program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}})]
{
    func embed128<ios18>(tensor<int32, [1, 128, 17]> input_ids) {
            tensor<int32, [3]> input_1_begin_0 = const()[name = string("input_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_1_end_0 = const()[name = string("input_1_end_0"), val = tensor<int32, [3]>([1, 128, 1])];
            tensor<bool, [3]> input_1_end_mask_0 = const()[name = string("input_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> input_1_squeeze_mask_0 = const()[name = string("input_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 128]> input_1 = slice_by_index(begin = input_1_begin_0, end = input_1_end_0, end_mask = input_1_end_mask_0, squeeze_mask = input_1_squeeze_mask_0, x = input_ids)[name = string("input_1")];
            int32 embeddings_1_batch_dims_0 = const()[name = string("embeddings_1_batch_dims_0"), val = int32(0)];
            bool embeddings_1_validate_indices_0 = const()[name = string("embeddings_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [16384, 768]> text_embedding_weight_to_fp16 = const()[name = string("text_embedding_weight_to_fp16"), val = tensor<fp16, [16384, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40558976)))];
            string input_1_to_int16_dtype_0 = const()[name = string("input_1_to_int16_dtype_0"), val = string("int16")];
            string cast_18_dtype_0 = const()[name = string("cast_18_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, 128]> input_1_to_int16 = cast(dtype = input_1_to_int16_dtype_0, x = input_1)[name = string("cast_53")];
            tensor<int32, [1, 128]> cast_18 = cast(dtype = cast_18_dtype_0, x = input_1_to_int16)[name = string("cast_52")];
            tensor<bool, [1, 128]> greater_equal_0 = greater_equal(x = cast_18, 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(16384)];
            tensor<int32, [1, 128]> add_0 = add(x = cast_18, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 128]> select_0 = select(a = cast_18, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 embeddings_1_cast_fp16_cast_uint16_axis_0 = const()[name = string("embeddings_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, 128]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_51")];
            tensor<fp16, [1, 128, 768]> embeddings_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = embeddings_1_cast_fp16_cast_uint16_axis_0, batch_dims = embeddings_1_batch_dims_0, indices = select_0_to_int16, validate_indices = embeddings_1_validate_indices_0, x = text_embedding_weight_to_fp16)[name = string("embeddings_1_cast_fp16_cast_uint16_cast_uint16")];
            tensor<int32, [3]> channel_ids_1_begin_0 = const()[name = string("channel_ids_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> channel_ids_1_end_0 = const()[name = string("channel_ids_1_end_0"), val = tensor<int32, [3]>([1, 128, 2])];
            tensor<bool, [3]> channel_ids_1_end_mask_0 = const()[name = string("channel_ids_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_1_squeeze_mask_0 = const()[name = string("channel_ids_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 128]> channel_ids_1 = slice_by_index(begin = channel_ids_1_begin_0, end = channel_ids_1_end_0, end_mask = channel_ids_1_end_mask_0, squeeze_mask = channel_ids_1_squeeze_mask_0, x = input_ids)[name = string("channel_ids_1")];
            int32 var_70 = const()[name = string("op_70"), val = int32(1024)];
            tensor<bool, [1, 128]> valid_1 = not_equal(x = channel_ids_1, y = var_70)[name = string("valid_1")];
            tensor<int32, [1, 128]> var_77 = const()[name = string("op_77"), val = tensor<int32, [1, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65724864)))];
            tensor<int32, [1, 128]> input_3 = select(a = channel_ids_1, b = var_77, cond = valid_1)[name = string("input_3")];
            int32 var_82_axis_0 = const()[name = string("op_82_axis_0"), val = int32(0)];
            int32 var_82_batch_dims_0 = const()[name = string("op_82_batch_dims_0"), val = int32(0)];
            bool var_82_validate_indices_0 = const()[name = string("op_82_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_0_weight_to_fp16 = const()[name = string("audio_embeddings_0_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14183488)))];
            string input_3_to_uint16_dtype_0 = const()[name = string("input_3_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 128]> input_3_to_uint16 = cast(dtype = input_3_to_uint16_dtype_0, x = input_3)[name = string("cast_50")];
            tensor<fp16, [1, 128, 768]> var_82_cast_fp16_cast_uint16 = gather(axis = var_82_axis_0, batch_dims = var_82_batch_dims_0, indices = input_3_to_uint16, validate_indices = var_82_validate_indices_0, x = audio_embeddings_0_weight_to_fp16)[name = string("op_82_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_84_axes_0 = const()[name = string("op_84_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 128, 1]> var_84 = expand_dims(axes = var_84_axes_0, x = valid_1)[name = string("op_84")];
            string var_84_promoted_to_fp16_dtype_0 = const()[name = string("op_84_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 1]> var_84_to_fp16 = cast(dtype = var_84_promoted_to_fp16_dtype_0, x = var_84)[name = string("cast_49")];
            tensor<fp16, [1, 128, 768]> var_85_cast_fp16 = mul(x = var_82_cast_fp16_cast_uint16, y = var_84_to_fp16)[name = string("op_85_cast_fp16")];
            tensor<fp16, [1, 128, 768]> embeddings_3_cast_fp16 = add(x = embeddings_1_cast_fp16_cast_uint16_cast_uint16, y = var_85_cast_fp16)[name = string("embeddings_3_cast_fp16")];
            tensor<int32, [3]> channel_ids_3_begin_0 = const()[name = string("channel_ids_3_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
            tensor<int32, [3]> channel_ids_3_end_0 = const()[name = string("channel_ids_3_end_0"), val = tensor<int32, [3]>([1, 128, 3])];
            tensor<bool, [3]> channel_ids_3_end_mask_0 = const()[name = string("channel_ids_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_3_squeeze_mask_0 = const()[name = string("channel_ids_3_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 128]> channel_ids_3 = slice_by_index(begin = channel_ids_3_begin_0, end = channel_ids_3_end_0, end_mask = channel_ids_3_end_mask_0, squeeze_mask = channel_ids_3_squeeze_mask_0, x = input_ids)[name = string("channel_ids_3")];
            int32 var_101 = const()[name = string("op_101"), val = int32(1024)];
            tensor<bool, [1, 128]> valid_3 = not_equal(x = channel_ids_3, y = var_101)[name = string("valid_3")];
            tensor<int32, [1, 128]> input_5 = select(a = channel_ids_3, b = var_77, cond = valid_3)[name = string("input_5")];
            int32 var_113_axis_0 = const()[name = string("op_113_axis_0"), val = int32(0)];
            int32 var_113_batch_dims_0 = const()[name = string("op_113_batch_dims_0"), val = int32(0)];
            bool var_113_validate_indices_0 = const()[name = string("op_113_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_1_weight_to_fp16 = const()[name = string("audio_embeddings_1_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15758912)))];
            string input_5_to_uint16_dtype_0 = const()[name = string("input_5_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 128]> input_5_to_uint16 = cast(dtype = input_5_to_uint16_dtype_0, x = input_5)[name = string("cast_48")];
            tensor<fp16, [1, 128, 768]> var_113_cast_fp16_cast_uint16 = gather(axis = var_113_axis_0, batch_dims = var_113_batch_dims_0, indices = input_5_to_uint16, validate_indices = var_113_validate_indices_0, x = audio_embeddings_1_weight_to_fp16)[name = string("op_113_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_115_axes_0 = const()[name = string("op_115_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 128, 1]> var_115 = expand_dims(axes = var_115_axes_0, x = valid_3)[name = string("op_115")];
            string var_115_promoted_to_fp16_dtype_0 = const()[name = string("op_115_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 1]> var_115_to_fp16 = cast(dtype = var_115_promoted_to_fp16_dtype_0, x = var_115)[name = string("cast_47")];
            tensor<fp16, [1, 128, 768]> var_116_cast_fp16 = mul(x = var_113_cast_fp16_cast_uint16, y = var_115_to_fp16)[name = string("op_116_cast_fp16")];
            tensor<fp16, [1, 128, 768]> embeddings_5_cast_fp16 = add(x = embeddings_3_cast_fp16, y = var_116_cast_fp16)[name = string("embeddings_5_cast_fp16")];
            tensor<int32, [3]> channel_ids_5_begin_0 = const()[name = string("channel_ids_5_begin_0"), val = tensor<int32, [3]>([0, 0, 3])];
            tensor<int32, [3]> channel_ids_5_end_0 = const()[name = string("channel_ids_5_end_0"), val = tensor<int32, [3]>([1, 128, 4])];
            tensor<bool, [3]> channel_ids_5_end_mask_0 = const()[name = string("channel_ids_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_5_squeeze_mask_0 = const()[name = string("channel_ids_5_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 128]> channel_ids_5 = slice_by_index(begin = channel_ids_5_begin_0, end = channel_ids_5_end_0, end_mask = channel_ids_5_end_mask_0, squeeze_mask = channel_ids_5_squeeze_mask_0, x = input_ids)[name = string("channel_ids_5")];
            int32 var_132 = const()[name = string("op_132"), val = int32(1024)];
            tensor<bool, [1, 128]> valid_5 = not_equal(x = channel_ids_5, y = var_132)[name = string("valid_5")];
            tensor<int32, [1, 128]> input_7 = select(a = channel_ids_5, b = var_77, cond = valid_5)[name = string("input_7")];
            int32 var_144_axis_0 = const()[name = string("op_144_axis_0"), val = int32(0)];
            int32 var_144_batch_dims_0 = const()[name = string("op_144_batch_dims_0"), val = int32(0)];
            bool var_144_validate_indices_0 = const()[name = string("op_144_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_2_weight_to_fp16 = const()[name = string("audio_embeddings_2_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17332224)))];
            string input_7_to_uint16_dtype_0 = const()[name = string("input_7_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 128]> input_7_to_uint16 = cast(dtype = input_7_to_uint16_dtype_0, x = input_7)[name = string("cast_46")];
            tensor<fp16, [1, 128, 768]> var_144_cast_fp16_cast_uint16 = gather(axis = var_144_axis_0, batch_dims = var_144_batch_dims_0, indices = input_7_to_uint16, validate_indices = var_144_validate_indices_0, x = audio_embeddings_2_weight_to_fp16)[name = string("op_144_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_146_axes_0 = const()[name = string("op_146_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 128, 1]> var_146 = expand_dims(axes = var_146_axes_0, x = valid_5)[name = string("op_146")];
            string var_146_promoted_to_fp16_dtype_0 = const()[name = string("op_146_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 1]> var_146_to_fp16 = cast(dtype = var_146_promoted_to_fp16_dtype_0, x = var_146)[name = string("cast_45")];
            tensor<fp16, [1, 128, 768]> var_147_cast_fp16 = mul(x = var_144_cast_fp16_cast_uint16, y = var_146_to_fp16)[name = string("op_147_cast_fp16")];
            tensor<fp16, [1, 128, 768]> embeddings_7_cast_fp16 = add(x = embeddings_5_cast_fp16, y = var_147_cast_fp16)[name = string("embeddings_7_cast_fp16")];
            tensor<int32, [3]> channel_ids_7_begin_0 = const()[name = string("channel_ids_7_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> channel_ids_7_end_0 = const()[name = string("channel_ids_7_end_0"), val = tensor<int32, [3]>([1, 128, 5])];
            tensor<bool, [3]> channel_ids_7_end_mask_0 = const()[name = string("channel_ids_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_7_squeeze_mask_0 = const()[name = string("channel_ids_7_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 128]> channel_ids_7 = slice_by_index(begin = channel_ids_7_begin_0, end = channel_ids_7_end_0, end_mask = channel_ids_7_end_mask_0, squeeze_mask = channel_ids_7_squeeze_mask_0, x = input_ids)[name = string("channel_ids_7")];
            int32 var_163 = const()[name = string("op_163"), val = int32(1024)];
            tensor<bool, [1, 128]> valid_7 = not_equal(x = channel_ids_7, y = var_163)[name = string("valid_7")];
            tensor<int32, [1, 128]> input_9 = select(a = channel_ids_7, b = var_77, cond = valid_7)[name = string("input_9")];
            int32 var_175_axis_0 = const()[name = string("op_175_axis_0"), val = int32(0)];
            int32 var_175_batch_dims_0 = const()[name = string("op_175_batch_dims_0"), val = int32(0)];
            bool var_175_validate_indices_0 = const()[name = string("op_175_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_3_weight_to_fp16 = const()[name = string("audio_embeddings_3_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18905536)))];
            string input_9_to_uint16_dtype_0 = const()[name = string("input_9_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 128]> input_9_to_uint16 = cast(dtype = input_9_to_uint16_dtype_0, x = input_9)[name = string("cast_44")];
            tensor<fp16, [1, 128, 768]> var_175_cast_fp16_cast_uint16 = gather(axis = var_175_axis_0, batch_dims = var_175_batch_dims_0, indices = input_9_to_uint16, validate_indices = var_175_validate_indices_0, x = audio_embeddings_3_weight_to_fp16)[name = string("op_175_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_177_axes_0 = const()[name = string("op_177_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 128, 1]> var_177 = expand_dims(axes = var_177_axes_0, x = valid_7)[name = string("op_177")];
            string var_177_promoted_to_fp16_dtype_0 = const()[name = string("op_177_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 1]> var_177_to_fp16 = cast(dtype = var_177_promoted_to_fp16_dtype_0, x = var_177)[name = string("cast_43")];
            tensor<fp16, [1, 128, 768]> var_178_cast_fp16 = mul(x = var_175_cast_fp16_cast_uint16, y = var_177_to_fp16)[name = string("op_178_cast_fp16")];
            tensor<fp16, [1, 128, 768]> embeddings_9_cast_fp16 = add(x = embeddings_7_cast_fp16, y = var_178_cast_fp16)[name = string("embeddings_9_cast_fp16")];
            tensor<int32, [3]> channel_ids_9_begin_0 = const()[name = string("channel_ids_9_begin_0"), val = tensor<int32, [3]>([0, 0, 5])];
            tensor<int32, [3]> channel_ids_9_end_0 = const()[name = string("channel_ids_9_end_0"), val = tensor<int32, [3]>([1, 128, 6])];
            tensor<bool, [3]> channel_ids_9_end_mask_0 = const()[name = string("channel_ids_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_9_squeeze_mask_0 = const()[name = string("channel_ids_9_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 128]> channel_ids_9 = slice_by_index(begin = channel_ids_9_begin_0, end = channel_ids_9_end_0, end_mask = channel_ids_9_end_mask_0, squeeze_mask = channel_ids_9_squeeze_mask_0, x = input_ids)[name = string("channel_ids_9")];
            int32 var_194 = const()[name = string("op_194"), val = int32(1024)];
            tensor<bool, [1, 128]> valid_9 = not_equal(x = channel_ids_9, y = var_194)[name = string("valid_9")];
            tensor<int32, [1, 128]> input_11 = select(a = channel_ids_9, b = var_77, cond = valid_9)[name = string("input_11")];
            int32 var_206_axis_0 = const()[name = string("op_206_axis_0"), val = int32(0)];
            int32 var_206_batch_dims_0 = const()[name = string("op_206_batch_dims_0"), val = int32(0)];
            bool var_206_validate_indices_0 = const()[name = string("op_206_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_4_weight_to_fp16 = const()[name = string("audio_embeddings_4_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20478848)))];
            string input_11_to_uint16_dtype_0 = const()[name = string("input_11_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 128]> input_11_to_uint16 = cast(dtype = input_11_to_uint16_dtype_0, x = input_11)[name = string("cast_42")];
            tensor<fp16, [1, 128, 768]> var_206_cast_fp16_cast_uint16 = gather(axis = var_206_axis_0, batch_dims = var_206_batch_dims_0, indices = input_11_to_uint16, validate_indices = var_206_validate_indices_0, x = audio_embeddings_4_weight_to_fp16)[name = string("op_206_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_208_axes_0 = const()[name = string("op_208_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 128, 1]> var_208 = expand_dims(axes = var_208_axes_0, x = valid_9)[name = string("op_208")];
            string var_208_promoted_to_fp16_dtype_0 = const()[name = string("op_208_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 1]> var_208_to_fp16 = cast(dtype = var_208_promoted_to_fp16_dtype_0, x = var_208)[name = string("cast_41")];
            tensor<fp16, [1, 128, 768]> var_209_cast_fp16 = mul(x = var_206_cast_fp16_cast_uint16, y = var_208_to_fp16)[name = string("op_209_cast_fp16")];
            tensor<fp16, [1, 128, 768]> embeddings_11_cast_fp16 = add(x = embeddings_9_cast_fp16, y = var_209_cast_fp16)[name = string("embeddings_11_cast_fp16")];
            tensor<int32, [3]> channel_ids_11_begin_0 = const()[name = string("channel_ids_11_begin_0"), val = tensor<int32, [3]>([0, 0, 6])];
            tensor<int32, [3]> channel_ids_11_end_0 = const()[name = string("channel_ids_11_end_0"), val = tensor<int32, [3]>([1, 128, 7])];
            tensor<bool, [3]> channel_ids_11_end_mask_0 = const()[name = string("channel_ids_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_11_squeeze_mask_0 = const()[name = string("channel_ids_11_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 128]> channel_ids_11 = slice_by_index(begin = channel_ids_11_begin_0, end = channel_ids_11_end_0, end_mask = channel_ids_11_end_mask_0, squeeze_mask = channel_ids_11_squeeze_mask_0, x = input_ids)[name = string("channel_ids_11")];
            int32 var_225 = const()[name = string("op_225"), val = int32(1024)];
            tensor<bool, [1, 128]> valid_11 = not_equal(x = channel_ids_11, y = var_225)[name = string("valid_11")];
            tensor<int32, [1, 128]> input_13 = select(a = channel_ids_11, b = var_77, cond = valid_11)[name = string("input_13")];
            int32 var_237_axis_0 = const()[name = string("op_237_axis_0"), val = int32(0)];
            int32 var_237_batch_dims_0 = const()[name = string("op_237_batch_dims_0"), val = int32(0)];
            bool var_237_validate_indices_0 = const()[name = string("op_237_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_5_weight_to_fp16 = const()[name = string("audio_embeddings_5_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22052160)))];
            string input_13_to_uint16_dtype_0 = const()[name = string("input_13_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 128]> input_13_to_uint16 = cast(dtype = input_13_to_uint16_dtype_0, x = input_13)[name = string("cast_40")];
            tensor<fp16, [1, 128, 768]> var_237_cast_fp16_cast_uint16 = gather(axis = var_237_axis_0, batch_dims = var_237_batch_dims_0, indices = input_13_to_uint16, validate_indices = var_237_validate_indices_0, x = audio_embeddings_5_weight_to_fp16)[name = string("op_237_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_239_axes_0 = const()[name = string("op_239_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 128, 1]> var_239 = expand_dims(axes = var_239_axes_0, x = valid_11)[name = string("op_239")];
            string var_239_promoted_to_fp16_dtype_0 = const()[name = string("op_239_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 1]> var_239_to_fp16 = cast(dtype = var_239_promoted_to_fp16_dtype_0, x = var_239)[name = string("cast_39")];
            tensor<fp16, [1, 128, 768]> var_240_cast_fp16 = mul(x = var_237_cast_fp16_cast_uint16, y = var_239_to_fp16)[name = string("op_240_cast_fp16")];
            tensor<fp16, [1, 128, 768]> embeddings_13_cast_fp16 = add(x = embeddings_11_cast_fp16, y = var_240_cast_fp16)[name = string("embeddings_13_cast_fp16")];
            tensor<int32, [3]> channel_ids_13_begin_0 = const()[name = string("channel_ids_13_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
            tensor<int32, [3]> channel_ids_13_end_0 = const()[name = string("channel_ids_13_end_0"), val = tensor<int32, [3]>([1, 128, 8])];
            tensor<bool, [3]> channel_ids_13_end_mask_0 = const()[name = string("channel_ids_13_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_13_squeeze_mask_0 = const()[name = string("channel_ids_13_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 128]> channel_ids_13 = slice_by_index(begin = channel_ids_13_begin_0, end = channel_ids_13_end_0, end_mask = channel_ids_13_end_mask_0, squeeze_mask = channel_ids_13_squeeze_mask_0, x = input_ids)[name = string("channel_ids_13")];
            int32 var_256 = const()[name = string("op_256"), val = int32(1024)];
            tensor<bool, [1, 128]> valid_13 = not_equal(x = channel_ids_13, y = var_256)[name = string("valid_13")];
            tensor<int32, [1, 128]> input_15 = select(a = channel_ids_13, b = var_77, cond = valid_13)[name = string("input_15")];
            int32 var_268_axis_0 = const()[name = string("op_268_axis_0"), val = int32(0)];
            int32 var_268_batch_dims_0 = const()[name = string("op_268_batch_dims_0"), val = int32(0)];
            bool var_268_validate_indices_0 = const()[name = string("op_268_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_6_weight_to_fp16 = const()[name = string("audio_embeddings_6_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23625472)))];
            string input_15_to_uint16_dtype_0 = const()[name = string("input_15_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 128]> input_15_to_uint16 = cast(dtype = input_15_to_uint16_dtype_0, x = input_15)[name = string("cast_38")];
            tensor<fp16, [1, 128, 768]> var_268_cast_fp16_cast_uint16 = gather(axis = var_268_axis_0, batch_dims = var_268_batch_dims_0, indices = input_15_to_uint16, validate_indices = var_268_validate_indices_0, x = audio_embeddings_6_weight_to_fp16)[name = string("op_268_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_270_axes_0 = const()[name = string("op_270_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 128, 1]> var_270 = expand_dims(axes = var_270_axes_0, x = valid_13)[name = string("op_270")];
            string var_270_promoted_to_fp16_dtype_0 = const()[name = string("op_270_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 1]> var_270_to_fp16 = cast(dtype = var_270_promoted_to_fp16_dtype_0, x = var_270)[name = string("cast_37")];
            tensor<fp16, [1, 128, 768]> var_271_cast_fp16 = mul(x = var_268_cast_fp16_cast_uint16, y = var_270_to_fp16)[name = string("op_271_cast_fp16")];
            tensor<fp16, [1, 128, 768]> embeddings_15_cast_fp16 = add(x = embeddings_13_cast_fp16, y = var_271_cast_fp16)[name = string("embeddings_15_cast_fp16")];
            tensor<int32, [3]> channel_ids_15_begin_0 = const()[name = string("channel_ids_15_begin_0"), val = tensor<int32, [3]>([0, 0, 8])];
            tensor<int32, [3]> channel_ids_15_end_0 = const()[name = string("channel_ids_15_end_0"), val = tensor<int32, [3]>([1, 128, 9])];
            tensor<bool, [3]> channel_ids_15_end_mask_0 = const()[name = string("channel_ids_15_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_15_squeeze_mask_0 = const()[name = string("channel_ids_15_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 128]> channel_ids_15 = slice_by_index(begin = channel_ids_15_begin_0, end = channel_ids_15_end_0, end_mask = channel_ids_15_end_mask_0, squeeze_mask = channel_ids_15_squeeze_mask_0, x = input_ids)[name = string("channel_ids_15")];
            int32 var_287 = const()[name = string("op_287"), val = int32(1024)];
            tensor<bool, [1, 128]> valid_15 = not_equal(x = channel_ids_15, y = var_287)[name = string("valid_15")];
            tensor<int32, [1, 128]> input_17 = select(a = channel_ids_15, b = var_77, cond = valid_15)[name = string("input_17")];
            int32 var_299_axis_0 = const()[name = string("op_299_axis_0"), val = int32(0)];
            int32 var_299_batch_dims_0 = const()[name = string("op_299_batch_dims_0"), val = int32(0)];
            bool var_299_validate_indices_0 = const()[name = string("op_299_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_7_weight_to_fp16 = const()[name = string("audio_embeddings_7_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25198784)))];
            string input_17_to_uint16_dtype_0 = const()[name = string("input_17_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 128]> input_17_to_uint16 = cast(dtype = input_17_to_uint16_dtype_0, x = input_17)[name = string("cast_36")];
            tensor<fp16, [1, 128, 768]> var_299_cast_fp16_cast_uint16 = gather(axis = var_299_axis_0, batch_dims = var_299_batch_dims_0, indices = input_17_to_uint16, validate_indices = var_299_validate_indices_0, x = audio_embeddings_7_weight_to_fp16)[name = string("op_299_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_301_axes_0 = const()[name = string("op_301_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 128, 1]> var_301 = expand_dims(axes = var_301_axes_0, x = valid_15)[name = string("op_301")];
            string var_301_promoted_to_fp16_dtype_0 = const()[name = string("op_301_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 1]> var_301_to_fp16 = cast(dtype = var_301_promoted_to_fp16_dtype_0, x = var_301)[name = string("cast_35")];
            tensor<fp16, [1, 128, 768]> var_302_cast_fp16 = mul(x = var_299_cast_fp16_cast_uint16, y = var_301_to_fp16)[name = string("op_302_cast_fp16")];
            tensor<fp16, [1, 128, 768]> embeddings_17_cast_fp16 = add(x = embeddings_15_cast_fp16, y = var_302_cast_fp16)[name = string("embeddings_17_cast_fp16")];
            tensor<int32, [3]> channel_ids_17_begin_0 = const()[name = string("channel_ids_17_begin_0"), val = tensor<int32, [3]>([0, 0, 9])];
            tensor<int32, [3]> channel_ids_17_end_0 = const()[name = string("channel_ids_17_end_0"), val = tensor<int32, [3]>([1, 128, 10])];
            tensor<bool, [3]> channel_ids_17_end_mask_0 = const()[name = string("channel_ids_17_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_17_squeeze_mask_0 = const()[name = string("channel_ids_17_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 128]> channel_ids_17 = slice_by_index(begin = channel_ids_17_begin_0, end = channel_ids_17_end_0, end_mask = channel_ids_17_end_mask_0, squeeze_mask = channel_ids_17_squeeze_mask_0, x = input_ids)[name = string("channel_ids_17")];
            int32 var_318 = const()[name = string("op_318"), val = int32(1024)];
            tensor<bool, [1, 128]> valid_17 = not_equal(x = channel_ids_17, y = var_318)[name = string("valid_17")];
            tensor<int32, [1, 128]> input_19 = select(a = channel_ids_17, b = var_77, cond = valid_17)[name = string("input_19")];
            int32 var_330_axis_0 = const()[name = string("op_330_axis_0"), val = int32(0)];
            int32 var_330_batch_dims_0 = const()[name = string("op_330_batch_dims_0"), val = int32(0)];
            bool var_330_validate_indices_0 = const()[name = string("op_330_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_8_weight_to_fp16 = const()[name = string("audio_embeddings_8_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26772096)))];
            string input_19_to_uint16_dtype_0 = const()[name = string("input_19_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 128]> input_19_to_uint16 = cast(dtype = input_19_to_uint16_dtype_0, x = input_19)[name = string("cast_34")];
            tensor<fp16, [1, 128, 768]> var_330_cast_fp16_cast_uint16 = gather(axis = var_330_axis_0, batch_dims = var_330_batch_dims_0, indices = input_19_to_uint16, validate_indices = var_330_validate_indices_0, x = audio_embeddings_8_weight_to_fp16)[name = string("op_330_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_332_axes_0 = const()[name = string("op_332_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 128, 1]> var_332 = expand_dims(axes = var_332_axes_0, x = valid_17)[name = string("op_332")];
            string var_332_promoted_to_fp16_dtype_0 = const()[name = string("op_332_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 1]> var_332_to_fp16 = cast(dtype = var_332_promoted_to_fp16_dtype_0, x = var_332)[name = string("cast_33")];
            tensor<fp16, [1, 128, 768]> var_333_cast_fp16 = mul(x = var_330_cast_fp16_cast_uint16, y = var_332_to_fp16)[name = string("op_333_cast_fp16")];
            tensor<fp16, [1, 128, 768]> embeddings_19_cast_fp16 = add(x = embeddings_17_cast_fp16, y = var_333_cast_fp16)[name = string("embeddings_19_cast_fp16")];
            tensor<int32, [3]> channel_ids_19_begin_0 = const()[name = string("channel_ids_19_begin_0"), val = tensor<int32, [3]>([0, 0, 10])];
            tensor<int32, [3]> channel_ids_19_end_0 = const()[name = string("channel_ids_19_end_0"), val = tensor<int32, [3]>([1, 128, 11])];
            tensor<bool, [3]> channel_ids_19_end_mask_0 = const()[name = string("channel_ids_19_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_19_squeeze_mask_0 = const()[name = string("channel_ids_19_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 128]> channel_ids_19 = slice_by_index(begin = channel_ids_19_begin_0, end = channel_ids_19_end_0, end_mask = channel_ids_19_end_mask_0, squeeze_mask = channel_ids_19_squeeze_mask_0, x = input_ids)[name = string("channel_ids_19")];
            int32 var_349 = const()[name = string("op_349"), val = int32(1024)];
            tensor<bool, [1, 128]> valid_19 = not_equal(x = channel_ids_19, y = var_349)[name = string("valid_19")];
            tensor<int32, [1, 128]> input_21 = select(a = channel_ids_19, b = var_77, cond = valid_19)[name = string("input_21")];
            int32 var_361_axis_0 = const()[name = string("op_361_axis_0"), val = int32(0)];
            int32 var_361_batch_dims_0 = const()[name = string("op_361_batch_dims_0"), val = int32(0)];
            bool var_361_validate_indices_0 = const()[name = string("op_361_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_9_weight_to_fp16 = const()[name = string("audio_embeddings_9_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28345408)))];
            string input_21_to_uint16_dtype_0 = const()[name = string("input_21_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 128]> input_21_to_uint16 = cast(dtype = input_21_to_uint16_dtype_0, x = input_21)[name = string("cast_32")];
            tensor<fp16, [1, 128, 768]> var_361_cast_fp16_cast_uint16 = gather(axis = var_361_axis_0, batch_dims = var_361_batch_dims_0, indices = input_21_to_uint16, validate_indices = var_361_validate_indices_0, x = audio_embeddings_9_weight_to_fp16)[name = string("op_361_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_363_axes_0 = const()[name = string("op_363_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 128, 1]> var_363 = expand_dims(axes = var_363_axes_0, x = valid_19)[name = string("op_363")];
            string var_363_promoted_to_fp16_dtype_0 = const()[name = string("op_363_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 1]> var_363_to_fp16 = cast(dtype = var_363_promoted_to_fp16_dtype_0, x = var_363)[name = string("cast_31")];
            tensor<fp16, [1, 128, 768]> var_364_cast_fp16 = mul(x = var_361_cast_fp16_cast_uint16, y = var_363_to_fp16)[name = string("op_364_cast_fp16")];
            tensor<fp16, [1, 128, 768]> embeddings_21_cast_fp16 = add(x = embeddings_19_cast_fp16, y = var_364_cast_fp16)[name = string("embeddings_21_cast_fp16")];
            tensor<int32, [3]> channel_ids_21_begin_0 = const()[name = string("channel_ids_21_begin_0"), val = tensor<int32, [3]>([0, 0, 11])];
            tensor<int32, [3]> channel_ids_21_end_0 = const()[name = string("channel_ids_21_end_0"), val = tensor<int32, [3]>([1, 128, 12])];
            tensor<bool, [3]> channel_ids_21_end_mask_0 = const()[name = string("channel_ids_21_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_21_squeeze_mask_0 = const()[name = string("channel_ids_21_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 128]> channel_ids_21 = slice_by_index(begin = channel_ids_21_begin_0, end = channel_ids_21_end_0, end_mask = channel_ids_21_end_mask_0, squeeze_mask = channel_ids_21_squeeze_mask_0, x = input_ids)[name = string("channel_ids_21")];
            int32 var_380 = const()[name = string("op_380"), val = int32(1024)];
            tensor<bool, [1, 128]> valid_21 = not_equal(x = channel_ids_21, y = var_380)[name = string("valid_21")];
            tensor<int32, [1, 128]> input_23 = select(a = channel_ids_21, b = var_77, cond = valid_21)[name = string("input_23")];
            int32 var_392_axis_0 = const()[name = string("op_392_axis_0"), val = int32(0)];
            int32 var_392_batch_dims_0 = const()[name = string("op_392_batch_dims_0"), val = int32(0)];
            bool var_392_validate_indices_0 = const()[name = string("op_392_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_10_weight_to_fp16 = const()[name = string("audio_embeddings_10_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29918720)))];
            string input_23_to_uint16_dtype_0 = const()[name = string("input_23_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 128]> input_23_to_uint16 = cast(dtype = input_23_to_uint16_dtype_0, x = input_23)[name = string("cast_30")];
            tensor<fp16, [1, 128, 768]> var_392_cast_fp16_cast_uint16 = gather(axis = var_392_axis_0, batch_dims = var_392_batch_dims_0, indices = input_23_to_uint16, validate_indices = var_392_validate_indices_0, x = audio_embeddings_10_weight_to_fp16)[name = string("op_392_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_394_axes_0 = const()[name = string("op_394_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 128, 1]> var_394 = expand_dims(axes = var_394_axes_0, x = valid_21)[name = string("op_394")];
            string var_394_promoted_to_fp16_dtype_0 = const()[name = string("op_394_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 1]> var_394_to_fp16 = cast(dtype = var_394_promoted_to_fp16_dtype_0, x = var_394)[name = string("cast_29")];
            tensor<fp16, [1, 128, 768]> var_395_cast_fp16 = mul(x = var_392_cast_fp16_cast_uint16, y = var_394_to_fp16)[name = string("op_395_cast_fp16")];
            tensor<fp16, [1, 128, 768]> embeddings_23_cast_fp16 = add(x = embeddings_21_cast_fp16, y = var_395_cast_fp16)[name = string("embeddings_23_cast_fp16")];
            tensor<int32, [3]> channel_ids_23_begin_0 = const()[name = string("channel_ids_23_begin_0"), val = tensor<int32, [3]>([0, 0, 12])];
            tensor<int32, [3]> channel_ids_23_end_0 = const()[name = string("channel_ids_23_end_0"), val = tensor<int32, [3]>([1, 128, 13])];
            tensor<bool, [3]> channel_ids_23_end_mask_0 = const()[name = string("channel_ids_23_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_23_squeeze_mask_0 = const()[name = string("channel_ids_23_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 128]> channel_ids_23 = slice_by_index(begin = channel_ids_23_begin_0, end = channel_ids_23_end_0, end_mask = channel_ids_23_end_mask_0, squeeze_mask = channel_ids_23_squeeze_mask_0, x = input_ids)[name = string("channel_ids_23")];
            int32 var_411 = const()[name = string("op_411"), val = int32(1024)];
            tensor<bool, [1, 128]> valid_23 = not_equal(x = channel_ids_23, y = var_411)[name = string("valid_23")];
            tensor<int32, [1, 128]> input_25 = select(a = channel_ids_23, b = var_77, cond = valid_23)[name = string("input_25")];
            int32 var_423_axis_0 = const()[name = string("op_423_axis_0"), val = int32(0)];
            int32 var_423_batch_dims_0 = const()[name = string("op_423_batch_dims_0"), val = int32(0)];
            bool var_423_validate_indices_0 = const()[name = string("op_423_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_11_weight_to_fp16 = const()[name = string("audio_embeddings_11_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31492032)))];
            string input_25_to_uint16_dtype_0 = const()[name = string("input_25_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 128]> input_25_to_uint16 = cast(dtype = input_25_to_uint16_dtype_0, x = input_25)[name = string("cast_28")];
            tensor<fp16, [1, 128, 768]> var_423_cast_fp16_cast_uint16 = gather(axis = var_423_axis_0, batch_dims = var_423_batch_dims_0, indices = input_25_to_uint16, validate_indices = var_423_validate_indices_0, x = audio_embeddings_11_weight_to_fp16)[name = string("op_423_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_425_axes_0 = const()[name = string("op_425_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 128, 1]> var_425 = expand_dims(axes = var_425_axes_0, x = valid_23)[name = string("op_425")];
            string var_425_promoted_to_fp16_dtype_0 = const()[name = string("op_425_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 1]> var_425_to_fp16 = cast(dtype = var_425_promoted_to_fp16_dtype_0, x = var_425)[name = string("cast_27")];
            tensor<fp16, [1, 128, 768]> var_426_cast_fp16 = mul(x = var_423_cast_fp16_cast_uint16, y = var_425_to_fp16)[name = string("op_426_cast_fp16")];
            tensor<fp16, [1, 128, 768]> embeddings_25_cast_fp16 = add(x = embeddings_23_cast_fp16, y = var_426_cast_fp16)[name = string("embeddings_25_cast_fp16")];
            tensor<int32, [3]> channel_ids_25_begin_0 = const()[name = string("channel_ids_25_begin_0"), val = tensor<int32, [3]>([0, 0, 13])];
            tensor<int32, [3]> channel_ids_25_end_0 = const()[name = string("channel_ids_25_end_0"), val = tensor<int32, [3]>([1, 128, 14])];
            tensor<bool, [3]> channel_ids_25_end_mask_0 = const()[name = string("channel_ids_25_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_25_squeeze_mask_0 = const()[name = string("channel_ids_25_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 128]> channel_ids_25 = slice_by_index(begin = channel_ids_25_begin_0, end = channel_ids_25_end_0, end_mask = channel_ids_25_end_mask_0, squeeze_mask = channel_ids_25_squeeze_mask_0, x = input_ids)[name = string("channel_ids_25")];
            int32 var_442 = const()[name = string("op_442"), val = int32(1024)];
            tensor<bool, [1, 128]> valid_25 = not_equal(x = channel_ids_25, y = var_442)[name = string("valid_25")];
            tensor<int32, [1, 128]> input_27 = select(a = channel_ids_25, b = var_77, cond = valid_25)[name = string("input_27")];
            int32 var_454_axis_0 = const()[name = string("op_454_axis_0"), val = int32(0)];
            int32 var_454_batch_dims_0 = const()[name = string("op_454_batch_dims_0"), val = int32(0)];
            bool var_454_validate_indices_0 = const()[name = string("op_454_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_12_weight_to_fp16 = const()[name = string("audio_embeddings_12_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33065344)))];
            string input_27_to_uint16_dtype_0 = const()[name = string("input_27_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 128]> input_27_to_uint16 = cast(dtype = input_27_to_uint16_dtype_0, x = input_27)[name = string("cast_26")];
            tensor<fp16, [1, 128, 768]> var_454_cast_fp16_cast_uint16 = gather(axis = var_454_axis_0, batch_dims = var_454_batch_dims_0, indices = input_27_to_uint16, validate_indices = var_454_validate_indices_0, x = audio_embeddings_12_weight_to_fp16)[name = string("op_454_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_456_axes_0 = const()[name = string("op_456_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 128, 1]> var_456 = expand_dims(axes = var_456_axes_0, x = valid_25)[name = string("op_456")];
            string var_456_promoted_to_fp16_dtype_0 = const()[name = string("op_456_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 1]> var_456_to_fp16 = cast(dtype = var_456_promoted_to_fp16_dtype_0, x = var_456)[name = string("cast_25")];
            tensor<fp16, [1, 128, 768]> var_457_cast_fp16 = mul(x = var_454_cast_fp16_cast_uint16, y = var_456_to_fp16)[name = string("op_457_cast_fp16")];
            tensor<fp16, [1, 128, 768]> embeddings_27_cast_fp16 = add(x = embeddings_25_cast_fp16, y = var_457_cast_fp16)[name = string("embeddings_27_cast_fp16")];
            tensor<int32, [3]> channel_ids_27_begin_0 = const()[name = string("channel_ids_27_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
            tensor<int32, [3]> channel_ids_27_end_0 = const()[name = string("channel_ids_27_end_0"), val = tensor<int32, [3]>([1, 128, 15])];
            tensor<bool, [3]> channel_ids_27_end_mask_0 = const()[name = string("channel_ids_27_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_27_squeeze_mask_0 = const()[name = string("channel_ids_27_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 128]> channel_ids_27 = slice_by_index(begin = channel_ids_27_begin_0, end = channel_ids_27_end_0, end_mask = channel_ids_27_end_mask_0, squeeze_mask = channel_ids_27_squeeze_mask_0, x = input_ids)[name = string("channel_ids_27")];
            int32 var_473 = const()[name = string("op_473"), val = int32(1024)];
            tensor<bool, [1, 128]> valid_27 = not_equal(x = channel_ids_27, y = var_473)[name = string("valid_27")];
            tensor<int32, [1, 128]> input_29 = select(a = channel_ids_27, b = var_77, cond = valid_27)[name = string("input_29")];
            int32 var_485_axis_0 = const()[name = string("op_485_axis_0"), val = int32(0)];
            int32 var_485_batch_dims_0 = const()[name = string("op_485_batch_dims_0"), val = int32(0)];
            bool var_485_validate_indices_0 = const()[name = string("op_485_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_13_weight_to_fp16 = const()[name = string("audio_embeddings_13_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34638656)))];
            string input_29_to_uint16_dtype_0 = const()[name = string("input_29_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 128]> input_29_to_uint16 = cast(dtype = input_29_to_uint16_dtype_0, x = input_29)[name = string("cast_24")];
            tensor<fp16, [1, 128, 768]> var_485_cast_fp16_cast_uint16 = gather(axis = var_485_axis_0, batch_dims = var_485_batch_dims_0, indices = input_29_to_uint16, validate_indices = var_485_validate_indices_0, x = audio_embeddings_13_weight_to_fp16)[name = string("op_485_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_487_axes_0 = const()[name = string("op_487_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 128, 1]> var_487 = expand_dims(axes = var_487_axes_0, x = valid_27)[name = string("op_487")];
            string var_487_promoted_to_fp16_dtype_0 = const()[name = string("op_487_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 1]> var_487_to_fp16 = cast(dtype = var_487_promoted_to_fp16_dtype_0, x = var_487)[name = string("cast_23")];
            tensor<fp16, [1, 128, 768]> var_488_cast_fp16 = mul(x = var_485_cast_fp16_cast_uint16, y = var_487_to_fp16)[name = string("op_488_cast_fp16")];
            tensor<fp16, [1, 128, 768]> embeddings_29_cast_fp16 = add(x = embeddings_27_cast_fp16, y = var_488_cast_fp16)[name = string("embeddings_29_cast_fp16")];
            tensor<int32, [3]> channel_ids_29_begin_0 = const()[name = string("channel_ids_29_begin_0"), val = tensor<int32, [3]>([0, 0, 15])];
            tensor<int32, [3]> channel_ids_29_end_0 = const()[name = string("channel_ids_29_end_0"), val = tensor<int32, [3]>([1, 128, 16])];
            tensor<bool, [3]> channel_ids_29_end_mask_0 = const()[name = string("channel_ids_29_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_29_squeeze_mask_0 = const()[name = string("channel_ids_29_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 128]> channel_ids_29 = slice_by_index(begin = channel_ids_29_begin_0, end = channel_ids_29_end_0, end_mask = channel_ids_29_end_mask_0, squeeze_mask = channel_ids_29_squeeze_mask_0, x = input_ids)[name = string("channel_ids_29")];
            int32 var_504 = const()[name = string("op_504"), val = int32(1024)];
            tensor<bool, [1, 128]> valid_29 = not_equal(x = channel_ids_29, y = var_504)[name = string("valid_29")];
            tensor<int32, [1, 128]> input_31 = select(a = channel_ids_29, b = var_77, cond = valid_29)[name = string("input_31")];
            int32 var_516_axis_0 = const()[name = string("op_516_axis_0"), val = int32(0)];
            int32 var_516_batch_dims_0 = const()[name = string("op_516_batch_dims_0"), val = int32(0)];
            bool var_516_validate_indices_0 = const()[name = string("op_516_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_14_weight_to_fp16 = const()[name = string("audio_embeddings_14_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36211968)))];
            string input_31_to_uint16_dtype_0 = const()[name = string("input_31_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 128]> input_31_to_uint16 = cast(dtype = input_31_to_uint16_dtype_0, x = input_31)[name = string("cast_22")];
            tensor<fp16, [1, 128, 768]> var_516_cast_fp16_cast_uint16 = gather(axis = var_516_axis_0, batch_dims = var_516_batch_dims_0, indices = input_31_to_uint16, validate_indices = var_516_validate_indices_0, x = audio_embeddings_14_weight_to_fp16)[name = string("op_516_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_518_axes_0 = const()[name = string("op_518_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 128, 1]> var_518 = expand_dims(axes = var_518_axes_0, x = valid_29)[name = string("op_518")];
            string var_518_promoted_to_fp16_dtype_0 = const()[name = string("op_518_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 1]> var_518_to_fp16 = cast(dtype = var_518_promoted_to_fp16_dtype_0, x = var_518)[name = string("cast_21")];
            tensor<fp16, [1, 128, 768]> var_519_cast_fp16 = mul(x = var_516_cast_fp16_cast_uint16, y = var_518_to_fp16)[name = string("op_519_cast_fp16")];
            tensor<fp16, [1, 128, 768]> embeddings_cast_fp16 = add(x = embeddings_29_cast_fp16, y = var_519_cast_fp16)[name = string("embeddings_cast_fp16")];
            tensor<int32, [3]> channel_ids_begin_0 = const()[name = string("channel_ids_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> channel_ids_end_0 = const()[name = string("channel_ids_end_0"), val = tensor<int32, [3]>([1, 128, 17])];
            tensor<bool, [3]> channel_ids_end_mask_0 = const()[name = string("channel_ids_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_squeeze_mask_0 = const()[name = string("channel_ids_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 128]> channel_ids = slice_by_index(begin = channel_ids_begin_0, end = channel_ids_end_0, end_mask = channel_ids_end_mask_0, squeeze_mask = channel_ids_squeeze_mask_0, x = input_ids)[name = string("channel_ids")];
            int32 var_535 = const()[name = string("op_535"), val = int32(1024)];
            tensor<bool, [1, 128]> valid = not_equal(x = channel_ids, y = var_535)[name = string("valid")];
            tensor<int32, [1, 128]> input = select(a = channel_ids, b = var_77, cond = valid)[name = string("input")];
            int32 var_547_axis_0 = const()[name = string("op_547_axis_0"), val = int32(0)];
            int32 var_547_batch_dims_0 = const()[name = string("op_547_batch_dims_0"), val = int32(0)];
            bool var_547_validate_indices_0 = const()[name = string("op_547_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_15_weight_to_fp16 = const()[name = string("audio_embeddings_15_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37785280)))];
            string input_to_uint16_dtype_0 = const()[name = string("input_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 128]> input_to_uint16 = cast(dtype = input_to_uint16_dtype_0, x = input)[name = string("cast_20")];
            tensor<fp16, [1, 128, 768]> var_547_cast_fp16_cast_uint16 = gather(axis = var_547_axis_0, batch_dims = var_547_batch_dims_0, indices = input_to_uint16, validate_indices = var_547_validate_indices_0, x = audio_embeddings_15_weight_to_fp16)[name = string("op_547_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_549_axes_0 = const()[name = string("op_549_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 128, 1]> var_549 = expand_dims(axes = var_549_axes_0, x = valid)[name = string("op_549")];
            string var_549_promoted_to_fp16_dtype_0 = const()[name = string("op_549_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128, 1]> var_549_to_fp16 = cast(dtype = var_549_promoted_to_fp16_dtype_0, x = var_549)[name = string("cast_19")];
            tensor<fp16, [1, 128, 768]> var_550_cast_fp16 = mul(x = var_547_cast_fp16_cast_uint16, y = var_549_to_fp16)[name = string("op_550_cast_fp16")];
            tensor<fp16, [1, 128, 768]> embeddings = add(x = embeddings_cast_fp16, y = var_550_cast_fp16)[name = string("op_552_cast_fp16")];
        } -> (embeddings);
    func embed256<ios18>(tensor<int32, [1, 256, 17]> input_ids) {
            tensor<int32, [3]> input_1_begin_0 = const()[name = string("input_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> input_1_end_0 = const()[name = string("input_1_end_0"), val = tensor<int32, [3]>([1, 256, 1])];
            tensor<bool, [3]> input_1_end_mask_0 = const()[name = string("input_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> input_1_squeeze_mask_0 = const()[name = string("input_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 256]> input_1 = slice_by_index(begin = input_1_begin_0, end = input_1_end_0, end_mask = input_1_end_mask_0, squeeze_mask = input_1_squeeze_mask_0, x = input_ids)[name = string("input_1")];
            int32 embeddings_1_batch_dims_0 = const()[name = string("embeddings_1_batch_dims_0"), val = int32(0)];
            bool embeddings_1_validate_indices_0 = const()[name = string("embeddings_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [16384, 768]> text_embedding_weight_to_fp16 = const()[name = string("text_embedding_weight_to_fp16"), val = tensor<fp16, [16384, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40558976)))];
            string input_1_to_int16_dtype_0 = const()[name = string("input_1_to_int16_dtype_0"), val = string("int16")];
            string cast_18_dtype_0 = const()[name = string("cast_18_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, 256]> input_1_to_int16 = cast(dtype = input_1_to_int16_dtype_0, x = input_1)[name = string("cast_53")];
            tensor<int32, [1, 256]> cast_18 = cast(dtype = cast_18_dtype_0, x = input_1_to_int16)[name = string("cast_52")];
            tensor<bool, [1, 256]> greater_equal_0 = greater_equal(x = cast_18, 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(16384)];
            tensor<int32, [1, 256]> add_0 = add(x = cast_18, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 256]> select_0 = select(a = cast_18, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 embeddings_1_cast_fp16_cast_uint16_axis_0 = const()[name = string("embeddings_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, 256]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_51")];
            tensor<fp16, [1, 256, 768]> embeddings_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = embeddings_1_cast_fp16_cast_uint16_axis_0, batch_dims = embeddings_1_batch_dims_0, indices = select_0_to_int16, validate_indices = embeddings_1_validate_indices_0, x = text_embedding_weight_to_fp16)[name = string("embeddings_1_cast_fp16_cast_uint16_cast_uint16")];
            tensor<int32, [3]> channel_ids_1_begin_0 = const()[name = string("channel_ids_1_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
            tensor<int32, [3]> channel_ids_1_end_0 = const()[name = string("channel_ids_1_end_0"), val = tensor<int32, [3]>([1, 256, 2])];
            tensor<bool, [3]> channel_ids_1_end_mask_0 = const()[name = string("channel_ids_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_1_squeeze_mask_0 = const()[name = string("channel_ids_1_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 256]> channel_ids_1 = slice_by_index(begin = channel_ids_1_begin_0, end = channel_ids_1_end_0, end_mask = channel_ids_1_end_mask_0, squeeze_mask = channel_ids_1_squeeze_mask_0, x = input_ids)[name = string("channel_ids_1")];
            int32 var_70 = const()[name = string("op_70"), val = int32(1024)];
            tensor<bool, [1, 256]> valid_1 = not_equal(x = channel_ids_1, y = var_70)[name = string("valid_1")];
            tensor<int32, [1, 256]> var_77 = const()[name = string("op_77"), val = tensor<int32, [1, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65725440)))];
            tensor<int32, [1, 256]> input_3 = select(a = channel_ids_1, b = var_77, cond = valid_1)[name = string("input_3")];
            int32 var_82_axis_0 = const()[name = string("op_82_axis_0"), val = int32(0)];
            int32 var_82_batch_dims_0 = const()[name = string("op_82_batch_dims_0"), val = int32(0)];
            bool var_82_validate_indices_0 = const()[name = string("op_82_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_0_weight_to_fp16 = const()[name = string("audio_embeddings_0_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14183488)))];
            string input_3_to_uint16_dtype_0 = const()[name = string("input_3_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 256]> input_3_to_uint16 = cast(dtype = input_3_to_uint16_dtype_0, x = input_3)[name = string("cast_50")];
            tensor<fp16, [1, 256, 768]> var_82_cast_fp16_cast_uint16 = gather(axis = var_82_axis_0, batch_dims = var_82_batch_dims_0, indices = input_3_to_uint16, validate_indices = var_82_validate_indices_0, x = audio_embeddings_0_weight_to_fp16)[name = string("op_82_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_84_axes_0 = const()[name = string("op_84_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 256, 1]> var_84 = expand_dims(axes = var_84_axes_0, x = valid_1)[name = string("op_84")];
            string var_84_promoted_to_fp16_dtype_0 = const()[name = string("op_84_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256, 1]> var_84_to_fp16 = cast(dtype = var_84_promoted_to_fp16_dtype_0, x = var_84)[name = string("cast_49")];
            tensor<fp16, [1, 256, 768]> var_85_cast_fp16 = mul(x = var_82_cast_fp16_cast_uint16, y = var_84_to_fp16)[name = string("op_85_cast_fp16")];
            tensor<fp16, [1, 256, 768]> embeddings_3_cast_fp16 = add(x = embeddings_1_cast_fp16_cast_uint16_cast_uint16, y = var_85_cast_fp16)[name = string("embeddings_3_cast_fp16")];
            tensor<int32, [3]> channel_ids_3_begin_0 = const()[name = string("channel_ids_3_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
            tensor<int32, [3]> channel_ids_3_end_0 = const()[name = string("channel_ids_3_end_0"), val = tensor<int32, [3]>([1, 256, 3])];
            tensor<bool, [3]> channel_ids_3_end_mask_0 = const()[name = string("channel_ids_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_3_squeeze_mask_0 = const()[name = string("channel_ids_3_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 256]> channel_ids_3 = slice_by_index(begin = channel_ids_3_begin_0, end = channel_ids_3_end_0, end_mask = channel_ids_3_end_mask_0, squeeze_mask = channel_ids_3_squeeze_mask_0, x = input_ids)[name = string("channel_ids_3")];
            int32 var_101 = const()[name = string("op_101"), val = int32(1024)];
            tensor<bool, [1, 256]> valid_3 = not_equal(x = channel_ids_3, y = var_101)[name = string("valid_3")];
            tensor<int32, [1, 256]> input_5 = select(a = channel_ids_3, b = var_77, cond = valid_3)[name = string("input_5")];
            int32 var_113_axis_0 = const()[name = string("op_113_axis_0"), val = int32(0)];
            int32 var_113_batch_dims_0 = const()[name = string("op_113_batch_dims_0"), val = int32(0)];
            bool var_113_validate_indices_0 = const()[name = string("op_113_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_1_weight_to_fp16 = const()[name = string("audio_embeddings_1_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15758912)))];
            string input_5_to_uint16_dtype_0 = const()[name = string("input_5_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 256]> input_5_to_uint16 = cast(dtype = input_5_to_uint16_dtype_0, x = input_5)[name = string("cast_48")];
            tensor<fp16, [1, 256, 768]> var_113_cast_fp16_cast_uint16 = gather(axis = var_113_axis_0, batch_dims = var_113_batch_dims_0, indices = input_5_to_uint16, validate_indices = var_113_validate_indices_0, x = audio_embeddings_1_weight_to_fp16)[name = string("op_113_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_115_axes_0 = const()[name = string("op_115_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 256, 1]> var_115 = expand_dims(axes = var_115_axes_0, x = valid_3)[name = string("op_115")];
            string var_115_promoted_to_fp16_dtype_0 = const()[name = string("op_115_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256, 1]> var_115_to_fp16 = cast(dtype = var_115_promoted_to_fp16_dtype_0, x = var_115)[name = string("cast_47")];
            tensor<fp16, [1, 256, 768]> var_116_cast_fp16 = mul(x = var_113_cast_fp16_cast_uint16, y = var_115_to_fp16)[name = string("op_116_cast_fp16")];
            tensor<fp16, [1, 256, 768]> embeddings_5_cast_fp16 = add(x = embeddings_3_cast_fp16, y = var_116_cast_fp16)[name = string("embeddings_5_cast_fp16")];
            tensor<int32, [3]> channel_ids_5_begin_0 = const()[name = string("channel_ids_5_begin_0"), val = tensor<int32, [3]>([0, 0, 3])];
            tensor<int32, [3]> channel_ids_5_end_0 = const()[name = string("channel_ids_5_end_0"), val = tensor<int32, [3]>([1, 256, 4])];
            tensor<bool, [3]> channel_ids_5_end_mask_0 = const()[name = string("channel_ids_5_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_5_squeeze_mask_0 = const()[name = string("channel_ids_5_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 256]> channel_ids_5 = slice_by_index(begin = channel_ids_5_begin_0, end = channel_ids_5_end_0, end_mask = channel_ids_5_end_mask_0, squeeze_mask = channel_ids_5_squeeze_mask_0, x = input_ids)[name = string("channel_ids_5")];
            int32 var_132 = const()[name = string("op_132"), val = int32(1024)];
            tensor<bool, [1, 256]> valid_5 = not_equal(x = channel_ids_5, y = var_132)[name = string("valid_5")];
            tensor<int32, [1, 256]> input_7 = select(a = channel_ids_5, b = var_77, cond = valid_5)[name = string("input_7")];
            int32 var_144_axis_0 = const()[name = string("op_144_axis_0"), val = int32(0)];
            int32 var_144_batch_dims_0 = const()[name = string("op_144_batch_dims_0"), val = int32(0)];
            bool var_144_validate_indices_0 = const()[name = string("op_144_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_2_weight_to_fp16 = const()[name = string("audio_embeddings_2_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17332224)))];
            string input_7_to_uint16_dtype_0 = const()[name = string("input_7_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 256]> input_7_to_uint16 = cast(dtype = input_7_to_uint16_dtype_0, x = input_7)[name = string("cast_46")];
            tensor<fp16, [1, 256, 768]> var_144_cast_fp16_cast_uint16 = gather(axis = var_144_axis_0, batch_dims = var_144_batch_dims_0, indices = input_7_to_uint16, validate_indices = var_144_validate_indices_0, x = audio_embeddings_2_weight_to_fp16)[name = string("op_144_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_146_axes_0 = const()[name = string("op_146_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 256, 1]> var_146 = expand_dims(axes = var_146_axes_0, x = valid_5)[name = string("op_146")];
            string var_146_promoted_to_fp16_dtype_0 = const()[name = string("op_146_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256, 1]> var_146_to_fp16 = cast(dtype = var_146_promoted_to_fp16_dtype_0, x = var_146)[name = string("cast_45")];
            tensor<fp16, [1, 256, 768]> var_147_cast_fp16 = mul(x = var_144_cast_fp16_cast_uint16, y = var_146_to_fp16)[name = string("op_147_cast_fp16")];
            tensor<fp16, [1, 256, 768]> embeddings_7_cast_fp16 = add(x = embeddings_5_cast_fp16, y = var_147_cast_fp16)[name = string("embeddings_7_cast_fp16")];
            tensor<int32, [3]> channel_ids_7_begin_0 = const()[name = string("channel_ids_7_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> channel_ids_7_end_0 = const()[name = string("channel_ids_7_end_0"), val = tensor<int32, [3]>([1, 256, 5])];
            tensor<bool, [3]> channel_ids_7_end_mask_0 = const()[name = string("channel_ids_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_7_squeeze_mask_0 = const()[name = string("channel_ids_7_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 256]> channel_ids_7 = slice_by_index(begin = channel_ids_7_begin_0, end = channel_ids_7_end_0, end_mask = channel_ids_7_end_mask_0, squeeze_mask = channel_ids_7_squeeze_mask_0, x = input_ids)[name = string("channel_ids_7")];
            int32 var_163 = const()[name = string("op_163"), val = int32(1024)];
            tensor<bool, [1, 256]> valid_7 = not_equal(x = channel_ids_7, y = var_163)[name = string("valid_7")];
            tensor<int32, [1, 256]> input_9 = select(a = channel_ids_7, b = var_77, cond = valid_7)[name = string("input_9")];
            int32 var_175_axis_0 = const()[name = string("op_175_axis_0"), val = int32(0)];
            int32 var_175_batch_dims_0 = const()[name = string("op_175_batch_dims_0"), val = int32(0)];
            bool var_175_validate_indices_0 = const()[name = string("op_175_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_3_weight_to_fp16 = const()[name = string("audio_embeddings_3_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18905536)))];
            string input_9_to_uint16_dtype_0 = const()[name = string("input_9_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 256]> input_9_to_uint16 = cast(dtype = input_9_to_uint16_dtype_0, x = input_9)[name = string("cast_44")];
            tensor<fp16, [1, 256, 768]> var_175_cast_fp16_cast_uint16 = gather(axis = var_175_axis_0, batch_dims = var_175_batch_dims_0, indices = input_9_to_uint16, validate_indices = var_175_validate_indices_0, x = audio_embeddings_3_weight_to_fp16)[name = string("op_175_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_177_axes_0 = const()[name = string("op_177_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 256, 1]> var_177 = expand_dims(axes = var_177_axes_0, x = valid_7)[name = string("op_177")];
            string var_177_promoted_to_fp16_dtype_0 = const()[name = string("op_177_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256, 1]> var_177_to_fp16 = cast(dtype = var_177_promoted_to_fp16_dtype_0, x = var_177)[name = string("cast_43")];
            tensor<fp16, [1, 256, 768]> var_178_cast_fp16 = mul(x = var_175_cast_fp16_cast_uint16, y = var_177_to_fp16)[name = string("op_178_cast_fp16")];
            tensor<fp16, [1, 256, 768]> embeddings_9_cast_fp16 = add(x = embeddings_7_cast_fp16, y = var_178_cast_fp16)[name = string("embeddings_9_cast_fp16")];
            tensor<int32, [3]> channel_ids_9_begin_0 = const()[name = string("channel_ids_9_begin_0"), val = tensor<int32, [3]>([0, 0, 5])];
            tensor<int32, [3]> channel_ids_9_end_0 = const()[name = string("channel_ids_9_end_0"), val = tensor<int32, [3]>([1, 256, 6])];
            tensor<bool, [3]> channel_ids_9_end_mask_0 = const()[name = string("channel_ids_9_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_9_squeeze_mask_0 = const()[name = string("channel_ids_9_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 256]> channel_ids_9 = slice_by_index(begin = channel_ids_9_begin_0, end = channel_ids_9_end_0, end_mask = channel_ids_9_end_mask_0, squeeze_mask = channel_ids_9_squeeze_mask_0, x = input_ids)[name = string("channel_ids_9")];
            int32 var_194 = const()[name = string("op_194"), val = int32(1024)];
            tensor<bool, [1, 256]> valid_9 = not_equal(x = channel_ids_9, y = var_194)[name = string("valid_9")];
            tensor<int32, [1, 256]> input_11 = select(a = channel_ids_9, b = var_77, cond = valid_9)[name = string("input_11")];
            int32 var_206_axis_0 = const()[name = string("op_206_axis_0"), val = int32(0)];
            int32 var_206_batch_dims_0 = const()[name = string("op_206_batch_dims_0"), val = int32(0)];
            bool var_206_validate_indices_0 = const()[name = string("op_206_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_4_weight_to_fp16 = const()[name = string("audio_embeddings_4_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20478848)))];
            string input_11_to_uint16_dtype_0 = const()[name = string("input_11_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 256]> input_11_to_uint16 = cast(dtype = input_11_to_uint16_dtype_0, x = input_11)[name = string("cast_42")];
            tensor<fp16, [1, 256, 768]> var_206_cast_fp16_cast_uint16 = gather(axis = var_206_axis_0, batch_dims = var_206_batch_dims_0, indices = input_11_to_uint16, validate_indices = var_206_validate_indices_0, x = audio_embeddings_4_weight_to_fp16)[name = string("op_206_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_208_axes_0 = const()[name = string("op_208_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 256, 1]> var_208 = expand_dims(axes = var_208_axes_0, x = valid_9)[name = string("op_208")];
            string var_208_promoted_to_fp16_dtype_0 = const()[name = string("op_208_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256, 1]> var_208_to_fp16 = cast(dtype = var_208_promoted_to_fp16_dtype_0, x = var_208)[name = string("cast_41")];
            tensor<fp16, [1, 256, 768]> var_209_cast_fp16 = mul(x = var_206_cast_fp16_cast_uint16, y = var_208_to_fp16)[name = string("op_209_cast_fp16")];
            tensor<fp16, [1, 256, 768]> embeddings_11_cast_fp16 = add(x = embeddings_9_cast_fp16, y = var_209_cast_fp16)[name = string("embeddings_11_cast_fp16")];
            tensor<int32, [3]> channel_ids_11_begin_0 = const()[name = string("channel_ids_11_begin_0"), val = tensor<int32, [3]>([0, 0, 6])];
            tensor<int32, [3]> channel_ids_11_end_0 = const()[name = string("channel_ids_11_end_0"), val = tensor<int32, [3]>([1, 256, 7])];
            tensor<bool, [3]> channel_ids_11_end_mask_0 = const()[name = string("channel_ids_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_11_squeeze_mask_0 = const()[name = string("channel_ids_11_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 256]> channel_ids_11 = slice_by_index(begin = channel_ids_11_begin_0, end = channel_ids_11_end_0, end_mask = channel_ids_11_end_mask_0, squeeze_mask = channel_ids_11_squeeze_mask_0, x = input_ids)[name = string("channel_ids_11")];
            int32 var_225 = const()[name = string("op_225"), val = int32(1024)];
            tensor<bool, [1, 256]> valid_11 = not_equal(x = channel_ids_11, y = var_225)[name = string("valid_11")];
            tensor<int32, [1, 256]> input_13 = select(a = channel_ids_11, b = var_77, cond = valid_11)[name = string("input_13")];
            int32 var_237_axis_0 = const()[name = string("op_237_axis_0"), val = int32(0)];
            int32 var_237_batch_dims_0 = const()[name = string("op_237_batch_dims_0"), val = int32(0)];
            bool var_237_validate_indices_0 = const()[name = string("op_237_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_5_weight_to_fp16 = const()[name = string("audio_embeddings_5_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22052160)))];
            string input_13_to_uint16_dtype_0 = const()[name = string("input_13_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 256]> input_13_to_uint16 = cast(dtype = input_13_to_uint16_dtype_0, x = input_13)[name = string("cast_40")];
            tensor<fp16, [1, 256, 768]> var_237_cast_fp16_cast_uint16 = gather(axis = var_237_axis_0, batch_dims = var_237_batch_dims_0, indices = input_13_to_uint16, validate_indices = var_237_validate_indices_0, x = audio_embeddings_5_weight_to_fp16)[name = string("op_237_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_239_axes_0 = const()[name = string("op_239_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 256, 1]> var_239 = expand_dims(axes = var_239_axes_0, x = valid_11)[name = string("op_239")];
            string var_239_promoted_to_fp16_dtype_0 = const()[name = string("op_239_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256, 1]> var_239_to_fp16 = cast(dtype = var_239_promoted_to_fp16_dtype_0, x = var_239)[name = string("cast_39")];
            tensor<fp16, [1, 256, 768]> var_240_cast_fp16 = mul(x = var_237_cast_fp16_cast_uint16, y = var_239_to_fp16)[name = string("op_240_cast_fp16")];
