y = Np_55_cast_fp16)[name = string("op_2357_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_75_cast_fp16 = add(x = Minv_73_cast_fp16, y = var_2357_cast_fp16)[name = string("Minv_75_cast_fp16")];
            bool Np_57_transpose_x_0 = const()[name = string("Np_57_transpose_x_0"), val = bool(false)];
            bool Np_57_transpose_y_0 = const()[name = string("Np_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_57_cast_fp16 = matmul(transpose_x = Np_57_transpose_x_0, transpose_y = Np_57_transpose_y_0, x = Np_55_cast_fp16, y = Npa_37_cast_fp16)[name = string("Np_57_cast_fp16")];
            bool Npa_39_transpose_x_0 = const()[name = string("Npa_39_transpose_x_0"), val = bool(false)];
            bool Npa_39_transpose_y_0 = const()[name = string("Npa_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_39_cast_fp16 = matmul(transpose_x = Npa_39_transpose_x_0, transpose_y = Npa_39_transpose_y_0, x = Npa_37_cast_fp16, y = Np_55_cast_fp16)[name = string("Npa_39_cast_fp16")];
            bool var_2362_transpose_x_0 = const()[name = string("op_2362_transpose_x_0"), val = bool(false)];
            bool var_2362_transpose_y_0 = const()[name = string("op_2362_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2362_cast_fp16 = matmul(transpose_x = var_2362_transpose_x_0, transpose_y = var_2362_transpose_y_0, x = Minv_75_cast_fp16, y = Np_57_cast_fp16)[name = string("op_2362_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_77_cast_fp16 = add(x = Minv_75_cast_fp16, y = var_2362_cast_fp16)[name = string("Minv_77_cast_fp16")];
            bool Np_59_transpose_x_0 = const()[name = string("Np_59_transpose_x_0"), val = bool(false)];
            bool Np_59_transpose_y_0 = const()[name = string("Np_59_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_59_cast_fp16 = matmul(transpose_x = Np_59_transpose_x_0, transpose_y = Np_59_transpose_y_0, x = Np_57_cast_fp16, y = Npa_39_cast_fp16)[name = string("Np_59_cast_fp16")];
            bool var_2366_transpose_x_0 = const()[name = string("op_2366_transpose_x_0"), val = bool(false)];
            bool var_2366_transpose_y_0 = const()[name = string("op_2366_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2366_cast_fp16 = matmul(transpose_x = var_2366_transpose_x_0, transpose_y = var_2366_transpose_y_0, x = Minv_77_cast_fp16, y = Np_59_cast_fp16)[name = string("op_2366_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_79_cast_fp16 = add(x = Minv_77_cast_fp16, y = var_2366_cast_fp16)[name = string("Minv_79_cast_fp16")];
            bool u_19_transpose_x_0 = const()[name = string("u_19_transpose_x_0"), val = bool(false)];
            bool u_19_transpose_y_0 = const()[name = string("u_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_19_cast_fp16 = matmul(transpose_x = u_19_transpose_x_0, transpose_y = u_19_transpose_y_0, x = Minv_79_cast_fp16, y = vb_23_cast_fp16)[name = string("u_19_cast_fp16")];
            bool var_2373_transpose_x_1 = const()[name = string("op_2373_transpose_x_1"), val = bool(false)];
            bool var_2373_transpose_y_1 = const()[name = string("op_2373_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_2373_cast_fp16 = matmul(transpose_x = var_2373_transpose_x_1, transpose_y = var_2373_transpose_y_1, x = q_35_cast_fp16, y = k_35_cast_fp16)[name = string("op_2373_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_19_cast_fp16 = mul(x = var_2373_cast_fp16, y = D_19_cast_fp16)[name = string("A_19_cast_fp16")];
            tensor<fp16, [16]> var_2375_bias_0_to_fp16 = const()[name = string("op_2375_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_2375_cast_fp16 = linear(bias = var_2375_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_23_cast_fp16)[name = string("op_2375_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_15_cast_fp16 = exp(x = var_2375_cast_fp16)[name = string("cum_15_cast_fp16")];
            tensor<int32, [1]> var_2378_axes_0 = const()[name = string("op_2378_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_2378_cast_fp16 = expand_dims(axes = var_2378_axes_0, x = cum_15_cast_fp16)[name = string("op_2378_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2379_cast_fp16 = mul(x = kb_19_cast_fp16, y = var_2378_cast_fp16)[name = string("op_2379_cast_fp16")];
            bool w_19_transpose_x_0 = const()[name = string("w_19_transpose_x_0"), val = bool(false)];
            bool w_19_transpose_y_0 = const()[name = string("w_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_19_cast_fp16 = matmul(transpose_x = w_19_transpose_x_0, transpose_y = w_19_transpose_y_0, x = Minv_79_cast_fp16, y = var_2379_cast_fp16)[name = string("w_19_cast_fp16")];
            bool var_2381_transpose_x_0 = const()[name = string("op_2381_transpose_x_0"), val = bool(false)];
            bool var_2381_transpose_y_0 = const()[name = string("op_2381_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2381_cast_fp16 = matmul(transpose_x = var_2381_transpose_x_0, transpose_y = var_2381_transpose_y_0, x = w_19_cast_fp16, y = S_15_cast_fp16)[name = string("op_2381_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_15_cast_fp16 = sub(x = u_19_cast_fp16, y = var_2381_cast_fp16)[name = string("vnew_15_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2386_cast_fp16 = mul(x = q_35_cast_fp16, y = var_2378_cast_fp16)[name = string("op_2386_cast_fp16")];
            bool var_2387_transpose_x_0 = const()[name = string("op_2387_transpose_x_0"), val = bool(false)];
            bool var_2387_transpose_y_0 = const()[name = string("op_2387_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2387_cast_fp16 = matmul(transpose_x = var_2387_transpose_x_0, transpose_y = var_2387_transpose_y_0, x = var_2386_cast_fp16, y = S_15_cast_fp16)[name = string("op_2387_cast_fp16")];
            bool var_2388_transpose_x_0 = const()[name = string("op_2388_transpose_x_0"), val = bool(false)];
            bool var_2388_transpose_y_0 = const()[name = string("op_2388_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2388_cast_fp16 = matmul(transpose_x = var_2388_transpose_x_0, transpose_y = var_2388_transpose_y_0, x = A_19_cast_fp16, y = vnew_15_cast_fp16)[name = string("op_2388_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2390_cast_fp16 = add(x = var_2387_cast_fp16, y = var_2388_cast_fp16)[name = string("op_2390_cast_fp16")];
            tensor<fp16, [16]> var_2391_bias_0_to_fp16 = const()[name = string("op_2391_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_2391_cast_fp16 = linear(bias = var_2391_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_23_cast_fp16)[name = string("op_2391_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_17_cast_fp16 = exp(x = var_2391_cast_fp16)[name = string("tail_17_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_33_to_fp16 = const()[name = string("transpose_33_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_2393_bias_0_to_fp16 = const()[name = string("op_2393_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_2393_cast_fp16 = linear(bias = var_2393_bias_0_to_fp16, weight = transpose_33_to_fp16, x = gdec_23_cast_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_2394_cast_fp16 = exp(x = var_2393_cast_fp16)[name = string("op_2394_cast_fp16")];
            tensor<int32, [1]> last_13_axes_0 = const()[name = string("last_13_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_13_cast_fp16 = expand_dims(axes = last_13_axes_0, x = var_2394_cast_fp16)[name = string("last_13_cast_fp16")];
            tensor<int32, [1]> var_2398_axes_0 = const()[name = string("op_2398_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_2398_cast_fp16 = expand_dims(axes = var_2398_axes_0, x = tail_17_cast_fp16)[name = string("op_2398_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2399_cast_fp16 = mul(x = k_35_cast_fp16, y = var_2398_cast_fp16)[name = string("op_2399_cast_fp16")];
            bool upd_13_transpose_x_1 = const()[name = string("upd_13_transpose_x_1"), val = bool(true)];
            bool upd_13_transpose_y_1 = const()[name = string("upd_13_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_13_cast_fp16 = matmul(transpose_x = upd_13_transpose_x_1, transpose_y = upd_13_transpose_y_1, x = var_2399_cast_fp16, y = vnew_15_cast_fp16)[name = string("upd_13_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_2404_cast_fp16 = mul(x = S_15_cast_fp16, y = last_13_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_17_cast_fp16 = add(x = var_2404_cast_fp16, y = upd_13_cast_fp16)[name = string("S_17_cast_fp16")];
            tensor<int32, [4]> q_37_begin_0 = const()[name = string("q_37_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> q_37_end_0 = const()[name = string("q_37_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> q_37_end_mask_0 = const()[name = string("q_37_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, x = q_27_cast_fp16)[name = string("q_37_cast_fp16")];
            tensor<int32, [4]> k_37_begin_0 = const()[name = string("k_37_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> k_37_end_0 = const()[name = string("k_37_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> k_37_end_mask_0 = const()[name = string("k_37_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, x = k_27_cast_fp16)[name = string("k_37_cast_fp16")];
            tensor<int32, [4]> kb_21_begin_0 = const()[name = string("kb_21_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> kb_21_end_0 = const()[name = string("kb_21_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> kb_21_end_mask_0 = const()[name = string("kb_21_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_21_cast_fp16 = slice_by_index(begin = kb_21_begin_0, end = kb_21_end_0, end_mask = kb_21_end_mask_0, x = var_1845_cast_fp16)[name = string("kb_21_cast_fp16")];
            tensor<int32, [4]> vb_25_begin_0 = const()[name = string("vb_25_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> vb_25_end_0 = const()[name = string("vb_25_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> vb_25_end_mask_0 = const()[name = string("vb_25_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_25_cast_fp16 = slice_by_index(begin = vb_25_begin_0, end = vb_25_end_0, end_mask = vb_25_end_mask_0, x = vb_15_cast_fp16)[name = string("vb_25_cast_fp16")];
            tensor<int32, [3]> gdec_25_begin_0 = const()[name = string("gdec_25_begin_0"), val = tensor<int32, [3]>([0, 0, 64])];
            tensor<int32, [3]> gdec_25_end_0 = const()[name = string("gdec_25_end_0"), val = tensor<int32, [3]>([3, 4, 80])];
            tensor<bool, [3]> gdec_25_end_mask_0 = const()[name = string("gdec_25_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_25_cast_fp16 = slice_by_index(begin = gdec_25_begin_0, end = gdec_25_end_0, end_mask = gdec_25_end_mask_0, x = gdec_15_cast_fp16)[name = string("gdec_25_cast_fp16")];
            tensor<int32, [2]> var_2487 = const()[name = string("op_2487"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_2488_cast_fp16 = reshape(shape = var_2487, x = gdec_25_cast_fp16)[name = string("op_2488_cast_fp16")];
            tensor<fp16, [12, 256]> var_2489_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2488_cast_fp16)[name = string("op_2489_cast_fp16")];
            tensor<int32, [4]> var_2494 = const()[name = string("op_2494"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_21_cast_fp16 = reshape(shape = var_2494, x = var_2489_cast_fp16)[name = string("P_21_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2496_cast_fp16 = exp(x = P_21_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_21_cast_fp16 = mul(x = var_2496_cast_fp16, y = tril_incl_to_fp16)[name = string("D_21_cast_fp16")];
            bool kk_21_transpose_x_1 = const()[name = string("kk_21_transpose_x_1"), val = bool(false)];
            bool kk_21_transpose_y_1 = const()[name = string("kk_21_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_21_cast_fp16 = matmul(transpose_x = kk_21_transpose_x_1, transpose_y = kk_21_transpose_y_1, x = kb_21_cast_fp16, y = k_37_cast_fp16)[name = string("kk_21_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_2502_cast_fp16 = mul(x = kk_21_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_2502_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2503_cast_fp16 = mul(x = var_2502_cast_fp16, y = D_21_cast_fp16)[name = string("op_2503_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_21_cast_fp16 = mul(x = var_2503_cast_fp16, y = cum_tail_to_fp16)[name = string("N_21_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2505_cast_fp16 = mul(x = D_21_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2505_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_21_cast_fp16 = mul(x = kk_21_cast_fp16, y = var_2505_cast_fp16)[name = string("Na_21_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_81_cast_fp16 = add(x = eye_to_fp16, y = N_21_cast_fp16)[name = string("Minv_81_cast_fp16")];
            bool Np_61_transpose_x_0 = const()[name = string("Np_61_transpose_x_0"), val = bool(false)];
            bool Np_61_transpose_y_0 = const()[name = string("Np_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_61_cast_fp16 = matmul(transpose_x = Np_61_transpose_x_0, transpose_y = Np_61_transpose_y_0, x = N_21_cast_fp16, y = Na_21_cast_fp16)[name = string("Np_61_cast_fp16")];
            bool Npa_41_transpose_x_0 = const()[name = string("Npa_41_transpose_x_0"), val = bool(false)];
            bool Npa_41_transpose_y_0 = const()[name = string("Npa_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_41_cast_fp16 = matmul(transpose_x = Npa_41_transpose_x_0, transpose_y = Npa_41_transpose_y_0, x = Na_21_cast_fp16, y = N_21_cast_fp16)[name = string("Npa_41_cast_fp16")];
            bool var_2511_transpose_x_0 = const()[name = string("op_2511_transpose_x_0"), val = bool(false)];
            bool var_2511_transpose_y_0 = const()[name = string("op_2511_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2511_cast_fp16 = matmul(transpose_x = var_2511_transpose_x_0, transpose_y = var_2511_transpose_y_0, x = Minv_81_cast_fp16, y = Np_61_cast_fp16)[name = string("op_2511_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_83_cast_fp16 = add(x = Minv_81_cast_fp16, y = var_2511_cast_fp16)[name = string("Minv_83_cast_fp16")];
            bool Np_63_transpose_x_0 = const()[name = string("Np_63_transpose_x_0"), val = bool(false)];
            bool Np_63_transpose_y_0 = const()[name = string("Np_63_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_63_cast_fp16 = matmul(transpose_x = Np_63_transpose_x_0, transpose_y = Np_63_transpose_y_0, x = Np_61_cast_fp16, y = Npa_41_cast_fp16)[name = string("Np_63_cast_fp16")];
            bool Npa_43_transpose_x_0 = const()[name = string("Npa_43_transpose_x_0"), val = bool(false)];
            bool Npa_43_transpose_y_0 = const()[name = string("Npa_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_43_cast_fp16 = matmul(transpose_x = Npa_43_transpose_x_0, transpose_y = Npa_43_transpose_y_0, x = Npa_41_cast_fp16, y = Np_61_cast_fp16)[name = string("Npa_43_cast_fp16")];
            bool var_2516_transpose_x_0 = const()[name = string("op_2516_transpose_x_0"), val = bool(false)];
            bool var_2516_transpose_y_0 = const()[name = string("op_2516_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2516_cast_fp16 = matmul(transpose_x = var_2516_transpose_x_0, transpose_y = var_2516_transpose_y_0, x = Minv_83_cast_fp16, y = Np_63_cast_fp16)[name = string("op_2516_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_85_cast_fp16 = add(x = Minv_83_cast_fp16, y = var_2516_cast_fp16)[name = string("Minv_85_cast_fp16")];
            bool Np_65_transpose_x_0 = const()[name = string("Np_65_transpose_x_0"), val = bool(false)];
            bool Np_65_transpose_y_0 = const()[name = string("Np_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_65_cast_fp16 = matmul(transpose_x = Np_65_transpose_x_0, transpose_y = Np_65_transpose_y_0, x = Np_63_cast_fp16, y = Npa_43_cast_fp16)[name = string("Np_65_cast_fp16")];
            bool var_2520_transpose_x_0 = const()[name = string("op_2520_transpose_x_0"), val = bool(false)];
            bool var_2520_transpose_y_0 = const()[name = string("op_2520_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2520_cast_fp16 = matmul(transpose_x = var_2520_transpose_x_0, transpose_y = var_2520_transpose_y_0, x = Minv_85_cast_fp16, y = Np_65_cast_fp16)[name = string("op_2520_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_87_cast_fp16 = add(x = Minv_85_cast_fp16, y = var_2520_cast_fp16)[name = string("Minv_87_cast_fp16")];
            bool u_21_transpose_x_0 = const()[name = string("u_21_transpose_x_0"), val = bool(false)];
            bool u_21_transpose_y_0 = const()[name = string("u_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_21_cast_fp16 = matmul(transpose_x = u_21_transpose_x_0, transpose_y = u_21_transpose_y_0, x = Minv_87_cast_fp16, y = vb_25_cast_fp16)[name = string("u_21_cast_fp16")];
            bool var_2527_transpose_x_1 = const()[name = string("op_2527_transpose_x_1"), val = bool(false)];
            bool var_2527_transpose_y_1 = const()[name = string("op_2527_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_2527_cast_fp16 = matmul(transpose_x = var_2527_transpose_x_1, transpose_y = var_2527_transpose_y_1, x = q_37_cast_fp16, y = k_37_cast_fp16)[name = string("op_2527_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_21_cast_fp16 = mul(x = var_2527_cast_fp16, y = D_21_cast_fp16)[name = string("A_21_cast_fp16")];
            tensor<fp16, [16]> var_2529_bias_0_to_fp16 = const()[name = string("op_2529_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_2529_cast_fp16 = linear(bias = var_2529_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_25_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_17_cast_fp16 = exp(x = var_2529_cast_fp16)[name = string("cum_17_cast_fp16")];
            tensor<int32, [1]> var_2532_axes_0 = const()[name = string("op_2532_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_2532_cast_fp16 = expand_dims(axes = var_2532_axes_0, x = cum_17_cast_fp16)[name = string("op_2532_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2533_cast_fp16 = mul(x = kb_21_cast_fp16, y = var_2532_cast_fp16)[name = string("op_2533_cast_fp16")];
            bool w_21_transpose_x_0 = const()[name = string("w_21_transpose_x_0"), val = bool(false)];
            bool w_21_transpose_y_0 = const()[name = string("w_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_21_cast_fp16 = matmul(transpose_x = w_21_transpose_x_0, transpose_y = w_21_transpose_y_0, x = Minv_87_cast_fp16, y = var_2533_cast_fp16)[name = string("w_21_cast_fp16")];
            bool var_2535_transpose_x_0 = const()[name = string("op_2535_transpose_x_0"), val = bool(false)];
            bool var_2535_transpose_y_0 = const()[name = string("op_2535_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2535_cast_fp16 = matmul(transpose_x = var_2535_transpose_x_0, transpose_y = var_2535_transpose_y_0, x = w_21_cast_fp16, y = S_17_cast_fp16)[name = string("op_2535_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_17_cast_fp16 = sub(x = u_21_cast_fp16, y = var_2535_cast_fp16)[name = string("vnew_17_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2540_cast_fp16 = mul(x = q_37_cast_fp16, y = var_2532_cast_fp16)[name = string("op_2540_cast_fp16")];
            bool var_2541_transpose_x_0 = const()[name = string("op_2541_transpose_x_0"), val = bool(false)];
            bool var_2541_transpose_y_0 = const()[name = string("op_2541_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2541_cast_fp16 = matmul(transpose_x = var_2541_transpose_x_0, transpose_y = var_2541_transpose_y_0, x = var_2540_cast_fp16, y = S_17_cast_fp16)[name = string("op_2541_cast_fp16")];
            bool var_2542_transpose_x_0 = const()[name = string("op_2542_transpose_x_0"), val = bool(false)];
            bool var_2542_transpose_y_0 = const()[name = string("op_2542_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2542_cast_fp16 = matmul(transpose_x = var_2542_transpose_x_0, transpose_y = var_2542_transpose_y_0, x = A_21_cast_fp16, y = vnew_17_cast_fp16)[name = string("op_2542_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2544_cast_fp16 = add(x = var_2541_cast_fp16, y = var_2542_cast_fp16)[name = string("op_2544_cast_fp16")];
            tensor<fp16, [16]> var_2545_bias_0_to_fp16 = const()[name = string("op_2545_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_2545_cast_fp16 = linear(bias = var_2545_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_25_cast_fp16)[name = string("op_2545_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_19_cast_fp16 = exp(x = var_2545_cast_fp16)[name = string("tail_19_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_37_to_fp16 = const()[name = string("transpose_37_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_2547_bias_0_to_fp16 = const()[name = string("op_2547_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_2547_cast_fp16 = linear(bias = var_2547_bias_0_to_fp16, weight = transpose_37_to_fp16, x = gdec_25_cast_fp16)[name = string("op_2547_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_2548_cast_fp16 = exp(x = var_2547_cast_fp16)[name = string("op_2548_cast_fp16")];
            tensor<int32, [1]> last_15_axes_0 = const()[name = string("last_15_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_15_cast_fp16 = expand_dims(axes = last_15_axes_0, x = var_2548_cast_fp16)[name = string("last_15_cast_fp16")];
            tensor<int32, [1]> var_2552_axes_0 = const()[name = string("op_2552_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_2552_cast_fp16 = expand_dims(axes = var_2552_axes_0, x = tail_19_cast_fp16)[name = string("op_2552_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2553_cast_fp16 = mul(x = k_37_cast_fp16, y = var_2552_cast_fp16)[name = string("op_2553_cast_fp16")];
            bool upd_15_transpose_x_1 = const()[name = string("upd_15_transpose_x_1"), val = bool(true)];
            bool upd_15_transpose_y_1 = const()[name = string("upd_15_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_15_cast_fp16 = matmul(transpose_x = upd_15_transpose_x_1, transpose_y = upd_15_transpose_y_1, x = var_2553_cast_fp16, y = vnew_17_cast_fp16)[name = string("upd_15_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_2558_cast_fp16 = mul(x = S_17_cast_fp16, y = last_15_cast_fp16)[name = string("op_2558_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_19_cast_fp16 = add(x = var_2558_cast_fp16, y = upd_15_cast_fp16)[name = string("S_19_cast_fp16")];
            tensor<int32, [4]> q_39_begin_0 = const()[name = string("q_39_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> q_39_end_0 = const()[name = string("q_39_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_39_end_mask_0 = const()[name = string("q_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_39_cast_fp16 = slice_by_index(begin = q_39_begin_0, end = q_39_end_0, end_mask = q_39_end_mask_0, x = q_27_cast_fp16)[name = string("q_39_cast_fp16")];
            tensor<int32, [4]> k_39_begin_0 = const()[name = string("k_39_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> k_39_end_0 = const()[name = string("k_39_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_39_end_mask_0 = const()[name = string("k_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_39_cast_fp16 = slice_by_index(begin = k_39_begin_0, end = k_39_end_0, end_mask = k_39_end_mask_0, x = k_27_cast_fp16)[name = string("k_39_cast_fp16")];
            tensor<int32, [4]> kb_23_begin_0 = const()[name = string("kb_23_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> kb_23_end_0 = const()[name = string("kb_23_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_23_end_mask_0 = const()[name = string("kb_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_23_cast_fp16 = slice_by_index(begin = kb_23_begin_0, end = kb_23_end_0, end_mask = kb_23_end_mask_0, x = var_1845_cast_fp16)[name = string("kb_23_cast_fp16")];
            tensor<int32, [4]> vb_27_begin_0 = const()[name = string("vb_27_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> vb_27_end_0 = const()[name = string("vb_27_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_27_end_mask_0 = const()[name = string("vb_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_27_cast_fp16 = slice_by_index(begin = vb_27_begin_0, end = vb_27_end_0, end_mask = vb_27_end_mask_0, x = vb_15_cast_fp16)[name = string("vb_27_cast_fp16")];
            tensor<int32, [3]> gdec_27_begin_0 = const()[name = string("gdec_27_begin_0"), val = tensor<int32, [3]>([0, 0, 80])];
            tensor<int32, [3]> gdec_27_end_0 = const()[name = string("gdec_27_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_27_end_mask_0 = const()[name = string("gdec_27_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_27_cast_fp16 = slice_by_index(begin = gdec_27_begin_0, end = gdec_27_end_0, end_mask = gdec_27_end_mask_0, x = gdec_15_cast_fp16)[name = string("gdec_27_cast_fp16")];
            tensor<int32, [2]> var_2641 = const()[name = string("op_2641"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_2642_cast_fp16 = reshape(shape = var_2641, x = gdec_27_cast_fp16)[name = string("op_2642_cast_fp16")];
            tensor<fp16, [12, 256]> var_2643_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_2642_cast_fp16)[name = string("op_2643_cast_fp16")];
            tensor<int32, [4]> var_2648 = const()[name = string("op_2648"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_23_cast_fp16 = reshape(shape = var_2648, x = var_2643_cast_fp16)[name = string("P_23_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2650_cast_fp16 = exp(x = P_23_cast_fp16)[name = string("op_2650_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_23_cast_fp16 = mul(x = var_2650_cast_fp16, y = tril_incl_to_fp16)[name = string("D_23_cast_fp16")];
            bool kk_23_transpose_x_1 = const()[name = string("kk_23_transpose_x_1"), val = bool(false)];
            bool kk_23_transpose_y_1 = const()[name = string("kk_23_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_23_cast_fp16 = matmul(transpose_x = kk_23_transpose_x_1, transpose_y = kk_23_transpose_y_1, x = kb_23_cast_fp16, y = k_39_cast_fp16)[name = string("kk_23_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_2656_cast_fp16 = mul(x = kk_23_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_2656_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2657_cast_fp16 = mul(x = var_2656_cast_fp16, y = D_23_cast_fp16)[name = string("op_2657_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_23_cast_fp16 = mul(x = var_2657_cast_fp16, y = cum_tail_to_fp16)[name = string("N_23_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_2659_cast_fp16 = mul(x = D_23_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_2659_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_23_cast_fp16 = mul(x = kk_23_cast_fp16, y = var_2659_cast_fp16)[name = string("Na_23_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_89_cast_fp16 = add(x = eye_to_fp16, y = N_23_cast_fp16)[name = string("Minv_89_cast_fp16")];
            bool Np_67_transpose_x_0 = const()[name = string("Np_67_transpose_x_0"), val = bool(false)];
            bool Np_67_transpose_y_0 = const()[name = string("Np_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_67_cast_fp16 = matmul(transpose_x = Np_67_transpose_x_0, transpose_y = Np_67_transpose_y_0, x = N_23_cast_fp16, y = Na_23_cast_fp16)[name = string("Np_67_cast_fp16")];
            bool Npa_45_transpose_x_0 = const()[name = string("Npa_45_transpose_x_0"), val = bool(false)];
            bool Npa_45_transpose_y_0 = const()[name = string("Npa_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_45_cast_fp16 = matmul(transpose_x = Npa_45_transpose_x_0, transpose_y = Npa_45_transpose_y_0, x = Na_23_cast_fp16, y = N_23_cast_fp16)[name = string("Npa_45_cast_fp16")];
            bool var_2665_transpose_x_0 = const()[name = string("op_2665_transpose_x_0"), val = bool(false)];
            bool var_2665_transpose_y_0 = const()[name = string("op_2665_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2665_cast_fp16 = matmul(transpose_x = var_2665_transpose_x_0, transpose_y = var_2665_transpose_y_0, x = Minv_89_cast_fp16, y = Np_67_cast_fp16)[name = string("op_2665_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_91_cast_fp16 = add(x = Minv_89_cast_fp16, y = var_2665_cast_fp16)[name = string("Minv_91_cast_fp16")];
            bool Np_69_transpose_x_0 = const()[name = string("Np_69_transpose_x_0"), val = bool(false)];
            bool Np_69_transpose_y_0 = const()[name = string("Np_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_69_cast_fp16 = matmul(transpose_x = Np_69_transpose_x_0, transpose_y = Np_69_transpose_y_0, x = Np_67_cast_fp16, y = Npa_45_cast_fp16)[name = string("Np_69_cast_fp16")];
            bool Npa_47_transpose_x_0 = const()[name = string("Npa_47_transpose_x_0"), val = bool(false)];
            bool Npa_47_transpose_y_0 = const()[name = string("Npa_47_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_47_cast_fp16 = matmul(transpose_x = Npa_47_transpose_x_0, transpose_y = Npa_47_transpose_y_0, x = Npa_45_cast_fp16, y = Np_67_cast_fp16)[name = string("Npa_47_cast_fp16")];
            bool var_2670_transpose_x_0 = const()[name = string("op_2670_transpose_x_0"), val = bool(false)];
            bool var_2670_transpose_y_0 = const()[name = string("op_2670_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2670_cast_fp16 = matmul(transpose_x = var_2670_transpose_x_0, transpose_y = var_2670_transpose_y_0, x = Minv_91_cast_fp16, y = Np_69_cast_fp16)[name = string("op_2670_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_93_cast_fp16 = add(x = Minv_91_cast_fp16, y = var_2670_cast_fp16)[name = string("Minv_93_cast_fp16")];
            bool Np_71_transpose_x_0 = const()[name = string("Np_71_transpose_x_0"), val = bool(false)];
            bool Np_71_transpose_y_0 = const()[name = string("Np_71_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_71_cast_fp16 = matmul(transpose_x = Np_71_transpose_x_0, transpose_y = Np_71_transpose_y_0, x = Np_69_cast_fp16, y = Npa_47_cast_fp16)[name = string("Np_71_cast_fp16")];
            bool var_2674_transpose_x_0 = const()[name = string("op_2674_transpose_x_0"), val = bool(false)];
            bool var_2674_transpose_y_0 = const()[name = string("op_2674_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_2674_cast_fp16 = matmul(transpose_x = var_2674_transpose_x_0, transpose_y = var_2674_transpose_y_0, x = Minv_93_cast_fp16, y = Np_71_cast_fp16)[name = string("op_2674_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_95_cast_fp16 = add(x = Minv_93_cast_fp16, y = var_2674_cast_fp16)[name = string("Minv_95_cast_fp16")];
            bool u_23_transpose_x_0 = const()[name = string("u_23_transpose_x_0"), val = bool(false)];
            bool u_23_transpose_y_0 = const()[name = string("u_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_23_cast_fp16 = matmul(transpose_x = u_23_transpose_x_0, transpose_y = u_23_transpose_y_0, x = Minv_95_cast_fp16, y = vb_27_cast_fp16)[name = string("u_23_cast_fp16")];
            bool var_2681_transpose_x_1 = const()[name = string("op_2681_transpose_x_1"), val = bool(false)];
            bool var_2681_transpose_y_1 = const()[name = string("op_2681_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_2681_cast_fp16 = matmul(transpose_x = var_2681_transpose_x_1, transpose_y = var_2681_transpose_y_1, x = q_39_cast_fp16, y = k_39_cast_fp16)[name = string("op_2681_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_23_cast_fp16 = mul(x = var_2681_cast_fp16, y = D_23_cast_fp16)[name = string("A_23_cast_fp16")];
            tensor<fp16, [16]> var_2683_bias_0_to_fp16 = const()[name = string("op_2683_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_2683_cast_fp16 = linear(bias = var_2683_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_27_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_19_cast_fp16 = exp(x = var_2683_cast_fp16)[name = string("cum_19_cast_fp16")];
            tensor<int32, [1]> var_2686_axes_0 = const()[name = string("op_2686_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_2686_cast_fp16 = expand_dims(axes = var_2686_axes_0, x = cum_19_cast_fp16)[name = string("op_2686_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2687_cast_fp16 = mul(x = kb_23_cast_fp16, y = var_2686_cast_fp16)[name = string("op_2687_cast_fp16")];
            bool w_23_transpose_x_0 = const()[name = string("w_23_transpose_x_0"), val = bool(false)];
            bool w_23_transpose_y_0 = const()[name = string("w_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_23_cast_fp16 = matmul(transpose_x = w_23_transpose_x_0, transpose_y = w_23_transpose_y_0, x = Minv_95_cast_fp16, y = var_2687_cast_fp16)[name = string("w_23_cast_fp16")];
            bool var_2689_transpose_x_0 = const()[name = string("op_2689_transpose_x_0"), val = bool(false)];
            bool var_2689_transpose_y_0 = const()[name = string("op_2689_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2689_cast_fp16 = matmul(transpose_x = var_2689_transpose_x_0, transpose_y = var_2689_transpose_y_0, x = w_23_cast_fp16, y = S_19_cast_fp16)[name = string("op_2689_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_19_cast_fp16 = sub(x = u_23_cast_fp16, y = var_2689_cast_fp16)[name = string("vnew_19_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_2694_cast_fp16 = mul(x = q_39_cast_fp16, y = var_2686_cast_fp16)[name = string("op_2694_cast_fp16")];
            bool var_2695_transpose_x_0 = const()[name = string("op_2695_transpose_x_0"), val = bool(false)];
            bool var_2695_transpose_y_0 = const()[name = string("op_2695_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2695_cast_fp16 = matmul(transpose_x = var_2695_transpose_x_0, transpose_y = var_2695_transpose_y_0, x = var_2694_cast_fp16, y = S_19_cast_fp16)[name = string("op_2695_cast_fp16")];
            bool var_2696_transpose_x_0 = const()[name = string("op_2696_transpose_x_0"), val = bool(false)];
            bool var_2696_transpose_y_0 = const()[name = string("op_2696_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_2696_cast_fp16 = matmul(transpose_x = var_2696_transpose_x_0, transpose_y = var_2696_transpose_y_0, x = A_23_cast_fp16, y = vnew_19_cast_fp16)[name = string("op_2696_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_3_cast_fp16 = add(x = var_2695_cast_fp16, y = var_2696_cast_fp16)[name = string("oc_3_cast_fp16")];
            int32 var_2700 = const()[name = string("op_2700"), val = int32(2)];
            bool o_7_interleave_0 = const()[name = string("o_7_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 96, 128]> o_7_cast_fp16 = concat(axis = var_2700, interleave = o_7_interleave_0, values = (var_1935_cast_fp16, var_2082_cast_fp16, var_2236_cast_fp16, var_2390_cast_fp16, var_2544_cast_fp16, oc_3_cast_fp16))[name = string("o_7_cast_fp16")];
            tensor<int32, [4]> var_2706 = const()[name = string("op_2706"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2712 = const()[name = string("op_2712"), val = tensor<int32, [4]>([3, 96, 4, 128])];
            tensor<fp16, [3, 96, 4, 128]> input_23_cast_fp16 = reshape(shape = var_2712, x = z_3_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> o_9_cast_fp16 = transpose(perm = var_2706, x = o_7_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [3, 96, 4, 128]> var_2714_cast_fp16 = mul(x = o_9_cast_fp16, y = o_9_cast_fp16)[name = string("op_2714_cast_fp16")];
            tensor<int32, [1]> var_2719_axes_0 = const()[name = string("op_2719_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2719_keep_dims_0 = const()[name = string("op_2719_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 4, 1]> var_2719_cast_fp16 = reduce_mean(axes = var_2719_axes_0, keep_dims = var_2719_keep_dims_0, x = var_2714_cast_fp16)[name = string("op_2719_cast_fp16")];
            fp16 var_2721_to_fp16 = const()[name = string("op_2721_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 96, 4, 1]> var_2722_cast_fp16 = add(x = var_2719_cast_fp16, y = var_2721_to_fp16)[name = string("op_2722_cast_fp16")];
            fp32 var_2723_epsilon_0 = const()[name = string("op_2723_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 4, 1]> var_2723_cast_fp16 = rsqrt(epsilon = var_2723_epsilon_0, x = var_2722_cast_fp16)[name = string("op_2723_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> var_2724_cast_fp16 = mul(x = o_9_cast_fp16, y = var_2723_cast_fp16)[name = string("op_2724_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> var_2726_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> o_11_cast_fp16 = mul(x = var_2724_cast_fp16, y = var_2726_cast_fp16)[name = string("o_11_cast_fp16")];
            tensor<int32, [3]> var_2731 = const()[name = string("op_2731"), val = tensor<int32, [3]>([3, 96, 512])];
            tensor<fp16, [3, 96, 512]> input_25_cast_fp16 = reshape(shape = var_2731, x = o_11_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [640, 512]> layers_1_out_weight_to_fp16 = const()[name = string("layers_1_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34984256)))];
            tensor<fp16, [3, 96, 640]> linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_out_weight_to_fp16, x = input_25_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [3, 96, 640]> x_13_cast_fp16 = add(x = x_9_cast_fp16, y = linear_5_cast_fp16)[name = string("x_13_cast_fp16")];
            fp16 var_2744_to_fp16 = const()[name = string("op_2744_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> x_15_cast_fp16 = mul(x = x_13_cast_fp16, y = var_2744_to_fp16)[name = string("x_15_cast_fp16")];
            fp16 var_2742_promoted_to_fp16 = const()[name = string("op_2742_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> var_2746_cast_fp16 = pow(x = x_15_cast_fp16, y = var_2742_promoted_to_fp16)[name = string("op_2746_cast_fp16")];
            tensor<int32, [1]> var_2748_axes_0 = const()[name = string("op_2748_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2748_keep_dims_0 = const()[name = string("op_2748_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 1]> var_2748_cast_fp16 = reduce_mean(axes = var_2748_axes_0, keep_dims = var_2748_keep_dims_0, x = var_2746_cast_fp16)[name = string("op_2748_cast_fp16")];
            fp16 var_2749_to_fp16 = const()[name = string("op_2749_to_fp16"), val = fp16(0x1.0cp-18)];
            tensor<fp16, [3, 96, 1]> var_2750_cast_fp16 = add(x = var_2748_cast_fp16, y = var_2749_to_fp16)[name = string("op_2750_cast_fp16")];
            fp32 var_2751_epsilon_0 = const()[name = string("op_2751_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 1]> var_2751_cast_fp16 = rsqrt(epsilon = var_2751_epsilon_0, x = var_2750_cast_fp16)[name = string("op_2751_cast_fp16")];
            tensor<fp16, [3, 96, 640]> var_2752_cast_fp16 = mul(x = x_15_cast_fp16, y = var_2751_cast_fp16)[name = string("op_2752_cast_fp16")];
            tensor<fp16, [640]> layers_1_n2_w1_to_fp16 = const()[name = string("layers_1_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35639680)))];
            tensor<fp16, [3, 96, 640]> input_27_cast_fp16 = mul(x = var_2752_cast_fp16, y = layers_1_n2_w1_to_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_1_gate_up_weight_to_fp16 = const()[name = string("layers_1_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35641024)))];
            tensor<fp16, [3, 96, 5120]> linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_1_gate_up_weight_to_fp16, x = input_27_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<int32, [3]> input_29_begin_0 = const()[name = string("input_29_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_29_end_0 = const()[name = string("input_29_end_0"), val = tensor<int32, [3]>([3, 96, 2560])];
            tensor<bool, [3]> input_29_end_mask_0 = const()[name = string("input_29_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 2560]> input_29_cast_fp16 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = linear_6_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [3, 96, 2560]> var_2762_cast_fp16 = silu(x = input_29_cast_fp16)[name = string("op_2762_cast_fp16")];
            tensor<int32, [3]> var_2767_begin_0 = const()[name = string("op_2767_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_2767_end_0 = const()[name = string("op_2767_end_0"), val = tensor<int32, [3]>([3, 96, 5120])];
            tensor<bool, [3]> var_2767_end_mask_0 = const()[name = string("op_2767_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 2560]> var_2767_cast_fp16 = slice_by_index(begin = var_2767_begin_0, end = var_2767_end_0, end_mask = var_2767_end_mask_0, x = linear_6_cast_fp16)[name = string("op_2767_cast_fp16")];
            tensor<fp16, [3, 96, 2560]> input_31_cast_fp16 = mul(x = var_2762_cast_fp16, y = var_2767_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_1_down_weight_to_fp16 = const()[name = string("layers_1_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42194688)))];
            tensor<fp16, [3, 96, 640]> linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_down_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [3, 96, 640]> x_17_cast_fp16 = add(x = x_13_cast_fp16, y = linear_7_cast_fp16)[name = string("x_17_cast_fp16")];
            fp16 var_2778_promoted_to_fp16 = const()[name = string("op_2778_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> var_2782_cast_fp16 = pow(x = x_17_cast_fp16, y = var_2778_promoted_to_fp16)[name = string("op_2782_cast_fp16")];
            tensor<int32, [1]> var_2784_axes_0 = const()[name = string("op_2784_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2784_keep_dims_0 = const()[name = string("op_2784_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 1]> var_2784_cast_fp16 = reduce_mean(axes = var_2784_axes_0, keep_dims = var_2784_keep_dims_0, x = var_2782_cast_fp16)[name = string("op_2784_cast_fp16")];
            fp16 var_2785_to_fp16 = const()[name = string("op_2785_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 96, 1]> var_2786_cast_fp16 = add(x = var_2784_cast_fp16, y = var_2785_to_fp16)[name = string("op_2786_cast_fp16")];
            fp32 var_2787_epsilon_0 = const()[name = string("op_2787_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 1]> var_2787_cast_fp16 = rsqrt(epsilon = var_2787_epsilon_0, x = var_2786_cast_fp16)[name = string("op_2787_cast_fp16")];
            tensor<fp16, [3, 96, 640]> var_2788_cast_fp16 = mul(x = x_17_cast_fp16, y = var_2787_cast_fp16)[name = string("op_2788_cast_fp16")];
            tensor<fp16, [640]> layers_2_n1_w1_to_fp16 = const()[name = string("layers_2_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45471552)))];
            tensor<fp16, [3, 96, 640]> input_33_cast_fp16 = mul(x = var_2788_cast_fp16, y = layers_2_n1_w1_to_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_2_qkvzba_weight_to_fp16 = const()[name = string("layers_2_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45472896)))];
            tensor<fp16, [3, 96, 1544]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_qkvzba_weight_to_fp16, x = input_33_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [3]> qkv_5_begin_0 = const()[name = string("qkv_5_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_5_end_0 = const()[name = string("qkv_5_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> qkv_5_end_mask_0 = const()[name = string("qkv_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 1024]> qkv_5_cast_fp16 = slice_by_index(begin = qkv_5_begin_0, end = qkv_5_end_0, end_mask = qkv_5_end_mask_0, x = linear_8_cast_fp16)[name = string("qkv_5_cast_fp16")];
            tensor<int32, [3]> z_5_begin_0 = const()[name = string("z_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_5_end_0 = const()[name = string("z_5_end_0"), val = tensor<int32, [3]>([3, 96, 1536])];
            tensor<bool, [3]> z_5_end_mask_0 = const()[name = string("z_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 512]> z_5_cast_fp16 = slice_by_index(begin = z_5_begin_0, end = z_5_end_0, end_mask = z_5_end_mask_0, x = linear_8_cast_fp16)[name = string("z_5_cast_fp16")];
            tensor<int32, [3]> b_5_begin_0 = const()[name = string("b_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_5_end_0 = const()[name = string("b_5_end_0"), val = tensor<int32, [3]>([3, 96, 1540])];
            tensor<bool, [3]> b_5_end_mask_0 = const()[name = string("b_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 4]> b_5_cast_fp16 = slice_by_index(begin = b_5_begin_0, end = b_5_end_0, end_mask = b_5_end_mask_0, x = linear_8_cast_fp16)[name = string("b_5_cast_fp16")];
            tensor<int32, [3]> a_5_begin_0 = const()[name = string("a_5_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_5_end_0 = const()[name = string("a_5_end_0"), val = tensor<int32, [3]>([3, 96, 1])];
            tensor<bool, [3]> a_5_end_mask_0 = const()[name = string("a_5_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 4]> a_5_cast_fp16 = slice_by_index(begin = a_5_begin_0, end = a_5_end_0, end_mask = a_5_end_mask_0, x = linear_8_cast_fp16)[name = string("a_5_cast_fp16")];
            int32 var_2827 = const()[name = string("op_2827"), val = int32(1)];
            bool xp_5_interleave_0 = const()[name = string("xp_5_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 99, 1024]> xp_5_cast_fp16 = concat(axis = var_2827, interleave = xp_5_interleave_0, values = (zpad_to_fp16, qkv_5_cast_fp16))[name = string("xp_5_cast_fp16")];
            tensor<int32, [3]> var_2838_begin_0 = const()[name = string("op_2838_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_2838_end_0 = const()[name = string("op_2838_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_2838_end_mask_0 = const()[name = string("op_2838_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 1024]> var_2838_cast_fp16 = slice_by_index(begin = var_2838_begin_0, end = var_2838_end_0, end_mask = var_2838_end_mask_0, x = xp_5_cast_fp16)[name = string("op_2838_cast_fp16")];
            tensor<fp16, [1024]> var_2846_to_fp16 = const()[name = string("op_2846_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47449280)))];
            tensor<fp16, [3, 96, 1024]> var_2847_cast_fp16 = mul(x = var_2838_cast_fp16, y = var_2846_to_fp16)[name = string("op_2847_cast_fp16")];
            tensor<int32, [3]> var_2857_begin_0 = const()[name = string("op_2857_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_2857_end_0 = const()[name = string("op_2857_end_0"), val = tensor<int32, [3]>([3, 98, 1024])];
            tensor<bool, [3]> var_2857_end_mask_0 = const()[name = string("op_2857_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_2857_cast_fp16 = slice_by_index(begin = var_2857_begin_0, end = var_2857_end_0, end_mask = var_2857_end_mask_0, x = xp_5_cast_fp16)[name = string("op_2857_cast_fp16")];
            tensor<fp16, [1024]> var_2865_to_fp16 = const()[name = string("op_2865_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47451392)))];
            tensor<fp16, [3, 96, 1024]> var_2866_cast_fp16 = mul(x = var_2857_cast_fp16, y = var_2865_to_fp16)[name = string("op_2866_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> var_2868_cast_fp16 = add(x = var_2847_cast_fp16, y = var_2866_cast_fp16)[name = string("op_2868_cast_fp16")];
            tensor<int32, [3]> var_2878_begin_0 = const()[name = string("op_2878_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2878_end_0 = const()[name = string("op_2878_end_0"), val = tensor<int32, [3]>([3, 97, 1024])];
            tensor<bool, [3]> var_2878_end_mask_0 = const()[name = string("op_2878_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_2878_cast_fp16 = slice_by_index(begin = var_2878_begin_0, end = var_2878_end_0, end_mask = var_2878_end_mask_0, x = xp_5_cast_fp16)[name = string("op_2878_cast_fp16")];
            tensor<fp16, [1024]> var_2886_to_fp16 = const()[name = string("op_2886_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47453504)))];
            tensor<fp16, [3, 96, 1024]> var_2887_cast_fp16 = mul(x = var_2878_cast_fp16, y = var_2886_to_fp16)[name = string("op_2887_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> var_2889_cast_fp16 = add(x = var_2868_cast_fp16, y = var_2887_cast_fp16)[name = string("op_2889_cast_fp16")];
            tensor<int32, [3]> var_2899_begin_0 = const()[name = string("op_2899_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2899_end_0 = const()[name = string("op_2899_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> var_2899_end_mask_0 = const()[name = string("op_2899_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_2899_cast_fp16 = slice_by_index(begin = var_2899_begin_0, end = var_2899_end_0, end_mask = var_2899_end_mask_0, x = xp_5_cast_fp16)[name = string("op_2899_cast_fp16")];
            tensor<fp16, [1024]> var_2907_to_fp16 = const()[name = string("op_2907_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47455616)))];
            tensor<fp16, [3, 96, 1024]> var_2908_cast_fp16 = mul(x = var_2899_cast_fp16, y = var_2907_to_fp16)[name = string("op_2908_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> input_35_cast_fp16 = add(x = var_2889_cast_fp16, y = var_2908_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> c_5_cast_fp16 = silu(x = input_35_cast_fp16)[name = string("c_5_cast_fp16")];
            tensor<int32, [3]> var_2916_begin_0 = const()[name = string("op_2916_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2916_end_0 = const()[name = string("op_2916_end_0"), val = tensor<int32, [3]>([3, 96, 256])];
            tensor<bool, [3]> var_2916_end_mask_0 = const()[name = string("op_2916_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 256]> var_2916_cast_fp16 = slice_by_index(begin = var_2916_begin_0, end = var_2916_end_0, end_mask = var_2916_end_mask_0, x = c_5_cast_fp16)[name = string("op_2916_cast_fp16")];
            tensor<int32, [4]> var_2921 = const()[name = string("op_2921"), val = tensor<int32, [4]>([3, 96, 2, 128])];
            tensor<fp16, [3, 96, 2, 128]> q_41_cast_fp16 = reshape(shape = var_2921, x = var_2916_cast_fp16)[name = string("q_41_cast_fp16")];
            tensor<int32, [3]> var_2927_begin_0 = const()[name = string("op_2927_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_2927_end_0 = const()[name = string("op_2927_end_0"), val = tensor<int32, [3]>([3, 96, 512])];
            tensor<bool, [3]> var_2927_end_mask_0 = const()[name = string("op_2927_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 256]> var_2927_cast_fp16 = slice_by_index(begin = var_2927_begin_0, end = var_2927_end_0, end_mask = var_2927_end_mask_0, x = c_5_cast_fp16)[name = string("op_2927_cast_fp16")];
            tensor<int32, [4]> var_2932 = const()[name = string("op_2932"), val = tensor<int32, [4]>([3, 96, 2, 128])];
            tensor<fp16, [3, 96, 2, 128]> k_41_cast_fp16 = reshape(shape = var_2932, x = var_2927_cast_fp16)[name = string("k_41_cast_fp16")];
            tensor<int32, [3]> var_2938_begin_0 = const()[name = string("op_2938_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_2938_end_0 = const()[name = string("op_2938_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> var_2938_end_mask_0 = const()[name = string("op_2938_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 512]> var_2938_cast_fp16 = slice_by_index(begin = var_2938_begin_0, end = var_2938_end_0, end_mask = var_2938_end_mask_0, x = c_5_cast_fp16)[name = string("op_2938_cast_fp16")];
            tensor<int32, [4]> var_2943 = const()[name = string("op_2943"), val = tensor<int32, [4]>([3, 96, 4, 128])];
            tensor<fp16, [3, 96, 4, 128]> var_2944_cast_fp16 = reshape(shape = var_2943, x = var_2938_cast_fp16)[name = string("op_2944_cast_fp16")];
            tensor<int32, [4]> var_2949 = const()[name = string("op_2949"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2951_to_fp16 = const()[name = string("op_2951_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 96, 2, 128]> q_43_cast_fp16 = mul(x = q_41_cast_fp16, y = var_2951_to_fp16)[name = string("q_43_cast_fp16")];
            fp16 var_2953_to_fp16 = const()[name = string("op_2953_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 96, 2, 128]> k_43_cast_fp16 = mul(x = k_41_cast_fp16, y = var_2953_to_fp16)[name = string("k_43_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_2955_cast_fp16 = mul(x = q_43_cast_fp16, y = q_43_cast_fp16)[name = string("op_2955_cast_fp16")];
            tensor<int32, [1]> var_2960_axes_0 = const()[name = string("op_2960_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2960_keep_dims_0 = const()[name = string("op_2960_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 2, 1]> var_2960_cast_fp16 = reduce_mean(axes = var_2960_axes_0, keep_dims = var_2960_keep_dims_0, x = var_2955_cast_fp16)[name = string("op_2960_cast_fp16")];
            fp16 var_2961_promoted_to_fp16 = const()[name = string("op_2961_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 96, 2, 1]> var_2962_cast_fp16 = mul(x = var_2960_cast_fp16, y = var_2961_promoted_to_fp16)[name = string("op_2962_cast_fp16")];
            fp16 var_2964_to_fp16 = const()[name = string("op_2964_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 96, 2, 1]> var_2965_cast_fp16 = add(x = var_2962_cast_fp16, y = var_2964_to_fp16)[name = string("op_2965_cast_fp16")];
            fp32 var_2966_epsilon_0 = const()[name = string("op_2966_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 2, 1]> var_2966_cast_fp16 = rsqrt(epsilon = var_2966_epsilon_0, x = var_2965_cast_fp16)[name = string("op_2966_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_2967_cast_fp16 = mul(x = q_43_cast_fp16, y = var_2966_cast_fp16)[name = string("op_2967_cast_fp16")];
            fp16 var_2968_to_fp16 = const()[name = string("op_2968_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 96, 2, 128]> q_45_cast_fp16 = mul(x = var_2967_cast_fp16, y = var_2968_to_fp16)[name = string("q_45_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_2970_cast_fp16 = mul(x = k_43_cast_fp16, y = k_43_cast_fp16)[name = string("op_2970_cast_fp16")];
            tensor<int32, [1]> var_2975_axes_0 = const()[name = string("op_2975_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2975_keep_dims_0 = const()[name = string("op_2975_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 2, 1]> var_2975_cast_fp16 = reduce_mean(axes = var_2975_axes_0, keep_dims = var_2975_keep_dims_0, x = var_2970_cast_fp16)[name = string("op_2975_cast_fp16")];
            fp16 var_2976_promoted_to_fp16 = const()[name = string("op_2976_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 96, 2, 1]> var_2977_cast_fp16 = mul(x = var_2975_cast_fp16, y = var_2976_promoted_to_fp16)[name = string("op_2977_cast_fp16")];
            fp16 var_2979_to_fp16 = const()[name = string("op_2979_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 96, 2, 1]> var_2980_cast_fp16 = add(x = var_2977_cast_fp16, y = var_2979_to_fp16)[name = string("op_2980_cast_fp16")];
            fp32 var_2981_epsilon_0 = const()[name = string("op_2981_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 2, 1]> var_2981_cast_fp16 = rsqrt(epsilon = var_2981_epsilon_0, x = var_2980_cast_fp16)[name = string("op_2981_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> k_45_cast_fp16 = mul(x = k_43_cast_fp16, y = var_2981_cast_fp16)[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> var_2987 = const()[name = string("op_2987"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_3000_axes_0 = const()[name = string("op_3000_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 96, 128]> var_2988_cast_fp16 = transpose(perm = var_2987, x = q_45_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [3, 2, 1, 96, 128]> var_3000_cast_fp16 = expand_dims(axes = var_3000_axes_0, x = var_2988_cast_fp16)[name = string("op_3000_cast_fp16")];
            tensor<int32, [5]> var_3008_reps_0 = const()[name = string("op_3008_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 96, 128]> var_3008_cast_fp16 = tile(reps = var_3008_reps_0, x = var_3000_cast_fp16)[name = string("op_3008_cast_fp16")];
            tensor<int32, [4]> var_3013 = const()[name = string("op_3013"), val = tensor<int32, [4]>([3, 4, 96, 128])];
            tensor<fp16, [3, 4, 96, 128]> q_47_cast_fp16 = reshape(shape = var_3013, x = var_3008_cast_fp16)[name = string("q_47_cast_fp16")];
            tensor<int32, [4]> var_3019 = const()[name = string("op_3019"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_3032_axes_0 = const()[name = string("op_3032_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 96, 128]> var_3020_cast_fp16 = transpose(perm = var_3019, x = k_45_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [3, 2, 1, 96, 128]> var_3032_cast_fp16 = expand_dims(axes = var_3032_axes_0, x = var_3020_cast_fp16)[name = string("op_3032_cast_fp16")];
            tensor<int32, [5]> var_3040_reps_0 = const()[name = string("op_3040_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 96, 128]> var_3040_cast_fp16 = tile(reps = var_3040_reps_0, x = var_3032_cast_fp16)[name = string("op_3040_cast_fp16")];
            tensor<int32, [4]> var_3045 = const()[name = string("op_3045"), val = tensor<int32, [4]>([3, 4, 96, 128])];
            tensor<fp16, [3, 4, 96, 128]> k_47_cast_fp16 = reshape(shape = var_3045, x = var_3040_cast_fp16)[name = string("k_47_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_3047_cast_fp16 = sigmoid(x = b_5_cast_fp16)[name = string("op_3047_cast_fp16")];
            tensor<int32, [3]> var_3051 = const()[name = string("op_3051"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_2_dt_bias_to_fp16 = const()[name = string("layers_2_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 96, 4]> input_37_cast_fp16 = add(x = a_5_cast_fp16, y = layers_2_dt_bias_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_3055_cast_fp16 = relu(x = input_37_cast_fp16)[name = string("op_3055_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_3056_cast_fp16 = abs(x = input_37_cast_fp16)[name = string("op_3056_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 96, 4]> var_3057_cast_fp16 = mul(x = var_3056_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_3057_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_3058_cast_fp16 = exp(x = var_3057_cast_fp16)[name = string("op_3058_cast_fp16")];
            fp16 var_3060_to_fp16 = const()[name = string("op_3060_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 96, 4]> var_3061_cast_fp16 = add(x = var_3058_cast_fp16, y = var_3060_to_fp16)[name = string("op_3061_cast_fp16")];
            fp32 var_3062_epsilon_0 = const()[name = string("op_3062_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 96, 4]> var_3062_cast_fp16 = log(epsilon = var_3062_epsilon_0, x = var_3061_cast_fp16)[name = string("op_3062_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_3064_cast_fp16 = add(x = var_3055_cast_fp16, y = var_3062_cast_fp16)[name = string("op_3064_cast_fp16")];
            tensor<fp16, [4]> layers_2_negA_to_fp16 = const()[name = string("layers_2_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.c0cp+3, -0x1.3e4p+3, -0x1.568p+1, -0x1.9acp+2])];
            tensor<fp16, [3, 96, 4]> var_3065_cast_fp16 = mul(x = layers_2_negA_to_fp16, y = var_3064_cast_fp16)[name = string("op_3065_cast_fp16")];
            tensor<int32, [3]> var_3069 = const()[name = string("op_3069"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_3072_axes_0 = const()[name = string("op_3072_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 96]> beta_5_cast_fp16 = transpose(perm = var_3051, x = var_3047_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [3, 4, 96, 1]> var_3072_cast_fp16 = expand_dims(axes = var_3072_axes_0, x = beta_5_cast_fp16)[name = string("op_3072_cast_fp16")];
            fp16 var_3073_to_fp16 = const()[name = string("op_3073_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 96, 1]> var_3074_cast_fp16 = mul(x = var_3072_cast_fp16, y = var_3073_to_fp16)[name = string("op_3074_cast_fp16")];
            tensor<fp16, [3, 4, 96, 128]> v_5_cast_fp16 = transpose(perm = var_2949, x = var_2944_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [3, 4, 96, 128]> vb_29_cast_fp16 = mul(x = v_5_cast_fp16, y = var_3074_cast_fp16)[name = string("vb_29_cast_fp16")];
            tensor<int32, [4]> q_49_begin_0 = const()[name = string("q_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_49_end_0 = const()[name = string("q_49_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_49_end_mask_0 = const()[name = string("q_49_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, x = q_47_cast_fp16)[name = string("q_49_cast_fp16")];
            tensor<int32, [4]> k_49_begin_0 = const()[name = string("k_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_49_end_0 = const()[name = string("k_49_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_49_end_mask_0 = const()[name = string("k_49_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, x = k_47_cast_fp16)[name = string("k_49_cast_fp16")];
            tensor<fp16, [3, 4, 96, 128]> var_3108_cast_fp16 = mul(x = k_47_cast_fp16, y = var_3072_cast_fp16)[name = string("op_3108_cast_fp16")];
            tensor<int32, [4]> kb_25_begin_0 = const()[name = string("kb_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_25_end_0 = const()[name = string("kb_25_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_25_end_mask_0 = const()[name = string("kb_25_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_25_cast_fp16 = slice_by_index(begin = kb_25_begin_0, end = kb_25_end_0, end_mask = kb_25_end_mask_0, x = var_3108_cast_fp16)[name = string("kb_25_cast_fp16")];
            tensor<int32, [4]> vb_31_begin_0 = const()[name = string("vb_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_31_end_0 = const()[name = string("vb_31_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_31_end_mask_0 = const()[name = string("vb_31_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_31_cast_fp16 = slice_by_index(begin = vb_31_begin_0, end = vb_31_end_0, end_mask = vb_31_end_mask_0, x = vb_29_cast_fp16)[name = string("vb_31_cast_fp16")];
            tensor<int32, [3]> gdec_31_begin_0 = const()[name = string("gdec_31_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_31_end_0 = const()[name = string("gdec_31_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_31_end_mask_0 = const()[name = string("gdec_31_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 96]> gdec_29_cast_fp16 = transpose(perm = var_3069, x = var_3065_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [3, 4, 16]> gdec_31_cast_fp16 = slice_by_index(begin = gdec_31_begin_0, end = gdec_31_end_0, end_mask = gdec_31_end_mask_0, x = gdec_29_cast_fp16)[name = string("gdec_31_cast_fp16")];
            tensor<int32, [2]> var_3156 = const()[name = string("op_3156"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_3157_cast_fp16 = reshape(shape = var_3156, x = gdec_31_cast_fp16)[name = string("op_3157_cast_fp16")];
            tensor<fp16, [12, 256]> var_3158_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3157_cast_fp16)[name = string("op_3158_cast_fp16")];
            tensor<int32, [4]> var_3163 = const()[name = string("op_3163"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_25_cast_fp16 = reshape(shape = var_3163, x = var_3158_cast_fp16)[name = string("P_25_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3165_cast_fp16 = exp(x = P_25_cast_fp16)[name = string("op_3165_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_25_cast_fp16 = mul(x = var_3165_cast_fp16, y = tril_incl_to_fp16)[name = string("D_25_cast_fp16")];
            bool kk_25_transpose_x_1 = const()[name = string("kk_25_transpose_x_1"), val = bool(false)];
            bool kk_25_transpose_y_1 = const()[name = string("kk_25_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_25_cast_fp16 = matmul(transpose_x = kk_25_transpose_x_1, transpose_y = kk_25_transpose_y_1, x = kb_25_cast_fp16, y = k_49_cast_fp16)[name = string("kk_25_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_3171_cast_fp16 = mul(x = kk_25_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_3171_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3172_cast_fp16 = mul(x = var_3171_cast_fp16, y = D_25_cast_fp16)[name = string("op_3172_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_25_cast_fp16 = mul(x = var_3172_cast_fp16, y = cum_tail_to_fp16)[name = string("N_25_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3174_cast_fp16 = mul(x = D_25_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3174_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_25_cast_fp16 = mul(x = kk_25_cast_fp16, y = var_3174_cast_fp16)[name = string("Na_25_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_97_cast_fp16 = add(x = eye_to_fp16, y = N_25_cast_fp16)[name = string("Minv_97_cast_fp16")];
            bool Np_73_transpose_x_0 = const()[name = string("Np_73_transpose_x_0"), val = bool(false)];
            bool Np_73_transpose_y_0 = const()[name = string("Np_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_73_cast_fp16 = matmul(transpose_x = Np_73_transpose_x_0, transpose_y = Np_73_transpose_y_0, x = N_25_cast_fp16, y = Na_25_cast_fp16)[name = string("Np_73_cast_fp16")];
            bool Npa_49_transpose_x_0 = const()[name = string("Npa_49_transpose_x_0"), val = bool(false)];
            bool Npa_49_transpose_y_0 = const()[name = string("Npa_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_49_cast_fp16 = matmul(transpose_x = Npa_49_transpose_x_0, transpose_y = Npa_49_transpose_y_0, x = Na_25_cast_fp16, y = N_25_cast_fp16)[name = string("Npa_49_cast_fp16")];