            tensor<fp16, [1, 256, 768]> embeddings_13_cast_fp16 = add(x = embeddings_11_cast_fp16, y = var_240_cast_fp16)[name = string("embeddings_13_cast_fp16")];
            tensor<int32, [3]> channel_ids_13_begin_0 = const()[name = string("channel_ids_13_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
            tensor<int32, [3]> channel_ids_13_end_0 = const()[name = string("channel_ids_13_end_0"), val = tensor<int32, [3]>([1, 256, 8])];
            tensor<bool, [3]> channel_ids_13_end_mask_0 = const()[name = string("channel_ids_13_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_13_squeeze_mask_0 = const()[name = string("channel_ids_13_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 256]> channel_ids_13 = slice_by_index(begin = channel_ids_13_begin_0, end = channel_ids_13_end_0, end_mask = channel_ids_13_end_mask_0, squeeze_mask = channel_ids_13_squeeze_mask_0, x = input_ids)[name = string("channel_ids_13")];
            int32 var_256 = const()[name = string("op_256"), val = int32(1024)];
            tensor<bool, [1, 256]> valid_13 = not_equal(x = channel_ids_13, y = var_256)[name = string("valid_13")];
            tensor<int32, [1, 256]> input_15 = select(a = channel_ids_13, b = var_77, cond = valid_13)[name = string("input_15")];
            int32 var_268_axis_0 = const()[name = string("op_268_axis_0"), val = int32(0)];
            int32 var_268_batch_dims_0 = const()[name = string("op_268_batch_dims_0"), val = int32(0)];
            bool var_268_validate_indices_0 = const()[name = string("op_268_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_6_weight_to_fp16 = const()[name = string("audio_embeddings_6_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23625472)))];
            string input_15_to_uint16_dtype_0 = const()[name = string("input_15_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 256]> input_15_to_uint16 = cast(dtype = input_15_to_uint16_dtype_0, x = input_15)[name = string("cast_38")];
            tensor<fp16, [1, 256, 768]> var_268_cast_fp16_cast_uint16 = gather(axis = var_268_axis_0, batch_dims = var_268_batch_dims_0, indices = input_15_to_uint16, validate_indices = var_268_validate_indices_0, x = audio_embeddings_6_weight_to_fp16)[name = string("op_268_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_270_axes_0 = const()[name = string("op_270_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 256, 1]> var_270 = expand_dims(axes = var_270_axes_0, x = valid_13)[name = string("op_270")];
            string var_270_promoted_to_fp16_dtype_0 = const()[name = string("op_270_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256, 1]> var_270_to_fp16 = cast(dtype = var_270_promoted_to_fp16_dtype_0, x = var_270)[name = string("cast_37")];
            tensor<fp16, [1, 256, 768]> var_271_cast_fp16 = mul(x = var_268_cast_fp16_cast_uint16, y = var_270_to_fp16)[name = string("op_271_cast_fp16")];
            tensor<fp16, [1, 256, 768]> embeddings_15_cast_fp16 = add(x = embeddings_13_cast_fp16, y = var_271_cast_fp16)[name = string("embeddings_15_cast_fp16")];
            tensor<int32, [3]> channel_ids_15_begin_0 = const()[name = string("channel_ids_15_begin_0"), val = tensor<int32, [3]>([0, 0, 8])];
            tensor<int32, [3]> channel_ids_15_end_0 = const()[name = string("channel_ids_15_end_0"), val = tensor<int32, [3]>([1, 256, 9])];
            tensor<bool, [3]> channel_ids_15_end_mask_0 = const()[name = string("channel_ids_15_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_15_squeeze_mask_0 = const()[name = string("channel_ids_15_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 256]> channel_ids_15 = slice_by_index(begin = channel_ids_15_begin_0, end = channel_ids_15_end_0, end_mask = channel_ids_15_end_mask_0, squeeze_mask = channel_ids_15_squeeze_mask_0, x = input_ids)[name = string("channel_ids_15")];
            int32 var_287 = const()[name = string("op_287"), val = int32(1024)];
            tensor<bool, [1, 256]> valid_15 = not_equal(x = channel_ids_15, y = var_287)[name = string("valid_15")];
            tensor<int32, [1, 256]> input_17 = select(a = channel_ids_15, b = var_77, cond = valid_15)[name = string("input_17")];
            int32 var_299_axis_0 = const()[name = string("op_299_axis_0"), val = int32(0)];
            int32 var_299_batch_dims_0 = const()[name = string("op_299_batch_dims_0"), val = int32(0)];
            bool var_299_validate_indices_0 = const()[name = string("op_299_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_7_weight_to_fp16 = const()[name = string("audio_embeddings_7_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25198784)))];
            string input_17_to_uint16_dtype_0 = const()[name = string("input_17_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 256]> input_17_to_uint16 = cast(dtype = input_17_to_uint16_dtype_0, x = input_17)[name = string("cast_36")];
            tensor<fp16, [1, 256, 768]> var_299_cast_fp16_cast_uint16 = gather(axis = var_299_axis_0, batch_dims = var_299_batch_dims_0, indices = input_17_to_uint16, validate_indices = var_299_validate_indices_0, x = audio_embeddings_7_weight_to_fp16)[name = string("op_299_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_301_axes_0 = const()[name = string("op_301_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 256, 1]> var_301 = expand_dims(axes = var_301_axes_0, x = valid_15)[name = string("op_301")];
            string var_301_promoted_to_fp16_dtype_0 = const()[name = string("op_301_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256, 1]> var_301_to_fp16 = cast(dtype = var_301_promoted_to_fp16_dtype_0, x = var_301)[name = string("cast_35")];
            tensor<fp16, [1, 256, 768]> var_302_cast_fp16 = mul(x = var_299_cast_fp16_cast_uint16, y = var_301_to_fp16)[name = string("op_302_cast_fp16")];
            tensor<fp16, [1, 256, 768]> embeddings_17_cast_fp16 = add(x = embeddings_15_cast_fp16, y = var_302_cast_fp16)[name = string("embeddings_17_cast_fp16")];
            tensor<int32, [3]> channel_ids_17_begin_0 = const()[name = string("channel_ids_17_begin_0"), val = tensor<int32, [3]>([0, 0, 9])];
            tensor<int32, [3]> channel_ids_17_end_0 = const()[name = string("channel_ids_17_end_0"), val = tensor<int32, [3]>([1, 256, 10])];
            tensor<bool, [3]> channel_ids_17_end_mask_0 = const()[name = string("channel_ids_17_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_17_squeeze_mask_0 = const()[name = string("channel_ids_17_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 256]> channel_ids_17 = slice_by_index(begin = channel_ids_17_begin_0, end = channel_ids_17_end_0, end_mask = channel_ids_17_end_mask_0, squeeze_mask = channel_ids_17_squeeze_mask_0, x = input_ids)[name = string("channel_ids_17")];
            int32 var_318 = const()[name = string("op_318"), val = int32(1024)];
            tensor<bool, [1, 256]> valid_17 = not_equal(x = channel_ids_17, y = var_318)[name = string("valid_17")];
            tensor<int32, [1, 256]> input_19 = select(a = channel_ids_17, b = var_77, cond = valid_17)[name = string("input_19")];
            int32 var_330_axis_0 = const()[name = string("op_330_axis_0"), val = int32(0)];
            int32 var_330_batch_dims_0 = const()[name = string("op_330_batch_dims_0"), val = int32(0)];
            bool var_330_validate_indices_0 = const()[name = string("op_330_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_8_weight_to_fp16 = const()[name = string("audio_embeddings_8_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26772096)))];
            string input_19_to_uint16_dtype_0 = const()[name = string("input_19_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 256]> input_19_to_uint16 = cast(dtype = input_19_to_uint16_dtype_0, x = input_19)[name = string("cast_34")];
            tensor<fp16, [1, 256, 768]> var_330_cast_fp16_cast_uint16 = gather(axis = var_330_axis_0, batch_dims = var_330_batch_dims_0, indices = input_19_to_uint16, validate_indices = var_330_validate_indices_0, x = audio_embeddings_8_weight_to_fp16)[name = string("op_330_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_332_axes_0 = const()[name = string("op_332_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 256, 1]> var_332 = expand_dims(axes = var_332_axes_0, x = valid_17)[name = string("op_332")];
            string var_332_promoted_to_fp16_dtype_0 = const()[name = string("op_332_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256, 1]> var_332_to_fp16 = cast(dtype = var_332_promoted_to_fp16_dtype_0, x = var_332)[name = string("cast_33")];
            tensor<fp16, [1, 256, 768]> var_333_cast_fp16 = mul(x = var_330_cast_fp16_cast_uint16, y = var_332_to_fp16)[name = string("op_333_cast_fp16")];
            tensor<fp16, [1, 256, 768]> embeddings_19_cast_fp16 = add(x = embeddings_17_cast_fp16, y = var_333_cast_fp16)[name = string("embeddings_19_cast_fp16")];
            tensor<int32, [3]> channel_ids_19_begin_0 = const()[name = string("channel_ids_19_begin_0"), val = tensor<int32, [3]>([0, 0, 10])];
            tensor<int32, [3]> channel_ids_19_end_0 = const()[name = string("channel_ids_19_end_0"), val = tensor<int32, [3]>([1, 256, 11])];
            tensor<bool, [3]> channel_ids_19_end_mask_0 = const()[name = string("channel_ids_19_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_19_squeeze_mask_0 = const()[name = string("channel_ids_19_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 256]> channel_ids_19 = slice_by_index(begin = channel_ids_19_begin_0, end = channel_ids_19_end_0, end_mask = channel_ids_19_end_mask_0, squeeze_mask = channel_ids_19_squeeze_mask_0, x = input_ids)[name = string("channel_ids_19")];
            int32 var_349 = const()[name = string("op_349"), val = int32(1024)];
            tensor<bool, [1, 256]> valid_19 = not_equal(x = channel_ids_19, y = var_349)[name = string("valid_19")];
            tensor<int32, [1, 256]> input_21 = select(a = channel_ids_19, b = var_77, cond = valid_19)[name = string("input_21")];
            int32 var_361_axis_0 = const()[name = string("op_361_axis_0"), val = int32(0)];
            int32 var_361_batch_dims_0 = const()[name = string("op_361_batch_dims_0"), val = int32(0)];
            bool var_361_validate_indices_0 = const()[name = string("op_361_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_9_weight_to_fp16 = const()[name = string("audio_embeddings_9_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28345408)))];
            string input_21_to_uint16_dtype_0 = const()[name = string("input_21_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 256]> input_21_to_uint16 = cast(dtype = input_21_to_uint16_dtype_0, x = input_21)[name = string("cast_32")];
            tensor<fp16, [1, 256, 768]> var_361_cast_fp16_cast_uint16 = gather(axis = var_361_axis_0, batch_dims = var_361_batch_dims_0, indices = input_21_to_uint16, validate_indices = var_361_validate_indices_0, x = audio_embeddings_9_weight_to_fp16)[name = string("op_361_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_363_axes_0 = const()[name = string("op_363_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 256, 1]> var_363 = expand_dims(axes = var_363_axes_0, x = valid_19)[name = string("op_363")];
            string var_363_promoted_to_fp16_dtype_0 = const()[name = string("op_363_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256, 1]> var_363_to_fp16 = cast(dtype = var_363_promoted_to_fp16_dtype_0, x = var_363)[name = string("cast_31")];
            tensor<fp16, [1, 256, 768]> var_364_cast_fp16 = mul(x = var_361_cast_fp16_cast_uint16, y = var_363_to_fp16)[name = string("op_364_cast_fp16")];
            tensor<fp16, [1, 256, 768]> embeddings_21_cast_fp16 = add(x = embeddings_19_cast_fp16, y = var_364_cast_fp16)[name = string("embeddings_21_cast_fp16")];
            tensor<int32, [3]> channel_ids_21_begin_0 = const()[name = string("channel_ids_21_begin_0"), val = tensor<int32, [3]>([0, 0, 11])];
            tensor<int32, [3]> channel_ids_21_end_0 = const()[name = string("channel_ids_21_end_0"), val = tensor<int32, [3]>([1, 256, 12])];
            tensor<bool, [3]> channel_ids_21_end_mask_0 = const()[name = string("channel_ids_21_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_21_squeeze_mask_0 = const()[name = string("channel_ids_21_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 256]> channel_ids_21 = slice_by_index(begin = channel_ids_21_begin_0, end = channel_ids_21_end_0, end_mask = channel_ids_21_end_mask_0, squeeze_mask = channel_ids_21_squeeze_mask_0, x = input_ids)[name = string("channel_ids_21")];
            int32 var_380 = const()[name = string("op_380"), val = int32(1024)];
            tensor<bool, [1, 256]> valid_21 = not_equal(x = channel_ids_21, y = var_380)[name = string("valid_21")];
            tensor<int32, [1, 256]> input_23 = select(a = channel_ids_21, b = var_77, cond = valid_21)[name = string("input_23")];
            int32 var_392_axis_0 = const()[name = string("op_392_axis_0"), val = int32(0)];
            int32 var_392_batch_dims_0 = const()[name = string("op_392_batch_dims_0"), val = int32(0)];
            bool var_392_validate_indices_0 = const()[name = string("op_392_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_10_weight_to_fp16 = const()[name = string("audio_embeddings_10_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29918720)))];
            string input_23_to_uint16_dtype_0 = const()[name = string("input_23_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 256]> input_23_to_uint16 = cast(dtype = input_23_to_uint16_dtype_0, x = input_23)[name = string("cast_30")];
            tensor<fp16, [1, 256, 768]> var_392_cast_fp16_cast_uint16 = gather(axis = var_392_axis_0, batch_dims = var_392_batch_dims_0, indices = input_23_to_uint16, validate_indices = var_392_validate_indices_0, x = audio_embeddings_10_weight_to_fp16)[name = string("op_392_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_394_axes_0 = const()[name = string("op_394_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 256, 1]> var_394 = expand_dims(axes = var_394_axes_0, x = valid_21)[name = string("op_394")];
            string var_394_promoted_to_fp16_dtype_0 = const()[name = string("op_394_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256, 1]> var_394_to_fp16 = cast(dtype = var_394_promoted_to_fp16_dtype_0, x = var_394)[name = string("cast_29")];
            tensor<fp16, [1, 256, 768]> var_395_cast_fp16 = mul(x = var_392_cast_fp16_cast_uint16, y = var_394_to_fp16)[name = string("op_395_cast_fp16")];
            tensor<fp16, [1, 256, 768]> embeddings_23_cast_fp16 = add(x = embeddings_21_cast_fp16, y = var_395_cast_fp16)[name = string("embeddings_23_cast_fp16")];
            tensor<int32, [3]> channel_ids_23_begin_0 = const()[name = string("channel_ids_23_begin_0"), val = tensor<int32, [3]>([0, 0, 12])];
            tensor<int32, [3]> channel_ids_23_end_0 = const()[name = string("channel_ids_23_end_0"), val = tensor<int32, [3]>([1, 256, 13])];
            tensor<bool, [3]> channel_ids_23_end_mask_0 = const()[name = string("channel_ids_23_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_23_squeeze_mask_0 = const()[name = string("channel_ids_23_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 256]> channel_ids_23 = slice_by_index(begin = channel_ids_23_begin_0, end = channel_ids_23_end_0, end_mask = channel_ids_23_end_mask_0, squeeze_mask = channel_ids_23_squeeze_mask_0, x = input_ids)[name = string("channel_ids_23")];
            int32 var_411 = const()[name = string("op_411"), val = int32(1024)];
            tensor<bool, [1, 256]> valid_23 = not_equal(x = channel_ids_23, y = var_411)[name = string("valid_23")];
            tensor<int32, [1, 256]> input_25 = select(a = channel_ids_23, b = var_77, cond = valid_23)[name = string("input_25")];
            int32 var_423_axis_0 = const()[name = string("op_423_axis_0"), val = int32(0)];
            int32 var_423_batch_dims_0 = const()[name = string("op_423_batch_dims_0"), val = int32(0)];
            bool var_423_validate_indices_0 = const()[name = string("op_423_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_11_weight_to_fp16 = const()[name = string("audio_embeddings_11_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31492032)))];
            string input_25_to_uint16_dtype_0 = const()[name = string("input_25_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 256]> input_25_to_uint16 = cast(dtype = input_25_to_uint16_dtype_0, x = input_25)[name = string("cast_28")];
            tensor<fp16, [1, 256, 768]> var_423_cast_fp16_cast_uint16 = gather(axis = var_423_axis_0, batch_dims = var_423_batch_dims_0, indices = input_25_to_uint16, validate_indices = var_423_validate_indices_0, x = audio_embeddings_11_weight_to_fp16)[name = string("op_423_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_425_axes_0 = const()[name = string("op_425_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 256, 1]> var_425 = expand_dims(axes = var_425_axes_0, x = valid_23)[name = string("op_425")];
            string var_425_promoted_to_fp16_dtype_0 = const()[name = string("op_425_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256, 1]> var_425_to_fp16 = cast(dtype = var_425_promoted_to_fp16_dtype_0, x = var_425)[name = string("cast_27")];
            tensor<fp16, [1, 256, 768]> var_426_cast_fp16 = mul(x = var_423_cast_fp16_cast_uint16, y = var_425_to_fp16)[name = string("op_426_cast_fp16")];
            tensor<fp16, [1, 256, 768]> embeddings_25_cast_fp16 = add(x = embeddings_23_cast_fp16, y = var_426_cast_fp16)[name = string("embeddings_25_cast_fp16")];
            tensor<int32, [3]> channel_ids_25_begin_0 = const()[name = string("channel_ids_25_begin_0"), val = tensor<int32, [3]>([0, 0, 13])];
            tensor<int32, [3]> channel_ids_25_end_0 = const()[name = string("channel_ids_25_end_0"), val = tensor<int32, [3]>([1, 256, 14])];
            tensor<bool, [3]> channel_ids_25_end_mask_0 = const()[name = string("channel_ids_25_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_25_squeeze_mask_0 = const()[name = string("channel_ids_25_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 256]> channel_ids_25 = slice_by_index(begin = channel_ids_25_begin_0, end = channel_ids_25_end_0, end_mask = channel_ids_25_end_mask_0, squeeze_mask = channel_ids_25_squeeze_mask_0, x = input_ids)[name = string("channel_ids_25")];
            int32 var_442 = const()[name = string("op_442"), val = int32(1024)];
            tensor<bool, [1, 256]> valid_25 = not_equal(x = channel_ids_25, y = var_442)[name = string("valid_25")];
            tensor<int32, [1, 256]> input_27 = select(a = channel_ids_25, b = var_77, cond = valid_25)[name = string("input_27")];
            int32 var_454_axis_0 = const()[name = string("op_454_axis_0"), val = int32(0)];
            int32 var_454_batch_dims_0 = const()[name = string("op_454_batch_dims_0"), val = int32(0)];
            bool var_454_validate_indices_0 = const()[name = string("op_454_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_12_weight_to_fp16 = const()[name = string("audio_embeddings_12_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33065344)))];
            string input_27_to_uint16_dtype_0 = const()[name = string("input_27_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 256]> input_27_to_uint16 = cast(dtype = input_27_to_uint16_dtype_0, x = input_27)[name = string("cast_26")];
            tensor<fp16, [1, 256, 768]> var_454_cast_fp16_cast_uint16 = gather(axis = var_454_axis_0, batch_dims = var_454_batch_dims_0, indices = input_27_to_uint16, validate_indices = var_454_validate_indices_0, x = audio_embeddings_12_weight_to_fp16)[name = string("op_454_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_456_axes_0 = const()[name = string("op_456_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 256, 1]> var_456 = expand_dims(axes = var_456_axes_0, x = valid_25)[name = string("op_456")];
            string var_456_promoted_to_fp16_dtype_0 = const()[name = string("op_456_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256, 1]> var_456_to_fp16 = cast(dtype = var_456_promoted_to_fp16_dtype_0, x = var_456)[name = string("cast_25")];
            tensor<fp16, [1, 256, 768]> var_457_cast_fp16 = mul(x = var_454_cast_fp16_cast_uint16, y = var_456_to_fp16)[name = string("op_457_cast_fp16")];
            tensor<fp16, [1, 256, 768]> embeddings_27_cast_fp16 = add(x = embeddings_25_cast_fp16, y = var_457_cast_fp16)[name = string("embeddings_27_cast_fp16")];
            tensor<int32, [3]> channel_ids_27_begin_0 = const()[name = string("channel_ids_27_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
            tensor<int32, [3]> channel_ids_27_end_0 = const()[name = string("channel_ids_27_end_0"), val = tensor<int32, [3]>([1, 256, 15])];
            tensor<bool, [3]> channel_ids_27_end_mask_0 = const()[name = string("channel_ids_27_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_27_squeeze_mask_0 = const()[name = string("channel_ids_27_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 256]> channel_ids_27 = slice_by_index(begin = channel_ids_27_begin_0, end = channel_ids_27_end_0, end_mask = channel_ids_27_end_mask_0, squeeze_mask = channel_ids_27_squeeze_mask_0, x = input_ids)[name = string("channel_ids_27")];
            int32 var_473 = const()[name = string("op_473"), val = int32(1024)];
            tensor<bool, [1, 256]> valid_27 = not_equal(x = channel_ids_27, y = var_473)[name = string("valid_27")];
            tensor<int32, [1, 256]> input_29 = select(a = channel_ids_27, b = var_77, cond = valid_27)[name = string("input_29")];
            int32 var_485_axis_0 = const()[name = string("op_485_axis_0"), val = int32(0)];
            int32 var_485_batch_dims_0 = const()[name = string("op_485_batch_dims_0"), val = int32(0)];
            bool var_485_validate_indices_0 = const()[name = string("op_485_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_13_weight_to_fp16 = const()[name = string("audio_embeddings_13_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34638656)))];
            string input_29_to_uint16_dtype_0 = const()[name = string("input_29_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 256]> input_29_to_uint16 = cast(dtype = input_29_to_uint16_dtype_0, x = input_29)[name = string("cast_24")];
            tensor<fp16, [1, 256, 768]> var_485_cast_fp16_cast_uint16 = gather(axis = var_485_axis_0, batch_dims = var_485_batch_dims_0, indices = input_29_to_uint16, validate_indices = var_485_validate_indices_0, x = audio_embeddings_13_weight_to_fp16)[name = string("op_485_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_487_axes_0 = const()[name = string("op_487_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 256, 1]> var_487 = expand_dims(axes = var_487_axes_0, x = valid_27)[name = string("op_487")];
            string var_487_promoted_to_fp16_dtype_0 = const()[name = string("op_487_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256, 1]> var_487_to_fp16 = cast(dtype = var_487_promoted_to_fp16_dtype_0, x = var_487)[name = string("cast_23")];
            tensor<fp16, [1, 256, 768]> var_488_cast_fp16 = mul(x = var_485_cast_fp16_cast_uint16, y = var_487_to_fp16)[name = string("op_488_cast_fp16")];
            tensor<fp16, [1, 256, 768]> embeddings_29_cast_fp16 = add(x = embeddings_27_cast_fp16, y = var_488_cast_fp16)[name = string("embeddings_29_cast_fp16")];
            tensor<int32, [3]> channel_ids_29_begin_0 = const()[name = string("channel_ids_29_begin_0"), val = tensor<int32, [3]>([0, 0, 15])];
            tensor<int32, [3]> channel_ids_29_end_0 = const()[name = string("channel_ids_29_end_0"), val = tensor<int32, [3]>([1, 256, 16])];
            tensor<bool, [3]> channel_ids_29_end_mask_0 = const()[name = string("channel_ids_29_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_29_squeeze_mask_0 = const()[name = string("channel_ids_29_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 256]> channel_ids_29 = slice_by_index(begin = channel_ids_29_begin_0, end = channel_ids_29_end_0, end_mask = channel_ids_29_end_mask_0, squeeze_mask = channel_ids_29_squeeze_mask_0, x = input_ids)[name = string("channel_ids_29")];
            int32 var_504 = const()[name = string("op_504"), val = int32(1024)];
            tensor<bool, [1, 256]> valid_29 = not_equal(x = channel_ids_29, y = var_504)[name = string("valid_29")];
            tensor<int32, [1, 256]> input_31 = select(a = channel_ids_29, b = var_77, cond = valid_29)[name = string("input_31")];
            int32 var_516_axis_0 = const()[name = string("op_516_axis_0"), val = int32(0)];
            int32 var_516_batch_dims_0 = const()[name = string("op_516_batch_dims_0"), val = int32(0)];
            bool var_516_validate_indices_0 = const()[name = string("op_516_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_14_weight_to_fp16 = const()[name = string("audio_embeddings_14_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36211968)))];
            string input_31_to_uint16_dtype_0 = const()[name = string("input_31_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 256]> input_31_to_uint16 = cast(dtype = input_31_to_uint16_dtype_0, x = input_31)[name = string("cast_22")];
            tensor<fp16, [1, 256, 768]> var_516_cast_fp16_cast_uint16 = gather(axis = var_516_axis_0, batch_dims = var_516_batch_dims_0, indices = input_31_to_uint16, validate_indices = var_516_validate_indices_0, x = audio_embeddings_14_weight_to_fp16)[name = string("op_516_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_518_axes_0 = const()[name = string("op_518_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 256, 1]> var_518 = expand_dims(axes = var_518_axes_0, x = valid_29)[name = string("op_518")];
            string var_518_promoted_to_fp16_dtype_0 = const()[name = string("op_518_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256, 1]> var_518_to_fp16 = cast(dtype = var_518_promoted_to_fp16_dtype_0, x = var_518)[name = string("cast_21")];
            tensor<fp16, [1, 256, 768]> var_519_cast_fp16 = mul(x = var_516_cast_fp16_cast_uint16, y = var_518_to_fp16)[name = string("op_519_cast_fp16")];
            tensor<fp16, [1, 256, 768]> embeddings_cast_fp16 = add(x = embeddings_29_cast_fp16, y = var_519_cast_fp16)[name = string("embeddings_cast_fp16")];
            tensor<int32, [3]> channel_ids_begin_0 = const()[name = string("channel_ids_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> channel_ids_end_0 = const()[name = string("channel_ids_end_0"), val = tensor<int32, [3]>([1, 256, 17])];
            tensor<bool, [3]> channel_ids_end_mask_0 = const()[name = string("channel_ids_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> channel_ids_squeeze_mask_0 = const()[name = string("channel_ids_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<int32, [1, 256]> channel_ids = slice_by_index(begin = channel_ids_begin_0, end = channel_ids_end_0, end_mask = channel_ids_end_mask_0, squeeze_mask = channel_ids_squeeze_mask_0, x = input_ids)[name = string("channel_ids")];
            int32 var_535 = const()[name = string("op_535"), val = int32(1024)];
            tensor<bool, [1, 256]> valid = not_equal(x = channel_ids, y = var_535)[name = string("valid")];
            tensor<int32, [1, 256]> input = select(a = channel_ids, b = var_77, cond = valid)[name = string("input")];
            int32 var_547_axis_0 = const()[name = string("op_547_axis_0"), val = int32(0)];
            int32 var_547_batch_dims_0 = const()[name = string("op_547_batch_dims_0"), val = int32(0)];
            bool var_547_validate_indices_0 = const()[name = string("op_547_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 768]> audio_embeddings_15_weight_to_fp16 = const()[name = string("audio_embeddings_15_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37785280)))];
            string input_to_uint16_dtype_0 = const()[name = string("input_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 256]> input_to_uint16 = cast(dtype = input_to_uint16_dtype_0, x = input)[name = string("cast_20")];
            tensor<fp16, [1, 256, 768]> var_547_cast_fp16_cast_uint16 = gather(axis = var_547_axis_0, batch_dims = var_547_batch_dims_0, indices = input_to_uint16, validate_indices = var_547_validate_indices_0, x = audio_embeddings_15_weight_to_fp16)[name = string("op_547_cast_fp16_cast_uint16")];
            tensor<int32, [1]> var_549_axes_0 = const()[name = string("op_549_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, [1, 256, 1]> var_549 = expand_dims(axes = var_549_axes_0, x = valid)[name = string("op_549")];
            string var_549_promoted_to_fp16_dtype_0 = const()[name = string("op_549_promoted_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256, 1]> var_549_to_fp16 = cast(dtype = var_549_promoted_to_fp16_dtype_0, x = var_549)[name = string("cast_19")];
            tensor<fp16, [1, 256, 768]> var_550_cast_fp16 = mul(x = var_547_cast_fp16_cast_uint16, y = var_549_to_fp16)[name = string("op_550_cast_fp16")];
            tensor<fp16, [1, 256, 768]> embeddings = add(x = embeddings_cast_fp16, y = var_550_cast_fp16)[name = string("op_552_cast_fp16")];
        } -> (embeddings);
    func sample<ios18>(tensor<fp32, [1]> assistant_random, tensor<fp32, [1, 16]> audio_random, tensor<fp16, [1, 768]> global_hidden, tensor<int32, [1, 16, 1024]> repetition_seen) {
            tensor<int32, [3]> var_267 = const()[name = string("op_267"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 768]> input_1_cast_fp16 = reshape(shape = var_267, x = global_hidden)[name = string("input_1_cast_fp16")];
            tensor<int32, [1]> input_3_axes_0 = const()[name = string("input_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> block_ln_1_weight_to_fp16 = const()[name = string("block_ln_1_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            tensor<fp16, [768]> block_ln_1_bias_to_fp16 = const()[name = string("block_ln_1_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1664)))];
            fp16 var_270_to_fp16 = const()[name = string("op_270_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = block_ln_1_bias_to_fp16, epsilon = var_270_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [2304, 768]> block_attn_c_attn_weight_to_fp16 = const()[name = string("block_attn_c_attn_weight_to_fp16"), val = tensor<fp16, [2304, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3264)))];
            tensor<fp16, [2304]> block_attn_c_attn_bias_to_fp16 = const()[name = string("block_attn_c_attn_bias_to_fp16"), val = tensor<fp16, [2304]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3542272)))];
            tensor<fp16, [1, 1, 2304]> linear_0_cast_fp16 = linear(bias = block_attn_c_attn_bias_to_fp16, weight = block_attn_c_attn_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [3]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [3]>([768, 768, 768])];
            int32 var_281_axis_0 = const()[name = string("op_281_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 768]> var_281_cast_fp16_0, tensor<fp16, [1, 1, 768]> var_281_cast_fp16_1, tensor<fp16, [1, 1, 768]> var_281_cast_fp16_2 = split(axis = var_281_axis_0, split_sizes = tile_0, x = linear_0_cast_fp16)[name = string("op_281_cast_fp16")];
            tensor<int32, [4]> var_289 = const()[name = string("op_289"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> query_3_cast_fp16 = reshape(shape = var_289, x = var_281_cast_fp16_0)[name = string("query_3_cast_fp16")];
            tensor<int32, [4]> var_295 = const()[name = string("op_295"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> key_3_cast_fp16 = reshape(shape = var_295, x = var_281_cast_fp16_1)[name = string("key_3_cast_fp16")];
            tensor<int32, [4]> var_301 = const()[name = string("op_301"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> value_3_cast_fp16 = reshape(shape = var_301, x = var_281_cast_fp16_2)[name = string("value_3_cast_fp16")];
            tensor<int32, [4]> var_308_begin_0 = const()[name = string("op_308_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_308_end_0 = const()[name = string("op_308_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_308_end_mask_0 = const()[name = string("op_308_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_308_stride_0 = const()[name = string("op_308_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_308_cast_fp16 = slice_by_index(begin = var_308_begin_0, end = var_308_end_0, end_mask = var_308_end_mask_0, stride = var_308_stride_0, x = query_3_cast_fp16)[name = string("op_308_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_309_cast_fp16 = mul(x = var_308_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_309_cast_fp16")];
            tensor<int32, [4]> var_314_begin_0 = const()[name = string("op_314_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_314_end_0 = const()[name = string("op_314_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_314_end_mask_0 = const()[name = string("op_314_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_314_stride_0 = const()[name = string("op_314_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_314_cast_fp16 = slice_by_index(begin = var_314_begin_0, end = var_314_end_0, end_mask = var_314_end_mask_0, stride = var_314_stride_0, x = query_3_cast_fp16)[name = string("op_314_cast_fp16")];
            int32 var_317_axis_0 = const()[name = string("op_317_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_317_cast_fp16 = stack(axis = var_317_axis_0, values = (var_309_cast_fp16, var_314_cast_fp16))[name = string("op_317_cast_fp16")];
            tensor<int32, [4]> shape_1 = const()[name = string("shape_1"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_318_cast_fp16 = reshape(shape = shape_1, x = var_317_cast_fp16)[name = string("op_318_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_sin_0_to_fp16 = const()[name = string("rope_sin_0_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3546944)))];
            tensor<fp16, [1, 1, 12, 64]> var_319_cast_fp16 = mul(x = var_318_cast_fp16, y = rope_sin_0_to_fp16)[name = string("op_319_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> query_5_cast_fp16 = add(x = query_3_cast_fp16, y = var_319_cast_fp16)[name = string("query_5_cast_fp16")];
            tensor<int32, [4]> var_327_begin_0 = const()[name = string("op_327_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_327_end_0 = const()[name = string("op_327_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_327_end_mask_0 = const()[name = string("op_327_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_327_stride_0 = const()[name = string("op_327_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_327_cast_fp16 = slice_by_index(begin = var_327_begin_0, end = var_327_end_0, end_mask = var_327_end_mask_0, stride = var_327_stride_0, x = key_3_cast_fp16)[name = string("op_327_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_328_cast_fp16 = mul(x = var_327_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_328_cast_fp16")];
            tensor<int32, [4]> var_333_begin_0 = const()[name = string("op_333_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_333_end_0 = const()[name = string("op_333_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_333_end_mask_0 = const()[name = string("op_333_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_333_stride_0 = const()[name = string("op_333_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_333_cast_fp16 = slice_by_index(begin = var_333_begin_0, end = var_333_end_0, end_mask = var_333_end_mask_0, stride = var_333_stride_0, x = key_3_cast_fp16)[name = string("op_333_cast_fp16")];
            int32 var_336_axis_0 = const()[name = string("op_336_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_336_cast_fp16 = stack(axis = var_336_axis_0, values = (var_328_cast_fp16, var_333_cast_fp16))[name = string("op_336_cast_fp16")];
            tensor<int32, [4]> shape_2 = const()[name = string("shape_2"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_337_cast_fp16 = reshape(shape = shape_2, x = var_336_cast_fp16)[name = string("op_337_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_338_cast_fp16 = mul(x = var_337_cast_fp16, y = rope_sin_0_to_fp16)[name = string("op_338_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> key_5_cast_fp16 = add(x = key_3_cast_fp16, y = var_338_cast_fp16)[name = string("key_5_cast_fp16")];
            tensor<int32, [4]> var_357 = const()[name = string("op_357"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_362_transpose_x_1 = const()[name = string("op_362_transpose_x_1"), val = bool(false)];
            bool var_362_transpose_y_1 = const()[name = string("op_362_transpose_y_1"), val = bool(true)];
            tensor<int32, [4]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_2_perm_0 = const()[name = string("transpose_2_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 12, 1, 64]> transpose_2 = transpose(perm = transpose_2_perm_0, x = key_5_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 12, 1, 64]> transpose_1 = transpose(perm = transpose_1_perm_0, x = query_5_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 12, 1, 1]> var_362_cast_fp16 = matmul(transpose_x = var_362_transpose_x_1, transpose_y = var_362_transpose_y_1, x = transpose_1, y = transpose_2)[name = string("op_362_cast_fp16")];
            fp16 var_363_to_fp16 = const()[name = string("op_363_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 1, 1]> scores_1_cast_fp16 = mul(x = var_362_cast_fp16, y = var_363_to_fp16)[name = string("scores_1_cast_fp16")];
            int32 var_365 = const()[name = string("op_365"), val = int32(-1)];
            tensor<fp16, [1, 12, 1, 1]> probabilities_1_cast_fp16 = softmax(axis = var_365, x = scores_1_cast_fp16)[name = string("probabilities_1_cast_fp16")];
            bool attention_1_transpose_x_0 = const()[name = string("attention_1_transpose_x_0"), val = bool(false)];
            bool attention_1_transpose_y_0 = const()[name = string("attention_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> value_5_cast_fp16 = transpose(perm = var_357, x = value_3_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 12, 1, 64]> attention_1_cast_fp16 = matmul(transpose_x = attention_1_transpose_x_0, transpose_y = attention_1_transpose_y_0, x = probabilities_1_cast_fp16, y = value_5_cast_fp16)[name = string("attention_1_cast_fp16")];
            tensor<int32, [4]> var_373 = const()[name = string("op_373"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_378 = const()[name = string("op_378"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 12, 64]> var_374_cast_fp16 = transpose(perm = var_373, x = attention_1_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 1, 768]> input_5_cast_fp16 = reshape(shape = var_378, x = var_374_cast_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [768, 768]> block_attn_c_proj_weight_to_fp16 = const()[name = string("block_attn_c_proj_weight_to_fp16"), val = tensor<fp16, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3547136)))];
            tensor<fp16, [768]> block_attn_c_proj_bias_to_fp16 = const()[name = string("block_attn_c_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4726848)))];
            tensor<fp16, [1, 1, 768]> linear_1_cast_fp16 = linear(bias = block_attn_c_proj_bias_to_fp16, weight = block_attn_c_proj_weight_to_fp16, x = input_5_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_7_cast_fp16 = add(x = input_1_cast_fp16, y = linear_1_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<int32, [1]> input_9_axes_0 = const()[name = string("input_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> block_ln_2_bias_to_fp16 = const()[name = string("block_ln_2_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4728448)))];
            fp16 var_386_to_fp16 = const()[name = string("op_386_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = block_ln_2_bias_to_fp16, epsilon = var_386_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [3072, 768]> block_mlp_fc_in_weight_to_fp16 = const()[name = string("block_mlp_fc_in_weight_to_fp16"), val = tensor<fp16, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4730048)))];
            tensor<fp16, [3072]> block_mlp_fc_in_bias_to_fp16 = const()[name = string("block_mlp_fc_in_bias_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9448704)))];
            tensor<fp16, [1, 1, 3072]> linear_2_cast_fp16 = linear(bias = block_mlp_fc_in_bias_to_fp16, weight = block_mlp_fc_in_weight_to_fp16, x = input_9_cast_fp16)[name = string("linear_2_cast_fp16")];
            string input_13_mode_0 = const()[name = string("input_13_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1, 3072]> input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_2_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [768, 3072]> block_mlp_fc_out_weight_to_fp16 = const()[name = string("block_mlp_fc_out_weight_to_fp16"), val = tensor<fp16, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9454912)))];
            tensor<fp16, [768]> block_mlp_fc_out_bias_to_fp16 = const()[name = string("block_mlp_fc_out_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14173568)))];
            tensor<fp16, [1, 1, 768]> linear_3_cast_fp16 = linear(bias = block_mlp_fc_out_bias_to_fp16, weight = block_mlp_fc_out_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_17_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = string("input_17_cast_fp16")];
            tensor<int32, [1]> var_425_axes_0 = const()[name = string("op_425_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [768]> final_norm_bias_to_fp16 = const()[name = string("final_norm_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14175168)))];
            fp16 var_420_to_fp16 = const()[name = string("op_420_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> var_425_cast_fp16 = layer_norm(axes = var_425_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_420_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_17_cast_fp16)[name = string("op_425_cast_fp16")];
            tensor<int32, [3]> var_433_begin_0 = const()[name = string("op_433_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_433_end_0 = const()[name = string("op_433_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_433_end_mask_0 = const()[name = string("op_433_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_433_squeeze_mask_0 = const()[name = string("op_433_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_433_cast_fp16 = slice_by_index(begin = var_433_begin_0, end = var_433_end_0, end_mask = var_433_end_mask_0, squeeze_mask = var_433_squeeze_mask_0, x = var_425_cast_fp16)[name = string("op_433_cast_fp16")];
            tensor<fp16, [2, 768]> candidate_rows_to_fp16 = const()[name = string("candidate_rows_to_fp16"), val = tensor<fp16, [2, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14176768)))];
            tensor<fp16, [2]> candidate_scores_bias_0_to_fp16 = const()[name = string("candidate_scores_bias_0_to_fp16"), val = tensor<fp16, [2]>([0x0p+0, 0x0p+0])];
            tensor<fp16, [1, 2]> candidate_scores_cast_fp16 = linear(bias = candidate_scores_bias_0_to_fp16, weight = candidate_rows_to_fp16, x = var_433_cast_fp16)[name = string("candidate_scores_cast_fp16")];
            int32 var_452 = const()[name = string("op_452"), val = int32(1)];
            tensor<fp16, [1, 2]> var_454_cast_fp16 = softmax(axis = var_452, x = candidate_scores_cast_fp16)[name = string("op_454_cast_fp16")];
            tensor<int32, [2]> assistant_probability_begin_0 = const()[name = string("assistant_probability_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [2]> assistant_probability_end_0 = const()[name = string("assistant_probability_end_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<bool, [2]> assistant_probability_end_mask_0 = const()[name = string("assistant_probability_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<bool, [2]> assistant_probability_squeeze_mask_0 = const()[name = string("assistant_probability_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [1]> assistant_probability_cast_fp16 = slice_by_index(begin = assistant_probability_begin_0, end = assistant_probability_end_0, end_mask = assistant_probability_end_mask_0, squeeze_mask = assistant_probability_squeeze_mask_0, x = var_454_cast_fp16)[name = string("assistant_probability_cast_fp16")];
            string assistant_random_to_fp16_dtype_0 = const()[name = string("assistant_random_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1]> assistant_random_to_fp16 = cast(dtype = assistant_random_to_fp16_dtype_0, x = assistant_random)[name = string("cast_384")];
            tensor<bool, [1]> should_continue_cast_fp16 = less_equal(x = assistant_random_to_fp16, y = assistant_probability_cast_fp16)[name = string("should_continue_cast_fp16")];
            tensor<int32, [2]> var_469 = const()[name = string("op_469"), val = tensor<int32, [2]>([1, 1])];
            tensor<bool, [1, 1]> var_470 = reshape(shape = var_469, x = should_continue_cast_fp16)[name = string("op_470")];
            tensor<int32, [2]> var_470_after_broadcast_reps_0 = const()[name = string("op_470_after_broadcast_reps_0"), val = tensor<int32, [2]>([1, 768])];
            tensor<bool, [1, 768]> var_470_after_broadcast = tile(reps = var_470_after_broadcast_reps_0, x = var_470)[name = string("op_470_after_broadcast")];
            tensor<fp16, [1, 768]> var_474_to_fp16 = const()[name = string("op_474_to_fp16"), val = tensor<fp16, [1, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14179904)))];
            tensor<fp16, [1, 768]> var_478_to_fp16 = const()[name = string("op_478_to_fp16"), val = tensor<fp16, [1, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14181504)))];
            tensor<fp16, [1, 768]> next_text_embed_cast_fp16 = select(a = var_474_to_fp16, b = var_478_to_fp16, cond = var_470_after_broadcast)[name = string("next_text_embed_cast_fp16")];
            tensor<int32, [3]> var_487 = const()[name = string("op_487"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 768]> input_19_cast_fp16 = reshape(shape = var_487, x = next_text_embed_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<int32, [1]> input_21_axes_0 = const()[name = string("input_21_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_490_to_fp16 = const()[name = string("op_490_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = block_ln_1_bias_to_fp16, epsilon = var_490_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_19_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 1, 2304]> linear_4_cast_fp16 = linear(bias = block_attn_c_attn_bias_to_fp16, weight = block_attn_c_attn_weight_to_fp16, x = input_21_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<int32, [3]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [3]>([768, 768, 768])];
            int32 var_501_axis_0 = const()[name = string("op_501_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 768]> var_501_cast_fp16_0, tensor<fp16, [1, 1, 768]> var_501_cast_fp16_1, tensor<fp16, [1, 1, 768]> var_501_cast_fp16_2 = split(axis = var_501_axis_0, split_sizes = tile_1, x = linear_4_cast_fp16)[name = string("op_501_cast_fp16")];
            tensor<int32, [4]> var_509 = const()[name = string("op_509"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> query_11_cast_fp16 = reshape(shape = var_509, x = var_501_cast_fp16_0)[name = string("query_11_cast_fp16")];
            tensor<int32, [4]> var_515 = const()[name = string("op_515"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> key_11_cast_fp16 = reshape(shape = var_515, x = var_501_cast_fp16_1)[name = string("key_11_cast_fp16")];
            tensor<int32, [4]> var_521 = const()[name = string("op_521"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> value_9_cast_fp16 = reshape(shape = var_521, x = var_501_cast_fp16_2)[name = string("value_9_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_cos_1_to_fp16 = const()[name = string("rope_cos_1_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14183104)))];
            tensor<fp16, [1, 1, 12, 64]> var_523_cast_fp16 = mul(x = query_11_cast_fp16, y = rope_cos_1_to_fp16)[name = string("op_523_cast_fp16")];
            tensor<int32, [4]> var_528_begin_0 = const()[name = string("op_528_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_528_end_0 = const()[name = string("op_528_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_528_end_mask_0 = const()[name = string("op_528_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_528_stride_0 = const()[name = string("op_528_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_528_cast_fp16 = slice_by_index(begin = var_528_begin_0, end = var_528_end_0, end_mask = var_528_end_mask_0, stride = var_528_stride_0, x = query_11_cast_fp16)[name = string("op_528_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_529_cast_fp16 = mul(x = var_528_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_529_cast_fp16")];
            tensor<int32, [4]> var_534_begin_0 = const()[name = string("op_534_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_534_end_0 = const()[name = string("op_534_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_534_end_mask_0 = const()[name = string("op_534_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_534_stride_0 = const()[name = string("op_534_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_534_cast_fp16 = slice_by_index(begin = var_534_begin_0, end = var_534_end_0, end_mask = var_534_end_mask_0, stride = var_534_stride_0, x = query_11_cast_fp16)[name = string("op_534_cast_fp16")];
            int32 var_537_axis_0 = const()[name = string("op_537_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_537_cast_fp16 = stack(axis = var_537_axis_0, values = (var_529_cast_fp16, var_534_cast_fp16))[name = string("op_537_cast_fp16")];
            tensor<int32, [4]> shape_4 = const()[name = string("shape_4"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_538_cast_fp16 = reshape(shape = shape_4, x = var_537_cast_fp16)[name = string("op_538_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_sin_1_to_fp16 = const()[name = string("rope_sin_1_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14183296)))];
            tensor<fp16, [1, 1, 12, 64]> var_539_cast_fp16 = mul(x = var_538_cast_fp16, y = rope_sin_1_to_fp16)[name = string("op_539_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> query_13_cast_fp16 = add(x = var_523_cast_fp16, y = var_539_cast_fp16)[name = string("query_13_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_542_cast_fp16 = mul(x = key_11_cast_fp16, y = rope_cos_1_to_fp16)[name = string("op_542_cast_fp16")];
            tensor<int32, [4]> var_547_begin_0 = const()[name = string("op_547_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_547_end_0 = const()[name = string("op_547_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_547_end_mask_0 = const()[name = string("op_547_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_547_stride_0 = const()[name = string("op_547_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_547_cast_fp16 = slice_by_index(begin = var_547_begin_0, end = var_547_end_0, end_mask = var_547_end_mask_0, stride = var_547_stride_0, x = key_11_cast_fp16)[name = string("op_547_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_548_cast_fp16 = mul(x = var_547_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_548_cast_fp16")];
            tensor<int32, [4]> var_553_begin_0 = const()[name = string("op_553_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_553_end_0 = const()[name = string("op_553_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_553_end_mask_0 = const()[name = string("op_553_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_553_stride_0 = const()[name = string("op_553_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_553_cast_fp16 = slice_by_index(begin = var_553_begin_0, end = var_553_end_0, end_mask = var_553_end_mask_0, stride = var_553_stride_0, x = key_11_cast_fp16)[name = string("op_553_cast_fp16")];
            int32 var_556_axis_0 = const()[name = string("op_556_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_556_cast_fp16 = stack(axis = var_556_axis_0, values = (var_548_cast_fp16, var_553_cast_fp16))[name = string("op_556_cast_fp16")];
            tensor<int32, [4]> shape_5 = const()[name = string("shape_5"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_557_cast_fp16 = reshape(shape = shape_5, x = var_556_cast_fp16)[name = string("op_557_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_558_cast_fp16 = mul(x = var_557_cast_fp16, y = rope_sin_1_to_fp16)[name = string("op_558_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> key_13_cast_fp16 = add(x = var_542_cast_fp16, y = var_558_cast_fp16)[name = string("key_13_cast_fp16")];
            tensor<int32, [4]> var_565 = const()[name = string("op_565"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_577 = const()[name = string("op_577"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 var_580 = const()[name = string("op_580"), val = int32(2)];
            bool key_17_interleave_0 = const()[name = string("key_17_interleave_0"), val = bool(false)];
            tensor<int32, [4]> transpose_4_perm_0 = const()[name = string("transpose_4_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 64]> transpose_4 = transpose(perm = transpose_4_perm_0, x = key_13_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 12, 2, 64]> key_17_cast_fp16 = concat(axis = var_580, interleave = key_17_interleave_0, values = (transpose_2, transpose_4))[name = string("key_17_cast_fp16")];
            int32 var_583 = const()[name = string("op_583"), val = int32(2)];
            bool value_13_interleave_0 = const()[name = string("value_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> value_11_cast_fp16 = transpose(perm = var_577, x = value_9_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 12, 2, 64]> value_13_cast_fp16 = concat(axis = var_583, interleave = value_13_interleave_0, values = (value_5_cast_fp16, value_11_cast_fp16))[name = string("value_13_cast_fp16")];
            bool var_588_transpose_x_1 = const()[name = string("op_588_transpose_x_1"), val = bool(false)];
            bool var_588_transpose_y_1 = const()[name = string("op_588_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 12, 1, 64]> query_15_cast_fp16 = transpose(perm = var_565, x = query_13_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 12, 1, 2]> var_588_cast_fp16 = matmul(transpose_x = var_588_transpose_x_1, transpose_y = var_588_transpose_y_1, x = query_15_cast_fp16, y = key_17_cast_fp16)[name = string("op_588_cast_fp16")];
            fp16 var_589_to_fp16 = const()[name = string("op_589_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 1, 2]> scores_3_cast_fp16 = mul(x = var_588_cast_fp16, y = var_589_to_fp16)[name = string("scores_3_cast_fp16")];
            int32 var_591 = const()[name = string("op_591"), val = int32(-1)];
            tensor<fp16, [1, 12, 1, 2]> probabilities_3_cast_fp16 = softmax(axis = var_591, x = scores_3_cast_fp16)[name = string("probabilities_3_cast_fp16")];
            bool attention_3_transpose_x_0 = const()[name = string("attention_3_transpose_x_0"), val = bool(false)];
            bool attention_3_transpose_y_0 = const()[name = string("attention_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> attention_3_cast_fp16 = matmul(transpose_x = attention_3_transpose_x_0, transpose_y = attention_3_transpose_y_0, x = probabilities_3_cast_fp16, y = value_13_cast_fp16)[name = string("attention_3_cast_fp16")];
            tensor<int32, [4]> var_599 = const()[name = string("op_599"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_604 = const()[name = string("op_604"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 12, 64]> var_600_cast_fp16 = transpose(perm = var_599, x = attention_3_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 1, 768]> input_23_cast_fp16 = reshape(shape = var_604, x = var_600_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_5_cast_fp16 = linear(bias = block_attn_c_proj_bias_to_fp16, weight = block_attn_c_proj_weight_to_fp16, x = input_23_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_5_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<int32, [1]> input_27_axes_0 = const()[name = string("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_612_to_fp16 = const()[name = string("op_612_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = block_ln_2_bias_to_fp16, epsilon = var_612_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_25_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [1, 1, 3072]> linear_6_cast_fp16 = linear(bias = block_mlp_fc_in_bias_to_fp16, weight = block_mlp_fc_in_weight_to_fp16, x = input_27_cast_fp16)[name = string("linear_6_cast_fp16")];
            string input_31_mode_0 = const()[name = string("input_31_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1, 3072]> input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = linear_6_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_7_cast_fp16 = linear(bias = block_mlp_fc_out_bias_to_fp16, weight = block_mlp_fc_out_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_35_cast_fp16 = add(x = input_25_cast_fp16, y = linear_7_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<int32, [1]> var_651_axes_0 = const()[name = string("op_651_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_646_to_fp16 = const()[name = string("op_646_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> var_651_cast_fp16 = layer_norm(axes = var_651_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_646_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_35_cast_fp16)[name = string("op_651_cast_fp16")];
            tensor<int32, [3]> var_659_begin_0 = const()[name = string("op_659_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_659_end_0 = const()[name = string("op_659_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_659_end_mask_0 = const()[name = string("op_659_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_659_squeeze_mask_0 = const()[name = string("op_659_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_659_cast_fp16 = slice_by_index(begin = var_659_begin_0, end = var_659_end_0, end_mask = var_659_end_mask_0, squeeze_mask = var_659_squeeze_mask_0, x = var_651_cast_fp16)[name = string("op_659_cast_fp16")];
            tensor<fp16, [1024, 768]> audio_embeddings_0_weight_to_fp16 = const()[name = string("audio_embeddings_0_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14183488)))];
            tensor<fp16, [1024]> linear_8_bias_0_to_fp16 = const()[name = string("linear_8_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15756416)))];
            tensor<fp16, [1, 1024]> linear_8_cast_fp16 = linear(bias = linear_8_bias_0_to_fp16, weight = audio_embeddings_0_weight_to_fp16, x = var_659_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [3]> var_674_begin_0 = const()[name = string("op_674_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_674_end_0 = const()[name = string("op_674_end_0"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<bool, [3]> var_674_end_mask_0 = const()[name = string("op_674_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_674_squeeze_mask_0 = const()[name = string("op_674_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<int32, [1, 1024]> var_674 = slice_by_index(begin = var_674_begin_0, end = var_674_end_0, end_mask = var_674_end_mask_0, squeeze_mask = var_674_squeeze_mask_0, x = repetition_seen)[name = string("op_674")];
            tensor<int32, [2]> random_u_1_begin_0 = const()[name = string("random_u_1_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [2]> random_u_1_end_0 = const()[name = string("random_u_1_end_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<bool, [2]> random_u_1_end_mask_0 = const()[name = string("random_u_1_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<bool, [2]> random_u_1_squeeze_mask_0 = const()[name = string("random_u_1_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
            string audio_random_to_fp16_dtype_0 = const()[name = string("audio_random_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 16]> audio_random_to_fp16 = cast(dtype = audio_random_to_fp16_dtype_0, x = audio_random)[name = string("cast_383")];
            tensor<fp16, [1]> random_u_1_cast_fp16 = slice_by_index(begin = random_u_1_begin_0, end = random_u_1_end_0, end_mask = random_u_1_end_mask_0, squeeze_mask = random_u_1_squeeze_mask_0, x = audio_random_to_fp16)[name = string("random_u_1_cast_fp16")];
            fp16 var_688_promoted_to_fp16 = const()[name = string("op_688_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [1, 1024]> var_689_cast_fp16 = less(x = linear_8_cast_fp16, y = var_688_promoted_to_fp16)[name = string("op_689_cast_fp16")];
            fp16 var_690_to_fp16 = const()[name = string("op_690_to_fp16"), val = fp16(0x1.334p+0)];
            tensor<fp16, [1, 1024]> var_691_cast_fp16 = mul(x = linear_8_cast_fp16, y = var_690_to_fp16)[name = string("op_691_cast_fp16")];
            fp16 _inversed_693_y_0_to_fp16 = const()[name = string("_inversed_693_y_0_to_fp16"), val = fp16(0x1.aacp-1)];
            tensor<fp16, [1, 1024]> _inversed_693_cast_fp16 = mul(x = linear_8_cast_fp16, y = _inversed_693_y_0_to_fp16)[name = string("_inversed_693_cast_fp16")];
            tensor<fp16, [1, 1024]> penalized_1_cast_fp16 = select(a = var_691_cast_fp16, b = _inversed_693_cast_fp16, cond = var_689_cast_fp16)[name = string("penalized_1_cast_fp16")];
            string var_699_dtype_0 = const()[name = string("op_699_dtype_0"), val = string("bool")];
            tensor<bool, [1, 1024]> var_699 = cast(dtype = var_699_dtype_0, x = var_674)[name = string("cast_382")];
            tensor<fp16, [1, 1024]> var_700_cast_fp16 = select(a = penalized_1_cast_fp16, b = linear_8_cast_fp16, cond = var_699)[name = string("op_700_cast_fp16")];
            fp16 _inversed_scores_5_y_0_to_fp16 = const()[name = string("_inversed_scores_5_y_0_to_fp16"), val = fp16(0x1.4p+0)];
            tensor<fp16, [1, 1024]> _inversed_scores_5_cast_fp16 = mul(x = var_700_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = string("_inversed_scores_5_cast_fp16")];
            int32 var_703 = const()[name = string("op_703"), val = int32(25)];
            int32 top_scores_1_axis_0 = const()[name = string("top_scores_1_axis_0"), val = int32(1)];
            bool top_scores_1_ascending_0 = const()[name = string("top_scores_1_ascending_0"), val = bool(false)];
            bool top_scores_1_sort_0 = const()[name = string("top_scores_1_sort_0"), val = bool(true)];
            bool top_scores_1_return_indices_0 = const()[name = string("top_scores_1_return_indices_0"), val = bool(true)];
            string top_scores_1_cast_fp16_cast_uint16_output_indices_dtype_0 = const()[name = string("top_scores_1_cast_fp16_cast_uint16_output_indices_dtype_0"), val = string("uint16")];
            tensor<fp16, [1, 25]> top_scores_1_cast_fp16_cast_uint16_0, tensor<uint16, [1, 25]> top_scores_1_cast_fp16_cast_uint16_1 = topk(ascending = top_scores_1_ascending_0, axis = top_scores_1_axis_0, k = var_703, output_indices_dtype = top_scores_1_cast_fp16_cast_uint16_output_indices_dtype_0, return_indices = top_scores_1_return_indices_0, sort = top_scores_1_sort_0, x = _inversed_scores_5_cast_fp16)[name = string("top_scores_1_cast_fp16_cast_uint16")];
            int32 var_709 = const()[name = string("op_709"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_5_cast_fp16 = softmax(axis = var_709, x = top_scores_1_cast_fp16_cast_uint16_0)[name = string("probabilities_5_cast_fp16")];
            int32 var_712 = const()[name = string("op_712"), val = int32(1)];
            bool cumulative_1_exclusive_0 = const()[name = string("cumulative_1_exclusive_0"), val = bool(false)];
            bool cumulative_1_reverse_0 = const()[name = string("cumulative_1_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_1_cast_fp16 = cumsum(axis = var_712, exclusive = cumulative_1_exclusive_0, reverse = cumulative_1_reverse_0, x = probabilities_5_cast_fp16)[name = string("cumulative_1_cast_fp16")];
            tensor<fp16, [1, 25]> var_716_cast_fp16 = sub(x = cumulative_1_cast_fp16, y = probabilities_5_cast_fp16)[name = string("op_716_cast_fp16")];
            fp16 var_717_to_fp16 = const()[name = string("op_717_to_fp16"), val = fp16(0x1.e68p-1)];
            tensor<bool, [1, 25]> keep_1_cast_fp16 = less(x = var_716_cast_fp16, y = var_717_to_fp16)[name = string("keep_1_cast_fp16")];
            fp16 var_725_value_0_to_fp16 = const()[name = string("op_725_value_0_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 25]> var_725_cast_fp16 = fill_like(ref_tensor = top_scores_1_cast_fp16_cast_uint16_0, value = var_725_value_0_to_fp16)[name = string("op_725_cast_fp16")];
            tensor<fp16, [1, 25]> top_scores_3_cast_fp16 = select(a = top_scores_1_cast_fp16_cast_uint16_0, b = var_725_cast_fp16, cond = keep_1_cast_fp16)[name = string("top_scores_3_cast_fp16")];
            int32 var_727 = const()[name = string("op_727"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_7_cast_fp16 = softmax(axis = var_727, x = top_scores_3_cast_fp16)[name = string("probabilities_7_cast_fp16")];
            int32 var_730 = const()[name = string("op_730"), val = int32(1)];
            bool cumulative_3_exclusive_0 = const()[name = string("cumulative_3_exclusive_0"), val = bool(false)];
            bool cumulative_3_reverse_0 = const()[name = string("cumulative_3_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_3_cast_fp16 = cumsum(axis = var_730, exclusive = cumulative_3_exclusive_0, reverse = cumulative_3_reverse_0, x = probabilities_7_cast_fp16)[name = string("cumulative_3_cast_fp16")];
            tensor<int32, [2]> var_735 = const()[name = string("op_735"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_736_cast_fp16 = reshape(shape = var_735, x = random_u_1_cast_fp16)[name = string("op_736_cast_fp16")];
            tensor<bool, [1, 25]> var_737_cast_fp16 = less(x = cumulative_3_cast_fp16, y = var_736_cast_fp16)[name = string("op_737_cast_fp16")];
            string var_742_dtype_0 = const()[name = string("op_742_dtype_0"), val = string("int32")];
            tensor<int32, [1]> selected_1_axes_0 = const()[name = string("selected_1_axes_0"), val = tensor<int32, [1]>([1])];
            bool selected_1_keep_dims_0 = const()[name = string("selected_1_keep_dims_0"), val = bool(false)];
            tensor<int32, [1, 25]> var_742 = cast(dtype = var_742_dtype_0, x = var_737_cast_fp16)[name = string("cast_381")];
            tensor<int32, [1]> selected_1 = reduce_sum(axes = selected_1_axes_0, keep_dims = selected_1_keep_dims_0, x = var_742)[name = string("selected_1")];
            string selected_1_promoted_to_fp16_dtype_0 = const()[name = string("selected_1_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_8_to_fp16 = const()[name = string("const_8_to_fp16"), val = fp16(-inf)];
            fp16 var_749_promoted_to_fp16 = const()[name = string("op_749_promoted_to_fp16"), val = fp16(0x1.8p+4)];
            tensor<fp16, [1]> selected_1_to_fp16 = cast(dtype = selected_1_promoted_to_fp16_dtype_0, x = selected_1)[name = string("cast_380")];
            tensor<fp16, [1]> clip_0_cast_fp16 = clip(alpha = const_8_to_fp16, beta = var_749_promoted_to_fp16, x = selected_1_to_fp16)[name = string("clip_0_cast_fp16")];
            string selected_3_dtype_0 = const()[name = string("selected_3_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_758 = const()[name = string("op_758"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> clip_0_cast_fp16_to_int32 = cast(dtype = selected_3_dtype_0, x = clip_0_cast_fp16)[name = string("cast_379")];
            tensor<int32, [1, 1]> var_759 = reshape(shape = var_758, x = clip_0_cast_fp16_to_int32)[name = string("op_759")];
            int32 var_760 = const()[name = string("op_760"), val = int32(1)];
            bool var_762_validate_indices_0 = const()[name = string("op_762_validate_indices_0"), val = bool(false)];
            string var_759_to_uint16_dtype_0 = const()[name = string("op_759_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 1]> var_759_to_uint16 = cast(dtype = var_759_to_uint16_dtype_0, x = var_759)[name = string("cast_378")];
            tensor<uint16, [1, 1]> var_762_cast_uint16 = gather_along_axis(axis = var_760, indices = var_759_to_uint16, validate_indices = var_762_validate_indices_0, x = top_scores_1_cast_fp16_cast_uint16_1)[name = string("op_762_cast_uint16")];
            string var_762_cast_uint16_to_int32_dtype_0 = const()[name = string("op_762_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [1]> var_764 = const()[name = string("op_764"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_762_cast_uint16_to_int32 = cast(dtype = var_762_cast_uint16_to_int32_dtype_0, x = var_762_cast_uint16)[name = string("cast_377")];
            tensor<int32, [1]> token_1 = reshape(shape = var_764, x = var_762_cast_uint16_to_int32)[name = string("token_1")];
            int32 var_774_axis_0 = const()[name = string("op_774_axis_0"), val = int32(0)];
            int32 var_774_batch_dims_0 = const()[name = string("op_774_batch_dims_0"), val = int32(0)];
            bool var_774_validate_indices_0 = const()[name = string("op_774_validate_indices_0"), val = bool(false)];
            string token_1_to_uint16_dtype_0 = const()[name = string("token_1_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> token_1_to_uint16 = cast(dtype = token_1_to_uint16_dtype_0, x = token_1)[name = string("cast_376")];
            tensor<fp16, [1, 768]> var_774_cast_fp16_cast_uint16 = gather(axis = var_774_axis_0, batch_dims = var_774_batch_dims_0, indices = token_1_to_uint16, validate_indices = var_774_validate_indices_0, x = audio_embeddings_0_weight_to_fp16)[name = string("op_774_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 768]> next_global_embed_3_cast_fp16 = add(x = var_474_to_fp16, y = var_774_cast_fp16_cast_uint16)[name = string("next_global_embed_3_cast_fp16")];
            tensor<int32, [3]> var_784 = const()[name = string("op_784"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 768]> input_39_cast_fp16 = reshape(shape = var_784, x = var_774_cast_fp16_cast_uint16)[name = string("input_39_cast_fp16")];
            tensor<int32, [1]> input_41_axes_0 = const()[name = string("input_41_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_787_to_fp16 = const()[name = string("op_787_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = block_ln_1_bias_to_fp16, epsilon = var_787_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_39_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [1, 1, 2304]> linear_9_cast_fp16 = linear(bias = block_attn_c_attn_bias_to_fp16, weight = block_attn_c_attn_weight_to_fp16, x = input_41_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [3]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [3]>([768, 768, 768])];
            int32 var_798_axis_0 = const()[name = string("op_798_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 768]> var_798_cast_fp16_0, tensor<fp16, [1, 1, 768]> var_798_cast_fp16_1, tensor<fp16, [1, 1, 768]> var_798_cast_fp16_2 = split(axis = var_798_axis_0, split_sizes = tile_2, x = linear_9_cast_fp16)[name = string("op_798_cast_fp16")];
            tensor<int32, [4]> var_806 = const()[name = string("op_806"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> query_19_cast_fp16 = reshape(shape = var_806, x = var_798_cast_fp16_0)[name = string("query_19_cast_fp16")];
            tensor<int32, [4]> var_812 = const()[name = string("op_812"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> key_21_cast_fp16 = reshape(shape = var_812, x = var_798_cast_fp16_1)[name = string("key_21_cast_fp16")];
            tensor<int32, [4]> var_818 = const()[name = string("op_818"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> value_17_cast_fp16 = reshape(shape = var_818, x = var_798_cast_fp16_2)[name = string("value_17_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_cos_2_to_fp16 = const()[name = string("rope_cos_2_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15758528)))];
            tensor<fp16, [1, 1, 12, 64]> var_820_cast_fp16 = mul(x = query_19_cast_fp16, y = rope_cos_2_to_fp16)[name = string("op_820_cast_fp16")];
            tensor<int32, [4]> var_825_begin_0 = const()[name = string("op_825_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_825_end_0 = const()[name = string("op_825_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_825_end_mask_0 = const()[name = string("op_825_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_825_stride_0 = const()[name = string("op_825_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_825_cast_fp16 = slice_by_index(begin = var_825_begin_0, end = var_825_end_0, end_mask = var_825_end_mask_0, stride = var_825_stride_0, x = query_19_cast_fp16)[name = string("op_825_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_826_cast_fp16 = mul(x = var_825_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_826_cast_fp16")];
            tensor<int32, [4]> var_831_begin_0 = const()[name = string("op_831_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_831_end_0 = const()[name = string("op_831_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_831_end_mask_0 = const()[name = string("op_831_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_831_stride_0 = const()[name = string("op_831_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_831_cast_fp16 = slice_by_index(begin = var_831_begin_0, end = var_831_end_0, end_mask = var_831_end_mask_0, stride = var_831_stride_0, x = query_19_cast_fp16)[name = string("op_831_cast_fp16")];
            int32 var_834_axis_0 = const()[name = string("op_834_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_834_cast_fp16 = stack(axis = var_834_axis_0, values = (var_826_cast_fp16, var_831_cast_fp16))[name = string("op_834_cast_fp16")];
            tensor<int32, [4]> shape_7 = const()[name = string("shape_7"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_835_cast_fp16 = reshape(shape = shape_7, x = var_834_cast_fp16)[name = string("op_835_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_sin_2_to_fp16 = const()[name = string("rope_sin_2_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15758720)))];
            tensor<fp16, [1, 1, 12, 64]> var_836_cast_fp16 = mul(x = var_835_cast_fp16, y = rope_sin_2_to_fp16)[name = string("op_836_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> query_21_cast_fp16 = add(x = var_820_cast_fp16, y = var_836_cast_fp16)[name = string("query_21_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_839_cast_fp16 = mul(x = key_21_cast_fp16, y = rope_cos_2_to_fp16)[name = string("op_839_cast_fp16")];
            tensor<int32, [4]> var_844_begin_0 = const()[name = string("op_844_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_844_end_0 = const()[name = string("op_844_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_844_end_mask_0 = const()[name = string("op_844_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_844_stride_0 = const()[name = string("op_844_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_844_cast_fp16 = slice_by_index(begin = var_844_begin_0, end = var_844_end_0, end_mask = var_844_end_mask_0, stride = var_844_stride_0, x = key_21_cast_fp16)[name = string("op_844_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_845_cast_fp16 = mul(x = var_844_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_845_cast_fp16")];
            tensor<int32, [4]> var_850_begin_0 = const()[name = string("op_850_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_850_end_0 = const()[name = string("op_850_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_850_end_mask_0 = const()[name = string("op_850_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_850_stride_0 = const()[name = string("op_850_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_850_cast_fp16 = slice_by_index(begin = var_850_begin_0, end = var_850_end_0, end_mask = var_850_end_mask_0, stride = var_850_stride_0, x = key_21_cast_fp16)[name = string("op_850_cast_fp16")];
            int32 var_853_axis_0 = const()[name = string("op_853_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_853_cast_fp16 = stack(axis = var_853_axis_0, values = (var_845_cast_fp16, var_850_cast_fp16))[name = string("op_853_cast_fp16")];
            tensor<int32, [4]> shape_8 = const()[name = string("shape_8"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_854_cast_fp16 = reshape(shape = shape_8, x = var_853_cast_fp16)[name = string("op_854_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_855_cast_fp16 = mul(x = var_854_cast_fp16, y = rope_sin_2_to_fp16)[name = string("op_855_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> key_23_cast_fp16 = add(x = var_839_cast_fp16, y = var_855_cast_fp16)[name = string("key_23_cast_fp16")];
            tensor<int32, [4]> var_862 = const()[name = string("op_862"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_868 = const()[name = string("op_868"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_874 = const()[name = string("op_874"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 var_877 = const()[name = string("op_877"), val = int32(2)];
            bool key_27_interleave_0 = const()[name = string("key_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> key_25_cast_fp16 = transpose(perm = var_868, x = key_23_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 12, 3, 64]> key_27_cast_fp16 = concat(axis = var_877, interleave = key_27_interleave_0, values = (key_17_cast_fp16, key_25_cast_fp16))[name = string("key_27_cast_fp16")];
            int32 var_880 = const()[name = string("op_880"), val = int32(2)];
            bool value_21_interleave_0 = const()[name = string("value_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> value_19_cast_fp16 = transpose(perm = var_874, x = value_17_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 12, 3, 64]> value_21_cast_fp16 = concat(axis = var_880, interleave = value_21_interleave_0, values = (value_13_cast_fp16, value_19_cast_fp16))[name = string("value_21_cast_fp16")];
            bool var_885_transpose_x_1 = const()[name = string("op_885_transpose_x_1"), val = bool(false)];
            bool var_885_transpose_y_1 = const()[name = string("op_885_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 12, 1, 64]> query_23_cast_fp16 = transpose(perm = var_862, x = query_21_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 12, 1, 3]> var_885_cast_fp16 = matmul(transpose_x = var_885_transpose_x_1, transpose_y = var_885_transpose_y_1, x = query_23_cast_fp16, y = key_27_cast_fp16)[name = string("op_885_cast_fp16")];
            fp16 var_886_to_fp16 = const()[name = string("op_886_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 1, 3]> scores_7_cast_fp16 = mul(x = var_885_cast_fp16, y = var_886_to_fp16)[name = string("scores_7_cast_fp16")];
            int32 var_888 = const()[name = string("op_888"), val = int32(-1)];
            tensor<fp16, [1, 12, 1, 3]> probabilities_9_cast_fp16 = softmax(axis = var_888, x = scores_7_cast_fp16)[name = string("probabilities_9_cast_fp16")];
            bool attention_5_transpose_x_0 = const()[name = string("attention_5_transpose_x_0"), val = bool(false)];
            bool attention_5_transpose_y_0 = const()[name = string("attention_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> attention_5_cast_fp16 = matmul(transpose_x = attention_5_transpose_x_0, transpose_y = attention_5_transpose_y_0, x = probabilities_9_cast_fp16, y = value_21_cast_fp16)[name = string("attention_5_cast_fp16")];
            tensor<int32, [4]> var_896 = const()[name = string("op_896"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_901 = const()[name = string("op_901"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 12, 64]> var_897_cast_fp16 = transpose(perm = var_896, x = attention_5_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 1, 768]> input_43_cast_fp16 = reshape(shape = var_901, x = var_897_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_10_cast_fp16 = linear(bias = block_attn_c_proj_bias_to_fp16, weight = block_attn_c_proj_weight_to_fp16, x = input_43_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_10_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<int32, [1]> input_47_axes_0 = const()[name = string("input_47_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_909_to_fp16 = const()[name = string("op_909_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = block_ln_2_bias_to_fp16, epsilon = var_909_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_45_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [1, 1, 3072]> linear_11_cast_fp16 = linear(bias = block_mlp_fc_in_bias_to_fp16, weight = block_mlp_fc_in_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_11_cast_fp16")];
            string input_51_mode_0 = const()[name = string("input_51_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1, 3072]> input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = linear_11_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_12_cast_fp16 = linear(bias = block_mlp_fc_out_bias_to_fp16, weight = block_mlp_fc_out_weight_to_fp16, x = input_51_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_55_cast_fp16 = add(x = input_45_cast_fp16, y = linear_12_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<int32, [1]> var_948_axes_0 = const()[name = string("op_948_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_943_to_fp16 = const()[name = string("op_943_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> var_948_cast_fp16 = layer_norm(axes = var_948_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_943_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_55_cast_fp16)[name = string("op_948_cast_fp16")];
            tensor<int32, [3]> var_956_begin_0 = const()[name = string("op_956_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_956_end_0 = const()[name = string("op_956_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_956_end_mask_0 = const()[name = string("op_956_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_956_squeeze_mask_0 = const()[name = string("op_956_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_956_cast_fp16 = slice_by_index(begin = var_956_begin_0, end = var_956_end_0, end_mask = var_956_end_mask_0, squeeze_mask = var_956_squeeze_mask_0, x = var_948_cast_fp16)[name = string("op_956_cast_fp16")];
            tensor<fp16, [1024, 768]> audio_embeddings_1_weight_to_fp16 = const()[name = string("audio_embeddings_1_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15758912)))];
            tensor<fp16, [1, 1024]> linear_13_cast_fp16 = linear(bias = linear_8_bias_0_to_fp16, weight = audio_embeddings_1_weight_to_fp16, x = var_956_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<int32, [3]> var_971_begin_0 = const()[name = string("op_971_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_971_end_0 = const()[name = string("op_971_end_0"), val = tensor<int32, [3]>([1, 2, 1024])];
            tensor<bool, [3]> var_971_end_mask_0 = const()[name = string("op_971_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_971_squeeze_mask_0 = const()[name = string("op_971_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<int32, [1, 1024]> var_971 = slice_by_index(begin = var_971_begin_0, end = var_971_end_0, end_mask = var_971_end_mask_0, squeeze_mask = var_971_squeeze_mask_0, x = repetition_seen)[name = string("op_971")];
            tensor<int32, [2]> random_u_3_begin_0 = const()[name = string("random_u_3_begin_0"), val = tensor<int32, [2]>([0, 1])];
            tensor<int32, [2]> random_u_3_end_0 = const()[name = string("random_u_3_end_0"), val = tensor<int32, [2]>([1, 2])];
            tensor<bool, [2]> random_u_3_end_mask_0 = const()[name = string("random_u_3_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<bool, [2]> random_u_3_squeeze_mask_0 = const()[name = string("random_u_3_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [1]> random_u_3_cast_fp16 = slice_by_index(begin = random_u_3_begin_0, end = random_u_3_end_0, end_mask = random_u_3_end_mask_0, squeeze_mask = random_u_3_squeeze_mask_0, x = audio_random_to_fp16)[name = string("random_u_3_cast_fp16")];
            fp16 var_985_promoted_to_fp16 = const()[name = string("op_985_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [1, 1024]> var_986_cast_fp16 = less(x = linear_13_cast_fp16, y = var_985_promoted_to_fp16)[name = string("op_986_cast_fp16")];
            fp16 var_987_to_fp16 = const()[name = string("op_987_to_fp16"), val = fp16(0x1.334p+0)];
            tensor<fp16, [1, 1024]> var_988_cast_fp16 = mul(x = linear_13_cast_fp16, y = var_987_to_fp16)[name = string("op_988_cast_fp16")];
            fp16 _inversed_990_y_0_to_fp16 = const()[name = string("_inversed_990_y_0_to_fp16"), val = fp16(0x1.aacp-1)];
            tensor<fp16, [1, 1024]> _inversed_990_cast_fp16 = mul(x = linear_13_cast_fp16, y = _inversed_990_y_0_to_fp16)[name = string("_inversed_990_cast_fp16")];
            tensor<fp16, [1, 1024]> penalized_3_cast_fp16 = select(a = var_988_cast_fp16, b = _inversed_990_cast_fp16, cond = var_986_cast_fp16)[name = string("penalized_3_cast_fp16")];
            string var_996_dtype_0 = const()[name = string("op_996_dtype_0"), val = string("bool")];
            tensor<bool, [1, 1024]> var_996 = cast(dtype = var_996_dtype_0, x = var_971)[name = string("cast_375")];
            tensor<fp16, [1, 1024]> var_997_cast_fp16 = select(a = penalized_3_cast_fp16, b = linear_13_cast_fp16, cond = var_996)[name = string("op_997_cast_fp16")];
            fp16 _inversed_scores_9_y_0_to_fp16 = const()[name = string("_inversed_scores_9_y_0_to_fp16"), val = fp16(0x1.4p+0)];
            tensor<fp16, [1, 1024]> _inversed_scores_9_cast_fp16 = mul(x = var_997_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = string("_inversed_scores_9_cast_fp16")];
            int32 var_1000 = const()[name = string("op_1000"), val = int32(25)];
            int32 top_scores_5_axis_0 = const()[name = string("top_scores_5_axis_0"), val = int32(1)];
            bool top_scores_5_ascending_0 = const()[name = string("top_scores_5_ascending_0"), val = bool(false)];
            bool top_scores_5_sort_0 = const()[name = string("top_scores_5_sort_0"), val = bool(true)];
            bool top_scores_5_return_indices_0 = const()[name = string("top_scores_5_return_indices_0"), val = bool(true)];
            string top_scores_5_cast_fp16_cast_uint16_output_indices_dtype_0 = const()[name = string("top_scores_5_cast_fp16_cast_uint16_output_indices_dtype_0"), val = string("uint16")];
            tensor<fp16, [1, 25]> top_scores_5_cast_fp16_cast_uint16_0, tensor<uint16, [1, 25]> top_scores_5_cast_fp16_cast_uint16_1 = topk(ascending = top_scores_5_ascending_0, axis = top_scores_5_axis_0, k = var_1000, output_indices_dtype = top_scores_5_cast_fp16_cast_uint16_output_indices_dtype_0, return_indices = top_scores_5_return_indices_0, sort = top_scores_5_sort_0, x = _inversed_scores_9_cast_fp16)[name = string("top_scores_5_cast_fp16_cast_uint16")];
            int32 var_1006 = const()[name = string("op_1006"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_11_cast_fp16 = softmax(axis = var_1006, x = top_scores_5_cast_fp16_cast_uint16_0)[name = string("probabilities_11_cast_fp16")];
            int32 var_1009 = const()[name = string("op_1009"), val = int32(1)];
            bool cumulative_5_exclusive_0 = const()[name = string("cumulative_5_exclusive_0"), val = bool(false)];
            bool cumulative_5_reverse_0 = const()[name = string("cumulative_5_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_5_cast_fp16 = cumsum(axis = var_1009, exclusive = cumulative_5_exclusive_0, reverse = cumulative_5_reverse_0, x = probabilities_11_cast_fp16)[name = string("cumulative_5_cast_fp16")];
            tensor<fp16, [1, 25]> var_1013_cast_fp16 = sub(x = cumulative_5_cast_fp16, y = probabilities_11_cast_fp16)[name = string("op_1013_cast_fp16")];
            fp16 var_1014_to_fp16 = const()[name = string("op_1014_to_fp16"), val = fp16(0x1.e68p-1)];
            tensor<bool, [1, 25]> keep_3_cast_fp16 = less(x = var_1013_cast_fp16, y = var_1014_to_fp16)[name = string("keep_3_cast_fp16")];
            fp16 var_1022_value_0_to_fp16 = const()[name = string("op_1022_value_0_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 25]> var_1022_cast_fp16 = fill_like(ref_tensor = top_scores_5_cast_fp16_cast_uint16_0, value = var_1022_value_0_to_fp16)[name = string("op_1022_cast_fp16")];
            tensor<fp16, [1, 25]> top_scores_7_cast_fp16 = select(a = top_scores_5_cast_fp16_cast_uint16_0, b = var_1022_cast_fp16, cond = keep_3_cast_fp16)[name = string("top_scores_7_cast_fp16")];
            int32 var_1024 = const()[name = string("op_1024"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_13_cast_fp16 = softmax(axis = var_1024, x = top_scores_7_cast_fp16)[name = string("probabilities_13_cast_fp16")];
            int32 var_1027 = const()[name = string("op_1027"), val = int32(1)];
            bool cumulative_7_exclusive_0 = const()[name = string("cumulative_7_exclusive_0"), val = bool(false)];
            bool cumulative_7_reverse_0 = const()[name = string("cumulative_7_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_7_cast_fp16 = cumsum(axis = var_1027, exclusive = cumulative_7_exclusive_0, reverse = cumulative_7_reverse_0, x = probabilities_13_cast_fp16)[name = string("cumulative_7_cast_fp16")];
            tensor<int32, [2]> var_1032 = const()[name = string("op_1032"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_1033_cast_fp16 = reshape(shape = var_1032, x = random_u_3_cast_fp16)[name = string("op_1033_cast_fp16")];
            tensor<bool, [1, 25]> var_1034_cast_fp16 = less(x = cumulative_7_cast_fp16, y = var_1033_cast_fp16)[name = string("op_1034_cast_fp16")];
            string var_1039_dtype_0 = const()[name = string("op_1039_dtype_0"), val = string("int32")];
            tensor<int32, [1]> selected_5_axes_0 = const()[name = string("selected_5_axes_0"), val = tensor<int32, [1]>([1])];
            bool selected_5_keep_dims_0 = const()[name = string("selected_5_keep_dims_0"), val = bool(false)];
            tensor<int32, [1, 25]> var_1039 = cast(dtype = var_1039_dtype_0, x = var_1034_cast_fp16)[name = string("cast_374")];
            tensor<int32, [1]> selected_5 = reduce_sum(axes = selected_5_axes_0, keep_dims = selected_5_keep_dims_0, x = var_1039)[name = string("selected_5")];
            string selected_5_promoted_to_fp16_dtype_0 = const()[name = string("selected_5_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_13_to_fp16 = const()[name = string("const_13_to_fp16"), val = fp16(-inf)];
            fp16 var_1046_promoted_to_fp16 = const()[name = string("op_1046_promoted_to_fp16"), val = fp16(0x1.8p+4)];
            tensor<fp16, [1]> selected_5_to_fp16 = cast(dtype = selected_5_promoted_to_fp16_dtype_0, x = selected_5)[name = string("cast_373")];
            tensor<fp16, [1]> clip_1_cast_fp16 = clip(alpha = const_13_to_fp16, beta = var_1046_promoted_to_fp16, x = selected_5_to_fp16)[name = string("clip_1_cast_fp16")];
            string selected_7_dtype_0 = const()[name = string("selected_7_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_1055 = const()[name = string("op_1055"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> clip_1_cast_fp16_to_int32 = cast(dtype = selected_7_dtype_0, x = clip_1_cast_fp16)[name = string("cast_372")];
            tensor<int32, [1, 1]> var_1056 = reshape(shape = var_1055, x = clip_1_cast_fp16_to_int32)[name = string("op_1056")];
            int32 var_1057 = const()[name = string("op_1057"), val = int32(1)];
            bool var_1059_validate_indices_0 = const()[name = string("op_1059_validate_indices_0"), val = bool(false)];
            string var_1056_to_uint16_dtype_0 = const()[name = string("op_1056_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 1]> var_1056_to_uint16 = cast(dtype = var_1056_to_uint16_dtype_0, x = var_1056)[name = string("cast_371")];
            tensor<uint16, [1, 1]> var_1059_cast_uint16 = gather_along_axis(axis = var_1057, indices = var_1056_to_uint16, validate_indices = var_1059_validate_indices_0, x = top_scores_5_cast_fp16_cast_uint16_1)[name = string("op_1059_cast_uint16")];
            string var_1059_cast_uint16_to_int32_dtype_0 = const()[name = string("op_1059_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [1]> var_1061 = const()[name = string("op_1061"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_1059_cast_uint16_to_int32 = cast(dtype = var_1059_cast_uint16_to_int32_dtype_0, x = var_1059_cast_uint16)[name = string("cast_370")];
            tensor<int32, [1]> token_3 = reshape(shape = var_1061, x = var_1059_cast_uint16_to_int32)[name = string("token_3")];
            int32 var_1071_axis_0 = const()[name = string("op_1071_axis_0"), val = int32(0)];
            int32 var_1071_batch_dims_0 = const()[name = string("op_1071_batch_dims_0"), val = int32(0)];
            bool var_1071_validate_indices_0 = const()[name = string("op_1071_validate_indices_0"), val = bool(false)];
            string token_3_to_uint16_dtype_0 = const()[name = string("token_3_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> token_3_to_uint16 = cast(dtype = token_3_to_uint16_dtype_0, x = token_3)[name = string("cast_369")];
            tensor<fp16, [1, 768]> var_1071_cast_fp16_cast_uint16 = gather(axis = var_1071_axis_0, batch_dims = var_1071_batch_dims_0, indices = token_3_to_uint16, validate_indices = var_1071_validate_indices_0, x = audio_embeddings_1_weight_to_fp16)[name = string("op_1071_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 768]> next_global_embed_5_cast_fp16 = add(x = next_global_embed_3_cast_fp16, y = var_1071_cast_fp16_cast_uint16)[name = string("next_global_embed_5_cast_fp16")];
            tensor<int32, [3]> var_1081 = const()[name = string("op_1081"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 768]> input_59_cast_fp16 = reshape(shape = var_1081, x = var_1071_cast_fp16_cast_uint16)[name = string("input_59_cast_fp16")];
            tensor<int32, [1]> input_61_axes_0 = const()[name = string("input_61_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_1084_to_fp16 = const()[name = string("op_1084_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_61_cast_fp16 = layer_norm(axes = input_61_axes_0, beta = block_ln_1_bias_to_fp16, epsilon = var_1084_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_59_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [1, 1, 2304]> linear_14_cast_fp16 = linear(bias = block_attn_c_attn_bias_to_fp16, weight = block_attn_c_attn_weight_to_fp16, x = input_61_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [3]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [3]>([768, 768, 768])];
            int32 var_1095_axis_0 = const()[name = string("op_1095_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 768]> var_1095_cast_fp16_0, tensor<fp16, [1, 1, 768]> var_1095_cast_fp16_1, tensor<fp16, [1, 1, 768]> var_1095_cast_fp16_2 = split(axis = var_1095_axis_0, split_sizes = tile_3, x = linear_14_cast_fp16)[name = string("op_1095_cast_fp16")];
            tensor<int32, [4]> var_1103 = const()[name = string("op_1103"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> query_27_cast_fp16 = reshape(shape = var_1103, x = var_1095_cast_fp16_0)[name = string("query_27_cast_fp16")];
            tensor<int32, [4]> var_1109 = const()[name = string("op_1109"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> key_31_cast_fp16 = reshape(shape = var_1109, x = var_1095_cast_fp16_1)[name = string("key_31_cast_fp16")];
            tensor<int32, [4]> var_1115 = const()[name = string("op_1115"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> value_25_cast_fp16 = reshape(shape = var_1115, x = var_1095_cast_fp16_2)[name = string("value_25_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_cos_3_to_fp16 = const()[name = string("rope_cos_3_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17331840)))];
            tensor<fp16, [1, 1, 12, 64]> var_1117_cast_fp16 = mul(x = query_27_cast_fp16, y = rope_cos_3_to_fp16)[name = string("op_1117_cast_fp16")];
            tensor<int32, [4]> var_1122_begin_0 = const()[name = string("op_1122_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_1122_end_0 = const()[name = string("op_1122_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_1122_end_mask_0 = const()[name = string("op_1122_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1122_stride_0 = const()[name = string("op_1122_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_1122_cast_fp16 = slice_by_index(begin = var_1122_begin_0, end = var_1122_end_0, end_mask = var_1122_end_mask_0, stride = var_1122_stride_0, x = query_27_cast_fp16)[name = string("op_1122_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_1123_cast_fp16 = mul(x = var_1122_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_1123_cast_fp16")];
            tensor<int32, [4]> var_1128_begin_0 = const()[name = string("op_1128_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1128_end_0 = const()[name = string("op_1128_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_1128_end_mask_0 = const()[name = string("op_1128_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1128_stride_0 = const()[name = string("op_1128_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_1128_cast_fp16 = slice_by_index(begin = var_1128_begin_0, end = var_1128_end_0, end_mask = var_1128_end_mask_0, stride = var_1128_stride_0, x = query_27_cast_fp16)[name = string("op_1128_cast_fp16")];
            int32 var_1131_axis_0 = const()[name = string("op_1131_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_1131_cast_fp16 = stack(axis = var_1131_axis_0, values = (var_1123_cast_fp16, var_1128_cast_fp16))[name = string("op_1131_cast_fp16")];
            tensor<int32, [4]> shape_10 = const()[name = string("shape_10"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_1132_cast_fp16 = reshape(shape = shape_10, x = var_1131_cast_fp16)[name = string("op_1132_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_sin_3_to_fp16 = const()[name = string("rope_sin_3_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17332032)))];
            tensor<fp16, [1, 1, 12, 64]> var_1133_cast_fp16 = mul(x = var_1132_cast_fp16, y = rope_sin_3_to_fp16)[name = string("op_1133_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> query_29_cast_fp16 = add(x = var_1117_cast_fp16, y = var_1133_cast_fp16)[name = string("query_29_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_1136_cast_fp16 = mul(x = key_31_cast_fp16, y = rope_cos_3_to_fp16)[name = string("op_1136_cast_fp16")];
            tensor<int32, [4]> var_1141_begin_0 = const()[name = string("op_1141_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_1141_end_0 = const()[name = string("op_1141_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_1141_end_mask_0 = const()[name = string("op_1141_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1141_stride_0 = const()[name = string("op_1141_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_1141_cast_fp16 = slice_by_index(begin = var_1141_begin_0, end = var_1141_end_0, end_mask = var_1141_end_mask_0, stride = var_1141_stride_0, x = key_31_cast_fp16)[name = string("op_1141_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_1142_cast_fp16 = mul(x = var_1141_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_1142_cast_fp16")];
            tensor<int32, [4]> var_1147_begin_0 = const()[name = string("op_1147_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1147_end_0 = const()[name = string("op_1147_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_1147_end_mask_0 = const()[name = string("op_1147_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1147_stride_0 = const()[name = string("op_1147_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_1147_cast_fp16 = slice_by_index(begin = var_1147_begin_0, end = var_1147_end_0, end_mask = var_1147_end_mask_0, stride = var_1147_stride_0, x = key_31_cast_fp16)[name = string("op_1147_cast_fp16")];
            int32 var_1150_axis_0 = const()[name = string("op_1150_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_1150_cast_fp16 = stack(axis = var_1150_axis_0, values = (var_1142_cast_fp16, var_1147_cast_fp16))[name = string("op_1150_cast_fp16")];
            tensor<int32, [4]> shape_11 = const()[name = string("shape_11"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_1151_cast_fp16 = reshape(shape = shape_11, x = var_1150_cast_fp16)[name = string("op_1151_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_1152_cast_fp16 = mul(x = var_1151_cast_fp16, y = rope_sin_3_to_fp16)[name = string("op_1152_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> key_33_cast_fp16 = add(x = var_1136_cast_fp16, y = var_1152_cast_fp16)[name = string("key_33_cast_fp16")];
            tensor<int32, [4]> var_1159 = const()[name = string("op_1159"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1165 = const()[name = string("op_1165"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1171 = const()[name = string("op_1171"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 var_1174 = const()[name = string("op_1174"), val = int32(2)];
            bool key_37_interleave_0 = const()[name = string("key_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> key_35_cast_fp16 = transpose(perm = var_1165, x = key_33_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 12, 4, 64]> key_37_cast_fp16 = concat(axis = var_1174, interleave = key_37_interleave_0, values = (key_27_cast_fp16, key_35_cast_fp16))[name = string("key_37_cast_fp16")];
            int32 var_1177 = const()[name = string("op_1177"), val = int32(2)];
            bool value_29_interleave_0 = const()[name = string("value_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> value_27_cast_fp16 = transpose(perm = var_1171, x = value_25_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 12, 4, 64]> value_29_cast_fp16 = concat(axis = var_1177, interleave = value_29_interleave_0, values = (value_21_cast_fp16, value_27_cast_fp16))[name = string("value_29_cast_fp16")];
            bool var_1182_transpose_x_1 = const()[name = string("op_1182_transpose_x_1"), val = bool(false)];
            bool var_1182_transpose_y_1 = const()[name = string("op_1182_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 12, 1, 64]> query_31_cast_fp16 = transpose(perm = var_1159, x = query_29_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 12, 1, 4]> var_1182_cast_fp16 = matmul(transpose_x = var_1182_transpose_x_1, transpose_y = var_1182_transpose_y_1, x = query_31_cast_fp16, y = key_37_cast_fp16)[name = string("op_1182_cast_fp16")];
            fp16 var_1183_to_fp16 = const()[name = string("op_1183_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 1, 4]> scores_11_cast_fp16 = mul(x = var_1182_cast_fp16, y = var_1183_to_fp16)[name = string("scores_11_cast_fp16")];
            int32 var_1185 = const()[name = string("op_1185"), val = int32(-1)];
            tensor<fp16, [1, 12, 1, 4]> probabilities_15_cast_fp16 = softmax(axis = var_1185, x = scores_11_cast_fp16)[name = string("probabilities_15_cast_fp16")];
            bool attention_7_transpose_x_0 = const()[name = string("attention_7_transpose_x_0"), val = bool(false)];
            bool attention_7_transpose_y_0 = const()[name = string("attention_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> attention_7_cast_fp16 = matmul(transpose_x = attention_7_transpose_x_0, transpose_y = attention_7_transpose_y_0, x = probabilities_15_cast_fp16, y = value_29_cast_fp16)[name = string("attention_7_cast_fp16")];
            tensor<int32, [4]> var_1193 = const()[name = string("op_1193"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1198 = const()[name = string("op_1198"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 12, 64]> var_1194_cast_fp16 = transpose(perm = var_1193, x = attention_7_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 1, 768]> input_63_cast_fp16 = reshape(shape = var_1198, x = var_1194_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_15_cast_fp16 = linear(bias = block_attn_c_proj_bias_to_fp16, weight = block_attn_c_proj_weight_to_fp16, x = input_63_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_65_cast_fp16 = add(x = input_59_cast_fp16, y = linear_15_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<int32, [1]> input_67_axes_0 = const()[name = string("input_67_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_1206_to_fp16 = const()[name = string("op_1206_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = block_ln_2_bias_to_fp16, epsilon = var_1206_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_65_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [1, 1, 3072]> linear_16_cast_fp16 = linear(bias = block_mlp_fc_in_bias_to_fp16, weight = block_mlp_fc_in_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_16_cast_fp16")];
            string input_71_mode_0 = const()[name = string("input_71_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1, 3072]> input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_16_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_17_cast_fp16 = linear(bias = block_mlp_fc_out_bias_to_fp16, weight = block_mlp_fc_out_weight_to_fp16, x = input_71_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_75_cast_fp16 = add(x = input_65_cast_fp16, y = linear_17_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<int32, [1]> var_1245_axes_0 = const()[name = string("op_1245_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_1240_to_fp16 = const()[name = string("op_1240_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> var_1245_cast_fp16 = layer_norm(axes = var_1245_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_1240_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_75_cast_fp16)[name = string("op_1245_cast_fp16")];
            tensor<int32, [3]> var_1253_begin_0 = const()[name = string("op_1253_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1253_end_0 = const()[name = string("op_1253_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_1253_end_mask_0 = const()[name = string("op_1253_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_1253_squeeze_mask_0 = const()[name = string("op_1253_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_1253_cast_fp16 = slice_by_index(begin = var_1253_begin_0, end = var_1253_end_0, end_mask = var_1253_end_mask_0, squeeze_mask = var_1253_squeeze_mask_0, x = var_1245_cast_fp16)[name = string("op_1253_cast_fp16")];
            tensor<fp16, [1024, 768]> audio_embeddings_2_weight_to_fp16 = const()[name = string("audio_embeddings_2_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17332224)))];
            tensor<fp16, [1, 1024]> linear_18_cast_fp16 = linear(bias = linear_8_bias_0_to_fp16, weight = audio_embeddings_2_weight_to_fp16, x = var_1253_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<int32, [3]> var_1268_begin_0 = const()[name = string("op_1268_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
            tensor<int32, [3]> var_1268_end_0 = const()[name = string("op_1268_end_0"), val = tensor<int32, [3]>([1, 3, 1024])];
            tensor<bool, [3]> var_1268_end_mask_0 = const()[name = string("op_1268_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_1268_squeeze_mask_0 = const()[name = string("op_1268_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<int32, [1, 1024]> var_1268 = slice_by_index(begin = var_1268_begin_0, end = var_1268_end_0, end_mask = var_1268_end_mask_0, squeeze_mask = var_1268_squeeze_mask_0, x = repetition_seen)[name = string("op_1268")];
            tensor<int32, [2]> random_u_5_begin_0 = const()[name = string("random_u_5_begin_0"), val = tensor<int32, [2]>([0, 2])];
            tensor<int32, [2]> random_u_5_end_0 = const()[name = string("random_u_5_end_0"), val = tensor<int32, [2]>([1, 3])];
            tensor<bool, [2]> random_u_5_end_mask_0 = const()[name = string("random_u_5_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<bool, [2]> random_u_5_squeeze_mask_0 = const()[name = string("random_u_5_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [1]> random_u_5_cast_fp16 = slice_by_index(begin = random_u_5_begin_0, end = random_u_5_end_0, end_mask = random_u_5_end_mask_0, squeeze_mask = random_u_5_squeeze_mask_0, x = audio_random_to_fp16)[name = string("random_u_5_cast_fp16")];
            fp16 var_1282_promoted_to_fp16 = const()[name = string("op_1282_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [1, 1024]> var_1283_cast_fp16 = less(x = linear_18_cast_fp16, y = var_1282_promoted_to_fp16)[name = string("op_1283_cast_fp16")];
            fp16 var_1284_to_fp16 = const()[name = string("op_1284_to_fp16"), val = fp16(0x1.334p+0)];
            tensor<fp16, [1, 1024]> var_1285_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_1284_to_fp16)[name = string("op_1285_cast_fp16")];
            fp16 _inversed_1287_y_0_to_fp16 = const()[name = string("_inversed_1287_y_0_to_fp16"), val = fp16(0x1.aacp-1)];
            tensor<fp16, [1, 1024]> _inversed_1287_cast_fp16 = mul(x = linear_18_cast_fp16, y = _inversed_1287_y_0_to_fp16)[name = string("_inversed_1287_cast_fp16")];
            tensor<fp16, [1, 1024]> penalized_5_cast_fp16 = select(a = var_1285_cast_fp16, b = _inversed_1287_cast_fp16, cond = var_1283_cast_fp16)[name = string("penalized_5_cast_fp16")];
            string var_1293_dtype_0 = const()[name = string("op_1293_dtype_0"), val = string("bool")];
            tensor<bool, [1, 1024]> var_1293 = cast(dtype = var_1293_dtype_0, x = var_1268)[name = string("cast_368")];
            tensor<fp16, [1, 1024]> var_1294_cast_fp16 = select(a = penalized_5_cast_fp16, b = linear_18_cast_fp16, cond = var_1293)[name = string("op_1294_cast_fp16")];
            fp16 _inversed_scores_13_y_0_to_fp16 = const()[name = string("_inversed_scores_13_y_0_to_fp16"), val = fp16(0x1.4p+0)];
            tensor<fp16, [1, 1024]> _inversed_scores_13_cast_fp16 = mul(x = var_1294_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = string("_inversed_scores_13_cast_fp16")];
            int32 var_1297 = const()[name = string("op_1297"), val = int32(25)];
            int32 top_scores_9_axis_0 = const()[name = string("top_scores_9_axis_0"), val = int32(1)];
            bool top_scores_9_ascending_0 = const()[name = string("top_scores_9_ascending_0"), val = bool(false)];
            bool top_scores_9_sort_0 = const()[name = string("top_scores_9_sort_0"), val = bool(true)];
            bool top_scores_9_return_indices_0 = const()[name = string("top_scores_9_return_indices_0"), val = bool(true)];
            string top_scores_9_cast_fp16_cast_uint16_output_indices_dtype_0 = const()[name = string("top_scores_9_cast_fp16_cast_uint16_output_indices_dtype_0"), val = string("uint16")];
            tensor<fp16, [1, 25]> top_scores_9_cast_fp16_cast_uint16_0, tensor<uint16, [1, 25]> top_scores_9_cast_fp16_cast_uint16_1 = topk(ascending = top_scores_9_ascending_0, axis = top_scores_9_axis_0, k = var_1297, output_indices_dtype = top_scores_9_cast_fp16_cast_uint16_output_indices_dtype_0, return_indices = top_scores_9_return_indices_0, sort = top_scores_9_sort_0, x = _inversed_scores_13_cast_fp16)[name = string("top_scores_9_cast_fp16_cast_uint16")];
            int32 var_1303 = const()[name = string("op_1303"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_17_cast_fp16 = softmax(axis = var_1303, x = top_scores_9_cast_fp16_cast_uint16_0)[name = string("probabilities_17_cast_fp16")];
            int32 var_1306 = const()[name = string("op_1306"), val = int32(1)];
            bool cumulative_9_exclusive_0 = const()[name = string("cumulative_9_exclusive_0"), val = bool(false)];
            bool cumulative_9_reverse_0 = const()[name = string("cumulative_9_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_9_cast_fp16 = cumsum(axis = var_1306, exclusive = cumulative_9_exclusive_0, reverse = cumulative_9_reverse_0, x = probabilities_17_cast_fp16)[name = string("cumulative_9_cast_fp16")];
            tensor<fp16, [1, 25]> var_1310_cast_fp16 = sub(x = cumulative_9_cast_fp16, y = probabilities_17_cast_fp16)[name = string("op_1310_cast_fp16")];
            fp16 var_1311_to_fp16 = const()[name = string("op_1311_to_fp16"), val = fp16(0x1.e68p-1)];
            tensor<bool, [1, 25]> keep_5_cast_fp16 = less(x = var_1310_cast_fp16, y = var_1311_to_fp16)[name = string("keep_5_cast_fp16")];
            fp16 var_1319_value_0_to_fp16 = const()[name = string("op_1319_value_0_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 25]> var_1319_cast_fp16 = fill_like(ref_tensor = top_scores_9_cast_fp16_cast_uint16_0, value = var_1319_value_0_to_fp16)[name = string("op_1319_cast_fp16")];
            tensor<fp16, [1, 25]> top_scores_11_cast_fp16 = select(a = top_scores_9_cast_fp16_cast_uint16_0, b = var_1319_cast_fp16, cond = keep_5_cast_fp16)[name = string("top_scores_11_cast_fp16")];
            int32 var_1321 = const()[name = string("op_1321"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_19_cast_fp16 = softmax(axis = var_1321, x = top_scores_11_cast_fp16)[name = string("probabilities_19_cast_fp16")];
            int32 var_1324 = const()[name = string("op_1324"), val = int32(1)];
            bool cumulative_11_exclusive_0 = const()[name = string("cumulative_11_exclusive_0"), val = bool(false)];
            bool cumulative_11_reverse_0 = const()[name = string("cumulative_11_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_11_cast_fp16 = cumsum(axis = var_1324, exclusive = cumulative_11_exclusive_0, reverse = cumulative_11_reverse_0, x = probabilities_19_cast_fp16)[name = string("cumulative_11_cast_fp16")];
            tensor<int32, [2]> var_1329 = const()[name = string("op_1329"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_1330_cast_fp16 = reshape(shape = var_1329, x = random_u_5_cast_fp16)[name = string("op_1330_cast_fp16")];
            tensor<bool, [1, 25]> var_1331_cast_fp16 = less(x = cumulative_11_cast_fp16, y = var_1330_cast_fp16)[name = string("op_1331_cast_fp16")];
            string var_1336_dtype_0 = const()[name = string("op_1336_dtype_0"), val = string("int32")];
            tensor<int32, [1]> selected_9_axes_0 = const()[name = string("selected_9_axes_0"), val = tensor<int32, [1]>([1])];
            bool selected_9_keep_dims_0 = const()[name = string("selected_9_keep_dims_0"), val = bool(false)];
            tensor<int32, [1, 25]> var_1336 = cast(dtype = var_1336_dtype_0, x = var_1331_cast_fp16)[name = string("cast_367")];
            tensor<int32, [1]> selected_9 = reduce_sum(axes = selected_9_axes_0, keep_dims = selected_9_keep_dims_0, x = var_1336)[name = string("selected_9")];
            string selected_9_promoted_to_fp16_dtype_0 = const()[name = string("selected_9_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_18_to_fp16 = const()[name = string("const_18_to_fp16"), val = fp16(-inf)];
            fp16 var_1343_promoted_to_fp16 = const()[name = string("op_1343_promoted_to_fp16"), val = fp16(0x1.8p+4)];
            tensor<fp16, [1]> selected_9_to_fp16 = cast(dtype = selected_9_promoted_to_fp16_dtype_0, x = selected_9)[name = string("cast_366")];
            tensor<fp16, [1]> clip_2_cast_fp16 = clip(alpha = const_18_to_fp16, beta = var_1343_promoted_to_fp16, x = selected_9_to_fp16)[name = string("clip_2_cast_fp16")];
            string selected_11_dtype_0 = const()[name = string("selected_11_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_1352 = const()[name = string("op_1352"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> clip_2_cast_fp16_to_int32 = cast(dtype = selected_11_dtype_0, x = clip_2_cast_fp16)[name = string("cast_365")];
            tensor<int32, [1, 1]> var_1353 = reshape(shape = var_1352, x = clip_2_cast_fp16_to_int32)[name = string("op_1353")];
            int32 var_1354 = const()[name = string("op_1354"), val = int32(1)];
            bool var_1356_validate_indices_0 = const()[name = string("op_1356_validate_indices_0"), val = bool(false)];
            string var_1353_to_uint16_dtype_0 = const()[name = string("op_1353_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 1]> var_1353_to_uint16 = cast(dtype = var_1353_to_uint16_dtype_0, x = var_1353)[name = string("cast_364")];
            tensor<uint16, [1, 1]> var_1356_cast_uint16 = gather_along_axis(axis = var_1354, indices = var_1353_to_uint16, validate_indices = var_1356_validate_indices_0, x = top_scores_9_cast_fp16_cast_uint16_1)[name = string("op_1356_cast_uint16")];
            string var_1356_cast_uint16_to_int32_dtype_0 = const()[name = string("op_1356_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [1]> var_1358 = const()[name = string("op_1358"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_1356_cast_uint16_to_int32 = cast(dtype = var_1356_cast_uint16_to_int32_dtype_0, x = var_1356_cast_uint16)[name = string("cast_363")];
            tensor<int32, [1]> token_5 = reshape(shape = var_1358, x = var_1356_cast_uint16_to_int32)[name = string("token_5")];
            int32 var_1368_axis_0 = const()[name = string("op_1368_axis_0"), val = int32(0)];
            int32 var_1368_batch_dims_0 = const()[name = string("op_1368_batch_dims_0"), val = int32(0)];
            bool var_1368_validate_indices_0 = const()[name = string("op_1368_validate_indices_0"), val = bool(false)];
            string token_5_to_uint16_dtype_0 = const()[name = string("token_5_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> token_5_to_uint16 = cast(dtype = token_5_to_uint16_dtype_0, x = token_5)[name = string("cast_362")];
            tensor<fp16, [1, 768]> var_1368_cast_fp16_cast_uint16 = gather(axis = var_1368_axis_0, batch_dims = var_1368_batch_dims_0, indices = token_5_to_uint16, validate_indices = var_1368_validate_indices_0, x = audio_embeddings_2_weight_to_fp16)[name = string("op_1368_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 768]> next_global_embed_7_cast_fp16 = add(x = next_global_embed_5_cast_fp16, y = var_1368_cast_fp16_cast_uint16)[name = string("next_global_embed_7_cast_fp16")];
            tensor<int32, [3]> var_1378 = const()[name = string("op_1378"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 768]> input_79_cast_fp16 = reshape(shape = var_1378, x = var_1368_cast_fp16_cast_uint16)[name = string("input_79_cast_fp16")];
            tensor<int32, [1]> input_81_axes_0 = const()[name = string("input_81_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_1381_to_fp16 = const()[name = string("op_1381_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = block_ln_1_bias_to_fp16, epsilon = var_1381_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_79_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [1, 1, 2304]> linear_19_cast_fp16 = linear(bias = block_attn_c_attn_bias_to_fp16, weight = block_attn_c_attn_weight_to_fp16, x = input_81_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<int32, [3]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [3]>([768, 768, 768])];
            int32 var_1392_axis_0 = const()[name = string("op_1392_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 768]> var_1392_cast_fp16_0, tensor<fp16, [1, 1, 768]> var_1392_cast_fp16_1, tensor<fp16, [1, 1, 768]> var_1392_cast_fp16_2 = split(axis = var_1392_axis_0, split_sizes = tile_4, x = linear_19_cast_fp16)[name = string("op_1392_cast_fp16")];
            tensor<int32, [4]> var_1400 = const()[name = string("op_1400"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> query_35_cast_fp16 = reshape(shape = var_1400, x = var_1392_cast_fp16_0)[name = string("query_35_cast_fp16")];
            tensor<int32, [4]> var_1406 = const()[name = string("op_1406"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> key_41_cast_fp16 = reshape(shape = var_1406, x = var_1392_cast_fp16_1)[name = string("key_41_cast_fp16")];
            tensor<int32, [4]> var_1412 = const()[name = string("op_1412"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> value_33_cast_fp16 = reshape(shape = var_1412, x = var_1392_cast_fp16_2)[name = string("value_33_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_cos_4_to_fp16 = const()[name = string("rope_cos_4_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18905152)))];
            tensor<fp16, [1, 1, 12, 64]> var_1414_cast_fp16 = mul(x = query_35_cast_fp16, y = rope_cos_4_to_fp16)[name = string("op_1414_cast_fp16")];
            tensor<int32, [4]> var_1419_begin_0 = const()[name = string("op_1419_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_1419_end_0 = const()[name = string("op_1419_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_1419_end_mask_0 = const()[name = string("op_1419_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1419_stride_0 = const()[name = string("op_1419_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_1419_cast_fp16 = slice_by_index(begin = var_1419_begin_0, end = var_1419_end_0, end_mask = var_1419_end_mask_0, stride = var_1419_stride_0, x = query_35_cast_fp16)[name = string("op_1419_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_1420_cast_fp16 = mul(x = var_1419_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_1420_cast_fp16")];
            tensor<int32, [4]> var_1425_begin_0 = const()[name = string("op_1425_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1425_end_0 = const()[name = string("op_1425_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_1425_end_mask_0 = const()[name = string("op_1425_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1425_stride_0 = const()[name = string("op_1425_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_1425_cast_fp16 = slice_by_index(begin = var_1425_begin_0, end = var_1425_end_0, end_mask = var_1425_end_mask_0, stride = var_1425_stride_0, x = query_35_cast_fp16)[name = string("op_1425_cast_fp16")];
            int32 var_1428_axis_0 = const()[name = string("op_1428_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_1428_cast_fp16 = stack(axis = var_1428_axis_0, values = (var_1420_cast_fp16, var_1425_cast_fp16))[name = string("op_1428_cast_fp16")];
            tensor<int32, [4]> shape_13 = const()[name = string("shape_13"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_1429_cast_fp16 = reshape(shape = shape_13, x = var_1428_cast_fp16)[name = string("op_1429_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_sin_4_to_fp16 = const()[name = string("rope_sin_4_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18905344)))];
            tensor<fp16, [1, 1, 12, 64]> var_1430_cast_fp16 = mul(x = var_1429_cast_fp16, y = rope_sin_4_to_fp16)[name = string("op_1430_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> query_37_cast_fp16 = add(x = var_1414_cast_fp16, y = var_1430_cast_fp16)[name = string("query_37_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_1433_cast_fp16 = mul(x = key_41_cast_fp16, y = rope_cos_4_to_fp16)[name = string("op_1433_cast_fp16")];
            tensor<int32, [4]> var_1438_begin_0 = const()[name = string("op_1438_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_1438_end_0 = const()[name = string("op_1438_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_1438_end_mask_0 = const()[name = string("op_1438_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1438_stride_0 = const()[name = string("op_1438_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_1438_cast_fp16 = slice_by_index(begin = var_1438_begin_0, end = var_1438_end_0, end_mask = var_1438_end_mask_0, stride = var_1438_stride_0, x = key_41_cast_fp16)[name = string("op_1438_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_1439_cast_fp16 = mul(x = var_1438_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1439_cast_fp16")];
            tensor<int32, [4]> var_1444_begin_0 = const()[name = string("op_1444_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1444_end_0 = const()[name = string("op_1444_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_1444_end_mask_0 = const()[name = string("op_1444_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1444_stride_0 = const()[name = string("op_1444_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_1444_cast_fp16 = slice_by_index(begin = var_1444_begin_0, end = var_1444_end_0, end_mask = var_1444_end_mask_0, stride = var_1444_stride_0, x = key_41_cast_fp16)[name = string("op_1444_cast_fp16")];
            int32 var_1447_axis_0 = const()[name = string("op_1447_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_1447_cast_fp16 = stack(axis = var_1447_axis_0, values = (var_1439_cast_fp16, var_1444_cast_fp16))[name = string("op_1447_cast_fp16")];
            tensor<int32, [4]> shape_14 = const()[name = string("shape_14"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_1448_cast_fp16 = reshape(shape = shape_14, x = var_1447_cast_fp16)[name = string("op_1448_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_1449_cast_fp16 = mul(x = var_1448_cast_fp16, y = rope_sin_4_to_fp16)[name = string("op_1449_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> key_43_cast_fp16 = add(x = var_1433_cast_fp16, y = var_1449_cast_fp16)[name = string("key_43_cast_fp16")];
            tensor<int32, [4]> var_1456 = const()[name = string("op_1456"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1462 = const()[name = string("op_1462"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1468 = const()[name = string("op_1468"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 var_1471 = const()[name = string("op_1471"), val = int32(2)];
            bool key_47_interleave_0 = const()[name = string("key_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> key_45_cast_fp16 = transpose(perm = var_1462, x = key_43_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 12, 5, 64]> key_47_cast_fp16 = concat(axis = var_1471, interleave = key_47_interleave_0, values = (key_37_cast_fp16, key_45_cast_fp16))[name = string("key_47_cast_fp16")];
            int32 var_1474 = const()[name = string("op_1474"), val = int32(2)];
            bool value_37_interleave_0 = const()[name = string("value_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> value_35_cast_fp16 = transpose(perm = var_1468, x = value_33_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 12, 5, 64]> value_37_cast_fp16 = concat(axis = var_1474, interleave = value_37_interleave_0, values = (value_29_cast_fp16, value_35_cast_fp16))[name = string("value_37_cast_fp16")];
            bool var_1479_transpose_x_1 = const()[name = string("op_1479_transpose_x_1"), val = bool(false)];
            bool var_1479_transpose_y_1 = const()[name = string("op_1479_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 12, 1, 64]> query_39_cast_fp16 = transpose(perm = var_1456, x = query_37_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 12, 1, 5]> var_1479_cast_fp16 = matmul(transpose_x = var_1479_transpose_x_1, transpose_y = var_1479_transpose_y_1, x = query_39_cast_fp16, y = key_47_cast_fp16)[name = string("op_1479_cast_fp16")];
            fp16 var_1480_to_fp16 = const()[name = string("op_1480_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 1, 5]> scores_15_cast_fp16 = mul(x = var_1479_cast_fp16, y = var_1480_to_fp16)[name = string("scores_15_cast_fp16")];
            int32 var_1482 = const()[name = string("op_1482"), val = int32(-1)];
            tensor<fp16, [1, 12, 1, 5]> probabilities_21_cast_fp16 = softmax(axis = var_1482, x = scores_15_cast_fp16)[name = string("probabilities_21_cast_fp16")];
            bool attention_9_transpose_x_0 = const()[name = string("attention_9_transpose_x_0"), val = bool(false)];
            bool attention_9_transpose_y_0 = const()[name = string("attention_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> attention_9_cast_fp16 = matmul(transpose_x = attention_9_transpose_x_0, transpose_y = attention_9_transpose_y_0, x = probabilities_21_cast_fp16, y = value_37_cast_fp16)[name = string("attention_9_cast_fp16")];
            tensor<int32, [4]> var_1490 = const()[name = string("op_1490"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1495 = const()[name = string("op_1495"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 12, 64]> var_1491_cast_fp16 = transpose(perm = var_1490, x = attention_9_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 1, 768]> input_83_cast_fp16 = reshape(shape = var_1495, x = var_1491_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_20_cast_fp16 = linear(bias = block_attn_c_proj_bias_to_fp16, weight = block_attn_c_proj_weight_to_fp16, x = input_83_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_85_cast_fp16 = add(x = input_79_cast_fp16, y = linear_20_cast_fp16)[name = string("input_85_cast_fp16")];
            tensor<int32, [1]> input_87_axes_0 = const()[name = string("input_87_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_1503_to_fp16 = const()[name = string("op_1503_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = block_ln_2_bias_to_fp16, epsilon = var_1503_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_85_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [1, 1, 3072]> linear_21_cast_fp16 = linear(bias = block_mlp_fc_in_bias_to_fp16, weight = block_mlp_fc_in_weight_to_fp16, x = input_87_cast_fp16)[name = string("linear_21_cast_fp16")];
            string input_91_mode_0 = const()[name = string("input_91_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1, 3072]> input_91_cast_fp16 = gelu(mode = input_91_mode_0, x = linear_21_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_22_cast_fp16 = linear(bias = block_mlp_fc_out_bias_to_fp16, weight = block_mlp_fc_out_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_95_cast_fp16 = add(x = input_85_cast_fp16, y = linear_22_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<int32, [1]> var_1542_axes_0 = const()[name = string("op_1542_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_1537_to_fp16 = const()[name = string("op_1537_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> var_1542_cast_fp16 = layer_norm(axes = var_1542_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_1537_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_95_cast_fp16)[name = string("op_1542_cast_fp16")];
            tensor<int32, [3]> var_1550_begin_0 = const()[name = string("op_1550_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1550_end_0 = const()[name = string("op_1550_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_1550_end_mask_0 = const()[name = string("op_1550_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_1550_squeeze_mask_0 = const()[name = string("op_1550_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_1550_cast_fp16 = slice_by_index(begin = var_1550_begin_0, end = var_1550_end_0, end_mask = var_1550_end_mask_0, squeeze_mask = var_1550_squeeze_mask_0, x = var_1542_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<fp16, [1024, 768]> audio_embeddings_3_weight_to_fp16 = const()[name = string("audio_embeddings_3_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18905536)))];
            tensor<fp16, [1, 1024]> linear_23_cast_fp16 = linear(bias = linear_8_bias_0_to_fp16, weight = audio_embeddings_3_weight_to_fp16, x = var_1550_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [3]> var_1565_begin_0 = const()[name = string("op_1565_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
            tensor<int32, [3]> var_1565_end_0 = const()[name = string("op_1565_end_0"), val = tensor<int32, [3]>([1, 4, 1024])];
            tensor<bool, [3]> var_1565_end_mask_0 = const()[name = string("op_1565_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_1565_squeeze_mask_0 = const()[name = string("op_1565_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<int32, [1, 1024]> var_1565 = slice_by_index(begin = var_1565_begin_0, end = var_1565_end_0, end_mask = var_1565_end_mask_0, squeeze_mask = var_1565_squeeze_mask_0, x = repetition_seen)[name = string("op_1565")];
            tensor<int32, [2]> random_u_7_begin_0 = const()[name = string("random_u_7_begin_0"), val = tensor<int32, [2]>([0, 3])];
            tensor<int32, [2]> random_u_7_end_0 = const()[name = string("random_u_7_end_0"), val = tensor<int32, [2]>([1, 4])];
            tensor<bool, [2]> random_u_7_end_mask_0 = const()[name = string("random_u_7_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<bool, [2]> random_u_7_squeeze_mask_0 = const()[name = string("random_u_7_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [1]> random_u_7_cast_fp16 = slice_by_index(begin = random_u_7_begin_0, end = random_u_7_end_0, end_mask = random_u_7_end_mask_0, squeeze_mask = random_u_7_squeeze_mask_0, x = audio_random_to_fp16)[name = string("random_u_7_cast_fp16")];
            fp16 var_1579_promoted_to_fp16 = const()[name = string("op_1579_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [1, 1024]> var_1580_cast_fp16 = less(x = linear_23_cast_fp16, y = var_1579_promoted_to_fp16)[name = string("op_1580_cast_fp16")];
            fp16 var_1581_to_fp16 = const()[name = string("op_1581_to_fp16"), val = fp16(0x1.334p+0)];
            tensor<fp16, [1, 1024]> var_1582_cast_fp16 = mul(x = linear_23_cast_fp16, y = var_1581_to_fp16)[name = string("op_1582_cast_fp16")];
            fp16 _inversed_1584_y_0_to_fp16 = const()[name = string("_inversed_1584_y_0_to_fp16"), val = fp16(0x1.aacp-1)];
            tensor<fp16, [1, 1024]> _inversed_1584_cast_fp16 = mul(x = linear_23_cast_fp16, y = _inversed_1584_y_0_to_fp16)[name = string("_inversed_1584_cast_fp16")];
            tensor<fp16, [1, 1024]> penalized_7_cast_fp16 = select(a = var_1582_cast_fp16, b = _inversed_1584_cast_fp16, cond = var_1580_cast_fp16)[name = string("penalized_7_cast_fp16")];
            string var_1590_dtype_0 = const()[name = string("op_1590_dtype_0"), val = string("bool")];
            tensor<bool, [1, 1024]> var_1590 = cast(dtype = var_1590_dtype_0, x = var_1565)[name = string("cast_361")];
            tensor<fp16, [1, 1024]> var_1591_cast_fp16 = select(a = penalized_7_cast_fp16, b = linear_23_cast_fp16, cond = var_1590)[name = string("op_1591_cast_fp16")];
            fp16 _inversed_scores_17_y_0_to_fp16 = const()[name = string("_inversed_scores_17_y_0_to_fp16"), val = fp16(0x1.4p+0)];
            tensor<fp16, [1, 1024]> _inversed_scores_17_cast_fp16 = mul(x = var_1591_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = string("_inversed_scores_17_cast_fp16")];
            int32 var_1594 = const()[name = string("op_1594"), val = int32(25)];
            int32 top_scores_13_axis_0 = const()[name = string("top_scores_13_axis_0"), val = int32(1)];
            bool top_scores_13_ascending_0 = const()[name = string("top_scores_13_ascending_0"), val = bool(false)];
            bool top_scores_13_sort_0 = const()[name = string("top_scores_13_sort_0"), val = bool(true)];
            bool top_scores_13_return_indices_0 = const()[name = string("top_scores_13_return_indices_0"), val = bool(true)];
            string top_scores_13_cast_fp16_cast_uint16_output_indices_dtype_0 = const()[name = string("top_scores_13_cast_fp16_cast_uint16_output_indices_dtype_0"), val = string("uint16")];
            tensor<fp16, [1, 25]> top_scores_13_cast_fp16_cast_uint16_0, tensor<uint16, [1, 25]> top_scores_13_cast_fp16_cast_uint16_1 = topk(ascending = top_scores_13_ascending_0, axis = top_scores_13_axis_0, k = var_1594, output_indices_dtype = top_scores_13_cast_fp16_cast_uint16_output_indices_dtype_0, return_indices = top_scores_13_return_indices_0, sort = top_scores_13_sort_0, x = _inversed_scores_17_cast_fp16)[name = string("top_scores_13_cast_fp16_cast_uint16")];
            int32 var_1600 = const()[name = string("op_1600"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_23_cast_fp16 = softmax(axis = var_1600, x = top_scores_13_cast_fp16_cast_uint16_0)[name = string("probabilities_23_cast_fp16")];
            int32 var_1603 = const()[name = string("op_1603"), val = int32(1)];
            bool cumulative_13_exclusive_0 = const()[name = string("cumulative_13_exclusive_0"), val = bool(false)];
            bool cumulative_13_reverse_0 = const()[name = string("cumulative_13_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_13_cast_fp16 = cumsum(axis = var_1603, exclusive = cumulative_13_exclusive_0, reverse = cumulative_13_reverse_0, x = probabilities_23_cast_fp16)[name = string("cumulative_13_cast_fp16")];
            tensor<fp16, [1, 25]> var_1607_cast_fp16 = sub(x = cumulative_13_cast_fp16, y = probabilities_23_cast_fp16)[name = string("op_1607_cast_fp16")];
            fp16 var_1608_to_fp16 = const()[name = string("op_1608_to_fp16"), val = fp16(0x1.e68p-1)];
            tensor<bool, [1, 25]> keep_7_cast_fp16 = less(x = var_1607_cast_fp16, y = var_1608_to_fp16)[name = string("keep_7_cast_fp16")];
            fp16 var_1616_value_0_to_fp16 = const()[name = string("op_1616_value_0_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 25]> var_1616_cast_fp16 = fill_like(ref_tensor = top_scores_13_cast_fp16_cast_uint16_0, value = var_1616_value_0_to_fp16)[name = string("op_1616_cast_fp16")];