            bool var_3180_transpose_x_0 = const()[name = string("op_3180_transpose_x_0"), val = bool(false)];
            bool var_3180_transpose_y_0 = const()[name = string("op_3180_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3180_cast_fp16 = matmul(transpose_x = var_3180_transpose_x_0, transpose_y = var_3180_transpose_y_0, x = Minv_97_cast_fp16, y = Np_73_cast_fp16)[name = string("op_3180_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_99_cast_fp16 = add(x = Minv_97_cast_fp16, y = var_3180_cast_fp16)[name = string("Minv_99_cast_fp16")];
            bool Np_75_transpose_x_0 = const()[name = string("Np_75_transpose_x_0"), val = bool(false)];
            bool Np_75_transpose_y_0 = const()[name = string("Np_75_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_75_cast_fp16 = matmul(transpose_x = Np_75_transpose_x_0, transpose_y = Np_75_transpose_y_0, x = Np_73_cast_fp16, y = Npa_49_cast_fp16)[name = string("Np_75_cast_fp16")];
            bool Npa_51_transpose_x_0 = const()[name = string("Npa_51_transpose_x_0"), val = bool(false)];
            bool Npa_51_transpose_y_0 = const()[name = string("Npa_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_51_cast_fp16 = matmul(transpose_x = Npa_51_transpose_x_0, transpose_y = Npa_51_transpose_y_0, x = Npa_49_cast_fp16, y = Np_73_cast_fp16)[name = string("Npa_51_cast_fp16")];
            bool var_3185_transpose_x_0 = const()[name = string("op_3185_transpose_x_0"), val = bool(false)];
            bool var_3185_transpose_y_0 = const()[name = string("op_3185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3185_cast_fp16 = matmul(transpose_x = var_3185_transpose_x_0, transpose_y = var_3185_transpose_y_0, x = Minv_99_cast_fp16, y = Np_75_cast_fp16)[name = string("op_3185_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_101_cast_fp16 = add(x = Minv_99_cast_fp16, y = var_3185_cast_fp16)[name = string("Minv_101_cast_fp16")];
            bool Np_77_transpose_x_0 = const()[name = string("Np_77_transpose_x_0"), val = bool(false)];
            bool Np_77_transpose_y_0 = const()[name = string("Np_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_77_cast_fp16 = matmul(transpose_x = Np_77_transpose_x_0, transpose_y = Np_77_transpose_y_0, x = Np_75_cast_fp16, y = Npa_51_cast_fp16)[name = string("Np_77_cast_fp16")];
            bool var_3189_transpose_x_0 = const()[name = string("op_3189_transpose_x_0"), val = bool(false)];
            bool var_3189_transpose_y_0 = const()[name = string("op_3189_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3189_cast_fp16 = matmul(transpose_x = var_3189_transpose_x_0, transpose_y = var_3189_transpose_y_0, x = Minv_101_cast_fp16, y = Np_77_cast_fp16)[name = string("op_3189_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_103_cast_fp16 = add(x = Minv_101_cast_fp16, y = var_3189_cast_fp16)[name = string("Minv_103_cast_fp16")];
            bool u_25_transpose_x_0 = const()[name = string("u_25_transpose_x_0"), val = bool(false)];
            bool u_25_transpose_y_0 = const()[name = string("u_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_25_cast_fp16 = matmul(transpose_x = u_25_transpose_x_0, transpose_y = u_25_transpose_y_0, x = Minv_103_cast_fp16, y = vb_31_cast_fp16)[name = string("u_25_cast_fp16")];
            bool var_3196_transpose_x_1 = const()[name = string("op_3196_transpose_x_1"), val = bool(false)];
            bool var_3196_transpose_y_1 = const()[name = string("op_3196_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_3196_cast_fp16 = matmul(transpose_x = var_3196_transpose_x_1, transpose_y = var_3196_transpose_y_1, x = q_49_cast_fp16, y = k_49_cast_fp16)[name = string("op_3196_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_25_cast_fp16 = mul(x = var_3196_cast_fp16, y = D_25_cast_fp16)[name = string("A_25_cast_fp16")];
            bool var_3198_transpose_x_0 = const()[name = string("op_3198_transpose_x_0"), val = bool(false)];
            bool var_3198_transpose_y_0 = const()[name = string("op_3198_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3198_cast_fp16 = matmul(transpose_x = var_3198_transpose_x_0, transpose_y = var_3198_transpose_y_0, x = A_25_cast_fp16, y = u_25_cast_fp16)[name = string("op_3198_cast_fp16")];
            tensor<fp16, [16]> var_3199_bias_0_to_fp16 = const()[name = string("op_3199_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_3199_cast_fp16 = linear(bias = var_3199_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_31_cast_fp16)[name = string("op_3199_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_21_cast_fp16 = exp(x = var_3199_cast_fp16)[name = string("tail_21_cast_fp16")];
            tensor<int32, [1]> var_3202_axes_0 = const()[name = string("op_3202_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_3202_cast_fp16 = expand_dims(axes = var_3202_axes_0, x = tail_21_cast_fp16)[name = string("op_3202_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3203_cast_fp16 = mul(x = k_49_cast_fp16, y = var_3202_cast_fp16)[name = string("op_3203_cast_fp16")];
            bool S_21_transpose_x_1 = const()[name = string("S_21_transpose_x_1"), val = bool(true)];
            bool S_21_transpose_y_1 = const()[name = string("S_21_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_21_cast_fp16 = matmul(transpose_x = S_21_transpose_x_1, transpose_y = S_21_transpose_y_1, x = var_3203_cast_fp16, y = u_25_cast_fp16)[name = string("S_21_cast_fp16")];
            tensor<int32, [4]> q_51_begin_0 = const()[name = string("q_51_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_51_end_0 = const()[name = string("q_51_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> q_51_end_mask_0 = const()[name = string("q_51_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_51_cast_fp16 = slice_by_index(begin = q_51_begin_0, end = q_51_end_0, end_mask = q_51_end_mask_0, x = q_47_cast_fp16)[name = string("q_51_cast_fp16")];
            tensor<int32, [4]> k_51_begin_0 = const()[name = string("k_51_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_51_end_0 = const()[name = string("k_51_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> k_51_end_mask_0 = const()[name = string("k_51_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_51_cast_fp16 = slice_by_index(begin = k_51_begin_0, end = k_51_end_0, end_mask = k_51_end_mask_0, x = k_47_cast_fp16)[name = string("k_51_cast_fp16")];
            tensor<int32, [4]> kb_27_begin_0 = const()[name = string("kb_27_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_27_end_0 = const()[name = string("kb_27_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> kb_27_end_mask_0 = const()[name = string("kb_27_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_27_cast_fp16 = slice_by_index(begin = kb_27_begin_0, end = kb_27_end_0, end_mask = kb_27_end_mask_0, x = var_3108_cast_fp16)[name = string("kb_27_cast_fp16")];
            tensor<int32, [4]> vb_33_begin_0 = const()[name = string("vb_33_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_33_end_0 = const()[name = string("vb_33_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> vb_33_end_mask_0 = const()[name = string("vb_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_33_cast_fp16 = slice_by_index(begin = vb_33_begin_0, end = vb_33_end_0, end_mask = vb_33_end_mask_0, x = vb_29_cast_fp16)[name = string("vb_33_cast_fp16")];
            tensor<int32, [3]> gdec_33_begin_0 = const()[name = string("gdec_33_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_33_end_0 = const()[name = string("gdec_33_end_0"), val = tensor<int32, [3]>([3, 4, 32])];
            tensor<bool, [3]> gdec_33_end_mask_0 = const()[name = string("gdec_33_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_33_cast_fp16 = slice_by_index(begin = gdec_33_begin_0, end = gdec_33_end_0, end_mask = gdec_33_end_mask_0, x = gdec_29_cast_fp16)[name = string("gdec_33_cast_fp16")];
            tensor<int32, [2]> var_3288 = const()[name = string("op_3288"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_3289_cast_fp16 = reshape(shape = var_3288, x = gdec_33_cast_fp16)[name = string("op_3289_cast_fp16")];
            tensor<fp16, [12, 256]> var_3290_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3289_cast_fp16)[name = string("op_3290_cast_fp16")];
            tensor<int32, [4]> var_3295 = const()[name = string("op_3295"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_27_cast_fp16 = reshape(shape = var_3295, x = var_3290_cast_fp16)[name = string("P_27_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3297_cast_fp16 = exp(x = P_27_cast_fp16)[name = string("op_3297_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_27_cast_fp16 = mul(x = var_3297_cast_fp16, y = tril_incl_to_fp16)[name = string("D_27_cast_fp16")];
            bool kk_27_transpose_x_1 = const()[name = string("kk_27_transpose_x_1"), val = bool(false)];
            bool kk_27_transpose_y_1 = const()[name = string("kk_27_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_27_cast_fp16 = matmul(transpose_x = kk_27_transpose_x_1, transpose_y = kk_27_transpose_y_1, x = kb_27_cast_fp16, y = k_51_cast_fp16)[name = string("kk_27_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_3303_cast_fp16 = mul(x = kk_27_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_3303_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3304_cast_fp16 = mul(x = var_3303_cast_fp16, y = D_27_cast_fp16)[name = string("op_3304_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_27_cast_fp16 = mul(x = var_3304_cast_fp16, y = cum_tail_to_fp16)[name = string("N_27_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3306_cast_fp16 = mul(x = D_27_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3306_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_27_cast_fp16 = mul(x = kk_27_cast_fp16, y = var_3306_cast_fp16)[name = string("Na_27_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_105_cast_fp16 = add(x = eye_to_fp16, y = N_27_cast_fp16)[name = string("Minv_105_cast_fp16")];
            bool Np_79_transpose_x_0 = const()[name = string("Np_79_transpose_x_0"), val = bool(false)];
            bool Np_79_transpose_y_0 = const()[name = string("Np_79_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_79_cast_fp16 = matmul(transpose_x = Np_79_transpose_x_0, transpose_y = Np_79_transpose_y_0, x = N_27_cast_fp16, y = Na_27_cast_fp16)[name = string("Np_79_cast_fp16")];
            bool Npa_53_transpose_x_0 = const()[name = string("Npa_53_transpose_x_0"), val = bool(false)];
            bool Npa_53_transpose_y_0 = const()[name = string("Npa_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_53_cast_fp16 = matmul(transpose_x = Npa_53_transpose_x_0, transpose_y = Npa_53_transpose_y_0, x = Na_27_cast_fp16, y = N_27_cast_fp16)[name = string("Npa_53_cast_fp16")];
            bool var_3312_transpose_x_0 = const()[name = string("op_3312_transpose_x_0"), val = bool(false)];
            bool var_3312_transpose_y_0 = const()[name = string("op_3312_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3312_cast_fp16 = matmul(transpose_x = var_3312_transpose_x_0, transpose_y = var_3312_transpose_y_0, x = Minv_105_cast_fp16, y = Np_79_cast_fp16)[name = string("op_3312_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_107_cast_fp16 = add(x = Minv_105_cast_fp16, y = var_3312_cast_fp16)[name = string("Minv_107_cast_fp16")];
            bool Np_81_transpose_x_0 = const()[name = string("Np_81_transpose_x_0"), val = bool(false)];
            bool Np_81_transpose_y_0 = const()[name = string("Np_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_81_cast_fp16 = matmul(transpose_x = Np_81_transpose_x_0, transpose_y = Np_81_transpose_y_0, x = Np_79_cast_fp16, y = Npa_53_cast_fp16)[name = string("Np_81_cast_fp16")];
            bool Npa_55_transpose_x_0 = const()[name = string("Npa_55_transpose_x_0"), val = bool(false)];
            bool Npa_55_transpose_y_0 = const()[name = string("Npa_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_55_cast_fp16 = matmul(transpose_x = Npa_55_transpose_x_0, transpose_y = Npa_55_transpose_y_0, x = Npa_53_cast_fp16, y = Np_79_cast_fp16)[name = string("Npa_55_cast_fp16")];
            bool var_3317_transpose_x_0 = const()[name = string("op_3317_transpose_x_0"), val = bool(false)];
            bool var_3317_transpose_y_0 = const()[name = string("op_3317_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3317_cast_fp16 = matmul(transpose_x = var_3317_transpose_x_0, transpose_y = var_3317_transpose_y_0, x = Minv_107_cast_fp16, y = Np_81_cast_fp16)[name = string("op_3317_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_109_cast_fp16 = add(x = Minv_107_cast_fp16, y = var_3317_cast_fp16)[name = string("Minv_109_cast_fp16")];
            bool Np_83_transpose_x_0 = const()[name = string("Np_83_transpose_x_0"), val = bool(false)];
            bool Np_83_transpose_y_0 = const()[name = string("Np_83_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_83_cast_fp16 = matmul(transpose_x = Np_83_transpose_x_0, transpose_y = Np_83_transpose_y_0, x = Np_81_cast_fp16, y = Npa_55_cast_fp16)[name = string("Np_83_cast_fp16")];
            bool var_3321_transpose_x_0 = const()[name = string("op_3321_transpose_x_0"), val = bool(false)];
            bool var_3321_transpose_y_0 = const()[name = string("op_3321_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3321_cast_fp16 = matmul(transpose_x = var_3321_transpose_x_0, transpose_y = var_3321_transpose_y_0, x = Minv_109_cast_fp16, y = Np_83_cast_fp16)[name = string("op_3321_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_111_cast_fp16 = add(x = Minv_109_cast_fp16, y = var_3321_cast_fp16)[name = string("Minv_111_cast_fp16")];
            bool u_27_transpose_x_0 = const()[name = string("u_27_transpose_x_0"), val = bool(false)];
            bool u_27_transpose_y_0 = const()[name = string("u_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_27_cast_fp16 = matmul(transpose_x = u_27_transpose_x_0, transpose_y = u_27_transpose_y_0, x = Minv_111_cast_fp16, y = vb_33_cast_fp16)[name = string("u_27_cast_fp16")];
            bool var_3328_transpose_x_1 = const()[name = string("op_3328_transpose_x_1"), val = bool(false)];
            bool var_3328_transpose_y_1 = const()[name = string("op_3328_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_3328_cast_fp16 = matmul(transpose_x = var_3328_transpose_x_1, transpose_y = var_3328_transpose_y_1, x = q_51_cast_fp16, y = k_51_cast_fp16)[name = string("op_3328_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_27_cast_fp16 = mul(x = var_3328_cast_fp16, y = D_27_cast_fp16)[name = string("A_27_cast_fp16")];
            tensor<fp16, [16]> var_3330_bias_0_to_fp16 = const()[name = string("op_3330_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_3330_cast_fp16 = linear(bias = var_3330_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_33_cast_fp16)[name = string("op_3330_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_21_cast_fp16 = exp(x = var_3330_cast_fp16)[name = string("cum_21_cast_fp16")];
            tensor<int32, [1]> var_3333_axes_0 = const()[name = string("op_3333_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_3333_cast_fp16 = expand_dims(axes = var_3333_axes_0, x = cum_21_cast_fp16)[name = string("op_3333_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3334_cast_fp16 = mul(x = kb_27_cast_fp16, y = var_3333_cast_fp16)[name = string("op_3334_cast_fp16")];
            bool w_27_transpose_x_0 = const()[name = string("w_27_transpose_x_0"), val = bool(false)];
            bool w_27_transpose_y_0 = const()[name = string("w_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_27_cast_fp16 = matmul(transpose_x = w_27_transpose_x_0, transpose_y = w_27_transpose_y_0, x = Minv_111_cast_fp16, y = var_3334_cast_fp16)[name = string("w_27_cast_fp16")];
            bool var_3336_transpose_x_0 = const()[name = string("op_3336_transpose_x_0"), val = bool(false)];
            bool var_3336_transpose_y_0 = const()[name = string("op_3336_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3336_cast_fp16 = matmul(transpose_x = var_3336_transpose_x_0, transpose_y = var_3336_transpose_y_0, x = w_27_cast_fp16, y = S_21_cast_fp16)[name = string("op_3336_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_21_cast_fp16 = sub(x = u_27_cast_fp16, y = var_3336_cast_fp16)[name = string("vnew_21_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3341_cast_fp16 = mul(x = q_51_cast_fp16, y = var_3333_cast_fp16)[name = string("op_3341_cast_fp16")];
            bool var_3342_transpose_x_0 = const()[name = string("op_3342_transpose_x_0"), val = bool(false)];
            bool var_3342_transpose_y_0 = const()[name = string("op_3342_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3342_cast_fp16 = matmul(transpose_x = var_3342_transpose_x_0, transpose_y = var_3342_transpose_y_0, x = var_3341_cast_fp16, y = S_21_cast_fp16)[name = string("op_3342_cast_fp16")];
            bool var_3343_transpose_x_0 = const()[name = string("op_3343_transpose_x_0"), val = bool(false)];
            bool var_3343_transpose_y_0 = const()[name = string("op_3343_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3343_cast_fp16 = matmul(transpose_x = var_3343_transpose_x_0, transpose_y = var_3343_transpose_y_0, x = A_27_cast_fp16, y = vnew_21_cast_fp16)[name = string("op_3343_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3345_cast_fp16 = add(x = var_3342_cast_fp16, y = var_3343_cast_fp16)[name = string("op_3345_cast_fp16")];
            tensor<fp16, [16]> var_3346_bias_0_to_fp16 = const()[name = string("op_3346_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_3346_cast_fp16 = linear(bias = var_3346_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_33_cast_fp16)[name = string("op_3346_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_23_cast_fp16 = exp(x = var_3346_cast_fp16)[name = string("tail_23_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_45_to_fp16 = const()[name = string("transpose_45_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_3348_bias_0_to_fp16 = const()[name = string("op_3348_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_3348_cast_fp16 = linear(bias = var_3348_bias_0_to_fp16, weight = transpose_45_to_fp16, x = gdec_33_cast_fp16)[name = string("op_3348_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_3349_cast_fp16 = exp(x = var_3348_cast_fp16)[name = string("op_3349_cast_fp16")];
            tensor<int32, [1]> last_17_axes_0 = const()[name = string("last_17_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_17_cast_fp16 = expand_dims(axes = last_17_axes_0, x = var_3349_cast_fp16)[name = string("last_17_cast_fp16")];
            tensor<int32, [1]> var_3353_axes_0 = const()[name = string("op_3353_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_3353_cast_fp16 = expand_dims(axes = var_3353_axes_0, x = tail_23_cast_fp16)[name = string("op_3353_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3354_cast_fp16 = mul(x = k_51_cast_fp16, y = var_3353_cast_fp16)[name = string("op_3354_cast_fp16")];
            bool upd_17_transpose_x_1 = const()[name = string("upd_17_transpose_x_1"), val = bool(true)];
            bool upd_17_transpose_y_1 = const()[name = string("upd_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_17_cast_fp16 = matmul(transpose_x = upd_17_transpose_x_1, transpose_y = upd_17_transpose_y_1, x = var_3354_cast_fp16, y = vnew_21_cast_fp16)[name = string("upd_17_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_3359_cast_fp16 = mul(x = S_21_cast_fp16, y = last_17_cast_fp16)[name = string("op_3359_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_23_cast_fp16 = add(x = var_3359_cast_fp16, y = upd_17_cast_fp16)[name = string("S_23_cast_fp16")];
            tensor<int32, [4]> q_53_begin_0 = const()[name = string("q_53_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_53_end_0 = const()[name = string("q_53_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> q_53_end_mask_0 = const()[name = string("q_53_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_53_cast_fp16 = slice_by_index(begin = q_53_begin_0, end = q_53_end_0, end_mask = q_53_end_mask_0, x = q_47_cast_fp16)[name = string("q_53_cast_fp16")];
            tensor<int32, [4]> k_53_begin_0 = const()[name = string("k_53_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_53_end_0 = const()[name = string("k_53_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> k_53_end_mask_0 = const()[name = string("k_53_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_53_cast_fp16 = slice_by_index(begin = k_53_begin_0, end = k_53_end_0, end_mask = k_53_end_mask_0, x = k_47_cast_fp16)[name = string("k_53_cast_fp16")];
            tensor<int32, [4]> kb_29_begin_0 = const()[name = string("kb_29_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_29_end_0 = const()[name = string("kb_29_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> kb_29_end_mask_0 = const()[name = string("kb_29_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_29_cast_fp16 = slice_by_index(begin = kb_29_begin_0, end = kb_29_end_0, end_mask = kb_29_end_mask_0, x = var_3108_cast_fp16)[name = string("kb_29_cast_fp16")];
            tensor<int32, [4]> vb_35_begin_0 = const()[name = string("vb_35_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_35_end_0 = const()[name = string("vb_35_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> vb_35_end_mask_0 = const()[name = string("vb_35_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_35_cast_fp16 = slice_by_index(begin = vb_35_begin_0, end = vb_35_end_0, end_mask = vb_35_end_mask_0, x = vb_29_cast_fp16)[name = string("vb_35_cast_fp16")];
            tensor<int32, [3]> gdec_35_begin_0 = const()[name = string("gdec_35_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_35_end_0 = const()[name = string("gdec_35_end_0"), val = tensor<int32, [3]>([3, 4, 48])];
            tensor<bool, [3]> gdec_35_end_mask_0 = const()[name = string("gdec_35_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_35_cast_fp16 = slice_by_index(begin = gdec_35_begin_0, end = gdec_35_end_0, end_mask = gdec_35_end_mask_0, x = gdec_29_cast_fp16)[name = string("gdec_35_cast_fp16")];
            tensor<int32, [2]> var_3442 = const()[name = string("op_3442"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_3443_cast_fp16 = reshape(shape = var_3442, x = gdec_35_cast_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [12, 256]> var_3444_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3443_cast_fp16)[name = string("op_3444_cast_fp16")];
            tensor<int32, [4]> var_3449 = const()[name = string("op_3449"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_29_cast_fp16 = reshape(shape = var_3449, x = var_3444_cast_fp16)[name = string("P_29_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3451_cast_fp16 = exp(x = P_29_cast_fp16)[name = string("op_3451_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_29_cast_fp16 = mul(x = var_3451_cast_fp16, y = tril_incl_to_fp16)[name = string("D_29_cast_fp16")];
            bool kk_29_transpose_x_1 = const()[name = string("kk_29_transpose_x_1"), val = bool(false)];
            bool kk_29_transpose_y_1 = const()[name = string("kk_29_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_29_cast_fp16 = matmul(transpose_x = kk_29_transpose_x_1, transpose_y = kk_29_transpose_y_1, x = kb_29_cast_fp16, y = k_53_cast_fp16)[name = string("kk_29_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_3457_cast_fp16 = mul(x = kk_29_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_3457_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3458_cast_fp16 = mul(x = var_3457_cast_fp16, y = D_29_cast_fp16)[name = string("op_3458_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_29_cast_fp16 = mul(x = var_3458_cast_fp16, y = cum_tail_to_fp16)[name = string("N_29_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3460_cast_fp16 = mul(x = D_29_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3460_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_29_cast_fp16 = mul(x = kk_29_cast_fp16, y = var_3460_cast_fp16)[name = string("Na_29_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_113_cast_fp16 = add(x = eye_to_fp16, y = N_29_cast_fp16)[name = string("Minv_113_cast_fp16")];
            bool Np_85_transpose_x_0 = const()[name = string("Np_85_transpose_x_0"), val = bool(false)];
            bool Np_85_transpose_y_0 = const()[name = string("Np_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_85_cast_fp16 = matmul(transpose_x = Np_85_transpose_x_0, transpose_y = Np_85_transpose_y_0, x = N_29_cast_fp16, y = Na_29_cast_fp16)[name = string("Np_85_cast_fp16")];
            bool Npa_57_transpose_x_0 = const()[name = string("Npa_57_transpose_x_0"), val = bool(false)];
            bool Npa_57_transpose_y_0 = const()[name = string("Npa_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_57_cast_fp16 = matmul(transpose_x = Npa_57_transpose_x_0, transpose_y = Npa_57_transpose_y_0, x = Na_29_cast_fp16, y = N_29_cast_fp16)[name = string("Npa_57_cast_fp16")];
            bool var_3466_transpose_x_0 = const()[name = string("op_3466_transpose_x_0"), val = bool(false)];
            bool var_3466_transpose_y_0 = const()[name = string("op_3466_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3466_cast_fp16 = matmul(transpose_x = var_3466_transpose_x_0, transpose_y = var_3466_transpose_y_0, x = Minv_113_cast_fp16, y = Np_85_cast_fp16)[name = string("op_3466_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_115_cast_fp16 = add(x = Minv_113_cast_fp16, y = var_3466_cast_fp16)[name = string("Minv_115_cast_fp16")];
            bool Np_87_transpose_x_0 = const()[name = string("Np_87_transpose_x_0"), val = bool(false)];
            bool Np_87_transpose_y_0 = const()[name = string("Np_87_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_87_cast_fp16 = matmul(transpose_x = Np_87_transpose_x_0, transpose_y = Np_87_transpose_y_0, x = Np_85_cast_fp16, y = Npa_57_cast_fp16)[name = string("Np_87_cast_fp16")];
            bool Npa_59_transpose_x_0 = const()[name = string("Npa_59_transpose_x_0"), val = bool(false)];
            bool Npa_59_transpose_y_0 = const()[name = string("Npa_59_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_59_cast_fp16 = matmul(transpose_x = Npa_59_transpose_x_0, transpose_y = Npa_59_transpose_y_0, x = Npa_57_cast_fp16, y = Np_85_cast_fp16)[name = string("Npa_59_cast_fp16")];
            bool var_3471_transpose_x_0 = const()[name = string("op_3471_transpose_x_0"), val = bool(false)];
            bool var_3471_transpose_y_0 = const()[name = string("op_3471_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3471_cast_fp16 = matmul(transpose_x = var_3471_transpose_x_0, transpose_y = var_3471_transpose_y_0, x = Minv_115_cast_fp16, y = Np_87_cast_fp16)[name = string("op_3471_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_117_cast_fp16 = add(x = Minv_115_cast_fp16, y = var_3471_cast_fp16)[name = string("Minv_117_cast_fp16")];
            bool Np_89_transpose_x_0 = const()[name = string("Np_89_transpose_x_0"), val = bool(false)];
            bool Np_89_transpose_y_0 = const()[name = string("Np_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_89_cast_fp16 = matmul(transpose_x = Np_89_transpose_x_0, transpose_y = Np_89_transpose_y_0, x = Np_87_cast_fp16, y = Npa_59_cast_fp16)[name = string("Np_89_cast_fp16")];
            bool var_3475_transpose_x_0 = const()[name = string("op_3475_transpose_x_0"), val = bool(false)];
            bool var_3475_transpose_y_0 = const()[name = string("op_3475_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3475_cast_fp16 = matmul(transpose_x = var_3475_transpose_x_0, transpose_y = var_3475_transpose_y_0, x = Minv_117_cast_fp16, y = Np_89_cast_fp16)[name = string("op_3475_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_119_cast_fp16 = add(x = Minv_117_cast_fp16, y = var_3475_cast_fp16)[name = string("Minv_119_cast_fp16")];
            bool u_29_transpose_x_0 = const()[name = string("u_29_transpose_x_0"), val = bool(false)];
            bool u_29_transpose_y_0 = const()[name = string("u_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_29_cast_fp16 = matmul(transpose_x = u_29_transpose_x_0, transpose_y = u_29_transpose_y_0, x = Minv_119_cast_fp16, y = vb_35_cast_fp16)[name = string("u_29_cast_fp16")];
            bool var_3482_transpose_x_1 = const()[name = string("op_3482_transpose_x_1"), val = bool(false)];
            bool var_3482_transpose_y_1 = const()[name = string("op_3482_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_3482_cast_fp16 = matmul(transpose_x = var_3482_transpose_x_1, transpose_y = var_3482_transpose_y_1, x = q_53_cast_fp16, y = k_53_cast_fp16)[name = string("op_3482_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_29_cast_fp16 = mul(x = var_3482_cast_fp16, y = D_29_cast_fp16)[name = string("A_29_cast_fp16")];
            tensor<fp16, [16]> var_3484_bias_0_to_fp16 = const()[name = string("op_3484_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_3484_cast_fp16 = linear(bias = var_3484_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_35_cast_fp16)[name = string("op_3484_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_23_cast_fp16 = exp(x = var_3484_cast_fp16)[name = string("cum_23_cast_fp16")];
            tensor<int32, [1]> var_3487_axes_0 = const()[name = string("op_3487_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_3487_cast_fp16 = expand_dims(axes = var_3487_axes_0, x = cum_23_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3488_cast_fp16 = mul(x = kb_29_cast_fp16, y = var_3487_cast_fp16)[name = string("op_3488_cast_fp16")];
            bool w_29_transpose_x_0 = const()[name = string("w_29_transpose_x_0"), val = bool(false)];
            bool w_29_transpose_y_0 = const()[name = string("w_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_29_cast_fp16 = matmul(transpose_x = w_29_transpose_x_0, transpose_y = w_29_transpose_y_0, x = Minv_119_cast_fp16, y = var_3488_cast_fp16)[name = string("w_29_cast_fp16")];
            bool var_3490_transpose_x_0 = const()[name = string("op_3490_transpose_x_0"), val = bool(false)];
            bool var_3490_transpose_y_0 = const()[name = string("op_3490_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3490_cast_fp16 = matmul(transpose_x = var_3490_transpose_x_0, transpose_y = var_3490_transpose_y_0, x = w_29_cast_fp16, y = S_23_cast_fp16)[name = string("op_3490_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_23_cast_fp16 = sub(x = u_29_cast_fp16, y = var_3490_cast_fp16)[name = string("vnew_23_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3495_cast_fp16 = mul(x = q_53_cast_fp16, y = var_3487_cast_fp16)[name = string("op_3495_cast_fp16")];
            bool var_3496_transpose_x_0 = const()[name = string("op_3496_transpose_x_0"), val = bool(false)];
            bool var_3496_transpose_y_0 = const()[name = string("op_3496_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3496_cast_fp16 = matmul(transpose_x = var_3496_transpose_x_0, transpose_y = var_3496_transpose_y_0, x = var_3495_cast_fp16, y = S_23_cast_fp16)[name = string("op_3496_cast_fp16")];
            bool var_3497_transpose_x_0 = const()[name = string("op_3497_transpose_x_0"), val = bool(false)];
            bool var_3497_transpose_y_0 = const()[name = string("op_3497_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3497_cast_fp16 = matmul(transpose_x = var_3497_transpose_x_0, transpose_y = var_3497_transpose_y_0, x = A_29_cast_fp16, y = vnew_23_cast_fp16)[name = string("op_3497_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3499_cast_fp16 = add(x = var_3496_cast_fp16, y = var_3497_cast_fp16)[name = string("op_3499_cast_fp16")];
            tensor<fp16, [16]> var_3500_bias_0_to_fp16 = const()[name = string("op_3500_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_3500_cast_fp16 = linear(bias = var_3500_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_35_cast_fp16)[name = string("op_3500_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_25_cast_fp16 = exp(x = var_3500_cast_fp16)[name = string("tail_25_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_49_to_fp16 = const()[name = string("transpose_49_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_3502_bias_0_to_fp16 = const()[name = string("op_3502_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_3502_cast_fp16 = linear(bias = var_3502_bias_0_to_fp16, weight = transpose_49_to_fp16, x = gdec_35_cast_fp16)[name = string("op_3502_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_3503_cast_fp16 = exp(x = var_3502_cast_fp16)[name = string("op_3503_cast_fp16")];
            tensor<int32, [1]> last_19_axes_0 = const()[name = string("last_19_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_19_cast_fp16 = expand_dims(axes = last_19_axes_0, x = var_3503_cast_fp16)[name = string("last_19_cast_fp16")];
            tensor<int32, [1]> var_3507_axes_0 = const()[name = string("op_3507_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_3507_cast_fp16 = expand_dims(axes = var_3507_axes_0, x = tail_25_cast_fp16)[name = string("op_3507_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3508_cast_fp16 = mul(x = k_53_cast_fp16, y = var_3507_cast_fp16)[name = string("op_3508_cast_fp16")];
            bool upd_19_transpose_x_1 = const()[name = string("upd_19_transpose_x_1"), val = bool(true)];
            bool upd_19_transpose_y_1 = const()[name = string("upd_19_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_19_cast_fp16 = matmul(transpose_x = upd_19_transpose_x_1, transpose_y = upd_19_transpose_y_1, x = var_3508_cast_fp16, y = vnew_23_cast_fp16)[name = string("upd_19_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_3513_cast_fp16 = mul(x = S_23_cast_fp16, y = last_19_cast_fp16)[name = string("op_3513_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_25_cast_fp16 = add(x = var_3513_cast_fp16, y = upd_19_cast_fp16)[name = string("S_25_cast_fp16")];
            tensor<int32, [4]> q_55_begin_0 = const()[name = string("q_55_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_55_end_0 = const()[name = string("q_55_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> q_55_end_mask_0 = const()[name = string("q_55_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, x = q_47_cast_fp16)[name = string("q_55_cast_fp16")];
            tensor<int32, [4]> k_55_begin_0 = const()[name = string("k_55_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_55_end_0 = const()[name = string("k_55_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> k_55_end_mask_0 = const()[name = string("k_55_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_55_cast_fp16 = slice_by_index(begin = k_55_begin_0, end = k_55_end_0, end_mask = k_55_end_mask_0, x = k_47_cast_fp16)[name = string("k_55_cast_fp16")];
            tensor<int32, [4]> kb_31_begin_0 = const()[name = string("kb_31_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_31_end_0 = const()[name = string("kb_31_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> kb_31_end_mask_0 = const()[name = string("kb_31_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_31_cast_fp16 = slice_by_index(begin = kb_31_begin_0, end = kb_31_end_0, end_mask = kb_31_end_mask_0, x = var_3108_cast_fp16)[name = string("kb_31_cast_fp16")];
            tensor<int32, [4]> vb_37_begin_0 = const()[name = string("vb_37_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_37_end_0 = const()[name = string("vb_37_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> vb_37_end_mask_0 = const()[name = string("vb_37_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_37_cast_fp16 = slice_by_index(begin = vb_37_begin_0, end = vb_37_end_0, end_mask = vb_37_end_mask_0, x = vb_29_cast_fp16)[name = string("vb_37_cast_fp16")];
            tensor<int32, [3]> gdec_37_begin_0 = const()[name = string("gdec_37_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_37_end_0 = const()[name = string("gdec_37_end_0"), val = tensor<int32, [3]>([3, 4, 64])];
            tensor<bool, [3]> gdec_37_end_mask_0 = const()[name = string("gdec_37_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_37_cast_fp16 = slice_by_index(begin = gdec_37_begin_0, end = gdec_37_end_0, end_mask = gdec_37_end_mask_0, x = gdec_29_cast_fp16)[name = string("gdec_37_cast_fp16")];
            tensor<int32, [2]> var_3596 = const()[name = string("op_3596"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_3597_cast_fp16 = reshape(shape = var_3596, x = gdec_37_cast_fp16)[name = string("op_3597_cast_fp16")];
            tensor<fp16, [12, 256]> var_3598_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3597_cast_fp16)[name = string("op_3598_cast_fp16")];
            tensor<int32, [4]> var_3603 = const()[name = string("op_3603"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_31_cast_fp16 = reshape(shape = var_3603, x = var_3598_cast_fp16)[name = string("P_31_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3605_cast_fp16 = exp(x = P_31_cast_fp16)[name = string("op_3605_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_31_cast_fp16 = mul(x = var_3605_cast_fp16, y = tril_incl_to_fp16)[name = string("D_31_cast_fp16")];
            bool kk_31_transpose_x_1 = const()[name = string("kk_31_transpose_x_1"), val = bool(false)];
            bool kk_31_transpose_y_1 = const()[name = string("kk_31_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_31_cast_fp16 = matmul(transpose_x = kk_31_transpose_x_1, transpose_y = kk_31_transpose_y_1, x = kb_31_cast_fp16, y = k_55_cast_fp16)[name = string("kk_31_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_3611_cast_fp16 = mul(x = kk_31_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_3611_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3612_cast_fp16 = mul(x = var_3611_cast_fp16, y = D_31_cast_fp16)[name = string("op_3612_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_31_cast_fp16 = mul(x = var_3612_cast_fp16, y = cum_tail_to_fp16)[name = string("N_31_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3614_cast_fp16 = mul(x = D_31_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3614_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_31_cast_fp16 = mul(x = kk_31_cast_fp16, y = var_3614_cast_fp16)[name = string("Na_31_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_121_cast_fp16 = add(x = eye_to_fp16, y = N_31_cast_fp16)[name = string("Minv_121_cast_fp16")];
            bool Np_91_transpose_x_0 = const()[name = string("Np_91_transpose_x_0"), val = bool(false)];
            bool Np_91_transpose_y_0 = const()[name = string("Np_91_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_91_cast_fp16 = matmul(transpose_x = Np_91_transpose_x_0, transpose_y = Np_91_transpose_y_0, x = N_31_cast_fp16, y = Na_31_cast_fp16)[name = string("Np_91_cast_fp16")];
            bool Npa_61_transpose_x_0 = const()[name = string("Npa_61_transpose_x_0"), val = bool(false)];
            bool Npa_61_transpose_y_0 = const()[name = string("Npa_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_61_cast_fp16 = matmul(transpose_x = Npa_61_transpose_x_0, transpose_y = Npa_61_transpose_y_0, x = Na_31_cast_fp16, y = N_31_cast_fp16)[name = string("Npa_61_cast_fp16")];
            bool var_3620_transpose_x_0 = const()[name = string("op_3620_transpose_x_0"), val = bool(false)];
            bool var_3620_transpose_y_0 = const()[name = string("op_3620_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3620_cast_fp16 = matmul(transpose_x = var_3620_transpose_x_0, transpose_y = var_3620_transpose_y_0, x = Minv_121_cast_fp16, y = Np_91_cast_fp16)[name = string("op_3620_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_123_cast_fp16 = add(x = Minv_121_cast_fp16, y = var_3620_cast_fp16)[name = string("Minv_123_cast_fp16")];
            bool Np_93_transpose_x_0 = const()[name = string("Np_93_transpose_x_0"), val = bool(false)];
            bool Np_93_transpose_y_0 = const()[name = string("Np_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_93_cast_fp16 = matmul(transpose_x = Np_93_transpose_x_0, transpose_y = Np_93_transpose_y_0, x = Np_91_cast_fp16, y = Npa_61_cast_fp16)[name = string("Np_93_cast_fp16")];
            bool Npa_63_transpose_x_0 = const()[name = string("Npa_63_transpose_x_0"), val = bool(false)];
            bool Npa_63_transpose_y_0 = const()[name = string("Npa_63_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_63_cast_fp16 = matmul(transpose_x = Npa_63_transpose_x_0, transpose_y = Npa_63_transpose_y_0, x = Npa_61_cast_fp16, y = Np_91_cast_fp16)[name = string("Npa_63_cast_fp16")];
            bool var_3625_transpose_x_0 = const()[name = string("op_3625_transpose_x_0"), val = bool(false)];
            bool var_3625_transpose_y_0 = const()[name = string("op_3625_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3625_cast_fp16 = matmul(transpose_x = var_3625_transpose_x_0, transpose_y = var_3625_transpose_y_0, x = Minv_123_cast_fp16, y = Np_93_cast_fp16)[name = string("op_3625_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_125_cast_fp16 = add(x = Minv_123_cast_fp16, y = var_3625_cast_fp16)[name = string("Minv_125_cast_fp16")];
            bool Np_95_transpose_x_0 = const()[name = string("Np_95_transpose_x_0"), val = bool(false)];
            bool Np_95_transpose_y_0 = const()[name = string("Np_95_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_95_cast_fp16 = matmul(transpose_x = Np_95_transpose_x_0, transpose_y = Np_95_transpose_y_0, x = Np_93_cast_fp16, y = Npa_63_cast_fp16)[name = string("Np_95_cast_fp16")];
            bool var_3629_transpose_x_0 = const()[name = string("op_3629_transpose_x_0"), val = bool(false)];
            bool var_3629_transpose_y_0 = const()[name = string("op_3629_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3629_cast_fp16 = matmul(transpose_x = var_3629_transpose_x_0, transpose_y = var_3629_transpose_y_0, x = Minv_125_cast_fp16, y = Np_95_cast_fp16)[name = string("op_3629_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_127_cast_fp16 = add(x = Minv_125_cast_fp16, y = var_3629_cast_fp16)[name = string("Minv_127_cast_fp16")];
            bool u_31_transpose_x_0 = const()[name = string("u_31_transpose_x_0"), val = bool(false)];
            bool u_31_transpose_y_0 = const()[name = string("u_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_31_cast_fp16 = matmul(transpose_x = u_31_transpose_x_0, transpose_y = u_31_transpose_y_0, x = Minv_127_cast_fp16, y = vb_37_cast_fp16)[name = string("u_31_cast_fp16")];
            bool var_3636_transpose_x_1 = const()[name = string("op_3636_transpose_x_1"), val = bool(false)];
            bool var_3636_transpose_y_1 = const()[name = string("op_3636_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_3636_cast_fp16 = matmul(transpose_x = var_3636_transpose_x_1, transpose_y = var_3636_transpose_y_1, x = q_55_cast_fp16, y = k_55_cast_fp16)[name = string("op_3636_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_31_cast_fp16 = mul(x = var_3636_cast_fp16, y = D_31_cast_fp16)[name = string("A_31_cast_fp16")];
            tensor<fp16, [16]> var_3638_bias_0_to_fp16 = const()[name = string("op_3638_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_3638_cast_fp16 = linear(bias = var_3638_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_37_cast_fp16)[name = string("op_3638_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_25_cast_fp16 = exp(x = var_3638_cast_fp16)[name = string("cum_25_cast_fp16")];
            tensor<int32, [1]> var_3641_axes_0 = const()[name = string("op_3641_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_3641_cast_fp16 = expand_dims(axes = var_3641_axes_0, x = cum_25_cast_fp16)[name = string("op_3641_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3642_cast_fp16 = mul(x = kb_31_cast_fp16, y = var_3641_cast_fp16)[name = string("op_3642_cast_fp16")];
            bool w_31_transpose_x_0 = const()[name = string("w_31_transpose_x_0"), val = bool(false)];
            bool w_31_transpose_y_0 = const()[name = string("w_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_31_cast_fp16 = matmul(transpose_x = w_31_transpose_x_0, transpose_y = w_31_transpose_y_0, x = Minv_127_cast_fp16, y = var_3642_cast_fp16)[name = string("w_31_cast_fp16")];
            bool var_3644_transpose_x_0 = const()[name = string("op_3644_transpose_x_0"), val = bool(false)];
            bool var_3644_transpose_y_0 = const()[name = string("op_3644_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3644_cast_fp16 = matmul(transpose_x = var_3644_transpose_x_0, transpose_y = var_3644_transpose_y_0, x = w_31_cast_fp16, y = S_25_cast_fp16)[name = string("op_3644_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_25_cast_fp16 = sub(x = u_31_cast_fp16, y = var_3644_cast_fp16)[name = string("vnew_25_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3649_cast_fp16 = mul(x = q_55_cast_fp16, y = var_3641_cast_fp16)[name = string("op_3649_cast_fp16")];
            bool var_3650_transpose_x_0 = const()[name = string("op_3650_transpose_x_0"), val = bool(false)];
            bool var_3650_transpose_y_0 = const()[name = string("op_3650_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3650_cast_fp16 = matmul(transpose_x = var_3650_transpose_x_0, transpose_y = var_3650_transpose_y_0, x = var_3649_cast_fp16, y = S_25_cast_fp16)[name = string("op_3650_cast_fp16")];
            bool var_3651_transpose_x_0 = const()[name = string("op_3651_transpose_x_0"), val = bool(false)];
            bool var_3651_transpose_y_0 = const()[name = string("op_3651_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3651_cast_fp16 = matmul(transpose_x = var_3651_transpose_x_0, transpose_y = var_3651_transpose_y_0, x = A_31_cast_fp16, y = vnew_25_cast_fp16)[name = string("op_3651_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3653_cast_fp16 = add(x = var_3650_cast_fp16, y = var_3651_cast_fp16)[name = string("op_3653_cast_fp16")];
            tensor<fp16, [16]> var_3654_bias_0_to_fp16 = const()[name = string("op_3654_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_3654_cast_fp16 = linear(bias = var_3654_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_37_cast_fp16)[name = string("op_3654_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_27_cast_fp16 = exp(x = var_3654_cast_fp16)[name = string("tail_27_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_53_to_fp16 = const()[name = string("transpose_53_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_3656_bias_0_to_fp16 = const()[name = string("op_3656_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_3656_cast_fp16 = linear(bias = var_3656_bias_0_to_fp16, weight = transpose_53_to_fp16, x = gdec_37_cast_fp16)[name = string("op_3656_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_3657_cast_fp16 = exp(x = var_3656_cast_fp16)[name = string("op_3657_cast_fp16")];
            tensor<int32, [1]> last_21_axes_0 = const()[name = string("last_21_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_21_cast_fp16 = expand_dims(axes = last_21_axes_0, x = var_3657_cast_fp16)[name = string("last_21_cast_fp16")];
            tensor<int32, [1]> var_3661_axes_0 = const()[name = string("op_3661_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_3661_cast_fp16 = expand_dims(axes = var_3661_axes_0, x = tail_27_cast_fp16)[name = string("op_3661_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3662_cast_fp16 = mul(x = k_55_cast_fp16, y = var_3661_cast_fp16)[name = string("op_3662_cast_fp16")];
            bool upd_21_transpose_x_1 = const()[name = string("upd_21_transpose_x_1"), val = bool(true)];
            bool upd_21_transpose_y_1 = const()[name = string("upd_21_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_21_cast_fp16 = matmul(transpose_x = upd_21_transpose_x_1, transpose_y = upd_21_transpose_y_1, x = var_3662_cast_fp16, y = vnew_25_cast_fp16)[name = string("upd_21_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_3667_cast_fp16 = mul(x = S_25_cast_fp16, y = last_21_cast_fp16)[name = string("op_3667_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_27_cast_fp16 = add(x = var_3667_cast_fp16, y = upd_21_cast_fp16)[name = string("S_27_cast_fp16")];
            tensor<int32, [4]> q_57_begin_0 = const()[name = string("q_57_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> q_57_end_0 = const()[name = string("q_57_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> q_57_end_mask_0 = const()[name = string("q_57_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_57_cast_fp16 = slice_by_index(begin = q_57_begin_0, end = q_57_end_0, end_mask = q_57_end_mask_0, x = q_47_cast_fp16)[name = string("q_57_cast_fp16")];
            tensor<int32, [4]> k_57_begin_0 = const()[name = string("k_57_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> k_57_end_0 = const()[name = string("k_57_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> k_57_end_mask_0 = const()[name = string("k_57_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_57_cast_fp16 = slice_by_index(begin = k_57_begin_0, end = k_57_end_0, end_mask = k_57_end_mask_0, x = k_47_cast_fp16)[name = string("k_57_cast_fp16")];
            tensor<int32, [4]> kb_33_begin_0 = const()[name = string("kb_33_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> kb_33_end_0 = const()[name = string("kb_33_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> kb_33_end_mask_0 = const()[name = string("kb_33_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_33_cast_fp16 = slice_by_index(begin = kb_33_begin_0, end = kb_33_end_0, end_mask = kb_33_end_mask_0, x = var_3108_cast_fp16)[name = string("kb_33_cast_fp16")];
            tensor<int32, [4]> vb_39_begin_0 = const()[name = string("vb_39_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> vb_39_end_0 = const()[name = string("vb_39_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> vb_39_end_mask_0 = const()[name = string("vb_39_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_39_cast_fp16 = slice_by_index(begin = vb_39_begin_0, end = vb_39_end_0, end_mask = vb_39_end_mask_0, x = vb_29_cast_fp16)[name = string("vb_39_cast_fp16")];
            tensor<int32, [3]> gdec_39_begin_0 = const()[name = string("gdec_39_begin_0"), val = tensor<int32, [3]>([0, 0, 64])];
            tensor<int32, [3]> gdec_39_end_0 = const()[name = string("gdec_39_end_0"), val = tensor<int32, [3]>([3, 4, 80])];
            tensor<bool, [3]> gdec_39_end_mask_0 = const()[name = string("gdec_39_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_39_cast_fp16 = slice_by_index(begin = gdec_39_begin_0, end = gdec_39_end_0, end_mask = gdec_39_end_mask_0, x = gdec_29_cast_fp16)[name = string("gdec_39_cast_fp16")];
            tensor<int32, [2]> var_3750 = const()[name = string("op_3750"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_3751_cast_fp16 = reshape(shape = var_3750, x = gdec_39_cast_fp16)[name = string("op_3751_cast_fp16")];
            tensor<fp16, [12, 256]> var_3752_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3751_cast_fp16)[name = string("op_3752_cast_fp16")];
            tensor<int32, [4]> var_3757 = const()[name = string("op_3757"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_33_cast_fp16 = reshape(shape = var_3757, x = var_3752_cast_fp16)[name = string("P_33_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3759_cast_fp16 = exp(x = P_33_cast_fp16)[name = string("op_3759_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_33_cast_fp16 = mul(x = var_3759_cast_fp16, y = tril_incl_to_fp16)[name = string("D_33_cast_fp16")];
            bool kk_33_transpose_x_1 = const()[name = string("kk_33_transpose_x_1"), val = bool(false)];
            bool kk_33_transpose_y_1 = const()[name = string("kk_33_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_33_cast_fp16 = matmul(transpose_x = kk_33_transpose_x_1, transpose_y = kk_33_transpose_y_1, x = kb_33_cast_fp16, y = k_57_cast_fp16)[name = string("kk_33_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_3765_cast_fp16 = mul(x = kk_33_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_3765_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3766_cast_fp16 = mul(x = var_3765_cast_fp16, y = D_33_cast_fp16)[name = string("op_3766_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_33_cast_fp16 = mul(x = var_3766_cast_fp16, y = cum_tail_to_fp16)[name = string("N_33_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3768_cast_fp16 = mul(x = D_33_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3768_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_33_cast_fp16 = mul(x = kk_33_cast_fp16, y = var_3768_cast_fp16)[name = string("Na_33_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_129_cast_fp16 = add(x = eye_to_fp16, y = N_33_cast_fp16)[name = string("Minv_129_cast_fp16")];
            bool Np_97_transpose_x_0 = const()[name = string("Np_97_transpose_x_0"), val = bool(false)];
            bool Np_97_transpose_y_0 = const()[name = string("Np_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_97_cast_fp16 = matmul(transpose_x = Np_97_transpose_x_0, transpose_y = Np_97_transpose_y_0, x = N_33_cast_fp16, y = Na_33_cast_fp16)[name = string("Np_97_cast_fp16")];
            bool Npa_65_transpose_x_0 = const()[name = string("Npa_65_transpose_x_0"), val = bool(false)];
            bool Npa_65_transpose_y_0 = const()[name = string("Npa_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_65_cast_fp16 = matmul(transpose_x = Npa_65_transpose_x_0, transpose_y = Npa_65_transpose_y_0, x = Na_33_cast_fp16, y = N_33_cast_fp16)[name = string("Npa_65_cast_fp16")];
            bool var_3774_transpose_x_0 = const()[name = string("op_3774_transpose_x_0"), val = bool(false)];
            bool var_3774_transpose_y_0 = const()[name = string("op_3774_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3774_cast_fp16 = matmul(transpose_x = var_3774_transpose_x_0, transpose_y = var_3774_transpose_y_0, x = Minv_129_cast_fp16, y = Np_97_cast_fp16)[name = string("op_3774_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_131_cast_fp16 = add(x = Minv_129_cast_fp16, y = var_3774_cast_fp16)[name = string("Minv_131_cast_fp16")];
            bool Np_99_transpose_x_0 = const()[name = string("Np_99_transpose_x_0"), val = bool(false)];
            bool Np_99_transpose_y_0 = const()[name = string("Np_99_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_99_cast_fp16 = matmul(transpose_x = Np_99_transpose_x_0, transpose_y = Np_99_transpose_y_0, x = Np_97_cast_fp16, y = Npa_65_cast_fp16)[name = string("Np_99_cast_fp16")];
            bool Npa_67_transpose_x_0 = const()[name = string("Npa_67_transpose_x_0"), val = bool(false)];
            bool Npa_67_transpose_y_0 = const()[name = string("Npa_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_67_cast_fp16 = matmul(transpose_x = Npa_67_transpose_x_0, transpose_y = Npa_67_transpose_y_0, x = Npa_65_cast_fp16, y = Np_97_cast_fp16)[name = string("Npa_67_cast_fp16")];
            bool var_3779_transpose_x_0 = const()[name = string("op_3779_transpose_x_0"), val = bool(false)];
            bool var_3779_transpose_y_0 = const()[name = string("op_3779_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3779_cast_fp16 = matmul(transpose_x = var_3779_transpose_x_0, transpose_y = var_3779_transpose_y_0, x = Minv_131_cast_fp16, y = Np_99_cast_fp16)[name = string("op_3779_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_133_cast_fp16 = add(x = Minv_131_cast_fp16, y = var_3779_cast_fp16)[name = string("Minv_133_cast_fp16")];
            bool Np_101_transpose_x_0 = const()[name = string("Np_101_transpose_x_0"), val = bool(false)];
            bool Np_101_transpose_y_0 = const()[name = string("Np_101_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_101_cast_fp16 = matmul(transpose_x = Np_101_transpose_x_0, transpose_y = Np_101_transpose_y_0, x = Np_99_cast_fp16, y = Npa_67_cast_fp16)[name = string("Np_101_cast_fp16")];
            bool var_3783_transpose_x_0 = const()[name = string("op_3783_transpose_x_0"), val = bool(false)];
            bool var_3783_transpose_y_0 = const()[name = string("op_3783_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3783_cast_fp16 = matmul(transpose_x = var_3783_transpose_x_0, transpose_y = var_3783_transpose_y_0, x = Minv_133_cast_fp16, y = Np_101_cast_fp16)[name = string("op_3783_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_135_cast_fp16 = add(x = Minv_133_cast_fp16, y = var_3783_cast_fp16)[name = string("Minv_135_cast_fp16")];
            bool u_33_transpose_x_0 = const()[name = string("u_33_transpose_x_0"), val = bool(false)];
            bool u_33_transpose_y_0 = const()[name = string("u_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_33_cast_fp16 = matmul(transpose_x = u_33_transpose_x_0, transpose_y = u_33_transpose_y_0, x = Minv_135_cast_fp16, y = vb_39_cast_fp16)[name = string("u_33_cast_fp16")];
            bool var_3790_transpose_x_1 = const()[name = string("op_3790_transpose_x_1"), val = bool(false)];
            bool var_3790_transpose_y_1 = const()[name = string("op_3790_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_3790_cast_fp16 = matmul(transpose_x = var_3790_transpose_x_1, transpose_y = var_3790_transpose_y_1, x = q_57_cast_fp16, y = k_57_cast_fp16)[name = string("op_3790_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_33_cast_fp16 = mul(x = var_3790_cast_fp16, y = D_33_cast_fp16)[name = string("A_33_cast_fp16")];
            tensor<fp16, [16]> var_3792_bias_0_to_fp16 = const()[name = string("op_3792_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_3792_cast_fp16 = linear(bias = var_3792_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_39_cast_fp16)[name = string("op_3792_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_27_cast_fp16 = exp(x = var_3792_cast_fp16)[name = string("cum_27_cast_fp16")];
            tensor<int32, [1]> var_3795_axes_0 = const()[name = string("op_3795_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_3795_cast_fp16 = expand_dims(axes = var_3795_axes_0, x = cum_27_cast_fp16)[name = string("op_3795_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3796_cast_fp16 = mul(x = kb_33_cast_fp16, y = var_3795_cast_fp16)[name = string("op_3796_cast_fp16")];
            bool w_33_transpose_x_0 = const()[name = string("w_33_transpose_x_0"), val = bool(false)];
            bool w_33_transpose_y_0 = const()[name = string("w_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_33_cast_fp16 = matmul(transpose_x = w_33_transpose_x_0, transpose_y = w_33_transpose_y_0, x = Minv_135_cast_fp16, y = var_3796_cast_fp16)[name = string("w_33_cast_fp16")];
            bool var_3798_transpose_x_0 = const()[name = string("op_3798_transpose_x_0"), val = bool(false)];
            bool var_3798_transpose_y_0 = const()[name = string("op_3798_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3798_cast_fp16 = matmul(transpose_x = var_3798_transpose_x_0, transpose_y = var_3798_transpose_y_0, x = w_33_cast_fp16, y = S_27_cast_fp16)[name = string("op_3798_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_27_cast_fp16 = sub(x = u_33_cast_fp16, y = var_3798_cast_fp16)[name = string("vnew_27_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3803_cast_fp16 = mul(x = q_57_cast_fp16, y = var_3795_cast_fp16)[name = string("op_3803_cast_fp16")];
            bool var_3804_transpose_x_0 = const()[name = string("op_3804_transpose_x_0"), val = bool(false)];
            bool var_3804_transpose_y_0 = const()[name = string("op_3804_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3804_cast_fp16 = matmul(transpose_x = var_3804_transpose_x_0, transpose_y = var_3804_transpose_y_0, x = var_3803_cast_fp16, y = S_27_cast_fp16)[name = string("op_3804_cast_fp16")];
            bool var_3805_transpose_x_0 = const()[name = string("op_3805_transpose_x_0"), val = bool(false)];
            bool var_3805_transpose_y_0 = const()[name = string("op_3805_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3805_cast_fp16 = matmul(transpose_x = var_3805_transpose_x_0, transpose_y = var_3805_transpose_y_0, x = A_33_cast_fp16, y = vnew_27_cast_fp16)[name = string("op_3805_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3807_cast_fp16 = add(x = var_3804_cast_fp16, y = var_3805_cast_fp16)[name = string("op_3807_cast_fp16")];
            tensor<fp16, [16]> var_3808_bias_0_to_fp16 = const()[name = string("op_3808_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_3808_cast_fp16 = linear(bias = var_3808_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_39_cast_fp16)[name = string("op_3808_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_29_cast_fp16 = exp(x = var_3808_cast_fp16)[name = string("tail_29_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_57_to_fp16 = const()[name = string("transpose_57_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_3810_bias_0_to_fp16 = const()[name = string("op_3810_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_3810_cast_fp16 = linear(bias = var_3810_bias_0_to_fp16, weight = transpose_57_to_fp16, x = gdec_39_cast_fp16)[name = string("op_3810_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_3811_cast_fp16 = exp(x = var_3810_cast_fp16)[name = string("op_3811_cast_fp16")];
            tensor<int32, [1]> last_23_axes_0 = const()[name = string("last_23_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_23_cast_fp16 = expand_dims(axes = last_23_axes_0, x = var_3811_cast_fp16)[name = string("last_23_cast_fp16")];
            tensor<int32, [1]> var_3815_axes_0 = const()[name = string("op_3815_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_3815_cast_fp16 = expand_dims(axes = var_3815_axes_0, x = tail_29_cast_fp16)[name = string("op_3815_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3816_cast_fp16 = mul(x = k_57_cast_fp16, y = var_3815_cast_fp16)[name = string("op_3816_cast_fp16")];
            bool upd_23_transpose_x_1 = const()[name = string("upd_23_transpose_x_1"), val = bool(true)];
            bool upd_23_transpose_y_1 = const()[name = string("upd_23_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_23_cast_fp16 = matmul(transpose_x = upd_23_transpose_x_1, transpose_y = upd_23_transpose_y_1, x = var_3816_cast_fp16, y = vnew_27_cast_fp16)[name = string("upd_23_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_3821_cast_fp16 = mul(x = S_27_cast_fp16, y = last_23_cast_fp16)[name = string("op_3821_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_29_cast_fp16 = add(x = var_3821_cast_fp16, y = upd_23_cast_fp16)[name = string("S_29_cast_fp16")];
            tensor<int32, [4]> q_59_begin_0 = const()[name = string("q_59_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> q_59_end_0 = const()[name = string("q_59_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_59_end_mask_0 = const()[name = string("q_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_59_cast_fp16 = slice_by_index(begin = q_59_begin_0, end = q_59_end_0, end_mask = q_59_end_mask_0, x = q_47_cast_fp16)[name = string("q_59_cast_fp16")];
            tensor<int32, [4]> k_59_begin_0 = const()[name = string("k_59_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> k_59_end_0 = const()[name = string("k_59_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_59_end_mask_0 = const()[name = string("k_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_59_cast_fp16 = slice_by_index(begin = k_59_begin_0, end = k_59_end_0, end_mask = k_59_end_mask_0, x = k_47_cast_fp16)[name = string("k_59_cast_fp16")];
            tensor<int32, [4]> kb_35_begin_0 = const()[name = string("kb_35_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> kb_35_end_0 = const()[name = string("kb_35_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_35_end_mask_0 = const()[name = string("kb_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_35_cast_fp16 = slice_by_index(begin = kb_35_begin_0, end = kb_35_end_0, end_mask = kb_35_end_mask_0, x = var_3108_cast_fp16)[name = string("kb_35_cast_fp16")];
            tensor<int32, [4]> vb_41_begin_0 = const()[name = string("vb_41_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> vb_41_end_0 = const()[name = string("vb_41_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_41_end_mask_0 = const()[name = string("vb_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_41_cast_fp16 = slice_by_index(begin = vb_41_begin_0, end = vb_41_end_0, end_mask = vb_41_end_mask_0, x = vb_29_cast_fp16)[name = string("vb_41_cast_fp16")];
            tensor<int32, [3]> gdec_41_begin_0 = const()[name = string("gdec_41_begin_0"), val = tensor<int32, [3]>([0, 0, 80])];
            tensor<int32, [3]> gdec_41_end_0 = const()[name = string("gdec_41_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_41_end_mask_0 = const()[name = string("gdec_41_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_41_cast_fp16 = slice_by_index(begin = gdec_41_begin_0, end = gdec_41_end_0, end_mask = gdec_41_end_mask_0, x = gdec_29_cast_fp16)[name = string("gdec_41_cast_fp16")];
            tensor<int32, [2]> var_3904 = const()[name = string("op_3904"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_3905_cast_fp16 = reshape(shape = var_3904, x = gdec_41_cast_fp16)[name = string("op_3905_cast_fp16")];
            tensor<fp16, [12, 256]> var_3906_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_3905_cast_fp16)[name = string("op_3906_cast_fp16")];
            tensor<int32, [4]> var_3911 = const()[name = string("op_3911"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_35_cast_fp16 = reshape(shape = var_3911, x = var_3906_cast_fp16)[name = string("P_35_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3913_cast_fp16 = exp(x = P_35_cast_fp16)[name = string("op_3913_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_35_cast_fp16 = mul(x = var_3913_cast_fp16, y = tril_incl_to_fp16)[name = string("D_35_cast_fp16")];