            tensor<fp16, [1, 25]> top_scores_15_cast_fp16 = select(a = top_scores_13_cast_fp16_cast_uint16_0, b = var_1616_cast_fp16, cond = keep_7_cast_fp16)[name = string("top_scores_15_cast_fp16")];
            int32 var_1618 = const()[name = string("op_1618"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_25_cast_fp16 = softmax(axis = var_1618, x = top_scores_15_cast_fp16)[name = string("probabilities_25_cast_fp16")];
            int32 var_1621 = const()[name = string("op_1621"), val = int32(1)];
            bool cumulative_15_exclusive_0 = const()[name = string("cumulative_15_exclusive_0"), val = bool(false)];
            bool cumulative_15_reverse_0 = const()[name = string("cumulative_15_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_15_cast_fp16 = cumsum(axis = var_1621, exclusive = cumulative_15_exclusive_0, reverse = cumulative_15_reverse_0, x = probabilities_25_cast_fp16)[name = string("cumulative_15_cast_fp16")];
            tensor<int32, [2]> var_1626 = const()[name = string("op_1626"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_1627_cast_fp16 = reshape(shape = var_1626, x = random_u_7_cast_fp16)[name = string("op_1627_cast_fp16")];
            tensor<bool, [1, 25]> var_1628_cast_fp16 = less(x = cumulative_15_cast_fp16, y = var_1627_cast_fp16)[name = string("op_1628_cast_fp16")];
            string var_1633_dtype_0 = const()[name = string("op_1633_dtype_0"), val = string("int32")];
            tensor<int32, [1]> selected_13_axes_0 = const()[name = string("selected_13_axes_0"), val = tensor<int32, [1]>([1])];
            bool selected_13_keep_dims_0 = const()[name = string("selected_13_keep_dims_0"), val = bool(false)];
            tensor<int32, [1, 25]> var_1633 = cast(dtype = var_1633_dtype_0, x = var_1628_cast_fp16)[name = string("cast_360")];
            tensor<int32, [1]> selected_13 = reduce_sum(axes = selected_13_axes_0, keep_dims = selected_13_keep_dims_0, x = var_1633)[name = string("selected_13")];
            string selected_13_promoted_to_fp16_dtype_0 = const()[name = string("selected_13_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_23_to_fp16 = const()[name = string("const_23_to_fp16"), val = fp16(-inf)];
            fp16 var_1640_promoted_to_fp16 = const()[name = string("op_1640_promoted_to_fp16"), val = fp16(0x1.8p+4)];
            tensor<fp16, [1]> selected_13_to_fp16 = cast(dtype = selected_13_promoted_to_fp16_dtype_0, x = selected_13)[name = string("cast_359")];
            tensor<fp16, [1]> clip_3_cast_fp16 = clip(alpha = const_23_to_fp16, beta = var_1640_promoted_to_fp16, x = selected_13_to_fp16)[name = string("clip_3_cast_fp16")];
            string selected_15_dtype_0 = const()[name = string("selected_15_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_1649 = const()[name = string("op_1649"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> clip_3_cast_fp16_to_int32 = cast(dtype = selected_15_dtype_0, x = clip_3_cast_fp16)[name = string("cast_358")];
            tensor<int32, [1, 1]> var_1650 = reshape(shape = var_1649, x = clip_3_cast_fp16_to_int32)[name = string("op_1650")];
            int32 var_1651 = const()[name = string("op_1651"), val = int32(1)];
            bool var_1653_validate_indices_0 = const()[name = string("op_1653_validate_indices_0"), val = bool(false)];
            string var_1650_to_uint16_dtype_0 = const()[name = string("op_1650_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 1]> var_1650_to_uint16 = cast(dtype = var_1650_to_uint16_dtype_0, x = var_1650)[name = string("cast_357")];
            tensor<uint16, [1, 1]> var_1653_cast_uint16 = gather_along_axis(axis = var_1651, indices = var_1650_to_uint16, validate_indices = var_1653_validate_indices_0, x = top_scores_13_cast_fp16_cast_uint16_1)[name = string("op_1653_cast_uint16")];
            string var_1653_cast_uint16_to_int32_dtype_0 = const()[name = string("op_1653_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [1]> var_1655 = const()[name = string("op_1655"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_1653_cast_uint16_to_int32 = cast(dtype = var_1653_cast_uint16_to_int32_dtype_0, x = var_1653_cast_uint16)[name = string("cast_356")];
            tensor<int32, [1]> token_7 = reshape(shape = var_1655, x = var_1653_cast_uint16_to_int32)[name = string("token_7")];
            int32 var_1665_axis_0 = const()[name = string("op_1665_axis_0"), val = int32(0)];
            int32 var_1665_batch_dims_0 = const()[name = string("op_1665_batch_dims_0"), val = int32(0)];
            bool var_1665_validate_indices_0 = const()[name = string("op_1665_validate_indices_0"), val = bool(false)];
            string token_7_to_uint16_dtype_0 = const()[name = string("token_7_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> token_7_to_uint16 = cast(dtype = token_7_to_uint16_dtype_0, x = token_7)[name = string("cast_355")];
            tensor<fp16, [1, 768]> var_1665_cast_fp16_cast_uint16 = gather(axis = var_1665_axis_0, batch_dims = var_1665_batch_dims_0, indices = token_7_to_uint16, validate_indices = var_1665_validate_indices_0, x = audio_embeddings_3_weight_to_fp16)[name = string("op_1665_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 768]> next_global_embed_9_cast_fp16 = add(x = next_global_embed_7_cast_fp16, y = var_1665_cast_fp16_cast_uint16)[name = string("next_global_embed_9_cast_fp16")];
            tensor<int32, [3]> var_1675 = const()[name = string("op_1675"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 768]> input_99_cast_fp16 = reshape(shape = var_1675, x = var_1665_cast_fp16_cast_uint16)[name = string("input_99_cast_fp16")];
            tensor<int32, [1]> input_101_axes_0 = const()[name = string("input_101_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_1678_to_fp16 = const()[name = string("op_1678_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = block_ln_1_bias_to_fp16, epsilon = var_1678_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_99_cast_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [1, 1, 2304]> linear_24_cast_fp16 = linear(bias = block_attn_c_attn_bias_to_fp16, weight = block_attn_c_attn_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<int32, [3]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [3]>([768, 768, 768])];
            int32 var_1689_axis_0 = const()[name = string("op_1689_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 768]> var_1689_cast_fp16_0, tensor<fp16, [1, 1, 768]> var_1689_cast_fp16_1, tensor<fp16, [1, 1, 768]> var_1689_cast_fp16_2 = split(axis = var_1689_axis_0, split_sizes = tile_5, x = linear_24_cast_fp16)[name = string("op_1689_cast_fp16")];
            tensor<int32, [4]> var_1697 = const()[name = string("op_1697"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> query_43_cast_fp16 = reshape(shape = var_1697, x = var_1689_cast_fp16_0)[name = string("query_43_cast_fp16")];
            tensor<int32, [4]> var_1703 = const()[name = string("op_1703"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> key_51_cast_fp16 = reshape(shape = var_1703, x = var_1689_cast_fp16_1)[name = string("key_51_cast_fp16")];
            tensor<int32, [4]> var_1709 = const()[name = string("op_1709"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> value_41_cast_fp16 = reshape(shape = var_1709, x = var_1689_cast_fp16_2)[name = string("value_41_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_cos_5_to_fp16 = const()[name = string("rope_cos_5_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20478464)))];
            tensor<fp16, [1, 1, 12, 64]> var_1711_cast_fp16 = mul(x = query_43_cast_fp16, y = rope_cos_5_to_fp16)[name = string("op_1711_cast_fp16")];
            tensor<int32, [4]> var_1716_begin_0 = const()[name = string("op_1716_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_1716_end_0 = const()[name = string("op_1716_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_1716_end_mask_0 = const()[name = string("op_1716_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1716_stride_0 = const()[name = string("op_1716_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_1716_cast_fp16 = slice_by_index(begin = var_1716_begin_0, end = var_1716_end_0, end_mask = var_1716_end_mask_0, stride = var_1716_stride_0, x = query_43_cast_fp16)[name = string("op_1716_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_1717_cast_fp16 = mul(x = var_1716_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_1717_cast_fp16")];
            tensor<int32, [4]> var_1722_begin_0 = const()[name = string("op_1722_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1722_end_0 = const()[name = string("op_1722_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_1722_end_mask_0 = const()[name = string("op_1722_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1722_stride_0 = const()[name = string("op_1722_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_1722_cast_fp16 = slice_by_index(begin = var_1722_begin_0, end = var_1722_end_0, end_mask = var_1722_end_mask_0, stride = var_1722_stride_0, x = query_43_cast_fp16)[name = string("op_1722_cast_fp16")];
            int32 var_1725_axis_0 = const()[name = string("op_1725_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_1725_cast_fp16 = stack(axis = var_1725_axis_0, values = (var_1717_cast_fp16, var_1722_cast_fp16))[name = string("op_1725_cast_fp16")];
            tensor<int32, [4]> shape_16 = const()[name = string("shape_16"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_1726_cast_fp16 = reshape(shape = shape_16, x = var_1725_cast_fp16)[name = string("op_1726_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_sin_5_to_fp16 = const()[name = string("rope_sin_5_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20478656)))];
            tensor<fp16, [1, 1, 12, 64]> var_1727_cast_fp16 = mul(x = var_1726_cast_fp16, y = rope_sin_5_to_fp16)[name = string("op_1727_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> query_45_cast_fp16 = add(x = var_1711_cast_fp16, y = var_1727_cast_fp16)[name = string("query_45_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_1730_cast_fp16 = mul(x = key_51_cast_fp16, y = rope_cos_5_to_fp16)[name = string("op_1730_cast_fp16")];
            tensor<int32, [4]> var_1735_begin_0 = const()[name = string("op_1735_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_1735_end_0 = const()[name = string("op_1735_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_1735_end_mask_0 = const()[name = string("op_1735_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1735_stride_0 = const()[name = string("op_1735_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_1735_cast_fp16 = slice_by_index(begin = var_1735_begin_0, end = var_1735_end_0, end_mask = var_1735_end_mask_0, stride = var_1735_stride_0, x = key_51_cast_fp16)[name = string("op_1735_cast_fp16")];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_1736_cast_fp16 = mul(x = var_1735_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_1736_cast_fp16")];
            tensor<int32, [4]> var_1741_begin_0 = const()[name = string("op_1741_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1741_end_0 = const()[name = string("op_1741_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_1741_end_mask_0 = const()[name = string("op_1741_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1741_stride_0 = const()[name = string("op_1741_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_1741_cast_fp16 = slice_by_index(begin = var_1741_begin_0, end = var_1741_end_0, end_mask = var_1741_end_mask_0, stride = var_1741_stride_0, x = key_51_cast_fp16)[name = string("op_1741_cast_fp16")];
            int32 var_1744_axis_0 = const()[name = string("op_1744_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_1744_cast_fp16 = stack(axis = var_1744_axis_0, values = (var_1736_cast_fp16, var_1741_cast_fp16))[name = string("op_1744_cast_fp16")];
            tensor<int32, [4]> shape_17 = const()[name = string("shape_17"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_1745_cast_fp16 = reshape(shape = shape_17, x = var_1744_cast_fp16)[name = string("op_1745_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_1746_cast_fp16 = mul(x = var_1745_cast_fp16, y = rope_sin_5_to_fp16)[name = string("op_1746_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> key_53_cast_fp16 = add(x = var_1730_cast_fp16, y = var_1746_cast_fp16)[name = string("key_53_cast_fp16")];
            tensor<int32, [4]> var_1753 = const()[name = string("op_1753"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1759 = const()[name = string("op_1759"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1765 = const()[name = string("op_1765"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 var_1768 = const()[name = string("op_1768"), val = int32(2)];
            bool key_57_interleave_0 = const()[name = string("key_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> key_55_cast_fp16 = transpose(perm = var_1759, x = key_53_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 12, 6, 64]> key_57_cast_fp16 = concat(axis = var_1768, interleave = key_57_interleave_0, values = (key_47_cast_fp16, key_55_cast_fp16))[name = string("key_57_cast_fp16")];
            int32 var_1771 = const()[name = string("op_1771"), val = int32(2)];
            bool value_45_interleave_0 = const()[name = string("value_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> value_43_cast_fp16 = transpose(perm = var_1765, x = value_41_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 12, 6, 64]> value_45_cast_fp16 = concat(axis = var_1771, interleave = value_45_interleave_0, values = (value_37_cast_fp16, value_43_cast_fp16))[name = string("value_45_cast_fp16")];
            bool var_1776_transpose_x_1 = const()[name = string("op_1776_transpose_x_1"), val = bool(false)];
            bool var_1776_transpose_y_1 = const()[name = string("op_1776_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 12, 1, 64]> query_47_cast_fp16 = transpose(perm = var_1753, x = query_45_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 12, 1, 6]> var_1776_cast_fp16 = matmul(transpose_x = var_1776_transpose_x_1, transpose_y = var_1776_transpose_y_1, x = query_47_cast_fp16, y = key_57_cast_fp16)[name = string("op_1776_cast_fp16")];
            fp16 var_1777_to_fp16 = const()[name = string("op_1777_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 1, 6]> scores_19_cast_fp16 = mul(x = var_1776_cast_fp16, y = var_1777_to_fp16)[name = string("scores_19_cast_fp16")];
            int32 var_1779 = const()[name = string("op_1779"), val = int32(-1)];
            tensor<fp16, [1, 12, 1, 6]> probabilities_27_cast_fp16 = softmax(axis = var_1779, x = scores_19_cast_fp16)[name = string("probabilities_27_cast_fp16")];
            bool attention_11_transpose_x_0 = const()[name = string("attention_11_transpose_x_0"), val = bool(false)];
            bool attention_11_transpose_y_0 = const()[name = string("attention_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> attention_11_cast_fp16 = matmul(transpose_x = attention_11_transpose_x_0, transpose_y = attention_11_transpose_y_0, x = probabilities_27_cast_fp16, y = value_45_cast_fp16)[name = string("attention_11_cast_fp16")];
            tensor<int32, [4]> var_1787 = const()[name = string("op_1787"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1792 = const()[name = string("op_1792"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 12, 64]> var_1788_cast_fp16 = transpose(perm = var_1787, x = attention_11_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 1, 768]> input_103_cast_fp16 = reshape(shape = var_1792, x = var_1788_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_25_cast_fp16 = linear(bias = block_attn_c_proj_bias_to_fp16, weight = block_attn_c_proj_weight_to_fp16, x = input_103_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_105_cast_fp16 = add(x = input_99_cast_fp16, y = linear_25_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<int32, [1]> input_107_axes_0 = const()[name = string("input_107_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_1800_to_fp16 = const()[name = string("op_1800_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = block_ln_2_bias_to_fp16, epsilon = var_1800_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_105_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [1, 1, 3072]> linear_26_cast_fp16 = linear(bias = block_mlp_fc_in_bias_to_fp16, weight = block_mlp_fc_in_weight_to_fp16, x = input_107_cast_fp16)[name = string("linear_26_cast_fp16")];
            string input_111_mode_0 = const()[name = string("input_111_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1, 3072]> input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = linear_26_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_27_cast_fp16 = linear(bias = block_mlp_fc_out_bias_to_fp16, weight = block_mlp_fc_out_weight_to_fp16, x = input_111_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_115_cast_fp16 = add(x = input_105_cast_fp16, y = linear_27_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<int32, [1]> var_1839_axes_0 = const()[name = string("op_1839_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_1834_to_fp16 = const()[name = string("op_1834_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> var_1839_cast_fp16 = layer_norm(axes = var_1839_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_1834_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_115_cast_fp16)[name = string("op_1839_cast_fp16")];
            tensor<int32, [3]> var_1847_begin_0 = const()[name = string("op_1847_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_1847_end_0 = const()[name = string("op_1847_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_1847_end_mask_0 = const()[name = string("op_1847_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_1847_squeeze_mask_0 = const()[name = string("op_1847_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_1847_cast_fp16 = slice_by_index(begin = var_1847_begin_0, end = var_1847_end_0, end_mask = var_1847_end_mask_0, squeeze_mask = var_1847_squeeze_mask_0, x = var_1839_cast_fp16)[name = string("op_1847_cast_fp16")];
            tensor<fp16, [1024, 768]> audio_embeddings_4_weight_to_fp16 = const()[name = string("audio_embeddings_4_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20478848)))];
            tensor<fp16, [1, 1024]> linear_28_cast_fp16 = linear(bias = linear_8_bias_0_to_fp16, weight = audio_embeddings_4_weight_to_fp16, x = var_1847_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [3]> var_1862_begin_0 = const()[name = string("op_1862_begin_0"), val = tensor<int32, [3]>([0, 4, 0])];
            tensor<int32, [3]> var_1862_end_0 = const()[name = string("op_1862_end_0"), val = tensor<int32, [3]>([1, 5, 1024])];
            tensor<bool, [3]> var_1862_end_mask_0 = const()[name = string("op_1862_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_1862_squeeze_mask_0 = const()[name = string("op_1862_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<int32, [1, 1024]> var_1862 = slice_by_index(begin = var_1862_begin_0, end = var_1862_end_0, end_mask = var_1862_end_mask_0, squeeze_mask = var_1862_squeeze_mask_0, x = repetition_seen)[name = string("op_1862")];
            tensor<int32, [2]> random_u_9_begin_0 = const()[name = string("random_u_9_begin_0"), val = tensor<int32, [2]>([0, 4])];
            tensor<int32, [2]> random_u_9_end_0 = const()[name = string("random_u_9_end_0"), val = tensor<int32, [2]>([1, 5])];
            tensor<bool, [2]> random_u_9_end_mask_0 = const()[name = string("random_u_9_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<bool, [2]> random_u_9_squeeze_mask_0 = const()[name = string("random_u_9_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [1]> random_u_9_cast_fp16 = slice_by_index(begin = random_u_9_begin_0, end = random_u_9_end_0, end_mask = random_u_9_end_mask_0, squeeze_mask = random_u_9_squeeze_mask_0, x = audio_random_to_fp16)[name = string("random_u_9_cast_fp16")];
            fp16 var_1876_promoted_to_fp16 = const()[name = string("op_1876_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [1, 1024]> var_1877_cast_fp16 = less(x = linear_28_cast_fp16, y = var_1876_promoted_to_fp16)[name = string("op_1877_cast_fp16")];
            fp16 var_1878_to_fp16 = const()[name = string("op_1878_to_fp16"), val = fp16(0x1.334p+0)];
            tensor<fp16, [1, 1024]> var_1879_cast_fp16 = mul(x = linear_28_cast_fp16, y = var_1878_to_fp16)[name = string("op_1879_cast_fp16")];
            fp16 _inversed_1881_y_0_to_fp16 = const()[name = string("_inversed_1881_y_0_to_fp16"), val = fp16(0x1.aacp-1)];
            tensor<fp16, [1, 1024]> _inversed_1881_cast_fp16 = mul(x = linear_28_cast_fp16, y = _inversed_1881_y_0_to_fp16)[name = string("_inversed_1881_cast_fp16")];
            tensor<fp16, [1, 1024]> penalized_9_cast_fp16 = select(a = var_1879_cast_fp16, b = _inversed_1881_cast_fp16, cond = var_1877_cast_fp16)[name = string("penalized_9_cast_fp16")];
            string var_1887_dtype_0 = const()[name = string("op_1887_dtype_0"), val = string("bool")];
            tensor<bool, [1, 1024]> var_1887 = cast(dtype = var_1887_dtype_0, x = var_1862)[name = string("cast_354")];
            tensor<fp16, [1, 1024]> var_1888_cast_fp16 = select(a = penalized_9_cast_fp16, b = linear_28_cast_fp16, cond = var_1887)[name = string("op_1888_cast_fp16")];
            fp16 _inversed_scores_21_y_0_to_fp16 = const()[name = string("_inversed_scores_21_y_0_to_fp16"), val = fp16(0x1.4p+0)];
            tensor<fp16, [1, 1024]> _inversed_scores_21_cast_fp16 = mul(x = var_1888_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = string("_inversed_scores_21_cast_fp16")];
            int32 var_1891 = const()[name = string("op_1891"), val = int32(25)];
            int32 top_scores_17_axis_0 = const()[name = string("top_scores_17_axis_0"), val = int32(1)];
            bool top_scores_17_ascending_0 = const()[name = string("top_scores_17_ascending_0"), val = bool(false)];
            bool top_scores_17_sort_0 = const()[name = string("top_scores_17_sort_0"), val = bool(true)];
            bool top_scores_17_return_indices_0 = const()[name = string("top_scores_17_return_indices_0"), val = bool(true)];
            string top_scores_17_cast_fp16_cast_uint16_output_indices_dtype_0 = const()[name = string("top_scores_17_cast_fp16_cast_uint16_output_indices_dtype_0"), val = string("uint16")];
            tensor<fp16, [1, 25]> top_scores_17_cast_fp16_cast_uint16_0, tensor<uint16, [1, 25]> top_scores_17_cast_fp16_cast_uint16_1 = topk(ascending = top_scores_17_ascending_0, axis = top_scores_17_axis_0, k = var_1891, output_indices_dtype = top_scores_17_cast_fp16_cast_uint16_output_indices_dtype_0, return_indices = top_scores_17_return_indices_0, sort = top_scores_17_sort_0, x = _inversed_scores_21_cast_fp16)[name = string("top_scores_17_cast_fp16_cast_uint16")];
            int32 var_1897 = const()[name = string("op_1897"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_29_cast_fp16 = softmax(axis = var_1897, x = top_scores_17_cast_fp16_cast_uint16_0)[name = string("probabilities_29_cast_fp16")];
            int32 var_1900 = const()[name = string("op_1900"), val = int32(1)];
            bool cumulative_17_exclusive_0 = const()[name = string("cumulative_17_exclusive_0"), val = bool(false)];
            bool cumulative_17_reverse_0 = const()[name = string("cumulative_17_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_17_cast_fp16 = cumsum(axis = var_1900, exclusive = cumulative_17_exclusive_0, reverse = cumulative_17_reverse_0, x = probabilities_29_cast_fp16)[name = string("cumulative_17_cast_fp16")];
            tensor<fp16, [1, 25]> var_1904_cast_fp16 = sub(x = cumulative_17_cast_fp16, y = probabilities_29_cast_fp16)[name = string("op_1904_cast_fp16")];
            fp16 var_1905_to_fp16 = const()[name = string("op_1905_to_fp16"), val = fp16(0x1.e68p-1)];
            tensor<bool, [1, 25]> keep_9_cast_fp16 = less(x = var_1904_cast_fp16, y = var_1905_to_fp16)[name = string("keep_9_cast_fp16")];
            fp16 var_1913_value_0_to_fp16 = const()[name = string("op_1913_value_0_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 25]> var_1913_cast_fp16 = fill_like(ref_tensor = top_scores_17_cast_fp16_cast_uint16_0, value = var_1913_value_0_to_fp16)[name = string("op_1913_cast_fp16")];
            tensor<fp16, [1, 25]> top_scores_19_cast_fp16 = select(a = top_scores_17_cast_fp16_cast_uint16_0, b = var_1913_cast_fp16, cond = keep_9_cast_fp16)[name = string("top_scores_19_cast_fp16")];
            int32 var_1915 = const()[name = string("op_1915"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_31_cast_fp16 = softmax(axis = var_1915, x = top_scores_19_cast_fp16)[name = string("probabilities_31_cast_fp16")];
            int32 var_1918 = const()[name = string("op_1918"), val = int32(1)];
            bool cumulative_19_exclusive_0 = const()[name = string("cumulative_19_exclusive_0"), val = bool(false)];
            bool cumulative_19_reverse_0 = const()[name = string("cumulative_19_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_19_cast_fp16 = cumsum(axis = var_1918, exclusive = cumulative_19_exclusive_0, reverse = cumulative_19_reverse_0, x = probabilities_31_cast_fp16)[name = string("cumulative_19_cast_fp16")];
            tensor<int32, [2]> var_1923 = const()[name = string("op_1923"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_1924_cast_fp16 = reshape(shape = var_1923, x = random_u_9_cast_fp16)[name = string("op_1924_cast_fp16")];
            tensor<bool, [1, 25]> var_1925_cast_fp16 = less(x = cumulative_19_cast_fp16, y = var_1924_cast_fp16)[name = string("op_1925_cast_fp16")];
            string var_1930_dtype_0 = const()[name = string("op_1930_dtype_0"), val = string("int32")];
            tensor<int32, [1]> selected_17_axes_0 = const()[name = string("selected_17_axes_0"), val = tensor<int32, [1]>([1])];
            bool selected_17_keep_dims_0 = const()[name = string("selected_17_keep_dims_0"), val = bool(false)];
            tensor<int32, [1, 25]> var_1930 = cast(dtype = var_1930_dtype_0, x = var_1925_cast_fp16)[name = string("cast_353")];
            tensor<int32, [1]> selected_17 = reduce_sum(axes = selected_17_axes_0, keep_dims = selected_17_keep_dims_0, x = var_1930)[name = string("selected_17")];
            string selected_17_promoted_to_fp16_dtype_0 = const()[name = string("selected_17_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_28_to_fp16 = const()[name = string("const_28_to_fp16"), val = fp16(-inf)];
            fp16 var_1937_promoted_to_fp16 = const()[name = string("op_1937_promoted_to_fp16"), val = fp16(0x1.8p+4)];
            tensor<fp16, [1]> selected_17_to_fp16 = cast(dtype = selected_17_promoted_to_fp16_dtype_0, x = selected_17)[name = string("cast_352")];
            tensor<fp16, [1]> clip_4_cast_fp16 = clip(alpha = const_28_to_fp16, beta = var_1937_promoted_to_fp16, x = selected_17_to_fp16)[name = string("clip_4_cast_fp16")];
            string selected_19_dtype_0 = const()[name = string("selected_19_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_1946 = const()[name = string("op_1946"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> clip_4_cast_fp16_to_int32 = cast(dtype = selected_19_dtype_0, x = clip_4_cast_fp16)[name = string("cast_351")];
            tensor<int32, [1, 1]> var_1947 = reshape(shape = var_1946, x = clip_4_cast_fp16_to_int32)[name = string("op_1947")];
            int32 var_1948 = const()[name = string("op_1948"), val = int32(1)];
            bool var_1950_validate_indices_0 = const()[name = string("op_1950_validate_indices_0"), val = bool(false)];
            string var_1947_to_uint16_dtype_0 = const()[name = string("op_1947_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 1]> var_1947_to_uint16 = cast(dtype = var_1947_to_uint16_dtype_0, x = var_1947)[name = string("cast_350")];
            tensor<uint16, [1, 1]> var_1950_cast_uint16 = gather_along_axis(axis = var_1948, indices = var_1947_to_uint16, validate_indices = var_1950_validate_indices_0, x = top_scores_17_cast_fp16_cast_uint16_1)[name = string("op_1950_cast_uint16")];
            string var_1950_cast_uint16_to_int32_dtype_0 = const()[name = string("op_1950_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [1]> var_1952 = const()[name = string("op_1952"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_1950_cast_uint16_to_int32 = cast(dtype = var_1950_cast_uint16_to_int32_dtype_0, x = var_1950_cast_uint16)[name = string("cast_349")];
            tensor<int32, [1]> token_9 = reshape(shape = var_1952, x = var_1950_cast_uint16_to_int32)[name = string("token_9")];
            int32 var_1962_axis_0 = const()[name = string("op_1962_axis_0"), val = int32(0)];
            int32 var_1962_batch_dims_0 = const()[name = string("op_1962_batch_dims_0"), val = int32(0)];
            bool var_1962_validate_indices_0 = const()[name = string("op_1962_validate_indices_0"), val = bool(false)];
            string token_9_to_uint16_dtype_0 = const()[name = string("token_9_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> token_9_to_uint16 = cast(dtype = token_9_to_uint16_dtype_0, x = token_9)[name = string("cast_348")];
            tensor<fp16, [1, 768]> var_1962_cast_fp16_cast_uint16 = gather(axis = var_1962_axis_0, batch_dims = var_1962_batch_dims_0, indices = token_9_to_uint16, validate_indices = var_1962_validate_indices_0, x = audio_embeddings_4_weight_to_fp16)[name = string("op_1962_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 768]> next_global_embed_11_cast_fp16 = add(x = next_global_embed_9_cast_fp16, y = var_1962_cast_fp16_cast_uint16)[name = string("next_global_embed_11_cast_fp16")];
            tensor<int32, [3]> var_1972 = const()[name = string("op_1972"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 768]> input_119_cast_fp16 = reshape(shape = var_1972, x = var_1962_cast_fp16_cast_uint16)[name = string("input_119_cast_fp16")];
            tensor<int32, [1]> input_121_axes_0 = const()[name = string("input_121_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_1975_to_fp16 = const()[name = string("op_1975_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_121_cast_fp16 = layer_norm(axes = input_121_axes_0, beta = block_ln_1_bias_to_fp16, epsilon = var_1975_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_119_cast_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [1, 1, 2304]> linear_29_cast_fp16 = linear(bias = block_attn_c_attn_bias_to_fp16, weight = block_attn_c_attn_weight_to_fp16, x = input_121_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [3]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [3]>([768, 768, 768])];
            int32 var_1986_axis_0 = const()[name = string("op_1986_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 768]> var_1986_cast_fp16_0, tensor<fp16, [1, 1, 768]> var_1986_cast_fp16_1, tensor<fp16, [1, 1, 768]> var_1986_cast_fp16_2 = split(axis = var_1986_axis_0, split_sizes = tile_6, x = linear_29_cast_fp16)[name = string("op_1986_cast_fp16")];
            tensor<int32, [4]> var_1994 = const()[name = string("op_1994"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> query_51_cast_fp16 = reshape(shape = var_1994, x = var_1986_cast_fp16_0)[name = string("query_51_cast_fp16")];
            tensor<int32, [4]> var_2000 = const()[name = string("op_2000"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> key_61_cast_fp16 = reshape(shape = var_2000, x = var_1986_cast_fp16_1)[name = string("key_61_cast_fp16")];
            tensor<int32, [4]> var_2006 = const()[name = string("op_2006"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> value_49_cast_fp16 = reshape(shape = var_2006, x = var_1986_cast_fp16_2)[name = string("value_49_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_cos_6_to_fp16 = const()[name = string("rope_cos_6_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22051776)))];
            tensor<fp16, [1, 1, 12, 64]> var_2008_cast_fp16 = mul(x = query_51_cast_fp16, y = rope_cos_6_to_fp16)[name = string("op_2008_cast_fp16")];
            tensor<int32, [4]> var_2013_begin_0 = const()[name = string("op_2013_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_2013_end_0 = const()[name = string("op_2013_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_2013_end_mask_0 = const()[name = string("op_2013_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2013_stride_0 = const()[name = string("op_2013_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_2013_cast_fp16 = slice_by_index(begin = var_2013_begin_0, end = var_2013_end_0, end_mask = var_2013_end_mask_0, stride = var_2013_stride_0, x = query_51_cast_fp16)[name = string("op_2013_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_2014_cast_fp16 = mul(x = var_2013_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_2014_cast_fp16")];
            tensor<int32, [4]> var_2019_begin_0 = const()[name = string("op_2019_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2019_end_0 = const()[name = string("op_2019_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_2019_end_mask_0 = const()[name = string("op_2019_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2019_stride_0 = const()[name = string("op_2019_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_2019_cast_fp16 = slice_by_index(begin = var_2019_begin_0, end = var_2019_end_0, end_mask = var_2019_end_mask_0, stride = var_2019_stride_0, x = query_51_cast_fp16)[name = string("op_2019_cast_fp16")];
            int32 var_2022_axis_0 = const()[name = string("op_2022_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_2022_cast_fp16 = stack(axis = var_2022_axis_0, values = (var_2014_cast_fp16, var_2019_cast_fp16))[name = string("op_2022_cast_fp16")];
            tensor<int32, [4]> shape_19 = const()[name = string("shape_19"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_2023_cast_fp16 = reshape(shape = shape_19, x = var_2022_cast_fp16)[name = string("op_2023_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_sin_6_to_fp16 = const()[name = string("rope_sin_6_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22051968)))];
            tensor<fp16, [1, 1, 12, 64]> var_2024_cast_fp16 = mul(x = var_2023_cast_fp16, y = rope_sin_6_to_fp16)[name = string("op_2024_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> query_53_cast_fp16 = add(x = var_2008_cast_fp16, y = var_2024_cast_fp16)[name = string("query_53_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_2027_cast_fp16 = mul(x = key_61_cast_fp16, y = rope_cos_6_to_fp16)[name = string("op_2027_cast_fp16")];
            tensor<int32, [4]> var_2032_begin_0 = const()[name = string("op_2032_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_2032_end_0 = const()[name = string("op_2032_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_2032_end_mask_0 = const()[name = string("op_2032_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2032_stride_0 = const()[name = string("op_2032_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_2032_cast_fp16 = slice_by_index(begin = var_2032_begin_0, end = var_2032_end_0, end_mask = var_2032_end_mask_0, stride = var_2032_stride_0, x = key_61_cast_fp16)[name = string("op_2032_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_2033_cast_fp16 = mul(x = var_2032_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_2033_cast_fp16")];
            tensor<int32, [4]> var_2038_begin_0 = const()[name = string("op_2038_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2038_end_0 = const()[name = string("op_2038_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_2038_end_mask_0 = const()[name = string("op_2038_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2038_stride_0 = const()[name = string("op_2038_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_2038_cast_fp16 = slice_by_index(begin = var_2038_begin_0, end = var_2038_end_0, end_mask = var_2038_end_mask_0, stride = var_2038_stride_0, x = key_61_cast_fp16)[name = string("op_2038_cast_fp16")];
            int32 var_2041_axis_0 = const()[name = string("op_2041_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_2041_cast_fp16 = stack(axis = var_2041_axis_0, values = (var_2033_cast_fp16, var_2038_cast_fp16))[name = string("op_2041_cast_fp16")];
            tensor<int32, [4]> shape_20 = const()[name = string("shape_20"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_2042_cast_fp16 = reshape(shape = shape_20, x = var_2041_cast_fp16)[name = string("op_2042_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_2043_cast_fp16 = mul(x = var_2042_cast_fp16, y = rope_sin_6_to_fp16)[name = string("op_2043_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> key_63_cast_fp16 = add(x = var_2027_cast_fp16, y = var_2043_cast_fp16)[name = string("key_63_cast_fp16")];
            tensor<int32, [4]> var_2050 = const()[name = string("op_2050"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2056 = const()[name = string("op_2056"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2062 = const()[name = string("op_2062"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 var_2065 = const()[name = string("op_2065"), val = int32(2)];
            bool key_67_interleave_0 = const()[name = string("key_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> key_65_cast_fp16 = transpose(perm = var_2056, x = key_63_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 12, 7, 64]> key_67_cast_fp16 = concat(axis = var_2065, interleave = key_67_interleave_0, values = (key_57_cast_fp16, key_65_cast_fp16))[name = string("key_67_cast_fp16")];
            int32 var_2068 = const()[name = string("op_2068"), val = int32(2)];
            bool value_53_interleave_0 = const()[name = string("value_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> value_51_cast_fp16 = transpose(perm = var_2062, x = value_49_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 12, 7, 64]> value_53_cast_fp16 = concat(axis = var_2068, interleave = value_53_interleave_0, values = (value_45_cast_fp16, value_51_cast_fp16))[name = string("value_53_cast_fp16")];
            bool var_2073_transpose_x_1 = const()[name = string("op_2073_transpose_x_1"), val = bool(false)];
            bool var_2073_transpose_y_1 = const()[name = string("op_2073_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 12, 1, 64]> query_55_cast_fp16 = transpose(perm = var_2050, x = query_53_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 12, 1, 7]> var_2073_cast_fp16 = matmul(transpose_x = var_2073_transpose_x_1, transpose_y = var_2073_transpose_y_1, x = query_55_cast_fp16, y = key_67_cast_fp16)[name = string("op_2073_cast_fp16")];
            fp16 var_2074_to_fp16 = const()[name = string("op_2074_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 1, 7]> scores_23_cast_fp16 = mul(x = var_2073_cast_fp16, y = var_2074_to_fp16)[name = string("scores_23_cast_fp16")];
            int32 var_2076 = const()[name = string("op_2076"), val = int32(-1)];
            tensor<fp16, [1, 12, 1, 7]> probabilities_33_cast_fp16 = softmax(axis = var_2076, x = scores_23_cast_fp16)[name = string("probabilities_33_cast_fp16")];
            bool attention_13_transpose_x_0 = const()[name = string("attention_13_transpose_x_0"), val = bool(false)];
            bool attention_13_transpose_y_0 = const()[name = string("attention_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> attention_13_cast_fp16 = matmul(transpose_x = attention_13_transpose_x_0, transpose_y = attention_13_transpose_y_0, x = probabilities_33_cast_fp16, y = value_53_cast_fp16)[name = string("attention_13_cast_fp16")];
            tensor<int32, [4]> var_2084 = const()[name = string("op_2084"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2089 = const()[name = string("op_2089"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 12, 64]> var_2085_cast_fp16 = transpose(perm = var_2084, x = attention_13_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 1, 768]> input_123_cast_fp16 = reshape(shape = var_2089, x = var_2085_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_30_cast_fp16 = linear(bias = block_attn_c_proj_bias_to_fp16, weight = block_attn_c_proj_weight_to_fp16, x = input_123_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_125_cast_fp16 = add(x = input_119_cast_fp16, y = linear_30_cast_fp16)[name = string("input_125_cast_fp16")];
            tensor<int32, [1]> input_127_axes_0 = const()[name = string("input_127_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_2097_to_fp16 = const()[name = string("op_2097_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_127_cast_fp16 = layer_norm(axes = input_127_axes_0, beta = block_ln_2_bias_to_fp16, epsilon = var_2097_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_125_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [1, 1, 3072]> linear_31_cast_fp16 = linear(bias = block_mlp_fc_in_bias_to_fp16, weight = block_mlp_fc_in_weight_to_fp16, x = input_127_cast_fp16)[name = string("linear_31_cast_fp16")];
            string input_131_mode_0 = const()[name = string("input_131_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1, 3072]> input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = linear_31_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_32_cast_fp16 = linear(bias = block_mlp_fc_out_bias_to_fp16, weight = block_mlp_fc_out_weight_to_fp16, x = input_131_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_135_cast_fp16 = add(x = input_125_cast_fp16, y = linear_32_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<int32, [1]> var_2136_axes_0 = const()[name = string("op_2136_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_2131_to_fp16 = const()[name = string("op_2131_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> var_2136_cast_fp16 = layer_norm(axes = var_2136_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_2131_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_135_cast_fp16)[name = string("op_2136_cast_fp16")];
            tensor<int32, [3]> var_2144_begin_0 = const()[name = string("op_2144_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2144_end_0 = const()[name = string("op_2144_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_2144_end_mask_0 = const()[name = string("op_2144_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_2144_squeeze_mask_0 = const()[name = string("op_2144_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_2144_cast_fp16 = slice_by_index(begin = var_2144_begin_0, end = var_2144_end_0, end_mask = var_2144_end_mask_0, squeeze_mask = var_2144_squeeze_mask_0, x = var_2136_cast_fp16)[name = string("op_2144_cast_fp16")];
            tensor<fp16, [1024, 768]> audio_embeddings_5_weight_to_fp16 = const()[name = string("audio_embeddings_5_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22052160)))];
            tensor<fp16, [1, 1024]> linear_33_cast_fp16 = linear(bias = linear_8_bias_0_to_fp16, weight = audio_embeddings_5_weight_to_fp16, x = var_2144_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<int32, [3]> var_2159_begin_0 = const()[name = string("op_2159_begin_0"), val = tensor<int32, [3]>([0, 5, 0])];
            tensor<int32, [3]> var_2159_end_0 = const()[name = string("op_2159_end_0"), val = tensor<int32, [3]>([1, 6, 1024])];
            tensor<bool, [3]> var_2159_end_mask_0 = const()[name = string("op_2159_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_2159_squeeze_mask_0 = const()[name = string("op_2159_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<int32, [1, 1024]> var_2159 = slice_by_index(begin = var_2159_begin_0, end = var_2159_end_0, end_mask = var_2159_end_mask_0, squeeze_mask = var_2159_squeeze_mask_0, x = repetition_seen)[name = string("op_2159")];
            tensor<int32, [2]> random_u_11_begin_0 = const()[name = string("random_u_11_begin_0"), val = tensor<int32, [2]>([0, 5])];
            tensor<int32, [2]> random_u_11_end_0 = const()[name = string("random_u_11_end_0"), val = tensor<int32, [2]>([1, 6])];
            tensor<bool, [2]> random_u_11_end_mask_0 = const()[name = string("random_u_11_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<bool, [2]> random_u_11_squeeze_mask_0 = const()[name = string("random_u_11_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [1]> random_u_11_cast_fp16 = slice_by_index(begin = random_u_11_begin_0, end = random_u_11_end_0, end_mask = random_u_11_end_mask_0, squeeze_mask = random_u_11_squeeze_mask_0, x = audio_random_to_fp16)[name = string("random_u_11_cast_fp16")];
            fp16 var_2173_promoted_to_fp16 = const()[name = string("op_2173_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [1, 1024]> var_2174_cast_fp16 = less(x = linear_33_cast_fp16, y = var_2173_promoted_to_fp16)[name = string("op_2174_cast_fp16")];
            fp16 var_2175_to_fp16 = const()[name = string("op_2175_to_fp16"), val = fp16(0x1.334p+0)];
            tensor<fp16, [1, 1024]> var_2176_cast_fp16 = mul(x = linear_33_cast_fp16, y = var_2175_to_fp16)[name = string("op_2176_cast_fp16")];
            fp16 _inversed_2178_y_0_to_fp16 = const()[name = string("_inversed_2178_y_0_to_fp16"), val = fp16(0x1.aacp-1)];
            tensor<fp16, [1, 1024]> _inversed_2178_cast_fp16 = mul(x = linear_33_cast_fp16, y = _inversed_2178_y_0_to_fp16)[name = string("_inversed_2178_cast_fp16")];
            tensor<fp16, [1, 1024]> penalized_11_cast_fp16 = select(a = var_2176_cast_fp16, b = _inversed_2178_cast_fp16, cond = var_2174_cast_fp16)[name = string("penalized_11_cast_fp16")];
            string var_2184_dtype_0 = const()[name = string("op_2184_dtype_0"), val = string("bool")];
            tensor<bool, [1, 1024]> var_2184 = cast(dtype = var_2184_dtype_0, x = var_2159)[name = string("cast_347")];
            tensor<fp16, [1, 1024]> var_2185_cast_fp16 = select(a = penalized_11_cast_fp16, b = linear_33_cast_fp16, cond = var_2184)[name = string("op_2185_cast_fp16")];
            fp16 _inversed_scores_25_y_0_to_fp16 = const()[name = string("_inversed_scores_25_y_0_to_fp16"), val = fp16(0x1.4p+0)];
            tensor<fp16, [1, 1024]> _inversed_scores_25_cast_fp16 = mul(x = var_2185_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = string("_inversed_scores_25_cast_fp16")];
            int32 var_2188 = const()[name = string("op_2188"), val = int32(25)];
            int32 top_scores_21_axis_0 = const()[name = string("top_scores_21_axis_0"), val = int32(1)];
            bool top_scores_21_ascending_0 = const()[name = string("top_scores_21_ascending_0"), val = bool(false)];
            bool top_scores_21_sort_0 = const()[name = string("top_scores_21_sort_0"), val = bool(true)];
            bool top_scores_21_return_indices_0 = const()[name = string("top_scores_21_return_indices_0"), val = bool(true)];
            string top_scores_21_cast_fp16_cast_uint16_output_indices_dtype_0 = const()[name = string("top_scores_21_cast_fp16_cast_uint16_output_indices_dtype_0"), val = string("uint16")];
            tensor<fp16, [1, 25]> top_scores_21_cast_fp16_cast_uint16_0, tensor<uint16, [1, 25]> top_scores_21_cast_fp16_cast_uint16_1 = topk(ascending = top_scores_21_ascending_0, axis = top_scores_21_axis_0, k = var_2188, output_indices_dtype = top_scores_21_cast_fp16_cast_uint16_output_indices_dtype_0, return_indices = top_scores_21_return_indices_0, sort = top_scores_21_sort_0, x = _inversed_scores_25_cast_fp16)[name = string("top_scores_21_cast_fp16_cast_uint16")];
            int32 var_2194 = const()[name = string("op_2194"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_35_cast_fp16 = softmax(axis = var_2194, x = top_scores_21_cast_fp16_cast_uint16_0)[name = string("probabilities_35_cast_fp16")];
            int32 var_2197 = const()[name = string("op_2197"), val = int32(1)];
            bool cumulative_21_exclusive_0 = const()[name = string("cumulative_21_exclusive_0"), val = bool(false)];
            bool cumulative_21_reverse_0 = const()[name = string("cumulative_21_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_21_cast_fp16 = cumsum(axis = var_2197, exclusive = cumulative_21_exclusive_0, reverse = cumulative_21_reverse_0, x = probabilities_35_cast_fp16)[name = string("cumulative_21_cast_fp16")];
            tensor<fp16, [1, 25]> var_2201_cast_fp16 = sub(x = cumulative_21_cast_fp16, y = probabilities_35_cast_fp16)[name = string("op_2201_cast_fp16")];
            fp16 var_2202_to_fp16 = const()[name = string("op_2202_to_fp16"), val = fp16(0x1.e68p-1)];
            tensor<bool, [1, 25]> keep_11_cast_fp16 = less(x = var_2201_cast_fp16, y = var_2202_to_fp16)[name = string("keep_11_cast_fp16")];
            fp16 var_2210_value_0_to_fp16 = const()[name = string("op_2210_value_0_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 25]> var_2210_cast_fp16 = fill_like(ref_tensor = top_scores_21_cast_fp16_cast_uint16_0, value = var_2210_value_0_to_fp16)[name = string("op_2210_cast_fp16")];
            tensor<fp16, [1, 25]> top_scores_23_cast_fp16 = select(a = top_scores_21_cast_fp16_cast_uint16_0, b = var_2210_cast_fp16, cond = keep_11_cast_fp16)[name = string("top_scores_23_cast_fp16")];
            int32 var_2212 = const()[name = string("op_2212"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_37_cast_fp16 = softmax(axis = var_2212, x = top_scores_23_cast_fp16)[name = string("probabilities_37_cast_fp16")];
            int32 var_2215 = const()[name = string("op_2215"), val = int32(1)];
            bool cumulative_23_exclusive_0 = const()[name = string("cumulative_23_exclusive_0"), val = bool(false)];
            bool cumulative_23_reverse_0 = const()[name = string("cumulative_23_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_23_cast_fp16 = cumsum(axis = var_2215, exclusive = cumulative_23_exclusive_0, reverse = cumulative_23_reverse_0, x = probabilities_37_cast_fp16)[name = string("cumulative_23_cast_fp16")];
            tensor<int32, [2]> var_2220 = const()[name = string("op_2220"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_2221_cast_fp16 = reshape(shape = var_2220, x = random_u_11_cast_fp16)[name = string("op_2221_cast_fp16")];
            tensor<bool, [1, 25]> var_2222_cast_fp16 = less(x = cumulative_23_cast_fp16, y = var_2221_cast_fp16)[name = string("op_2222_cast_fp16")];
            string var_2227_dtype_0 = const()[name = string("op_2227_dtype_0"), val = string("int32")];
            tensor<int32, [1]> selected_21_axes_0 = const()[name = string("selected_21_axes_0"), val = tensor<int32, [1]>([1])];
            bool selected_21_keep_dims_0 = const()[name = string("selected_21_keep_dims_0"), val = bool(false)];
            tensor<int32, [1, 25]> var_2227 = cast(dtype = var_2227_dtype_0, x = var_2222_cast_fp16)[name = string("cast_346")];
            tensor<int32, [1]> selected_21 = reduce_sum(axes = selected_21_axes_0, keep_dims = selected_21_keep_dims_0, x = var_2227)[name = string("selected_21")];
            string selected_21_promoted_to_fp16_dtype_0 = const()[name = string("selected_21_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_33_to_fp16 = const()[name = string("const_33_to_fp16"), val = fp16(-inf)];
            fp16 var_2234_promoted_to_fp16 = const()[name = string("op_2234_promoted_to_fp16"), val = fp16(0x1.8p+4)];
            tensor<fp16, [1]> selected_21_to_fp16 = cast(dtype = selected_21_promoted_to_fp16_dtype_0, x = selected_21)[name = string("cast_345")];
            tensor<fp16, [1]> clip_5_cast_fp16 = clip(alpha = const_33_to_fp16, beta = var_2234_promoted_to_fp16, x = selected_21_to_fp16)[name = string("clip_5_cast_fp16")];
            string selected_23_dtype_0 = const()[name = string("selected_23_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_2243 = const()[name = string("op_2243"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> clip_5_cast_fp16_to_int32 = cast(dtype = selected_23_dtype_0, x = clip_5_cast_fp16)[name = string("cast_344")];
            tensor<int32, [1, 1]> var_2244 = reshape(shape = var_2243, x = clip_5_cast_fp16_to_int32)[name = string("op_2244")];
            int32 var_2245 = const()[name = string("op_2245"), val = int32(1)];
            bool var_2247_validate_indices_0 = const()[name = string("op_2247_validate_indices_0"), val = bool(false)];
            string var_2244_to_uint16_dtype_0 = const()[name = string("op_2244_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 1]> var_2244_to_uint16 = cast(dtype = var_2244_to_uint16_dtype_0, x = var_2244)[name = string("cast_343")];
            tensor<uint16, [1, 1]> var_2247_cast_uint16 = gather_along_axis(axis = var_2245, indices = var_2244_to_uint16, validate_indices = var_2247_validate_indices_0, x = top_scores_21_cast_fp16_cast_uint16_1)[name = string("op_2247_cast_uint16")];
            string var_2247_cast_uint16_to_int32_dtype_0 = const()[name = string("op_2247_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [1]> var_2249 = const()[name = string("op_2249"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_2247_cast_uint16_to_int32 = cast(dtype = var_2247_cast_uint16_to_int32_dtype_0, x = var_2247_cast_uint16)[name = string("cast_342")];
            tensor<int32, [1]> token_11 = reshape(shape = var_2249, x = var_2247_cast_uint16_to_int32)[name = string("token_11")];
            int32 var_2259_axis_0 = const()[name = string("op_2259_axis_0"), val = int32(0)];
            int32 var_2259_batch_dims_0 = const()[name = string("op_2259_batch_dims_0"), val = int32(0)];
            bool var_2259_validate_indices_0 = const()[name = string("op_2259_validate_indices_0"), val = bool(false)];
            string token_11_to_uint16_dtype_0 = const()[name = string("token_11_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> token_11_to_uint16 = cast(dtype = token_11_to_uint16_dtype_0, x = token_11)[name = string("cast_341")];
            tensor<fp16, [1, 768]> var_2259_cast_fp16_cast_uint16 = gather(axis = var_2259_axis_0, batch_dims = var_2259_batch_dims_0, indices = token_11_to_uint16, validate_indices = var_2259_validate_indices_0, x = audio_embeddings_5_weight_to_fp16)[name = string("op_2259_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 768]> next_global_embed_13_cast_fp16 = add(x = next_global_embed_11_cast_fp16, y = var_2259_cast_fp16_cast_uint16)[name = string("next_global_embed_13_cast_fp16")];
            tensor<int32, [3]> var_2269 = const()[name = string("op_2269"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 768]> input_139_cast_fp16 = reshape(shape = var_2269, x = var_2259_cast_fp16_cast_uint16)[name = string("input_139_cast_fp16")];
            tensor<int32, [1]> input_141_axes_0 = const()[name = string("input_141_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_2272_to_fp16 = const()[name = string("op_2272_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_141_cast_fp16 = layer_norm(axes = input_141_axes_0, beta = block_ln_1_bias_to_fp16, epsilon = var_2272_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_139_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [1, 1, 2304]> linear_34_cast_fp16 = linear(bias = block_attn_c_attn_bias_to_fp16, weight = block_attn_c_attn_weight_to_fp16, x = input_141_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<int32, [3]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [3]>([768, 768, 768])];
            int32 var_2283_axis_0 = const()[name = string("op_2283_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 768]> var_2283_cast_fp16_0, tensor<fp16, [1, 1, 768]> var_2283_cast_fp16_1, tensor<fp16, [1, 1, 768]> var_2283_cast_fp16_2 = split(axis = var_2283_axis_0, split_sizes = tile_7, x = linear_34_cast_fp16)[name = string("op_2283_cast_fp16")];
            tensor<int32, [4]> var_2291 = const()[name = string("op_2291"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> query_59_cast_fp16 = reshape(shape = var_2291, x = var_2283_cast_fp16_0)[name = string("query_59_cast_fp16")];
            tensor<int32, [4]> var_2297 = const()[name = string("op_2297"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> key_71_cast_fp16 = reshape(shape = var_2297, x = var_2283_cast_fp16_1)[name = string("key_71_cast_fp16")];
            tensor<int32, [4]> var_2303 = const()[name = string("op_2303"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> value_57_cast_fp16 = reshape(shape = var_2303, x = var_2283_cast_fp16_2)[name = string("value_57_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_cos_7_to_fp16 = const()[name = string("rope_cos_7_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23625088)))];
            tensor<fp16, [1, 1, 12, 64]> var_2305_cast_fp16 = mul(x = query_59_cast_fp16, y = rope_cos_7_to_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [4]> var_2310_begin_0 = const()[name = string("op_2310_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_2310_end_0 = const()[name = string("op_2310_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_2310_end_mask_0 = const()[name = string("op_2310_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2310_stride_0 = const()[name = string("op_2310_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_2310_cast_fp16 = slice_by_index(begin = var_2310_begin_0, end = var_2310_end_0, end_mask = var_2310_end_mask_0, stride = var_2310_stride_0, x = query_59_cast_fp16)[name = string("op_2310_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_2311_cast_fp16 = mul(x = var_2310_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_2311_cast_fp16")];
            tensor<int32, [4]> var_2316_begin_0 = const()[name = string("op_2316_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2316_end_0 = const()[name = string("op_2316_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_2316_end_mask_0 = const()[name = string("op_2316_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2316_stride_0 = const()[name = string("op_2316_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_2316_cast_fp16 = slice_by_index(begin = var_2316_begin_0, end = var_2316_end_0, end_mask = var_2316_end_mask_0, stride = var_2316_stride_0, x = query_59_cast_fp16)[name = string("op_2316_cast_fp16")];
            int32 var_2319_axis_0 = const()[name = string("op_2319_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_2319_cast_fp16 = stack(axis = var_2319_axis_0, values = (var_2311_cast_fp16, var_2316_cast_fp16))[name = string("op_2319_cast_fp16")];
            tensor<int32, [4]> shape_22 = const()[name = string("shape_22"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_2320_cast_fp16 = reshape(shape = shape_22, x = var_2319_cast_fp16)[name = string("op_2320_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_sin_7_to_fp16 = const()[name = string("rope_sin_7_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23625280)))];
            tensor<fp16, [1, 1, 12, 64]> var_2321_cast_fp16 = mul(x = var_2320_cast_fp16, y = rope_sin_7_to_fp16)[name = string("op_2321_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> query_61_cast_fp16 = add(x = var_2305_cast_fp16, y = var_2321_cast_fp16)[name = string("query_61_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_2324_cast_fp16 = mul(x = key_71_cast_fp16, y = rope_cos_7_to_fp16)[name = string("op_2324_cast_fp16")];
            tensor<int32, [4]> var_2329_begin_0 = const()[name = string("op_2329_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_2329_end_0 = const()[name = string("op_2329_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_2329_end_mask_0 = const()[name = string("op_2329_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2329_stride_0 = const()[name = string("op_2329_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_2329_cast_fp16 = slice_by_index(begin = var_2329_begin_0, end = var_2329_end_0, end_mask = var_2329_end_mask_0, stride = var_2329_stride_0, x = key_71_cast_fp16)[name = string("op_2329_cast_fp16")];
            fp16 const_37_promoted_to_fp16 = const()[name = string("const_37_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_2330_cast_fp16 = mul(x = var_2329_cast_fp16, y = const_37_promoted_to_fp16)[name = string("op_2330_cast_fp16")];
            tensor<int32, [4]> var_2335_begin_0 = const()[name = string("op_2335_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2335_end_0 = const()[name = string("op_2335_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_2335_end_mask_0 = const()[name = string("op_2335_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2335_stride_0 = const()[name = string("op_2335_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_2335_cast_fp16 = slice_by_index(begin = var_2335_begin_0, end = var_2335_end_0, end_mask = var_2335_end_mask_0, stride = var_2335_stride_0, x = key_71_cast_fp16)[name = string("op_2335_cast_fp16")];
            int32 var_2338_axis_0 = const()[name = string("op_2338_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_2338_cast_fp16 = stack(axis = var_2338_axis_0, values = (var_2330_cast_fp16, var_2335_cast_fp16))[name = string("op_2338_cast_fp16")];
            tensor<int32, [4]> shape_23 = const()[name = string("shape_23"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_2339_cast_fp16 = reshape(shape = shape_23, x = var_2338_cast_fp16)[name = string("op_2339_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_2340_cast_fp16 = mul(x = var_2339_cast_fp16, y = rope_sin_7_to_fp16)[name = string("op_2340_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> key_73_cast_fp16 = add(x = var_2324_cast_fp16, y = var_2340_cast_fp16)[name = string("key_73_cast_fp16")];
            tensor<int32, [4]> var_2347 = const()[name = string("op_2347"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2353 = const()[name = string("op_2353"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2359 = const()[name = string("op_2359"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 var_2362 = const()[name = string("op_2362"), val = int32(2)];
            bool key_77_interleave_0 = const()[name = string("key_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> key_75_cast_fp16 = transpose(perm = var_2353, x = key_73_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 12, 8, 64]> key_77_cast_fp16 = concat(axis = var_2362, interleave = key_77_interleave_0, values = (key_67_cast_fp16, key_75_cast_fp16))[name = string("key_77_cast_fp16")];
            int32 var_2365 = const()[name = string("op_2365"), val = int32(2)];
            bool value_61_interleave_0 = const()[name = string("value_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> value_59_cast_fp16 = transpose(perm = var_2359, x = value_57_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 12, 8, 64]> value_61_cast_fp16 = concat(axis = var_2365, interleave = value_61_interleave_0, values = (value_53_cast_fp16, value_59_cast_fp16))[name = string("value_61_cast_fp16")];
            bool var_2370_transpose_x_1 = const()[name = string("op_2370_transpose_x_1"), val = bool(false)];
            bool var_2370_transpose_y_1 = const()[name = string("op_2370_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 12, 1, 64]> query_63_cast_fp16 = transpose(perm = var_2347, x = query_61_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 12, 1, 8]> var_2370_cast_fp16 = matmul(transpose_x = var_2370_transpose_x_1, transpose_y = var_2370_transpose_y_1, x = query_63_cast_fp16, y = key_77_cast_fp16)[name = string("op_2370_cast_fp16")];
            fp16 var_2371_to_fp16 = const()[name = string("op_2371_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 1, 8]> scores_27_cast_fp16 = mul(x = var_2370_cast_fp16, y = var_2371_to_fp16)[name = string("scores_27_cast_fp16")];
            int32 var_2373 = const()[name = string("op_2373"), val = int32(-1)];
            tensor<fp16, [1, 12, 1, 8]> probabilities_39_cast_fp16 = softmax(axis = var_2373, x = scores_27_cast_fp16)[name = string("probabilities_39_cast_fp16")];
            bool attention_15_transpose_x_0 = const()[name = string("attention_15_transpose_x_0"), val = bool(false)];
            bool attention_15_transpose_y_0 = const()[name = string("attention_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> attention_15_cast_fp16 = matmul(transpose_x = attention_15_transpose_x_0, transpose_y = attention_15_transpose_y_0, x = probabilities_39_cast_fp16, y = value_61_cast_fp16)[name = string("attention_15_cast_fp16")];
            tensor<int32, [4]> var_2381 = const()[name = string("op_2381"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2386 = const()[name = string("op_2386"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 12, 64]> var_2382_cast_fp16 = transpose(perm = var_2381, x = attention_15_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 1, 768]> input_143_cast_fp16 = reshape(shape = var_2386, x = var_2382_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_35_cast_fp16 = linear(bias = block_attn_c_proj_bias_to_fp16, weight = block_attn_c_proj_weight_to_fp16, x = input_143_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_145_cast_fp16 = add(x = input_139_cast_fp16, y = linear_35_cast_fp16)[name = string("input_145_cast_fp16")];
            tensor<int32, [1]> input_147_axes_0 = const()[name = string("input_147_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_2394_to_fp16 = const()[name = string("op_2394_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = block_ln_2_bias_to_fp16, epsilon = var_2394_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_145_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [1, 1, 3072]> linear_36_cast_fp16 = linear(bias = block_mlp_fc_in_bias_to_fp16, weight = block_mlp_fc_in_weight_to_fp16, x = input_147_cast_fp16)[name = string("linear_36_cast_fp16")];
            string input_151_mode_0 = const()[name = string("input_151_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1, 3072]> input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = linear_36_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_37_cast_fp16 = linear(bias = block_mlp_fc_out_bias_to_fp16, weight = block_mlp_fc_out_weight_to_fp16, x = input_151_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_155_cast_fp16 = add(x = input_145_cast_fp16, y = linear_37_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<int32, [1]> var_2433_axes_0 = const()[name = string("op_2433_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_2428_to_fp16 = const()[name = string("op_2428_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> var_2433_cast_fp16 = layer_norm(axes = var_2433_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_2428_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_155_cast_fp16)[name = string("op_2433_cast_fp16")];
            tensor<int32, [3]> var_2441_begin_0 = const()[name = string("op_2441_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2441_end_0 = const()[name = string("op_2441_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_2441_end_mask_0 = const()[name = string("op_2441_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_2441_squeeze_mask_0 = const()[name = string("op_2441_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_2441_cast_fp16 = slice_by_index(begin = var_2441_begin_0, end = var_2441_end_0, end_mask = var_2441_end_mask_0, squeeze_mask = var_2441_squeeze_mask_0, x = var_2433_cast_fp16)[name = string("op_2441_cast_fp16")];
            tensor<fp16, [1024, 768]> audio_embeddings_6_weight_to_fp16 = const()[name = string("audio_embeddings_6_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23625472)))];
            tensor<fp16, [1, 1024]> linear_38_cast_fp16 = linear(bias = linear_8_bias_0_to_fp16, weight = audio_embeddings_6_weight_to_fp16, x = var_2441_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [3]> var_2456_begin_0 = const()[name = string("op_2456_begin_0"), val = tensor<int32, [3]>([0, 6, 0])];
            tensor<int32, [3]> var_2456_end_0 = const()[name = string("op_2456_end_0"), val = tensor<int32, [3]>([1, 7, 1024])];
            tensor<bool, [3]> var_2456_end_mask_0 = const()[name = string("op_2456_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_2456_squeeze_mask_0 = const()[name = string("op_2456_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<int32, [1, 1024]> var_2456 = slice_by_index(begin = var_2456_begin_0, end = var_2456_end_0, end_mask = var_2456_end_mask_0, squeeze_mask = var_2456_squeeze_mask_0, x = repetition_seen)[name = string("op_2456")];
            tensor<int32, [2]> random_u_13_begin_0 = const()[name = string("random_u_13_begin_0"), val = tensor<int32, [2]>([0, 6])];
            tensor<int32, [2]> random_u_13_end_0 = const()[name = string("random_u_13_end_0"), val = tensor<int32, [2]>([1, 7])];