            bool kk_35_transpose_x_1 = const()[name = string("kk_35_transpose_x_1"), val = bool(false)];
            bool kk_35_transpose_y_1 = const()[name = string("kk_35_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_35_cast_fp16 = matmul(transpose_x = kk_35_transpose_x_1, transpose_y = kk_35_transpose_y_1, x = kb_35_cast_fp16, y = k_59_cast_fp16)[name = string("kk_35_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_3919_cast_fp16 = mul(x = kk_35_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_3919_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3920_cast_fp16 = mul(x = var_3919_cast_fp16, y = D_35_cast_fp16)[name = string("op_3920_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_35_cast_fp16 = mul(x = var_3920_cast_fp16, y = cum_tail_to_fp16)[name = string("N_35_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_3922_cast_fp16 = mul(x = D_35_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_3922_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_35_cast_fp16 = mul(x = kk_35_cast_fp16, y = var_3922_cast_fp16)[name = string("Na_35_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_137_cast_fp16 = add(x = eye_to_fp16, y = N_35_cast_fp16)[name = string("Minv_137_cast_fp16")];
            bool Np_103_transpose_x_0 = const()[name = string("Np_103_transpose_x_0"), val = bool(false)];
            bool Np_103_transpose_y_0 = const()[name = string("Np_103_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_103_cast_fp16 = matmul(transpose_x = Np_103_transpose_x_0, transpose_y = Np_103_transpose_y_0, x = N_35_cast_fp16, y = Na_35_cast_fp16)[name = string("Np_103_cast_fp16")];
            bool Npa_69_transpose_x_0 = const()[name = string("Npa_69_transpose_x_0"), val = bool(false)];
            bool Npa_69_transpose_y_0 = const()[name = string("Npa_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_69_cast_fp16 = matmul(transpose_x = Npa_69_transpose_x_0, transpose_y = Npa_69_transpose_y_0, x = Na_35_cast_fp16, y = N_35_cast_fp16)[name = string("Npa_69_cast_fp16")];
            bool var_3928_transpose_x_0 = const()[name = string("op_3928_transpose_x_0"), val = bool(false)];
            bool var_3928_transpose_y_0 = const()[name = string("op_3928_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3928_cast_fp16 = matmul(transpose_x = var_3928_transpose_x_0, transpose_y = var_3928_transpose_y_0, x = Minv_137_cast_fp16, y = Np_103_cast_fp16)[name = string("op_3928_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_139_cast_fp16 = add(x = Minv_137_cast_fp16, y = var_3928_cast_fp16)[name = string("Minv_139_cast_fp16")];
            bool Np_105_transpose_x_0 = const()[name = string("Np_105_transpose_x_0"), val = bool(false)];
            bool Np_105_transpose_y_0 = const()[name = string("Np_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_105_cast_fp16 = matmul(transpose_x = Np_105_transpose_x_0, transpose_y = Np_105_transpose_y_0, x = Np_103_cast_fp16, y = Npa_69_cast_fp16)[name = string("Np_105_cast_fp16")];
            bool Npa_71_transpose_x_0 = const()[name = string("Npa_71_transpose_x_0"), val = bool(false)];
            bool Npa_71_transpose_y_0 = const()[name = string("Npa_71_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_71_cast_fp16 = matmul(transpose_x = Npa_71_transpose_x_0, transpose_y = Npa_71_transpose_y_0, x = Npa_69_cast_fp16, y = Np_103_cast_fp16)[name = string("Npa_71_cast_fp16")];
            bool var_3933_transpose_x_0 = const()[name = string("op_3933_transpose_x_0"), val = bool(false)];
            bool var_3933_transpose_y_0 = const()[name = string("op_3933_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3933_cast_fp16 = matmul(transpose_x = var_3933_transpose_x_0, transpose_y = var_3933_transpose_y_0, x = Minv_139_cast_fp16, y = Np_105_cast_fp16)[name = string("op_3933_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_141_cast_fp16 = add(x = Minv_139_cast_fp16, y = var_3933_cast_fp16)[name = string("Minv_141_cast_fp16")];
            bool Np_107_transpose_x_0 = const()[name = string("Np_107_transpose_x_0"), val = bool(false)];
            bool Np_107_transpose_y_0 = const()[name = string("Np_107_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_107_cast_fp16 = matmul(transpose_x = Np_107_transpose_x_0, transpose_y = Np_107_transpose_y_0, x = Np_105_cast_fp16, y = Npa_71_cast_fp16)[name = string("Np_107_cast_fp16")];
            bool var_3937_transpose_x_0 = const()[name = string("op_3937_transpose_x_0"), val = bool(false)];
            bool var_3937_transpose_y_0 = const()[name = string("op_3937_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_3937_cast_fp16 = matmul(transpose_x = var_3937_transpose_x_0, transpose_y = var_3937_transpose_y_0, x = Minv_141_cast_fp16, y = Np_107_cast_fp16)[name = string("op_3937_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_143_cast_fp16 = add(x = Minv_141_cast_fp16, y = var_3937_cast_fp16)[name = string("Minv_143_cast_fp16")];
            bool u_35_transpose_x_0 = const()[name = string("u_35_transpose_x_0"), val = bool(false)];
            bool u_35_transpose_y_0 = const()[name = string("u_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_35_cast_fp16 = matmul(transpose_x = u_35_transpose_x_0, transpose_y = u_35_transpose_y_0, x = Minv_143_cast_fp16, y = vb_41_cast_fp16)[name = string("u_35_cast_fp16")];
            bool var_3944_transpose_x_1 = const()[name = string("op_3944_transpose_x_1"), val = bool(false)];
            bool var_3944_transpose_y_1 = const()[name = string("op_3944_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_3944_cast_fp16 = matmul(transpose_x = var_3944_transpose_x_1, transpose_y = var_3944_transpose_y_1, x = q_59_cast_fp16, y = k_59_cast_fp16)[name = string("op_3944_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_35_cast_fp16 = mul(x = var_3944_cast_fp16, y = D_35_cast_fp16)[name = string("A_35_cast_fp16")];
            tensor<fp16, [16]> var_3946_bias_0_to_fp16 = const()[name = string("op_3946_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_3946_cast_fp16 = linear(bias = var_3946_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_41_cast_fp16)[name = string("op_3946_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_29_cast_fp16 = exp(x = var_3946_cast_fp16)[name = string("cum_29_cast_fp16")];
            tensor<int32, [1]> var_3949_axes_0 = const()[name = string("op_3949_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_3949_cast_fp16 = expand_dims(axes = var_3949_axes_0, x = cum_29_cast_fp16)[name = string("op_3949_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3950_cast_fp16 = mul(x = kb_35_cast_fp16, y = var_3949_cast_fp16)[name = string("op_3950_cast_fp16")];
            bool w_35_transpose_x_0 = const()[name = string("w_35_transpose_x_0"), val = bool(false)];
            bool w_35_transpose_y_0 = const()[name = string("w_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_35_cast_fp16 = matmul(transpose_x = w_35_transpose_x_0, transpose_y = w_35_transpose_y_0, x = Minv_143_cast_fp16, y = var_3950_cast_fp16)[name = string("w_35_cast_fp16")];
            bool var_3952_transpose_x_0 = const()[name = string("op_3952_transpose_x_0"), val = bool(false)];
            bool var_3952_transpose_y_0 = const()[name = string("op_3952_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3952_cast_fp16 = matmul(transpose_x = var_3952_transpose_x_0, transpose_y = var_3952_transpose_y_0, x = w_35_cast_fp16, y = S_29_cast_fp16)[name = string("op_3952_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_29_cast_fp16 = sub(x = u_35_cast_fp16, y = var_3952_cast_fp16)[name = string("vnew_29_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_3957_cast_fp16 = mul(x = q_59_cast_fp16, y = var_3949_cast_fp16)[name = string("op_3957_cast_fp16")];
            bool var_3958_transpose_x_0 = const()[name = string("op_3958_transpose_x_0"), val = bool(false)];
            bool var_3958_transpose_y_0 = const()[name = string("op_3958_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3958_cast_fp16 = matmul(transpose_x = var_3958_transpose_x_0, transpose_y = var_3958_transpose_y_0, x = var_3957_cast_fp16, y = S_29_cast_fp16)[name = string("op_3958_cast_fp16")];
            bool var_3959_transpose_x_0 = const()[name = string("op_3959_transpose_x_0"), val = bool(false)];
            bool var_3959_transpose_y_0 = const()[name = string("op_3959_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_3959_cast_fp16 = matmul(transpose_x = var_3959_transpose_x_0, transpose_y = var_3959_transpose_y_0, x = A_35_cast_fp16, y = vnew_29_cast_fp16)[name = string("op_3959_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_5_cast_fp16 = add(x = var_3958_cast_fp16, y = var_3959_cast_fp16)[name = string("oc_5_cast_fp16")];
            int32 var_3963 = const()[name = string("op_3963"), val = int32(2)];
            bool o_13_interleave_0 = const()[name = string("o_13_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 96, 128]> o_13_cast_fp16 = concat(axis = var_3963, interleave = o_13_interleave_0, values = (var_3198_cast_fp16, var_3345_cast_fp16, var_3499_cast_fp16, var_3653_cast_fp16, var_3807_cast_fp16, oc_5_cast_fp16))[name = string("o_13_cast_fp16")];
            tensor<int32, [4]> var_3969 = const()[name = string("op_3969"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3975 = const()[name = string("op_3975"), val = tensor<int32, [4]>([3, 96, 4, 128])];
            tensor<fp16, [3, 96, 4, 128]> input_39_cast_fp16 = reshape(shape = var_3975, x = z_5_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> o_15_cast_fp16 = transpose(perm = var_3969, x = o_13_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [3, 96, 4, 128]> var_3977_cast_fp16 = mul(x = o_15_cast_fp16, y = o_15_cast_fp16)[name = string("op_3977_cast_fp16")];
            tensor<int32, [1]> var_3982_axes_0 = const()[name = string("op_3982_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3982_keep_dims_0 = const()[name = string("op_3982_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 4, 1]> var_3982_cast_fp16 = reduce_mean(axes = var_3982_axes_0, keep_dims = var_3982_keep_dims_0, x = var_3977_cast_fp16)[name = string("op_3982_cast_fp16")];
            fp16 var_3984_to_fp16 = const()[name = string("op_3984_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 96, 4, 1]> var_3985_cast_fp16 = add(x = var_3982_cast_fp16, y = var_3984_to_fp16)[name = string("op_3985_cast_fp16")];
            fp32 var_3986_epsilon_0 = const()[name = string("op_3986_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 4, 1]> var_3986_cast_fp16 = rsqrt(epsilon = var_3986_epsilon_0, x = var_3985_cast_fp16)[name = string("op_3986_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> var_3987_cast_fp16 = mul(x = o_15_cast_fp16, y = var_3986_cast_fp16)[name = string("op_3987_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> var_3989_cast_fp16 = silu(x = input_39_cast_fp16)[name = string("op_3989_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> o_17_cast_fp16 = mul(x = var_3987_cast_fp16, y = var_3989_cast_fp16)[name = string("o_17_cast_fp16")];
            tensor<int32, [3]> var_3994 = const()[name = string("op_3994"), val = tensor<int32, [3]>([3, 96, 512])];
            tensor<fp16, [3, 96, 512]> input_41_cast_fp16 = reshape(shape = var_3994, x = o_17_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [640, 512]> layers_2_out_weight_to_fp16 = const()[name = string("layers_2_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47457728)))];
            tensor<fp16, [3, 96, 640]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_out_weight_to_fp16, x = input_41_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [3, 96, 640]> x_21_cast_fp16 = add(x = x_17_cast_fp16, y = linear_9_cast_fp16)[name = string("x_21_cast_fp16")];
            fp16 var_4005_promoted_to_fp16 = const()[name = string("op_4005_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> var_4009_cast_fp16 = pow(x = x_21_cast_fp16, y = var_4005_promoted_to_fp16)[name = string("op_4009_cast_fp16")];
            tensor<int32, [1]> var_4011_axes_0 = const()[name = string("op_4011_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4011_keep_dims_0 = const()[name = string("op_4011_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 1]> var_4011_cast_fp16 = reduce_mean(axes = var_4011_axes_0, keep_dims = var_4011_keep_dims_0, x = var_4009_cast_fp16)[name = string("op_4011_cast_fp16")];
            fp16 var_4012_to_fp16 = const()[name = string("op_4012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 96, 1]> var_4013_cast_fp16 = add(x = var_4011_cast_fp16, y = var_4012_to_fp16)[name = string("op_4013_cast_fp16")];
            fp32 var_4014_epsilon_0 = const()[name = string("op_4014_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 1]> var_4014_cast_fp16 = rsqrt(epsilon = var_4014_epsilon_0, x = var_4013_cast_fp16)[name = string("op_4014_cast_fp16")];
            tensor<fp16, [3, 96, 640]> var_4015_cast_fp16 = mul(x = x_21_cast_fp16, y = var_4014_cast_fp16)[name = string("op_4015_cast_fp16")];
            tensor<fp16, [640]> layers_2_n2_w1_to_fp16 = const()[name = string("layers_2_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48113152)))];
            tensor<fp16, [3, 96, 640]> input_43_cast_fp16 = mul(x = var_4015_cast_fp16, y = layers_2_n2_w1_to_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_2_gate_up_weight_to_fp16 = const()[name = string("layers_2_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48114496)))];
            tensor<fp16, [3, 96, 5120]> linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_2_gate_up_weight_to_fp16, x = input_43_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [3]> input_45_begin_0 = const()[name = string("input_45_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_45_end_0 = const()[name = string("input_45_end_0"), val = tensor<int32, [3]>([3, 96, 2560])];
            tensor<bool, [3]> input_45_end_mask_0 = const()[name = string("input_45_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 2560]> input_45_cast_fp16 = slice_by_index(begin = input_45_begin_0, end = input_45_end_0, end_mask = input_45_end_mask_0, x = linear_10_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [3, 96, 2560]> var_4025_cast_fp16 = silu(x = input_45_cast_fp16)[name = string("op_4025_cast_fp16")];
            tensor<int32, [3]> var_4030_begin_0 = const()[name = string("op_4030_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_4030_end_0 = const()[name = string("op_4030_end_0"), val = tensor<int32, [3]>([3, 96, 5120])];
            tensor<bool, [3]> var_4030_end_mask_0 = const()[name = string("op_4030_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 2560]> var_4030_cast_fp16 = slice_by_index(begin = var_4030_begin_0, end = var_4030_end_0, end_mask = var_4030_end_mask_0, x = linear_10_cast_fp16)[name = string("op_4030_cast_fp16")];
            tensor<fp16, [3, 96, 2560]> input_47_cast_fp16 = mul(x = var_4025_cast_fp16, y = var_4030_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_2_down_weight_to_fp16 = const()[name = string("layers_2_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54668160)))];
            tensor<fp16, [3, 96, 640]> linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_down_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [3, 96, 640]> x_25_cast_fp16 = add(x = x_21_cast_fp16, y = linear_11_cast_fp16)[name = string("x_25_cast_fp16")];
            fp16 var_4043_to_fp16 = const()[name = string("op_4043_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 96, 640]> x_27_cast_fp16 = mul(x = x_25_cast_fp16, y = var_4043_to_fp16)[name = string("x_27_cast_fp16")];
            fp16 var_4041_promoted_to_fp16 = const()[name = string("op_4041_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> var_4045_cast_fp16 = pow(x = x_27_cast_fp16, y = var_4041_promoted_to_fp16)[name = string("op_4045_cast_fp16")];
            tensor<int32, [1]> var_4047_axes_0 = const()[name = string("op_4047_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4047_keep_dims_0 = const()[name = string("op_4047_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 1]> var_4047_cast_fp16 = reduce_mean(axes = var_4047_axes_0, keep_dims = var_4047_keep_dims_0, x = var_4045_cast_fp16)[name = string("op_4047_cast_fp16")];
            fp16 var_4048_to_fp16 = const()[name = string("op_4048_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 96, 1]> var_4049_cast_fp16 = add(x = var_4047_cast_fp16, y = var_4048_to_fp16)[name = string("op_4049_cast_fp16")];
            fp32 var_4050_epsilon_0 = const()[name = string("op_4050_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 1]> var_4050_cast_fp16 = rsqrt(epsilon = var_4050_epsilon_0, x = var_4049_cast_fp16)[name = string("op_4050_cast_fp16")];
            tensor<fp16, [3, 96, 640]> var_4051_cast_fp16 = mul(x = x_27_cast_fp16, y = var_4050_cast_fp16)[name = string("op_4051_cast_fp16")];
            tensor<fp16, [640]> layers_3_n1_w1_to_fp16 = const()[name = string("layers_3_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57945024)))];
            tensor<fp16, [3, 96, 640]> input_49_cast_fp16 = mul(x = var_4051_cast_fp16, y = layers_3_n1_w1_to_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1280, 640]> layers_3_qkv_weight_to_fp16 = const()[name = string("layers_3_qkv_weight_to_fp16"), val = tensor<fp16, [1280, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57946368)))];
            tensor<fp16, [1280]> linear_12_bias_0_to_fp16 = const()[name = string("linear_12_bias_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59584832)))];
            tensor<fp16, [3, 96, 1280]> linear_12_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = layers_3_qkv_weight_to_fp16, x = input_49_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [3]> var_4060_begin_0 = const()[name = string("op_4060_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_4060_end_0 = const()[name = string("op_4060_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> var_4060_end_mask_0 = const()[name = string("op_4060_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 1024]> var_4060_cast_fp16 = slice_by_index(begin = var_4060_begin_0, end = var_4060_end_0, end_mask = var_4060_end_mask_0, x = linear_12_cast_fp16)[name = string("op_4060_cast_fp16")];
            tensor<int32, [4]> var_4065 = const()[name = string("op_4065"), val = tensor<int32, [4]>([3, 96, 8, 128])];
            tensor<fp16, [3, 96, 8, 128]> qg_1_cast_fp16 = reshape(shape = var_4065, x = var_4060_cast_fp16)[name = string("qg_1_cast_fp16")];
            tensor<int32, [4]> x_29_begin_0 = const()[name = string("x_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_29_end_0 = const()[name = string("x_29_end_0"), val = tensor<int32, [4]>([3, 96, 8, 64])];
            tensor<bool, [4]> x_29_end_mask_0 = const()[name = string("x_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 96, 8, 64]> x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, x = qg_1_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [4]> var_4076_begin_0 = const()[name = string("op_4076_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> var_4076_end_0 = const()[name = string("op_4076_end_0"), val = tensor<int32, [4]>([3, 96, 8, 128])];
            tensor<bool, [4]> var_4076_end_mask_0 = const()[name = string("op_4076_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 96, 8, 64]> var_4076_cast_fp16 = slice_by_index(begin = var_4076_begin_0, end = var_4076_end_0, end_mask = var_4076_end_mask_0, x = qg_1_cast_fp16)[name = string("op_4076_cast_fp16")];
            tensor<int32, [3]> var_4080 = const()[name = string("op_4080"), val = tensor<int32, [3]>([3, 96, 512])];
            tensor<fp16, [3, 96, 512]> gate_1_cast_fp16 = reshape(shape = var_4080, x = var_4076_cast_fp16)[name = string("gate_1_cast_fp16")];
            tensor<int32, [3]> var_4086_begin_0 = const()[name = string("op_4086_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> var_4086_end_0 = const()[name = string("op_4086_end_0"), val = tensor<int32, [3]>([3, 96, 1152])];
            tensor<bool, [3]> var_4086_end_mask_0 = const()[name = string("op_4086_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 128]> var_4086_cast_fp16 = slice_by_index(begin = var_4086_begin_0, end = var_4086_end_0, end_mask = var_4086_end_mask_0, x = linear_12_cast_fp16)[name = string("op_4086_cast_fp16")];
            tensor<int32, [4]> var_4091 = const()[name = string("op_4091"), val = tensor<int32, [4]>([3, 96, 2, 64])];
            tensor<fp16, [3, 96, 2, 64]> x_35_cast_fp16 = reshape(shape = var_4091, x = var_4086_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<int32, [3]> var_4097_begin_0 = const()[name = string("op_4097_begin_0"), val = tensor<int32, [3]>([0, 0, 1152])];
            tensor<int32, [3]> var_4097_end_0 = const()[name = string("op_4097_end_0"), val = tensor<int32, [3]>([3, 96, 1280])];
            tensor<bool, [3]> var_4097_end_mask_0 = const()[name = string("op_4097_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 128]> var_4097_cast_fp16 = slice_by_index(begin = var_4097_begin_0, end = var_4097_end_0, end_mask = var_4097_end_mask_0, x = linear_12_cast_fp16)[name = string("op_4097_cast_fp16")];
            tensor<int32, [4]> var_4102 = const()[name = string("op_4102"), val = tensor<int32, [4]>([3, 96, 2, 64])];
            tensor<fp16, [3, 96, 2, 64]> var_4103_cast_fp16 = reshape(shape = var_4102, x = var_4097_cast_fp16)[name = string("op_4103_cast_fp16")];
            tensor<int32, [4]> var_4108 = const()[name = string("op_4108"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4114_promoted_to_fp16 = const()[name = string("op_4114_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 8, 64]> var_4118_cast_fp16 = pow(x = x_29_cast_fp16, y = var_4114_promoted_to_fp16)[name = string("op_4118_cast_fp16")];
            tensor<int32, [1]> var_4120_axes_0 = const()[name = string("op_4120_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4120_keep_dims_0 = const()[name = string("op_4120_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 8, 1]> var_4120_cast_fp16 = reduce_mean(axes = var_4120_axes_0, keep_dims = var_4120_keep_dims_0, x = var_4118_cast_fp16)[name = string("op_4120_cast_fp16")];
            fp16 var_4121_to_fp16 = const()[name = string("op_4121_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 96, 8, 1]> var_4122_cast_fp16 = add(x = var_4120_cast_fp16, y = var_4121_to_fp16)[name = string("op_4122_cast_fp16")];
            fp32 var_4123_epsilon_0 = const()[name = string("op_4123_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 8, 1]> var_4123_cast_fp16 = rsqrt(epsilon = var_4123_epsilon_0, x = var_4122_cast_fp16)[name = string("op_4123_cast_fp16")];
            tensor<fp16, [3, 96, 8, 64]> var_4124_cast_fp16 = mul(x = x_29_cast_fp16, y = var_4123_cast_fp16)[name = string("op_4124_cast_fp16")];
            tensor<fp16, [64]> layers_3_qn_w1_to_fp16 = const()[name = string("layers_3_qn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59587456)))];
            tensor<fp16, [3, 96, 8, 64]> var_4125_cast_fp16 = mul(x = var_4124_cast_fp16, y = layers_3_qn_w1_to_fp16)[name = string("op_4125_cast_fp16")];
            tensor<int32, [4]> var_4130 = const()[name = string("op_4130"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_1_begin_0 = const()[name = string("r_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_1_end_0 = const()[name = string("r_1_end_0"), val = tensor<int32, [4]>([3, 8, 96, 16])];
            tensor<bool, [4]> r_1_end_mask_0 = const()[name = string("r_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 8, 96, 64]> x_33_cast_fp16 = transpose(perm = var_4130, x = var_4125_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [3, 8, 96, 16]> r_1_cast_fp16 = slice_by_index(begin = r_1_begin_0, end = r_1_end_0, end_mask = r_1_end_mask_0, x = x_33_cast_fp16)[name = string("r_1_cast_fp16")];
            tensor<int32, [4]> p_1_begin_0 = const()[name = string("p_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_1_end_0 = const()[name = string("p_1_end_0"), val = tensor<int32, [4]>([3, 8, 96, 64])];
            tensor<bool, [4]> p_1_end_mask_0 = const()[name = string("p_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 8, 96, 48]> p_1_cast_fp16 = slice_by_index(begin = p_1_begin_0, end = p_1_end_0, end_mask = p_1_end_mask_0, x = x_33_cast_fp16)[name = string("p_1_cast_fp16")];
            tensor<int32, [4]> var_4146_begin_0 = const()[name = string("op_4146_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_4146_end_0 = const()[name = string("op_4146_end_0"), val = tensor<int32, [4]>([3, 8, 96, 16])];
            tensor<bool, [4]> var_4146_end_mask_0 = const()[name = string("op_4146_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 8, 96, 8]> var_4146_cast_fp16 = slice_by_index(begin = var_4146_begin_0, end = var_4146_end_0, end_mask = var_4146_end_mask_0, x = r_1_cast_fp16)[name = string("op_4146_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 8, 96, 8]> var_4147_cast_fp16 = mul(x = var_4146_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_4147_cast_fp16")];
            tensor<int32, [4]> var_4152_begin_0 = const()[name = string("op_4152_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4152_end_0 = const()[name = string("op_4152_end_0"), val = tensor<int32, [4]>([3, 8, 96, 8])];
            tensor<bool, [4]> var_4152_end_mask_0 = const()[name = string("op_4152_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 8, 96, 8]> var_4152_cast_fp16 = slice_by_index(begin = var_4152_begin_0, end = var_4152_end_0, end_mask = var_4152_end_mask_0, x = r_1_cast_fp16)[name = string("op_4152_cast_fp16")];
            int32 var_4154 = const()[name = string("op_4154"), val = int32(-1)];
            bool rh_1_interleave_0 = const()[name = string("rh_1_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 8, 96, 16]> rh_1_cast_fp16 = concat(axis = var_4154, interleave = rh_1_interleave_0, values = (var_4147_cast_fp16, var_4152_cast_fp16))[name = string("rh_1_cast_fp16")];
            tensor<fp16, [96, 16]> cos_to_fp16 = const()[name = string("cos_to_fp16"), val = tensor<fp16, [96, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119675776)))];
            tensor<fp16, [3, 8, 96, 16]> var_4156_cast_fp16 = mul(x = r_1_cast_fp16, y = cos_to_fp16)[name = string("op_4156_cast_fp16")];
            tensor<fp16, [96, 16]> sin_to_fp16 = const()[name = string("sin_to_fp16"), val = tensor<fp16, [96, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119678912)))];
            tensor<fp16, [3, 8, 96, 16]> var_4157_cast_fp16 = mul(x = rh_1_cast_fp16, y = sin_to_fp16)[name = string("op_4157_cast_fp16")];
            tensor<fp16, [3, 8, 96, 16]> var_4159_cast_fp16 = add(x = var_4156_cast_fp16, y = var_4157_cast_fp16)[name = string("op_4159_cast_fp16")];
            int32 var_4161 = const()[name = string("op_4161"), val = int32(-1)];
            bool q_61_interleave_0 = const()[name = string("q_61_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 8, 96, 64]> q_61_cast_fp16 = concat(axis = var_4161, interleave = q_61_interleave_0, values = (var_4159_cast_fp16, p_1_cast_fp16))[name = string("q_61_cast_fp16")];
            fp16 var_4167_promoted_to_fp16 = const()[name = string("op_4167_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 2, 64]> var_4171_cast_fp16 = pow(x = x_35_cast_fp16, y = var_4167_promoted_to_fp16)[name = string("op_4171_cast_fp16")];
            tensor<int32, [1]> var_4173_axes_0 = const()[name = string("op_4173_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4173_keep_dims_0 = const()[name = string("op_4173_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 2, 1]> var_4173_cast_fp16 = reduce_mean(axes = var_4173_axes_0, keep_dims = var_4173_keep_dims_0, x = var_4171_cast_fp16)[name = string("op_4173_cast_fp16")];
            fp16 var_4174_to_fp16 = const()[name = string("op_4174_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 96, 2, 1]> var_4175_cast_fp16 = add(x = var_4173_cast_fp16, y = var_4174_to_fp16)[name = string("op_4175_cast_fp16")];
            fp32 var_4176_epsilon_0 = const()[name = string("op_4176_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 2, 1]> var_4176_cast_fp16 = rsqrt(epsilon = var_4176_epsilon_0, x = var_4175_cast_fp16)[name = string("op_4176_cast_fp16")];
            tensor<fp16, [3, 96, 2, 64]> var_4177_cast_fp16 = mul(x = x_35_cast_fp16, y = var_4176_cast_fp16)[name = string("op_4177_cast_fp16")];
            tensor<fp16, [64]> layers_3_kn_w1_to_fp16 = const()[name = string("layers_3_kn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59588800)))];
            tensor<fp16, [3, 96, 2, 64]> var_4178_cast_fp16 = mul(x = var_4177_cast_fp16, y = layers_3_kn_w1_to_fp16)[name = string("op_4178_cast_fp16")];
            tensor<int32, [4]> var_4183 = const()[name = string("op_4183"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_3_begin_0 = const()[name = string("r_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_3_end_0 = const()[name = string("r_3_end_0"), val = tensor<int32, [4]>([3, 2, 96, 16])];
            tensor<bool, [4]> r_3_end_mask_0 = const()[name = string("r_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 2, 96, 64]> x_39_cast_fp16 = transpose(perm = var_4183, x = var_4178_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [3, 2, 96, 16]> r_3_cast_fp16 = slice_by_index(begin = r_3_begin_0, end = r_3_end_0, end_mask = r_3_end_mask_0, x = x_39_cast_fp16)[name = string("r_3_cast_fp16")];
            tensor<int32, [4]> p_3_begin_0 = const()[name = string("p_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_3_end_0 = const()[name = string("p_3_end_0"), val = tensor<int32, [4]>([3, 2, 96, 64])];
            tensor<bool, [4]> p_3_end_mask_0 = const()[name = string("p_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 2, 96, 48]> p_3_cast_fp16 = slice_by_index(begin = p_3_begin_0, end = p_3_end_0, end_mask = p_3_end_mask_0, x = x_39_cast_fp16)[name = string("p_3_cast_fp16")];
            tensor<int32, [4]> var_4199_begin_0 = const()[name = string("op_4199_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_4199_end_0 = const()[name = string("op_4199_end_0"), val = tensor<int32, [4]>([3, 2, 96, 16])];
            tensor<bool, [4]> var_4199_end_mask_0 = const()[name = string("op_4199_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 2, 96, 8]> var_4199_cast_fp16 = slice_by_index(begin = var_4199_begin_0, end = var_4199_end_0, end_mask = var_4199_end_mask_0, x = r_3_cast_fp16)[name = string("op_4199_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 2, 96, 8]> var_4200_cast_fp16 = mul(x = var_4199_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_4200_cast_fp16")];
            tensor<int32, [4]> var_4205_begin_0 = const()[name = string("op_4205_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4205_end_0 = const()[name = string("op_4205_end_0"), val = tensor<int32, [4]>([3, 2, 96, 8])];
            tensor<bool, [4]> var_4205_end_mask_0 = const()[name = string("op_4205_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 2, 96, 8]> var_4205_cast_fp16 = slice_by_index(begin = var_4205_begin_0, end = var_4205_end_0, end_mask = var_4205_end_mask_0, x = r_3_cast_fp16)[name = string("op_4205_cast_fp16")];
            int32 var_4207 = const()[name = string("op_4207"), val = int32(-1)];
            bool rh_3_interleave_0 = const()[name = string("rh_3_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 2, 96, 16]> rh_3_cast_fp16 = concat(axis = var_4207, interleave = rh_3_interleave_0, values = (var_4200_cast_fp16, var_4205_cast_fp16))[name = string("rh_3_cast_fp16")];
            tensor<fp16, [3, 2, 96, 16]> var_4209_cast_fp16 = mul(x = r_3_cast_fp16, y = cos_to_fp16)[name = string("op_4209_cast_fp16")];
            tensor<fp16, [3, 2, 96, 16]> var_4210_cast_fp16 = mul(x = rh_3_cast_fp16, y = sin_to_fp16)[name = string("op_4210_cast_fp16")];
            tensor<fp16, [3, 2, 96, 16]> var_4212_cast_fp16 = add(x = var_4209_cast_fp16, y = var_4210_cast_fp16)[name = string("op_4212_cast_fp16")];
            int32 var_4214 = const()[name = string("op_4214"), val = int32(-1)];
            bool k_61_interleave_0 = const()[name = string("k_61_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 2, 96, 64]> k_61_cast_fp16 = concat(axis = var_4214, interleave = k_61_interleave_0, values = (var_4212_cast_fp16, p_3_cast_fp16))[name = string("k_61_cast_fp16")];
            tensor<int32, [1]> var_4227_axes_0 = const()[name = string("op_4227_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 1, 96, 64]> var_4227_cast_fp16 = expand_dims(axes = var_4227_axes_0, x = k_61_cast_fp16)[name = string("op_4227_cast_fp16")];
            tensor<int32, [5]> var_4235_reps_0 = const()[name = string("op_4235_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [3, 2, 4, 96, 64]> var_4235_cast_fp16 = tile(reps = var_4235_reps_0, x = var_4227_cast_fp16)[name = string("op_4235_cast_fp16")];
            tensor<int32, [4]> var_4240 = const()[name = string("op_4240"), val = tensor<int32, [4]>([3, 8, 96, 64])];
            tensor<fp16, [3, 8, 96, 64]> k_63_cast_fp16 = reshape(shape = var_4240, x = var_4235_cast_fp16)[name = string("k_63_cast_fp16")];
            tensor<int32, [1]> var_4253_axes_0 = const()[name = string("op_4253_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 96, 64]> v_7_cast_fp16 = transpose(perm = var_4108, x = var_4103_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [3, 2, 1, 96, 64]> var_4253_cast_fp16 = expand_dims(axes = var_4253_axes_0, x = v_7_cast_fp16)[name = string("op_4253_cast_fp16")];
            tensor<int32, [5]> var_4261_reps_0 = const()[name = string("op_4261_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [3, 2, 4, 96, 64]> var_4261_cast_fp16 = tile(reps = var_4261_reps_0, x = var_4253_cast_fp16)[name = string("op_4261_cast_fp16")];
            tensor<int32, [4]> var_4266 = const()[name = string("op_4266"), val = tensor<int32, [4]>([3, 8, 96, 64])];
            tensor<fp16, [3, 8, 96, 64]> v_9_cast_fp16 = reshape(shape = var_4266, x = var_4261_cast_fp16)[name = string("v_9_cast_fp16")];
            bool var_4271_transpose_x_1 = const()[name = string("op_4271_transpose_x_1"), val = bool(false)];
            bool var_4271_transpose_y_1 = const()[name = string("op_4271_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 8, 96, 96]> var_4271_cast_fp16 = matmul(transpose_x = var_4271_transpose_x_1, transpose_y = var_4271_transpose_y_1, x = q_61_cast_fp16, y = k_63_cast_fp16)[name = string("op_4271_cast_fp16")];
            fp16 var_4272_to_fp16 = const()[name = string("op_4272_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [3, 8, 96, 96]> var_4273_cast_fp16 = mul(x = var_4271_cast_fp16, y = var_4272_to_fp16)[name = string("op_4273_cast_fp16")];
            tensor<fp16, [96, 96]> causal_add_to_fp16 = const()[name = string("causal_add_to_fp16"), val = tensor<fp16, [96, 96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119682048)))];
            tensor<fp16, [3, 8, 96, 96]> s_1_cast_fp16 = add(x = var_4273_cast_fp16, y = causal_add_to_fp16)[name = string("s_1_cast_fp16")];
            int32 var_4276 = const()[name = string("op_4276"), val = int32(-1)];
            tensor<fp16, [3, 8, 96, 96]> var_4278_cast_fp16 = softmax(axis = var_4276, x = s_1_cast_fp16)[name = string("op_4278_cast_fp16")];
            bool o_19_transpose_x_0 = const()[name = string("o_19_transpose_x_0"), val = bool(false)];
            bool o_19_transpose_y_0 = const()[name = string("o_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 8, 96, 64]> o_19_cast_fp16 = matmul(transpose_x = o_19_transpose_x_0, transpose_y = o_19_transpose_y_0, x = var_4278_cast_fp16, y = v_9_cast_fp16)[name = string("o_19_cast_fp16")];
            tensor<int32, [4]> var_4284 = const()[name = string("op_4284"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4289 = const()[name = string("op_4289"), val = tensor<int32, [3]>([3, 96, 512])];
            tensor<fp16, [3, 96, 8, 64]> var_4285_cast_fp16 = transpose(perm = var_4284, x = o_19_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [3, 96, 512]> var_4290_cast_fp16 = reshape(shape = var_4289, x = var_4285_cast_fp16)[name = string("op_4290_cast_fp16")];
            tensor<fp16, [3, 96, 512]> var_4291_cast_fp16 = sigmoid(x = gate_1_cast_fp16)[name = string("op_4291_cast_fp16")];
            tensor<fp16, [3, 96, 512]> input_51_cast_fp16 = mul(x = var_4290_cast_fp16, y = var_4291_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [640, 512]> layers_3_o_weight_to_fp16 = const()[name = string("layers_3_o_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59589568)))];
            tensor<fp16, [3, 96, 640]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_o_weight_to_fp16, x = input_51_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [3, 96, 640]> x_41_cast_fp16 = add(x = x_25_cast_fp16, y = linear_13_cast_fp16)[name = string("x_41_cast_fp16")];
            fp16 var_4304_to_fp16 = const()[name = string("op_4304_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 96, 640]> x_43_cast_fp16 = mul(x = x_41_cast_fp16, y = var_4304_to_fp16)[name = string("x_43_cast_fp16")];
            fp16 var_4302_promoted_to_fp16 = const()[name = string("op_4302_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> var_4306_cast_fp16 = pow(x = x_43_cast_fp16, y = var_4302_promoted_to_fp16)[name = string("op_4306_cast_fp16")];
            tensor<int32, [1]> var_4308_axes_0 = const()[name = string("op_4308_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4308_keep_dims_0 = const()[name = string("op_4308_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 1]> var_4308_cast_fp16 = reduce_mean(axes = var_4308_axes_0, keep_dims = var_4308_keep_dims_0, x = var_4306_cast_fp16)[name = string("op_4308_cast_fp16")];
            fp16 var_4309_to_fp16 = const()[name = string("op_4309_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 96, 1]> var_4310_cast_fp16 = add(x = var_4308_cast_fp16, y = var_4309_to_fp16)[name = string("op_4310_cast_fp16")];
            fp32 var_4311_epsilon_0 = const()[name = string("op_4311_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 1]> var_4311_cast_fp16 = rsqrt(epsilon = var_4311_epsilon_0, x = var_4310_cast_fp16)[name = string("op_4311_cast_fp16")];
            tensor<fp16, [3, 96, 640]> var_4312_cast_fp16 = mul(x = x_43_cast_fp16, y = var_4311_cast_fp16)[name = string("op_4312_cast_fp16")];
            tensor<fp16, [640]> layers_3_n2_w1_to_fp16 = const()[name = string("layers_3_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60244992)))];
            tensor<fp16, [3, 96, 640]> input_53_cast_fp16 = mul(x = var_4312_cast_fp16, y = layers_3_n2_w1_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_3_gate_up_weight_to_fp16 = const()[name = string("layers_3_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60246336)))];
            tensor<fp16, [3, 96, 5120]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_3_gate_up_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [3]> input_55_begin_0 = const()[name = string("input_55_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_55_end_0 = const()[name = string("input_55_end_0"), val = tensor<int32, [3]>([3, 96, 2560])];
            tensor<bool, [3]> input_55_end_mask_0 = const()[name = string("input_55_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 2560]> input_55_cast_fp16 = slice_by_index(begin = input_55_begin_0, end = input_55_end_0, end_mask = input_55_end_mask_0, x = linear_14_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [3, 96, 2560]> var_4322_cast_fp16 = silu(x = input_55_cast_fp16)[name = string("op_4322_cast_fp16")];
            tensor<int32, [3]> var_4327_begin_0 = const()[name = string("op_4327_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_4327_end_0 = const()[name = string("op_4327_end_0"), val = tensor<int32, [3]>([3, 96, 5120])];
            tensor<bool, [3]> var_4327_end_mask_0 = const()[name = string("op_4327_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 2560]> var_4327_cast_fp16 = slice_by_index(begin = var_4327_begin_0, end = var_4327_end_0, end_mask = var_4327_end_mask_0, x = linear_14_cast_fp16)[name = string("op_4327_cast_fp16")];
            tensor<fp16, [3, 96, 2560]> input_57_cast_fp16 = mul(x = var_4322_cast_fp16, y = var_4327_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_3_down_weight_to_fp16 = const()[name = string("layers_3_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66800000)))];
            tensor<fp16, [3, 96, 640]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_down_weight_to_fp16, x = input_57_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [3, 96, 640]> x_45_cast_fp16 = add(x = x_41_cast_fp16, y = linear_15_cast_fp16)[name = string("x_45_cast_fp16")];
            fp16 var_4340_to_fp16 = const()[name = string("op_4340_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 96, 640]> x_47_cast_fp16 = mul(x = x_45_cast_fp16, y = var_4340_to_fp16)[name = string("x_47_cast_fp16")];
            fp16 var_4338_promoted_to_fp16 = const()[name = string("op_4338_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> var_4342_cast_fp16 = pow(x = x_47_cast_fp16, y = var_4338_promoted_to_fp16)[name = string("op_4342_cast_fp16")];
            tensor<int32, [1]> var_4344_axes_0 = const()[name = string("op_4344_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4344_keep_dims_0 = const()[name = string("op_4344_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 1]> var_4344_cast_fp16 = reduce_mean(axes = var_4344_axes_0, keep_dims = var_4344_keep_dims_0, x = var_4342_cast_fp16)[name = string("op_4344_cast_fp16")];
            fp16 var_4345_to_fp16 = const()[name = string("op_4345_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 96, 1]> var_4346_cast_fp16 = add(x = var_4344_cast_fp16, y = var_4345_to_fp16)[name = string("op_4346_cast_fp16")];
            fp32 var_4347_epsilon_0 = const()[name = string("op_4347_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 1]> var_4347_cast_fp16 = rsqrt(epsilon = var_4347_epsilon_0, x = var_4346_cast_fp16)[name = string("op_4347_cast_fp16")];
            tensor<fp16, [3, 96, 640]> var_4348_cast_fp16 = mul(x = x_47_cast_fp16, y = var_4347_cast_fp16)[name = string("op_4348_cast_fp16")];
            tensor<fp16, [640]> layers_4_n1_w1_to_fp16 = const()[name = string("layers_4_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70076864)))];
            tensor<fp16, [3, 96, 640]> input_59_cast_fp16 = mul(x = var_4348_cast_fp16, y = layers_4_n1_w1_to_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_4_qkvzba_weight_to_fp16 = const()[name = string("layers_4_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70078208)))];
            tensor<fp16, [3, 96, 1544]> linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_qkvzba_weight_to_fp16, x = input_59_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [3]> qkv_7_begin_0 = const()[name = string("qkv_7_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_7_end_0 = const()[name = string("qkv_7_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> qkv_7_end_mask_0 = const()[name = string("qkv_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 1024]> qkv_7_cast_fp16 = slice_by_index(begin = qkv_7_begin_0, end = qkv_7_end_0, end_mask = qkv_7_end_mask_0, x = linear_16_cast_fp16)[name = string("qkv_7_cast_fp16")];
            tensor<int32, [3]> z_7_begin_0 = const()[name = string("z_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_7_end_0 = const()[name = string("z_7_end_0"), val = tensor<int32, [3]>([3, 96, 1536])];
            tensor<bool, [3]> z_7_end_mask_0 = const()[name = string("z_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 512]> z_7_cast_fp16 = slice_by_index(begin = z_7_begin_0, end = z_7_end_0, end_mask = z_7_end_mask_0, x = linear_16_cast_fp16)[name = string("z_7_cast_fp16")];
            tensor<int32, [3]> b_7_begin_0 = const()[name = string("b_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_7_end_0 = const()[name = string("b_7_end_0"), val = tensor<int32, [3]>([3, 96, 1540])];
            tensor<bool, [3]> b_7_end_mask_0 = const()[name = string("b_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 4]> b_7_cast_fp16 = slice_by_index(begin = b_7_begin_0, end = b_7_end_0, end_mask = b_7_end_mask_0, x = linear_16_cast_fp16)[name = string("b_7_cast_fp16")];
            tensor<int32, [3]> a_7_begin_0 = const()[name = string("a_7_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_7_end_0 = const()[name = string("a_7_end_0"), val = tensor<int32, [3]>([3, 96, 1])];
            tensor<bool, [3]> a_7_end_mask_0 = const()[name = string("a_7_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 4]> a_7_cast_fp16 = slice_by_index(begin = a_7_begin_0, end = a_7_end_0, end_mask = a_7_end_mask_0, x = linear_16_cast_fp16)[name = string("a_7_cast_fp16")];
            int32 var_4387 = const()[name = string("op_4387"), val = int32(1)];
            bool xp_7_interleave_0 = const()[name = string("xp_7_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 99, 1024]> xp_7_cast_fp16 = concat(axis = var_4387, interleave = xp_7_interleave_0, values = (zpad_to_fp16, qkv_7_cast_fp16))[name = string("xp_7_cast_fp16")];
            tensor<int32, [3]> var_4398_begin_0 = const()[name = string("op_4398_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_4398_end_0 = const()[name = string("op_4398_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_4398_end_mask_0 = const()[name = string("op_4398_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 1024]> var_4398_cast_fp16 = slice_by_index(begin = var_4398_begin_0, end = var_4398_end_0, end_mask = var_4398_end_mask_0, x = xp_7_cast_fp16)[name = string("op_4398_cast_fp16")];
            tensor<fp16, [1024]> var_4406_to_fp16 = const()[name = string("op_4406_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72054592)))];
            tensor<fp16, [3, 96, 1024]> var_4407_cast_fp16 = mul(x = var_4398_cast_fp16, y = var_4406_to_fp16)[name = string("op_4407_cast_fp16")];
            tensor<int32, [3]> var_4417_begin_0 = const()[name = string("op_4417_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_4417_end_0 = const()[name = string("op_4417_end_0"), val = tensor<int32, [3]>([3, 98, 1024])];
            tensor<bool, [3]> var_4417_end_mask_0 = const()[name = string("op_4417_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_4417_cast_fp16 = slice_by_index(begin = var_4417_begin_0, end = var_4417_end_0, end_mask = var_4417_end_mask_0, x = xp_7_cast_fp16)[name = string("op_4417_cast_fp16")];
            tensor<fp16, [1024]> var_4425_to_fp16 = const()[name = string("op_4425_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72056704)))];
            tensor<fp16, [3, 96, 1024]> var_4426_cast_fp16 = mul(x = var_4417_cast_fp16, y = var_4425_to_fp16)[name = string("op_4426_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> var_4428_cast_fp16 = add(x = var_4407_cast_fp16, y = var_4426_cast_fp16)[name = string("op_4428_cast_fp16")];
            tensor<int32, [3]> var_4438_begin_0 = const()[name = string("op_4438_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_4438_end_0 = const()[name = string("op_4438_end_0"), val = tensor<int32, [3]>([3, 97, 1024])];
            tensor<bool, [3]> var_4438_end_mask_0 = const()[name = string("op_4438_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_4438_cast_fp16 = slice_by_index(begin = var_4438_begin_0, end = var_4438_end_0, end_mask = var_4438_end_mask_0, x = xp_7_cast_fp16)[name = string("op_4438_cast_fp16")];
            tensor<fp16, [1024]> var_4446_to_fp16 = const()[name = string("op_4446_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72058816)))];
            tensor<fp16, [3, 96, 1024]> var_4447_cast_fp16 = mul(x = var_4438_cast_fp16, y = var_4446_to_fp16)[name = string("op_4447_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> var_4449_cast_fp16 = add(x = var_4428_cast_fp16, y = var_4447_cast_fp16)[name = string("op_4449_cast_fp16")];
            tensor<int32, [3]> var_4459_begin_0 = const()[name = string("op_4459_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_4459_end_0 = const()[name = string("op_4459_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> var_4459_end_mask_0 = const()[name = string("op_4459_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_4459_cast_fp16 = slice_by_index(begin = var_4459_begin_0, end = var_4459_end_0, end_mask = var_4459_end_mask_0, x = xp_7_cast_fp16)[name = string("op_4459_cast_fp16")];
            tensor<fp16, [1024]> var_4467_to_fp16 = const()[name = string("op_4467_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72060928)))];
            tensor<fp16, [3, 96, 1024]> var_4468_cast_fp16 = mul(x = var_4459_cast_fp16, y = var_4467_to_fp16)[name = string("op_4468_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> input_61_cast_fp16 = add(x = var_4449_cast_fp16, y = var_4468_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> c_7_cast_fp16 = silu(x = input_61_cast_fp16)[name = string("c_7_cast_fp16")];
            tensor<int32, [3]> var_4476_begin_0 = const()[name = string("op_4476_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_4476_end_0 = const()[name = string("op_4476_end_0"), val = tensor<int32, [3]>([3, 96, 256])];
            tensor<bool, [3]> var_4476_end_mask_0 = const()[name = string("op_4476_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 256]> var_4476_cast_fp16 = slice_by_index(begin = var_4476_begin_0, end = var_4476_end_0, end_mask = var_4476_end_mask_0, x = c_7_cast_fp16)[name = string("op_4476_cast_fp16")];
            tensor<int32, [4]> var_4481 = const()[name = string("op_4481"), val = tensor<int32, [4]>([3, 96, 2, 128])];
            tensor<fp16, [3, 96, 2, 128]> q_63_cast_fp16 = reshape(shape = var_4481, x = var_4476_cast_fp16)[name = string("q_63_cast_fp16")];
            tensor<int32, [3]> var_4487_begin_0 = const()[name = string("op_4487_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_4487_end_0 = const()[name = string("op_4487_end_0"), val = tensor<int32, [3]>([3, 96, 512])];
            tensor<bool, [3]> var_4487_end_mask_0 = const()[name = string("op_4487_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 256]> var_4487_cast_fp16 = slice_by_index(begin = var_4487_begin_0, end = var_4487_end_0, end_mask = var_4487_end_mask_0, x = c_7_cast_fp16)[name = string("op_4487_cast_fp16")];
            tensor<int32, [4]> var_4492 = const()[name = string("op_4492"), val = tensor<int32, [4]>([3, 96, 2, 128])];
            tensor<fp16, [3, 96, 2, 128]> k_65_cast_fp16 = reshape(shape = var_4492, x = var_4487_cast_fp16)[name = string("k_65_cast_fp16")];
            tensor<int32, [3]> var_4498_begin_0 = const()[name = string("op_4498_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_4498_end_0 = const()[name = string("op_4498_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> var_4498_end_mask_0 = const()[name = string("op_4498_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 512]> var_4498_cast_fp16 = slice_by_index(begin = var_4498_begin_0, end = var_4498_end_0, end_mask = var_4498_end_mask_0, x = c_7_cast_fp16)[name = string("op_4498_cast_fp16")];
            tensor<int32, [4]> var_4503 = const()[name = string("op_4503"), val = tensor<int32, [4]>([3, 96, 4, 128])];
            tensor<fp16, [3, 96, 4, 128]> var_4504_cast_fp16 = reshape(shape = var_4503, x = var_4498_cast_fp16)[name = string("op_4504_cast_fp16")];
            tensor<int32, [4]> var_4509 = const()[name = string("op_4509"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4511_to_fp16 = const()[name = string("op_4511_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 96, 2, 128]> q_65_cast_fp16 = mul(x = q_63_cast_fp16, y = var_4511_to_fp16)[name = string("q_65_cast_fp16")];
            fp16 var_4513_to_fp16 = const()[name = string("op_4513_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 96, 2, 128]> k_67_cast_fp16 = mul(x = k_65_cast_fp16, y = var_4513_to_fp16)[name = string("k_67_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_4515_cast_fp16 = mul(x = q_65_cast_fp16, y = q_65_cast_fp16)[name = string("op_4515_cast_fp16")];
            tensor<int32, [1]> var_4520_axes_0 = const()[name = string("op_4520_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4520_keep_dims_0 = const()[name = string("op_4520_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 2, 1]> var_4520_cast_fp16 = reduce_mean(axes = var_4520_axes_0, keep_dims = var_4520_keep_dims_0, x = var_4515_cast_fp16)[name = string("op_4520_cast_fp16")];
            fp16 var_4521_promoted_to_fp16 = const()[name = string("op_4521_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 96, 2, 1]> var_4522_cast_fp16 = mul(x = var_4520_cast_fp16, y = var_4521_promoted_to_fp16)[name = string("op_4522_cast_fp16")];
            fp16 var_4524_to_fp16 = const()[name = string("op_4524_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 96, 2, 1]> var_4525_cast_fp16 = add(x = var_4522_cast_fp16, y = var_4524_to_fp16)[name = string("op_4525_cast_fp16")];
            fp32 var_4526_epsilon_0 = const()[name = string("op_4526_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 2, 1]> var_4526_cast_fp16 = rsqrt(epsilon = var_4526_epsilon_0, x = var_4525_cast_fp16)[name = string("op_4526_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_4527_cast_fp16 = mul(x = q_65_cast_fp16, y = var_4526_cast_fp16)[name = string("op_4527_cast_fp16")];
            fp16 var_4528_to_fp16 = const()[name = string("op_4528_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 96, 2, 128]> q_67_cast_fp16 = mul(x = var_4527_cast_fp16, y = var_4528_to_fp16)[name = string("q_67_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_4530_cast_fp16 = mul(x = k_67_cast_fp16, y = k_67_cast_fp16)[name = string("op_4530_cast_fp16")];
            tensor<int32, [1]> var_4535_axes_0 = const()[name = string("op_4535_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4535_keep_dims_0 = const()[name = string("op_4535_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 2, 1]> var_4535_cast_fp16 = reduce_mean(axes = var_4535_axes_0, keep_dims = var_4535_keep_dims_0, x = var_4530_cast_fp16)[name = string("op_4535_cast_fp16")];
            fp16 var_4536_promoted_to_fp16 = const()[name = string("op_4536_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 96, 2, 1]> var_4537_cast_fp16 = mul(x = var_4535_cast_fp16, y = var_4536_promoted_to_fp16)[name = string("op_4537_cast_fp16")];
            fp16 var_4539_to_fp16 = const()[name = string("op_4539_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 96, 2, 1]> var_4540_cast_fp16 = add(x = var_4537_cast_fp16, y = var_4539_to_fp16)[name = string("op_4540_cast_fp16")];
            fp32 var_4541_epsilon_0 = const()[name = string("op_4541_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 2, 1]> var_4541_cast_fp16 = rsqrt(epsilon = var_4541_epsilon_0, x = var_4540_cast_fp16)[name = string("op_4541_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> k_69_cast_fp16 = mul(x = k_67_cast_fp16, y = var_4541_cast_fp16)[name = string("k_69_cast_fp16")];
            tensor<int32, [4]> var_4547 = const()[name = string("op_4547"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_4560_axes_0 = const()[name = string("op_4560_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 96, 128]> var_4548_cast_fp16 = transpose(perm = var_4547, x = q_67_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [3, 2, 1, 96, 128]> var_4560_cast_fp16 = expand_dims(axes = var_4560_axes_0, x = var_4548_cast_fp16)[name = string("op_4560_cast_fp16")];
            tensor<int32, [5]> var_4568_reps_0 = const()[name = string("op_4568_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 96, 128]> var_4568_cast_fp16 = tile(reps = var_4568_reps_0, x = var_4560_cast_fp16)[name = string("op_4568_cast_fp16")];
            tensor<int32, [4]> var_4573 = const()[name = string("op_4573"), val = tensor<int32, [4]>([3, 4, 96, 128])];
            tensor<fp16, [3, 4, 96, 128]> q_69_cast_fp16 = reshape(shape = var_4573, x = var_4568_cast_fp16)[name = string("q_69_cast_fp16")];
            tensor<int32, [4]> var_4579 = const()[name = string("op_4579"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_4592_axes_0 = const()[name = string("op_4592_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 96, 128]> var_4580_cast_fp16 = transpose(perm = var_4579, x = k_69_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [3, 2, 1, 96, 128]> var_4592_cast_fp16 = expand_dims(axes = var_4592_axes_0, x = var_4580_cast_fp16)[name = string("op_4592_cast_fp16")];
            tensor<int32, [5]> var_4600_reps_0 = const()[name = string("op_4600_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 96, 128]> var_4600_cast_fp16 = tile(reps = var_4600_reps_0, x = var_4592_cast_fp16)[name = string("op_4600_cast_fp16")];
            tensor<int32, [4]> var_4605 = const()[name = string("op_4605"), val = tensor<int32, [4]>([3, 4, 96, 128])];
            tensor<fp16, [3, 4, 96, 128]> k_71_cast_fp16 = reshape(shape = var_4605, x = var_4600_cast_fp16)[name = string("k_71_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_4607_cast_fp16 = sigmoid(x = b_7_cast_fp16)[name = string("op_4607_cast_fp16")];
            tensor<int32, [3]> var_4611 = const()[name = string("op_4611"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_4_dt_bias_to_fp16 = const()[name = string("layers_4_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 96, 4]> input_63_cast_fp16 = add(x = a_7_cast_fp16, y = layers_4_dt_bias_to_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_4615_cast_fp16 = relu(x = input_63_cast_fp16)[name = string("op_4615_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_4616_cast_fp16 = abs(x = input_63_cast_fp16)[name = string("op_4616_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 96, 4]> var_4617_cast_fp16 = mul(x = var_4616_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_4617_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_4618_cast_fp16 = exp(x = var_4617_cast_fp16)[name = string("op_4618_cast_fp16")];
            fp16 var_4620_to_fp16 = const()[name = string("op_4620_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 96, 4]> var_4621_cast_fp16 = add(x = var_4618_cast_fp16, y = var_4620_to_fp16)[name = string("op_4621_cast_fp16")];
            fp32 var_4622_epsilon_0 = const()[name = string("op_4622_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 96, 4]> var_4622_cast_fp16 = log(epsilon = var_4622_epsilon_0, x = var_4621_cast_fp16)[name = string("op_4622_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_4624_cast_fp16 = add(x = var_4615_cast_fp16, y = var_4622_cast_fp16)[name = string("op_4624_cast_fp16")];
            tensor<fp16, [4]> layers_4_negA_to_fp16 = const()[name = string("layers_4_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.348p+3, -0x1.734p+2, -0x1.5a8p-1, -0x1.264p+3])];
            tensor<fp16, [3, 96, 4]> var_4625_cast_fp16 = mul(x = layers_4_negA_to_fp16, y = var_4624_cast_fp16)[name = string("op_4625_cast_fp16")];
            tensor<int32, [3]> var_4629 = const()[name = string("op_4629"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_4632_axes_0 = const()[name = string("op_4632_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 96]> beta_7_cast_fp16 = transpose(perm = var_4611, x = var_4607_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [3, 4, 96, 1]> var_4632_cast_fp16 = expand_dims(axes = var_4632_axes_0, x = beta_7_cast_fp16)[name = string("op_4632_cast_fp16")];
            fp16 var_4633_to_fp16 = const()[name = string("op_4633_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 96, 1]> var_4634_cast_fp16 = mul(x = var_4632_cast_fp16, y = var_4633_to_fp16)[name = string("op_4634_cast_fp16")];
            tensor<fp16, [3, 4, 96, 128]> v_11_cast_fp16 = transpose(perm = var_4509, x = var_4504_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [3, 4, 96, 128]> vb_43_cast_fp16 = mul(x = v_11_cast_fp16, y = var_4634_cast_fp16)[name = string("vb_43_cast_fp16")];
            tensor<int32, [4]> q_71_begin_0 = const()[name = string("q_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_71_end_0 = const()[name = string("q_71_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_71_end_mask_0 = const()[name = string("q_71_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_71_cast_fp16 = slice_by_index(begin = q_71_begin_0, end = q_71_end_0, end_mask = q_71_end_mask_0, x = q_69_cast_fp16)[name = string("q_71_cast_fp16")];
            tensor<int32, [4]> k_73_begin_0 = const()[name = string("k_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_73_end_0 = const()[name = string("k_73_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_73_end_mask_0 = const()[name = string("k_73_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, x = k_71_cast_fp16)[name = string("k_73_cast_fp16")];
            tensor<fp16, [3, 4, 96, 128]> var_4668_cast_fp16 = mul(x = k_71_cast_fp16, y = var_4632_cast_fp16)[name = string("op_4668_cast_fp16")];
            tensor<int32, [4]> kb_37_begin_0 = const()[name = string("kb_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_37_end_0 = const()[name = string("kb_37_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_37_end_mask_0 = const()[name = string("kb_37_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_37_cast_fp16 = slice_by_index(begin = kb_37_begin_0, end = kb_37_end_0, end_mask = kb_37_end_mask_0, x = var_4668_cast_fp16)[name = string("kb_37_cast_fp16")];
            tensor<int32, [4]> vb_45_begin_0 = const()[name = string("vb_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_45_end_0 = const()[name = string("vb_45_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_45_end_mask_0 = const()[name = string("vb_45_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_45_cast_fp16 = slice_by_index(begin = vb_45_begin_0, end = vb_45_end_0, end_mask = vb_45_end_mask_0, x = vb_43_cast_fp16)[name = string("vb_45_cast_fp16")];
            tensor<int32, [3]> gdec_45_begin_0 = const()[name = string("gdec_45_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_45_end_0 = const()[name = string("gdec_45_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_45_end_mask_0 = const()[name = string("gdec_45_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 96]> gdec_43_cast_fp16 = transpose(perm = var_4629, x = var_4625_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [3, 4, 16]> gdec_45_cast_fp16 = slice_by_index(begin = gdec_45_begin_0, end = gdec_45_end_0, end_mask = gdec_45_end_mask_0, x = gdec_43_cast_fp16)[name = string("gdec_45_cast_fp16")];
            tensor<int32, [2]> var_4716 = const()[name = string("op_4716"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_4717_cast_fp16 = reshape(shape = var_4716, x = gdec_45_cast_fp16)[name = string("op_4717_cast_fp16")];
            tensor<fp16, [12, 256]> var_4718_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_4717_cast_fp16)[name = string("op_4718_cast_fp16")];
            tensor<int32, [4]> var_4723 = const()[name = string("op_4723"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_37_cast_fp16 = reshape(shape = var_4723, x = var_4718_cast_fp16)[name = string("P_37_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4725_cast_fp16 = exp(x = P_37_cast_fp16)[name = string("op_4725_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_37_cast_fp16 = mul(x = var_4725_cast_fp16, y = tril_incl_to_fp16)[name = string("D_37_cast_fp16")];
            bool kk_37_transpose_x_1 = const()[name = string("kk_37_transpose_x_1"), val = bool(false)];
            bool kk_37_transpose_y_1 = const()[name = string("kk_37_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_37_cast_fp16 = matmul(transpose_x = kk_37_transpose_x_1, transpose_y = kk_37_transpose_y_1, x = kb_37_cast_fp16, y = k_73_cast_fp16)[name = string("kk_37_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_4731_cast_fp16 = mul(x = kk_37_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_4731_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4732_cast_fp16 = mul(x = var_4731_cast_fp16, y = D_37_cast_fp16)[name = string("op_4732_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_37_cast_fp16 = mul(x = var_4732_cast_fp16, y = cum_tail_to_fp16)[name = string("N_37_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4734_cast_fp16 = mul(x = D_37_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_4734_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_37_cast_fp16 = mul(x = kk_37_cast_fp16, y = var_4734_cast_fp16)[name = string("Na_37_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_145_cast_fp16 = add(x = eye_to_fp16, y = N_37_cast_fp16)[name = string("Minv_145_cast_fp16")];
            bool Np_109_transpose_x_0 = const()[name = string("Np_109_transpose_x_0"), val = bool(false)];
            bool Np_109_transpose_y_0 = const()[name = string("Np_109_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_109_cast_fp16 = matmul(transpose_x = Np_109_transpose_x_0, transpose_y = Np_109_transpose_y_0, x = N_37_cast_fp16, y = Na_37_cast_fp16)[name = string("Np_109_cast_fp16")];
            bool Npa_73_transpose_x_0 = const()[name = string("Npa_73_transpose_x_0"), val = bool(false)];
            bool Npa_73_transpose_y_0 = const()[name = string("Npa_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_73_cast_fp16 = matmul(transpose_x = Npa_73_transpose_x_0, transpose_y = Npa_73_transpose_y_0, x = Na_37_cast_fp16, y = N_37_cast_fp16)[name = string("Npa_73_cast_fp16")];
            bool var_4740_transpose_x_0 = const()[name = string("op_4740_transpose_x_0"), val = bool(false)];
            bool var_4740_transpose_y_0 = const()[name = string("op_4740_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4740_cast_fp16 = matmul(transpose_x = var_4740_transpose_x_0, transpose_y = var_4740_transpose_y_0, x = Minv_145_cast_fp16, y = Np_109_cast_fp16)[name = string("op_4740_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_147_cast_fp16 = add(x = Minv_145_cast_fp16, y = var_4740_cast_fp16)[name = string("Minv_147_cast_fp16")];
            bool Np_111_transpose_x_0 = const()[name = string("Np_111_transpose_x_0"), val = bool(false)];
            bool Np_111_transpose_y_0 = const()[name = string("Np_111_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_111_cast_fp16 = matmul(transpose_x = Np_111_transpose_x_0, transpose_y = Np_111_transpose_y_0, x = Np_109_cast_fp16, y = Npa_73_cast_fp16)[name = string("Np_111_cast_fp16")];
            bool Npa_75_transpose_x_0 = const()[name = string("Npa_75_transpose_x_0"), val = bool(false)];
            bool Npa_75_transpose_y_0 = const()[name = string("Npa_75_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_75_cast_fp16 = matmul(transpose_x = Npa_75_transpose_x_0, transpose_y = Npa_75_transpose_y_0, x = Npa_73_cast_fp16, y = Np_109_cast_fp16)[name = string("Npa_75_cast_fp16")];
            bool var_4745_transpose_x_0 = const()[name = string("op_4745_transpose_x_0"), val = bool(false)];
            bool var_4745_transpose_y_0 = const()[name = string("op_4745_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4745_cast_fp16 = matmul(transpose_x = var_4745_transpose_x_0, transpose_y = var_4745_transpose_y_0, x = Minv_147_cast_fp16, y = Np_111_cast_fp16)[name = string("op_4745_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_149_cast_fp16 = add(x = Minv_147_cast_fp16, y = var_4745_cast_fp16)[name = string("Minv_149_cast_fp16")];
            bool Np_113_transpose_x_0 = const()[name = string("Np_113_transpose_x_0"), val = bool(false)];
            bool Np_113_transpose_y_0 = const()[name = string("Np_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_113_cast_fp16 = matmul(transpose_x = Np_113_transpose_x_0, transpose_y = Np_113_transpose_y_0, x = Np_111_cast_fp16, y = Npa_75_cast_fp16)[name = string("Np_113_cast_fp16")];
            bool var_4749_transpose_x_0 = const()[name = string("op_4749_transpose_x_0"), val = bool(false)];
            bool var_4749_transpose_y_0 = const()[name = string("op_4749_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4749_cast_fp16 = matmul(transpose_x = var_4749_transpose_x_0, transpose_y = var_4749_transpose_y_0, x = Minv_149_cast_fp16, y = Np_113_cast_fp16)[name = string("op_4749_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_151_cast_fp16 = add(x = Minv_149_cast_fp16, y = var_4749_cast_fp16)[name = string("Minv_151_cast_fp16")];
            bool u_37_transpose_x_0 = const()[name = string("u_37_transpose_x_0"), val = bool(false)];
            bool u_37_transpose_y_0 = const()[name = string("u_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_37_cast_fp16 = matmul(transpose_x = u_37_transpose_x_0, transpose_y = u_37_transpose_y_0, x = Minv_151_cast_fp16, y = vb_45_cast_fp16)[name = string("u_37_cast_fp16")];
            bool var_4756_transpose_x_1 = const()[name = string("op_4756_transpose_x_1"), val = bool(false)];
            bool var_4756_transpose_y_1 = const()[name = string("op_4756_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_4756_cast_fp16 = matmul(transpose_x = var_4756_transpose_x_1, transpose_y = var_4756_transpose_y_1, x = q_71_cast_fp16, y = k_73_cast_fp16)[name = string("op_4756_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_37_cast_fp16 = mul(x = var_4756_cast_fp16, y = D_37_cast_fp16)[name = string("A_37_cast_fp16")];
            bool var_4758_transpose_x_0 = const()[name = string("op_4758_transpose_x_0"), val = bool(false)];
            bool var_4758_transpose_y_0 = const()[name = string("op_4758_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4758_cast_fp16 = matmul(transpose_x = var_4758_transpose_x_0, transpose_y = var_4758_transpose_y_0, x = A_37_cast_fp16, y = u_37_cast_fp16)[name = string("op_4758_cast_fp16")];
            tensor<fp16, [16]> var_4759_bias_0_to_fp16 = const()[name = string("op_4759_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_4759_cast_fp16 = linear(bias = var_4759_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_45_cast_fp16)[name = string("op_4759_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_31_cast_fp16 = exp(x = var_4759_cast_fp16)[name = string("tail_31_cast_fp16")];
            tensor<int32, [1]> var_4762_axes_0 = const()[name = string("op_4762_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_4762_cast_fp16 = expand_dims(axes = var_4762_axes_0, x = tail_31_cast_fp16)[name = string("op_4762_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4763_cast_fp16 = mul(x = k_73_cast_fp16, y = var_4762_cast_fp16)[name = string("op_4763_cast_fp16")];
            bool S_31_transpose_x_1 = const()[name = string("S_31_transpose_x_1"), val = bool(true)];
            bool S_31_transpose_y_1 = const()[name = string("S_31_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_31_cast_fp16 = matmul(transpose_x = S_31_transpose_x_1, transpose_y = S_31_transpose_y_1, x = var_4763_cast_fp16, y = u_37_cast_fp16)[name = string("S_31_cast_fp16")];
            tensor<int32, [4]> q_73_begin_0 = const()[name = string("q_73_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_73_end_0 = const()[name = string("q_73_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> q_73_end_mask_0 = const()[name = string("q_73_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, x = q_69_cast_fp16)[name = string("q_73_cast_fp16")];
            tensor<int32, [4]> k_75_begin_0 = const()[name = string("k_75_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_75_end_0 = const()[name = string("k_75_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> k_75_end_mask_0 = const()[name = string("k_75_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_75_cast_fp16 = slice_by_index(begin = k_75_begin_0, end = k_75_end_0, end_mask = k_75_end_mask_0, x = k_71_cast_fp16)[name = string("k_75_cast_fp16")];
            tensor<int32, [4]> kb_39_begin_0 = const()[name = string("kb_39_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_39_end_0 = const()[name = string("kb_39_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> kb_39_end_mask_0 = const()[name = string("kb_39_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_39_cast_fp16 = slice_by_index(begin = kb_39_begin_0, end = kb_39_end_0, end_mask = kb_39_end_mask_0, x = var_4668_cast_fp16)[name = string("kb_39_cast_fp16")];
            tensor<int32, [4]> vb_47_begin_0 = const()[name = string("vb_47_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_47_end_0 = const()[name = string("vb_47_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> vb_47_end_mask_0 = const()[name = string("vb_47_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_47_cast_fp16 = slice_by_index(begin = vb_47_begin_0, end = vb_47_end_0, end_mask = vb_47_end_mask_0, x = vb_43_cast_fp16)[name = string("vb_47_cast_fp16")];
            tensor<int32, [3]> gdec_47_begin_0 = const()[name = string("gdec_47_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_47_end_0 = const()[name = string("gdec_47_end_0"), val = tensor<int32, [3]>([3, 4, 32])];
            tensor<bool, [3]> gdec_47_end_mask_0 = const()[name = string("gdec_47_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_47_cast_fp16 = slice_by_index(begin = gdec_47_begin_0, end = gdec_47_end_0, end_mask = gdec_47_end_mask_0, x = gdec_43_cast_fp16)[name = string("gdec_47_cast_fp16")];
            tensor<int32, [2]> var_4848 = const()[name = string("op_4848"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_4849_cast_fp16 = reshape(shape = var_4848, x = gdec_47_cast_fp16)[name = string("op_4849_cast_fp16")];
            tensor<fp16, [12, 256]> var_4850_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_4849_cast_fp16)[name = string("op_4850_cast_fp16")];
            tensor<int32, [4]> var_4855 = const()[name = string("op_4855"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_39_cast_fp16 = reshape(shape = var_4855, x = var_4850_cast_fp16)[name = string("P_39_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4857_cast_fp16 = exp(x = P_39_cast_fp16)[name = string("op_4857_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_39_cast_fp16 = mul(x = var_4857_cast_fp16, y = tril_incl_to_fp16)[name = string("D_39_cast_fp16")];
            bool kk_39_transpose_x_1 = const()[name = string("kk_39_transpose_x_1"), val = bool(false)];
            bool kk_39_transpose_y_1 = const()[name = string("kk_39_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_39_cast_fp16 = matmul(transpose_x = kk_39_transpose_x_1, transpose_y = kk_39_transpose_y_1, x = kb_39_cast_fp16, y = k_75_cast_fp16)[name = string("kk_39_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_4863_cast_fp16 = mul(x = kk_39_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_4863_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4864_cast_fp16 = mul(x = var_4863_cast_fp16, y = D_39_cast_fp16)[name = string("op_4864_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_39_cast_fp16 = mul(x = var_4864_cast_fp16, y = cum_tail_to_fp16)[name = string("N_39_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_4866_cast_fp16 = mul(x = D_39_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_4866_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_39_cast_fp16 = mul(x = kk_39_cast_fp16, y = var_4866_cast_fp16)[name = string("Na_39_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_153_cast_fp16 = add(x = eye_to_fp16, y = N_39_cast_fp16)[name = string("Minv_153_cast_fp16")];
            bool Np_115_transpose_x_0 = const()[name = string("Np_115_transpose_x_0"), val = bool(false)];
            bool Np_115_transpose_y_0 = const()[name = string("Np_115_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_115_cast_fp16 = matmul(transpose_x = Np_115_transpose_x_0, transpose_y = Np_115_transpose_y_0, x = N_39_cast_fp16, y = Na_39_cast_fp16)[name = string("Np_115_cast_fp16")];
            bool Npa_77_transpose_x_0 = const()[name = string("Npa_77_transpose_x_0"), val = bool(false)];
            bool Npa_77_transpose_y_0 = const()[name = string("Npa_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_77_cast_fp16 = matmul(transpose_x = Npa_77_transpose_x_0, transpose_y = Npa_77_transpose_y_0, x = Na_39_cast_fp16, y = N_39_cast_fp16)[name = string("Npa_77_cast_fp16")];
            bool var_4872_transpose_x_0 = const()[name = string("op_4872_transpose_x_0"), val = bool(false)];
            bool var_4872_transpose_y_0 = const()[name = string("op_4872_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4872_cast_fp16 = matmul(transpose_x = var_4872_transpose_x_0, transpose_y = var_4872_transpose_y_0, x = Minv_153_cast_fp16, y = Np_115_cast_fp16)[name = string("op_4872_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_155_cast_fp16 = add(x = Minv_153_cast_fp16, y = var_4872_cast_fp16)[name = string("Minv_155_cast_fp16")];
            bool Np_117_transpose_x_0 = const()[name = string("Np_117_transpose_x_0"), val = bool(false)];
            bool Np_117_transpose_y_0 = const()[name = string("Np_117_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_117_cast_fp16 = matmul(transpose_x = Np_117_transpose_x_0, transpose_y = Np_117_transpose_y_0, x = Np_115_cast_fp16, y = Npa_77_cast_fp16)[name = string("Np_117_cast_fp16")];
            bool Npa_79_transpose_x_0 = const()[name = string("Npa_79_transpose_x_0"), val = bool(false)];
            bool Npa_79_transpose_y_0 = const()[name = string("Npa_79_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_79_cast_fp16 = matmul(transpose_x = Npa_79_transpose_x_0, transpose_y = Npa_79_transpose_y_0, x = Npa_77_cast_fp16, y = Np_115_cast_fp16)[name = string("Npa_79_cast_fp16")];
            bool var_4877_transpose_x_0 = const()[name = string("op_4877_transpose_x_0"), val = bool(false)];
            bool var_4877_transpose_y_0 = const()[name = string("op_4877_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4877_cast_fp16 = matmul(transpose_x = var_4877_transpose_x_0, transpose_y = var_4877_transpose_y_0, x = Minv_155_cast_fp16, y = Np_117_cast_fp16)[name = string("op_4877_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_157_cast_fp16 = add(x = Minv_155_cast_fp16, y = var_4877_cast_fp16)[name = string("Minv_157_cast_fp16")];
            bool Np_119_transpose_x_0 = const()[name = string("Np_119_transpose_x_0"), val = bool(false)];
            bool Np_119_transpose_y_0 = const()[name = string("Np_119_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_119_cast_fp16 = matmul(transpose_x = Np_119_transpose_x_0, transpose_y = Np_119_transpose_y_0, x = Np_117_cast_fp16, y = Npa_79_cast_fp16)[name = string("Np_119_cast_fp16")];
            bool var_4881_transpose_x_0 = const()[name = string("op_4881_transpose_x_0"), val = bool(false)];
            bool var_4881_transpose_y_0 = const()[name = string("op_4881_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_4881_cast_fp16 = matmul(transpose_x = var_4881_transpose_x_0, transpose_y = var_4881_transpose_y_0, x = Minv_157_cast_fp16, y = Np_119_cast_fp16)[name = string("op_4881_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_159_cast_fp16 = add(x = Minv_157_cast_fp16, y = var_4881_cast_fp16)[name = string("Minv_159_cast_fp16")];
            bool u_39_transpose_x_0 = const()[name = string("u_39_transpose_x_0"), val = bool(false)];
            bool u_39_transpose_y_0 = const()[name = string("u_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_39_cast_fp16 = matmul(transpose_x = u_39_transpose_x_0, transpose_y = u_39_transpose_y_0, x = Minv_159_cast_fp16, y = vb_47_cast_fp16)[name = string("u_39_cast_fp16")];
            bool var_4888_transpose_x_1 = const()[name = string("op_4888_transpose_x_1"), val = bool(false)];
            bool var_4888_transpose_y_1 = const()[name = string("op_4888_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_4888_cast_fp16 = matmul(transpose_x = var_4888_transpose_x_1, transpose_y = var_4888_transpose_y_1, x = q_73_cast_fp16, y = k_75_cast_fp16)[name = string("op_4888_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_39_cast_fp16 = mul(x = var_4888_cast_fp16, y = D_39_cast_fp16)[name = string("A_39_cast_fp16")];
            tensor<fp16, [16]> var_4890_bias_0_to_fp16 = const()[name = string("op_4890_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_4890_cast_fp16 = linear(bias = var_4890_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_47_cast_fp16)[name = string("op_4890_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_31_cast_fp16 = exp(x = var_4890_cast_fp16)[name = string("cum_31_cast_fp16")];
            tensor<int32, [1]> var_4893_axes_0 = const()[name = string("op_4893_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_4893_cast_fp16 = expand_dims(axes = var_4893_axes_0, x = cum_31_cast_fp16)[name = string("op_4893_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4894_cast_fp16 = mul(x = kb_39_cast_fp16, y = var_4893_cast_fp16)[name = string("op_4894_cast_fp16")];
            bool w_39_transpose_x_0 = const()[name = string("w_39_transpose_x_0"), val = bool(false)];
            bool w_39_transpose_y_0 = const()[name = string("w_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_39_cast_fp16 = matmul(transpose_x = w_39_transpose_x_0, transpose_y = w_39_transpose_y_0, x = Minv_159_cast_fp16, y = var_4894_cast_fp16)[name = string("w_39_cast_fp16")];
            bool var_4896_transpose_x_0 = const()[name = string("op_4896_transpose_x_0"), val = bool(false)];
            bool var_4896_transpose_y_0 = const()[name = string("op_4896_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4896_cast_fp16 = matmul(transpose_x = var_4896_transpose_x_0, transpose_y = var_4896_transpose_y_0, x = w_39_cast_fp16, y = S_31_cast_fp16)[name = string("op_4896_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_31_cast_fp16 = sub(x = u_39_cast_fp16, y = var_4896_cast_fp16)[name = string("vnew_31_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4901_cast_fp16 = mul(x = q_73_cast_fp16, y = var_4893_cast_fp16)[name = string("op_4901_cast_fp16")];
            bool var_4902_transpose_x_0 = const()[name = string("op_4902_transpose_x_0"), val = bool(false)];
            bool var_4902_transpose_y_0 = const()[name = string("op_4902_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4902_cast_fp16 = matmul(transpose_x = var_4902_transpose_x_0, transpose_y = var_4902_transpose_y_0, x = var_4901_cast_fp16, y = S_31_cast_fp16)[name = string("op_4902_cast_fp16")];
            bool var_4903_transpose_x_0 = const()[name = string("op_4903_transpose_x_0"), val = bool(false)];
            bool var_4903_transpose_y_0 = const()[name = string("op_4903_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_4903_cast_fp16 = matmul(transpose_x = var_4903_transpose_x_0, transpose_y = var_4903_transpose_y_0, x = A_39_cast_fp16, y = vnew_31_cast_fp16)[name = string("op_4903_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4905_cast_fp16 = add(x = var_4902_cast_fp16, y = var_4903_cast_fp16)[name = string("op_4905_cast_fp16")];
            tensor<fp16, [16]> var_4906_bias_0_to_fp16 = const()[name = string("op_4906_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_4906_cast_fp16 = linear(bias = var_4906_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_47_cast_fp16)[name = string("op_4906_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_33_cast_fp16 = exp(x = var_4906_cast_fp16)[name = string("tail_33_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_65_to_fp16 = const()[name = string("transpose_65_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_4908_bias_0_to_fp16 = const()[name = string("op_4908_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_4908_cast_fp16 = linear(bias = var_4908_bias_0_to_fp16, weight = transpose_65_to_fp16, x = gdec_47_cast_fp16)[name = string("op_4908_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_4909_cast_fp16 = exp(x = var_4908_cast_fp16)[name = string("op_4909_cast_fp16")];
            tensor<int32, [1]> last_25_axes_0 = const()[name = string("last_25_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_25_cast_fp16 = expand_dims(axes = last_25_axes_0, x = var_4909_cast_fp16)[name = string("last_25_cast_fp16")];
            tensor<int32, [1]> var_4913_axes_0 = const()[name = string("op_4913_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_4913_cast_fp16 = expand_dims(axes = var_4913_axes_0, x = tail_33_cast_fp16)[name = string("op_4913_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_4914_cast_fp16 = mul(x = k_75_cast_fp16, y = var_4913_cast_fp16)[name = string("op_4914_cast_fp16")];
            bool upd_25_transpose_x_1 = const()[name = string("upd_25_transpose_x_1"), val = bool(true)];
            bool upd_25_transpose_y_1 = const()[name = string("upd_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_25_cast_fp16 = matmul(transpose_x = upd_25_transpose_x_1, transpose_y = upd_25_transpose_y_1, x = var_4914_cast_fp16, y = vnew_31_cast_fp16)[name = string("upd_25_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_4919_cast_fp16 = mul(x = S_31_cast_fp16, y = last_25_cast_fp16)[name = string("op_4919_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_33_cast_fp16 = add(x = var_4919_cast_fp16, y = upd_25_cast_fp16)[name = string("S_33_cast_fp16")];
            tensor<int32, [4]> q_75_begin_0 = const()[name = string("q_75_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_75_end_0 = const()[name = string("q_75_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> q_75_end_mask_0 = const()[name = string("q_75_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_75_cast_fp16 = slice_by_index(begin = q_75_begin_0, end = q_75_end_0, end_mask = q_75_end_mask_0, x = q_69_cast_fp16)[name = string("q_75_cast_fp16")];
            tensor<int32, [4]> k_77_begin_0 = const()[name = string("k_77_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_77_end_0 = const()[name = string("k_77_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> k_77_end_mask_0 = const()[name = string("k_77_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, x = k_71_cast_fp16)[name = string("k_77_cast_fp16")];
            tensor<int32, [4]> kb_41_begin_0 = const()[name = string("kb_41_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_41_end_0 = const()[name = string("kb_41_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> kb_41_end_mask_0 = const()[name = string("kb_41_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_41_cast_fp16 = slice_by_index(begin = kb_41_begin_0, end = kb_41_end_0, end_mask = kb_41_end_mask_0, x = var_4668_cast_fp16)[name = string("kb_41_cast_fp16")];
            tensor<int32, [4]> vb_49_begin_0 = const()[name = string("vb_49_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_49_end_0 = const()[name = string("vb_49_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> vb_49_end_mask_0 = const()[name = string("vb_49_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_49_cast_fp16 = slice_by_index(begin = vb_49_begin_0, end = vb_49_end_0, end_mask = vb_49_end_mask_0, x = vb_43_cast_fp16)[name = string("vb_49_cast_fp16")];
            tensor<int32, [3]> gdec_49_begin_0 = const()[name = string("gdec_49_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_49_end_0 = const()[name = string("gdec_49_end_0"), val = tensor<int32, [3]>([3, 4, 48])];
            tensor<bool, [3]> gdec_49_end_mask_0 = const()[name = string("gdec_49_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_49_cast_fp16 = slice_by_index(begin = gdec_49_begin_0, end = gdec_49_end_0, end_mask = gdec_49_end_mask_0, x = gdec_43_cast_fp16)[name = string("gdec_49_cast_fp16")];
            tensor<int32, [2]> var_5002 = const()[name = string("op_5002"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_5003_cast_fp16 = reshape(shape = var_5002, x = gdec_49_cast_fp16)[name = string("op_5003_cast_fp16")];
            tensor<fp16, [12, 256]> var_5004_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5003_cast_fp16)[name = string("op_5004_cast_fp16")];
            tensor<int32, [4]> var_5009 = const()[name = string("op_5009"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_41_cast_fp16 = reshape(shape = var_5009, x = var_5004_cast_fp16)[name = string("P_41_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5011_cast_fp16 = exp(x = P_41_cast_fp16)[name = string("op_5011_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_41_cast_fp16 = mul(x = var_5011_cast_fp16, y = tril_incl_to_fp16)[name = string("D_41_cast_fp16")];
            bool kk_41_transpose_x_1 = const()[name = string("kk_41_transpose_x_1"), val = bool(false)];
            bool kk_41_transpose_y_1 = const()[name = string("kk_41_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_41_cast_fp16 = matmul(transpose_x = kk_41_transpose_x_1, transpose_y = kk_41_transpose_y_1, x = kb_41_cast_fp16, y = k_77_cast_fp16)[name = string("kk_41_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_5017_cast_fp16 = mul(x = kk_41_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_5017_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5018_cast_fp16 = mul(x = var_5017_cast_fp16, y = D_41_cast_fp16)[name = string("op_5018_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_41_cast_fp16 = mul(x = var_5018_cast_fp16, y = cum_tail_to_fp16)[name = string("N_41_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5020_cast_fp16 = mul(x = D_41_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5020_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_41_cast_fp16 = mul(x = kk_41_cast_fp16, y = var_5020_cast_fp16)[name = string("Na_41_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_161_cast_fp16 = add(x = eye_to_fp16, y = N_41_cast_fp16)[name = string("Minv_161_cast_fp16")];
            bool Np_121_transpose_x_0 = const()[name = string("Np_121_transpose_x_0"), val = bool(false)];
            bool Np_121_transpose_y_0 = const()[name = string("Np_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_121_cast_fp16 = matmul(transpose_x = Np_121_transpose_x_0, transpose_y = Np_121_transpose_y_0, x = N_41_cast_fp16, y = Na_41_cast_fp16)[name = string("Np_121_cast_fp16")];
            bool Npa_81_transpose_x_0 = const()[name = string("Npa_81_transpose_x_0"), val = bool(false)];
            bool Npa_81_transpose_y_0 = const()[name = string("Npa_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_81_cast_fp16 = matmul(transpose_x = Npa_81_transpose_x_0, transpose_y = Npa_81_transpose_y_0, x = Na_41_cast_fp16, y = N_41_cast_fp16)[name = string("Npa_81_cast_fp16")];
            bool var_5026_transpose_x_0 = const()[name = string("op_5026_transpose_x_0"), val = bool(false)];
            bool var_5026_transpose_y_0 = const()[name = string("op_5026_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5026_cast_fp16 = matmul(transpose_x = var_5026_transpose_x_0, transpose_y = var_5026_transpose_y_0, x = Minv_161_cast_fp16, y = Np_121_cast_fp16)[name = string("op_5026_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_163_cast_fp16 = add(x = Minv_161_cast_fp16, y = var_5026_cast_fp16)[name = string("Minv_163_cast_fp16")];
            bool Np_123_transpose_x_0 = const()[name = string("Np_123_transpose_x_0"), val = bool(false)];
            bool Np_123_transpose_y_0 = const()[name = string("Np_123_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_123_cast_fp16 = matmul(transpose_x = Np_123_transpose_x_0, transpose_y = Np_123_transpose_y_0, x = Np_121_cast_fp16, y = Npa_81_cast_fp16)[name = string("Np_123_cast_fp16")];
            bool Npa_83_transpose_x_0 = const()[name = string("Npa_83_transpose_x_0"), val = bool(false)];
            bool Npa_83_transpose_y_0 = const()[name = string("Npa_83_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_83_cast_fp16 = matmul(transpose_x = Npa_83_transpose_x_0, transpose_y = Npa_83_transpose_y_0, x = Npa_81_cast_fp16, y = Np_121_cast_fp16)[name = string("Npa_83_cast_fp16")];
            bool var_5031_transpose_x_0 = const()[name = string("op_5031_transpose_x_0"), val = bool(false)];
            bool var_5031_transpose_y_0 = const()[name = string("op_5031_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5031_cast_fp16 = matmul(transpose_x = var_5031_transpose_x_0, transpose_y = var_5031_transpose_y_0, x = Minv_163_cast_fp16, y = Np_123_cast_fp16)[name = string("op_5031_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_165_cast_fp16 = add(x = Minv_163_cast_fp16, y = var_5031_cast_fp16)[name = string("Minv_165_cast_fp16")];
            bool Np_125_transpose_x_0 = const()[name = string("Np_125_transpose_x_0"), val = bool(false)];
            bool Np_125_transpose_y_0 = const()[name = string("Np_125_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_125_cast_fp16 = matmul(transpose_x = Np_125_transpose_x_0, transpose_y = Np_125_transpose_y_0, x = Np_123_cast_fp16, y = Npa_83_cast_fp16)[name = string("Np_125_cast_fp16")];
            bool var_5035_transpose_x_0 = const()[name = string("op_5035_transpose_x_0"), val = bool(false)];
            bool var_5035_transpose_y_0 = const()[name = string("op_5035_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5035_cast_fp16 = matmul(transpose_x = var_5035_transpose_x_0, transpose_y = var_5035_transpose_y_0, x = Minv_165_cast_fp16, y = Np_125_cast_fp16)[name = string("op_5035_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_167_cast_fp16 = add(x = Minv_165_cast_fp16, y = var_5035_cast_fp16)[name = string("Minv_167_cast_fp16")];
            bool u_41_transpose_x_0 = const()[name = string("u_41_transpose_x_0"), val = bool(false)];
            bool u_41_transpose_y_0 = const()[name = string("u_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_41_cast_fp16 = matmul(transpose_x = u_41_transpose_x_0, transpose_y = u_41_transpose_y_0, x = Minv_167_cast_fp16, y = vb_49_cast_fp16)[name = string("u_41_cast_fp16")];
            bool var_5042_transpose_x_1 = const()[name = string("op_5042_transpose_x_1"), val = bool(false)];
            bool var_5042_transpose_y_1 = const()[name = string("op_5042_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_5042_cast_fp16 = matmul(transpose_x = var_5042_transpose_x_1, transpose_y = var_5042_transpose_y_1, x = q_75_cast_fp16, y = k_77_cast_fp16)[name = string("op_5042_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_41_cast_fp16 = mul(x = var_5042_cast_fp16, y = D_41_cast_fp16)[name = string("A_41_cast_fp16")];
            tensor<fp16, [16]> var_5044_bias_0_to_fp16 = const()[name = string("op_5044_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_5044_cast_fp16 = linear(bias = var_5044_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_49_cast_fp16)[name = string("op_5044_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_33_cast_fp16 = exp(x = var_5044_cast_fp16)[name = string("cum_33_cast_fp16")];
            tensor<int32, [1]> var_5047_axes_0 = const()[name = string("op_5047_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_5047_cast_fp16 = expand_dims(axes = var_5047_axes_0, x = cum_33_cast_fp16)[name = string("op_5047_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5048_cast_fp16 = mul(x = kb_41_cast_fp16, y = var_5047_cast_fp16)[name = string("op_5048_cast_fp16")];
            bool w_41_transpose_x_0 = const()[name = string("w_41_transpose_x_0"), val = bool(false)];
            bool w_41_transpose_y_0 = const()[name = string("w_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_41_cast_fp16 = matmul(transpose_x = w_41_transpose_x_0, transpose_y = w_41_transpose_y_0, x = Minv_167_cast_fp16, y = var_5048_cast_fp16)[name = string("w_41_cast_fp16")];
            bool var_5050_transpose_x_0 = const()[name = string("op_5050_transpose_x_0"), val = bool(false)];
            bool var_5050_transpose_y_0 = const()[name = string("op_5050_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5050_cast_fp16 = matmul(transpose_x = var_5050_transpose_x_0, transpose_y = var_5050_transpose_y_0, x = w_41_cast_fp16, y = S_33_cast_fp16)[name = string("op_5050_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_33_cast_fp16 = sub(x = u_41_cast_fp16, y = var_5050_cast_fp16)[name = string("vnew_33_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5055_cast_fp16 = mul(x = q_75_cast_fp16, y = var_5047_cast_fp16)[name = string("op_5055_cast_fp16")];
            bool var_5056_transpose_x_0 = const()[name = string("op_5056_transpose_x_0"), val = bool(false)];
            bool var_5056_transpose_y_0 = const()[name = string("op_5056_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5056_cast_fp16 = matmul(transpose_x = var_5056_transpose_x_0, transpose_y = var_5056_transpose_y_0, x = var_5055_cast_fp16, y = S_33_cast_fp16)[name = string("op_5056_cast_fp16")];
            bool var_5057_transpose_x_0 = const()[name = string("op_5057_transpose_x_0"), val = bool(false)];
            bool var_5057_transpose_y_0 = const()[name = string("op_5057_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5057_cast_fp16 = matmul(transpose_x = var_5057_transpose_x_0, transpose_y = var_5057_transpose_y_0, x = A_41_cast_fp16, y = vnew_33_cast_fp16)[name = string("op_5057_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5059_cast_fp16 = add(x = var_5056_cast_fp16, y = var_5057_cast_fp16)[name = string("op_5059_cast_fp16")];
            tensor<fp16, [16]> var_5060_bias_0_to_fp16 = const()[name = string("op_5060_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_5060_cast_fp16 = linear(bias = var_5060_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_49_cast_fp16)[name = string("op_5060_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_35_cast_fp16 = exp(x = var_5060_cast_fp16)[name = string("tail_35_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_69_to_fp16 = const()[name = string("transpose_69_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_5062_bias_0_to_fp16 = const()[name = string("op_5062_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_5062_cast_fp16 = linear(bias = var_5062_bias_0_to_fp16, weight = transpose_69_to_fp16, x = gdec_49_cast_fp16)[name = string("op_5062_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_5063_cast_fp16 = exp(x = var_5062_cast_fp16)[name = string("op_5063_cast_fp16")];
            tensor<int32, [1]> last_27_axes_0 = const()[name = string("last_27_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_27_cast_fp16 = expand_dims(axes = last_27_axes_0, x = var_5063_cast_fp16)[name = string("last_27_cast_fp16")];
            tensor<int32, [1]> var_5067_axes_0 = const()[name = string("op_5067_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_5067_cast_fp16 = expand_dims(axes = var_5067_axes_0, x = tail_35_cast_fp16)[name = string("op_5067_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5068_cast_fp16 = mul(x = k_77_cast_fp16, y = var_5067_cast_fp16)[name = string("op_5068_cast_fp16")];
            bool upd_27_transpose_x_1 = const()[name = string("upd_27_transpose_x_1"), val = bool(true)];
            bool upd_27_transpose_y_1 = const()[name = string("upd_27_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_27_cast_fp16 = matmul(transpose_x = upd_27_transpose_x_1, transpose_y = upd_27_transpose_y_1, x = var_5068_cast_fp16, y = vnew_33_cast_fp16)[name = string("upd_27_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_5073_cast_fp16 = mul(x = S_33_cast_fp16, y = last_27_cast_fp16)[name = string("op_5073_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_35_cast_fp16 = add(x = var_5073_cast_fp16, y = upd_27_cast_fp16)[name = string("S_35_cast_fp16")];
            tensor<int32, [4]> q_77_begin_0 = const()[name = string("q_77_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_77_end_0 = const()[name = string("q_77_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> q_77_end_mask_0 = const()[name = string("q_77_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_77_cast_fp16 = slice_by_index(begin = q_77_begin_0, end = q_77_end_0, end_mask = q_77_end_mask_0, x = q_69_cast_fp16)[name = string("q_77_cast_fp16")];
            tensor<int32, [4]> k_79_begin_0 = const()[name = string("k_79_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_79_end_0 = const()[name = string("k_79_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> k_79_end_mask_0 = const()[name = string("k_79_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_79_cast_fp16 = slice_by_index(begin = k_79_begin_0, end = k_79_end_0, end_mask = k_79_end_mask_0, x = k_71_cast_fp16)[name = string("k_79_cast_fp16")];
            tensor<int32, [4]> kb_43_begin_0 = const()[name = string("kb_43_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_43_end_0 = const()[name = string("kb_43_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> kb_43_end_mask_0 = const()[name = string("kb_43_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_43_cast_fp16 = slice_by_index(begin = kb_43_begin_0, end = kb_43_end_0, end_mask = kb_43_end_mask_0, x = var_4668_cast_fp16)[name = string("kb_43_cast_fp16")];
            tensor<int32, [4]> vb_51_begin_0 = const()[name = string("vb_51_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_51_end_0 = const()[name = string("vb_51_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> vb_51_end_mask_0 = const()[name = string("vb_51_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_51_cast_fp16 = slice_by_index(begin = vb_51_begin_0, end = vb_51_end_0, end_mask = vb_51_end_mask_0, x = vb_43_cast_fp16)[name = string("vb_51_cast_fp16")];
            tensor<int32, [3]> gdec_51_begin_0 = const()[name = string("gdec_51_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_51_end_0 = const()[name = string("gdec_51_end_0"), val = tensor<int32, [3]>([3, 4, 64])];
            tensor<bool, [3]> gdec_51_end_mask_0 = const()[name = string("gdec_51_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_51_cast_fp16 = slice_by_index(begin = gdec_51_begin_0, end = gdec_51_end_0, end_mask = gdec_51_end_mask_0, x = gdec_43_cast_fp16)[name = string("gdec_51_cast_fp16")];
            tensor<int32, [2]> var_5156 = const()[name = string("op_5156"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_5157_cast_fp16 = reshape(shape = var_5156, x = gdec_51_cast_fp16)[name = string("op_5157_cast_fp16")];
            tensor<fp16, [12, 256]> var_5158_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5157_cast_fp16)[name = string("op_5158_cast_fp16")];
            tensor<int32, [4]> var_5163 = const()[name = string("op_5163"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_43_cast_fp16 = reshape(shape = var_5163, x = var_5158_cast_fp16)[name = string("P_43_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5165_cast_fp16 = exp(x = P_43_cast_fp16)[name = string("op_5165_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_43_cast_fp16 = mul(x = var_5165_cast_fp16, y = tril_incl_to_fp16)[name = string("D_43_cast_fp16")];
            bool kk_43_transpose_x_1 = const()[name = string("kk_43_transpose_x_1"), val = bool(false)];
            bool kk_43_transpose_y_1 = const()[name = string("kk_43_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_43_cast_fp16 = matmul(transpose_x = kk_43_transpose_x_1, transpose_y = kk_43_transpose_y_1, x = kb_43_cast_fp16, y = k_79_cast_fp16)[name = string("kk_43_cast_fp16")];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_5171_cast_fp16 = mul(x = kk_43_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_5171_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5172_cast_fp16 = mul(x = var_5171_cast_fp16, y = D_43_cast_fp16)[name = string("op_5172_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_43_cast_fp16 = mul(x = var_5172_cast_fp16, y = cum_tail_to_fp16)[name = string("N_43_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5174_cast_fp16 = mul(x = D_43_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5174_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_43_cast_fp16 = mul(x = kk_43_cast_fp16, y = var_5174_cast_fp16)[name = string("Na_43_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_169_cast_fp16 = add(x = eye_to_fp16, y = N_43_cast_fp16)[name = string("Minv_169_cast_fp16")];
            bool Np_127_transpose_x_0 = const()[name = string("Np_127_transpose_x_0"), val = bool(false)];
            bool Np_127_transpose_y_0 = const()[name = string("Np_127_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_127_cast_fp16 = matmul(transpose_x = Np_127_transpose_x_0, transpose_y = Np_127_transpose_y_0, x = N_43_cast_fp16, y = Na_43_cast_fp16)[name = string("Np_127_cast_fp16")];
            bool Npa_85_transpose_x_0 = const()[name = string("Npa_85_transpose_x_0"), val = bool(false)];
            bool Npa_85_transpose_y_0 = const()[name = string("Npa_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_85_cast_fp16 = matmul(transpose_x = Npa_85_transpose_x_0, transpose_y = Npa_85_transpose_y_0, x = Na_43_cast_fp16, y = N_43_cast_fp16)[name = string("Npa_85_cast_fp16")];
            bool var_5180_transpose_x_0 = const()[name = string("op_5180_transpose_x_0"), val = bool(false)];
            bool var_5180_transpose_y_0 = const()[name = string("op_5180_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5180_cast_fp16 = matmul(transpose_x = var_5180_transpose_x_0, transpose_y = var_5180_transpose_y_0, x = Minv_169_cast_fp16, y = Np_127_cast_fp16)[name = string("op_5180_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_171_cast_fp16 = add(x = Minv_169_cast_fp16, y = var_5180_cast_fp16)[name = string("Minv_171_cast_fp16")];
            bool Np_129_transpose_x_0 = const()[name = string("Np_129_transpose_x_0"), val = bool(false)];
            bool Np_129_transpose_y_0 = const()[name = string("Np_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_129_cast_fp16 = matmul(transpose_x = Np_129_transpose_x_0, transpose_y = Np_129_transpose_y_0, x = Np_127_cast_fp16, y = Npa_85_cast_fp16)[name = string("Np_129_cast_fp16")];
            bool Npa_87_transpose_x_0 = const()[name = string("Npa_87_transpose_x_0"), val = bool(false)];
            bool Npa_87_transpose_y_0 = const()[name = string("Npa_87_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_87_cast_fp16 = matmul(transpose_x = Npa_87_transpose_x_0, transpose_y = Npa_87_transpose_y_0, x = Npa_85_cast_fp16, y = Np_127_cast_fp16)[name = string("Npa_87_cast_fp16")];
            bool var_5185_transpose_x_0 = const()[name = string("op_5185_transpose_x_0"), val = bool(false)];
            bool var_5185_transpose_y_0 = const()[name = string("op_5185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5185_cast_fp16 = matmul(transpose_x = var_5185_transpose_x_0, transpose_y = var_5185_transpose_y_0, x = Minv_171_cast_fp16, y = Np_129_cast_fp16)[name = string("op_5185_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_173_cast_fp16 = add(x = Minv_171_cast_fp16, y = var_5185_cast_fp16)[name = string("Minv_173_cast_fp16")];
            bool Np_131_transpose_x_0 = const()[name = string("Np_131_transpose_x_0"), val = bool(false)];
            bool Np_131_transpose_y_0 = const()[name = string("Np_131_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_131_cast_fp16 = matmul(transpose_x = Np_131_transpose_x_0, transpose_y = Np_131_transpose_y_0, x = Np_129_cast_fp16, y = Npa_87_cast_fp16)[name = string("Np_131_cast_fp16")];
            bool var_5189_transpose_x_0 = const()[name = string("op_5189_transpose_x_0"), val = bool(false)];
            bool var_5189_transpose_y_0 = const()[name = string("op_5189_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5189_cast_fp16 = matmul(transpose_x = var_5189_transpose_x_0, transpose_y = var_5189_transpose_y_0, x = Minv_173_cast_fp16, y = Np_131_cast_fp16)[name = string("op_5189_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_175_cast_fp16 = add(x = Minv_173_cast_fp16, y = var_5189_cast_fp16)[name = string("Minv_175_cast_fp16")];
            bool u_43_transpose_x_0 = const()[name = string("u_43_transpose_x_0"), val = bool(false)];
            bool u_43_transpose_y_0 = const()[name = string("u_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_43_cast_fp16 = matmul(transpose_x = u_43_transpose_x_0, transpose_y = u_43_transpose_y_0, x = Minv_175_cast_fp16, y = vb_51_cast_fp16)[name = string("u_43_cast_fp16")];
            bool var_5196_transpose_x_1 = const()[name = string("op_5196_transpose_x_1"), val = bool(false)];
            bool var_5196_transpose_y_1 = const()[name = string("op_5196_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_5196_cast_fp16 = matmul(transpose_x = var_5196_transpose_x_1, transpose_y = var_5196_transpose_y_1, x = q_77_cast_fp16, y = k_79_cast_fp16)[name = string("op_5196_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_43_cast_fp16 = mul(x = var_5196_cast_fp16, y = D_43_cast_fp16)[name = string("A_43_cast_fp16")];
            tensor<fp16, [16]> var_5198_bias_0_to_fp16 = const()[name = string("op_5198_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_5198_cast_fp16 = linear(bias = var_5198_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_51_cast_fp16)[name = string("op_5198_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_35_cast_fp16 = exp(x = var_5198_cast_fp16)[name = string("cum_35_cast_fp16")];
            tensor<int32, [1]> var_5201_axes_0 = const()[name = string("op_5201_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_5201_cast_fp16 = expand_dims(axes = var_5201_axes_0, x = cum_35_cast_fp16)[name = string("op_5201_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5202_cast_fp16 = mul(x = kb_43_cast_fp16, y = var_5201_cast_fp16)[name = string("op_5202_cast_fp16")];
            bool w_43_transpose_x_0 = const()[name = string("w_43_transpose_x_0"), val = bool(false)];
            bool w_43_transpose_y_0 = const()[name = string("w_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_43_cast_fp16 = matmul(transpose_x = w_43_transpose_x_0, transpose_y = w_43_transpose_y_0, x = Minv_175_cast_fp16, y = var_5202_cast_fp16)[name = string("w_43_cast_fp16")];
            bool var_5204_transpose_x_0 = const()[name = string("op_5204_transpose_x_0"), val = bool(false)];
            bool var_5204_transpose_y_0 = const()[name = string("op_5204_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5204_cast_fp16 = matmul(transpose_x = var_5204_transpose_x_0, transpose_y = var_5204_transpose_y_0, x = w_43_cast_fp16, y = S_35_cast_fp16)[name = string("op_5204_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_35_cast_fp16 = sub(x = u_43_cast_fp16, y = var_5204_cast_fp16)[name = string("vnew_35_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5209_cast_fp16 = mul(x = q_77_cast_fp16, y = var_5201_cast_fp16)[name = string("op_5209_cast_fp16")];
            bool var_5210_transpose_x_0 = const()[name = string("op_5210_transpose_x_0"), val = bool(false)];
            bool var_5210_transpose_y_0 = const()[name = string("op_5210_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5210_cast_fp16 = matmul(transpose_x = var_5210_transpose_x_0, transpose_y = var_5210_transpose_y_0, x = var_5209_cast_fp16, y = S_35_cast_fp16)[name = string("op_5210_cast_fp16")];
            bool var_5211_transpose_x_0 = const()[name = string("op_5211_transpose_x_0"), val = bool(false)];
            bool var_5211_transpose_y_0 = const()[name = string("op_5211_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5211_cast_fp16 = matmul(transpose_x = var_5211_transpose_x_0, transpose_y = var_5211_transpose_y_0, x = A_43_cast_fp16, y = vnew_35_cast_fp16)[name = string("op_5211_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5213_cast_fp16 = add(x = var_5210_cast_fp16, y = var_5211_cast_fp16)[name = string("op_5213_cast_fp16")];
            tensor<fp16, [16]> var_5214_bias_0_to_fp16 = const()[name = string("op_5214_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_5214_cast_fp16 = linear(bias = var_5214_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_51_cast_fp16)[name = string("op_5214_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_37_cast_fp16 = exp(x = var_5214_cast_fp16)[name = string("tail_37_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_73_to_fp16 = const()[name = string("transpose_73_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_5216_bias_0_to_fp16 = const()[name = string("op_5216_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_5216_cast_fp16 = linear(bias = var_5216_bias_0_to_fp16, weight = transpose_73_to_fp16, x = gdec_51_cast_fp16)[name = string("op_5216_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_5217_cast_fp16 = exp(x = var_5216_cast_fp16)[name = string("op_5217_cast_fp16")];
            tensor<int32, [1]> last_29_axes_0 = const()[name = string("last_29_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_29_cast_fp16 = expand_dims(axes = last_29_axes_0, x = var_5217_cast_fp16)[name = string("last_29_cast_fp16")];
            tensor<int32, [1]> var_5221_axes_0 = const()[name = string("op_5221_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_5221_cast_fp16 = expand_dims(axes = var_5221_axes_0, x = tail_37_cast_fp16)[name = string("op_5221_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5222_cast_fp16 = mul(x = k_79_cast_fp16, y = var_5221_cast_fp16)[name = string("op_5222_cast_fp16")];
            bool upd_29_transpose_x_1 = const()[name = string("upd_29_transpose_x_1"), val = bool(true)];
            bool upd_29_transpose_y_1 = const()[name = string("upd_29_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_29_cast_fp16 = matmul(transpose_x = upd_29_transpose_x_1, transpose_y = upd_29_transpose_y_1, x = var_5222_cast_fp16, y = vnew_35_cast_fp16)[name = string("upd_29_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_5227_cast_fp16 = mul(x = S_35_cast_fp16, y = last_29_cast_fp16)[name = string("op_5227_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_37_cast_fp16 = add(x = var_5227_cast_fp16, y = upd_29_cast_fp16)[name = string("S_37_cast_fp16")];
            tensor<int32, [4]> q_79_begin_0 = const()[name = string("q_79_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> q_79_end_0 = const()[name = string("q_79_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> q_79_end_mask_0 = const()[name = string("q_79_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, x = q_69_cast_fp16)[name = string("q_79_cast_fp16")];
            tensor<int32, [4]> k_81_begin_0 = const()[name = string("k_81_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> k_81_end_0 = const()[name = string("k_81_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> k_81_end_mask_0 = const()[name = string("k_81_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, x = k_71_cast_fp16)[name = string("k_81_cast_fp16")];
            tensor<int32, [4]> kb_45_begin_0 = const()[name = string("kb_45_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> kb_45_end_0 = const()[name = string("kb_45_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> kb_45_end_mask_0 = const()[name = string("kb_45_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_45_cast_fp16 = slice_by_index(begin = kb_45_begin_0, end = kb_45_end_0, end_mask = kb_45_end_mask_0, x = var_4668_cast_fp16)[name = string("kb_45_cast_fp16")];
            tensor<int32, [4]> vb_53_begin_0 = const()[name = string("vb_53_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> vb_53_end_0 = const()[name = string("vb_53_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> vb_53_end_mask_0 = const()[name = string("vb_53_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_53_cast_fp16 = slice_by_index(begin = vb_53_begin_0, end = vb_53_end_0, end_mask = vb_53_end_mask_0, x = vb_43_cast_fp16)[name = string("vb_53_cast_fp16")];
            tensor<int32, [3]> gdec_53_begin_0 = const()[name = string("gdec_53_begin_0"), val = tensor<int32, [3]>([0, 0, 64])];
            tensor<int32, [3]> gdec_53_end_0 = const()[name = string("gdec_53_end_0"), val = tensor<int32, [3]>([3, 4, 80])];
            tensor<bool, [3]> gdec_53_end_mask_0 = const()[name = string("gdec_53_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_53_cast_fp16 = slice_by_index(begin = gdec_53_begin_0, end = gdec_53_end_0, end_mask = gdec_53_end_mask_0, x = gdec_43_cast_fp16)[name = string("gdec_53_cast_fp16")];
            tensor<int32, [2]> var_5310 = const()[name = string("op_5310"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_5311_cast_fp16 = reshape(shape = var_5310, x = gdec_53_cast_fp16)[name = string("op_5311_cast_fp16")];
            tensor<fp16, [12, 256]> var_5312_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5311_cast_fp16)[name = string("op_5312_cast_fp16")];
            tensor<int32, [4]> var_5317 = const()[name = string("op_5317"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_45_cast_fp16 = reshape(shape = var_5317, x = var_5312_cast_fp16)[name = string("P_45_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5319_cast_fp16 = exp(x = P_45_cast_fp16)[name = string("op_5319_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_45_cast_fp16 = mul(x = var_5319_cast_fp16, y = tril_incl_to_fp16)[name = string("D_45_cast_fp16")];
            bool kk_45_transpose_x_1 = const()[name = string("kk_45_transpose_x_1"), val = bool(false)];
            bool kk_45_transpose_y_1 = const()[name = string("kk_45_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_45_cast_fp16 = matmul(transpose_x = kk_45_transpose_x_1, transpose_y = kk_45_transpose_y_1, x = kb_45_cast_fp16, y = k_81_cast_fp16)[name = string("kk_45_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_5325_cast_fp16 = mul(x = kk_45_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_5325_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5326_cast_fp16 = mul(x = var_5325_cast_fp16, y = D_45_cast_fp16)[name = string("op_5326_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_45_cast_fp16 = mul(x = var_5326_cast_fp16, y = cum_tail_to_fp16)[name = string("N_45_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5328_cast_fp16 = mul(x = D_45_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5328_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_45_cast_fp16 = mul(x = kk_45_cast_fp16, y = var_5328_cast_fp16)[name = string("Na_45_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_177_cast_fp16 = add(x = eye_to_fp16, y = N_45_cast_fp16)[name = string("Minv_177_cast_fp16")];
            bool Np_133_transpose_x_0 = const()[name = string("Np_133_transpose_x_0"), val = bool(false)];
            bool Np_133_transpose_y_0 = const()[name = string("Np_133_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_133_cast_fp16 = matmul(transpose_x = Np_133_transpose_x_0, transpose_y = Np_133_transpose_y_0, x = N_45_cast_fp16, y = Na_45_cast_fp16)[name = string("Np_133_cast_fp16")];
            bool Npa_89_transpose_x_0 = const()[name = string("Npa_89_transpose_x_0"), val = bool(false)];
            bool Npa_89_transpose_y_0 = const()[name = string("Npa_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_89_cast_fp16 = matmul(transpose_x = Npa_89_transpose_x_0, transpose_y = Npa_89_transpose_y_0, x = Na_45_cast_fp16, y = N_45_cast_fp16)[name = string("Npa_89_cast_fp16")];
            bool var_5334_transpose_x_0 = const()[name = string("op_5334_transpose_x_0"), val = bool(false)];
            bool var_5334_transpose_y_0 = const()[name = string("op_5334_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5334_cast_fp16 = matmul(transpose_x = var_5334_transpose_x_0, transpose_y = var_5334_transpose_y_0, x = Minv_177_cast_fp16, y = Np_133_cast_fp16)[name = string("op_5334_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_179_cast_fp16 = add(x = Minv_177_cast_fp16, y = var_5334_cast_fp16)[name = string("Minv_179_cast_fp16")];
            bool Np_135_transpose_x_0 = const()[name = string("Np_135_transpose_x_0"), val = bool(false)];
            bool Np_135_transpose_y_0 = const()[name = string("Np_135_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_135_cast_fp16 = matmul(transpose_x = Np_135_transpose_x_0, transpose_y = Np_135_transpose_y_0, x = Np_133_cast_fp16, y = Npa_89_cast_fp16)[name = string("Np_135_cast_fp16")];
            bool Npa_91_transpose_x_0 = const()[name = string("Npa_91_transpose_x_0"), val = bool(false)];
            bool Npa_91_transpose_y_0 = const()[name = string("Npa_91_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_91_cast_fp16 = matmul(transpose_x = Npa_91_transpose_x_0, transpose_y = Npa_91_transpose_y_0, x = Npa_89_cast_fp16, y = Np_133_cast_fp16)[name = string("Npa_91_cast_fp16")];
            bool var_5339_transpose_x_0 = const()[name = string("op_5339_transpose_x_0"), val = bool(false)];
            bool var_5339_transpose_y_0 = const()[name = string("op_5339_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5339_cast_fp16 = matmul(transpose_x = var_5339_transpose_x_0, transpose_y = var_5339_transpose_y_0, x = Minv_179_cast_fp16, y = Np_135_cast_fp16)[name = string("op_5339_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_181_cast_fp16 = add(x = Minv_179_cast_fp16, y = var_5339_cast_fp16)[name = string("Minv_181_cast_fp16")];
            bool Np_137_transpose_x_0 = const()[name = string("Np_137_transpose_x_0"), val = bool(false)];
            bool Np_137_transpose_y_0 = const()[name = string("Np_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_137_cast_fp16 = matmul(transpose_x = Np_137_transpose_x_0, transpose_y = Np_137_transpose_y_0, x = Np_135_cast_fp16, y = Npa_91_cast_fp16)[name = string("Np_137_cast_fp16")];
            bool var_5343_transpose_x_0 = const()[name = string("op_5343_transpose_x_0"), val = bool(false)];
            bool var_5343_transpose_y_0 = const()[name = string("op_5343_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5343_cast_fp16 = matmul(transpose_x = var_5343_transpose_x_0, transpose_y = var_5343_transpose_y_0, x = Minv_181_cast_fp16, y = Np_137_cast_fp16)[name = string("op_5343_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_183_cast_fp16 = add(x = Minv_181_cast_fp16, y = var_5343_cast_fp16)[name = string("Minv_183_cast_fp16")];
            bool u_45_transpose_x_0 = const()[name = string("u_45_transpose_x_0"), val = bool(false)];
            bool u_45_transpose_y_0 = const()[name = string("u_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_45_cast_fp16 = matmul(transpose_x = u_45_transpose_x_0, transpose_y = u_45_transpose_y_0, x = Minv_183_cast_fp16, y = vb_53_cast_fp16)[name = string("u_45_cast_fp16")];
            bool var_5350_transpose_x_1 = const()[name = string("op_5350_transpose_x_1"), val = bool(false)];
            bool var_5350_transpose_y_1 = const()[name = string("op_5350_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_5350_cast_fp16 = matmul(transpose_x = var_5350_transpose_x_1, transpose_y = var_5350_transpose_y_1, x = q_79_cast_fp16, y = k_81_cast_fp16)[name = string("op_5350_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_45_cast_fp16 = mul(x = var_5350_cast_fp16, y = D_45_cast_fp16)[name = string("A_45_cast_fp16")];
            tensor<fp16, [16]> var_5352_bias_0_to_fp16 = const()[name = string("op_5352_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_5352_cast_fp16 = linear(bias = var_5352_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_53_cast_fp16)[name = string("op_5352_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_37_cast_fp16 = exp(x = var_5352_cast_fp16)[name = string("cum_37_cast_fp16")];
            tensor<int32, [1]> var_5355_axes_0 = const()[name = string("op_5355_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_5355_cast_fp16 = expand_dims(axes = var_5355_axes_0, x = cum_37_cast_fp16)[name = string("op_5355_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5356_cast_fp16 = mul(x = kb_45_cast_fp16, y = var_5355_cast_fp16)[name = string("op_5356_cast_fp16")];
            bool w_45_transpose_x_0 = const()[name = string("w_45_transpose_x_0"), val = bool(false)];
            bool w_45_transpose_y_0 = const()[name = string("w_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_45_cast_fp16 = matmul(transpose_x = w_45_transpose_x_0, transpose_y = w_45_transpose_y_0, x = Minv_183_cast_fp16, y = var_5356_cast_fp16)[name = string("w_45_cast_fp16")];
            bool var_5358_transpose_x_0 = const()[name = string("op_5358_transpose_x_0"), val = bool(false)];
            bool var_5358_transpose_y_0 = const()[name = string("op_5358_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5358_cast_fp16 = matmul(transpose_x = var_5358_transpose_x_0, transpose_y = var_5358_transpose_y_0, x = w_45_cast_fp16, y = S_37_cast_fp16)[name = string("op_5358_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_37_cast_fp16 = sub(x = u_45_cast_fp16, y = var_5358_cast_fp16)[name = string("vnew_37_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5363_cast_fp16 = mul(x = q_79_cast_fp16, y = var_5355_cast_fp16)[name = string("op_5363_cast_fp16")];
            bool var_5364_transpose_x_0 = const()[name = string("op_5364_transpose_x_0"), val = bool(false)];
            bool var_5364_transpose_y_0 = const()[name = string("op_5364_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5364_cast_fp16 = matmul(transpose_x = var_5364_transpose_x_0, transpose_y = var_5364_transpose_y_0, x = var_5363_cast_fp16, y = S_37_cast_fp16)[name = string("op_5364_cast_fp16")];
            bool var_5365_transpose_x_0 = const()[name = string("op_5365_transpose_x_0"), val = bool(false)];
            bool var_5365_transpose_y_0 = const()[name = string("op_5365_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5365_cast_fp16 = matmul(transpose_x = var_5365_transpose_x_0, transpose_y = var_5365_transpose_y_0, x = A_45_cast_fp16, y = vnew_37_cast_fp16)[name = string("op_5365_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5367_cast_fp16 = add(x = var_5364_cast_fp16, y = var_5365_cast_fp16)[name = string("op_5367_cast_fp16")];
            tensor<fp16, [16]> var_5368_bias_0_to_fp16 = const()[name = string("op_5368_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_5368_cast_fp16 = linear(bias = var_5368_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_53_cast_fp16)[name = string("op_5368_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_39_cast_fp16 = exp(x = var_5368_cast_fp16)[name = string("tail_39_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_77_to_fp16 = const()[name = string("transpose_77_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_5370_bias_0_to_fp16 = const()[name = string("op_5370_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_5370_cast_fp16 = linear(bias = var_5370_bias_0_to_fp16, weight = transpose_77_to_fp16, x = gdec_53_cast_fp16)[name = string("op_5370_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_5371_cast_fp16 = exp(x = var_5370_cast_fp16)[name = string("op_5371_cast_fp16")];
            tensor<int32, [1]> last_31_axes_0 = const()[name = string("last_31_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_31_cast_fp16 = expand_dims(axes = last_31_axes_0, x = var_5371_cast_fp16)[name = string("last_31_cast_fp16")];
            tensor<int32, [1]> var_5375_axes_0 = const()[name = string("op_5375_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_5375_cast_fp16 = expand_dims(axes = var_5375_axes_0, x = tail_39_cast_fp16)[name = string("op_5375_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5376_cast_fp16 = mul(x = k_81_cast_fp16, y = var_5375_cast_fp16)[name = string("op_5376_cast_fp16")];
            bool upd_31_transpose_x_1 = const()[name = string("upd_31_transpose_x_1"), val = bool(true)];
            bool upd_31_transpose_y_1 = const()[name = string("upd_31_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_31_cast_fp16 = matmul(transpose_x = upd_31_transpose_x_1, transpose_y = upd_31_transpose_y_1, x = var_5376_cast_fp16, y = vnew_37_cast_fp16)[name = string("upd_31_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_5381_cast_fp16 = mul(x = S_37_cast_fp16, y = last_31_cast_fp16)[name = string("op_5381_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_39_cast_fp16 = add(x = var_5381_cast_fp16, y = upd_31_cast_fp16)[name = string("S_39_cast_fp16")];
            tensor<int32, [4]> q_81_begin_0 = const()[name = string("q_81_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> q_81_end_0 = const()[name = string("q_81_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_81_end_mask_0 = const()[name = string("q_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_81_cast_fp16 = slice_by_index(begin = q_81_begin_0, end = q_81_end_0, end_mask = q_81_end_mask_0, x = q_69_cast_fp16)[name = string("q_81_cast_fp16")];
            tensor<int32, [4]> k_83_begin_0 = const()[name = string("k_83_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> k_83_end_0 = const()[name = string("k_83_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_83_end_mask_0 = const()[name = string("k_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_83_cast_fp16 = slice_by_index(begin = k_83_begin_0, end = k_83_end_0, end_mask = k_83_end_mask_0, x = k_71_cast_fp16)[name = string("k_83_cast_fp16")];
            tensor<int32, [4]> kb_47_begin_0 = const()[name = string("kb_47_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> kb_47_end_0 = const()[name = string("kb_47_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_47_end_mask_0 = const()[name = string("kb_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_47_cast_fp16 = slice_by_index(begin = kb_47_begin_0, end = kb_47_end_0, end_mask = kb_47_end_mask_0, x = var_4668_cast_fp16)[name = string("kb_47_cast_fp16")];
            tensor<int32, [4]> vb_55_begin_0 = const()[name = string("vb_55_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> vb_55_end_0 = const()[name = string("vb_55_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_55_end_mask_0 = const()[name = string("vb_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_55_cast_fp16 = slice_by_index(begin = vb_55_begin_0, end = vb_55_end_0, end_mask = vb_55_end_mask_0, x = vb_43_cast_fp16)[name = string("vb_55_cast_fp16")];
            tensor<int32, [3]> gdec_55_begin_0 = const()[name = string("gdec_55_begin_0"), val = tensor<int32, [3]>([0, 0, 80])];
            tensor<int32, [3]> gdec_55_end_0 = const()[name = string("gdec_55_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_55_end_mask_0 = const()[name = string("gdec_55_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_55_cast_fp16 = slice_by_index(begin = gdec_55_begin_0, end = gdec_55_end_0, end_mask = gdec_55_end_mask_0, x = gdec_43_cast_fp16)[name = string("gdec_55_cast_fp16")];
            tensor<int32, [2]> var_5464 = const()[name = string("op_5464"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_5465_cast_fp16 = reshape(shape = var_5464, x = gdec_55_cast_fp16)[name = string("op_5465_cast_fp16")];
            tensor<fp16, [12, 256]> var_5466_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5465_cast_fp16)[name = string("op_5466_cast_fp16")];
            tensor<int32, [4]> var_5471 = const()[name = string("op_5471"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_47_cast_fp16 = reshape(shape = var_5471, x = var_5466_cast_fp16)[name = string("P_47_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5473_cast_fp16 = exp(x = P_47_cast_fp16)[name = string("op_5473_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_47_cast_fp16 = mul(x = var_5473_cast_fp16, y = tril_incl_to_fp16)[name = string("D_47_cast_fp16")];