            tensor<bool, [2]> random_u_13_end_mask_0 = const()[name = string("random_u_13_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<bool, [2]> random_u_13_squeeze_mask_0 = const()[name = string("random_u_13_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [1]> random_u_13_cast_fp16 = slice_by_index(begin = random_u_13_begin_0, end = random_u_13_end_0, end_mask = random_u_13_end_mask_0, squeeze_mask = random_u_13_squeeze_mask_0, x = audio_random_to_fp16)[name = string("random_u_13_cast_fp16")];
            fp16 var_2470_promoted_to_fp16 = const()[name = string("op_2470_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [1, 1024]> var_2471_cast_fp16 = less(x = linear_38_cast_fp16, y = var_2470_promoted_to_fp16)[name = string("op_2471_cast_fp16")];
            fp16 var_2472_to_fp16 = const()[name = string("op_2472_to_fp16"), val = fp16(0x1.334p+0)];
            tensor<fp16, [1, 1024]> var_2473_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_2472_to_fp16)[name = string("op_2473_cast_fp16")];
            fp16 _inversed_2475_y_0_to_fp16 = const()[name = string("_inversed_2475_y_0_to_fp16"), val = fp16(0x1.aacp-1)];
            tensor<fp16, [1, 1024]> _inversed_2475_cast_fp16 = mul(x = linear_38_cast_fp16, y = _inversed_2475_y_0_to_fp16)[name = string("_inversed_2475_cast_fp16")];
            tensor<fp16, [1, 1024]> penalized_13_cast_fp16 = select(a = var_2473_cast_fp16, b = _inversed_2475_cast_fp16, cond = var_2471_cast_fp16)[name = string("penalized_13_cast_fp16")];
            string var_2481_dtype_0 = const()[name = string("op_2481_dtype_0"), val = string("bool")];
            tensor<bool, [1, 1024]> var_2481 = cast(dtype = var_2481_dtype_0, x = var_2456)[name = string("cast_340")];
            tensor<fp16, [1, 1024]> var_2482_cast_fp16 = select(a = penalized_13_cast_fp16, b = linear_38_cast_fp16, cond = var_2481)[name = string("op_2482_cast_fp16")];
            fp16 _inversed_scores_29_y_0_to_fp16 = const()[name = string("_inversed_scores_29_y_0_to_fp16"), val = fp16(0x1.4p+0)];
            tensor<fp16, [1, 1024]> _inversed_scores_29_cast_fp16 = mul(x = var_2482_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = string("_inversed_scores_29_cast_fp16")];
            int32 var_2485 = const()[name = string("op_2485"), val = int32(25)];
            int32 top_scores_25_axis_0 = const()[name = string("top_scores_25_axis_0"), val = int32(1)];
            bool top_scores_25_ascending_0 = const()[name = string("top_scores_25_ascending_0"), val = bool(false)];
            bool top_scores_25_sort_0 = const()[name = string("top_scores_25_sort_0"), val = bool(true)];
            bool top_scores_25_return_indices_0 = const()[name = string("top_scores_25_return_indices_0"), val = bool(true)];
            string top_scores_25_cast_fp16_cast_uint16_output_indices_dtype_0 = const()[name = string("top_scores_25_cast_fp16_cast_uint16_output_indices_dtype_0"), val = string("uint16")];
            tensor<fp16, [1, 25]> top_scores_25_cast_fp16_cast_uint16_0, tensor<uint16, [1, 25]> top_scores_25_cast_fp16_cast_uint16_1 = topk(ascending = top_scores_25_ascending_0, axis = top_scores_25_axis_0, k = var_2485, output_indices_dtype = top_scores_25_cast_fp16_cast_uint16_output_indices_dtype_0, return_indices = top_scores_25_return_indices_0, sort = top_scores_25_sort_0, x = _inversed_scores_29_cast_fp16)[name = string("top_scores_25_cast_fp16_cast_uint16")];
            int32 var_2491 = const()[name = string("op_2491"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_41_cast_fp16 = softmax(axis = var_2491, x = top_scores_25_cast_fp16_cast_uint16_0)[name = string("probabilities_41_cast_fp16")];
            int32 var_2494 = const()[name = string("op_2494"), val = int32(1)];
            bool cumulative_25_exclusive_0 = const()[name = string("cumulative_25_exclusive_0"), val = bool(false)];
            bool cumulative_25_reverse_0 = const()[name = string("cumulative_25_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_25_cast_fp16 = cumsum(axis = var_2494, exclusive = cumulative_25_exclusive_0, reverse = cumulative_25_reverse_0, x = probabilities_41_cast_fp16)[name = string("cumulative_25_cast_fp16")];
            tensor<fp16, [1, 25]> var_2498_cast_fp16 = sub(x = cumulative_25_cast_fp16, y = probabilities_41_cast_fp16)[name = string("op_2498_cast_fp16")];
            fp16 var_2499_to_fp16 = const()[name = string("op_2499_to_fp16"), val = fp16(0x1.e68p-1)];
            tensor<bool, [1, 25]> keep_13_cast_fp16 = less(x = var_2498_cast_fp16, y = var_2499_to_fp16)[name = string("keep_13_cast_fp16")];
            fp16 var_2507_value_0_to_fp16 = const()[name = string("op_2507_value_0_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 25]> var_2507_cast_fp16 = fill_like(ref_tensor = top_scores_25_cast_fp16_cast_uint16_0, value = var_2507_value_0_to_fp16)[name = string("op_2507_cast_fp16")];
            tensor<fp16, [1, 25]> top_scores_27_cast_fp16 = select(a = top_scores_25_cast_fp16_cast_uint16_0, b = var_2507_cast_fp16, cond = keep_13_cast_fp16)[name = string("top_scores_27_cast_fp16")];
            int32 var_2509 = const()[name = string("op_2509"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_43_cast_fp16 = softmax(axis = var_2509, x = top_scores_27_cast_fp16)[name = string("probabilities_43_cast_fp16")];
            int32 var_2512 = const()[name = string("op_2512"), val = int32(1)];
            bool cumulative_27_exclusive_0 = const()[name = string("cumulative_27_exclusive_0"), val = bool(false)];
            bool cumulative_27_reverse_0 = const()[name = string("cumulative_27_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_27_cast_fp16 = cumsum(axis = var_2512, exclusive = cumulative_27_exclusive_0, reverse = cumulative_27_reverse_0, x = probabilities_43_cast_fp16)[name = string("cumulative_27_cast_fp16")];
            tensor<int32, [2]> var_2517 = const()[name = string("op_2517"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_2518_cast_fp16 = reshape(shape = var_2517, x = random_u_13_cast_fp16)[name = string("op_2518_cast_fp16")];
            tensor<bool, [1, 25]> var_2519_cast_fp16 = less(x = cumulative_27_cast_fp16, y = var_2518_cast_fp16)[name = string("op_2519_cast_fp16")];
            string var_2524_dtype_0 = const()[name = string("op_2524_dtype_0"), val = string("int32")];
            tensor<int32, [1]> selected_25_axes_0 = const()[name = string("selected_25_axes_0"), val = tensor<int32, [1]>([1])];
            bool selected_25_keep_dims_0 = const()[name = string("selected_25_keep_dims_0"), val = bool(false)];
            tensor<int32, [1, 25]> var_2524 = cast(dtype = var_2524_dtype_0, x = var_2519_cast_fp16)[name = string("cast_339")];
            tensor<int32, [1]> selected_25 = reduce_sum(axes = selected_25_axes_0, keep_dims = selected_25_keep_dims_0, x = var_2524)[name = string("selected_25")];
            string selected_25_promoted_to_fp16_dtype_0 = const()[name = string("selected_25_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_38_to_fp16 = const()[name = string("const_38_to_fp16"), val = fp16(-inf)];
            fp16 var_2531_promoted_to_fp16 = const()[name = string("op_2531_promoted_to_fp16"), val = fp16(0x1.8p+4)];
            tensor<fp16, [1]> selected_25_to_fp16 = cast(dtype = selected_25_promoted_to_fp16_dtype_0, x = selected_25)[name = string("cast_338")];
            tensor<fp16, [1]> clip_6_cast_fp16 = clip(alpha = const_38_to_fp16, beta = var_2531_promoted_to_fp16, x = selected_25_to_fp16)[name = string("clip_6_cast_fp16")];
            string selected_27_dtype_0 = const()[name = string("selected_27_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_2540 = const()[name = string("op_2540"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> clip_6_cast_fp16_to_int32 = cast(dtype = selected_27_dtype_0, x = clip_6_cast_fp16)[name = string("cast_337")];
            tensor<int32, [1, 1]> var_2541 = reshape(shape = var_2540, x = clip_6_cast_fp16_to_int32)[name = string("op_2541")];
            int32 var_2542 = const()[name = string("op_2542"), val = int32(1)];
            bool var_2544_validate_indices_0 = const()[name = string("op_2544_validate_indices_0"), val = bool(false)];
            string var_2541_to_uint16_dtype_0 = const()[name = string("op_2541_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 1]> var_2541_to_uint16 = cast(dtype = var_2541_to_uint16_dtype_0, x = var_2541)[name = string("cast_336")];
            tensor<uint16, [1, 1]> var_2544_cast_uint16 = gather_along_axis(axis = var_2542, indices = var_2541_to_uint16, validate_indices = var_2544_validate_indices_0, x = top_scores_25_cast_fp16_cast_uint16_1)[name = string("op_2544_cast_uint16")];
            string var_2544_cast_uint16_to_int32_dtype_0 = const()[name = string("op_2544_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [1]> var_2546 = const()[name = string("op_2546"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_2544_cast_uint16_to_int32 = cast(dtype = var_2544_cast_uint16_to_int32_dtype_0, x = var_2544_cast_uint16)[name = string("cast_335")];
            tensor<int32, [1]> token_13 = reshape(shape = var_2546, x = var_2544_cast_uint16_to_int32)[name = string("token_13")];
            int32 var_2556_axis_0 = const()[name = string("op_2556_axis_0"), val = int32(0)];
            int32 var_2556_batch_dims_0 = const()[name = string("op_2556_batch_dims_0"), val = int32(0)];
            bool var_2556_validate_indices_0 = const()[name = string("op_2556_validate_indices_0"), val = bool(false)];
            string token_13_to_uint16_dtype_0 = const()[name = string("token_13_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> token_13_to_uint16 = cast(dtype = token_13_to_uint16_dtype_0, x = token_13)[name = string("cast_334")];
            tensor<fp16, [1, 768]> var_2556_cast_fp16_cast_uint16 = gather(axis = var_2556_axis_0, batch_dims = var_2556_batch_dims_0, indices = token_13_to_uint16, validate_indices = var_2556_validate_indices_0, x = audio_embeddings_6_weight_to_fp16)[name = string("op_2556_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 768]> next_global_embed_15_cast_fp16 = add(x = next_global_embed_13_cast_fp16, y = var_2556_cast_fp16_cast_uint16)[name = string("next_global_embed_15_cast_fp16")];
            tensor<int32, [3]> var_2566 = const()[name = string("op_2566"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 768]> input_159_cast_fp16 = reshape(shape = var_2566, x = var_2556_cast_fp16_cast_uint16)[name = string("input_159_cast_fp16")];
            tensor<int32, [1]> input_161_axes_0 = const()[name = string("input_161_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_2569_to_fp16 = const()[name = string("op_2569_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_161_cast_fp16 = layer_norm(axes = input_161_axes_0, beta = block_ln_1_bias_to_fp16, epsilon = var_2569_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_159_cast_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [1, 1, 2304]> linear_39_cast_fp16 = linear(bias = block_attn_c_attn_bias_to_fp16, weight = block_attn_c_attn_weight_to_fp16, x = input_161_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<int32, [3]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [3]>([768, 768, 768])];
            int32 var_2580_axis_0 = const()[name = string("op_2580_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 768]> var_2580_cast_fp16_0, tensor<fp16, [1, 1, 768]> var_2580_cast_fp16_1, tensor<fp16, [1, 1, 768]> var_2580_cast_fp16_2 = split(axis = var_2580_axis_0, split_sizes = tile_8, x = linear_39_cast_fp16)[name = string("op_2580_cast_fp16")];
            tensor<int32, [4]> var_2588 = const()[name = string("op_2588"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> query_67_cast_fp16 = reshape(shape = var_2588, x = var_2580_cast_fp16_0)[name = string("query_67_cast_fp16")];
            tensor<int32, [4]> var_2594 = const()[name = string("op_2594"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> key_81_cast_fp16 = reshape(shape = var_2594, x = var_2580_cast_fp16_1)[name = string("key_81_cast_fp16")];
            tensor<int32, [4]> var_2600 = const()[name = string("op_2600"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> value_65_cast_fp16 = reshape(shape = var_2600, x = var_2580_cast_fp16_2)[name = string("value_65_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_cos_8_to_fp16 = const()[name = string("rope_cos_8_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25198400)))];
            tensor<fp16, [1, 1, 12, 64]> var_2602_cast_fp16 = mul(x = query_67_cast_fp16, y = rope_cos_8_to_fp16)[name = string("op_2602_cast_fp16")];
            tensor<int32, [4]> var_2607_begin_0 = const()[name = string("op_2607_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_2607_end_0 = const()[name = string("op_2607_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_2607_end_mask_0 = const()[name = string("op_2607_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2607_stride_0 = const()[name = string("op_2607_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_2607_cast_fp16 = slice_by_index(begin = var_2607_begin_0, end = var_2607_end_0, end_mask = var_2607_end_mask_0, stride = var_2607_stride_0, x = query_67_cast_fp16)[name = string("op_2607_cast_fp16")];
            fp16 const_41_promoted_to_fp16 = const()[name = string("const_41_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_2608_cast_fp16 = mul(x = var_2607_cast_fp16, y = const_41_promoted_to_fp16)[name = string("op_2608_cast_fp16")];
            tensor<int32, [4]> var_2613_begin_0 = const()[name = string("op_2613_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2613_end_0 = const()[name = string("op_2613_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_2613_end_mask_0 = const()[name = string("op_2613_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2613_stride_0 = const()[name = string("op_2613_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_2613_cast_fp16 = slice_by_index(begin = var_2613_begin_0, end = var_2613_end_0, end_mask = var_2613_end_mask_0, stride = var_2613_stride_0, x = query_67_cast_fp16)[name = string("op_2613_cast_fp16")];
            int32 var_2616_axis_0 = const()[name = string("op_2616_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_2616_cast_fp16 = stack(axis = var_2616_axis_0, values = (var_2608_cast_fp16, var_2613_cast_fp16))[name = string("op_2616_cast_fp16")];
            tensor<int32, [4]> shape_25 = const()[name = string("shape_25"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_2617_cast_fp16 = reshape(shape = shape_25, x = var_2616_cast_fp16)[name = string("op_2617_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_sin_8_to_fp16 = const()[name = string("rope_sin_8_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25198592)))];
            tensor<fp16, [1, 1, 12, 64]> var_2618_cast_fp16 = mul(x = var_2617_cast_fp16, y = rope_sin_8_to_fp16)[name = string("op_2618_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> query_69_cast_fp16 = add(x = var_2602_cast_fp16, y = var_2618_cast_fp16)[name = string("query_69_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_2621_cast_fp16 = mul(x = key_81_cast_fp16, y = rope_cos_8_to_fp16)[name = string("op_2621_cast_fp16")];
            tensor<int32, [4]> var_2626_begin_0 = const()[name = string("op_2626_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_2626_end_0 = const()[name = string("op_2626_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_2626_end_mask_0 = const()[name = string("op_2626_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2626_stride_0 = const()[name = string("op_2626_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_2626_cast_fp16 = slice_by_index(begin = var_2626_begin_0, end = var_2626_end_0, end_mask = var_2626_end_mask_0, stride = var_2626_stride_0, x = key_81_cast_fp16)[name = string("op_2626_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_2627_cast_fp16 = mul(x = var_2626_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_2627_cast_fp16")];
            tensor<int32, [4]> var_2632_begin_0 = const()[name = string("op_2632_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2632_end_0 = const()[name = string("op_2632_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_2632_end_mask_0 = const()[name = string("op_2632_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2632_stride_0 = const()[name = string("op_2632_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_2632_cast_fp16 = slice_by_index(begin = var_2632_begin_0, end = var_2632_end_0, end_mask = var_2632_end_mask_0, stride = var_2632_stride_0, x = key_81_cast_fp16)[name = string("op_2632_cast_fp16")];
            int32 var_2635_axis_0 = const()[name = string("op_2635_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_2635_cast_fp16 = stack(axis = var_2635_axis_0, values = (var_2627_cast_fp16, var_2632_cast_fp16))[name = string("op_2635_cast_fp16")];
            tensor<int32, [4]> shape_26 = const()[name = string("shape_26"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_2636_cast_fp16 = reshape(shape = shape_26, x = var_2635_cast_fp16)[name = string("op_2636_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_2637_cast_fp16 = mul(x = var_2636_cast_fp16, y = rope_sin_8_to_fp16)[name = string("op_2637_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> key_83_cast_fp16 = add(x = var_2621_cast_fp16, y = var_2637_cast_fp16)[name = string("key_83_cast_fp16")];
            tensor<int32, [4]> var_2644 = const()[name = string("op_2644"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2650 = const()[name = string("op_2650"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2656 = const()[name = string("op_2656"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 var_2659 = const()[name = string("op_2659"), val = int32(2)];
            bool key_87_interleave_0 = const()[name = string("key_87_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> key_85_cast_fp16 = transpose(perm = var_2650, x = key_83_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 12, 9, 64]> key_87_cast_fp16 = concat(axis = var_2659, interleave = key_87_interleave_0, values = (key_77_cast_fp16, key_85_cast_fp16))[name = string("key_87_cast_fp16")];
            int32 var_2662 = const()[name = string("op_2662"), val = int32(2)];
            bool value_69_interleave_0 = const()[name = string("value_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> value_67_cast_fp16 = transpose(perm = var_2656, x = value_65_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 12, 9, 64]> value_69_cast_fp16 = concat(axis = var_2662, interleave = value_69_interleave_0, values = (value_61_cast_fp16, value_67_cast_fp16))[name = string("value_69_cast_fp16")];
            bool var_2667_transpose_x_1 = const()[name = string("op_2667_transpose_x_1"), val = bool(false)];
            bool var_2667_transpose_y_1 = const()[name = string("op_2667_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 12, 1, 64]> query_71_cast_fp16 = transpose(perm = var_2644, x = query_69_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 12, 1, 9]> var_2667_cast_fp16 = matmul(transpose_x = var_2667_transpose_x_1, transpose_y = var_2667_transpose_y_1, x = query_71_cast_fp16, y = key_87_cast_fp16)[name = string("op_2667_cast_fp16")];
            fp16 var_2668_to_fp16 = const()[name = string("op_2668_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 1, 9]> scores_31_cast_fp16 = mul(x = var_2667_cast_fp16, y = var_2668_to_fp16)[name = string("scores_31_cast_fp16")];
            int32 var_2670 = const()[name = string("op_2670"), val = int32(-1)];
            tensor<fp16, [1, 12, 1, 9]> probabilities_45_cast_fp16 = softmax(axis = var_2670, x = scores_31_cast_fp16)[name = string("probabilities_45_cast_fp16")];
            bool attention_17_transpose_x_0 = const()[name = string("attention_17_transpose_x_0"), val = bool(false)];
            bool attention_17_transpose_y_0 = const()[name = string("attention_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> attention_17_cast_fp16 = matmul(transpose_x = attention_17_transpose_x_0, transpose_y = attention_17_transpose_y_0, x = probabilities_45_cast_fp16, y = value_69_cast_fp16)[name = string("attention_17_cast_fp16")];
            tensor<int32, [4]> var_2678 = const()[name = string("op_2678"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2683 = const()[name = string("op_2683"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 12, 64]> var_2679_cast_fp16 = transpose(perm = var_2678, x = attention_17_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 1, 768]> input_163_cast_fp16 = reshape(shape = var_2683, x = var_2679_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_40_cast_fp16 = linear(bias = block_attn_c_proj_bias_to_fp16, weight = block_attn_c_proj_weight_to_fp16, x = input_163_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_165_cast_fp16 = add(x = input_159_cast_fp16, y = linear_40_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<int32, [1]> input_167_axes_0 = const()[name = string("input_167_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_2691_to_fp16 = const()[name = string("op_2691_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = block_ln_2_bias_to_fp16, epsilon = var_2691_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_165_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [1, 1, 3072]> linear_41_cast_fp16 = linear(bias = block_mlp_fc_in_bias_to_fp16, weight = block_mlp_fc_in_weight_to_fp16, x = input_167_cast_fp16)[name = string("linear_41_cast_fp16")];
            string input_171_mode_0 = const()[name = string("input_171_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1, 3072]> input_171_cast_fp16 = gelu(mode = input_171_mode_0, x = linear_41_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_42_cast_fp16 = linear(bias = block_mlp_fc_out_bias_to_fp16, weight = block_mlp_fc_out_weight_to_fp16, x = input_171_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_175_cast_fp16 = add(x = input_165_cast_fp16, y = linear_42_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<int32, [1]> var_2730_axes_0 = const()[name = string("op_2730_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_2725_to_fp16 = const()[name = string("op_2725_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> var_2730_cast_fp16 = layer_norm(axes = var_2730_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_2725_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_175_cast_fp16)[name = string("op_2730_cast_fp16")];
            tensor<int32, [3]> var_2738_begin_0 = const()[name = string("op_2738_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_2738_end_0 = const()[name = string("op_2738_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_2738_end_mask_0 = const()[name = string("op_2738_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_2738_squeeze_mask_0 = const()[name = string("op_2738_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_2738_cast_fp16 = slice_by_index(begin = var_2738_begin_0, end = var_2738_end_0, end_mask = var_2738_end_mask_0, squeeze_mask = var_2738_squeeze_mask_0, x = var_2730_cast_fp16)[name = string("op_2738_cast_fp16")];
            tensor<fp16, [1024, 768]> audio_embeddings_7_weight_to_fp16 = const()[name = string("audio_embeddings_7_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25198784)))];
            tensor<fp16, [1, 1024]> linear_43_cast_fp16 = linear(bias = linear_8_bias_0_to_fp16, weight = audio_embeddings_7_weight_to_fp16, x = var_2738_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [3]> var_2753_begin_0 = const()[name = string("op_2753_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
            tensor<int32, [3]> var_2753_end_0 = const()[name = string("op_2753_end_0"), val = tensor<int32, [3]>([1, 8, 1024])];
            tensor<bool, [3]> var_2753_end_mask_0 = const()[name = string("op_2753_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_2753_squeeze_mask_0 = const()[name = string("op_2753_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<int32, [1, 1024]> var_2753 = slice_by_index(begin = var_2753_begin_0, end = var_2753_end_0, end_mask = var_2753_end_mask_0, squeeze_mask = var_2753_squeeze_mask_0, x = repetition_seen)[name = string("op_2753")];
            tensor<int32, [2]> random_u_15_begin_0 = const()[name = string("random_u_15_begin_0"), val = tensor<int32, [2]>([0, 7])];
            tensor<int32, [2]> random_u_15_end_0 = const()[name = string("random_u_15_end_0"), val = tensor<int32, [2]>([1, 8])];
            tensor<bool, [2]> random_u_15_end_mask_0 = const()[name = string("random_u_15_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<bool, [2]> random_u_15_squeeze_mask_0 = const()[name = string("random_u_15_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [1]> random_u_15_cast_fp16 = slice_by_index(begin = random_u_15_begin_0, end = random_u_15_end_0, end_mask = random_u_15_end_mask_0, squeeze_mask = random_u_15_squeeze_mask_0, x = audio_random_to_fp16)[name = string("random_u_15_cast_fp16")];
            fp16 var_2767_promoted_to_fp16 = const()[name = string("op_2767_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [1, 1024]> var_2768_cast_fp16 = less(x = linear_43_cast_fp16, y = var_2767_promoted_to_fp16)[name = string("op_2768_cast_fp16")];
            fp16 var_2769_to_fp16 = const()[name = string("op_2769_to_fp16"), val = fp16(0x1.334p+0)];
            tensor<fp16, [1, 1024]> var_2770_cast_fp16 = mul(x = linear_43_cast_fp16, y = var_2769_to_fp16)[name = string("op_2770_cast_fp16")];
            fp16 _inversed_2772_y_0_to_fp16 = const()[name = string("_inversed_2772_y_0_to_fp16"), val = fp16(0x1.aacp-1)];
            tensor<fp16, [1, 1024]> _inversed_2772_cast_fp16 = mul(x = linear_43_cast_fp16, y = _inversed_2772_y_0_to_fp16)[name = string("_inversed_2772_cast_fp16")];
            tensor<fp16, [1, 1024]> penalized_15_cast_fp16 = select(a = var_2770_cast_fp16, b = _inversed_2772_cast_fp16, cond = var_2768_cast_fp16)[name = string("penalized_15_cast_fp16")];
            string var_2778_dtype_0 = const()[name = string("op_2778_dtype_0"), val = string("bool")];
            tensor<bool, [1, 1024]> var_2778 = cast(dtype = var_2778_dtype_0, x = var_2753)[name = string("cast_333")];
            tensor<fp16, [1, 1024]> var_2779_cast_fp16 = select(a = penalized_15_cast_fp16, b = linear_43_cast_fp16, cond = var_2778)[name = string("op_2779_cast_fp16")];
            fp16 _inversed_scores_33_y_0_to_fp16 = const()[name = string("_inversed_scores_33_y_0_to_fp16"), val = fp16(0x1.4p+0)];
            tensor<fp16, [1, 1024]> _inversed_scores_33_cast_fp16 = mul(x = var_2779_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = string("_inversed_scores_33_cast_fp16")];
            int32 var_2782 = const()[name = string("op_2782"), val = int32(25)];
            int32 top_scores_29_axis_0 = const()[name = string("top_scores_29_axis_0"), val = int32(1)];
            bool top_scores_29_ascending_0 = const()[name = string("top_scores_29_ascending_0"), val = bool(false)];
            bool top_scores_29_sort_0 = const()[name = string("top_scores_29_sort_0"), val = bool(true)];
            bool top_scores_29_return_indices_0 = const()[name = string("top_scores_29_return_indices_0"), val = bool(true)];
            string top_scores_29_cast_fp16_cast_uint16_output_indices_dtype_0 = const()[name = string("top_scores_29_cast_fp16_cast_uint16_output_indices_dtype_0"), val = string("uint16")];
            tensor<fp16, [1, 25]> top_scores_29_cast_fp16_cast_uint16_0, tensor<uint16, [1, 25]> top_scores_29_cast_fp16_cast_uint16_1 = topk(ascending = top_scores_29_ascending_0, axis = top_scores_29_axis_0, k = var_2782, output_indices_dtype = top_scores_29_cast_fp16_cast_uint16_output_indices_dtype_0, return_indices = top_scores_29_return_indices_0, sort = top_scores_29_sort_0, x = _inversed_scores_33_cast_fp16)[name = string("top_scores_29_cast_fp16_cast_uint16")];
            int32 var_2788 = const()[name = string("op_2788"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_47_cast_fp16 = softmax(axis = var_2788, x = top_scores_29_cast_fp16_cast_uint16_0)[name = string("probabilities_47_cast_fp16")];
            int32 var_2791 = const()[name = string("op_2791"), val = int32(1)];
            bool cumulative_29_exclusive_0 = const()[name = string("cumulative_29_exclusive_0"), val = bool(false)];
            bool cumulative_29_reverse_0 = const()[name = string("cumulative_29_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_29_cast_fp16 = cumsum(axis = var_2791, exclusive = cumulative_29_exclusive_0, reverse = cumulative_29_reverse_0, x = probabilities_47_cast_fp16)[name = string("cumulative_29_cast_fp16")];
            tensor<fp16, [1, 25]> var_2795_cast_fp16 = sub(x = cumulative_29_cast_fp16, y = probabilities_47_cast_fp16)[name = string("op_2795_cast_fp16")];
            fp16 var_2796_to_fp16 = const()[name = string("op_2796_to_fp16"), val = fp16(0x1.e68p-1)];
            tensor<bool, [1, 25]> keep_15_cast_fp16 = less(x = var_2795_cast_fp16, y = var_2796_to_fp16)[name = string("keep_15_cast_fp16")];
            fp16 var_2804_value_0_to_fp16 = const()[name = string("op_2804_value_0_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 25]> var_2804_cast_fp16 = fill_like(ref_tensor = top_scores_29_cast_fp16_cast_uint16_0, value = var_2804_value_0_to_fp16)[name = string("op_2804_cast_fp16")];
            tensor<fp16, [1, 25]> top_scores_31_cast_fp16 = select(a = top_scores_29_cast_fp16_cast_uint16_0, b = var_2804_cast_fp16, cond = keep_15_cast_fp16)[name = string("top_scores_31_cast_fp16")];
            int32 var_2806 = const()[name = string("op_2806"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_49_cast_fp16 = softmax(axis = var_2806, x = top_scores_31_cast_fp16)[name = string("probabilities_49_cast_fp16")];
            int32 var_2809 = const()[name = string("op_2809"), val = int32(1)];
            bool cumulative_31_exclusive_0 = const()[name = string("cumulative_31_exclusive_0"), val = bool(false)];
            bool cumulative_31_reverse_0 = const()[name = string("cumulative_31_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_31_cast_fp16 = cumsum(axis = var_2809, exclusive = cumulative_31_exclusive_0, reverse = cumulative_31_reverse_0, x = probabilities_49_cast_fp16)[name = string("cumulative_31_cast_fp16")];
            tensor<int32, [2]> var_2814 = const()[name = string("op_2814"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_2815_cast_fp16 = reshape(shape = var_2814, x = random_u_15_cast_fp16)[name = string("op_2815_cast_fp16")];
            tensor<bool, [1, 25]> var_2816_cast_fp16 = less(x = cumulative_31_cast_fp16, y = var_2815_cast_fp16)[name = string("op_2816_cast_fp16")];
            string var_2821_dtype_0 = const()[name = string("op_2821_dtype_0"), val = string("int32")];
            tensor<int32, [1]> selected_29_axes_0 = const()[name = string("selected_29_axes_0"), val = tensor<int32, [1]>([1])];
            bool selected_29_keep_dims_0 = const()[name = string("selected_29_keep_dims_0"), val = bool(false)];
            tensor<int32, [1, 25]> var_2821 = cast(dtype = var_2821_dtype_0, x = var_2816_cast_fp16)[name = string("cast_332")];
            tensor<int32, [1]> selected_29 = reduce_sum(axes = selected_29_axes_0, keep_dims = selected_29_keep_dims_0, x = var_2821)[name = string("selected_29")];
            string selected_29_promoted_to_fp16_dtype_0 = const()[name = string("selected_29_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_43_to_fp16 = const()[name = string("const_43_to_fp16"), val = fp16(-inf)];
            fp16 var_2828_promoted_to_fp16 = const()[name = string("op_2828_promoted_to_fp16"), val = fp16(0x1.8p+4)];
            tensor<fp16, [1]> selected_29_to_fp16 = cast(dtype = selected_29_promoted_to_fp16_dtype_0, x = selected_29)[name = string("cast_331")];
            tensor<fp16, [1]> clip_7_cast_fp16 = clip(alpha = const_43_to_fp16, beta = var_2828_promoted_to_fp16, x = selected_29_to_fp16)[name = string("clip_7_cast_fp16")];
            string selected_31_dtype_0 = const()[name = string("selected_31_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_2837 = const()[name = string("op_2837"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> clip_7_cast_fp16_to_int32 = cast(dtype = selected_31_dtype_0, x = clip_7_cast_fp16)[name = string("cast_330")];
            tensor<int32, [1, 1]> var_2838 = reshape(shape = var_2837, x = clip_7_cast_fp16_to_int32)[name = string("op_2838")];
            int32 var_2839 = const()[name = string("op_2839"), val = int32(1)];
            bool var_2841_validate_indices_0 = const()[name = string("op_2841_validate_indices_0"), val = bool(false)];
            string var_2838_to_uint16_dtype_0 = const()[name = string("op_2838_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 1]> var_2838_to_uint16 = cast(dtype = var_2838_to_uint16_dtype_0, x = var_2838)[name = string("cast_329")];
            tensor<uint16, [1, 1]> var_2841_cast_uint16 = gather_along_axis(axis = var_2839, indices = var_2838_to_uint16, validate_indices = var_2841_validate_indices_0, x = top_scores_29_cast_fp16_cast_uint16_1)[name = string("op_2841_cast_uint16")];
            string var_2841_cast_uint16_to_int32_dtype_0 = const()[name = string("op_2841_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [1]> var_2843 = const()[name = string("op_2843"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_2841_cast_uint16_to_int32 = cast(dtype = var_2841_cast_uint16_to_int32_dtype_0, x = var_2841_cast_uint16)[name = string("cast_328")];
            tensor<int32, [1]> token_15 = reshape(shape = var_2843, x = var_2841_cast_uint16_to_int32)[name = string("token_15")];
            int32 var_2853_axis_0 = const()[name = string("op_2853_axis_0"), val = int32(0)];
            int32 var_2853_batch_dims_0 = const()[name = string("op_2853_batch_dims_0"), val = int32(0)];
            bool var_2853_validate_indices_0 = const()[name = string("op_2853_validate_indices_0"), val = bool(false)];
            string token_15_to_uint16_dtype_0 = const()[name = string("token_15_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> token_15_to_uint16 = cast(dtype = token_15_to_uint16_dtype_0, x = token_15)[name = string("cast_327")];
            tensor<fp16, [1, 768]> var_2853_cast_fp16_cast_uint16 = gather(axis = var_2853_axis_0, batch_dims = var_2853_batch_dims_0, indices = token_15_to_uint16, validate_indices = var_2853_validate_indices_0, x = audio_embeddings_7_weight_to_fp16)[name = string("op_2853_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 768]> next_global_embed_17_cast_fp16 = add(x = next_global_embed_15_cast_fp16, y = var_2853_cast_fp16_cast_uint16)[name = string("next_global_embed_17_cast_fp16")];
            tensor<int32, [3]> var_2863 = const()[name = string("op_2863"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 768]> input_179_cast_fp16 = reshape(shape = var_2863, x = var_2853_cast_fp16_cast_uint16)[name = string("input_179_cast_fp16")];
            tensor<int32, [1]> input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_2866_to_fp16 = const()[name = string("op_2866_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = block_ln_1_bias_to_fp16, epsilon = var_2866_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_179_cast_fp16)[name = string("input_181_cast_fp16")];
            tensor<fp16, [1, 1, 2304]> linear_44_cast_fp16 = linear(bias = block_attn_c_attn_bias_to_fp16, weight = block_attn_c_attn_weight_to_fp16, x = input_181_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [3]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [3]>([768, 768, 768])];
            int32 var_2877_axis_0 = const()[name = string("op_2877_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 768]> var_2877_cast_fp16_0, tensor<fp16, [1, 1, 768]> var_2877_cast_fp16_1, tensor<fp16, [1, 1, 768]> var_2877_cast_fp16_2 = split(axis = var_2877_axis_0, split_sizes = tile_9, x = linear_44_cast_fp16)[name = string("op_2877_cast_fp16")];
            tensor<int32, [4]> var_2885 = const()[name = string("op_2885"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> query_75_cast_fp16 = reshape(shape = var_2885, x = var_2877_cast_fp16_0)[name = string("query_75_cast_fp16")];
            tensor<int32, [4]> var_2891 = const()[name = string("op_2891"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> key_91_cast_fp16 = reshape(shape = var_2891, x = var_2877_cast_fp16_1)[name = string("key_91_cast_fp16")];
            tensor<int32, [4]> var_2897 = const()[name = string("op_2897"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> value_73_cast_fp16 = reshape(shape = var_2897, x = var_2877_cast_fp16_2)[name = string("value_73_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_cos_9_to_fp16 = const()[name = string("rope_cos_9_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26771712)))];
            tensor<fp16, [1, 1, 12, 64]> var_2899_cast_fp16 = mul(x = query_75_cast_fp16, y = rope_cos_9_to_fp16)[name = string("op_2899_cast_fp16")];
            tensor<int32, [4]> var_2904_begin_0 = const()[name = string("op_2904_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_2904_end_0 = const()[name = string("op_2904_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_2904_end_mask_0 = const()[name = string("op_2904_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2904_stride_0 = const()[name = string("op_2904_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_2904_cast_fp16 = slice_by_index(begin = var_2904_begin_0, end = var_2904_end_0, end_mask = var_2904_end_mask_0, stride = var_2904_stride_0, x = query_75_cast_fp16)[name = string("op_2904_cast_fp16")];
            fp16 const_46_promoted_to_fp16 = const()[name = string("const_46_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_2905_cast_fp16 = mul(x = var_2904_cast_fp16, y = const_46_promoted_to_fp16)[name = string("op_2905_cast_fp16")];
            tensor<int32, [4]> var_2910_begin_0 = const()[name = string("op_2910_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2910_end_0 = const()[name = string("op_2910_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_2910_end_mask_0 = const()[name = string("op_2910_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2910_stride_0 = const()[name = string("op_2910_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_2910_cast_fp16 = slice_by_index(begin = var_2910_begin_0, end = var_2910_end_0, end_mask = var_2910_end_mask_0, stride = var_2910_stride_0, x = query_75_cast_fp16)[name = string("op_2910_cast_fp16")];
            int32 var_2913_axis_0 = const()[name = string("op_2913_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_2913_cast_fp16 = stack(axis = var_2913_axis_0, values = (var_2905_cast_fp16, var_2910_cast_fp16))[name = string("op_2913_cast_fp16")];
            tensor<int32, [4]> shape_28 = const()[name = string("shape_28"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_2914_cast_fp16 = reshape(shape = shape_28, x = var_2913_cast_fp16)[name = string("op_2914_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_sin_9_to_fp16 = const()[name = string("rope_sin_9_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26771904)))];
            tensor<fp16, [1, 1, 12, 64]> var_2915_cast_fp16 = mul(x = var_2914_cast_fp16, y = rope_sin_9_to_fp16)[name = string("op_2915_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> query_77_cast_fp16 = add(x = var_2899_cast_fp16, y = var_2915_cast_fp16)[name = string("query_77_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_2918_cast_fp16 = mul(x = key_91_cast_fp16, y = rope_cos_9_to_fp16)[name = string("op_2918_cast_fp16")];
            tensor<int32, [4]> var_2923_begin_0 = const()[name = string("op_2923_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_2923_end_0 = const()[name = string("op_2923_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_2923_end_mask_0 = const()[name = string("op_2923_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2923_stride_0 = const()[name = string("op_2923_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_2923_cast_fp16 = slice_by_index(begin = var_2923_begin_0, end = var_2923_end_0, end_mask = var_2923_end_mask_0, stride = var_2923_stride_0, x = key_91_cast_fp16)[name = string("op_2923_cast_fp16")];
            fp16 const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_2924_cast_fp16 = mul(x = var_2923_cast_fp16, y = const_47_promoted_to_fp16)[name = string("op_2924_cast_fp16")];
            tensor<int32, [4]> var_2929_begin_0 = const()[name = string("op_2929_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2929_end_0 = const()[name = string("op_2929_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_2929_end_mask_0 = const()[name = string("op_2929_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2929_stride_0 = const()[name = string("op_2929_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_2929_cast_fp16 = slice_by_index(begin = var_2929_begin_0, end = var_2929_end_0, end_mask = var_2929_end_mask_0, stride = var_2929_stride_0, x = key_91_cast_fp16)[name = string("op_2929_cast_fp16")];
            int32 var_2932_axis_0 = const()[name = string("op_2932_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_2932_cast_fp16 = stack(axis = var_2932_axis_0, values = (var_2924_cast_fp16, var_2929_cast_fp16))[name = string("op_2932_cast_fp16")];
            tensor<int32, [4]> shape_29 = const()[name = string("shape_29"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_2933_cast_fp16 = reshape(shape = shape_29, x = var_2932_cast_fp16)[name = string("op_2933_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_2934_cast_fp16 = mul(x = var_2933_cast_fp16, y = rope_sin_9_to_fp16)[name = string("op_2934_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> key_93_cast_fp16 = add(x = var_2918_cast_fp16, y = var_2934_cast_fp16)[name = string("key_93_cast_fp16")];
            tensor<int32, [4]> var_2941 = const()[name = string("op_2941"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2947 = const()[name = string("op_2947"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2953 = const()[name = string("op_2953"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 var_2956 = const()[name = string("op_2956"), val = int32(2)];
            bool key_97_interleave_0 = const()[name = string("key_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> key_95_cast_fp16 = transpose(perm = var_2947, x = key_93_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 12, 10, 64]> key_97_cast_fp16 = concat(axis = var_2956, interleave = key_97_interleave_0, values = (key_87_cast_fp16, key_95_cast_fp16))[name = string("key_97_cast_fp16")];
            int32 var_2959 = const()[name = string("op_2959"), val = int32(2)];
            bool value_77_interleave_0 = const()[name = string("value_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> value_75_cast_fp16 = transpose(perm = var_2953, x = value_73_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 12, 10, 64]> value_77_cast_fp16 = concat(axis = var_2959, interleave = value_77_interleave_0, values = (value_69_cast_fp16, value_75_cast_fp16))[name = string("value_77_cast_fp16")];
            bool var_2964_transpose_x_1 = const()[name = string("op_2964_transpose_x_1"), val = bool(false)];
            bool var_2964_transpose_y_1 = const()[name = string("op_2964_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 12, 1, 64]> query_79_cast_fp16 = transpose(perm = var_2941, x = query_77_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 12, 1, 10]> var_2964_cast_fp16 = matmul(transpose_x = var_2964_transpose_x_1, transpose_y = var_2964_transpose_y_1, x = query_79_cast_fp16, y = key_97_cast_fp16)[name = string("op_2964_cast_fp16")];
            fp16 var_2965_to_fp16 = const()[name = string("op_2965_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 1, 10]> scores_35_cast_fp16 = mul(x = var_2964_cast_fp16, y = var_2965_to_fp16)[name = string("scores_35_cast_fp16")];
            int32 var_2967 = const()[name = string("op_2967"), val = int32(-1)];
            tensor<fp16, [1, 12, 1, 10]> probabilities_51_cast_fp16 = softmax(axis = var_2967, x = scores_35_cast_fp16)[name = string("probabilities_51_cast_fp16")];
            bool attention_19_transpose_x_0 = const()[name = string("attention_19_transpose_x_0"), val = bool(false)];
            bool attention_19_transpose_y_0 = const()[name = string("attention_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> attention_19_cast_fp16 = matmul(transpose_x = attention_19_transpose_x_0, transpose_y = attention_19_transpose_y_0, x = probabilities_51_cast_fp16, y = value_77_cast_fp16)[name = string("attention_19_cast_fp16")];
            tensor<int32, [4]> var_2975 = const()[name = string("op_2975"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2980 = const()[name = string("op_2980"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 12, 64]> var_2976_cast_fp16 = transpose(perm = var_2975, x = attention_19_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 1, 768]> input_183_cast_fp16 = reshape(shape = var_2980, x = var_2976_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_45_cast_fp16 = linear(bias = block_attn_c_proj_bias_to_fp16, weight = block_attn_c_proj_weight_to_fp16, x = input_183_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_185_cast_fp16 = add(x = input_179_cast_fp16, y = linear_45_cast_fp16)[name = string("input_185_cast_fp16")];
            tensor<int32, [1]> input_187_axes_0 = const()[name = string("input_187_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_2988_to_fp16 = const()[name = string("op_2988_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_187_cast_fp16 = layer_norm(axes = input_187_axes_0, beta = block_ln_2_bias_to_fp16, epsilon = var_2988_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_185_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [1, 1, 3072]> linear_46_cast_fp16 = linear(bias = block_mlp_fc_in_bias_to_fp16, weight = block_mlp_fc_in_weight_to_fp16, x = input_187_cast_fp16)[name = string("linear_46_cast_fp16")];
            string input_191_mode_0 = const()[name = string("input_191_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1, 3072]> input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_46_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_47_cast_fp16 = linear(bias = block_mlp_fc_out_bias_to_fp16, weight = block_mlp_fc_out_weight_to_fp16, x = input_191_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_195_cast_fp16 = add(x = input_185_cast_fp16, y = linear_47_cast_fp16)[name = string("input_195_cast_fp16")];
            tensor<int32, [1]> var_3027_axes_0 = const()[name = string("op_3027_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_3022_to_fp16 = const()[name = string("op_3022_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> var_3027_cast_fp16 = layer_norm(axes = var_3027_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_3022_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_195_cast_fp16)[name = string("op_3027_cast_fp16")];
            tensor<int32, [3]> var_3035_begin_0 = const()[name = string("op_3035_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3035_end_0 = const()[name = string("op_3035_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_3035_end_mask_0 = const()[name = string("op_3035_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_3035_squeeze_mask_0 = const()[name = string("op_3035_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_3035_cast_fp16 = slice_by_index(begin = var_3035_begin_0, end = var_3035_end_0, end_mask = var_3035_end_mask_0, squeeze_mask = var_3035_squeeze_mask_0, x = var_3027_cast_fp16)[name = string("op_3035_cast_fp16")];
            tensor<fp16, [1024, 768]> audio_embeddings_8_weight_to_fp16 = const()[name = string("audio_embeddings_8_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26772096)))];
            tensor<fp16, [1, 1024]> linear_48_cast_fp16 = linear(bias = linear_8_bias_0_to_fp16, weight = audio_embeddings_8_weight_to_fp16, x = var_3035_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<int32, [3]> var_3050_begin_0 = const()[name = string("op_3050_begin_0"), val = tensor<int32, [3]>([0, 8, 0])];
            tensor<int32, [3]> var_3050_end_0 = const()[name = string("op_3050_end_0"), val = tensor<int32, [3]>([1, 9, 1024])];
            tensor<bool, [3]> var_3050_end_mask_0 = const()[name = string("op_3050_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_3050_squeeze_mask_0 = const()[name = string("op_3050_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<int32, [1, 1024]> var_3050 = slice_by_index(begin = var_3050_begin_0, end = var_3050_end_0, end_mask = var_3050_end_mask_0, squeeze_mask = var_3050_squeeze_mask_0, x = repetition_seen)[name = string("op_3050")];
            tensor<int32, [2]> random_u_17_begin_0 = const()[name = string("random_u_17_begin_0"), val = tensor<int32, [2]>([0, 8])];
            tensor<int32, [2]> random_u_17_end_0 = const()[name = string("random_u_17_end_0"), val = tensor<int32, [2]>([1, 9])];
            tensor<bool, [2]> random_u_17_end_mask_0 = const()[name = string("random_u_17_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<bool, [2]> random_u_17_squeeze_mask_0 = const()[name = string("random_u_17_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [1]> random_u_17_cast_fp16 = slice_by_index(begin = random_u_17_begin_0, end = random_u_17_end_0, end_mask = random_u_17_end_mask_0, squeeze_mask = random_u_17_squeeze_mask_0, x = audio_random_to_fp16)[name = string("random_u_17_cast_fp16")];
            fp16 var_3064_promoted_to_fp16 = const()[name = string("op_3064_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [1, 1024]> var_3065_cast_fp16 = less(x = linear_48_cast_fp16, y = var_3064_promoted_to_fp16)[name = string("op_3065_cast_fp16")];
            fp16 var_3066_to_fp16 = const()[name = string("op_3066_to_fp16"), val = fp16(0x1.334p+0)];
            tensor<fp16, [1, 1024]> var_3067_cast_fp16 = mul(x = linear_48_cast_fp16, y = var_3066_to_fp16)[name = string("op_3067_cast_fp16")];
            fp16 _inversed_3069_y_0_to_fp16 = const()[name = string("_inversed_3069_y_0_to_fp16"), val = fp16(0x1.aacp-1)];
            tensor<fp16, [1, 1024]> _inversed_3069_cast_fp16 = mul(x = linear_48_cast_fp16, y = _inversed_3069_y_0_to_fp16)[name = string("_inversed_3069_cast_fp16")];
            tensor<fp16, [1, 1024]> penalized_17_cast_fp16 = select(a = var_3067_cast_fp16, b = _inversed_3069_cast_fp16, cond = var_3065_cast_fp16)[name = string("penalized_17_cast_fp16")];
            string var_3075_dtype_0 = const()[name = string("op_3075_dtype_0"), val = string("bool")];
            tensor<bool, [1, 1024]> var_3075 = cast(dtype = var_3075_dtype_0, x = var_3050)[name = string("cast_326")];
            tensor<fp16, [1, 1024]> var_3076_cast_fp16 = select(a = penalized_17_cast_fp16, b = linear_48_cast_fp16, cond = var_3075)[name = string("op_3076_cast_fp16")];
            fp16 _inversed_scores_37_y_0_to_fp16 = const()[name = string("_inversed_scores_37_y_0_to_fp16"), val = fp16(0x1.4p+0)];
            tensor<fp16, [1, 1024]> _inversed_scores_37_cast_fp16 = mul(x = var_3076_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = string("_inversed_scores_37_cast_fp16")];
            int32 var_3079 = const()[name = string("op_3079"), val = int32(25)];
            int32 top_scores_33_axis_0 = const()[name = string("top_scores_33_axis_0"), val = int32(1)];
            bool top_scores_33_ascending_0 = const()[name = string("top_scores_33_ascending_0"), val = bool(false)];
            bool top_scores_33_sort_0 = const()[name = string("top_scores_33_sort_0"), val = bool(true)];
            bool top_scores_33_return_indices_0 = const()[name = string("top_scores_33_return_indices_0"), val = bool(true)];
            string top_scores_33_cast_fp16_cast_uint16_output_indices_dtype_0 = const()[name = string("top_scores_33_cast_fp16_cast_uint16_output_indices_dtype_0"), val = string("uint16")];
            tensor<fp16, [1, 25]> top_scores_33_cast_fp16_cast_uint16_0, tensor<uint16, [1, 25]> top_scores_33_cast_fp16_cast_uint16_1 = topk(ascending = top_scores_33_ascending_0, axis = top_scores_33_axis_0, k = var_3079, output_indices_dtype = top_scores_33_cast_fp16_cast_uint16_output_indices_dtype_0, return_indices = top_scores_33_return_indices_0, sort = top_scores_33_sort_0, x = _inversed_scores_37_cast_fp16)[name = string("top_scores_33_cast_fp16_cast_uint16")];
            int32 var_3085 = const()[name = string("op_3085"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_53_cast_fp16 = softmax(axis = var_3085, x = top_scores_33_cast_fp16_cast_uint16_0)[name = string("probabilities_53_cast_fp16")];
            int32 var_3088 = const()[name = string("op_3088"), val = int32(1)];
            bool cumulative_33_exclusive_0 = const()[name = string("cumulative_33_exclusive_0"), val = bool(false)];
            bool cumulative_33_reverse_0 = const()[name = string("cumulative_33_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_33_cast_fp16 = cumsum(axis = var_3088, exclusive = cumulative_33_exclusive_0, reverse = cumulative_33_reverse_0, x = probabilities_53_cast_fp16)[name = string("cumulative_33_cast_fp16")];
            tensor<fp16, [1, 25]> var_3092_cast_fp16 = sub(x = cumulative_33_cast_fp16, y = probabilities_53_cast_fp16)[name = string("op_3092_cast_fp16")];
            fp16 var_3093_to_fp16 = const()[name = string("op_3093_to_fp16"), val = fp16(0x1.e68p-1)];
            tensor<bool, [1, 25]> keep_17_cast_fp16 = less(x = var_3092_cast_fp16, y = var_3093_to_fp16)[name = string("keep_17_cast_fp16")];
            fp16 var_3101_value_0_to_fp16 = const()[name = string("op_3101_value_0_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 25]> var_3101_cast_fp16 = fill_like(ref_tensor = top_scores_33_cast_fp16_cast_uint16_0, value = var_3101_value_0_to_fp16)[name = string("op_3101_cast_fp16")];
            tensor<fp16, [1, 25]> top_scores_35_cast_fp16 = select(a = top_scores_33_cast_fp16_cast_uint16_0, b = var_3101_cast_fp16, cond = keep_17_cast_fp16)[name = string("top_scores_35_cast_fp16")];
            int32 var_3103 = const()[name = string("op_3103"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_55_cast_fp16 = softmax(axis = var_3103, x = top_scores_35_cast_fp16)[name = string("probabilities_55_cast_fp16")];
            int32 var_3106 = const()[name = string("op_3106"), val = int32(1)];
            bool cumulative_35_exclusive_0 = const()[name = string("cumulative_35_exclusive_0"), val = bool(false)];
            bool cumulative_35_reverse_0 = const()[name = string("cumulative_35_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_35_cast_fp16 = cumsum(axis = var_3106, exclusive = cumulative_35_exclusive_0, reverse = cumulative_35_reverse_0, x = probabilities_55_cast_fp16)[name = string("cumulative_35_cast_fp16")];
            tensor<int32, [2]> var_3111 = const()[name = string("op_3111"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_3112_cast_fp16 = reshape(shape = var_3111, x = random_u_17_cast_fp16)[name = string("op_3112_cast_fp16")];
            tensor<bool, [1, 25]> var_3113_cast_fp16 = less(x = cumulative_35_cast_fp16, y = var_3112_cast_fp16)[name = string("op_3113_cast_fp16")];
            string var_3118_dtype_0 = const()[name = string("op_3118_dtype_0"), val = string("int32")];
            tensor<int32, [1]> selected_33_axes_0 = const()[name = string("selected_33_axes_0"), val = tensor<int32, [1]>([1])];
            bool selected_33_keep_dims_0 = const()[name = string("selected_33_keep_dims_0"), val = bool(false)];
            tensor<int32, [1, 25]> var_3118 = cast(dtype = var_3118_dtype_0, x = var_3113_cast_fp16)[name = string("cast_325")];
            tensor<int32, [1]> selected_33 = reduce_sum(axes = selected_33_axes_0, keep_dims = selected_33_keep_dims_0, x = var_3118)[name = string("selected_33")];
            string selected_33_promoted_to_fp16_dtype_0 = const()[name = string("selected_33_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_48_to_fp16 = const()[name = string("const_48_to_fp16"), val = fp16(-inf)];
            fp16 var_3125_promoted_to_fp16 = const()[name = string("op_3125_promoted_to_fp16"), val = fp16(0x1.8p+4)];
            tensor<fp16, [1]> selected_33_to_fp16 = cast(dtype = selected_33_promoted_to_fp16_dtype_0, x = selected_33)[name = string("cast_324")];
            tensor<fp16, [1]> clip_8_cast_fp16 = clip(alpha = const_48_to_fp16, beta = var_3125_promoted_to_fp16, x = selected_33_to_fp16)[name = string("clip_8_cast_fp16")];
            string selected_35_dtype_0 = const()[name = string("selected_35_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_3134 = const()[name = string("op_3134"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> clip_8_cast_fp16_to_int32 = cast(dtype = selected_35_dtype_0, x = clip_8_cast_fp16)[name = string("cast_323")];
            tensor<int32, [1, 1]> var_3135 = reshape(shape = var_3134, x = clip_8_cast_fp16_to_int32)[name = string("op_3135")];
            int32 var_3136 = const()[name = string("op_3136"), val = int32(1)];
            bool var_3138_validate_indices_0 = const()[name = string("op_3138_validate_indices_0"), val = bool(false)];
            string var_3135_to_uint16_dtype_0 = const()[name = string("op_3135_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 1]> var_3135_to_uint16 = cast(dtype = var_3135_to_uint16_dtype_0, x = var_3135)[name = string("cast_322")];
            tensor<uint16, [1, 1]> var_3138_cast_uint16 = gather_along_axis(axis = var_3136, indices = var_3135_to_uint16, validate_indices = var_3138_validate_indices_0, x = top_scores_33_cast_fp16_cast_uint16_1)[name = string("op_3138_cast_uint16")];
            string var_3138_cast_uint16_to_int32_dtype_0 = const()[name = string("op_3138_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [1]> var_3140 = const()[name = string("op_3140"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_3138_cast_uint16_to_int32 = cast(dtype = var_3138_cast_uint16_to_int32_dtype_0, x = var_3138_cast_uint16)[name = string("cast_321")];
            tensor<int32, [1]> token_17 = reshape(shape = var_3140, x = var_3138_cast_uint16_to_int32)[name = string("token_17")];
            int32 var_3150_axis_0 = const()[name = string("op_3150_axis_0"), val = int32(0)];
            int32 var_3150_batch_dims_0 = const()[name = string("op_3150_batch_dims_0"), val = int32(0)];
            bool var_3150_validate_indices_0 = const()[name = string("op_3150_validate_indices_0"), val = bool(false)];
            string token_17_to_uint16_dtype_0 = const()[name = string("token_17_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> token_17_to_uint16 = cast(dtype = token_17_to_uint16_dtype_0, x = token_17)[name = string("cast_320")];
            tensor<fp16, [1, 768]> var_3150_cast_fp16_cast_uint16 = gather(axis = var_3150_axis_0, batch_dims = var_3150_batch_dims_0, indices = token_17_to_uint16, validate_indices = var_3150_validate_indices_0, x = audio_embeddings_8_weight_to_fp16)[name = string("op_3150_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 768]> next_global_embed_19_cast_fp16 = add(x = next_global_embed_17_cast_fp16, y = var_3150_cast_fp16_cast_uint16)[name = string("next_global_embed_19_cast_fp16")];
            tensor<int32, [3]> var_3160 = const()[name = string("op_3160"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 768]> input_199_cast_fp16 = reshape(shape = var_3160, x = var_3150_cast_fp16_cast_uint16)[name = string("input_199_cast_fp16")];
            tensor<int32, [1]> input_201_axes_0 = const()[name = string("input_201_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_3163_to_fp16 = const()[name = string("op_3163_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_201_cast_fp16 = layer_norm(axes = input_201_axes_0, beta = block_ln_1_bias_to_fp16, epsilon = var_3163_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_199_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<fp16, [1, 1, 2304]> linear_49_cast_fp16 = linear(bias = block_attn_c_attn_bias_to_fp16, weight = block_attn_c_attn_weight_to_fp16, x = input_201_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [3]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [3]>([768, 768, 768])];
            int32 var_3174_axis_0 = const()[name = string("op_3174_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 768]> var_3174_cast_fp16_0, tensor<fp16, [1, 1, 768]> var_3174_cast_fp16_1, tensor<fp16, [1, 1, 768]> var_3174_cast_fp16_2 = split(axis = var_3174_axis_0, split_sizes = tile_10, x = linear_49_cast_fp16)[name = string("op_3174_cast_fp16")];
            tensor<int32, [4]> var_3182 = const()[name = string("op_3182"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> query_83_cast_fp16 = reshape(shape = var_3182, x = var_3174_cast_fp16_0)[name = string("query_83_cast_fp16")];
            tensor<int32, [4]> var_3188 = const()[name = string("op_3188"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> key_101_cast_fp16 = reshape(shape = var_3188, x = var_3174_cast_fp16_1)[name = string("key_101_cast_fp16")];
            tensor<int32, [4]> var_3194 = const()[name = string("op_3194"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> value_81_cast_fp16 = reshape(shape = var_3194, x = var_3174_cast_fp16_2)[name = string("value_81_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_cos_10_to_fp16 = const()[name = string("rope_cos_10_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28345024)))];
            tensor<fp16, [1, 1, 12, 64]> var_3196_cast_fp16 = mul(x = query_83_cast_fp16, y = rope_cos_10_to_fp16)[name = string("op_3196_cast_fp16")];
            tensor<int32, [4]> var_3201_begin_0 = const()[name = string("op_3201_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_3201_end_0 = const()[name = string("op_3201_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_3201_end_mask_0 = const()[name = string("op_3201_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3201_stride_0 = const()[name = string("op_3201_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_3201_cast_fp16 = slice_by_index(begin = var_3201_begin_0, end = var_3201_end_0, end_mask = var_3201_end_mask_0, stride = var_3201_stride_0, x = query_83_cast_fp16)[name = string("op_3201_cast_fp16")];
            fp16 const_51_promoted_to_fp16 = const()[name = string("const_51_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_3202_cast_fp16 = mul(x = var_3201_cast_fp16, y = const_51_promoted_to_fp16)[name = string("op_3202_cast_fp16")];
            tensor<int32, [4]> var_3207_begin_0 = const()[name = string("op_3207_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3207_end_0 = const()[name = string("op_3207_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_3207_end_mask_0 = const()[name = string("op_3207_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3207_stride_0 = const()[name = string("op_3207_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_3207_cast_fp16 = slice_by_index(begin = var_3207_begin_0, end = var_3207_end_0, end_mask = var_3207_end_mask_0, stride = var_3207_stride_0, x = query_83_cast_fp16)[name = string("op_3207_cast_fp16")];
            int32 var_3210_axis_0 = const()[name = string("op_3210_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_3210_cast_fp16 = stack(axis = var_3210_axis_0, values = (var_3202_cast_fp16, var_3207_cast_fp16))[name = string("op_3210_cast_fp16")];
            tensor<int32, [4]> shape_31 = const()[name = string("shape_31"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_3211_cast_fp16 = reshape(shape = shape_31, x = var_3210_cast_fp16)[name = string("op_3211_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_sin_10_to_fp16 = const()[name = string("rope_sin_10_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28345216)))];
            tensor<fp16, [1, 1, 12, 64]> var_3212_cast_fp16 = mul(x = var_3211_cast_fp16, y = rope_sin_10_to_fp16)[name = string("op_3212_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> query_85_cast_fp16 = add(x = var_3196_cast_fp16, y = var_3212_cast_fp16)[name = string("query_85_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_3215_cast_fp16 = mul(x = key_101_cast_fp16, y = rope_cos_10_to_fp16)[name = string("op_3215_cast_fp16")];
            tensor<int32, [4]> var_3220_begin_0 = const()[name = string("op_3220_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_3220_end_0 = const()[name = string("op_3220_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_3220_end_mask_0 = const()[name = string("op_3220_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3220_stride_0 = const()[name = string("op_3220_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_3220_cast_fp16 = slice_by_index(begin = var_3220_begin_0, end = var_3220_end_0, end_mask = var_3220_end_mask_0, stride = var_3220_stride_0, x = key_101_cast_fp16)[name = string("op_3220_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_3221_cast_fp16 = mul(x = var_3220_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_3221_cast_fp16")];
            tensor<int32, [4]> var_3226_begin_0 = const()[name = string("op_3226_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3226_end_0 = const()[name = string("op_3226_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_3226_end_mask_0 = const()[name = string("op_3226_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3226_stride_0 = const()[name = string("op_3226_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_3226_cast_fp16 = slice_by_index(begin = var_3226_begin_0, end = var_3226_end_0, end_mask = var_3226_end_mask_0, stride = var_3226_stride_0, x = key_101_cast_fp16)[name = string("op_3226_cast_fp16")];
            int32 var_3229_axis_0 = const()[name = string("op_3229_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_3229_cast_fp16 = stack(axis = var_3229_axis_0, values = (var_3221_cast_fp16, var_3226_cast_fp16))[name = string("op_3229_cast_fp16")];
            tensor<int32, [4]> shape_32 = const()[name = string("shape_32"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_3230_cast_fp16 = reshape(shape = shape_32, x = var_3229_cast_fp16)[name = string("op_3230_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_3231_cast_fp16 = mul(x = var_3230_cast_fp16, y = rope_sin_10_to_fp16)[name = string("op_3231_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> key_103_cast_fp16 = add(x = var_3215_cast_fp16, y = var_3231_cast_fp16)[name = string("key_103_cast_fp16")];
            tensor<int32, [4]> var_3238 = const()[name = string("op_3238"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3244 = const()[name = string("op_3244"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3250 = const()[name = string("op_3250"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 var_3253 = const()[name = string("op_3253"), val = int32(2)];
            bool key_107_interleave_0 = const()[name = string("key_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> key_105_cast_fp16 = transpose(perm = var_3244, x = key_103_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 12, 11, 64]> key_107_cast_fp16 = concat(axis = var_3253, interleave = key_107_interleave_0, values = (key_97_cast_fp16, key_105_cast_fp16))[name = string("key_107_cast_fp16")];
            int32 var_3256 = const()[name = string("op_3256"), val = int32(2)];
            bool value_85_interleave_0 = const()[name = string("value_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> value_83_cast_fp16 = transpose(perm = var_3250, x = value_81_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 12, 11, 64]> value_85_cast_fp16 = concat(axis = var_3256, interleave = value_85_interleave_0, values = (value_77_cast_fp16, value_83_cast_fp16))[name = string("value_85_cast_fp16")];
            bool var_3261_transpose_x_1 = const()[name = string("op_3261_transpose_x_1"), val = bool(false)];
            bool var_3261_transpose_y_1 = const()[name = string("op_3261_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 12, 1, 64]> query_87_cast_fp16 = transpose(perm = var_3238, x = query_85_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 12, 1, 11]> var_3261_cast_fp16 = matmul(transpose_x = var_3261_transpose_x_1, transpose_y = var_3261_transpose_y_1, x = query_87_cast_fp16, y = key_107_cast_fp16)[name = string("op_3261_cast_fp16")];
            fp16 var_3262_to_fp16 = const()[name = string("op_3262_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 1, 11]> scores_39_cast_fp16 = mul(x = var_3261_cast_fp16, y = var_3262_to_fp16)[name = string("scores_39_cast_fp16")];
            int32 var_3264 = const()[name = string("op_3264"), val = int32(-1)];