            bool kk_47_transpose_x_1 = const()[name = string("kk_47_transpose_x_1"), val = bool(false)];
            bool kk_47_transpose_y_1 = const()[name = string("kk_47_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_47_cast_fp16 = matmul(transpose_x = kk_47_transpose_x_1, transpose_y = kk_47_transpose_y_1, x = kb_47_cast_fp16, y = k_83_cast_fp16)[name = string("kk_47_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_5479_cast_fp16 = mul(x = kk_47_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_5479_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5480_cast_fp16 = mul(x = var_5479_cast_fp16, y = D_47_cast_fp16)[name = string("op_5480_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_47_cast_fp16 = mul(x = var_5480_cast_fp16, y = cum_tail_to_fp16)[name = string("N_47_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5482_cast_fp16 = mul(x = D_47_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5482_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_47_cast_fp16 = mul(x = kk_47_cast_fp16, y = var_5482_cast_fp16)[name = string("Na_47_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_185_cast_fp16 = add(x = eye_to_fp16, y = N_47_cast_fp16)[name = string("Minv_185_cast_fp16")];
            bool Np_139_transpose_x_0 = const()[name = string("Np_139_transpose_x_0"), val = bool(false)];
            bool Np_139_transpose_y_0 = const()[name = string("Np_139_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_139_cast_fp16 = matmul(transpose_x = Np_139_transpose_x_0, transpose_y = Np_139_transpose_y_0, x = N_47_cast_fp16, y = Na_47_cast_fp16)[name = string("Np_139_cast_fp16")];
            bool Npa_93_transpose_x_0 = const()[name = string("Npa_93_transpose_x_0"), val = bool(false)];
            bool Npa_93_transpose_y_0 = const()[name = string("Npa_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_93_cast_fp16 = matmul(transpose_x = Npa_93_transpose_x_0, transpose_y = Npa_93_transpose_y_0, x = Na_47_cast_fp16, y = N_47_cast_fp16)[name = string("Npa_93_cast_fp16")];
            bool var_5488_transpose_x_0 = const()[name = string("op_5488_transpose_x_0"), val = bool(false)];
            bool var_5488_transpose_y_0 = const()[name = string("op_5488_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5488_cast_fp16 = matmul(transpose_x = var_5488_transpose_x_0, transpose_y = var_5488_transpose_y_0, x = Minv_185_cast_fp16, y = Np_139_cast_fp16)[name = string("op_5488_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_187_cast_fp16 = add(x = Minv_185_cast_fp16, y = var_5488_cast_fp16)[name = string("Minv_187_cast_fp16")];
            bool Np_141_transpose_x_0 = const()[name = string("Np_141_transpose_x_0"), val = bool(false)];
            bool Np_141_transpose_y_0 = const()[name = string("Np_141_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_141_cast_fp16 = matmul(transpose_x = Np_141_transpose_x_0, transpose_y = Np_141_transpose_y_0, x = Np_139_cast_fp16, y = Npa_93_cast_fp16)[name = string("Np_141_cast_fp16")];
            bool Npa_95_transpose_x_0 = const()[name = string("Npa_95_transpose_x_0"), val = bool(false)];
            bool Npa_95_transpose_y_0 = const()[name = string("Npa_95_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_95_cast_fp16 = matmul(transpose_x = Npa_95_transpose_x_0, transpose_y = Npa_95_transpose_y_0, x = Npa_93_cast_fp16, y = Np_139_cast_fp16)[name = string("Npa_95_cast_fp16")];
            bool var_5493_transpose_x_0 = const()[name = string("op_5493_transpose_x_0"), val = bool(false)];
            bool var_5493_transpose_y_0 = const()[name = string("op_5493_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5493_cast_fp16 = matmul(transpose_x = var_5493_transpose_x_0, transpose_y = var_5493_transpose_y_0, x = Minv_187_cast_fp16, y = Np_141_cast_fp16)[name = string("op_5493_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_189_cast_fp16 = add(x = Minv_187_cast_fp16, y = var_5493_cast_fp16)[name = string("Minv_189_cast_fp16")];
            bool Np_143_transpose_x_0 = const()[name = string("Np_143_transpose_x_0"), val = bool(false)];
            bool Np_143_transpose_y_0 = const()[name = string("Np_143_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_143_cast_fp16 = matmul(transpose_x = Np_143_transpose_x_0, transpose_y = Np_143_transpose_y_0, x = Np_141_cast_fp16, y = Npa_95_cast_fp16)[name = string("Np_143_cast_fp16")];
            bool var_5497_transpose_x_0 = const()[name = string("op_5497_transpose_x_0"), val = bool(false)];
            bool var_5497_transpose_y_0 = const()[name = string("op_5497_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_5497_cast_fp16 = matmul(transpose_x = var_5497_transpose_x_0, transpose_y = var_5497_transpose_y_0, x = Minv_189_cast_fp16, y = Np_143_cast_fp16)[name = string("op_5497_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_191_cast_fp16 = add(x = Minv_189_cast_fp16, y = var_5497_cast_fp16)[name = string("Minv_191_cast_fp16")];
            bool u_47_transpose_x_0 = const()[name = string("u_47_transpose_x_0"), val = bool(false)];
            bool u_47_transpose_y_0 = const()[name = string("u_47_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_47_cast_fp16 = matmul(transpose_x = u_47_transpose_x_0, transpose_y = u_47_transpose_y_0, x = Minv_191_cast_fp16, y = vb_55_cast_fp16)[name = string("u_47_cast_fp16")];
            bool var_5504_transpose_x_1 = const()[name = string("op_5504_transpose_x_1"), val = bool(false)];
            bool var_5504_transpose_y_1 = const()[name = string("op_5504_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_5504_cast_fp16 = matmul(transpose_x = var_5504_transpose_x_1, transpose_y = var_5504_transpose_y_1, x = q_81_cast_fp16, y = k_83_cast_fp16)[name = string("op_5504_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_47_cast_fp16 = mul(x = var_5504_cast_fp16, y = D_47_cast_fp16)[name = string("A_47_cast_fp16")];
            tensor<fp16, [16]> var_5506_bias_0_to_fp16 = const()[name = string("op_5506_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_5506_cast_fp16 = linear(bias = var_5506_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_55_cast_fp16)[name = string("op_5506_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_39_cast_fp16 = exp(x = var_5506_cast_fp16)[name = string("cum_39_cast_fp16")];
            tensor<int32, [1]> var_5509_axes_0 = const()[name = string("op_5509_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_5509_cast_fp16 = expand_dims(axes = var_5509_axes_0, x = cum_39_cast_fp16)[name = string("op_5509_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5510_cast_fp16 = mul(x = kb_47_cast_fp16, y = var_5509_cast_fp16)[name = string("op_5510_cast_fp16")];
            bool w_47_transpose_x_0 = const()[name = string("w_47_transpose_x_0"), val = bool(false)];
            bool w_47_transpose_y_0 = const()[name = string("w_47_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_47_cast_fp16 = matmul(transpose_x = w_47_transpose_x_0, transpose_y = w_47_transpose_y_0, x = Minv_191_cast_fp16, y = var_5510_cast_fp16)[name = string("w_47_cast_fp16")];
            bool var_5512_transpose_x_0 = const()[name = string("op_5512_transpose_x_0"), val = bool(false)];
            bool var_5512_transpose_y_0 = const()[name = string("op_5512_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5512_cast_fp16 = matmul(transpose_x = var_5512_transpose_x_0, transpose_y = var_5512_transpose_y_0, x = w_47_cast_fp16, y = S_39_cast_fp16)[name = string("op_5512_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_39_cast_fp16 = sub(x = u_47_cast_fp16, y = var_5512_cast_fp16)[name = string("vnew_39_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_5517_cast_fp16 = mul(x = q_81_cast_fp16, y = var_5509_cast_fp16)[name = string("op_5517_cast_fp16")];
            bool var_5518_transpose_x_0 = const()[name = string("op_5518_transpose_x_0"), val = bool(false)];
            bool var_5518_transpose_y_0 = const()[name = string("op_5518_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5518_cast_fp16 = matmul(transpose_x = var_5518_transpose_x_0, transpose_y = var_5518_transpose_y_0, x = var_5517_cast_fp16, y = S_39_cast_fp16)[name = string("op_5518_cast_fp16")];
            bool var_5519_transpose_x_0 = const()[name = string("op_5519_transpose_x_0"), val = bool(false)];
            bool var_5519_transpose_y_0 = const()[name = string("op_5519_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_5519_cast_fp16 = matmul(transpose_x = var_5519_transpose_x_0, transpose_y = var_5519_transpose_y_0, x = A_47_cast_fp16, y = vnew_39_cast_fp16)[name = string("op_5519_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_7_cast_fp16 = add(x = var_5518_cast_fp16, y = var_5519_cast_fp16)[name = string("oc_7_cast_fp16")];
            int32 var_5523 = const()[name = string("op_5523"), val = int32(2)];
            bool o_21_interleave_0 = const()[name = string("o_21_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 96, 128]> o_21_cast_fp16 = concat(axis = var_5523, interleave = o_21_interleave_0, values = (var_4758_cast_fp16, var_4905_cast_fp16, var_5059_cast_fp16, var_5213_cast_fp16, var_5367_cast_fp16, oc_7_cast_fp16))[name = string("o_21_cast_fp16")];
            tensor<int32, [4]> var_5529 = const()[name = string("op_5529"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_5535 = const()[name = string("op_5535"), val = tensor<int32, [4]>([3, 96, 4, 128])];
            tensor<fp16, [3, 96, 4, 128]> input_65_cast_fp16 = reshape(shape = var_5535, x = z_7_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> o_23_cast_fp16 = transpose(perm = var_5529, x = o_21_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [3, 96, 4, 128]> var_5537_cast_fp16 = mul(x = o_23_cast_fp16, y = o_23_cast_fp16)[name = string("op_5537_cast_fp16")];
            tensor<int32, [1]> var_5542_axes_0 = const()[name = string("op_5542_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5542_keep_dims_0 = const()[name = string("op_5542_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 4, 1]> var_5542_cast_fp16 = reduce_mean(axes = var_5542_axes_0, keep_dims = var_5542_keep_dims_0, x = var_5537_cast_fp16)[name = string("op_5542_cast_fp16")];
            fp16 var_5544_to_fp16 = const()[name = string("op_5544_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 96, 4, 1]> var_5545_cast_fp16 = add(x = var_5542_cast_fp16, y = var_5544_to_fp16)[name = string("op_5545_cast_fp16")];
            fp32 var_5546_epsilon_0 = const()[name = string("op_5546_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 4, 1]> var_5546_cast_fp16 = rsqrt(epsilon = var_5546_epsilon_0, x = var_5545_cast_fp16)[name = string("op_5546_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> var_5547_cast_fp16 = mul(x = o_23_cast_fp16, y = var_5546_cast_fp16)[name = string("op_5547_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> var_5549_cast_fp16 = silu(x = input_65_cast_fp16)[name = string("op_5549_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> o_25_cast_fp16 = mul(x = var_5547_cast_fp16, y = var_5549_cast_fp16)[name = string("o_25_cast_fp16")];
            tensor<int32, [3]> var_5554 = const()[name = string("op_5554"), val = tensor<int32, [3]>([3, 96, 512])];
            tensor<fp16, [3, 96, 512]> input_67_cast_fp16 = reshape(shape = var_5554, x = o_25_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [640, 512]> layers_4_out_weight_to_fp16 = const()[name = string("layers_4_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72063040)))];
            tensor<fp16, [3, 96, 640]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_out_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [3, 96, 640]> x_49_cast_fp16 = add(x = x_45_cast_fp16, y = linear_17_cast_fp16)[name = string("x_49_cast_fp16")];
            fp16 var_5567_to_fp16 = const()[name = string("op_5567_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 96, 640]> x_51_cast_fp16 = mul(x = x_49_cast_fp16, y = var_5567_to_fp16)[name = string("x_51_cast_fp16")];
            fp16 var_5565_promoted_to_fp16 = const()[name = string("op_5565_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> var_5569_cast_fp16 = pow(x = x_51_cast_fp16, y = var_5565_promoted_to_fp16)[name = string("op_5569_cast_fp16")];
            tensor<int32, [1]> var_5571_axes_0 = const()[name = string("op_5571_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5571_keep_dims_0 = const()[name = string("op_5571_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 1]> var_5571_cast_fp16 = reduce_mean(axes = var_5571_axes_0, keep_dims = var_5571_keep_dims_0, x = var_5569_cast_fp16)[name = string("op_5571_cast_fp16")];
            fp16 var_5572_to_fp16 = const()[name = string("op_5572_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 96, 1]> var_5573_cast_fp16 = add(x = var_5571_cast_fp16, y = var_5572_to_fp16)[name = string("op_5573_cast_fp16")];
            fp32 var_5574_epsilon_0 = const()[name = string("op_5574_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 1]> var_5574_cast_fp16 = rsqrt(epsilon = var_5574_epsilon_0, x = var_5573_cast_fp16)[name = string("op_5574_cast_fp16")];
            tensor<fp16, [3, 96, 640]> var_5575_cast_fp16 = mul(x = x_51_cast_fp16, y = var_5574_cast_fp16)[name = string("op_5575_cast_fp16")];
            tensor<fp16, [640]> layers_4_n2_w1_to_fp16 = const()[name = string("layers_4_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72718464)))];
            tensor<fp16, [3, 96, 640]> input_69_cast_fp16 = mul(x = var_5575_cast_fp16, y = layers_4_n2_w1_to_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_4_gate_up_weight_to_fp16 = const()[name = string("layers_4_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72719808)))];
            tensor<fp16, [3, 96, 5120]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_4_gate_up_weight_to_fp16, x = input_69_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<int32, [3]> input_71_begin_0 = const()[name = string("input_71_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_71_end_0 = const()[name = string("input_71_end_0"), val = tensor<int32, [3]>([3, 96, 2560])];
            tensor<bool, [3]> input_71_end_mask_0 = const()[name = string("input_71_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 2560]> input_71_cast_fp16 = slice_by_index(begin = input_71_begin_0, end = input_71_end_0, end_mask = input_71_end_mask_0, x = linear_18_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [3, 96, 2560]> var_5585_cast_fp16 = silu(x = input_71_cast_fp16)[name = string("op_5585_cast_fp16")];
            tensor<int32, [3]> var_5590_begin_0 = const()[name = string("op_5590_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_5590_end_0 = const()[name = string("op_5590_end_0"), val = tensor<int32, [3]>([3, 96, 5120])];
            tensor<bool, [3]> var_5590_end_mask_0 = const()[name = string("op_5590_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 2560]> var_5590_cast_fp16 = slice_by_index(begin = var_5590_begin_0, end = var_5590_end_0, end_mask = var_5590_end_mask_0, x = linear_18_cast_fp16)[name = string("op_5590_cast_fp16")];
            tensor<fp16, [3, 96, 2560]> input_73_cast_fp16 = mul(x = var_5585_cast_fp16, y = var_5590_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_4_down_weight_to_fp16 = const()[name = string("layers_4_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79273472)))];
            tensor<fp16, [3, 96, 640]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_down_weight_to_fp16, x = input_73_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [3, 96, 640]> x_53_cast_fp16 = add(x = x_49_cast_fp16, y = linear_19_cast_fp16)[name = string("x_53_cast_fp16")];
            fp16 var_5603_to_fp16 = const()[name = string("op_5603_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 96, 640]> x_55_cast_fp16 = mul(x = x_53_cast_fp16, y = var_5603_to_fp16)[name = string("x_55_cast_fp16")];
            fp16 var_5601_promoted_to_fp16 = const()[name = string("op_5601_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> var_5605_cast_fp16 = pow(x = x_55_cast_fp16, y = var_5601_promoted_to_fp16)[name = string("op_5605_cast_fp16")];
            tensor<int32, [1]> var_5607_axes_0 = const()[name = string("op_5607_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5607_keep_dims_0 = const()[name = string("op_5607_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 1]> var_5607_cast_fp16 = reduce_mean(axes = var_5607_axes_0, keep_dims = var_5607_keep_dims_0, x = var_5605_cast_fp16)[name = string("op_5607_cast_fp16")];
            fp16 var_5608_to_fp16 = const()[name = string("op_5608_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 96, 1]> var_5609_cast_fp16 = add(x = var_5607_cast_fp16, y = var_5608_to_fp16)[name = string("op_5609_cast_fp16")];
            fp32 var_5610_epsilon_0 = const()[name = string("op_5610_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 1]> var_5610_cast_fp16 = rsqrt(epsilon = var_5610_epsilon_0, x = var_5609_cast_fp16)[name = string("op_5610_cast_fp16")];
            tensor<fp16, [3, 96, 640]> var_5611_cast_fp16 = mul(x = x_55_cast_fp16, y = var_5610_cast_fp16)[name = string("op_5611_cast_fp16")];
            tensor<fp16, [640]> layers_5_n1_w1_to_fp16 = const()[name = string("layers_5_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82550336)))];
            tensor<fp16, [3, 96, 640]> input_75_cast_fp16 = mul(x = var_5611_cast_fp16, y = layers_5_n1_w1_to_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_5_qkvzba_weight_to_fp16 = const()[name = string("layers_5_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82551680)))];
            tensor<fp16, [3, 96, 1544]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_qkvzba_weight_to_fp16, x = input_75_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [3]> qkv_9_begin_0 = const()[name = string("qkv_9_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_9_end_0 = const()[name = string("qkv_9_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> qkv_9_end_mask_0 = const()[name = string("qkv_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 1024]> qkv_9_cast_fp16 = slice_by_index(begin = qkv_9_begin_0, end = qkv_9_end_0, end_mask = qkv_9_end_mask_0, x = linear_20_cast_fp16)[name = string("qkv_9_cast_fp16")];
            tensor<int32, [3]> z_9_begin_0 = const()[name = string("z_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_9_end_0 = const()[name = string("z_9_end_0"), val = tensor<int32, [3]>([3, 96, 1536])];
            tensor<bool, [3]> z_9_end_mask_0 = const()[name = string("z_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 512]> z_9_cast_fp16 = slice_by_index(begin = z_9_begin_0, end = z_9_end_0, end_mask = z_9_end_mask_0, x = linear_20_cast_fp16)[name = string("z_9_cast_fp16")];
            tensor<int32, [3]> b_9_begin_0 = const()[name = string("b_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_9_end_0 = const()[name = string("b_9_end_0"), val = tensor<int32, [3]>([3, 96, 1540])];
            tensor<bool, [3]> b_9_end_mask_0 = const()[name = string("b_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 4]> b_9_cast_fp16 = slice_by_index(begin = b_9_begin_0, end = b_9_end_0, end_mask = b_9_end_mask_0, x = linear_20_cast_fp16)[name = string("b_9_cast_fp16")];
            tensor<int32, [3]> a_9_begin_0 = const()[name = string("a_9_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_9_end_0 = const()[name = string("a_9_end_0"), val = tensor<int32, [3]>([3, 96, 1])];
            tensor<bool, [3]> a_9_end_mask_0 = const()[name = string("a_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 4]> a_9_cast_fp16 = slice_by_index(begin = a_9_begin_0, end = a_9_end_0, end_mask = a_9_end_mask_0, x = linear_20_cast_fp16)[name = string("a_9_cast_fp16")];
            int32 var_5650 = const()[name = string("op_5650"), val = int32(1)];
            bool xp_9_interleave_0 = const()[name = string("xp_9_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 99, 1024]> xp_9_cast_fp16 = concat(axis = var_5650, interleave = xp_9_interleave_0, values = (zpad_to_fp16, qkv_9_cast_fp16))[name = string("xp_9_cast_fp16")];
            tensor<int32, [3]> var_5661_begin_0 = const()[name = string("op_5661_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_5661_end_0 = const()[name = string("op_5661_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_5661_end_mask_0 = const()[name = string("op_5661_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 1024]> var_5661_cast_fp16 = slice_by_index(begin = var_5661_begin_0, end = var_5661_end_0, end_mask = var_5661_end_mask_0, x = xp_9_cast_fp16)[name = string("op_5661_cast_fp16")];
            tensor<fp16, [1024]> var_5669_to_fp16 = const()[name = string("op_5669_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84528064)))];
            tensor<fp16, [3, 96, 1024]> var_5670_cast_fp16 = mul(x = var_5661_cast_fp16, y = var_5669_to_fp16)[name = string("op_5670_cast_fp16")];
            tensor<int32, [3]> var_5680_begin_0 = const()[name = string("op_5680_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_5680_end_0 = const()[name = string("op_5680_end_0"), val = tensor<int32, [3]>([3, 98, 1024])];
            tensor<bool, [3]> var_5680_end_mask_0 = const()[name = string("op_5680_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_5680_cast_fp16 = slice_by_index(begin = var_5680_begin_0, end = var_5680_end_0, end_mask = var_5680_end_mask_0, x = xp_9_cast_fp16)[name = string("op_5680_cast_fp16")];
            tensor<fp16, [1024]> var_5688_to_fp16 = const()[name = string("op_5688_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84530176)))];
            tensor<fp16, [3, 96, 1024]> var_5689_cast_fp16 = mul(x = var_5680_cast_fp16, y = var_5688_to_fp16)[name = string("op_5689_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> var_5691_cast_fp16 = add(x = var_5670_cast_fp16, y = var_5689_cast_fp16)[name = string("op_5691_cast_fp16")];
            tensor<int32, [3]> var_5701_begin_0 = const()[name = string("op_5701_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_5701_end_0 = const()[name = string("op_5701_end_0"), val = tensor<int32, [3]>([3, 97, 1024])];
            tensor<bool, [3]> var_5701_end_mask_0 = const()[name = string("op_5701_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_5701_cast_fp16 = slice_by_index(begin = var_5701_begin_0, end = var_5701_end_0, end_mask = var_5701_end_mask_0, x = xp_9_cast_fp16)[name = string("op_5701_cast_fp16")];
            tensor<fp16, [1024]> var_5709_to_fp16 = const()[name = string("op_5709_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84532288)))];
            tensor<fp16, [3, 96, 1024]> var_5710_cast_fp16 = mul(x = var_5701_cast_fp16, y = var_5709_to_fp16)[name = string("op_5710_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> var_5712_cast_fp16 = add(x = var_5691_cast_fp16, y = var_5710_cast_fp16)[name = string("op_5712_cast_fp16")];
            tensor<int32, [3]> var_5722_begin_0 = const()[name = string("op_5722_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_5722_end_0 = const()[name = string("op_5722_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> var_5722_end_mask_0 = const()[name = string("op_5722_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_5722_cast_fp16 = slice_by_index(begin = var_5722_begin_0, end = var_5722_end_0, end_mask = var_5722_end_mask_0, x = xp_9_cast_fp16)[name = string("op_5722_cast_fp16")];
            tensor<fp16, [1024]> var_5730_to_fp16 = const()[name = string("op_5730_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84534400)))];
            tensor<fp16, [3, 96, 1024]> var_5731_cast_fp16 = mul(x = var_5722_cast_fp16, y = var_5730_to_fp16)[name = string("op_5731_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> input_77_cast_fp16 = add(x = var_5712_cast_fp16, y = var_5731_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> c_9_cast_fp16 = silu(x = input_77_cast_fp16)[name = string("c_9_cast_fp16")];
            tensor<int32, [3]> var_5739_begin_0 = const()[name = string("op_5739_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_5739_end_0 = const()[name = string("op_5739_end_0"), val = tensor<int32, [3]>([3, 96, 256])];
            tensor<bool, [3]> var_5739_end_mask_0 = const()[name = string("op_5739_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 256]> var_5739_cast_fp16 = slice_by_index(begin = var_5739_begin_0, end = var_5739_end_0, end_mask = var_5739_end_mask_0, x = c_9_cast_fp16)[name = string("op_5739_cast_fp16")];
            tensor<int32, [4]> var_5744 = const()[name = string("op_5744"), val = tensor<int32, [4]>([3, 96, 2, 128])];
            tensor<fp16, [3, 96, 2, 128]> q_83_cast_fp16 = reshape(shape = var_5744, x = var_5739_cast_fp16)[name = string("q_83_cast_fp16")];
            tensor<int32, [3]> var_5750_begin_0 = const()[name = string("op_5750_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_5750_end_0 = const()[name = string("op_5750_end_0"), val = tensor<int32, [3]>([3, 96, 512])];
            tensor<bool, [3]> var_5750_end_mask_0 = const()[name = string("op_5750_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 256]> var_5750_cast_fp16 = slice_by_index(begin = var_5750_begin_0, end = var_5750_end_0, end_mask = var_5750_end_mask_0, x = c_9_cast_fp16)[name = string("op_5750_cast_fp16")];
            tensor<int32, [4]> var_5755 = const()[name = string("op_5755"), val = tensor<int32, [4]>([3, 96, 2, 128])];
            tensor<fp16, [3, 96, 2, 128]> k_85_cast_fp16 = reshape(shape = var_5755, x = var_5750_cast_fp16)[name = string("k_85_cast_fp16")];
            tensor<int32, [3]> var_5761_begin_0 = const()[name = string("op_5761_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_5761_end_0 = const()[name = string("op_5761_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> var_5761_end_mask_0 = const()[name = string("op_5761_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 512]> var_5761_cast_fp16 = slice_by_index(begin = var_5761_begin_0, end = var_5761_end_0, end_mask = var_5761_end_mask_0, x = c_9_cast_fp16)[name = string("op_5761_cast_fp16")];
            tensor<int32, [4]> var_5766 = const()[name = string("op_5766"), val = tensor<int32, [4]>([3, 96, 4, 128])];
            tensor<fp16, [3, 96, 4, 128]> var_5767_cast_fp16 = reshape(shape = var_5766, x = var_5761_cast_fp16)[name = string("op_5767_cast_fp16")];
            tensor<int32, [4]> var_5772 = const()[name = string("op_5772"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5774_to_fp16 = const()[name = string("op_5774_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 96, 2, 128]> q_85_cast_fp16 = mul(x = q_83_cast_fp16, y = var_5774_to_fp16)[name = string("q_85_cast_fp16")];
            fp16 var_5776_to_fp16 = const()[name = string("op_5776_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 96, 2, 128]> k_87_cast_fp16 = mul(x = k_85_cast_fp16, y = var_5776_to_fp16)[name = string("k_87_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_5778_cast_fp16 = mul(x = q_85_cast_fp16, y = q_85_cast_fp16)[name = string("op_5778_cast_fp16")];
            tensor<int32, [1]> var_5783_axes_0 = const()[name = string("op_5783_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5783_keep_dims_0 = const()[name = string("op_5783_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 2, 1]> var_5783_cast_fp16 = reduce_mean(axes = var_5783_axes_0, keep_dims = var_5783_keep_dims_0, x = var_5778_cast_fp16)[name = string("op_5783_cast_fp16")];
            fp16 var_5784_promoted_to_fp16 = const()[name = string("op_5784_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 96, 2, 1]> var_5785_cast_fp16 = mul(x = var_5783_cast_fp16, y = var_5784_promoted_to_fp16)[name = string("op_5785_cast_fp16")];
            fp16 var_5787_to_fp16 = const()[name = string("op_5787_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 96, 2, 1]> var_5788_cast_fp16 = add(x = var_5785_cast_fp16, y = var_5787_to_fp16)[name = string("op_5788_cast_fp16")];
            fp32 var_5789_epsilon_0 = const()[name = string("op_5789_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 2, 1]> var_5789_cast_fp16 = rsqrt(epsilon = var_5789_epsilon_0, x = var_5788_cast_fp16)[name = string("op_5789_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_5790_cast_fp16 = mul(x = q_85_cast_fp16, y = var_5789_cast_fp16)[name = string("op_5790_cast_fp16")];
            fp16 var_5791_to_fp16 = const()[name = string("op_5791_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 96, 2, 128]> q_87_cast_fp16 = mul(x = var_5790_cast_fp16, y = var_5791_to_fp16)[name = string("q_87_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_5793_cast_fp16 = mul(x = k_87_cast_fp16, y = k_87_cast_fp16)[name = string("op_5793_cast_fp16")];
            tensor<int32, [1]> var_5798_axes_0 = const()[name = string("op_5798_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5798_keep_dims_0 = const()[name = string("op_5798_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 2, 1]> var_5798_cast_fp16 = reduce_mean(axes = var_5798_axes_0, keep_dims = var_5798_keep_dims_0, x = var_5793_cast_fp16)[name = string("op_5798_cast_fp16")];
            fp16 var_5799_promoted_to_fp16 = const()[name = string("op_5799_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 96, 2, 1]> var_5800_cast_fp16 = mul(x = var_5798_cast_fp16, y = var_5799_promoted_to_fp16)[name = string("op_5800_cast_fp16")];
            fp16 var_5802_to_fp16 = const()[name = string("op_5802_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 96, 2, 1]> var_5803_cast_fp16 = add(x = var_5800_cast_fp16, y = var_5802_to_fp16)[name = string("op_5803_cast_fp16")];
            fp32 var_5804_epsilon_0 = const()[name = string("op_5804_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 2, 1]> var_5804_cast_fp16 = rsqrt(epsilon = var_5804_epsilon_0, x = var_5803_cast_fp16)[name = string("op_5804_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> k_89_cast_fp16 = mul(x = k_87_cast_fp16, y = var_5804_cast_fp16)[name = string("k_89_cast_fp16")];
            tensor<int32, [4]> var_5810 = const()[name = string("op_5810"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_5823_axes_0 = const()[name = string("op_5823_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 96, 128]> var_5811_cast_fp16 = transpose(perm = var_5810, x = q_87_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [3, 2, 1, 96, 128]> var_5823_cast_fp16 = expand_dims(axes = var_5823_axes_0, x = var_5811_cast_fp16)[name = string("op_5823_cast_fp16")];
            tensor<int32, [5]> var_5831_reps_0 = const()[name = string("op_5831_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 96, 128]> var_5831_cast_fp16 = tile(reps = var_5831_reps_0, x = var_5823_cast_fp16)[name = string("op_5831_cast_fp16")];
            tensor<int32, [4]> var_5836 = const()[name = string("op_5836"), val = tensor<int32, [4]>([3, 4, 96, 128])];
            tensor<fp16, [3, 4, 96, 128]> q_89_cast_fp16 = reshape(shape = var_5836, x = var_5831_cast_fp16)[name = string("q_89_cast_fp16")];
            tensor<int32, [4]> var_5842 = const()[name = string("op_5842"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_5855_axes_0 = const()[name = string("op_5855_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 96, 128]> var_5843_cast_fp16 = transpose(perm = var_5842, x = k_89_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [3, 2, 1, 96, 128]> var_5855_cast_fp16 = expand_dims(axes = var_5855_axes_0, x = var_5843_cast_fp16)[name = string("op_5855_cast_fp16")];
            tensor<int32, [5]> var_5863_reps_0 = const()[name = string("op_5863_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 96, 128]> var_5863_cast_fp16 = tile(reps = var_5863_reps_0, x = var_5855_cast_fp16)[name = string("op_5863_cast_fp16")];
            tensor<int32, [4]> var_5868 = const()[name = string("op_5868"), val = tensor<int32, [4]>([3, 4, 96, 128])];
            tensor<fp16, [3, 4, 96, 128]> k_91_cast_fp16 = reshape(shape = var_5868, x = var_5863_cast_fp16)[name = string("k_91_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_5870_cast_fp16 = sigmoid(x = b_9_cast_fp16)[name = string("op_5870_cast_fp16")];
            tensor<int32, [3]> var_5874 = const()[name = string("op_5874"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_5_dt_bias_to_fp16 = const()[name = string("layers_5_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 96, 4]> input_79_cast_fp16 = add(x = a_9_cast_fp16, y = layers_5_dt_bias_to_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_5878_cast_fp16 = relu(x = input_79_cast_fp16)[name = string("op_5878_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_5879_cast_fp16 = abs(x = input_79_cast_fp16)[name = string("op_5879_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 96, 4]> var_5880_cast_fp16 = mul(x = var_5879_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_5880_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_5881_cast_fp16 = exp(x = var_5880_cast_fp16)[name = string("op_5881_cast_fp16")];
            fp16 var_5883_to_fp16 = const()[name = string("op_5883_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 96, 4]> var_5884_cast_fp16 = add(x = var_5881_cast_fp16, y = var_5883_to_fp16)[name = string("op_5884_cast_fp16")];
            fp32 var_5885_epsilon_0 = const()[name = string("op_5885_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 96, 4]> var_5885_cast_fp16 = log(epsilon = var_5885_epsilon_0, x = var_5884_cast_fp16)[name = string("op_5885_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_5887_cast_fp16 = add(x = var_5878_cast_fp16, y = var_5885_cast_fp16)[name = string("op_5887_cast_fp16")];
            tensor<fp16, [4]> layers_5_negA_to_fp16 = const()[name = string("layers_5_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.2f8p+3, -0x1.ae8p+2, -0x1.548p+2, -0x1.63p+3])];
            tensor<fp16, [3, 96, 4]> var_5888_cast_fp16 = mul(x = layers_5_negA_to_fp16, y = var_5887_cast_fp16)[name = string("op_5888_cast_fp16")];
            tensor<int32, [3]> var_5892 = const()[name = string("op_5892"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_5895_axes_0 = const()[name = string("op_5895_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 96]> beta_9_cast_fp16 = transpose(perm = var_5874, x = var_5870_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [3, 4, 96, 1]> var_5895_cast_fp16 = expand_dims(axes = var_5895_axes_0, x = beta_9_cast_fp16)[name = string("op_5895_cast_fp16")];
            fp16 var_5896_to_fp16 = const()[name = string("op_5896_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 96, 1]> var_5897_cast_fp16 = mul(x = var_5895_cast_fp16, y = var_5896_to_fp16)[name = string("op_5897_cast_fp16")];
            tensor<fp16, [3, 4, 96, 128]> v_13_cast_fp16 = transpose(perm = var_5772, x = var_5767_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [3, 4, 96, 128]> vb_57_cast_fp16 = mul(x = v_13_cast_fp16, y = var_5897_cast_fp16)[name = string("vb_57_cast_fp16")];
            tensor<int32, [4]> q_91_begin_0 = const()[name = string("q_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_91_end_0 = const()[name = string("q_91_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_91_end_mask_0 = const()[name = string("q_91_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, x = q_89_cast_fp16)[name = string("q_91_cast_fp16")];
            tensor<int32, [4]> k_93_begin_0 = const()[name = string("k_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_93_end_0 = const()[name = string("k_93_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_93_end_mask_0 = const()[name = string("k_93_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, x = k_91_cast_fp16)[name = string("k_93_cast_fp16")];
            tensor<fp16, [3, 4, 96, 128]> var_5931_cast_fp16 = mul(x = k_91_cast_fp16, y = var_5895_cast_fp16)[name = string("op_5931_cast_fp16")];
            tensor<int32, [4]> kb_49_begin_0 = const()[name = string("kb_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_49_end_0 = const()[name = string("kb_49_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_49_end_mask_0 = const()[name = string("kb_49_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_49_cast_fp16 = slice_by_index(begin = kb_49_begin_0, end = kb_49_end_0, end_mask = kb_49_end_mask_0, x = var_5931_cast_fp16)[name = string("kb_49_cast_fp16")];
            tensor<int32, [4]> vb_59_begin_0 = const()[name = string("vb_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_59_end_0 = const()[name = string("vb_59_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_59_end_mask_0 = const()[name = string("vb_59_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_59_cast_fp16 = slice_by_index(begin = vb_59_begin_0, end = vb_59_end_0, end_mask = vb_59_end_mask_0, x = vb_57_cast_fp16)[name = string("vb_59_cast_fp16")];
            tensor<int32, [3]> gdec_59_begin_0 = const()[name = string("gdec_59_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_59_end_0 = const()[name = string("gdec_59_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_59_end_mask_0 = const()[name = string("gdec_59_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 96]> gdec_57_cast_fp16 = transpose(perm = var_5892, x = var_5888_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [3, 4, 16]> gdec_59_cast_fp16 = slice_by_index(begin = gdec_59_begin_0, end = gdec_59_end_0, end_mask = gdec_59_end_mask_0, x = gdec_57_cast_fp16)[name = string("gdec_59_cast_fp16")];
            tensor<int32, [2]> var_5979 = const()[name = string("op_5979"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_5980_cast_fp16 = reshape(shape = var_5979, x = gdec_59_cast_fp16)[name = string("op_5980_cast_fp16")];
            tensor<fp16, [12, 256]> var_5981_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_5980_cast_fp16)[name = string("op_5981_cast_fp16")];
            tensor<int32, [4]> var_5986 = const()[name = string("op_5986"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_49_cast_fp16 = reshape(shape = var_5986, x = var_5981_cast_fp16)[name = string("P_49_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5988_cast_fp16 = exp(x = P_49_cast_fp16)[name = string("op_5988_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_49_cast_fp16 = mul(x = var_5988_cast_fp16, y = tril_incl_to_fp16)[name = string("D_49_cast_fp16")];
            bool kk_49_transpose_x_1 = const()[name = string("kk_49_transpose_x_1"), val = bool(false)];
            bool kk_49_transpose_y_1 = const()[name = string("kk_49_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_49_cast_fp16 = matmul(transpose_x = kk_49_transpose_x_1, transpose_y = kk_49_transpose_y_1, x = kb_49_cast_fp16, y = k_93_cast_fp16)[name = string("kk_49_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_5994_cast_fp16 = mul(x = kk_49_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_5994_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5995_cast_fp16 = mul(x = var_5994_cast_fp16, y = D_49_cast_fp16)[name = string("op_5995_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_49_cast_fp16 = mul(x = var_5995_cast_fp16, y = cum_tail_to_fp16)[name = string("N_49_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_5997_cast_fp16 = mul(x = D_49_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_5997_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_49_cast_fp16 = mul(x = kk_49_cast_fp16, y = var_5997_cast_fp16)[name = string("Na_49_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_193_cast_fp16 = add(x = eye_to_fp16, y = N_49_cast_fp16)[name = string("Minv_193_cast_fp16")];
            bool Np_145_transpose_x_0 = const()[name = string("Np_145_transpose_x_0"), val = bool(false)];
            bool Np_145_transpose_y_0 = const()[name = string("Np_145_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_145_cast_fp16 = matmul(transpose_x = Np_145_transpose_x_0, transpose_y = Np_145_transpose_y_0, x = N_49_cast_fp16, y = Na_49_cast_fp16)[name = string("Np_145_cast_fp16")];
            bool Npa_97_transpose_x_0 = const()[name = string("Npa_97_transpose_x_0"), val = bool(false)];
            bool Npa_97_transpose_y_0 = const()[name = string("Npa_97_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_97_cast_fp16 = matmul(transpose_x = Npa_97_transpose_x_0, transpose_y = Npa_97_transpose_y_0, x = Na_49_cast_fp16, y = N_49_cast_fp16)[name = string("Npa_97_cast_fp16")];
            bool var_6003_transpose_x_0 = const()[name = string("op_6003_transpose_x_0"), val = bool(false)];
            bool var_6003_transpose_y_0 = const()[name = string("op_6003_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6003_cast_fp16 = matmul(transpose_x = var_6003_transpose_x_0, transpose_y = var_6003_transpose_y_0, x = Minv_193_cast_fp16, y = Np_145_cast_fp16)[name = string("op_6003_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_195_cast_fp16 = add(x = Minv_193_cast_fp16, y = var_6003_cast_fp16)[name = string("Minv_195_cast_fp16")];
            bool Np_147_transpose_x_0 = const()[name = string("Np_147_transpose_x_0"), val = bool(false)];
            bool Np_147_transpose_y_0 = const()[name = string("Np_147_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_147_cast_fp16 = matmul(transpose_x = Np_147_transpose_x_0, transpose_y = Np_147_transpose_y_0, x = Np_145_cast_fp16, y = Npa_97_cast_fp16)[name = string("Np_147_cast_fp16")];
            bool Npa_99_transpose_x_0 = const()[name = string("Npa_99_transpose_x_0"), val = bool(false)];
            bool Npa_99_transpose_y_0 = const()[name = string("Npa_99_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_99_cast_fp16 = matmul(transpose_x = Npa_99_transpose_x_0, transpose_y = Npa_99_transpose_y_0, x = Npa_97_cast_fp16, y = Np_145_cast_fp16)[name = string("Npa_99_cast_fp16")];
            bool var_6008_transpose_x_0 = const()[name = string("op_6008_transpose_x_0"), val = bool(false)];
            bool var_6008_transpose_y_0 = const()[name = string("op_6008_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6008_cast_fp16 = matmul(transpose_x = var_6008_transpose_x_0, transpose_y = var_6008_transpose_y_0, x = Minv_195_cast_fp16, y = Np_147_cast_fp16)[name = string("op_6008_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_197_cast_fp16 = add(x = Minv_195_cast_fp16, y = var_6008_cast_fp16)[name = string("Minv_197_cast_fp16")];
            bool Np_149_transpose_x_0 = const()[name = string("Np_149_transpose_x_0"), val = bool(false)];
            bool Np_149_transpose_y_0 = const()[name = string("Np_149_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_149_cast_fp16 = matmul(transpose_x = Np_149_transpose_x_0, transpose_y = Np_149_transpose_y_0, x = Np_147_cast_fp16, y = Npa_99_cast_fp16)[name = string("Np_149_cast_fp16")];
            bool var_6012_transpose_x_0 = const()[name = string("op_6012_transpose_x_0"), val = bool(false)];
            bool var_6012_transpose_y_0 = const()[name = string("op_6012_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6012_cast_fp16 = matmul(transpose_x = var_6012_transpose_x_0, transpose_y = var_6012_transpose_y_0, x = Minv_197_cast_fp16, y = Np_149_cast_fp16)[name = string("op_6012_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_199_cast_fp16 = add(x = Minv_197_cast_fp16, y = var_6012_cast_fp16)[name = string("Minv_199_cast_fp16")];
            bool u_49_transpose_x_0 = const()[name = string("u_49_transpose_x_0"), val = bool(false)];
            bool u_49_transpose_y_0 = const()[name = string("u_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_49_cast_fp16 = matmul(transpose_x = u_49_transpose_x_0, transpose_y = u_49_transpose_y_0, x = Minv_199_cast_fp16, y = vb_59_cast_fp16)[name = string("u_49_cast_fp16")];
            bool var_6019_transpose_x_1 = const()[name = string("op_6019_transpose_x_1"), val = bool(false)];
            bool var_6019_transpose_y_1 = const()[name = string("op_6019_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_6019_cast_fp16 = matmul(transpose_x = var_6019_transpose_x_1, transpose_y = var_6019_transpose_y_1, x = q_91_cast_fp16, y = k_93_cast_fp16)[name = string("op_6019_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_49_cast_fp16 = mul(x = var_6019_cast_fp16, y = D_49_cast_fp16)[name = string("A_49_cast_fp16")];
            bool var_6021_transpose_x_0 = const()[name = string("op_6021_transpose_x_0"), val = bool(false)];
            bool var_6021_transpose_y_0 = const()[name = string("op_6021_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6021_cast_fp16 = matmul(transpose_x = var_6021_transpose_x_0, transpose_y = var_6021_transpose_y_0, x = A_49_cast_fp16, y = u_49_cast_fp16)[name = string("op_6021_cast_fp16")];
            tensor<fp16, [16]> var_6022_bias_0_to_fp16 = const()[name = string("op_6022_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_6022_cast_fp16 = linear(bias = var_6022_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_59_cast_fp16)[name = string("op_6022_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_41_cast_fp16 = exp(x = var_6022_cast_fp16)[name = string("tail_41_cast_fp16")];
            tensor<int32, [1]> var_6025_axes_0 = const()[name = string("op_6025_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_6025_cast_fp16 = expand_dims(axes = var_6025_axes_0, x = tail_41_cast_fp16)[name = string("op_6025_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6026_cast_fp16 = mul(x = k_93_cast_fp16, y = var_6025_cast_fp16)[name = string("op_6026_cast_fp16")];
            bool S_41_transpose_x_1 = const()[name = string("S_41_transpose_x_1"), val = bool(true)];
            bool S_41_transpose_y_1 = const()[name = string("S_41_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_41_cast_fp16 = matmul(transpose_x = S_41_transpose_x_1, transpose_y = S_41_transpose_y_1, x = var_6026_cast_fp16, y = u_49_cast_fp16)[name = string("S_41_cast_fp16")];
            tensor<int32, [4]> q_93_begin_0 = const()[name = string("q_93_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_93_end_0 = const()[name = string("q_93_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> q_93_end_mask_0 = const()[name = string("q_93_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_93_cast_fp16 = slice_by_index(begin = q_93_begin_0, end = q_93_end_0, end_mask = q_93_end_mask_0, x = q_89_cast_fp16)[name = string("q_93_cast_fp16")];
            tensor<int32, [4]> k_95_begin_0 = const()[name = string("k_95_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_95_end_0 = const()[name = string("k_95_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> k_95_end_mask_0 = const()[name = string("k_95_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_95_cast_fp16 = slice_by_index(begin = k_95_begin_0, end = k_95_end_0, end_mask = k_95_end_mask_0, x = k_91_cast_fp16)[name = string("k_95_cast_fp16")];
            tensor<int32, [4]> kb_51_begin_0 = const()[name = string("kb_51_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_51_end_0 = const()[name = string("kb_51_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> kb_51_end_mask_0 = const()[name = string("kb_51_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_51_cast_fp16 = slice_by_index(begin = kb_51_begin_0, end = kb_51_end_0, end_mask = kb_51_end_mask_0, x = var_5931_cast_fp16)[name = string("kb_51_cast_fp16")];
            tensor<int32, [4]> vb_61_begin_0 = const()[name = string("vb_61_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_61_end_0 = const()[name = string("vb_61_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> vb_61_end_mask_0 = const()[name = string("vb_61_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_61_cast_fp16 = slice_by_index(begin = vb_61_begin_0, end = vb_61_end_0, end_mask = vb_61_end_mask_0, x = vb_57_cast_fp16)[name = string("vb_61_cast_fp16")];
            tensor<int32, [3]> gdec_61_begin_0 = const()[name = string("gdec_61_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_61_end_0 = const()[name = string("gdec_61_end_0"), val = tensor<int32, [3]>([3, 4, 32])];
            tensor<bool, [3]> gdec_61_end_mask_0 = const()[name = string("gdec_61_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_61_cast_fp16 = slice_by_index(begin = gdec_61_begin_0, end = gdec_61_end_0, end_mask = gdec_61_end_mask_0, x = gdec_57_cast_fp16)[name = string("gdec_61_cast_fp16")];
            tensor<int32, [2]> var_6111 = const()[name = string("op_6111"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_6112_cast_fp16 = reshape(shape = var_6111, x = gdec_61_cast_fp16)[name = string("op_6112_cast_fp16")];
            tensor<fp16, [12, 256]> var_6113_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_6112_cast_fp16)[name = string("op_6113_cast_fp16")];
            tensor<int32, [4]> var_6118 = const()[name = string("op_6118"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_51_cast_fp16 = reshape(shape = var_6118, x = var_6113_cast_fp16)[name = string("P_51_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6120_cast_fp16 = exp(x = P_51_cast_fp16)[name = string("op_6120_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_51_cast_fp16 = mul(x = var_6120_cast_fp16, y = tril_incl_to_fp16)[name = string("D_51_cast_fp16")];
            bool kk_51_transpose_x_1 = const()[name = string("kk_51_transpose_x_1"), val = bool(false)];
            bool kk_51_transpose_y_1 = const()[name = string("kk_51_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_51_cast_fp16 = matmul(transpose_x = kk_51_transpose_x_1, transpose_y = kk_51_transpose_y_1, x = kb_51_cast_fp16, y = k_95_cast_fp16)[name = string("kk_51_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_6126_cast_fp16 = mul(x = kk_51_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_6126_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6127_cast_fp16 = mul(x = var_6126_cast_fp16, y = D_51_cast_fp16)[name = string("op_6127_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_51_cast_fp16 = mul(x = var_6127_cast_fp16, y = cum_tail_to_fp16)[name = string("N_51_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6129_cast_fp16 = mul(x = D_51_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_6129_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_51_cast_fp16 = mul(x = kk_51_cast_fp16, y = var_6129_cast_fp16)[name = string("Na_51_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_201_cast_fp16 = add(x = eye_to_fp16, y = N_51_cast_fp16)[name = string("Minv_201_cast_fp16")];
            bool Np_151_transpose_x_0 = const()[name = string("Np_151_transpose_x_0"), val = bool(false)];
            bool Np_151_transpose_y_0 = const()[name = string("Np_151_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_151_cast_fp16 = matmul(transpose_x = Np_151_transpose_x_0, transpose_y = Np_151_transpose_y_0, x = N_51_cast_fp16, y = Na_51_cast_fp16)[name = string("Np_151_cast_fp16")];
            bool Npa_101_transpose_x_0 = const()[name = string("Npa_101_transpose_x_0"), val = bool(false)];
            bool Npa_101_transpose_y_0 = const()[name = string("Npa_101_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_101_cast_fp16 = matmul(transpose_x = Npa_101_transpose_x_0, transpose_y = Npa_101_transpose_y_0, x = Na_51_cast_fp16, y = N_51_cast_fp16)[name = string("Npa_101_cast_fp16")];
            bool var_6135_transpose_x_0 = const()[name = string("op_6135_transpose_x_0"), val = bool(false)];
            bool var_6135_transpose_y_0 = const()[name = string("op_6135_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6135_cast_fp16 = matmul(transpose_x = var_6135_transpose_x_0, transpose_y = var_6135_transpose_y_0, x = Minv_201_cast_fp16, y = Np_151_cast_fp16)[name = string("op_6135_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_203_cast_fp16 = add(x = Minv_201_cast_fp16, y = var_6135_cast_fp16)[name = string("Minv_203_cast_fp16")];
            bool Np_153_transpose_x_0 = const()[name = string("Np_153_transpose_x_0"), val = bool(false)];
            bool Np_153_transpose_y_0 = const()[name = string("Np_153_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_153_cast_fp16 = matmul(transpose_x = Np_153_transpose_x_0, transpose_y = Np_153_transpose_y_0, x = Np_151_cast_fp16, y = Npa_101_cast_fp16)[name = string("Np_153_cast_fp16")];
            bool Npa_103_transpose_x_0 = const()[name = string("Npa_103_transpose_x_0"), val = bool(false)];
            bool Npa_103_transpose_y_0 = const()[name = string("Npa_103_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_103_cast_fp16 = matmul(transpose_x = Npa_103_transpose_x_0, transpose_y = Npa_103_transpose_y_0, x = Npa_101_cast_fp16, y = Np_151_cast_fp16)[name = string("Npa_103_cast_fp16")];
            bool var_6140_transpose_x_0 = const()[name = string("op_6140_transpose_x_0"), val = bool(false)];
            bool var_6140_transpose_y_0 = const()[name = string("op_6140_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6140_cast_fp16 = matmul(transpose_x = var_6140_transpose_x_0, transpose_y = var_6140_transpose_y_0, x = Minv_203_cast_fp16, y = Np_153_cast_fp16)[name = string("op_6140_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_205_cast_fp16 = add(x = Minv_203_cast_fp16, y = var_6140_cast_fp16)[name = string("Minv_205_cast_fp16")];
            bool Np_155_transpose_x_0 = const()[name = string("Np_155_transpose_x_0"), val = bool(false)];
            bool Np_155_transpose_y_0 = const()[name = string("Np_155_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_155_cast_fp16 = matmul(transpose_x = Np_155_transpose_x_0, transpose_y = Np_155_transpose_y_0, x = Np_153_cast_fp16, y = Npa_103_cast_fp16)[name = string("Np_155_cast_fp16")];
            bool var_6144_transpose_x_0 = const()[name = string("op_6144_transpose_x_0"), val = bool(false)];
            bool var_6144_transpose_y_0 = const()[name = string("op_6144_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6144_cast_fp16 = matmul(transpose_x = var_6144_transpose_x_0, transpose_y = var_6144_transpose_y_0, x = Minv_205_cast_fp16, y = Np_155_cast_fp16)[name = string("op_6144_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_207_cast_fp16 = add(x = Minv_205_cast_fp16, y = var_6144_cast_fp16)[name = string("Minv_207_cast_fp16")];
            bool u_51_transpose_x_0 = const()[name = string("u_51_transpose_x_0"), val = bool(false)];
            bool u_51_transpose_y_0 = const()[name = string("u_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_51_cast_fp16 = matmul(transpose_x = u_51_transpose_x_0, transpose_y = u_51_transpose_y_0, x = Minv_207_cast_fp16, y = vb_61_cast_fp16)[name = string("u_51_cast_fp16")];
            bool var_6151_transpose_x_1 = const()[name = string("op_6151_transpose_x_1"), val = bool(false)];
            bool var_6151_transpose_y_1 = const()[name = string("op_6151_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_6151_cast_fp16 = matmul(transpose_x = var_6151_transpose_x_1, transpose_y = var_6151_transpose_y_1, x = q_93_cast_fp16, y = k_95_cast_fp16)[name = string("op_6151_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_51_cast_fp16 = mul(x = var_6151_cast_fp16, y = D_51_cast_fp16)[name = string("A_51_cast_fp16")];
            tensor<fp16, [16]> var_6153_bias_0_to_fp16 = const()[name = string("op_6153_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_6153_cast_fp16 = linear(bias = var_6153_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_61_cast_fp16)[name = string("op_6153_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_41_cast_fp16 = exp(x = var_6153_cast_fp16)[name = string("cum_41_cast_fp16")];
            tensor<int32, [1]> var_6156_axes_0 = const()[name = string("op_6156_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_6156_cast_fp16 = expand_dims(axes = var_6156_axes_0, x = cum_41_cast_fp16)[name = string("op_6156_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6157_cast_fp16 = mul(x = kb_51_cast_fp16, y = var_6156_cast_fp16)[name = string("op_6157_cast_fp16")];
            bool w_51_transpose_x_0 = const()[name = string("w_51_transpose_x_0"), val = bool(false)];
            bool w_51_transpose_y_0 = const()[name = string("w_51_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_51_cast_fp16 = matmul(transpose_x = w_51_transpose_x_0, transpose_y = w_51_transpose_y_0, x = Minv_207_cast_fp16, y = var_6157_cast_fp16)[name = string("w_51_cast_fp16")];
            bool var_6159_transpose_x_0 = const()[name = string("op_6159_transpose_x_0"), val = bool(false)];
            bool var_6159_transpose_y_0 = const()[name = string("op_6159_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6159_cast_fp16 = matmul(transpose_x = var_6159_transpose_x_0, transpose_y = var_6159_transpose_y_0, x = w_51_cast_fp16, y = S_41_cast_fp16)[name = string("op_6159_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_41_cast_fp16 = sub(x = u_51_cast_fp16, y = var_6159_cast_fp16)[name = string("vnew_41_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6164_cast_fp16 = mul(x = q_93_cast_fp16, y = var_6156_cast_fp16)[name = string("op_6164_cast_fp16")];
            bool var_6165_transpose_x_0 = const()[name = string("op_6165_transpose_x_0"), val = bool(false)];
            bool var_6165_transpose_y_0 = const()[name = string("op_6165_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6165_cast_fp16 = matmul(transpose_x = var_6165_transpose_x_0, transpose_y = var_6165_transpose_y_0, x = var_6164_cast_fp16, y = S_41_cast_fp16)[name = string("op_6165_cast_fp16")];
            bool var_6166_transpose_x_0 = const()[name = string("op_6166_transpose_x_0"), val = bool(false)];
            bool var_6166_transpose_y_0 = const()[name = string("op_6166_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6166_cast_fp16 = matmul(transpose_x = var_6166_transpose_x_0, transpose_y = var_6166_transpose_y_0, x = A_51_cast_fp16, y = vnew_41_cast_fp16)[name = string("op_6166_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6168_cast_fp16 = add(x = var_6165_cast_fp16, y = var_6166_cast_fp16)[name = string("op_6168_cast_fp16")];
            tensor<fp16, [16]> var_6169_bias_0_to_fp16 = const()[name = string("op_6169_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_6169_cast_fp16 = linear(bias = var_6169_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_61_cast_fp16)[name = string("op_6169_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_43_cast_fp16 = exp(x = var_6169_cast_fp16)[name = string("tail_43_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_85_to_fp16 = const()[name = string("transpose_85_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_6171_bias_0_to_fp16 = const()[name = string("op_6171_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_6171_cast_fp16 = linear(bias = var_6171_bias_0_to_fp16, weight = transpose_85_to_fp16, x = gdec_61_cast_fp16)[name = string("op_6171_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_6172_cast_fp16 = exp(x = var_6171_cast_fp16)[name = string("op_6172_cast_fp16")];
            tensor<int32, [1]> last_33_axes_0 = const()[name = string("last_33_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_33_cast_fp16 = expand_dims(axes = last_33_axes_0, x = var_6172_cast_fp16)[name = string("last_33_cast_fp16")];
            tensor<int32, [1]> var_6176_axes_0 = const()[name = string("op_6176_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_6176_cast_fp16 = expand_dims(axes = var_6176_axes_0, x = tail_43_cast_fp16)[name = string("op_6176_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6177_cast_fp16 = mul(x = k_95_cast_fp16, y = var_6176_cast_fp16)[name = string("op_6177_cast_fp16")];
            bool upd_33_transpose_x_1 = const()[name = string("upd_33_transpose_x_1"), val = bool(true)];
            bool upd_33_transpose_y_1 = const()[name = string("upd_33_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_33_cast_fp16 = matmul(transpose_x = upd_33_transpose_x_1, transpose_y = upd_33_transpose_y_1, x = var_6177_cast_fp16, y = vnew_41_cast_fp16)[name = string("upd_33_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_6182_cast_fp16 = mul(x = S_41_cast_fp16, y = last_33_cast_fp16)[name = string("op_6182_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_43_cast_fp16 = add(x = var_6182_cast_fp16, y = upd_33_cast_fp16)[name = string("S_43_cast_fp16")];
            tensor<int32, [4]> q_95_begin_0 = const()[name = string("q_95_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_95_end_0 = const()[name = string("q_95_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> q_95_end_mask_0 = const()[name = string("q_95_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_95_cast_fp16 = slice_by_index(begin = q_95_begin_0, end = q_95_end_0, end_mask = q_95_end_mask_0, x = q_89_cast_fp16)[name = string("q_95_cast_fp16")];
            tensor<int32, [4]> k_97_begin_0 = const()[name = string("k_97_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_97_end_0 = const()[name = string("k_97_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> k_97_end_mask_0 = const()[name = string("k_97_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_97_cast_fp16 = slice_by_index(begin = k_97_begin_0, end = k_97_end_0, end_mask = k_97_end_mask_0, x = k_91_cast_fp16)[name = string("k_97_cast_fp16")];
            tensor<int32, [4]> kb_53_begin_0 = const()[name = string("kb_53_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_53_end_0 = const()[name = string("kb_53_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> kb_53_end_mask_0 = const()[name = string("kb_53_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_53_cast_fp16 = slice_by_index(begin = kb_53_begin_0, end = kb_53_end_0, end_mask = kb_53_end_mask_0, x = var_5931_cast_fp16)[name = string("kb_53_cast_fp16")];
            tensor<int32, [4]> vb_63_begin_0 = const()[name = string("vb_63_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_63_end_0 = const()[name = string("vb_63_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> vb_63_end_mask_0 = const()[name = string("vb_63_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_63_cast_fp16 = slice_by_index(begin = vb_63_begin_0, end = vb_63_end_0, end_mask = vb_63_end_mask_0, x = vb_57_cast_fp16)[name = string("vb_63_cast_fp16")];
            tensor<int32, [3]> gdec_63_begin_0 = const()[name = string("gdec_63_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_63_end_0 = const()[name = string("gdec_63_end_0"), val = tensor<int32, [3]>([3, 4, 48])];
            tensor<bool, [3]> gdec_63_end_mask_0 = const()[name = string("gdec_63_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_63_cast_fp16 = slice_by_index(begin = gdec_63_begin_0, end = gdec_63_end_0, end_mask = gdec_63_end_mask_0, x = gdec_57_cast_fp16)[name = string("gdec_63_cast_fp16")];
            tensor<int32, [2]> var_6265 = const()[name = string("op_6265"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_6266_cast_fp16 = reshape(shape = var_6265, x = gdec_63_cast_fp16)[name = string("op_6266_cast_fp16")];
            tensor<fp16, [12, 256]> var_6267_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_6266_cast_fp16)[name = string("op_6267_cast_fp16")];
            tensor<int32, [4]> var_6272 = const()[name = string("op_6272"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_53_cast_fp16 = reshape(shape = var_6272, x = var_6267_cast_fp16)[name = string("P_53_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6274_cast_fp16 = exp(x = P_53_cast_fp16)[name = string("op_6274_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_53_cast_fp16 = mul(x = var_6274_cast_fp16, y = tril_incl_to_fp16)[name = string("D_53_cast_fp16")];
            bool kk_53_transpose_x_1 = const()[name = string("kk_53_transpose_x_1"), val = bool(false)];
            bool kk_53_transpose_y_1 = const()[name = string("kk_53_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_53_cast_fp16 = matmul(transpose_x = kk_53_transpose_x_1, transpose_y = kk_53_transpose_y_1, x = kb_53_cast_fp16, y = k_97_cast_fp16)[name = string("kk_53_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_6280_cast_fp16 = mul(x = kk_53_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_6280_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6281_cast_fp16 = mul(x = var_6280_cast_fp16, y = D_53_cast_fp16)[name = string("op_6281_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_53_cast_fp16 = mul(x = var_6281_cast_fp16, y = cum_tail_to_fp16)[name = string("N_53_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6283_cast_fp16 = mul(x = D_53_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_6283_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_53_cast_fp16 = mul(x = kk_53_cast_fp16, y = var_6283_cast_fp16)[name = string("Na_53_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_209_cast_fp16 = add(x = eye_to_fp16, y = N_53_cast_fp16)[name = string("Minv_209_cast_fp16")];
            bool Np_157_transpose_x_0 = const()[name = string("Np_157_transpose_x_0"), val = bool(false)];
            bool Np_157_transpose_y_0 = const()[name = string("Np_157_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_157_cast_fp16 = matmul(transpose_x = Np_157_transpose_x_0, transpose_y = Np_157_transpose_y_0, x = N_53_cast_fp16, y = Na_53_cast_fp16)[name = string("Np_157_cast_fp16")];
            bool Npa_105_transpose_x_0 = const()[name = string("Npa_105_transpose_x_0"), val = bool(false)];
            bool Npa_105_transpose_y_0 = const()[name = string("Npa_105_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_105_cast_fp16 = matmul(transpose_x = Npa_105_transpose_x_0, transpose_y = Npa_105_transpose_y_0, x = Na_53_cast_fp16, y = N_53_cast_fp16)[name = string("Npa_105_cast_fp16")];
            bool var_6289_transpose_x_0 = const()[name = string("op_6289_transpose_x_0"), val = bool(false)];
            bool var_6289_transpose_y_0 = const()[name = string("op_6289_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6289_cast_fp16 = matmul(transpose_x = var_6289_transpose_x_0, transpose_y = var_6289_transpose_y_0, x = Minv_209_cast_fp16, y = Np_157_cast_fp16)[name = string("op_6289_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_211_cast_fp16 = add(x = Minv_209_cast_fp16, y = var_6289_cast_fp16)[name = string("Minv_211_cast_fp16")];
            bool Np_159_transpose_x_0 = const()[name = string("Np_159_transpose_x_0"), val = bool(false)];
            bool Np_159_transpose_y_0 = const()[name = string("Np_159_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_159_cast_fp16 = matmul(transpose_x = Np_159_transpose_x_0, transpose_y = Np_159_transpose_y_0, x = Np_157_cast_fp16, y = Npa_105_cast_fp16)[name = string("Np_159_cast_fp16")];
            bool Npa_107_transpose_x_0 = const()[name = string("Npa_107_transpose_x_0"), val = bool(false)];
            bool Npa_107_transpose_y_0 = const()[name = string("Npa_107_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_107_cast_fp16 = matmul(transpose_x = Npa_107_transpose_x_0, transpose_y = Npa_107_transpose_y_0, x = Npa_105_cast_fp16, y = Np_157_cast_fp16)[name = string("Npa_107_cast_fp16")];
            bool var_6294_transpose_x_0 = const()[name = string("op_6294_transpose_x_0"), val = bool(false)];
            bool var_6294_transpose_y_0 = const()[name = string("op_6294_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6294_cast_fp16 = matmul(transpose_x = var_6294_transpose_x_0, transpose_y = var_6294_transpose_y_0, x = Minv_211_cast_fp16, y = Np_159_cast_fp16)[name = string("op_6294_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_213_cast_fp16 = add(x = Minv_211_cast_fp16, y = var_6294_cast_fp16)[name = string("Minv_213_cast_fp16")];
            bool Np_161_transpose_x_0 = const()[name = string("Np_161_transpose_x_0"), val = bool(false)];
            bool Np_161_transpose_y_0 = const()[name = string("Np_161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_161_cast_fp16 = matmul(transpose_x = Np_161_transpose_x_0, transpose_y = Np_161_transpose_y_0, x = Np_159_cast_fp16, y = Npa_107_cast_fp16)[name = string("Np_161_cast_fp16")];
            bool var_6298_transpose_x_0 = const()[name = string("op_6298_transpose_x_0"), val = bool(false)];
            bool var_6298_transpose_y_0 = const()[name = string("op_6298_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6298_cast_fp16 = matmul(transpose_x = var_6298_transpose_x_0, transpose_y = var_6298_transpose_y_0, x = Minv_213_cast_fp16, y = Np_161_cast_fp16)[name = string("op_6298_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_215_cast_fp16 = add(x = Minv_213_cast_fp16, y = var_6298_cast_fp16)[name = string("Minv_215_cast_fp16")];
            bool u_53_transpose_x_0 = const()[name = string("u_53_transpose_x_0"), val = bool(false)];
            bool u_53_transpose_y_0 = const()[name = string("u_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_53_cast_fp16 = matmul(transpose_x = u_53_transpose_x_0, transpose_y = u_53_transpose_y_0, x = Minv_215_cast_fp16, y = vb_63_cast_fp16)[name = string("u_53_cast_fp16")];
            bool var_6305_transpose_x_1 = const()[name = string("op_6305_transpose_x_1"), val = bool(false)];
            bool var_6305_transpose_y_1 = const()[name = string("op_6305_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_6305_cast_fp16 = matmul(transpose_x = var_6305_transpose_x_1, transpose_y = var_6305_transpose_y_1, x = q_95_cast_fp16, y = k_97_cast_fp16)[name = string("op_6305_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_53_cast_fp16 = mul(x = var_6305_cast_fp16, y = D_53_cast_fp16)[name = string("A_53_cast_fp16")];
            tensor<fp16, [16]> var_6307_bias_0_to_fp16 = const()[name = string("op_6307_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_6307_cast_fp16 = linear(bias = var_6307_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_63_cast_fp16)[name = string("op_6307_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_43_cast_fp16 = exp(x = var_6307_cast_fp16)[name = string("cum_43_cast_fp16")];
            tensor<int32, [1]> var_6310_axes_0 = const()[name = string("op_6310_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_6310_cast_fp16 = expand_dims(axes = var_6310_axes_0, x = cum_43_cast_fp16)[name = string("op_6310_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6311_cast_fp16 = mul(x = kb_53_cast_fp16, y = var_6310_cast_fp16)[name = string("op_6311_cast_fp16")];
            bool w_53_transpose_x_0 = const()[name = string("w_53_transpose_x_0"), val = bool(false)];
            bool w_53_transpose_y_0 = const()[name = string("w_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_53_cast_fp16 = matmul(transpose_x = w_53_transpose_x_0, transpose_y = w_53_transpose_y_0, x = Minv_215_cast_fp16, y = var_6311_cast_fp16)[name = string("w_53_cast_fp16")];
            bool var_6313_transpose_x_0 = const()[name = string("op_6313_transpose_x_0"), val = bool(false)];
            bool var_6313_transpose_y_0 = const()[name = string("op_6313_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6313_cast_fp16 = matmul(transpose_x = var_6313_transpose_x_0, transpose_y = var_6313_transpose_y_0, x = w_53_cast_fp16, y = S_43_cast_fp16)[name = string("op_6313_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_43_cast_fp16 = sub(x = u_53_cast_fp16, y = var_6313_cast_fp16)[name = string("vnew_43_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6318_cast_fp16 = mul(x = q_95_cast_fp16, y = var_6310_cast_fp16)[name = string("op_6318_cast_fp16")];
            bool var_6319_transpose_x_0 = const()[name = string("op_6319_transpose_x_0"), val = bool(false)];
            bool var_6319_transpose_y_0 = const()[name = string("op_6319_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6319_cast_fp16 = matmul(transpose_x = var_6319_transpose_x_0, transpose_y = var_6319_transpose_y_0, x = var_6318_cast_fp16, y = S_43_cast_fp16)[name = string("op_6319_cast_fp16")];
            bool var_6320_transpose_x_0 = const()[name = string("op_6320_transpose_x_0"), val = bool(false)];
            bool var_6320_transpose_y_0 = const()[name = string("op_6320_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6320_cast_fp16 = matmul(transpose_x = var_6320_transpose_x_0, transpose_y = var_6320_transpose_y_0, x = A_53_cast_fp16, y = vnew_43_cast_fp16)[name = string("op_6320_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6322_cast_fp16 = add(x = var_6319_cast_fp16, y = var_6320_cast_fp16)[name = string("op_6322_cast_fp16")];
            tensor<fp16, [16]> var_6323_bias_0_to_fp16 = const()[name = string("op_6323_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_6323_cast_fp16 = linear(bias = var_6323_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_63_cast_fp16)[name = string("op_6323_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_45_cast_fp16 = exp(x = var_6323_cast_fp16)[name = string("tail_45_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_89_to_fp16 = const()[name = string("transpose_89_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_6325_bias_0_to_fp16 = const()[name = string("op_6325_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_6325_cast_fp16 = linear(bias = var_6325_bias_0_to_fp16, weight = transpose_89_to_fp16, x = gdec_63_cast_fp16)[name = string("op_6325_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_6326_cast_fp16 = exp(x = var_6325_cast_fp16)[name = string("op_6326_cast_fp16")];
            tensor<int32, [1]> last_35_axes_0 = const()[name = string("last_35_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_35_cast_fp16 = expand_dims(axes = last_35_axes_0, x = var_6326_cast_fp16)[name = string("last_35_cast_fp16")];
            tensor<int32, [1]> var_6330_axes_0 = const()[name = string("op_6330_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_6330_cast_fp16 = expand_dims(axes = var_6330_axes_0, x = tail_45_cast_fp16)[name = string("op_6330_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6331_cast_fp16 = mul(x = k_97_cast_fp16, y = var_6330_cast_fp16)[name = string("op_6331_cast_fp16")];
            bool upd_35_transpose_x_1 = const()[name = string("upd_35_transpose_x_1"), val = bool(true)];
            bool upd_35_transpose_y_1 = const()[name = string("upd_35_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_35_cast_fp16 = matmul(transpose_x = upd_35_transpose_x_1, transpose_y = upd_35_transpose_y_1, x = var_6331_cast_fp16, y = vnew_43_cast_fp16)[name = string("upd_35_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_6336_cast_fp16 = mul(x = S_43_cast_fp16, y = last_35_cast_fp16)[name = string("op_6336_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_45_cast_fp16 = add(x = var_6336_cast_fp16, y = upd_35_cast_fp16)[name = string("S_45_cast_fp16")];
            tensor<int32, [4]> q_97_begin_0 = const()[name = string("q_97_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_97_end_0 = const()[name = string("q_97_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> q_97_end_mask_0 = const()[name = string("q_97_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, x = q_89_cast_fp16)[name = string("q_97_cast_fp16")];
            tensor<int32, [4]> k_99_begin_0 = const()[name = string("k_99_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_99_end_0 = const()[name = string("k_99_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> k_99_end_mask_0 = const()[name = string("k_99_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_99_cast_fp16 = slice_by_index(begin = k_99_begin_0, end = k_99_end_0, end_mask = k_99_end_mask_0, x = k_91_cast_fp16)[name = string("k_99_cast_fp16")];
            tensor<int32, [4]> kb_55_begin_0 = const()[name = string("kb_55_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_55_end_0 = const()[name = string("kb_55_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> kb_55_end_mask_0 = const()[name = string("kb_55_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_55_cast_fp16 = slice_by_index(begin = kb_55_begin_0, end = kb_55_end_0, end_mask = kb_55_end_mask_0, x = var_5931_cast_fp16)[name = string("kb_55_cast_fp16")];
            tensor<int32, [4]> vb_65_begin_0 = const()[name = string("vb_65_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_65_end_0 = const()[name = string("vb_65_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> vb_65_end_mask_0 = const()[name = string("vb_65_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_65_cast_fp16 = slice_by_index(begin = vb_65_begin_0, end = vb_65_end_0, end_mask = vb_65_end_mask_0, x = vb_57_cast_fp16)[name = string("vb_65_cast_fp16")];
            tensor<int32, [3]> gdec_65_begin_0 = const()[name = string("gdec_65_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_65_end_0 = const()[name = string("gdec_65_end_0"), val = tensor<int32, [3]>([3, 4, 64])];
            tensor<bool, [3]> gdec_65_end_mask_0 = const()[name = string("gdec_65_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_65_cast_fp16 = slice_by_index(begin = gdec_65_begin_0, end = gdec_65_end_0, end_mask = gdec_65_end_mask_0, x = gdec_57_cast_fp16)[name = string("gdec_65_cast_fp16")];
            tensor<int32, [2]> var_6419 = const()[name = string("op_6419"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_6420_cast_fp16 = reshape(shape = var_6419, x = gdec_65_cast_fp16)[name = string("op_6420_cast_fp16")];
            tensor<fp16, [12, 256]> var_6421_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_6420_cast_fp16)[name = string("op_6421_cast_fp16")];
            tensor<int32, [4]> var_6426 = const()[name = string("op_6426"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_55_cast_fp16 = reshape(shape = var_6426, x = var_6421_cast_fp16)[name = string("P_55_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6428_cast_fp16 = exp(x = P_55_cast_fp16)[name = string("op_6428_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_55_cast_fp16 = mul(x = var_6428_cast_fp16, y = tril_incl_to_fp16)[name = string("D_55_cast_fp16")];
            bool kk_55_transpose_x_1 = const()[name = string("kk_55_transpose_x_1"), val = bool(false)];
            bool kk_55_transpose_y_1 = const()[name = string("kk_55_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_55_cast_fp16 = matmul(transpose_x = kk_55_transpose_x_1, transpose_y = kk_55_transpose_y_1, x = kb_55_cast_fp16, y = k_99_cast_fp16)[name = string("kk_55_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_6434_cast_fp16 = mul(x = kk_55_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_6434_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6435_cast_fp16 = mul(x = var_6434_cast_fp16, y = D_55_cast_fp16)[name = string("op_6435_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_55_cast_fp16 = mul(x = var_6435_cast_fp16, y = cum_tail_to_fp16)[name = string("N_55_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6437_cast_fp16 = mul(x = D_55_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_6437_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_55_cast_fp16 = mul(x = kk_55_cast_fp16, y = var_6437_cast_fp16)[name = string("Na_55_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_217_cast_fp16 = add(x = eye_to_fp16, y = N_55_cast_fp16)[name = string("Minv_217_cast_fp16")];
            bool Np_163_transpose_x_0 = const()[name = string("Np_163_transpose_x_0"), val = bool(false)];
            bool Np_163_transpose_y_0 = const()[name = string("Np_163_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_163_cast_fp16 = matmul(transpose_x = Np_163_transpose_x_0, transpose_y = Np_163_transpose_y_0, x = N_55_cast_fp16, y = Na_55_cast_fp16)[name = string("Np_163_cast_fp16")];
            bool Npa_109_transpose_x_0 = const()[name = string("Npa_109_transpose_x_0"), val = bool(false)];
            bool Npa_109_transpose_y_0 = const()[name = string("Npa_109_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_109_cast_fp16 = matmul(transpose_x = Npa_109_transpose_x_0, transpose_y = Npa_109_transpose_y_0, x = Na_55_cast_fp16, y = N_55_cast_fp16)[name = string("Npa_109_cast_fp16")];
            bool var_6443_transpose_x_0 = const()[name = string("op_6443_transpose_x_0"), val = bool(false)];
            bool var_6443_transpose_y_0 = const()[name = string("op_6443_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6443_cast_fp16 = matmul(transpose_x = var_6443_transpose_x_0, transpose_y = var_6443_transpose_y_0, x = Minv_217_cast_fp16, y = Np_163_cast_fp16)[name = string("op_6443_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_219_cast_fp16 = add(x = Minv_217_cast_fp16, y = var_6443_cast_fp16)[name = string("Minv_219_cast_fp16")];
            bool Np_165_transpose_x_0 = const()[name = string("Np_165_transpose_x_0"), val = bool(false)];
            bool Np_165_transpose_y_0 = const()[name = string("Np_165_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_165_cast_fp16 = matmul(transpose_x = Np_165_transpose_x_0, transpose_y = Np_165_transpose_y_0, x = Np_163_cast_fp16, y = Npa_109_cast_fp16)[name = string("Np_165_cast_fp16")];
            bool Npa_111_transpose_x_0 = const()[name = string("Npa_111_transpose_x_0"), val = bool(false)];
            bool Npa_111_transpose_y_0 = const()[name = string("Npa_111_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_111_cast_fp16 = matmul(transpose_x = Npa_111_transpose_x_0, transpose_y = Npa_111_transpose_y_0, x = Npa_109_cast_fp16, y = Np_163_cast_fp16)[name = string("Npa_111_cast_fp16")];
            bool var_6448_transpose_x_0 = const()[name = string("op_6448_transpose_x_0"), val = bool(false)];
            bool var_6448_transpose_y_0 = const()[name = string("op_6448_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6448_cast_fp16 = matmul(transpose_x = var_6448_transpose_x_0, transpose_y = var_6448_transpose_y_0, x = Minv_219_cast_fp16, y = Np_165_cast_fp16)[name = string("op_6448_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_221_cast_fp16 = add(x = Minv_219_cast_fp16, y = var_6448_cast_fp16)[name = string("Minv_221_cast_fp16")];
            bool Np_167_transpose_x_0 = const()[name = string("Np_167_transpose_x_0"), val = bool(false)];
            bool Np_167_transpose_y_0 = const()[name = string("Np_167_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_167_cast_fp16 = matmul(transpose_x = Np_167_transpose_x_0, transpose_y = Np_167_transpose_y_0, x = Np_165_cast_fp16, y = Npa_111_cast_fp16)[name = string("Np_167_cast_fp16")];
            bool var_6452_transpose_x_0 = const()[name = string("op_6452_transpose_x_0"), val = bool(false)];
            bool var_6452_transpose_y_0 = const()[name = string("op_6452_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6452_cast_fp16 = matmul(transpose_x = var_6452_transpose_x_0, transpose_y = var_6452_transpose_y_0, x = Minv_221_cast_fp16, y = Np_167_cast_fp16)[name = string("op_6452_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_223_cast_fp16 = add(x = Minv_221_cast_fp16, y = var_6452_cast_fp16)[name = string("Minv_223_cast_fp16")];
            bool u_55_transpose_x_0 = const()[name = string("u_55_transpose_x_0"), val = bool(false)];
            bool u_55_transpose_y_0 = const()[name = string("u_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_55_cast_fp16 = matmul(transpose_x = u_55_transpose_x_0, transpose_y = u_55_transpose_y_0, x = Minv_223_cast_fp16, y = vb_65_cast_fp16)[name = string("u_55_cast_fp16")];
            bool var_6459_transpose_x_1 = const()[name = string("op_6459_transpose_x_1"), val = bool(false)];
            bool var_6459_transpose_y_1 = const()[name = string("op_6459_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_6459_cast_fp16 = matmul(transpose_x = var_6459_transpose_x_1, transpose_y = var_6459_transpose_y_1, x = q_97_cast_fp16, y = k_99_cast_fp16)[name = string("op_6459_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_55_cast_fp16 = mul(x = var_6459_cast_fp16, y = D_55_cast_fp16)[name = string("A_55_cast_fp16")];
            tensor<fp16, [16]> var_6461_bias_0_to_fp16 = const()[name = string("op_6461_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_6461_cast_fp16 = linear(bias = var_6461_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_65_cast_fp16)[name = string("op_6461_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_45_cast_fp16 = exp(x = var_6461_cast_fp16)[name = string("cum_45_cast_fp16")];
            tensor<int32, [1]> var_6464_axes_0 = const()[name = string("op_6464_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_6464_cast_fp16 = expand_dims(axes = var_6464_axes_0, x = cum_45_cast_fp16)[name = string("op_6464_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6465_cast_fp16 = mul(x = kb_55_cast_fp16, y = var_6464_cast_fp16)[name = string("op_6465_cast_fp16")];
            bool w_55_transpose_x_0 = const()[name = string("w_55_transpose_x_0"), val = bool(false)];
            bool w_55_transpose_y_0 = const()[name = string("w_55_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_55_cast_fp16 = matmul(transpose_x = w_55_transpose_x_0, transpose_y = w_55_transpose_y_0, x = Minv_223_cast_fp16, y = var_6465_cast_fp16)[name = string("w_55_cast_fp16")];
            bool var_6467_transpose_x_0 = const()[name = string("op_6467_transpose_x_0"), val = bool(false)];
            bool var_6467_transpose_y_0 = const()[name = string("op_6467_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6467_cast_fp16 = matmul(transpose_x = var_6467_transpose_x_0, transpose_y = var_6467_transpose_y_0, x = w_55_cast_fp16, y = S_45_cast_fp16)[name = string("op_6467_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_45_cast_fp16 = sub(x = u_55_cast_fp16, y = var_6467_cast_fp16)[name = string("vnew_45_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6472_cast_fp16 = mul(x = q_97_cast_fp16, y = var_6464_cast_fp16)[name = string("op_6472_cast_fp16")];
            bool var_6473_transpose_x_0 = const()[name = string("op_6473_transpose_x_0"), val = bool(false)];
            bool var_6473_transpose_y_0 = const()[name = string("op_6473_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6473_cast_fp16 = matmul(transpose_x = var_6473_transpose_x_0, transpose_y = var_6473_transpose_y_0, x = var_6472_cast_fp16, y = S_45_cast_fp16)[name = string("op_6473_cast_fp16")];
            bool var_6474_transpose_x_0 = const()[name = string("op_6474_transpose_x_0"), val = bool(false)];
            bool var_6474_transpose_y_0 = const()[name = string("op_6474_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6474_cast_fp16 = matmul(transpose_x = var_6474_transpose_x_0, transpose_y = var_6474_transpose_y_0, x = A_55_cast_fp16, y = vnew_45_cast_fp16)[name = string("op_6474_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6476_cast_fp16 = add(x = var_6473_cast_fp16, y = var_6474_cast_fp16)[name = string("op_6476_cast_fp16")];
            tensor<fp16, [16]> var_6477_bias_0_to_fp16 = const()[name = string("op_6477_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_6477_cast_fp16 = linear(bias = var_6477_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_65_cast_fp16)[name = string("op_6477_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_47_cast_fp16 = exp(x = var_6477_cast_fp16)[name = string("tail_47_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_93_to_fp16 = const()[name = string("transpose_93_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_6479_bias_0_to_fp16 = const()[name = string("op_6479_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_6479_cast_fp16 = linear(bias = var_6479_bias_0_to_fp16, weight = transpose_93_to_fp16, x = gdec_65_cast_fp16)[name = string("op_6479_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_6480_cast_fp16 = exp(x = var_6479_cast_fp16)[name = string("op_6480_cast_fp16")];
            tensor<int32, [1]> last_37_axes_0 = const()[name = string("last_37_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_37_cast_fp16 = expand_dims(axes = last_37_axes_0, x = var_6480_cast_fp16)[name = string("last_37_cast_fp16")];
            tensor<int32, [1]> var_6484_axes_0 = const()[name = string("op_6484_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_6484_cast_fp16 = expand_dims(axes = var_6484_axes_0, x = tail_47_cast_fp16)[name = string("op_6484_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6485_cast_fp16 = mul(x = k_99_cast_fp16, y = var_6484_cast_fp16)[name = string("op_6485_cast_fp16")];
            bool upd_37_transpose_x_1 = const()[name = string("upd_37_transpose_x_1"), val = bool(true)];
            bool upd_37_transpose_y_1 = const()[name = string("upd_37_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_37_cast_fp16 = matmul(transpose_x = upd_37_transpose_x_1, transpose_y = upd_37_transpose_y_1, x = var_6485_cast_fp16, y = vnew_45_cast_fp16)[name = string("upd_37_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_6490_cast_fp16 = mul(x = S_45_cast_fp16, y = last_37_cast_fp16)[name = string("op_6490_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_47_cast_fp16 = add(x = var_6490_cast_fp16, y = upd_37_cast_fp16)[name = string("S_47_cast_fp16")];
            tensor<int32, [4]> q_99_begin_0 = const()[name = string("q_99_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> q_99_end_0 = const()[name = string("q_99_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> q_99_end_mask_0 = const()[name = string("q_99_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_99_cast_fp16 = slice_by_index(begin = q_99_begin_0, end = q_99_end_0, end_mask = q_99_end_mask_0, x = q_89_cast_fp16)[name = string("q_99_cast_fp16")];
            tensor<int32, [4]> k_101_begin_0 = const()[name = string("k_101_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> k_101_end_0 = const()[name = string("k_101_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> k_101_end_mask_0 = const()[name = string("k_101_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_101_cast_fp16 = slice_by_index(begin = k_101_begin_0, end = k_101_end_0, end_mask = k_101_end_mask_0, x = k_91_cast_fp16)[name = string("k_101_cast_fp16")];
            tensor<int32, [4]> kb_57_begin_0 = const()[name = string("kb_57_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> kb_57_end_0 = const()[name = string("kb_57_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> kb_57_end_mask_0 = const()[name = string("kb_57_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_57_cast_fp16 = slice_by_index(begin = kb_57_begin_0, end = kb_57_end_0, end_mask = kb_57_end_mask_0, x = var_5931_cast_fp16)[name = string("kb_57_cast_fp16")];
            tensor<int32, [4]> vb_67_begin_0 = const()[name = string("vb_67_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> vb_67_end_0 = const()[name = string("vb_67_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> vb_67_end_mask_0 = const()[name = string("vb_67_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_67_cast_fp16 = slice_by_index(begin = vb_67_begin_0, end = vb_67_end_0, end_mask = vb_67_end_mask_0, x = vb_57_cast_fp16)[name = string("vb_67_cast_fp16")];
            tensor<int32, [3]> gdec_67_begin_0 = const()[name = string("gdec_67_begin_0"), val = tensor<int32, [3]>([0, 0, 64])];
            tensor<int32, [3]> gdec_67_end_0 = const()[name = string("gdec_67_end_0"), val = tensor<int32, [3]>([3, 4, 80])];
            tensor<bool, [3]> gdec_67_end_mask_0 = const()[name = string("gdec_67_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_67_cast_fp16 = slice_by_index(begin = gdec_67_begin_0, end = gdec_67_end_0, end_mask = gdec_67_end_mask_0, x = gdec_57_cast_fp16)[name = string("gdec_67_cast_fp16")];
            tensor<int32, [2]> var_6573 = const()[name = string("op_6573"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_6574_cast_fp16 = reshape(shape = var_6573, x = gdec_67_cast_fp16)[name = string("op_6574_cast_fp16")];
            tensor<fp16, [12, 256]> var_6575_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_6574_cast_fp16)[name = string("op_6575_cast_fp16")];
            tensor<int32, [4]> var_6580 = const()[name = string("op_6580"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_57_cast_fp16 = reshape(shape = var_6580, x = var_6575_cast_fp16)[name = string("P_57_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6582_cast_fp16 = exp(x = P_57_cast_fp16)[name = string("op_6582_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_57_cast_fp16 = mul(x = var_6582_cast_fp16, y = tril_incl_to_fp16)[name = string("D_57_cast_fp16")];
            bool kk_57_transpose_x_1 = const()[name = string("kk_57_transpose_x_1"), val = bool(false)];
            bool kk_57_transpose_y_1 = const()[name = string("kk_57_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_57_cast_fp16 = matmul(transpose_x = kk_57_transpose_x_1, transpose_y = kk_57_transpose_y_1, x = kb_57_cast_fp16, y = k_101_cast_fp16)[name = string("kk_57_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_6588_cast_fp16 = mul(x = kk_57_cast_fp16, y = const_35_promoted_to_fp16)[name = string("op_6588_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6589_cast_fp16 = mul(x = var_6588_cast_fp16, y = D_57_cast_fp16)[name = string("op_6589_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_57_cast_fp16 = mul(x = var_6589_cast_fp16, y = cum_tail_to_fp16)[name = string("N_57_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6591_cast_fp16 = mul(x = D_57_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_6591_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_57_cast_fp16 = mul(x = kk_57_cast_fp16, y = var_6591_cast_fp16)[name = string("Na_57_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_225_cast_fp16 = add(x = eye_to_fp16, y = N_57_cast_fp16)[name = string("Minv_225_cast_fp16")];
            bool Np_169_transpose_x_0 = const()[name = string("Np_169_transpose_x_0"), val = bool(false)];
            bool Np_169_transpose_y_0 = const()[name = string("Np_169_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_169_cast_fp16 = matmul(transpose_x = Np_169_transpose_x_0, transpose_y = Np_169_transpose_y_0, x = N_57_cast_fp16, y = Na_57_cast_fp16)[name = string("Np_169_cast_fp16")];
            bool Npa_113_transpose_x_0 = const()[name = string("Npa_113_transpose_x_0"), val = bool(false)];
            bool Npa_113_transpose_y_0 = const()[name = string("Npa_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_113_cast_fp16 = matmul(transpose_x = Npa_113_transpose_x_0, transpose_y = Npa_113_transpose_y_0, x = Na_57_cast_fp16, y = N_57_cast_fp16)[name = string("Npa_113_cast_fp16")];
            bool var_6597_transpose_x_0 = const()[name = string("op_6597_transpose_x_0"), val = bool(false)];
            bool var_6597_transpose_y_0 = const()[name = string("op_6597_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6597_cast_fp16 = matmul(transpose_x = var_6597_transpose_x_0, transpose_y = var_6597_transpose_y_0, x = Minv_225_cast_fp16, y = Np_169_cast_fp16)[name = string("op_6597_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_227_cast_fp16 = add(x = Minv_225_cast_fp16, y = var_6597_cast_fp16)[name = string("Minv_227_cast_fp16")];
            bool Np_171_transpose_x_0 = const()[name = string("Np_171_transpose_x_0"), val = bool(false)];
            bool Np_171_transpose_y_0 = const()[name = string("Np_171_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_171_cast_fp16 = matmul(transpose_x = Np_171_transpose_x_0, transpose_y = Np_171_transpose_y_0, x = Np_169_cast_fp16, y = Npa_113_cast_fp16)[name = string("Np_171_cast_fp16")];
            bool Npa_115_transpose_x_0 = const()[name = string("Npa_115_transpose_x_0"), val = bool(false)];
            bool Npa_115_transpose_y_0 = const()[name = string("Npa_115_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_115_cast_fp16 = matmul(transpose_x = Npa_115_transpose_x_0, transpose_y = Npa_115_transpose_y_0, x = Npa_113_cast_fp16, y = Np_169_cast_fp16)[name = string("Npa_115_cast_fp16")];
            bool var_6602_transpose_x_0 = const()[name = string("op_6602_transpose_x_0"), val = bool(false)];
            bool var_6602_transpose_y_0 = const()[name = string("op_6602_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6602_cast_fp16 = matmul(transpose_x = var_6602_transpose_x_0, transpose_y = var_6602_transpose_y_0, x = Minv_227_cast_fp16, y = Np_171_cast_fp16)[name = string("op_6602_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_229_cast_fp16 = add(x = Minv_227_cast_fp16, y = var_6602_cast_fp16)[name = string("Minv_229_cast_fp16")];
            bool Np_173_transpose_x_0 = const()[name = string("Np_173_transpose_x_0"), val = bool(false)];
            bool Np_173_transpose_y_0 = const()[name = string("Np_173_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_173_cast_fp16 = matmul(transpose_x = Np_173_transpose_x_0, transpose_y = Np_173_transpose_y_0, x = Np_171_cast_fp16, y = Npa_115_cast_fp16)[name = string("Np_173_cast_fp16")];
            bool var_6606_transpose_x_0 = const()[name = string("op_6606_transpose_x_0"), val = bool(false)];
            bool var_6606_transpose_y_0 = const()[name = string("op_6606_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6606_cast_fp16 = matmul(transpose_x = var_6606_transpose_x_0, transpose_y = var_6606_transpose_y_0, x = Minv_229_cast_fp16, y = Np_173_cast_fp16)[name = string("op_6606_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_231_cast_fp16 = add(x = Minv_229_cast_fp16, y = var_6606_cast_fp16)[name = string("Minv_231_cast_fp16")];
            bool u_57_transpose_x_0 = const()[name = string("u_57_transpose_x_0"), val = bool(false)];
            bool u_57_transpose_y_0 = const()[name = string("u_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_57_cast_fp16 = matmul(transpose_x = u_57_transpose_x_0, transpose_y = u_57_transpose_y_0, x = Minv_231_cast_fp16, y = vb_67_cast_fp16)[name = string("u_57_cast_fp16")];
            bool var_6613_transpose_x_1 = const()[name = string("op_6613_transpose_x_1"), val = bool(false)];
            bool var_6613_transpose_y_1 = const()[name = string("op_6613_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_6613_cast_fp16 = matmul(transpose_x = var_6613_transpose_x_1, transpose_y = var_6613_transpose_y_1, x = q_99_cast_fp16, y = k_101_cast_fp16)[name = string("op_6613_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_57_cast_fp16 = mul(x = var_6613_cast_fp16, y = D_57_cast_fp16)[name = string("A_57_cast_fp16")];
            tensor<fp16, [16]> var_6615_bias_0_to_fp16 = const()[name = string("op_6615_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_6615_cast_fp16 = linear(bias = var_6615_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_67_cast_fp16)[name = string("op_6615_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_47_cast_fp16 = exp(x = var_6615_cast_fp16)[name = string("cum_47_cast_fp16")];
            tensor<int32, [1]> var_6618_axes_0 = const()[name = string("op_6618_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_6618_cast_fp16 = expand_dims(axes = var_6618_axes_0, x = cum_47_cast_fp16)[name = string("op_6618_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6619_cast_fp16 = mul(x = kb_57_cast_fp16, y = var_6618_cast_fp16)[name = string("op_6619_cast_fp16")];
            bool w_57_transpose_x_0 = const()[name = string("w_57_transpose_x_0"), val = bool(false)];
            bool w_57_transpose_y_0 = const()[name = string("w_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_57_cast_fp16 = matmul(transpose_x = w_57_transpose_x_0, transpose_y = w_57_transpose_y_0, x = Minv_231_cast_fp16, y = var_6619_cast_fp16)[name = string("w_57_cast_fp16")];
            bool var_6621_transpose_x_0 = const()[name = string("op_6621_transpose_x_0"), val = bool(false)];
            bool var_6621_transpose_y_0 = const()[name = string("op_6621_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6621_cast_fp16 = matmul(transpose_x = var_6621_transpose_x_0, transpose_y = var_6621_transpose_y_0, x = w_57_cast_fp16, y = S_47_cast_fp16)[name = string("op_6621_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_47_cast_fp16 = sub(x = u_57_cast_fp16, y = var_6621_cast_fp16)[name = string("vnew_47_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6626_cast_fp16 = mul(x = q_99_cast_fp16, y = var_6618_cast_fp16)[name = string("op_6626_cast_fp16")];
            bool var_6627_transpose_x_0 = const()[name = string("op_6627_transpose_x_0"), val = bool(false)];
            bool var_6627_transpose_y_0 = const()[name = string("op_6627_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6627_cast_fp16 = matmul(transpose_x = var_6627_transpose_x_0, transpose_y = var_6627_transpose_y_0, x = var_6626_cast_fp16, y = S_47_cast_fp16)[name = string("op_6627_cast_fp16")];
            bool var_6628_transpose_x_0 = const()[name = string("op_6628_transpose_x_0"), val = bool(false)];
            bool var_6628_transpose_y_0 = const()[name = string("op_6628_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6628_cast_fp16 = matmul(transpose_x = var_6628_transpose_x_0, transpose_y = var_6628_transpose_y_0, x = A_57_cast_fp16, y = vnew_47_cast_fp16)[name = string("op_6628_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6630_cast_fp16 = add(x = var_6627_cast_fp16, y = var_6628_cast_fp16)[name = string("op_6630_cast_fp16")];
            tensor<fp16, [16]> var_6631_bias_0_to_fp16 = const()[name = string("op_6631_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_6631_cast_fp16 = linear(bias = var_6631_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_67_cast_fp16)[name = string("op_6631_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_49_cast_fp16 = exp(x = var_6631_cast_fp16)[name = string("tail_49_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_97_to_fp16 = const()[name = string("transpose_97_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_6633_bias_0_to_fp16 = const()[name = string("op_6633_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_6633_cast_fp16 = linear(bias = var_6633_bias_0_to_fp16, weight = transpose_97_to_fp16, x = gdec_67_cast_fp16)[name = string("op_6633_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_6634_cast_fp16 = exp(x = var_6633_cast_fp16)[name = string("op_6634_cast_fp16")];
            tensor<int32, [1]> last_39_axes_0 = const()[name = string("last_39_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_39_cast_fp16 = expand_dims(axes = last_39_axes_0, x = var_6634_cast_fp16)[name = string("last_39_cast_fp16")];
            tensor<int32, [1]> var_6638_axes_0 = const()[name = string("op_6638_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_6638_cast_fp16 = expand_dims(axes = var_6638_axes_0, x = tail_49_cast_fp16)[name = string("op_6638_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6639_cast_fp16 = mul(x = k_101_cast_fp16, y = var_6638_cast_fp16)[name = string("op_6639_cast_fp16")];
            bool upd_39_transpose_x_1 = const()[name = string("upd_39_transpose_x_1"), val = bool(true)];
            bool upd_39_transpose_y_1 = const()[name = string("upd_39_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_39_cast_fp16 = matmul(transpose_x = upd_39_transpose_x_1, transpose_y = upd_39_transpose_y_1, x = var_6639_cast_fp16, y = vnew_47_cast_fp16)[name = string("upd_39_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_6644_cast_fp16 = mul(x = S_47_cast_fp16, y = last_39_cast_fp16)[name = string("op_6644_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_49_cast_fp16 = add(x = var_6644_cast_fp16, y = upd_39_cast_fp16)[name = string("S_49_cast_fp16")];
            tensor<int32, [4]> q_101_begin_0 = const()[name = string("q_101_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> q_101_end_0 = const()[name = string("q_101_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_101_end_mask_0 = const()[name = string("q_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_101_cast_fp16 = slice_by_index(begin = q_101_begin_0, end = q_101_end_0, end_mask = q_101_end_mask_0, x = q_89_cast_fp16)[name = string("q_101_cast_fp16")];
            tensor<int32, [4]> k_103_begin_0 = const()[name = string("k_103_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> k_103_end_0 = const()[name = string("k_103_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_103_end_mask_0 = const()[name = string("k_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_103_cast_fp16 = slice_by_index(begin = k_103_begin_0, end = k_103_end_0, end_mask = k_103_end_mask_0, x = k_91_cast_fp16)[name = string("k_103_cast_fp16")];
            tensor<int32, [4]> kb_59_begin_0 = const()[name = string("kb_59_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> kb_59_end_0 = const()[name = string("kb_59_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_59_end_mask_0 = const()[name = string("kb_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_59_cast_fp16 = slice_by_index(begin = kb_59_begin_0, end = kb_59_end_0, end_mask = kb_59_end_mask_0, x = var_5931_cast_fp16)[name = string("kb_59_cast_fp16")];
            tensor<int32, [4]> vb_69_begin_0 = const()[name = string("vb_69_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> vb_69_end_0 = const()[name = string("vb_69_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_69_end_mask_0 = const()[name = string("vb_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_69_cast_fp16 = slice_by_index(begin = vb_69_begin_0, end = vb_69_end_0, end_mask = vb_69_end_mask_0, x = vb_57_cast_fp16)[name = string("vb_69_cast_fp16")];
            tensor<int32, [3]> gdec_69_begin_0 = const()[name = string("gdec_69_begin_0"), val = tensor<int32, [3]>([0, 0, 80])];
            tensor<int32, [3]> gdec_69_end_0 = const()[name = string("gdec_69_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_69_end_mask_0 = const()[name = string("gdec_69_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_69_cast_fp16 = slice_by_index(begin = gdec_69_begin_0, end = gdec_69_end_0, end_mask = gdec_69_end_mask_0, x = gdec_57_cast_fp16)[name = string("gdec_69_cast_fp16")];
            tensor<int32, [2]> var_6727 = const()[name = string("op_6727"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_6728_cast_fp16 = reshape(shape = var_6727, x = gdec_69_cast_fp16)[name = string("op_6728_cast_fp16")];
            tensor<fp16, [12, 256]> var_6729_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_6728_cast_fp16)[name = string("op_6729_cast_fp16")];
            tensor<int32, [4]> var_6734 = const()[name = string("op_6734"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_59_cast_fp16 = reshape(shape = var_6734, x = var_6729_cast_fp16)[name = string("P_59_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6736_cast_fp16 = exp(x = P_59_cast_fp16)[name = string("op_6736_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_59_cast_fp16 = mul(x = var_6736_cast_fp16, y = tril_incl_to_fp16)[name = string("D_59_cast_fp16")];
            bool kk_59_transpose_x_1 = const()[name = string("kk_59_transpose_x_1"), val = bool(false)];
            bool kk_59_transpose_y_1 = const()[name = string("kk_59_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_59_cast_fp16 = matmul(transpose_x = kk_59_transpose_x_1, transpose_y = kk_59_transpose_y_1, x = kb_59_cast_fp16, y = k_103_cast_fp16)[name = string("kk_59_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_6742_cast_fp16 = mul(x = kk_59_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_6742_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6743_cast_fp16 = mul(x = var_6742_cast_fp16, y = D_59_cast_fp16)[name = string("op_6743_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_59_cast_fp16 = mul(x = var_6743_cast_fp16, y = cum_tail_to_fp16)[name = string("N_59_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_6745_cast_fp16 = mul(x = D_59_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_6745_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_59_cast_fp16 = mul(x = kk_59_cast_fp16, y = var_6745_cast_fp16)[name = string("Na_59_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_233_cast_fp16 = add(x = eye_to_fp16, y = N_59_cast_fp16)[name = string("Minv_233_cast_fp16")];
            bool Np_175_transpose_x_0 = const()[name = string("Np_175_transpose_x_0"), val = bool(false)];
            bool Np_175_transpose_y_0 = const()[name = string("Np_175_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_175_cast_fp16 = matmul(transpose_x = Np_175_transpose_x_0, transpose_y = Np_175_transpose_y_0, x = N_59_cast_fp16, y = Na_59_cast_fp16)[name = string("Np_175_cast_fp16")];
            bool Npa_117_transpose_x_0 = const()[name = string("Npa_117_transpose_x_0"), val = bool(false)];
            bool Npa_117_transpose_y_0 = const()[name = string("Npa_117_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_117_cast_fp16 = matmul(transpose_x = Npa_117_transpose_x_0, transpose_y = Npa_117_transpose_y_0, x = Na_59_cast_fp16, y = N_59_cast_fp16)[name = string("Npa_117_cast_fp16")];
            bool var_6751_transpose_x_0 = const()[name = string("op_6751_transpose_x_0"), val = bool(false)];
            bool var_6751_transpose_y_0 = const()[name = string("op_6751_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6751_cast_fp16 = matmul(transpose_x = var_6751_transpose_x_0, transpose_y = var_6751_transpose_y_0, x = Minv_233_cast_fp16, y = Np_175_cast_fp16)[name = string("op_6751_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_235_cast_fp16 = add(x = Minv_233_cast_fp16, y = var_6751_cast_fp16)[name = string("Minv_235_cast_fp16")];
            bool Np_177_transpose_x_0 = const()[name = string("Np_177_transpose_x_0"), val = bool(false)];
            bool Np_177_transpose_y_0 = const()[name = string("Np_177_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_177_cast_fp16 = matmul(transpose_x = Np_177_transpose_x_0, transpose_y = Np_177_transpose_y_0, x = Np_175_cast_fp16, y = Npa_117_cast_fp16)[name = string("Np_177_cast_fp16")];
            bool Npa_119_transpose_x_0 = const()[name = string("Npa_119_transpose_x_0"), val = bool(false)];
            bool Npa_119_transpose_y_0 = const()[name = string("Npa_119_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_119_cast_fp16 = matmul(transpose_x = Npa_119_transpose_x_0, transpose_y = Npa_119_transpose_y_0, x = Npa_117_cast_fp16, y = Np_175_cast_fp16)[name = string("Npa_119_cast_fp16")];
            bool var_6756_transpose_x_0 = const()[name = string("op_6756_transpose_x_0"), val = bool(false)];
            bool var_6756_transpose_y_0 = const()[name = string("op_6756_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6756_cast_fp16 = matmul(transpose_x = var_6756_transpose_x_0, transpose_y = var_6756_transpose_y_0, x = Minv_235_cast_fp16, y = Np_177_cast_fp16)[name = string("op_6756_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_237_cast_fp16 = add(x = Minv_235_cast_fp16, y = var_6756_cast_fp16)[name = string("Minv_237_cast_fp16")];
            bool Np_179_transpose_x_0 = const()[name = string("Np_179_transpose_x_0"), val = bool(false)];
            bool Np_179_transpose_y_0 = const()[name = string("Np_179_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_179_cast_fp16 = matmul(transpose_x = Np_179_transpose_x_0, transpose_y = Np_179_transpose_y_0, x = Np_177_cast_fp16, y = Npa_119_cast_fp16)[name = string("Np_179_cast_fp16")];
            bool var_6760_transpose_x_0 = const()[name = string("op_6760_transpose_x_0"), val = bool(false)];
            bool var_6760_transpose_y_0 = const()[name = string("op_6760_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_6760_cast_fp16 = matmul(transpose_x = var_6760_transpose_x_0, transpose_y = var_6760_transpose_y_0, x = Minv_237_cast_fp16, y = Np_179_cast_fp16)[name = string("op_6760_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_239_cast_fp16 = add(x = Minv_237_cast_fp16, y = var_6760_cast_fp16)[name = string("Minv_239_cast_fp16")];
            bool u_59_transpose_x_0 = const()[name = string("u_59_transpose_x_0"), val = bool(false)];
            bool u_59_transpose_y_0 = const()[name = string("u_59_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_59_cast_fp16 = matmul(transpose_x = u_59_transpose_x_0, transpose_y = u_59_transpose_y_0, x = Minv_239_cast_fp16, y = vb_69_cast_fp16)[name = string("u_59_cast_fp16")];
            bool var_6767_transpose_x_1 = const()[name = string("op_6767_transpose_x_1"), val = bool(false)];
            bool var_6767_transpose_y_1 = const()[name = string("op_6767_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_6767_cast_fp16 = matmul(transpose_x = var_6767_transpose_x_1, transpose_y = var_6767_transpose_y_1, x = q_101_cast_fp16, y = k_103_cast_fp16)[name = string("op_6767_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_59_cast_fp16 = mul(x = var_6767_cast_fp16, y = D_59_cast_fp16)[name = string("A_59_cast_fp16")];
            tensor<fp16, [16]> var_6769_bias_0_to_fp16 = const()[name = string("op_6769_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_6769_cast_fp16 = linear(bias = var_6769_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_69_cast_fp16)[name = string("op_6769_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_49_cast_fp16 = exp(x = var_6769_cast_fp16)[name = string("cum_49_cast_fp16")];
            tensor<int32, [1]> var_6772_axes_0 = const()[name = string("op_6772_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_6772_cast_fp16 = expand_dims(axes = var_6772_axes_0, x = cum_49_cast_fp16)[name = string("op_6772_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6773_cast_fp16 = mul(x = kb_59_cast_fp16, y = var_6772_cast_fp16)[name = string("op_6773_cast_fp16")];
            bool w_59_transpose_x_0 = const()[name = string("w_59_transpose_x_0"), val = bool(false)];
            bool w_59_transpose_y_0 = const()[name = string("w_59_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_59_cast_fp16 = matmul(transpose_x = w_59_transpose_x_0, transpose_y = w_59_transpose_y_0, x = Minv_239_cast_fp16, y = var_6773_cast_fp16)[name = string("w_59_cast_fp16")];
            bool var_6775_transpose_x_0 = const()[name = string("op_6775_transpose_x_0"), val = bool(false)];
            bool var_6775_transpose_y_0 = const()[name = string("op_6775_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6775_cast_fp16 = matmul(transpose_x = var_6775_transpose_x_0, transpose_y = var_6775_transpose_y_0, x = w_59_cast_fp16, y = S_49_cast_fp16)[name = string("op_6775_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_49_cast_fp16 = sub(x = u_59_cast_fp16, y = var_6775_cast_fp16)[name = string("vnew_49_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_6780_cast_fp16 = mul(x = q_101_cast_fp16, y = var_6772_cast_fp16)[name = string("op_6780_cast_fp16")];
            bool var_6781_transpose_x_0 = const()[name = string("op_6781_transpose_x_0"), val = bool(false)];
            bool var_6781_transpose_y_0 = const()[name = string("op_6781_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6781_cast_fp16 = matmul(transpose_x = var_6781_transpose_x_0, transpose_y = var_6781_transpose_y_0, x = var_6780_cast_fp16, y = S_49_cast_fp16)[name = string("op_6781_cast_fp16")];
            bool var_6782_transpose_x_0 = const()[name = string("op_6782_transpose_x_0"), val = bool(false)];
            bool var_6782_transpose_y_0 = const()[name = string("op_6782_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_6782_cast_fp16 = matmul(transpose_x = var_6782_transpose_x_0, transpose_y = var_6782_transpose_y_0, x = A_59_cast_fp16, y = vnew_49_cast_fp16)[name = string("op_6782_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_9_cast_fp16 = add(x = var_6781_cast_fp16, y = var_6782_cast_fp16)[name = string("oc_9_cast_fp16")];
            int32 var_6786 = const()[name = string("op_6786"), val = int32(2)];
            bool o_27_interleave_0 = const()[name = string("o_27_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 96, 128]> o_27_cast_fp16 = concat(axis = var_6786, interleave = o_27_interleave_0, values = (var_6021_cast_fp16, var_6168_cast_fp16, var_6322_cast_fp16, var_6476_cast_fp16, var_6630_cast_fp16, oc_9_cast_fp16))[name = string("o_27_cast_fp16")];
            tensor<int32, [4]> var_6792 = const()[name = string("op_6792"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_6798 = const()[name = string("op_6798"), val = tensor<int32, [4]>([3, 96, 4, 128])];
            tensor<fp16, [3, 96, 4, 128]> input_81_cast_fp16 = reshape(shape = var_6798, x = z_9_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> o_29_cast_fp16 = transpose(perm = var_6792, x = o_27_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [3, 96, 4, 128]> var_6800_cast_fp16 = mul(x = o_29_cast_fp16, y = o_29_cast_fp16)[name = string("op_6800_cast_fp16")];
            tensor<int32, [1]> var_6805_axes_0 = const()[name = string("op_6805_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6805_keep_dims_0 = const()[name = string("op_6805_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 4, 1]> var_6805_cast_fp16 = reduce_mean(axes = var_6805_axes_0, keep_dims = var_6805_keep_dims_0, x = var_6800_cast_fp16)[name = string("op_6805_cast_fp16")];
            fp16 var_6807_to_fp16 = const()[name = string("op_6807_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 96, 4, 1]> var_6808_cast_fp16 = add(x = var_6805_cast_fp16, y = var_6807_to_fp16)[name = string("op_6808_cast_fp16")];
            fp32 var_6809_epsilon_0 = const()[name = string("op_6809_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 4, 1]> var_6809_cast_fp16 = rsqrt(epsilon = var_6809_epsilon_0, x = var_6808_cast_fp16)[name = string("op_6809_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> var_6810_cast_fp16 = mul(x = o_29_cast_fp16, y = var_6809_cast_fp16)[name = string("op_6810_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> var_6812_cast_fp16 = silu(x = input_81_cast_fp16)[name = string("op_6812_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> o_31_cast_fp16 = mul(x = var_6810_cast_fp16, y = var_6812_cast_fp16)[name = string("o_31_cast_fp16")];
            tensor<int32, [3]> var_6817 = const()[name = string("op_6817"), val = tensor<int32, [3]>([3, 96, 512])];
            tensor<fp16, [3, 96, 512]> input_83_cast_fp16 = reshape(shape = var_6817, x = o_31_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [640, 512]> layers_5_out_weight_to_fp16 = const()[name = string("layers_5_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84536512)))];
            tensor<fp16, [3, 96, 640]> linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_out_weight_to_fp16, x = input_83_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [3, 96, 640]> x_57_cast_fp16 = add(x = x_53_cast_fp16, y = linear_21_cast_fp16)[name = string("x_57_cast_fp16")];
            fp16 var_6830_to_fp16 = const()[name = string("op_6830_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [3, 96, 640]> x_59_cast_fp16 = mul(x = x_57_cast_fp16, y = var_6830_to_fp16)[name = string("x_59_cast_fp16")];
            fp16 var_6828_promoted_to_fp16 = const()[name = string("op_6828_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> var_6832_cast_fp16 = pow(x = x_59_cast_fp16, y = var_6828_promoted_to_fp16)[name = string("op_6832_cast_fp16")];
            tensor<int32, [1]> var_6834_axes_0 = const()[name = string("op_6834_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6834_keep_dims_0 = const()[name = string("op_6834_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 1]> var_6834_cast_fp16 = reduce_mean(axes = var_6834_axes_0, keep_dims = var_6834_keep_dims_0, x = var_6832_cast_fp16)[name = string("op_6834_cast_fp16")];
            fp16 var_6835_to_fp16 = const()[name = string("op_6835_to_fp16"), val = fp16(0x1p-22)];
            tensor<fp16, [3, 96, 1]> var_6836_cast_fp16 = add(x = var_6834_cast_fp16, y = var_6835_to_fp16)[name = string("op_6836_cast_fp16")];
            fp32 var_6837_epsilon_0 = const()[name = string("op_6837_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 1]> var_6837_cast_fp16 = rsqrt(epsilon = var_6837_epsilon_0, x = var_6836_cast_fp16)[name = string("op_6837_cast_fp16")];
            tensor<fp16, [3, 96, 640]> var_6838_cast_fp16 = mul(x = x_59_cast_fp16, y = var_6837_cast_fp16)[name = string("op_6838_cast_fp16")];
            tensor<fp16, [640]> layers_5_n2_w1_to_fp16 = const()[name = string("layers_5_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85191936)))];
            tensor<fp16, [3, 96, 640]> input_85_cast_fp16 = mul(x = var_6838_cast_fp16, y = layers_5_n2_w1_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_5_gate_up_weight_to_fp16 = const()[name = string("layers_5_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85193280)))];
            tensor<fp16, [3, 96, 5120]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_5_gate_up_weight_to_fp16, x = input_85_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [3]> input_87_begin_0 = const()[name = string("input_87_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_87_end_0 = const()[name = string("input_87_end_0"), val = tensor<int32, [3]>([3, 96, 2560])];
            tensor<bool, [3]> input_87_end_mask_0 = const()[name = string("input_87_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 2560]> input_87_cast_fp16 = slice_by_index(begin = input_87_begin_0, end = input_87_end_0, end_mask = input_87_end_mask_0, x = linear_22_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [3, 96, 2560]> var_6848_cast_fp16 = silu(x = input_87_cast_fp16)[name = string("op_6848_cast_fp16")];
            tensor<int32, [3]> var_6853_begin_0 = const()[name = string("op_6853_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_6853_end_0 = const()[name = string("op_6853_end_0"), val = tensor<int32, [3]>([3, 96, 5120])];
            tensor<bool, [3]> var_6853_end_mask_0 = const()[name = string("op_6853_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 2560]> var_6853_cast_fp16 = slice_by_index(begin = var_6853_begin_0, end = var_6853_end_0, end_mask = var_6853_end_mask_0, x = linear_22_cast_fp16)[name = string("op_6853_cast_fp16")];
            tensor<fp16, [3, 96, 2560]> input_89_cast_fp16 = mul(x = var_6848_cast_fp16, y = var_6853_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_5_down_weight_to_fp16 = const()[name = string("layers_5_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91746944)))];
            tensor<fp16, [3, 96, 640]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_down_weight_to_fp16, x = input_89_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<fp16, [3, 96, 640]> x_61_cast_fp16 = add(x = x_57_cast_fp16, y = linear_23_cast_fp16)[name = string("x_61_cast_fp16")];
            fp16 var_6866_to_fp16 = const()[name = string("op_6866_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [3, 96, 640]> x_63_cast_fp16 = mul(x = x_61_cast_fp16, y = var_6866_to_fp16)[name = string("x_63_cast_fp16")];
            fp16 var_6864_promoted_to_fp16 = const()[name = string("op_6864_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> var_6868_cast_fp16 = pow(x = x_63_cast_fp16, y = var_6864_promoted_to_fp16)[name = string("op_6868_cast_fp16")];
            tensor<int32, [1]> var_6870_axes_0 = const()[name = string("op_6870_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6870_keep_dims_0 = const()[name = string("op_6870_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 1]> var_6870_cast_fp16 = reduce_mean(axes = var_6870_axes_0, keep_dims = var_6870_keep_dims_0, x = var_6868_cast_fp16)[name = string("op_6870_cast_fp16")];
            fp16 var_6871_to_fp16 = const()[name = string("op_6871_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 96, 1]> var_6872_cast_fp16 = add(x = var_6870_cast_fp16, y = var_6871_to_fp16)[name = string("op_6872_cast_fp16")];
            fp32 var_6873_epsilon_0 = const()[name = string("op_6873_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 1]> var_6873_cast_fp16 = rsqrt(epsilon = var_6873_epsilon_0, x = var_6872_cast_fp16)[name = string("op_6873_cast_fp16")];
            tensor<fp16, [3, 96, 640]> var_6874_cast_fp16 = mul(x = x_63_cast_fp16, y = var_6873_cast_fp16)[name = string("op_6874_cast_fp16")];
            tensor<fp16, [640]> layers_6_n1_w1_to_fp16 = const()[name = string("layers_6_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95023808)))];
            tensor<fp16, [3, 96, 640]> input_91_cast_fp16 = mul(x = var_6874_cast_fp16, y = layers_6_n1_w1_to_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [1544, 640]> layers_6_qkvzba_weight_to_fp16 = const()[name = string("layers_6_qkvzba_weight_to_fp16"), val = tensor<fp16, [1544, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95025152)))];
            tensor<fp16, [3, 96, 1544]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_qkvzba_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<int32, [3]> qkv_11_begin_0 = const()[name = string("qkv_11_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> qkv_11_end_0 = const()[name = string("qkv_11_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> qkv_11_end_mask_0 = const()[name = string("qkv_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 1024]> qkv_11_cast_fp16 = slice_by_index(begin = qkv_11_begin_0, end = qkv_11_end_0, end_mask = qkv_11_end_mask_0, x = linear_24_cast_fp16)[name = string("qkv_11_cast_fp16")];
            tensor<int32, [3]> z_begin_0 = const()[name = string("z_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> z_end_0 = const()[name = string("z_end_0"), val = tensor<int32, [3]>([3, 96, 1536])];
            tensor<bool, [3]> z_end_mask_0 = const()[name = string("z_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 512]> z_cast_fp16 = slice_by_index(begin = z_begin_0, end = z_end_0, end_mask = z_end_mask_0, x = linear_24_cast_fp16)[name = string("z_cast_fp16")];
            tensor<int32, [3]> b_begin_0 = const()[name = string("b_begin_0"), val = tensor<int32, [3]>([0, 0, 1536])];
            tensor<int32, [3]> b_end_0 = const()[name = string("b_end_0"), val = tensor<int32, [3]>([3, 96, 1540])];
            tensor<bool, [3]> b_end_mask_0 = const()[name = string("b_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 4]> b_cast_fp16 = slice_by_index(begin = b_begin_0, end = b_end_0, end_mask = b_end_mask_0, x = linear_24_cast_fp16)[name = string("b_cast_fp16")];
            tensor<int32, [3]> a_begin_0 = const()[name = string("a_begin_0"), val = tensor<int32, [3]>([0, 0, 1540])];
            tensor<int32, [3]> a_end_0 = const()[name = string("a_end_0"), val = tensor<int32, [3]>([3, 96, 1])];
            tensor<bool, [3]> a_end_mask_0 = const()[name = string("a_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 4]> a_cast_fp16 = slice_by_index(begin = a_begin_0, end = a_end_0, end_mask = a_end_mask_0, x = linear_24_cast_fp16)[name = string("a_cast_fp16")];
            int32 var_6913 = const()[name = string("op_6913"), val = int32(1)];
            bool xp_interleave_0 = const()[name = string("xp_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 99, 1024]> xp_cast_fp16 = concat(axis = var_6913, interleave = xp_interleave_0, values = (zpad_to_fp16, qkv_11_cast_fp16))[name = string("xp_cast_fp16")];
            tensor<int32, [3]> var_6924_begin_0 = const()[name = string("op_6924_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_6924_end_0 = const()[name = string("op_6924_end_0"), val = tensor<int32, [3]>([3, 1, 1024])];
            tensor<bool, [3]> var_6924_end_mask_0 = const()[name = string("op_6924_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 1024]> var_6924_cast_fp16 = slice_by_index(begin = var_6924_begin_0, end = var_6924_end_0, end_mask = var_6924_end_mask_0, x = xp_cast_fp16)[name = string("op_6924_cast_fp16")];
            tensor<fp16, [1024]> var_6932_to_fp16 = const()[name = string("op_6932_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97001536)))];
            tensor<fp16, [3, 96, 1024]> var_6933_cast_fp16 = mul(x = var_6924_cast_fp16, y = var_6932_to_fp16)[name = string("op_6933_cast_fp16")];
            tensor<int32, [3]> var_6943_begin_0 = const()[name = string("op_6943_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_6943_end_0 = const()[name = string("op_6943_end_0"), val = tensor<int32, [3]>([3, 98, 1024])];
            tensor<bool, [3]> var_6943_end_mask_0 = const()[name = string("op_6943_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_6943_cast_fp16 = slice_by_index(begin = var_6943_begin_0, end = var_6943_end_0, end_mask = var_6943_end_mask_0, x = xp_cast_fp16)[name = string("op_6943_cast_fp16")];
            tensor<fp16, [1024]> var_6951_to_fp16 = const()[name = string("op_6951_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97003648)))];
            tensor<fp16, [3, 96, 1024]> var_6952_cast_fp16 = mul(x = var_6943_cast_fp16, y = var_6951_to_fp16)[name = string("op_6952_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> var_6954_cast_fp16 = add(x = var_6933_cast_fp16, y = var_6952_cast_fp16)[name = string("op_6954_cast_fp16")];
            tensor<int32, [3]> var_6964_begin_0 = const()[name = string("op_6964_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_6964_end_0 = const()[name = string("op_6964_end_0"), val = tensor<int32, [3]>([3, 97, 1024])];
            tensor<bool, [3]> var_6964_end_mask_0 = const()[name = string("op_6964_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_6964_cast_fp16 = slice_by_index(begin = var_6964_begin_0, end = var_6964_end_0, end_mask = var_6964_end_mask_0, x = xp_cast_fp16)[name = string("op_6964_cast_fp16")];
            tensor<fp16, [1024]> var_6972_to_fp16 = const()[name = string("op_6972_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97005760)))];
            tensor<fp16, [3, 96, 1024]> var_6973_cast_fp16 = mul(x = var_6964_cast_fp16, y = var_6972_to_fp16)[name = string("op_6973_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> var_6975_cast_fp16 = add(x = var_6954_cast_fp16, y = var_6973_cast_fp16)[name = string("op_6975_cast_fp16")];
            tensor<int32, [3]> var_6985_begin_0 = const()[name = string("op_6985_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_6985_end_0 = const()[name = string("op_6985_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> var_6985_end_mask_0 = const()[name = string("op_6985_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 96, 1024]> var_6985_cast_fp16 = slice_by_index(begin = var_6985_begin_0, end = var_6985_end_0, end_mask = var_6985_end_mask_0, x = xp_cast_fp16)[name = string("op_6985_cast_fp16")];
            tensor<fp16, [1024]> var_6993_to_fp16 = const()[name = string("op_6993_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97007872)))];
            tensor<fp16, [3, 96, 1024]> var_6994_cast_fp16 = mul(x = var_6985_cast_fp16, y = var_6993_to_fp16)[name = string("op_6994_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> input_93_cast_fp16 = add(x = var_6975_cast_fp16, y = var_6994_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [3, 96, 1024]> c_cast_fp16 = silu(x = input_93_cast_fp16)[name = string("c_cast_fp16")];
            tensor<int32, [3]> var_7002_begin_0 = const()[name = string("op_7002_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_7002_end_0 = const()[name = string("op_7002_end_0"), val = tensor<int32, [3]>([3, 96, 256])];
            tensor<bool, [3]> var_7002_end_mask_0 = const()[name = string("op_7002_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 256]> var_7002_cast_fp16 = slice_by_index(begin = var_7002_begin_0, end = var_7002_end_0, end_mask = var_7002_end_mask_0, x = c_cast_fp16)[name = string("op_7002_cast_fp16")];
            tensor<int32, [4]> var_7007 = const()[name = string("op_7007"), val = tensor<int32, [4]>([3, 96, 2, 128])];