            tensor<fp16, [1, 12, 1, 11]> probabilities_57_cast_fp16 = softmax(axis = var_3264, x = scores_39_cast_fp16)[name = string("probabilities_57_cast_fp16")];
            bool attention_21_transpose_x_0 = const()[name = string("attention_21_transpose_x_0"), val = bool(false)];
            bool attention_21_transpose_y_0 = const()[name = string("attention_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> attention_21_cast_fp16 = matmul(transpose_x = attention_21_transpose_x_0, transpose_y = attention_21_transpose_y_0, x = probabilities_57_cast_fp16, y = value_85_cast_fp16)[name = string("attention_21_cast_fp16")];
            tensor<int32, [4]> var_3272 = const()[name = string("op_3272"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3277 = const()[name = string("op_3277"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 12, 64]> var_3273_cast_fp16 = transpose(perm = var_3272, x = attention_21_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 1, 768]> input_203_cast_fp16 = reshape(shape = var_3277, x = var_3273_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_50_cast_fp16 = linear(bias = block_attn_c_proj_bias_to_fp16, weight = block_attn_c_proj_weight_to_fp16, x = input_203_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_205_cast_fp16 = add(x = input_199_cast_fp16, y = linear_50_cast_fp16)[name = string("input_205_cast_fp16")];
            tensor<int32, [1]> input_207_axes_0 = const()[name = string("input_207_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_3285_to_fp16 = const()[name = string("op_3285_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = block_ln_2_bias_to_fp16, epsilon = var_3285_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")];
            tensor<fp16, [1, 1, 3072]> linear_51_cast_fp16 = linear(bias = block_mlp_fc_in_bias_to_fp16, weight = block_mlp_fc_in_weight_to_fp16, x = input_207_cast_fp16)[name = string("linear_51_cast_fp16")];
            string input_211_mode_0 = const()[name = string("input_211_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1, 3072]> input_211_cast_fp16 = gelu(mode = input_211_mode_0, x = linear_51_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_52_cast_fp16 = linear(bias = block_mlp_fc_out_bias_to_fp16, weight = block_mlp_fc_out_weight_to_fp16, x = input_211_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_215_cast_fp16 = add(x = input_205_cast_fp16, y = linear_52_cast_fp16)[name = string("input_215_cast_fp16")];
            tensor<int32, [1]> var_3324_axes_0 = const()[name = string("op_3324_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_3319_to_fp16 = const()[name = string("op_3319_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> var_3324_cast_fp16 = layer_norm(axes = var_3324_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_3319_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_215_cast_fp16)[name = string("op_3324_cast_fp16")];
            tensor<int32, [3]> var_3332_begin_0 = const()[name = string("op_3332_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3332_end_0 = const()[name = string("op_3332_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_3332_end_mask_0 = const()[name = string("op_3332_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_3332_squeeze_mask_0 = const()[name = string("op_3332_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_3332_cast_fp16 = slice_by_index(begin = var_3332_begin_0, end = var_3332_end_0, end_mask = var_3332_end_mask_0, squeeze_mask = var_3332_squeeze_mask_0, x = var_3324_cast_fp16)[name = string("op_3332_cast_fp16")];
            tensor<fp16, [1024, 768]> audio_embeddings_9_weight_to_fp16 = const()[name = string("audio_embeddings_9_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28345408)))];
            tensor<fp16, [1, 1024]> linear_53_cast_fp16 = linear(bias = linear_8_bias_0_to_fp16, weight = audio_embeddings_9_weight_to_fp16, x = var_3332_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<int32, [3]> var_3347_begin_0 = const()[name = string("op_3347_begin_0"), val = tensor<int32, [3]>([0, 9, 0])];
            tensor<int32, [3]> var_3347_end_0 = const()[name = string("op_3347_end_0"), val = tensor<int32, [3]>([1, 10, 1024])];
            tensor<bool, [3]> var_3347_end_mask_0 = const()[name = string("op_3347_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_3347_squeeze_mask_0 = const()[name = string("op_3347_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<int32, [1, 1024]> var_3347 = slice_by_index(begin = var_3347_begin_0, end = var_3347_end_0, end_mask = var_3347_end_mask_0, squeeze_mask = var_3347_squeeze_mask_0, x = repetition_seen)[name = string("op_3347")];
            tensor<int32, [2]> random_u_19_begin_0 = const()[name = string("random_u_19_begin_0"), val = tensor<int32, [2]>([0, 9])];
            tensor<int32, [2]> random_u_19_end_0 = const()[name = string("random_u_19_end_0"), val = tensor<int32, [2]>([1, 10])];
            tensor<bool, [2]> random_u_19_end_mask_0 = const()[name = string("random_u_19_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<bool, [2]> random_u_19_squeeze_mask_0 = const()[name = string("random_u_19_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [1]> random_u_19_cast_fp16 = slice_by_index(begin = random_u_19_begin_0, end = random_u_19_end_0, end_mask = random_u_19_end_mask_0, squeeze_mask = random_u_19_squeeze_mask_0, x = audio_random_to_fp16)[name = string("random_u_19_cast_fp16")];
            fp16 var_3361_promoted_to_fp16 = const()[name = string("op_3361_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [1, 1024]> var_3362_cast_fp16 = less(x = linear_53_cast_fp16, y = var_3361_promoted_to_fp16)[name = string("op_3362_cast_fp16")];
            fp16 var_3363_to_fp16 = const()[name = string("op_3363_to_fp16"), val = fp16(0x1.334p+0)];
            tensor<fp16, [1, 1024]> var_3364_cast_fp16 = mul(x = linear_53_cast_fp16, y = var_3363_to_fp16)[name = string("op_3364_cast_fp16")];
            fp16 _inversed_3366_y_0_to_fp16 = const()[name = string("_inversed_3366_y_0_to_fp16"), val = fp16(0x1.aacp-1)];
            tensor<fp16, [1, 1024]> _inversed_3366_cast_fp16 = mul(x = linear_53_cast_fp16, y = _inversed_3366_y_0_to_fp16)[name = string("_inversed_3366_cast_fp16")];
            tensor<fp16, [1, 1024]> penalized_19_cast_fp16 = select(a = var_3364_cast_fp16, b = _inversed_3366_cast_fp16, cond = var_3362_cast_fp16)[name = string("penalized_19_cast_fp16")];
            string var_3372_dtype_0 = const()[name = string("op_3372_dtype_0"), val = string("bool")];
            tensor<bool, [1, 1024]> var_3372 = cast(dtype = var_3372_dtype_0, x = var_3347)[name = string("cast_319")];
            tensor<fp16, [1, 1024]> var_3373_cast_fp16 = select(a = penalized_19_cast_fp16, b = linear_53_cast_fp16, cond = var_3372)[name = string("op_3373_cast_fp16")];
            fp16 _inversed_scores_41_y_0_to_fp16 = const()[name = string("_inversed_scores_41_y_0_to_fp16"), val = fp16(0x1.4p+0)];
            tensor<fp16, [1, 1024]> _inversed_scores_41_cast_fp16 = mul(x = var_3373_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = string("_inversed_scores_41_cast_fp16")];
            int32 var_3376 = const()[name = string("op_3376"), val = int32(25)];
            int32 top_scores_37_axis_0 = const()[name = string("top_scores_37_axis_0"), val = int32(1)];
            bool top_scores_37_ascending_0 = const()[name = string("top_scores_37_ascending_0"), val = bool(false)];
            bool top_scores_37_sort_0 = const()[name = string("top_scores_37_sort_0"), val = bool(true)];
            bool top_scores_37_return_indices_0 = const()[name = string("top_scores_37_return_indices_0"), val = bool(true)];
            string top_scores_37_cast_fp16_cast_uint16_output_indices_dtype_0 = const()[name = string("top_scores_37_cast_fp16_cast_uint16_output_indices_dtype_0"), val = string("uint16")];
            tensor<fp16, [1, 25]> top_scores_37_cast_fp16_cast_uint16_0, tensor<uint16, [1, 25]> top_scores_37_cast_fp16_cast_uint16_1 = topk(ascending = top_scores_37_ascending_0, axis = top_scores_37_axis_0, k = var_3376, output_indices_dtype = top_scores_37_cast_fp16_cast_uint16_output_indices_dtype_0, return_indices = top_scores_37_return_indices_0, sort = top_scores_37_sort_0, x = _inversed_scores_41_cast_fp16)[name = string("top_scores_37_cast_fp16_cast_uint16")];
            int32 var_3382 = const()[name = string("op_3382"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_59_cast_fp16 = softmax(axis = var_3382, x = top_scores_37_cast_fp16_cast_uint16_0)[name = string("probabilities_59_cast_fp16")];
            int32 var_3385 = const()[name = string("op_3385"), val = int32(1)];
            bool cumulative_37_exclusive_0 = const()[name = string("cumulative_37_exclusive_0"), val = bool(false)];
            bool cumulative_37_reverse_0 = const()[name = string("cumulative_37_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_37_cast_fp16 = cumsum(axis = var_3385, exclusive = cumulative_37_exclusive_0, reverse = cumulative_37_reverse_0, x = probabilities_59_cast_fp16)[name = string("cumulative_37_cast_fp16")];
            tensor<fp16, [1, 25]> var_3389_cast_fp16 = sub(x = cumulative_37_cast_fp16, y = probabilities_59_cast_fp16)[name = string("op_3389_cast_fp16")];
            fp16 var_3390_to_fp16 = const()[name = string("op_3390_to_fp16"), val = fp16(0x1.e68p-1)];
            tensor<bool, [1, 25]> keep_19_cast_fp16 = less(x = var_3389_cast_fp16, y = var_3390_to_fp16)[name = string("keep_19_cast_fp16")];
            fp16 var_3398_value_0_to_fp16 = const()[name = string("op_3398_value_0_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 25]> var_3398_cast_fp16 = fill_like(ref_tensor = top_scores_37_cast_fp16_cast_uint16_0, value = var_3398_value_0_to_fp16)[name = string("op_3398_cast_fp16")];
            tensor<fp16, [1, 25]> top_scores_39_cast_fp16 = select(a = top_scores_37_cast_fp16_cast_uint16_0, b = var_3398_cast_fp16, cond = keep_19_cast_fp16)[name = string("top_scores_39_cast_fp16")];
            int32 var_3400 = const()[name = string("op_3400"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_61_cast_fp16 = softmax(axis = var_3400, x = top_scores_39_cast_fp16)[name = string("probabilities_61_cast_fp16")];
            int32 var_3403 = const()[name = string("op_3403"), val = int32(1)];
            bool cumulative_39_exclusive_0 = const()[name = string("cumulative_39_exclusive_0"), val = bool(false)];
            bool cumulative_39_reverse_0 = const()[name = string("cumulative_39_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_39_cast_fp16 = cumsum(axis = var_3403, exclusive = cumulative_39_exclusive_0, reverse = cumulative_39_reverse_0, x = probabilities_61_cast_fp16)[name = string("cumulative_39_cast_fp16")];
            tensor<int32, [2]> var_3408 = const()[name = string("op_3408"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_3409_cast_fp16 = reshape(shape = var_3408, x = random_u_19_cast_fp16)[name = string("op_3409_cast_fp16")];
            tensor<bool, [1, 25]> var_3410_cast_fp16 = less(x = cumulative_39_cast_fp16, y = var_3409_cast_fp16)[name = string("op_3410_cast_fp16")];
            string var_3415_dtype_0 = const()[name = string("op_3415_dtype_0"), val = string("int32")];
            tensor<int32, [1]> selected_37_axes_0 = const()[name = string("selected_37_axes_0"), val = tensor<int32, [1]>([1])];
            bool selected_37_keep_dims_0 = const()[name = string("selected_37_keep_dims_0"), val = bool(false)];
            tensor<int32, [1, 25]> var_3415 = cast(dtype = var_3415_dtype_0, x = var_3410_cast_fp16)[name = string("cast_318")];
            tensor<int32, [1]> selected_37 = reduce_sum(axes = selected_37_axes_0, keep_dims = selected_37_keep_dims_0, x = var_3415)[name = string("selected_37")];
            string selected_37_promoted_to_fp16_dtype_0 = const()[name = string("selected_37_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_53_to_fp16 = const()[name = string("const_53_to_fp16"), val = fp16(-inf)];
            fp16 var_3422_promoted_to_fp16 = const()[name = string("op_3422_promoted_to_fp16"), val = fp16(0x1.8p+4)];
            tensor<fp16, [1]> selected_37_to_fp16 = cast(dtype = selected_37_promoted_to_fp16_dtype_0, x = selected_37)[name = string("cast_317")];
            tensor<fp16, [1]> clip_9_cast_fp16 = clip(alpha = const_53_to_fp16, beta = var_3422_promoted_to_fp16, x = selected_37_to_fp16)[name = string("clip_9_cast_fp16")];
            string selected_39_dtype_0 = const()[name = string("selected_39_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_3431 = const()[name = string("op_3431"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> clip_9_cast_fp16_to_int32 = cast(dtype = selected_39_dtype_0, x = clip_9_cast_fp16)[name = string("cast_316")];
            tensor<int32, [1, 1]> var_3432 = reshape(shape = var_3431, x = clip_9_cast_fp16_to_int32)[name = string("op_3432")];
            int32 var_3433 = const()[name = string("op_3433"), val = int32(1)];
            bool var_3435_validate_indices_0 = const()[name = string("op_3435_validate_indices_0"), val = bool(false)];
            string var_3432_to_uint16_dtype_0 = const()[name = string("op_3432_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 1]> var_3432_to_uint16 = cast(dtype = var_3432_to_uint16_dtype_0, x = var_3432)[name = string("cast_315")];
            tensor<uint16, [1, 1]> var_3435_cast_uint16 = gather_along_axis(axis = var_3433, indices = var_3432_to_uint16, validate_indices = var_3435_validate_indices_0, x = top_scores_37_cast_fp16_cast_uint16_1)[name = string("op_3435_cast_uint16")];
            string var_3435_cast_uint16_to_int32_dtype_0 = const()[name = string("op_3435_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [1]> var_3437 = const()[name = string("op_3437"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_3435_cast_uint16_to_int32 = cast(dtype = var_3435_cast_uint16_to_int32_dtype_0, x = var_3435_cast_uint16)[name = string("cast_314")];
            tensor<int32, [1]> token_19 = reshape(shape = var_3437, x = var_3435_cast_uint16_to_int32)[name = string("token_19")];
            int32 var_3447_axis_0 = const()[name = string("op_3447_axis_0"), val = int32(0)];
            int32 var_3447_batch_dims_0 = const()[name = string("op_3447_batch_dims_0"), val = int32(0)];
            bool var_3447_validate_indices_0 = const()[name = string("op_3447_validate_indices_0"), val = bool(false)];
            string token_19_to_uint16_dtype_0 = const()[name = string("token_19_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> token_19_to_uint16 = cast(dtype = token_19_to_uint16_dtype_0, x = token_19)[name = string("cast_313")];
            tensor<fp16, [1, 768]> var_3447_cast_fp16_cast_uint16 = gather(axis = var_3447_axis_0, batch_dims = var_3447_batch_dims_0, indices = token_19_to_uint16, validate_indices = var_3447_validate_indices_0, x = audio_embeddings_9_weight_to_fp16)[name = string("op_3447_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 768]> next_global_embed_21_cast_fp16 = add(x = next_global_embed_19_cast_fp16, y = var_3447_cast_fp16_cast_uint16)[name = string("next_global_embed_21_cast_fp16")];
            tensor<int32, [3]> var_3457 = const()[name = string("op_3457"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 768]> input_219_cast_fp16 = reshape(shape = var_3457, x = var_3447_cast_fp16_cast_uint16)[name = string("input_219_cast_fp16")];
            tensor<int32, [1]> input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_3460_to_fp16 = const()[name = string("op_3460_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = block_ln_1_bias_to_fp16, epsilon = var_3460_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [1, 1, 2304]> linear_54_cast_fp16 = linear(bias = block_attn_c_attn_bias_to_fp16, weight = block_attn_c_attn_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<int32, [3]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [3]>([768, 768, 768])];
            int32 var_3471_axis_0 = const()[name = string("op_3471_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 768]> var_3471_cast_fp16_0, tensor<fp16, [1, 1, 768]> var_3471_cast_fp16_1, tensor<fp16, [1, 1, 768]> var_3471_cast_fp16_2 = split(axis = var_3471_axis_0, split_sizes = tile_11, x = linear_54_cast_fp16)[name = string("op_3471_cast_fp16")];
            tensor<int32, [4]> var_3479 = const()[name = string("op_3479"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> query_91_cast_fp16 = reshape(shape = var_3479, x = var_3471_cast_fp16_0)[name = string("query_91_cast_fp16")];
            tensor<int32, [4]> var_3485 = const()[name = string("op_3485"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> key_111_cast_fp16 = reshape(shape = var_3485, x = var_3471_cast_fp16_1)[name = string("key_111_cast_fp16")];
            tensor<int32, [4]> var_3491 = const()[name = string("op_3491"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> value_89_cast_fp16 = reshape(shape = var_3491, x = var_3471_cast_fp16_2)[name = string("value_89_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_cos_11_to_fp16 = const()[name = string("rope_cos_11_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29918336)))];
            tensor<fp16, [1, 1, 12, 64]> var_3493_cast_fp16 = mul(x = query_91_cast_fp16, y = rope_cos_11_to_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [4]> var_3498_begin_0 = const()[name = string("op_3498_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_3498_end_0 = const()[name = string("op_3498_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_3498_end_mask_0 = const()[name = string("op_3498_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3498_stride_0 = const()[name = string("op_3498_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_3498_cast_fp16 = slice_by_index(begin = var_3498_begin_0, end = var_3498_end_0, end_mask = var_3498_end_mask_0, stride = var_3498_stride_0, x = query_91_cast_fp16)[name = string("op_3498_cast_fp16")];
            fp16 const_56_promoted_to_fp16 = const()[name = string("const_56_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_3499_cast_fp16 = mul(x = var_3498_cast_fp16, y = const_56_promoted_to_fp16)[name = string("op_3499_cast_fp16")];
            tensor<int32, [4]> var_3504_begin_0 = const()[name = string("op_3504_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3504_end_0 = const()[name = string("op_3504_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_3504_end_mask_0 = const()[name = string("op_3504_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3504_stride_0 = const()[name = string("op_3504_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_3504_cast_fp16 = slice_by_index(begin = var_3504_begin_0, end = var_3504_end_0, end_mask = var_3504_end_mask_0, stride = var_3504_stride_0, x = query_91_cast_fp16)[name = string("op_3504_cast_fp16")];
            int32 var_3507_axis_0 = const()[name = string("op_3507_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_3507_cast_fp16 = stack(axis = var_3507_axis_0, values = (var_3499_cast_fp16, var_3504_cast_fp16))[name = string("op_3507_cast_fp16")];
            tensor<int32, [4]> shape_34 = const()[name = string("shape_34"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_3508_cast_fp16 = reshape(shape = shape_34, x = var_3507_cast_fp16)[name = string("op_3508_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_sin_11_to_fp16 = const()[name = string("rope_sin_11_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29918528)))];
            tensor<fp16, [1, 1, 12, 64]> var_3509_cast_fp16 = mul(x = var_3508_cast_fp16, y = rope_sin_11_to_fp16)[name = string("op_3509_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> query_93_cast_fp16 = add(x = var_3493_cast_fp16, y = var_3509_cast_fp16)[name = string("query_93_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_3512_cast_fp16 = mul(x = key_111_cast_fp16, y = rope_cos_11_to_fp16)[name = string("op_3512_cast_fp16")];
            tensor<int32, [4]> var_3517_begin_0 = const()[name = string("op_3517_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_3517_end_0 = const()[name = string("op_3517_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_3517_end_mask_0 = const()[name = string("op_3517_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3517_stride_0 = const()[name = string("op_3517_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_3517_cast_fp16 = slice_by_index(begin = var_3517_begin_0, end = var_3517_end_0, end_mask = var_3517_end_mask_0, stride = var_3517_stride_0, x = key_111_cast_fp16)[name = string("op_3517_cast_fp16")];
            fp16 const_57_promoted_to_fp16 = const()[name = string("const_57_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_3518_cast_fp16 = mul(x = var_3517_cast_fp16, y = const_57_promoted_to_fp16)[name = string("op_3518_cast_fp16")];
            tensor<int32, [4]> var_3523_begin_0 = const()[name = string("op_3523_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3523_end_0 = const()[name = string("op_3523_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_3523_end_mask_0 = const()[name = string("op_3523_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3523_stride_0 = const()[name = string("op_3523_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_3523_cast_fp16 = slice_by_index(begin = var_3523_begin_0, end = var_3523_end_0, end_mask = var_3523_end_mask_0, stride = var_3523_stride_0, x = key_111_cast_fp16)[name = string("op_3523_cast_fp16")];
            int32 var_3526_axis_0 = const()[name = string("op_3526_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_3526_cast_fp16 = stack(axis = var_3526_axis_0, values = (var_3518_cast_fp16, var_3523_cast_fp16))[name = string("op_3526_cast_fp16")];
            tensor<int32, [4]> shape_35 = const()[name = string("shape_35"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_3527_cast_fp16 = reshape(shape = shape_35, x = var_3526_cast_fp16)[name = string("op_3527_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_3528_cast_fp16 = mul(x = var_3527_cast_fp16, y = rope_sin_11_to_fp16)[name = string("op_3528_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> key_113_cast_fp16 = add(x = var_3512_cast_fp16, y = var_3528_cast_fp16)[name = string("key_113_cast_fp16")];
            tensor<int32, [4]> var_3535 = const()[name = string("op_3535"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3541 = const()[name = string("op_3541"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3547 = const()[name = string("op_3547"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 var_3550 = const()[name = string("op_3550"), val = int32(2)];
            bool key_117_interleave_0 = const()[name = string("key_117_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> key_115_cast_fp16 = transpose(perm = var_3541, x = key_113_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 12, 12, 64]> key_117_cast_fp16 = concat(axis = var_3550, interleave = key_117_interleave_0, values = (key_107_cast_fp16, key_115_cast_fp16))[name = string("key_117_cast_fp16")];
            int32 var_3553 = const()[name = string("op_3553"), val = int32(2)];
            bool value_93_interleave_0 = const()[name = string("value_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> value_91_cast_fp16 = transpose(perm = var_3547, x = value_89_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 12, 12, 64]> value_93_cast_fp16 = concat(axis = var_3553, interleave = value_93_interleave_0, values = (value_85_cast_fp16, value_91_cast_fp16))[name = string("value_93_cast_fp16")];
            bool var_3558_transpose_x_1 = const()[name = string("op_3558_transpose_x_1"), val = bool(false)];
            bool var_3558_transpose_y_1 = const()[name = string("op_3558_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 12, 1, 64]> query_95_cast_fp16 = transpose(perm = var_3535, x = query_93_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 12, 1, 12]> var_3558_cast_fp16 = matmul(transpose_x = var_3558_transpose_x_1, transpose_y = var_3558_transpose_y_1, x = query_95_cast_fp16, y = key_117_cast_fp16)[name = string("op_3558_cast_fp16")];
            fp16 var_3559_to_fp16 = const()[name = string("op_3559_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 1, 12]> scores_43_cast_fp16 = mul(x = var_3558_cast_fp16, y = var_3559_to_fp16)[name = string("scores_43_cast_fp16")];
            int32 var_3561 = const()[name = string("op_3561"), val = int32(-1)];
            tensor<fp16, [1, 12, 1, 12]> probabilities_63_cast_fp16 = softmax(axis = var_3561, x = scores_43_cast_fp16)[name = string("probabilities_63_cast_fp16")];
            bool attention_23_transpose_x_0 = const()[name = string("attention_23_transpose_x_0"), val = bool(false)];
            bool attention_23_transpose_y_0 = const()[name = string("attention_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> attention_23_cast_fp16 = matmul(transpose_x = attention_23_transpose_x_0, transpose_y = attention_23_transpose_y_0, x = probabilities_63_cast_fp16, y = value_93_cast_fp16)[name = string("attention_23_cast_fp16")];
            tensor<int32, [4]> var_3569 = const()[name = string("op_3569"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3574 = const()[name = string("op_3574"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 12, 64]> var_3570_cast_fp16 = transpose(perm = var_3569, x = attention_23_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 1, 768]> input_223_cast_fp16 = reshape(shape = var_3574, x = var_3570_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_55_cast_fp16 = linear(bias = block_attn_c_proj_bias_to_fp16, weight = block_attn_c_proj_weight_to_fp16, x = input_223_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_225_cast_fp16 = add(x = input_219_cast_fp16, y = linear_55_cast_fp16)[name = string("input_225_cast_fp16")];
            tensor<int32, [1]> input_227_axes_0 = const()[name = string("input_227_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_3582_to_fp16 = const()[name = string("op_3582_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = block_ln_2_bias_to_fp16, epsilon = var_3582_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_225_cast_fp16)[name = string("input_227_cast_fp16")];
            tensor<fp16, [1, 1, 3072]> linear_56_cast_fp16 = linear(bias = block_mlp_fc_in_bias_to_fp16, weight = block_mlp_fc_in_weight_to_fp16, x = input_227_cast_fp16)[name = string("linear_56_cast_fp16")];
            string input_231_mode_0 = const()[name = string("input_231_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1, 3072]> input_231_cast_fp16 = gelu(mode = input_231_mode_0, x = linear_56_cast_fp16)[name = string("input_231_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_57_cast_fp16 = linear(bias = block_mlp_fc_out_bias_to_fp16, weight = block_mlp_fc_out_weight_to_fp16, x = input_231_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_235_cast_fp16 = add(x = input_225_cast_fp16, y = linear_57_cast_fp16)[name = string("input_235_cast_fp16")];
            tensor<int32, [1]> var_3621_axes_0 = const()[name = string("op_3621_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_3616_to_fp16 = const()[name = string("op_3616_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> var_3621_cast_fp16 = layer_norm(axes = var_3621_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_3616_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_235_cast_fp16)[name = string("op_3621_cast_fp16")];
            tensor<int32, [3]> var_3629_begin_0 = const()[name = string("op_3629_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3629_end_0 = const()[name = string("op_3629_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_3629_end_mask_0 = const()[name = string("op_3629_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_3629_squeeze_mask_0 = const()[name = string("op_3629_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_3629_cast_fp16 = slice_by_index(begin = var_3629_begin_0, end = var_3629_end_0, end_mask = var_3629_end_mask_0, squeeze_mask = var_3629_squeeze_mask_0, x = var_3621_cast_fp16)[name = string("op_3629_cast_fp16")];
            tensor<fp16, [1024, 768]> audio_embeddings_10_weight_to_fp16 = const()[name = string("audio_embeddings_10_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29918720)))];
            tensor<fp16, [1, 1024]> linear_58_cast_fp16 = linear(bias = linear_8_bias_0_to_fp16, weight = audio_embeddings_10_weight_to_fp16, x = var_3629_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [3]> var_3644_begin_0 = const()[name = string("op_3644_begin_0"), val = tensor<int32, [3]>([0, 10, 0])];
            tensor<int32, [3]> var_3644_end_0 = const()[name = string("op_3644_end_0"), val = tensor<int32, [3]>([1, 11, 1024])];
            tensor<bool, [3]> var_3644_end_mask_0 = const()[name = string("op_3644_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_3644_squeeze_mask_0 = const()[name = string("op_3644_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<int32, [1, 1024]> var_3644 = slice_by_index(begin = var_3644_begin_0, end = var_3644_end_0, end_mask = var_3644_end_mask_0, squeeze_mask = var_3644_squeeze_mask_0, x = repetition_seen)[name = string("op_3644")];
            tensor<int32, [2]> random_u_21_begin_0 = const()[name = string("random_u_21_begin_0"), val = tensor<int32, [2]>([0, 10])];
            tensor<int32, [2]> random_u_21_end_0 = const()[name = string("random_u_21_end_0"), val = tensor<int32, [2]>([1, 11])];
            tensor<bool, [2]> random_u_21_end_mask_0 = const()[name = string("random_u_21_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<bool, [2]> random_u_21_squeeze_mask_0 = const()[name = string("random_u_21_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [1]> random_u_21_cast_fp16 = slice_by_index(begin = random_u_21_begin_0, end = random_u_21_end_0, end_mask = random_u_21_end_mask_0, squeeze_mask = random_u_21_squeeze_mask_0, x = audio_random_to_fp16)[name = string("random_u_21_cast_fp16")];
            fp16 var_3658_promoted_to_fp16 = const()[name = string("op_3658_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [1, 1024]> var_3659_cast_fp16 = less(x = linear_58_cast_fp16, y = var_3658_promoted_to_fp16)[name = string("op_3659_cast_fp16")];
            fp16 var_3660_to_fp16 = const()[name = string("op_3660_to_fp16"), val = fp16(0x1.334p+0)];
            tensor<fp16, [1, 1024]> var_3661_cast_fp16 = mul(x = linear_58_cast_fp16, y = var_3660_to_fp16)[name = string("op_3661_cast_fp16")];
            fp16 _inversed_3663_y_0_to_fp16 = const()[name = string("_inversed_3663_y_0_to_fp16"), val = fp16(0x1.aacp-1)];
            tensor<fp16, [1, 1024]> _inversed_3663_cast_fp16 = mul(x = linear_58_cast_fp16, y = _inversed_3663_y_0_to_fp16)[name = string("_inversed_3663_cast_fp16")];
            tensor<fp16, [1, 1024]> penalized_21_cast_fp16 = select(a = var_3661_cast_fp16, b = _inversed_3663_cast_fp16, cond = var_3659_cast_fp16)[name = string("penalized_21_cast_fp16")];
            string var_3669_dtype_0 = const()[name = string("op_3669_dtype_0"), val = string("bool")];
            tensor<bool, [1, 1024]> var_3669 = cast(dtype = var_3669_dtype_0, x = var_3644)[name = string("cast_312")];
            tensor<fp16, [1, 1024]> var_3670_cast_fp16 = select(a = penalized_21_cast_fp16, b = linear_58_cast_fp16, cond = var_3669)[name = string("op_3670_cast_fp16")];
            fp16 _inversed_scores_45_y_0_to_fp16 = const()[name = string("_inversed_scores_45_y_0_to_fp16"), val = fp16(0x1.4p+0)];
            tensor<fp16, [1, 1024]> _inversed_scores_45_cast_fp16 = mul(x = var_3670_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = string("_inversed_scores_45_cast_fp16")];
            int32 var_3673 = const()[name = string("op_3673"), val = int32(25)];
            int32 top_scores_41_axis_0 = const()[name = string("top_scores_41_axis_0"), val = int32(1)];
            bool top_scores_41_ascending_0 = const()[name = string("top_scores_41_ascending_0"), val = bool(false)];
            bool top_scores_41_sort_0 = const()[name = string("top_scores_41_sort_0"), val = bool(true)];
            bool top_scores_41_return_indices_0 = const()[name = string("top_scores_41_return_indices_0"), val = bool(true)];
            string top_scores_41_cast_fp16_cast_uint16_output_indices_dtype_0 = const()[name = string("top_scores_41_cast_fp16_cast_uint16_output_indices_dtype_0"), val = string("uint16")];
            tensor<fp16, [1, 25]> top_scores_41_cast_fp16_cast_uint16_0, tensor<uint16, [1, 25]> top_scores_41_cast_fp16_cast_uint16_1 = topk(ascending = top_scores_41_ascending_0, axis = top_scores_41_axis_0, k = var_3673, output_indices_dtype = top_scores_41_cast_fp16_cast_uint16_output_indices_dtype_0, return_indices = top_scores_41_return_indices_0, sort = top_scores_41_sort_0, x = _inversed_scores_45_cast_fp16)[name = string("top_scores_41_cast_fp16_cast_uint16")];
            int32 var_3679 = const()[name = string("op_3679"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_65_cast_fp16 = softmax(axis = var_3679, x = top_scores_41_cast_fp16_cast_uint16_0)[name = string("probabilities_65_cast_fp16")];
            int32 var_3682 = const()[name = string("op_3682"), val = int32(1)];
            bool cumulative_41_exclusive_0 = const()[name = string("cumulative_41_exclusive_0"), val = bool(false)];
            bool cumulative_41_reverse_0 = const()[name = string("cumulative_41_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_41_cast_fp16 = cumsum(axis = var_3682, exclusive = cumulative_41_exclusive_0, reverse = cumulative_41_reverse_0, x = probabilities_65_cast_fp16)[name = string("cumulative_41_cast_fp16")];
            tensor<fp16, [1, 25]> var_3686_cast_fp16 = sub(x = cumulative_41_cast_fp16, y = probabilities_65_cast_fp16)[name = string("op_3686_cast_fp16")];
            fp16 var_3687_to_fp16 = const()[name = string("op_3687_to_fp16"), val = fp16(0x1.e68p-1)];
            tensor<bool, [1, 25]> keep_21_cast_fp16 = less(x = var_3686_cast_fp16, y = var_3687_to_fp16)[name = string("keep_21_cast_fp16")];
            fp16 var_3695_value_0_to_fp16 = const()[name = string("op_3695_value_0_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 25]> var_3695_cast_fp16 = fill_like(ref_tensor = top_scores_41_cast_fp16_cast_uint16_0, value = var_3695_value_0_to_fp16)[name = string("op_3695_cast_fp16")];
            tensor<fp16, [1, 25]> top_scores_43_cast_fp16 = select(a = top_scores_41_cast_fp16_cast_uint16_0, b = var_3695_cast_fp16, cond = keep_21_cast_fp16)[name = string("top_scores_43_cast_fp16")];
            int32 var_3697 = const()[name = string("op_3697"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_67_cast_fp16 = softmax(axis = var_3697, x = top_scores_43_cast_fp16)[name = string("probabilities_67_cast_fp16")];
            int32 var_3700 = const()[name = string("op_3700"), val = int32(1)];
            bool cumulative_43_exclusive_0 = const()[name = string("cumulative_43_exclusive_0"), val = bool(false)];
            bool cumulative_43_reverse_0 = const()[name = string("cumulative_43_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_43_cast_fp16 = cumsum(axis = var_3700, exclusive = cumulative_43_exclusive_0, reverse = cumulative_43_reverse_0, x = probabilities_67_cast_fp16)[name = string("cumulative_43_cast_fp16")];
            tensor<int32, [2]> var_3705 = const()[name = string("op_3705"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_3706_cast_fp16 = reshape(shape = var_3705, x = random_u_21_cast_fp16)[name = string("op_3706_cast_fp16")];
            tensor<bool, [1, 25]> var_3707_cast_fp16 = less(x = cumulative_43_cast_fp16, y = var_3706_cast_fp16)[name = string("op_3707_cast_fp16")];
            string var_3712_dtype_0 = const()[name = string("op_3712_dtype_0"), val = string("int32")];
            tensor<int32, [1]> selected_41_axes_0 = const()[name = string("selected_41_axes_0"), val = tensor<int32, [1]>([1])];
            bool selected_41_keep_dims_0 = const()[name = string("selected_41_keep_dims_0"), val = bool(false)];
            tensor<int32, [1, 25]> var_3712 = cast(dtype = var_3712_dtype_0, x = var_3707_cast_fp16)[name = string("cast_311")];
            tensor<int32, [1]> selected_41 = reduce_sum(axes = selected_41_axes_0, keep_dims = selected_41_keep_dims_0, x = var_3712)[name = string("selected_41")];
            string selected_41_promoted_to_fp16_dtype_0 = const()[name = string("selected_41_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_58_to_fp16 = const()[name = string("const_58_to_fp16"), val = fp16(-inf)];
            fp16 var_3719_promoted_to_fp16 = const()[name = string("op_3719_promoted_to_fp16"), val = fp16(0x1.8p+4)];
            tensor<fp16, [1]> selected_41_to_fp16 = cast(dtype = selected_41_promoted_to_fp16_dtype_0, x = selected_41)[name = string("cast_310")];
            tensor<fp16, [1]> clip_10_cast_fp16 = clip(alpha = const_58_to_fp16, beta = var_3719_promoted_to_fp16, x = selected_41_to_fp16)[name = string("clip_10_cast_fp16")];
            string selected_43_dtype_0 = const()[name = string("selected_43_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_3728 = const()[name = string("op_3728"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> clip_10_cast_fp16_to_int32 = cast(dtype = selected_43_dtype_0, x = clip_10_cast_fp16)[name = string("cast_309")];
            tensor<int32, [1, 1]> var_3729 = reshape(shape = var_3728, x = clip_10_cast_fp16_to_int32)[name = string("op_3729")];
            int32 var_3730 = const()[name = string("op_3730"), val = int32(1)];
            bool var_3732_validate_indices_0 = const()[name = string("op_3732_validate_indices_0"), val = bool(false)];
            string var_3729_to_uint16_dtype_0 = const()[name = string("op_3729_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 1]> var_3729_to_uint16 = cast(dtype = var_3729_to_uint16_dtype_0, x = var_3729)[name = string("cast_308")];
            tensor<uint16, [1, 1]> var_3732_cast_uint16 = gather_along_axis(axis = var_3730, indices = var_3729_to_uint16, validate_indices = var_3732_validate_indices_0, x = top_scores_41_cast_fp16_cast_uint16_1)[name = string("op_3732_cast_uint16")];
            string var_3732_cast_uint16_to_int32_dtype_0 = const()[name = string("op_3732_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [1]> var_3734 = const()[name = string("op_3734"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_3732_cast_uint16_to_int32 = cast(dtype = var_3732_cast_uint16_to_int32_dtype_0, x = var_3732_cast_uint16)[name = string("cast_307")];
            tensor<int32, [1]> token_21 = reshape(shape = var_3734, x = var_3732_cast_uint16_to_int32)[name = string("token_21")];
            int32 var_3744_axis_0 = const()[name = string("op_3744_axis_0"), val = int32(0)];
            int32 var_3744_batch_dims_0 = const()[name = string("op_3744_batch_dims_0"), val = int32(0)];
            bool var_3744_validate_indices_0 = const()[name = string("op_3744_validate_indices_0"), val = bool(false)];
            string token_21_to_uint16_dtype_0 = const()[name = string("token_21_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> token_21_to_uint16 = cast(dtype = token_21_to_uint16_dtype_0, x = token_21)[name = string("cast_306")];
            tensor<fp16, [1, 768]> var_3744_cast_fp16_cast_uint16 = gather(axis = var_3744_axis_0, batch_dims = var_3744_batch_dims_0, indices = token_21_to_uint16, validate_indices = var_3744_validate_indices_0, x = audio_embeddings_10_weight_to_fp16)[name = string("op_3744_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 768]> next_global_embed_23_cast_fp16 = add(x = next_global_embed_21_cast_fp16, y = var_3744_cast_fp16_cast_uint16)[name = string("next_global_embed_23_cast_fp16")];
            tensor<int32, [3]> var_3754 = const()[name = string("op_3754"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 768]> input_239_cast_fp16 = reshape(shape = var_3754, x = var_3744_cast_fp16_cast_uint16)[name = string("input_239_cast_fp16")];
            tensor<int32, [1]> input_241_axes_0 = const()[name = string("input_241_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_3757_to_fp16 = const()[name = string("op_3757_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_241_cast_fp16 = layer_norm(axes = input_241_axes_0, beta = block_ln_1_bias_to_fp16, epsilon = var_3757_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_239_cast_fp16)[name = string("input_241_cast_fp16")];
            tensor<fp16, [1, 1, 2304]> linear_59_cast_fp16 = linear(bias = block_attn_c_attn_bias_to_fp16, weight = block_attn_c_attn_weight_to_fp16, x = input_241_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<int32, [3]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [3]>([768, 768, 768])];
            int32 var_3768_axis_0 = const()[name = string("op_3768_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 768]> var_3768_cast_fp16_0, tensor<fp16, [1, 1, 768]> var_3768_cast_fp16_1, tensor<fp16, [1, 1, 768]> var_3768_cast_fp16_2 = split(axis = var_3768_axis_0, split_sizes = tile_12, x = linear_59_cast_fp16)[name = string("op_3768_cast_fp16")];
            tensor<int32, [4]> var_3776 = const()[name = string("op_3776"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> query_99_cast_fp16 = reshape(shape = var_3776, x = var_3768_cast_fp16_0)[name = string("query_99_cast_fp16")];
            tensor<int32, [4]> var_3782 = const()[name = string("op_3782"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> key_121_cast_fp16 = reshape(shape = var_3782, x = var_3768_cast_fp16_1)[name = string("key_121_cast_fp16")];
            tensor<int32, [4]> var_3788 = const()[name = string("op_3788"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> value_97_cast_fp16 = reshape(shape = var_3788, x = var_3768_cast_fp16_2)[name = string("value_97_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_cos_12_to_fp16 = const()[name = string("rope_cos_12_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31491648)))];
            tensor<fp16, [1, 1, 12, 64]> var_3790_cast_fp16 = mul(x = query_99_cast_fp16, y = rope_cos_12_to_fp16)[name = string("op_3790_cast_fp16")];
            tensor<int32, [4]> var_3795_begin_0 = const()[name = string("op_3795_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_3795_end_0 = const()[name = string("op_3795_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_3795_end_mask_0 = const()[name = string("op_3795_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3795_stride_0 = const()[name = string("op_3795_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_3795_cast_fp16 = slice_by_index(begin = var_3795_begin_0, end = var_3795_end_0, end_mask = var_3795_end_mask_0, stride = var_3795_stride_0, x = query_99_cast_fp16)[name = string("op_3795_cast_fp16")];
            fp16 const_61_promoted_to_fp16 = const()[name = string("const_61_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_3796_cast_fp16 = mul(x = var_3795_cast_fp16, y = const_61_promoted_to_fp16)[name = string("op_3796_cast_fp16")];
            tensor<int32, [4]> var_3801_begin_0 = const()[name = string("op_3801_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3801_end_0 = const()[name = string("op_3801_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_3801_end_mask_0 = const()[name = string("op_3801_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3801_stride_0 = const()[name = string("op_3801_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_3801_cast_fp16 = slice_by_index(begin = var_3801_begin_0, end = var_3801_end_0, end_mask = var_3801_end_mask_0, stride = var_3801_stride_0, x = query_99_cast_fp16)[name = string("op_3801_cast_fp16")];
            int32 var_3804_axis_0 = const()[name = string("op_3804_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_3804_cast_fp16 = stack(axis = var_3804_axis_0, values = (var_3796_cast_fp16, var_3801_cast_fp16))[name = string("op_3804_cast_fp16")];
            tensor<int32, [4]> shape_37 = const()[name = string("shape_37"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_3805_cast_fp16 = reshape(shape = shape_37, x = var_3804_cast_fp16)[name = string("op_3805_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_sin_12_to_fp16 = const()[name = string("rope_sin_12_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31491840)))];
            tensor<fp16, [1, 1, 12, 64]> var_3806_cast_fp16 = mul(x = var_3805_cast_fp16, y = rope_sin_12_to_fp16)[name = string("op_3806_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> query_101_cast_fp16 = add(x = var_3790_cast_fp16, y = var_3806_cast_fp16)[name = string("query_101_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_3809_cast_fp16 = mul(x = key_121_cast_fp16, y = rope_cos_12_to_fp16)[name = string("op_3809_cast_fp16")];
            tensor<int32, [4]> var_3814_begin_0 = const()[name = string("op_3814_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_3814_end_0 = const()[name = string("op_3814_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_3814_end_mask_0 = const()[name = string("op_3814_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3814_stride_0 = const()[name = string("op_3814_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_3814_cast_fp16 = slice_by_index(begin = var_3814_begin_0, end = var_3814_end_0, end_mask = var_3814_end_mask_0, stride = var_3814_stride_0, x = key_121_cast_fp16)[name = string("op_3814_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_3815_cast_fp16 = mul(x = var_3814_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_3815_cast_fp16")];
            tensor<int32, [4]> var_3820_begin_0 = const()[name = string("op_3820_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3820_end_0 = const()[name = string("op_3820_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_3820_end_mask_0 = const()[name = string("op_3820_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3820_stride_0 = const()[name = string("op_3820_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_3820_cast_fp16 = slice_by_index(begin = var_3820_begin_0, end = var_3820_end_0, end_mask = var_3820_end_mask_0, stride = var_3820_stride_0, x = key_121_cast_fp16)[name = string("op_3820_cast_fp16")];
            int32 var_3823_axis_0 = const()[name = string("op_3823_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_3823_cast_fp16 = stack(axis = var_3823_axis_0, values = (var_3815_cast_fp16, var_3820_cast_fp16))[name = string("op_3823_cast_fp16")];
            tensor<int32, [4]> shape_38 = const()[name = string("shape_38"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_3824_cast_fp16 = reshape(shape = shape_38, x = var_3823_cast_fp16)[name = string("op_3824_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_3825_cast_fp16 = mul(x = var_3824_cast_fp16, y = rope_sin_12_to_fp16)[name = string("op_3825_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> key_123_cast_fp16 = add(x = var_3809_cast_fp16, y = var_3825_cast_fp16)[name = string("key_123_cast_fp16")];
            tensor<int32, [4]> var_3832 = const()[name = string("op_3832"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3838 = const()[name = string("op_3838"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3844 = const()[name = string("op_3844"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 var_3847 = const()[name = string("op_3847"), val = int32(2)];
            bool key_127_interleave_0 = const()[name = string("key_127_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> key_125_cast_fp16 = transpose(perm = var_3838, x = key_123_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 12, 13, 64]> key_127_cast_fp16 = concat(axis = var_3847, interleave = key_127_interleave_0, values = (key_117_cast_fp16, key_125_cast_fp16))[name = string("key_127_cast_fp16")];
            int32 var_3850 = const()[name = string("op_3850"), val = int32(2)];
            bool value_101_interleave_0 = const()[name = string("value_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> value_99_cast_fp16 = transpose(perm = var_3844, x = value_97_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 12, 13, 64]> value_101_cast_fp16 = concat(axis = var_3850, interleave = value_101_interleave_0, values = (value_93_cast_fp16, value_99_cast_fp16))[name = string("value_101_cast_fp16")];
            bool var_3855_transpose_x_1 = const()[name = string("op_3855_transpose_x_1"), val = bool(false)];
            bool var_3855_transpose_y_1 = const()[name = string("op_3855_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 12, 1, 64]> query_103_cast_fp16 = transpose(perm = var_3832, x = query_101_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 12, 1, 13]> var_3855_cast_fp16 = matmul(transpose_x = var_3855_transpose_x_1, transpose_y = var_3855_transpose_y_1, x = query_103_cast_fp16, y = key_127_cast_fp16)[name = string("op_3855_cast_fp16")];
            fp16 var_3856_to_fp16 = const()[name = string("op_3856_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 1, 13]> scores_47_cast_fp16 = mul(x = var_3855_cast_fp16, y = var_3856_to_fp16)[name = string("scores_47_cast_fp16")];
            int32 var_3858 = const()[name = string("op_3858"), val = int32(-1)];
            tensor<fp16, [1, 12, 1, 13]> probabilities_69_cast_fp16 = softmax(axis = var_3858, x = scores_47_cast_fp16)[name = string("probabilities_69_cast_fp16")];
            bool attention_25_transpose_x_0 = const()[name = string("attention_25_transpose_x_0"), val = bool(false)];
            bool attention_25_transpose_y_0 = const()[name = string("attention_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> attention_25_cast_fp16 = matmul(transpose_x = attention_25_transpose_x_0, transpose_y = attention_25_transpose_y_0, x = probabilities_69_cast_fp16, y = value_101_cast_fp16)[name = string("attention_25_cast_fp16")];
            tensor<int32, [4]> var_3866 = const()[name = string("op_3866"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3871 = const()[name = string("op_3871"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 12, 64]> var_3867_cast_fp16 = transpose(perm = var_3866, x = attention_25_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 1, 768]> input_243_cast_fp16 = reshape(shape = var_3871, x = var_3867_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_60_cast_fp16 = linear(bias = block_attn_c_proj_bias_to_fp16, weight = block_attn_c_proj_weight_to_fp16, x = input_243_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_245_cast_fp16 = add(x = input_239_cast_fp16, y = linear_60_cast_fp16)[name = string("input_245_cast_fp16")];
            tensor<int32, [1]> input_247_axes_0 = const()[name = string("input_247_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_3879_to_fp16 = const()[name = string("op_3879_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_247_cast_fp16 = layer_norm(axes = input_247_axes_0, beta = block_ln_2_bias_to_fp16, epsilon = var_3879_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_245_cast_fp16)[name = string("input_247_cast_fp16")];
            tensor<fp16, [1, 1, 3072]> linear_61_cast_fp16 = linear(bias = block_mlp_fc_in_bias_to_fp16, weight = block_mlp_fc_in_weight_to_fp16, x = input_247_cast_fp16)[name = string("linear_61_cast_fp16")];
            string input_251_mode_0 = const()[name = string("input_251_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1, 3072]> input_251_cast_fp16 = gelu(mode = input_251_mode_0, x = linear_61_cast_fp16)[name = string("input_251_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_62_cast_fp16 = linear(bias = block_mlp_fc_out_bias_to_fp16, weight = block_mlp_fc_out_weight_to_fp16, x = input_251_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_62_cast_fp16)[name = string("input_255_cast_fp16")];
            tensor<int32, [1]> var_3918_axes_0 = const()[name = string("op_3918_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_3913_to_fp16 = const()[name = string("op_3913_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> var_3918_cast_fp16 = layer_norm(axes = var_3918_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_3913_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_255_cast_fp16)[name = string("op_3918_cast_fp16")];
            tensor<int32, [3]> var_3926_begin_0 = const()[name = string("op_3926_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_3926_end_0 = const()[name = string("op_3926_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_3926_end_mask_0 = const()[name = string("op_3926_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_3926_squeeze_mask_0 = const()[name = string("op_3926_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_3926_cast_fp16 = slice_by_index(begin = var_3926_begin_0, end = var_3926_end_0, end_mask = var_3926_end_mask_0, squeeze_mask = var_3926_squeeze_mask_0, x = var_3918_cast_fp16)[name = string("op_3926_cast_fp16")];
            tensor<fp16, [1024, 768]> audio_embeddings_11_weight_to_fp16 = const()[name = string("audio_embeddings_11_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31492032)))];
            tensor<fp16, [1, 1024]> linear_63_cast_fp16 = linear(bias = linear_8_bias_0_to_fp16, weight = audio_embeddings_11_weight_to_fp16, x = var_3926_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [3]> var_3941_begin_0 = const()[name = string("op_3941_begin_0"), val = tensor<int32, [3]>([0, 11, 0])];
            tensor<int32, [3]> var_3941_end_0 = const()[name = string("op_3941_end_0"), val = tensor<int32, [3]>([1, 12, 1024])];
            tensor<bool, [3]> var_3941_end_mask_0 = const()[name = string("op_3941_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_3941_squeeze_mask_0 = const()[name = string("op_3941_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<int32, [1, 1024]> var_3941 = slice_by_index(begin = var_3941_begin_0, end = var_3941_end_0, end_mask = var_3941_end_mask_0, squeeze_mask = var_3941_squeeze_mask_0, x = repetition_seen)[name = string("op_3941")];
            tensor<int32, [2]> random_u_23_begin_0 = const()[name = string("random_u_23_begin_0"), val = tensor<int32, [2]>([0, 11])];
            tensor<int32, [2]> random_u_23_end_0 = const()[name = string("random_u_23_end_0"), val = tensor<int32, [2]>([1, 12])];
            tensor<bool, [2]> random_u_23_end_mask_0 = const()[name = string("random_u_23_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<bool, [2]> random_u_23_squeeze_mask_0 = const()[name = string("random_u_23_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [1]> random_u_23_cast_fp16 = slice_by_index(begin = random_u_23_begin_0, end = random_u_23_end_0, end_mask = random_u_23_end_mask_0, squeeze_mask = random_u_23_squeeze_mask_0, x = audio_random_to_fp16)[name = string("random_u_23_cast_fp16")];
            fp16 var_3955_promoted_to_fp16 = const()[name = string("op_3955_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [1, 1024]> var_3956_cast_fp16 = less(x = linear_63_cast_fp16, y = var_3955_promoted_to_fp16)[name = string("op_3956_cast_fp16")];
            fp16 var_3957_to_fp16 = const()[name = string("op_3957_to_fp16"), val = fp16(0x1.334p+0)];
            tensor<fp16, [1, 1024]> var_3958_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_3957_to_fp16)[name = string("op_3958_cast_fp16")];
            fp16 _inversed_3960_y_0_to_fp16 = const()[name = string("_inversed_3960_y_0_to_fp16"), val = fp16(0x1.aacp-1)];
            tensor<fp16, [1, 1024]> _inversed_3960_cast_fp16 = mul(x = linear_63_cast_fp16, y = _inversed_3960_y_0_to_fp16)[name = string("_inversed_3960_cast_fp16")];
            tensor<fp16, [1, 1024]> penalized_23_cast_fp16 = select(a = var_3958_cast_fp16, b = _inversed_3960_cast_fp16, cond = var_3956_cast_fp16)[name = string("penalized_23_cast_fp16")];
            string var_3966_dtype_0 = const()[name = string("op_3966_dtype_0"), val = string("bool")];
            tensor<bool, [1, 1024]> var_3966 = cast(dtype = var_3966_dtype_0, x = var_3941)[name = string("cast_305")];
            tensor<fp16, [1, 1024]> var_3967_cast_fp16 = select(a = penalized_23_cast_fp16, b = linear_63_cast_fp16, cond = var_3966)[name = string("op_3967_cast_fp16")];
            fp16 _inversed_scores_49_y_0_to_fp16 = const()[name = string("_inversed_scores_49_y_0_to_fp16"), val = fp16(0x1.4p+0)];
            tensor<fp16, [1, 1024]> _inversed_scores_49_cast_fp16 = mul(x = var_3967_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = string("_inversed_scores_49_cast_fp16")];
            int32 var_3970 = const()[name = string("op_3970"), val = int32(25)];
            int32 top_scores_45_axis_0 = const()[name = string("top_scores_45_axis_0"), val = int32(1)];
            bool top_scores_45_ascending_0 = const()[name = string("top_scores_45_ascending_0"), val = bool(false)];
            bool top_scores_45_sort_0 = const()[name = string("top_scores_45_sort_0"), val = bool(true)];
            bool top_scores_45_return_indices_0 = const()[name = string("top_scores_45_return_indices_0"), val = bool(true)];
            string top_scores_45_cast_fp16_cast_uint16_output_indices_dtype_0 = const()[name = string("top_scores_45_cast_fp16_cast_uint16_output_indices_dtype_0"), val = string("uint16")];
            tensor<fp16, [1, 25]> top_scores_45_cast_fp16_cast_uint16_0, tensor<uint16, [1, 25]> top_scores_45_cast_fp16_cast_uint16_1 = topk(ascending = top_scores_45_ascending_0, axis = top_scores_45_axis_0, k = var_3970, output_indices_dtype = top_scores_45_cast_fp16_cast_uint16_output_indices_dtype_0, return_indices = top_scores_45_return_indices_0, sort = top_scores_45_sort_0, x = _inversed_scores_49_cast_fp16)[name = string("top_scores_45_cast_fp16_cast_uint16")];
            int32 var_3976 = const()[name = string("op_3976"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_71_cast_fp16 = softmax(axis = var_3976, x = top_scores_45_cast_fp16_cast_uint16_0)[name = string("probabilities_71_cast_fp16")];
            int32 var_3979 = const()[name = string("op_3979"), val = int32(1)];
            bool cumulative_45_exclusive_0 = const()[name = string("cumulative_45_exclusive_0"), val = bool(false)];
            bool cumulative_45_reverse_0 = const()[name = string("cumulative_45_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_45_cast_fp16 = cumsum(axis = var_3979, exclusive = cumulative_45_exclusive_0, reverse = cumulative_45_reverse_0, x = probabilities_71_cast_fp16)[name = string("cumulative_45_cast_fp16")];
            tensor<fp16, [1, 25]> var_3983_cast_fp16 = sub(x = cumulative_45_cast_fp16, y = probabilities_71_cast_fp16)[name = string("op_3983_cast_fp16")];
            fp16 var_3984_to_fp16 = const()[name = string("op_3984_to_fp16"), val = fp16(0x1.e68p-1)];
            tensor<bool, [1, 25]> keep_23_cast_fp16 = less(x = var_3983_cast_fp16, y = var_3984_to_fp16)[name = string("keep_23_cast_fp16")];
            fp16 var_3992_value_0_to_fp16 = const()[name = string("op_3992_value_0_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 25]> var_3992_cast_fp16 = fill_like(ref_tensor = top_scores_45_cast_fp16_cast_uint16_0, value = var_3992_value_0_to_fp16)[name = string("op_3992_cast_fp16")];
            tensor<fp16, [1, 25]> top_scores_47_cast_fp16 = select(a = top_scores_45_cast_fp16_cast_uint16_0, b = var_3992_cast_fp16, cond = keep_23_cast_fp16)[name = string("top_scores_47_cast_fp16")];
            int32 var_3994 = const()[name = string("op_3994"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_73_cast_fp16 = softmax(axis = var_3994, x = top_scores_47_cast_fp16)[name = string("probabilities_73_cast_fp16")];
            int32 var_3997 = const()[name = string("op_3997"), val = int32(1)];
            bool cumulative_47_exclusive_0 = const()[name = string("cumulative_47_exclusive_0"), val = bool(false)];
            bool cumulative_47_reverse_0 = const()[name = string("cumulative_47_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_47_cast_fp16 = cumsum(axis = var_3997, exclusive = cumulative_47_exclusive_0, reverse = cumulative_47_reverse_0, x = probabilities_73_cast_fp16)[name = string("cumulative_47_cast_fp16")];
            tensor<int32, [2]> var_4002 = const()[name = string("op_4002"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_4003_cast_fp16 = reshape(shape = var_4002, x = random_u_23_cast_fp16)[name = string("op_4003_cast_fp16")];
            tensor<bool, [1, 25]> var_4004_cast_fp16 = less(x = cumulative_47_cast_fp16, y = var_4003_cast_fp16)[name = string("op_4004_cast_fp16")];
            string var_4009_dtype_0 = const()[name = string("op_4009_dtype_0"), val = string("int32")];
            tensor<int32, [1]> selected_45_axes_0 = const()[name = string("selected_45_axes_0"), val = tensor<int32, [1]>([1])];
            bool selected_45_keep_dims_0 = const()[name = string("selected_45_keep_dims_0"), val = bool(false)];
            tensor<int32, [1, 25]> var_4009 = cast(dtype = var_4009_dtype_0, x = var_4004_cast_fp16)[name = string("cast_304")];
            tensor<int32, [1]> selected_45 = reduce_sum(axes = selected_45_axes_0, keep_dims = selected_45_keep_dims_0, x = var_4009)[name = string("selected_45")];
            string selected_45_promoted_to_fp16_dtype_0 = const()[name = string("selected_45_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_63_to_fp16 = const()[name = string("const_63_to_fp16"), val = fp16(-inf)];
            fp16 var_4016_promoted_to_fp16 = const()[name = string("op_4016_promoted_to_fp16"), val = fp16(0x1.8p+4)];
            tensor<fp16, [1]> selected_45_to_fp16 = cast(dtype = selected_45_promoted_to_fp16_dtype_0, x = selected_45)[name = string("cast_303")];
            tensor<fp16, [1]> clip_11_cast_fp16 = clip(alpha = const_63_to_fp16, beta = var_4016_promoted_to_fp16, x = selected_45_to_fp16)[name = string("clip_11_cast_fp16")];
            string selected_47_dtype_0 = const()[name = string("selected_47_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_4025 = const()[name = string("op_4025"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> clip_11_cast_fp16_to_int32 = cast(dtype = selected_47_dtype_0, x = clip_11_cast_fp16)[name = string("cast_302")];
            tensor<int32, [1, 1]> var_4026 = reshape(shape = var_4025, x = clip_11_cast_fp16_to_int32)[name = string("op_4026")];
            int32 var_4027 = const()[name = string("op_4027"), val = int32(1)];
            bool var_4029_validate_indices_0 = const()[name = string("op_4029_validate_indices_0"), val = bool(false)];
            string var_4026_to_uint16_dtype_0 = const()[name = string("op_4026_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 1]> var_4026_to_uint16 = cast(dtype = var_4026_to_uint16_dtype_0, x = var_4026)[name = string("cast_301")];
            tensor<uint16, [1, 1]> var_4029_cast_uint16 = gather_along_axis(axis = var_4027, indices = var_4026_to_uint16, validate_indices = var_4029_validate_indices_0, x = top_scores_45_cast_fp16_cast_uint16_1)[name = string("op_4029_cast_uint16")];
            string var_4029_cast_uint16_to_int32_dtype_0 = const()[name = string("op_4029_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [1]> var_4031 = const()[name = string("op_4031"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_4029_cast_uint16_to_int32 = cast(dtype = var_4029_cast_uint16_to_int32_dtype_0, x = var_4029_cast_uint16)[name = string("cast_300")];
            tensor<int32, [1]> token_23 = reshape(shape = var_4031, x = var_4029_cast_uint16_to_int32)[name = string("token_23")];
            int32 var_4041_axis_0 = const()[name = string("op_4041_axis_0"), val = int32(0)];
            int32 var_4041_batch_dims_0 = const()[name = string("op_4041_batch_dims_0"), val = int32(0)];
            bool var_4041_validate_indices_0 = const()[name = string("op_4041_validate_indices_0"), val = bool(false)];
            string token_23_to_uint16_dtype_0 = const()[name = string("token_23_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> token_23_to_uint16 = cast(dtype = token_23_to_uint16_dtype_0, x = token_23)[name = string("cast_299")];
            tensor<fp16, [1, 768]> var_4041_cast_fp16_cast_uint16 = gather(axis = var_4041_axis_0, batch_dims = var_4041_batch_dims_0, indices = token_23_to_uint16, validate_indices = var_4041_validate_indices_0, x = audio_embeddings_11_weight_to_fp16)[name = string("op_4041_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 768]> next_global_embed_25_cast_fp16 = add(x = next_global_embed_23_cast_fp16, y = var_4041_cast_fp16_cast_uint16)[name = string("next_global_embed_25_cast_fp16")];
            tensor<int32, [3]> var_4051 = const()[name = string("op_4051"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 768]> input_259_cast_fp16 = reshape(shape = var_4051, x = var_4041_cast_fp16_cast_uint16)[name = string("input_259_cast_fp16")];
            tensor<int32, [1]> input_261_axes_0 = const()[name = string("input_261_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_4054_to_fp16 = const()[name = string("op_4054_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_261_cast_fp16 = layer_norm(axes = input_261_axes_0, beta = block_ln_1_bias_to_fp16, epsilon = var_4054_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_259_cast_fp16)[name = string("input_261_cast_fp16")];
            tensor<fp16, [1, 1, 2304]> linear_64_cast_fp16 = linear(bias = block_attn_c_attn_bias_to_fp16, weight = block_attn_c_attn_weight_to_fp16, x = input_261_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [3]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [3]>([768, 768, 768])];
            int32 var_4065_axis_0 = const()[name = string("op_4065_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 768]> var_4065_cast_fp16_0, tensor<fp16, [1, 1, 768]> var_4065_cast_fp16_1, tensor<fp16, [1, 1, 768]> var_4065_cast_fp16_2 = split(axis = var_4065_axis_0, split_sizes = tile_13, x = linear_64_cast_fp16)[name = string("op_4065_cast_fp16")];
            tensor<int32, [4]> var_4073 = const()[name = string("op_4073"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> query_107_cast_fp16 = reshape(shape = var_4073, x = var_4065_cast_fp16_0)[name = string("query_107_cast_fp16")];
            tensor<int32, [4]> var_4079 = const()[name = string("op_4079"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> key_131_cast_fp16 = reshape(shape = var_4079, x = var_4065_cast_fp16_1)[name = string("key_131_cast_fp16")];
            tensor<int32, [4]> var_4085 = const()[name = string("op_4085"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> value_105_cast_fp16 = reshape(shape = var_4085, x = var_4065_cast_fp16_2)[name = string("value_105_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_cos_13_to_fp16 = const()[name = string("rope_cos_13_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33064960)))];
            tensor<fp16, [1, 1, 12, 64]> var_4087_cast_fp16 = mul(x = query_107_cast_fp16, y = rope_cos_13_to_fp16)[name = string("op_4087_cast_fp16")];
            tensor<int32, [4]> var_4092_begin_0 = const()[name = string("op_4092_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_4092_end_0 = const()[name = string("op_4092_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_4092_end_mask_0 = const()[name = string("op_4092_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4092_stride_0 = const()[name = string("op_4092_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_4092_cast_fp16 = slice_by_index(begin = var_4092_begin_0, end = var_4092_end_0, end_mask = var_4092_end_mask_0, stride = var_4092_stride_0, x = query_107_cast_fp16)[name = string("op_4092_cast_fp16")];
            fp16 const_66_promoted_to_fp16 = const()[name = string("const_66_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_4093_cast_fp16 = mul(x = var_4092_cast_fp16, y = const_66_promoted_to_fp16)[name = string("op_4093_cast_fp16")];
            tensor<int32, [4]> var_4098_begin_0 = const()[name = string("op_4098_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4098_end_0 = const()[name = string("op_4098_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_4098_end_mask_0 = const()[name = string("op_4098_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4098_stride_0 = const()[name = string("op_4098_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_4098_cast_fp16 = slice_by_index(begin = var_4098_begin_0, end = var_4098_end_0, end_mask = var_4098_end_mask_0, stride = var_4098_stride_0, x = query_107_cast_fp16)[name = string("op_4098_cast_fp16")];
            int32 var_4101_axis_0 = const()[name = string("op_4101_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_4101_cast_fp16 = stack(axis = var_4101_axis_0, values = (var_4093_cast_fp16, var_4098_cast_fp16))[name = string("op_4101_cast_fp16")];