            tensor<fp16, [3, 96, 2, 128]> q_103_cast_fp16 = reshape(shape = var_7007, x = var_7002_cast_fp16)[name = string("q_103_cast_fp16")];
            tensor<int32, [3]> var_7013_begin_0 = const()[name = string("op_7013_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> var_7013_end_0 = const()[name = string("op_7013_end_0"), val = tensor<int32, [3]>([3, 96, 512])];
            tensor<bool, [3]> var_7013_end_mask_0 = const()[name = string("op_7013_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 256]> var_7013_cast_fp16 = slice_by_index(begin = var_7013_begin_0, end = var_7013_end_0, end_mask = var_7013_end_mask_0, x = c_cast_fp16)[name = string("op_7013_cast_fp16")];
            tensor<int32, [4]> var_7018 = const()[name = string("op_7018"), val = tensor<int32, [4]>([3, 96, 2, 128])];
            tensor<fp16, [3, 96, 2, 128]> k_105_cast_fp16 = reshape(shape = var_7018, x = var_7013_cast_fp16)[name = string("k_105_cast_fp16")];
            tensor<int32, [3]> var_7024_begin_0 = const()[name = string("op_7024_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> var_7024_end_0 = const()[name = string("op_7024_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> var_7024_end_mask_0 = const()[name = string("op_7024_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 512]> var_7024_cast_fp16 = slice_by_index(begin = var_7024_begin_0, end = var_7024_end_0, end_mask = var_7024_end_mask_0, x = c_cast_fp16)[name = string("op_7024_cast_fp16")];
            tensor<int32, [4]> var_7029 = const()[name = string("op_7029"), val = tensor<int32, [4]>([3, 96, 4, 128])];
            tensor<fp16, [3, 96, 4, 128]> var_7030_cast_fp16 = reshape(shape = var_7029, x = var_7024_cast_fp16)[name = string("op_7030_cast_fp16")];
            tensor<int32, [4]> var_7035 = const()[name = string("op_7035"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7037_to_fp16 = const()[name = string("op_7037_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 96, 2, 128]> q_105_cast_fp16 = mul(x = q_103_cast_fp16, y = var_7037_to_fp16)[name = string("q_105_cast_fp16")];
            fp16 var_7039_to_fp16 = const()[name = string("op_7039_to_fp16"), val = fp16(0x1p+4)];
            tensor<fp16, [3, 96, 2, 128]> k_107_cast_fp16 = mul(x = k_105_cast_fp16, y = var_7039_to_fp16)[name = string("k_107_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_7041_cast_fp16 = mul(x = q_105_cast_fp16, y = q_105_cast_fp16)[name = string("op_7041_cast_fp16")];
            tensor<int32, [1]> var_7046_axes_0 = const()[name = string("op_7046_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7046_keep_dims_0 = const()[name = string("op_7046_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 2, 1]> var_7046_cast_fp16 = reduce_mean(axes = var_7046_axes_0, keep_dims = var_7046_keep_dims_0, x = var_7041_cast_fp16)[name = string("op_7046_cast_fp16")];
            fp16 var_7047_promoted_to_fp16 = const()[name = string("op_7047_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 96, 2, 1]> var_7048_cast_fp16 = mul(x = var_7046_cast_fp16, y = var_7047_promoted_to_fp16)[name = string("op_7048_cast_fp16")];
            fp16 var_7050_to_fp16 = const()[name = string("op_7050_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 96, 2, 1]> var_7051_cast_fp16 = add(x = var_7048_cast_fp16, y = var_7050_to_fp16)[name = string("op_7051_cast_fp16")];
            fp32 var_7052_epsilon_0 = const()[name = string("op_7052_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 2, 1]> var_7052_cast_fp16 = rsqrt(epsilon = var_7052_epsilon_0, x = var_7051_cast_fp16)[name = string("op_7052_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_7053_cast_fp16 = mul(x = q_105_cast_fp16, y = var_7052_cast_fp16)[name = string("op_7053_cast_fp16")];
            fp16 var_7054_to_fp16 = const()[name = string("op_7054_to_fp16"), val = fp16(0x1.6ap+0)];
            tensor<fp16, [3, 96, 2, 128]> q_107_cast_fp16 = mul(x = var_7053_cast_fp16, y = var_7054_to_fp16)[name = string("q_107_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> var_7056_cast_fp16 = mul(x = k_107_cast_fp16, y = k_107_cast_fp16)[name = string("op_7056_cast_fp16")];
            tensor<int32, [1]> var_7061_axes_0 = const()[name = string("op_7061_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7061_keep_dims_0 = const()[name = string("op_7061_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 2, 1]> var_7061_cast_fp16 = reduce_mean(axes = var_7061_axes_0, keep_dims = var_7061_keep_dims_0, x = var_7056_cast_fp16)[name = string("op_7061_cast_fp16")];
            fp16 var_7062_promoted_to_fp16 = const()[name = string("op_7062_promoted_to_fp16"), val = fp16(0x1p+7)];
            tensor<fp16, [3, 96, 2, 1]> var_7063_cast_fp16 = mul(x = var_7061_cast_fp16, y = var_7062_promoted_to_fp16)[name = string("op_7063_cast_fp16")];
            fp16 var_7065_to_fp16 = const()[name = string("op_7065_to_fp16"), val = fp16(0x1.0c8p-12)];
            tensor<fp16, [3, 96, 2, 1]> var_7066_cast_fp16 = add(x = var_7063_cast_fp16, y = var_7065_to_fp16)[name = string("op_7066_cast_fp16")];
            fp32 var_7067_epsilon_0 = const()[name = string("op_7067_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 2, 1]> var_7067_cast_fp16 = rsqrt(epsilon = var_7067_epsilon_0, x = var_7066_cast_fp16)[name = string("op_7067_cast_fp16")];
            tensor<fp16, [3, 96, 2, 128]> k_109_cast_fp16 = mul(x = k_107_cast_fp16, y = var_7067_cast_fp16)[name = string("k_109_cast_fp16")];
            tensor<int32, [4]> var_7073 = const()[name = string("op_7073"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_7086_axes_0 = const()[name = string("op_7086_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 96, 128]> var_7074_cast_fp16 = transpose(perm = var_7073, x = q_107_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [3, 2, 1, 96, 128]> var_7086_cast_fp16 = expand_dims(axes = var_7086_axes_0, x = var_7074_cast_fp16)[name = string("op_7086_cast_fp16")];
            tensor<int32, [5]> var_7094_reps_0 = const()[name = string("op_7094_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 96, 128]> var_7094_cast_fp16 = tile(reps = var_7094_reps_0, x = var_7086_cast_fp16)[name = string("op_7094_cast_fp16")];
            tensor<int32, [4]> var_7099 = const()[name = string("op_7099"), val = tensor<int32, [4]>([3, 4, 96, 128])];
            tensor<fp16, [3, 4, 96, 128]> q_109_cast_fp16 = reshape(shape = var_7099, x = var_7094_cast_fp16)[name = string("q_109_cast_fp16")];
            tensor<int32, [4]> var_7105 = const()[name = string("op_7105"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> var_7118_axes_0 = const()[name = string("op_7118_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 96, 128]> var_7106_cast_fp16 = transpose(perm = var_7105, x = k_109_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [3, 2, 1, 96, 128]> var_7118_cast_fp16 = expand_dims(axes = var_7118_axes_0, x = var_7106_cast_fp16)[name = string("op_7118_cast_fp16")];
            tensor<int32, [5]> var_7126_reps_0 = const()[name = string("op_7126_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [3, 2, 2, 96, 128]> var_7126_cast_fp16 = tile(reps = var_7126_reps_0, x = var_7118_cast_fp16)[name = string("op_7126_cast_fp16")];
            tensor<int32, [4]> var_7131 = const()[name = string("op_7131"), val = tensor<int32, [4]>([3, 4, 96, 128])];
            tensor<fp16, [3, 4, 96, 128]> k_111_cast_fp16 = reshape(shape = var_7131, x = var_7126_cast_fp16)[name = string("k_111_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_7133_cast_fp16 = sigmoid(x = b_cast_fp16)[name = string("op_7133_cast_fp16")];
            tensor<int32, [3]> var_7137 = const()[name = string("op_7137"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [4]> layers_6_dt_bias_to_fp16 = const()[name = string("layers_6_dt_bias_to_fp16"), val = tensor<fp16, [4]>([0x1p+0, 0x1p+0, 0x1p+0, 0x1p+0])];
            tensor<fp16, [3, 96, 4]> input_95_cast_fp16 = add(x = a_cast_fp16, y = layers_6_dt_bias_to_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_7141_cast_fp16 = relu(x = input_95_cast_fp16)[name = string("op_7141_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_7142_cast_fp16 = abs(x = input_95_cast_fp16)[name = string("op_7142_cast_fp16")];
            fp16 const_37_promoted_to_fp16 = const()[name = string("const_37_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 96, 4]> var_7143_cast_fp16 = mul(x = var_7142_cast_fp16, y = const_37_promoted_to_fp16)[name = string("op_7143_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_7144_cast_fp16 = exp(x = var_7143_cast_fp16)[name = string("op_7144_cast_fp16")];
            fp16 var_7146_to_fp16 = const()[name = string("op_7146_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [3, 96, 4]> var_7147_cast_fp16 = add(x = var_7144_cast_fp16, y = var_7146_to_fp16)[name = string("op_7147_cast_fp16")];
            fp32 var_7148_epsilon_0 = const()[name = string("op_7148_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 96, 4]> var_7148_cast_fp16 = log(epsilon = var_7148_epsilon_0, x = var_7147_cast_fp16)[name = string("op_7148_cast_fp16")];
            tensor<fp16, [3, 96, 4]> var_7150_cast_fp16 = add(x = var_7141_cast_fp16, y = var_7148_cast_fp16)[name = string("op_7150_cast_fp16")];
            tensor<fp16, [4]> layers_6_negA_to_fp16 = const()[name = string("layers_6_negA_to_fp16"), val = tensor<fp16, [4]>([-0x1.07cp+1, -0x1.eccp+3, -0x1.63p+3, -0x1.728p+1])];
            tensor<fp16, [3, 96, 4]> var_7151_cast_fp16 = mul(x = layers_6_negA_to_fp16, y = var_7150_cast_fp16)[name = string("op_7151_cast_fp16")];
            tensor<int32, [3]> var_7155 = const()[name = string("op_7155"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> var_7158_axes_0 = const()[name = string("op_7158_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 96]> beta_cast_fp16 = transpose(perm = var_7137, x = var_7133_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [3, 4, 96, 1]> var_7158_cast_fp16 = expand_dims(axes = var_7158_axes_0, x = beta_cast_fp16)[name = string("op_7158_cast_fp16")];
            fp16 var_7159_to_fp16 = const()[name = string("op_7159_to_fp16"), val = fp16(0x1p+8)];
            tensor<fp16, [3, 4, 96, 1]> var_7160_cast_fp16 = mul(x = var_7158_cast_fp16, y = var_7159_to_fp16)[name = string("op_7160_cast_fp16")];
            tensor<fp16, [3, 4, 96, 128]> v_15_cast_fp16 = transpose(perm = var_7035, x = var_7030_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [3, 4, 96, 128]> vb_71_cast_fp16 = mul(x = v_15_cast_fp16, y = var_7160_cast_fp16)[name = string("vb_71_cast_fp16")];
            tensor<int32, [4]> q_111_begin_0 = const()[name = string("q_111_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> q_111_end_0 = const()[name = string("q_111_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> q_111_end_mask_0 = const()[name = string("q_111_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_111_cast_fp16 = slice_by_index(begin = q_111_begin_0, end = q_111_end_0, end_mask = q_111_end_mask_0, x = q_109_cast_fp16)[name = string("q_111_cast_fp16")];
            tensor<int32, [4]> k_113_begin_0 = const()[name = string("k_113_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> k_113_end_0 = const()[name = string("k_113_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> k_113_end_mask_0 = const()[name = string("k_113_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_113_cast_fp16 = slice_by_index(begin = k_113_begin_0, end = k_113_end_0, end_mask = k_113_end_mask_0, x = k_111_cast_fp16)[name = string("k_113_cast_fp16")];
            tensor<fp16, [3, 4, 96, 128]> var_7194_cast_fp16 = mul(x = k_111_cast_fp16, y = var_7158_cast_fp16)[name = string("op_7194_cast_fp16")];
            tensor<int32, [4]> kb_61_begin_0 = const()[name = string("kb_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> kb_61_end_0 = const()[name = string("kb_61_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> kb_61_end_mask_0 = const()[name = string("kb_61_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_61_cast_fp16 = slice_by_index(begin = kb_61_begin_0, end = kb_61_end_0, end_mask = kb_61_end_mask_0, x = var_7194_cast_fp16)[name = string("kb_61_cast_fp16")];
            tensor<int32, [4]> vb_73_begin_0 = const()[name = string("vb_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> vb_73_end_0 = const()[name = string("vb_73_end_0"), val = tensor<int32, [4]>([3, 4, 16, 128])];
            tensor<bool, [4]> vb_73_end_mask_0 = const()[name = string("vb_73_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_73_cast_fp16 = slice_by_index(begin = vb_73_begin_0, end = vb_73_end_0, end_mask = vb_73_end_mask_0, x = vb_71_cast_fp16)[name = string("vb_73_cast_fp16")];
            tensor<int32, [3]> gdec_73_begin_0 = const()[name = string("gdec_73_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> gdec_73_end_0 = const()[name = string("gdec_73_end_0"), val = tensor<int32, [3]>([3, 4, 16])];
            tensor<bool, [3]> gdec_73_end_mask_0 = const()[name = string("gdec_73_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 96]> gdec_71_cast_fp16 = transpose(perm = var_7155, x = var_7151_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [3, 4, 16]> gdec_73_cast_fp16 = slice_by_index(begin = gdec_73_begin_0, end = gdec_73_end_0, end_mask = gdec_73_end_mask_0, x = gdec_71_cast_fp16)[name = string("gdec_73_cast_fp16")];
            tensor<int32, [2]> var_7242 = const()[name = string("op_7242"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_7243_cast_fp16 = reshape(shape = var_7242, x = gdec_73_cast_fp16)[name = string("op_7243_cast_fp16")];
            tensor<fp16, [12, 256]> var_7244_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_7243_cast_fp16)[name = string("op_7244_cast_fp16")];
            tensor<int32, [4]> var_7249 = const()[name = string("op_7249"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_61_cast_fp16 = reshape(shape = var_7249, x = var_7244_cast_fp16)[name = string("P_61_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_7251_cast_fp16 = exp(x = P_61_cast_fp16)[name = string("op_7251_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_61_cast_fp16 = mul(x = var_7251_cast_fp16, y = tril_incl_to_fp16)[name = string("D_61_cast_fp16")];
            bool kk_61_transpose_x_1 = const()[name = string("kk_61_transpose_x_1"), val = bool(false)];
            bool kk_61_transpose_y_1 = const()[name = string("kk_61_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_61_cast_fp16 = matmul(transpose_x = kk_61_transpose_x_1, transpose_y = kk_61_transpose_y_1, x = kb_61_cast_fp16, y = k_113_cast_fp16)[name = string("kk_61_cast_fp16")];
            fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_7257_cast_fp16 = mul(x = kk_61_cast_fp16, y = const_38_promoted_to_fp16)[name = string("op_7257_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_7258_cast_fp16 = mul(x = var_7257_cast_fp16, y = D_61_cast_fp16)[name = string("op_7258_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_61_cast_fp16 = mul(x = var_7258_cast_fp16, y = cum_tail_to_fp16)[name = string("N_61_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_7260_cast_fp16 = mul(x = D_61_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_7260_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_61_cast_fp16 = mul(x = kk_61_cast_fp16, y = var_7260_cast_fp16)[name = string("Na_61_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_241_cast_fp16 = add(x = eye_to_fp16, y = N_61_cast_fp16)[name = string("Minv_241_cast_fp16")];
            bool Np_181_transpose_x_0 = const()[name = string("Np_181_transpose_x_0"), val = bool(false)];
            bool Np_181_transpose_y_0 = const()[name = string("Np_181_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_181_cast_fp16 = matmul(transpose_x = Np_181_transpose_x_0, transpose_y = Np_181_transpose_y_0, x = N_61_cast_fp16, y = Na_61_cast_fp16)[name = string("Np_181_cast_fp16")];
            bool Npa_121_transpose_x_0 = const()[name = string("Npa_121_transpose_x_0"), val = bool(false)];
            bool Npa_121_transpose_y_0 = const()[name = string("Npa_121_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_121_cast_fp16 = matmul(transpose_x = Npa_121_transpose_x_0, transpose_y = Npa_121_transpose_y_0, x = Na_61_cast_fp16, y = N_61_cast_fp16)[name = string("Npa_121_cast_fp16")];
            bool var_7266_transpose_x_0 = const()[name = string("op_7266_transpose_x_0"), val = bool(false)];
            bool var_7266_transpose_y_0 = const()[name = string("op_7266_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_7266_cast_fp16 = matmul(transpose_x = var_7266_transpose_x_0, transpose_y = var_7266_transpose_y_0, x = Minv_241_cast_fp16, y = Np_181_cast_fp16)[name = string("op_7266_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_243_cast_fp16 = add(x = Minv_241_cast_fp16, y = var_7266_cast_fp16)[name = string("Minv_243_cast_fp16")];
            bool Np_183_transpose_x_0 = const()[name = string("Np_183_transpose_x_0"), val = bool(false)];
            bool Np_183_transpose_y_0 = const()[name = string("Np_183_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_183_cast_fp16 = matmul(transpose_x = Np_183_transpose_x_0, transpose_y = Np_183_transpose_y_0, x = Np_181_cast_fp16, y = Npa_121_cast_fp16)[name = string("Np_183_cast_fp16")];
            bool Npa_123_transpose_x_0 = const()[name = string("Npa_123_transpose_x_0"), val = bool(false)];
            bool Npa_123_transpose_y_0 = const()[name = string("Npa_123_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_123_cast_fp16 = matmul(transpose_x = Npa_123_transpose_x_0, transpose_y = Npa_123_transpose_y_0, x = Npa_121_cast_fp16, y = Np_181_cast_fp16)[name = string("Npa_123_cast_fp16")];
            bool var_7271_transpose_x_0 = const()[name = string("op_7271_transpose_x_0"), val = bool(false)];
            bool var_7271_transpose_y_0 = const()[name = string("op_7271_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_7271_cast_fp16 = matmul(transpose_x = var_7271_transpose_x_0, transpose_y = var_7271_transpose_y_0, x = Minv_243_cast_fp16, y = Np_183_cast_fp16)[name = string("op_7271_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_245_cast_fp16 = add(x = Minv_243_cast_fp16, y = var_7271_cast_fp16)[name = string("Minv_245_cast_fp16")];
            bool Np_185_transpose_x_0 = const()[name = string("Np_185_transpose_x_0"), val = bool(false)];
            bool Np_185_transpose_y_0 = const()[name = string("Np_185_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_185_cast_fp16 = matmul(transpose_x = Np_185_transpose_x_0, transpose_y = Np_185_transpose_y_0, x = Np_183_cast_fp16, y = Npa_123_cast_fp16)[name = string("Np_185_cast_fp16")];
            bool var_7275_transpose_x_0 = const()[name = string("op_7275_transpose_x_0"), val = bool(false)];
            bool var_7275_transpose_y_0 = const()[name = string("op_7275_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_7275_cast_fp16 = matmul(transpose_x = var_7275_transpose_x_0, transpose_y = var_7275_transpose_y_0, x = Minv_245_cast_fp16, y = Np_185_cast_fp16)[name = string("op_7275_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_247_cast_fp16 = add(x = Minv_245_cast_fp16, y = var_7275_cast_fp16)[name = string("Minv_247_cast_fp16")];
            bool u_61_transpose_x_0 = const()[name = string("u_61_transpose_x_0"), val = bool(false)];
            bool u_61_transpose_y_0 = const()[name = string("u_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_61_cast_fp16 = matmul(transpose_x = u_61_transpose_x_0, transpose_y = u_61_transpose_y_0, x = Minv_247_cast_fp16, y = vb_73_cast_fp16)[name = string("u_61_cast_fp16")];
            bool var_7282_transpose_x_1 = const()[name = string("op_7282_transpose_x_1"), val = bool(false)];
            bool var_7282_transpose_y_1 = const()[name = string("op_7282_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_7282_cast_fp16 = matmul(transpose_x = var_7282_transpose_x_1, transpose_y = var_7282_transpose_y_1, x = q_111_cast_fp16, y = k_113_cast_fp16)[name = string("op_7282_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_61_cast_fp16 = mul(x = var_7282_cast_fp16, y = D_61_cast_fp16)[name = string("A_61_cast_fp16")];
            bool var_7284_transpose_x_0 = const()[name = string("op_7284_transpose_x_0"), val = bool(false)];
            bool var_7284_transpose_y_0 = const()[name = string("op_7284_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_7284_cast_fp16 = matmul(transpose_x = var_7284_transpose_x_0, transpose_y = var_7284_transpose_y_0, x = A_61_cast_fp16, y = u_61_cast_fp16)[name = string("op_7284_cast_fp16")];
            tensor<fp16, [16]> var_7285_bias_0_to_fp16 = const()[name = string("op_7285_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_7285_cast_fp16 = linear(bias = var_7285_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_73_cast_fp16)[name = string("op_7285_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_51_cast_fp16 = exp(x = var_7285_cast_fp16)[name = string("tail_51_cast_fp16")];
            tensor<int32, [1]> var_7288_axes_0 = const()[name = string("op_7288_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_7288_cast_fp16 = expand_dims(axes = var_7288_axes_0, x = tail_51_cast_fp16)[name = string("op_7288_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_7289_cast_fp16 = mul(x = k_113_cast_fp16, y = var_7288_cast_fp16)[name = string("op_7289_cast_fp16")];
            bool S_51_transpose_x_1 = const()[name = string("S_51_transpose_x_1"), val = bool(true)];
            bool S_51_transpose_y_1 = const()[name = string("S_51_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> S_51_cast_fp16 = matmul(transpose_x = S_51_transpose_x_1, transpose_y = S_51_transpose_y_1, x = var_7289_cast_fp16, y = u_61_cast_fp16)[name = string("S_51_cast_fp16")];
            tensor<int32, [4]> q_113_begin_0 = const()[name = string("q_113_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> q_113_end_0 = const()[name = string("q_113_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> q_113_end_mask_0 = const()[name = string("q_113_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_113_cast_fp16 = slice_by_index(begin = q_113_begin_0, end = q_113_end_0, end_mask = q_113_end_mask_0, x = q_109_cast_fp16)[name = string("q_113_cast_fp16")];
            tensor<int32, [4]> k_115_begin_0 = const()[name = string("k_115_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> k_115_end_0 = const()[name = string("k_115_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> k_115_end_mask_0 = const()[name = string("k_115_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_115_cast_fp16 = slice_by_index(begin = k_115_begin_0, end = k_115_end_0, end_mask = k_115_end_mask_0, x = k_111_cast_fp16)[name = string("k_115_cast_fp16")];
            tensor<int32, [4]> kb_63_begin_0 = const()[name = string("kb_63_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> kb_63_end_0 = const()[name = string("kb_63_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> kb_63_end_mask_0 = const()[name = string("kb_63_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_63_cast_fp16 = slice_by_index(begin = kb_63_begin_0, end = kb_63_end_0, end_mask = kb_63_end_mask_0, x = var_7194_cast_fp16)[name = string("kb_63_cast_fp16")];
            tensor<int32, [4]> vb_75_begin_0 = const()[name = string("vb_75_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> vb_75_end_0 = const()[name = string("vb_75_end_0"), val = tensor<int32, [4]>([3, 4, 32, 128])];
            tensor<bool, [4]> vb_75_end_mask_0 = const()[name = string("vb_75_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_75_cast_fp16 = slice_by_index(begin = vb_75_begin_0, end = vb_75_end_0, end_mask = vb_75_end_mask_0, x = vb_71_cast_fp16)[name = string("vb_75_cast_fp16")];
            tensor<int32, [3]> gdec_75_begin_0 = const()[name = string("gdec_75_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> gdec_75_end_0 = const()[name = string("gdec_75_end_0"), val = tensor<int32, [3]>([3, 4, 32])];
            tensor<bool, [3]> gdec_75_end_mask_0 = const()[name = string("gdec_75_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_75_cast_fp16 = slice_by_index(begin = gdec_75_begin_0, end = gdec_75_end_0, end_mask = gdec_75_end_mask_0, x = gdec_71_cast_fp16)[name = string("gdec_75_cast_fp16")];
            tensor<int32, [2]> var_7374 = const()[name = string("op_7374"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_7375_cast_fp16 = reshape(shape = var_7374, x = gdec_75_cast_fp16)[name = string("op_7375_cast_fp16")];
            tensor<fp16, [12, 256]> var_7376_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_7375_cast_fp16)[name = string("op_7376_cast_fp16")];
            tensor<int32, [4]> var_7381 = const()[name = string("op_7381"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_63_cast_fp16 = reshape(shape = var_7381, x = var_7376_cast_fp16)[name = string("P_63_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_7383_cast_fp16 = exp(x = P_63_cast_fp16)[name = string("op_7383_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_63_cast_fp16 = mul(x = var_7383_cast_fp16, y = tril_incl_to_fp16)[name = string("D_63_cast_fp16")];
            bool kk_63_transpose_x_1 = const()[name = string("kk_63_transpose_x_1"), val = bool(false)];
            bool kk_63_transpose_y_1 = const()[name = string("kk_63_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_63_cast_fp16 = matmul(transpose_x = kk_63_transpose_x_1, transpose_y = kk_63_transpose_y_1, x = kb_63_cast_fp16, y = k_115_cast_fp16)[name = string("kk_63_cast_fp16")];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_7389_cast_fp16 = mul(x = kk_63_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_7389_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_7390_cast_fp16 = mul(x = var_7389_cast_fp16, y = D_63_cast_fp16)[name = string("op_7390_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_63_cast_fp16 = mul(x = var_7390_cast_fp16, y = cum_tail_to_fp16)[name = string("N_63_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_7392_cast_fp16 = mul(x = D_63_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_7392_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_63_cast_fp16 = mul(x = kk_63_cast_fp16, y = var_7392_cast_fp16)[name = string("Na_63_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_249_cast_fp16 = add(x = eye_to_fp16, y = N_63_cast_fp16)[name = string("Minv_249_cast_fp16")];
            bool Np_187_transpose_x_0 = const()[name = string("Np_187_transpose_x_0"), val = bool(false)];
            bool Np_187_transpose_y_0 = const()[name = string("Np_187_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_187_cast_fp16 = matmul(transpose_x = Np_187_transpose_x_0, transpose_y = Np_187_transpose_y_0, x = N_63_cast_fp16, y = Na_63_cast_fp16)[name = string("Np_187_cast_fp16")];
            bool Npa_125_transpose_x_0 = const()[name = string("Npa_125_transpose_x_0"), val = bool(false)];
            bool Npa_125_transpose_y_0 = const()[name = string("Npa_125_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_125_cast_fp16 = matmul(transpose_x = Npa_125_transpose_x_0, transpose_y = Npa_125_transpose_y_0, x = Na_63_cast_fp16, y = N_63_cast_fp16)[name = string("Npa_125_cast_fp16")];
            bool var_7398_transpose_x_0 = const()[name = string("op_7398_transpose_x_0"), val = bool(false)];
            bool var_7398_transpose_y_0 = const()[name = string("op_7398_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_7398_cast_fp16 = matmul(transpose_x = var_7398_transpose_x_0, transpose_y = var_7398_transpose_y_0, x = Minv_249_cast_fp16, y = Np_187_cast_fp16)[name = string("op_7398_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_251_cast_fp16 = add(x = Minv_249_cast_fp16, y = var_7398_cast_fp16)[name = string("Minv_251_cast_fp16")];
            bool Np_189_transpose_x_0 = const()[name = string("Np_189_transpose_x_0"), val = bool(false)];
            bool Np_189_transpose_y_0 = const()[name = string("Np_189_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_189_cast_fp16 = matmul(transpose_x = Np_189_transpose_x_0, transpose_y = Np_189_transpose_y_0, x = Np_187_cast_fp16, y = Npa_125_cast_fp16)[name = string("Np_189_cast_fp16")];
            bool Npa_127_transpose_x_0 = const()[name = string("Npa_127_transpose_x_0"), val = bool(false)];
            bool Npa_127_transpose_y_0 = const()[name = string("Npa_127_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_127_cast_fp16 = matmul(transpose_x = Npa_127_transpose_x_0, transpose_y = Npa_127_transpose_y_0, x = Npa_125_cast_fp16, y = Np_187_cast_fp16)[name = string("Npa_127_cast_fp16")];
            bool var_7403_transpose_x_0 = const()[name = string("op_7403_transpose_x_0"), val = bool(false)];
            bool var_7403_transpose_y_0 = const()[name = string("op_7403_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_7403_cast_fp16 = matmul(transpose_x = var_7403_transpose_x_0, transpose_y = var_7403_transpose_y_0, x = Minv_251_cast_fp16, y = Np_189_cast_fp16)[name = string("op_7403_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_253_cast_fp16 = add(x = Minv_251_cast_fp16, y = var_7403_cast_fp16)[name = string("Minv_253_cast_fp16")];
            bool Np_191_transpose_x_0 = const()[name = string("Np_191_transpose_x_0"), val = bool(false)];
            bool Np_191_transpose_y_0 = const()[name = string("Np_191_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_191_cast_fp16 = matmul(transpose_x = Np_191_transpose_x_0, transpose_y = Np_191_transpose_y_0, x = Np_189_cast_fp16, y = Npa_127_cast_fp16)[name = string("Np_191_cast_fp16")];
            bool var_7407_transpose_x_0 = const()[name = string("op_7407_transpose_x_0"), val = bool(false)];
            bool var_7407_transpose_y_0 = const()[name = string("op_7407_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_7407_cast_fp16 = matmul(transpose_x = var_7407_transpose_x_0, transpose_y = var_7407_transpose_y_0, x = Minv_253_cast_fp16, y = Np_191_cast_fp16)[name = string("op_7407_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_255_cast_fp16 = add(x = Minv_253_cast_fp16, y = var_7407_cast_fp16)[name = string("Minv_255_cast_fp16")];
            bool u_63_transpose_x_0 = const()[name = string("u_63_transpose_x_0"), val = bool(false)];
            bool u_63_transpose_y_0 = const()[name = string("u_63_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_63_cast_fp16 = matmul(transpose_x = u_63_transpose_x_0, transpose_y = u_63_transpose_y_0, x = Minv_255_cast_fp16, y = vb_75_cast_fp16)[name = string("u_63_cast_fp16")];
            bool var_7414_transpose_x_1 = const()[name = string("op_7414_transpose_x_1"), val = bool(false)];
            bool var_7414_transpose_y_1 = const()[name = string("op_7414_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_7414_cast_fp16 = matmul(transpose_x = var_7414_transpose_x_1, transpose_y = var_7414_transpose_y_1, x = q_113_cast_fp16, y = k_115_cast_fp16)[name = string("op_7414_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_63_cast_fp16 = mul(x = var_7414_cast_fp16, y = D_63_cast_fp16)[name = string("A_63_cast_fp16")];
            tensor<fp16, [16]> var_7416_bias_0_to_fp16 = const()[name = string("op_7416_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_7416_cast_fp16 = linear(bias = var_7416_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_75_cast_fp16)[name = string("op_7416_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_51_cast_fp16 = exp(x = var_7416_cast_fp16)[name = string("cum_51_cast_fp16")];
            tensor<int32, [1]> var_7419_axes_0 = const()[name = string("op_7419_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_7419_cast_fp16 = expand_dims(axes = var_7419_axes_0, x = cum_51_cast_fp16)[name = string("op_7419_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_7420_cast_fp16 = mul(x = kb_63_cast_fp16, y = var_7419_cast_fp16)[name = string("op_7420_cast_fp16")];
            bool w_63_transpose_x_0 = const()[name = string("w_63_transpose_x_0"), val = bool(false)];
            bool w_63_transpose_y_0 = const()[name = string("w_63_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_63_cast_fp16 = matmul(transpose_x = w_63_transpose_x_0, transpose_y = w_63_transpose_y_0, x = Minv_255_cast_fp16, y = var_7420_cast_fp16)[name = string("w_63_cast_fp16")];
            bool var_7422_transpose_x_0 = const()[name = string("op_7422_transpose_x_0"), val = bool(false)];
            bool var_7422_transpose_y_0 = const()[name = string("op_7422_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_7422_cast_fp16 = matmul(transpose_x = var_7422_transpose_x_0, transpose_y = var_7422_transpose_y_0, x = w_63_cast_fp16, y = S_51_cast_fp16)[name = string("op_7422_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_51_cast_fp16 = sub(x = u_63_cast_fp16, y = var_7422_cast_fp16)[name = string("vnew_51_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_7427_cast_fp16 = mul(x = q_113_cast_fp16, y = var_7419_cast_fp16)[name = string("op_7427_cast_fp16")];
            bool var_7428_transpose_x_0 = const()[name = string("op_7428_transpose_x_0"), val = bool(false)];
            bool var_7428_transpose_y_0 = const()[name = string("op_7428_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_7428_cast_fp16 = matmul(transpose_x = var_7428_transpose_x_0, transpose_y = var_7428_transpose_y_0, x = var_7427_cast_fp16, y = S_51_cast_fp16)[name = string("op_7428_cast_fp16")];
            bool var_7429_transpose_x_0 = const()[name = string("op_7429_transpose_x_0"), val = bool(false)];
            bool var_7429_transpose_y_0 = const()[name = string("op_7429_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_7429_cast_fp16 = matmul(transpose_x = var_7429_transpose_x_0, transpose_y = var_7429_transpose_y_0, x = A_63_cast_fp16, y = vnew_51_cast_fp16)[name = string("op_7429_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_7431_cast_fp16 = add(x = var_7428_cast_fp16, y = var_7429_cast_fp16)[name = string("op_7431_cast_fp16")];
            tensor<fp16, [16]> var_7432_bias_0_to_fp16 = const()[name = string("op_7432_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_7432_cast_fp16 = linear(bias = var_7432_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_75_cast_fp16)[name = string("op_7432_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_53_cast_fp16 = exp(x = var_7432_cast_fp16)[name = string("tail_53_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_105_to_fp16 = const()[name = string("transpose_105_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_7434_bias_0_to_fp16 = const()[name = string("op_7434_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_7434_cast_fp16 = linear(bias = var_7434_bias_0_to_fp16, weight = transpose_105_to_fp16, x = gdec_75_cast_fp16)[name = string("op_7434_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_7435_cast_fp16 = exp(x = var_7434_cast_fp16)[name = string("op_7435_cast_fp16")];
            tensor<int32, [1]> last_41_axes_0 = const()[name = string("last_41_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_41_cast_fp16 = expand_dims(axes = last_41_axes_0, x = var_7435_cast_fp16)[name = string("last_41_cast_fp16")];
            tensor<int32, [1]> var_7439_axes_0 = const()[name = string("op_7439_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_7439_cast_fp16 = expand_dims(axes = var_7439_axes_0, x = tail_53_cast_fp16)[name = string("op_7439_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_7440_cast_fp16 = mul(x = k_115_cast_fp16, y = var_7439_cast_fp16)[name = string("op_7440_cast_fp16")];
            bool upd_41_transpose_x_1 = const()[name = string("upd_41_transpose_x_1"), val = bool(true)];
            bool upd_41_transpose_y_1 = const()[name = string("upd_41_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_41_cast_fp16 = matmul(transpose_x = upd_41_transpose_x_1, transpose_y = upd_41_transpose_y_1, x = var_7440_cast_fp16, y = vnew_51_cast_fp16)[name = string("upd_41_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_7445_cast_fp16 = mul(x = S_51_cast_fp16, y = last_41_cast_fp16)[name = string("op_7445_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_53_cast_fp16 = add(x = var_7445_cast_fp16, y = upd_41_cast_fp16)[name = string("S_53_cast_fp16")];
            tensor<int32, [4]> q_115_begin_0 = const()[name = string("q_115_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> q_115_end_0 = const()[name = string("q_115_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> q_115_end_mask_0 = const()[name = string("q_115_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_115_cast_fp16 = slice_by_index(begin = q_115_begin_0, end = q_115_end_0, end_mask = q_115_end_mask_0, x = q_109_cast_fp16)[name = string("q_115_cast_fp16")];
            tensor<int32, [4]> k_117_begin_0 = const()[name = string("k_117_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> k_117_end_0 = const()[name = string("k_117_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> k_117_end_mask_0 = const()[name = string("k_117_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_117_cast_fp16 = slice_by_index(begin = k_117_begin_0, end = k_117_end_0, end_mask = k_117_end_mask_0, x = k_111_cast_fp16)[name = string("k_117_cast_fp16")];
            tensor<int32, [4]> kb_65_begin_0 = const()[name = string("kb_65_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> kb_65_end_0 = const()[name = string("kb_65_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> kb_65_end_mask_0 = const()[name = string("kb_65_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_65_cast_fp16 = slice_by_index(begin = kb_65_begin_0, end = kb_65_end_0, end_mask = kb_65_end_mask_0, x = var_7194_cast_fp16)[name = string("kb_65_cast_fp16")];
            tensor<int32, [4]> vb_77_begin_0 = const()[name = string("vb_77_begin_0"), val = tensor<int32, [4]>([0, 0, 32, 0])];
            tensor<int32, [4]> vb_77_end_0 = const()[name = string("vb_77_end_0"), val = tensor<int32, [4]>([3, 4, 48, 128])];
            tensor<bool, [4]> vb_77_end_mask_0 = const()[name = string("vb_77_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_77_cast_fp16 = slice_by_index(begin = vb_77_begin_0, end = vb_77_end_0, end_mask = vb_77_end_mask_0, x = vb_71_cast_fp16)[name = string("vb_77_cast_fp16")];
            tensor<int32, [3]> gdec_77_begin_0 = const()[name = string("gdec_77_begin_0"), val = tensor<int32, [3]>([0, 0, 32])];
            tensor<int32, [3]> gdec_77_end_0 = const()[name = string("gdec_77_end_0"), val = tensor<int32, [3]>([3, 4, 48])];
            tensor<bool, [3]> gdec_77_end_mask_0 = const()[name = string("gdec_77_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_77_cast_fp16 = slice_by_index(begin = gdec_77_begin_0, end = gdec_77_end_0, end_mask = gdec_77_end_mask_0, x = gdec_71_cast_fp16)[name = string("gdec_77_cast_fp16")];
            tensor<int32, [2]> var_7528 = const()[name = string("op_7528"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_7529_cast_fp16 = reshape(shape = var_7528, x = gdec_77_cast_fp16)[name = string("op_7529_cast_fp16")];
            tensor<fp16, [12, 256]> var_7530_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_7529_cast_fp16)[name = string("op_7530_cast_fp16")];
            tensor<int32, [4]> var_7535 = const()[name = string("op_7535"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_65_cast_fp16 = reshape(shape = var_7535, x = var_7530_cast_fp16)[name = string("P_65_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_7537_cast_fp16 = exp(x = P_65_cast_fp16)[name = string("op_7537_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_65_cast_fp16 = mul(x = var_7537_cast_fp16, y = tril_incl_to_fp16)[name = string("D_65_cast_fp16")];
            bool kk_65_transpose_x_1 = const()[name = string("kk_65_transpose_x_1"), val = bool(false)];
            bool kk_65_transpose_y_1 = const()[name = string("kk_65_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_65_cast_fp16 = matmul(transpose_x = kk_65_transpose_x_1, transpose_y = kk_65_transpose_y_1, x = kb_65_cast_fp16, y = k_117_cast_fp16)[name = string("kk_65_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_7543_cast_fp16 = mul(x = kk_65_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_7543_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_7544_cast_fp16 = mul(x = var_7543_cast_fp16, y = D_65_cast_fp16)[name = string("op_7544_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_65_cast_fp16 = mul(x = var_7544_cast_fp16, y = cum_tail_to_fp16)[name = string("N_65_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_7546_cast_fp16 = mul(x = D_65_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_7546_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_65_cast_fp16 = mul(x = kk_65_cast_fp16, y = var_7546_cast_fp16)[name = string("Na_65_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_257_cast_fp16 = add(x = eye_to_fp16, y = N_65_cast_fp16)[name = string("Minv_257_cast_fp16")];
            bool Np_193_transpose_x_0 = const()[name = string("Np_193_transpose_x_0"), val = bool(false)];
            bool Np_193_transpose_y_0 = const()[name = string("Np_193_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_193_cast_fp16 = matmul(transpose_x = Np_193_transpose_x_0, transpose_y = Np_193_transpose_y_0, x = N_65_cast_fp16, y = Na_65_cast_fp16)[name = string("Np_193_cast_fp16")];
            bool Npa_129_transpose_x_0 = const()[name = string("Npa_129_transpose_x_0"), val = bool(false)];
            bool Npa_129_transpose_y_0 = const()[name = string("Npa_129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_129_cast_fp16 = matmul(transpose_x = Npa_129_transpose_x_0, transpose_y = Npa_129_transpose_y_0, x = Na_65_cast_fp16, y = N_65_cast_fp16)[name = string("Npa_129_cast_fp16")];
            bool var_7552_transpose_x_0 = const()[name = string("op_7552_transpose_x_0"), val = bool(false)];
            bool var_7552_transpose_y_0 = const()[name = string("op_7552_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_7552_cast_fp16 = matmul(transpose_x = var_7552_transpose_x_0, transpose_y = var_7552_transpose_y_0, x = Minv_257_cast_fp16, y = Np_193_cast_fp16)[name = string("op_7552_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_259_cast_fp16 = add(x = Minv_257_cast_fp16, y = var_7552_cast_fp16)[name = string("Minv_259_cast_fp16")];
            bool Np_195_transpose_x_0 = const()[name = string("Np_195_transpose_x_0"), val = bool(false)];
            bool Np_195_transpose_y_0 = const()[name = string("Np_195_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_195_cast_fp16 = matmul(transpose_x = Np_195_transpose_x_0, transpose_y = Np_195_transpose_y_0, x = Np_193_cast_fp16, y = Npa_129_cast_fp16)[name = string("Np_195_cast_fp16")];
            bool Npa_131_transpose_x_0 = const()[name = string("Npa_131_transpose_x_0"), val = bool(false)];
            bool Npa_131_transpose_y_0 = const()[name = string("Npa_131_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_131_cast_fp16 = matmul(transpose_x = Npa_131_transpose_x_0, transpose_y = Npa_131_transpose_y_0, x = Npa_129_cast_fp16, y = Np_193_cast_fp16)[name = string("Npa_131_cast_fp16")];
            bool var_7557_transpose_x_0 = const()[name = string("op_7557_transpose_x_0"), val = bool(false)];
            bool var_7557_transpose_y_0 = const()[name = string("op_7557_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_7557_cast_fp16 = matmul(transpose_x = var_7557_transpose_x_0, transpose_y = var_7557_transpose_y_0, x = Minv_259_cast_fp16, y = Np_195_cast_fp16)[name = string("op_7557_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_261_cast_fp16 = add(x = Minv_259_cast_fp16, y = var_7557_cast_fp16)[name = string("Minv_261_cast_fp16")];
            bool Np_197_transpose_x_0 = const()[name = string("Np_197_transpose_x_0"), val = bool(false)];
            bool Np_197_transpose_y_0 = const()[name = string("Np_197_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_197_cast_fp16 = matmul(transpose_x = Np_197_transpose_x_0, transpose_y = Np_197_transpose_y_0, x = Np_195_cast_fp16, y = Npa_131_cast_fp16)[name = string("Np_197_cast_fp16")];
            bool var_7561_transpose_x_0 = const()[name = string("op_7561_transpose_x_0"), val = bool(false)];
            bool var_7561_transpose_y_0 = const()[name = string("op_7561_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_7561_cast_fp16 = matmul(transpose_x = var_7561_transpose_x_0, transpose_y = var_7561_transpose_y_0, x = Minv_261_cast_fp16, y = Np_197_cast_fp16)[name = string("op_7561_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_263_cast_fp16 = add(x = Minv_261_cast_fp16, y = var_7561_cast_fp16)[name = string("Minv_263_cast_fp16")];
            bool u_65_transpose_x_0 = const()[name = string("u_65_transpose_x_0"), val = bool(false)];
            bool u_65_transpose_y_0 = const()[name = string("u_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_65_cast_fp16 = matmul(transpose_x = u_65_transpose_x_0, transpose_y = u_65_transpose_y_0, x = Minv_263_cast_fp16, y = vb_77_cast_fp16)[name = string("u_65_cast_fp16")];
            bool var_7568_transpose_x_1 = const()[name = string("op_7568_transpose_x_1"), val = bool(false)];
            bool var_7568_transpose_y_1 = const()[name = string("op_7568_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_7568_cast_fp16 = matmul(transpose_x = var_7568_transpose_x_1, transpose_y = var_7568_transpose_y_1, x = q_115_cast_fp16, y = k_117_cast_fp16)[name = string("op_7568_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_65_cast_fp16 = mul(x = var_7568_cast_fp16, y = D_65_cast_fp16)[name = string("A_65_cast_fp16")];
            tensor<fp16, [16]> var_7570_bias_0_to_fp16 = const()[name = string("op_7570_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_7570_cast_fp16 = linear(bias = var_7570_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_77_cast_fp16)[name = string("op_7570_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_53_cast_fp16 = exp(x = var_7570_cast_fp16)[name = string("cum_53_cast_fp16")];
            tensor<int32, [1]> var_7573_axes_0 = const()[name = string("op_7573_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_7573_cast_fp16 = expand_dims(axes = var_7573_axes_0, x = cum_53_cast_fp16)[name = string("op_7573_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_7574_cast_fp16 = mul(x = kb_65_cast_fp16, y = var_7573_cast_fp16)[name = string("op_7574_cast_fp16")];
            bool w_65_transpose_x_0 = const()[name = string("w_65_transpose_x_0"), val = bool(false)];
            bool w_65_transpose_y_0 = const()[name = string("w_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_65_cast_fp16 = matmul(transpose_x = w_65_transpose_x_0, transpose_y = w_65_transpose_y_0, x = Minv_263_cast_fp16, y = var_7574_cast_fp16)[name = string("w_65_cast_fp16")];
            bool var_7576_transpose_x_0 = const()[name = string("op_7576_transpose_x_0"), val = bool(false)];
            bool var_7576_transpose_y_0 = const()[name = string("op_7576_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_7576_cast_fp16 = matmul(transpose_x = var_7576_transpose_x_0, transpose_y = var_7576_transpose_y_0, x = w_65_cast_fp16, y = S_53_cast_fp16)[name = string("op_7576_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_53_cast_fp16 = sub(x = u_65_cast_fp16, y = var_7576_cast_fp16)[name = string("vnew_53_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_7581_cast_fp16 = mul(x = q_115_cast_fp16, y = var_7573_cast_fp16)[name = string("op_7581_cast_fp16")];
            bool var_7582_transpose_x_0 = const()[name = string("op_7582_transpose_x_0"), val = bool(false)];
            bool var_7582_transpose_y_0 = const()[name = string("op_7582_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_7582_cast_fp16 = matmul(transpose_x = var_7582_transpose_x_0, transpose_y = var_7582_transpose_y_0, x = var_7581_cast_fp16, y = S_53_cast_fp16)[name = string("op_7582_cast_fp16")];
            bool var_7583_transpose_x_0 = const()[name = string("op_7583_transpose_x_0"), val = bool(false)];
            bool var_7583_transpose_y_0 = const()[name = string("op_7583_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_7583_cast_fp16 = matmul(transpose_x = var_7583_transpose_x_0, transpose_y = var_7583_transpose_y_0, x = A_65_cast_fp16, y = vnew_53_cast_fp16)[name = string("op_7583_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_7585_cast_fp16 = add(x = var_7582_cast_fp16, y = var_7583_cast_fp16)[name = string("op_7585_cast_fp16")];
            tensor<fp16, [16]> var_7586_bias_0_to_fp16 = const()[name = string("op_7586_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_7586_cast_fp16 = linear(bias = var_7586_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_77_cast_fp16)[name = string("op_7586_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_55_cast_fp16 = exp(x = var_7586_cast_fp16)[name = string("tail_55_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_109_to_fp16 = const()[name = string("transpose_109_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_7588_bias_0_to_fp16 = const()[name = string("op_7588_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_7588_cast_fp16 = linear(bias = var_7588_bias_0_to_fp16, weight = transpose_109_to_fp16, x = gdec_77_cast_fp16)[name = string("op_7588_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_7589_cast_fp16 = exp(x = var_7588_cast_fp16)[name = string("op_7589_cast_fp16")];
            tensor<int32, [1]> last_43_axes_0 = const()[name = string("last_43_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_43_cast_fp16 = expand_dims(axes = last_43_axes_0, x = var_7589_cast_fp16)[name = string("last_43_cast_fp16")];
            tensor<int32, [1]> var_7593_axes_0 = const()[name = string("op_7593_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_7593_cast_fp16 = expand_dims(axes = var_7593_axes_0, x = tail_55_cast_fp16)[name = string("op_7593_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_7594_cast_fp16 = mul(x = k_117_cast_fp16, y = var_7593_cast_fp16)[name = string("op_7594_cast_fp16")];
            bool upd_43_transpose_x_1 = const()[name = string("upd_43_transpose_x_1"), val = bool(true)];
            bool upd_43_transpose_y_1 = const()[name = string("upd_43_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_43_cast_fp16 = matmul(transpose_x = upd_43_transpose_x_1, transpose_y = upd_43_transpose_y_1, x = var_7594_cast_fp16, y = vnew_53_cast_fp16)[name = string("upd_43_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_7599_cast_fp16 = mul(x = S_53_cast_fp16, y = last_43_cast_fp16)[name = string("op_7599_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_55_cast_fp16 = add(x = var_7599_cast_fp16, y = upd_43_cast_fp16)[name = string("S_55_cast_fp16")];
            tensor<int32, [4]> q_117_begin_0 = const()[name = string("q_117_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> q_117_end_0 = const()[name = string("q_117_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> q_117_end_mask_0 = const()[name = string("q_117_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_117_cast_fp16 = slice_by_index(begin = q_117_begin_0, end = q_117_end_0, end_mask = q_117_end_mask_0, x = q_109_cast_fp16)[name = string("q_117_cast_fp16")];
            tensor<int32, [4]> k_119_begin_0 = const()[name = string("k_119_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> k_119_end_0 = const()[name = string("k_119_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> k_119_end_mask_0 = const()[name = string("k_119_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_119_cast_fp16 = slice_by_index(begin = k_119_begin_0, end = k_119_end_0, end_mask = k_119_end_mask_0, x = k_111_cast_fp16)[name = string("k_119_cast_fp16")];
            tensor<int32, [4]> kb_67_begin_0 = const()[name = string("kb_67_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> kb_67_end_0 = const()[name = string("kb_67_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> kb_67_end_mask_0 = const()[name = string("kb_67_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_67_cast_fp16 = slice_by_index(begin = kb_67_begin_0, end = kb_67_end_0, end_mask = kb_67_end_mask_0, x = var_7194_cast_fp16)[name = string("kb_67_cast_fp16")];
            tensor<int32, [4]> vb_79_begin_0 = const()[name = string("vb_79_begin_0"), val = tensor<int32, [4]>([0, 0, 48, 0])];
            tensor<int32, [4]> vb_79_end_0 = const()[name = string("vb_79_end_0"), val = tensor<int32, [4]>([3, 4, 64, 128])];
            tensor<bool, [4]> vb_79_end_mask_0 = const()[name = string("vb_79_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_79_cast_fp16 = slice_by_index(begin = vb_79_begin_0, end = vb_79_end_0, end_mask = vb_79_end_mask_0, x = vb_71_cast_fp16)[name = string("vb_79_cast_fp16")];
            tensor<int32, [3]> gdec_79_begin_0 = const()[name = string("gdec_79_begin_0"), val = tensor<int32, [3]>([0, 0, 48])];
            tensor<int32, [3]> gdec_79_end_0 = const()[name = string("gdec_79_end_0"), val = tensor<int32, [3]>([3, 4, 64])];
            tensor<bool, [3]> gdec_79_end_mask_0 = const()[name = string("gdec_79_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_79_cast_fp16 = slice_by_index(begin = gdec_79_begin_0, end = gdec_79_end_0, end_mask = gdec_79_end_mask_0, x = gdec_71_cast_fp16)[name = string("gdec_79_cast_fp16")];
            tensor<int32, [2]> var_7682 = const()[name = string("op_7682"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_7683_cast_fp16 = reshape(shape = var_7682, x = gdec_79_cast_fp16)[name = string("op_7683_cast_fp16")];
            tensor<fp16, [12, 256]> var_7684_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_7683_cast_fp16)[name = string("op_7684_cast_fp16")];
            tensor<int32, [4]> var_7689 = const()[name = string("op_7689"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_67_cast_fp16 = reshape(shape = var_7689, x = var_7684_cast_fp16)[name = string("P_67_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_7691_cast_fp16 = exp(x = P_67_cast_fp16)[name = string("op_7691_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_67_cast_fp16 = mul(x = var_7691_cast_fp16, y = tril_incl_to_fp16)[name = string("D_67_cast_fp16")];
            bool kk_67_transpose_x_1 = const()[name = string("kk_67_transpose_x_1"), val = bool(false)];
            bool kk_67_transpose_y_1 = const()[name = string("kk_67_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_67_cast_fp16 = matmul(transpose_x = kk_67_transpose_x_1, transpose_y = kk_67_transpose_y_1, x = kb_67_cast_fp16, y = k_119_cast_fp16)[name = string("kk_67_cast_fp16")];
            fp16 const_41_promoted_to_fp16 = const()[name = string("const_41_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_7697_cast_fp16 = mul(x = kk_67_cast_fp16, y = const_41_promoted_to_fp16)[name = string("op_7697_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_7698_cast_fp16 = mul(x = var_7697_cast_fp16, y = D_67_cast_fp16)[name = string("op_7698_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_67_cast_fp16 = mul(x = var_7698_cast_fp16, y = cum_tail_to_fp16)[name = string("N_67_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_7700_cast_fp16 = mul(x = D_67_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_7700_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_67_cast_fp16 = mul(x = kk_67_cast_fp16, y = var_7700_cast_fp16)[name = string("Na_67_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_265_cast_fp16 = add(x = eye_to_fp16, y = N_67_cast_fp16)[name = string("Minv_265_cast_fp16")];
            bool Np_199_transpose_x_0 = const()[name = string("Np_199_transpose_x_0"), val = bool(false)];
            bool Np_199_transpose_y_0 = const()[name = string("Np_199_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_199_cast_fp16 = matmul(transpose_x = Np_199_transpose_x_0, transpose_y = Np_199_transpose_y_0, x = N_67_cast_fp16, y = Na_67_cast_fp16)[name = string("Np_199_cast_fp16")];
            bool Npa_133_transpose_x_0 = const()[name = string("Npa_133_transpose_x_0"), val = bool(false)];
            bool Npa_133_transpose_y_0 = const()[name = string("Npa_133_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_133_cast_fp16 = matmul(transpose_x = Npa_133_transpose_x_0, transpose_y = Npa_133_transpose_y_0, x = Na_67_cast_fp16, y = N_67_cast_fp16)[name = string("Npa_133_cast_fp16")];
            bool var_7706_transpose_x_0 = const()[name = string("op_7706_transpose_x_0"), val = bool(false)];
            bool var_7706_transpose_y_0 = const()[name = string("op_7706_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_7706_cast_fp16 = matmul(transpose_x = var_7706_transpose_x_0, transpose_y = var_7706_transpose_y_0, x = Minv_265_cast_fp16, y = Np_199_cast_fp16)[name = string("op_7706_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_267_cast_fp16 = add(x = Minv_265_cast_fp16, y = var_7706_cast_fp16)[name = string("Minv_267_cast_fp16")];
            bool Np_201_transpose_x_0 = const()[name = string("Np_201_transpose_x_0"), val = bool(false)];
            bool Np_201_transpose_y_0 = const()[name = string("Np_201_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_201_cast_fp16 = matmul(transpose_x = Np_201_transpose_x_0, transpose_y = Np_201_transpose_y_0, x = Np_199_cast_fp16, y = Npa_133_cast_fp16)[name = string("Np_201_cast_fp16")];
            bool Npa_135_transpose_x_0 = const()[name = string("Npa_135_transpose_x_0"), val = bool(false)];
            bool Npa_135_transpose_y_0 = const()[name = string("Npa_135_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_135_cast_fp16 = matmul(transpose_x = Npa_135_transpose_x_0, transpose_y = Npa_135_transpose_y_0, x = Npa_133_cast_fp16, y = Np_199_cast_fp16)[name = string("Npa_135_cast_fp16")];
            bool var_7711_transpose_x_0 = const()[name = string("op_7711_transpose_x_0"), val = bool(false)];
            bool var_7711_transpose_y_0 = const()[name = string("op_7711_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_7711_cast_fp16 = matmul(transpose_x = var_7711_transpose_x_0, transpose_y = var_7711_transpose_y_0, x = Minv_267_cast_fp16, y = Np_201_cast_fp16)[name = string("op_7711_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_269_cast_fp16 = add(x = Minv_267_cast_fp16, y = var_7711_cast_fp16)[name = string("Minv_269_cast_fp16")];
            bool Np_203_transpose_x_0 = const()[name = string("Np_203_transpose_x_0"), val = bool(false)];
            bool Np_203_transpose_y_0 = const()[name = string("Np_203_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_203_cast_fp16 = matmul(transpose_x = Np_203_transpose_x_0, transpose_y = Np_203_transpose_y_0, x = Np_201_cast_fp16, y = Npa_135_cast_fp16)[name = string("Np_203_cast_fp16")];
            bool var_7715_transpose_x_0 = const()[name = string("op_7715_transpose_x_0"), val = bool(false)];
            bool var_7715_transpose_y_0 = const()[name = string("op_7715_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_7715_cast_fp16 = matmul(transpose_x = var_7715_transpose_x_0, transpose_y = var_7715_transpose_y_0, x = Minv_269_cast_fp16, y = Np_203_cast_fp16)[name = string("op_7715_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_271_cast_fp16 = add(x = Minv_269_cast_fp16, y = var_7715_cast_fp16)[name = string("Minv_271_cast_fp16")];
            bool u_67_transpose_x_0 = const()[name = string("u_67_transpose_x_0"), val = bool(false)];
            bool u_67_transpose_y_0 = const()[name = string("u_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_67_cast_fp16 = matmul(transpose_x = u_67_transpose_x_0, transpose_y = u_67_transpose_y_0, x = Minv_271_cast_fp16, y = vb_79_cast_fp16)[name = string("u_67_cast_fp16")];
            bool var_7722_transpose_x_1 = const()[name = string("op_7722_transpose_x_1"), val = bool(false)];
            bool var_7722_transpose_y_1 = const()[name = string("op_7722_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_7722_cast_fp16 = matmul(transpose_x = var_7722_transpose_x_1, transpose_y = var_7722_transpose_y_1, x = q_117_cast_fp16, y = k_119_cast_fp16)[name = string("op_7722_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_67_cast_fp16 = mul(x = var_7722_cast_fp16, y = D_67_cast_fp16)[name = string("A_67_cast_fp16")];
            tensor<fp16, [16]> var_7724_bias_0_to_fp16 = const()[name = string("op_7724_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_7724_cast_fp16 = linear(bias = var_7724_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_79_cast_fp16)[name = string("op_7724_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_55_cast_fp16 = exp(x = var_7724_cast_fp16)[name = string("cum_55_cast_fp16")];
            tensor<int32, [1]> var_7727_axes_0 = const()[name = string("op_7727_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_7727_cast_fp16 = expand_dims(axes = var_7727_axes_0, x = cum_55_cast_fp16)[name = string("op_7727_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_7728_cast_fp16 = mul(x = kb_67_cast_fp16, y = var_7727_cast_fp16)[name = string("op_7728_cast_fp16")];
            bool w_67_transpose_x_0 = const()[name = string("w_67_transpose_x_0"), val = bool(false)];
            bool w_67_transpose_y_0 = const()[name = string("w_67_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_67_cast_fp16 = matmul(transpose_x = w_67_transpose_x_0, transpose_y = w_67_transpose_y_0, x = Minv_271_cast_fp16, y = var_7728_cast_fp16)[name = string("w_67_cast_fp16")];
            bool var_7730_transpose_x_0 = const()[name = string("op_7730_transpose_x_0"), val = bool(false)];
            bool var_7730_transpose_y_0 = const()[name = string("op_7730_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_7730_cast_fp16 = matmul(transpose_x = var_7730_transpose_x_0, transpose_y = var_7730_transpose_y_0, x = w_67_cast_fp16, y = S_55_cast_fp16)[name = string("op_7730_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_55_cast_fp16 = sub(x = u_67_cast_fp16, y = var_7730_cast_fp16)[name = string("vnew_55_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_7735_cast_fp16 = mul(x = q_117_cast_fp16, y = var_7727_cast_fp16)[name = string("op_7735_cast_fp16")];
            bool var_7736_transpose_x_0 = const()[name = string("op_7736_transpose_x_0"), val = bool(false)];
            bool var_7736_transpose_y_0 = const()[name = string("op_7736_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_7736_cast_fp16 = matmul(transpose_x = var_7736_transpose_x_0, transpose_y = var_7736_transpose_y_0, x = var_7735_cast_fp16, y = S_55_cast_fp16)[name = string("op_7736_cast_fp16")];
            bool var_7737_transpose_x_0 = const()[name = string("op_7737_transpose_x_0"), val = bool(false)];
            bool var_7737_transpose_y_0 = const()[name = string("op_7737_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_7737_cast_fp16 = matmul(transpose_x = var_7737_transpose_x_0, transpose_y = var_7737_transpose_y_0, x = A_67_cast_fp16, y = vnew_55_cast_fp16)[name = string("op_7737_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_7739_cast_fp16 = add(x = var_7736_cast_fp16, y = var_7737_cast_fp16)[name = string("op_7739_cast_fp16")];
            tensor<fp16, [16]> var_7740_bias_0_to_fp16 = const()[name = string("op_7740_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_7740_cast_fp16 = linear(bias = var_7740_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_79_cast_fp16)[name = string("op_7740_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_57_cast_fp16 = exp(x = var_7740_cast_fp16)[name = string("tail_57_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_113_to_fp16 = const()[name = string("transpose_113_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_7742_bias_0_to_fp16 = const()[name = string("op_7742_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_7742_cast_fp16 = linear(bias = var_7742_bias_0_to_fp16, weight = transpose_113_to_fp16, x = gdec_79_cast_fp16)[name = string("op_7742_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_7743_cast_fp16 = exp(x = var_7742_cast_fp16)[name = string("op_7743_cast_fp16")];
            tensor<int32, [1]> last_45_axes_0 = const()[name = string("last_45_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_45_cast_fp16 = expand_dims(axes = last_45_axes_0, x = var_7743_cast_fp16)[name = string("last_45_cast_fp16")];
            tensor<int32, [1]> var_7747_axes_0 = const()[name = string("op_7747_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_7747_cast_fp16 = expand_dims(axes = var_7747_axes_0, x = tail_57_cast_fp16)[name = string("op_7747_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_7748_cast_fp16 = mul(x = k_119_cast_fp16, y = var_7747_cast_fp16)[name = string("op_7748_cast_fp16")];
            bool upd_45_transpose_x_1 = const()[name = string("upd_45_transpose_x_1"), val = bool(true)];
            bool upd_45_transpose_y_1 = const()[name = string("upd_45_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_45_cast_fp16 = matmul(transpose_x = upd_45_transpose_x_1, transpose_y = upd_45_transpose_y_1, x = var_7748_cast_fp16, y = vnew_55_cast_fp16)[name = string("upd_45_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_7753_cast_fp16 = mul(x = S_55_cast_fp16, y = last_45_cast_fp16)[name = string("op_7753_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_57_cast_fp16 = add(x = var_7753_cast_fp16, y = upd_45_cast_fp16)[name = string("S_57_cast_fp16")];
            tensor<int32, [4]> q_119_begin_0 = const()[name = string("q_119_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> q_119_end_0 = const()[name = string("q_119_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> q_119_end_mask_0 = const()[name = string("q_119_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> q_119_cast_fp16 = slice_by_index(begin = q_119_begin_0, end = q_119_end_0, end_mask = q_119_end_mask_0, x = q_109_cast_fp16)[name = string("q_119_cast_fp16")];