            tensor<int32, [4]> shape_40 = const()[name = string("shape_40"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_4102_cast_fp16 = reshape(shape = shape_40, x = var_4101_cast_fp16)[name = string("op_4102_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_sin_13_to_fp16 = const()[name = string("rope_sin_13_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33065152)))];
            tensor<fp16, [1, 1, 12, 64]> var_4103_cast_fp16 = mul(x = var_4102_cast_fp16, y = rope_sin_13_to_fp16)[name = string("op_4103_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> query_109_cast_fp16 = add(x = var_4087_cast_fp16, y = var_4103_cast_fp16)[name = string("query_109_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_4106_cast_fp16 = mul(x = key_131_cast_fp16, y = rope_cos_13_to_fp16)[name = string("op_4106_cast_fp16")];
            tensor<int32, [4]> var_4111_begin_0 = const()[name = string("op_4111_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_4111_end_0 = const()[name = string("op_4111_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_4111_end_mask_0 = const()[name = string("op_4111_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4111_stride_0 = const()[name = string("op_4111_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_4111_cast_fp16 = slice_by_index(begin = var_4111_begin_0, end = var_4111_end_0, end_mask = var_4111_end_mask_0, stride = var_4111_stride_0, x = key_131_cast_fp16)[name = string("op_4111_cast_fp16")];
            fp16 const_67_promoted_to_fp16 = const()[name = string("const_67_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_4112_cast_fp16 = mul(x = var_4111_cast_fp16, y = const_67_promoted_to_fp16)[name = string("op_4112_cast_fp16")];
            tensor<int32, [4]> var_4117_begin_0 = const()[name = string("op_4117_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4117_end_0 = const()[name = string("op_4117_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_4117_end_mask_0 = const()[name = string("op_4117_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4117_stride_0 = const()[name = string("op_4117_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_4117_cast_fp16 = slice_by_index(begin = var_4117_begin_0, end = var_4117_end_0, end_mask = var_4117_end_mask_0, stride = var_4117_stride_0, x = key_131_cast_fp16)[name = string("op_4117_cast_fp16")];
            int32 var_4120_axis_0 = const()[name = string("op_4120_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_4120_cast_fp16 = stack(axis = var_4120_axis_0, values = (var_4112_cast_fp16, var_4117_cast_fp16))[name = string("op_4120_cast_fp16")];
            tensor<int32, [4]> shape_41 = const()[name = string("shape_41"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_4121_cast_fp16 = reshape(shape = shape_41, x = var_4120_cast_fp16)[name = string("op_4121_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_4122_cast_fp16 = mul(x = var_4121_cast_fp16, y = rope_sin_13_to_fp16)[name = string("op_4122_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> key_133_cast_fp16 = add(x = var_4106_cast_fp16, y = var_4122_cast_fp16)[name = string("key_133_cast_fp16")];
            tensor<int32, [4]> var_4129 = const()[name = string("op_4129"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_4135 = const()[name = string("op_4135"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_4141 = const()[name = string("op_4141"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 var_4144 = const()[name = string("op_4144"), val = int32(2)];
            bool key_137_interleave_0 = const()[name = string("key_137_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> key_135_cast_fp16 = transpose(perm = var_4135, x = key_133_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 12, 14, 64]> key_137_cast_fp16 = concat(axis = var_4144, interleave = key_137_interleave_0, values = (key_127_cast_fp16, key_135_cast_fp16))[name = string("key_137_cast_fp16")];
            int32 var_4147 = const()[name = string("op_4147"), val = int32(2)];
            bool value_109_interleave_0 = const()[name = string("value_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> value_107_cast_fp16 = transpose(perm = var_4141, x = value_105_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 12, 14, 64]> value_109_cast_fp16 = concat(axis = var_4147, interleave = value_109_interleave_0, values = (value_101_cast_fp16, value_107_cast_fp16))[name = string("value_109_cast_fp16")];
            bool var_4152_transpose_x_1 = const()[name = string("op_4152_transpose_x_1"), val = bool(false)];
            bool var_4152_transpose_y_1 = const()[name = string("op_4152_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 12, 1, 64]> query_111_cast_fp16 = transpose(perm = var_4129, x = query_109_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 12, 1, 14]> var_4152_cast_fp16 = matmul(transpose_x = var_4152_transpose_x_1, transpose_y = var_4152_transpose_y_1, x = query_111_cast_fp16, y = key_137_cast_fp16)[name = string("op_4152_cast_fp16")];
            fp16 var_4153_to_fp16 = const()[name = string("op_4153_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 1, 14]> scores_51_cast_fp16 = mul(x = var_4152_cast_fp16, y = var_4153_to_fp16)[name = string("scores_51_cast_fp16")];
            int32 var_4155 = const()[name = string("op_4155"), val = int32(-1)];
            tensor<fp16, [1, 12, 1, 14]> probabilities_75_cast_fp16 = softmax(axis = var_4155, x = scores_51_cast_fp16)[name = string("probabilities_75_cast_fp16")];
            bool attention_27_transpose_x_0 = const()[name = string("attention_27_transpose_x_0"), val = bool(false)];
            bool attention_27_transpose_y_0 = const()[name = string("attention_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> attention_27_cast_fp16 = matmul(transpose_x = attention_27_transpose_x_0, transpose_y = attention_27_transpose_y_0, x = probabilities_75_cast_fp16, y = value_109_cast_fp16)[name = string("attention_27_cast_fp16")];
            tensor<int32, [4]> var_4163 = const()[name = string("op_4163"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4168 = const()[name = string("op_4168"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 12, 64]> var_4164_cast_fp16 = transpose(perm = var_4163, x = attention_27_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 1, 768]> input_263_cast_fp16 = reshape(shape = var_4168, x = var_4164_cast_fp16)[name = string("input_263_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_65_cast_fp16 = linear(bias = block_attn_c_proj_bias_to_fp16, weight = block_attn_c_proj_weight_to_fp16, x = input_263_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_265_cast_fp16 = add(x = input_259_cast_fp16, y = linear_65_cast_fp16)[name = string("input_265_cast_fp16")];
            tensor<int32, [1]> input_267_axes_0 = const()[name = string("input_267_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_4176_to_fp16 = const()[name = string("op_4176_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_267_cast_fp16 = layer_norm(axes = input_267_axes_0, beta = block_ln_2_bias_to_fp16, epsilon = var_4176_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_265_cast_fp16)[name = string("input_267_cast_fp16")];
            tensor<fp16, [1, 1, 3072]> linear_66_cast_fp16 = linear(bias = block_mlp_fc_in_bias_to_fp16, weight = block_mlp_fc_in_weight_to_fp16, x = input_267_cast_fp16)[name = string("linear_66_cast_fp16")];
            string input_271_mode_0 = const()[name = string("input_271_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1, 3072]> input_271_cast_fp16 = gelu(mode = input_271_mode_0, x = linear_66_cast_fp16)[name = string("input_271_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_67_cast_fp16 = linear(bias = block_mlp_fc_out_bias_to_fp16, weight = block_mlp_fc_out_weight_to_fp16, x = input_271_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_275_cast_fp16 = add(x = input_265_cast_fp16, y = linear_67_cast_fp16)[name = string("input_275_cast_fp16")];
            tensor<int32, [1]> var_4215_axes_0 = const()[name = string("op_4215_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_4210_to_fp16 = const()[name = string("op_4210_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> var_4215_cast_fp16 = layer_norm(axes = var_4215_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_4210_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_275_cast_fp16)[name = string("op_4215_cast_fp16")];
            tensor<int32, [3]> var_4223_begin_0 = const()[name = string("op_4223_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_4223_end_0 = const()[name = string("op_4223_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_4223_end_mask_0 = const()[name = string("op_4223_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_4223_squeeze_mask_0 = const()[name = string("op_4223_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_4223_cast_fp16 = slice_by_index(begin = var_4223_begin_0, end = var_4223_end_0, end_mask = var_4223_end_mask_0, squeeze_mask = var_4223_squeeze_mask_0, x = var_4215_cast_fp16)[name = string("op_4223_cast_fp16")];
            tensor<fp16, [1024, 768]> audio_embeddings_12_weight_to_fp16 = const()[name = string("audio_embeddings_12_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33065344)))];
            tensor<fp16, [1, 1024]> linear_68_cast_fp16 = linear(bias = linear_8_bias_0_to_fp16, weight = audio_embeddings_12_weight_to_fp16, x = var_4223_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<int32, [3]> var_4238_begin_0 = const()[name = string("op_4238_begin_0"), val = tensor<int32, [3]>([0, 12, 0])];
            tensor<int32, [3]> var_4238_end_0 = const()[name = string("op_4238_end_0"), val = tensor<int32, [3]>([1, 13, 1024])];
            tensor<bool, [3]> var_4238_end_mask_0 = const()[name = string("op_4238_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_4238_squeeze_mask_0 = const()[name = string("op_4238_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<int32, [1, 1024]> var_4238 = slice_by_index(begin = var_4238_begin_0, end = var_4238_end_0, end_mask = var_4238_end_mask_0, squeeze_mask = var_4238_squeeze_mask_0, x = repetition_seen)[name = string("op_4238")];
            tensor<int32, [2]> random_u_25_begin_0 = const()[name = string("random_u_25_begin_0"), val = tensor<int32, [2]>([0, 12])];
            tensor<int32, [2]> random_u_25_end_0 = const()[name = string("random_u_25_end_0"), val = tensor<int32, [2]>([1, 13])];
            tensor<bool, [2]> random_u_25_end_mask_0 = const()[name = string("random_u_25_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<bool, [2]> random_u_25_squeeze_mask_0 = const()[name = string("random_u_25_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [1]> random_u_25_cast_fp16 = slice_by_index(begin = random_u_25_begin_0, end = random_u_25_end_0, end_mask = random_u_25_end_mask_0, squeeze_mask = random_u_25_squeeze_mask_0, x = audio_random_to_fp16)[name = string("random_u_25_cast_fp16")];
            fp16 var_4252_promoted_to_fp16 = const()[name = string("op_4252_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [1, 1024]> var_4253_cast_fp16 = less(x = linear_68_cast_fp16, y = var_4252_promoted_to_fp16)[name = string("op_4253_cast_fp16")];
            fp16 var_4254_to_fp16 = const()[name = string("op_4254_to_fp16"), val = fp16(0x1.334p+0)];
            tensor<fp16, [1, 1024]> var_4255_cast_fp16 = mul(x = linear_68_cast_fp16, y = var_4254_to_fp16)[name = string("op_4255_cast_fp16")];
            fp16 _inversed_4257_y_0_to_fp16 = const()[name = string("_inversed_4257_y_0_to_fp16"), val = fp16(0x1.aacp-1)];
            tensor<fp16, [1, 1024]> _inversed_4257_cast_fp16 = mul(x = linear_68_cast_fp16, y = _inversed_4257_y_0_to_fp16)[name = string("_inversed_4257_cast_fp16")];
            tensor<fp16, [1, 1024]> penalized_25_cast_fp16 = select(a = var_4255_cast_fp16, b = _inversed_4257_cast_fp16, cond = var_4253_cast_fp16)[name = string("penalized_25_cast_fp16")];
            string var_4263_dtype_0 = const()[name = string("op_4263_dtype_0"), val = string("bool")];
            tensor<bool, [1, 1024]> var_4263 = cast(dtype = var_4263_dtype_0, x = var_4238)[name = string("cast_298")];
            tensor<fp16, [1, 1024]> var_4264_cast_fp16 = select(a = penalized_25_cast_fp16, b = linear_68_cast_fp16, cond = var_4263)[name = string("op_4264_cast_fp16")];
            fp16 _inversed_scores_53_y_0_to_fp16 = const()[name = string("_inversed_scores_53_y_0_to_fp16"), val = fp16(0x1.4p+0)];
            tensor<fp16, [1, 1024]> _inversed_scores_53_cast_fp16 = mul(x = var_4264_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = string("_inversed_scores_53_cast_fp16")];
            int32 var_4267 = const()[name = string("op_4267"), val = int32(25)];
            int32 top_scores_49_axis_0 = const()[name = string("top_scores_49_axis_0"), val = int32(1)];
            bool top_scores_49_ascending_0 = const()[name = string("top_scores_49_ascending_0"), val = bool(false)];
            bool top_scores_49_sort_0 = const()[name = string("top_scores_49_sort_0"), val = bool(true)];
            bool top_scores_49_return_indices_0 = const()[name = string("top_scores_49_return_indices_0"), val = bool(true)];
            string top_scores_49_cast_fp16_cast_uint16_output_indices_dtype_0 = const()[name = string("top_scores_49_cast_fp16_cast_uint16_output_indices_dtype_0"), val = string("uint16")];
            tensor<fp16, [1, 25]> top_scores_49_cast_fp16_cast_uint16_0, tensor<uint16, [1, 25]> top_scores_49_cast_fp16_cast_uint16_1 = topk(ascending = top_scores_49_ascending_0, axis = top_scores_49_axis_0, k = var_4267, output_indices_dtype = top_scores_49_cast_fp16_cast_uint16_output_indices_dtype_0, return_indices = top_scores_49_return_indices_0, sort = top_scores_49_sort_0, x = _inversed_scores_53_cast_fp16)[name = string("top_scores_49_cast_fp16_cast_uint16")];
            int32 var_4273 = const()[name = string("op_4273"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_77_cast_fp16 = softmax(axis = var_4273, x = top_scores_49_cast_fp16_cast_uint16_0)[name = string("probabilities_77_cast_fp16")];
            int32 var_4276 = const()[name = string("op_4276"), val = int32(1)];
            bool cumulative_49_exclusive_0 = const()[name = string("cumulative_49_exclusive_0"), val = bool(false)];
            bool cumulative_49_reverse_0 = const()[name = string("cumulative_49_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_49_cast_fp16 = cumsum(axis = var_4276, exclusive = cumulative_49_exclusive_0, reverse = cumulative_49_reverse_0, x = probabilities_77_cast_fp16)[name = string("cumulative_49_cast_fp16")];
            tensor<fp16, [1, 25]> var_4280_cast_fp16 = sub(x = cumulative_49_cast_fp16, y = probabilities_77_cast_fp16)[name = string("op_4280_cast_fp16")];
            fp16 var_4281_to_fp16 = const()[name = string("op_4281_to_fp16"), val = fp16(0x1.e68p-1)];
            tensor<bool, [1, 25]> keep_25_cast_fp16 = less(x = var_4280_cast_fp16, y = var_4281_to_fp16)[name = string("keep_25_cast_fp16")];
            fp16 var_4289_value_0_to_fp16 = const()[name = string("op_4289_value_0_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 25]> var_4289_cast_fp16 = fill_like(ref_tensor = top_scores_49_cast_fp16_cast_uint16_0, value = var_4289_value_0_to_fp16)[name = string("op_4289_cast_fp16")];
            tensor<fp16, [1, 25]> top_scores_51_cast_fp16 = select(a = top_scores_49_cast_fp16_cast_uint16_0, b = var_4289_cast_fp16, cond = keep_25_cast_fp16)[name = string("top_scores_51_cast_fp16")];
            int32 var_4291 = const()[name = string("op_4291"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_79_cast_fp16 = softmax(axis = var_4291, x = top_scores_51_cast_fp16)[name = string("probabilities_79_cast_fp16")];
            int32 var_4294 = const()[name = string("op_4294"), val = int32(1)];
            bool cumulative_51_exclusive_0 = const()[name = string("cumulative_51_exclusive_0"), val = bool(false)];
            bool cumulative_51_reverse_0 = const()[name = string("cumulative_51_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_51_cast_fp16 = cumsum(axis = var_4294, exclusive = cumulative_51_exclusive_0, reverse = cumulative_51_reverse_0, x = probabilities_79_cast_fp16)[name = string("cumulative_51_cast_fp16")];
            tensor<int32, [2]> var_4299 = const()[name = string("op_4299"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_4300_cast_fp16 = reshape(shape = var_4299, x = random_u_25_cast_fp16)[name = string("op_4300_cast_fp16")];
            tensor<bool, [1, 25]> var_4301_cast_fp16 = less(x = cumulative_51_cast_fp16, y = var_4300_cast_fp16)[name = string("op_4301_cast_fp16")];
            string var_4306_dtype_0 = const()[name = string("op_4306_dtype_0"), val = string("int32")];
            tensor<int32, [1]> selected_49_axes_0 = const()[name = string("selected_49_axes_0"), val = tensor<int32, [1]>([1])];
            bool selected_49_keep_dims_0 = const()[name = string("selected_49_keep_dims_0"), val = bool(false)];
            tensor<int32, [1, 25]> var_4306 = cast(dtype = var_4306_dtype_0, x = var_4301_cast_fp16)[name = string("cast_297")];
            tensor<int32, [1]> selected_49 = reduce_sum(axes = selected_49_axes_0, keep_dims = selected_49_keep_dims_0, x = var_4306)[name = string("selected_49")];
            string selected_49_promoted_to_fp16_dtype_0 = const()[name = string("selected_49_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_68_to_fp16 = const()[name = string("const_68_to_fp16"), val = fp16(-inf)];
            fp16 var_4313_promoted_to_fp16 = const()[name = string("op_4313_promoted_to_fp16"), val = fp16(0x1.8p+4)];
            tensor<fp16, [1]> selected_49_to_fp16 = cast(dtype = selected_49_promoted_to_fp16_dtype_0, x = selected_49)[name = string("cast_296")];
            tensor<fp16, [1]> clip_12_cast_fp16 = clip(alpha = const_68_to_fp16, beta = var_4313_promoted_to_fp16, x = selected_49_to_fp16)[name = string("clip_12_cast_fp16")];
            string selected_51_dtype_0 = const()[name = string("selected_51_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_4322 = const()[name = string("op_4322"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> clip_12_cast_fp16_to_int32 = cast(dtype = selected_51_dtype_0, x = clip_12_cast_fp16)[name = string("cast_295")];
            tensor<int32, [1, 1]> var_4323 = reshape(shape = var_4322, x = clip_12_cast_fp16_to_int32)[name = string("op_4323")];
            int32 var_4324 = const()[name = string("op_4324"), val = int32(1)];
            bool var_4326_validate_indices_0 = const()[name = string("op_4326_validate_indices_0"), val = bool(false)];
            string var_4323_to_uint16_dtype_0 = const()[name = string("op_4323_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 1]> var_4323_to_uint16 = cast(dtype = var_4323_to_uint16_dtype_0, x = var_4323)[name = string("cast_294")];
            tensor<uint16, [1, 1]> var_4326_cast_uint16 = gather_along_axis(axis = var_4324, indices = var_4323_to_uint16, validate_indices = var_4326_validate_indices_0, x = top_scores_49_cast_fp16_cast_uint16_1)[name = string("op_4326_cast_uint16")];
            string var_4326_cast_uint16_to_int32_dtype_0 = const()[name = string("op_4326_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [1]> var_4328 = const()[name = string("op_4328"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_4326_cast_uint16_to_int32 = cast(dtype = var_4326_cast_uint16_to_int32_dtype_0, x = var_4326_cast_uint16)[name = string("cast_293")];
            tensor<int32, [1]> token_25 = reshape(shape = var_4328, x = var_4326_cast_uint16_to_int32)[name = string("token_25")];
            int32 var_4338_axis_0 = const()[name = string("op_4338_axis_0"), val = int32(0)];
            int32 var_4338_batch_dims_0 = const()[name = string("op_4338_batch_dims_0"), val = int32(0)];
            bool var_4338_validate_indices_0 = const()[name = string("op_4338_validate_indices_0"), val = bool(false)];
            string token_25_to_uint16_dtype_0 = const()[name = string("token_25_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> token_25_to_uint16 = cast(dtype = token_25_to_uint16_dtype_0, x = token_25)[name = string("cast_292")];
            tensor<fp16, [1, 768]> var_4338_cast_fp16_cast_uint16 = gather(axis = var_4338_axis_0, batch_dims = var_4338_batch_dims_0, indices = token_25_to_uint16, validate_indices = var_4338_validate_indices_0, x = audio_embeddings_12_weight_to_fp16)[name = string("op_4338_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 768]> next_global_embed_27_cast_fp16 = add(x = next_global_embed_25_cast_fp16, y = var_4338_cast_fp16_cast_uint16)[name = string("next_global_embed_27_cast_fp16")];
            tensor<int32, [3]> var_4348 = const()[name = string("op_4348"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 768]> input_279_cast_fp16 = reshape(shape = var_4348, x = var_4338_cast_fp16_cast_uint16)[name = string("input_279_cast_fp16")];
            tensor<int32, [1]> input_281_axes_0 = const()[name = string("input_281_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_4351_to_fp16 = const()[name = string("op_4351_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_281_cast_fp16 = layer_norm(axes = input_281_axes_0, beta = block_ln_1_bias_to_fp16, epsilon = var_4351_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_279_cast_fp16)[name = string("input_281_cast_fp16")];
            tensor<fp16, [1, 1, 2304]> linear_69_cast_fp16 = linear(bias = block_attn_c_attn_bias_to_fp16, weight = block_attn_c_attn_weight_to_fp16, x = input_281_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<int32, [3]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [3]>([768, 768, 768])];
            int32 var_4362_axis_0 = const()[name = string("op_4362_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 768]> var_4362_cast_fp16_0, tensor<fp16, [1, 1, 768]> var_4362_cast_fp16_1, tensor<fp16, [1, 1, 768]> var_4362_cast_fp16_2 = split(axis = var_4362_axis_0, split_sizes = tile_14, x = linear_69_cast_fp16)[name = string("op_4362_cast_fp16")];
            tensor<int32, [4]> var_4370 = const()[name = string("op_4370"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> query_115_cast_fp16 = reshape(shape = var_4370, x = var_4362_cast_fp16_0)[name = string("query_115_cast_fp16")];
            tensor<int32, [4]> var_4376 = const()[name = string("op_4376"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> key_141_cast_fp16 = reshape(shape = var_4376, x = var_4362_cast_fp16_1)[name = string("key_141_cast_fp16")];
            tensor<int32, [4]> var_4382 = const()[name = string("op_4382"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> value_113_cast_fp16 = reshape(shape = var_4382, x = var_4362_cast_fp16_2)[name = string("value_113_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_cos_14_to_fp16 = const()[name = string("rope_cos_14_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34638272)))];
            tensor<fp16, [1, 1, 12, 64]> var_4384_cast_fp16 = mul(x = query_115_cast_fp16, y = rope_cos_14_to_fp16)[name = string("op_4384_cast_fp16")];
            tensor<int32, [4]> var_4389_begin_0 = const()[name = string("op_4389_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_4389_end_0 = const()[name = string("op_4389_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_4389_end_mask_0 = const()[name = string("op_4389_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4389_stride_0 = const()[name = string("op_4389_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_4389_cast_fp16 = slice_by_index(begin = var_4389_begin_0, end = var_4389_end_0, end_mask = var_4389_end_mask_0, stride = var_4389_stride_0, x = query_115_cast_fp16)[name = string("op_4389_cast_fp16")];
            fp16 const_71_promoted_to_fp16 = const()[name = string("const_71_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_4390_cast_fp16 = mul(x = var_4389_cast_fp16, y = const_71_promoted_to_fp16)[name = string("op_4390_cast_fp16")];
            tensor<int32, [4]> var_4395_begin_0 = const()[name = string("op_4395_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4395_end_0 = const()[name = string("op_4395_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_4395_end_mask_0 = const()[name = string("op_4395_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4395_stride_0 = const()[name = string("op_4395_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_4395_cast_fp16 = slice_by_index(begin = var_4395_begin_0, end = var_4395_end_0, end_mask = var_4395_end_mask_0, stride = var_4395_stride_0, x = query_115_cast_fp16)[name = string("op_4395_cast_fp16")];
            int32 var_4398_axis_0 = const()[name = string("op_4398_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_4398_cast_fp16 = stack(axis = var_4398_axis_0, values = (var_4390_cast_fp16, var_4395_cast_fp16))[name = string("op_4398_cast_fp16")];
            tensor<int32, [4]> shape_43 = const()[name = string("shape_43"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_4399_cast_fp16 = reshape(shape = shape_43, x = var_4398_cast_fp16)[name = string("op_4399_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_sin_14_to_fp16 = const()[name = string("rope_sin_14_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34638464)))];
            tensor<fp16, [1, 1, 12, 64]> var_4400_cast_fp16 = mul(x = var_4399_cast_fp16, y = rope_sin_14_to_fp16)[name = string("op_4400_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> query_117_cast_fp16 = add(x = var_4384_cast_fp16, y = var_4400_cast_fp16)[name = string("query_117_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_4403_cast_fp16 = mul(x = key_141_cast_fp16, y = rope_cos_14_to_fp16)[name = string("op_4403_cast_fp16")];
            tensor<int32, [4]> var_4408_begin_0 = const()[name = string("op_4408_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_4408_end_0 = const()[name = string("op_4408_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_4408_end_mask_0 = const()[name = string("op_4408_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4408_stride_0 = const()[name = string("op_4408_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_4408_cast_fp16 = slice_by_index(begin = var_4408_begin_0, end = var_4408_end_0, end_mask = var_4408_end_mask_0, stride = var_4408_stride_0, x = key_141_cast_fp16)[name = string("op_4408_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_4409_cast_fp16 = mul(x = var_4408_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_4409_cast_fp16")];
            tensor<int32, [4]> var_4414_begin_0 = const()[name = string("op_4414_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4414_end_0 = const()[name = string("op_4414_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_4414_end_mask_0 = const()[name = string("op_4414_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4414_stride_0 = const()[name = string("op_4414_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_4414_cast_fp16 = slice_by_index(begin = var_4414_begin_0, end = var_4414_end_0, end_mask = var_4414_end_mask_0, stride = var_4414_stride_0, x = key_141_cast_fp16)[name = string("op_4414_cast_fp16")];
            int32 var_4417_axis_0 = const()[name = string("op_4417_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_4417_cast_fp16 = stack(axis = var_4417_axis_0, values = (var_4409_cast_fp16, var_4414_cast_fp16))[name = string("op_4417_cast_fp16")];
            tensor<int32, [4]> shape_44 = const()[name = string("shape_44"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_4418_cast_fp16 = reshape(shape = shape_44, x = var_4417_cast_fp16)[name = string("op_4418_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_4419_cast_fp16 = mul(x = var_4418_cast_fp16, y = rope_sin_14_to_fp16)[name = string("op_4419_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> key_143_cast_fp16 = add(x = var_4403_cast_fp16, y = var_4419_cast_fp16)[name = string("key_143_cast_fp16")];
            tensor<int32, [4]> var_4426 = const()[name = string("op_4426"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_4432 = const()[name = string("op_4432"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_4438 = const()[name = string("op_4438"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 var_4441 = const()[name = string("op_4441"), val = int32(2)];
            bool key_147_interleave_0 = const()[name = string("key_147_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> key_145_cast_fp16 = transpose(perm = var_4432, x = key_143_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 12, 15, 64]> key_147_cast_fp16 = concat(axis = var_4441, interleave = key_147_interleave_0, values = (key_137_cast_fp16, key_145_cast_fp16))[name = string("key_147_cast_fp16")];
            int32 var_4444 = const()[name = string("op_4444"), val = int32(2)];
            bool value_117_interleave_0 = const()[name = string("value_117_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> value_115_cast_fp16 = transpose(perm = var_4438, x = value_113_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 12, 15, 64]> value_117_cast_fp16 = concat(axis = var_4444, interleave = value_117_interleave_0, values = (value_109_cast_fp16, value_115_cast_fp16))[name = string("value_117_cast_fp16")];
            bool var_4449_transpose_x_1 = const()[name = string("op_4449_transpose_x_1"), val = bool(false)];
            bool var_4449_transpose_y_1 = const()[name = string("op_4449_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 12, 1, 64]> query_119_cast_fp16 = transpose(perm = var_4426, x = query_117_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 12, 1, 15]> var_4449_cast_fp16 = matmul(transpose_x = var_4449_transpose_x_1, transpose_y = var_4449_transpose_y_1, x = query_119_cast_fp16, y = key_147_cast_fp16)[name = string("op_4449_cast_fp16")];
            fp16 var_4450_to_fp16 = const()[name = string("op_4450_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 1, 15]> scores_55_cast_fp16 = mul(x = var_4449_cast_fp16, y = var_4450_to_fp16)[name = string("scores_55_cast_fp16")];
            int32 var_4452 = const()[name = string("op_4452"), val = int32(-1)];
            tensor<fp16, [1, 12, 1, 15]> probabilities_81_cast_fp16 = softmax(axis = var_4452, x = scores_55_cast_fp16)[name = string("probabilities_81_cast_fp16")];
            bool attention_29_transpose_x_0 = const()[name = string("attention_29_transpose_x_0"), val = bool(false)];
            bool attention_29_transpose_y_0 = const()[name = string("attention_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> attention_29_cast_fp16 = matmul(transpose_x = attention_29_transpose_x_0, transpose_y = attention_29_transpose_y_0, x = probabilities_81_cast_fp16, y = value_117_cast_fp16)[name = string("attention_29_cast_fp16")];
            tensor<int32, [4]> var_4460 = const()[name = string("op_4460"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4465 = const()[name = string("op_4465"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 12, 64]> var_4461_cast_fp16 = transpose(perm = var_4460, x = attention_29_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 1, 768]> input_283_cast_fp16 = reshape(shape = var_4465, x = var_4461_cast_fp16)[name = string("input_283_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_70_cast_fp16 = linear(bias = block_attn_c_proj_bias_to_fp16, weight = block_attn_c_proj_weight_to_fp16, x = input_283_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_285_cast_fp16 = add(x = input_279_cast_fp16, y = linear_70_cast_fp16)[name = string("input_285_cast_fp16")];
            tensor<int32, [1]> input_287_axes_0 = const()[name = string("input_287_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_4473_to_fp16 = const()[name = string("op_4473_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = block_ln_2_bias_to_fp16, epsilon = var_4473_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_285_cast_fp16)[name = string("input_287_cast_fp16")];
            tensor<fp16, [1, 1, 3072]> linear_71_cast_fp16 = linear(bias = block_mlp_fc_in_bias_to_fp16, weight = block_mlp_fc_in_weight_to_fp16, x = input_287_cast_fp16)[name = string("linear_71_cast_fp16")];
            string input_291_mode_0 = const()[name = string("input_291_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1, 3072]> input_291_cast_fp16 = gelu(mode = input_291_mode_0, x = linear_71_cast_fp16)[name = string("input_291_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_72_cast_fp16 = linear(bias = block_mlp_fc_out_bias_to_fp16, weight = block_mlp_fc_out_weight_to_fp16, x = input_291_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_295_cast_fp16 = add(x = input_285_cast_fp16, y = linear_72_cast_fp16)[name = string("input_295_cast_fp16")];
            tensor<int32, [1]> var_4512_axes_0 = const()[name = string("op_4512_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_4507_to_fp16 = const()[name = string("op_4507_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> var_4512_cast_fp16 = layer_norm(axes = var_4512_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_4507_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_295_cast_fp16)[name = string("op_4512_cast_fp16")];
            tensor<int32, [3]> var_4520_begin_0 = const()[name = string("op_4520_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_4520_end_0 = const()[name = string("op_4520_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_4520_end_mask_0 = const()[name = string("op_4520_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_4520_squeeze_mask_0 = const()[name = string("op_4520_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_4520_cast_fp16 = slice_by_index(begin = var_4520_begin_0, end = var_4520_end_0, end_mask = var_4520_end_mask_0, squeeze_mask = var_4520_squeeze_mask_0, x = var_4512_cast_fp16)[name = string("op_4520_cast_fp16")];
            tensor<fp16, [1024, 768]> audio_embeddings_13_weight_to_fp16 = const()[name = string("audio_embeddings_13_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34638656)))];
            tensor<fp16, [1, 1024]> linear_73_cast_fp16 = linear(bias = linear_8_bias_0_to_fp16, weight = audio_embeddings_13_weight_to_fp16, x = var_4520_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<int32, [3]> var_4535_begin_0 = const()[name = string("op_4535_begin_0"), val = tensor<int32, [3]>([0, 13, 0])];
            tensor<int32, [3]> var_4535_end_0 = const()[name = string("op_4535_end_0"), val = tensor<int32, [3]>([1, 14, 1024])];
            tensor<bool, [3]> var_4535_end_mask_0 = const()[name = string("op_4535_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_4535_squeeze_mask_0 = const()[name = string("op_4535_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<int32, [1, 1024]> var_4535 = slice_by_index(begin = var_4535_begin_0, end = var_4535_end_0, end_mask = var_4535_end_mask_0, squeeze_mask = var_4535_squeeze_mask_0, x = repetition_seen)[name = string("op_4535")];
            tensor<int32, [2]> random_u_27_begin_0 = const()[name = string("random_u_27_begin_0"), val = tensor<int32, [2]>([0, 13])];
            tensor<int32, [2]> random_u_27_end_0 = const()[name = string("random_u_27_end_0"), val = tensor<int32, [2]>([1, 14])];
            tensor<bool, [2]> random_u_27_end_mask_0 = const()[name = string("random_u_27_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<bool, [2]> random_u_27_squeeze_mask_0 = const()[name = string("random_u_27_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [1]> random_u_27_cast_fp16 = slice_by_index(begin = random_u_27_begin_0, end = random_u_27_end_0, end_mask = random_u_27_end_mask_0, squeeze_mask = random_u_27_squeeze_mask_0, x = audio_random_to_fp16)[name = string("random_u_27_cast_fp16")];
            fp16 var_4549_promoted_to_fp16 = const()[name = string("op_4549_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [1, 1024]> var_4550_cast_fp16 = less(x = linear_73_cast_fp16, y = var_4549_promoted_to_fp16)[name = string("op_4550_cast_fp16")];
            fp16 var_4551_to_fp16 = const()[name = string("op_4551_to_fp16"), val = fp16(0x1.334p+0)];
            tensor<fp16, [1, 1024]> var_4552_cast_fp16 = mul(x = linear_73_cast_fp16, y = var_4551_to_fp16)[name = string("op_4552_cast_fp16")];
            fp16 _inversed_4554_y_0_to_fp16 = const()[name = string("_inversed_4554_y_0_to_fp16"), val = fp16(0x1.aacp-1)];
            tensor<fp16, [1, 1024]> _inversed_4554_cast_fp16 = mul(x = linear_73_cast_fp16, y = _inversed_4554_y_0_to_fp16)[name = string("_inversed_4554_cast_fp16")];
            tensor<fp16, [1, 1024]> penalized_27_cast_fp16 = select(a = var_4552_cast_fp16, b = _inversed_4554_cast_fp16, cond = var_4550_cast_fp16)[name = string("penalized_27_cast_fp16")];
            string var_4560_dtype_0 = const()[name = string("op_4560_dtype_0"), val = string("bool")];
            tensor<bool, [1, 1024]> var_4560 = cast(dtype = var_4560_dtype_0, x = var_4535)[name = string("cast_291")];
            tensor<fp16, [1, 1024]> var_4561_cast_fp16 = select(a = penalized_27_cast_fp16, b = linear_73_cast_fp16, cond = var_4560)[name = string("op_4561_cast_fp16")];
            fp16 _inversed_scores_57_y_0_to_fp16 = const()[name = string("_inversed_scores_57_y_0_to_fp16"), val = fp16(0x1.4p+0)];
            tensor<fp16, [1, 1024]> _inversed_scores_57_cast_fp16 = mul(x = var_4561_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = string("_inversed_scores_57_cast_fp16")];
            int32 var_4564 = const()[name = string("op_4564"), val = int32(25)];
            int32 top_scores_53_axis_0 = const()[name = string("top_scores_53_axis_0"), val = int32(1)];
            bool top_scores_53_ascending_0 = const()[name = string("top_scores_53_ascending_0"), val = bool(false)];
            bool top_scores_53_sort_0 = const()[name = string("top_scores_53_sort_0"), val = bool(true)];
            bool top_scores_53_return_indices_0 = const()[name = string("top_scores_53_return_indices_0"), val = bool(true)];
            string top_scores_53_cast_fp16_cast_uint16_output_indices_dtype_0 = const()[name = string("top_scores_53_cast_fp16_cast_uint16_output_indices_dtype_0"), val = string("uint16")];
            tensor<fp16, [1, 25]> top_scores_53_cast_fp16_cast_uint16_0, tensor<uint16, [1, 25]> top_scores_53_cast_fp16_cast_uint16_1 = topk(ascending = top_scores_53_ascending_0, axis = top_scores_53_axis_0, k = var_4564, output_indices_dtype = top_scores_53_cast_fp16_cast_uint16_output_indices_dtype_0, return_indices = top_scores_53_return_indices_0, sort = top_scores_53_sort_0, x = _inversed_scores_57_cast_fp16)[name = string("top_scores_53_cast_fp16_cast_uint16")];
            int32 var_4570 = const()[name = string("op_4570"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_83_cast_fp16 = softmax(axis = var_4570, x = top_scores_53_cast_fp16_cast_uint16_0)[name = string("probabilities_83_cast_fp16")];
            int32 var_4573 = const()[name = string("op_4573"), val = int32(1)];
            bool cumulative_53_exclusive_0 = const()[name = string("cumulative_53_exclusive_0"), val = bool(false)];
            bool cumulative_53_reverse_0 = const()[name = string("cumulative_53_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_53_cast_fp16 = cumsum(axis = var_4573, exclusive = cumulative_53_exclusive_0, reverse = cumulative_53_reverse_0, x = probabilities_83_cast_fp16)[name = string("cumulative_53_cast_fp16")];
            tensor<fp16, [1, 25]> var_4577_cast_fp16 = sub(x = cumulative_53_cast_fp16, y = probabilities_83_cast_fp16)[name = string("op_4577_cast_fp16")];
            fp16 var_4578_to_fp16 = const()[name = string("op_4578_to_fp16"), val = fp16(0x1.e68p-1)];
            tensor<bool, [1, 25]> keep_27_cast_fp16 = less(x = var_4577_cast_fp16, y = var_4578_to_fp16)[name = string("keep_27_cast_fp16")];
            fp16 var_4586_value_0_to_fp16 = const()[name = string("op_4586_value_0_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 25]> var_4586_cast_fp16 = fill_like(ref_tensor = top_scores_53_cast_fp16_cast_uint16_0, value = var_4586_value_0_to_fp16)[name = string("op_4586_cast_fp16")];
            tensor<fp16, [1, 25]> top_scores_55_cast_fp16 = select(a = top_scores_53_cast_fp16_cast_uint16_0, b = var_4586_cast_fp16, cond = keep_27_cast_fp16)[name = string("top_scores_55_cast_fp16")];
            int32 var_4588 = const()[name = string("op_4588"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_85_cast_fp16 = softmax(axis = var_4588, x = top_scores_55_cast_fp16)[name = string("probabilities_85_cast_fp16")];
            int32 var_4591 = const()[name = string("op_4591"), val = int32(1)];
            bool cumulative_55_exclusive_0 = const()[name = string("cumulative_55_exclusive_0"), val = bool(false)];
            bool cumulative_55_reverse_0 = const()[name = string("cumulative_55_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_55_cast_fp16 = cumsum(axis = var_4591, exclusive = cumulative_55_exclusive_0, reverse = cumulative_55_reverse_0, x = probabilities_85_cast_fp16)[name = string("cumulative_55_cast_fp16")];
            tensor<int32, [2]> var_4596 = const()[name = string("op_4596"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_4597_cast_fp16 = reshape(shape = var_4596, x = random_u_27_cast_fp16)[name = string("op_4597_cast_fp16")];
            tensor<bool, [1, 25]> var_4598_cast_fp16 = less(x = cumulative_55_cast_fp16, y = var_4597_cast_fp16)[name = string("op_4598_cast_fp16")];
            string var_4603_dtype_0 = const()[name = string("op_4603_dtype_0"), val = string("int32")];
            tensor<int32, [1]> selected_53_axes_0 = const()[name = string("selected_53_axes_0"), val = tensor<int32, [1]>([1])];
            bool selected_53_keep_dims_0 = const()[name = string("selected_53_keep_dims_0"), val = bool(false)];
            tensor<int32, [1, 25]> var_4603 = cast(dtype = var_4603_dtype_0, x = var_4598_cast_fp16)[name = string("cast_290")];
            tensor<int32, [1]> selected_53 = reduce_sum(axes = selected_53_axes_0, keep_dims = selected_53_keep_dims_0, x = var_4603)[name = string("selected_53")];
            string selected_53_promoted_to_fp16_dtype_0 = const()[name = string("selected_53_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_73_to_fp16 = const()[name = string("const_73_to_fp16"), val = fp16(-inf)];
            fp16 var_4610_promoted_to_fp16 = const()[name = string("op_4610_promoted_to_fp16"), val = fp16(0x1.8p+4)];
            tensor<fp16, [1]> selected_53_to_fp16 = cast(dtype = selected_53_promoted_to_fp16_dtype_0, x = selected_53)[name = string("cast_289")];
            tensor<fp16, [1]> clip_13_cast_fp16 = clip(alpha = const_73_to_fp16, beta = var_4610_promoted_to_fp16, x = selected_53_to_fp16)[name = string("clip_13_cast_fp16")];
            string selected_55_dtype_0 = const()[name = string("selected_55_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_4619 = const()[name = string("op_4619"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> clip_13_cast_fp16_to_int32 = cast(dtype = selected_55_dtype_0, x = clip_13_cast_fp16)[name = string("cast_288")];
            tensor<int32, [1, 1]> var_4620 = reshape(shape = var_4619, x = clip_13_cast_fp16_to_int32)[name = string("op_4620")];
            int32 var_4621 = const()[name = string("op_4621"), val = int32(1)];
            bool var_4623_validate_indices_0 = const()[name = string("op_4623_validate_indices_0"), val = bool(false)];
            string var_4620_to_uint16_dtype_0 = const()[name = string("op_4620_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 1]> var_4620_to_uint16 = cast(dtype = var_4620_to_uint16_dtype_0, x = var_4620)[name = string("cast_287")];
            tensor<uint16, [1, 1]> var_4623_cast_uint16 = gather_along_axis(axis = var_4621, indices = var_4620_to_uint16, validate_indices = var_4623_validate_indices_0, x = top_scores_53_cast_fp16_cast_uint16_1)[name = string("op_4623_cast_uint16")];
            string var_4623_cast_uint16_to_int32_dtype_0 = const()[name = string("op_4623_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [1]> var_4625 = const()[name = string("op_4625"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_4623_cast_uint16_to_int32 = cast(dtype = var_4623_cast_uint16_to_int32_dtype_0, x = var_4623_cast_uint16)[name = string("cast_286")];
            tensor<int32, [1]> token_27 = reshape(shape = var_4625, x = var_4623_cast_uint16_to_int32)[name = string("token_27")];
            int32 var_4635_axis_0 = const()[name = string("op_4635_axis_0"), val = int32(0)];
            int32 var_4635_batch_dims_0 = const()[name = string("op_4635_batch_dims_0"), val = int32(0)];
            bool var_4635_validate_indices_0 = const()[name = string("op_4635_validate_indices_0"), val = bool(false)];
            string token_27_to_uint16_dtype_0 = const()[name = string("token_27_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> token_27_to_uint16 = cast(dtype = token_27_to_uint16_dtype_0, x = token_27)[name = string("cast_285")];
            tensor<fp16, [1, 768]> var_4635_cast_fp16_cast_uint16 = gather(axis = var_4635_axis_0, batch_dims = var_4635_batch_dims_0, indices = token_27_to_uint16, validate_indices = var_4635_validate_indices_0, x = audio_embeddings_13_weight_to_fp16)[name = string("op_4635_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 768]> next_global_embed_29_cast_fp16 = add(x = next_global_embed_27_cast_fp16, y = var_4635_cast_fp16_cast_uint16)[name = string("next_global_embed_29_cast_fp16")];
            tensor<int32, [3]> var_4645 = const()[name = string("op_4645"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 768]> input_299_cast_fp16 = reshape(shape = var_4645, x = var_4635_cast_fp16_cast_uint16)[name = string("input_299_cast_fp16")];
            tensor<int32, [1]> input_301_axes_0 = const()[name = string("input_301_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_4648_to_fp16 = const()[name = string("op_4648_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_301_cast_fp16 = layer_norm(axes = input_301_axes_0, beta = block_ln_1_bias_to_fp16, epsilon = var_4648_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_299_cast_fp16)[name = string("input_301_cast_fp16")];
            tensor<fp16, [1, 1, 2304]> linear_74_cast_fp16 = linear(bias = block_attn_c_attn_bias_to_fp16, weight = block_attn_c_attn_weight_to_fp16, x = input_301_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<int32, [3]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [3]>([768, 768, 768])];
            int32 var_4659_axis_0 = const()[name = string("op_4659_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 768]> var_4659_cast_fp16_0, tensor<fp16, [1, 1, 768]> var_4659_cast_fp16_1, tensor<fp16, [1, 1, 768]> var_4659_cast_fp16_2 = split(axis = var_4659_axis_0, split_sizes = tile_15, x = linear_74_cast_fp16)[name = string("op_4659_cast_fp16")];
            tensor<int32, [4]> var_4667 = const()[name = string("op_4667"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> query_123_cast_fp16 = reshape(shape = var_4667, x = var_4659_cast_fp16_0)[name = string("query_123_cast_fp16")];
            tensor<int32, [4]> var_4673 = const()[name = string("op_4673"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> key_151_cast_fp16 = reshape(shape = var_4673, x = var_4659_cast_fp16_1)[name = string("key_151_cast_fp16")];
            tensor<int32, [4]> var_4679 = const()[name = string("op_4679"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> value_121_cast_fp16 = reshape(shape = var_4679, x = var_4659_cast_fp16_2)[name = string("value_121_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_cos_15_to_fp16 = const()[name = string("rope_cos_15_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36211584)))];
            tensor<fp16, [1, 1, 12, 64]> var_4681_cast_fp16 = mul(x = query_123_cast_fp16, y = rope_cos_15_to_fp16)[name = string("op_4681_cast_fp16")];
            tensor<int32, [4]> var_4686_begin_0 = const()[name = string("op_4686_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_4686_end_0 = const()[name = string("op_4686_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_4686_end_mask_0 = const()[name = string("op_4686_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4686_stride_0 = const()[name = string("op_4686_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_4686_cast_fp16 = slice_by_index(begin = var_4686_begin_0, end = var_4686_end_0, end_mask = var_4686_end_mask_0, stride = var_4686_stride_0, x = query_123_cast_fp16)[name = string("op_4686_cast_fp16")];
            fp16 const_76_promoted_to_fp16 = const()[name = string("const_76_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_4687_cast_fp16 = mul(x = var_4686_cast_fp16, y = const_76_promoted_to_fp16)[name = string("op_4687_cast_fp16")];
            tensor<int32, [4]> var_4692_begin_0 = const()[name = string("op_4692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4692_end_0 = const()[name = string("op_4692_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_4692_end_mask_0 = const()[name = string("op_4692_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4692_stride_0 = const()[name = string("op_4692_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_4692_cast_fp16 = slice_by_index(begin = var_4692_begin_0, end = var_4692_end_0, end_mask = var_4692_end_mask_0, stride = var_4692_stride_0, x = query_123_cast_fp16)[name = string("op_4692_cast_fp16")];
            int32 var_4695_axis_0 = const()[name = string("op_4695_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_4695_cast_fp16 = stack(axis = var_4695_axis_0, values = (var_4687_cast_fp16, var_4692_cast_fp16))[name = string("op_4695_cast_fp16")];
            tensor<int32, [4]> shape_46 = const()[name = string("shape_46"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_4696_cast_fp16 = reshape(shape = shape_46, x = var_4695_cast_fp16)[name = string("op_4696_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_sin_15_to_fp16 = const()[name = string("rope_sin_15_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36211776)))];
            tensor<fp16, [1, 1, 12, 64]> var_4697_cast_fp16 = mul(x = var_4696_cast_fp16, y = rope_sin_15_to_fp16)[name = string("op_4697_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> query_125_cast_fp16 = add(x = var_4681_cast_fp16, y = var_4697_cast_fp16)[name = string("query_125_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_4700_cast_fp16 = mul(x = key_151_cast_fp16, y = rope_cos_15_to_fp16)[name = string("op_4700_cast_fp16")];
            tensor<int32, [4]> var_4705_begin_0 = const()[name = string("op_4705_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_4705_end_0 = const()[name = string("op_4705_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_4705_end_mask_0 = const()[name = string("op_4705_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4705_stride_0 = const()[name = string("op_4705_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_4705_cast_fp16 = slice_by_index(begin = var_4705_begin_0, end = var_4705_end_0, end_mask = var_4705_end_mask_0, stride = var_4705_stride_0, x = key_151_cast_fp16)[name = string("op_4705_cast_fp16")];
            fp16 const_77_promoted_to_fp16 = const()[name = string("const_77_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_4706_cast_fp16 = mul(x = var_4705_cast_fp16, y = const_77_promoted_to_fp16)[name = string("op_4706_cast_fp16")];
            tensor<int32, [4]> var_4711_begin_0 = const()[name = string("op_4711_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4711_end_0 = const()[name = string("op_4711_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_4711_end_mask_0 = const()[name = string("op_4711_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4711_stride_0 = const()[name = string("op_4711_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_4711_cast_fp16 = slice_by_index(begin = var_4711_begin_0, end = var_4711_end_0, end_mask = var_4711_end_mask_0, stride = var_4711_stride_0, x = key_151_cast_fp16)[name = string("op_4711_cast_fp16")];
            int32 var_4714_axis_0 = const()[name = string("op_4714_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_4714_cast_fp16 = stack(axis = var_4714_axis_0, values = (var_4706_cast_fp16, var_4711_cast_fp16))[name = string("op_4714_cast_fp16")];
            tensor<int32, [4]> shape_47 = const()[name = string("shape_47"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_4715_cast_fp16 = reshape(shape = shape_47, x = var_4714_cast_fp16)[name = string("op_4715_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_4716_cast_fp16 = mul(x = var_4715_cast_fp16, y = rope_sin_15_to_fp16)[name = string("op_4716_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> key_153_cast_fp16 = add(x = var_4700_cast_fp16, y = var_4716_cast_fp16)[name = string("key_153_cast_fp16")];
            tensor<int32, [4]> var_4723 = const()[name = string("op_4723"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_4729 = const()[name = string("op_4729"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_4735 = const()[name = string("op_4735"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 var_4738 = const()[name = string("op_4738"), val = int32(2)];
            bool key_157_interleave_0 = const()[name = string("key_157_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> key_155_cast_fp16 = transpose(perm = var_4729, x = key_153_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 12, 16, 64]> key_157_cast_fp16 = concat(axis = var_4738, interleave = key_157_interleave_0, values = (key_147_cast_fp16, key_155_cast_fp16))[name = string("key_157_cast_fp16")];
            int32 var_4741 = const()[name = string("op_4741"), val = int32(2)];
            bool value_125_interleave_0 = const()[name = string("value_125_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> value_123_cast_fp16 = transpose(perm = var_4735, x = value_121_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 12, 16, 64]> value_125_cast_fp16 = concat(axis = var_4741, interleave = value_125_interleave_0, values = (value_117_cast_fp16, value_123_cast_fp16))[name = string("value_125_cast_fp16")];
            bool var_4746_transpose_x_1 = const()[name = string("op_4746_transpose_x_1"), val = bool(false)];
            bool var_4746_transpose_y_1 = const()[name = string("op_4746_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 12, 1, 64]> query_127_cast_fp16 = transpose(perm = var_4723, x = query_125_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 12, 1, 16]> var_4746_cast_fp16 = matmul(transpose_x = var_4746_transpose_x_1, transpose_y = var_4746_transpose_y_1, x = query_127_cast_fp16, y = key_157_cast_fp16)[name = string("op_4746_cast_fp16")];
            fp16 var_4747_to_fp16 = const()[name = string("op_4747_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 1, 16]> scores_59_cast_fp16 = mul(x = var_4746_cast_fp16, y = var_4747_to_fp16)[name = string("scores_59_cast_fp16")];
            int32 var_4749 = const()[name = string("op_4749"), val = int32(-1)];
            tensor<fp16, [1, 12, 1, 16]> probabilities_87_cast_fp16 = softmax(axis = var_4749, x = scores_59_cast_fp16)[name = string("probabilities_87_cast_fp16")];
            bool attention_31_transpose_x_0 = const()[name = string("attention_31_transpose_x_0"), val = bool(false)];
            bool attention_31_transpose_y_0 = const()[name = string("attention_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> attention_31_cast_fp16 = matmul(transpose_x = attention_31_transpose_x_0, transpose_y = attention_31_transpose_y_0, x = probabilities_87_cast_fp16, y = value_125_cast_fp16)[name = string("attention_31_cast_fp16")];
            tensor<int32, [4]> var_4757 = const()[name = string("op_4757"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4762 = const()[name = string("op_4762"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 12, 64]> var_4758_cast_fp16 = transpose(perm = var_4757, x = attention_31_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 1, 768]> input_303_cast_fp16 = reshape(shape = var_4762, x = var_4758_cast_fp16)[name = string("input_303_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_75_cast_fp16 = linear(bias = block_attn_c_proj_bias_to_fp16, weight = block_attn_c_proj_weight_to_fp16, x = input_303_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_305_cast_fp16 = add(x = input_299_cast_fp16, y = linear_75_cast_fp16)[name = string("input_305_cast_fp16")];
            tensor<int32, [1]> input_307_axes_0 = const()[name = string("input_307_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_4770_to_fp16 = const()[name = string("op_4770_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_307_cast_fp16 = layer_norm(axes = input_307_axes_0, beta = block_ln_2_bias_to_fp16, epsilon = var_4770_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_305_cast_fp16)[name = string("input_307_cast_fp16")];
            tensor<fp16, [1, 1, 3072]> linear_76_cast_fp16 = linear(bias = block_mlp_fc_in_bias_to_fp16, weight = block_mlp_fc_in_weight_to_fp16, x = input_307_cast_fp16)[name = string("linear_76_cast_fp16")];
            string input_311_mode_0 = const()[name = string("input_311_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1, 3072]> input_311_cast_fp16 = gelu(mode = input_311_mode_0, x = linear_76_cast_fp16)[name = string("input_311_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_77_cast_fp16 = linear(bias = block_mlp_fc_out_bias_to_fp16, weight = block_mlp_fc_out_weight_to_fp16, x = input_311_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_315_cast_fp16 = add(x = input_305_cast_fp16, y = linear_77_cast_fp16)[name = string("input_315_cast_fp16")];
            tensor<int32, [1]> var_4809_axes_0 = const()[name = string("op_4809_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_4804_to_fp16 = const()[name = string("op_4804_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> var_4809_cast_fp16 = layer_norm(axes = var_4809_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_4804_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_315_cast_fp16)[name = string("op_4809_cast_fp16")];
            tensor<int32, [3]> var_4817_begin_0 = const()[name = string("op_4817_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_4817_end_0 = const()[name = string("op_4817_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_4817_end_mask_0 = const()[name = string("op_4817_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_4817_squeeze_mask_0 = const()[name = string("op_4817_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_4817_cast_fp16 = slice_by_index(begin = var_4817_begin_0, end = var_4817_end_0, end_mask = var_4817_end_mask_0, squeeze_mask = var_4817_squeeze_mask_0, x = var_4809_cast_fp16)[name = string("op_4817_cast_fp16")];
            tensor<fp16, [1024, 768]> audio_embeddings_14_weight_to_fp16 = const()[name = string("audio_embeddings_14_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36211968)))];
            tensor<fp16, [1, 1024]> linear_78_cast_fp16 = linear(bias = linear_8_bias_0_to_fp16, weight = audio_embeddings_14_weight_to_fp16, x = var_4817_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [3]> var_4832_begin_0 = const()[name = string("op_4832_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
            tensor<int32, [3]> var_4832_end_0 = const()[name = string("op_4832_end_0"), val = tensor<int32, [3]>([1, 15, 1024])];
            tensor<bool, [3]> var_4832_end_mask_0 = const()[name = string("op_4832_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_4832_squeeze_mask_0 = const()[name = string("op_4832_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<int32, [1, 1024]> var_4832 = slice_by_index(begin = var_4832_begin_0, end = var_4832_end_0, end_mask = var_4832_end_mask_0, squeeze_mask = var_4832_squeeze_mask_0, x = repetition_seen)[name = string("op_4832")];
            tensor<int32, [2]> random_u_29_begin_0 = const()[name = string("random_u_29_begin_0"), val = tensor<int32, [2]>([0, 14])];
            tensor<int32, [2]> random_u_29_end_0 = const()[name = string("random_u_29_end_0"), val = tensor<int32, [2]>([1, 15])];
            tensor<bool, [2]> random_u_29_end_mask_0 = const()[name = string("random_u_29_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<bool, [2]> random_u_29_squeeze_mask_0 = const()[name = string("random_u_29_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [1]> random_u_29_cast_fp16 = slice_by_index(begin = random_u_29_begin_0, end = random_u_29_end_0, end_mask = random_u_29_end_mask_0, squeeze_mask = random_u_29_squeeze_mask_0, x = audio_random_to_fp16)[name = string("random_u_29_cast_fp16")];
            fp16 var_4846_promoted_to_fp16 = const()[name = string("op_4846_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [1, 1024]> var_4847_cast_fp16 = less(x = linear_78_cast_fp16, y = var_4846_promoted_to_fp16)[name = string("op_4847_cast_fp16")];
            fp16 var_4848_to_fp16 = const()[name = string("op_4848_to_fp16"), val = fp16(0x1.334p+0)];
            tensor<fp16, [1, 1024]> var_4849_cast_fp16 = mul(x = linear_78_cast_fp16, y = var_4848_to_fp16)[name = string("op_4849_cast_fp16")];
            fp16 _inversed_4851_y_0_to_fp16 = const()[name = string("_inversed_4851_y_0_to_fp16"), val = fp16(0x1.aacp-1)];
            tensor<fp16, [1, 1024]> _inversed_4851_cast_fp16 = mul(x = linear_78_cast_fp16, y = _inversed_4851_y_0_to_fp16)[name = string("_inversed_4851_cast_fp16")];
            tensor<fp16, [1, 1024]> penalized_29_cast_fp16 = select(a = var_4849_cast_fp16, b = _inversed_4851_cast_fp16, cond = var_4847_cast_fp16)[name = string("penalized_29_cast_fp16")];
            string var_4857_dtype_0 = const()[name = string("op_4857_dtype_0"), val = string("bool")];
            tensor<bool, [1, 1024]> var_4857 = cast(dtype = var_4857_dtype_0, x = var_4832)[name = string("cast_284")];
            tensor<fp16, [1, 1024]> var_4858_cast_fp16 = select(a = penalized_29_cast_fp16, b = linear_78_cast_fp16, cond = var_4857)[name = string("op_4858_cast_fp16")];
            fp16 _inversed_scores_61_y_0_to_fp16 = const()[name = string("_inversed_scores_61_y_0_to_fp16"), val = fp16(0x1.4p+0)];
            tensor<fp16, [1, 1024]> _inversed_scores_61_cast_fp16 = mul(x = var_4858_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = string("_inversed_scores_61_cast_fp16")];
            int32 var_4861 = const()[name = string("op_4861"), val = int32(25)];
            int32 top_scores_57_axis_0 = const()[name = string("top_scores_57_axis_0"), val = int32(1)];
            bool top_scores_57_ascending_0 = const()[name = string("top_scores_57_ascending_0"), val = bool(false)];
            bool top_scores_57_sort_0 = const()[name = string("top_scores_57_sort_0"), val = bool(true)];
            bool top_scores_57_return_indices_0 = const()[name = string("top_scores_57_return_indices_0"), val = bool(true)];
            string top_scores_57_cast_fp16_cast_uint16_output_indices_dtype_0 = const()[name = string("top_scores_57_cast_fp16_cast_uint16_output_indices_dtype_0"), val = string("uint16")];
            tensor<fp16, [1, 25]> top_scores_57_cast_fp16_cast_uint16_0, tensor<uint16, [1, 25]> top_scores_57_cast_fp16_cast_uint16_1 = topk(ascending = top_scores_57_ascending_0, axis = top_scores_57_axis_0, k = var_4861, output_indices_dtype = top_scores_57_cast_fp16_cast_uint16_output_indices_dtype_0, return_indices = top_scores_57_return_indices_0, sort = top_scores_57_sort_0, x = _inversed_scores_61_cast_fp16)[name = string("top_scores_57_cast_fp16_cast_uint16")];
            int32 var_4867 = const()[name = string("op_4867"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_89_cast_fp16 = softmax(axis = var_4867, x = top_scores_57_cast_fp16_cast_uint16_0)[name = string("probabilities_89_cast_fp16")];
            int32 var_4870 = const()[name = string("op_4870"), val = int32(1)];
            bool cumulative_57_exclusive_0 = const()[name = string("cumulative_57_exclusive_0"), val = bool(false)];
            bool cumulative_57_reverse_0 = const()[name = string("cumulative_57_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_57_cast_fp16 = cumsum(axis = var_4870, exclusive = cumulative_57_exclusive_0, reverse = cumulative_57_reverse_0, x = probabilities_89_cast_fp16)[name = string("cumulative_57_cast_fp16")];
            tensor<fp16, [1, 25]> var_4874_cast_fp16 = sub(x = cumulative_57_cast_fp16, y = probabilities_89_cast_fp16)[name = string("op_4874_cast_fp16")];
            fp16 var_4875_to_fp16 = const()[name = string("op_4875_to_fp16"), val = fp16(0x1.e68p-1)];
            tensor<bool, [1, 25]> keep_29_cast_fp16 = less(x = var_4874_cast_fp16, y = var_4875_to_fp16)[name = string("keep_29_cast_fp16")];
            fp16 var_4883_value_0_to_fp16 = const()[name = string("op_4883_value_0_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 25]> var_4883_cast_fp16 = fill_like(ref_tensor = top_scores_57_cast_fp16_cast_uint16_0, value = var_4883_value_0_to_fp16)[name = string("op_4883_cast_fp16")];
            tensor<fp16, [1, 25]> top_scores_59_cast_fp16 = select(a = top_scores_57_cast_fp16_cast_uint16_0, b = var_4883_cast_fp16, cond = keep_29_cast_fp16)[name = string("top_scores_59_cast_fp16")];
            int32 var_4885 = const()[name = string("op_4885"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_91_cast_fp16 = softmax(axis = var_4885, x = top_scores_59_cast_fp16)[name = string("probabilities_91_cast_fp16")];
            int32 var_4888 = const()[name = string("op_4888"), val = int32(1)];
            bool cumulative_59_exclusive_0 = const()[name = string("cumulative_59_exclusive_0"), val = bool(false)];
            bool cumulative_59_reverse_0 = const()[name = string("cumulative_59_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_59_cast_fp16 = cumsum(axis = var_4888, exclusive = cumulative_59_exclusive_0, reverse = cumulative_59_reverse_0, x = probabilities_91_cast_fp16)[name = string("cumulative_59_cast_fp16")];
            tensor<int32, [2]> var_4893 = const()[name = string("op_4893"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_4894_cast_fp16 = reshape(shape = var_4893, x = random_u_29_cast_fp16)[name = string("op_4894_cast_fp16")];
            tensor<bool, [1, 25]> var_4895_cast_fp16 = less(x = cumulative_59_cast_fp16, y = var_4894_cast_fp16)[name = string("op_4895_cast_fp16")];
            string var_4900_dtype_0 = const()[name = string("op_4900_dtype_0"), val = string("int32")];
            tensor<int32, [1]> selected_57_axes_0 = const()[name = string("selected_57_axes_0"), val = tensor<int32, [1]>([1])];
            bool selected_57_keep_dims_0 = const()[name = string("selected_57_keep_dims_0"), val = bool(false)];
            tensor<int32, [1, 25]> var_4900 = cast(dtype = var_4900_dtype_0, x = var_4895_cast_fp16)[name = string("cast_283")];
            tensor<int32, [1]> selected_57 = reduce_sum(axes = selected_57_axes_0, keep_dims = selected_57_keep_dims_0, x = var_4900)[name = string("selected_57")];
            string selected_57_promoted_to_fp16_dtype_0 = const()[name = string("selected_57_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_78_to_fp16 = const()[name = string("const_78_to_fp16"), val = fp16(-inf)];
            fp16 var_4907_promoted_to_fp16 = const()[name = string("op_4907_promoted_to_fp16"), val = fp16(0x1.8p+4)];
            tensor<fp16, [1]> selected_57_to_fp16 = cast(dtype = selected_57_promoted_to_fp16_dtype_0, x = selected_57)[name = string("cast_282")];
            tensor<fp16, [1]> clip_14_cast_fp16 = clip(alpha = const_78_to_fp16, beta = var_4907_promoted_to_fp16, x = selected_57_to_fp16)[name = string("clip_14_cast_fp16")];
            string selected_59_dtype_0 = const()[name = string("selected_59_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_4916 = const()[name = string("op_4916"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> clip_14_cast_fp16_to_int32 = cast(dtype = selected_59_dtype_0, x = clip_14_cast_fp16)[name = string("cast_281")];
            tensor<int32, [1, 1]> var_4917 = reshape(shape = var_4916, x = clip_14_cast_fp16_to_int32)[name = string("op_4917")];
            int32 var_4918 = const()[name = string("op_4918"), val = int32(1)];
            bool var_4920_validate_indices_0 = const()[name = string("op_4920_validate_indices_0"), val = bool(false)];
            string var_4917_to_uint16_dtype_0 = const()[name = string("op_4917_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 1]> var_4917_to_uint16 = cast(dtype = var_4917_to_uint16_dtype_0, x = var_4917)[name = string("cast_280")];
            tensor<uint16, [1, 1]> var_4920_cast_uint16 = gather_along_axis(axis = var_4918, indices = var_4917_to_uint16, validate_indices = var_4920_validate_indices_0, x = top_scores_57_cast_fp16_cast_uint16_1)[name = string("op_4920_cast_uint16")];
            string var_4920_cast_uint16_to_int32_dtype_0 = const()[name = string("op_4920_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [1]> var_4922 = const()[name = string("op_4922"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_4920_cast_uint16_to_int32 = cast(dtype = var_4920_cast_uint16_to_int32_dtype_0, x = var_4920_cast_uint16)[name = string("cast_279")];
            tensor<int32, [1]> token_29 = reshape(shape = var_4922, x = var_4920_cast_uint16_to_int32)[name = string("token_29")];
            int32 var_4932_axis_0 = const()[name = string("op_4932_axis_0"), val = int32(0)];
            int32 var_4932_batch_dims_0 = const()[name = string("op_4932_batch_dims_0"), val = int32(0)];
            bool var_4932_validate_indices_0 = const()[name = string("op_4932_validate_indices_0"), val = bool(false)];
            string token_29_to_uint16_dtype_0 = const()[name = string("token_29_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> token_29_to_uint16 = cast(dtype = token_29_to_uint16_dtype_0, x = token_29)[name = string("cast_278")];
            tensor<fp16, [1, 768]> var_4932_cast_fp16_cast_uint16 = gather(axis = var_4932_axis_0, batch_dims = var_4932_batch_dims_0, indices = token_29_to_uint16, validate_indices = var_4932_validate_indices_0, x = audio_embeddings_14_weight_to_fp16)[name = string("op_4932_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 768]> next_global_embed_cast_fp16 = add(x = next_global_embed_29_cast_fp16, y = var_4932_cast_fp16_cast_uint16)[name = string("next_global_embed_cast_fp16")];
            tensor<int32, [3]> var_4942 = const()[name = string("op_4942"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 768]> input_319_cast_fp16 = reshape(shape = var_4942, x = var_4932_cast_fp16_cast_uint16)[name = string("input_319_cast_fp16")];
            tensor<int32, [1]> input_321_axes_0 = const()[name = string("input_321_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_4945_to_fp16 = const()[name = string("op_4945_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_321_cast_fp16 = layer_norm(axes = input_321_axes_0, beta = block_ln_1_bias_to_fp16, epsilon = var_4945_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_319_cast_fp16)[name = string("input_321_cast_fp16")];
            tensor<fp16, [1, 1, 2304]> linear_79_cast_fp16 = linear(bias = block_attn_c_attn_bias_to_fp16, weight = block_attn_c_attn_weight_to_fp16, x = input_321_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [3]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [3]>([768, 768, 768])];
            int32 var_4956_axis_0 = const()[name = string("op_4956_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 768]> var_4956_cast_fp16_0, tensor<fp16, [1, 1, 768]> var_4956_cast_fp16_1, tensor<fp16, [1, 1, 768]> var_4956_cast_fp16_2 = split(axis = var_4956_axis_0, split_sizes = tile_16, x = linear_79_cast_fp16)[name = string("op_4956_cast_fp16")];
            tensor<int32, [4]> var_4964 = const()[name = string("op_4964"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> query_131_cast_fp16 = reshape(shape = var_4964, x = var_4956_cast_fp16_0)[name = string("query_131_cast_fp16")];
            tensor<int32, [4]> var_4970 = const()[name = string("op_4970"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> key_161_cast_fp16 = reshape(shape = var_4970, x = var_4956_cast_fp16_1)[name = string("key_161_cast_fp16")];
            tensor<int32, [4]> var_4976 = const()[name = string("op_4976"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> value_129_cast_fp16 = reshape(shape = var_4976, x = var_4956_cast_fp16_2)[name = string("value_129_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_cos_16_to_fp16 = const()[name = string("rope_cos_16_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37784896)))];
            tensor<fp16, [1, 1, 12, 64]> var_4978_cast_fp16 = mul(x = query_131_cast_fp16, y = rope_cos_16_to_fp16)[name = string("op_4978_cast_fp16")];
            tensor<int32, [4]> var_4983_begin_0 = const()[name = string("op_4983_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_4983_end_0 = const()[name = string("op_4983_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_4983_end_mask_0 = const()[name = string("op_4983_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4983_stride_0 = const()[name = string("op_4983_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_4983_cast_fp16 = slice_by_index(begin = var_4983_begin_0, end = var_4983_end_0, end_mask = var_4983_end_mask_0, stride = var_4983_stride_0, x = query_131_cast_fp16)[name = string("op_4983_cast_fp16")];
            fp16 const_81_promoted_to_fp16 = const()[name = string("const_81_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_4984_cast_fp16 = mul(x = var_4983_cast_fp16, y = const_81_promoted_to_fp16)[name = string("op_4984_cast_fp16")];
            tensor<int32, [4]> var_4989_begin_0 = const()[name = string("op_4989_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4989_end_0 = const()[name = string("op_4989_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_4989_end_mask_0 = const()[name = string("op_4989_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4989_stride_0 = const()[name = string("op_4989_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_4989_cast_fp16 = slice_by_index(begin = var_4989_begin_0, end = var_4989_end_0, end_mask = var_4989_end_mask_0, stride = var_4989_stride_0, x = query_131_cast_fp16)[name = string("op_4989_cast_fp16")];
            int32 var_4992_axis_0 = const()[name = string("op_4992_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_4992_cast_fp16 = stack(axis = var_4992_axis_0, values = (var_4984_cast_fp16, var_4989_cast_fp16))[name = string("op_4992_cast_fp16")];
            tensor<int32, [4]> shape_49 = const()[name = string("shape_49"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_4993_cast_fp16 = reshape(shape = shape_49, x = var_4992_cast_fp16)[name = string("op_4993_cast_fp16")];
            tensor<fp16, [1, 1, 1, 64]> rope_sin_16_to_fp16 = const()[name = string("rope_sin_16_to_fp16"), val = tensor<fp16, [1, 1, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37785088)))];
            tensor<fp16, [1, 1, 12, 64]> var_4994_cast_fp16 = mul(x = var_4993_cast_fp16, y = rope_sin_16_to_fp16)[name = string("op_4994_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> query_133_cast_fp16 = add(x = var_4978_cast_fp16, y = var_4994_cast_fp16)[name = string("query_133_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_4997_cast_fp16 = mul(x = key_161_cast_fp16, y = rope_cos_16_to_fp16)[name = string("op_4997_cast_fp16")];
            tensor<int32, [4]> var_5002_begin_0 = const()[name = string("op_5002_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])];
            tensor<int32, [4]> var_5002_end_0 = const()[name = string("op_5002_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_5002_end_mask_0 = const()[name = string("op_5002_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5002_stride_0 = const()[name = string("op_5002_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_5002_cast_fp16 = slice_by_index(begin = var_5002_begin_0, end = var_5002_end_0, end_mask = var_5002_end_mask_0, stride = var_5002_stride_0, x = key_161_cast_fp16)[name = string("op_5002_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 12, 32]> var_5003_cast_fp16 = mul(x = var_5002_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_5003_cast_fp16")];
            tensor<int32, [4]> var_5008_begin_0 = const()[name = string("op_5008_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5008_end_0 = const()[name = string("op_5008_end_0"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<bool, [4]> var_5008_end_mask_0 = const()[name = string("op_5008_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5008_stride_0 = const()[name = string("op_5008_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 12, 32]> var_5008_cast_fp16 = slice_by_index(begin = var_5008_begin_0, end = var_5008_end_0, end_mask = var_5008_end_mask_0, stride = var_5008_stride_0, x = key_161_cast_fp16)[name = string("op_5008_cast_fp16")];
            int32 var_5011_axis_0 = const()[name = string("op_5011_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 1, 12, 32, 2]> var_5011_cast_fp16 = stack(axis = var_5011_axis_0, values = (var_5003_cast_fp16, var_5008_cast_fp16))[name = string("op_5011_cast_fp16")];
            tensor<int32, [4]> shape_50 = const()[name = string("shape_50"), val = tensor<int32, [4]>([1, 1, 12, 64])];
            tensor<fp16, [1, 1, 12, 64]> var_5012_cast_fp16 = reshape(shape = shape_50, x = var_5011_cast_fp16)[name = string("op_5012_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> var_5013_cast_fp16 = mul(x = var_5012_cast_fp16, y = rope_sin_16_to_fp16)[name = string("op_5013_cast_fp16")];
            tensor<fp16, [1, 1, 12, 64]> key_163_cast_fp16 = add(x = var_4997_cast_fp16, y = var_5013_cast_fp16)[name = string("key_163_cast_fp16")];
            tensor<int32, [4]> var_5020 = const()[name = string("op_5020"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_5026 = const()[name = string("op_5026"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_5032 = const()[name = string("op_5032"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            int32 var_5035 = const()[name = string("op_5035"), val = int32(2)];
            bool key_interleave_0 = const()[name = string("key_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> key_165_cast_fp16 = transpose(perm = var_5026, x = key_163_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 12, 17, 64]> key_cast_fp16 = concat(axis = var_5035, interleave = key_interleave_0, values = (key_157_cast_fp16, key_165_cast_fp16))[name = string("key_cast_fp16")];
            int32 var_5038 = const()[name = string("op_5038"), val = int32(2)];
            bool value_interleave_0 = const()[name = string("value_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> value_131_cast_fp16 = transpose(perm = var_5032, x = value_129_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 12, 17, 64]> value_cast_fp16 = concat(axis = var_5038, interleave = value_interleave_0, values = (value_125_cast_fp16, value_131_cast_fp16))[name = string("value_cast_fp16")];
            bool var_5043_transpose_x_1 = const()[name = string("op_5043_transpose_x_1"), val = bool(false)];
            bool var_5043_transpose_y_1 = const()[name = string("op_5043_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 12, 1, 64]> query_cast_fp16 = transpose(perm = var_5020, x = query_133_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 12, 1, 17]> var_5043_cast_fp16 = matmul(transpose_x = var_5043_transpose_x_1, transpose_y = var_5043_transpose_y_1, x = query_cast_fp16, y = key_cast_fp16)[name = string("op_5043_cast_fp16")];
            fp16 var_5044_to_fp16 = const()[name = string("op_5044_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 12, 1, 17]> scores_63_cast_fp16 = mul(x = var_5043_cast_fp16, y = var_5044_to_fp16)[name = string("scores_63_cast_fp16")];
            int32 var_5046 = const()[name = string("op_5046"), val = int32(-1)];
            tensor<fp16, [1, 12, 1, 17]> probabilities_93_cast_fp16 = softmax(axis = var_5046, x = scores_63_cast_fp16)[name = string("probabilities_93_cast_fp16")];
            bool attention_transpose_x_0 = const()[name = string("attention_transpose_x_0"), val = bool(false)];
            bool attention_transpose_y_0 = const()[name = string("attention_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1, 64]> attention_cast_fp16 = matmul(transpose_x = attention_transpose_x_0, transpose_y = attention_transpose_y_0, x = probabilities_93_cast_fp16, y = value_cast_fp16)[name = string("attention_cast_fp16")];
            tensor<int32, [4]> var_5054 = const()[name = string("op_5054"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5059 = const()[name = string("op_5059"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<fp16, [1, 1, 12, 64]> var_5055_cast_fp16 = transpose(perm = var_5054, x = attention_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 1, 768]> input_323_cast_fp16 = reshape(shape = var_5059, x = var_5055_cast_fp16)[name = string("input_323_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_80_cast_fp16 = linear(bias = block_attn_c_proj_bias_to_fp16, weight = block_attn_c_proj_weight_to_fp16, x = input_323_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_325_cast_fp16 = add(x = input_319_cast_fp16, y = linear_80_cast_fp16)[name = string("input_325_cast_fp16")];
            tensor<int32, [1]> input_327_axes_0 = const()[name = string("input_327_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_5067_to_fp16 = const()[name = string("op_5067_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> input_327_cast_fp16 = layer_norm(axes = input_327_axes_0, beta = block_ln_2_bias_to_fp16, epsilon = var_5067_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_325_cast_fp16)[name = string("input_327_cast_fp16")];
            tensor<fp16, [1, 1, 3072]> linear_81_cast_fp16 = linear(bias = block_mlp_fc_in_bias_to_fp16, weight = block_mlp_fc_in_weight_to_fp16, x = input_327_cast_fp16)[name = string("linear_81_cast_fp16")];
            string input_331_mode_0 = const()[name = string("input_331_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1, 3072]> input_331_cast_fp16 = gelu(mode = input_331_mode_0, x = linear_81_cast_fp16)[name = string("input_331_cast_fp16")];
            tensor<fp16, [1, 1, 768]> linear_82_cast_fp16 = linear(bias = block_mlp_fc_out_bias_to_fp16, weight = block_mlp_fc_out_weight_to_fp16, x = input_331_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, 1, 768]> input_335_cast_fp16 = add(x = input_325_cast_fp16, y = linear_82_cast_fp16)[name = string("input_335_cast_fp16")];
            tensor<int32, [1]> var_5106_axes_0 = const()[name = string("op_5106_axes_0"), val = tensor<int32, [1]>([-1])];
            fp16 var_5101_to_fp16 = const()[name = string("op_5101_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1, 768]> var_5106_cast_fp16 = layer_norm(axes = var_5106_axes_0, beta = final_norm_bias_to_fp16, epsilon = var_5101_to_fp16, gamma = block_ln_1_weight_to_fp16, x = input_335_cast_fp16)[name = string("op_5106_cast_fp16")];
            tensor<int32, [3]> var_5114_begin_0 = const()[name = string("op_5114_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_5114_end_0 = const()[name = string("op_5114_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_5114_end_mask_0 = const()[name = string("op_5114_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_5114_squeeze_mask_0 = const()[name = string("op_5114_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_5114_cast_fp16 = slice_by_index(begin = var_5114_begin_0, end = var_5114_end_0, end_mask = var_5114_end_mask_0, squeeze_mask = var_5114_squeeze_mask_0, x = var_5106_cast_fp16)[name = string("op_5114_cast_fp16")];
            tensor<fp16, [1024, 768]> audio_embeddings_15_weight_to_fp16 = const()[name = string("audio_embeddings_15_weight_to_fp16"), val = tensor<fp16, [1024, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37785280)))];
            tensor<fp16, [1, 1024]> linear_83_cast_fp16 = linear(bias = linear_8_bias_0_to_fp16, weight = audio_embeddings_15_weight_to_fp16, x = var_5114_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<int32, [3]> var_5129_begin_0 = const()[name = string("op_5129_begin_0"), val = tensor<int32, [3]>([0, 15, 0])];
            tensor<int32, [3]> var_5129_end_0 = const()[name = string("op_5129_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<bool, [3]> var_5129_end_mask_0 = const()[name = string("op_5129_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_5129_squeeze_mask_0 = const()[name = string("op_5129_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<int32, [1, 1024]> var_5129 = slice_by_index(begin = var_5129_begin_0, end = var_5129_end_0, end_mask = var_5129_end_mask_0, squeeze_mask = var_5129_squeeze_mask_0, x = repetition_seen)[name = string("op_5129")];
            tensor<int32, [2]> random_u_begin_0 = const()[name = string("random_u_begin_0"), val = tensor<int32, [2]>([0, 15])];
            tensor<int32, [2]> random_u_end_0 = const()[name = string("random_u_end_0"), val = tensor<int32, [2]>([1, 16])];
            tensor<bool, [2]> random_u_end_mask_0 = const()[name = string("random_u_end_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<bool, [2]> random_u_squeeze_mask_0 = const()[name = string("random_u_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<fp16, [1]> random_u_cast_fp16 = slice_by_index(begin = random_u_begin_0, end = random_u_end_0, end_mask = random_u_end_mask_0, squeeze_mask = random_u_squeeze_mask_0, x = audio_random_to_fp16)[name = string("random_u_cast_fp16")];
            fp16 var_5143_promoted_to_fp16 = const()[name = string("op_5143_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [1, 1024]> var_5144_cast_fp16 = less(x = linear_83_cast_fp16, y = var_5143_promoted_to_fp16)[name = string("op_5144_cast_fp16")];
            fp16 var_5145_to_fp16 = const()[name = string("op_5145_to_fp16"), val = fp16(0x1.334p+0)];
            tensor<fp16, [1, 1024]> var_5146_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_5145_to_fp16)[name = string("op_5146_cast_fp16")];
            fp16 _inversed_5148_y_0_to_fp16 = const()[name = string("_inversed_5148_y_0_to_fp16"), val = fp16(0x1.aacp-1)];
            tensor<fp16, [1, 1024]> _inversed_5148_cast_fp16 = mul(x = linear_83_cast_fp16, y = _inversed_5148_y_0_to_fp16)[name = string("_inversed_5148_cast_fp16")];
            tensor<fp16, [1, 1024]> penalized_cast_fp16 = select(a = var_5146_cast_fp16, b = _inversed_5148_cast_fp16, cond = var_5144_cast_fp16)[name = string("penalized_cast_fp16")];
            string var_5154_dtype_0 = const()[name = string("op_5154_dtype_0"), val = string("bool")];
            tensor<bool, [1, 1024]> var_5154 = cast(dtype = var_5154_dtype_0, x = var_5129)[name = string("cast_277")];
            tensor<fp16, [1, 1024]> var_5155_cast_fp16 = select(a = penalized_cast_fp16, b = linear_83_cast_fp16, cond = var_5154)[name = string("op_5155_cast_fp16")];
            fp16 _inversed_scores_y_0_to_fp16 = const()[name = string("_inversed_scores_y_0_to_fp16"), val = fp16(0x1.4p+0)];
            tensor<fp16, [1, 1024]> _inversed_scores_cast_fp16 = mul(x = var_5155_cast_fp16, y = _inversed_scores_y_0_to_fp16)[name = string("_inversed_scores_cast_fp16")];
            int32 var_5158 = const()[name = string("op_5158"), val = int32(25)];
            int32 top_scores_61_axis_0 = const()[name = string("top_scores_61_axis_0"), val = int32(1)];
            bool top_scores_61_ascending_0 = const()[name = string("top_scores_61_ascending_0"), val = bool(false)];
            bool top_scores_61_sort_0 = const()[name = string("top_scores_61_sort_0"), val = bool(true)];
            bool top_scores_61_return_indices_0 = const()[name = string("top_scores_61_return_indices_0"), val = bool(true)];
            string top_scores_61_cast_fp16_cast_uint16_output_indices_dtype_0 = const()[name = string("top_scores_61_cast_fp16_cast_uint16_output_indices_dtype_0"), val = string("uint16")];
            tensor<fp16, [1, 25]> top_scores_61_cast_fp16_cast_uint16_0, tensor<uint16, [1, 25]> top_scores_61_cast_fp16_cast_uint16_1 = topk(ascending = top_scores_61_ascending_0, axis = top_scores_61_axis_0, k = var_5158, output_indices_dtype = top_scores_61_cast_fp16_cast_uint16_output_indices_dtype_0, return_indices = top_scores_61_return_indices_0, sort = top_scores_61_sort_0, x = _inversed_scores_cast_fp16)[name = string("top_scores_61_cast_fp16_cast_uint16")];
            int32 var_5164 = const()[name = string("op_5164"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_95_cast_fp16 = softmax(axis = var_5164, x = top_scores_61_cast_fp16_cast_uint16_0)[name = string("probabilities_95_cast_fp16")];
            int32 var_5167 = const()[name = string("op_5167"), val = int32(1)];
            bool cumulative_61_exclusive_0 = const()[name = string("cumulative_61_exclusive_0"), val = bool(false)];
            bool cumulative_61_reverse_0 = const()[name = string("cumulative_61_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_61_cast_fp16 = cumsum(axis = var_5167, exclusive = cumulative_61_exclusive_0, reverse = cumulative_61_reverse_0, x = probabilities_95_cast_fp16)[name = string("cumulative_61_cast_fp16")];
            tensor<fp16, [1, 25]> var_5171_cast_fp16 = sub(x = cumulative_61_cast_fp16, y = probabilities_95_cast_fp16)[name = string("op_5171_cast_fp16")];
            fp16 var_5172_to_fp16 = const()[name = string("op_5172_to_fp16"), val = fp16(0x1.e68p-1)];
            tensor<bool, [1, 25]> keep_cast_fp16 = less(x = var_5171_cast_fp16, y = var_5172_to_fp16)[name = string("keep_cast_fp16")];
            fp16 var_5180_value_0_to_fp16 = const()[name = string("op_5180_value_0_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 25]> var_5180_cast_fp16 = fill_like(ref_tensor = top_scores_61_cast_fp16_cast_uint16_0, value = var_5180_value_0_to_fp16)[name = string("op_5180_cast_fp16")];
            tensor<fp16, [1, 25]> top_scores_cast_fp16 = select(a = top_scores_61_cast_fp16_cast_uint16_0, b = var_5180_cast_fp16, cond = keep_cast_fp16)[name = string("top_scores_cast_fp16")];
            int32 var_5182 = const()[name = string("op_5182"), val = int32(1)];
            tensor<fp16, [1, 25]> probabilities_cast_fp16 = softmax(axis = var_5182, x = top_scores_cast_fp16)[name = string("probabilities_cast_fp16")];
            int32 var_5185 = const()[name = string("op_5185"), val = int32(1)];
            bool cumulative_exclusive_0 = const()[name = string("cumulative_exclusive_0"), val = bool(false)];
            bool cumulative_reverse_0 = const()[name = string("cumulative_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 25]> cumulative_cast_fp16 = cumsum(axis = var_5185, exclusive = cumulative_exclusive_0, reverse = cumulative_reverse_0, x = probabilities_cast_fp16)[name = string("cumulative_cast_fp16")];
            tensor<int32, [2]> var_5190 = const()[name = string("op_5190"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_5191_cast_fp16 = reshape(shape = var_5190, x = random_u_cast_fp16)[name = string("op_5191_cast_fp16")];
            tensor<bool, [1, 25]> var_5192_cast_fp16 = less(x = cumulative_cast_fp16, y = var_5191_cast_fp16)[name = string("op_5192_cast_fp16")];
            string var_5197_dtype_0 = const()[name = string("op_5197_dtype_0"), val = string("int32")];
            tensor<int32, [1]> selected_61_axes_0 = const()[name = string("selected_61_axes_0"), val = tensor<int32, [1]>([1])];
            bool selected_61_keep_dims_0 = const()[name = string("selected_61_keep_dims_0"), val = bool(false)];
            tensor<int32, [1, 25]> var_5197 = cast(dtype = var_5197_dtype_0, x = var_5192_cast_fp16)[name = string("cast_276")];
            tensor<int32, [1]> selected_61 = reduce_sum(axes = selected_61_axes_0, keep_dims = selected_61_keep_dims_0, x = var_5197)[name = string("selected_61")];
            string selected_61_promoted_to_fp16_dtype_0 = const()[name = string("selected_61_promoted_to_fp16_dtype_0"), val = string("fp16")];
            fp16 const_83_to_fp16 = const()[name = string("const_83_to_fp16"), val = fp16(-inf)];
            fp16 var_5204_promoted_to_fp16 = const()[name = string("op_5204_promoted_to_fp16"), val = fp16(0x1.8p+4)];
            tensor<fp16, [1]> selected_61_to_fp16 = cast(dtype = selected_61_promoted_to_fp16_dtype_0, x = selected_61)[name = string("cast_275")];
            tensor<fp16, [1]> clip_15_cast_fp16 = clip(alpha = const_83_to_fp16, beta = var_5204_promoted_to_fp16, x = selected_61_to_fp16)[name = string("clip_15_cast_fp16")];
            string selected_dtype_0 = const()[name = string("selected_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_5213 = const()[name = string("op_5213"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> clip_15_cast_fp16_to_int32 = cast(dtype = selected_dtype_0, x = clip_15_cast_fp16)[name = string("cast_274")];
            tensor<int32, [1, 1]> var_5214 = reshape(shape = var_5213, x = clip_15_cast_fp16_to_int32)[name = string("op_5214")];
            int32 var_5215 = const()[name = string("op_5215"), val = int32(1)];
            bool var_5217_validate_indices_0 = const()[name = string("op_5217_validate_indices_0"), val = bool(false)];
            string var_5214_to_uint16_dtype_0 = const()[name = string("op_5214_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 1]> var_5214_to_uint16 = cast(dtype = var_5214_to_uint16_dtype_0, x = var_5214)[name = string("cast_273")];
            tensor<uint16, [1, 1]> var_5217_cast_uint16 = gather_along_axis(axis = var_5215, indices = var_5214_to_uint16, validate_indices = var_5217_validate_indices_0, x = top_scores_61_cast_fp16_cast_uint16_1)[name = string("op_5217_cast_uint16")];
            string var_5217_cast_uint16_to_int32_dtype_0 = const()[name = string("op_5217_cast_uint16_to_int32_dtype_0"), val = string("int32")];
            tensor<int32, [1]> var_5219 = const()[name = string("op_5219"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_5217_cast_uint16_to_int32 = cast(dtype = var_5217_cast_uint16_to_int32_dtype_0, x = var_5217_cast_uint16)[name = string("cast_272")];
            tensor<int32, [1]> token = reshape(shape = var_5219, x = var_5217_cast_uint16_to_int32)[name = string("token")];
            int32 var_5229_axis_0 = const()[name = string("op_5229_axis_0"), val = int32(0)];
            int32 var_5229_batch_dims_0 = const()[name = string("op_5229_batch_dims_0"), val = int32(0)];
            bool var_5229_validate_indices_0 = const()[name = string("op_5229_validate_indices_0"), val = bool(false)];
            string token_to_uint16_dtype_0 = const()[name = string("token_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> token_to_uint16 = cast(dtype = token_to_uint16_dtype_0, x = token)[name = string("cast_271")];
            tensor<fp16, [1, 768]> var_5229_cast_fp16_cast_uint16 = gather(axis = var_5229_axis_0, batch_dims = var_5229_batch_dims_0, indices = token_to_uint16, validate_indices = var_5229_validate_indices_0, x = audio_embeddings_15_weight_to_fp16)[name = string("op_5229_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 768]> next_global_embedding = add(x = next_global_embed_cast_fp16, y = var_5229_cast_fp16_cast_uint16)[name = string("op_5235_cast_fp16")];
            int32 var_5238_axis_0 = const()[name = string("op_5238_axis_0"), val = int32(1)];
            tensor<int32, [1, 16]> frame_tokens = stack(axis = var_5238_axis_0, values = (token_1, token_3, token_5, token_7, token_9, token_11, token_13, token_15, token_17, token_19, token_21, token_23, token_25, token_27, token_29, token))[name = string("op_5238")];
            int32 var_5256_axis_0 = const()[name = string("op_5256_axis_0"), val = int32(0)];
            int32 var_5256_batch_dims_0 = const()[name = string("op_5256_batch_dims_0"), val = int32(0)];
            bool var_5256_validate_indices_0 = const()[name = string("op_5256_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 8]> codec_quantizer_quantizers_0_codebook_weight_to_fp16 = const()[name = string("codec_quantizer_quantizers_0_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39358208)))];
            tensor<fp16, [1, 8]> var_5256_cast_fp16_cast_uint16 = gather(axis = var_5256_axis_0, batch_dims = var_5256_batch_dims_0, indices = token_1_to_uint16, validate_indices = var_5256_validate_indices_0, x = codec_quantizer_quantizers_0_codebook_weight_to_fp16)[name = string("op_5256_cast_fp16_cast_uint16")];
            tensor<int32, [3]> var_5260 = const()[name = string("op_5260"), val = tensor<int32, [3]>([1, 8, 1])];
            tensor<fp16, [1, 8, 1]> input_341_cast_fp16 = reshape(shape = var_5260, x = var_5256_cast_fp16_cast_uint16)[name = string("input_341_cast_fp16")];
            string var_5276_pad_type_0 = const()[name = string("op_5276_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_5276_strides_0 = const()[name = string("op_5276_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5276_pad_0 = const()[name = string("op_5276_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5276_dilations_0 = const()[name = string("op_5276_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_5276_groups_0 = const()[name = string("op_5276_groups_0"), val = int32(1)];
            tensor<fp16, [512, 8, 1]> weight_41_to_fp16 = const()[name = string("weight_41_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39374656)))];
            tensor<fp16, [512]> codec_quantizer_quantizers_0_out_proj_bias_to_fp16 = const()[name = string("codec_quantizer_quantizers_0_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39382912)))];
            tensor<fp16, [1, 512, 1]> var_5276_cast_fp16 = conv(bias = codec_quantizer_quantizers_0_out_proj_bias_to_fp16, dilations = var_5276_dilations_0, groups = var_5276_groups_0, pad = var_5276_pad_0, pad_type = var_5276_pad_type_0, strides = var_5276_strides_0, weight = weight_41_to_fp16, x = input_341_cast_fp16)[name = string("op_5276_cast_fp16")];
            int32 var_5287_axis_0 = const()[name = string("op_5287_axis_0"), val = int32(0)];
            int32 var_5287_batch_dims_0 = const()[name = string("op_5287_batch_dims_0"), val = int32(0)];
            bool var_5287_validate_indices_0 = const()[name = string("op_5287_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 8]> codec_quantizer_quantizers_1_codebook_weight_to_fp16 = const()[name = string("codec_quantizer_quantizers_1_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39384000)))];
            tensor<fp16, [1, 8]> var_5287_cast_fp16_cast_uint16 = gather(axis = var_5287_axis_0, batch_dims = var_5287_batch_dims_0, indices = token_3_to_uint16, validate_indices = var_5287_validate_indices_0, x = codec_quantizer_quantizers_1_codebook_weight_to_fp16)[name = string("op_5287_cast_fp16_cast_uint16")];
            tensor<int32, [3]> var_5291 = const()[name = string("op_5291"), val = tensor<int32, [3]>([1, 8, 1])];
            tensor<fp16, [1, 8, 1]> input_345_cast_fp16 = reshape(shape = var_5291, x = var_5287_cast_fp16_cast_uint16)[name = string("input_345_cast_fp16")];
            string var_5307_pad_type_0 = const()[name = string("op_5307_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_5307_strides_0 = const()[name = string("op_5307_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5307_pad_0 = const()[name = string("op_5307_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5307_dilations_0 = const()[name = string("op_5307_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_5307_groups_0 = const()[name = string("op_5307_groups_0"), val = int32(1)];
            tensor<fp16, [512, 8, 1]> weight_45_to_fp16 = const()[name = string("weight_45_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39400448)))];
            tensor<fp16, [512]> codec_quantizer_quantizers_1_out_proj_bias_to_fp16 = const()[name = string("codec_quantizer_quantizers_1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39408704)))];
            tensor<fp16, [1, 512, 1]> var_5307_cast_fp16 = conv(bias = codec_quantizer_quantizers_1_out_proj_bias_to_fp16, dilations = var_5307_dilations_0, groups = var_5307_groups_0, pad = var_5307_pad_0, pad_type = var_5307_pad_type_0, strides = var_5307_strides_0, weight = weight_45_to_fp16, x = input_345_cast_fp16)[name = string("op_5307_cast_fp16")];
            tensor<fp16, [1, 512, 1]> codec_hidden_5_cast_fp16 = add(x = var_5276_cast_fp16, y = var_5307_cast_fp16)[name = string("codec_hidden_5_cast_fp16")];
            int32 var_5318_axis_0 = const()[name = string("op_5318_axis_0"), val = int32(0)];
            int32 var_5318_batch_dims_0 = const()[name = string("op_5318_batch_dims_0"), val = int32(0)];
            bool var_5318_validate_indices_0 = const()[name = string("op_5318_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 8]> codec_quantizer_quantizers_2_codebook_weight_to_fp16 = const()[name = string("codec_quantizer_quantizers_2_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39409792)))];
            tensor<fp16, [1, 8]> var_5318_cast_fp16_cast_uint16 = gather(axis = var_5318_axis_0, batch_dims = var_5318_batch_dims_0, indices = token_5_to_uint16, validate_indices = var_5318_validate_indices_0, x = codec_quantizer_quantizers_2_codebook_weight_to_fp16)[name = string("op_5318_cast_fp16_cast_uint16")];
            tensor<int32, [3]> var_5322 = const()[name = string("op_5322"), val = tensor<int32, [3]>([1, 8, 1])];
            tensor<fp16, [1, 8, 1]> input_349_cast_fp16 = reshape(shape = var_5322, x = var_5318_cast_fp16_cast_uint16)[name = string("input_349_cast_fp16")];
            string var_5338_pad_type_0 = const()[name = string("op_5338_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_5338_strides_0 = const()[name = string("op_5338_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5338_pad_0 = const()[name = string("op_5338_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5338_dilations_0 = const()[name = string("op_5338_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_5338_groups_0 = const()[name = string("op_5338_groups_0"), val = int32(1)];
            tensor<fp16, [512, 8, 1]> weight_49_to_fp16 = const()[name = string("weight_49_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39426240)))];
            tensor<fp16, [512]> codec_quantizer_quantizers_2_out_proj_bias_to_fp16 = const()[name = string("codec_quantizer_quantizers_2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39434496)))];
            tensor<fp16, [1, 512, 1]> var_5338_cast_fp16 = conv(bias = codec_quantizer_quantizers_2_out_proj_bias_to_fp16, dilations = var_5338_dilations_0, groups = var_5338_groups_0, pad = var_5338_pad_0, pad_type = var_5338_pad_type_0, strides = var_5338_strides_0, weight = weight_49_to_fp16, x = input_349_cast_fp16)[name = string("op_5338_cast_fp16")];
            tensor<fp16, [1, 512, 1]> codec_hidden_7_cast_fp16 = add(x = codec_hidden_5_cast_fp16, y = var_5338_cast_fp16)[name = string("codec_hidden_7_cast_fp16")];
            int32 var_5349_axis_0 = const()[name = string("op_5349_axis_0"), val = int32(0)];
            int32 var_5349_batch_dims_0 = const()[name = string("op_5349_batch_dims_0"), val = int32(0)];
            bool var_5349_validate_indices_0 = const()[name = string("op_5349_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 8]> codec_quantizer_quantizers_3_codebook_weight_to_fp16 = const()[name = string("codec_quantizer_quantizers_3_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39435584)))];
            tensor<fp16, [1, 8]> var_5349_cast_fp16_cast_uint16 = gather(axis = var_5349_axis_0, batch_dims = var_5349_batch_dims_0, indices = token_7_to_uint16, validate_indices = var_5349_validate_indices_0, x = codec_quantizer_quantizers_3_codebook_weight_to_fp16)[name = string("op_5349_cast_fp16_cast_uint16")];
            tensor<int32, [3]> var_5353 = const()[name = string("op_5353"), val = tensor<int32, [3]>([1, 8, 1])];
            tensor<fp16, [1, 8, 1]> input_353_cast_fp16 = reshape(shape = var_5353, x = var_5349_cast_fp16_cast_uint16)[name = string("input_353_cast_fp16")];
            string var_5369_pad_type_0 = const()[name = string("op_5369_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_5369_strides_0 = const()[name = string("op_5369_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5369_pad_0 = const()[name = string("op_5369_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5369_dilations_0 = const()[name = string("op_5369_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_5369_groups_0 = const()[name = string("op_5369_groups_0"), val = int32(1)];
            tensor<fp16, [512, 8, 1]> weight_53_to_fp16 = const()[name = string("weight_53_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39452032)))];
            tensor<fp16, [512]> codec_quantizer_quantizers_3_out_proj_bias_to_fp16 = const()[name = string("codec_quantizer_quantizers_3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39460288)))];
            tensor<fp16, [1, 512, 1]> var_5369_cast_fp16 = conv(bias = codec_quantizer_quantizers_3_out_proj_bias_to_fp16, dilations = var_5369_dilations_0, groups = var_5369_groups_0, pad = var_5369_pad_0, pad_type = var_5369_pad_type_0, strides = var_5369_strides_0, weight = weight_53_to_fp16, x = input_353_cast_fp16)[name = string("op_5369_cast_fp16")];
            tensor<fp16, [1, 512, 1]> codec_hidden_9_cast_fp16 = add(x = codec_hidden_7_cast_fp16, y = var_5369_cast_fp16)[name = string("codec_hidden_9_cast_fp16")];
            int32 var_5380_axis_0 = const()[name = string("op_5380_axis_0"), val = int32(0)];
            int32 var_5380_batch_dims_0 = const()[name = string("op_5380_batch_dims_0"), val = int32(0)];
            bool var_5380_validate_indices_0 = const()[name = string("op_5380_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 8]> codec_quantizer_quantizers_4_codebook_weight_to_fp16 = const()[name = string("codec_quantizer_quantizers_4_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39461376)))];
            tensor<fp16, [1, 8]> var_5380_cast_fp16_cast_uint16 = gather(axis = var_5380_axis_0, batch_dims = var_5380_batch_dims_0, indices = token_9_to_uint16, validate_indices = var_5380_validate_indices_0, x = codec_quantizer_quantizers_4_codebook_weight_to_fp16)[name = string("op_5380_cast_fp16_cast_uint16")];
            tensor<int32, [3]> var_5384 = const()[name = string("op_5384"), val = tensor<int32, [3]>([1, 8, 1])];
            tensor<fp16, [1, 8, 1]> input_357_cast_fp16 = reshape(shape = var_5384, x = var_5380_cast_fp16_cast_uint16)[name = string("input_357_cast_fp16")];
            string var_5400_pad_type_0 = const()[name = string("op_5400_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_5400_strides_0 = const()[name = string("op_5400_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5400_pad_0 = const()[name = string("op_5400_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5400_dilations_0 = const()[name = string("op_5400_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_5400_groups_0 = const()[name = string("op_5400_groups_0"), val = int32(1)];
            tensor<fp16, [512, 8, 1]> weight_57_to_fp16 = const()[name = string("weight_57_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39477824)))];
            tensor<fp16, [512]> codec_quantizer_quantizers_4_out_proj_bias_to_fp16 = const()[name = string("codec_quantizer_quantizers_4_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39486080)))];
            tensor<fp16, [1, 512, 1]> var_5400_cast_fp16 = conv(bias = codec_quantizer_quantizers_4_out_proj_bias_to_fp16, dilations = var_5400_dilations_0, groups = var_5400_groups_0, pad = var_5400_pad_0, pad_type = var_5400_pad_type_0, strides = var_5400_strides_0, weight = weight_57_to_fp16, x = input_357_cast_fp16)[name = string("op_5400_cast_fp16")];
            tensor<fp16, [1, 512, 1]> codec_hidden_11_cast_fp16 = add(x = codec_hidden_9_cast_fp16, y = var_5400_cast_fp16)[name = string("codec_hidden_11_cast_fp16")];
            int32 var_5411_axis_0 = const()[name = string("op_5411_axis_0"), val = int32(0)];
            int32 var_5411_batch_dims_0 = const()[name = string("op_5411_batch_dims_0"), val = int32(0)];
            bool var_5411_validate_indices_0 = const()[name = string("op_5411_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 8]> codec_quantizer_quantizers_5_codebook_weight_to_fp16 = const()[name = string("codec_quantizer_quantizers_5_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39487168)))];
            tensor<fp16, [1, 8]> var_5411_cast_fp16_cast_uint16 = gather(axis = var_5411_axis_0, batch_dims = var_5411_batch_dims_0, indices = token_11_to_uint16, validate_indices = var_5411_validate_indices_0, x = codec_quantizer_quantizers_5_codebook_weight_to_fp16)[name = string("op_5411_cast_fp16_cast_uint16")];
            tensor<int32, [3]> var_5415 = const()[name = string("op_5415"), val = tensor<int32, [3]>([1, 8, 1])];
            tensor<fp16, [1, 8, 1]> input_361_cast_fp16 = reshape(shape = var_5415, x = var_5411_cast_fp16_cast_uint16)[name = string("input_361_cast_fp16")];
            string var_5431_pad_type_0 = const()[name = string("op_5431_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_5431_strides_0 = const()[name = string("op_5431_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5431_pad_0 = const()[name = string("op_5431_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5431_dilations_0 = const()[name = string("op_5431_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_5431_groups_0 = const()[name = string("op_5431_groups_0"), val = int32(1)];
            tensor<fp16, [512, 8, 1]> weight_61_to_fp16 = const()[name = string("weight_61_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39503616)))];
            tensor<fp16, [512]> codec_quantizer_quantizers_5_out_proj_bias_to_fp16 = const()[name = string("codec_quantizer_quantizers_5_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39511872)))];
            tensor<fp16, [1, 512, 1]> var_5431_cast_fp16 = conv(bias = codec_quantizer_quantizers_5_out_proj_bias_to_fp16, dilations = var_5431_dilations_0, groups = var_5431_groups_0, pad = var_5431_pad_0, pad_type = var_5431_pad_type_0, strides = var_5431_strides_0, weight = weight_61_to_fp16, x = input_361_cast_fp16)[name = string("op_5431_cast_fp16")];
            tensor<fp16, [1, 512, 1]> codec_hidden_13_cast_fp16 = add(x = codec_hidden_11_cast_fp16, y = var_5431_cast_fp16)[name = string("codec_hidden_13_cast_fp16")];
            int32 var_5442_axis_0 = const()[name = string("op_5442_axis_0"), val = int32(0)];
            int32 var_5442_batch_dims_0 = const()[name = string("op_5442_batch_dims_0"), val = int32(0)];
            bool var_5442_validate_indices_0 = const()[name = string("op_5442_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 8]> codec_quantizer_quantizers_6_codebook_weight_to_fp16 = const()[name = string("codec_quantizer_quantizers_6_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39512960)))];
            tensor<fp16, [1, 8]> var_5442_cast_fp16_cast_uint16 = gather(axis = var_5442_axis_0, batch_dims = var_5442_batch_dims_0, indices = token_13_to_uint16, validate_indices = var_5442_validate_indices_0, x = codec_quantizer_quantizers_6_codebook_weight_to_fp16)[name = string("op_5442_cast_fp16_cast_uint16")];
            tensor<int32, [3]> var_5446 = const()[name = string("op_5446"), val = tensor<int32, [3]>([1, 8, 1])];
            tensor<fp16, [1, 8, 1]> input_365_cast_fp16 = reshape(shape = var_5446, x = var_5442_cast_fp16_cast_uint16)[name = string("input_365_cast_fp16")];
            string var_5462_pad_type_0 = const()[name = string("op_5462_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_5462_strides_0 = const()[name = string("op_5462_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5462_pad_0 = const()[name = string("op_5462_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5462_dilations_0 = const()[name = string("op_5462_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_5462_groups_0 = const()[name = string("op_5462_groups_0"), val = int32(1)];
            tensor<fp16, [512, 8, 1]> weight_65_to_fp16 = const()[name = string("weight_65_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39529408)))];
            tensor<fp16, [512]> codec_quantizer_quantizers_6_out_proj_bias_to_fp16 = const()[name = string("codec_quantizer_quantizers_6_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39537664)))];
            tensor<fp16, [1, 512, 1]> var_5462_cast_fp16 = conv(bias = codec_quantizer_quantizers_6_out_proj_bias_to_fp16, dilations = var_5462_dilations_0, groups = var_5462_groups_0, pad = var_5462_pad_0, pad_type = var_5462_pad_type_0, strides = var_5462_strides_0, weight = weight_65_to_fp16, x = input_365_cast_fp16)[name = string("op_5462_cast_fp16")];
            tensor<fp16, [1, 512, 1]> codec_hidden_15_cast_fp16 = add(x = codec_hidden_13_cast_fp16, y = var_5462_cast_fp16)[name = string("codec_hidden_15_cast_fp16")];
            int32 var_5473_axis_0 = const()[name = string("op_5473_axis_0"), val = int32(0)];
            int32 var_5473_batch_dims_0 = const()[name = string("op_5473_batch_dims_0"), val = int32(0)];
            bool var_5473_validate_indices_0 = const()[name = string("op_5473_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 8]> codec_quantizer_quantizers_7_codebook_weight_to_fp16 = const()[name = string("codec_quantizer_quantizers_7_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39538752)))];
            tensor<fp16, [1, 8]> var_5473_cast_fp16_cast_uint16 = gather(axis = var_5473_axis_0, batch_dims = var_5473_batch_dims_0, indices = token_15_to_uint16, validate_indices = var_5473_validate_indices_0, x = codec_quantizer_quantizers_7_codebook_weight_to_fp16)[name = string("op_5473_cast_fp16_cast_uint16")];
            tensor<int32, [3]> var_5477 = const()[name = string("op_5477"), val = tensor<int32, [3]>([1, 8, 1])];
            tensor<fp16, [1, 8, 1]> input_369_cast_fp16 = reshape(shape = var_5477, x = var_5473_cast_fp16_cast_uint16)[name = string("input_369_cast_fp16")];
            string var_5493_pad_type_0 = const()[name = string("op_5493_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_5493_strides_0 = const()[name = string("op_5493_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5493_pad_0 = const()[name = string("op_5493_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5493_dilations_0 = const()[name = string("op_5493_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_5493_groups_0 = const()[name = string("op_5493_groups_0"), val = int32(1)];
            tensor<fp16, [512, 8, 1]> weight_69_to_fp16 = const()[name = string("weight_69_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39555200)))];
            tensor<fp16, [512]> codec_quantizer_quantizers_7_out_proj_bias_to_fp16 = const()[name = string("codec_quantizer_quantizers_7_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39563456)))];
            tensor<fp16, [1, 512, 1]> var_5493_cast_fp16 = conv(bias = codec_quantizer_quantizers_7_out_proj_bias_to_fp16, dilations = var_5493_dilations_0, groups = var_5493_groups_0, pad = var_5493_pad_0, pad_type = var_5493_pad_type_0, strides = var_5493_strides_0, weight = weight_69_to_fp16, x = input_369_cast_fp16)[name = string("op_5493_cast_fp16")];
            tensor<fp16, [1, 512, 1]> codec_hidden_17_cast_fp16 = add(x = codec_hidden_15_cast_fp16, y = var_5493_cast_fp16)[name = string("codec_hidden_17_cast_fp16")];
            int32 var_5504_axis_0 = const()[name = string("op_5504_axis_0"), val = int32(0)];
            int32 var_5504_batch_dims_0 = const()[name = string("op_5504_batch_dims_0"), val = int32(0)];
            bool var_5504_validate_indices_0 = const()[name = string("op_5504_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 8]> codec_quantizer_quantizers_8_codebook_weight_to_fp16 = const()[name = string("codec_quantizer_quantizers_8_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39564544)))];
            tensor<fp16, [1, 8]> var_5504_cast_fp16_cast_uint16 = gather(axis = var_5504_axis_0, batch_dims = var_5504_batch_dims_0, indices = token_17_to_uint16, validate_indices = var_5504_validate_indices_0, x = codec_quantizer_quantizers_8_codebook_weight_to_fp16)[name = string("op_5504_cast_fp16_cast_uint16")];
            tensor<int32, [3]> var_5508 = const()[name = string("op_5508"), val = tensor<int32, [3]>([1, 8, 1])];
            tensor<fp16, [1, 8, 1]> input_373_cast_fp16 = reshape(shape = var_5508, x = var_5504_cast_fp16_cast_uint16)[name = string("input_373_cast_fp16")];
            string var_5524_pad_type_0 = const()[name = string("op_5524_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_5524_strides_0 = const()[name = string("op_5524_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5524_pad_0 = const()[name = string("op_5524_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5524_dilations_0 = const()[name = string("op_5524_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_5524_groups_0 = const()[name = string("op_5524_groups_0"), val = int32(1)];
            tensor<fp16, [512, 8, 1]> weight_73_to_fp16 = const()[name = string("weight_73_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39580992)))];
            tensor<fp16, [512]> codec_quantizer_quantizers_8_out_proj_bias_to_fp16 = const()[name = string("codec_quantizer_quantizers_8_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39589248)))];
            tensor<fp16, [1, 512, 1]> var_5524_cast_fp16 = conv(bias = codec_quantizer_quantizers_8_out_proj_bias_to_fp16, dilations = var_5524_dilations_0, groups = var_5524_groups_0, pad = var_5524_pad_0, pad_type = var_5524_pad_type_0, strides = var_5524_strides_0, weight = weight_73_to_fp16, x = input_373_cast_fp16)[name = string("op_5524_cast_fp16")];
            tensor<fp16, [1, 512, 1]> codec_hidden_19_cast_fp16 = add(x = codec_hidden_17_cast_fp16, y = var_5524_cast_fp16)[name = string("codec_hidden_19_cast_fp16")];
            int32 var_5535_axis_0 = const()[name = string("op_5535_axis_0"), val = int32(0)];
            int32 var_5535_batch_dims_0 = const()[name = string("op_5535_batch_dims_0"), val = int32(0)];
            bool var_5535_validate_indices_0 = const()[name = string("op_5535_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 8]> codec_quantizer_quantizers_9_codebook_weight_to_fp16 = const()[name = string("codec_quantizer_quantizers_9_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39590336)))];
            tensor<fp16, [1, 8]> var_5535_cast_fp16_cast_uint16 = gather(axis = var_5535_axis_0, batch_dims = var_5535_batch_dims_0, indices = token_19_to_uint16, validate_indices = var_5535_validate_indices_0, x = codec_quantizer_quantizers_9_codebook_weight_to_fp16)[name = string("op_5535_cast_fp16_cast_uint16")];
            tensor<int32, [3]> var_5539 = const()[name = string("op_5539"), val = tensor<int32, [3]>([1, 8, 1])];
            tensor<fp16, [1, 8, 1]> input_377_cast_fp16 = reshape(shape = var_5539, x = var_5535_cast_fp16_cast_uint16)[name = string("input_377_cast_fp16")];
            string var_5555_pad_type_0 = const()[name = string("op_5555_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_5555_strides_0 = const()[name = string("op_5555_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5555_pad_0 = const()[name = string("op_5555_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5555_dilations_0 = const()[name = string("op_5555_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_5555_groups_0 = const()[name = string("op_5555_groups_0"), val = int32(1)];
            tensor<fp16, [512, 8, 1]> weight_77_to_fp16 = const()[name = string("weight_77_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39606784)))];
            tensor<fp16, [512]> codec_quantizer_quantizers_9_out_proj_bias_to_fp16 = const()[name = string("codec_quantizer_quantizers_9_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39615040)))];
            tensor<fp16, [1, 512, 1]> var_5555_cast_fp16 = conv(bias = codec_quantizer_quantizers_9_out_proj_bias_to_fp16, dilations = var_5555_dilations_0, groups = var_5555_groups_0, pad = var_5555_pad_0, pad_type = var_5555_pad_type_0, strides = var_5555_strides_0, weight = weight_77_to_fp16, x = input_377_cast_fp16)[name = string("op_5555_cast_fp16")];
            tensor<fp16, [1, 512, 1]> codec_hidden_21_cast_fp16 = add(x = codec_hidden_19_cast_fp16, y = var_5555_cast_fp16)[name = string("codec_hidden_21_cast_fp16")];
            int32 var_5566_axis_0 = const()[name = string("op_5566_axis_0"), val = int32(0)];
            int32 var_5566_batch_dims_0 = const()[name = string("op_5566_batch_dims_0"), val = int32(0)];
            bool var_5566_validate_indices_0 = const()[name = string("op_5566_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 8]> codec_quantizer_quantizers_10_codebook_weight_to_fp16 = const()[name = string("codec_quantizer_quantizers_10_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39616128)))];
            tensor<fp16, [1, 8]> var_5566_cast_fp16_cast_uint16 = gather(axis = var_5566_axis_0, batch_dims = var_5566_batch_dims_0, indices = token_21_to_uint16, validate_indices = var_5566_validate_indices_0, x = codec_quantizer_quantizers_10_codebook_weight_to_fp16)[name = string("op_5566_cast_fp16_cast_uint16")];
            tensor<int32, [3]> var_5570 = const()[name = string("op_5570"), val = tensor<int32, [3]>([1, 8, 1])];
            tensor<fp16, [1, 8, 1]> input_381_cast_fp16 = reshape(shape = var_5570, x = var_5566_cast_fp16_cast_uint16)[name = string("input_381_cast_fp16")];
            string var_5586_pad_type_0 = const()[name = string("op_5586_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_5586_strides_0 = const()[name = string("op_5586_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5586_pad_0 = const()[name = string("op_5586_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5586_dilations_0 = const()[name = string("op_5586_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_5586_groups_0 = const()[name = string("op_5586_groups_0"), val = int32(1)];
            tensor<fp16, [512, 8, 1]> weight_81_to_fp16 = const()[name = string("weight_81_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39632576)))];
            tensor<fp16, [512]> codec_quantizer_quantizers_10_out_proj_bias_to_fp16 = const()[name = string("codec_quantizer_quantizers_10_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39640832)))];
            tensor<fp16, [1, 512, 1]> var_5586_cast_fp16 = conv(bias = codec_quantizer_quantizers_10_out_proj_bias_to_fp16, dilations = var_5586_dilations_0, groups = var_5586_groups_0, pad = var_5586_pad_0, pad_type = var_5586_pad_type_0, strides = var_5586_strides_0, weight = weight_81_to_fp16, x = input_381_cast_fp16)[name = string("op_5586_cast_fp16")];
            tensor<fp16, [1, 512, 1]> codec_hidden_23_cast_fp16 = add(x = codec_hidden_21_cast_fp16, y = var_5586_cast_fp16)[name = string("codec_hidden_23_cast_fp16")];
            int32 var_5597_axis_0 = const()[name = string("op_5597_axis_0"), val = int32(0)];
            int32 var_5597_batch_dims_0 = const()[name = string("op_5597_batch_dims_0"), val = int32(0)];
            bool var_5597_validate_indices_0 = const()[name = string("op_5597_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 8]> codec_quantizer_quantizers_11_codebook_weight_to_fp16 = const()[name = string("codec_quantizer_quantizers_11_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39641920)))];
            tensor<fp16, [1, 8]> var_5597_cast_fp16_cast_uint16 = gather(axis = var_5597_axis_0, batch_dims = var_5597_batch_dims_0, indices = token_23_to_uint16, validate_indices = var_5597_validate_indices_0, x = codec_quantizer_quantizers_11_codebook_weight_to_fp16)[name = string("op_5597_cast_fp16_cast_uint16")];
            tensor<int32, [3]> var_5601 = const()[name = string("op_5601"), val = tensor<int32, [3]>([1, 8, 1])];
            tensor<fp16, [1, 8, 1]> input_385_cast_fp16 = reshape(shape = var_5601, x = var_5597_cast_fp16_cast_uint16)[name = string("input_385_cast_fp16")];
            string var_5617_pad_type_0 = const()[name = string("op_5617_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_5617_strides_0 = const()[name = string("op_5617_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5617_pad_0 = const()[name = string("op_5617_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5617_dilations_0 = const()[name = string("op_5617_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_5617_groups_0 = const()[name = string("op_5617_groups_0"), val = int32(1)];
            tensor<fp16, [512, 8, 1]> weight_85_to_fp16 = const()[name = string("weight_85_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39658368)))];
            tensor<fp16, [512]> codec_quantizer_quantizers_11_out_proj_bias_to_fp16 = const()[name = string("codec_quantizer_quantizers_11_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39666624)))];
            tensor<fp16, [1, 512, 1]> var_5617_cast_fp16 = conv(bias = codec_quantizer_quantizers_11_out_proj_bias_to_fp16, dilations = var_5617_dilations_0, groups = var_5617_groups_0, pad = var_5617_pad_0, pad_type = var_5617_pad_type_0, strides = var_5617_strides_0, weight = weight_85_to_fp16, x = input_385_cast_fp16)[name = string("op_5617_cast_fp16")];
            tensor<fp16, [1, 512, 1]> codec_hidden_25_cast_fp16 = add(x = codec_hidden_23_cast_fp16, y = var_5617_cast_fp16)[name = string("codec_hidden_25_cast_fp16")];
            int32 var_5628_axis_0 = const()[name = string("op_5628_axis_0"), val = int32(0)];
            int32 var_5628_batch_dims_0 = const()[name = string("op_5628_batch_dims_0"), val = int32(0)];
            bool var_5628_validate_indices_0 = const()[name = string("op_5628_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 8]> codec_quantizer_quantizers_12_codebook_weight_to_fp16 = const()[name = string("codec_quantizer_quantizers_12_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39667712)))];
            tensor<fp16, [1, 8]> var_5628_cast_fp16_cast_uint16 = gather(axis = var_5628_axis_0, batch_dims = var_5628_batch_dims_0, indices = token_25_to_uint16, validate_indices = var_5628_validate_indices_0, x = codec_quantizer_quantizers_12_codebook_weight_to_fp16)[name = string("op_5628_cast_fp16_cast_uint16")];
            tensor<int32, [3]> var_5632 = const()[name = string("op_5632"), val = tensor<int32, [3]>([1, 8, 1])];
            tensor<fp16, [1, 8, 1]> input_389_cast_fp16 = reshape(shape = var_5632, x = var_5628_cast_fp16_cast_uint16)[name = string("input_389_cast_fp16")];
            string var_5648_pad_type_0 = const()[name = string("op_5648_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_5648_strides_0 = const()[name = string("op_5648_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5648_pad_0 = const()[name = string("op_5648_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5648_dilations_0 = const()[name = string("op_5648_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_5648_groups_0 = const()[name = string("op_5648_groups_0"), val = int32(1)];
            tensor<fp16, [512, 8, 1]> weight_89_to_fp16 = const()[name = string("weight_89_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39684160)))];
            tensor<fp16, [512]> codec_quantizer_quantizers_12_out_proj_bias_to_fp16 = const()[name = string("codec_quantizer_quantizers_12_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39692416)))];
            tensor<fp16, [1, 512, 1]> var_5648_cast_fp16 = conv(bias = codec_quantizer_quantizers_12_out_proj_bias_to_fp16, dilations = var_5648_dilations_0, groups = var_5648_groups_0, pad = var_5648_pad_0, pad_type = var_5648_pad_type_0, strides = var_5648_strides_0, weight = weight_89_to_fp16, x = input_389_cast_fp16)[name = string("op_5648_cast_fp16")];
            tensor<fp16, [1, 512, 1]> codec_hidden_27_cast_fp16 = add(x = codec_hidden_25_cast_fp16, y = var_5648_cast_fp16)[name = string("codec_hidden_27_cast_fp16")];
            int32 var_5659_axis_0 = const()[name = string("op_5659_axis_0"), val = int32(0)];
            int32 var_5659_batch_dims_0 = const()[name = string("op_5659_batch_dims_0"), val = int32(0)];
            bool var_5659_validate_indices_0 = const()[name = string("op_5659_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 8]> codec_quantizer_quantizers_13_codebook_weight_to_fp16 = const()[name = string("codec_quantizer_quantizers_13_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39693504)))];
            tensor<fp16, [1, 8]> var_5659_cast_fp16_cast_uint16 = gather(axis = var_5659_axis_0, batch_dims = var_5659_batch_dims_0, indices = token_27_to_uint16, validate_indices = var_5659_validate_indices_0, x = codec_quantizer_quantizers_13_codebook_weight_to_fp16)[name = string("op_5659_cast_fp16_cast_uint16")];
            tensor<int32, [3]> var_5663 = const()[name = string("op_5663"), val = tensor<int32, [3]>([1, 8, 1])];
            tensor<fp16, [1, 8, 1]> input_393_cast_fp16 = reshape(shape = var_5663, x = var_5659_cast_fp16_cast_uint16)[name = string("input_393_cast_fp16")];
            string var_5679_pad_type_0 = const()[name = string("op_5679_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_5679_strides_0 = const()[name = string("op_5679_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5679_pad_0 = const()[name = string("op_5679_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5679_dilations_0 = const()[name = string("op_5679_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_5679_groups_0 = const()[name = string("op_5679_groups_0"), val = int32(1)];
            tensor<fp16, [512, 8, 1]> weight_93_to_fp16 = const()[name = string("weight_93_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39709952)))];
            tensor<fp16, [512]> codec_quantizer_quantizers_13_out_proj_bias_to_fp16 = const()[name = string("codec_quantizer_quantizers_13_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39718208)))];
            tensor<fp16, [1, 512, 1]> var_5679_cast_fp16 = conv(bias = codec_quantizer_quantizers_13_out_proj_bias_to_fp16, dilations = var_5679_dilations_0, groups = var_5679_groups_0, pad = var_5679_pad_0, pad_type = var_5679_pad_type_0, strides = var_5679_strides_0, weight = weight_93_to_fp16, x = input_393_cast_fp16)[name = string("op_5679_cast_fp16")];
            tensor<fp16, [1, 512, 1]> codec_hidden_29_cast_fp16 = add(x = codec_hidden_27_cast_fp16, y = var_5679_cast_fp16)[name = string("codec_hidden_29_cast_fp16")];
            int32 var_5690_axis_0 = const()[name = string("op_5690_axis_0"), val = int32(0)];
            int32 var_5690_batch_dims_0 = const()[name = string("op_5690_batch_dims_0"), val = int32(0)];
            bool var_5690_validate_indices_0 = const()[name = string("op_5690_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 8]> codec_quantizer_quantizers_14_codebook_weight_to_fp16 = const()[name = string("codec_quantizer_quantizers_14_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39719296)))];
            tensor<fp16, [1, 8]> var_5690_cast_fp16_cast_uint16 = gather(axis = var_5690_axis_0, batch_dims = var_5690_batch_dims_0, indices = token_29_to_uint16, validate_indices = var_5690_validate_indices_0, x = codec_quantizer_quantizers_14_codebook_weight_to_fp16)[name = string("op_5690_cast_fp16_cast_uint16")];
            tensor<int32, [3]> var_5694 = const()[name = string("op_5694"), val = tensor<int32, [3]>([1, 8, 1])];
            tensor<fp16, [1, 8, 1]> input_397_cast_fp16 = reshape(shape = var_5694, x = var_5690_cast_fp16_cast_uint16)[name = string("input_397_cast_fp16")];
            string var_5710_pad_type_0 = const()[name = string("op_5710_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_5710_strides_0 = const()[name = string("op_5710_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5710_pad_0 = const()[name = string("op_5710_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5710_dilations_0 = const()[name = string("op_5710_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_5710_groups_0 = const()[name = string("op_5710_groups_0"), val = int32(1)];
            tensor<fp16, [512, 8, 1]> weight_97_to_fp16 = const()[name = string("weight_97_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39735744)))];
            tensor<fp16, [512]> codec_quantizer_quantizers_14_out_proj_bias_to_fp16 = const()[name = string("codec_quantizer_quantizers_14_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39744000)))];
            tensor<fp16, [1, 512, 1]> var_5710_cast_fp16 = conv(bias = codec_quantizer_quantizers_14_out_proj_bias_to_fp16, dilations = var_5710_dilations_0, groups = var_5710_groups_0, pad = var_5710_pad_0, pad_type = var_5710_pad_type_0, strides = var_5710_strides_0, weight = weight_97_to_fp16, x = input_397_cast_fp16)[name = string("op_5710_cast_fp16")];
            tensor<fp16, [1, 512, 1]> codec_hidden_cast_fp16 = add(x = codec_hidden_29_cast_fp16, y = var_5710_cast_fp16)[name = string("codec_hidden_cast_fp16")];
            int32 var_5721_axis_0 = const()[name = string("op_5721_axis_0"), val = int32(0)];
            int32 var_5721_batch_dims_0 = const()[name = string("op_5721_batch_dims_0"), val = int32(0)];
            bool var_5721_validate_indices_0 = const()[name = string("op_5721_validate_indices_0"), val = bool(false)];
            tensor<fp16, [1024, 8]> codec_quantizer_quantizers_15_codebook_weight_to_fp16 = const()[name = string("codec_quantizer_quantizers_15_codebook_weight_to_fp16"), val = tensor<fp16, [1024, 8]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39745088)))];
            tensor<fp16, [1, 8]> var_5721_cast_fp16_cast_uint16 = gather(axis = var_5721_axis_0, batch_dims = var_5721_batch_dims_0, indices = token_to_uint16, validate_indices = var_5721_validate_indices_0, x = codec_quantizer_quantizers_15_codebook_weight_to_fp16)[name = string("op_5721_cast_fp16_cast_uint16")];
            tensor<int32, [3]> var_5725 = const()[name = string("op_5725"), val = tensor<int32, [3]>([1, 8, 1])];
            tensor<fp16, [1, 8, 1]> input_401_cast_fp16 = reshape(shape = var_5725, x = var_5721_cast_fp16_cast_uint16)[name = string("input_401_cast_fp16")];
            string var_5741_pad_type_0 = const()[name = string("op_5741_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_5741_strides_0 = const()[name = string("op_5741_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5741_pad_0 = const()[name = string("op_5741_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5741_dilations_0 = const()[name = string("op_5741_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_5741_groups_0 = const()[name = string("op_5741_groups_0"), val = int32(1)];
            tensor<fp16, [512, 8, 1]> weight_101_to_fp16 = const()[name = string("weight_101_to_fp16"), val = tensor<fp16, [512, 8, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39761536)))];
            tensor<fp16, [512]> codec_quantizer_quantizers_15_out_proj_bias_to_fp16 = const()[name = string("codec_quantizer_quantizers_15_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39769792)))];
            tensor<fp16, [1, 512, 1]> var_5741_cast_fp16 = conv(bias = codec_quantizer_quantizers_15_out_proj_bias_to_fp16, dilations = var_5741_dilations_0, groups = var_5741_groups_0, pad = var_5741_pad_0, pad_type = var_5741_pad_type_0, strides = var_5741_strides_0, weight = weight_101_to_fp16, x = input_401_cast_fp16)[name = string("op_5741_cast_fp16")];
            tensor<fp16, [1, 512, 1]> input_cast_fp16 = add(x = codec_hidden_cast_fp16, y = var_5741_cast_fp16)[name = string("input_cast_fp16")];
            string var_5758_pad_type_0 = const()[name = string("op_5758_pad_type_0"), val = string("valid")];
            tensor<int32, [1]> var_5758_strides_0 = const()[name = string("op_5758_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5758_pad_0 = const()[name = string("op_5758_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5758_dilations_0 = const()[name = string("op_5758_dilations_0"), val = tensor<int32, [1]>([1])];
            int32 var_5758_groups_0 = const()[name = string("op_5758_groups_0"), val = int32(1)];
            tensor<fp16, [768, 512, 1]> weight_103_to_fp16 = const()[name = string("weight_103_to_fp16"), val = tensor<fp16, [768, 512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39770880)))];
            tensor<fp16, [768]> codec_quantizer_output_proj_bias_to_fp16 = const()[name = string("codec_quantizer_output_proj_bias_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40557376)))];
            tensor<fp16, [1, 768, 1]> codec_latent = conv(bias = codec_quantizer_output_proj_bias_to_fp16, dilations = var_5758_dilations_0, groups = var_5758_groups_0, pad = var_5758_pad_0, pad_type = var_5758_pad_type_0, strides = var_5758_strides_0, weight = weight_103_to_fp16, x = input_cast_fp16)[name = string("op_5758_cast_fp16")];
            string var_5763_dtype_0 = const()[name = string("op_5763_dtype_0"), val = string("int32")];
            tensor<int32, [2]> var_5766 = const()[name = string("op_5766"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1]> var_5763 = cast(dtype = var_5763_dtype_0, x = should_continue_cast_fp16)[name = string("cast_270")];
            tensor<int32, [1, 1]> should_continue = reshape(shape = var_5766, x = var_5763)[name = string("op_5767")];
        } -> (should_continue, frame_tokens, next_global_embedding, codec_latent);
}