            tensor<int32, [4]> k_121_begin_0 = const()[name = string("k_121_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> k_121_end_0 = const()[name = string("k_121_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> k_121_end_mask_0 = const()[name = string("k_121_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> k_121_cast_fp16 = slice_by_index(begin = k_121_begin_0, end = k_121_end_0, end_mask = k_121_end_mask_0, x = k_111_cast_fp16)[name = string("k_121_cast_fp16")];
            tensor<int32, [4]> kb_69_begin_0 = const()[name = string("kb_69_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> kb_69_end_0 = const()[name = string("kb_69_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> kb_69_end_mask_0 = const()[name = string("kb_69_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_69_cast_fp16 = slice_by_index(begin = kb_69_begin_0, end = kb_69_end_0, end_mask = kb_69_end_mask_0, x = var_7194_cast_fp16)[name = string("kb_69_cast_fp16")];
            tensor<int32, [4]> vb_81_begin_0 = const()[name = string("vb_81_begin_0"), val = tensor<int32, [4]>([0, 0, 64, 0])];
            tensor<int32, [4]> vb_81_end_0 = const()[name = string("vb_81_end_0"), val = tensor<int32, [4]>([3, 4, 80, 128])];
            tensor<bool, [4]> vb_81_end_mask_0 = const()[name = string("vb_81_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_81_cast_fp16 = slice_by_index(begin = vb_81_begin_0, end = vb_81_end_0, end_mask = vb_81_end_mask_0, x = vb_71_cast_fp16)[name = string("vb_81_cast_fp16")];
            tensor<int32, [3]> gdec_81_begin_0 = const()[name = string("gdec_81_begin_0"), val = tensor<int32, [3]>([0, 0, 64])];
            tensor<int32, [3]> gdec_81_end_0 = const()[name = string("gdec_81_end_0"), val = tensor<int32, [3]>([3, 4, 80])];
            tensor<bool, [3]> gdec_81_end_mask_0 = const()[name = string("gdec_81_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 4, 16]> gdec_81_cast_fp16 = slice_by_index(begin = gdec_81_begin_0, end = gdec_81_end_0, end_mask = gdec_81_end_mask_0, x = gdec_71_cast_fp16)[name = string("gdec_81_cast_fp16")];
            tensor<int32, [2]> var_7836 = const()[name = string("op_7836"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_7837_cast_fp16 = reshape(shape = var_7836, x = gdec_81_cast_fp16)[name = string("op_7837_cast_fp16")];
            tensor<fp16, [12, 256]> var_7838_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_7837_cast_fp16)[name = string("op_7838_cast_fp16")];
            tensor<int32, [4]> var_7843 = const()[name = string("op_7843"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_69_cast_fp16 = reshape(shape = var_7843, x = var_7838_cast_fp16)[name = string("P_69_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_7845_cast_fp16 = exp(x = P_69_cast_fp16)[name = string("op_7845_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_69_cast_fp16 = mul(x = var_7845_cast_fp16, y = tril_incl_to_fp16)[name = string("D_69_cast_fp16")];
            bool kk_69_transpose_x_1 = const()[name = string("kk_69_transpose_x_1"), val = bool(false)];
            bool kk_69_transpose_y_1 = const()[name = string("kk_69_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_69_cast_fp16 = matmul(transpose_x = kk_69_transpose_x_1, transpose_y = kk_69_transpose_y_1, x = kb_69_cast_fp16, y = k_121_cast_fp16)[name = string("kk_69_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_7851_cast_fp16 = mul(x = kk_69_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_7851_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_7852_cast_fp16 = mul(x = var_7851_cast_fp16, y = D_69_cast_fp16)[name = string("op_7852_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_69_cast_fp16 = mul(x = var_7852_cast_fp16, y = cum_tail_to_fp16)[name = string("N_69_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_7854_cast_fp16 = mul(x = D_69_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_7854_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_69_cast_fp16 = mul(x = kk_69_cast_fp16, y = var_7854_cast_fp16)[name = string("Na_69_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_273_cast_fp16 = add(x = eye_to_fp16, y = N_69_cast_fp16)[name = string("Minv_273_cast_fp16")];
            bool Np_205_transpose_x_0 = const()[name = string("Np_205_transpose_x_0"), val = bool(false)];
            bool Np_205_transpose_y_0 = const()[name = string("Np_205_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_205_cast_fp16 = matmul(transpose_x = Np_205_transpose_x_0, transpose_y = Np_205_transpose_y_0, x = N_69_cast_fp16, y = Na_69_cast_fp16)[name = string("Np_205_cast_fp16")];
            bool Npa_137_transpose_x_0 = const()[name = string("Npa_137_transpose_x_0"), val = bool(false)];
            bool Npa_137_transpose_y_0 = const()[name = string("Npa_137_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_137_cast_fp16 = matmul(transpose_x = Npa_137_transpose_x_0, transpose_y = Npa_137_transpose_y_0, x = Na_69_cast_fp16, y = N_69_cast_fp16)[name = string("Npa_137_cast_fp16")];
            bool var_7860_transpose_x_0 = const()[name = string("op_7860_transpose_x_0"), val = bool(false)];
            bool var_7860_transpose_y_0 = const()[name = string("op_7860_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_7860_cast_fp16 = matmul(transpose_x = var_7860_transpose_x_0, transpose_y = var_7860_transpose_y_0, x = Minv_273_cast_fp16, y = Np_205_cast_fp16)[name = string("op_7860_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_275_cast_fp16 = add(x = Minv_273_cast_fp16, y = var_7860_cast_fp16)[name = string("Minv_275_cast_fp16")];
            bool Np_207_transpose_x_0 = const()[name = string("Np_207_transpose_x_0"), val = bool(false)];
            bool Np_207_transpose_y_0 = const()[name = string("Np_207_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_207_cast_fp16 = matmul(transpose_x = Np_207_transpose_x_0, transpose_y = Np_207_transpose_y_0, x = Np_205_cast_fp16, y = Npa_137_cast_fp16)[name = string("Np_207_cast_fp16")];
            bool Npa_139_transpose_x_0 = const()[name = string("Npa_139_transpose_x_0"), val = bool(false)];
            bool Npa_139_transpose_y_0 = const()[name = string("Npa_139_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_139_cast_fp16 = matmul(transpose_x = Npa_139_transpose_x_0, transpose_y = Npa_139_transpose_y_0, x = Npa_137_cast_fp16, y = Np_205_cast_fp16)[name = string("Npa_139_cast_fp16")];
            bool var_7865_transpose_x_0 = const()[name = string("op_7865_transpose_x_0"), val = bool(false)];
            bool var_7865_transpose_y_0 = const()[name = string("op_7865_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_7865_cast_fp16 = matmul(transpose_x = var_7865_transpose_x_0, transpose_y = var_7865_transpose_y_0, x = Minv_275_cast_fp16, y = Np_207_cast_fp16)[name = string("op_7865_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_277_cast_fp16 = add(x = Minv_275_cast_fp16, y = var_7865_cast_fp16)[name = string("Minv_277_cast_fp16")];
            bool Np_209_transpose_x_0 = const()[name = string("Np_209_transpose_x_0"), val = bool(false)];
            bool Np_209_transpose_y_0 = const()[name = string("Np_209_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_209_cast_fp16 = matmul(transpose_x = Np_209_transpose_x_0, transpose_y = Np_209_transpose_y_0, x = Np_207_cast_fp16, y = Npa_139_cast_fp16)[name = string("Np_209_cast_fp16")];
            bool var_7869_transpose_x_0 = const()[name = string("op_7869_transpose_x_0"), val = bool(false)];
            bool var_7869_transpose_y_0 = const()[name = string("op_7869_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_7869_cast_fp16 = matmul(transpose_x = var_7869_transpose_x_0, transpose_y = var_7869_transpose_y_0, x = Minv_277_cast_fp16, y = Np_209_cast_fp16)[name = string("op_7869_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_279_cast_fp16 = add(x = Minv_277_cast_fp16, y = var_7869_cast_fp16)[name = string("Minv_279_cast_fp16")];
            bool u_69_transpose_x_0 = const()[name = string("u_69_transpose_x_0"), val = bool(false)];
            bool u_69_transpose_y_0 = const()[name = string("u_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_69_cast_fp16 = matmul(transpose_x = u_69_transpose_x_0, transpose_y = u_69_transpose_y_0, x = Minv_279_cast_fp16, y = vb_81_cast_fp16)[name = string("u_69_cast_fp16")];
            bool var_7876_transpose_x_1 = const()[name = string("op_7876_transpose_x_1"), val = bool(false)];
            bool var_7876_transpose_y_1 = const()[name = string("op_7876_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_7876_cast_fp16 = matmul(transpose_x = var_7876_transpose_x_1, transpose_y = var_7876_transpose_y_1, x = q_119_cast_fp16, y = k_121_cast_fp16)[name = string("op_7876_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_69_cast_fp16 = mul(x = var_7876_cast_fp16, y = D_69_cast_fp16)[name = string("A_69_cast_fp16")];
            tensor<fp16, [16]> var_7878_bias_0_to_fp16 = const()[name = string("op_7878_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_7878_cast_fp16 = linear(bias = var_7878_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_81_cast_fp16)[name = string("op_7878_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_57_cast_fp16 = exp(x = var_7878_cast_fp16)[name = string("cum_57_cast_fp16")];
            tensor<int32, [1]> var_7881_axes_0 = const()[name = string("op_7881_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_7881_cast_fp16 = expand_dims(axes = var_7881_axes_0, x = cum_57_cast_fp16)[name = string("op_7881_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_7882_cast_fp16 = mul(x = kb_69_cast_fp16, y = var_7881_cast_fp16)[name = string("op_7882_cast_fp16")];
            bool w_69_transpose_x_0 = const()[name = string("w_69_transpose_x_0"), val = bool(false)];
            bool w_69_transpose_y_0 = const()[name = string("w_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_69_cast_fp16 = matmul(transpose_x = w_69_transpose_x_0, transpose_y = w_69_transpose_y_0, x = Minv_279_cast_fp16, y = var_7882_cast_fp16)[name = string("w_69_cast_fp16")];
            bool var_7884_transpose_x_0 = const()[name = string("op_7884_transpose_x_0"), val = bool(false)];
            bool var_7884_transpose_y_0 = const()[name = string("op_7884_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_7884_cast_fp16 = matmul(transpose_x = var_7884_transpose_x_0, transpose_y = var_7884_transpose_y_0, x = w_69_cast_fp16, y = S_57_cast_fp16)[name = string("op_7884_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_57_cast_fp16 = sub(x = u_69_cast_fp16, y = var_7884_cast_fp16)[name = string("vnew_57_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_7889_cast_fp16 = mul(x = q_119_cast_fp16, y = var_7881_cast_fp16)[name = string("op_7889_cast_fp16")];
            bool var_7890_transpose_x_0 = const()[name = string("op_7890_transpose_x_0"), val = bool(false)];
            bool var_7890_transpose_y_0 = const()[name = string("op_7890_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_7890_cast_fp16 = matmul(transpose_x = var_7890_transpose_x_0, transpose_y = var_7890_transpose_y_0, x = var_7889_cast_fp16, y = S_57_cast_fp16)[name = string("op_7890_cast_fp16")];
            bool var_7891_transpose_x_0 = const()[name = string("op_7891_transpose_x_0"), val = bool(false)];
            bool var_7891_transpose_y_0 = const()[name = string("op_7891_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_7891_cast_fp16 = matmul(transpose_x = var_7891_transpose_x_0, transpose_y = var_7891_transpose_y_0, x = A_69_cast_fp16, y = vnew_57_cast_fp16)[name = string("op_7891_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_7893_cast_fp16 = add(x = var_7890_cast_fp16, y = var_7891_cast_fp16)[name = string("op_7893_cast_fp16")];
            tensor<fp16, [16]> var_7894_bias_0_to_fp16 = const()[name = string("op_7894_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_7894_cast_fp16 = linear(bias = var_7894_bias_0_to_fp16, weight = transpose_1_to_fp16, x = gdec_81_cast_fp16)[name = string("op_7894_cast_fp16")];
            tensor<fp16, [3, 4, 16]> tail_cast_fp16 = exp(x = var_7894_cast_fp16)[name = string("tail_cast_fp16")];
            tensor<fp16, [1, 16]> transpose_117_to_fp16 = const()[name = string("transpose_117_to_fp16"), val = tensor<fp16, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119663168)))];
            tensor<fp16, [1]> var_7896_bias_0_to_fp16 = const()[name = string("op_7896_bias_0_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [3, 4, 1]> var_7896_cast_fp16 = linear(bias = var_7896_bias_0_to_fp16, weight = transpose_117_to_fp16, x = gdec_81_cast_fp16)[name = string("op_7896_cast_fp16")];
            tensor<fp16, [3, 4, 1]> var_7897_cast_fp16 = exp(x = var_7896_cast_fp16)[name = string("op_7897_cast_fp16")];
            tensor<int32, [1]> last_axes_0 = const()[name = string("last_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 1, 1]> last_cast_fp16 = expand_dims(axes = last_axes_0, x = var_7897_cast_fp16)[name = string("last_cast_fp16")];
            tensor<int32, [1]> var_7901_axes_0 = const()[name = string("op_7901_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_7901_cast_fp16 = expand_dims(axes = var_7901_axes_0, x = tail_cast_fp16)[name = string("op_7901_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_7902_cast_fp16 = mul(x = k_121_cast_fp16, y = var_7901_cast_fp16)[name = string("op_7902_cast_fp16")];
            bool upd_transpose_x_1 = const()[name = string("upd_transpose_x_1"), val = bool(true)];
            bool upd_transpose_y_1 = const()[name = string("upd_transpose_y_1"), val = bool(false)];
            tensor<fp16, [3, 4, 128, 128]> upd_cast_fp16 = matmul(transpose_x = upd_transpose_x_1, transpose_y = upd_transpose_y_1, x = var_7902_cast_fp16, y = vnew_57_cast_fp16)[name = string("upd_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> var_7907_cast_fp16 = mul(x = S_57_cast_fp16, y = last_cast_fp16)[name = string("op_7907_cast_fp16")];
            tensor<fp16, [3, 4, 128, 128]> S_cast_fp16 = add(x = var_7907_cast_fp16, y = upd_cast_fp16)[name = string("S_cast_fp16")];
            tensor<int32, [4]> q_121_begin_0 = const()[name = string("q_121_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> q_121_end_0 = const()[name = string("q_121_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> q_121_end_mask_0 = const()[name = string("q_121_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> q_121_cast_fp16 = slice_by_index(begin = q_121_begin_0, end = q_121_end_0, end_mask = q_121_end_mask_0, x = q_109_cast_fp16)[name = string("q_121_cast_fp16")];
            tensor<int32, [4]> k_123_begin_0 = const()[name = string("k_123_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> k_123_end_0 = const()[name = string("k_123_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> k_123_end_mask_0 = const()[name = string("k_123_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> k_123_cast_fp16 = slice_by_index(begin = k_123_begin_0, end = k_123_end_0, end_mask = k_123_end_mask_0, x = k_111_cast_fp16)[name = string("k_123_cast_fp16")];
            tensor<int32, [4]> kb_begin_0 = const()[name = string("kb_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> kb_end_0 = const()[name = string("kb_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> kb_end_mask_0 = const()[name = string("kb_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> kb_cast_fp16 = slice_by_index(begin = kb_begin_0, end = kb_end_0, end_mask = kb_end_mask_0, x = var_7194_cast_fp16)[name = string("kb_cast_fp16")];
            tensor<int32, [4]> vb_begin_0 = const()[name = string("vb_begin_0"), val = tensor<int32, [4]>([0, 0, 80, 0])];
            tensor<int32, [4]> vb_end_0 = const()[name = string("vb_end_0"), val = tensor<int32, [4]>([3, 4, 1, 128])];
            tensor<bool, [4]> vb_end_mask_0 = const()[name = string("vb_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 4, 16, 128]> vb_cast_fp16 = slice_by_index(begin = vb_begin_0, end = vb_end_0, end_mask = vb_end_mask_0, x = vb_71_cast_fp16)[name = string("vb_cast_fp16")];
            tensor<int32, [3]> gdec_begin_0 = const()[name = string("gdec_begin_0"), val = tensor<int32, [3]>([0, 0, 80])];
            tensor<int32, [3]> gdec_end_0 = const()[name = string("gdec_end_0"), val = tensor<int32, [3]>([3, 4, 1])];
            tensor<bool, [3]> gdec_end_mask_0 = const()[name = string("gdec_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 4, 16]> gdec_cast_fp16 = slice_by_index(begin = gdec_begin_0, end = gdec_end_0, end_mask = gdec_end_mask_0, x = gdec_71_cast_fp16)[name = string("gdec_cast_fp16")];
            tensor<int32, [2]> var_7990 = const()[name = string("op_7990"), val = tensor<int32, [2]>([12, 16])];
            tensor<fp16, [12, 16]> var_7991_cast_fp16 = reshape(shape = var_7990, x = gdec_cast_fp16)[name = string("op_7991_cast_fp16")];
            tensor<fp16, [12, 256]> var_7992_cast_fp16 = linear(bias = var_632_bias_0_to_fp16, weight = transpose_0_to_fp16, x = var_7991_cast_fp16)[name = string("op_7992_cast_fp16")];
            tensor<int32, [4]> var_7997 = const()[name = string("op_7997"), val = tensor<int32, [4]>([3, 4, 16, 16])];
            tensor<fp16, [3, 4, 16, 16]> P_cast_fp16 = reshape(shape = var_7997, x = var_7992_cast_fp16)[name = string("P_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_7999_cast_fp16 = exp(x = P_cast_fp16)[name = string("op_7999_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> D_cast_fp16 = mul(x = var_7999_cast_fp16, y = tril_incl_to_fp16)[name = string("D_cast_fp16")];
            bool kk_transpose_x_1 = const()[name = string("kk_transpose_x_1"), val = bool(false)];
            bool kk_transpose_y_1 = const()[name = string("kk_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> kk_cast_fp16 = matmul(transpose_x = kk_transpose_x_1, transpose_y = kk_transpose_y_1, x = kb_cast_fp16, y = k_123_cast_fp16)[name = string("kk_cast_fp16")];
            fp16 const_43_promoted_to_fp16 = const()[name = string("const_43_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 4, 16, 16]> var_8005_cast_fp16 = mul(x = kk_cast_fp16, y = const_43_promoted_to_fp16)[name = string("op_8005_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_8006_cast_fp16 = mul(x = var_8005_cast_fp16, y = D_cast_fp16)[name = string("op_8006_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> N_cast_fp16 = mul(x = var_8006_cast_fp16, y = cum_tail_to_fp16)[name = string("N_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> var_8008_cast_fp16 = mul(x = D_cast_fp16, y = neg_tril_strict_to_fp16)[name = string("op_8008_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Na_cast_fp16 = mul(x = kk_cast_fp16, y = var_8008_cast_fp16)[name = string("Na_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_281_cast_fp16 = add(x = eye_to_fp16, y = N_cast_fp16)[name = string("Minv_281_cast_fp16")];
            bool Np_211_transpose_x_0 = const()[name = string("Np_211_transpose_x_0"), val = bool(false)];
            bool Np_211_transpose_y_0 = const()[name = string("Np_211_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_211_cast_fp16 = matmul(transpose_x = Np_211_transpose_x_0, transpose_y = Np_211_transpose_y_0, x = N_cast_fp16, y = Na_cast_fp16)[name = string("Np_211_cast_fp16")];
            bool Npa_141_transpose_x_0 = const()[name = string("Npa_141_transpose_x_0"), val = bool(false)];
            bool Npa_141_transpose_y_0 = const()[name = string("Npa_141_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_141_cast_fp16 = matmul(transpose_x = Npa_141_transpose_x_0, transpose_y = Npa_141_transpose_y_0, x = Na_cast_fp16, y = N_cast_fp16)[name = string("Npa_141_cast_fp16")];
            bool var_8014_transpose_x_0 = const()[name = string("op_8014_transpose_x_0"), val = bool(false)];
            bool var_8014_transpose_y_0 = const()[name = string("op_8014_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_8014_cast_fp16 = matmul(transpose_x = var_8014_transpose_x_0, transpose_y = var_8014_transpose_y_0, x = Minv_281_cast_fp16, y = Np_211_cast_fp16)[name = string("op_8014_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_283_cast_fp16 = add(x = Minv_281_cast_fp16, y = var_8014_cast_fp16)[name = string("Minv_283_cast_fp16")];
            bool Np_213_transpose_x_0 = const()[name = string("Np_213_transpose_x_0"), val = bool(false)];
            bool Np_213_transpose_y_0 = const()[name = string("Np_213_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_213_cast_fp16 = matmul(transpose_x = Np_213_transpose_x_0, transpose_y = Np_213_transpose_y_0, x = Np_211_cast_fp16, y = Npa_141_cast_fp16)[name = string("Np_213_cast_fp16")];
            bool Npa_transpose_x_0 = const()[name = string("Npa_transpose_x_0"), val = bool(false)];
            bool Npa_transpose_y_0 = const()[name = string("Npa_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Npa_cast_fp16 = matmul(transpose_x = Npa_transpose_x_0, transpose_y = Npa_transpose_y_0, x = Npa_141_cast_fp16, y = Np_211_cast_fp16)[name = string("Npa_cast_fp16")];
            bool var_8019_transpose_x_0 = const()[name = string("op_8019_transpose_x_0"), val = bool(false)];
            bool var_8019_transpose_y_0 = const()[name = string("op_8019_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_8019_cast_fp16 = matmul(transpose_x = var_8019_transpose_x_0, transpose_y = var_8019_transpose_y_0, x = Minv_283_cast_fp16, y = Np_213_cast_fp16)[name = string("op_8019_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_285_cast_fp16 = add(x = Minv_283_cast_fp16, y = var_8019_cast_fp16)[name = string("Minv_285_cast_fp16")];
            bool Np_transpose_x_0 = const()[name = string("Np_transpose_x_0"), val = bool(false)];
            bool Np_transpose_y_0 = const()[name = string("Np_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> Np_cast_fp16 = matmul(transpose_x = Np_transpose_x_0, transpose_y = Np_transpose_y_0, x = Np_213_cast_fp16, y = Npa_cast_fp16)[name = string("Np_cast_fp16")];
            bool var_8023_transpose_x_0 = const()[name = string("op_8023_transpose_x_0"), val = bool(false)];
            bool var_8023_transpose_y_0 = const()[name = string("op_8023_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 16]> var_8023_cast_fp16 = matmul(transpose_x = var_8023_transpose_x_0, transpose_y = var_8023_transpose_y_0, x = Minv_285_cast_fp16, y = Np_cast_fp16)[name = string("op_8023_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> Minv_cast_fp16 = add(x = Minv_285_cast_fp16, y = var_8023_cast_fp16)[name = string("Minv_cast_fp16")];
            bool u_transpose_x_0 = const()[name = string("u_transpose_x_0"), val = bool(false)];
            bool u_transpose_y_0 = const()[name = string("u_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> u_cast_fp16 = matmul(transpose_x = u_transpose_x_0, transpose_y = u_transpose_y_0, x = Minv_cast_fp16, y = vb_cast_fp16)[name = string("u_cast_fp16")];
            bool var_8030_transpose_x_1 = const()[name = string("op_8030_transpose_x_1"), val = bool(false)];
            bool var_8030_transpose_y_1 = const()[name = string("op_8030_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 4, 16, 16]> var_8030_cast_fp16 = matmul(transpose_x = var_8030_transpose_x_1, transpose_y = var_8030_transpose_y_1, x = q_121_cast_fp16, y = k_123_cast_fp16)[name = string("op_8030_cast_fp16")];
            tensor<fp16, [3, 4, 16, 16]> A_cast_fp16 = mul(x = var_8030_cast_fp16, y = D_cast_fp16)[name = string("A_cast_fp16")];
            tensor<fp16, [16]> var_8032_bias_0_to_fp16 = const()[name = string("op_8032_bias_0_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119658752)))];
            tensor<fp16, [3, 4, 16]> var_8032_cast_fp16 = linear(bias = var_8032_bias_0_to_fp16, weight = tril_incl_to_fp16, x = gdec_cast_fp16)[name = string("op_8032_cast_fp16")];
            tensor<fp16, [3, 4, 16]> cum_cast_fp16 = exp(x = var_8032_cast_fp16)[name = string("cum_cast_fp16")];
            tensor<int32, [1]> var_8035_axes_0 = const()[name = string("op_8035_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [3, 4, 16, 1]> var_8035_cast_fp16 = expand_dims(axes = var_8035_axes_0, x = cum_cast_fp16)[name = string("op_8035_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_8036_cast_fp16 = mul(x = kb_cast_fp16, y = var_8035_cast_fp16)[name = string("op_8036_cast_fp16")];
            bool w_transpose_x_0 = const()[name = string("w_transpose_x_0"), val = bool(false)];
            bool w_transpose_y_0 = const()[name = string("w_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> w_cast_fp16 = matmul(transpose_x = w_transpose_x_0, transpose_y = w_transpose_y_0, x = Minv_cast_fp16, y = var_8036_cast_fp16)[name = string("w_cast_fp16")];
            bool var_8038_transpose_x_0 = const()[name = string("op_8038_transpose_x_0"), val = bool(false)];
            bool var_8038_transpose_y_0 = const()[name = string("op_8038_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_8038_cast_fp16 = matmul(transpose_x = var_8038_transpose_x_0, transpose_y = var_8038_transpose_y_0, x = w_cast_fp16, y = S_cast_fp16)[name = string("op_8038_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> vnew_cast_fp16 = sub(x = u_cast_fp16, y = var_8038_cast_fp16)[name = string("vnew_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> var_8043_cast_fp16 = mul(x = q_121_cast_fp16, y = var_8035_cast_fp16)[name = string("op_8043_cast_fp16")];
            bool var_8044_transpose_x_0 = const()[name = string("op_8044_transpose_x_0"), val = bool(false)];
            bool var_8044_transpose_y_0 = const()[name = string("op_8044_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_8044_cast_fp16 = matmul(transpose_x = var_8044_transpose_x_0, transpose_y = var_8044_transpose_y_0, x = var_8043_cast_fp16, y = S_cast_fp16)[name = string("op_8044_cast_fp16")];
            bool var_8045_transpose_x_0 = const()[name = string("op_8045_transpose_x_0"), val = bool(false)];
            bool var_8045_transpose_y_0 = const()[name = string("op_8045_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 4, 16, 128]> var_8045_cast_fp16 = matmul(transpose_x = var_8045_transpose_x_0, transpose_y = var_8045_transpose_y_0, x = A_cast_fp16, y = vnew_cast_fp16)[name = string("op_8045_cast_fp16")];
            tensor<fp16, [3, 4, 16, 128]> oc_cast_fp16 = add(x = var_8044_cast_fp16, y = var_8045_cast_fp16)[name = string("oc_cast_fp16")];
            int32 var_8049 = const()[name = string("op_8049"), val = int32(2)];
            bool o_33_interleave_0 = const()[name = string("o_33_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 4, 96, 128]> o_33_cast_fp16 = concat(axis = var_8049, interleave = o_33_interleave_0, values = (var_7284_cast_fp16, var_7431_cast_fp16, var_7585_cast_fp16, var_7739_cast_fp16, var_7893_cast_fp16, oc_cast_fp16))[name = string("o_33_cast_fp16")];
            tensor<int32, [4]> var_8055 = const()[name = string("op_8055"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_8061 = const()[name = string("op_8061"), val = tensor<int32, [4]>([3, 96, 4, 128])];
            tensor<fp16, [3, 96, 4, 128]> input_97_cast_fp16 = reshape(shape = var_8061, x = z_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> o_35_cast_fp16 = transpose(perm = var_8055, x = o_33_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [3, 96, 4, 128]> var_8063_cast_fp16 = mul(x = o_35_cast_fp16, y = o_35_cast_fp16)[name = string("op_8063_cast_fp16")];
            tensor<int32, [1]> var_8068_axes_0 = const()[name = string("op_8068_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8068_keep_dims_0 = const()[name = string("op_8068_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 4, 1]> var_8068_cast_fp16 = reduce_mean(axes = var_8068_axes_0, keep_dims = var_8068_keep_dims_0, x = var_8063_cast_fp16)[name = string("op_8068_cast_fp16")];
            fp16 var_8070_to_fp16 = const()[name = string("op_8070_to_fp16"), val = fp16(0x1.0c8p+4)];
            tensor<fp16, [3, 96, 4, 1]> var_8071_cast_fp16 = add(x = var_8068_cast_fp16, y = var_8070_to_fp16)[name = string("op_8071_cast_fp16")];
            fp32 var_8072_epsilon_0 = const()[name = string("op_8072_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 4, 1]> var_8072_cast_fp16 = rsqrt(epsilon = var_8072_epsilon_0, x = var_8071_cast_fp16)[name = string("op_8072_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> var_8073_cast_fp16 = mul(x = o_35_cast_fp16, y = var_8072_cast_fp16)[name = string("op_8073_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> var_8075_cast_fp16 = silu(x = input_97_cast_fp16)[name = string("op_8075_cast_fp16")];
            tensor<fp16, [3, 96, 4, 128]> o_37_cast_fp16 = mul(x = var_8073_cast_fp16, y = var_8075_cast_fp16)[name = string("o_37_cast_fp16")];
            tensor<int32, [3]> var_8080 = const()[name = string("op_8080"), val = tensor<int32, [3]>([3, 96, 512])];
            tensor<fp16, [3, 96, 512]> input_99_cast_fp16 = reshape(shape = var_8080, x = o_37_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [640, 512]> layers_6_out_weight_to_fp16 = const()[name = string("layers_6_out_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97009984)))];
            tensor<fp16, [3, 96, 640]> linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_out_weight_to_fp16, x = input_99_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [3, 96, 640]> x_65_cast_fp16 = add(x = x_61_cast_fp16, y = linear_25_cast_fp16)[name = string("x_65_cast_fp16")];
            fp16 var_8093_to_fp16 = const()[name = string("op_8093_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [3, 96, 640]> x_67_cast_fp16 = mul(x = x_65_cast_fp16, y = var_8093_to_fp16)[name = string("x_67_cast_fp16")];
            fp16 var_8091_promoted_to_fp16 = const()[name = string("op_8091_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> var_8095_cast_fp16 = pow(x = x_67_cast_fp16, y = var_8091_promoted_to_fp16)[name = string("op_8095_cast_fp16")];
            tensor<int32, [1]> var_8097_axes_0 = const()[name = string("op_8097_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8097_keep_dims_0 = const()[name = string("op_8097_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 1]> var_8097_cast_fp16 = reduce_mean(axes = var_8097_axes_0, keep_dims = var_8097_keep_dims_0, x = var_8095_cast_fp16)[name = string("op_8097_cast_fp16")];
            fp16 var_8098_to_fp16 = const()[name = string("op_8098_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 96, 1]> var_8099_cast_fp16 = add(x = var_8097_cast_fp16, y = var_8098_to_fp16)[name = string("op_8099_cast_fp16")];
            fp32 var_8100_epsilon_0 = const()[name = string("op_8100_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 1]> var_8100_cast_fp16 = rsqrt(epsilon = var_8100_epsilon_0, x = var_8099_cast_fp16)[name = string("op_8100_cast_fp16")];
            tensor<fp16, [3, 96, 640]> var_8101_cast_fp16 = mul(x = x_67_cast_fp16, y = var_8100_cast_fp16)[name = string("op_8101_cast_fp16")];
            tensor<fp16, [640]> layers_6_n2_w1_to_fp16 = const()[name = string("layers_6_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97665408)))];
            tensor<fp16, [3, 96, 640]> input_101_cast_fp16 = mul(x = var_8101_cast_fp16, y = layers_6_n2_w1_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_6_gate_up_weight_to_fp16 = const()[name = string("layers_6_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97666752)))];
            tensor<fp16, [3, 96, 5120]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_6_gate_up_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<int32, [3]> input_103_begin_0 = const()[name = string("input_103_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_103_end_0 = const()[name = string("input_103_end_0"), val = tensor<int32, [3]>([3, 96, 2560])];
            tensor<bool, [3]> input_103_end_mask_0 = const()[name = string("input_103_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 2560]> input_103_cast_fp16 = slice_by_index(begin = input_103_begin_0, end = input_103_end_0, end_mask = input_103_end_mask_0, x = linear_26_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [3, 96, 2560]> var_8111_cast_fp16 = silu(x = input_103_cast_fp16)[name = string("op_8111_cast_fp16")];
            tensor<int32, [3]> var_8116_begin_0 = const()[name = string("op_8116_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_8116_end_0 = const()[name = string("op_8116_end_0"), val = tensor<int32, [3]>([3, 96, 5120])];
            tensor<bool, [3]> var_8116_end_mask_0 = const()[name = string("op_8116_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 2560]> var_8116_cast_fp16 = slice_by_index(begin = var_8116_begin_0, end = var_8116_end_0, end_mask = var_8116_end_mask_0, x = linear_26_cast_fp16)[name = string("op_8116_cast_fp16")];
            tensor<fp16, [3, 96, 2560]> input_105_cast_fp16 = mul(x = var_8111_cast_fp16, y = var_8116_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_6_down_weight_to_fp16 = const()[name = string("layers_6_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104220416)))];
            tensor<fp16, [3, 96, 640]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_down_weight_to_fp16, x = input_105_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [3, 96, 640]> x_69_cast_fp16 = add(x = x_65_cast_fp16, y = linear_27_cast_fp16)[name = string("x_69_cast_fp16")];
            fp16 var_8129_to_fp16 = const()[name = string("op_8129_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [3, 96, 640]> x_71_cast_fp16 = mul(x = x_69_cast_fp16, y = var_8129_to_fp16)[name = string("x_71_cast_fp16")];
            fp16 var_8127_promoted_to_fp16 = const()[name = string("op_8127_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> var_8131_cast_fp16 = pow(x = x_71_cast_fp16, y = var_8127_promoted_to_fp16)[name = string("op_8131_cast_fp16")];
            tensor<int32, [1]> var_8133_axes_0 = const()[name = string("op_8133_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8133_keep_dims_0 = const()[name = string("op_8133_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 1]> var_8133_cast_fp16 = reduce_mean(axes = var_8133_axes_0, keep_dims = var_8133_keep_dims_0, x = var_8131_cast_fp16)[name = string("op_8133_cast_fp16")];
            fp16 var_8134_to_fp16 = const()[name = string("op_8134_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 96, 1]> var_8135_cast_fp16 = add(x = var_8133_cast_fp16, y = var_8134_to_fp16)[name = string("op_8135_cast_fp16")];
            fp32 var_8136_epsilon_0 = const()[name = string("op_8136_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 1]> var_8136_cast_fp16 = rsqrt(epsilon = var_8136_epsilon_0, x = var_8135_cast_fp16)[name = string("op_8136_cast_fp16")];
            tensor<fp16, [3, 96, 640]> var_8137_cast_fp16 = mul(x = x_71_cast_fp16, y = var_8136_cast_fp16)[name = string("op_8137_cast_fp16")];
            tensor<fp16, [640]> layers_7_n1_w1_to_fp16 = const()[name = string("layers_7_n1_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107497280)))];
            tensor<fp16, [3, 96, 640]> input_107_cast_fp16 = mul(x = var_8137_cast_fp16, y = layers_7_n1_w1_to_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [1280, 640]> layers_7_qkv_weight_to_fp16 = const()[name = string("layers_7_qkv_weight_to_fp16"), val = tensor<fp16, [1280, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107498624)))];
            tensor<fp16, [3, 96, 1280]> linear_28_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = layers_7_qkv_weight_to_fp16, x = input_107_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [3]> var_8146_begin_0 = const()[name = string("op_8146_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_8146_end_0 = const()[name = string("op_8146_end_0"), val = tensor<int32, [3]>([3, 96, 1024])];
            tensor<bool, [3]> var_8146_end_mask_0 = const()[name = string("op_8146_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 1024]> var_8146_cast_fp16 = slice_by_index(begin = var_8146_begin_0, end = var_8146_end_0, end_mask = var_8146_end_mask_0, x = linear_28_cast_fp16)[name = string("op_8146_cast_fp16")];
            tensor<int32, [4]> var_8151 = const()[name = string("op_8151"), val = tensor<int32, [4]>([3, 96, 8, 128])];
            tensor<fp16, [3, 96, 8, 128]> qg_cast_fp16 = reshape(shape = var_8151, x = var_8146_cast_fp16)[name = string("qg_cast_fp16")];
            tensor<int32, [4]> x_73_begin_0 = const()[name = string("x_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x_73_end_0 = const()[name = string("x_73_end_0"), val = tensor<int32, [4]>([3, 96, 8, 64])];
            tensor<bool, [4]> x_73_end_mask_0 = const()[name = string("x_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 96, 8, 64]> x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, x = qg_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<int32, [4]> var_8162_begin_0 = const()[name = string("op_8162_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> var_8162_end_0 = const()[name = string("op_8162_end_0"), val = tensor<int32, [4]>([3, 96, 8, 128])];
            tensor<bool, [4]> var_8162_end_mask_0 = const()[name = string("op_8162_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 96, 8, 64]> var_8162_cast_fp16 = slice_by_index(begin = var_8162_begin_0, end = var_8162_end_0, end_mask = var_8162_end_mask_0, x = qg_cast_fp16)[name = string("op_8162_cast_fp16")];
            tensor<int32, [3]> var_8166 = const()[name = string("op_8166"), val = tensor<int32, [3]>([3, 96, 512])];
            tensor<fp16, [3, 96, 512]> gate_cast_fp16 = reshape(shape = var_8166, x = var_8162_cast_fp16)[name = string("gate_cast_fp16")];
            tensor<int32, [3]> var_8172_begin_0 = const()[name = string("op_8172_begin_0"), val = tensor<int32, [3]>([0, 0, 1024])];
            tensor<int32, [3]> var_8172_end_0 = const()[name = string("op_8172_end_0"), val = tensor<int32, [3]>([3, 96, 1152])];
            tensor<bool, [3]> var_8172_end_mask_0 = const()[name = string("op_8172_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 128]> var_8172_cast_fp16 = slice_by_index(begin = var_8172_begin_0, end = var_8172_end_0, end_mask = var_8172_end_mask_0, x = linear_28_cast_fp16)[name = string("op_8172_cast_fp16")];
            tensor<int32, [4]> var_8177 = const()[name = string("op_8177"), val = tensor<int32, [4]>([3, 96, 2, 64])];
            tensor<fp16, [3, 96, 2, 64]> x_79_cast_fp16 = reshape(shape = var_8177, x = var_8172_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<int32, [3]> var_8183_begin_0 = const()[name = string("op_8183_begin_0"), val = tensor<int32, [3]>([0, 0, 1152])];
            tensor<int32, [3]> var_8183_end_0 = const()[name = string("op_8183_end_0"), val = tensor<int32, [3]>([3, 96, 1280])];
            tensor<bool, [3]> var_8183_end_mask_0 = const()[name = string("op_8183_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 128]> var_8183_cast_fp16 = slice_by_index(begin = var_8183_begin_0, end = var_8183_end_0, end_mask = var_8183_end_mask_0, x = linear_28_cast_fp16)[name = string("op_8183_cast_fp16")];
            tensor<int32, [4]> var_8188 = const()[name = string("op_8188"), val = tensor<int32, [4]>([3, 96, 2, 64])];
            tensor<fp16, [3, 96, 2, 64]> var_8189_cast_fp16 = reshape(shape = var_8188, x = var_8183_cast_fp16)[name = string("op_8189_cast_fp16")];
            tensor<int32, [4]> var_8194 = const()[name = string("op_8194"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8200_promoted_to_fp16 = const()[name = string("op_8200_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 8, 64]> var_8204_cast_fp16 = pow(x = x_73_cast_fp16, y = var_8200_promoted_to_fp16)[name = string("op_8204_cast_fp16")];
            tensor<int32, [1]> var_8206_axes_0 = const()[name = string("op_8206_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8206_keep_dims_0 = const()[name = string("op_8206_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 8, 1]> var_8206_cast_fp16 = reduce_mean(axes = var_8206_axes_0, keep_dims = var_8206_keep_dims_0, x = var_8204_cast_fp16)[name = string("op_8206_cast_fp16")];
            fp16 var_8207_to_fp16 = const()[name = string("op_8207_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 96, 8, 1]> var_8208_cast_fp16 = add(x = var_8206_cast_fp16, y = var_8207_to_fp16)[name = string("op_8208_cast_fp16")];
            fp32 var_8209_epsilon_0 = const()[name = string("op_8209_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 8, 1]> var_8209_cast_fp16 = rsqrt(epsilon = var_8209_epsilon_0, x = var_8208_cast_fp16)[name = string("op_8209_cast_fp16")];
            tensor<fp16, [3, 96, 8, 64]> var_8210_cast_fp16 = mul(x = x_73_cast_fp16, y = var_8209_cast_fp16)[name = string("op_8210_cast_fp16")];
            tensor<fp16, [64]> layers_7_qn_w1_to_fp16 = const()[name = string("layers_7_qn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137088)))];
            tensor<fp16, [3, 96, 8, 64]> var_8211_cast_fp16 = mul(x = var_8210_cast_fp16, y = layers_7_qn_w1_to_fp16)[name = string("op_8211_cast_fp16")];
            tensor<int32, [4]> var_8216 = const()[name = string("op_8216"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_5_begin_0 = const()[name = string("r_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_5_end_0 = const()[name = string("r_5_end_0"), val = tensor<int32, [4]>([3, 8, 96, 16])];
            tensor<bool, [4]> r_5_end_mask_0 = const()[name = string("r_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 8, 96, 64]> x_77_cast_fp16 = transpose(perm = var_8216, x = var_8211_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [3, 8, 96, 16]> r_5_cast_fp16 = slice_by_index(begin = r_5_begin_0, end = r_5_end_0, end_mask = r_5_end_mask_0, x = x_77_cast_fp16)[name = string("r_5_cast_fp16")];
            tensor<int32, [4]> p_5_begin_0 = const()[name = string("p_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_5_end_0 = const()[name = string("p_5_end_0"), val = tensor<int32, [4]>([3, 8, 96, 64])];
            tensor<bool, [4]> p_5_end_mask_0 = const()[name = string("p_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 8, 96, 48]> p_5_cast_fp16 = slice_by_index(begin = p_5_begin_0, end = p_5_end_0, end_mask = p_5_end_mask_0, x = x_77_cast_fp16)[name = string("p_5_cast_fp16")];
            tensor<int32, [4]> var_8232_begin_0 = const()[name = string("op_8232_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_8232_end_0 = const()[name = string("op_8232_end_0"), val = tensor<int32, [4]>([3, 8, 96, 16])];
            tensor<bool, [4]> var_8232_end_mask_0 = const()[name = string("op_8232_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 8, 96, 8]> var_8232_cast_fp16 = slice_by_index(begin = var_8232_begin_0, end = var_8232_end_0, end_mask = var_8232_end_mask_0, x = r_5_cast_fp16)[name = string("op_8232_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 8, 96, 8]> var_8233_cast_fp16 = mul(x = var_8232_cast_fp16, y = const_44_promoted_to_fp16)[name = string("op_8233_cast_fp16")];
            tensor<int32, [4]> var_8238_begin_0 = const()[name = string("op_8238_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8238_end_0 = const()[name = string("op_8238_end_0"), val = tensor<int32, [4]>([3, 8, 96, 8])];
            tensor<bool, [4]> var_8238_end_mask_0 = const()[name = string("op_8238_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 8, 96, 8]> var_8238_cast_fp16 = slice_by_index(begin = var_8238_begin_0, end = var_8238_end_0, end_mask = var_8238_end_mask_0, x = r_5_cast_fp16)[name = string("op_8238_cast_fp16")];
            int32 var_8240 = const()[name = string("op_8240"), val = int32(-1)];
            bool rh_5_interleave_0 = const()[name = string("rh_5_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 8, 96, 16]> rh_5_cast_fp16 = concat(axis = var_8240, interleave = rh_5_interleave_0, values = (var_8233_cast_fp16, var_8238_cast_fp16))[name = string("rh_5_cast_fp16")];
            tensor<fp16, [3, 8, 96, 16]> var_8242_cast_fp16 = mul(x = r_5_cast_fp16, y = cos_to_fp16)[name = string("op_8242_cast_fp16")];
            tensor<fp16, [3, 8, 96, 16]> var_8243_cast_fp16 = mul(x = rh_5_cast_fp16, y = sin_to_fp16)[name = string("op_8243_cast_fp16")];
            tensor<fp16, [3, 8, 96, 16]> var_8245_cast_fp16 = add(x = var_8242_cast_fp16, y = var_8243_cast_fp16)[name = string("op_8245_cast_fp16")];
            int32 var_8247 = const()[name = string("op_8247"), val = int32(-1)];
            bool q_interleave_0 = const()[name = string("q_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 8, 96, 64]> q_cast_fp16 = concat(axis = var_8247, interleave = q_interleave_0, values = (var_8245_cast_fp16, p_5_cast_fp16))[name = string("q_cast_fp16")];
            fp16 var_8253_promoted_to_fp16 = const()[name = string("op_8253_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 2, 64]> var_8257_cast_fp16 = pow(x = x_79_cast_fp16, y = var_8253_promoted_to_fp16)[name = string("op_8257_cast_fp16")];
            tensor<int32, [1]> var_8259_axes_0 = const()[name = string("op_8259_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8259_keep_dims_0 = const()[name = string("op_8259_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 2, 1]> var_8259_cast_fp16 = reduce_mean(axes = var_8259_axes_0, keep_dims = var_8259_keep_dims_0, x = var_8257_cast_fp16)[name = string("op_8259_cast_fp16")];
            fp16 var_8260_to_fp16 = const()[name = string("op_8260_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [3, 96, 2, 1]> var_8261_cast_fp16 = add(x = var_8259_cast_fp16, y = var_8260_to_fp16)[name = string("op_8261_cast_fp16")];
            fp32 var_8262_epsilon_0 = const()[name = string("op_8262_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 2, 1]> var_8262_cast_fp16 = rsqrt(epsilon = var_8262_epsilon_0, x = var_8261_cast_fp16)[name = string("op_8262_cast_fp16")];
            tensor<fp16, [3, 96, 2, 64]> var_8263_cast_fp16 = mul(x = x_79_cast_fp16, y = var_8262_cast_fp16)[name = string("op_8263_cast_fp16")];
            tensor<fp16, [64]> layers_7_kn_w1_to_fp16 = const()[name = string("layers_7_kn_w1_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137280)))];
            tensor<fp16, [3, 96, 2, 64]> var_8264_cast_fp16 = mul(x = var_8263_cast_fp16, y = layers_7_kn_w1_to_fp16)[name = string("op_8264_cast_fp16")];
            tensor<int32, [4]> var_8269 = const()[name = string("op_8269"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> r_begin_0 = const()[name = string("r_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> r_end_0 = const()[name = string("r_end_0"), val = tensor<int32, [4]>([3, 2, 96, 16])];
            tensor<bool, [4]> r_end_mask_0 = const()[name = string("r_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 2, 96, 64]> x_83_cast_fp16 = transpose(perm = var_8269, x = var_8264_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [3, 2, 96, 16]> r_cast_fp16 = slice_by_index(begin = r_begin_0, end = r_end_0, end_mask = r_end_mask_0, x = x_83_cast_fp16)[name = string("r_cast_fp16")];
            tensor<int32, [4]> p_begin_0 = const()[name = string("p_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> p_end_0 = const()[name = string("p_end_0"), val = tensor<int32, [4]>([3, 2, 96, 64])];
            tensor<bool, [4]> p_end_mask_0 = const()[name = string("p_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 2, 96, 48]> p_cast_fp16 = slice_by_index(begin = p_begin_0, end = p_end_0, end_mask = p_end_mask_0, x = x_83_cast_fp16)[name = string("p_cast_fp16")];
            tensor<int32, [4]> var_8285_begin_0 = const()[name = string("op_8285_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8])];
            tensor<int32, [4]> var_8285_end_0 = const()[name = string("op_8285_end_0"), val = tensor<int32, [4]>([3, 2, 96, 16])];
            tensor<bool, [4]> var_8285_end_mask_0 = const()[name = string("op_8285_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [3, 2, 96, 8]> var_8285_cast_fp16 = slice_by_index(begin = var_8285_begin_0, end = var_8285_end_0, end_mask = var_8285_end_mask_0, x = r_cast_fp16)[name = string("op_8285_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [3, 2, 96, 8]> var_8286_cast_fp16 = mul(x = var_8285_cast_fp16, y = const_45_promoted_to_fp16)[name = string("op_8286_cast_fp16")];
            tensor<int32, [4]> var_8291_begin_0 = const()[name = string("op_8291_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8291_end_0 = const()[name = string("op_8291_end_0"), val = tensor<int32, [4]>([3, 2, 96, 8])];
            tensor<bool, [4]> var_8291_end_mask_0 = const()[name = string("op_8291_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [3, 2, 96, 8]> var_8291_cast_fp16 = slice_by_index(begin = var_8291_begin_0, end = var_8291_end_0, end_mask = var_8291_end_mask_0, x = r_cast_fp16)[name = string("op_8291_cast_fp16")];
            int32 var_8293 = const()[name = string("op_8293"), val = int32(-1)];
            bool rh_interleave_0 = const()[name = string("rh_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 2, 96, 16]> rh_cast_fp16 = concat(axis = var_8293, interleave = rh_interleave_0, values = (var_8286_cast_fp16, var_8291_cast_fp16))[name = string("rh_cast_fp16")];
            tensor<fp16, [3, 2, 96, 16]> var_8295_cast_fp16 = mul(x = r_cast_fp16, y = cos_to_fp16)[name = string("op_8295_cast_fp16")];
            tensor<fp16, [3, 2, 96, 16]> var_8296_cast_fp16 = mul(x = rh_cast_fp16, y = sin_to_fp16)[name = string("op_8296_cast_fp16")];
            tensor<fp16, [3, 2, 96, 16]> var_8298_cast_fp16 = add(x = var_8295_cast_fp16, y = var_8296_cast_fp16)[name = string("op_8298_cast_fp16")];
            int32 var_8300 = const()[name = string("op_8300"), val = int32(-1)];
            bool k_125_interleave_0 = const()[name = string("k_125_interleave_0"), val = bool(false)];
            tensor<fp16, [3, 2, 96, 64]> k_125_cast_fp16 = concat(axis = var_8300, interleave = k_125_interleave_0, values = (var_8298_cast_fp16, p_cast_fp16))[name = string("k_125_cast_fp16")];
            tensor<int32, [1]> var_8313_axes_0 = const()[name = string("op_8313_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 1, 96, 64]> var_8313_cast_fp16 = expand_dims(axes = var_8313_axes_0, x = k_125_cast_fp16)[name = string("op_8313_cast_fp16")];
            tensor<int32, [5]> var_8321_reps_0 = const()[name = string("op_8321_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [3, 2, 4, 96, 64]> var_8321_cast_fp16 = tile(reps = var_8321_reps_0, x = var_8313_cast_fp16)[name = string("op_8321_cast_fp16")];
            tensor<int32, [4]> var_8326 = const()[name = string("op_8326"), val = tensor<int32, [4]>([3, 8, 96, 64])];
            tensor<fp16, [3, 8, 96, 64]> k_cast_fp16 = reshape(shape = var_8326, x = var_8321_cast_fp16)[name = string("k_cast_fp16")];
            tensor<int32, [1]> var_8339_axes_0 = const()[name = string("op_8339_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [3, 2, 96, 64]> v_17_cast_fp16 = transpose(perm = var_8194, x = var_8189_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [3, 2, 1, 96, 64]> var_8339_cast_fp16 = expand_dims(axes = var_8339_axes_0, x = v_17_cast_fp16)[name = string("op_8339_cast_fp16")];
            tensor<int32, [5]> var_8347_reps_0 = const()[name = string("op_8347_reps_0"), val = tensor<int32, [5]>([1, 1, 4, 1, 1])];
            tensor<fp16, [3, 2, 4, 96, 64]> var_8347_cast_fp16 = tile(reps = var_8347_reps_0, x = var_8339_cast_fp16)[name = string("op_8347_cast_fp16")];
            tensor<int32, [4]> var_8352 = const()[name = string("op_8352"), val = tensor<int32, [4]>([3, 8, 96, 64])];
            tensor<fp16, [3, 8, 96, 64]> v_cast_fp16 = reshape(shape = var_8352, x = var_8347_cast_fp16)[name = string("v_cast_fp16")];
            bool var_8357_transpose_x_1 = const()[name = string("op_8357_transpose_x_1"), val = bool(false)];
            bool var_8357_transpose_y_1 = const()[name = string("op_8357_transpose_y_1"), val = bool(true)];
            tensor<fp16, [3, 8, 96, 96]> var_8357_cast_fp16 = matmul(transpose_x = var_8357_transpose_x_1, transpose_y = var_8357_transpose_y_1, x = q_cast_fp16, y = k_cast_fp16)[name = string("op_8357_cast_fp16")];
            fp16 var_8358_to_fp16 = const()[name = string("op_8358_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [3, 8, 96, 96]> var_8359_cast_fp16 = mul(x = var_8357_cast_fp16, y = var_8358_to_fp16)[name = string("op_8359_cast_fp16")];
            tensor<fp16, [3, 8, 96, 96]> s_cast_fp16 = add(x = var_8359_cast_fp16, y = causal_add_to_fp16)[name = string("s_cast_fp16")];
            int32 var_8362 = const()[name = string("op_8362"), val = int32(-1)];
            tensor<fp16, [3, 8, 96, 96]> var_8364_cast_fp16 = softmax(axis = var_8362, x = s_cast_fp16)[name = string("op_8364_cast_fp16")];
            bool o_39_transpose_x_0 = const()[name = string("o_39_transpose_x_0"), val = bool(false)];
            bool o_39_transpose_y_0 = const()[name = string("o_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [3, 8, 96, 64]> o_39_cast_fp16 = matmul(transpose_x = o_39_transpose_x_0, transpose_y = o_39_transpose_y_0, x = var_8364_cast_fp16, y = v_cast_fp16)[name = string("o_39_cast_fp16")];
            tensor<int32, [4]> var_8370 = const()[name = string("op_8370"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8375 = const()[name = string("op_8375"), val = tensor<int32, [3]>([3, 96, 512])];
            tensor<fp16, [3, 96, 8, 64]> var_8371_cast_fp16 = transpose(perm = var_8370, x = o_39_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [3, 96, 512]> var_8376_cast_fp16 = reshape(shape = var_8375, x = var_8371_cast_fp16)[name = string("op_8376_cast_fp16")];
            tensor<fp16, [3, 96, 512]> var_8377_cast_fp16 = sigmoid(x = gate_cast_fp16)[name = string("op_8377_cast_fp16")];
            tensor<fp16, [3, 96, 512]> input_109_cast_fp16 = mul(x = var_8376_cast_fp16, y = var_8377_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [640, 512]> layers_7_o_weight_to_fp16 = const()[name = string("layers_7_o_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109137472)))];
            tensor<fp16, [3, 96, 640]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_o_weight_to_fp16, x = input_109_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<fp16, [3, 96, 640]> x_85_cast_fp16 = add(x = x_69_cast_fp16, y = linear_29_cast_fp16)[name = string("x_85_cast_fp16")];
            fp16 var_8390_to_fp16 = const()[name = string("op_8390_to_fp16"), val = fp16(0x1p-2)];
            tensor<fp16, [3, 96, 640]> x_87_cast_fp16 = mul(x = x_85_cast_fp16, y = var_8390_to_fp16)[name = string("x_87_cast_fp16")];
            fp16 var_8388_promoted_to_fp16 = const()[name = string("op_8388_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> var_8392_cast_fp16 = pow(x = x_87_cast_fp16, y = var_8388_promoted_to_fp16)[name = string("op_8392_cast_fp16")];
            tensor<int32, [1]> var_8394_axes_0 = const()[name = string("op_8394_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8394_keep_dims_0 = const()[name = string("op_8394_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 1]> var_8394_cast_fp16 = reduce_mean(axes = var_8394_axes_0, keep_dims = var_8394_keep_dims_0, x = var_8392_cast_fp16)[name = string("op_8394_cast_fp16")];
            fp16 var_8395_to_fp16 = const()[name = string("op_8395_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 96, 1]> var_8396_cast_fp16 = add(x = var_8394_cast_fp16, y = var_8395_to_fp16)[name = string("op_8396_cast_fp16")];
            fp32 var_8397_epsilon_0 = const()[name = string("op_8397_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 1]> var_8397_cast_fp16 = rsqrt(epsilon = var_8397_epsilon_0, x = var_8396_cast_fp16)[name = string("op_8397_cast_fp16")];
            tensor<fp16, [3, 96, 640]> var_8398_cast_fp16 = mul(x = x_87_cast_fp16, y = var_8397_cast_fp16)[name = string("op_8398_cast_fp16")];
            tensor<fp16, [640]> layers_7_n2_w1_to_fp16 = const()[name = string("layers_7_n2_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109792896)))];
            tensor<fp16, [3, 96, 640]> input_111_cast_fp16 = mul(x = var_8398_cast_fp16, y = layers_7_n2_w1_to_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [5120, 640]> layers_7_gate_up_weight_to_fp16 = const()[name = string("layers_7_gate_up_weight_to_fp16"), val = tensor<fp16, [5120, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109794240)))];
            tensor<fp16, [3, 96, 5120]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = layers_7_gate_up_weight_to_fp16, x = input_111_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [3]> input_113_begin_0 = const()[name = string("input_113_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_113_end_0 = const()[name = string("input_113_end_0"), val = tensor<int32, [3]>([3, 96, 2560])];
            tensor<bool, [3]> input_113_end_mask_0 = const()[name = string("input_113_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [3, 96, 2560]> input_113_cast_fp16 = slice_by_index(begin = input_113_begin_0, end = input_113_end_0, end_mask = input_113_end_mask_0, x = linear_30_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [3, 96, 2560]> var_8408_cast_fp16 = silu(x = input_113_cast_fp16)[name = string("op_8408_cast_fp16")];
            tensor<int32, [3]> var_8413_begin_0 = const()[name = string("op_8413_begin_0"), val = tensor<int32, [3]>([0, 0, 2560])];
            tensor<int32, [3]> var_8413_end_0 = const()[name = string("op_8413_end_0"), val = tensor<int32, [3]>([3, 96, 5120])];
            tensor<bool, [3]> var_8413_end_mask_0 = const()[name = string("op_8413_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [3, 96, 2560]> var_8413_cast_fp16 = slice_by_index(begin = var_8413_begin_0, end = var_8413_end_0, end_mask = var_8413_end_mask_0, x = linear_30_cast_fp16)[name = string("op_8413_cast_fp16")];
            tensor<fp16, [3, 96, 2560]> input_115_cast_fp16 = mul(x = var_8408_cast_fp16, y = var_8413_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [640, 2560]> layers_7_down_weight_to_fp16 = const()[name = string("layers_7_down_weight_to_fp16"), val = tensor<fp16, [640, 2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116347904)))];
            tensor<fp16, [3, 96, 640]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_down_weight_to_fp16, x = input_115_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [3, 96, 640]> x_89_cast_fp16 = add(x = x_85_cast_fp16, y = linear_31_cast_fp16)[name = string("x_89_cast_fp16")];
            fp16 var_8426_to_fp16 = const()[name = string("op_8426_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [3, 96, 640]> x_cast_fp16 = mul(x = x_89_cast_fp16, y = var_8426_to_fp16)[name = string("x_cast_fp16")];
            fp16 var_8424_promoted_to_fp16 = const()[name = string("op_8424_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [3, 96, 640]> var_8428_cast_fp16 = pow(x = x_cast_fp16, y = var_8424_promoted_to_fp16)[name = string("op_8428_cast_fp16")];
            tensor<int32, [1]> var_8430_axes_0 = const()[name = string("op_8430_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8430_keep_dims_0 = const()[name = string("op_8430_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 96, 1]> var_8430_cast_fp16 = reduce_mean(axes = var_8430_axes_0, keep_dims = var_8430_keep_dims_0, x = var_8428_cast_fp16)[name = string("op_8430_cast_fp16")];
            fp16 var_8431_to_fp16 = const()[name = string("op_8431_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [3, 96, 1]> var_8432_cast_fp16 = add(x = var_8430_cast_fp16, y = var_8431_to_fp16)[name = string("op_8432_cast_fp16")];
            fp32 var_8433_epsilon_0 = const()[name = string("op_8433_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [3, 96, 1]> var_8433_cast_fp16 = rsqrt(epsilon = var_8433_epsilon_0, x = var_8432_cast_fp16)[name = string("op_8433_cast_fp16")];
            tensor<fp16, [3, 96, 640]> var_8434_cast_fp16 = mul(x = x_cast_fp16, y = var_8433_cast_fp16)[name = string("op_8434_cast_fp16")];
            tensor<fp16, [640]> norm_w1_to_fp16 = const()[name = string("norm_w1_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119624768)))];
            tensor<fp16, [3, 96, 640]> var_8435_cast_fp16 = mul(x = var_8434_cast_fp16, y = norm_w1_to_fp16)[name = string("op_8435_cast_fp16")];
            tensor<int32, [3]> h_begin_0 = const()[name = string("h_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> h_end_0 = const()[name = string("h_end_0"), val = tensor<int32, [3]>([3, 95, 640])];
            tensor<bool, [3]> h_end_mask_0 = const()[name = string("h_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [3, 95, 640]> h_cast_fp16 = slice_by_index(begin = h_begin_0, end = h_end_0, end_mask = h_end_mask_0, x = var_8435_cast_fp16)[name = string("h_cast_fp16")];
            tensor<fp16, [16000]> logits_bias_0_to_fp16 = const()[name = string("logits_bias_0_to_fp16"), val = tensor<fp16, [16000]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119626112)))];
            tensor<fp16, [3, 95, 16000]> logits_cast_fp16 = linear(bias = logits_bias_0_to_fp16, weight = emb_to_fp16, x = h_cast_fp16)[name = string("logits_cast_fp16")];
            tensor<int32, [1]> reduce_max_0_axes_0 = const()[name = string("reduce_max_0_axes_0"), val = tensor<int32, [1]>([-1])];
            bool reduce_max_0_keep_dims_0 = const()[name = string("reduce_max_0_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 95, 1]> reduce_max_0_cast_fp16 = reduce_max(axes = reduce_max_0_axes_0, keep_dims = reduce_max_0_keep_dims_0, x = logits_cast_fp16)[name = string("reduce_max_0_cast_fp16")];
            tensor<fp16, [3, 95, 16000]> var_8453_cast_fp16 = sub(x = logits_cast_fp16, y = reduce_max_0_cast_fp16)[name = string("op_8453_cast_fp16")];
            tensor<fp16, [3, 95, 16000]> var_8454_cast_fp16 = exp(x = var_8453_cast_fp16)[name = string("op_8454_cast_fp16")];
            tensor<int32, [1]> var_8459_axes_0 = const()[name = string("op_8459_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8459_keep_dims_0 = const()[name = string("op_8459_keep_dims_0"), val = bool(true)];
            tensor<fp16, [3, 95, 1]> var_8459_cast_fp16 = reduce_sum(axes = var_8459_axes_0, keep_dims = var_8459_keep_dims_0, x = var_8454_cast_fp16)[name = string("op_8459_cast_fp16")];
            fp32 var_8460_epsilon_0 = const()[name = string("op_8460_epsilon_0"), val = fp32(0x1p-149)];
            tensor<fp16, [3, 95, 1]> var_8460_cast_fp16 = log(epsilon = var_8460_epsilon_0, x = var_8459_cast_fp16)[name = string("op_8460_cast_fp16")];
            tensor<fp16, [3, 95, 1]> var_8462_cast_fp16 = add(x = reduce_max_0_cast_fp16, y = var_8460_cast_fp16)[name = string("op_8462_cast_fp16")];
            tensor<int32, [3]> lse_begin_0 = const()[name = string("lse_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> lse_end_0 = const()[name = string("lse_end_0"), val = tensor<int32, [3]>([3, 95, 1])];
            tensor<bool, [3]> lse_end_mask_0 = const()[name = string("lse_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> lse_squeeze_mask_0 = const()[name = string("lse_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<fp16, [3, 95]> lse_cast_fp16 = slice_by_index(begin = lse_begin_0, end = lse_end_0, end_mask = lse_end_mask_0, squeeze_mask = lse_squeeze_mask_0, x = var_8462_cast_fp16)[name = string("lse_cast_fp16")];
            tensor<int32, [2]> input_begin_0 = const()[name = string("input_begin_0"), val = tensor<int32, [2]>([0, 1])];
            tensor<int32, [2]> input_end_0 = const()[name = string("input_end_0"), val = tensor<int32, [2]>([3, 96])];
            tensor<bool, [2]> input_end_mask_0 = const()[name = string("input_end_mask_0"), val = tensor<bool, [2]>([true, true])];
            tensor<int32, [3, 95]> input = slice_by_index(begin = input_begin_0, end = input_end_0, end_mask = input_end_mask_0, x = ids)[name = string("input")];
            int32 var_8479_axis_0 = const()[name = string("op_8479_axis_0"), val = int32(0)];
            int32 var_8479_batch_dims_0 = const()[name = string("op_8479_batch_dims_0"), val = int32(0)];
            bool var_8479_validate_indices_0 = const()[name = string("op_8479_validate_indices_0"), val = bool(false)];
            string input_to_uint16_dtype_0 = const()[name = string("input_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [3, 95]> input_to_uint16 = cast(dtype = input_to_uint16_dtype_0, x = input)[name = string("cast_132")];
            tensor<fp16, [3, 95, 640]> var_8479_cast_fp16_cast_uint16 = gather(axis = var_8479_axis_0, batch_dims = var_8479_batch_dims_0, indices = input_to_uint16, validate_indices = var_8479_validate_indices_0, x = emb_to_fp16)[name = string("op_8479_cast_fp16_cast_uint16")];
            tensor<fp16, [3, 95, 640]> var_8480_cast_fp16 = mul(x = h_cast_fp16, y = var_8479_cast_fp16_cast_uint16)[name = string("op_8480_cast_fp16")];
            tensor<int32, [1]> tgt_axes_0 = const()[name = string("tgt_axes_0"), val = tensor<int32, [1]>([-1])];
            bool tgt_keep_dims_0 = const()[name = string("tgt_keep_dims_0"), val = bool(false)];
            tensor<fp16, [3, 95]> tgt_cast_fp16 = reduce_sum(axes = tgt_axes_0, keep_dims = tgt_keep_dims_0, x = var_8480_cast_fp16)[name = string("tgt_cast_fp16")];
            tensor<fp16, [3, 95]> var_8487_cast_fp16 = sub(x = tgt_cast_fp16, y = lse_cast_fp16)[name = string("op_8487_cast_fp16")];
            string var_8487_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_8487_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [3, 95]> lp = cast(dtype = var_8487_cast_fp16_to_fp32_dtype_0, x = var_8487_cast_fp16)[name = string("cast_131")];
        } -> (lp);
}